From aac71e563eeb83cbe925c605455122b13f8c0fb4 Mon Sep 17 00:00:00 2001 From: Georgios Pinitas Date: Thu, 25 Feb 2021 18:57:41 +0000 Subject: Update int8 A55 kernels Signed-off-by: Georgios Pinitas Change-Id: I15a443806a9773bd18b212fe23d9e4d126948be4 Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5191 Reviewed-by: Michele Di Giorgio Comments-Addressed: Arm Jenkins Tested-by: Arm Jenkins --- .../kernels/a64_hybrid_s8qa_dot_4x16/a55.cpp | 392 +-- .../kernels/a64_hybrid_s8qs_dot_6x16/a55.cpp | 2354 +++++++++--------- .../kernels/a64_hybrid_s8s32_dot_6x16/a55.cpp | 2628 ++++++++++---------- .../kernels/a64_hybrid_u8qa_dot_4x16/a55.cpp | 392 +-- .../kernels/a64_hybrid_u8u32_dot_6x16/a55.cpp | 2628 ++++++++++---------- 5 files changed, 4197 insertions(+), 4197 deletions(-) diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_dot_4x16/a55.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_dot_4x16/a55.cpp index fb85506ee8..11aa05a9b7 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_dot_4x16/a55.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_dot_4x16/a55.cpp @@ -130,11 +130,11 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( "mov v5.d[1], x24\n" "ldr d7, [x10, #0x30]\n" "ldr x19, [x10, #0x38]\n" - "mov v6.d[1], x23\n" ".inst 0x4f80e0b1 // sdot v17.4s, v5.16b, v0.4b[0]\n" + "mov v6.d[1], x23\n" "ldr d8, [x10, #0x40]\n" - "mov v7.d[1], x19\n" ".inst 0x4f80e0d2 // sdot v18.4s, v6.16b, v0.4b[0]\n" + "mov v7.d[1], x19\n" "ldr x23, [x10, #0x48]\n" ".inst 0x4f80e0f3 // sdot v19.4s, v7.16b, v0.4b[0]\n" "ldr d9, [x10, #0x50]\n" @@ -142,11 +142,11 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( "mov v8.d[1], x23\n" "ldr d10, [x10, #0x60]\n" "ldr x23, [x10, #0x68]\n" - "mov v9.d[1], x19\n" ".inst 0x4fa0e110 // sdot v16.4s, v8.16b, v0.4b[1]\n" + "mov v9.d[1], x19\n" "ldr d4, [x10, #0x70]\n" - "mov v10.d[1], x23\n" ".inst 0x4fa0e131 // sdot v17.4s, v9.16b, v0.4b[1]\n" + "mov v10.d[1], x23\n" "ldr x19, [x10, #0x78]\n" ".inst 0x4fa0e152 // sdot v18.4s, v10.16b, v0.4b[1]\n" "ldr d5, [x10, #0x80]\n" @@ -154,11 +154,11 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( "mov v4.d[1], x19\n" "ldr d6, [x10, #0x90]\n" "ldr x23, [x10, #0x98]\n" - "mov v5.d[1], x24\n" ".inst 0x4fa0e093 // sdot v19.4s, v4.16b, v0.4b[1]\n" + "mov v5.d[1], x24\n" "ldr d7, [x10, #0xa0]\n" - "mov v6.d[1], x23\n" ".inst 0x4f80e8b0 // sdot v16.4s, v5.16b, v0.4b[2]\n" + "mov v6.d[1], x23\n" "ldr x19, [x10, #0xa8]\n" ".inst 0x4f80e8d1 // sdot v17.4s, v6.16b, v0.4b[2]\n" "ldr d8, [x10, #0xb0]\n" @@ -166,11 +166,11 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( "mov v7.d[1], x19\n" "ldr d9, [x10, #0xc0]\n" "ldr x19, [x10, #0xc8]\n" - "mov v8.d[1], x23\n" ".inst 0x4f80e8f2 // sdot v18.4s, v7.16b, v0.4b[2]\n" + "mov v8.d[1], x23\n" "ldr d10, [x10, #0xd0]\n" - "mov v9.d[1], x19\n" ".inst 0x4f80e913 // sdot v19.4s, v8.16b, v0.4b[2]\n" + "mov v9.d[1], x19\n" "ldr x23, [x10, #0xd8]\n" ".inst 0x4fa0e930 // sdot v16.4s, v9.16b, v0.4b[3]\n" "ldr d4, [x10, #0xe0]\n" @@ -179,10 +179,10 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( "ldr d5, [x10, #0xf0]\n" "ldr x24, [x10, #0xf8]\n" "add x10, x10, #0x100\n" - "mov v4.d[1], x19\n" ".inst 0x4fa0e951 // sdot v17.4s, v10.16b, v0.4b[3]\n" - "mov v5.d[1], x24\n" + "mov v4.d[1], x19\n" ".inst 0x4fa0e892 // sdot v18.4s, v4.16b, v0.4b[3]\n" + "mov v5.d[1], x24\n" ".inst 0x4fa0e8b3 // sdot v19.4s, v5.16b, v0.4b[3]\n" "tbnz %x[flags], #31, 8f\n" ".inst 0x4e8f940b // sdot v11.4s, v0.16b, v15.16b\n" @@ -208,24 +208,24 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( "ldr q10, [x10, #0x60]\n" ".inst 0x4fa0e110 // sdot v16.4s, v8.16b, v0.4b[1]\n" "ldr q4, [x10, #0x70]\n" - "ldr q5, [x10, #0x80]\n" ".inst 0x4fa0e131 // sdot v17.4s, v9.16b, v0.4b[1]\n" - "ldr q6, [x10, #0x90]\n" + "ldr q5, [x10, #0x80]\n" ".inst 0x4fa0e152 // sdot v18.4s, v10.16b, v0.4b[1]\n" - "ldr q7, [x10, #0xa0]\n" + "ldr q6, [x10, #0x90]\n" ".inst 0x4fa0e093 // sdot v19.4s, v4.16b, v0.4b[1]\n" - "ldr q8, [x10, #0xb0]\n" + "ldr q7, [x10, #0xa0]\n" ".inst 0x4f80e8b0 // sdot v16.4s, v5.16b, v0.4b[2]\n" - "ldr q9, [x10, #0xc0]\n" - "ldr q10, [x10, #0xd0]\n" + "ldr q8, [x10, #0xb0]\n" ".inst 0x4f80e8d1 // sdot v17.4s, v6.16b, v0.4b[2]\n" - "ldr q4, [x10, #0xe0]\n" + "ldr q9, [x10, #0xc0]\n" ".inst 0x4f80e8f2 // sdot v18.4s, v7.16b, v0.4b[2]\n" - "ldr q5, [x10, #0xf0]\n" + "ldr q10, [x10, #0xd0]\n" ".inst 0x4f80e913 // sdot v19.4s, v8.16b, v0.4b[2]\n" - "add x10, x10, #0x100\n" + "ldr q4, [x10, #0xe0]\n" ".inst 0x4fa0e930 // sdot v16.4s, v9.16b, v0.4b[3]\n" + "ldr q5, [x10, #0xf0]\n" ".inst 0x4fa0e951 // sdot v17.4s, v10.16b, v0.4b[3]\n" + "add x10, x10, #0x100\n" ".inst 0x4fa0e892 // sdot v18.4s, v4.16b, v0.4b[3]\n" ".inst 0x4fa0e8b3 // sdot v19.4s, v5.16b, v0.4b[3]\n" "tbnz %x[flags], #31, 10f\n" @@ -451,83 +451,83 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( "blt 39f\n" "37:" // Height 2: Multiply loop: Main loop head ".inst 0x4f80e090 // sdot v16.4s, v4.16b, v0.4b[0]\n" - ".inst 0x4f81e094 // sdot v20.4s, v4.16b, v1.4b[0]\n" "ldr d5, [x10, #0x10]\n" - "add x25, x25, #0x10\n" + ".inst 0x4f81e094 // sdot v20.4s, v4.16b, v1.4b[0]\n" "ldr x24, [x10, #0x18]\n" - "add x22, x22, #0x10\n" "ldr d6, [x10, #0x20]\n" + "add x25, x25, #0x10\n" "ldr x23, [x10, #0x28]\n" + "add x22, x22, #0x10\n" "mov v5.d[1], x24\n" "ldr d7, [x10, #0x30]\n" "ldr x19, [x10, #0x38]\n" - "mov v6.d[1], x23\n" ".inst 0x4f80e0b1 // sdot v17.4s, v5.16b, v0.4b[0]\n" + "mov v6.d[1], x23\n" ".inst 0x4f81e0b5 // sdot v21.4s, v5.16b, v1.4b[0]\n" "ldr d8, [x10, #0x40]\n" - "mov v7.d[1], x19\n" ".inst 0x4f80e0d2 // sdot v18.4s, v6.16b, v0.4b[0]\n" + "mov v7.d[1], x19\n" ".inst 0x4f81e0d6 // sdot v22.4s, v6.16b, v1.4b[0]\n" "ldr x23, [x10, #0x48]\n" ".inst 0x4f80e0f3 // sdot v19.4s, v7.16b, v0.4b[0]\n" - ".inst 0x4f81e0f7 // sdot v23.4s, v7.16b, v1.4b[0]\n" "ldr d9, [x10, #0x50]\n" - "mov v8.d[1], x23\n" + ".inst 0x4f81e0f7 // sdot v23.4s, v7.16b, v1.4b[0]\n" "ldr x19, [x10, #0x58]\n" + "mov v8.d[1], x23\n" "ldr d10, [x10, #0x60]\n" + "ldr x23, [x10, #0x68]\n" ".inst 0x4fa0e110 // sdot v16.4s, v8.16b, v0.4b[1]\n" - ".inst 0x4fa1e114 // sdot v20.4s, v8.16b, v1.4b[1]\n" "mov v9.d[1], x19\n" - "ldr x23, [x10, #0x68]\n" + ".inst 0x4fa1e114 // sdot v20.4s, v8.16b, v1.4b[1]\n" "ldr d4, [x10, #0x70]\n" ".inst 0x4fa0e131 // sdot v17.4s, v9.16b, v0.4b[1]\n" - ".inst 0x4fa1e135 // sdot v21.4s, v9.16b, v1.4b[1]\n" "mov v10.d[1], x23\n" + ".inst 0x4fa1e135 // sdot v21.4s, v9.16b, v1.4b[1]\n" "ldr x19, [x10, #0x78]\n" - "ldr d5, [x10, #0x80]\n" ".inst 0x4fa0e152 // sdot v18.4s, v10.16b, v0.4b[1]\n" + "ldr d5, [x10, #0x80]\n" ".inst 0x4fa1e156 // sdot v22.4s, v10.16b, v1.4b[1]\n" - "mov v4.d[1], x19\n" "ldr x24, [x10, #0x88]\n" + "mov v4.d[1], x19\n" "ldr d6, [x10, #0x90]\n" + "ldr x23, [x10, #0x98]\n" ".inst 0x4fa0e093 // sdot v19.4s, v4.16b, v0.4b[1]\n" - ".inst 0x4fa1e097 // sdot v23.4s, v4.16b, v1.4b[1]\n" "mov v5.d[1], x24\n" - "ldr x23, [x10, #0x98]\n" + ".inst 0x4fa1e097 // sdot v23.4s, v4.16b, v1.4b[1]\n" "ldr d7, [x10, #0xa0]\n" ".inst 0x4f80e8b0 // sdot v16.4s, v5.16b, v0.4b[2]\n" - ".inst 0x4f81e8b4 // sdot v20.4s, v5.16b, v1.4b[2]\n" "mov v6.d[1], x23\n" + ".inst 0x4f81e8b4 // sdot v20.4s, v5.16b, v1.4b[2]\n" "ldr x19, [x10, #0xa8]\n" - "ldr d8, [x10, #0xb0]\n" ".inst 0x4f80e8d1 // sdot v17.4s, v6.16b, v0.4b[2]\n" + "ldr d8, [x10, #0xb0]\n" ".inst 0x4f81e8d5 // sdot v21.4s, v6.16b, v1.4b[2]\n" - "mov v7.d[1], x19\n" "ldr x23, [x10, #0xb8]\n" + "mov v7.d[1], x19\n" "ldr d9, [x10, #0xc0]\n" + "ldr x19, [x10, #0xc8]\n" ".inst 0x4f80e8f2 // sdot v18.4s, v7.16b, v0.4b[2]\n" - ".inst 0x4f81e8f6 // sdot v22.4s, v7.16b, v1.4b[2]\n" "mov v8.d[1], x23\n" - "ldr x19, [x10, #0xc8]\n" + ".inst 0x4f81e8f6 // sdot v22.4s, v7.16b, v1.4b[2]\n" "ldr d10, [x10, #0xd0]\n" ".inst 0x4f80e913 // sdot v19.4s, v8.16b, v0.4b[2]\n" - ".inst 0x4f81e917 // sdot v23.4s, v8.16b, v1.4b[2]\n" "mov v9.d[1], x19\n" + ".inst 0x4f81e917 // sdot v23.4s, v8.16b, v1.4b[2]\n" "ldr x23, [x10, #0xd8]\n" - "ldr d4, [x10, #0xe0]\n" ".inst 0x4fa0e930 // sdot v16.4s, v9.16b, v0.4b[3]\n" + "ldr d4, [x10, #0xe0]\n" ".inst 0x4fa1e934 // sdot v20.4s, v9.16b, v1.4b[3]\n" - "mov v10.d[1], x23\n" "ldr x19, [x10, #0xe8]\n" + "mov v10.d[1], x23\n" "ldr d5, [x10, #0xf0]\n" - ".inst 0x4fa0e951 // sdot v17.4s, v10.16b, v0.4b[3]\n" - ".inst 0x4fa1e955 // sdot v21.4s, v10.16b, v1.4b[3]\n" - "mov v4.d[1], x19\n" "ldr x24, [x10, #0xf8]\n" "add x10, x10, #0x100\n" + ".inst 0x4fa0e951 // sdot v17.4s, v10.16b, v0.4b[3]\n" + "mov v4.d[1], x19\n" + ".inst 0x4fa1e955 // sdot v21.4s, v10.16b, v1.4b[3]\n" ".inst 0x4fa0e892 // sdot v18.4s, v4.16b, v0.4b[3]\n" - ".inst 0x4fa1e896 // sdot v22.4s, v4.16b, v1.4b[3]\n" "mov v5.d[1], x24\n" + ".inst 0x4fa1e896 // sdot v22.4s, v4.16b, v1.4b[3]\n" ".inst 0x4fa0e8b3 // sdot v19.4s, v5.16b, v0.4b[3]\n" ".inst 0x4fa1e8b7 // sdot v23.4s, v5.16b, v1.4b[3]\n" "tbnz %x[flags], #31, 38f\n" @@ -544,50 +544,50 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( "bge 37b\n" "39:" // Height 2: Multiply loop: Single iteration only ".inst 0x4f80e090 // sdot v16.4s, v4.16b, v0.4b[0]\n" - ".inst 0x4f81e094 // sdot v20.4s, v4.16b, v1.4b[0]\n" "ldr q5, [x10, #0x10]\n" - "sub x26, x26, #0x10\n" + ".inst 0x4f81e094 // sdot v20.4s, v4.16b, v1.4b[0]\n" "ldr q6, [x10, #0x20]\n" - "add x25, x25, #0x10\n" "ldr q7, [x10, #0x30]\n" - "add x22, x22, #0x10\n" + "sub x26, x26, #0x10\n" ".inst 0x4f80e0b1 // sdot v17.4s, v5.16b, v0.4b[0]\n" + "ldr q8, [x10, #0x40]\n" ".inst 0x4f81e0b5 // sdot v21.4s, v5.16b, v1.4b[0]\n" + "ldr q9, [x10, #0x50]\n" ".inst 0x4f80e0d2 // sdot v18.4s, v6.16b, v0.4b[0]\n" + "ldr q10, [x10, #0x60]\n" ".inst 0x4f81e0d6 // sdot v22.4s, v6.16b, v1.4b[0]\n" + "ldr q4, [x10, #0x70]\n" ".inst 0x4f80e0f3 // sdot v19.4s, v7.16b, v0.4b[0]\n" + "ldr q5, [x10, #0x80]\n" ".inst 0x4f81e0f7 // sdot v23.4s, v7.16b, v1.4b[0]\n" - "ldr q8, [x10, #0x40]\n" - "ldr q9, [x10, #0x50]\n" - "ldr q10, [x10, #0x60]\n" + "ldr q6, [x10, #0x90]\n" ".inst 0x4fa0e110 // sdot v16.4s, v8.16b, v0.4b[1]\n" + "ldr q7, [x10, #0xa0]\n" ".inst 0x4fa1e114 // sdot v20.4s, v8.16b, v1.4b[1]\n" + "ldr q8, [x10, #0xb0]\n" ".inst 0x4fa0e131 // sdot v17.4s, v9.16b, v0.4b[1]\n" + "add x25, x25, #0x10\n" ".inst 0x4fa1e135 // sdot v21.4s, v9.16b, v1.4b[1]\n" + "ldr q9, [x10, #0xc0]\n" ".inst 0x4fa0e152 // sdot v18.4s, v10.16b, v0.4b[1]\n" + "add x22, x22, #0x10\n" ".inst 0x4fa1e156 // sdot v22.4s, v10.16b, v1.4b[1]\n" - "ldr q4, [x10, #0x70]\n" - "ldr q5, [x10, #0x80]\n" - "ldr q6, [x10, #0x90]\n" + "ldr q10, [x10, #0xd0]\n" ".inst 0x4fa0e093 // sdot v19.4s, v4.16b, v0.4b[1]\n" ".inst 0x4fa1e097 // sdot v23.4s, v4.16b, v1.4b[1]\n" + "ldr q4, [x10, #0xe0]\n" ".inst 0x4f80e8b0 // sdot v16.4s, v5.16b, v0.4b[2]\n" ".inst 0x4f81e8b4 // sdot v20.4s, v5.16b, v1.4b[2]\n" + "ldr q5, [x10, #0xf0]\n" ".inst 0x4f80e8d1 // sdot v17.4s, v6.16b, v0.4b[2]\n" + "add x10, x10, #0x100\n" ".inst 0x4f81e8d5 // sdot v21.4s, v6.16b, v1.4b[2]\n" - "ldr q7, [x10, #0xa0]\n" - "ldr q8, [x10, #0xb0]\n" - "ldr q9, [x10, #0xc0]\n" ".inst 0x4f80e8f2 // sdot v18.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8f6 // sdot v22.4s, v7.16b, v1.4b[2]\n" ".inst 0x4f80e913 // sdot v19.4s, v8.16b, v0.4b[2]\n" ".inst 0x4f81e917 // sdot v23.4s, v8.16b, v1.4b[2]\n" ".inst 0x4fa0e930 // sdot v16.4s, v9.16b, v0.4b[3]\n" ".inst 0x4fa1e934 // sdot v20.4s, v9.16b, v1.4b[3]\n" - "ldr q10, [x10, #0xd0]\n" - "ldr q4, [x10, #0xe0]\n" - "ldr q5, [x10, #0xf0]\n" - "add x10, x10, #0x100\n" ".inst 0x4fa0e951 // sdot v17.4s, v10.16b, v0.4b[3]\n" ".inst 0x4fa1e955 // sdot v21.4s, v10.16b, v1.4b[3]\n" ".inst 0x4fa0e892 // sdot v18.4s, v4.16b, v0.4b[3]\n" @@ -617,13 +617,13 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( "cmp x26, #0x4\n" "ldr q8, [x10, #0x20]\n" ".inst 0x4f80e0d0 // sdot v16.4s, v6.16b, v0.4b[0]\n" + "ldr q9, [x10, #0x30]\n" ".inst 0x4f81e0d4 // sdot v20.4s, v6.16b, v1.4b[0]\n" + "add x10, x10, #0x40\n" ".inst 0x4f80e0f1 // sdot v17.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0f5 // sdot v21.4s, v7.16b, v1.4b[0]\n" ".inst 0x4f80e112 // sdot v18.4s, v8.16b, v0.4b[0]\n" ".inst 0x4f81e116 // sdot v22.4s, v8.16b, v1.4b[0]\n" - "ldr q9, [x10, #0x30]\n" - "add x10, x10, #0x40\n" ".inst 0x4f80e133 // sdot v19.4s, v9.16b, v0.4b[0]\n" ".inst 0x4f81e137 // sdot v23.4s, v9.16b, v1.4b[0]\n" "bge 42b\n" @@ -648,13 +648,13 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( "ldr q4, [x10, #0x10]\n" "ldr q5, [x10, #0x20]\n" ".inst 0x4f80e150 // sdot v16.4s, v10.16b, v0.4b[0]\n" + "ldr q6, [x10, #0x30]\n" ".inst 0x4f81e154 // sdot v20.4s, v10.16b, v1.4b[0]\n" + "add x10, x10, #0x40\n" ".inst 0x4f80e091 // sdot v17.4s, v4.16b, v0.4b[0]\n" ".inst 0x4f81e095 // sdot v21.4s, v4.16b, v1.4b[0]\n" ".inst 0x4f80e0b2 // sdot v18.4s, v5.16b, v0.4b[0]\n" ".inst 0x4f81e0b6 // sdot v22.4s, v5.16b, v1.4b[0]\n" - "ldr q6, [x10, #0x30]\n" - "add x10, x10, #0x40\n" ".inst 0x4f80e0d3 // sdot v19.4s, v6.16b, v0.4b[0]\n" ".inst 0x4f81e0d7 // sdot v23.4s, v6.16b, v1.4b[0]\n" "48:" // Height 2: Multiply loop: No odd multiplies @@ -907,97 +907,97 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( "blt 69f\n" "67:" // Height 3: Multiply loop: Main loop head ".inst 0x4f80e090 // sdot v16.4s, v4.16b, v0.4b[0]\n" - ".inst 0x4f81e094 // sdot v20.4s, v4.16b, v1.4b[0]\n" - ".inst 0x4f82e098 // sdot v24.4s, v4.16b, v2.4b[0]\n" "ldr d5, [x10, #0x10]\n" + ".inst 0x4f81e094 // sdot v20.4s, v4.16b, v1.4b[0]\n" "ldr x24, [x10, #0x18]\n" - "add x25, x25, #0x10\n" + ".inst 0x4f82e098 // sdot v24.4s, v4.16b, v2.4b[0]\n" "ldr d6, [x10, #0x20]\n" - "add x22, x22, #0x10\n" "ldr x23, [x10, #0x28]\n" - "add x21, x21, #0x10\n" + "add x25, x25, #0x10\n" "mov v5.d[1], x24\n" "ldr d7, [x10, #0x30]\n" "ldr x19, [x10, #0x38]\n" - "mov v6.d[1], x23\n" + "add x22, x22, #0x10\n" ".inst 0x4f80e0b1 // sdot v17.4s, v5.16b, v0.4b[0]\n" + "mov v6.d[1], x23\n" ".inst 0x4f81e0b5 // sdot v21.4s, v5.16b, v1.4b[0]\n" + "ldr d8, [x10, #0x40]\n" ".inst 0x4f82e0b9 // sdot v25.4s, v5.16b, v2.4b[0]\n" "mov v7.d[1], x19\n" ".inst 0x4f80e0d2 // sdot v18.4s, v6.16b, v0.4b[0]\n" - ".inst 0x4f81e0d6 // sdot v22.4s, v6.16b, v1.4b[0]\n" - ".inst 0x4f82e0da // sdot v26.4s, v6.16b, v2.4b[0]\n" - ".inst 0x4f80e0f3 // sdot v19.4s, v7.16b, v0.4b[0]\n" - ".inst 0x4f81e0f7 // sdot v23.4s, v7.16b, v1.4b[0]\n" - ".inst 0x4f82e0fb // sdot v27.4s, v7.16b, v2.4b[0]\n" - "ldr d8, [x10, #0x40]\n" "ldr x23, [x10, #0x48]\n" + ".inst 0x4f81e0d6 // sdot v22.4s, v6.16b, v1.4b[0]\n" "ldr d9, [x10, #0x50]\n" + ".inst 0x4f82e0da // sdot v26.4s, v6.16b, v2.4b[0]\n" "ldr x19, [x10, #0x58]\n" + ".inst 0x4f80e0f3 // sdot v19.4s, v7.16b, v0.4b[0]\n" "mov v8.d[1], x23\n" + ".inst 0x4f81e0f7 // sdot v23.4s, v7.16b, v1.4b[0]\n" "ldr d10, [x10, #0x60]\n" - "ldr x23, [x10, #0x68]\n" + ".inst 0x4f82e0fb // sdot v27.4s, v7.16b, v2.4b[0]\n" "mov v9.d[1], x19\n" ".inst 0x4fa0e110 // sdot v16.4s, v8.16b, v0.4b[1]\n" + "ldr x23, [x10, #0x68]\n" ".inst 0x4fa1e114 // sdot v20.4s, v8.16b, v1.4b[1]\n" + "ldr d4, [x10, #0x70]\n" ".inst 0x4fa2e118 // sdot v24.4s, v8.16b, v2.4b[1]\n" - "mov v10.d[1], x23\n" + "ldr x19, [x10, #0x78]\n" ".inst 0x4fa0e131 // sdot v17.4s, v9.16b, v0.4b[1]\n" + "mov v10.d[1], x23\n" ".inst 0x4fa1e135 // sdot v21.4s, v9.16b, v1.4b[1]\n" + "ldr d5, [x10, #0x80]\n" ".inst 0x4fa2e139 // sdot v25.4s, v9.16b, v2.4b[1]\n" + "mov v4.d[1], x19\n" ".inst 0x4fa0e152 // sdot v18.4s, v10.16b, v0.4b[1]\n" - ".inst 0x4fa1e156 // sdot v22.4s, v10.16b, v1.4b[1]\n" - ".inst 0x4fa2e15a // sdot v26.4s, v10.16b, v2.4b[1]\n" - "ldr d4, [x10, #0x70]\n" - "ldr x19, [x10, #0x78]\n" - "ldr d5, [x10, #0x80]\n" "ldr x24, [x10, #0x88]\n" - "mov v4.d[1], x19\n" + ".inst 0x4fa1e156 // sdot v22.4s, v10.16b, v1.4b[1]\n" "ldr d6, [x10, #0x90]\n" + ".inst 0x4fa2e15a // sdot v26.4s, v10.16b, v2.4b[1]\n" "ldr x23, [x10, #0x98]\n" - "mov v5.d[1], x24\n" ".inst 0x4fa0e093 // sdot v19.4s, v4.16b, v0.4b[1]\n" + "mov v5.d[1], x24\n" ".inst 0x4fa1e097 // sdot v23.4s, v4.16b, v1.4b[1]\n" + "ldr d7, [x10, #0xa0]\n" ".inst 0x4fa2e09b // sdot v27.4s, v4.16b, v2.4b[1]\n" "mov v6.d[1], x23\n" ".inst 0x4f80e8b0 // sdot v16.4s, v5.16b, v0.4b[2]\n" - ".inst 0x4f81e8b4 // sdot v20.4s, v5.16b, v1.4b[2]\n" - ".inst 0x4f82e8b8 // sdot v24.4s, v5.16b, v2.4b[2]\n" - ".inst 0x4f80e8d1 // sdot v17.4s, v6.16b, v0.4b[2]\n" - ".inst 0x4f81e8d5 // sdot v21.4s, v6.16b, v1.4b[2]\n" - ".inst 0x4f82e8d9 // sdot v25.4s, v6.16b, v2.4b[2]\n" - "ldr d7, [x10, #0xa0]\n" "ldr x19, [x10, #0xa8]\n" + ".inst 0x4f81e8b4 // sdot v20.4s, v5.16b, v1.4b[2]\n" "ldr d8, [x10, #0xb0]\n" + ".inst 0x4f82e8b8 // sdot v24.4s, v5.16b, v2.4b[2]\n" "ldr x23, [x10, #0xb8]\n" + ".inst 0x4f80e8d1 // sdot v17.4s, v6.16b, v0.4b[2]\n" "mov v7.d[1], x19\n" + ".inst 0x4f81e8d5 // sdot v21.4s, v6.16b, v1.4b[2]\n" "ldr d9, [x10, #0xc0]\n" - "ldr x19, [x10, #0xc8]\n" + ".inst 0x4f82e8d9 // sdot v25.4s, v6.16b, v2.4b[2]\n" "mov v8.d[1], x23\n" ".inst 0x4f80e8f2 // sdot v18.4s, v7.16b, v0.4b[2]\n" + "ldr x19, [x10, #0xc8]\n" ".inst 0x4f81e8f6 // sdot v22.4s, v7.16b, v1.4b[2]\n" + "ldr d10, [x10, #0xd0]\n" ".inst 0x4f82e8fa // sdot v26.4s, v7.16b, v2.4b[2]\n" - "mov v9.d[1], x19\n" + "ldr x23, [x10, #0xd8]\n" ".inst 0x4f80e913 // sdot v19.4s, v8.16b, v0.4b[2]\n" + "mov v9.d[1], x19\n" ".inst 0x4f81e917 // sdot v23.4s, v8.16b, v1.4b[2]\n" + "ldr d4, [x10, #0xe0]\n" ".inst 0x4f82e91b // sdot v27.4s, v8.16b, v2.4b[2]\n" + "mov v10.d[1], x23\n" ".inst 0x4fa0e930 // sdot v16.4s, v9.16b, v0.4b[3]\n" - ".inst 0x4fa1e934 // sdot v20.4s, v9.16b, v1.4b[3]\n" - ".inst 0x4fa2e938 // sdot v24.4s, v9.16b, v2.4b[3]\n" - "ldr d10, [x10, #0xd0]\n" - "ldr x23, [x10, #0xd8]\n" - "ldr d4, [x10, #0xe0]\n" "ldr x19, [x10, #0xe8]\n" - "mov v10.d[1], x23\n" + ".inst 0x4fa1e934 // sdot v20.4s, v9.16b, v1.4b[3]\n" "ldr d5, [x10, #0xf0]\n" + ".inst 0x4fa2e938 // sdot v24.4s, v9.16b, v2.4b[3]\n" "ldr x24, [x10, #0xf8]\n" - "add x10, x10, #0x100\n" - "mov v4.d[1], x19\n" ".inst 0x4fa0e951 // sdot v17.4s, v10.16b, v0.4b[3]\n" + "mov v4.d[1], x19\n" ".inst 0x4fa1e955 // sdot v21.4s, v10.16b, v1.4b[3]\n" + "add x21, x21, #0x10\n" ".inst 0x4fa2e959 // sdot v25.4s, v10.16b, v2.4b[3]\n" "mov v5.d[1], x24\n" ".inst 0x4fa0e892 // sdot v18.4s, v4.16b, v0.4b[3]\n" + "add x10, x10, #0x100\n" ".inst 0x4fa1e896 // sdot v22.4s, v4.16b, v1.4b[3]\n" ".inst 0x4fa2e89a // sdot v26.4s, v4.16b, v2.4b[3]\n" ".inst 0x4fa0e8b3 // sdot v19.4s, v5.16b, v0.4b[3]\n" @@ -1020,51 +1020,55 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( "bge 67b\n" "69:" // Height 3: Multiply loop: Single iteration only ".inst 0x4f80e090 // sdot v16.4s, v4.16b, v0.4b[0]\n" - ".inst 0x4f81e094 // sdot v20.4s, v4.16b, v1.4b[0]\n" - ".inst 0x4f82e098 // sdot v24.4s, v4.16b, v2.4b[0]\n" "ldr q5, [x10, #0x10]\n" + ".inst 0x4f81e094 // sdot v20.4s, v4.16b, v1.4b[0]\n" "ldr q6, [x10, #0x20]\n" - "sub x26, x26, #0x10\n" + ".inst 0x4f82e098 // sdot v24.4s, v4.16b, v2.4b[0]\n" "ldr q7, [x10, #0x30]\n" - "add x25, x25, #0x10\n" ".inst 0x4f80e0b1 // sdot v17.4s, v5.16b, v0.4b[0]\n" + "ldr q8, [x10, #0x40]\n" ".inst 0x4f81e0b5 // sdot v21.4s, v5.16b, v1.4b[0]\n" + "ldr q9, [x10, #0x50]\n" ".inst 0x4f82e0b9 // sdot v25.4s, v5.16b, v2.4b[0]\n" + "ldr q10, [x10, #0x60]\n" ".inst 0x4f80e0d2 // sdot v18.4s, v6.16b, v0.4b[0]\n" + "ldr q4, [x10, #0x70]\n" ".inst 0x4f81e0d6 // sdot v22.4s, v6.16b, v1.4b[0]\n" + "ldr q5, [x10, #0x80]\n" ".inst 0x4f82e0da // sdot v26.4s, v6.16b, v2.4b[0]\n" + "ldr q6, [x10, #0x90]\n" ".inst 0x4f80e0f3 // sdot v19.4s, v7.16b, v0.4b[0]\n" + "sub x26, x26, #0x10\n" ".inst 0x4f81e0f7 // sdot v23.4s, v7.16b, v1.4b[0]\n" + "add x25, x25, #0x10\n" ".inst 0x4f82e0fb // sdot v27.4s, v7.16b, v2.4b[0]\n" - "ldr q8, [x10, #0x40]\n" - "ldr q9, [x10, #0x50]\n" - "add x22, x22, #0x10\n" - "ldr q10, [x10, #0x60]\n" - "add x21, x21, #0x10\n" + "ldr q7, [x10, #0xa0]\n" ".inst 0x4fa0e110 // sdot v16.4s, v8.16b, v0.4b[1]\n" + "add x22, x22, #0x10\n" ".inst 0x4fa1e114 // sdot v20.4s, v8.16b, v1.4b[1]\n" + "add x21, x21, #0x10\n" ".inst 0x4fa2e118 // sdot v24.4s, v8.16b, v2.4b[1]\n" + "ldr q8, [x10, #0xb0]\n" ".inst 0x4fa0e131 // sdot v17.4s, v9.16b, v0.4b[1]\n" ".inst 0x4fa1e135 // sdot v21.4s, v9.16b, v1.4b[1]\n" ".inst 0x4fa2e139 // sdot v25.4s, v9.16b, v2.4b[1]\n" + "ldr q9, [x10, #0xc0]\n" ".inst 0x4fa0e152 // sdot v18.4s, v10.16b, v0.4b[1]\n" ".inst 0x4fa1e156 // sdot v22.4s, v10.16b, v1.4b[1]\n" ".inst 0x4fa2e15a // sdot v26.4s, v10.16b, v2.4b[1]\n" - "ldr q4, [x10, #0x70]\n" - "ldr q5, [x10, #0x80]\n" - "ldr q6, [x10, #0x90]\n" + "ldr q10, [x10, #0xd0]\n" ".inst 0x4fa0e093 // sdot v19.4s, v4.16b, v0.4b[1]\n" ".inst 0x4fa1e097 // sdot v23.4s, v4.16b, v1.4b[1]\n" ".inst 0x4fa2e09b // sdot v27.4s, v4.16b, v2.4b[1]\n" + "ldr q4, [x10, #0xe0]\n" ".inst 0x4f80e8b0 // sdot v16.4s, v5.16b, v0.4b[2]\n" ".inst 0x4f81e8b4 // sdot v20.4s, v5.16b, v1.4b[2]\n" ".inst 0x4f82e8b8 // sdot v24.4s, v5.16b, v2.4b[2]\n" + "ldr q5, [x10, #0xf0]\n" ".inst 0x4f80e8d1 // sdot v17.4s, v6.16b, v0.4b[2]\n" + "add x10, x10, #0x100\n" ".inst 0x4f81e8d5 // sdot v21.4s, v6.16b, v1.4b[2]\n" ".inst 0x4f82e8d9 // sdot v25.4s, v6.16b, v2.4b[2]\n" - "ldr q7, [x10, #0xa0]\n" - "ldr q8, [x10, #0xb0]\n" - "ldr q9, [x10, #0xc0]\n" ".inst 0x4f80e8f2 // sdot v18.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8f6 // sdot v22.4s, v7.16b, v1.4b[2]\n" ".inst 0x4f82e8fa // sdot v26.4s, v7.16b, v2.4b[2]\n" @@ -1074,10 +1078,6 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( ".inst 0x4fa0e930 // sdot v16.4s, v9.16b, v0.4b[3]\n" ".inst 0x4fa1e934 // sdot v20.4s, v9.16b, v1.4b[3]\n" ".inst 0x4fa2e938 // sdot v24.4s, v9.16b, v2.4b[3]\n" - "ldr q10, [x10, #0xd0]\n" - "ldr q4, [x10, #0xe0]\n" - "ldr q5, [x10, #0xf0]\n" - "add x10, x10, #0x100\n" ".inst 0x4fa0e951 // sdot v17.4s, v10.16b, v0.4b[3]\n" ".inst 0x4fa1e955 // sdot v21.4s, v10.16b, v1.4b[3]\n" ".inst 0x4fa2e959 // sdot v25.4s, v10.16b, v2.4b[3]\n" @@ -1114,7 +1114,9 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( "cmp x26, #0x4\n" "ldr q8, [x10, #0x20]\n" ".inst 0x4f80e0d0 // sdot v16.4s, v6.16b, v0.4b[0]\n" + "ldr q9, [x10, #0x30]\n" ".inst 0x4f81e0d4 // sdot v20.4s, v6.16b, v1.4b[0]\n" + "add x10, x10, #0x40\n" ".inst 0x4f82e0d8 // sdot v24.4s, v6.16b, v2.4b[0]\n" ".inst 0x4f80e0f1 // sdot v17.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0f5 // sdot v21.4s, v7.16b, v1.4b[0]\n" @@ -1122,8 +1124,6 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( ".inst 0x4f80e112 // sdot v18.4s, v8.16b, v0.4b[0]\n" ".inst 0x4f81e116 // sdot v22.4s, v8.16b, v1.4b[0]\n" ".inst 0x4f82e11a // sdot v26.4s, v8.16b, v2.4b[0]\n" - "ldr q9, [x10, #0x30]\n" - "add x10, x10, #0x40\n" ".inst 0x4f80e133 // sdot v19.4s, v9.16b, v0.4b[0]\n" ".inst 0x4f81e137 // sdot v23.4s, v9.16b, v1.4b[0]\n" ".inst 0x4f82e13b // sdot v27.4s, v9.16b, v2.4b[0]\n" @@ -1153,7 +1153,9 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( "ldr q4, [x10, #0x10]\n" "ldr q5, [x10, #0x20]\n" ".inst 0x4f80e150 // sdot v16.4s, v10.16b, v0.4b[0]\n" + "ldr q6, [x10, #0x30]\n" ".inst 0x4f81e154 // sdot v20.4s, v10.16b, v1.4b[0]\n" + "add x10, x10, #0x40\n" ".inst 0x4f82e158 // sdot v24.4s, v10.16b, v2.4b[0]\n" ".inst 0x4f80e091 // sdot v17.4s, v4.16b, v0.4b[0]\n" ".inst 0x4f81e095 // sdot v21.4s, v4.16b, v1.4b[0]\n" @@ -1161,8 +1163,6 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( ".inst 0x4f80e0b2 // sdot v18.4s, v5.16b, v0.4b[0]\n" ".inst 0x4f81e0b6 // sdot v22.4s, v5.16b, v1.4b[0]\n" ".inst 0x4f82e0ba // sdot v26.4s, v5.16b, v2.4b[0]\n" - "ldr q6, [x10, #0x30]\n" - "add x10, x10, #0x40\n" ".inst 0x4f80e0d3 // sdot v19.4s, v6.16b, v0.4b[0]\n" ".inst 0x4f81e0d7 // sdot v23.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0db // sdot v27.4s, v6.16b, v2.4b[0]\n" @@ -1492,110 +1492,110 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( "blt 99f\n" "97:" // Height 4: Multiply loop: Main loop head ".inst 0x4f80e090 // sdot v16.4s, v4.16b, v0.4b[0]\n" - ".inst 0x4f81e094 // sdot v20.4s, v4.16b, v1.4b[0]\n" - ".inst 0x4f82e098 // sdot v24.4s, v4.16b, v2.4b[0]\n" - ".inst 0x4f83e09c // sdot v28.4s, v4.16b, v3.4b[0]\n" "ldr d5, [x10, #0x10]\n" - "add x25, x25, #0x10\n" + ".inst 0x4f81e094 // sdot v20.4s, v4.16b, v1.4b[0]\n" "ldr x24, [x10, #0x18]\n" - "add x22, x22, #0x10\n" + ".inst 0x4f82e098 // sdot v24.4s, v4.16b, v2.4b[0]\n" "ldr d6, [x10, #0x20]\n" - "add x21, x21, #0x10\n" + ".inst 0x4f83e09c // sdot v28.4s, v4.16b, v3.4b[0]\n" "ldr x23, [x10, #0x28]\n" - "add x20, x20, #0x10\n" "mov v5.d[1], x24\n" "ldr d7, [x10, #0x30]\n" "ldr x19, [x10, #0x38]\n" - "mov v6.d[1], x23\n" + "add x25, x25, #0x10\n" ".inst 0x4f80e0b1 // sdot v17.4s, v5.16b, v0.4b[0]\n" + "mov v6.d[1], x23\n" ".inst 0x4f81e0b5 // sdot v21.4s, v5.16b, v1.4b[0]\n" + "ldr d8, [x10, #0x40]\n" ".inst 0x4f82e0b9 // sdot v25.4s, v5.16b, v2.4b[0]\n" "mov v7.d[1], x19\n" ".inst 0x4f83e0bd // sdot v29.4s, v5.16b, v3.4b[0]\n" + "ldr x23, [x10, #0x48]\n" ".inst 0x4f80e0d2 // sdot v18.4s, v6.16b, v0.4b[0]\n" + "ldr d9, [x10, #0x50]\n" ".inst 0x4f81e0d6 // sdot v22.4s, v6.16b, v1.4b[0]\n" + "ldr x19, [x10, #0x58]\n" ".inst 0x4f82e0da // sdot v26.4s, v6.16b, v2.4b[0]\n" + "mov v8.d[1], x23\n" ".inst 0x4f83e0de // sdot v30.4s, v6.16b, v3.4b[0]\n" + "ldr d10, [x10, #0x60]\n" ".inst 0x4f80e0f3 // sdot v19.4s, v7.16b, v0.4b[0]\n" + "mov v9.d[1], x19\n" ".inst 0x4f81e0f7 // sdot v23.4s, v7.16b, v1.4b[0]\n" + "ldr x23, [x10, #0x68]\n" ".inst 0x4f82e0fb // sdot v27.4s, v7.16b, v2.4b[0]\n" + "ldr d4, [x10, #0x70]\n" ".inst 0x4f83e0ff // sdot v31.4s, v7.16b, v3.4b[0]\n" - "ldr d8, [x10, #0x40]\n" - "ldr x23, [x10, #0x48]\n" - "ldr d9, [x10, #0x50]\n" - "ldr x19, [x10, #0x58]\n" - "mov v8.d[1], x23\n" - "ldr d10, [x10, #0x60]\n" - "ldr x23, [x10, #0x68]\n" - "mov v9.d[1], x19\n" + "ldr x19, [x10, #0x78]\n" ".inst 0x4fa0e110 // sdot v16.4s, v8.16b, v0.4b[1]\n" + "mov v10.d[1], x23\n" ".inst 0x4fa1e114 // sdot v20.4s, v8.16b, v1.4b[1]\n" + "ldr d5, [x10, #0x80]\n" ".inst 0x4fa2e118 // sdot v24.4s, v8.16b, v2.4b[1]\n" - "mov v10.d[1], x23\n" + "mov v4.d[1], x19\n" ".inst 0x4fa3e11c // sdot v28.4s, v8.16b, v3.4b[1]\n" + "ldr x24, [x10, #0x88]\n" ".inst 0x4fa0e131 // sdot v17.4s, v9.16b, v0.4b[1]\n" + "ldr d6, [x10, #0x90]\n" ".inst 0x4fa1e135 // sdot v21.4s, v9.16b, v1.4b[1]\n" + "ldr x23, [x10, #0x98]\n" ".inst 0x4fa2e139 // sdot v25.4s, v9.16b, v2.4b[1]\n" + "mov v5.d[1], x24\n" ".inst 0x4fa3e13d // sdot v29.4s, v9.16b, v3.4b[1]\n" + "ldr d7, [x10, #0xa0]\n" ".inst 0x4fa0e152 // sdot v18.4s, v10.16b, v0.4b[1]\n" + "mov v6.d[1], x23\n" ".inst 0x4fa1e156 // sdot v22.4s, v10.16b, v1.4b[1]\n" + "ldr x19, [x10, #0xa8]\n" ".inst 0x4fa2e15a // sdot v26.4s, v10.16b, v2.4b[1]\n" + "ldr d8, [x10, #0xb0]\n" ".inst 0x4fa3e15e // sdot v30.4s, v10.16b, v3.4b[1]\n" - "ldr d4, [x10, #0x70]\n" - "ldr x19, [x10, #0x78]\n" - "ldr d5, [x10, #0x80]\n" - "ldr x24, [x10, #0x88]\n" - "mov v4.d[1], x19\n" - "ldr d6, [x10, #0x90]\n" - "ldr x23, [x10, #0x98]\n" - "mov v5.d[1], x24\n" + "ldr x23, [x10, #0xb8]\n" ".inst 0x4fa0e093 // sdot v19.4s, v4.16b, v0.4b[1]\n" + "mov v7.d[1], x19\n" ".inst 0x4fa1e097 // sdot v23.4s, v4.16b, v1.4b[1]\n" + "ldr d9, [x10, #0xc0]\n" ".inst 0x4fa2e09b // sdot v27.4s, v4.16b, v2.4b[1]\n" - "mov v6.d[1], x23\n" + "mov v8.d[1], x23\n" ".inst 0x4fa3e09f // sdot v31.4s, v4.16b, v3.4b[1]\n" + "ldr x19, [x10, #0xc8]\n" ".inst 0x4f80e8b0 // sdot v16.4s, v5.16b, v0.4b[2]\n" + "ldr d10, [x10, #0xd0]\n" ".inst 0x4f81e8b4 // sdot v20.4s, v5.16b, v1.4b[2]\n" + "ldr x23, [x10, #0xd8]\n" ".inst 0x4f82e8b8 // sdot v24.4s, v5.16b, v2.4b[2]\n" + "mov v9.d[1], x19\n" ".inst 0x4f83e8bc // sdot v28.4s, v5.16b, v3.4b[2]\n" + "ldr d4, [x10, #0xe0]\n" ".inst 0x4f80e8d1 // sdot v17.4s, v6.16b, v0.4b[2]\n" + "mov v10.d[1], x23\n" ".inst 0x4f81e8d5 // sdot v21.4s, v6.16b, v1.4b[2]\n" + "ldr x19, [x10, #0xe8]\n" ".inst 0x4f82e8d9 // sdot v25.4s, v6.16b, v2.4b[2]\n" + "ldr d5, [x10, #0xf0]\n" ".inst 0x4f83e8dd // sdot v29.4s, v6.16b, v3.4b[2]\n" - "ldr d7, [x10, #0xa0]\n" - "ldr x19, [x10, #0xa8]\n" - "ldr d8, [x10, #0xb0]\n" - "ldr x23, [x10, #0xb8]\n" - "mov v7.d[1], x19\n" - "ldr d9, [x10, #0xc0]\n" - "ldr x19, [x10, #0xc8]\n" - "mov v8.d[1], x23\n" + "ldr x24, [x10, #0xf8]\n" ".inst 0x4f80e8f2 // sdot v18.4s, v7.16b, v0.4b[2]\n" + "mov v4.d[1], x19\n" ".inst 0x4f81e8f6 // sdot v22.4s, v7.16b, v1.4b[2]\n" + "add x22, x22, #0x10\n" ".inst 0x4f82e8fa // sdot v26.4s, v7.16b, v2.4b[2]\n" - "mov v9.d[1], x19\n" + "mov v5.d[1], x24\n" ".inst 0x4f83e8fe // sdot v30.4s, v7.16b, v3.4b[2]\n" + "add x21, x21, #0x10\n" ".inst 0x4f80e913 // sdot v19.4s, v8.16b, v0.4b[2]\n" + "add x20, x20, #0x10\n" ".inst 0x4f81e917 // sdot v23.4s, v8.16b, v1.4b[2]\n" + "add x10, x10, #0x100\n" ".inst 0x4f82e91b // sdot v27.4s, v8.16b, v2.4b[2]\n" ".inst 0x4f83e91f // sdot v31.4s, v8.16b, v3.4b[2]\n" ".inst 0x4fa0e930 // sdot v16.4s, v9.16b, v0.4b[3]\n" ".inst 0x4fa1e934 // sdot v20.4s, v9.16b, v1.4b[3]\n" ".inst 0x4fa2e938 // sdot v24.4s, v9.16b, v2.4b[3]\n" ".inst 0x4fa3e93c // sdot v28.4s, v9.16b, v3.4b[3]\n" - "ldr d10, [x10, #0xd0]\n" - "ldr x23, [x10, #0xd8]\n" - "ldr d4, [x10, #0xe0]\n" - "ldr x19, [x10, #0xe8]\n" - "mov v10.d[1], x23\n" - "ldr d5, [x10, #0xf0]\n" - "ldr x24, [x10, #0xf8]\n" - "add x10, x10, #0x100\n" - "mov v4.d[1], x19\n" ".inst 0x4fa0e951 // sdot v17.4s, v10.16b, v0.4b[3]\n" ".inst 0x4fa1e955 // sdot v21.4s, v10.16b, v1.4b[3]\n" ".inst 0x4fa2e959 // sdot v25.4s, v10.16b, v2.4b[3]\n" - "mov v5.d[1], x24\n" ".inst 0x4fa3e95d // sdot v29.4s, v10.16b, v3.4b[3]\n" ".inst 0x4fa0e892 // sdot v18.4s, v4.16b, v0.4b[3]\n" ".inst 0x4fa1e896 // sdot v22.4s, v4.16b, v1.4b[3]\n" @@ -1625,62 +1625,66 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( "bge 97b\n" "99:" // Height 4: Multiply loop: Single iteration only ".inst 0x4f80e090 // sdot v16.4s, v4.16b, v0.4b[0]\n" - ".inst 0x4f81e094 // sdot v20.4s, v4.16b, v1.4b[0]\n" - ".inst 0x4f82e098 // sdot v24.4s, v4.16b, v2.4b[0]\n" - ".inst 0x4f83e09c // sdot v28.4s, v4.16b, v3.4b[0]\n" "ldr q5, [x10, #0x10]\n" - "sub x26, x26, #0x10\n" + ".inst 0x4f81e094 // sdot v20.4s, v4.16b, v1.4b[0]\n" "ldr q6, [x10, #0x20]\n" - "add x25, x25, #0x10\n" + ".inst 0x4f82e098 // sdot v24.4s, v4.16b, v2.4b[0]\n" "ldr q7, [x10, #0x30]\n" - "add x22, x22, #0x10\n" + ".inst 0x4f83e09c // sdot v28.4s, v4.16b, v3.4b[0]\n" + "ldr q8, [x10, #0x40]\n" ".inst 0x4f80e0b1 // sdot v17.4s, v5.16b, v0.4b[0]\n" + "ldr q9, [x10, #0x50]\n" ".inst 0x4f81e0b5 // sdot v21.4s, v5.16b, v1.4b[0]\n" + "ldr q10, [x10, #0x60]\n" ".inst 0x4f82e0b9 // sdot v25.4s, v5.16b, v2.4b[0]\n" + "ldr q4, [x10, #0x70]\n" ".inst 0x4f83e0bd // sdot v29.4s, v5.16b, v3.4b[0]\n" + "ldr q5, [x10, #0x80]\n" ".inst 0x4f80e0d2 // sdot v18.4s, v6.16b, v0.4b[0]\n" + "sub x26, x26, #0x10\n" ".inst 0x4f81e0d6 // sdot v22.4s, v6.16b, v1.4b[0]\n" + "add x25, x25, #0x10\n" ".inst 0x4f82e0da // sdot v26.4s, v6.16b, v2.4b[0]\n" + "add x22, x22, #0x10\n" ".inst 0x4f83e0de // sdot v30.4s, v6.16b, v3.4b[0]\n" + "ldr q6, [x10, #0x90]\n" ".inst 0x4f80e0f3 // sdot v19.4s, v7.16b, v0.4b[0]\n" + "add x21, x21, #0x10\n" ".inst 0x4f81e0f7 // sdot v23.4s, v7.16b, v1.4b[0]\n" + "add x20, x20, #0x10\n" ".inst 0x4f82e0fb // sdot v27.4s, v7.16b, v2.4b[0]\n" ".inst 0x4f83e0ff // sdot v31.4s, v7.16b, v3.4b[0]\n" - "ldr q8, [x10, #0x40]\n" - "add x21, x21, #0x10\n" - "ldr q9, [x10, #0x50]\n" - "add x20, x20, #0x10\n" - "ldr q10, [x10, #0x60]\n" + "ldr q7, [x10, #0xa0]\n" ".inst 0x4fa0e110 // sdot v16.4s, v8.16b, v0.4b[1]\n" ".inst 0x4fa1e114 // sdot v20.4s, v8.16b, v1.4b[1]\n" ".inst 0x4fa2e118 // sdot v24.4s, v8.16b, v2.4b[1]\n" ".inst 0x4fa3e11c // sdot v28.4s, v8.16b, v3.4b[1]\n" + "ldr q8, [x10, #0xb0]\n" ".inst 0x4fa0e131 // sdot v17.4s, v9.16b, v0.4b[1]\n" ".inst 0x4fa1e135 // sdot v21.4s, v9.16b, v1.4b[1]\n" ".inst 0x4fa2e139 // sdot v25.4s, v9.16b, v2.4b[1]\n" ".inst 0x4fa3e13d // sdot v29.4s, v9.16b, v3.4b[1]\n" + "ldr q9, [x10, #0xc0]\n" ".inst 0x4fa0e152 // sdot v18.4s, v10.16b, v0.4b[1]\n" ".inst 0x4fa1e156 // sdot v22.4s, v10.16b, v1.4b[1]\n" ".inst 0x4fa2e15a // sdot v26.4s, v10.16b, v2.4b[1]\n" ".inst 0x4fa3e15e // sdot v30.4s, v10.16b, v3.4b[1]\n" - "ldr q4, [x10, #0x70]\n" - "ldr q5, [x10, #0x80]\n" - "ldr q6, [x10, #0x90]\n" + "ldr q10, [x10, #0xd0]\n" ".inst 0x4fa0e093 // sdot v19.4s, v4.16b, v0.4b[1]\n" ".inst 0x4fa1e097 // sdot v23.4s, v4.16b, v1.4b[1]\n" ".inst 0x4fa2e09b // sdot v27.4s, v4.16b, v2.4b[1]\n" ".inst 0x4fa3e09f // sdot v31.4s, v4.16b, v3.4b[1]\n" + "ldr q4, [x10, #0xe0]\n" ".inst 0x4f80e8b0 // sdot v16.4s, v5.16b, v0.4b[2]\n" ".inst 0x4f81e8b4 // sdot v20.4s, v5.16b, v1.4b[2]\n" ".inst 0x4f82e8b8 // sdot v24.4s, v5.16b, v2.4b[2]\n" ".inst 0x4f83e8bc // sdot v28.4s, v5.16b, v3.4b[2]\n" + "ldr q5, [x10, #0xf0]\n" ".inst 0x4f80e8d1 // sdot v17.4s, v6.16b, v0.4b[2]\n" + "add x10, x10, #0x100\n" ".inst 0x4f81e8d5 // sdot v21.4s, v6.16b, v1.4b[2]\n" ".inst 0x4f82e8d9 // sdot v25.4s, v6.16b, v2.4b[2]\n" ".inst 0x4f83e8dd // sdot v29.4s, v6.16b, v3.4b[2]\n" - "ldr q7, [x10, #0xa0]\n" - "ldr q8, [x10, #0xb0]\n" - "ldr q9, [x10, #0xc0]\n" ".inst 0x4f80e8f2 // sdot v18.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8f6 // sdot v22.4s, v7.16b, v1.4b[2]\n" ".inst 0x4f82e8fa // sdot v26.4s, v7.16b, v2.4b[2]\n" @@ -1693,10 +1697,6 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( ".inst 0x4fa1e934 // sdot v20.4s, v9.16b, v1.4b[3]\n" ".inst 0x4fa2e938 // sdot v24.4s, v9.16b, v2.4b[3]\n" ".inst 0x4fa3e93c // sdot v28.4s, v9.16b, v3.4b[3]\n" - "ldr q10, [x10, #0xd0]\n" - "ldr q4, [x10, #0xe0]\n" - "ldr q5, [x10, #0xf0]\n" - "add x10, x10, #0x100\n" ".inst 0x4fa0e951 // sdot v17.4s, v10.16b, v0.4b[3]\n" ".inst 0x4fa1e955 // sdot v21.4s, v10.16b, v1.4b[3]\n" ".inst 0x4fa2e959 // sdot v25.4s, v10.16b, v2.4b[3]\n" @@ -1740,7 +1740,9 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( "cmp x26, #0x4\n" "ldr q8, [x10, #0x20]\n" ".inst 0x4f80e0d0 // sdot v16.4s, v6.16b, v0.4b[0]\n" + "ldr q9, [x10, #0x30]\n" ".inst 0x4f81e0d4 // sdot v20.4s, v6.16b, v1.4b[0]\n" + "add x10, x10, #0x40\n" ".inst 0x4f82e0d8 // sdot v24.4s, v6.16b, v2.4b[0]\n" ".inst 0x4f83e0dc // sdot v28.4s, v6.16b, v3.4b[0]\n" ".inst 0x4f80e0f1 // sdot v17.4s, v7.16b, v0.4b[0]\n" @@ -1751,8 +1753,6 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( ".inst 0x4f81e116 // sdot v22.4s, v8.16b, v1.4b[0]\n" ".inst 0x4f82e11a // sdot v26.4s, v8.16b, v2.4b[0]\n" ".inst 0x4f83e11e // sdot v30.4s, v8.16b, v3.4b[0]\n" - "ldr q9, [x10, #0x30]\n" - "add x10, x10, #0x40\n" ".inst 0x4f80e133 // sdot v19.4s, v9.16b, v0.4b[0]\n" ".inst 0x4f81e137 // sdot v23.4s, v9.16b, v1.4b[0]\n" ".inst 0x4f82e13b // sdot v27.4s, v9.16b, v2.4b[0]\n" @@ -1787,7 +1787,9 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( "ldr q4, [x10, #0x10]\n" "ldr q5, [x10, #0x20]\n" ".inst 0x4f80e150 // sdot v16.4s, v10.16b, v0.4b[0]\n" + "ldr q6, [x10, #0x30]\n" ".inst 0x4f81e154 // sdot v20.4s, v10.16b, v1.4b[0]\n" + "add x10, x10, #0x40\n" ".inst 0x4f82e158 // sdot v24.4s, v10.16b, v2.4b[0]\n" ".inst 0x4f83e15c // sdot v28.4s, v10.16b, v3.4b[0]\n" ".inst 0x4f80e091 // sdot v17.4s, v4.16b, v0.4b[0]\n" @@ -1798,8 +1800,6 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( ".inst 0x4f81e0b6 // sdot v22.4s, v5.16b, v1.4b[0]\n" ".inst 0x4f82e0ba // sdot v26.4s, v5.16b, v2.4b[0]\n" ".inst 0x4f83e0be // sdot v30.4s, v5.16b, v3.4b[0]\n" - "ldr q6, [x10, #0x30]\n" - "add x10, x10, #0x40\n" ".inst 0x4f80e0d3 // sdot v19.4s, v6.16b, v0.4b[0]\n" ".inst 0x4f81e0d7 // sdot v23.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0db // sdot v27.4s, v6.16b, v2.4b[0]\n" diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16/a55.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16/a55.cpp index 6da3f7a9e3..6e3a00ed72 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16/a55.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16/a55.cpp @@ -95,213 +95,213 @@ void a64_hybrid_s8qs_dot_6x16_a55 ( "cmp %x[M], #0x2\n" "bgt 55f\n" "beq 28f\n" - "ldr x17, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n" - "ldr x16, [%x[args_ptr], %[offsetof_shift_ptr]]\n" - "mov x15, %x[col_bias]\n" - "ldr x14, [%x[args_ptr], %[offsetof_N]]\n" - "mov x13, %x[output_ptr]\n" - "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "ldr x6, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n" + "ldr x7, [%x[args_ptr], %[offsetof_shift_ptr]]\n" + "mov x8, %x[col_bias]\n" + "ldr x17, [%x[args_ptr], %[offsetof_N]]\n" + "mov x16, %x[output_ptr]\n" + "ldr x15, [%x[args_ptr], %[offsetof_B_ptr]]\n" "2:" // Height 1: Column loop "movi v8.4s, #0x0\n" "movi v9.4s, #0x0\n" "movi v10.4s, #0x0\n" "movi v11.4s, #0x0\n" "3:" // Height 1: setup done - "mov x11, #0x0\n" + "mov x14, #0x0\n" "4:" // Height 1: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w10, [x20, x11, LSL #0x2]\n" + "ldr w13, [x20, x14, LSL #0x2]\n" "tbz %x[flags], #3, 5f\n" - "ldr x20, [%x[input_ptr], x11, LSL #0x3]\n" + "ldr x20, [%x[input_ptr], x14, LSL #0x3]\n" "add x20, x20, x19, LSL #3\n" - "ldr x9, [x20, #0x0]\n" - "cbnz x11, 6f\n" + "ldr x12, [x20, #0x0]\n" + "cbnz x14, 6f\n" "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x9, x9, x19\n" + "add x12, x12, x19\n" "b 6f\n" "5:" // Height 1: setup direct input - "mov x9, %x[input_ptr]\n" + "mov x12, %x[input_ptr]\n" "6:" // Height 1: input setup done - "cmp x10, #0x10\n" + "cmp x13, #0x10\n" "blt 9f\n" - "ldr q0, [x9, #0x0]\n" - "ldr q6, [x12, #0x0]\n" - "cmp x10, #0x20\n" + "ldr q0, [x12, #0x0]\n" + "ldr q6, [x15, #0x0]\n" + "cmp x13, #0x20\n" "blt 8f\n" "7:" // Height 1: Multiply loop: Main loop head ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr d7, [x12, #0x10]\n" - "ldr x19, [x12, #0x18]\n" - "add x9, x9, #0x10\n" - "ldr d6, [x12, #0x20]\n" - "sub x10, x10, #0x10\n" - "ldr x28, [x12, #0x28]\n" - "cmp x10, #0x20\n" - "mov v7.d[1], x19\n" - "prfm pldl1keep, [x9, #0x80]\n" - "ldr x19, [x12, #0x38]\n" - "mov v6.d[1], x28\n" + "ldr d7, [x15, #0x10]\n" + "ldr x11, [x15, #0x18]\n" + "add x12, x12, #0x10\n" + "ldr d6, [x15, #0x20]\n" + "sub x13, x13, #0x10\n" + "ldr x10, [x15, #0x28]\n" + "cmp x13, #0x20\n" + "mov v7.d[1], x11\n" + "prfm pldl1keep, [x12, #0x80]\n" + "ldr x11, [x15, #0x38]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" - "ldr d7, [x12, #0x30]\n" + "mov v6.d[1], x10\n" + "ldr d7, [x15, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" - "ldr d6, [x12, #0x40]\n" - "ldr x28, [x12, #0x48]\n" - "mov v7.d[1], x19\n" - "ldr x19, [x12, #0x58]\n" - "ldr x27, [x9, #0x8]\n" - "mov v6.d[1], x28\n" + "ldr d6, [x15, #0x40]\n" + "ldr x10, [x15, #0x48]\n" + "mov v7.d[1], x11\n" + "ldr x11, [x15, #0x58]\n" + "ldr x9, [x12, #0x8]\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" - "ldr d7, [x12, #0x50]\n" + "mov v6.d[1], x10\n" + "ldr d7, [x15, #0x50]\n" ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n" - "ldr d6, [x12, #0x60]\n" - "ldr x28, [x12, #0x68]\n" - "mov v7.d[1], x19\n" - "ldr x19, [x12, #0x78]\n" - "mov v6.d[1], x28\n" + "ldr d6, [x15, #0x60]\n" + "ldr x10, [x15, #0x68]\n" + "mov v7.d[1], x11\n" + "ldr x11, [x15, #0x78]\n" ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n" - "ldr d7, [x12, #0x70]\n" + "mov v6.d[1], x10\n" + "ldr d7, [x15, #0x70]\n" ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n" - "ldr d6, [x12, #0x80]\n" - "ldr x28, [x12, #0x88]\n" - "mov v7.d[1], x19\n" - "ldr x19, [x12, #0x98]\n" - "mov v6.d[1], x28\n" + "ldr d6, [x15, #0x80]\n" + "ldr x10, [x15, #0x88]\n" + "mov v7.d[1], x11\n" + "ldr x11, [x15, #0x98]\n" ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n" - "ldr d7, [x12, #0x90]\n" + "mov v6.d[1], x10\n" + "ldr d7, [x15, #0x90]\n" ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n" - "ldr d6, [x12, #0xa0]\n" - "ldr x28, [x12, #0xa8]\n" - "mov v7.d[1], x19\n" - "ldr x19, [x12, #0xb8]\n" - "mov v6.d[1], x28\n" + "ldr d6, [x15, #0xa0]\n" + "ldr x10, [x15, #0xa8]\n" + "mov v7.d[1], x11\n" + "ldr x11, [x15, #0xb8]\n" ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n" - "ldr d7, [x12, #0xb0]\n" + "mov v6.d[1], x10\n" + "ldr d7, [x15, #0xb0]\n" ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n" - "ldr d6, [x12, #0xc0]\n" - "ldr x28, [x12, #0xc8]\n" - "mov v7.d[1], x19\n" - "ldr x19, [x12, #0xd8]\n" - "mov v6.d[1], x28\n" + "ldr d6, [x15, #0xc0]\n" + "ldr x10, [x15, #0xc8]\n" + "mov v7.d[1], x11\n" + "ldr x11, [x15, #0xd8]\n" ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n" - "ldr d7, [x12, #0xd0]\n" + "mov v6.d[1], x10\n" + "ldr d7, [x15, #0xd0]\n" ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n" - "ldr d6, [x12, #0xe0]\n" - "ldr x28, [x12, #0xe8]\n" - "mov v7.d[1], x19\n" - "ldr x19, [x12, #0xf8]\n" - "mov v6.d[1], x28\n" + "ldr d6, [x15, #0xe0]\n" + "ldr x10, [x15, #0xe8]\n" + "mov v7.d[1], x11\n" + "ldr x11, [x15, #0xf8]\n" ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n" - "ldr d7, [x12, #0xf0]\n" - "add x12, x12, #0x100\n" + "mov v6.d[1], x10\n" + "ldr d7, [x15, #0xf0]\n" + "add x15, x15, #0x100\n" ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n" - "ldr d6, [x12, #0x0]\n" - "ldr x28, [x12, #0x8]\n" - "mov v7.d[1], x19\n" - "mov v6.d[1], x28\n" + "ldr d6, [x15, #0x0]\n" + "ldr x10, [x15, #0x8]\n" + "mov v7.d[1], x11\n" ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n" - "ldr d0, [x9, #0x0]\n" - "mov v0.d[1], x27\n" + "mov v6.d[1], x10\n" + "ldr d0, [x12, #0x0]\n" + "mov v0.d[1], x9\n" "bge 7b\n" "8:" // Height 1: Multiply loop: Single iteration only ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr q7, [x12, #0x10]\n" - "ldr q6, [x12, #0x20]\n" - "sub x10, x10, #0x10\n" - "add x9, x9, #0x10\n" + "ldr q7, [x15, #0x10]\n" + "ldr q6, [x15, #0x20]\n" + "sub x13, x13, #0x10\n" + "add x12, x12, #0x10\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" - "prfm pldl1keep, [x9, #0x80]\n" + "prfm pldl1keep, [x12, #0x80]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" - "ldr q7, [x12, #0x30]\n" - "ldr q6, [x12, #0x40]\n" + "ldr q7, [x15, #0x30]\n" + "ldr q6, [x15, #0x40]\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" - "ldr q7, [x12, #0x50]\n" + "ldr q7, [x15, #0x50]\n" ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n" - "ldr q6, [x12, #0x60]\n" + "ldr q6, [x15, #0x60]\n" ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n" - "ldr q7, [x12, #0x70]\n" + "ldr q7, [x15, #0x70]\n" ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n" - "ldr q6, [x12, #0x80]\n" + "ldr q6, [x15, #0x80]\n" ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n" - "ldr q7, [x12, #0x90]\n" + "ldr q7, [x15, #0x90]\n" ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n" - "ldr q6, [x12, #0xa0]\n" + "ldr q6, [x15, #0xa0]\n" ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n" - "ldr q7, [x12, #0xb0]\n" + "ldr q7, [x15, #0xb0]\n" ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n" - "ldr q6, [x12, #0xc0]\n" + "ldr q6, [x15, #0xc0]\n" ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n" - "ldr q7, [x12, #0xd0]\n" + "ldr q7, [x15, #0xd0]\n" ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n" - "ldr q6, [x12, #0xe0]\n" + "ldr q6, [x15, #0xe0]\n" ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n" - "ldr q7, [x12, #0xf0]\n" + "ldr q7, [x15, #0xf0]\n" ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n" - "add x12, x12, #0x100\n" + "add x15, x15, #0x100\n" ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n" "9:" // Height 1: Multiply loop: Main loop skip - "cbz x10, 14f\n" - "cmp x10, #0x4\n" + "cbz x13, 14f\n" + "cmp x13, #0x4\n" "blt 11f\n" "10:" // Height 1: Multiply loop: Odd block loop - "ldr s0, [x9], #0x4\n" - "sub x10, x10, #0x4\n" - "ldr q6, [x12, #0x0]\n" - "cmp x10, #0x4\n" - "ldr q7, [x12, #0x10]\n" + "ldr s0, [x12], #0x4\n" + "sub x13, x13, #0x4\n" + "ldr q6, [x15, #0x0]\n" + "cmp x13, #0x4\n" + "ldr q7, [x15, #0x10]\n" ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr q6, [x12, #0x20]\n" + "ldr q6, [x15, #0x20]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" - "ldr q7, [x12, #0x30]\n" - "add x12, x12, #0x40\n" + "ldr q7, [x15, #0x30]\n" + "add x15, x15, #0x40\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" "bge 10b\n" - "cbz x10, 14f\n" + "cbz x13, 14f\n" "11:" // Height 1: Multiply loop: Skip odd blocks - "tbz x10, #1, 12f\n" - "ldr h0, [x9], #0x2\n" - "tbz x10, #0, 13f\n" - "ld1 { v0.b }[2], [x9]\n" + "tbz x13, #1, 12f\n" + "ldr h0, [x12], #0x2\n" + "tbz x13, #0, 13f\n" + "ld1 { v0.b }[2], [x12]\n" "b 13f\n" "12:" // Height 1: Multiply loop: Ragged operand read: partial_1_0 - "ldr b0, [x9, #0x0]\n" + "ldr b0, [x12, #0x0]\n" "13:" // Height 1: Multiply loop: Ragged operand read: Done - "ldr q6, [x12, #0x0]\n" - "ldr q7, [x12, #0x10]\n" + "ldr q6, [x15, #0x0]\n" + "ldr q7, [x15, #0x10]\n" ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr q6, [x12, #0x20]\n" + "ldr q6, [x15, #0x20]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" - "ldr q7, [x12, #0x30]\n" - "add x12, x12, #0x40\n" + "ldr q7, [x15, #0x30]\n" + "add x15, x15, #0x40\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" "14:" // Height 1: Multiply loop: No odd multiplies "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" - "add x11, x11, #0x1\n" - "cmp x11, x19\n" + "add x14, x14, #0x1\n" + "cmp x14, x19\n" "bne 4b\n" - "prfm pstl1keep, [x13, #0x0]\n" - "ldr q0, [x15, #0x0]\n" - "ldr q1, [x15, #0x10]\n" - "ldr q2, [x15, #0x20]\n" + "prfm pstl1keep, [x16, #0x0]\n" + "ldr q0, [x8, #0x0]\n" + "ldr q1, [x8, #0x10]\n" + "ldr q2, [x8, #0x20]\n" "add v8.4s, v8.4s, v0.4s\n" - "ldr q3, [x15, #0x30]\n" + "ldr q3, [x8, #0x30]\n" "add v9.4s, v9.4s, v1.4s\n" - "add x15, x15, #0x40\n" + "add x8, x8, #0x40\n" "add v10.4s, v10.4s, v2.4s\n" "add v11.4s, v11.4s, v3.4s\n" "tbz %x[flags], #4, 15f\n" - "ldr q0, [x16, #0x0]\n" - "ldr q4, [x17, #0x0]\n" - "ldr q1, [x16, #0x10]\n" - "ldr q5, [x17, #0x10]\n" - "ldr q2, [x16, #0x20]\n" - "ldr q6, [x17, #0x20]\n" - "ldr q3, [x16, #0x30]\n" - "add x16, x16, #0x40\n" - "ldr q7, [x17, #0x30]\n" - "add x17, x17, #0x40\n" + "ldr q0, [x7, #0x0]\n" + "ldr q4, [x6, #0x0]\n" + "ldr q1, [x7, #0x10]\n" + "ldr q5, [x6, #0x10]\n" + "ldr q2, [x7, #0x20]\n" + "ldr q6, [x6, #0x20]\n" + "ldr q3, [x7, #0x30]\n" + "add x7, x7, #0x40\n" + "ldr q7, [x6, #0x30]\n" + "add x6, x6, #0x40\n" "b 16f\n" "15:" // Height 1: per layer parameters "add x25, %x[qp], %[per_layer_right_shift]\n" @@ -342,7 +342,7 @@ void a64_hybrid_s8qs_dot_6x16_a55 ( "ld1r { v4.4s }, [x24]\n" "add x24, %x[qp], %[maxval]\n" "ld1r { v5.4s }, [x25]\n" - "cmp x14, #0x10\n" + "cmp x17, #0x10\n" "ld1r { v6.4s }, [x24]\n" "add v8.4s, v8.4s, v4.4s\n" "add v9.4s, v9.4s, v4.4s\n" @@ -360,65 +360,65 @@ void a64_hybrid_s8qs_dot_6x16_a55 ( "uzp1 v9.8h, v10.8h, v11.8h\n" "uzp1 v8.16b, v8.16b, v9.16b\n" "bge 26f\n" - "tbz x14, #3, 21f\n" - "str d8, [x13], #0x8\n" - "tbz x14, #2, 19f\n" - "st1 { v8.s }[2], [x13], #0x4\n" - "tbz x14, #1, 18f\n" - "st1 { v8.h }[6], [x13], #0x2\n" - "tbz x14, #0, 25f\n" - "st1 { v8.b }[14], [x13]\n" + "tbz x17, #3, 21f\n" + "str d8, [x16], #0x8\n" + "tbz x17, #2, 19f\n" + "st1 { v8.s }[2], [x16], #0x4\n" + "tbz x17, #1, 18f\n" + "st1 { v8.h }[6], [x16], #0x2\n" + "tbz x17, #0, 25f\n" + "st1 { v8.b }[14], [x16]\n" "b 25f\n" "18:" // Height 1: Partial direct writeback: partial_1_12 - "tbz x14, #0, 25f\n" - "st1 { v8.b }[12], [x13]\n" + "tbz x17, #0, 25f\n" + "st1 { v8.b }[12], [x16]\n" "b 25f\n" "19:" // Height 1: Partial direct writeback: partial_2_8 - "tbz x14, #1, 20f\n" - "st1 { v8.h }[4], [x13], #0x2\n" - "tbz x14, #0, 25f\n" - "st1 { v8.b }[10], [x13]\n" + "tbz x17, #1, 20f\n" + "st1 { v8.h }[4], [x16], #0x2\n" + "tbz x17, #0, 25f\n" + "st1 { v8.b }[10], [x16]\n" "b 25f\n" "20:" // Height 1: Partial direct writeback: partial_1_8 - "tbz x14, #0, 25f\n" - "st1 { v8.b }[8], [x13]\n" + "tbz x17, #0, 25f\n" + "st1 { v8.b }[8], [x16]\n" "b 25f\n" "21:" // Height 1: Partial direct writeback: partial_4_0 - "tbz x14, #2, 23f\n" - "str s8, [x13], #0x4\n" - "tbz x14, #1, 22f\n" - "st1 { v8.h }[2], [x13], #0x2\n" - "tbz x14, #0, 25f\n" - "st1 { v8.b }[6], [x13]\n" + "tbz x17, #2, 23f\n" + "str s8, [x16], #0x4\n" + "tbz x17, #1, 22f\n" + "st1 { v8.h }[2], [x16], #0x2\n" + "tbz x17, #0, 25f\n" + "st1 { v8.b }[6], [x16]\n" "b 25f\n" "22:" // Height 1: Partial direct writeback: partial_1_4 - "tbz x14, #0, 25f\n" - "st1 { v8.b }[4], [x13]\n" + "tbz x17, #0, 25f\n" + "st1 { v8.b }[4], [x16]\n" "b 25f\n" "23:" // Height 1: Partial direct writeback: partial_2_0 - "tbz x14, #1, 24f\n" - "str h8, [x13], #0x2\n" - "tbz x14, #0, 25f\n" - "st1 { v8.b }[2], [x13]\n" + "tbz x17, #1, 24f\n" + "str h8, [x16], #0x2\n" + "tbz x17, #0, 25f\n" + "st1 { v8.b }[2], [x16]\n" "b 25f\n" "24:" // Height 1: Partial direct writeback: partial_1_0 - "str b8, [x13, #0x0]\n" + "str b8, [x16, #0x0]\n" "25:" // Height 1: Partial direct writeback: Done "b 27f\n" "26:" // Height 1: Full writeback - "str q8, [x13, #0x0]\n" - "add x13, x13, #0x10\n" + "str q8, [x16, #0x0]\n" + "add x16, x16, #0x10\n" "27:" // Height 1: Writeback done - "subs x14, x14, #0x10\n" + "subs x17, x17, #0x10\n" "bgt 2b\n" "b 164f\n" "28:" // Height 2 - "ldr x17, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n" - "mov x15, %x[col_bias]\n" - "ldr x16, [%x[args_ptr], %[offsetof_shift_ptr]]\n" - "mov x13, %x[output_ptr]\n" - "ldr x14, [%x[args_ptr], %[offsetof_N]]\n" - "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "ldr x6, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n" + "mov x8, %x[col_bias]\n" + "ldr x7, [%x[args_ptr], %[offsetof_shift_ptr]]\n" + "mov x16, %x[output_ptr]\n" + "ldr x17, [%x[args_ptr], %[offsetof_N]]\n" + "ldr x15, [%x[args_ptr], %[offsetof_B_ptr]]\n" "29:" // Height 2: Column loop "movi v8.4s, #0x0\n" "movi v9.4s, #0x0\n" @@ -429,263 +429,263 @@ void a64_hybrid_s8qs_dot_6x16_a55 ( "movi v14.4s, #0x0\n" "movi v15.4s, #0x0\n" "30:" // Height 2: setup done - "mov x11, #0x0\n" + "mov x14, #0x0\n" "31:" // Height 2: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w10, [x20, x11, LSL #0x2]\n" + "ldr w13, [x20, x14, LSL #0x2]\n" "tbz %x[flags], #3, 32f\n" - "ldr x20, [%x[input_ptr], x11, LSL #0x3]\n" + "ldr x20, [%x[input_ptr], x14, LSL #0x3]\n" "add x20, x20, x19, LSL #3\n" - "ldr x9, [x20, #0x0]\n" - "ldr x26, [x20, #0x8]\n" - "cbnz x11, 33f\n" + "ldr x12, [x20, #0x0]\n" + "ldr x28, [x20, #0x8]\n" + "cbnz x14, 33f\n" "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x9, x9, x19\n" - "add x26, x26, x19\n" + "add x12, x12, x19\n" + "add x28, x28, x19\n" "b 33f\n" "32:" // Height 2: setup direct input - "mov x9, %x[input_ptr]\n" - "add x26, x9, x19\n" + "mov x12, %x[input_ptr]\n" + "add x28, x12, x19\n" "33:" // Height 2: input setup done - "cmp x10, #0x10\n" + "cmp x13, #0x10\n" "blt 36f\n" - "ldr q0, [x9, #0x0]\n" - "ldr q1, [x26, #0x0]\n" - "cmp x10, #0x20\n" - "ldr q6, [x12, #0x0]\n" + "ldr q0, [x12, #0x0]\n" + "ldr q1, [x28, #0x0]\n" + "cmp x13, #0x20\n" + "ldr q6, [x15, #0x0]\n" "blt 35f\n" "34:" // Height 2: Multiply loop: Main loop head ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" + "ldr d7, [x15, #0x10]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" - "ldr d7, [x12, #0x10]\n" - "add x9, x9, #0x10\n" - "ldr x19, [x12, #0x18]\n" - "add x26, x26, #0x10\n" - "prfm pldl1keep, [x9, #0x80]\n" - "sub x10, x10, #0x10\n" - "ldr d6, [x12, #0x20]\n" - "cmp x10, #0x20\n" - "mov v7.d[1], x19\n" - "ldr x28, [x12, #0x28]\n" - "ldr x19, [x12, #0x38]\n" + "ldr x11, [x15, #0x18]\n" + "ldr d6, [x15, #0x20]\n" + "add x12, x12, #0x10\n" + "ldr x10, [x15, #0x28]\n" + "add x28, x28, #0x10\n" + "mov v7.d[1], x11\n" + "prfm pldl1keep, [x12, #0x80]\n" + "prfm pldl1keep, [x28, #0x80]\n" + "sub x13, x13, #0x10\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" + "mov v6.d[1], x10\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" - "mov v6.d[1], x28\n" - "ldr d7, [x12, #0x30]\n" - "prfm pldl1keep, [x26, #0x80]\n" + "ldr d7, [x15, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "ldr x11, [x15, #0x38]\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" - "mov v7.d[1], x19\n" - "ldr d6, [x12, #0x40]\n" - "ldr x28, [x12, #0x48]\n" + "ldr d6, [x15, #0x40]\n" + "ldr x10, [x15, #0x48]\n" + "cmp x13, #0x20\n" + "mov v7.d[1], x11\n" + "ldr x11, [x15, #0x58]\n" + "ldr x9, [x12, #0x8]\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" + "mov v6.d[1], x10\n" ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" - "ldr d7, [x12, #0x50]\n" - "mov v6.d[1], x28\n" - "ldr x19, [x12, #0x58]\n" - "ldr x28, [x12, #0x68]\n" + "ldr d7, [x15, #0x50]\n" ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n" + "ldr x10, [x15, #0x68]\n" ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n" - "mov v7.d[1], x19\n" - "ldr d6, [x12, #0x60]\n" - "ldr x19, [x12, #0x78]\n" + "ldr d6, [x15, #0x60]\n" + "mov v7.d[1], x11\n" + "ldr x11, [x15, #0x78]\n" + "ldr x27, [x28, #0x8]\n" ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n" + "mov v6.d[1], x10\n" ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n" - "mov v6.d[1], x28\n" - "ldr d7, [x12, #0x70]\n" - "ldr x28, [x12, #0x88]\n" + "ldr d7, [x15, #0x70]\n" ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n" + "ldr x10, [x15, #0x88]\n" ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n" - "mov v7.d[1], x19\n" - "ldr d6, [x12, #0x80]\n" - "ldr x19, [x12, #0x98]\n" + "ldr d6, [x15, #0x80]\n" + "mov v7.d[1], x11\n" + "ldr x11, [x15, #0x98]\n" ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n" + "mov v6.d[1], x10\n" ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n" - "mov v6.d[1], x28\n" - "ldr d7, [x12, #0x90]\n" - "ldr x28, [x12, #0xa8]\n" + "ldr d7, [x15, #0x90]\n" ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n" + "ldr x10, [x15, #0xa8]\n" ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n" - "mov v7.d[1], x19\n" - "ldr d6, [x12, #0xa0]\n" - "ldr x19, [x12, #0xb8]\n" + "ldr d6, [x15, #0xa0]\n" + "mov v7.d[1], x11\n" + "ldr x11, [x15, #0xb8]\n" ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n" + "mov v6.d[1], x10\n" ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n" - "mov v6.d[1], x28\n" - "ldr d7, [x12, #0xb0]\n" - "ldr x28, [x12, #0xc8]\n" + "ldr d7, [x15, #0xb0]\n" ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n" + "ldr x10, [x15, #0xc8]\n" ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n" - "mov v7.d[1], x19\n" - "ldr d6, [x12, #0xc0]\n" - "ldr x19, [x12, #0xd8]\n" + "ldr d6, [x15, #0xc0]\n" + "mov v7.d[1], x11\n" + "ldr x11, [x15, #0xd8]\n" ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n" + "mov v6.d[1], x10\n" ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n" - "mov v6.d[1], x28\n" - "ldr d7, [x12, #0xd0]\n" - "ldr x28, [x12, #0xe8]\n" + "ldr d7, [x15, #0xd0]\n" ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n" + "ldr x10, [x15, #0xe8]\n" ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n" - "mov v7.d[1], x19\n" - "ldr d6, [x12, #0xe0]\n" - "ldr x19, [x12, #0xf8]\n" + "ldr d6, [x15, #0xe0]\n" + "mov v7.d[1], x11\n" + "ldr x11, [x15, #0xf8]\n" ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n" + "mov v6.d[1], x10\n" ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n" - "mov v6.d[1], x28\n" - "ldr d7, [x12, #0xf0]\n" - "ldr x27, [x9, #0x8]\n" - "add x12, x12, #0x100\n" + "ldr d7, [x15, #0xf0]\n" ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n" + "add x15, x15, #0x100\n" ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n" - "mov v7.d[1], x19\n" - "ldr x25, [x26, #0x8]\n" - "ldr d6, [x12, #0x0]\n" + "ldr d6, [x15, #0x0]\n" + "mov v7.d[1], x11\n" + "ldr x10, [x15, #0x8]\n" ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n" + "ldr d0, [x12, #0x0]\n" ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n" - "ldr d0, [x9, #0x0]\n" - "ldr d1, [x26, #0x0]\n" - "ldr x28, [x12, #0x8]\n" - "mov v0.d[1], x27\n" - "mov v1.d[1], x25\n" - "mov v6.d[1], x28\n" + "mov v6.d[1], x10\n" + "ldr d1, [x28, #0x0]\n" + "mov v0.d[1], x9\n" + "mov v1.d[1], x27\n" "bge 34b\n" "35:" // Height 2: Multiply loop: Single iteration only ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" + "ldr q7, [x15, #0x10]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" - "ldr q7, [x12, #0x10]\n" - "sub x10, x10, #0x10\n" - "ldr q6, [x12, #0x20]\n" - "add x9, x9, #0x10\n" - "prfm pldl1keep, [x9, #0x80]\n" - "add x26, x26, #0x10\n" + "ldr q6, [x15, #0x20]\n" + "sub x13, x13, #0x10\n" + "add x12, x12, #0x10\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" + "prfm pldl1keep, [x12, #0x80]\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" + "ldr q7, [x15, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "add x28, x28, #0x10\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" - "ldr q7, [x12, #0x30]\n" - "prfm pldl1keep, [x26, #0x80]\n" - "ldr q6, [x12, #0x40]\n" + "prfm pldl1keep, [x28, #0x80]\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" + "ldr q6, [x15, #0x40]\n" ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" - "ldr q7, [x12, #0x50]\n" + "ldr q7, [x15, #0x50]\n" ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n" ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n" - "ldr q6, [x12, #0x60]\n" + "ldr q6, [x15, #0x60]\n" ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n" ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n" - "ldr q7, [x12, #0x70]\n" + "ldr q7, [x15, #0x70]\n" ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n" ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n" - "ldr q6, [x12, #0x80]\n" + "ldr q6, [x15, #0x80]\n" ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n" ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n" - "ldr q7, [x12, #0x90]\n" + "ldr q7, [x15, #0x90]\n" ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n" ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n" - "ldr q6, [x12, #0xa0]\n" + "ldr q6, [x15, #0xa0]\n" ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n" - "ldr q7, [x12, #0xb0]\n" + "ldr q7, [x15, #0xb0]\n" ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n" ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n" - "ldr q6, [x12, #0xc0]\n" + "ldr q6, [x15, #0xc0]\n" ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n" - "ldr q7, [x12, #0xd0]\n" + "ldr q7, [x15, #0xd0]\n" ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n" ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n" - "ldr q6, [x12, #0xe0]\n" + "ldr q6, [x15, #0xe0]\n" ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n" ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n" - "ldr q7, [x12, #0xf0]\n" - "add x12, x12, #0x100\n" + "ldr q7, [x15, #0xf0]\n" ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n" + "add x15, x15, #0x100\n" ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n" ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n" ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n" "36:" // Height 2: Multiply loop: Main loop skip - "cbz x10, 41f\n" - "cmp x10, #0x4\n" + "cbz x13, 41f\n" + "cmp x13, #0x4\n" "blt 38f\n" "37:" // Height 2: Multiply loop: Odd block loop - "ldr s0, [x9], #0x4\n" - "sub x10, x10, #0x4\n" - "ldr s1, [x26], #0x4\n" - "cmp x10, #0x4\n" - "ldr q6, [x12, #0x0]\n" - "ldr q7, [x12, #0x10]\n" + "ldr s0, [x12], #0x4\n" + "sub x13, x13, #0x4\n" + "ldr s1, [x28], #0x4\n" + "cmp x13, #0x4\n" + "ldr q6, [x15, #0x0]\n" + "ldr q7, [x15, #0x10]\n" ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" + "ldr q6, [x15, #0x20]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" - "ldr q6, [x12, #0x20]\n" - "ldr q7, [x12, #0x30]\n" - "add x12, x12, #0x40\n" + "ldr q7, [x15, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "add x15, x15, #0x40\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" "bge 37b\n" - "cbz x10, 41f\n" + "cbz x13, 41f\n" "38:" // Height 2: Multiply loop: Skip odd blocks - "tbz x10, #1, 39f\n" - "ldr h0, [x9], #0x2\n" - "ldr h1, [x26], #0x2\n" - "tbz x10, #0, 40f\n" - "ld1 { v0.b }[2], [x9]\n" - "ld1 { v1.b }[2], [x26]\n" + "tbz x13, #1, 39f\n" + "ldr h0, [x12], #0x2\n" + "ldr h1, [x28], #0x2\n" + "tbz x13, #0, 40f\n" + "ld1 { v0.b }[2], [x12]\n" + "ld1 { v1.b }[2], [x28]\n" "b 40f\n" "39:" // Height 2: Multiply loop: Ragged operand read: partial_1_0 - "ldr b0, [x9, #0x0]\n" - "ldr b1, [x26, #0x0]\n" + "ldr b0, [x12, #0x0]\n" + "ldr b1, [x28, #0x0]\n" "40:" // Height 2: Multiply loop: Ragged operand read: Done - "ldr q6, [x12, #0x0]\n" - "ldr q7, [x12, #0x10]\n" + "ldr q6, [x15, #0x0]\n" + "ldr q7, [x15, #0x10]\n" ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" + "ldr q6, [x15, #0x20]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" - "ldr q6, [x12, #0x20]\n" - "ldr q7, [x12, #0x30]\n" - "add x12, x12, #0x40\n" + "ldr q7, [x15, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "add x15, x15, #0x40\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" "41:" // Height 2: Multiply loop: No odd multiplies "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" - "add x11, x11, #0x1\n" - "cmp x11, x19\n" + "add x14, x14, #0x1\n" + "cmp x14, x19\n" "bne 31b\n" "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "prfm pstl1keep, [x13, #0x0]\n" - "ldr q0, [x15, #0x0]\n" - "ldr q1, [x15, #0x10]\n" - "add x23, x13, x19\n" + "prfm pstl1keep, [x16, #0x0]\n" + "ldr q0, [x8, #0x0]\n" + "ldr q1, [x8, #0x10]\n" + "add x23, x16, x19\n" "prfm pstl1keep, [x23, #0x0]\n" "add v8.4s, v8.4s, v0.4s\n" "add v12.4s, v12.4s, v0.4s\n" "add v9.4s, v9.4s, v1.4s\n" "add v13.4s, v13.4s, v1.4s\n" - "ldr q2, [x15, #0x20]\n" - "ldr q3, [x15, #0x30]\n" - "add x15, x15, #0x40\n" + "ldr q2, [x8, #0x20]\n" + "ldr q3, [x8, #0x30]\n" + "add x8, x8, #0x40\n" "add v10.4s, v10.4s, v2.4s\n" "add v14.4s, v14.4s, v2.4s\n" "add v11.4s, v11.4s, v3.4s\n" "add v15.4s, v15.4s, v3.4s\n" "tbz %x[flags], #4, 42f\n" - "ldr q0, [x16, #0x0]\n" - "ldr q4, [x17, #0x0]\n" - "ldr q1, [x16, #0x10]\n" - "ldr q5, [x17, #0x10]\n" - "ldr q2, [x16, #0x20]\n" - "ldr q6, [x17, #0x20]\n" - "ldr q3, [x16, #0x30]\n" - "add x16, x16, #0x40\n" - "ldr q7, [x17, #0x30]\n" - "add x17, x17, #0x40\n" + "ldr q0, [x7, #0x0]\n" + "ldr q4, [x6, #0x0]\n" + "ldr q1, [x7, #0x10]\n" + "ldr q5, [x6, #0x10]\n" + "ldr q2, [x7, #0x20]\n" + "ldr q6, [x6, #0x20]\n" + "ldr q3, [x7, #0x30]\n" + "add x7, x7, #0x40\n" + "ldr q7, [x6, #0x30]\n" + "add x6, x6, #0x40\n" "b 43f\n" "42:" // Height 2: per layer parameters "add x25, %x[qp], %[per_layer_right_shift]\n" @@ -746,7 +746,7 @@ void a64_hybrid_s8qs_dot_6x16_a55 ( "ld1r { v4.4s }, [x24]\n" "add x24, %x[qp], %[maxval]\n" "ld1r { v5.4s }, [x25]\n" - "cmp x14, #0x10\n" + "cmp x17, #0x10\n" "ld1r { v6.4s }, [x24]\n" "add v8.4s, v8.4s, v4.4s\n" "add v9.4s, v9.4s, v4.4s\n" @@ -779,81 +779,81 @@ void a64_hybrid_s8qs_dot_6x16_a55 ( "uzp1 v8.16b, v8.16b, v9.16b\n" "uzp1 v12.16b, v12.16b, v13.16b\n" "bge 53f\n" - "tbz x14, #3, 48f\n" - "str d8, [x13], #0x8\n" + "tbz x17, #3, 48f\n" + "str d8, [x16], #0x8\n" "str d12, [x23], #0x8\n" - "tbz x14, #2, 46f\n" - "st1 { v8.s }[2], [x13], #0x4\n" + "tbz x17, #2, 46f\n" + "st1 { v8.s }[2], [x16], #0x4\n" "st1 { v12.s }[2], [x23], #0x4\n" - "tbz x14, #1, 45f\n" - "st1 { v8.h }[6], [x13], #0x2\n" + "tbz x17, #1, 45f\n" + "st1 { v8.h }[6], [x16], #0x2\n" "st1 { v12.h }[6], [x23], #0x2\n" - "tbz x14, #0, 52f\n" - "st1 { v8.b }[14], [x13]\n" + "tbz x17, #0, 52f\n" + "st1 { v8.b }[14], [x16]\n" "st1 { v12.b }[14], [x23]\n" "b 52f\n" "45:" // Height 2: Partial direct writeback: partial_1_12 - "tbz x14, #0, 52f\n" - "st1 { v8.b }[12], [x13]\n" + "tbz x17, #0, 52f\n" + "st1 { v8.b }[12], [x16]\n" "st1 { v12.b }[12], [x23]\n" "b 52f\n" "46:" // Height 2: Partial direct writeback: partial_2_8 - "tbz x14, #1, 47f\n" - "st1 { v8.h }[4], [x13], #0x2\n" + "tbz x17, #1, 47f\n" + "st1 { v8.h }[4], [x16], #0x2\n" "st1 { v12.h }[4], [x23], #0x2\n" - "tbz x14, #0, 52f\n" - "st1 { v8.b }[10], [x13]\n" + "tbz x17, #0, 52f\n" + "st1 { v8.b }[10], [x16]\n" "st1 { v12.b }[10], [x23]\n" "b 52f\n" "47:" // Height 2: Partial direct writeback: partial_1_8 - "tbz x14, #0, 52f\n" - "st1 { v8.b }[8], [x13]\n" + "tbz x17, #0, 52f\n" + "st1 { v8.b }[8], [x16]\n" "st1 { v12.b }[8], [x23]\n" "b 52f\n" "48:" // Height 2: Partial direct writeback: partial_4_0 - "tbz x14, #2, 50f\n" - "str s8, [x13], #0x4\n" + "tbz x17, #2, 50f\n" + "str s8, [x16], #0x4\n" "str s12, [x23], #0x4\n" - "tbz x14, #1, 49f\n" - "st1 { v8.h }[2], [x13], #0x2\n" + "tbz x17, #1, 49f\n" + "st1 { v8.h }[2], [x16], #0x2\n" "st1 { v12.h }[2], [x23], #0x2\n" - "tbz x14, #0, 52f\n" - "st1 { v8.b }[6], [x13]\n" + "tbz x17, #0, 52f\n" + "st1 { v8.b }[6], [x16]\n" "st1 { v12.b }[6], [x23]\n" "b 52f\n" "49:" // Height 2: Partial direct writeback: partial_1_4 - "tbz x14, #0, 52f\n" - "st1 { v8.b }[4], [x13]\n" + "tbz x17, #0, 52f\n" + "st1 { v8.b }[4], [x16]\n" "st1 { v12.b }[4], [x23]\n" "b 52f\n" "50:" // Height 2: Partial direct writeback: partial_2_0 - "tbz x14, #1, 51f\n" - "str h8, [x13], #0x2\n" + "tbz x17, #1, 51f\n" + "str h8, [x16], #0x2\n" "str h12, [x23], #0x2\n" - "tbz x14, #0, 52f\n" - "st1 { v8.b }[2], [x13]\n" + "tbz x17, #0, 52f\n" + "st1 { v8.b }[2], [x16]\n" "st1 { v12.b }[2], [x23]\n" "b 52f\n" "51:" // Height 2: Partial direct writeback: partial_1_0 - "str b8, [x13, #0x0]\n" + "str b8, [x16, #0x0]\n" "str b12, [x23, #0x0]\n" "52:" // Height 2: Partial direct writeback: Done "b 54f\n" "53:" // Height 2: Full writeback - "str q8, [x13, #0x0]\n" - "add x13, x13, #0x10\n" + "str q8, [x16, #0x0]\n" + "add x16, x16, #0x10\n" "str q12, [x23, #0x0]\n" "54:" // Height 2: Writeback done - "subs x14, x14, #0x10\n" + "subs x17, x17, #0x10\n" "bgt 29b\n" "b 164f\n" "55:" // Height 3 - "ldr x17, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n" - "mov x15, %x[col_bias]\n" - "ldr x16, [%x[args_ptr], %[offsetof_shift_ptr]]\n" - "mov x13, %x[output_ptr]\n" - "ldr x14, [%x[args_ptr], %[offsetof_N]]\n" - "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "ldr x6, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n" + "mov x8, %x[col_bias]\n" + "ldr x7, [%x[args_ptr], %[offsetof_shift_ptr]]\n" + "mov x16, %x[output_ptr]\n" + "ldr x17, [%x[args_ptr], %[offsetof_N]]\n" + "ldr x15, [%x[args_ptr], %[offsetof_B_ptr]]\n" "56:" // Height 3: Column loop "movi v8.4s, #0x0\n" "movi v9.4s, #0x0\n" @@ -868,280 +868,280 @@ void a64_hybrid_s8qs_dot_6x16_a55 ( "movi v18.4s, #0x0\n" "movi v19.4s, #0x0\n" "57:" // Height 3: setup done - "mov x11, #0x0\n" + "mov x14, #0x0\n" "58:" // Height 3: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w10, [x20, x11, LSL #0x2]\n" + "ldr w13, [x20, x14, LSL #0x2]\n" "tbz %x[flags], #3, 59f\n" - "ldr x20, [%x[input_ptr], x11, LSL #0x3]\n" + "ldr x20, [%x[input_ptr], x14, LSL #0x3]\n" "add x20, x20, x19, LSL #3\n" - "ldr x9, [x20, #0x0]\n" - "ldr x26, [x20, #0x8]\n" - "ldr x24, [x20, #0x10]\n" - "cbnz x11, 60f\n" + "ldr x12, [x20, #0x0]\n" + "ldr x28, [x20, #0x8]\n" + "ldr x26, [x20, #0x10]\n" + "cbnz x14, 60f\n" "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x9, x9, x19\n" + "add x12, x12, x19\n" + "add x28, x28, x19\n" "add x26, x26, x19\n" - "add x24, x24, x19\n" "b 60f\n" "59:" // Height 3: setup direct input - "mov x9, %x[input_ptr]\n" - "add x26, x9, x19\n" - "add x24, x26, x19\n" + "mov x12, %x[input_ptr]\n" + "add x28, x12, x19\n" + "add x26, x28, x19\n" "60:" // Height 3: input setup done - "cmp x10, #0x10\n" + "cmp x13, #0x10\n" "blt 63f\n" - "ldr q0, [x9, #0x0]\n" - "ldr q1, [x26, #0x0]\n" - "cmp x10, #0x20\n" - "ldr q2, [x24, #0x0]\n" - "ldr q6, [x12, #0x0]\n" + "ldr q0, [x12, #0x0]\n" + "ldr q1, [x28, #0x0]\n" + "cmp x13, #0x20\n" + "ldr q2, [x26, #0x0]\n" + "ldr q6, [x15, #0x0]\n" "blt 62f\n" "61:" // Height 3: Multiply loop: Main loop head ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" + "ldr d7, [x15, #0x10]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" + "ldr x11, [x15, #0x18]\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" - "ldr d7, [x12, #0x10]\n" - "ldr x19, [x12, #0x18]\n" - "add x9, x9, #0x10\n" - "ldr d6, [x12, #0x20]\n" - "add x26, x26, #0x10\n" - "ldr x28, [x12, #0x28]\n" - "add x24, x24, #0x10\n" - "mov v7.d[1], x19\n" - "prfm pldl1keep, [x9, #0x80]\n" - "ldr x19, [x12, #0x38]\n" - "sub x10, x10, #0x10\n" - "mov v6.d[1], x28\n" + "ldr d6, [x15, #0x20]\n" + "ldr x10, [x15, #0x28]\n" + "add x12, x12, #0x10\n" + "mov v7.d[1], x11\n" + "prfm pldl1keep, [x12, #0x80]\n" + "ldr x11, [x15, #0x38]\n" + "add x28, x28, #0x10\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" + "mov v6.d[1], x10\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" + "prfm pldl1keep, [x28, #0x80]\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" + "ldr d7, [x15, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "ldr x10, [x15, #0x48]\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" + "ldr x9, [x12, #0x8]\n" ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" - "ldr d7, [x12, #0x30]\n" - "prfm pldl1keep, [x26, #0x80]\n" - "cmp x10, #0x20\n" - "ldr d6, [x12, #0x40]\n" - "mov v7.d[1], x19\n" - "ldr x28, [x12, #0x48]\n" - "ldr x19, [x12, #0x58]\n" + "mov v7.d[1], x11\n" + "ldr d6, [x15, #0x40]\n" + "add x26, x26, #0x10\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" + "prfm pldl1keep, [x26, #0x80]\n" ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" - "mov v6.d[1], x28\n" + "ldr x11, [x15, #0x58]\n" ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" - "ldr d7, [x12, #0x50]\n" + "mov v6.d[1], x10\n" + "ldr d7, [x15, #0x50]\n" + "sub x13, x13, #0x10\n" ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n" + "ldr x10, [x15, #0x68]\n" ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n" + "ldr x27, [x28, #0x8]\n" ".inst 0x4fa2e0d0 // sdot v16.4s, v6.16b, v2.4b[1]\n" - "prfm pldl1keep, [x24, #0x80]\n" - "mov v7.d[1], x19\n" - "ldr d6, [x12, #0x60]\n" - "ldr x28, [x12, #0x68]\n" + "mov v7.d[1], x11\n" + "ldr d6, [x15, #0x60]\n" + "cmp x13, #0x20\n" ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n" + "ldr x11, [x15, #0x78]\n" ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n" + "ldr x25, [x26, #0x8]\n" ".inst 0x4fa2e0f1 // sdot v17.4s, v7.16b, v2.4b[1]\n" - "ldr d7, [x12, #0x70]\n" - "mov v6.d[1], x28\n" - "ldr x19, [x12, #0x78]\n" - "ldr x28, [x12, #0x88]\n" + "mov v6.d[1], x10\n" + "ldr d7, [x15, #0x70]\n" ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n" + "ldr x10, [x15, #0x88]\n" ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n" - "mov v7.d[1], x19\n" ".inst 0x4fa2e0d2 // sdot v18.4s, v6.16b, v2.4b[1]\n" - "ldr d6, [x12, #0x80]\n" + "mov v7.d[1], x11\n" + "ldr d6, [x15, #0x80]\n" ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n" + "ldr x11, [x15, #0x98]\n" ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n" ".inst 0x4fa2e0f3 // sdot v19.4s, v7.16b, v2.4b[1]\n" - "ldr d7, [x12, #0x90]\n" - "mov v6.d[1], x28\n" - "ldr x19, [x12, #0x98]\n" - "ldr x28, [x12, #0xa8]\n" + "mov v6.d[1], x10\n" + "ldr d7, [x15, #0x90]\n" ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n" + "ldr x10, [x15, #0xa8]\n" ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n" - "mov v7.d[1], x19\n" ".inst 0x4f82e8d0 // sdot v16.4s, v6.16b, v2.4b[2]\n" - "ldr d6, [x12, #0xa0]\n" + "mov v7.d[1], x11\n" + "ldr d6, [x15, #0xa0]\n" ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n" + "ldr x11, [x15, #0xb8]\n" ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n" ".inst 0x4f82e8f1 // sdot v17.4s, v7.16b, v2.4b[2]\n" - "ldr d7, [x12, #0xb0]\n" - "mov v6.d[1], x28\n" - "ldr x19, [x12, #0xb8]\n" - "ldr x28, [x12, #0xc8]\n" + "mov v6.d[1], x10\n" + "ldr d7, [x15, #0xb0]\n" ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n" + "ldr x10, [x15, #0xc8]\n" ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n" - "mov v7.d[1], x19\n" ".inst 0x4f82e8d2 // sdot v18.4s, v6.16b, v2.4b[2]\n" - "ldr d6, [x12, #0xc0]\n" + "mov v7.d[1], x11\n" + "ldr d6, [x15, #0xc0]\n" ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n" + "ldr x11, [x15, #0xd8]\n" ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n" ".inst 0x4f82e8f3 // sdot v19.4s, v7.16b, v2.4b[2]\n" - "ldr d7, [x12, #0xd0]\n" - "mov v6.d[1], x28\n" - "ldr x19, [x12, #0xd8]\n" - "ldr x28, [x12, #0xe8]\n" + "mov v6.d[1], x10\n" + "ldr d7, [x15, #0xd0]\n" ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n" + "ldr x10, [x15, #0xe8]\n" ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n" - "mov v7.d[1], x19\n" ".inst 0x4fa2e8d0 // sdot v16.4s, v6.16b, v2.4b[3]\n" - "ldr d6, [x12, #0xe0]\n" + "mov v7.d[1], x11\n" + "ldr d6, [x15, #0xe0]\n" ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n" + "ldr x11, [x15, #0xf8]\n" ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n" ".inst 0x4fa2e8f1 // sdot v17.4s, v7.16b, v2.4b[3]\n" - "ldr d7, [x12, #0xf0]\n" - "mov v6.d[1], x28\n" - "ldr x19, [x12, #0xf8]\n" - "ldr x27, [x9, #0x8]\n" - "add x12, x12, #0x100\n" + "mov v6.d[1], x10\n" + "ldr d7, [x15, #0xf0]\n" + "add x15, x15, #0x100\n" ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n" + "ldr x10, [x15, #0x8]\n" ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n" - "mov v7.d[1], x19\n" ".inst 0x4fa2e8d2 // sdot v18.4s, v6.16b, v2.4b[3]\n" - "ldr x25, [x26, #0x8]\n" + "mov v7.d[1], x11\n" + "ldr d6, [x15, #0x0]\n" ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n" + "ldr d0, [x12, #0x0]\n" ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n" + "ldr d1, [x28, #0x0]\n" ".inst 0x4fa2e8f3 // sdot v19.4s, v7.16b, v2.4b[3]\n" - "ldr d0, [x9, #0x0]\n" - "ldr d1, [x26, #0x0]\n" - "ldr d2, [x24, #0x0]\n" - "mov v0.d[1], x27\n" - "ldr x21, [x24, #0x8]\n" - "mov v1.d[1], x25\n" - "ldr d6, [x12, #0x0]\n" - "ldr x28, [x12, #0x8]\n" - "mov v2.d[1], x21\n" - "mov v6.d[1], x28\n" + "mov v6.d[1], x10\n" + "mov v0.d[1], x9\n" + "ldr d2, [x26, #0x0]\n" + "mov v1.d[1], x27\n" + "mov v2.d[1], x25\n" "bge 61b\n" "62:" // Height 3: Multiply loop: Single iteration only ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" + "ldr q7, [x15, #0x10]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" + "sub x13, x13, #0x10\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" - "ldr q7, [x12, #0x10]\n" - "ldr q6, [x12, #0x20]\n" - "sub x10, x10, #0x10\n" - "add x9, x9, #0x10\n" - "add x26, x26, #0x10\n" + "ldr q6, [x15, #0x20]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" + "add x12, x12, #0x10\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" + "prfm pldl1keep, [x12, #0x80]\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" + "ldr q7, [x15, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "add x28, x28, #0x10\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" + "prfm pldl1keep, [x28, #0x80]\n" ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" - "prfm pldl1keep, [x9, #0x80]\n" - "add x24, x24, #0x10\n" - "ldr q7, [x12, #0x30]\n" - "prfm pldl1keep, [x26, #0x80]\n" - "ldr q6, [x12, #0x40]\n" + "ldr q6, [x15, #0x40]\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" + "add x26, x26, #0x10\n" ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" + "prfm pldl1keep, [x26, #0x80]\n" ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" - "ldr q7, [x12, #0x50]\n" + "ldr q7, [x15, #0x50]\n" ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n" ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n" ".inst 0x4fa2e0d0 // sdot v16.4s, v6.16b, v2.4b[1]\n" - "prfm pldl1keep, [x24, #0x80]\n" + "ldr q6, [x15, #0x60]\n" ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n" ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n" ".inst 0x4fa2e0f1 // sdot v17.4s, v7.16b, v2.4b[1]\n" - "ldr q6, [x12, #0x60]\n" - "ldr q7, [x12, #0x70]\n" + "ldr q7, [x15, #0x70]\n" ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n" ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n" ".inst 0x4fa2e0d2 // sdot v18.4s, v6.16b, v2.4b[1]\n" + "ldr q6, [x15, #0x80]\n" ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n" ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n" ".inst 0x4fa2e0f3 // sdot v19.4s, v7.16b, v2.4b[1]\n" - "ldr q6, [x12, #0x80]\n" - "ldr q7, [x12, #0x90]\n" + "ldr q7, [x15, #0x90]\n" ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n" ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n" ".inst 0x4f82e8d0 // sdot v16.4s, v6.16b, v2.4b[2]\n" + "ldr q6, [x15, #0xa0]\n" ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n" ".inst 0x4f82e8f1 // sdot v17.4s, v7.16b, v2.4b[2]\n" - "ldr q6, [x12, #0xa0]\n" - "ldr q7, [x12, #0xb0]\n" + "ldr q7, [x15, #0xb0]\n" ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n" ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n" ".inst 0x4f82e8d2 // sdot v18.4s, v6.16b, v2.4b[2]\n" + "ldr q6, [x15, #0xc0]\n" ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n" ".inst 0x4f82e8f3 // sdot v19.4s, v7.16b, v2.4b[2]\n" - "ldr q6, [x12, #0xc0]\n" - "ldr q7, [x12, #0xd0]\n" + "ldr q7, [x15, #0xd0]\n" ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n" ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n" ".inst 0x4fa2e8d0 // sdot v16.4s, v6.16b, v2.4b[3]\n" + "ldr q6, [x15, #0xe0]\n" ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n" ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n" ".inst 0x4fa2e8f1 // sdot v17.4s, v7.16b, v2.4b[3]\n" - "ldr q6, [x12, #0xe0]\n" - "ldr q7, [x12, #0xf0]\n" - "add x12, x12, #0x100\n" + "ldr q7, [x15, #0xf0]\n" ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n" + "add x15, x15, #0x100\n" ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n" ".inst 0x4fa2e8d2 // sdot v18.4s, v6.16b, v2.4b[3]\n" ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n" ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n" ".inst 0x4fa2e8f3 // sdot v19.4s, v7.16b, v2.4b[3]\n" "63:" // Height 3: Multiply loop: Main loop skip - "cbz x10, 68f\n" - "cmp x10, #0x4\n" + "cbz x13, 68f\n" + "cmp x13, #0x4\n" "blt 65f\n" "64:" // Height 3: Multiply loop: Odd block loop - "ldr s0, [x9], #0x4\n" - "sub x10, x10, #0x4\n" - "ldr s1, [x26], #0x4\n" - "cmp x10, #0x4\n" - "ldr s2, [x24], #0x4\n" - "ldr q6, [x12, #0x0]\n" - "ldr q7, [x12, #0x10]\n" + "ldr s0, [x12], #0x4\n" + "sub x13, x13, #0x4\n" + "ldr s1, [x28], #0x4\n" + "cmp x13, #0x4\n" + "ldr s2, [x26], #0x4\n" + "ldr q6, [x15, #0x0]\n" + "ldr q7, [x15, #0x10]\n" ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" + "ldr q6, [x15, #0x20]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" - "ldr q6, [x12, #0x20]\n" - "ldr q7, [x12, #0x30]\n" - "add x12, x12, #0x40\n" + "ldr q7, [x15, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "add x15, x15, #0x40\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" "bge 64b\n" - "cbz x10, 68f\n" + "cbz x13, 68f\n" "65:" // Height 3: Multiply loop: Skip odd blocks - "tbz x10, #1, 66f\n" - "ldr h0, [x9], #0x2\n" - "ldr h1, [x26], #0x2\n" - "ldr h2, [x24], #0x2\n" - "tbz x10, #0, 67f\n" - "ld1 { v0.b }[2], [x9]\n" - "ld1 { v1.b }[2], [x26]\n" - "ld1 { v2.b }[2], [x24]\n" + "tbz x13, #1, 66f\n" + "ldr h0, [x12], #0x2\n" + "ldr h1, [x28], #0x2\n" + "ldr h2, [x26], #0x2\n" + "tbz x13, #0, 67f\n" + "ld1 { v0.b }[2], [x12]\n" + "ld1 { v1.b }[2], [x28]\n" + "ld1 { v2.b }[2], [x26]\n" "b 67f\n" "66:" // Height 3: Multiply loop: Ragged operand read: partial_1_0 - "ldr b0, [x9, #0x0]\n" - "ldr b1, [x26, #0x0]\n" - "ldr b2, [x24, #0x0]\n" + "ldr b0, [x12, #0x0]\n" + "ldr b1, [x28, #0x0]\n" + "ldr b2, [x26, #0x0]\n" "67:" // Height 3: Multiply loop: Ragged operand read: Done - "ldr q6, [x12, #0x0]\n" - "ldr q7, [x12, #0x10]\n" + "ldr q6, [x15, #0x0]\n" + "ldr q7, [x15, #0x10]\n" ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" + "ldr q6, [x15, #0x20]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" - "ldr q6, [x12, #0x20]\n" - "ldr q7, [x12, #0x30]\n" - "add x12, x12, #0x40\n" + "ldr q7, [x15, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "add x15, x15, #0x40\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" @@ -1149,14 +1149,14 @@ void a64_hybrid_s8qs_dot_6x16_a55 ( ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" "68:" // Height 3: Multiply loop: No odd multiplies "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" - "add x11, x11, #0x1\n" - "cmp x11, x19\n" + "add x14, x14, #0x1\n" + "cmp x14, x19\n" "bne 58b\n" "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "prfm pstl1keep, [x13, #0x0]\n" - "ldr q0, [x15, #0x0]\n" - "ldr q1, [x15, #0x10]\n" - "add x23, x13, x19\n" + "prfm pstl1keep, [x16, #0x0]\n" + "ldr q0, [x8, #0x0]\n" + "ldr q1, [x8, #0x10]\n" + "add x23, x16, x19\n" "prfm pstl1keep, [x23, #0x0]\n" "add x22, x23, x19\n" "add v8.4s, v8.4s, v0.4s\n" @@ -1166,9 +1166,9 @@ void a64_hybrid_s8qs_dot_6x16_a55 ( "add v16.4s, v16.4s, v0.4s\n" "add v17.4s, v17.4s, v1.4s\n" "prfm pstl1keep, [x22, #0x0]\n" - "ldr q2, [x15, #0x20]\n" - "ldr q3, [x15, #0x30]\n" - "add x15, x15, #0x40\n" + "ldr q2, [x8, #0x20]\n" + "ldr q3, [x8, #0x30]\n" + "add x8, x8, #0x40\n" "add v10.4s, v10.4s, v2.4s\n" "add v14.4s, v14.4s, v2.4s\n" "add v11.4s, v11.4s, v3.4s\n" @@ -1176,16 +1176,16 @@ void a64_hybrid_s8qs_dot_6x16_a55 ( "add v18.4s, v18.4s, v2.4s\n" "add v19.4s, v19.4s, v3.4s\n" "tbz %x[flags], #4, 69f\n" - "ldr q0, [x16, #0x0]\n" - "ldr q4, [x17, #0x0]\n" - "ldr q1, [x16, #0x10]\n" - "ldr q5, [x17, #0x10]\n" - "ldr q2, [x16, #0x20]\n" - "ldr q6, [x17, #0x20]\n" - "ldr q3, [x16, #0x30]\n" - "add x16, x16, #0x40\n" - "ldr q7, [x17, #0x30]\n" - "add x17, x17, #0x40\n" + "ldr q0, [x7, #0x0]\n" + "ldr q4, [x6, #0x0]\n" + "ldr q1, [x7, #0x10]\n" + "ldr q5, [x6, #0x10]\n" + "ldr q2, [x7, #0x20]\n" + "ldr q6, [x6, #0x20]\n" + "ldr q3, [x7, #0x30]\n" + "add x7, x7, #0x40\n" + "ldr q7, [x6, #0x30]\n" + "add x6, x6, #0x40\n" "b 70f\n" "69:" // Height 3: per layer parameters "add x25, %x[qp], %[per_layer_right_shift]\n" @@ -1266,7 +1266,7 @@ void a64_hybrid_s8qs_dot_6x16_a55 ( "ld1r { v4.4s }, [x24]\n" "add x24, %x[qp], %[maxval]\n" "ld1r { v5.4s }, [x25]\n" - "cmp x14, #0x10\n" + "cmp x17, #0x10\n" "ld1r { v6.4s }, [x24]\n" "add v8.4s, v8.4s, v4.4s\n" "add v9.4s, v9.4s, v4.4s\n" @@ -1314,97 +1314,97 @@ void a64_hybrid_s8qs_dot_6x16_a55 ( "uzp1 v17.8h, v18.8h, v19.8h\n" "uzp1 v16.16b, v16.16b, v17.16b\n" "bge 80f\n" - "tbz x14, #3, 75f\n" - "str d8, [x13], #0x8\n" + "tbz x17, #3, 75f\n" + "str d8, [x16], #0x8\n" "str d12, [x23], #0x8\n" "str d16, [x22], #0x8\n" - "tbz x14, #2, 73f\n" - "st1 { v8.s }[2], [x13], #0x4\n" + "tbz x17, #2, 73f\n" + "st1 { v8.s }[2], [x16], #0x4\n" "st1 { v12.s }[2], [x23], #0x4\n" "st1 { v16.s }[2], [x22], #0x4\n" - "tbz x14, #1, 72f\n" - "st1 { v8.h }[6], [x13], #0x2\n" + "tbz x17, #1, 72f\n" + "st1 { v8.h }[6], [x16], #0x2\n" "st1 { v12.h }[6], [x23], #0x2\n" "st1 { v16.h }[6], [x22], #0x2\n" - "tbz x14, #0, 79f\n" - "st1 { v8.b }[14], [x13]\n" + "tbz x17, #0, 79f\n" + "st1 { v8.b }[14], [x16]\n" "st1 { v12.b }[14], [x23]\n" "st1 { v16.b }[14], [x22]\n" "b 79f\n" "72:" // Height 3: Partial direct writeback: partial_1_12 - "tbz x14, #0, 79f\n" - "st1 { v8.b }[12], [x13]\n" + "tbz x17, #0, 79f\n" + "st1 { v8.b }[12], [x16]\n" "st1 { v12.b }[12], [x23]\n" "st1 { v16.b }[12], [x22]\n" "b 79f\n" "73:" // Height 3: Partial direct writeback: partial_2_8 - "tbz x14, #1, 74f\n" - "st1 { v8.h }[4], [x13], #0x2\n" + "tbz x17, #1, 74f\n" + "st1 { v8.h }[4], [x16], #0x2\n" "st1 { v12.h }[4], [x23], #0x2\n" "st1 { v16.h }[4], [x22], #0x2\n" - "tbz x14, #0, 79f\n" - "st1 { v8.b }[10], [x13]\n" + "tbz x17, #0, 79f\n" + "st1 { v8.b }[10], [x16]\n" "st1 { v12.b }[10], [x23]\n" "st1 { v16.b }[10], [x22]\n" "b 79f\n" "74:" // Height 3: Partial direct writeback: partial_1_8 - "tbz x14, #0, 79f\n" - "st1 { v8.b }[8], [x13]\n" + "tbz x17, #0, 79f\n" + "st1 { v8.b }[8], [x16]\n" "st1 { v12.b }[8], [x23]\n" "st1 { v16.b }[8], [x22]\n" "b 79f\n" "75:" // Height 3: Partial direct writeback: partial_4_0 - "tbz x14, #2, 77f\n" - "str s8, [x13], #0x4\n" + "tbz x17, #2, 77f\n" + "str s8, [x16], #0x4\n" "str s12, [x23], #0x4\n" "str s16, [x22], #0x4\n" - "tbz x14, #1, 76f\n" - "st1 { v8.h }[2], [x13], #0x2\n" + "tbz x17, #1, 76f\n" + "st1 { v8.h }[2], [x16], #0x2\n" "st1 { v12.h }[2], [x23], #0x2\n" "st1 { v16.h }[2], [x22], #0x2\n" - "tbz x14, #0, 79f\n" - "st1 { v8.b }[6], [x13]\n" + "tbz x17, #0, 79f\n" + "st1 { v8.b }[6], [x16]\n" "st1 { v12.b }[6], [x23]\n" "st1 { v16.b }[6], [x22]\n" "b 79f\n" "76:" // Height 3: Partial direct writeback: partial_1_4 - "tbz x14, #0, 79f\n" - "st1 { v8.b }[4], [x13]\n" + "tbz x17, #0, 79f\n" + "st1 { v8.b }[4], [x16]\n" "st1 { v12.b }[4], [x23]\n" "st1 { v16.b }[4], [x22]\n" "b 79f\n" "77:" // Height 3: Partial direct writeback: partial_2_0 - "tbz x14, #1, 78f\n" - "str h8, [x13], #0x2\n" + "tbz x17, #1, 78f\n" + "str h8, [x16], #0x2\n" "str h12, [x23], #0x2\n" "str h16, [x22], #0x2\n" - "tbz x14, #0, 79f\n" - "st1 { v8.b }[2], [x13]\n" + "tbz x17, #0, 79f\n" + "st1 { v8.b }[2], [x16]\n" "st1 { v12.b }[2], [x23]\n" "st1 { v16.b }[2], [x22]\n" "b 79f\n" "78:" // Height 3: Partial direct writeback: partial_1_0 - "str b8, [x13, #0x0]\n" + "str b8, [x16, #0x0]\n" "str b12, [x23, #0x0]\n" "str b16, [x22, #0x0]\n" "79:" // Height 3: Partial direct writeback: Done "b 81f\n" "80:" // Height 3: Full writeback - "str q8, [x13, #0x0]\n" - "add x13, x13, #0x10\n" + "str q8, [x16, #0x0]\n" + "add x16, x16, #0x10\n" "str q12, [x23, #0x0]\n" "str q16, [x22, #0x0]\n" "81:" // Height 3: Writeback done - "subs x14, x14, #0x10\n" + "subs x17, x17, #0x10\n" "bgt 56b\n" "b 164f\n" "82:" // Height 4 - "ldr x17, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n" - "mov x15, %x[col_bias]\n" - "ldr x16, [%x[args_ptr], %[offsetof_shift_ptr]]\n" - "mov x13, %x[output_ptr]\n" - "ldr x14, [%x[args_ptr], %[offsetof_N]]\n" - "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "ldr x6, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n" + "mov x8, %x[col_bias]\n" + "ldr x7, [%x[args_ptr], %[offsetof_shift_ptr]]\n" + "mov x16, %x[output_ptr]\n" + "ldr x17, [%x[args_ptr], %[offsetof_N]]\n" + "ldr x15, [%x[args_ptr], %[offsetof_B_ptr]]\n" "83:" // Height 4: Column loop "movi v8.4s, #0x0\n" "movi v9.4s, #0x0\n" @@ -1423,292 +1423,292 @@ void a64_hybrid_s8qs_dot_6x16_a55 ( "movi v22.4s, #0x0\n" "movi v23.4s, #0x0\n" "84:" // Height 4: setup done - "mov x11, #0x0\n" + "mov x14, #0x0\n" "85:" // Height 4: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w10, [x20, x11, LSL #0x2]\n" + "ldr w13, [x20, x14, LSL #0x2]\n" "tbz %x[flags], #3, 86f\n" - "ldr x20, [%x[input_ptr], x11, LSL #0x3]\n" + "ldr x20, [%x[input_ptr], x14, LSL #0x3]\n" "add x20, x20, x19, LSL #3\n" - "ldr x9, [x20, #0x0]\n" - "ldr x26, [x20, #0x8]\n" - "ldr x24, [x20, #0x10]\n" - "ldr x23, [x20, #0x18]\n" - "cbnz x11, 87f\n" + "ldr x12, [x20, #0x0]\n" + "ldr x28, [x20, #0x8]\n" + "ldr x26, [x20, #0x10]\n" + "ldr x24, [x20, #0x18]\n" + "cbnz x14, 87f\n" "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x9, x9, x19\n" + "add x12, x12, x19\n" + "add x28, x28, x19\n" "add x26, x26, x19\n" "add x24, x24, x19\n" - "add x23, x23, x19\n" "b 87f\n" "86:" // Height 4: setup direct input - "mov x9, %x[input_ptr]\n" - "add x26, x9, x19\n" + "mov x12, %x[input_ptr]\n" + "add x28, x12, x19\n" + "add x26, x28, x19\n" "add x24, x26, x19\n" - "add x23, x24, x19\n" "87:" // Height 4: input setup done - "cmp x10, #0x10\n" + "cmp x13, #0x10\n" "blt 90f\n" - "ldr q0, [x9, #0x0]\n" - "ldr q1, [x26, #0x0]\n" - "cmp x10, #0x20\n" - "ldr q2, [x24, #0x0]\n" - "ldr q3, [x23, #0x0]\n" - "ldr q6, [x12, #0x0]\n" + "ldr q0, [x12, #0x0]\n" + "ldr q1, [x28, #0x0]\n" + "cmp x13, #0x20\n" + "ldr q2, [x26, #0x0]\n" + "ldr q3, [x24, #0x0]\n" + "ldr q6, [x15, #0x0]\n" "blt 89f\n" "88:" // Height 4: Multiply loop: Main loop head ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" + "ldr d7, [x15, #0x10]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" + "ldr x11, [x15, #0x18]\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" + "ldr x10, [x15, #0x28]\n" ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n" - "ldr d7, [x12, #0x10]\n" - "add x9, x9, #0x10\n" - "ldr x19, [x12, #0x18]\n" - "add x26, x26, #0x10\n" - "prfm pldl1keep, [x9, #0x80]\n" - "add x24, x24, #0x10\n" - "ldr d6, [x12, #0x20]\n" - "add x23, x23, #0x10\n" - "mov v7.d[1], x19\n" - "ldr x28, [x12, #0x28]\n" - "ldr x19, [x12, #0x38]\n" - "sub x10, x10, #0x10\n" + "ldr d6, [x15, #0x20]\n" + "mov v7.d[1], x11\n" + "ldr x11, [x15, #0x38]\n" + "add x12, x12, #0x10\n" + "add x28, x28, #0x10\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" + "mov v6.d[1], x10\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" - "mov v6.d[1], x28\n" + "prfm pldl1keep, [x12, #0x80]\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" + "prfm pldl1keep, [x28, #0x80]\n" ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n" - "ldr d7, [x12, #0x30]\n" + "ldr d7, [x15, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "ldr x10, [x15, #0x48]\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" + "ldr x9, [x12, #0x8]\n" ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" + "mov v7.d[1], x11\n" ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n" - "mov v7.d[1], x19\n" - "prfm pldl1keep, [x26, #0x80]\n" - "ldr d6, [x12, #0x40]\n" - "cmp x10, #0x20\n" + "ldr d6, [x15, #0x40]\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" + "ldr x11, [x15, #0x58]\n" ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" + "ldr x27, [x28, #0x8]\n" ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" + "mov v6.d[1], x10\n" ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n" - "ldr x28, [x12, #0x48]\n" - "ldr d7, [x12, #0x50]\n" - "ldr x19, [x12, #0x58]\n" - "mov v6.d[1], x28\n" - "prfm pldl1keep, [x24, #0x80]\n" - "ldr x28, [x12, #0x68]\n" + "ldr d7, [x15, #0x50]\n" ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n" + "ldr x10, [x15, #0x68]\n" ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n" + "add x26, x26, #0x10\n" ".inst 0x4fa2e0d0 // sdot v16.4s, v6.16b, v2.4b[1]\n" + "mov v7.d[1], x11\n" ".inst 0x4fa3e0d4 // sdot v20.4s, v6.16b, v3.4b[1]\n" - "mov v7.d[1], x19\n" - "ldr d6, [x12, #0x60]\n" - "ldr x19, [x12, #0x78]\n" + "prfm pldl1keep, [x26, #0x80]\n" ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n" + "ldr d6, [x15, #0x60]\n" ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n" - "mov v6.d[1], x28\n" + "ldr x11, [x15, #0x78]\n" ".inst 0x4fa2e0f1 // sdot v17.4s, v7.16b, v2.4b[1]\n" + "ldr x25, [x26, #0x8]\n" ".inst 0x4fa3e0f5 // sdot v21.4s, v7.16b, v3.4b[1]\n" - "ldr d7, [x12, #0x70]\n" + "mov v6.d[1], x10\n" + "ldr d7, [x15, #0x70]\n" + "add x24, x24, #0x10\n" ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n" + "prfm pldl1keep, [x24, #0x80]\n" ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n" + "ldr x10, [x15, #0x88]\n" ".inst 0x4fa2e0d2 // sdot v18.4s, v6.16b, v2.4b[1]\n" + "mov v7.d[1], x11\n" ".inst 0x4fa3e0d6 // sdot v22.4s, v6.16b, v3.4b[1]\n" - "mov v7.d[1], x19\n" - "prfm pldl1keep, [x23, #0x80]\n" - "ldr d6, [x12, #0x80]\n" + "ldr d6, [x15, #0x80]\n" ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n" + "ldr x11, [x15, #0x98]\n" ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n" + "ldr x23, [x24, #0x8]\n" ".inst 0x4fa2e0f3 // sdot v19.4s, v7.16b, v2.4b[1]\n" + "mov v6.d[1], x10\n" ".inst 0x4fa3e0f7 // sdot v23.4s, v7.16b, v3.4b[1]\n" - "ldr x28, [x12, #0x88]\n" - "ldr d7, [x12, #0x90]\n" - "ldr x19, [x12, #0x98]\n" - "mov v6.d[1], x28\n" - "ldr x28, [x12, #0xa8]\n" - "ldr x27, [x9, #0x8]\n" + "ldr d7, [x15, #0x90]\n" ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n" + "ldr x10, [x15, #0xa8]\n" ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n" + "sub x13, x13, #0x10\n" ".inst 0x4f82e8d0 // sdot v16.4s, v6.16b, v2.4b[2]\n" + "mov v7.d[1], x11\n" ".inst 0x4f83e8d4 // sdot v20.4s, v6.16b, v3.4b[2]\n" - "mov v7.d[1], x19\n" - "ldr d6, [x12, #0xa0]\n" - "ldr x19, [x12, #0xb8]\n" + "ldr d6, [x15, #0xa0]\n" ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n" + "ldr x11, [x15, #0xb8]\n" ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n" - "mov v6.d[1], x28\n" + "cmp x13, #0x20\n" ".inst 0x4f82e8f1 // sdot v17.4s, v7.16b, v2.4b[2]\n" + "mov v6.d[1], x10\n" ".inst 0x4f83e8f5 // sdot v21.4s, v7.16b, v3.4b[2]\n" - "ldr d7, [x12, #0xb0]\n" + "ldr d7, [x15, #0xb0]\n" ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n" + "ldr x10, [x15, #0xc8]\n" ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n" ".inst 0x4f82e8d2 // sdot v18.4s, v6.16b, v2.4b[2]\n" + "mov v7.d[1], x11\n" ".inst 0x4f83e8d6 // sdot v22.4s, v6.16b, v3.4b[2]\n" - "mov v7.d[1], x19\n" - "ldr d6, [x12, #0xc0]\n" - "ldr x28, [x12, #0xc8]\n" + "ldr d6, [x15, #0xc0]\n" ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n" + "ldr x11, [x15, #0xd8]\n" ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n" ".inst 0x4f82e8f3 // sdot v19.4s, v7.16b, v2.4b[2]\n" + "mov v6.d[1], x10\n" ".inst 0x4f83e8f7 // sdot v23.4s, v7.16b, v3.4b[2]\n" - "mov v6.d[1], x28\n" - "ldr d7, [x12, #0xd0]\n" - "ldr x19, [x12, #0xd8]\n" + "ldr d7, [x15, #0xd0]\n" ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n" + "ldr x10, [x15, #0xe8]\n" ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n" ".inst 0x4fa2e8d0 // sdot v16.4s, v6.16b, v2.4b[3]\n" + "mov v7.d[1], x11\n" ".inst 0x4fa3e8d4 // sdot v20.4s, v6.16b, v3.4b[3]\n" - "mov v7.d[1], x19\n" - "ldr d6, [x12, #0xe0]\n" - "ldr x28, [x12, #0xe8]\n" + "ldr d6, [x15, #0xe0]\n" ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n" + "ldr x11, [x15, #0xf8]\n" ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n" ".inst 0x4fa2e8f1 // sdot v17.4s, v7.16b, v2.4b[3]\n" + "mov v6.d[1], x10\n" ".inst 0x4fa3e8f5 // sdot v21.4s, v7.16b, v3.4b[3]\n" - "mov v6.d[1], x28\n" - "ldr d7, [x12, #0xf0]\n" - "ldr x19, [x12, #0xf8]\n" - "add x12, x12, #0x100\n" + "ldr d7, [x15, #0xf0]\n" ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n" + "add x15, x15, #0x100\n" ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n" + "ldr x10, [x15, #0x8]\n" ".inst 0x4fa2e8d2 // sdot v18.4s, v6.16b, v2.4b[3]\n" + "mov v7.d[1], x11\n" ".inst 0x4fa3e8d6 // sdot v22.4s, v6.16b, v3.4b[3]\n" - "mov v7.d[1], x19\n" - "ldr d6, [x12, #0x0]\n" - "ldr x28, [x12, #0x8]\n" + "ldr d6, [x15, #0x0]\n" ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n" + "ldr d0, [x12, #0x0]\n" ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n" + "ldr d1, [x28, #0x0]\n" ".inst 0x4fa2e8f3 // sdot v19.4s, v7.16b, v2.4b[3]\n" + "mov v6.d[1], x10\n" ".inst 0x4fa3e8f7 // sdot v23.4s, v7.16b, v3.4b[3]\n" - "mov v6.d[1], x28\n" - "ldr d0, [x9, #0x0]\n" - "ldr d1, [x26, #0x0]\n" - "ldr x25, [x26, #0x8]\n" - "mov v0.d[1], x27\n" - "ldr d2, [x24, #0x0]\n" - "ldr x21, [x24, #0x8]\n" - "mov v1.d[1], x25\n" - "ldr d3, [x23, #0x0]\n" - "ldr x19, [x23, #0x8]\n" - "mov v2.d[1], x21\n" - "mov v3.d[1], x19\n" + "mov v0.d[1], x9\n" + "mov v1.d[1], x27\n" + "ldr d2, [x26, #0x0]\n" + "ldr d3, [x24, #0x0]\n" + "mov v2.d[1], x25\n" + "mov v3.d[1], x23\n" "bge 88b\n" "89:" // Height 4: Multiply loop: Single iteration only ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" + "ldr q7, [x15, #0x10]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" + "sub x13, x13, #0x10\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" + "add x12, x12, #0x10\n" ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n" - "ldr q7, [x12, #0x10]\n" - "sub x10, x10, #0x10\n" - "ldr q6, [x12, #0x20]\n" - "add x9, x9, #0x10\n" - "prfm pldl1keep, [x9, #0x80]\n" - "add x26, x26, #0x10\n" + "prfm pldl1keep, [x12, #0x80]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" + "ldr q6, [x15, #0x20]\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" + "add x28, x28, #0x10\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" + "prfm pldl1keep, [x28, #0x80]\n" ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n" + "ldr q7, [x15, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "add x26, x26, #0x10\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" + "prfm pldl1keep, [x26, #0x80]\n" ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" - ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n" - "ldr q7, [x12, #0x30]\n" "add x24, x24, #0x10\n" - "prfm pldl1keep, [x26, #0x80]\n" - "add x23, x23, #0x10\n" - "ldr q6, [x12, #0x40]\n" + ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n" + "prfm pldl1keep, [x24, #0x80]\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" + "ldr q6, [x15, #0x40]\n" ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n" + "ldr q7, [x15, #0x50]\n" ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n" ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n" ".inst 0x4fa2e0d0 // sdot v16.4s, v6.16b, v2.4b[1]\n" ".inst 0x4fa3e0d4 // sdot v20.4s, v6.16b, v3.4b[1]\n" - "ldr q7, [x12, #0x50]\n" - "prfm pldl1keep, [x24, #0x80]\n" - "ldr q6, [x12, #0x60]\n" + "ldr q6, [x15, #0x60]\n" ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n" ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n" ".inst 0x4fa2e0f1 // sdot v17.4s, v7.16b, v2.4b[1]\n" ".inst 0x4fa3e0f5 // sdot v21.4s, v7.16b, v3.4b[1]\n" + "ldr q7, [x15, #0x70]\n" ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n" ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n" ".inst 0x4fa2e0d2 // sdot v18.4s, v6.16b, v2.4b[1]\n" ".inst 0x4fa3e0d6 // sdot v22.4s, v6.16b, v3.4b[1]\n" - "ldr q7, [x12, #0x70]\n" - "prfm pldl1keep, [x23, #0x80]\n" - "ldr q6, [x12, #0x80]\n" + "ldr q6, [x15, #0x80]\n" ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n" ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n" ".inst 0x4fa2e0f3 // sdot v19.4s, v7.16b, v2.4b[1]\n" ".inst 0x4fa3e0f7 // sdot v23.4s, v7.16b, v3.4b[1]\n" + "ldr q7, [x15, #0x90]\n" ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n" ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n" ".inst 0x4f82e8d0 // sdot v16.4s, v6.16b, v2.4b[2]\n" ".inst 0x4f83e8d4 // sdot v20.4s, v6.16b, v3.4b[2]\n" - "ldr q7, [x12, #0x90]\n" - "ldr q6, [x12, #0xa0]\n" + "ldr q6, [x15, #0xa0]\n" ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n" ".inst 0x4f82e8f1 // sdot v17.4s, v7.16b, v2.4b[2]\n" ".inst 0x4f83e8f5 // sdot v21.4s, v7.16b, v3.4b[2]\n" + "ldr q7, [x15, #0xb0]\n" ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n" ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n" ".inst 0x4f82e8d2 // sdot v18.4s, v6.16b, v2.4b[2]\n" ".inst 0x4f83e8d6 // sdot v22.4s, v6.16b, v3.4b[2]\n" - "ldr q7, [x12, #0xb0]\n" - "ldr q6, [x12, #0xc0]\n" + "ldr q6, [x15, #0xc0]\n" ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n" ".inst 0x4f82e8f3 // sdot v19.4s, v7.16b, v2.4b[2]\n" ".inst 0x4f83e8f7 // sdot v23.4s, v7.16b, v3.4b[2]\n" + "ldr q7, [x15, #0xd0]\n" ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n" ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n" ".inst 0x4fa2e8d0 // sdot v16.4s, v6.16b, v2.4b[3]\n" ".inst 0x4fa3e8d4 // sdot v20.4s, v6.16b, v3.4b[3]\n" - "ldr q7, [x12, #0xd0]\n" - "ldr q6, [x12, #0xe0]\n" + "ldr q6, [x15, #0xe0]\n" ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n" ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n" ".inst 0x4fa2e8f1 // sdot v17.4s, v7.16b, v2.4b[3]\n" ".inst 0x4fa3e8f5 // sdot v21.4s, v7.16b, v3.4b[3]\n" + "ldr q7, [x15, #0xf0]\n" ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n" + "add x15, x15, #0x100\n" ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n" ".inst 0x4fa2e8d2 // sdot v18.4s, v6.16b, v2.4b[3]\n" ".inst 0x4fa3e8d6 // sdot v22.4s, v6.16b, v3.4b[3]\n" - "ldr q7, [x12, #0xf0]\n" - "add x12, x12, #0x100\n" ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n" ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n" ".inst 0x4fa2e8f3 // sdot v19.4s, v7.16b, v2.4b[3]\n" ".inst 0x4fa3e8f7 // sdot v23.4s, v7.16b, v3.4b[3]\n" "90:" // Height 4: Multiply loop: Main loop skip - "cbz x10, 95f\n" - "cmp x10, #0x4\n" + "cbz x13, 95f\n" + "cmp x13, #0x4\n" "blt 92f\n" "91:" // Height 4: Multiply loop: Odd block loop - "ldr s0, [x9], #0x4\n" - "sub x10, x10, #0x4\n" - "ldr s1, [x26], #0x4\n" - "cmp x10, #0x4\n" - "ldr s2, [x24], #0x4\n" - "ldr s3, [x23], #0x4\n" - "ldr q6, [x12, #0x0]\n" - "ldr q7, [x12, #0x10]\n" + "ldr s0, [x12], #0x4\n" + "sub x13, x13, #0x4\n" + "ldr s1, [x28], #0x4\n" + "cmp x13, #0x4\n" + "ldr s2, [x26], #0x4\n" + "ldr s3, [x24], #0x4\n" + "ldr q6, [x15, #0x0]\n" + "ldr q7, [x15, #0x10]\n" ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n" + "ldr q6, [x15, #0x20]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n" - "ldr q6, [x12, #0x20]\n" - "ldr q7, [x12, #0x30]\n" - "add x12, x12, #0x40\n" + "ldr q7, [x15, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "add x15, x15, #0x40\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n" @@ -1717,39 +1717,39 @@ void a64_hybrid_s8qs_dot_6x16_a55 ( ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n" "bge 91b\n" - "cbz x10, 95f\n" + "cbz x13, 95f\n" "92:" // Height 4: Multiply loop: Skip odd blocks - "tbz x10, #1, 93f\n" - "ldr h0, [x9], #0x2\n" - "ldr h1, [x26], #0x2\n" - "ldr h2, [x24], #0x2\n" - "ldr h3, [x23], #0x2\n" - "tbz x10, #0, 94f\n" - "ld1 { v0.b }[2], [x9]\n" - "ld1 { v1.b }[2], [x26]\n" - "ld1 { v2.b }[2], [x24]\n" - "ld1 { v3.b }[2], [x23]\n" + "tbz x13, #1, 93f\n" + "ldr h0, [x12], #0x2\n" + "ldr h1, [x28], #0x2\n" + "ldr h2, [x26], #0x2\n" + "ldr h3, [x24], #0x2\n" + "tbz x13, #0, 94f\n" + "ld1 { v0.b }[2], [x12]\n" + "ld1 { v1.b }[2], [x28]\n" + "ld1 { v2.b }[2], [x26]\n" + "ld1 { v3.b }[2], [x24]\n" "b 94f\n" "93:" // Height 4: Multiply loop: Ragged operand read: partial_1_0 - "ldr b0, [x9, #0x0]\n" - "ldr b1, [x26, #0x0]\n" - "ldr b2, [x24, #0x0]\n" - "ldr b3, [x23, #0x0]\n" + "ldr b0, [x12, #0x0]\n" + "ldr b1, [x28, #0x0]\n" + "ldr b2, [x26, #0x0]\n" + "ldr b3, [x24, #0x0]\n" "94:" // Height 4: Multiply loop: Ragged operand read: Done - "ldr q6, [x12, #0x0]\n" - "ldr q7, [x12, #0x10]\n" + "ldr q6, [x15, #0x0]\n" + "ldr q7, [x15, #0x10]\n" ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n" + "ldr q6, [x15, #0x20]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n" - "ldr q6, [x12, #0x20]\n" - "ldr q7, [x12, #0x30]\n" - "add x12, x12, #0x40\n" + "ldr q7, [x15, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "add x15, x15, #0x40\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n" @@ -1759,14 +1759,14 @@ void a64_hybrid_s8qs_dot_6x16_a55 ( ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n" "95:" // Height 4: Multiply loop: No odd multiplies "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" - "add x11, x11, #0x1\n" - "cmp x11, x19\n" + "add x14, x14, #0x1\n" + "cmp x14, x19\n" "bne 85b\n" "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "prfm pstl1keep, [x13, #0x0]\n" - "ldr q0, [x15, #0x0]\n" - "ldr q1, [x15, #0x10]\n" - "add x23, x13, x19\n" + "prfm pstl1keep, [x16, #0x0]\n" + "ldr q0, [x8, #0x0]\n" + "ldr q1, [x8, #0x10]\n" + "add x23, x16, x19\n" "prfm pstl1keep, [x23, #0x0]\n" "add x22, x23, x19\n" "add v8.4s, v8.4s, v0.4s\n" @@ -1780,9 +1780,9 @@ void a64_hybrid_s8qs_dot_6x16_a55 ( "prfm pstl1keep, [x22, #0x0]\n" "add x21, x22, x19\n" "prfm pstl1keep, [x21, #0x0]\n" - "ldr q2, [x15, #0x20]\n" - "ldr q3, [x15, #0x30]\n" - "add x15, x15, #0x40\n" + "ldr q2, [x8, #0x20]\n" + "ldr q3, [x8, #0x30]\n" + "add x8, x8, #0x40\n" "add v10.4s, v10.4s, v2.4s\n" "add v14.4s, v14.4s, v2.4s\n" "add v11.4s, v11.4s, v3.4s\n" @@ -1792,16 +1792,16 @@ void a64_hybrid_s8qs_dot_6x16_a55 ( "add v22.4s, v22.4s, v2.4s\n" "add v23.4s, v23.4s, v3.4s\n" "tbz %x[flags], #4, 96f\n" - "ldr q0, [x16, #0x0]\n" - "ldr q4, [x17, #0x0]\n" - "ldr q1, [x16, #0x10]\n" - "ldr q5, [x17, #0x10]\n" - "ldr q2, [x16, #0x20]\n" - "ldr q6, [x17, #0x20]\n" - "ldr q3, [x16, #0x30]\n" - "add x16, x16, #0x40\n" - "ldr q7, [x17, #0x30]\n" - "add x17, x17, #0x40\n" + "ldr q0, [x7, #0x0]\n" + "ldr q4, [x6, #0x0]\n" + "ldr q1, [x7, #0x10]\n" + "ldr q5, [x6, #0x10]\n" + "ldr q2, [x7, #0x20]\n" + "ldr q6, [x6, #0x20]\n" + "ldr q3, [x7, #0x30]\n" + "add x7, x7, #0x40\n" + "ldr q7, [x6, #0x30]\n" + "add x6, x6, #0x40\n" "b 97f\n" "96:" // Height 4: per layer parameters "add x25, %x[qp], %[per_layer_right_shift]\n" @@ -1902,7 +1902,7 @@ void a64_hybrid_s8qs_dot_6x16_a55 ( "ld1r { v4.4s }, [x24]\n" "add x24, %x[qp], %[maxval]\n" "ld1r { v5.4s }, [x25]\n" - "cmp x14, #0x10\n" + "cmp x17, #0x10\n" "ld1r { v6.4s }, [x24]\n" "add v8.4s, v8.4s, v4.4s\n" "add v9.4s, v9.4s, v4.4s\n" @@ -1965,113 +1965,113 @@ void a64_hybrid_s8qs_dot_6x16_a55 ( "uzp1 v16.16b, v16.16b, v17.16b\n" "uzp1 v20.16b, v20.16b, v21.16b\n" "bge 107f\n" - "tbz x14, #3, 102f\n" - "str d8, [x13], #0x8\n" + "tbz x17, #3, 102f\n" + "str d8, [x16], #0x8\n" "str d12, [x23], #0x8\n" "str d16, [x22], #0x8\n" "str d20, [x21], #0x8\n" - "tbz x14, #2, 100f\n" - "st1 { v8.s }[2], [x13], #0x4\n" + "tbz x17, #2, 100f\n" + "st1 { v8.s }[2], [x16], #0x4\n" "st1 { v12.s }[2], [x23], #0x4\n" "st1 { v16.s }[2], [x22], #0x4\n" "st1 { v20.s }[2], [x21], #0x4\n" - "tbz x14, #1, 99f\n" - "st1 { v8.h }[6], [x13], #0x2\n" + "tbz x17, #1, 99f\n" + "st1 { v8.h }[6], [x16], #0x2\n" "st1 { v12.h }[6], [x23], #0x2\n" "st1 { v16.h }[6], [x22], #0x2\n" "st1 { v20.h }[6], [x21], #0x2\n" - "tbz x14, #0, 106f\n" - "st1 { v8.b }[14], [x13]\n" + "tbz x17, #0, 106f\n" + "st1 { v8.b }[14], [x16]\n" "st1 { v12.b }[14], [x23]\n" "st1 { v16.b }[14], [x22]\n" "st1 { v20.b }[14], [x21]\n" "b 106f\n" "99:" // Height 4: Partial direct writeback: partial_1_12 - "tbz x14, #0, 106f\n" - "st1 { v8.b }[12], [x13]\n" + "tbz x17, #0, 106f\n" + "st1 { v8.b }[12], [x16]\n" "st1 { v12.b }[12], [x23]\n" "st1 { v16.b }[12], [x22]\n" "st1 { v20.b }[12], [x21]\n" "b 106f\n" "100:" // Height 4: Partial direct writeback: partial_2_8 - "tbz x14, #1, 101f\n" - "st1 { v8.h }[4], [x13], #0x2\n" + "tbz x17, #1, 101f\n" + "st1 { v8.h }[4], [x16], #0x2\n" "st1 { v12.h }[4], [x23], #0x2\n" "st1 { v16.h }[4], [x22], #0x2\n" "st1 { v20.h }[4], [x21], #0x2\n" - "tbz x14, #0, 106f\n" - "st1 { v8.b }[10], [x13]\n" + "tbz x17, #0, 106f\n" + "st1 { v8.b }[10], [x16]\n" "st1 { v12.b }[10], [x23]\n" "st1 { v16.b }[10], [x22]\n" "st1 { v20.b }[10], [x21]\n" "b 106f\n" "101:" // Height 4: Partial direct writeback: partial_1_8 - "tbz x14, #0, 106f\n" - "st1 { v8.b }[8], [x13]\n" + "tbz x17, #0, 106f\n" + "st1 { v8.b }[8], [x16]\n" "st1 { v12.b }[8], [x23]\n" "st1 { v16.b }[8], [x22]\n" "st1 { v20.b }[8], [x21]\n" "b 106f\n" "102:" // Height 4: Partial direct writeback: partial_4_0 - "tbz x14, #2, 104f\n" - "str s8, [x13], #0x4\n" + "tbz x17, #2, 104f\n" + "str s8, [x16], #0x4\n" "str s12, [x23], #0x4\n" "str s16, [x22], #0x4\n" "str s20, [x21], #0x4\n" - "tbz x14, #1, 103f\n" - "st1 { v8.h }[2], [x13], #0x2\n" + "tbz x17, #1, 103f\n" + "st1 { v8.h }[2], [x16], #0x2\n" "st1 { v12.h }[2], [x23], #0x2\n" "st1 { v16.h }[2], [x22], #0x2\n" "st1 { v20.h }[2], [x21], #0x2\n" - "tbz x14, #0, 106f\n" - "st1 { v8.b }[6], [x13]\n" + "tbz x17, #0, 106f\n" + "st1 { v8.b }[6], [x16]\n" "st1 { v12.b }[6], [x23]\n" "st1 { v16.b }[6], [x22]\n" "st1 { v20.b }[6], [x21]\n" "b 106f\n" "103:" // Height 4: Partial direct writeback: partial_1_4 - "tbz x14, #0, 106f\n" - "st1 { v8.b }[4], [x13]\n" + "tbz x17, #0, 106f\n" + "st1 { v8.b }[4], [x16]\n" "st1 { v12.b }[4], [x23]\n" "st1 { v16.b }[4], [x22]\n" "st1 { v20.b }[4], [x21]\n" "b 106f\n" "104:" // Height 4: Partial direct writeback: partial_2_0 - "tbz x14, #1, 105f\n" - "str h8, [x13], #0x2\n" + "tbz x17, #1, 105f\n" + "str h8, [x16], #0x2\n" "str h12, [x23], #0x2\n" "str h16, [x22], #0x2\n" "str h20, [x21], #0x2\n" - "tbz x14, #0, 106f\n" - "st1 { v8.b }[2], [x13]\n" + "tbz x17, #0, 106f\n" + "st1 { v8.b }[2], [x16]\n" "st1 { v12.b }[2], [x23]\n" "st1 { v16.b }[2], [x22]\n" "st1 { v20.b }[2], [x21]\n" "b 106f\n" "105:" // Height 4: Partial direct writeback: partial_1_0 - "str b8, [x13, #0x0]\n" + "str b8, [x16, #0x0]\n" "str b12, [x23, #0x0]\n" "str b16, [x22, #0x0]\n" "str b20, [x21, #0x0]\n" "106:" // Height 4: Partial direct writeback: Done "b 108f\n" "107:" // Height 4: Full writeback - "str q8, [x13, #0x0]\n" - "add x13, x13, #0x10\n" + "str q8, [x16, #0x0]\n" + "add x16, x16, #0x10\n" "str q12, [x23, #0x0]\n" "str q16, [x22, #0x0]\n" "str q20, [x21, #0x0]\n" "108:" // Height 4: Writeback done - "subs x14, x14, #0x10\n" + "subs x17, x17, #0x10\n" "bgt 83b\n" "b 164f\n" "109:" // Height 5 - "ldr x17, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n" - "mov x15, %x[col_bias]\n" - "ldr x16, [%x[args_ptr], %[offsetof_shift_ptr]]\n" - "mov x13, %x[output_ptr]\n" - "ldr x14, [%x[args_ptr], %[offsetof_N]]\n" - "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "ldr x6, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n" + "mov x8, %x[col_bias]\n" + "ldr x7, [%x[args_ptr], %[offsetof_shift_ptr]]\n" + "mov x16, %x[output_ptr]\n" + "ldr x17, [%x[args_ptr], %[offsetof_N]]\n" + "ldr x15, [%x[args_ptr], %[offsetof_B_ptr]]\n" "110:" // Height 5: Column loop "movi v8.4s, #0x0\n" "movi v9.4s, #0x0\n" @@ -2094,338 +2094,338 @@ void a64_hybrid_s8qs_dot_6x16_a55 ( "movi v26.4s, #0x0\n" "movi v27.4s, #0x0\n" "111:" // Height 5: setup done - "mov x11, #0x0\n" + "mov x14, #0x0\n" "112:" // Height 5: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w10, [x20, x11, LSL #0x2]\n" + "ldr w13, [x20, x14, LSL #0x2]\n" "tbz %x[flags], #3, 113f\n" - "ldr x20, [%x[input_ptr], x11, LSL #0x3]\n" + "ldr x20, [%x[input_ptr], x14, LSL #0x3]\n" "add x20, x20, x19, LSL #3\n" - "ldr x9, [x20, #0x0]\n" - "ldr x26, [x20, #0x8]\n" - "ldr x24, [x20, #0x10]\n" - "ldr x23, [x20, #0x18]\n" + "ldr x12, [x20, #0x0]\n" + "ldr x28, [x20, #0x8]\n" + "ldr x26, [x20, #0x10]\n" + "ldr x24, [x20, #0x18]\n" "ldr x22, [x20, #0x20]\n" - "cbnz x11, 114f\n" + "cbnz x14, 114f\n" "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x9, x9, x19\n" + "add x12, x12, x19\n" + "add x28, x28, x19\n" "add x26, x26, x19\n" "add x24, x24, x19\n" - "add x23, x23, x19\n" "add x22, x22, x19\n" "b 114f\n" "113:" // Height 5: setup direct input - "mov x9, %x[input_ptr]\n" - "add x26, x9, x19\n" + "mov x12, %x[input_ptr]\n" + "add x28, x12, x19\n" + "add x26, x28, x19\n" "add x24, x26, x19\n" - "add x23, x24, x19\n" - "add x22, x23, x19\n" + "add x22, x24, x19\n" "114:" // Height 5: input setup done - "cmp x10, #0x10\n" + "cmp x13, #0x10\n" "blt 117f\n" - "ldr q0, [x9, #0x0]\n" - "ldr q1, [x26, #0x0]\n" - "cmp x10, #0x20\n" - "ldr q2, [x24, #0x0]\n" - "ldr q3, [x23, #0x0]\n" + "ldr q0, [x12, #0x0]\n" + "ldr q1, [x28, #0x0]\n" + "cmp x13, #0x20\n" + "ldr q2, [x26, #0x0]\n" + "ldr q3, [x24, #0x0]\n" "ldr q4, [x22, #0x0]\n" - "ldr q6, [x12, #0x0]\n" + "ldr q6, [x15, #0x0]\n" "blt 116f\n" "115:" // Height 5: Multiply loop: Main loop head ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" + "ldr d7, [x15, #0x10]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" + "ldr x11, [x15, #0x18]\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" + "ldr x10, [x15, #0x28]\n" ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n" + "add x12, x12, #0x10\n" ".inst 0x4f84e0d8 // sdot v24.4s, v6.16b, v4.4b[0]\n" - "ldr d7, [x12, #0x10]\n" - "ldr x19, [x12, #0x18]\n" - "add x9, x9, #0x10\n" - "ldr d6, [x12, #0x20]\n" - "add x26, x26, #0x10\n" - "ldr x28, [x12, #0x28]\n" - "add x24, x24, #0x10\n" - "mov v7.d[1], x19\n" - "prfm pldl1keep, [x9, #0x80]\n" - "ldr x19, [x12, #0x38]\n" - "add x23, x23, #0x10\n" - "mov v6.d[1], x28\n" + "mov v7.d[1], x11\n" + "prfm pldl1keep, [x12, #0x80]\n" + "add x28, x28, #0x10\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" + "prfm pldl1keep, [x28, #0x80]\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" + "ldr d6, [x15, #0x20]\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" + "ldr x11, [x15, #0x38]\n" ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n" + "ldr x9, [x12, #0x8]\n" ".inst 0x4f84e0f9 // sdot v25.4s, v7.16b, v4.4b[0]\n" + "mov v6.d[1], x10\n" + "ldr d7, [x15, #0x30]\n" + "add x26, x26, #0x10\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "prfm pldl1keep, [x26, #0x80]\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" + "ldr x10, [x15, #0x48]\n" ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" + "mov v7.d[1], x11\n" ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n" + "ldr x11, [x15, #0x58]\n" ".inst 0x4f84e0da // sdot v26.4s, v6.16b, v4.4b[0]\n" - "ldr d7, [x12, #0x30]\n" - "prfm pldl1keep, [x26, #0x80]\n" - "add x22, x22, #0x10\n" - "ldr d6, [x12, #0x40]\n" - "sub x10, x10, #0x10\n" - "mov v7.d[1], x19\n" - "ldr x28, [x12, #0x48]\n" - "ldr x19, [x12, #0x58]\n" - "cmp x10, #0x20\n" + "ldr d6, [x15, #0x40]\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" + "ldr x27, [x28, #0x8]\n" ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" - "mov v6.d[1], x28\n" + "ldr x25, [x26, #0x8]\n" ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" + "mov v6.d[1], x10\n" ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n" + "ldr x10, [x15, #0x68]\n" ".inst 0x4f84e0fb // sdot v27.4s, v7.16b, v4.4b[0]\n" + "ldr d7, [x15, #0x50]\n" ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n" + "add x24, x24, #0x10\n" ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n" + "prfm pldl1keep, [x24, #0x80]\n" ".inst 0x4fa2e0d0 // sdot v16.4s, v6.16b, v2.4b[1]\n" + "mov v7.d[1], x11\n" ".inst 0x4fa3e0d4 // sdot v20.4s, v6.16b, v3.4b[1]\n" + "ldr x11, [x15, #0x78]\n" ".inst 0x4fa4e0d8 // sdot v24.4s, v6.16b, v4.4b[1]\n" - "ldr d7, [x12, #0x50]\n" - "prfm pldl1keep, [x24, #0x80]\n" - "ldr d6, [x12, #0x60]\n" - "mov v7.d[1], x19\n" - "ldr x28, [x12, #0x68]\n" - "ldr x19, [x12, #0x78]\n" + "ldr d6, [x15, #0x60]\n" ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n" + "ldr x23, [x24, #0x8]\n" ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n" - "mov v6.d[1], x28\n" + "add x22, x22, #0x10\n" ".inst 0x4fa2e0f1 // sdot v17.4s, v7.16b, v2.4b[1]\n" + "mov v6.d[1], x10\n" ".inst 0x4fa3e0f5 // sdot v21.4s, v7.16b, v3.4b[1]\n" + "prfm pldl1keep, [x22, #0x80]\n" ".inst 0x4fa4e0f9 // sdot v25.4s, v7.16b, v4.4b[1]\n" + "ldr d7, [x15, #0x70]\n" ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n" + "ldr x10, [x15, #0x88]\n" ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n" + "ldr x21, [x22, #0x8]\n" ".inst 0x4fa2e0d2 // sdot v18.4s, v6.16b, v2.4b[1]\n" + "mov v7.d[1], x11\n" ".inst 0x4fa3e0d6 // sdot v22.4s, v6.16b, v3.4b[1]\n" + "ldr x11, [x15, #0x98]\n" ".inst 0x4fa4e0da // sdot v26.4s, v6.16b, v4.4b[1]\n" - "ldr d7, [x12, #0x70]\n" - "prfm pldl1keep, [x23, #0x80]\n" - "ldr d6, [x12, #0x80]\n" - "mov v7.d[1], x19\n" - "ldr x28, [x12, #0x88]\n" - "ldr x19, [x12, #0x98]\n" + "ldr d6, [x15, #0x80]\n" ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n" + "sub x13, x13, #0x10\n" ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n" - "mov v6.d[1], x28\n" + "cmp x13, #0x20\n" ".inst 0x4fa2e0f3 // sdot v19.4s, v7.16b, v2.4b[1]\n" + "mov v6.d[1], x10\n" ".inst 0x4fa3e0f7 // sdot v23.4s, v7.16b, v3.4b[1]\n" + "ldr x10, [x15, #0xa8]\n" ".inst 0x4fa4e0fb // sdot v27.4s, v7.16b, v4.4b[1]\n" + "ldr d7, [x15, #0x90]\n" ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n" ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n" ".inst 0x4f82e8d0 // sdot v16.4s, v6.16b, v2.4b[2]\n" + "mov v7.d[1], x11\n" ".inst 0x4f83e8d4 // sdot v20.4s, v6.16b, v3.4b[2]\n" + "ldr x11, [x15, #0xb8]\n" ".inst 0x4f84e8d8 // sdot v24.4s, v6.16b, v4.4b[2]\n" - "ldr d7, [x12, #0x90]\n" - "prfm pldl1keep, [x22, #0x80]\n" - "ldr d6, [x12, #0xa0]\n" - "mov v7.d[1], x19\n" - "ldr x28, [x12, #0xa8]\n" - "ldr x19, [x12, #0xb8]\n" + "ldr d6, [x15, #0xa0]\n" ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n" - "mov v6.d[1], x28\n" ".inst 0x4f82e8f1 // sdot v17.4s, v7.16b, v2.4b[2]\n" + "mov v6.d[1], x10\n" ".inst 0x4f83e8f5 // sdot v21.4s, v7.16b, v3.4b[2]\n" + "ldr x10, [x15, #0xc8]\n" ".inst 0x4f84e8f9 // sdot v25.4s, v7.16b, v4.4b[2]\n" + "ldr d7, [x15, #0xb0]\n" ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n" ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n" ".inst 0x4f82e8d2 // sdot v18.4s, v6.16b, v2.4b[2]\n" + "mov v7.d[1], x11\n" ".inst 0x4f83e8d6 // sdot v22.4s, v6.16b, v3.4b[2]\n" + "ldr x11, [x15, #0xd8]\n" ".inst 0x4f84e8da // sdot v26.4s, v6.16b, v4.4b[2]\n" - "ldr d7, [x12, #0xb0]\n" - "ldr d6, [x12, #0xc0]\n" - "ldr x28, [x12, #0xc8]\n" - "mov v7.d[1], x19\n" - "ldr x19, [x12, #0xd8]\n" - "ldr x27, [x9, #0x8]\n" - "mov v6.d[1], x28\n" + "ldr d6, [x15, #0xc0]\n" ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n" ".inst 0x4f82e8f3 // sdot v19.4s, v7.16b, v2.4b[2]\n" + "mov v6.d[1], x10\n" ".inst 0x4f83e8f7 // sdot v23.4s, v7.16b, v3.4b[2]\n" + "ldr x10, [x15, #0xe8]\n" ".inst 0x4f84e8fb // sdot v27.4s, v7.16b, v4.4b[2]\n" + "ldr d7, [x15, #0xd0]\n" ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n" ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n" ".inst 0x4fa2e8d0 // sdot v16.4s, v6.16b, v2.4b[3]\n" + "mov v7.d[1], x11\n" ".inst 0x4fa3e8d4 // sdot v20.4s, v6.16b, v3.4b[3]\n" + "ldr x11, [x15, #0xf8]\n" ".inst 0x4fa4e8d8 // sdot v24.4s, v6.16b, v4.4b[3]\n" - "ldr d7, [x12, #0xd0]\n" - "ldr d6, [x12, #0xe0]\n" - "ldr x28, [x12, #0xe8]\n" - "mov v7.d[1], x19\n" - "ldr x19, [x12, #0xf8]\n" - "ldr x25, [x26, #0x8]\n" - "mov v6.d[1], x28\n" + "ldr d6, [x15, #0xe0]\n" ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n" ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n" ".inst 0x4fa2e8f1 // sdot v17.4s, v7.16b, v2.4b[3]\n" + "mov v6.d[1], x10\n" ".inst 0x4fa3e8f5 // sdot v21.4s, v7.16b, v3.4b[3]\n" ".inst 0x4fa4e8f9 // sdot v25.4s, v7.16b, v4.4b[3]\n" + "ldr d7, [x15, #0xf0]\n" ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n" + "add x15, x15, #0x100\n" ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n" + "ldr x10, [x15, #0x8]\n" ".inst 0x4fa2e8d2 // sdot v18.4s, v6.16b, v2.4b[3]\n" + "mov v7.d[1], x11\n" ".inst 0x4fa3e8d6 // sdot v22.4s, v6.16b, v3.4b[3]\n" ".inst 0x4fa4e8da // sdot v26.4s, v6.16b, v4.4b[3]\n" - "ldr d7, [x12, #0xf0]\n" - "ldr x21, [x24, #0x8]\n" - "add x12, x12, #0x100\n" - "ldr d6, [x12, #0x0]\n" - "mov v7.d[1], x19\n" - "ldr x28, [x12, #0x8]\n" - "ldr x19, [x23, #0x8]\n" + "ldr d6, [x15, #0x0]\n" ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n" + "ldr d0, [x12, #0x0]\n" ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n" - "mov v6.d[1], x28\n" + "ldr d1, [x28, #0x0]\n" ".inst 0x4fa2e8f3 // sdot v19.4s, v7.16b, v2.4b[3]\n" + "mov v6.d[1], x10\n" ".inst 0x4fa3e8f7 // sdot v23.4s, v7.16b, v3.4b[3]\n" + "mov v0.d[1], x9\n" ".inst 0x4fa4e8fb // sdot v27.4s, v7.16b, v4.4b[3]\n" - "ldr d0, [x9, #0x0]\n" - "ldr d1, [x26, #0x0]\n" - "ldr d2, [x24, #0x0]\n" - "mov v0.d[1], x27\n" - "ldr d3, [x23, #0x0]\n" - "mov v1.d[1], x25\n" + "mov v1.d[1], x27\n" + "ldr d2, [x26, #0x0]\n" + "ldr d3, [x24, #0x0]\n" "ldr d4, [x22, #0x0]\n" - "mov v2.d[1], x21\n" - "ldr x21, [x22, #0x8]\n" - "mov v3.d[1], x19\n" + "mov v2.d[1], x25\n" + "mov v3.d[1], x23\n" "mov v4.d[1], x21\n" "bge 115b\n" "116:" // Height 5: Multiply loop: Single iteration only ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" + "ldr q7, [x15, #0x10]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" + "sub x13, x13, #0x10\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" + "add x12, x12, #0x10\n" ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n" + "prfm pldl1keep, [x12, #0x80]\n" ".inst 0x4f84e0d8 // sdot v24.4s, v6.16b, v4.4b[0]\n" - "ldr q7, [x12, #0x10]\n" - "ldr q6, [x12, #0x20]\n" - "sub x10, x10, #0x10\n" - "add x9, x9, #0x10\n" - "add x26, x26, #0x10\n" + "ldr q6, [x15, #0x20]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" + "add x28, x28, #0x10\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" + "prfm pldl1keep, [x28, #0x80]\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" + "add x26, x26, #0x10\n" ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n" + "prfm pldl1keep, [x26, #0x80]\n" ".inst 0x4f84e0f9 // sdot v25.4s, v7.16b, v4.4b[0]\n" + "ldr q7, [x15, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "add x24, x24, #0x10\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" + "prfm pldl1keep, [x24, #0x80]\n" ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" + "add x22, x22, #0x10\n" ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n" + "prfm pldl1keep, [x22, #0x80]\n" ".inst 0x4f84e0da // sdot v26.4s, v6.16b, v4.4b[0]\n" - "prfm pldl1keep, [x9, #0x80]\n" - "add x24, x24, #0x10\n" - "ldr q7, [x12, #0x30]\n" - "add x23, x23, #0x10\n" - "prfm pldl1keep, [x26, #0x80]\n" - "add x22, x22, #0x10\n" - "ldr q6, [x12, #0x40]\n" + "ldr q6, [x15, #0x40]\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n" ".inst 0x4f84e0fb // sdot v27.4s, v7.16b, v4.4b[0]\n" + "ldr q7, [x15, #0x50]\n" ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n" ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n" ".inst 0x4fa2e0d0 // sdot v16.4s, v6.16b, v2.4b[1]\n" ".inst 0x4fa3e0d4 // sdot v20.4s, v6.16b, v3.4b[1]\n" ".inst 0x4fa4e0d8 // sdot v24.4s, v6.16b, v4.4b[1]\n" - "ldr q7, [x12, #0x50]\n" - "prfm pldl1keep, [x24, #0x80]\n" - "ldr q6, [x12, #0x60]\n" + "ldr q6, [x15, #0x60]\n" ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n" ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n" ".inst 0x4fa2e0f1 // sdot v17.4s, v7.16b, v2.4b[1]\n" ".inst 0x4fa3e0f5 // sdot v21.4s, v7.16b, v3.4b[1]\n" ".inst 0x4fa4e0f9 // sdot v25.4s, v7.16b, v4.4b[1]\n" + "ldr q7, [x15, #0x70]\n" ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n" ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n" ".inst 0x4fa2e0d2 // sdot v18.4s, v6.16b, v2.4b[1]\n" ".inst 0x4fa3e0d6 // sdot v22.4s, v6.16b, v3.4b[1]\n" ".inst 0x4fa4e0da // sdot v26.4s, v6.16b, v4.4b[1]\n" - "ldr q7, [x12, #0x70]\n" - "prfm pldl1keep, [x23, #0x80]\n" - "ldr q6, [x12, #0x80]\n" + "ldr q6, [x15, #0x80]\n" ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n" ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n" ".inst 0x4fa2e0f3 // sdot v19.4s, v7.16b, v2.4b[1]\n" ".inst 0x4fa3e0f7 // sdot v23.4s, v7.16b, v3.4b[1]\n" ".inst 0x4fa4e0fb // sdot v27.4s, v7.16b, v4.4b[1]\n" + "ldr q7, [x15, #0x90]\n" ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n" ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n" ".inst 0x4f82e8d0 // sdot v16.4s, v6.16b, v2.4b[2]\n" ".inst 0x4f83e8d4 // sdot v20.4s, v6.16b, v3.4b[2]\n" ".inst 0x4f84e8d8 // sdot v24.4s, v6.16b, v4.4b[2]\n" - "ldr q7, [x12, #0x90]\n" - "prfm pldl1keep, [x22, #0x80]\n" - "ldr q6, [x12, #0xa0]\n" + "ldr q6, [x15, #0xa0]\n" ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n" ".inst 0x4f82e8f1 // sdot v17.4s, v7.16b, v2.4b[2]\n" ".inst 0x4f83e8f5 // sdot v21.4s, v7.16b, v3.4b[2]\n" ".inst 0x4f84e8f9 // sdot v25.4s, v7.16b, v4.4b[2]\n" + "ldr q7, [x15, #0xb0]\n" ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n" ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n" ".inst 0x4f82e8d2 // sdot v18.4s, v6.16b, v2.4b[2]\n" ".inst 0x4f83e8d6 // sdot v22.4s, v6.16b, v3.4b[2]\n" ".inst 0x4f84e8da // sdot v26.4s, v6.16b, v4.4b[2]\n" - "ldr q7, [x12, #0xb0]\n" - "ldr q6, [x12, #0xc0]\n" + "ldr q6, [x15, #0xc0]\n" ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n" ".inst 0x4f82e8f3 // sdot v19.4s, v7.16b, v2.4b[2]\n" ".inst 0x4f83e8f7 // sdot v23.4s, v7.16b, v3.4b[2]\n" ".inst 0x4f84e8fb // sdot v27.4s, v7.16b, v4.4b[2]\n" + "ldr q7, [x15, #0xd0]\n" ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n" ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n" ".inst 0x4fa2e8d0 // sdot v16.4s, v6.16b, v2.4b[3]\n" ".inst 0x4fa3e8d4 // sdot v20.4s, v6.16b, v3.4b[3]\n" ".inst 0x4fa4e8d8 // sdot v24.4s, v6.16b, v4.4b[3]\n" - "ldr q7, [x12, #0xd0]\n" - "ldr q6, [x12, #0xe0]\n" + "ldr q6, [x15, #0xe0]\n" ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n" ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n" ".inst 0x4fa2e8f1 // sdot v17.4s, v7.16b, v2.4b[3]\n" ".inst 0x4fa3e8f5 // sdot v21.4s, v7.16b, v3.4b[3]\n" ".inst 0x4fa4e8f9 // sdot v25.4s, v7.16b, v4.4b[3]\n" + "ldr q7, [x15, #0xf0]\n" ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n" + "add x15, x15, #0x100\n" ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n" ".inst 0x4fa2e8d2 // sdot v18.4s, v6.16b, v2.4b[3]\n" ".inst 0x4fa3e8d6 // sdot v22.4s, v6.16b, v3.4b[3]\n" ".inst 0x4fa4e8da // sdot v26.4s, v6.16b, v4.4b[3]\n" - "ldr q7, [x12, #0xf0]\n" - "add x12, x12, #0x100\n" ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n" ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n" ".inst 0x4fa2e8f3 // sdot v19.4s, v7.16b, v2.4b[3]\n" ".inst 0x4fa3e8f7 // sdot v23.4s, v7.16b, v3.4b[3]\n" ".inst 0x4fa4e8fb // sdot v27.4s, v7.16b, v4.4b[3]\n" "117:" // Height 5: Multiply loop: Main loop skip - "cbz x10, 122f\n" - "cmp x10, #0x4\n" + "cbz x13, 122f\n" + "cmp x13, #0x4\n" "blt 119f\n" "118:" // Height 5: Multiply loop: Odd block loop - "ldr s0, [x9], #0x4\n" - "sub x10, x10, #0x4\n" - "ldr s1, [x26], #0x4\n" - "cmp x10, #0x4\n" - "ldr s2, [x24], #0x4\n" - "ldr s3, [x23], #0x4\n" + "ldr s0, [x12], #0x4\n" + "sub x13, x13, #0x4\n" + "ldr s1, [x28], #0x4\n" + "cmp x13, #0x4\n" + "ldr s2, [x26], #0x4\n" + "ldr s3, [x24], #0x4\n" "ldr s4, [x22], #0x4\n" - "ldr q6, [x12, #0x0]\n" - "ldr q7, [x12, #0x10]\n" + "ldr q6, [x15, #0x0]\n" + "ldr q7, [x15, #0x10]\n" ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n" ".inst 0x4f84e0d8 // sdot v24.4s, v6.16b, v4.4b[0]\n" + "ldr q6, [x15, #0x20]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n" ".inst 0x4f84e0f9 // sdot v25.4s, v7.16b, v4.4b[0]\n" - "ldr q6, [x12, #0x20]\n" - "ldr q7, [x12, #0x30]\n" - "add x12, x12, #0x40\n" + "ldr q7, [x15, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "add x15, x15, #0x40\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n" @@ -2436,44 +2436,44 @@ void a64_hybrid_s8qs_dot_6x16_a55 ( ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n" ".inst 0x4f84e0fb // sdot v27.4s, v7.16b, v4.4b[0]\n" "bge 118b\n" - "cbz x10, 122f\n" + "cbz x13, 122f\n" "119:" // Height 5: Multiply loop: Skip odd blocks - "tbz x10, #1, 120f\n" - "ldr h0, [x9], #0x2\n" - "ldr h1, [x26], #0x2\n" - "ldr h2, [x24], #0x2\n" - "ldr h3, [x23], #0x2\n" + "tbz x13, #1, 120f\n" + "ldr h0, [x12], #0x2\n" + "ldr h1, [x28], #0x2\n" + "ldr h2, [x26], #0x2\n" + "ldr h3, [x24], #0x2\n" "ldr h4, [x22], #0x2\n" - "tbz x10, #0, 121f\n" - "ld1 { v0.b }[2], [x9]\n" - "ld1 { v1.b }[2], [x26]\n" - "ld1 { v2.b }[2], [x24]\n" - "ld1 { v3.b }[2], [x23]\n" + "tbz x13, #0, 121f\n" + "ld1 { v0.b }[2], [x12]\n" + "ld1 { v1.b }[2], [x28]\n" + "ld1 { v2.b }[2], [x26]\n" + "ld1 { v3.b }[2], [x24]\n" "ld1 { v4.b }[2], [x22]\n" "b 121f\n" "120:" // Height 5: Multiply loop: Ragged operand read: partial_1_0 - "ldr b0, [x9, #0x0]\n" - "ldr b1, [x26, #0x0]\n" - "ldr b2, [x24, #0x0]\n" - "ldr b3, [x23, #0x0]\n" + "ldr b0, [x12, #0x0]\n" + "ldr b1, [x28, #0x0]\n" + "ldr b2, [x26, #0x0]\n" + "ldr b3, [x24, #0x0]\n" "ldr b4, [x22, #0x0]\n" "121:" // Height 5: Multiply loop: Ragged operand read: Done - "ldr q6, [x12, #0x0]\n" - "ldr q7, [x12, #0x10]\n" + "ldr q6, [x15, #0x0]\n" + "ldr q7, [x15, #0x10]\n" ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n" ".inst 0x4f84e0d8 // sdot v24.4s, v6.16b, v4.4b[0]\n" + "ldr q6, [x15, #0x20]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n" ".inst 0x4f84e0f9 // sdot v25.4s, v7.16b, v4.4b[0]\n" - "ldr q6, [x12, #0x20]\n" - "ldr q7, [x12, #0x30]\n" - "add x12, x12, #0x40\n" + "ldr q7, [x15, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "add x15, x15, #0x40\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n" @@ -2485,14 +2485,14 @@ void a64_hybrid_s8qs_dot_6x16_a55 ( ".inst 0x4f84e0fb // sdot v27.4s, v7.16b, v4.4b[0]\n" "122:" // Height 5: Multiply loop: No odd multiplies "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" - "add x11, x11, #0x1\n" - "cmp x11, x19\n" + "add x14, x14, #0x1\n" + "cmp x14, x19\n" "bne 112b\n" "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "prfm pstl1keep, [x13, #0x0]\n" - "ldr q0, [x15, #0x0]\n" - "ldr q1, [x15, #0x10]\n" - "add x23, x13, x19\n" + "prfm pstl1keep, [x16, #0x0]\n" + "ldr q0, [x8, #0x0]\n" + "ldr q1, [x8, #0x10]\n" + "add x23, x16, x19\n" "prfm pstl1keep, [x23, #0x0]\n" "add x22, x23, x19\n" "add v8.4s, v8.4s, v0.4s\n" @@ -2510,9 +2510,9 @@ void a64_hybrid_s8qs_dot_6x16_a55 ( "prfm pstl1keep, [x21, #0x0]\n" "add x20, x21, x19\n" "prfm pstl1keep, [x20, #0x0]\n" - "ldr q2, [x15, #0x20]\n" - "ldr q3, [x15, #0x30]\n" - "add x15, x15, #0x40\n" + "ldr q2, [x8, #0x20]\n" + "ldr q3, [x8, #0x30]\n" + "add x8, x8, #0x40\n" "add v10.4s, v10.4s, v2.4s\n" "add v14.4s, v14.4s, v2.4s\n" "add v11.4s, v11.4s, v3.4s\n" @@ -2524,16 +2524,16 @@ void a64_hybrid_s8qs_dot_6x16_a55 ( "add v26.4s, v26.4s, v2.4s\n" "add v27.4s, v27.4s, v3.4s\n" "tbz %x[flags], #4, 123f\n" - "ldr q0, [x16, #0x0]\n" - "ldr q4, [x17, #0x0]\n" - "ldr q1, [x16, #0x10]\n" - "ldr q5, [x17, #0x10]\n" - "ldr q2, [x16, #0x20]\n" - "ldr q6, [x17, #0x20]\n" - "ldr q3, [x16, #0x30]\n" - "add x16, x16, #0x40\n" - "ldr q7, [x17, #0x30]\n" - "add x17, x17, #0x40\n" + "ldr q0, [x7, #0x0]\n" + "ldr q4, [x6, #0x0]\n" + "ldr q1, [x7, #0x10]\n" + "ldr q5, [x6, #0x10]\n" + "ldr q2, [x7, #0x20]\n" + "ldr q6, [x6, #0x20]\n" + "ldr q3, [x7, #0x30]\n" + "add x7, x7, #0x40\n" + "ldr q7, [x6, #0x30]\n" + "add x6, x6, #0x40\n" "b 124f\n" "123:" // Height 5: per layer parameters "add x25, %x[qp], %[per_layer_right_shift]\n" @@ -2654,7 +2654,7 @@ void a64_hybrid_s8qs_dot_6x16_a55 ( "ld1r { v4.4s }, [x24]\n" "add x24, %x[qp], %[maxval]\n" "ld1r { v5.4s }, [x25]\n" - "cmp x14, #0x10\n" + "cmp x17, #0x10\n" "ld1r { v6.4s }, [x24]\n" "add v8.4s, v8.4s, v4.4s\n" "add v9.4s, v9.4s, v4.4s\n" @@ -2732,105 +2732,105 @@ void a64_hybrid_s8qs_dot_6x16_a55 ( "uzp1 v20.16b, v20.16b, v21.16b\n" "uzp1 v24.16b, v24.16b, v25.16b\n" "bge 134f\n" - "tbz x14, #3, 129f\n" - "str d8, [x13], #0x8\n" + "tbz x17, #3, 129f\n" + "str d8, [x16], #0x8\n" "str d12, [x23], #0x8\n" "str d16, [x22], #0x8\n" "str d20, [x21], #0x8\n" "str d24, [x20], #0x8\n" - "tbz x14, #2, 127f\n" - "st1 { v8.s }[2], [x13], #0x4\n" + "tbz x17, #2, 127f\n" + "st1 { v8.s }[2], [x16], #0x4\n" "st1 { v12.s }[2], [x23], #0x4\n" "st1 { v16.s }[2], [x22], #0x4\n" "st1 { v20.s }[2], [x21], #0x4\n" "st1 { v24.s }[2], [x20], #0x4\n" - "tbz x14, #1, 126f\n" - "st1 { v8.h }[6], [x13], #0x2\n" + "tbz x17, #1, 126f\n" + "st1 { v8.h }[6], [x16], #0x2\n" "st1 { v12.h }[6], [x23], #0x2\n" "st1 { v16.h }[6], [x22], #0x2\n" "st1 { v20.h }[6], [x21], #0x2\n" "st1 { v24.h }[6], [x20], #0x2\n" - "tbz x14, #0, 133f\n" - "st1 { v8.b }[14], [x13]\n" + "tbz x17, #0, 133f\n" + "st1 { v8.b }[14], [x16]\n" "st1 { v12.b }[14], [x23]\n" "st1 { v16.b }[14], [x22]\n" "st1 { v20.b }[14], [x21]\n" "st1 { v24.b }[14], [x20]\n" "b 133f\n" "126:" // Height 5: Partial direct writeback: partial_1_12 - "tbz x14, #0, 133f\n" - "st1 { v8.b }[12], [x13]\n" + "tbz x17, #0, 133f\n" + "st1 { v8.b }[12], [x16]\n" "st1 { v12.b }[12], [x23]\n" "st1 { v16.b }[12], [x22]\n" "st1 { v20.b }[12], [x21]\n" "st1 { v24.b }[12], [x20]\n" "b 133f\n" "127:" // Height 5: Partial direct writeback: partial_2_8 - "tbz x14, #1, 128f\n" - "st1 { v8.h }[4], [x13], #0x2\n" + "tbz x17, #1, 128f\n" + "st1 { v8.h }[4], [x16], #0x2\n" "st1 { v12.h }[4], [x23], #0x2\n" "st1 { v16.h }[4], [x22], #0x2\n" "st1 { v20.h }[4], [x21], #0x2\n" "st1 { v24.h }[4], [x20], #0x2\n" - "tbz x14, #0, 133f\n" - "st1 { v8.b }[10], [x13]\n" + "tbz x17, #0, 133f\n" + "st1 { v8.b }[10], [x16]\n" "st1 { v12.b }[10], [x23]\n" "st1 { v16.b }[10], [x22]\n" "st1 { v20.b }[10], [x21]\n" "st1 { v24.b }[10], [x20]\n" "b 133f\n" "128:" // Height 5: Partial direct writeback: partial_1_8 - "tbz x14, #0, 133f\n" - "st1 { v8.b }[8], [x13]\n" + "tbz x17, #0, 133f\n" + "st1 { v8.b }[8], [x16]\n" "st1 { v12.b }[8], [x23]\n" "st1 { v16.b }[8], [x22]\n" "st1 { v20.b }[8], [x21]\n" "st1 { v24.b }[8], [x20]\n" "b 133f\n" "129:" // Height 5: Partial direct writeback: partial_4_0 - "tbz x14, #2, 131f\n" - "str s8, [x13], #0x4\n" + "tbz x17, #2, 131f\n" + "str s8, [x16], #0x4\n" "str s12, [x23], #0x4\n" "str s16, [x22], #0x4\n" "str s20, [x21], #0x4\n" "str s24, [x20], #0x4\n" - "tbz x14, #1, 130f\n" - "st1 { v8.h }[2], [x13], #0x2\n" + "tbz x17, #1, 130f\n" + "st1 { v8.h }[2], [x16], #0x2\n" "st1 { v12.h }[2], [x23], #0x2\n" "st1 { v16.h }[2], [x22], #0x2\n" "st1 { v20.h }[2], [x21], #0x2\n" "st1 { v24.h }[2], [x20], #0x2\n" - "tbz x14, #0, 133f\n" - "st1 { v8.b }[6], [x13]\n" + "tbz x17, #0, 133f\n" + "st1 { v8.b }[6], [x16]\n" "st1 { v12.b }[6], [x23]\n" "st1 { v16.b }[6], [x22]\n" "st1 { v20.b }[6], [x21]\n" "st1 { v24.b }[6], [x20]\n" "b 133f\n" "130:" // Height 5: Partial direct writeback: partial_1_4 - "tbz x14, #0, 133f\n" - "st1 { v8.b }[4], [x13]\n" + "tbz x17, #0, 133f\n" + "st1 { v8.b }[4], [x16]\n" "st1 { v12.b }[4], [x23]\n" "st1 { v16.b }[4], [x22]\n" "st1 { v20.b }[4], [x21]\n" "st1 { v24.b }[4], [x20]\n" "b 133f\n" "131:" // Height 5: Partial direct writeback: partial_2_0 - "tbz x14, #1, 132f\n" - "str h8, [x13], #0x2\n" + "tbz x17, #1, 132f\n" + "str h8, [x16], #0x2\n" "str h12, [x23], #0x2\n" "str h16, [x22], #0x2\n" "str h20, [x21], #0x2\n" "str h24, [x20], #0x2\n" - "tbz x14, #0, 133f\n" - "st1 { v8.b }[2], [x13]\n" + "tbz x17, #0, 133f\n" + "st1 { v8.b }[2], [x16]\n" "st1 { v12.b }[2], [x23]\n" "st1 { v16.b }[2], [x22]\n" "st1 { v20.b }[2], [x21]\n" "st1 { v24.b }[2], [x20]\n" "b 133f\n" "132:" // Height 5: Partial direct writeback: partial_1_0 - "str b8, [x13, #0x0]\n" + "str b8, [x16, #0x0]\n" "str b12, [x23, #0x0]\n" "str b16, [x22, #0x0]\n" "str b20, [x21, #0x0]\n" @@ -2838,24 +2838,24 @@ void a64_hybrid_s8qs_dot_6x16_a55 ( "133:" // Height 5: Partial direct writeback: Done "b 135f\n" "134:" // Height 5: Full writeback - "str q8, [x13, #0x0]\n" - "add x13, x13, #0x10\n" + "str q8, [x16, #0x0]\n" + "add x16, x16, #0x10\n" "str q12, [x23, #0x0]\n" "str q16, [x22, #0x0]\n" "str q20, [x21, #0x0]\n" "str q24, [x20, #0x0]\n" "135:" // Height 5: Writeback done - "subs x14, x14, #0x10\n" + "subs x17, x17, #0x10\n" "bgt 110b\n" "b 164f\n" "136:" // Height 6 - "ldr x17, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n" - "mov x15, %x[col_bias]\n" - "ldr x16, [%x[args_ptr], %[offsetof_shift_ptr]]\n" - "mov x13, %x[output_ptr]\n" - "ldr x14, [%x[args_ptr], %[offsetof_N]]\n" + "ldr x6, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n" + "mov x8, %x[col_bias]\n" + "ldr x7, [%x[args_ptr], %[offsetof_shift_ptr]]\n" + "mov x16, %x[output_ptr]\n" + "ldr x17, [%x[args_ptr], %[offsetof_N]]\n" "mov x20, #0x6\n" - "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "ldr x15, [%x[args_ptr], %[offsetof_B_ptr]]\n" "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" "madd %x[output_ptr], x19, x20, %x[output_ptr]\n" "137:" // Height 6: Column loop @@ -2884,347 +2884,347 @@ void a64_hybrid_s8qs_dot_6x16_a55 ( "movi v30.4s, #0x0\n" "movi v31.4s, #0x0\n" "138:" // Height 6: setup done - "mov x11, #0x0\n" + "mov x14, #0x0\n" "139:" // Height 6: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w10, [x20, x11, LSL #0x2]\n" + "ldr w13, [x20, x14, LSL #0x2]\n" "tbz %x[flags], #3, 140f\n" - "ldr x20, [%x[input_ptr], x11, LSL #0x3]\n" + "ldr x20, [%x[input_ptr], x14, LSL #0x3]\n" "add x20, x20, x19, LSL #3\n" - "ldr x9, [x20, #0x0]\n" - "ldr x26, [x20, #0x8]\n" - "ldr x24, [x20, #0x10]\n" - "ldr x23, [x20, #0x18]\n" + "ldr x12, [x20, #0x0]\n" + "ldr x28, [x20, #0x8]\n" + "ldr x26, [x20, #0x10]\n" + "ldr x24, [x20, #0x18]\n" "ldr x22, [x20, #0x20]\n" "ldr x20, [x20, #0x28]\n" - "cbnz x11, 141f\n" + "cbnz x14, 141f\n" "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x9, x9, x19\n" + "add x12, x12, x19\n" + "add x28, x28, x19\n" "add x26, x26, x19\n" "add x24, x24, x19\n" - "add x23, x23, x19\n" "add x22, x22, x19\n" "add x20, x20, x19\n" "b 141f\n" "140:" // Height 6: setup direct input - "mov x9, %x[input_ptr]\n" - "add x26, x9, x19\n" + "mov x12, %x[input_ptr]\n" + "add x28, x12, x19\n" + "add x26, x28, x19\n" "add x24, x26, x19\n" - "add x23, x24, x19\n" - "add x22, x23, x19\n" + "add x22, x24, x19\n" "add x20, x22, x19\n" "141:" // Height 6: input setup done - "cmp x10, #0x10\n" + "cmp x13, #0x10\n" "blt 144f\n" - "ldr q0, [x9, #0x0]\n" - "ldr q1, [x26, #0x0]\n" - "cmp x10, #0x20\n" - "ldr q2, [x24, #0x0]\n" - "ldr q3, [x23, #0x0]\n" + "ldr q0, [x12, #0x0]\n" + "ldr q1, [x28, #0x0]\n" + "cmp x13, #0x20\n" + "ldr q2, [x26, #0x0]\n" + "ldr q3, [x24, #0x0]\n" "ldr q4, [x22, #0x0]\n" "ldr q5, [x20, #0x0]\n" - "ldr q6, [x12, #0x0]\n" + "ldr q6, [x15, #0x0]\n" "blt 143f\n" "142:" // Height 6: Multiply loop: Main loop head ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" + "ldr d7, [x15, #0x10]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" + "ldr x11, [x15, #0x18]\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" + "ldr x10, [x15, #0x28]\n" ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n" + "add x12, x12, #0x10\n" ".inst 0x4f84e0d8 // sdot v24.4s, v6.16b, v4.4b[0]\n" + "mov v7.d[1], x11\n" ".inst 0x4f85e0dc // sdot v28.4s, v6.16b, v5.4b[0]\n" - "ldr d7, [x12, #0x10]\n" - "add x9, x9, #0x10\n" - "ldr x19, [x12, #0x18]\n" - "add x26, x26, #0x10\n" - "prfm pldl1keep, [x9, #0x80]\n" - "add x24, x24, #0x10\n" - "ldr d6, [x12, #0x20]\n" - "add x23, x23, #0x10\n" - "mov v7.d[1], x19\n" - "ldr x28, [x12, #0x28]\n" - "ldr x19, [x12, #0x38]\n" - "add x22, x22, #0x10\n" + "prfm pldl1keep, [x12, #0x80]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" + "ldr d6, [x15, #0x20]\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" - "mov v6.d[1], x28\n" + "ldr x11, [x15, #0x38]\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" + "ldr x9, [x12, #0x8]\n" ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n" + "mov v6.d[1], x10\n" ".inst 0x4f84e0f9 // sdot v25.4s, v7.16b, v4.4b[0]\n" + "ldr x10, [x15, #0x48]\n" ".inst 0x4f85e0fd // sdot v29.4s, v7.16b, v5.4b[0]\n" + "ldr d7, [x15, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "add x28, x28, #0x10\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" + "prfm pldl1keep, [x28, #0x80]\n" ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" + "mov v7.d[1], x11\n" ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n" + "ldr x11, [x15, #0x58]\n" ".inst 0x4f84e0da // sdot v26.4s, v6.16b, v4.4b[0]\n" + "ldr x27, [x28, #0x8]\n" ".inst 0x4f85e0de // sdot v30.4s, v6.16b, v5.4b[0]\n" - "ldr d7, [x12, #0x30]\n" - "prfm pldl1keep, [x26, #0x80]\n" - "add x20, x20, #0x10\n" - "ldr d6, [x12, #0x40]\n" - "sub x10, x10, #0x10\n" - "mov v7.d[1], x19\n" - "ldr x28, [x12, #0x48]\n" - "ldr x19, [x12, #0x58]\n" - "cmp x10, #0x20\n" + "ldr d6, [x15, #0x40]\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" + "add x26, x26, #0x10\n" ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" - "mov v6.d[1], x28\n" + "prfm pldl1keep, [x26, #0x80]\n" ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" + "mov v6.d[1], x10\n" ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n" + "ldr x10, [x15, #0x68]\n" ".inst 0x4f84e0fb // sdot v27.4s, v7.16b, v4.4b[0]\n" + "ldr x25, [x26, #0x8]\n" ".inst 0x4f85e0ff // sdot v31.4s, v7.16b, v5.4b[0]\n" + "ldr d7, [x15, #0x50]\n" ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n" + "add x24, x24, #0x10\n" ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n" + "prfm pldl1keep, [x24, #0x80]\n" ".inst 0x4fa2e0d0 // sdot v16.4s, v6.16b, v2.4b[1]\n" + "mov v7.d[1], x11\n" ".inst 0x4fa3e0d4 // sdot v20.4s, v6.16b, v3.4b[1]\n" + "ldr x11, [x15, #0x78]\n" ".inst 0x4fa4e0d8 // sdot v24.4s, v6.16b, v4.4b[1]\n" + "ldr x23, [x24, #0x8]\n" ".inst 0x4fa5e0dc // sdot v28.4s, v6.16b, v5.4b[1]\n" - "ldr d7, [x12, #0x50]\n" - "prfm pldl1keep, [x24, #0x80]\n" - "ldr d6, [x12, #0x60]\n" - "mov v7.d[1], x19\n" - "ldr x28, [x12, #0x68]\n" - "ldr x19, [x12, #0x78]\n" + "ldr d6, [x15, #0x60]\n" ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n" + "add x22, x22, #0x10\n" ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n" - "mov v6.d[1], x28\n" + "prfm pldl1keep, [x22, #0x80]\n" ".inst 0x4fa2e0f1 // sdot v17.4s, v7.16b, v2.4b[1]\n" + "mov v6.d[1], x10\n" ".inst 0x4fa3e0f5 // sdot v21.4s, v7.16b, v3.4b[1]\n" + "ldr x10, [x15, #0x88]\n" ".inst 0x4fa4e0f9 // sdot v25.4s, v7.16b, v4.4b[1]\n" + "ldr x21, [x22, #0x8]\n" ".inst 0x4fa5e0fd // sdot v29.4s, v7.16b, v5.4b[1]\n" + "ldr d7, [x15, #0x70]\n" ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n" + "add x20, x20, #0x10\n" ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n" + "prfm pldl1keep, [x20, #0x80]\n" ".inst 0x4fa2e0d2 // sdot v18.4s, v6.16b, v2.4b[1]\n" + "mov v7.d[1], x11\n" ".inst 0x4fa3e0d6 // sdot v22.4s, v6.16b, v3.4b[1]\n" + "ldr x11, [x15, #0x98]\n" ".inst 0x4fa4e0da // sdot v26.4s, v6.16b, v4.4b[1]\n" + "sub x13, x13, #0x10\n" ".inst 0x4fa5e0de // sdot v30.4s, v6.16b, v5.4b[1]\n" - "ldr d7, [x12, #0x70]\n" - "prfm pldl1keep, [x23, #0x80]\n" - "ldr d6, [x12, #0x80]\n" - "mov v7.d[1], x19\n" - "ldr x28, [x12, #0x88]\n" - "ldr x19, [x12, #0x98]\n" + "ldr d6, [x15, #0x80]\n" ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n" + "cmp x13, #0x20\n" ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n" - "mov v6.d[1], x28\n" ".inst 0x4fa2e0f3 // sdot v19.4s, v7.16b, v2.4b[1]\n" + "mov v6.d[1], x10\n" ".inst 0x4fa3e0f7 // sdot v23.4s, v7.16b, v3.4b[1]\n" + "ldr x10, [x15, #0xa8]\n" ".inst 0x4fa4e0fb // sdot v27.4s, v7.16b, v4.4b[1]\n" ".inst 0x4fa5e0ff // sdot v31.4s, v7.16b, v5.4b[1]\n" + "ldr d7, [x15, #0x90]\n" ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n" ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n" ".inst 0x4f82e8d0 // sdot v16.4s, v6.16b, v2.4b[2]\n" + "mov v7.d[1], x11\n" ".inst 0x4f83e8d4 // sdot v20.4s, v6.16b, v3.4b[2]\n" + "ldr x11, [x15, #0xb8]\n" ".inst 0x4f84e8d8 // sdot v24.4s, v6.16b, v4.4b[2]\n" ".inst 0x4f85e8dc // sdot v28.4s, v6.16b, v5.4b[2]\n" - "ldr d7, [x12, #0x90]\n" - "prfm pldl1keep, [x22, #0x80]\n" - "ldr d6, [x12, #0xa0]\n" - "mov v7.d[1], x19\n" - "ldr x28, [x12, #0xa8]\n" - "ldr x19, [x12, #0xb8]\n" + "ldr d6, [x15, #0xa0]\n" ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n" - "mov v6.d[1], x28\n" ".inst 0x4f82e8f1 // sdot v17.4s, v7.16b, v2.4b[2]\n" + "mov v6.d[1], x10\n" ".inst 0x4f83e8f5 // sdot v21.4s, v7.16b, v3.4b[2]\n" + "ldr x10, [x15, #0xc8]\n" ".inst 0x4f84e8f9 // sdot v25.4s, v7.16b, v4.4b[2]\n" ".inst 0x4f85e8fd // sdot v29.4s, v7.16b, v5.4b[2]\n" + "ldr d7, [x15, #0xb0]\n" ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n" ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n" ".inst 0x4f82e8d2 // sdot v18.4s, v6.16b, v2.4b[2]\n" + "mov v7.d[1], x11\n" ".inst 0x4f83e8d6 // sdot v22.4s, v6.16b, v3.4b[2]\n" + "ldr x11, [x15, #0xd8]\n" ".inst 0x4f84e8da // sdot v26.4s, v6.16b, v4.4b[2]\n" ".inst 0x4f85e8de // sdot v30.4s, v6.16b, v5.4b[2]\n" - "ldr d7, [x12, #0xb0]\n" - "prfm pldl1keep, [x20, #0x80]\n" - "ldr d6, [x12, #0xc0]\n" - "mov v7.d[1], x19\n" - "ldr x28, [x12, #0xc8]\n" - "ldr x19, [x12, #0xd8]\n" + "ldr d6, [x15, #0xc0]\n" ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n" - "mov v6.d[1], x28\n" ".inst 0x4f82e8f3 // sdot v19.4s, v7.16b, v2.4b[2]\n" + "mov v6.d[1], x10\n" ".inst 0x4f83e8f7 // sdot v23.4s, v7.16b, v3.4b[2]\n" + "ldr x10, [x15, #0xe8]\n" ".inst 0x4f84e8fb // sdot v27.4s, v7.16b, v4.4b[2]\n" ".inst 0x4f85e8ff // sdot v31.4s, v7.16b, v5.4b[2]\n" + "ldr d7, [x15, #0xd0]\n" ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n" ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n" ".inst 0x4fa2e8d0 // sdot v16.4s, v6.16b, v2.4b[3]\n" + "mov v7.d[1], x11\n" ".inst 0x4fa3e8d4 // sdot v20.4s, v6.16b, v3.4b[3]\n" + "ldr x11, [x15, #0xf8]\n" ".inst 0x4fa4e8d8 // sdot v24.4s, v6.16b, v4.4b[3]\n" ".inst 0x4fa5e8dc // sdot v28.4s, v6.16b, v5.4b[3]\n" - "ldr d7, [x12, #0xd0]\n" - "ldr d6, [x12, #0xe0]\n" - "ldr x28, [x12, #0xe8]\n" - "mov v7.d[1], x19\n" - "ldr x19, [x12, #0xf8]\n" - "ldr x27, [x9, #0x8]\n" - "mov v6.d[1], x28\n" + "ldr d6, [x15, #0xe0]\n" ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n" ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n" ".inst 0x4fa2e8f1 // sdot v17.4s, v7.16b, v2.4b[3]\n" + "mov v6.d[1], x10\n" ".inst 0x4fa3e8f5 // sdot v21.4s, v7.16b, v3.4b[3]\n" + "ldr x19, [x20, #0x8]\n" ".inst 0x4fa4e8f9 // sdot v25.4s, v7.16b, v4.4b[3]\n" ".inst 0x4fa5e8fd // sdot v29.4s, v7.16b, v5.4b[3]\n" + "ldr d7, [x15, #0xf0]\n" ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n" + "add x15, x15, #0x100\n" ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n" ".inst 0x4fa2e8d2 // sdot v18.4s, v6.16b, v2.4b[3]\n" + "mov v7.d[1], x11\n" ".inst 0x4fa3e8d6 // sdot v22.4s, v6.16b, v3.4b[3]\n" + "ldr x10, [x15, #0x8]\n" ".inst 0x4fa4e8da // sdot v26.4s, v6.16b, v4.4b[3]\n" ".inst 0x4fa5e8de // sdot v30.4s, v6.16b, v5.4b[3]\n" - "ldr d7, [x12, #0xf0]\n" - "ldr x25, [x26, #0x8]\n" - "add x12, x12, #0x100\n" - "ldr d6, [x12, #0x0]\n" - "mov v7.d[1], x19\n" - "ldr x28, [x12, #0x8]\n" - "ldr x21, [x24, #0x8]\n" + "ldr d6, [x15, #0x0]\n" ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n" + "ldr d0, [x12, #0x0]\n" ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n" - "mov v6.d[1], x28\n" + "ldr d1, [x28, #0x0]\n" ".inst 0x4fa2e8f3 // sdot v19.4s, v7.16b, v2.4b[3]\n" + "mov v6.d[1], x10\n" ".inst 0x4fa3e8f7 // sdot v23.4s, v7.16b, v3.4b[3]\n" + "mov v0.d[1], x9\n" ".inst 0x4fa4e8fb // sdot v27.4s, v7.16b, v4.4b[3]\n" + "mov v1.d[1], x27\n" ".inst 0x4fa5e8ff // sdot v31.4s, v7.16b, v5.4b[3]\n" - "ldr d0, [x9, #0x0]\n" - "ldr d1, [x26, #0x0]\n" - "ldr d2, [x24, #0x0]\n" - "mov v0.d[1], x27\n" - "ldr d3, [x23, #0x0]\n" - "mov v1.d[1], x25\n" - "ldr x19, [x23, #0x8]\n" - "mov v2.d[1], x21\n" + "ldr d2, [x26, #0x0]\n" + "ldr d3, [x24, #0x0]\n" "ldr d4, [x22, #0x0]\n" - "ldr x21, [x22, #0x8]\n" - "mov v3.d[1], x19\n" + "mov v2.d[1], x25\n" "ldr d5, [x20, #0x0]\n" - "ldr x19, [x20, #0x8]\n" + "mov v3.d[1], x23\n" "mov v4.d[1], x21\n" "mov v5.d[1], x19\n" "bge 142b\n" "143:" // Height 6: Multiply loop: Single iteration only ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" + "ldr q7, [x15, #0x10]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" + "sub x13, x13, #0x10\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" + "add x12, x12, #0x10\n" ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n" + "prfm pldl1keep, [x12, #0x80]\n" ".inst 0x4f84e0d8 // sdot v24.4s, v6.16b, v4.4b[0]\n" + "add x28, x28, #0x10\n" ".inst 0x4f85e0dc // sdot v28.4s, v6.16b, v5.4b[0]\n" - "ldr q7, [x12, #0x10]\n" - "sub x10, x10, #0x10\n" - "ldr q6, [x12, #0x20]\n" - "add x9, x9, #0x10\n" - "prfm pldl1keep, [x9, #0x80]\n" - "add x26, x26, #0x10\n" + "prfm pldl1keep, [x28, #0x80]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" + "ldr q6, [x15, #0x20]\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" + "add x26, x26, #0x10\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" + "prfm pldl1keep, [x26, #0x80]\n" ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n" + "add x24, x24, #0x10\n" ".inst 0x4f84e0f9 // sdot v25.4s, v7.16b, v4.4b[0]\n" + "prfm pldl1keep, [x24, #0x80]\n" ".inst 0x4f85e0fd // sdot v29.4s, v7.16b, v5.4b[0]\n" + "ldr q7, [x15, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "add x22, x22, #0x10\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" + "prfm pldl1keep, [x22, #0x80]\n" ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" + "add x20, x20, #0x10\n" ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n" + "prfm pldl1keep, [x20, #0x80]\n" ".inst 0x4f84e0da // sdot v26.4s, v6.16b, v4.4b[0]\n" ".inst 0x4f85e0de // sdot v30.4s, v6.16b, v5.4b[0]\n" - "ldr q7, [x12, #0x30]\n" - "add x24, x24, #0x10\n" - "prfm pldl1keep, [x26, #0x80]\n" - "add x23, x23, #0x10\n" - "ldr q6, [x12, #0x40]\n" - "add x22, x22, #0x10\n" + "ldr q6, [x15, #0x40]\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n" ".inst 0x4f84e0fb // sdot v27.4s, v7.16b, v4.4b[0]\n" ".inst 0x4f85e0ff // sdot v31.4s, v7.16b, v5.4b[0]\n" + "ldr q7, [x15, #0x50]\n" ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n" ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n" ".inst 0x4fa2e0d0 // sdot v16.4s, v6.16b, v2.4b[1]\n" ".inst 0x4fa3e0d4 // sdot v20.4s, v6.16b, v3.4b[1]\n" ".inst 0x4fa4e0d8 // sdot v24.4s, v6.16b, v4.4b[1]\n" ".inst 0x4fa5e0dc // sdot v28.4s, v6.16b, v5.4b[1]\n" - "ldr q7, [x12, #0x50]\n" - "add x20, x20, #0x10\n" - "prfm pldl1keep, [x24, #0x80]\n" - "ldr q6, [x12, #0x60]\n" + "ldr q6, [x15, #0x60]\n" ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n" ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n" ".inst 0x4fa2e0f1 // sdot v17.4s, v7.16b, v2.4b[1]\n" ".inst 0x4fa3e0f5 // sdot v21.4s, v7.16b, v3.4b[1]\n" ".inst 0x4fa4e0f9 // sdot v25.4s, v7.16b, v4.4b[1]\n" ".inst 0x4fa5e0fd // sdot v29.4s, v7.16b, v5.4b[1]\n" + "ldr q7, [x15, #0x70]\n" ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n" ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n" ".inst 0x4fa2e0d2 // sdot v18.4s, v6.16b, v2.4b[1]\n" ".inst 0x4fa3e0d6 // sdot v22.4s, v6.16b, v3.4b[1]\n" ".inst 0x4fa4e0da // sdot v26.4s, v6.16b, v4.4b[1]\n" ".inst 0x4fa5e0de // sdot v30.4s, v6.16b, v5.4b[1]\n" - "ldr q7, [x12, #0x70]\n" - "prfm pldl1keep, [x23, #0x80]\n" - "ldr q6, [x12, #0x80]\n" + "ldr q6, [x15, #0x80]\n" ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n" ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n" ".inst 0x4fa2e0f3 // sdot v19.4s, v7.16b, v2.4b[1]\n" ".inst 0x4fa3e0f7 // sdot v23.4s, v7.16b, v3.4b[1]\n" ".inst 0x4fa4e0fb // sdot v27.4s, v7.16b, v4.4b[1]\n" ".inst 0x4fa5e0ff // sdot v31.4s, v7.16b, v5.4b[1]\n" + "ldr q7, [x15, #0x90]\n" ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n" ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n" ".inst 0x4f82e8d0 // sdot v16.4s, v6.16b, v2.4b[2]\n" ".inst 0x4f83e8d4 // sdot v20.4s, v6.16b, v3.4b[2]\n" ".inst 0x4f84e8d8 // sdot v24.4s, v6.16b, v4.4b[2]\n" ".inst 0x4f85e8dc // sdot v28.4s, v6.16b, v5.4b[2]\n" - "ldr q7, [x12, #0x90]\n" - "prfm pldl1keep, [x22, #0x80]\n" - "ldr q6, [x12, #0xa0]\n" + "ldr q6, [x15, #0xa0]\n" ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n" ".inst 0x4f82e8f1 // sdot v17.4s, v7.16b, v2.4b[2]\n" ".inst 0x4f83e8f5 // sdot v21.4s, v7.16b, v3.4b[2]\n" ".inst 0x4f84e8f9 // sdot v25.4s, v7.16b, v4.4b[2]\n" ".inst 0x4f85e8fd // sdot v29.4s, v7.16b, v5.4b[2]\n" + "ldr q7, [x15, #0xb0]\n" ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n" ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n" ".inst 0x4f82e8d2 // sdot v18.4s, v6.16b, v2.4b[2]\n" ".inst 0x4f83e8d6 // sdot v22.4s, v6.16b, v3.4b[2]\n" ".inst 0x4f84e8da // sdot v26.4s, v6.16b, v4.4b[2]\n" ".inst 0x4f85e8de // sdot v30.4s, v6.16b, v5.4b[2]\n" - "ldr q7, [x12, #0xb0]\n" - "prfm pldl1keep, [x20, #0x80]\n" - "ldr q6, [x12, #0xc0]\n" + "ldr q6, [x15, #0xc0]\n" ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n" ".inst 0x4f82e8f3 // sdot v19.4s, v7.16b, v2.4b[2]\n" ".inst 0x4f83e8f7 // sdot v23.4s, v7.16b, v3.4b[2]\n" ".inst 0x4f84e8fb // sdot v27.4s, v7.16b, v4.4b[2]\n" ".inst 0x4f85e8ff // sdot v31.4s, v7.16b, v5.4b[2]\n" + "ldr q7, [x15, #0xd0]\n" ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n" ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n" ".inst 0x4fa2e8d0 // sdot v16.4s, v6.16b, v2.4b[3]\n" ".inst 0x4fa3e8d4 // sdot v20.4s, v6.16b, v3.4b[3]\n" ".inst 0x4fa4e8d8 // sdot v24.4s, v6.16b, v4.4b[3]\n" ".inst 0x4fa5e8dc // sdot v28.4s, v6.16b, v5.4b[3]\n" - "ldr q7, [x12, #0xd0]\n" - "ldr q6, [x12, #0xe0]\n" + "ldr q6, [x15, #0xe0]\n" ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n" ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n" ".inst 0x4fa2e8f1 // sdot v17.4s, v7.16b, v2.4b[3]\n" ".inst 0x4fa3e8f5 // sdot v21.4s, v7.16b, v3.4b[3]\n" ".inst 0x4fa4e8f9 // sdot v25.4s, v7.16b, v4.4b[3]\n" ".inst 0x4fa5e8fd // sdot v29.4s, v7.16b, v5.4b[3]\n" + "ldr q7, [x15, #0xf0]\n" ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n" + "add x15, x15, #0x100\n" ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n" ".inst 0x4fa2e8d2 // sdot v18.4s, v6.16b, v2.4b[3]\n" ".inst 0x4fa3e8d6 // sdot v22.4s, v6.16b, v3.4b[3]\n" ".inst 0x4fa4e8da // sdot v26.4s, v6.16b, v4.4b[3]\n" ".inst 0x4fa5e8de // sdot v30.4s, v6.16b, v5.4b[3]\n" - "ldr q7, [x12, #0xf0]\n" - "add x12, x12, #0x100\n" ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n" ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n" ".inst 0x4fa2e8f3 // sdot v19.4s, v7.16b, v2.4b[3]\n" @@ -3232,36 +3232,36 @@ void a64_hybrid_s8qs_dot_6x16_a55 ( ".inst 0x4fa4e8fb // sdot v27.4s, v7.16b, v4.4b[3]\n" ".inst 0x4fa5e8ff // sdot v31.4s, v7.16b, v5.4b[3]\n" "144:" // Height 6: Multiply loop: Main loop skip - "cbz x10, 149f\n" - "cmp x10, #0x4\n" + "cbz x13, 149f\n" + "cmp x13, #0x4\n" "blt 146f\n" "145:" // Height 6: Multiply loop: Odd block loop - "ldr s0, [x9], #0x4\n" - "sub x10, x10, #0x4\n" - "ldr s1, [x26], #0x4\n" - "cmp x10, #0x4\n" - "ldr s2, [x24], #0x4\n" - "ldr s3, [x23], #0x4\n" + "ldr s0, [x12], #0x4\n" + "sub x13, x13, #0x4\n" + "ldr s1, [x28], #0x4\n" + "cmp x13, #0x4\n" + "ldr s2, [x26], #0x4\n" + "ldr s3, [x24], #0x4\n" "ldr s4, [x22], #0x4\n" "ldr s5, [x20], #0x4\n" - "ldr q6, [x12, #0x0]\n" - "ldr q7, [x12, #0x10]\n" + "ldr q6, [x15, #0x0]\n" + "ldr q7, [x15, #0x10]\n" ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n" ".inst 0x4f84e0d8 // sdot v24.4s, v6.16b, v4.4b[0]\n" ".inst 0x4f85e0dc // sdot v28.4s, v6.16b, v5.4b[0]\n" + "ldr q6, [x15, #0x20]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n" ".inst 0x4f84e0f9 // sdot v25.4s, v7.16b, v4.4b[0]\n" ".inst 0x4f85e0fd // sdot v29.4s, v7.16b, v5.4b[0]\n" - "ldr q6, [x12, #0x20]\n" - "ldr q7, [x12, #0x30]\n" - "add x12, x12, #0x40\n" + "ldr q7, [x15, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "add x15, x15, #0x40\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n" @@ -3274,49 +3274,49 @@ void a64_hybrid_s8qs_dot_6x16_a55 ( ".inst 0x4f84e0fb // sdot v27.4s, v7.16b, v4.4b[0]\n" ".inst 0x4f85e0ff // sdot v31.4s, v7.16b, v5.4b[0]\n" "bge 145b\n" - "cbz x10, 149f\n" + "cbz x13, 149f\n" "146:" // Height 6: Multiply loop: Skip odd blocks - "tbz x10, #1, 147f\n" - "ldr h0, [x9], #0x2\n" - "ldr h1, [x26], #0x2\n" - "ldr h2, [x24], #0x2\n" - "ldr h3, [x23], #0x2\n" + "tbz x13, #1, 147f\n" + "ldr h0, [x12], #0x2\n" + "ldr h1, [x28], #0x2\n" + "ldr h2, [x26], #0x2\n" + "ldr h3, [x24], #0x2\n" "ldr h4, [x22], #0x2\n" "ldr h5, [x20], #0x2\n" - "tbz x10, #0, 148f\n" - "ld1 { v0.b }[2], [x9]\n" - "ld1 { v1.b }[2], [x26]\n" - "ld1 { v2.b }[2], [x24]\n" - "ld1 { v3.b }[2], [x23]\n" + "tbz x13, #0, 148f\n" + "ld1 { v0.b }[2], [x12]\n" + "ld1 { v1.b }[2], [x28]\n" + "ld1 { v2.b }[2], [x26]\n" + "ld1 { v3.b }[2], [x24]\n" "ld1 { v4.b }[2], [x22]\n" "ld1 { v5.b }[2], [x20]\n" "b 148f\n" "147:" // Height 6: Multiply loop: Ragged operand read: partial_1_0 - "ldr b0, [x9, #0x0]\n" - "ldr b1, [x26, #0x0]\n" - "ldr b2, [x24, #0x0]\n" - "ldr b3, [x23, #0x0]\n" + "ldr b0, [x12, #0x0]\n" + "ldr b1, [x28, #0x0]\n" + "ldr b2, [x26, #0x0]\n" + "ldr b3, [x24, #0x0]\n" "ldr b4, [x22, #0x0]\n" "ldr b5, [x20, #0x0]\n" "148:" // Height 6: Multiply loop: Ragged operand read: Done - "ldr q6, [x12, #0x0]\n" - "ldr q7, [x12, #0x10]\n" + "ldr q6, [x15, #0x0]\n" + "ldr q7, [x15, #0x10]\n" ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n" ".inst 0x4f84e0d8 // sdot v24.4s, v6.16b, v4.4b[0]\n" ".inst 0x4f85e0dc // sdot v28.4s, v6.16b, v5.4b[0]\n" + "ldr q6, [x15, #0x20]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n" ".inst 0x4f84e0f9 // sdot v25.4s, v7.16b, v4.4b[0]\n" ".inst 0x4f85e0fd // sdot v29.4s, v7.16b, v5.4b[0]\n" - "ldr q6, [x12, #0x20]\n" - "ldr q7, [x12, #0x30]\n" - "add x12, x12, #0x40\n" + "ldr q7, [x15, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "add x15, x15, #0x40\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n" @@ -3330,14 +3330,14 @@ void a64_hybrid_s8qs_dot_6x16_a55 ( ".inst 0x4f85e0ff // sdot v31.4s, v7.16b, v5.4b[0]\n" "149:" // Height 6: Multiply loop: No odd multiplies "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" - "add x11, x11, #0x1\n" - "cmp x11, x19\n" + "add x14, x14, #0x1\n" + "cmp x14, x19\n" "bne 139b\n" "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "prfm pstl1keep, [x13, #0x0]\n" - "ldr q0, [x15, #0x0]\n" - "ldr q1, [x15, #0x10]\n" - "add x23, x13, x19\n" + "prfm pstl1keep, [x16, #0x0]\n" + "ldr q0, [x8, #0x0]\n" + "ldr q1, [x8, #0x10]\n" + "add x23, x16, x19\n" "prfm pstl1keep, [x23, #0x0]\n" "add x22, x23, x19\n" "add v8.4s, v8.4s, v0.4s\n" @@ -3359,9 +3359,9 @@ void a64_hybrid_s8qs_dot_6x16_a55 ( "prfm pstl1keep, [x20, #0x0]\n" "add x19, x20, x19\n" "prfm pstl1keep, [x19, #0x0]\n" - "ldr q2, [x15, #0x20]\n" - "ldr q3, [x15, #0x30]\n" - "add x15, x15, #0x40\n" + "ldr q2, [x8, #0x20]\n" + "ldr q3, [x8, #0x30]\n" + "add x8, x8, #0x40\n" "add v10.4s, v10.4s, v2.4s\n" "add v14.4s, v14.4s, v2.4s\n" "add v11.4s, v11.4s, v3.4s\n" @@ -3375,16 +3375,16 @@ void a64_hybrid_s8qs_dot_6x16_a55 ( "add v30.4s, v30.4s, v2.4s\n" "add v31.4s, v31.4s, v3.4s\n" "tbz %x[flags], #4, 150f\n" - "ldr q0, [x16, #0x0]\n" - "ldr q4, [x17, #0x0]\n" - "ldr q1, [x16, #0x10]\n" - "ldr q5, [x17, #0x10]\n" - "ldr q2, [x16, #0x20]\n" - "ldr q6, [x17, #0x20]\n" - "ldr q3, [x16, #0x30]\n" - "add x16, x16, #0x40\n" - "ldr q7, [x17, #0x30]\n" - "add x17, x17, #0x40\n" + "ldr q0, [x7, #0x0]\n" + "ldr q4, [x6, #0x0]\n" + "ldr q1, [x7, #0x10]\n" + "ldr q5, [x6, #0x10]\n" + "ldr q2, [x7, #0x20]\n" + "ldr q6, [x6, #0x20]\n" + "ldr q3, [x7, #0x30]\n" + "add x7, x7, #0x40\n" + "ldr q7, [x6, #0x30]\n" + "add x6, x6, #0x40\n" "b 151f\n" "150:" // Height 6: per layer parameters "add x25, %x[qp], %[per_layer_right_shift]\n" @@ -3525,7 +3525,7 @@ void a64_hybrid_s8qs_dot_6x16_a55 ( "ld1r { v4.4s }, [x24]\n" "add x24, %x[qp], %[maxval]\n" "ld1r { v5.4s }, [x25]\n" - "cmp x14, #0x10\n" + "cmp x17, #0x10\n" "ld1r { v6.4s }, [x24]\n" "add v8.4s, v8.4s, v4.4s\n" "add v9.4s, v9.4s, v4.4s\n" @@ -3618,29 +3618,29 @@ void a64_hybrid_s8qs_dot_6x16_a55 ( "uzp1 v29.8h, v30.8h, v31.8h\n" "uzp1 v28.16b, v28.16b, v29.16b\n" "bge 161f\n" - "tbz x14, #3, 156f\n" - "str d8, [x13], #0x8\n" + "tbz x17, #3, 156f\n" + "str d8, [x16], #0x8\n" "str d12, [x23], #0x8\n" "str d16, [x22], #0x8\n" "str d20, [x21], #0x8\n" "str d24, [x20], #0x8\n" "str d28, [x19], #0x8\n" - "tbz x14, #2, 154f\n" - "st1 { v8.s }[2], [x13], #0x4\n" + "tbz x17, #2, 154f\n" + "st1 { v8.s }[2], [x16], #0x4\n" "st1 { v12.s }[2], [x23], #0x4\n" "st1 { v16.s }[2], [x22], #0x4\n" "st1 { v20.s }[2], [x21], #0x4\n" "st1 { v24.s }[2], [x20], #0x4\n" "st1 { v28.s }[2], [x19], #0x4\n" - "tbz x14, #1, 153f\n" - "st1 { v8.h }[6], [x13], #0x2\n" + "tbz x17, #1, 153f\n" + "st1 { v8.h }[6], [x16], #0x2\n" "st1 { v12.h }[6], [x23], #0x2\n" "st1 { v16.h }[6], [x22], #0x2\n" "st1 { v20.h }[6], [x21], #0x2\n" "st1 { v24.h }[6], [x20], #0x2\n" "st1 { v28.h }[6], [x19], #0x2\n" - "tbz x14, #0, 160f\n" - "st1 { v8.b }[14], [x13]\n" + "tbz x17, #0, 160f\n" + "st1 { v8.b }[14], [x16]\n" "st1 { v12.b }[14], [x23]\n" "st1 { v16.b }[14], [x22]\n" "st1 { v20.b }[14], [x21]\n" @@ -3648,8 +3648,8 @@ void a64_hybrid_s8qs_dot_6x16_a55 ( "st1 { v28.b }[14], [x19]\n" "b 160f\n" "153:" // Height 6: Partial direct writeback: partial_1_12 - "tbz x14, #0, 160f\n" - "st1 { v8.b }[12], [x13]\n" + "tbz x17, #0, 160f\n" + "st1 { v8.b }[12], [x16]\n" "st1 { v12.b }[12], [x23]\n" "st1 { v16.b }[12], [x22]\n" "st1 { v20.b }[12], [x21]\n" @@ -3657,15 +3657,15 @@ void a64_hybrid_s8qs_dot_6x16_a55 ( "st1 { v28.b }[12], [x19]\n" "b 160f\n" "154:" // Height 6: Partial direct writeback: partial_2_8 - "tbz x14, #1, 155f\n" - "st1 { v8.h }[4], [x13], #0x2\n" + "tbz x17, #1, 155f\n" + "st1 { v8.h }[4], [x16], #0x2\n" "st1 { v12.h }[4], [x23], #0x2\n" "st1 { v16.h }[4], [x22], #0x2\n" "st1 { v20.h }[4], [x21], #0x2\n" "st1 { v24.h }[4], [x20], #0x2\n" "st1 { v28.h }[4], [x19], #0x2\n" - "tbz x14, #0, 160f\n" - "st1 { v8.b }[10], [x13]\n" + "tbz x17, #0, 160f\n" + "st1 { v8.b }[10], [x16]\n" "st1 { v12.b }[10], [x23]\n" "st1 { v16.b }[10], [x22]\n" "st1 { v20.b }[10], [x21]\n" @@ -3673,8 +3673,8 @@ void a64_hybrid_s8qs_dot_6x16_a55 ( "st1 { v28.b }[10], [x19]\n" "b 160f\n" "155:" // Height 6: Partial direct writeback: partial_1_8 - "tbz x14, #0, 160f\n" - "st1 { v8.b }[8], [x13]\n" + "tbz x17, #0, 160f\n" + "st1 { v8.b }[8], [x16]\n" "st1 { v12.b }[8], [x23]\n" "st1 { v16.b }[8], [x22]\n" "st1 { v20.b }[8], [x21]\n" @@ -3682,22 +3682,22 @@ void a64_hybrid_s8qs_dot_6x16_a55 ( "st1 { v28.b }[8], [x19]\n" "b 160f\n" "156:" // Height 6: Partial direct writeback: partial_4_0 - "tbz x14, #2, 158f\n" - "str s8, [x13], #0x4\n" + "tbz x17, #2, 158f\n" + "str s8, [x16], #0x4\n" "str s12, [x23], #0x4\n" "str s16, [x22], #0x4\n" "str s20, [x21], #0x4\n" "str s24, [x20], #0x4\n" "str s28, [x19], #0x4\n" - "tbz x14, #1, 157f\n" - "st1 { v8.h }[2], [x13], #0x2\n" + "tbz x17, #1, 157f\n" + "st1 { v8.h }[2], [x16], #0x2\n" "st1 { v12.h }[2], [x23], #0x2\n" "st1 { v16.h }[2], [x22], #0x2\n" "st1 { v20.h }[2], [x21], #0x2\n" "st1 { v24.h }[2], [x20], #0x2\n" "st1 { v28.h }[2], [x19], #0x2\n" - "tbz x14, #0, 160f\n" - "st1 { v8.b }[6], [x13]\n" + "tbz x17, #0, 160f\n" + "st1 { v8.b }[6], [x16]\n" "st1 { v12.b }[6], [x23]\n" "st1 { v16.b }[6], [x22]\n" "st1 { v20.b }[6], [x21]\n" @@ -3705,8 +3705,8 @@ void a64_hybrid_s8qs_dot_6x16_a55 ( "st1 { v28.b }[6], [x19]\n" "b 160f\n" "157:" // Height 6: Partial direct writeback: partial_1_4 - "tbz x14, #0, 160f\n" - "st1 { v8.b }[4], [x13]\n" + "tbz x17, #0, 160f\n" + "st1 { v8.b }[4], [x16]\n" "st1 { v12.b }[4], [x23]\n" "st1 { v16.b }[4], [x22]\n" "st1 { v20.b }[4], [x21]\n" @@ -3714,15 +3714,15 @@ void a64_hybrid_s8qs_dot_6x16_a55 ( "st1 { v28.b }[4], [x19]\n" "b 160f\n" "158:" // Height 6: Partial direct writeback: partial_2_0 - "tbz x14, #1, 159f\n" - "str h8, [x13], #0x2\n" + "tbz x17, #1, 159f\n" + "str h8, [x16], #0x2\n" "str h12, [x23], #0x2\n" "str h16, [x22], #0x2\n" "str h20, [x21], #0x2\n" "str h24, [x20], #0x2\n" "str h28, [x19], #0x2\n" - "tbz x14, #0, 160f\n" - "st1 { v8.b }[2], [x13]\n" + "tbz x17, #0, 160f\n" + "st1 { v8.b }[2], [x16]\n" "st1 { v12.b }[2], [x23]\n" "st1 { v16.b }[2], [x22]\n" "st1 { v20.b }[2], [x21]\n" @@ -3730,7 +3730,7 @@ void a64_hybrid_s8qs_dot_6x16_a55 ( "st1 { v28.b }[2], [x19]\n" "b 160f\n" "159:" // Height 6: Partial direct writeback: partial_1_0 - "str b8, [x13, #0x0]\n" + "str b8, [x16, #0x0]\n" "str b12, [x23, #0x0]\n" "str b16, [x22, #0x0]\n" "str b20, [x21, #0x0]\n" @@ -3739,15 +3739,15 @@ void a64_hybrid_s8qs_dot_6x16_a55 ( "160:" // Height 6: Partial direct writeback: Done "b 162f\n" "161:" // Height 6: Full writeback - "str q8, [x13, #0x0]\n" - "add x13, x13, #0x10\n" + "str q8, [x16, #0x0]\n" + "add x16, x16, #0x10\n" "str q12, [x23, #0x0]\n" "str q16, [x22, #0x0]\n" "str q20, [x21, #0x0]\n" "str q24, [x20, #0x0]\n" "str q28, [x19, #0x0]\n" "162:" // Height 6: Writeback done - "subs x14, x14, #0x10\n" + "subs x17, x17, #0x10\n" "bgt 137b\n" "subs %x[M], %x[M], #0x6\n" "beq 164f\n" @@ -3764,7 +3764,7 @@ void a64_hybrid_s8qs_dot_6x16_a55 ( : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr) : [args_ptr] "r" (&ka), [c_offset] "I" (offsetof(Requantize32, c_offset)), [col_bias] "r" (col_bias), [flags] "r" (flags), [maxval] "I" (offsetof(Requantize32, maxval)), [minval] "I" (offsetof(Requantize32, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_multiplier_ptr] "I" (offsetof(KernelArgs, multiplier_ptr)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_shift_ptr] "I" (offsetof(KernelArgs, shift_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)), [per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [qp] "r" (qp) - : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x6", "x7", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" ); } diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_6x16/a55.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_6x16/a55.cpp index 6b2be0a33a..3817785a79 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_6x16/a55.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_6x16/a55.cpp @@ -87,73 +87,73 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( "cmp %x[M], #0x2\n" "bgt 69f\n" "beq 35f\n" - "ldr x14, [%x[args_ptr], %[offsetof_N]]\n" - "ldr x13, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x12, %x[output_ptr]\n" + "ldr x17, [%x[args_ptr], %[offsetof_N]]\n" + "ldr x16, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "mov x15, %x[output_ptr]\n" "2:" // Height 1: Column loop "tbz %x[flags], #0, 12f\n" - "cmp x14, #0x10\n" + "cmp x17, #0x10\n" "bge 11f\n" - "tbz x14, #3, 6f\n" - "ld1 { v8.4s }, [x12], #0x10\n" - "ld1 { v9.4s }, [x12], #0x10\n" - "tbz x14, #2, 4f\n" - "ld1 { v10.4s }, [x12], #0x10\n" - "tbz x14, #1, 3f\n" + "tbz x17, #3, 6f\n" + "ld1 { v8.4s }, [x15], #0x10\n" + "ld1 { v9.4s }, [x15], #0x10\n" + "tbz x17, #2, 4f\n" + "ld1 { v10.4s }, [x15], #0x10\n" + "tbz x17, #1, 3f\n" "mov x24, #0x38\n" - "ldr d11, [x12], #0x8\n" - "tbz x14, #0, 10f\n" - "ld1 { v11.s }[2], [x12]\n" + "ldr d11, [x15], #0x8\n" + "tbz x17, #0, 10f\n" + "ld1 { v11.s }[2], [x15]\n" "b 10f\n" "3:" // Height 1: Partial accumulate: partial_1_12 "mov x24, #0x30\n" - "tbz x14, #0, 10f\n" - "ldr s11, [x12, #0x0]\n" + "tbz x17, #0, 10f\n" + "ldr s11, [x15, #0x0]\n" "b 10f\n" "4:" // Height 1: Partial accumulate: partial_2_8 - "tbz x14, #1, 5f\n" - "ldr d10, [x12], #0x8\n" + "tbz x17, #1, 5f\n" + "ldr d10, [x15], #0x8\n" "mov x24, #0x28\n" - "tbz x14, #0, 10f\n" - "ld1 { v10.s }[2], [x12]\n" + "tbz x17, #0, 10f\n" + "ld1 { v10.s }[2], [x15]\n" "b 10f\n" "5:" // Height 1: Partial accumulate: partial_1_8 "mov x24, #0x20\n" - "tbz x14, #0, 10f\n" - "ldr s10, [x12, #0x0]\n" + "tbz x17, #0, 10f\n" + "ldr s10, [x15, #0x0]\n" "b 10f\n" "6:" // Height 1: Partial accumulate: partial_4_0 - "tbz x14, #2, 8f\n" - "ld1 { v8.4s }, [x12], #0x10\n" - "tbz x14, #1, 7f\n" + "tbz x17, #2, 8f\n" + "ld1 { v8.4s }, [x15], #0x10\n" + "tbz x17, #1, 7f\n" "mov x24, #0x18\n" - "ldr d9, [x12], #0x8\n" - "tbz x14, #0, 10f\n" - "ld1 { v9.s }[2], [x12]\n" + "ldr d9, [x15], #0x8\n" + "tbz x17, #0, 10f\n" + "ld1 { v9.s }[2], [x15]\n" "b 10f\n" "7:" // Height 1: Partial accumulate: partial_1_4 "mov x24, #0x10\n" - "tbz x14, #0, 10f\n" - "ldr s9, [x12, #0x0]\n" + "tbz x17, #0, 10f\n" + "ldr s9, [x15, #0x0]\n" "b 10f\n" "8:" // Height 1: Partial accumulate: partial_2_0 - "tbz x14, #1, 9f\n" - "ldr d8, [x12], #0x8\n" + "tbz x17, #1, 9f\n" + "ldr d8, [x15], #0x8\n" "mov x24, #0x8\n" - "tbz x14, #0, 10f\n" - "ld1 { v8.s }[2], [x12]\n" + "tbz x17, #0, 10f\n" + "ld1 { v8.s }[2], [x15]\n" "b 10f\n" "9:" // Height 1: Partial accumulate: partial_1_0 - "ldr s8, [x12, #0x0]\n" + "ldr s8, [x15, #0x0]\n" "mov x24, #0x0\n" "10:" // Height 1: Partial accumulate: Done - "sub x12, x12, x24\n" + "sub x15, x15, x24\n" "b 13f\n" "11:" // Height 1: full accumulate - "ldr q8, [x12, #0x0]\n" - "ldr q9, [x12, #0x10]\n" - "ldr q10, [x12, #0x20]\n" - "ldr q11, [x12, #0x30]\n" + "ldr q8, [x15, #0x0]\n" + "ldr q9, [x15, #0x10]\n" + "ldr q10, [x15, #0x20]\n" + "ldr q11, [x15, #0x30]\n" "b 13f\n" "12:" // Height 1: no accumulate "movi v8.4s, #0x0\n" @@ -161,325 +161,325 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( "movi v10.4s, #0x0\n" "movi v11.4s, #0x0\n" "13:" // Height 1: setup done - "mov x11, #0x0\n" + "mov x14, #0x0\n" "14:" // Height 1: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w10, [x20, x11, LSL #0x2]\n" + "ldr w13, [x20, x14, LSL #0x2]\n" "tbz %x[flags], #3, 15f\n" - "ldr x20, [%x[input_ptr], x11, LSL #0x3]\n" + "ldr x20, [%x[input_ptr], x14, LSL #0x3]\n" "add x20, x20, x19, LSL #3\n" - "ldr x9, [x20, #0x0]\n" - "cbnz x11, 16f\n" + "ldr x12, [x20, #0x0]\n" + "cbnz x14, 16f\n" "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x9, x9, x19\n" + "add x12, x12, x19\n" "b 16f\n" "15:" // Height 1: setup direct input - "mov x9, %x[input_ptr]\n" + "mov x12, %x[input_ptr]\n" "16:" // Height 1: input setup done - "cmp x10, #0x10\n" + "cmp x13, #0x10\n" "blt 19f\n" - "ldr q0, [x9, #0x0]\n" - "ldr q6, [x13, #0x0]\n" - "cmp x10, #0x20\n" + "ldr q0, [x12, #0x0]\n" + "ldr q6, [x16, #0x0]\n" + "cmp x13, #0x20\n" "blt 18f\n" "17:" // Height 1: Multiply loop: Main loop head ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr d7, [x13, #0x10]\n" - "ldr x19, [x13, #0x18]\n" - "add x9, x9, #0x10\n" - "ldr d6, [x13, #0x20]\n" - "sub x10, x10, #0x10\n" - "ldr x28, [x13, #0x28]\n" - "cmp x10, #0x20\n" - "mov v7.d[1], x19\n" - "prfm pldl1keep, [x9, #0x80]\n" - "ldr x19, [x13, #0x38]\n" - "mov v6.d[1], x28\n" + "ldr d7, [x16, #0x10]\n" + "ldr x11, [x16, #0x18]\n" + "add x12, x12, #0x10\n" + "ldr d6, [x16, #0x20]\n" + "sub x13, x13, #0x10\n" + "ldr x10, [x16, #0x28]\n" + "cmp x13, #0x20\n" + "mov v7.d[1], x11\n" + "prfm pldl1keep, [x12, #0x80]\n" + "ldr x11, [x16, #0x38]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" - "ldr d7, [x13, #0x30]\n" + "mov v6.d[1], x10\n" + "ldr d7, [x16, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" - "ldr d6, [x13, #0x40]\n" - "ldr x28, [x13, #0x48]\n" - "mov v7.d[1], x19\n" - "ldr x19, [x13, #0x58]\n" - "ldr x27, [x9, #0x8]\n" - "mov v6.d[1], x28\n" + "ldr d6, [x16, #0x40]\n" + "ldr x10, [x16, #0x48]\n" + "mov v7.d[1], x11\n" + "ldr x11, [x16, #0x58]\n" + "ldr x9, [x12, #0x8]\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" - "ldr d7, [x13, #0x50]\n" + "mov v6.d[1], x10\n" + "ldr d7, [x16, #0x50]\n" ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n" - "ldr d6, [x13, #0x60]\n" - "ldr x28, [x13, #0x68]\n" - "mov v7.d[1], x19\n" - "ldr x19, [x13, #0x78]\n" - "mov v6.d[1], x28\n" + "ldr d6, [x16, #0x60]\n" + "ldr x10, [x16, #0x68]\n" + "mov v7.d[1], x11\n" + "ldr x11, [x16, #0x78]\n" ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n" - "ldr d7, [x13, #0x70]\n" + "mov v6.d[1], x10\n" + "ldr d7, [x16, #0x70]\n" ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n" - "ldr d6, [x13, #0x80]\n" - "ldr x28, [x13, #0x88]\n" - "mov v7.d[1], x19\n" - "ldr x19, [x13, #0x98]\n" - "mov v6.d[1], x28\n" + "ldr d6, [x16, #0x80]\n" + "ldr x10, [x16, #0x88]\n" + "mov v7.d[1], x11\n" + "ldr x11, [x16, #0x98]\n" ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n" - "ldr d7, [x13, #0x90]\n" + "mov v6.d[1], x10\n" + "ldr d7, [x16, #0x90]\n" ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n" - "ldr d6, [x13, #0xa0]\n" - "ldr x28, [x13, #0xa8]\n" - "mov v7.d[1], x19\n" - "ldr x19, [x13, #0xb8]\n" - "mov v6.d[1], x28\n" + "ldr d6, [x16, #0xa0]\n" + "ldr x10, [x16, #0xa8]\n" + "mov v7.d[1], x11\n" + "ldr x11, [x16, #0xb8]\n" ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n" - "ldr d7, [x13, #0xb0]\n" + "mov v6.d[1], x10\n" + "ldr d7, [x16, #0xb0]\n" ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n" - "ldr d6, [x13, #0xc0]\n" - "ldr x28, [x13, #0xc8]\n" - "mov v7.d[1], x19\n" - "ldr x19, [x13, #0xd8]\n" - "mov v6.d[1], x28\n" + "ldr d6, [x16, #0xc0]\n" + "ldr x10, [x16, #0xc8]\n" + "mov v7.d[1], x11\n" + "ldr x11, [x16, #0xd8]\n" ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n" - "ldr d7, [x13, #0xd0]\n" + "mov v6.d[1], x10\n" + "ldr d7, [x16, #0xd0]\n" ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n" - "ldr d6, [x13, #0xe0]\n" - "ldr x28, [x13, #0xe8]\n" - "mov v7.d[1], x19\n" - "ldr x19, [x13, #0xf8]\n" - "mov v6.d[1], x28\n" + "ldr d6, [x16, #0xe0]\n" + "ldr x10, [x16, #0xe8]\n" + "mov v7.d[1], x11\n" + "ldr x11, [x16, #0xf8]\n" ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n" - "ldr d7, [x13, #0xf0]\n" - "add x13, x13, #0x100\n" + "mov v6.d[1], x10\n" + "ldr d7, [x16, #0xf0]\n" + "add x16, x16, #0x100\n" ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n" - "ldr d6, [x13, #0x0]\n" - "ldr x28, [x13, #0x8]\n" - "mov v7.d[1], x19\n" - "mov v6.d[1], x28\n" + "ldr d6, [x16, #0x0]\n" + "ldr x10, [x16, #0x8]\n" + "mov v7.d[1], x11\n" ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n" - "ldr d0, [x9, #0x0]\n" - "mov v0.d[1], x27\n" + "mov v6.d[1], x10\n" + "ldr d0, [x12, #0x0]\n" + "mov v0.d[1], x9\n" "bge 17b\n" "18:" // Height 1: Multiply loop: Single iteration only ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr q7, [x13, #0x10]\n" - "ldr q6, [x13, #0x20]\n" - "sub x10, x10, #0x10\n" - "add x9, x9, #0x10\n" + "ldr q7, [x16, #0x10]\n" + "ldr q6, [x16, #0x20]\n" + "sub x13, x13, #0x10\n" + "add x12, x12, #0x10\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" - "prfm pldl1keep, [x9, #0x80]\n" + "prfm pldl1keep, [x12, #0x80]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" - "ldr q7, [x13, #0x30]\n" - "ldr q6, [x13, #0x40]\n" + "ldr q7, [x16, #0x30]\n" + "ldr q6, [x16, #0x40]\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" - "ldr q7, [x13, #0x50]\n" + "ldr q7, [x16, #0x50]\n" ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n" - "ldr q6, [x13, #0x60]\n" + "ldr q6, [x16, #0x60]\n" ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n" - "ldr q7, [x13, #0x70]\n" + "ldr q7, [x16, #0x70]\n" ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n" - "ldr q6, [x13, #0x80]\n" + "ldr q6, [x16, #0x80]\n" ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n" - "ldr q7, [x13, #0x90]\n" + "ldr q7, [x16, #0x90]\n" ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n" - "ldr q6, [x13, #0xa0]\n" + "ldr q6, [x16, #0xa0]\n" ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n" - "ldr q7, [x13, #0xb0]\n" + "ldr q7, [x16, #0xb0]\n" ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n" - "ldr q6, [x13, #0xc0]\n" + "ldr q6, [x16, #0xc0]\n" ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n" - "ldr q7, [x13, #0xd0]\n" + "ldr q7, [x16, #0xd0]\n" ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n" - "ldr q6, [x13, #0xe0]\n" + "ldr q6, [x16, #0xe0]\n" ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n" - "ldr q7, [x13, #0xf0]\n" + "ldr q7, [x16, #0xf0]\n" ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n" - "add x13, x13, #0x100\n" + "add x16, x16, #0x100\n" ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n" "19:" // Height 1: Multiply loop: Main loop skip - "cbz x10, 24f\n" - "cmp x10, #0x4\n" + "cbz x13, 24f\n" + "cmp x13, #0x4\n" "blt 21f\n" "20:" // Height 1: Multiply loop: Odd block loop - "ldr s0, [x9], #0x4\n" - "sub x10, x10, #0x4\n" - "ldr q6, [x13, #0x0]\n" - "cmp x10, #0x4\n" - "ldr q7, [x13, #0x10]\n" + "ldr s0, [x12], #0x4\n" + "sub x13, x13, #0x4\n" + "ldr q6, [x16, #0x0]\n" + "cmp x13, #0x4\n" + "ldr q7, [x16, #0x10]\n" ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr q6, [x13, #0x20]\n" + "ldr q6, [x16, #0x20]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" - "ldr q7, [x13, #0x30]\n" - "add x13, x13, #0x40\n" + "ldr q7, [x16, #0x30]\n" + "add x16, x16, #0x40\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" "bge 20b\n" - "cbz x10, 24f\n" + "cbz x13, 24f\n" "21:" // Height 1: Multiply loop: Skip odd blocks - "tbz x10, #1, 22f\n" - "ldr h0, [x9], #0x2\n" - "tbz x10, #0, 23f\n" - "ld1 { v0.b }[2], [x9]\n" + "tbz x13, #1, 22f\n" + "ldr h0, [x12], #0x2\n" + "tbz x13, #0, 23f\n" + "ld1 { v0.b }[2], [x12]\n" "b 23f\n" "22:" // Height 1: Multiply loop: Ragged operand read: partial_1_0 - "ldr b0, [x9, #0x0]\n" + "ldr b0, [x12, #0x0]\n" "23:" // Height 1: Multiply loop: Ragged operand read: Done - "ldr q6, [x13, #0x0]\n" - "ldr q7, [x13, #0x10]\n" + "ldr q6, [x16, #0x0]\n" + "ldr q7, [x16, #0x10]\n" ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr q6, [x13, #0x20]\n" + "ldr q6, [x16, #0x20]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" - "ldr q7, [x13, #0x30]\n" - "add x13, x13, #0x40\n" + "ldr q7, [x16, #0x30]\n" + "add x16, x16, #0x40\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" "24:" // Height 1: Multiply loop: No odd multiplies "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" - "add x11, x11, #0x1\n" - "cmp x11, x19\n" + "add x14, x14, #0x1\n" + "cmp x14, x19\n" "bne 14b\n" - "prfm pstl1keep, [x12, #0x0]\n" - "cmp x14, #0x10\n" + "prfm pstl1keep, [x15, #0x0]\n" + "cmp x17, #0x10\n" "bge 33f\n" - "tbz x14, #3, 28f\n" - "st1 { v8.4s }, [x12], #0x10\n" - "st1 { v9.4s }, [x12], #0x10\n" - "tbz x14, #2, 26f\n" - "st1 { v10.4s }, [x12], #0x10\n" - "tbz x14, #1, 25f\n" - "str d11, [x12], #0x8\n" - "tbz x14, #0, 32f\n" - "st1 { v11.s }[2], [x12]\n" + "tbz x17, #3, 28f\n" + "st1 { v8.4s }, [x15], #0x10\n" + "st1 { v9.4s }, [x15], #0x10\n" + "tbz x17, #2, 26f\n" + "st1 { v10.4s }, [x15], #0x10\n" + "tbz x17, #1, 25f\n" + "str d11, [x15], #0x8\n" + "tbz x17, #0, 32f\n" + "st1 { v11.s }[2], [x15]\n" "b 32f\n" "25:" // Height 1: Partial direct writeback: partial_1_12 - "tbz x14, #0, 32f\n" - "str s11, [x12, #0x0]\n" + "tbz x17, #0, 32f\n" + "str s11, [x15, #0x0]\n" "b 32f\n" "26:" // Height 1: Partial direct writeback: partial_2_8 - "tbz x14, #1, 27f\n" - "str d10, [x12], #0x8\n" - "tbz x14, #0, 32f\n" - "st1 { v10.s }[2], [x12]\n" + "tbz x17, #1, 27f\n" + "str d10, [x15], #0x8\n" + "tbz x17, #0, 32f\n" + "st1 { v10.s }[2], [x15]\n" "b 32f\n" "27:" // Height 1: Partial direct writeback: partial_1_8 - "tbz x14, #0, 32f\n" - "str s10, [x12, #0x0]\n" + "tbz x17, #0, 32f\n" + "str s10, [x15, #0x0]\n" "b 32f\n" "28:" // Height 1: Partial direct writeback: partial_4_0 - "tbz x14, #2, 30f\n" - "st1 { v8.4s }, [x12], #0x10\n" - "tbz x14, #1, 29f\n" - "str d9, [x12], #0x8\n" - "tbz x14, #0, 32f\n" - "st1 { v9.s }[2], [x12]\n" + "tbz x17, #2, 30f\n" + "st1 { v8.4s }, [x15], #0x10\n" + "tbz x17, #1, 29f\n" + "str d9, [x15], #0x8\n" + "tbz x17, #0, 32f\n" + "st1 { v9.s }[2], [x15]\n" "b 32f\n" "29:" // Height 1: Partial direct writeback: partial_1_4 - "tbz x14, #0, 32f\n" - "str s9, [x12, #0x0]\n" + "tbz x17, #0, 32f\n" + "str s9, [x15, #0x0]\n" "b 32f\n" "30:" // Height 1: Partial direct writeback: partial_2_0 - "tbz x14, #1, 31f\n" - "str d8, [x12], #0x8\n" - "tbz x14, #0, 32f\n" - "st1 { v8.s }[2], [x12]\n" + "tbz x17, #1, 31f\n" + "str d8, [x15], #0x8\n" + "tbz x17, #0, 32f\n" + "st1 { v8.s }[2], [x15]\n" "b 32f\n" "31:" // Height 1: Partial direct writeback: partial_1_0 - "str s8, [x12, #0x0]\n" + "str s8, [x15, #0x0]\n" "32:" // Height 1: Partial direct writeback: Done "b 34f\n" "33:" // Height 1: Full writeback - "str q8, [x12, #0x0]\n" - "str q9, [x12, #0x10]\n" - "str q10, [x12, #0x20]\n" - "str q11, [x12, #0x30]\n" - "add x12, x12, #0x40\n" + "str q8, [x15, #0x0]\n" + "str q9, [x15, #0x10]\n" + "str q10, [x15, #0x20]\n" + "str q11, [x15, #0x30]\n" + "add x15, x15, #0x40\n" "34:" // Height 1: Writeback done - "subs x14, x14, #0x10\n" + "subs x17, x17, #0x10\n" "bgt 2b\n" "b 206f\n" "35:" // Height 2 - "ldr x14, [%x[args_ptr], %[offsetof_N]]\n" - "mov x12, %x[output_ptr]\n" - "ldr x13, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "ldr x17, [%x[args_ptr], %[offsetof_N]]\n" + "mov x15, %x[output_ptr]\n" + "ldr x16, [%x[args_ptr], %[offsetof_B_ptr]]\n" "36:" // Height 2: Column loop "tbz %x[flags], #0, 46f\n" "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "cmp x14, #0x10\n" - "add x23, x12, x19, LSL #2\n" + "cmp x17, #0x10\n" + "add x23, x15, x19, LSL #2\n" "bge 45f\n" - "tbz x14, #3, 40f\n" - "ld1 { v8.4s }, [x12], #0x10\n" + "tbz x17, #3, 40f\n" + "ld1 { v8.4s }, [x15], #0x10\n" "ld1 { v12.4s }, [x23], #0x10\n" - "ld1 { v9.4s }, [x12], #0x10\n" + "ld1 { v9.4s }, [x15], #0x10\n" "ld1 { v13.4s }, [x23], #0x10\n" - "tbz x14, #2, 38f\n" - "ld1 { v10.4s }, [x12], #0x10\n" + "tbz x17, #2, 38f\n" + "ld1 { v10.4s }, [x15], #0x10\n" "ld1 { v14.4s }, [x23], #0x10\n" - "tbz x14, #1, 37f\n" + "tbz x17, #1, 37f\n" "mov x24, #0x38\n" - "ldr d11, [x12], #0x8\n" + "ldr d11, [x15], #0x8\n" "ldr d15, [x23], #0x8\n" - "tbz x14, #0, 44f\n" - "ld1 { v11.s }[2], [x12]\n" + "tbz x17, #0, 44f\n" + "ld1 { v11.s }[2], [x15]\n" "ld1 { v15.s }[2], [x23]\n" "b 44f\n" "37:" // Height 2: Partial accumulate: partial_1_12 "mov x24, #0x30\n" - "tbz x14, #0, 44f\n" - "ldr s11, [x12, #0x0]\n" + "tbz x17, #0, 44f\n" + "ldr s11, [x15, #0x0]\n" "ldr s15, [x23, #0x0]\n" "b 44f\n" "38:" // Height 2: Partial accumulate: partial_2_8 - "tbz x14, #1, 39f\n" - "ldr d10, [x12], #0x8\n" + "tbz x17, #1, 39f\n" + "ldr d10, [x15], #0x8\n" "ldr d14, [x23], #0x8\n" "mov x24, #0x28\n" - "tbz x14, #0, 44f\n" - "ld1 { v10.s }[2], [x12]\n" + "tbz x17, #0, 44f\n" + "ld1 { v10.s }[2], [x15]\n" "ld1 { v14.s }[2], [x23]\n" "b 44f\n" "39:" // Height 2: Partial accumulate: partial_1_8 "mov x24, #0x20\n" - "tbz x14, #0, 44f\n" - "ldr s10, [x12, #0x0]\n" + "tbz x17, #0, 44f\n" + "ldr s10, [x15, #0x0]\n" "ldr s14, [x23, #0x0]\n" "b 44f\n" "40:" // Height 2: Partial accumulate: partial_4_0 - "tbz x14, #2, 42f\n" - "ld1 { v8.4s }, [x12], #0x10\n" + "tbz x17, #2, 42f\n" + "ld1 { v8.4s }, [x15], #0x10\n" "ld1 { v12.4s }, [x23], #0x10\n" - "tbz x14, #1, 41f\n" + "tbz x17, #1, 41f\n" "mov x24, #0x18\n" - "ldr d9, [x12], #0x8\n" + "ldr d9, [x15], #0x8\n" "ldr d13, [x23], #0x8\n" - "tbz x14, #0, 44f\n" - "ld1 { v9.s }[2], [x12]\n" + "tbz x17, #0, 44f\n" + "ld1 { v9.s }[2], [x15]\n" "ld1 { v13.s }[2], [x23]\n" "b 44f\n" "41:" // Height 2: Partial accumulate: partial_1_4 "mov x24, #0x10\n" - "tbz x14, #0, 44f\n" - "ldr s9, [x12, #0x0]\n" + "tbz x17, #0, 44f\n" + "ldr s9, [x15, #0x0]\n" "ldr s13, [x23, #0x0]\n" "b 44f\n" "42:" // Height 2: Partial accumulate: partial_2_0 - "tbz x14, #1, 43f\n" - "ldr d8, [x12], #0x8\n" + "tbz x17, #1, 43f\n" + "ldr d8, [x15], #0x8\n" "ldr d12, [x23], #0x8\n" "mov x24, #0x8\n" - "tbz x14, #0, 44f\n" - "ld1 { v8.s }[2], [x12]\n" + "tbz x17, #0, 44f\n" + "ld1 { v8.s }[2], [x15]\n" "ld1 { v12.s }[2], [x23]\n" "b 44f\n" "43:" // Height 2: Partial accumulate: partial_1_0 - "ldr s8, [x12, #0x0]\n" + "ldr s8, [x15, #0x0]\n" "mov x24, #0x0\n" "ldr s12, [x23, #0x0]\n" "44:" // Height 2: Partial accumulate: Done - "sub x12, x12, x24\n" + "sub x15, x15, x24\n" "b 47f\n" "45:" // Height 2: full accumulate - "ldr q8, [x12, #0x0]\n" - "ldr q9, [x12, #0x10]\n" - "ldr q10, [x12, #0x20]\n" - "ldr q11, [x12, #0x30]\n" + "ldr q8, [x15, #0x0]\n" + "ldr q9, [x15, #0x10]\n" + "ldr q10, [x15, #0x20]\n" + "ldr q11, [x15, #0x30]\n" "ldr q12, [x23, #0x0]\n" "ldr q13, [x23, #0x10]\n" "ldr q14, [x23, #0x20]\n" @@ -495,420 +495,420 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( "movi v14.4s, #0x0\n" "movi v15.4s, #0x0\n" "47:" // Height 2: setup done - "mov x11, #0x0\n" + "mov x14, #0x0\n" "48:" // Height 2: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w10, [x20, x11, LSL #0x2]\n" + "ldr w13, [x20, x14, LSL #0x2]\n" "tbz %x[flags], #3, 49f\n" - "ldr x20, [%x[input_ptr], x11, LSL #0x3]\n" + "ldr x20, [%x[input_ptr], x14, LSL #0x3]\n" "add x20, x20, x19, LSL #3\n" - "ldr x9, [x20, #0x0]\n" - "ldr x26, [x20, #0x8]\n" - "cbnz x11, 50f\n" + "ldr x12, [x20, #0x0]\n" + "ldr x28, [x20, #0x8]\n" + "cbnz x14, 50f\n" "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x9, x9, x19\n" - "add x26, x26, x19\n" + "add x12, x12, x19\n" + "add x28, x28, x19\n" "b 50f\n" "49:" // Height 2: setup direct input - "mov x9, %x[input_ptr]\n" - "add x26, x9, x19\n" + "mov x12, %x[input_ptr]\n" + "add x28, x12, x19\n" "50:" // Height 2: input setup done - "cmp x10, #0x10\n" + "cmp x13, #0x10\n" "blt 53f\n" - "ldr q0, [x9, #0x0]\n" - "ldr q1, [x26, #0x0]\n" - "cmp x10, #0x20\n" - "ldr q6, [x13, #0x0]\n" + "ldr q0, [x12, #0x0]\n" + "ldr q1, [x28, #0x0]\n" + "cmp x13, #0x20\n" + "ldr q6, [x16, #0x0]\n" "blt 52f\n" "51:" // Height 2: Multiply loop: Main loop head ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" + "ldr d7, [x16, #0x10]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" - "ldr d7, [x13, #0x10]\n" - "add x9, x9, #0x10\n" - "ldr x19, [x13, #0x18]\n" - "add x26, x26, #0x10\n" - "prfm pldl1keep, [x9, #0x80]\n" - "sub x10, x10, #0x10\n" - "ldr d6, [x13, #0x20]\n" - "cmp x10, #0x20\n" - "mov v7.d[1], x19\n" - "ldr x28, [x13, #0x28]\n" - "ldr x19, [x13, #0x38]\n" + "ldr x11, [x16, #0x18]\n" + "ldr d6, [x16, #0x20]\n" + "add x12, x12, #0x10\n" + "ldr x10, [x16, #0x28]\n" + "add x28, x28, #0x10\n" + "mov v7.d[1], x11\n" + "prfm pldl1keep, [x12, #0x80]\n" + "prfm pldl1keep, [x28, #0x80]\n" + "sub x13, x13, #0x10\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" + "mov v6.d[1], x10\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" - "mov v6.d[1], x28\n" - "ldr d7, [x13, #0x30]\n" - "prfm pldl1keep, [x26, #0x80]\n" + "ldr d7, [x16, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "ldr x11, [x16, #0x38]\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" - "mov v7.d[1], x19\n" - "ldr d6, [x13, #0x40]\n" - "ldr x28, [x13, #0x48]\n" + "ldr d6, [x16, #0x40]\n" + "ldr x10, [x16, #0x48]\n" + "cmp x13, #0x20\n" + "mov v7.d[1], x11\n" + "ldr x11, [x16, #0x58]\n" + "ldr x9, [x12, #0x8]\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" + "mov v6.d[1], x10\n" ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" - "ldr d7, [x13, #0x50]\n" - "mov v6.d[1], x28\n" - "ldr x19, [x13, #0x58]\n" - "ldr x28, [x13, #0x68]\n" + "ldr d7, [x16, #0x50]\n" ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n" + "ldr x10, [x16, #0x68]\n" ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n" - "mov v7.d[1], x19\n" - "ldr d6, [x13, #0x60]\n" - "ldr x19, [x13, #0x78]\n" + "ldr d6, [x16, #0x60]\n" + "mov v7.d[1], x11\n" + "ldr x11, [x16, #0x78]\n" + "ldr x27, [x28, #0x8]\n" ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n" + "mov v6.d[1], x10\n" ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n" - "mov v6.d[1], x28\n" - "ldr d7, [x13, #0x70]\n" - "ldr x28, [x13, #0x88]\n" + "ldr d7, [x16, #0x70]\n" ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n" + "ldr x10, [x16, #0x88]\n" ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n" - "mov v7.d[1], x19\n" - "ldr d6, [x13, #0x80]\n" - "ldr x19, [x13, #0x98]\n" + "ldr d6, [x16, #0x80]\n" + "mov v7.d[1], x11\n" + "ldr x11, [x16, #0x98]\n" ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n" + "mov v6.d[1], x10\n" ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n" - "mov v6.d[1], x28\n" - "ldr d7, [x13, #0x90]\n" - "ldr x28, [x13, #0xa8]\n" + "ldr d7, [x16, #0x90]\n" ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n" + "ldr x10, [x16, #0xa8]\n" ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n" - "mov v7.d[1], x19\n" - "ldr d6, [x13, #0xa0]\n" - "ldr x19, [x13, #0xb8]\n" + "ldr d6, [x16, #0xa0]\n" + "mov v7.d[1], x11\n" + "ldr x11, [x16, #0xb8]\n" ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n" + "mov v6.d[1], x10\n" ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n" - "mov v6.d[1], x28\n" - "ldr d7, [x13, #0xb0]\n" - "ldr x28, [x13, #0xc8]\n" + "ldr d7, [x16, #0xb0]\n" ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n" + "ldr x10, [x16, #0xc8]\n" ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n" - "mov v7.d[1], x19\n" - "ldr d6, [x13, #0xc0]\n" - "ldr x19, [x13, #0xd8]\n" + "ldr d6, [x16, #0xc0]\n" + "mov v7.d[1], x11\n" + "ldr x11, [x16, #0xd8]\n" ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n" + "mov v6.d[1], x10\n" ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n" - "mov v6.d[1], x28\n" - "ldr d7, [x13, #0xd0]\n" - "ldr x28, [x13, #0xe8]\n" + "ldr d7, [x16, #0xd0]\n" ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n" + "ldr x10, [x16, #0xe8]\n" ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n" - "mov v7.d[1], x19\n" - "ldr d6, [x13, #0xe0]\n" - "ldr x19, [x13, #0xf8]\n" + "ldr d6, [x16, #0xe0]\n" + "mov v7.d[1], x11\n" + "ldr x11, [x16, #0xf8]\n" ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n" + "mov v6.d[1], x10\n" ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n" - "mov v6.d[1], x28\n" - "ldr d7, [x13, #0xf0]\n" - "ldr x27, [x9, #0x8]\n" - "add x13, x13, #0x100\n" + "ldr d7, [x16, #0xf0]\n" ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n" + "add x16, x16, #0x100\n" ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n" - "mov v7.d[1], x19\n" - "ldr x25, [x26, #0x8]\n" - "ldr d6, [x13, #0x0]\n" + "ldr d6, [x16, #0x0]\n" + "mov v7.d[1], x11\n" + "ldr x10, [x16, #0x8]\n" ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n" + "ldr d0, [x12, #0x0]\n" ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n" - "ldr d0, [x9, #0x0]\n" - "ldr d1, [x26, #0x0]\n" - "ldr x28, [x13, #0x8]\n" - "mov v0.d[1], x27\n" - "mov v1.d[1], x25\n" - "mov v6.d[1], x28\n" + "mov v6.d[1], x10\n" + "ldr d1, [x28, #0x0]\n" + "mov v0.d[1], x9\n" + "mov v1.d[1], x27\n" "bge 51b\n" "52:" // Height 2: Multiply loop: Single iteration only ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" + "ldr q7, [x16, #0x10]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" - "ldr q7, [x13, #0x10]\n" - "sub x10, x10, #0x10\n" - "ldr q6, [x13, #0x20]\n" - "add x9, x9, #0x10\n" - "prfm pldl1keep, [x9, #0x80]\n" - "add x26, x26, #0x10\n" + "ldr q6, [x16, #0x20]\n" + "sub x13, x13, #0x10\n" + "add x12, x12, #0x10\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" + "prfm pldl1keep, [x12, #0x80]\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" + "ldr q7, [x16, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "add x28, x28, #0x10\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" - "ldr q7, [x13, #0x30]\n" - "prfm pldl1keep, [x26, #0x80]\n" - "ldr q6, [x13, #0x40]\n" + "prfm pldl1keep, [x28, #0x80]\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" + "ldr q6, [x16, #0x40]\n" ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" - "ldr q7, [x13, #0x50]\n" + "ldr q7, [x16, #0x50]\n" ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n" ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n" - "ldr q6, [x13, #0x60]\n" + "ldr q6, [x16, #0x60]\n" ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n" ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n" - "ldr q7, [x13, #0x70]\n" + "ldr q7, [x16, #0x70]\n" ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n" ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n" - "ldr q6, [x13, #0x80]\n" + "ldr q6, [x16, #0x80]\n" ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n" ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n" - "ldr q7, [x13, #0x90]\n" + "ldr q7, [x16, #0x90]\n" ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n" ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n" - "ldr q6, [x13, #0xa0]\n" + "ldr q6, [x16, #0xa0]\n" ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n" - "ldr q7, [x13, #0xb0]\n" + "ldr q7, [x16, #0xb0]\n" ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n" ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n" - "ldr q6, [x13, #0xc0]\n" + "ldr q6, [x16, #0xc0]\n" ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n" - "ldr q7, [x13, #0xd0]\n" + "ldr q7, [x16, #0xd0]\n" ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n" ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n" - "ldr q6, [x13, #0xe0]\n" + "ldr q6, [x16, #0xe0]\n" ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n" ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n" - "ldr q7, [x13, #0xf0]\n" - "add x13, x13, #0x100\n" + "ldr q7, [x16, #0xf0]\n" ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n" + "add x16, x16, #0x100\n" ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n" ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n" ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n" "53:" // Height 2: Multiply loop: Main loop skip - "cbz x10, 58f\n" - "cmp x10, #0x4\n" + "cbz x13, 58f\n" + "cmp x13, #0x4\n" "blt 55f\n" "54:" // Height 2: Multiply loop: Odd block loop - "ldr s0, [x9], #0x4\n" - "sub x10, x10, #0x4\n" - "ldr s1, [x26], #0x4\n" - "cmp x10, #0x4\n" - "ldr q6, [x13, #0x0]\n" - "ldr q7, [x13, #0x10]\n" + "ldr s0, [x12], #0x4\n" + "sub x13, x13, #0x4\n" + "ldr s1, [x28], #0x4\n" + "cmp x13, #0x4\n" + "ldr q6, [x16, #0x0]\n" + "ldr q7, [x16, #0x10]\n" ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" + "ldr q6, [x16, #0x20]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" - "ldr q6, [x13, #0x20]\n" - "ldr q7, [x13, #0x30]\n" - "add x13, x13, #0x40\n" + "ldr q7, [x16, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "add x16, x16, #0x40\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" "bge 54b\n" - "cbz x10, 58f\n" + "cbz x13, 58f\n" "55:" // Height 2: Multiply loop: Skip odd blocks - "tbz x10, #1, 56f\n" - "ldr h0, [x9], #0x2\n" - "ldr h1, [x26], #0x2\n" - "tbz x10, #0, 57f\n" - "ld1 { v0.b }[2], [x9]\n" - "ld1 { v1.b }[2], [x26]\n" + "tbz x13, #1, 56f\n" + "ldr h0, [x12], #0x2\n" + "ldr h1, [x28], #0x2\n" + "tbz x13, #0, 57f\n" + "ld1 { v0.b }[2], [x12]\n" + "ld1 { v1.b }[2], [x28]\n" "b 57f\n" "56:" // Height 2: Multiply loop: Ragged operand read: partial_1_0 - "ldr b0, [x9, #0x0]\n" - "ldr b1, [x26, #0x0]\n" + "ldr b0, [x12, #0x0]\n" + "ldr b1, [x28, #0x0]\n" "57:" // Height 2: Multiply loop: Ragged operand read: Done - "ldr q6, [x13, #0x0]\n" - "ldr q7, [x13, #0x10]\n" + "ldr q6, [x16, #0x0]\n" + "ldr q7, [x16, #0x10]\n" ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" + "ldr q6, [x16, #0x20]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" - "ldr q6, [x13, #0x20]\n" - "ldr q7, [x13, #0x30]\n" - "add x13, x13, #0x40\n" + "ldr q7, [x16, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "add x16, x16, #0x40\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" "58:" // Height 2: Multiply loop: No odd multiplies "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" - "add x11, x11, #0x1\n" - "cmp x11, x19\n" + "add x14, x14, #0x1\n" + "cmp x14, x19\n" "bne 48b\n" "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "prfm pstl1keep, [x12, #0x0]\n" - "cmp x14, #0x10\n" - "add x23, x12, x19, LSL #2\n" + "prfm pstl1keep, [x15, #0x0]\n" + "cmp x17, #0x10\n" + "add x23, x15, x19, LSL #2\n" "prfm pstl1keep, [x23, #0x0]\n" "bge 67f\n" - "tbz x14, #3, 62f\n" - "st1 { v8.4s }, [x12], #0x10\n" - "st1 { v9.4s }, [x12], #0x10\n" + "tbz x17, #3, 62f\n" + "st1 { v8.4s }, [x15], #0x10\n" + "st1 { v9.4s }, [x15], #0x10\n" "st1 { v12.4s }, [x23], #0x10\n" "st1 { v13.4s }, [x23], #0x10\n" - "tbz x14, #2, 60f\n" - "st1 { v10.4s }, [x12], #0x10\n" + "tbz x17, #2, 60f\n" + "st1 { v10.4s }, [x15], #0x10\n" "st1 { v14.4s }, [x23], #0x10\n" - "tbz x14, #1, 59f\n" - "str d11, [x12], #0x8\n" + "tbz x17, #1, 59f\n" + "str d11, [x15], #0x8\n" "str d15, [x23], #0x8\n" - "tbz x14, #0, 66f\n" - "st1 { v11.s }[2], [x12]\n" + "tbz x17, #0, 66f\n" + "st1 { v11.s }[2], [x15]\n" "st1 { v15.s }[2], [x23]\n" "b 66f\n" "59:" // Height 2: Partial direct writeback: partial_1_12 - "tbz x14, #0, 66f\n" - "str s11, [x12, #0x0]\n" + "tbz x17, #0, 66f\n" + "str s11, [x15, #0x0]\n" "str s15, [x23, #0x0]\n" "b 66f\n" "60:" // Height 2: Partial direct writeback: partial_2_8 - "tbz x14, #1, 61f\n" - "str d10, [x12], #0x8\n" + "tbz x17, #1, 61f\n" + "str d10, [x15], #0x8\n" "str d14, [x23], #0x8\n" - "tbz x14, #0, 66f\n" - "st1 { v10.s }[2], [x12]\n" + "tbz x17, #0, 66f\n" + "st1 { v10.s }[2], [x15]\n" "st1 { v14.s }[2], [x23]\n" "b 66f\n" "61:" // Height 2: Partial direct writeback: partial_1_8 - "tbz x14, #0, 66f\n" - "str s10, [x12, #0x0]\n" + "tbz x17, #0, 66f\n" + "str s10, [x15, #0x0]\n" "str s14, [x23, #0x0]\n" "b 66f\n" "62:" // Height 2: Partial direct writeback: partial_4_0 - "tbz x14, #2, 64f\n" - "st1 { v8.4s }, [x12], #0x10\n" + "tbz x17, #2, 64f\n" + "st1 { v8.4s }, [x15], #0x10\n" "st1 { v12.4s }, [x23], #0x10\n" - "tbz x14, #1, 63f\n" - "str d9, [x12], #0x8\n" + "tbz x17, #1, 63f\n" + "str d9, [x15], #0x8\n" "str d13, [x23], #0x8\n" - "tbz x14, #0, 66f\n" - "st1 { v9.s }[2], [x12]\n" + "tbz x17, #0, 66f\n" + "st1 { v9.s }[2], [x15]\n" "st1 { v13.s }[2], [x23]\n" "b 66f\n" "63:" // Height 2: Partial direct writeback: partial_1_4 - "tbz x14, #0, 66f\n" - "str s9, [x12, #0x0]\n" + "tbz x17, #0, 66f\n" + "str s9, [x15, #0x0]\n" "str s13, [x23, #0x0]\n" "b 66f\n" "64:" // Height 2: Partial direct writeback: partial_2_0 - "tbz x14, #1, 65f\n" - "str d8, [x12], #0x8\n" + "tbz x17, #1, 65f\n" + "str d8, [x15], #0x8\n" "str d12, [x23], #0x8\n" - "tbz x14, #0, 66f\n" - "st1 { v8.s }[2], [x12]\n" + "tbz x17, #0, 66f\n" + "st1 { v8.s }[2], [x15]\n" "st1 { v12.s }[2], [x23]\n" "b 66f\n" "65:" // Height 2: Partial direct writeback: partial_1_0 - "str s8, [x12, #0x0]\n" + "str s8, [x15, #0x0]\n" "str s12, [x23, #0x0]\n" "66:" // Height 2: Partial direct writeback: Done "b 68f\n" "67:" // Height 2: Full writeback - "str q8, [x12, #0x0]\n" - "str q9, [x12, #0x10]\n" - "str q10, [x12, #0x20]\n" - "str q11, [x12, #0x30]\n" - "add x12, x12, #0x40\n" + "str q8, [x15, #0x0]\n" + "str q9, [x15, #0x10]\n" + "str q10, [x15, #0x20]\n" + "str q11, [x15, #0x30]\n" + "add x15, x15, #0x40\n" "str q12, [x23, #0x0]\n" "str q13, [x23, #0x10]\n" "str q14, [x23, #0x20]\n" "str q15, [x23, #0x30]\n" "68:" // Height 2: Writeback done - "subs x14, x14, #0x10\n" + "subs x17, x17, #0x10\n" "bgt 36b\n" "b 206f\n" "69:" // Height 3 - "ldr x14, [%x[args_ptr], %[offsetof_N]]\n" - "mov x12, %x[output_ptr]\n" - "ldr x13, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "ldr x17, [%x[args_ptr], %[offsetof_N]]\n" + "mov x15, %x[output_ptr]\n" + "ldr x16, [%x[args_ptr], %[offsetof_B_ptr]]\n" "70:" // Height 3: Column loop "tbz %x[flags], #0, 80f\n" "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "cmp x14, #0x10\n" - "add x23, x12, x19, LSL #2\n" + "cmp x17, #0x10\n" + "add x23, x15, x19, LSL #2\n" "add x22, x23, x19, LSL #2\n" "bge 79f\n" - "tbz x14, #3, 74f\n" - "ld1 { v8.4s }, [x12], #0x10\n" + "tbz x17, #3, 74f\n" + "ld1 { v8.4s }, [x15], #0x10\n" "ld1 { v12.4s }, [x23], #0x10\n" "ld1 { v16.4s }, [x22], #0x10\n" - "ld1 { v9.4s }, [x12], #0x10\n" + "ld1 { v9.4s }, [x15], #0x10\n" "ld1 { v13.4s }, [x23], #0x10\n" "ld1 { v17.4s }, [x22], #0x10\n" - "tbz x14, #2, 72f\n" - "ld1 { v10.4s }, [x12], #0x10\n" + "tbz x17, #2, 72f\n" + "ld1 { v10.4s }, [x15], #0x10\n" "ld1 { v14.4s }, [x23], #0x10\n" "ld1 { v18.4s }, [x22], #0x10\n" - "tbz x14, #1, 71f\n" - "ldr d11, [x12], #0x8\n" + "tbz x17, #1, 71f\n" + "ldr d11, [x15], #0x8\n" "mov x24, #0x38\n" "ldr d15, [x23], #0x8\n" "ldr d19, [x22], #0x8\n" - "tbz x14, #0, 78f\n" - "ld1 { v11.s }[2], [x12]\n" + "tbz x17, #0, 78f\n" + "ld1 { v11.s }[2], [x15]\n" "ld1 { v15.s }[2], [x23]\n" "ld1 { v19.s }[2], [x22]\n" "b 78f\n" "71:" // Height 3: Partial accumulate: partial_1_12 "mov x24, #0x30\n" - "tbz x14, #0, 78f\n" - "ldr s11, [x12, #0x0]\n" + "tbz x17, #0, 78f\n" + "ldr s11, [x15, #0x0]\n" "ldr s15, [x23, #0x0]\n" "ldr s19, [x22, #0x0]\n" "b 78f\n" "72:" // Height 3: Partial accumulate: partial_2_8 - "tbz x14, #1, 73f\n" - "ldr d10, [x12], #0x8\n" + "tbz x17, #1, 73f\n" + "ldr d10, [x15], #0x8\n" "ldr d14, [x23], #0x8\n" "mov x24, #0x28\n" "ldr d18, [x22], #0x8\n" - "tbz x14, #0, 78f\n" - "ld1 { v10.s }[2], [x12]\n" + "tbz x17, #0, 78f\n" + "ld1 { v10.s }[2], [x15]\n" "ld1 { v14.s }[2], [x23]\n" "ld1 { v18.s }[2], [x22]\n" "b 78f\n" "73:" // Height 3: Partial accumulate: partial_1_8 "mov x24, #0x20\n" - "tbz x14, #0, 78f\n" - "ldr s10, [x12, #0x0]\n" + "tbz x17, #0, 78f\n" + "ldr s10, [x15, #0x0]\n" "ldr s14, [x23, #0x0]\n" "ldr s18, [x22, #0x0]\n" "b 78f\n" "74:" // Height 3: Partial accumulate: partial_4_0 - "tbz x14, #2, 76f\n" - "ld1 { v8.4s }, [x12], #0x10\n" + "tbz x17, #2, 76f\n" + "ld1 { v8.4s }, [x15], #0x10\n" "ld1 { v12.4s }, [x23], #0x10\n" "ld1 { v16.4s }, [x22], #0x10\n" - "tbz x14, #1, 75f\n" - "ldr d9, [x12], #0x8\n" + "tbz x17, #1, 75f\n" + "ldr d9, [x15], #0x8\n" "mov x24, #0x18\n" "ldr d13, [x23], #0x8\n" "ldr d17, [x22], #0x8\n" - "tbz x14, #0, 78f\n" - "ld1 { v9.s }[2], [x12]\n" + "tbz x17, #0, 78f\n" + "ld1 { v9.s }[2], [x15]\n" "ld1 { v13.s }[2], [x23]\n" "ld1 { v17.s }[2], [x22]\n" "b 78f\n" "75:" // Height 3: Partial accumulate: partial_1_4 "mov x24, #0x10\n" - "tbz x14, #0, 78f\n" - "ldr s9, [x12, #0x0]\n" + "tbz x17, #0, 78f\n" + "ldr s9, [x15, #0x0]\n" "ldr s13, [x23, #0x0]\n" "ldr s17, [x22, #0x0]\n" "b 78f\n" "76:" // Height 3: Partial accumulate: partial_2_0 - "tbz x14, #1, 77f\n" - "ldr d8, [x12], #0x8\n" + "tbz x17, #1, 77f\n" + "ldr d8, [x15], #0x8\n" "ldr d12, [x23], #0x8\n" "mov x24, #0x8\n" "ldr d16, [x22], #0x8\n" - "tbz x14, #0, 78f\n" - "ld1 { v8.s }[2], [x12]\n" + "tbz x17, #0, 78f\n" + "ld1 { v8.s }[2], [x15]\n" "ld1 { v12.s }[2], [x23]\n" "ld1 { v16.s }[2], [x22]\n" "b 78f\n" "77:" // Height 3: Partial accumulate: partial_1_0 - "ldr s8, [x12, #0x0]\n" + "ldr s8, [x15, #0x0]\n" "mov x24, #0x0\n" "ldr s12, [x23, #0x0]\n" "ldr s16, [x22, #0x0]\n" "78:" // Height 3: Partial accumulate: Done - "sub x12, x12, x24\n" + "sub x15, x15, x24\n" "b 81f\n" "79:" // Height 3: full accumulate - "ldr q8, [x12, #0x0]\n" - "ldr q9, [x12, #0x10]\n" - "ldr q10, [x12, #0x20]\n" - "ldr q11, [x12, #0x30]\n" + "ldr q8, [x15, #0x0]\n" + "ldr q9, [x15, #0x10]\n" + "ldr q10, [x15, #0x20]\n" + "ldr q11, [x15, #0x30]\n" "ldr q12, [x23, #0x0]\n" "ldr q13, [x23, #0x10]\n" "ldr q14, [x23, #0x20]\n" @@ -932,280 +932,280 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( "movi v18.4s, #0x0\n" "movi v19.4s, #0x0\n" "81:" // Height 3: setup done - "mov x11, #0x0\n" + "mov x14, #0x0\n" "82:" // Height 3: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w10, [x20, x11, LSL #0x2]\n" + "ldr w13, [x20, x14, LSL #0x2]\n" "tbz %x[flags], #3, 83f\n" - "ldr x20, [%x[input_ptr], x11, LSL #0x3]\n" + "ldr x20, [%x[input_ptr], x14, LSL #0x3]\n" "add x20, x20, x19, LSL #3\n" - "ldr x9, [x20, #0x0]\n" - "ldr x26, [x20, #0x8]\n" - "ldr x24, [x20, #0x10]\n" - "cbnz x11, 84f\n" + "ldr x12, [x20, #0x0]\n" + "ldr x28, [x20, #0x8]\n" + "ldr x26, [x20, #0x10]\n" + "cbnz x14, 84f\n" "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x9, x9, x19\n" + "add x12, x12, x19\n" + "add x28, x28, x19\n" "add x26, x26, x19\n" - "add x24, x24, x19\n" "b 84f\n" "83:" // Height 3: setup direct input - "mov x9, %x[input_ptr]\n" - "add x26, x9, x19\n" - "add x24, x26, x19\n" + "mov x12, %x[input_ptr]\n" + "add x28, x12, x19\n" + "add x26, x28, x19\n" "84:" // Height 3: input setup done - "cmp x10, #0x10\n" + "cmp x13, #0x10\n" "blt 87f\n" - "ldr q0, [x9, #0x0]\n" - "ldr q1, [x26, #0x0]\n" - "cmp x10, #0x20\n" - "ldr q2, [x24, #0x0]\n" - "ldr q6, [x13, #0x0]\n" + "ldr q0, [x12, #0x0]\n" + "ldr q1, [x28, #0x0]\n" + "cmp x13, #0x20\n" + "ldr q2, [x26, #0x0]\n" + "ldr q6, [x16, #0x0]\n" "blt 86f\n" "85:" // Height 3: Multiply loop: Main loop head ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" + "ldr d7, [x16, #0x10]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" + "ldr x11, [x16, #0x18]\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" - "ldr d7, [x13, #0x10]\n" - "ldr x19, [x13, #0x18]\n" - "add x9, x9, #0x10\n" - "ldr d6, [x13, #0x20]\n" - "add x26, x26, #0x10\n" - "ldr x28, [x13, #0x28]\n" - "add x24, x24, #0x10\n" - "mov v7.d[1], x19\n" - "prfm pldl1keep, [x9, #0x80]\n" - "ldr x19, [x13, #0x38]\n" - "sub x10, x10, #0x10\n" - "mov v6.d[1], x28\n" + "ldr d6, [x16, #0x20]\n" + "ldr x10, [x16, #0x28]\n" + "add x12, x12, #0x10\n" + "mov v7.d[1], x11\n" + "prfm pldl1keep, [x12, #0x80]\n" + "ldr x11, [x16, #0x38]\n" + "add x28, x28, #0x10\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" + "mov v6.d[1], x10\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" + "prfm pldl1keep, [x28, #0x80]\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" + "ldr d7, [x16, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "ldr x10, [x16, #0x48]\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" + "ldr x9, [x12, #0x8]\n" ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" - "ldr d7, [x13, #0x30]\n" - "prfm pldl1keep, [x26, #0x80]\n" - "cmp x10, #0x20\n" - "ldr d6, [x13, #0x40]\n" - "mov v7.d[1], x19\n" - "ldr x28, [x13, #0x48]\n" - "ldr x19, [x13, #0x58]\n" + "mov v7.d[1], x11\n" + "ldr d6, [x16, #0x40]\n" + "add x26, x26, #0x10\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" + "prfm pldl1keep, [x26, #0x80]\n" ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" - "mov v6.d[1], x28\n" + "ldr x11, [x16, #0x58]\n" ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" - "ldr d7, [x13, #0x50]\n" + "mov v6.d[1], x10\n" + "ldr d7, [x16, #0x50]\n" + "sub x13, x13, #0x10\n" ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n" + "ldr x10, [x16, #0x68]\n" ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n" + "ldr x27, [x28, #0x8]\n" ".inst 0x4fa2e0d0 // sdot v16.4s, v6.16b, v2.4b[1]\n" - "prfm pldl1keep, [x24, #0x80]\n" - "mov v7.d[1], x19\n" - "ldr d6, [x13, #0x60]\n" - "ldr x28, [x13, #0x68]\n" + "mov v7.d[1], x11\n" + "ldr d6, [x16, #0x60]\n" + "cmp x13, #0x20\n" ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n" + "ldr x11, [x16, #0x78]\n" ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n" + "ldr x25, [x26, #0x8]\n" ".inst 0x4fa2e0f1 // sdot v17.4s, v7.16b, v2.4b[1]\n" - "ldr d7, [x13, #0x70]\n" - "mov v6.d[1], x28\n" - "ldr x19, [x13, #0x78]\n" - "ldr x28, [x13, #0x88]\n" + "mov v6.d[1], x10\n" + "ldr d7, [x16, #0x70]\n" ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n" + "ldr x10, [x16, #0x88]\n" ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n" - "mov v7.d[1], x19\n" ".inst 0x4fa2e0d2 // sdot v18.4s, v6.16b, v2.4b[1]\n" - "ldr d6, [x13, #0x80]\n" + "mov v7.d[1], x11\n" + "ldr d6, [x16, #0x80]\n" ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n" + "ldr x11, [x16, #0x98]\n" ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n" ".inst 0x4fa2e0f3 // sdot v19.4s, v7.16b, v2.4b[1]\n" - "ldr d7, [x13, #0x90]\n" - "mov v6.d[1], x28\n" - "ldr x19, [x13, #0x98]\n" - "ldr x28, [x13, #0xa8]\n" + "mov v6.d[1], x10\n" + "ldr d7, [x16, #0x90]\n" ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n" + "ldr x10, [x16, #0xa8]\n" ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n" - "mov v7.d[1], x19\n" ".inst 0x4f82e8d0 // sdot v16.4s, v6.16b, v2.4b[2]\n" - "ldr d6, [x13, #0xa0]\n" + "mov v7.d[1], x11\n" + "ldr d6, [x16, #0xa0]\n" ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n" + "ldr x11, [x16, #0xb8]\n" ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n" ".inst 0x4f82e8f1 // sdot v17.4s, v7.16b, v2.4b[2]\n" - "ldr d7, [x13, #0xb0]\n" - "mov v6.d[1], x28\n" - "ldr x19, [x13, #0xb8]\n" - "ldr x28, [x13, #0xc8]\n" + "mov v6.d[1], x10\n" + "ldr d7, [x16, #0xb0]\n" ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n" + "ldr x10, [x16, #0xc8]\n" ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n" - "mov v7.d[1], x19\n" ".inst 0x4f82e8d2 // sdot v18.4s, v6.16b, v2.4b[2]\n" - "ldr d6, [x13, #0xc0]\n" + "mov v7.d[1], x11\n" + "ldr d6, [x16, #0xc0]\n" ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n" + "ldr x11, [x16, #0xd8]\n" ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n" ".inst 0x4f82e8f3 // sdot v19.4s, v7.16b, v2.4b[2]\n" - "ldr d7, [x13, #0xd0]\n" - "mov v6.d[1], x28\n" - "ldr x19, [x13, #0xd8]\n" - "ldr x28, [x13, #0xe8]\n" + "mov v6.d[1], x10\n" + "ldr d7, [x16, #0xd0]\n" ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n" + "ldr x10, [x16, #0xe8]\n" ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n" - "mov v7.d[1], x19\n" ".inst 0x4fa2e8d0 // sdot v16.4s, v6.16b, v2.4b[3]\n" - "ldr d6, [x13, #0xe0]\n" + "mov v7.d[1], x11\n" + "ldr d6, [x16, #0xe0]\n" ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n" + "ldr x11, [x16, #0xf8]\n" ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n" ".inst 0x4fa2e8f1 // sdot v17.4s, v7.16b, v2.4b[3]\n" - "ldr d7, [x13, #0xf0]\n" - "mov v6.d[1], x28\n" - "ldr x19, [x13, #0xf8]\n" - "ldr x27, [x9, #0x8]\n" - "add x13, x13, #0x100\n" + "mov v6.d[1], x10\n" + "ldr d7, [x16, #0xf0]\n" + "add x16, x16, #0x100\n" ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n" + "ldr x10, [x16, #0x8]\n" ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n" - "mov v7.d[1], x19\n" ".inst 0x4fa2e8d2 // sdot v18.4s, v6.16b, v2.4b[3]\n" - "ldr x25, [x26, #0x8]\n" + "mov v7.d[1], x11\n" + "ldr d6, [x16, #0x0]\n" ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n" + "ldr d0, [x12, #0x0]\n" ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n" + "ldr d1, [x28, #0x0]\n" ".inst 0x4fa2e8f3 // sdot v19.4s, v7.16b, v2.4b[3]\n" - "ldr d0, [x9, #0x0]\n" - "ldr d1, [x26, #0x0]\n" - "ldr d2, [x24, #0x0]\n" - "mov v0.d[1], x27\n" - "ldr x21, [x24, #0x8]\n" - "mov v1.d[1], x25\n" - "ldr d6, [x13, #0x0]\n" - "ldr x28, [x13, #0x8]\n" - "mov v2.d[1], x21\n" - "mov v6.d[1], x28\n" + "mov v6.d[1], x10\n" + "mov v0.d[1], x9\n" + "ldr d2, [x26, #0x0]\n" + "mov v1.d[1], x27\n" + "mov v2.d[1], x25\n" "bge 85b\n" "86:" // Height 3: Multiply loop: Single iteration only ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" + "ldr q7, [x16, #0x10]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" + "sub x13, x13, #0x10\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" - "ldr q7, [x13, #0x10]\n" - "ldr q6, [x13, #0x20]\n" - "sub x10, x10, #0x10\n" - "add x9, x9, #0x10\n" - "add x26, x26, #0x10\n" + "ldr q6, [x16, #0x20]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" + "add x12, x12, #0x10\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" + "prfm pldl1keep, [x12, #0x80]\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" + "ldr q7, [x16, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "add x28, x28, #0x10\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" + "prfm pldl1keep, [x28, #0x80]\n" ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" - "prfm pldl1keep, [x9, #0x80]\n" - "add x24, x24, #0x10\n" - "ldr q7, [x13, #0x30]\n" - "prfm pldl1keep, [x26, #0x80]\n" - "ldr q6, [x13, #0x40]\n" + "ldr q6, [x16, #0x40]\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" + "add x26, x26, #0x10\n" ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" + "prfm pldl1keep, [x26, #0x80]\n" ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" - "ldr q7, [x13, #0x50]\n" + "ldr q7, [x16, #0x50]\n" ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n" ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n" ".inst 0x4fa2e0d0 // sdot v16.4s, v6.16b, v2.4b[1]\n" - "prfm pldl1keep, [x24, #0x80]\n" + "ldr q6, [x16, #0x60]\n" ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n" ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n" ".inst 0x4fa2e0f1 // sdot v17.4s, v7.16b, v2.4b[1]\n" - "ldr q6, [x13, #0x60]\n" - "ldr q7, [x13, #0x70]\n" + "ldr q7, [x16, #0x70]\n" ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n" ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n" ".inst 0x4fa2e0d2 // sdot v18.4s, v6.16b, v2.4b[1]\n" + "ldr q6, [x16, #0x80]\n" ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n" ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n" ".inst 0x4fa2e0f3 // sdot v19.4s, v7.16b, v2.4b[1]\n" - "ldr q6, [x13, #0x80]\n" - "ldr q7, [x13, #0x90]\n" + "ldr q7, [x16, #0x90]\n" ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n" ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n" ".inst 0x4f82e8d0 // sdot v16.4s, v6.16b, v2.4b[2]\n" + "ldr q6, [x16, #0xa0]\n" ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n" ".inst 0x4f82e8f1 // sdot v17.4s, v7.16b, v2.4b[2]\n" - "ldr q6, [x13, #0xa0]\n" - "ldr q7, [x13, #0xb0]\n" + "ldr q7, [x16, #0xb0]\n" ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n" ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n" ".inst 0x4f82e8d2 // sdot v18.4s, v6.16b, v2.4b[2]\n" + "ldr q6, [x16, #0xc0]\n" ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n" ".inst 0x4f82e8f3 // sdot v19.4s, v7.16b, v2.4b[2]\n" - "ldr q6, [x13, #0xc0]\n" - "ldr q7, [x13, #0xd0]\n" + "ldr q7, [x16, #0xd0]\n" ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n" ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n" ".inst 0x4fa2e8d0 // sdot v16.4s, v6.16b, v2.4b[3]\n" + "ldr q6, [x16, #0xe0]\n" ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n" ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n" ".inst 0x4fa2e8f1 // sdot v17.4s, v7.16b, v2.4b[3]\n" - "ldr q6, [x13, #0xe0]\n" - "ldr q7, [x13, #0xf0]\n" - "add x13, x13, #0x100\n" + "ldr q7, [x16, #0xf0]\n" ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n" + "add x16, x16, #0x100\n" ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n" ".inst 0x4fa2e8d2 // sdot v18.4s, v6.16b, v2.4b[3]\n" ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n" ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n" ".inst 0x4fa2e8f3 // sdot v19.4s, v7.16b, v2.4b[3]\n" "87:" // Height 3: Multiply loop: Main loop skip - "cbz x10, 92f\n" - "cmp x10, #0x4\n" + "cbz x13, 92f\n" + "cmp x13, #0x4\n" "blt 89f\n" "88:" // Height 3: Multiply loop: Odd block loop - "ldr s0, [x9], #0x4\n" - "sub x10, x10, #0x4\n" - "ldr s1, [x26], #0x4\n" - "cmp x10, #0x4\n" - "ldr s2, [x24], #0x4\n" - "ldr q6, [x13, #0x0]\n" - "ldr q7, [x13, #0x10]\n" + "ldr s0, [x12], #0x4\n" + "sub x13, x13, #0x4\n" + "ldr s1, [x28], #0x4\n" + "cmp x13, #0x4\n" + "ldr s2, [x26], #0x4\n" + "ldr q6, [x16, #0x0]\n" + "ldr q7, [x16, #0x10]\n" ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" + "ldr q6, [x16, #0x20]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" - "ldr q6, [x13, #0x20]\n" - "ldr q7, [x13, #0x30]\n" - "add x13, x13, #0x40\n" + "ldr q7, [x16, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "add x16, x16, #0x40\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" "bge 88b\n" - "cbz x10, 92f\n" + "cbz x13, 92f\n" "89:" // Height 3: Multiply loop: Skip odd blocks - "tbz x10, #1, 90f\n" - "ldr h0, [x9], #0x2\n" - "ldr h1, [x26], #0x2\n" - "ldr h2, [x24], #0x2\n" - "tbz x10, #0, 91f\n" - "ld1 { v0.b }[2], [x9]\n" - "ld1 { v1.b }[2], [x26]\n" - "ld1 { v2.b }[2], [x24]\n" + "tbz x13, #1, 90f\n" + "ldr h0, [x12], #0x2\n" + "ldr h1, [x28], #0x2\n" + "ldr h2, [x26], #0x2\n" + "tbz x13, #0, 91f\n" + "ld1 { v0.b }[2], [x12]\n" + "ld1 { v1.b }[2], [x28]\n" + "ld1 { v2.b }[2], [x26]\n" "b 91f\n" "90:" // Height 3: Multiply loop: Ragged operand read: partial_1_0 - "ldr b0, [x9, #0x0]\n" - "ldr b1, [x26, #0x0]\n" - "ldr b2, [x24, #0x0]\n" + "ldr b0, [x12, #0x0]\n" + "ldr b1, [x28, #0x0]\n" + "ldr b2, [x26, #0x0]\n" "91:" // Height 3: Multiply loop: Ragged operand read: Done - "ldr q6, [x13, #0x0]\n" - "ldr q7, [x13, #0x10]\n" + "ldr q6, [x16, #0x0]\n" + "ldr q7, [x16, #0x10]\n" ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" + "ldr q6, [x16, #0x20]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" - "ldr q6, [x13, #0x20]\n" - "ldr q7, [x13, #0x30]\n" - "add x13, x13, #0x40\n" + "ldr q7, [x16, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "add x16, x16, #0x40\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" @@ -1213,101 +1213,101 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" "92:" // Height 3: Multiply loop: No odd multiplies "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" - "add x11, x11, #0x1\n" - "cmp x11, x19\n" + "add x14, x14, #0x1\n" + "cmp x14, x19\n" "bne 82b\n" "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "prfm pstl1keep, [x12, #0x0]\n" - "cmp x14, #0x10\n" - "add x23, x12, x19, LSL #2\n" + "prfm pstl1keep, [x15, #0x0]\n" + "cmp x17, #0x10\n" + "add x23, x15, x19, LSL #2\n" "prfm pstl1keep, [x23, #0x0]\n" "add x22, x23, x19, LSL #2\n" "prfm pstl1keep, [x22, #0x0]\n" "bge 101f\n" - "tbz x14, #3, 96f\n" - "st1 { v8.4s }, [x12], #0x10\n" - "st1 { v9.4s }, [x12], #0x10\n" + "tbz x17, #3, 96f\n" + "st1 { v8.4s }, [x15], #0x10\n" + "st1 { v9.4s }, [x15], #0x10\n" "st1 { v12.4s }, [x23], #0x10\n" "st1 { v13.4s }, [x23], #0x10\n" "st1 { v16.4s }, [x22], #0x10\n" "st1 { v17.4s }, [x22], #0x10\n" - "tbz x14, #2, 94f\n" - "st1 { v10.4s }, [x12], #0x10\n" + "tbz x17, #2, 94f\n" + "st1 { v10.4s }, [x15], #0x10\n" "st1 { v14.4s }, [x23], #0x10\n" "st1 { v18.4s }, [x22], #0x10\n" - "tbz x14, #1, 93f\n" - "str d11, [x12], #0x8\n" + "tbz x17, #1, 93f\n" + "str d11, [x15], #0x8\n" "str d15, [x23], #0x8\n" "str d19, [x22], #0x8\n" - "tbz x14, #0, 100f\n" - "st1 { v11.s }[2], [x12]\n" + "tbz x17, #0, 100f\n" + "st1 { v11.s }[2], [x15]\n" "st1 { v15.s }[2], [x23]\n" "st1 { v19.s }[2], [x22]\n" "b 100f\n" "93:" // Height 3: Partial direct writeback: partial_1_12 - "tbz x14, #0, 100f\n" - "str s11, [x12, #0x0]\n" + "tbz x17, #0, 100f\n" + "str s11, [x15, #0x0]\n" "str s15, [x23, #0x0]\n" "str s19, [x22, #0x0]\n" "b 100f\n" "94:" // Height 3: Partial direct writeback: partial_2_8 - "tbz x14, #1, 95f\n" - "str d10, [x12], #0x8\n" + "tbz x17, #1, 95f\n" + "str d10, [x15], #0x8\n" "str d14, [x23], #0x8\n" "str d18, [x22], #0x8\n" - "tbz x14, #0, 100f\n" - "st1 { v10.s }[2], [x12]\n" + "tbz x17, #0, 100f\n" + "st1 { v10.s }[2], [x15]\n" "st1 { v14.s }[2], [x23]\n" "st1 { v18.s }[2], [x22]\n" "b 100f\n" "95:" // Height 3: Partial direct writeback: partial_1_8 - "tbz x14, #0, 100f\n" - "str s10, [x12, #0x0]\n" + "tbz x17, #0, 100f\n" + "str s10, [x15, #0x0]\n" "str s14, [x23, #0x0]\n" "str s18, [x22, #0x0]\n" "b 100f\n" "96:" // Height 3: Partial direct writeback: partial_4_0 - "tbz x14, #2, 98f\n" - "st1 { v8.4s }, [x12], #0x10\n" + "tbz x17, #2, 98f\n" + "st1 { v8.4s }, [x15], #0x10\n" "st1 { v12.4s }, [x23], #0x10\n" "st1 { v16.4s }, [x22], #0x10\n" - "tbz x14, #1, 97f\n" - "str d9, [x12], #0x8\n" + "tbz x17, #1, 97f\n" + "str d9, [x15], #0x8\n" "str d13, [x23], #0x8\n" "str d17, [x22], #0x8\n" - "tbz x14, #0, 100f\n" - "st1 { v9.s }[2], [x12]\n" + "tbz x17, #0, 100f\n" + "st1 { v9.s }[2], [x15]\n" "st1 { v13.s }[2], [x23]\n" "st1 { v17.s }[2], [x22]\n" "b 100f\n" "97:" // Height 3: Partial direct writeback: partial_1_4 - "tbz x14, #0, 100f\n" - "str s9, [x12, #0x0]\n" + "tbz x17, #0, 100f\n" + "str s9, [x15, #0x0]\n" "str s13, [x23, #0x0]\n" "str s17, [x22, #0x0]\n" "b 100f\n" "98:" // Height 3: Partial direct writeback: partial_2_0 - "tbz x14, #1, 99f\n" - "str d8, [x12], #0x8\n" + "tbz x17, #1, 99f\n" + "str d8, [x15], #0x8\n" "str d12, [x23], #0x8\n" "str d16, [x22], #0x8\n" - "tbz x14, #0, 100f\n" - "st1 { v8.s }[2], [x12]\n" + "tbz x17, #0, 100f\n" + "st1 { v8.s }[2], [x15]\n" "st1 { v12.s }[2], [x23]\n" "st1 { v16.s }[2], [x22]\n" "b 100f\n" "99:" // Height 3: Partial direct writeback: partial_1_0 - "str s8, [x12, #0x0]\n" + "str s8, [x15, #0x0]\n" "str s12, [x23, #0x0]\n" "str s16, [x22, #0x0]\n" "100:" // Height 3: Partial direct writeback: Done "b 102f\n" "101:" // Height 3: Full writeback - "str q8, [x12, #0x0]\n" - "str q9, [x12, #0x10]\n" - "str q10, [x12, #0x20]\n" - "str q11, [x12, #0x30]\n" - "add x12, x12, #0x40\n" + "str q8, [x15, #0x0]\n" + "str q9, [x15, #0x10]\n" + "str q10, [x15, #0x20]\n" + "str q11, [x15, #0x30]\n" + "add x15, x15, #0x40\n" "str q12, [x23, #0x0]\n" "str q13, [x23, #0x10]\n" "str q14, [x23, #0x20]\n" @@ -1317,129 +1317,129 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( "str q18, [x22, #0x20]\n" "str q19, [x22, #0x30]\n" "102:" // Height 3: Writeback done - "subs x14, x14, #0x10\n" + "subs x17, x17, #0x10\n" "bgt 70b\n" "b 206f\n" "103:" // Height 4 - "ldr x14, [%x[args_ptr], %[offsetof_N]]\n" - "mov x12, %x[output_ptr]\n" - "ldr x13, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "ldr x17, [%x[args_ptr], %[offsetof_N]]\n" + "mov x15, %x[output_ptr]\n" + "ldr x16, [%x[args_ptr], %[offsetof_B_ptr]]\n" "104:" // Height 4: Column loop "tbz %x[flags], #0, 114f\n" "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "cmp x14, #0x10\n" - "add x23, x12, x19, LSL #2\n" + "cmp x17, #0x10\n" + "add x23, x15, x19, LSL #2\n" "add x22, x23, x19, LSL #2\n" "add x21, x22, x19, LSL #2\n" "bge 113f\n" - "tbz x14, #3, 108f\n" - "ld1 { v8.4s }, [x12], #0x10\n" + "tbz x17, #3, 108f\n" + "ld1 { v8.4s }, [x15], #0x10\n" "ld1 { v12.4s }, [x23], #0x10\n" "ld1 { v16.4s }, [x22], #0x10\n" - "ld1 { v9.4s }, [x12], #0x10\n" + "ld1 { v9.4s }, [x15], #0x10\n" "ld1 { v13.4s }, [x23], #0x10\n" "ld1 { v17.4s }, [x22], #0x10\n" "ld1 { v20.4s }, [x21], #0x10\n" "ld1 { v21.4s }, [x21], #0x10\n" - "tbz x14, #2, 106f\n" - "ld1 { v10.4s }, [x12], #0x10\n" + "tbz x17, #2, 106f\n" + "ld1 { v10.4s }, [x15], #0x10\n" "ld1 { v14.4s }, [x23], #0x10\n" "ld1 { v18.4s }, [x22], #0x10\n" "ld1 { v22.4s }, [x21], #0x10\n" - "tbz x14, #1, 105f\n" - "ldr d11, [x12], #0x8\n" + "tbz x17, #1, 105f\n" + "ldr d11, [x15], #0x8\n" "mov x24, #0x38\n" "ldr d15, [x23], #0x8\n" "ldr d19, [x22], #0x8\n" "ldr d23, [x21], #0x8\n" - "tbz x14, #0, 112f\n" - "ld1 { v11.s }[2], [x12]\n" + "tbz x17, #0, 112f\n" + "ld1 { v11.s }[2], [x15]\n" "ld1 { v15.s }[2], [x23]\n" "ld1 { v19.s }[2], [x22]\n" "ld1 { v23.s }[2], [x21]\n" "b 112f\n" "105:" // Height 4: Partial accumulate: partial_1_12 "mov x24, #0x30\n" - "tbz x14, #0, 112f\n" - "ldr s11, [x12, #0x0]\n" + "tbz x17, #0, 112f\n" + "ldr s11, [x15, #0x0]\n" "ldr s15, [x23, #0x0]\n" "ldr s19, [x22, #0x0]\n" "ldr s23, [x21, #0x0]\n" "b 112f\n" "106:" // Height 4: Partial accumulate: partial_2_8 - "tbz x14, #1, 107f\n" - "ldr d10, [x12], #0x8\n" + "tbz x17, #1, 107f\n" + "ldr d10, [x15], #0x8\n" "ldr d14, [x23], #0x8\n" "mov x24, #0x28\n" "ldr d18, [x22], #0x8\n" "ldr d22, [x21], #0x8\n" - "tbz x14, #0, 112f\n" - "ld1 { v10.s }[2], [x12]\n" + "tbz x17, #0, 112f\n" + "ld1 { v10.s }[2], [x15]\n" "ld1 { v14.s }[2], [x23]\n" "ld1 { v18.s }[2], [x22]\n" "ld1 { v22.s }[2], [x21]\n" "b 112f\n" "107:" // Height 4: Partial accumulate: partial_1_8 "mov x24, #0x20\n" - "tbz x14, #0, 112f\n" - "ldr s10, [x12, #0x0]\n" + "tbz x17, #0, 112f\n" + "ldr s10, [x15, #0x0]\n" "ldr s14, [x23, #0x0]\n" "ldr s18, [x22, #0x0]\n" "ldr s22, [x21, #0x0]\n" "b 112f\n" "108:" // Height 4: Partial accumulate: partial_4_0 - "tbz x14, #2, 110f\n" - "ld1 { v8.4s }, [x12], #0x10\n" + "tbz x17, #2, 110f\n" + "ld1 { v8.4s }, [x15], #0x10\n" "ld1 { v12.4s }, [x23], #0x10\n" "ld1 { v16.4s }, [x22], #0x10\n" "ld1 { v20.4s }, [x21], #0x10\n" - "tbz x14, #1, 109f\n" - "ldr d9, [x12], #0x8\n" + "tbz x17, #1, 109f\n" + "ldr d9, [x15], #0x8\n" "mov x24, #0x18\n" "ldr d13, [x23], #0x8\n" "ldr d17, [x22], #0x8\n" "ldr d21, [x21], #0x8\n" - "tbz x14, #0, 112f\n" - "ld1 { v9.s }[2], [x12]\n" + "tbz x17, #0, 112f\n" + "ld1 { v9.s }[2], [x15]\n" "ld1 { v13.s }[2], [x23]\n" "ld1 { v17.s }[2], [x22]\n" "ld1 { v21.s }[2], [x21]\n" "b 112f\n" "109:" // Height 4: Partial accumulate: partial_1_4 "mov x24, #0x10\n" - "tbz x14, #0, 112f\n" - "ldr s9, [x12, #0x0]\n" + "tbz x17, #0, 112f\n" + "ldr s9, [x15, #0x0]\n" "ldr s13, [x23, #0x0]\n" "ldr s17, [x22, #0x0]\n" "ldr s21, [x21, #0x0]\n" "b 112f\n" "110:" // Height 4: Partial accumulate: partial_2_0 - "tbz x14, #1, 111f\n" - "ldr d8, [x12], #0x8\n" + "tbz x17, #1, 111f\n" + "ldr d8, [x15], #0x8\n" "ldr d12, [x23], #0x8\n" "mov x24, #0x8\n" "ldr d16, [x22], #0x8\n" "ldr d20, [x21], #0x8\n" - "tbz x14, #0, 112f\n" - "ld1 { v8.s }[2], [x12]\n" + "tbz x17, #0, 112f\n" + "ld1 { v8.s }[2], [x15]\n" "ld1 { v12.s }[2], [x23]\n" "ld1 { v16.s }[2], [x22]\n" "ld1 { v20.s }[2], [x21]\n" "b 112f\n" "111:" // Height 4: Partial accumulate: partial_1_0 - "ldr s8, [x12, #0x0]\n" + "ldr s8, [x15, #0x0]\n" "mov x24, #0x0\n" "ldr s12, [x23, #0x0]\n" "ldr s16, [x22, #0x0]\n" "ldr s20, [x21, #0x0]\n" "112:" // Height 4: Partial accumulate: Done - "sub x12, x12, x24\n" + "sub x15, x15, x24\n" "b 115f\n" "113:" // Height 4: full accumulate - "ldr q8, [x12, #0x0]\n" - "ldr q9, [x12, #0x10]\n" - "ldr q10, [x12, #0x20]\n" - "ldr q11, [x12, #0x30]\n" + "ldr q8, [x15, #0x0]\n" + "ldr q9, [x15, #0x10]\n" + "ldr q10, [x15, #0x20]\n" + "ldr q11, [x15, #0x30]\n" "ldr q12, [x23, #0x0]\n" "ldr q13, [x23, #0x10]\n" "ldr q14, [x23, #0x20]\n" @@ -1471,292 +1471,292 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( "movi v22.4s, #0x0\n" "movi v23.4s, #0x0\n" "115:" // Height 4: setup done - "mov x11, #0x0\n" + "mov x14, #0x0\n" "116:" // Height 4: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w10, [x20, x11, LSL #0x2]\n" + "ldr w13, [x20, x14, LSL #0x2]\n" "tbz %x[flags], #3, 117f\n" - "ldr x20, [%x[input_ptr], x11, LSL #0x3]\n" + "ldr x20, [%x[input_ptr], x14, LSL #0x3]\n" "add x20, x20, x19, LSL #3\n" - "ldr x9, [x20, #0x0]\n" - "ldr x26, [x20, #0x8]\n" - "ldr x24, [x20, #0x10]\n" - "ldr x23, [x20, #0x18]\n" - "cbnz x11, 118f\n" + "ldr x12, [x20, #0x0]\n" + "ldr x28, [x20, #0x8]\n" + "ldr x26, [x20, #0x10]\n" + "ldr x24, [x20, #0x18]\n" + "cbnz x14, 118f\n" "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x9, x9, x19\n" + "add x12, x12, x19\n" + "add x28, x28, x19\n" "add x26, x26, x19\n" "add x24, x24, x19\n" - "add x23, x23, x19\n" "b 118f\n" "117:" // Height 4: setup direct input - "mov x9, %x[input_ptr]\n" - "add x26, x9, x19\n" + "mov x12, %x[input_ptr]\n" + "add x28, x12, x19\n" + "add x26, x28, x19\n" "add x24, x26, x19\n" - "add x23, x24, x19\n" "118:" // Height 4: input setup done - "cmp x10, #0x10\n" + "cmp x13, #0x10\n" "blt 121f\n" - "ldr q0, [x9, #0x0]\n" - "ldr q1, [x26, #0x0]\n" - "cmp x10, #0x20\n" - "ldr q2, [x24, #0x0]\n" - "ldr q3, [x23, #0x0]\n" - "ldr q6, [x13, #0x0]\n" + "ldr q0, [x12, #0x0]\n" + "ldr q1, [x28, #0x0]\n" + "cmp x13, #0x20\n" + "ldr q2, [x26, #0x0]\n" + "ldr q3, [x24, #0x0]\n" + "ldr q6, [x16, #0x0]\n" "blt 120f\n" "119:" // Height 4: Multiply loop: Main loop head ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" + "ldr d7, [x16, #0x10]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" + "ldr x11, [x16, #0x18]\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" + "ldr x10, [x16, #0x28]\n" ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n" - "ldr d7, [x13, #0x10]\n" - "add x9, x9, #0x10\n" - "ldr x19, [x13, #0x18]\n" - "add x26, x26, #0x10\n" - "prfm pldl1keep, [x9, #0x80]\n" - "add x24, x24, #0x10\n" - "ldr d6, [x13, #0x20]\n" - "add x23, x23, #0x10\n" - "mov v7.d[1], x19\n" - "ldr x28, [x13, #0x28]\n" - "ldr x19, [x13, #0x38]\n" - "sub x10, x10, #0x10\n" + "ldr d6, [x16, #0x20]\n" + "mov v7.d[1], x11\n" + "ldr x11, [x16, #0x38]\n" + "add x12, x12, #0x10\n" + "add x28, x28, #0x10\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" + "mov v6.d[1], x10\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" - "mov v6.d[1], x28\n" + "prfm pldl1keep, [x12, #0x80]\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" + "prfm pldl1keep, [x28, #0x80]\n" ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n" - "ldr d7, [x13, #0x30]\n" + "ldr d7, [x16, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "ldr x10, [x16, #0x48]\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" + "ldr x9, [x12, #0x8]\n" ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" + "mov v7.d[1], x11\n" ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n" - "mov v7.d[1], x19\n" - "prfm pldl1keep, [x26, #0x80]\n" - "ldr d6, [x13, #0x40]\n" - "cmp x10, #0x20\n" + "ldr d6, [x16, #0x40]\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" + "ldr x11, [x16, #0x58]\n" ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" + "ldr x27, [x28, #0x8]\n" ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" + "mov v6.d[1], x10\n" ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n" - "ldr x28, [x13, #0x48]\n" - "ldr d7, [x13, #0x50]\n" - "ldr x19, [x13, #0x58]\n" - "mov v6.d[1], x28\n" - "prfm pldl1keep, [x24, #0x80]\n" - "ldr x28, [x13, #0x68]\n" + "ldr d7, [x16, #0x50]\n" ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n" + "ldr x10, [x16, #0x68]\n" ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n" + "add x26, x26, #0x10\n" ".inst 0x4fa2e0d0 // sdot v16.4s, v6.16b, v2.4b[1]\n" + "mov v7.d[1], x11\n" ".inst 0x4fa3e0d4 // sdot v20.4s, v6.16b, v3.4b[1]\n" - "mov v7.d[1], x19\n" - "ldr d6, [x13, #0x60]\n" - "ldr x19, [x13, #0x78]\n" + "prfm pldl1keep, [x26, #0x80]\n" ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n" + "ldr d6, [x16, #0x60]\n" ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n" - "mov v6.d[1], x28\n" + "ldr x11, [x16, #0x78]\n" ".inst 0x4fa2e0f1 // sdot v17.4s, v7.16b, v2.4b[1]\n" + "ldr x25, [x26, #0x8]\n" ".inst 0x4fa3e0f5 // sdot v21.4s, v7.16b, v3.4b[1]\n" - "ldr d7, [x13, #0x70]\n" + "mov v6.d[1], x10\n" + "ldr d7, [x16, #0x70]\n" + "add x24, x24, #0x10\n" ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n" + "prfm pldl1keep, [x24, #0x80]\n" ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n" + "ldr x10, [x16, #0x88]\n" ".inst 0x4fa2e0d2 // sdot v18.4s, v6.16b, v2.4b[1]\n" + "mov v7.d[1], x11\n" ".inst 0x4fa3e0d6 // sdot v22.4s, v6.16b, v3.4b[1]\n" - "mov v7.d[1], x19\n" - "prfm pldl1keep, [x23, #0x80]\n" - "ldr d6, [x13, #0x80]\n" + "ldr d6, [x16, #0x80]\n" ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n" + "ldr x11, [x16, #0x98]\n" ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n" + "ldr x23, [x24, #0x8]\n" ".inst 0x4fa2e0f3 // sdot v19.4s, v7.16b, v2.4b[1]\n" + "mov v6.d[1], x10\n" ".inst 0x4fa3e0f7 // sdot v23.4s, v7.16b, v3.4b[1]\n" - "ldr x28, [x13, #0x88]\n" - "ldr d7, [x13, #0x90]\n" - "ldr x19, [x13, #0x98]\n" - "mov v6.d[1], x28\n" - "ldr x28, [x13, #0xa8]\n" - "ldr x27, [x9, #0x8]\n" + "ldr d7, [x16, #0x90]\n" ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n" + "ldr x10, [x16, #0xa8]\n" ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n" + "sub x13, x13, #0x10\n" ".inst 0x4f82e8d0 // sdot v16.4s, v6.16b, v2.4b[2]\n" + "mov v7.d[1], x11\n" ".inst 0x4f83e8d4 // sdot v20.4s, v6.16b, v3.4b[2]\n" - "mov v7.d[1], x19\n" - "ldr d6, [x13, #0xa0]\n" - "ldr x19, [x13, #0xb8]\n" + "ldr d6, [x16, #0xa0]\n" ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n" + "ldr x11, [x16, #0xb8]\n" ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n" - "mov v6.d[1], x28\n" + "cmp x13, #0x20\n" ".inst 0x4f82e8f1 // sdot v17.4s, v7.16b, v2.4b[2]\n" + "mov v6.d[1], x10\n" ".inst 0x4f83e8f5 // sdot v21.4s, v7.16b, v3.4b[2]\n" - "ldr d7, [x13, #0xb0]\n" + "ldr d7, [x16, #0xb0]\n" ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n" + "ldr x10, [x16, #0xc8]\n" ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n" ".inst 0x4f82e8d2 // sdot v18.4s, v6.16b, v2.4b[2]\n" + "mov v7.d[1], x11\n" ".inst 0x4f83e8d6 // sdot v22.4s, v6.16b, v3.4b[2]\n" - "mov v7.d[1], x19\n" - "ldr d6, [x13, #0xc0]\n" - "ldr x28, [x13, #0xc8]\n" + "ldr d6, [x16, #0xc0]\n" ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n" + "ldr x11, [x16, #0xd8]\n" ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n" ".inst 0x4f82e8f3 // sdot v19.4s, v7.16b, v2.4b[2]\n" + "mov v6.d[1], x10\n" ".inst 0x4f83e8f7 // sdot v23.4s, v7.16b, v3.4b[2]\n" - "mov v6.d[1], x28\n" - "ldr d7, [x13, #0xd0]\n" - "ldr x19, [x13, #0xd8]\n" + "ldr d7, [x16, #0xd0]\n" ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n" + "ldr x10, [x16, #0xe8]\n" ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n" ".inst 0x4fa2e8d0 // sdot v16.4s, v6.16b, v2.4b[3]\n" + "mov v7.d[1], x11\n" ".inst 0x4fa3e8d4 // sdot v20.4s, v6.16b, v3.4b[3]\n" - "mov v7.d[1], x19\n" - "ldr d6, [x13, #0xe0]\n" - "ldr x28, [x13, #0xe8]\n" + "ldr d6, [x16, #0xe0]\n" ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n" + "ldr x11, [x16, #0xf8]\n" ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n" ".inst 0x4fa2e8f1 // sdot v17.4s, v7.16b, v2.4b[3]\n" + "mov v6.d[1], x10\n" ".inst 0x4fa3e8f5 // sdot v21.4s, v7.16b, v3.4b[3]\n" - "mov v6.d[1], x28\n" - "ldr d7, [x13, #0xf0]\n" - "ldr x19, [x13, #0xf8]\n" - "add x13, x13, #0x100\n" + "ldr d7, [x16, #0xf0]\n" ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n" + "add x16, x16, #0x100\n" ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n" + "ldr x10, [x16, #0x8]\n" ".inst 0x4fa2e8d2 // sdot v18.4s, v6.16b, v2.4b[3]\n" + "mov v7.d[1], x11\n" ".inst 0x4fa3e8d6 // sdot v22.4s, v6.16b, v3.4b[3]\n" - "mov v7.d[1], x19\n" - "ldr d6, [x13, #0x0]\n" - "ldr x28, [x13, #0x8]\n" + "ldr d6, [x16, #0x0]\n" ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n" + "ldr d0, [x12, #0x0]\n" ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n" + "ldr d1, [x28, #0x0]\n" ".inst 0x4fa2e8f3 // sdot v19.4s, v7.16b, v2.4b[3]\n" + "mov v6.d[1], x10\n" ".inst 0x4fa3e8f7 // sdot v23.4s, v7.16b, v3.4b[3]\n" - "mov v6.d[1], x28\n" - "ldr d0, [x9, #0x0]\n" - "ldr d1, [x26, #0x0]\n" - "ldr x25, [x26, #0x8]\n" - "mov v0.d[1], x27\n" - "ldr d2, [x24, #0x0]\n" - "ldr x21, [x24, #0x8]\n" - "mov v1.d[1], x25\n" - "ldr d3, [x23, #0x0]\n" - "ldr x19, [x23, #0x8]\n" - "mov v2.d[1], x21\n" - "mov v3.d[1], x19\n" + "mov v0.d[1], x9\n" + "mov v1.d[1], x27\n" + "ldr d2, [x26, #0x0]\n" + "ldr d3, [x24, #0x0]\n" + "mov v2.d[1], x25\n" + "mov v3.d[1], x23\n" "bge 119b\n" "120:" // Height 4: Multiply loop: Single iteration only ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" + "ldr q7, [x16, #0x10]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" + "sub x13, x13, #0x10\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" + "add x12, x12, #0x10\n" ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n" - "ldr q7, [x13, #0x10]\n" - "sub x10, x10, #0x10\n" - "ldr q6, [x13, #0x20]\n" - "add x9, x9, #0x10\n" - "prfm pldl1keep, [x9, #0x80]\n" - "add x26, x26, #0x10\n" + "prfm pldl1keep, [x12, #0x80]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" + "ldr q6, [x16, #0x20]\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" + "add x28, x28, #0x10\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" + "prfm pldl1keep, [x28, #0x80]\n" ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n" + "ldr q7, [x16, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "add x26, x26, #0x10\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" + "prfm pldl1keep, [x26, #0x80]\n" ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" - ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n" - "ldr q7, [x13, #0x30]\n" "add x24, x24, #0x10\n" - "prfm pldl1keep, [x26, #0x80]\n" - "add x23, x23, #0x10\n" - "ldr q6, [x13, #0x40]\n" + ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n" + "prfm pldl1keep, [x24, #0x80]\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" + "ldr q6, [x16, #0x40]\n" ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n" + "ldr q7, [x16, #0x50]\n" ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n" ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n" ".inst 0x4fa2e0d0 // sdot v16.4s, v6.16b, v2.4b[1]\n" ".inst 0x4fa3e0d4 // sdot v20.4s, v6.16b, v3.4b[1]\n" - "ldr q7, [x13, #0x50]\n" - "prfm pldl1keep, [x24, #0x80]\n" - "ldr q6, [x13, #0x60]\n" + "ldr q6, [x16, #0x60]\n" ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n" ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n" ".inst 0x4fa2e0f1 // sdot v17.4s, v7.16b, v2.4b[1]\n" ".inst 0x4fa3e0f5 // sdot v21.4s, v7.16b, v3.4b[1]\n" + "ldr q7, [x16, #0x70]\n" ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n" ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n" ".inst 0x4fa2e0d2 // sdot v18.4s, v6.16b, v2.4b[1]\n" ".inst 0x4fa3e0d6 // sdot v22.4s, v6.16b, v3.4b[1]\n" - "ldr q7, [x13, #0x70]\n" - "prfm pldl1keep, [x23, #0x80]\n" - "ldr q6, [x13, #0x80]\n" + "ldr q6, [x16, #0x80]\n" ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n" ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n" ".inst 0x4fa2e0f3 // sdot v19.4s, v7.16b, v2.4b[1]\n" ".inst 0x4fa3e0f7 // sdot v23.4s, v7.16b, v3.4b[1]\n" + "ldr q7, [x16, #0x90]\n" ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n" ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n" ".inst 0x4f82e8d0 // sdot v16.4s, v6.16b, v2.4b[2]\n" ".inst 0x4f83e8d4 // sdot v20.4s, v6.16b, v3.4b[2]\n" - "ldr q7, [x13, #0x90]\n" - "ldr q6, [x13, #0xa0]\n" + "ldr q6, [x16, #0xa0]\n" ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n" ".inst 0x4f82e8f1 // sdot v17.4s, v7.16b, v2.4b[2]\n" ".inst 0x4f83e8f5 // sdot v21.4s, v7.16b, v3.4b[2]\n" + "ldr q7, [x16, #0xb0]\n" ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n" ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n" ".inst 0x4f82e8d2 // sdot v18.4s, v6.16b, v2.4b[2]\n" ".inst 0x4f83e8d6 // sdot v22.4s, v6.16b, v3.4b[2]\n" - "ldr q7, [x13, #0xb0]\n" - "ldr q6, [x13, #0xc0]\n" + "ldr q6, [x16, #0xc0]\n" ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n" ".inst 0x4f82e8f3 // sdot v19.4s, v7.16b, v2.4b[2]\n" ".inst 0x4f83e8f7 // sdot v23.4s, v7.16b, v3.4b[2]\n" + "ldr q7, [x16, #0xd0]\n" ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n" ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n" ".inst 0x4fa2e8d0 // sdot v16.4s, v6.16b, v2.4b[3]\n" ".inst 0x4fa3e8d4 // sdot v20.4s, v6.16b, v3.4b[3]\n" - "ldr q7, [x13, #0xd0]\n" - "ldr q6, [x13, #0xe0]\n" + "ldr q6, [x16, #0xe0]\n" ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n" ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n" ".inst 0x4fa2e8f1 // sdot v17.4s, v7.16b, v2.4b[3]\n" ".inst 0x4fa3e8f5 // sdot v21.4s, v7.16b, v3.4b[3]\n" + "ldr q7, [x16, #0xf0]\n" ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n" + "add x16, x16, #0x100\n" ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n" ".inst 0x4fa2e8d2 // sdot v18.4s, v6.16b, v2.4b[3]\n" ".inst 0x4fa3e8d6 // sdot v22.4s, v6.16b, v3.4b[3]\n" - "ldr q7, [x13, #0xf0]\n" - "add x13, x13, #0x100\n" ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n" ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n" ".inst 0x4fa2e8f3 // sdot v19.4s, v7.16b, v2.4b[3]\n" ".inst 0x4fa3e8f7 // sdot v23.4s, v7.16b, v3.4b[3]\n" "121:" // Height 4: Multiply loop: Main loop skip - "cbz x10, 126f\n" - "cmp x10, #0x4\n" + "cbz x13, 126f\n" + "cmp x13, #0x4\n" "blt 123f\n" "122:" // Height 4: Multiply loop: Odd block loop - "ldr s0, [x9], #0x4\n" - "sub x10, x10, #0x4\n" - "ldr s1, [x26], #0x4\n" - "cmp x10, #0x4\n" - "ldr s2, [x24], #0x4\n" - "ldr s3, [x23], #0x4\n" - "ldr q6, [x13, #0x0]\n" - "ldr q7, [x13, #0x10]\n" + "ldr s0, [x12], #0x4\n" + "sub x13, x13, #0x4\n" + "ldr s1, [x28], #0x4\n" + "cmp x13, #0x4\n" + "ldr s2, [x26], #0x4\n" + "ldr s3, [x24], #0x4\n" + "ldr q6, [x16, #0x0]\n" + "ldr q7, [x16, #0x10]\n" ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n" + "ldr q6, [x16, #0x20]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n" - "ldr q6, [x13, #0x20]\n" - "ldr q7, [x13, #0x30]\n" - "add x13, x13, #0x40\n" + "ldr q7, [x16, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "add x16, x16, #0x40\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n" @@ -1765,39 +1765,39 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n" "bge 122b\n" - "cbz x10, 126f\n" + "cbz x13, 126f\n" "123:" // Height 4: Multiply loop: Skip odd blocks - "tbz x10, #1, 124f\n" - "ldr h0, [x9], #0x2\n" - "ldr h1, [x26], #0x2\n" - "ldr h2, [x24], #0x2\n" - "ldr h3, [x23], #0x2\n" - "tbz x10, #0, 125f\n" - "ld1 { v0.b }[2], [x9]\n" - "ld1 { v1.b }[2], [x26]\n" - "ld1 { v2.b }[2], [x24]\n" - "ld1 { v3.b }[2], [x23]\n" + "tbz x13, #1, 124f\n" + "ldr h0, [x12], #0x2\n" + "ldr h1, [x28], #0x2\n" + "ldr h2, [x26], #0x2\n" + "ldr h3, [x24], #0x2\n" + "tbz x13, #0, 125f\n" + "ld1 { v0.b }[2], [x12]\n" + "ld1 { v1.b }[2], [x28]\n" + "ld1 { v2.b }[2], [x26]\n" + "ld1 { v3.b }[2], [x24]\n" "b 125f\n" "124:" // Height 4: Multiply loop: Ragged operand read: partial_1_0 - "ldr b0, [x9, #0x0]\n" - "ldr b1, [x26, #0x0]\n" - "ldr b2, [x24, #0x0]\n" - "ldr b3, [x23, #0x0]\n" + "ldr b0, [x12, #0x0]\n" + "ldr b1, [x28, #0x0]\n" + "ldr b2, [x26, #0x0]\n" + "ldr b3, [x24, #0x0]\n" "125:" // Height 4: Multiply loop: Ragged operand read: Done - "ldr q6, [x13, #0x0]\n" - "ldr q7, [x13, #0x10]\n" + "ldr q6, [x16, #0x0]\n" + "ldr q7, [x16, #0x10]\n" ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n" + "ldr q6, [x16, #0x20]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n" - "ldr q6, [x13, #0x20]\n" - "ldr q7, [x13, #0x30]\n" - "add x13, x13, #0x40\n" + "ldr q7, [x16, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "add x16, x16, #0x40\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n" @@ -1807,119 +1807,119 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n" "126:" // Height 4: Multiply loop: No odd multiplies "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" - "add x11, x11, #0x1\n" - "cmp x11, x19\n" + "add x14, x14, #0x1\n" + "cmp x14, x19\n" "bne 116b\n" "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "prfm pstl1keep, [x12, #0x0]\n" - "cmp x14, #0x10\n" - "add x23, x12, x19, LSL #2\n" + "prfm pstl1keep, [x15, #0x0]\n" + "cmp x17, #0x10\n" + "add x23, x15, x19, LSL #2\n" "prfm pstl1keep, [x23, #0x0]\n" "add x22, x23, x19, LSL #2\n" "prfm pstl1keep, [x22, #0x0]\n" "add x21, x22, x19, LSL #2\n" "prfm pstl1keep, [x21, #0x0]\n" "bge 135f\n" - "tbz x14, #3, 130f\n" - "st1 { v8.4s }, [x12], #0x10\n" - "st1 { v9.4s }, [x12], #0x10\n" + "tbz x17, #3, 130f\n" + "st1 { v8.4s }, [x15], #0x10\n" + "st1 { v9.4s }, [x15], #0x10\n" "st1 { v12.4s }, [x23], #0x10\n" "st1 { v13.4s }, [x23], #0x10\n" "st1 { v16.4s }, [x22], #0x10\n" "st1 { v17.4s }, [x22], #0x10\n" "st1 { v20.4s }, [x21], #0x10\n" "st1 { v21.4s }, [x21], #0x10\n" - "tbz x14, #2, 128f\n" - "st1 { v10.4s }, [x12], #0x10\n" + "tbz x17, #2, 128f\n" + "st1 { v10.4s }, [x15], #0x10\n" "st1 { v14.4s }, [x23], #0x10\n" "st1 { v18.4s }, [x22], #0x10\n" "st1 { v22.4s }, [x21], #0x10\n" - "tbz x14, #1, 127f\n" - "str d11, [x12], #0x8\n" + "tbz x17, #1, 127f\n" + "str d11, [x15], #0x8\n" "str d15, [x23], #0x8\n" "str d19, [x22], #0x8\n" "str d23, [x21], #0x8\n" - "tbz x14, #0, 134f\n" - "st1 { v11.s }[2], [x12]\n" + "tbz x17, #0, 134f\n" + "st1 { v11.s }[2], [x15]\n" "st1 { v15.s }[2], [x23]\n" "st1 { v19.s }[2], [x22]\n" "st1 { v23.s }[2], [x21]\n" "b 134f\n" "127:" // Height 4: Partial direct writeback: partial_1_12 - "tbz x14, #0, 134f\n" - "str s11, [x12, #0x0]\n" + "tbz x17, #0, 134f\n" + "str s11, [x15, #0x0]\n" "str s15, [x23, #0x0]\n" "str s19, [x22, #0x0]\n" "str s23, [x21, #0x0]\n" "b 134f\n" "128:" // Height 4: Partial direct writeback: partial_2_8 - "tbz x14, #1, 129f\n" - "str d10, [x12], #0x8\n" + "tbz x17, #1, 129f\n" + "str d10, [x15], #0x8\n" "str d14, [x23], #0x8\n" "str d18, [x22], #0x8\n" "str d22, [x21], #0x8\n" - "tbz x14, #0, 134f\n" - "st1 { v10.s }[2], [x12]\n" + "tbz x17, #0, 134f\n" + "st1 { v10.s }[2], [x15]\n" "st1 { v14.s }[2], [x23]\n" "st1 { v18.s }[2], [x22]\n" "st1 { v22.s }[2], [x21]\n" "b 134f\n" "129:" // Height 4: Partial direct writeback: partial_1_8 - "tbz x14, #0, 134f\n" - "str s10, [x12, #0x0]\n" + "tbz x17, #0, 134f\n" + "str s10, [x15, #0x0]\n" "str s14, [x23, #0x0]\n" "str s18, [x22, #0x0]\n" "str s22, [x21, #0x0]\n" "b 134f\n" "130:" // Height 4: Partial direct writeback: partial_4_0 - "tbz x14, #2, 132f\n" - "st1 { v8.4s }, [x12], #0x10\n" + "tbz x17, #2, 132f\n" + "st1 { v8.4s }, [x15], #0x10\n" "st1 { v12.4s }, [x23], #0x10\n" "st1 { v16.4s }, [x22], #0x10\n" "st1 { v20.4s }, [x21], #0x10\n" - "tbz x14, #1, 131f\n" - "str d9, [x12], #0x8\n" + "tbz x17, #1, 131f\n" + "str d9, [x15], #0x8\n" "str d13, [x23], #0x8\n" "str d17, [x22], #0x8\n" "str d21, [x21], #0x8\n" - "tbz x14, #0, 134f\n" - "st1 { v9.s }[2], [x12]\n" + "tbz x17, #0, 134f\n" + "st1 { v9.s }[2], [x15]\n" "st1 { v13.s }[2], [x23]\n" "st1 { v17.s }[2], [x22]\n" "st1 { v21.s }[2], [x21]\n" "b 134f\n" "131:" // Height 4: Partial direct writeback: partial_1_4 - "tbz x14, #0, 134f\n" - "str s9, [x12, #0x0]\n" + "tbz x17, #0, 134f\n" + "str s9, [x15, #0x0]\n" "str s13, [x23, #0x0]\n" "str s17, [x22, #0x0]\n" "str s21, [x21, #0x0]\n" "b 134f\n" "132:" // Height 4: Partial direct writeback: partial_2_0 - "tbz x14, #1, 133f\n" - "str d8, [x12], #0x8\n" + "tbz x17, #1, 133f\n" + "str d8, [x15], #0x8\n" "str d12, [x23], #0x8\n" "str d16, [x22], #0x8\n" "str d20, [x21], #0x8\n" - "tbz x14, #0, 134f\n" - "st1 { v8.s }[2], [x12]\n" + "tbz x17, #0, 134f\n" + "st1 { v8.s }[2], [x15]\n" "st1 { v12.s }[2], [x23]\n" "st1 { v16.s }[2], [x22]\n" "st1 { v20.s }[2], [x21]\n" "b 134f\n" "133:" // Height 4: Partial direct writeback: partial_1_0 - "str s8, [x12, #0x0]\n" + "str s8, [x15, #0x0]\n" "str s12, [x23, #0x0]\n" "str s16, [x22, #0x0]\n" "str s20, [x21, #0x0]\n" "134:" // Height 4: Partial direct writeback: Done "b 136f\n" "135:" // Height 4: Full writeback - "str q8, [x12, #0x0]\n" - "str q9, [x12, #0x10]\n" - "str q10, [x12, #0x20]\n" - "str q11, [x12, #0x30]\n" - "add x12, x12, #0x40\n" + "str q8, [x15, #0x0]\n" + "str q9, [x15, #0x10]\n" + "str q10, [x15, #0x20]\n" + "str q11, [x15, #0x30]\n" + "add x15, x15, #0x40\n" "str q12, [x23, #0x0]\n" "str q13, [x23, #0x10]\n" "str q14, [x23, #0x20]\n" @@ -1933,48 +1933,48 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( "str q22, [x21, #0x20]\n" "str q23, [x21, #0x30]\n" "136:" // Height 4: Writeback done - "subs x14, x14, #0x10\n" + "subs x17, x17, #0x10\n" "bgt 104b\n" "b 206f\n" "137:" // Height 5 - "ldr x14, [%x[args_ptr], %[offsetof_N]]\n" - "mov x12, %x[output_ptr]\n" - "ldr x13, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "ldr x17, [%x[args_ptr], %[offsetof_N]]\n" + "mov x15, %x[output_ptr]\n" + "ldr x16, [%x[args_ptr], %[offsetof_B_ptr]]\n" "138:" // Height 5: Column loop "tbz %x[flags], #0, 148f\n" "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "cmp x14, #0x10\n" - "add x23, x12, x19, LSL #2\n" + "cmp x17, #0x10\n" + "add x23, x15, x19, LSL #2\n" "add x22, x23, x19, LSL #2\n" "add x21, x22, x19, LSL #2\n" "add x20, x21, x19, LSL #2\n" "bge 147f\n" - "tbz x14, #3, 142f\n" - "ld1 { v8.4s }, [x12], #0x10\n" + "tbz x17, #3, 142f\n" + "ld1 { v8.4s }, [x15], #0x10\n" "ld1 { v12.4s }, [x23], #0x10\n" "ld1 { v16.4s }, [x22], #0x10\n" - "ld1 { v9.4s }, [x12], #0x10\n" + "ld1 { v9.4s }, [x15], #0x10\n" "ld1 { v13.4s }, [x23], #0x10\n" "ld1 { v17.4s }, [x22], #0x10\n" "ld1 { v20.4s }, [x21], #0x10\n" "ld1 { v24.4s }, [x20], #0x10\n" "ld1 { v21.4s }, [x21], #0x10\n" "ld1 { v25.4s }, [x20], #0x10\n" - "tbz x14, #2, 140f\n" - "ld1 { v10.4s }, [x12], #0x10\n" + "tbz x17, #2, 140f\n" + "ld1 { v10.4s }, [x15], #0x10\n" "ld1 { v14.4s }, [x23], #0x10\n" "ld1 { v18.4s }, [x22], #0x10\n" "ld1 { v22.4s }, [x21], #0x10\n" "ld1 { v26.4s }, [x20], #0x10\n" - "tbz x14, #1, 139f\n" - "ldr d11, [x12], #0x8\n" + "tbz x17, #1, 139f\n" + "ldr d11, [x15], #0x8\n" "mov x24, #0x38\n" "ldr d15, [x23], #0x8\n" "ldr d19, [x22], #0x8\n" "ldr d23, [x21], #0x8\n" "ldr d27, [x20], #0x8\n" - "tbz x14, #0, 146f\n" - "ld1 { v11.s }[2], [x12]\n" + "tbz x17, #0, 146f\n" + "ld1 { v11.s }[2], [x15]\n" "ld1 { v15.s }[2], [x23]\n" "ld1 { v19.s }[2], [x22]\n" "ld1 { v23.s }[2], [x21]\n" @@ -1982,23 +1982,23 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( "b 146f\n" "139:" // Height 5: Partial accumulate: partial_1_12 "mov x24, #0x30\n" - "tbz x14, #0, 146f\n" - "ldr s11, [x12, #0x0]\n" + "tbz x17, #0, 146f\n" + "ldr s11, [x15, #0x0]\n" "ldr s15, [x23, #0x0]\n" "ldr s19, [x22, #0x0]\n" "ldr s23, [x21, #0x0]\n" "ldr s27, [x20, #0x0]\n" "b 146f\n" "140:" // Height 5: Partial accumulate: partial_2_8 - "tbz x14, #1, 141f\n" - "ldr d10, [x12], #0x8\n" + "tbz x17, #1, 141f\n" + "ldr d10, [x15], #0x8\n" "ldr d14, [x23], #0x8\n" "mov x24, #0x28\n" "ldr d18, [x22], #0x8\n" "ldr d22, [x21], #0x8\n" "ldr d26, [x20], #0x8\n" - "tbz x14, #0, 146f\n" - "ld1 { v10.s }[2], [x12]\n" + "tbz x17, #0, 146f\n" + "ld1 { v10.s }[2], [x15]\n" "ld1 { v14.s }[2], [x23]\n" "ld1 { v18.s }[2], [x22]\n" "ld1 { v22.s }[2], [x21]\n" @@ -2006,29 +2006,29 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( "b 146f\n" "141:" // Height 5: Partial accumulate: partial_1_8 "mov x24, #0x20\n" - "tbz x14, #0, 146f\n" - "ldr s10, [x12, #0x0]\n" + "tbz x17, #0, 146f\n" + "ldr s10, [x15, #0x0]\n" "ldr s14, [x23, #0x0]\n" "ldr s18, [x22, #0x0]\n" "ldr s22, [x21, #0x0]\n" "ldr s26, [x20, #0x0]\n" "b 146f\n" "142:" // Height 5: Partial accumulate: partial_4_0 - "tbz x14, #2, 144f\n" - "ld1 { v8.4s }, [x12], #0x10\n" + "tbz x17, #2, 144f\n" + "ld1 { v8.4s }, [x15], #0x10\n" "ld1 { v12.4s }, [x23], #0x10\n" "ld1 { v16.4s }, [x22], #0x10\n" "ld1 { v20.4s }, [x21], #0x10\n" "ld1 { v24.4s }, [x20], #0x10\n" - "tbz x14, #1, 143f\n" - "ldr d9, [x12], #0x8\n" + "tbz x17, #1, 143f\n" + "ldr d9, [x15], #0x8\n" "mov x24, #0x18\n" "ldr d13, [x23], #0x8\n" "ldr d17, [x22], #0x8\n" "ldr d21, [x21], #0x8\n" "ldr d25, [x20], #0x8\n" - "tbz x14, #0, 146f\n" - "ld1 { v9.s }[2], [x12]\n" + "tbz x17, #0, 146f\n" + "ld1 { v9.s }[2], [x15]\n" "ld1 { v13.s }[2], [x23]\n" "ld1 { v17.s }[2], [x22]\n" "ld1 { v21.s }[2], [x21]\n" @@ -2036,43 +2036,43 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( "b 146f\n" "143:" // Height 5: Partial accumulate: partial_1_4 "mov x24, #0x10\n" - "tbz x14, #0, 146f\n" - "ldr s9, [x12, #0x0]\n" + "tbz x17, #0, 146f\n" + "ldr s9, [x15, #0x0]\n" "ldr s13, [x23, #0x0]\n" "ldr s17, [x22, #0x0]\n" "ldr s21, [x21, #0x0]\n" "ldr s25, [x20, #0x0]\n" "b 146f\n" "144:" // Height 5: Partial accumulate: partial_2_0 - "tbz x14, #1, 145f\n" - "ldr d8, [x12], #0x8\n" + "tbz x17, #1, 145f\n" + "ldr d8, [x15], #0x8\n" "ldr d12, [x23], #0x8\n" "mov x24, #0x8\n" "ldr d16, [x22], #0x8\n" "ldr d20, [x21], #0x8\n" "ldr d24, [x20], #0x8\n" - "tbz x14, #0, 146f\n" - "ld1 { v8.s }[2], [x12]\n" + "tbz x17, #0, 146f\n" + "ld1 { v8.s }[2], [x15]\n" "ld1 { v12.s }[2], [x23]\n" "ld1 { v16.s }[2], [x22]\n" "ld1 { v20.s }[2], [x21]\n" "ld1 { v24.s }[2], [x20]\n" "b 146f\n" "145:" // Height 5: Partial accumulate: partial_1_0 - "ldr s8, [x12, #0x0]\n" + "ldr s8, [x15, #0x0]\n" "mov x24, #0x0\n" "ldr s12, [x23, #0x0]\n" "ldr s16, [x22, #0x0]\n" "ldr s20, [x21, #0x0]\n" "ldr s24, [x20, #0x0]\n" "146:" // Height 5: Partial accumulate: Done - "sub x12, x12, x24\n" + "sub x15, x15, x24\n" "b 149f\n" "147:" // Height 5: full accumulate - "ldr q8, [x12, #0x0]\n" - "ldr q9, [x12, #0x10]\n" - "ldr q10, [x12, #0x20]\n" - "ldr q11, [x12, #0x30]\n" + "ldr q8, [x15, #0x0]\n" + "ldr q9, [x15, #0x10]\n" + "ldr q10, [x15, #0x20]\n" + "ldr q11, [x15, #0x30]\n" "ldr q12, [x23, #0x0]\n" "ldr q13, [x23, #0x10]\n" "ldr q14, [x23, #0x20]\n" @@ -2112,338 +2112,338 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( "movi v26.4s, #0x0\n" "movi v27.4s, #0x0\n" "149:" // Height 5: setup done - "mov x11, #0x0\n" + "mov x14, #0x0\n" "150:" // Height 5: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w10, [x20, x11, LSL #0x2]\n" + "ldr w13, [x20, x14, LSL #0x2]\n" "tbz %x[flags], #3, 151f\n" - "ldr x20, [%x[input_ptr], x11, LSL #0x3]\n" + "ldr x20, [%x[input_ptr], x14, LSL #0x3]\n" "add x20, x20, x19, LSL #3\n" - "ldr x9, [x20, #0x0]\n" - "ldr x26, [x20, #0x8]\n" - "ldr x24, [x20, #0x10]\n" - "ldr x23, [x20, #0x18]\n" + "ldr x12, [x20, #0x0]\n" + "ldr x28, [x20, #0x8]\n" + "ldr x26, [x20, #0x10]\n" + "ldr x24, [x20, #0x18]\n" "ldr x22, [x20, #0x20]\n" - "cbnz x11, 152f\n" + "cbnz x14, 152f\n" "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x9, x9, x19\n" + "add x12, x12, x19\n" + "add x28, x28, x19\n" "add x26, x26, x19\n" "add x24, x24, x19\n" - "add x23, x23, x19\n" "add x22, x22, x19\n" "b 152f\n" "151:" // Height 5: setup direct input - "mov x9, %x[input_ptr]\n" - "add x26, x9, x19\n" + "mov x12, %x[input_ptr]\n" + "add x28, x12, x19\n" + "add x26, x28, x19\n" "add x24, x26, x19\n" - "add x23, x24, x19\n" - "add x22, x23, x19\n" + "add x22, x24, x19\n" "152:" // Height 5: input setup done - "cmp x10, #0x10\n" + "cmp x13, #0x10\n" "blt 155f\n" - "ldr q0, [x9, #0x0]\n" - "ldr q1, [x26, #0x0]\n" - "cmp x10, #0x20\n" - "ldr q2, [x24, #0x0]\n" - "ldr q3, [x23, #0x0]\n" + "ldr q0, [x12, #0x0]\n" + "ldr q1, [x28, #0x0]\n" + "cmp x13, #0x20\n" + "ldr q2, [x26, #0x0]\n" + "ldr q3, [x24, #0x0]\n" "ldr q4, [x22, #0x0]\n" - "ldr q6, [x13, #0x0]\n" + "ldr q6, [x16, #0x0]\n" "blt 154f\n" "153:" // Height 5: Multiply loop: Main loop head ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" + "ldr d7, [x16, #0x10]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" + "ldr x11, [x16, #0x18]\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" + "ldr x10, [x16, #0x28]\n" ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n" + "add x12, x12, #0x10\n" ".inst 0x4f84e0d8 // sdot v24.4s, v6.16b, v4.4b[0]\n" - "ldr d7, [x13, #0x10]\n" - "ldr x19, [x13, #0x18]\n" - "add x9, x9, #0x10\n" - "ldr d6, [x13, #0x20]\n" - "add x26, x26, #0x10\n" - "ldr x28, [x13, #0x28]\n" - "add x24, x24, #0x10\n" - "mov v7.d[1], x19\n" - "prfm pldl1keep, [x9, #0x80]\n" - "ldr x19, [x13, #0x38]\n" - "add x23, x23, #0x10\n" - "mov v6.d[1], x28\n" + "mov v7.d[1], x11\n" + "prfm pldl1keep, [x12, #0x80]\n" + "add x28, x28, #0x10\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" + "prfm pldl1keep, [x28, #0x80]\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" + "ldr d6, [x16, #0x20]\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" + "ldr x11, [x16, #0x38]\n" ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n" + "ldr x9, [x12, #0x8]\n" ".inst 0x4f84e0f9 // sdot v25.4s, v7.16b, v4.4b[0]\n" + "mov v6.d[1], x10\n" + "ldr d7, [x16, #0x30]\n" + "add x26, x26, #0x10\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "prfm pldl1keep, [x26, #0x80]\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" + "ldr x10, [x16, #0x48]\n" ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" + "mov v7.d[1], x11\n" ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n" + "ldr x11, [x16, #0x58]\n" ".inst 0x4f84e0da // sdot v26.4s, v6.16b, v4.4b[0]\n" - "ldr d7, [x13, #0x30]\n" - "prfm pldl1keep, [x26, #0x80]\n" - "add x22, x22, #0x10\n" - "ldr d6, [x13, #0x40]\n" - "sub x10, x10, #0x10\n" - "mov v7.d[1], x19\n" - "ldr x28, [x13, #0x48]\n" - "ldr x19, [x13, #0x58]\n" - "cmp x10, #0x20\n" + "ldr d6, [x16, #0x40]\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" + "ldr x27, [x28, #0x8]\n" ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" - "mov v6.d[1], x28\n" + "ldr x25, [x26, #0x8]\n" ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" + "mov v6.d[1], x10\n" ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n" + "ldr x10, [x16, #0x68]\n" ".inst 0x4f84e0fb // sdot v27.4s, v7.16b, v4.4b[0]\n" + "ldr d7, [x16, #0x50]\n" ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n" + "add x24, x24, #0x10\n" ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n" + "prfm pldl1keep, [x24, #0x80]\n" ".inst 0x4fa2e0d0 // sdot v16.4s, v6.16b, v2.4b[1]\n" + "mov v7.d[1], x11\n" ".inst 0x4fa3e0d4 // sdot v20.4s, v6.16b, v3.4b[1]\n" + "ldr x11, [x16, #0x78]\n" ".inst 0x4fa4e0d8 // sdot v24.4s, v6.16b, v4.4b[1]\n" - "ldr d7, [x13, #0x50]\n" - "prfm pldl1keep, [x24, #0x80]\n" - "ldr d6, [x13, #0x60]\n" - "mov v7.d[1], x19\n" - "ldr x28, [x13, #0x68]\n" - "ldr x19, [x13, #0x78]\n" + "ldr d6, [x16, #0x60]\n" ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n" + "ldr x23, [x24, #0x8]\n" ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n" - "mov v6.d[1], x28\n" + "add x22, x22, #0x10\n" ".inst 0x4fa2e0f1 // sdot v17.4s, v7.16b, v2.4b[1]\n" + "mov v6.d[1], x10\n" ".inst 0x4fa3e0f5 // sdot v21.4s, v7.16b, v3.4b[1]\n" + "prfm pldl1keep, [x22, #0x80]\n" ".inst 0x4fa4e0f9 // sdot v25.4s, v7.16b, v4.4b[1]\n" + "ldr d7, [x16, #0x70]\n" ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n" + "ldr x10, [x16, #0x88]\n" ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n" + "ldr x21, [x22, #0x8]\n" ".inst 0x4fa2e0d2 // sdot v18.4s, v6.16b, v2.4b[1]\n" + "mov v7.d[1], x11\n" ".inst 0x4fa3e0d6 // sdot v22.4s, v6.16b, v3.4b[1]\n" + "ldr x11, [x16, #0x98]\n" ".inst 0x4fa4e0da // sdot v26.4s, v6.16b, v4.4b[1]\n" - "ldr d7, [x13, #0x70]\n" - "prfm pldl1keep, [x23, #0x80]\n" - "ldr d6, [x13, #0x80]\n" - "mov v7.d[1], x19\n" - "ldr x28, [x13, #0x88]\n" - "ldr x19, [x13, #0x98]\n" + "ldr d6, [x16, #0x80]\n" ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n" + "sub x13, x13, #0x10\n" ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n" - "mov v6.d[1], x28\n" + "cmp x13, #0x20\n" ".inst 0x4fa2e0f3 // sdot v19.4s, v7.16b, v2.4b[1]\n" + "mov v6.d[1], x10\n" ".inst 0x4fa3e0f7 // sdot v23.4s, v7.16b, v3.4b[1]\n" + "ldr x10, [x16, #0xa8]\n" ".inst 0x4fa4e0fb // sdot v27.4s, v7.16b, v4.4b[1]\n" + "ldr d7, [x16, #0x90]\n" ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n" ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n" ".inst 0x4f82e8d0 // sdot v16.4s, v6.16b, v2.4b[2]\n" + "mov v7.d[1], x11\n" ".inst 0x4f83e8d4 // sdot v20.4s, v6.16b, v3.4b[2]\n" + "ldr x11, [x16, #0xb8]\n" ".inst 0x4f84e8d8 // sdot v24.4s, v6.16b, v4.4b[2]\n" - "ldr d7, [x13, #0x90]\n" - "prfm pldl1keep, [x22, #0x80]\n" - "ldr d6, [x13, #0xa0]\n" - "mov v7.d[1], x19\n" - "ldr x28, [x13, #0xa8]\n" - "ldr x19, [x13, #0xb8]\n" + "ldr d6, [x16, #0xa0]\n" ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n" - "mov v6.d[1], x28\n" ".inst 0x4f82e8f1 // sdot v17.4s, v7.16b, v2.4b[2]\n" + "mov v6.d[1], x10\n" ".inst 0x4f83e8f5 // sdot v21.4s, v7.16b, v3.4b[2]\n" + "ldr x10, [x16, #0xc8]\n" ".inst 0x4f84e8f9 // sdot v25.4s, v7.16b, v4.4b[2]\n" + "ldr d7, [x16, #0xb0]\n" ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n" ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n" ".inst 0x4f82e8d2 // sdot v18.4s, v6.16b, v2.4b[2]\n" + "mov v7.d[1], x11\n" ".inst 0x4f83e8d6 // sdot v22.4s, v6.16b, v3.4b[2]\n" + "ldr x11, [x16, #0xd8]\n" ".inst 0x4f84e8da // sdot v26.4s, v6.16b, v4.4b[2]\n" - "ldr d7, [x13, #0xb0]\n" - "ldr d6, [x13, #0xc0]\n" - "ldr x28, [x13, #0xc8]\n" - "mov v7.d[1], x19\n" - "ldr x19, [x13, #0xd8]\n" - "ldr x27, [x9, #0x8]\n" - "mov v6.d[1], x28\n" + "ldr d6, [x16, #0xc0]\n" ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n" ".inst 0x4f82e8f3 // sdot v19.4s, v7.16b, v2.4b[2]\n" + "mov v6.d[1], x10\n" ".inst 0x4f83e8f7 // sdot v23.4s, v7.16b, v3.4b[2]\n" + "ldr x10, [x16, #0xe8]\n" ".inst 0x4f84e8fb // sdot v27.4s, v7.16b, v4.4b[2]\n" + "ldr d7, [x16, #0xd0]\n" ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n" ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n" ".inst 0x4fa2e8d0 // sdot v16.4s, v6.16b, v2.4b[3]\n" + "mov v7.d[1], x11\n" ".inst 0x4fa3e8d4 // sdot v20.4s, v6.16b, v3.4b[3]\n" + "ldr x11, [x16, #0xf8]\n" ".inst 0x4fa4e8d8 // sdot v24.4s, v6.16b, v4.4b[3]\n" - "ldr d7, [x13, #0xd0]\n" - "ldr d6, [x13, #0xe0]\n" - "ldr x28, [x13, #0xe8]\n" - "mov v7.d[1], x19\n" - "ldr x19, [x13, #0xf8]\n" - "ldr x25, [x26, #0x8]\n" - "mov v6.d[1], x28\n" + "ldr d6, [x16, #0xe0]\n" ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n" ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n" ".inst 0x4fa2e8f1 // sdot v17.4s, v7.16b, v2.4b[3]\n" + "mov v6.d[1], x10\n" ".inst 0x4fa3e8f5 // sdot v21.4s, v7.16b, v3.4b[3]\n" ".inst 0x4fa4e8f9 // sdot v25.4s, v7.16b, v4.4b[3]\n" + "ldr d7, [x16, #0xf0]\n" ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n" + "add x16, x16, #0x100\n" ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n" + "ldr x10, [x16, #0x8]\n" ".inst 0x4fa2e8d2 // sdot v18.4s, v6.16b, v2.4b[3]\n" + "mov v7.d[1], x11\n" ".inst 0x4fa3e8d6 // sdot v22.4s, v6.16b, v3.4b[3]\n" ".inst 0x4fa4e8da // sdot v26.4s, v6.16b, v4.4b[3]\n" - "ldr d7, [x13, #0xf0]\n" - "ldr x21, [x24, #0x8]\n" - "add x13, x13, #0x100\n" - "ldr d6, [x13, #0x0]\n" - "mov v7.d[1], x19\n" - "ldr x28, [x13, #0x8]\n" - "ldr x19, [x23, #0x8]\n" + "ldr d6, [x16, #0x0]\n" ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n" + "ldr d0, [x12, #0x0]\n" ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n" - "mov v6.d[1], x28\n" + "ldr d1, [x28, #0x0]\n" ".inst 0x4fa2e8f3 // sdot v19.4s, v7.16b, v2.4b[3]\n" + "mov v6.d[1], x10\n" ".inst 0x4fa3e8f7 // sdot v23.4s, v7.16b, v3.4b[3]\n" + "mov v0.d[1], x9\n" ".inst 0x4fa4e8fb // sdot v27.4s, v7.16b, v4.4b[3]\n" - "ldr d0, [x9, #0x0]\n" - "ldr d1, [x26, #0x0]\n" - "ldr d2, [x24, #0x0]\n" - "mov v0.d[1], x27\n" - "ldr d3, [x23, #0x0]\n" - "mov v1.d[1], x25\n" + "mov v1.d[1], x27\n" + "ldr d2, [x26, #0x0]\n" + "ldr d3, [x24, #0x0]\n" "ldr d4, [x22, #0x0]\n" - "mov v2.d[1], x21\n" - "ldr x21, [x22, #0x8]\n" - "mov v3.d[1], x19\n" + "mov v2.d[1], x25\n" + "mov v3.d[1], x23\n" "mov v4.d[1], x21\n" "bge 153b\n" "154:" // Height 5: Multiply loop: Single iteration only ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" + "ldr q7, [x16, #0x10]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" + "sub x13, x13, #0x10\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" + "add x12, x12, #0x10\n" ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n" + "prfm pldl1keep, [x12, #0x80]\n" ".inst 0x4f84e0d8 // sdot v24.4s, v6.16b, v4.4b[0]\n" - "ldr q7, [x13, #0x10]\n" - "ldr q6, [x13, #0x20]\n" - "sub x10, x10, #0x10\n" - "add x9, x9, #0x10\n" - "add x26, x26, #0x10\n" + "ldr q6, [x16, #0x20]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" + "add x28, x28, #0x10\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" + "prfm pldl1keep, [x28, #0x80]\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" + "add x26, x26, #0x10\n" ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n" + "prfm pldl1keep, [x26, #0x80]\n" ".inst 0x4f84e0f9 // sdot v25.4s, v7.16b, v4.4b[0]\n" + "ldr q7, [x16, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "add x24, x24, #0x10\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" + "prfm pldl1keep, [x24, #0x80]\n" ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" + "add x22, x22, #0x10\n" ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n" + "prfm pldl1keep, [x22, #0x80]\n" ".inst 0x4f84e0da // sdot v26.4s, v6.16b, v4.4b[0]\n" - "prfm pldl1keep, [x9, #0x80]\n" - "add x24, x24, #0x10\n" - "ldr q7, [x13, #0x30]\n" - "add x23, x23, #0x10\n" - "prfm pldl1keep, [x26, #0x80]\n" - "add x22, x22, #0x10\n" - "ldr q6, [x13, #0x40]\n" + "ldr q6, [x16, #0x40]\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n" ".inst 0x4f84e0fb // sdot v27.4s, v7.16b, v4.4b[0]\n" + "ldr q7, [x16, #0x50]\n" ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n" ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n" ".inst 0x4fa2e0d0 // sdot v16.4s, v6.16b, v2.4b[1]\n" ".inst 0x4fa3e0d4 // sdot v20.4s, v6.16b, v3.4b[1]\n" ".inst 0x4fa4e0d8 // sdot v24.4s, v6.16b, v4.4b[1]\n" - "ldr q7, [x13, #0x50]\n" - "prfm pldl1keep, [x24, #0x80]\n" - "ldr q6, [x13, #0x60]\n" + "ldr q6, [x16, #0x60]\n" ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n" ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n" ".inst 0x4fa2e0f1 // sdot v17.4s, v7.16b, v2.4b[1]\n" ".inst 0x4fa3e0f5 // sdot v21.4s, v7.16b, v3.4b[1]\n" ".inst 0x4fa4e0f9 // sdot v25.4s, v7.16b, v4.4b[1]\n" + "ldr q7, [x16, #0x70]\n" ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n" ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n" ".inst 0x4fa2e0d2 // sdot v18.4s, v6.16b, v2.4b[1]\n" ".inst 0x4fa3e0d6 // sdot v22.4s, v6.16b, v3.4b[1]\n" ".inst 0x4fa4e0da // sdot v26.4s, v6.16b, v4.4b[1]\n" - "ldr q7, [x13, #0x70]\n" - "prfm pldl1keep, [x23, #0x80]\n" - "ldr q6, [x13, #0x80]\n" + "ldr q6, [x16, #0x80]\n" ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n" ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n" ".inst 0x4fa2e0f3 // sdot v19.4s, v7.16b, v2.4b[1]\n" ".inst 0x4fa3e0f7 // sdot v23.4s, v7.16b, v3.4b[1]\n" ".inst 0x4fa4e0fb // sdot v27.4s, v7.16b, v4.4b[1]\n" + "ldr q7, [x16, #0x90]\n" ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n" ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n" ".inst 0x4f82e8d0 // sdot v16.4s, v6.16b, v2.4b[2]\n" ".inst 0x4f83e8d4 // sdot v20.4s, v6.16b, v3.4b[2]\n" ".inst 0x4f84e8d8 // sdot v24.4s, v6.16b, v4.4b[2]\n" - "ldr q7, [x13, #0x90]\n" - "prfm pldl1keep, [x22, #0x80]\n" - "ldr q6, [x13, #0xa0]\n" + "ldr q6, [x16, #0xa0]\n" ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n" ".inst 0x4f82e8f1 // sdot v17.4s, v7.16b, v2.4b[2]\n" ".inst 0x4f83e8f5 // sdot v21.4s, v7.16b, v3.4b[2]\n" ".inst 0x4f84e8f9 // sdot v25.4s, v7.16b, v4.4b[2]\n" + "ldr q7, [x16, #0xb0]\n" ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n" ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n" ".inst 0x4f82e8d2 // sdot v18.4s, v6.16b, v2.4b[2]\n" ".inst 0x4f83e8d6 // sdot v22.4s, v6.16b, v3.4b[2]\n" ".inst 0x4f84e8da // sdot v26.4s, v6.16b, v4.4b[2]\n" - "ldr q7, [x13, #0xb0]\n" - "ldr q6, [x13, #0xc0]\n" + "ldr q6, [x16, #0xc0]\n" ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n" ".inst 0x4f82e8f3 // sdot v19.4s, v7.16b, v2.4b[2]\n" ".inst 0x4f83e8f7 // sdot v23.4s, v7.16b, v3.4b[2]\n" ".inst 0x4f84e8fb // sdot v27.4s, v7.16b, v4.4b[2]\n" + "ldr q7, [x16, #0xd0]\n" ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n" ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n" ".inst 0x4fa2e8d0 // sdot v16.4s, v6.16b, v2.4b[3]\n" ".inst 0x4fa3e8d4 // sdot v20.4s, v6.16b, v3.4b[3]\n" ".inst 0x4fa4e8d8 // sdot v24.4s, v6.16b, v4.4b[3]\n" - "ldr q7, [x13, #0xd0]\n" - "ldr q6, [x13, #0xe0]\n" + "ldr q6, [x16, #0xe0]\n" ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n" ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n" ".inst 0x4fa2e8f1 // sdot v17.4s, v7.16b, v2.4b[3]\n" ".inst 0x4fa3e8f5 // sdot v21.4s, v7.16b, v3.4b[3]\n" ".inst 0x4fa4e8f9 // sdot v25.4s, v7.16b, v4.4b[3]\n" + "ldr q7, [x16, #0xf0]\n" ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n" + "add x16, x16, #0x100\n" ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n" ".inst 0x4fa2e8d2 // sdot v18.4s, v6.16b, v2.4b[3]\n" ".inst 0x4fa3e8d6 // sdot v22.4s, v6.16b, v3.4b[3]\n" ".inst 0x4fa4e8da // sdot v26.4s, v6.16b, v4.4b[3]\n" - "ldr q7, [x13, #0xf0]\n" - "add x13, x13, #0x100\n" ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n" ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n" ".inst 0x4fa2e8f3 // sdot v19.4s, v7.16b, v2.4b[3]\n" ".inst 0x4fa3e8f7 // sdot v23.4s, v7.16b, v3.4b[3]\n" ".inst 0x4fa4e8fb // sdot v27.4s, v7.16b, v4.4b[3]\n" "155:" // Height 5: Multiply loop: Main loop skip - "cbz x10, 160f\n" - "cmp x10, #0x4\n" + "cbz x13, 160f\n" + "cmp x13, #0x4\n" "blt 157f\n" "156:" // Height 5: Multiply loop: Odd block loop - "ldr s0, [x9], #0x4\n" - "sub x10, x10, #0x4\n" - "ldr s1, [x26], #0x4\n" - "cmp x10, #0x4\n" - "ldr s2, [x24], #0x4\n" - "ldr s3, [x23], #0x4\n" + "ldr s0, [x12], #0x4\n" + "sub x13, x13, #0x4\n" + "ldr s1, [x28], #0x4\n" + "cmp x13, #0x4\n" + "ldr s2, [x26], #0x4\n" + "ldr s3, [x24], #0x4\n" "ldr s4, [x22], #0x4\n" - "ldr q6, [x13, #0x0]\n" - "ldr q7, [x13, #0x10]\n" + "ldr q6, [x16, #0x0]\n" + "ldr q7, [x16, #0x10]\n" ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n" ".inst 0x4f84e0d8 // sdot v24.4s, v6.16b, v4.4b[0]\n" + "ldr q6, [x16, #0x20]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n" ".inst 0x4f84e0f9 // sdot v25.4s, v7.16b, v4.4b[0]\n" - "ldr q6, [x13, #0x20]\n" - "ldr q7, [x13, #0x30]\n" - "add x13, x13, #0x40\n" + "ldr q7, [x16, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "add x16, x16, #0x40\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n" @@ -2454,44 +2454,44 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n" ".inst 0x4f84e0fb // sdot v27.4s, v7.16b, v4.4b[0]\n" "bge 156b\n" - "cbz x10, 160f\n" + "cbz x13, 160f\n" "157:" // Height 5: Multiply loop: Skip odd blocks - "tbz x10, #1, 158f\n" - "ldr h0, [x9], #0x2\n" - "ldr h1, [x26], #0x2\n" - "ldr h2, [x24], #0x2\n" - "ldr h3, [x23], #0x2\n" + "tbz x13, #1, 158f\n" + "ldr h0, [x12], #0x2\n" + "ldr h1, [x28], #0x2\n" + "ldr h2, [x26], #0x2\n" + "ldr h3, [x24], #0x2\n" "ldr h4, [x22], #0x2\n" - "tbz x10, #0, 159f\n" - "ld1 { v0.b }[2], [x9]\n" - "ld1 { v1.b }[2], [x26]\n" - "ld1 { v2.b }[2], [x24]\n" - "ld1 { v3.b }[2], [x23]\n" + "tbz x13, #0, 159f\n" + "ld1 { v0.b }[2], [x12]\n" + "ld1 { v1.b }[2], [x28]\n" + "ld1 { v2.b }[2], [x26]\n" + "ld1 { v3.b }[2], [x24]\n" "ld1 { v4.b }[2], [x22]\n" "b 159f\n" "158:" // Height 5: Multiply loop: Ragged operand read: partial_1_0 - "ldr b0, [x9, #0x0]\n" - "ldr b1, [x26, #0x0]\n" - "ldr b2, [x24, #0x0]\n" - "ldr b3, [x23, #0x0]\n" + "ldr b0, [x12, #0x0]\n" + "ldr b1, [x28, #0x0]\n" + "ldr b2, [x26, #0x0]\n" + "ldr b3, [x24, #0x0]\n" "ldr b4, [x22, #0x0]\n" "159:" // Height 5: Multiply loop: Ragged operand read: Done - "ldr q6, [x13, #0x0]\n" - "ldr q7, [x13, #0x10]\n" + "ldr q6, [x16, #0x0]\n" + "ldr q7, [x16, #0x10]\n" ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n" ".inst 0x4f84e0d8 // sdot v24.4s, v6.16b, v4.4b[0]\n" + "ldr q6, [x16, #0x20]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n" ".inst 0x4f84e0f9 // sdot v25.4s, v7.16b, v4.4b[0]\n" - "ldr q6, [x13, #0x20]\n" - "ldr q7, [x13, #0x30]\n" - "add x13, x13, #0x40\n" + "ldr q7, [x16, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "add x16, x16, #0x40\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n" @@ -2503,13 +2503,13 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( ".inst 0x4f84e0fb // sdot v27.4s, v7.16b, v4.4b[0]\n" "160:" // Height 5: Multiply loop: No odd multiplies "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" - "add x11, x11, #0x1\n" - "cmp x11, x19\n" + "add x14, x14, #0x1\n" + "cmp x14, x19\n" "bne 150b\n" "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "prfm pstl1keep, [x12, #0x0]\n" - "cmp x14, #0x10\n" - "add x23, x12, x19, LSL #2\n" + "prfm pstl1keep, [x15, #0x0]\n" + "cmp x17, #0x10\n" + "add x23, x15, x19, LSL #2\n" "prfm pstl1keep, [x23, #0x0]\n" "add x22, x23, x19, LSL #2\n" "prfm pstl1keep, [x22, #0x0]\n" @@ -2518,9 +2518,9 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( "add x20, x21, x19, LSL #2\n" "prfm pstl1keep, [x20, #0x0]\n" "bge 169f\n" - "tbz x14, #3, 164f\n" - "st1 { v8.4s }, [x12], #0x10\n" - "st1 { v9.4s }, [x12], #0x10\n" + "tbz x17, #3, 164f\n" + "st1 { v8.4s }, [x15], #0x10\n" + "st1 { v9.4s }, [x15], #0x10\n" "st1 { v12.4s }, [x23], #0x10\n" "st1 { v13.4s }, [x23], #0x10\n" "st1 { v16.4s }, [x22], #0x10\n" @@ -2529,99 +2529,99 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( "st1 { v21.4s }, [x21], #0x10\n" "st1 { v24.4s }, [x20], #0x10\n" "st1 { v25.4s }, [x20], #0x10\n" - "tbz x14, #2, 162f\n" - "st1 { v10.4s }, [x12], #0x10\n" + "tbz x17, #2, 162f\n" + "st1 { v10.4s }, [x15], #0x10\n" "st1 { v14.4s }, [x23], #0x10\n" "st1 { v18.4s }, [x22], #0x10\n" "st1 { v22.4s }, [x21], #0x10\n" "st1 { v26.4s }, [x20], #0x10\n" - "tbz x14, #1, 161f\n" - "str d11, [x12], #0x8\n" + "tbz x17, #1, 161f\n" + "str d11, [x15], #0x8\n" "str d15, [x23], #0x8\n" "str d19, [x22], #0x8\n" "str d23, [x21], #0x8\n" "str d27, [x20], #0x8\n" - "tbz x14, #0, 168f\n" - "st1 { v11.s }[2], [x12]\n" + "tbz x17, #0, 168f\n" + "st1 { v11.s }[2], [x15]\n" "st1 { v15.s }[2], [x23]\n" "st1 { v19.s }[2], [x22]\n" "st1 { v23.s }[2], [x21]\n" "st1 { v27.s }[2], [x20]\n" "b 168f\n" "161:" // Height 5: Partial direct writeback: partial_1_12 - "tbz x14, #0, 168f\n" - "str s11, [x12, #0x0]\n" + "tbz x17, #0, 168f\n" + "str s11, [x15, #0x0]\n" "str s15, [x23, #0x0]\n" "str s19, [x22, #0x0]\n" "str s23, [x21, #0x0]\n" "str s27, [x20, #0x0]\n" "b 168f\n" "162:" // Height 5: Partial direct writeback: partial_2_8 - "tbz x14, #1, 163f\n" - "str d10, [x12], #0x8\n" + "tbz x17, #1, 163f\n" + "str d10, [x15], #0x8\n" "str d14, [x23], #0x8\n" "str d18, [x22], #0x8\n" "str d22, [x21], #0x8\n" "str d26, [x20], #0x8\n" - "tbz x14, #0, 168f\n" - "st1 { v10.s }[2], [x12]\n" + "tbz x17, #0, 168f\n" + "st1 { v10.s }[2], [x15]\n" "st1 { v14.s }[2], [x23]\n" "st1 { v18.s }[2], [x22]\n" "st1 { v22.s }[2], [x21]\n" "st1 { v26.s }[2], [x20]\n" "b 168f\n" "163:" // Height 5: Partial direct writeback: partial_1_8 - "tbz x14, #0, 168f\n" - "str s10, [x12, #0x0]\n" + "tbz x17, #0, 168f\n" + "str s10, [x15, #0x0]\n" "str s14, [x23, #0x0]\n" "str s18, [x22, #0x0]\n" "str s22, [x21, #0x0]\n" "str s26, [x20, #0x0]\n" "b 168f\n" "164:" // Height 5: Partial direct writeback: partial_4_0 - "tbz x14, #2, 166f\n" - "st1 { v8.4s }, [x12], #0x10\n" + "tbz x17, #2, 166f\n" + "st1 { v8.4s }, [x15], #0x10\n" "st1 { v12.4s }, [x23], #0x10\n" "st1 { v16.4s }, [x22], #0x10\n" "st1 { v20.4s }, [x21], #0x10\n" "st1 { v24.4s }, [x20], #0x10\n" - "tbz x14, #1, 165f\n" - "str d9, [x12], #0x8\n" + "tbz x17, #1, 165f\n" + "str d9, [x15], #0x8\n" "str d13, [x23], #0x8\n" "str d17, [x22], #0x8\n" "str d21, [x21], #0x8\n" "str d25, [x20], #0x8\n" - "tbz x14, #0, 168f\n" - "st1 { v9.s }[2], [x12]\n" + "tbz x17, #0, 168f\n" + "st1 { v9.s }[2], [x15]\n" "st1 { v13.s }[2], [x23]\n" "st1 { v17.s }[2], [x22]\n" "st1 { v21.s }[2], [x21]\n" "st1 { v25.s }[2], [x20]\n" "b 168f\n" "165:" // Height 5: Partial direct writeback: partial_1_4 - "tbz x14, #0, 168f\n" - "str s9, [x12, #0x0]\n" + "tbz x17, #0, 168f\n" + "str s9, [x15, #0x0]\n" "str s13, [x23, #0x0]\n" "str s17, [x22, #0x0]\n" "str s21, [x21, #0x0]\n" "str s25, [x20, #0x0]\n" "b 168f\n" "166:" // Height 5: Partial direct writeback: partial_2_0 - "tbz x14, #1, 167f\n" - "str d8, [x12], #0x8\n" + "tbz x17, #1, 167f\n" + "str d8, [x15], #0x8\n" "str d12, [x23], #0x8\n" "str d16, [x22], #0x8\n" "str d20, [x21], #0x8\n" "str d24, [x20], #0x8\n" - "tbz x14, #0, 168f\n" - "st1 { v8.s }[2], [x12]\n" + "tbz x17, #0, 168f\n" + "st1 { v8.s }[2], [x15]\n" "st1 { v12.s }[2], [x23]\n" "st1 { v16.s }[2], [x22]\n" "st1 { v20.s }[2], [x21]\n" "st1 { v24.s }[2], [x20]\n" "b 168f\n" "167:" // Height 5: Partial direct writeback: partial_1_0 - "str s8, [x12, #0x0]\n" + "str s8, [x15, #0x0]\n" "str s12, [x23, #0x0]\n" "str s16, [x22, #0x0]\n" "str s20, [x21, #0x0]\n" @@ -2629,11 +2629,11 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( "168:" // Height 5: Partial direct writeback: Done "b 170f\n" "169:" // Height 5: Full writeback - "str q8, [x12, #0x0]\n" - "str q9, [x12, #0x10]\n" - "str q10, [x12, #0x20]\n" - "str q11, [x12, #0x30]\n" - "add x12, x12, #0x40\n" + "str q8, [x15, #0x0]\n" + "str q9, [x15, #0x10]\n" + "str q10, [x15, #0x20]\n" + "str q11, [x15, #0x30]\n" + "add x15, x15, #0x40\n" "str q12, [x23, #0x0]\n" "str q13, [x23, #0x10]\n" "str q14, [x23, #0x20]\n" @@ -2651,31 +2651,31 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( "str q26, [x20, #0x20]\n" "str q27, [x20, #0x30]\n" "170:" // Height 5: Writeback done - "subs x14, x14, #0x10\n" + "subs x17, x17, #0x10\n" "bgt 138b\n" "b 206f\n" "171:" // Height 6 - "ldr x14, [%x[args_ptr], %[offsetof_N]]\n" - "mov x12, %x[output_ptr]\n" - "ldr x13, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "ldr x17, [%x[args_ptr], %[offsetof_N]]\n" + "mov x15, %x[output_ptr]\n" + "ldr x16, [%x[args_ptr], %[offsetof_B_ptr]]\n" "mov x20, #0x18\n" "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" "madd %x[output_ptr], x19, x20, %x[output_ptr]\n" "172:" // Height 6: Column loop "tbz %x[flags], #0, 182f\n" "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "cmp x14, #0x10\n" - "add x23, x12, x19, LSL #2\n" + "cmp x17, #0x10\n" + "add x23, x15, x19, LSL #2\n" "add x22, x23, x19, LSL #2\n" "add x21, x22, x19, LSL #2\n" "add x20, x21, x19, LSL #2\n" "add x19, x20, x19, LSL #2\n" "bge 181f\n" - "tbz x14, #3, 176f\n" - "ld1 { v8.4s }, [x12], #0x10\n" + "tbz x17, #3, 176f\n" + "ld1 { v8.4s }, [x15], #0x10\n" "ld1 { v12.4s }, [x23], #0x10\n" "ld1 { v16.4s }, [x22], #0x10\n" - "ld1 { v9.4s }, [x12], #0x10\n" + "ld1 { v9.4s }, [x15], #0x10\n" "ld1 { v13.4s }, [x23], #0x10\n" "ld1 { v17.4s }, [x22], #0x10\n" "ld1 { v20.4s }, [x21], #0x10\n" @@ -2684,23 +2684,23 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( "ld1 { v21.4s }, [x21], #0x10\n" "ld1 { v25.4s }, [x20], #0x10\n" "ld1 { v29.4s }, [x19], #0x10\n" - "tbz x14, #2, 174f\n" - "ld1 { v10.4s }, [x12], #0x10\n" + "tbz x17, #2, 174f\n" + "ld1 { v10.4s }, [x15], #0x10\n" "ld1 { v14.4s }, [x23], #0x10\n" "ld1 { v18.4s }, [x22], #0x10\n" "ld1 { v22.4s }, [x21], #0x10\n" "ld1 { v26.4s }, [x20], #0x10\n" "ld1 { v30.4s }, [x19], #0x10\n" - "tbz x14, #1, 173f\n" - "ldr d11, [x12], #0x8\n" + "tbz x17, #1, 173f\n" + "ldr d11, [x15], #0x8\n" "mov x24, #0x38\n" "ldr d15, [x23], #0x8\n" "ldr d19, [x22], #0x8\n" "ldr d23, [x21], #0x8\n" "ldr d27, [x20], #0x8\n" "ldr d31, [x19], #0x8\n" - "tbz x14, #0, 180f\n" - "ld1 { v11.s }[2], [x12]\n" + "tbz x17, #0, 180f\n" + "ld1 { v11.s }[2], [x15]\n" "ld1 { v15.s }[2], [x23]\n" "ld1 { v19.s }[2], [x22]\n" "ld1 { v23.s }[2], [x21]\n" @@ -2709,8 +2709,8 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( "b 180f\n" "173:" // Height 6: Partial accumulate: partial_1_12 "mov x24, #0x30\n" - "tbz x14, #0, 180f\n" - "ldr s11, [x12, #0x0]\n" + "tbz x17, #0, 180f\n" + "ldr s11, [x15, #0x0]\n" "ldr s15, [x23, #0x0]\n" "ldr s19, [x22, #0x0]\n" "ldr s23, [x21, #0x0]\n" @@ -2718,16 +2718,16 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( "ldr s31, [x19, #0x0]\n" "b 180f\n" "174:" // Height 6: Partial accumulate: partial_2_8 - "tbz x14, #1, 175f\n" - "ldr d10, [x12], #0x8\n" + "tbz x17, #1, 175f\n" + "ldr d10, [x15], #0x8\n" "ldr d14, [x23], #0x8\n" "mov x24, #0x28\n" "ldr d18, [x22], #0x8\n" "ldr d22, [x21], #0x8\n" "ldr d26, [x20], #0x8\n" "ldr d30, [x19], #0x8\n" - "tbz x14, #0, 180f\n" - "ld1 { v10.s }[2], [x12]\n" + "tbz x17, #0, 180f\n" + "ld1 { v10.s }[2], [x15]\n" "ld1 { v14.s }[2], [x23]\n" "ld1 { v18.s }[2], [x22]\n" "ld1 { v22.s }[2], [x21]\n" @@ -2736,8 +2736,8 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( "b 180f\n" "175:" // Height 6: Partial accumulate: partial_1_8 "mov x24, #0x20\n" - "tbz x14, #0, 180f\n" - "ldr s10, [x12, #0x0]\n" + "tbz x17, #0, 180f\n" + "ldr s10, [x15, #0x0]\n" "ldr s14, [x23, #0x0]\n" "ldr s18, [x22, #0x0]\n" "ldr s22, [x21, #0x0]\n" @@ -2745,23 +2745,23 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( "ldr s30, [x19, #0x0]\n" "b 180f\n" "176:" // Height 6: Partial accumulate: partial_4_0 - "tbz x14, #2, 178f\n" - "ld1 { v8.4s }, [x12], #0x10\n" + "tbz x17, #2, 178f\n" + "ld1 { v8.4s }, [x15], #0x10\n" "ld1 { v12.4s }, [x23], #0x10\n" "ld1 { v16.4s }, [x22], #0x10\n" "ld1 { v20.4s }, [x21], #0x10\n" "ld1 { v24.4s }, [x20], #0x10\n" "ld1 { v28.4s }, [x19], #0x10\n" - "tbz x14, #1, 177f\n" - "ldr d9, [x12], #0x8\n" + "tbz x17, #1, 177f\n" + "ldr d9, [x15], #0x8\n" "mov x24, #0x18\n" "ldr d13, [x23], #0x8\n" "ldr d17, [x22], #0x8\n" "ldr d21, [x21], #0x8\n" "ldr d25, [x20], #0x8\n" "ldr d29, [x19], #0x8\n" - "tbz x14, #0, 180f\n" - "ld1 { v9.s }[2], [x12]\n" + "tbz x17, #0, 180f\n" + "ld1 { v9.s }[2], [x15]\n" "ld1 { v13.s }[2], [x23]\n" "ld1 { v17.s }[2], [x22]\n" "ld1 { v21.s }[2], [x21]\n" @@ -2770,8 +2770,8 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( "b 180f\n" "177:" // Height 6: Partial accumulate: partial_1_4 "mov x24, #0x10\n" - "tbz x14, #0, 180f\n" - "ldr s9, [x12, #0x0]\n" + "tbz x17, #0, 180f\n" + "ldr s9, [x15, #0x0]\n" "ldr s13, [x23, #0x0]\n" "ldr s17, [x22, #0x0]\n" "ldr s21, [x21, #0x0]\n" @@ -2779,16 +2779,16 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( "ldr s29, [x19, #0x0]\n" "b 180f\n" "178:" // Height 6: Partial accumulate: partial_2_0 - "tbz x14, #1, 179f\n" - "ldr d8, [x12], #0x8\n" + "tbz x17, #1, 179f\n" + "ldr d8, [x15], #0x8\n" "ldr d12, [x23], #0x8\n" "mov x24, #0x8\n" "ldr d16, [x22], #0x8\n" "ldr d20, [x21], #0x8\n" "ldr d24, [x20], #0x8\n" "ldr d28, [x19], #0x8\n" - "tbz x14, #0, 180f\n" - "ld1 { v8.s }[2], [x12]\n" + "tbz x17, #0, 180f\n" + "ld1 { v8.s }[2], [x15]\n" "ld1 { v12.s }[2], [x23]\n" "ld1 { v16.s }[2], [x22]\n" "ld1 { v20.s }[2], [x21]\n" @@ -2796,7 +2796,7 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( "ld1 { v28.s }[2], [x19]\n" "b 180f\n" "179:" // Height 6: Partial accumulate: partial_1_0 - "ldr s8, [x12, #0x0]\n" + "ldr s8, [x15, #0x0]\n" "mov x24, #0x0\n" "ldr s12, [x23, #0x0]\n" "ldr s16, [x22, #0x0]\n" @@ -2804,13 +2804,13 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( "ldr s24, [x20, #0x0]\n" "ldr s28, [x19, #0x0]\n" "180:" // Height 6: Partial accumulate: Done - "sub x12, x12, x24\n" + "sub x15, x15, x24\n" "b 183f\n" "181:" // Height 6: full accumulate - "ldr q8, [x12, #0x0]\n" - "ldr q9, [x12, #0x10]\n" - "ldr q10, [x12, #0x20]\n" - "ldr q11, [x12, #0x30]\n" + "ldr q8, [x15, #0x0]\n" + "ldr q9, [x15, #0x10]\n" + "ldr q10, [x15, #0x20]\n" + "ldr q11, [x15, #0x30]\n" "ldr q12, [x23, #0x0]\n" "ldr q13, [x23, #0x10]\n" "ldr q14, [x23, #0x20]\n" @@ -2858,347 +2858,347 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( "movi v30.4s, #0x0\n" "movi v31.4s, #0x0\n" "183:" // Height 6: setup done - "mov x11, #0x0\n" + "mov x14, #0x0\n" "184:" // Height 6: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w10, [x20, x11, LSL #0x2]\n" + "ldr w13, [x20, x14, LSL #0x2]\n" "tbz %x[flags], #3, 185f\n" - "ldr x20, [%x[input_ptr], x11, LSL #0x3]\n" + "ldr x20, [%x[input_ptr], x14, LSL #0x3]\n" "add x20, x20, x19, LSL #3\n" - "ldr x9, [x20, #0x0]\n" - "ldr x26, [x20, #0x8]\n" - "ldr x24, [x20, #0x10]\n" - "ldr x23, [x20, #0x18]\n" + "ldr x12, [x20, #0x0]\n" + "ldr x28, [x20, #0x8]\n" + "ldr x26, [x20, #0x10]\n" + "ldr x24, [x20, #0x18]\n" "ldr x22, [x20, #0x20]\n" "ldr x20, [x20, #0x28]\n" - "cbnz x11, 186f\n" + "cbnz x14, 186f\n" "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x9, x9, x19\n" + "add x12, x12, x19\n" + "add x28, x28, x19\n" "add x26, x26, x19\n" "add x24, x24, x19\n" - "add x23, x23, x19\n" "add x22, x22, x19\n" "add x20, x20, x19\n" "b 186f\n" "185:" // Height 6: setup direct input - "mov x9, %x[input_ptr]\n" - "add x26, x9, x19\n" + "mov x12, %x[input_ptr]\n" + "add x28, x12, x19\n" + "add x26, x28, x19\n" "add x24, x26, x19\n" - "add x23, x24, x19\n" - "add x22, x23, x19\n" + "add x22, x24, x19\n" "add x20, x22, x19\n" "186:" // Height 6: input setup done - "cmp x10, #0x10\n" + "cmp x13, #0x10\n" "blt 189f\n" - "ldr q0, [x9, #0x0]\n" - "ldr q1, [x26, #0x0]\n" - "cmp x10, #0x20\n" - "ldr q2, [x24, #0x0]\n" - "ldr q3, [x23, #0x0]\n" + "ldr q0, [x12, #0x0]\n" + "ldr q1, [x28, #0x0]\n" + "cmp x13, #0x20\n" + "ldr q2, [x26, #0x0]\n" + "ldr q3, [x24, #0x0]\n" "ldr q4, [x22, #0x0]\n" "ldr q5, [x20, #0x0]\n" - "ldr q6, [x13, #0x0]\n" + "ldr q6, [x16, #0x0]\n" "blt 188f\n" "187:" // Height 6: Multiply loop: Main loop head ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" + "ldr d7, [x16, #0x10]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" + "ldr x11, [x16, #0x18]\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" + "ldr x10, [x16, #0x28]\n" ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n" + "add x12, x12, #0x10\n" ".inst 0x4f84e0d8 // sdot v24.4s, v6.16b, v4.4b[0]\n" + "mov v7.d[1], x11\n" ".inst 0x4f85e0dc // sdot v28.4s, v6.16b, v5.4b[0]\n" - "ldr d7, [x13, #0x10]\n" - "add x9, x9, #0x10\n" - "ldr x19, [x13, #0x18]\n" - "add x26, x26, #0x10\n" - "prfm pldl1keep, [x9, #0x80]\n" - "add x24, x24, #0x10\n" - "ldr d6, [x13, #0x20]\n" - "add x23, x23, #0x10\n" - "mov v7.d[1], x19\n" - "ldr x28, [x13, #0x28]\n" - "ldr x19, [x13, #0x38]\n" - "add x22, x22, #0x10\n" + "prfm pldl1keep, [x12, #0x80]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" + "ldr d6, [x16, #0x20]\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" - "mov v6.d[1], x28\n" + "ldr x11, [x16, #0x38]\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" + "ldr x9, [x12, #0x8]\n" ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n" + "mov v6.d[1], x10\n" ".inst 0x4f84e0f9 // sdot v25.4s, v7.16b, v4.4b[0]\n" + "ldr x10, [x16, #0x48]\n" ".inst 0x4f85e0fd // sdot v29.4s, v7.16b, v5.4b[0]\n" + "ldr d7, [x16, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "add x28, x28, #0x10\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" + "prfm pldl1keep, [x28, #0x80]\n" ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" + "mov v7.d[1], x11\n" ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n" + "ldr x11, [x16, #0x58]\n" ".inst 0x4f84e0da // sdot v26.4s, v6.16b, v4.4b[0]\n" + "ldr x27, [x28, #0x8]\n" ".inst 0x4f85e0de // sdot v30.4s, v6.16b, v5.4b[0]\n" - "ldr d7, [x13, #0x30]\n" - "prfm pldl1keep, [x26, #0x80]\n" - "add x20, x20, #0x10\n" - "ldr d6, [x13, #0x40]\n" - "sub x10, x10, #0x10\n" - "mov v7.d[1], x19\n" - "ldr x28, [x13, #0x48]\n" - "ldr x19, [x13, #0x58]\n" - "cmp x10, #0x20\n" + "ldr d6, [x16, #0x40]\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" + "add x26, x26, #0x10\n" ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" - "mov v6.d[1], x28\n" + "prfm pldl1keep, [x26, #0x80]\n" ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" + "mov v6.d[1], x10\n" ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n" + "ldr x10, [x16, #0x68]\n" ".inst 0x4f84e0fb // sdot v27.4s, v7.16b, v4.4b[0]\n" + "ldr x25, [x26, #0x8]\n" ".inst 0x4f85e0ff // sdot v31.4s, v7.16b, v5.4b[0]\n" + "ldr d7, [x16, #0x50]\n" ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n" + "add x24, x24, #0x10\n" ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n" + "prfm pldl1keep, [x24, #0x80]\n" ".inst 0x4fa2e0d0 // sdot v16.4s, v6.16b, v2.4b[1]\n" + "mov v7.d[1], x11\n" ".inst 0x4fa3e0d4 // sdot v20.4s, v6.16b, v3.4b[1]\n" + "ldr x11, [x16, #0x78]\n" ".inst 0x4fa4e0d8 // sdot v24.4s, v6.16b, v4.4b[1]\n" + "ldr x23, [x24, #0x8]\n" ".inst 0x4fa5e0dc // sdot v28.4s, v6.16b, v5.4b[1]\n" - "ldr d7, [x13, #0x50]\n" - "prfm pldl1keep, [x24, #0x80]\n" - "ldr d6, [x13, #0x60]\n" - "mov v7.d[1], x19\n" - "ldr x28, [x13, #0x68]\n" - "ldr x19, [x13, #0x78]\n" + "ldr d6, [x16, #0x60]\n" ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n" + "add x22, x22, #0x10\n" ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n" - "mov v6.d[1], x28\n" + "prfm pldl1keep, [x22, #0x80]\n" ".inst 0x4fa2e0f1 // sdot v17.4s, v7.16b, v2.4b[1]\n" + "mov v6.d[1], x10\n" ".inst 0x4fa3e0f5 // sdot v21.4s, v7.16b, v3.4b[1]\n" + "ldr x10, [x16, #0x88]\n" ".inst 0x4fa4e0f9 // sdot v25.4s, v7.16b, v4.4b[1]\n" + "ldr x21, [x22, #0x8]\n" ".inst 0x4fa5e0fd // sdot v29.4s, v7.16b, v5.4b[1]\n" + "ldr d7, [x16, #0x70]\n" ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n" + "add x20, x20, #0x10\n" ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n" + "prfm pldl1keep, [x20, #0x80]\n" ".inst 0x4fa2e0d2 // sdot v18.4s, v6.16b, v2.4b[1]\n" + "mov v7.d[1], x11\n" ".inst 0x4fa3e0d6 // sdot v22.4s, v6.16b, v3.4b[1]\n" + "ldr x11, [x16, #0x98]\n" ".inst 0x4fa4e0da // sdot v26.4s, v6.16b, v4.4b[1]\n" + "ldr x19, [x20, #0x8]\n" ".inst 0x4fa5e0de // sdot v30.4s, v6.16b, v5.4b[1]\n" - "ldr d7, [x13, #0x70]\n" - "prfm pldl1keep, [x23, #0x80]\n" - "ldr d6, [x13, #0x80]\n" - "mov v7.d[1], x19\n" - "ldr x28, [x13, #0x88]\n" - "ldr x19, [x13, #0x98]\n" + "ldr d6, [x16, #0x80]\n" ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n" + "sub x13, x13, #0x10\n" ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n" - "mov v6.d[1], x28\n" + "cmp x13, #0x20\n" ".inst 0x4fa2e0f3 // sdot v19.4s, v7.16b, v2.4b[1]\n" + "mov v6.d[1], x10\n" ".inst 0x4fa3e0f7 // sdot v23.4s, v7.16b, v3.4b[1]\n" + "ldr x10, [x16, #0xa8]\n" ".inst 0x4fa4e0fb // sdot v27.4s, v7.16b, v4.4b[1]\n" ".inst 0x4fa5e0ff // sdot v31.4s, v7.16b, v5.4b[1]\n" + "ldr d7, [x16, #0x90]\n" ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n" ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n" ".inst 0x4f82e8d0 // sdot v16.4s, v6.16b, v2.4b[2]\n" + "mov v7.d[1], x11\n" ".inst 0x4f83e8d4 // sdot v20.4s, v6.16b, v3.4b[2]\n" + "ldr x11, [x16, #0xb8]\n" ".inst 0x4f84e8d8 // sdot v24.4s, v6.16b, v4.4b[2]\n" ".inst 0x4f85e8dc // sdot v28.4s, v6.16b, v5.4b[2]\n" - "ldr d7, [x13, #0x90]\n" - "prfm pldl1keep, [x22, #0x80]\n" - "ldr d6, [x13, #0xa0]\n" - "mov v7.d[1], x19\n" - "ldr x28, [x13, #0xa8]\n" - "ldr x19, [x13, #0xb8]\n" + "ldr d6, [x16, #0xa0]\n" ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n" - "mov v6.d[1], x28\n" ".inst 0x4f82e8f1 // sdot v17.4s, v7.16b, v2.4b[2]\n" + "mov v6.d[1], x10\n" ".inst 0x4f83e8f5 // sdot v21.4s, v7.16b, v3.4b[2]\n" + "ldr x10, [x16, #0xc8]\n" ".inst 0x4f84e8f9 // sdot v25.4s, v7.16b, v4.4b[2]\n" ".inst 0x4f85e8fd // sdot v29.4s, v7.16b, v5.4b[2]\n" + "ldr d7, [x16, #0xb0]\n" ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n" ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n" ".inst 0x4f82e8d2 // sdot v18.4s, v6.16b, v2.4b[2]\n" + "mov v7.d[1], x11\n" ".inst 0x4f83e8d6 // sdot v22.4s, v6.16b, v3.4b[2]\n" + "ldr x11, [x16, #0xd8]\n" ".inst 0x4f84e8da // sdot v26.4s, v6.16b, v4.4b[2]\n" ".inst 0x4f85e8de // sdot v30.4s, v6.16b, v5.4b[2]\n" - "ldr d7, [x13, #0xb0]\n" - "prfm pldl1keep, [x20, #0x80]\n" - "ldr d6, [x13, #0xc0]\n" - "mov v7.d[1], x19\n" - "ldr x28, [x13, #0xc8]\n" - "ldr x19, [x13, #0xd8]\n" + "ldr d6, [x16, #0xc0]\n" ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n" - "mov v6.d[1], x28\n" ".inst 0x4f82e8f3 // sdot v19.4s, v7.16b, v2.4b[2]\n" + "mov v6.d[1], x10\n" ".inst 0x4f83e8f7 // sdot v23.4s, v7.16b, v3.4b[2]\n" + "ldr x10, [x16, #0xe8]\n" ".inst 0x4f84e8fb // sdot v27.4s, v7.16b, v4.4b[2]\n" ".inst 0x4f85e8ff // sdot v31.4s, v7.16b, v5.4b[2]\n" + "ldr d7, [x16, #0xd0]\n" ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n" ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n" ".inst 0x4fa2e8d0 // sdot v16.4s, v6.16b, v2.4b[3]\n" + "mov v7.d[1], x11\n" ".inst 0x4fa3e8d4 // sdot v20.4s, v6.16b, v3.4b[3]\n" + "ldr x11, [x16, #0xf8]\n" ".inst 0x4fa4e8d8 // sdot v24.4s, v6.16b, v4.4b[3]\n" ".inst 0x4fa5e8dc // sdot v28.4s, v6.16b, v5.4b[3]\n" - "ldr d7, [x13, #0xd0]\n" - "ldr d6, [x13, #0xe0]\n" - "ldr x28, [x13, #0xe8]\n" - "mov v7.d[1], x19\n" - "ldr x19, [x13, #0xf8]\n" - "ldr x27, [x9, #0x8]\n" - "mov v6.d[1], x28\n" + "ldr d6, [x16, #0xe0]\n" ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n" ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n" ".inst 0x4fa2e8f1 // sdot v17.4s, v7.16b, v2.4b[3]\n" + "mov v6.d[1], x10\n" ".inst 0x4fa3e8f5 // sdot v21.4s, v7.16b, v3.4b[3]\n" ".inst 0x4fa4e8f9 // sdot v25.4s, v7.16b, v4.4b[3]\n" ".inst 0x4fa5e8fd // sdot v29.4s, v7.16b, v5.4b[3]\n" + "ldr d7, [x16, #0xf0]\n" ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n" + "add x16, x16, #0x100\n" ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n" + "ldr x10, [x16, #0x8]\n" ".inst 0x4fa2e8d2 // sdot v18.4s, v6.16b, v2.4b[3]\n" + "mov v7.d[1], x11\n" ".inst 0x4fa3e8d6 // sdot v22.4s, v6.16b, v3.4b[3]\n" ".inst 0x4fa4e8da // sdot v26.4s, v6.16b, v4.4b[3]\n" ".inst 0x4fa5e8de // sdot v30.4s, v6.16b, v5.4b[3]\n" - "ldr d7, [x13, #0xf0]\n" - "ldr x25, [x26, #0x8]\n" - "add x13, x13, #0x100\n" - "ldr d6, [x13, #0x0]\n" - "mov v7.d[1], x19\n" - "ldr x28, [x13, #0x8]\n" - "ldr x21, [x24, #0x8]\n" + "ldr d6, [x16, #0x0]\n" ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n" + "ldr d0, [x12, #0x0]\n" ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n" - "mov v6.d[1], x28\n" + "ldr d1, [x28, #0x0]\n" ".inst 0x4fa2e8f3 // sdot v19.4s, v7.16b, v2.4b[3]\n" + "mov v6.d[1], x10\n" ".inst 0x4fa3e8f7 // sdot v23.4s, v7.16b, v3.4b[3]\n" + "mov v0.d[1], x9\n" ".inst 0x4fa4e8fb // sdot v27.4s, v7.16b, v4.4b[3]\n" + "mov v1.d[1], x27\n" ".inst 0x4fa5e8ff // sdot v31.4s, v7.16b, v5.4b[3]\n" - "ldr d0, [x9, #0x0]\n" - "ldr d1, [x26, #0x0]\n" - "ldr d2, [x24, #0x0]\n" - "mov v0.d[1], x27\n" - "ldr d3, [x23, #0x0]\n" - "mov v1.d[1], x25\n" - "ldr x19, [x23, #0x8]\n" - "mov v2.d[1], x21\n" + "ldr d2, [x26, #0x0]\n" + "ldr d3, [x24, #0x0]\n" "ldr d4, [x22, #0x0]\n" - "ldr x21, [x22, #0x8]\n" - "mov v3.d[1], x19\n" + "mov v2.d[1], x25\n" "ldr d5, [x20, #0x0]\n" - "ldr x19, [x20, #0x8]\n" + "mov v3.d[1], x23\n" "mov v4.d[1], x21\n" "mov v5.d[1], x19\n" "bge 187b\n" "188:" // Height 6: Multiply loop: Single iteration only ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" + "ldr q7, [x16, #0x10]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" + "sub x13, x13, #0x10\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" + "add x12, x12, #0x10\n" ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n" + "prfm pldl1keep, [x12, #0x80]\n" ".inst 0x4f84e0d8 // sdot v24.4s, v6.16b, v4.4b[0]\n" + "add x28, x28, #0x10\n" ".inst 0x4f85e0dc // sdot v28.4s, v6.16b, v5.4b[0]\n" - "ldr q7, [x13, #0x10]\n" - "sub x10, x10, #0x10\n" - "ldr q6, [x13, #0x20]\n" - "add x9, x9, #0x10\n" - "prfm pldl1keep, [x9, #0x80]\n" - "add x26, x26, #0x10\n" + "prfm pldl1keep, [x28, #0x80]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" + "ldr q6, [x16, #0x20]\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" + "add x26, x26, #0x10\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" + "prfm pldl1keep, [x26, #0x80]\n" ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n" + "add x24, x24, #0x10\n" ".inst 0x4f84e0f9 // sdot v25.4s, v7.16b, v4.4b[0]\n" + "prfm pldl1keep, [x24, #0x80]\n" ".inst 0x4f85e0fd // sdot v29.4s, v7.16b, v5.4b[0]\n" + "ldr q7, [x16, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "add x22, x22, #0x10\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" + "prfm pldl1keep, [x22, #0x80]\n" ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" + "add x20, x20, #0x10\n" ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n" + "prfm pldl1keep, [x20, #0x80]\n" ".inst 0x4f84e0da // sdot v26.4s, v6.16b, v4.4b[0]\n" ".inst 0x4f85e0de // sdot v30.4s, v6.16b, v5.4b[0]\n" - "ldr q7, [x13, #0x30]\n" - "add x24, x24, #0x10\n" - "prfm pldl1keep, [x26, #0x80]\n" - "add x23, x23, #0x10\n" - "ldr q6, [x13, #0x40]\n" - "add x22, x22, #0x10\n" + "ldr q6, [x16, #0x40]\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n" ".inst 0x4f84e0fb // sdot v27.4s, v7.16b, v4.4b[0]\n" ".inst 0x4f85e0ff // sdot v31.4s, v7.16b, v5.4b[0]\n" + "ldr q7, [x16, #0x50]\n" ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n" ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n" ".inst 0x4fa2e0d0 // sdot v16.4s, v6.16b, v2.4b[1]\n" ".inst 0x4fa3e0d4 // sdot v20.4s, v6.16b, v3.4b[1]\n" ".inst 0x4fa4e0d8 // sdot v24.4s, v6.16b, v4.4b[1]\n" ".inst 0x4fa5e0dc // sdot v28.4s, v6.16b, v5.4b[1]\n" - "ldr q7, [x13, #0x50]\n" - "add x20, x20, #0x10\n" - "prfm pldl1keep, [x24, #0x80]\n" - "ldr q6, [x13, #0x60]\n" + "ldr q6, [x16, #0x60]\n" ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n" ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n" ".inst 0x4fa2e0f1 // sdot v17.4s, v7.16b, v2.4b[1]\n" ".inst 0x4fa3e0f5 // sdot v21.4s, v7.16b, v3.4b[1]\n" ".inst 0x4fa4e0f9 // sdot v25.4s, v7.16b, v4.4b[1]\n" ".inst 0x4fa5e0fd // sdot v29.4s, v7.16b, v5.4b[1]\n" + "ldr q7, [x16, #0x70]\n" ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n" ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n" ".inst 0x4fa2e0d2 // sdot v18.4s, v6.16b, v2.4b[1]\n" ".inst 0x4fa3e0d6 // sdot v22.4s, v6.16b, v3.4b[1]\n" ".inst 0x4fa4e0da // sdot v26.4s, v6.16b, v4.4b[1]\n" ".inst 0x4fa5e0de // sdot v30.4s, v6.16b, v5.4b[1]\n" - "ldr q7, [x13, #0x70]\n" - "prfm pldl1keep, [x23, #0x80]\n" - "ldr q6, [x13, #0x80]\n" + "ldr q6, [x16, #0x80]\n" ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n" ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n" ".inst 0x4fa2e0f3 // sdot v19.4s, v7.16b, v2.4b[1]\n" ".inst 0x4fa3e0f7 // sdot v23.4s, v7.16b, v3.4b[1]\n" ".inst 0x4fa4e0fb // sdot v27.4s, v7.16b, v4.4b[1]\n" ".inst 0x4fa5e0ff // sdot v31.4s, v7.16b, v5.4b[1]\n" + "ldr q7, [x16, #0x90]\n" ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n" ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n" ".inst 0x4f82e8d0 // sdot v16.4s, v6.16b, v2.4b[2]\n" ".inst 0x4f83e8d4 // sdot v20.4s, v6.16b, v3.4b[2]\n" ".inst 0x4f84e8d8 // sdot v24.4s, v6.16b, v4.4b[2]\n" ".inst 0x4f85e8dc // sdot v28.4s, v6.16b, v5.4b[2]\n" - "ldr q7, [x13, #0x90]\n" - "prfm pldl1keep, [x22, #0x80]\n" - "ldr q6, [x13, #0xa0]\n" + "ldr q6, [x16, #0xa0]\n" ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n" ".inst 0x4f82e8f1 // sdot v17.4s, v7.16b, v2.4b[2]\n" ".inst 0x4f83e8f5 // sdot v21.4s, v7.16b, v3.4b[2]\n" ".inst 0x4f84e8f9 // sdot v25.4s, v7.16b, v4.4b[2]\n" ".inst 0x4f85e8fd // sdot v29.4s, v7.16b, v5.4b[2]\n" + "ldr q7, [x16, #0xb0]\n" ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n" ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n" ".inst 0x4f82e8d2 // sdot v18.4s, v6.16b, v2.4b[2]\n" ".inst 0x4f83e8d6 // sdot v22.4s, v6.16b, v3.4b[2]\n" ".inst 0x4f84e8da // sdot v26.4s, v6.16b, v4.4b[2]\n" ".inst 0x4f85e8de // sdot v30.4s, v6.16b, v5.4b[2]\n" - "ldr q7, [x13, #0xb0]\n" - "prfm pldl1keep, [x20, #0x80]\n" - "ldr q6, [x13, #0xc0]\n" + "ldr q6, [x16, #0xc0]\n" ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n" ".inst 0x4f82e8f3 // sdot v19.4s, v7.16b, v2.4b[2]\n" ".inst 0x4f83e8f7 // sdot v23.4s, v7.16b, v3.4b[2]\n" ".inst 0x4f84e8fb // sdot v27.4s, v7.16b, v4.4b[2]\n" ".inst 0x4f85e8ff // sdot v31.4s, v7.16b, v5.4b[2]\n" + "ldr q7, [x16, #0xd0]\n" ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n" ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n" ".inst 0x4fa2e8d0 // sdot v16.4s, v6.16b, v2.4b[3]\n" ".inst 0x4fa3e8d4 // sdot v20.4s, v6.16b, v3.4b[3]\n" ".inst 0x4fa4e8d8 // sdot v24.4s, v6.16b, v4.4b[3]\n" ".inst 0x4fa5e8dc // sdot v28.4s, v6.16b, v5.4b[3]\n" - "ldr q7, [x13, #0xd0]\n" - "ldr q6, [x13, #0xe0]\n" + "ldr q6, [x16, #0xe0]\n" ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n" ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n" ".inst 0x4fa2e8f1 // sdot v17.4s, v7.16b, v2.4b[3]\n" ".inst 0x4fa3e8f5 // sdot v21.4s, v7.16b, v3.4b[3]\n" ".inst 0x4fa4e8f9 // sdot v25.4s, v7.16b, v4.4b[3]\n" ".inst 0x4fa5e8fd // sdot v29.4s, v7.16b, v5.4b[3]\n" + "ldr q7, [x16, #0xf0]\n" ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n" + "add x16, x16, #0x100\n" ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n" ".inst 0x4fa2e8d2 // sdot v18.4s, v6.16b, v2.4b[3]\n" ".inst 0x4fa3e8d6 // sdot v22.4s, v6.16b, v3.4b[3]\n" ".inst 0x4fa4e8da // sdot v26.4s, v6.16b, v4.4b[3]\n" ".inst 0x4fa5e8de // sdot v30.4s, v6.16b, v5.4b[3]\n" - "ldr q7, [x13, #0xf0]\n" - "add x13, x13, #0x100\n" ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n" ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n" ".inst 0x4fa2e8f3 // sdot v19.4s, v7.16b, v2.4b[3]\n" @@ -3206,36 +3206,36 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( ".inst 0x4fa4e8fb // sdot v27.4s, v7.16b, v4.4b[3]\n" ".inst 0x4fa5e8ff // sdot v31.4s, v7.16b, v5.4b[3]\n" "189:" // Height 6: Multiply loop: Main loop skip - "cbz x10, 194f\n" - "cmp x10, #0x4\n" + "cbz x13, 194f\n" + "cmp x13, #0x4\n" "blt 191f\n" "190:" // Height 6: Multiply loop: Odd block loop - "ldr s0, [x9], #0x4\n" - "sub x10, x10, #0x4\n" - "ldr s1, [x26], #0x4\n" - "cmp x10, #0x4\n" - "ldr s2, [x24], #0x4\n" - "ldr s3, [x23], #0x4\n" + "ldr s0, [x12], #0x4\n" + "sub x13, x13, #0x4\n" + "ldr s1, [x28], #0x4\n" + "cmp x13, #0x4\n" + "ldr s2, [x26], #0x4\n" + "ldr s3, [x24], #0x4\n" "ldr s4, [x22], #0x4\n" "ldr s5, [x20], #0x4\n" - "ldr q6, [x13, #0x0]\n" - "ldr q7, [x13, #0x10]\n" + "ldr q6, [x16, #0x0]\n" + "ldr q7, [x16, #0x10]\n" ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n" ".inst 0x4f84e0d8 // sdot v24.4s, v6.16b, v4.4b[0]\n" ".inst 0x4f85e0dc // sdot v28.4s, v6.16b, v5.4b[0]\n" + "ldr q6, [x16, #0x20]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n" ".inst 0x4f84e0f9 // sdot v25.4s, v7.16b, v4.4b[0]\n" ".inst 0x4f85e0fd // sdot v29.4s, v7.16b, v5.4b[0]\n" - "ldr q6, [x13, #0x20]\n" - "ldr q7, [x13, #0x30]\n" - "add x13, x13, #0x40\n" + "ldr q7, [x16, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "add x16, x16, #0x40\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n" @@ -3248,49 +3248,49 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( ".inst 0x4f84e0fb // sdot v27.4s, v7.16b, v4.4b[0]\n" ".inst 0x4f85e0ff // sdot v31.4s, v7.16b, v5.4b[0]\n" "bge 190b\n" - "cbz x10, 194f\n" + "cbz x13, 194f\n" "191:" // Height 6: Multiply loop: Skip odd blocks - "tbz x10, #1, 192f\n" - "ldr h0, [x9], #0x2\n" - "ldr h1, [x26], #0x2\n" - "ldr h2, [x24], #0x2\n" - "ldr h3, [x23], #0x2\n" + "tbz x13, #1, 192f\n" + "ldr h0, [x12], #0x2\n" + "ldr h1, [x28], #0x2\n" + "ldr h2, [x26], #0x2\n" + "ldr h3, [x24], #0x2\n" "ldr h4, [x22], #0x2\n" "ldr h5, [x20], #0x2\n" - "tbz x10, #0, 193f\n" - "ld1 { v0.b }[2], [x9]\n" - "ld1 { v1.b }[2], [x26]\n" - "ld1 { v2.b }[2], [x24]\n" - "ld1 { v3.b }[2], [x23]\n" + "tbz x13, #0, 193f\n" + "ld1 { v0.b }[2], [x12]\n" + "ld1 { v1.b }[2], [x28]\n" + "ld1 { v2.b }[2], [x26]\n" + "ld1 { v3.b }[2], [x24]\n" "ld1 { v4.b }[2], [x22]\n" "ld1 { v5.b }[2], [x20]\n" "b 193f\n" "192:" // Height 6: Multiply loop: Ragged operand read: partial_1_0 - "ldr b0, [x9, #0x0]\n" - "ldr b1, [x26, #0x0]\n" - "ldr b2, [x24, #0x0]\n" - "ldr b3, [x23, #0x0]\n" + "ldr b0, [x12, #0x0]\n" + "ldr b1, [x28, #0x0]\n" + "ldr b2, [x26, #0x0]\n" + "ldr b3, [x24, #0x0]\n" "ldr b4, [x22, #0x0]\n" "ldr b5, [x20, #0x0]\n" "193:" // Height 6: Multiply loop: Ragged operand read: Done - "ldr q6, [x13, #0x0]\n" - "ldr q7, [x13, #0x10]\n" + "ldr q6, [x16, #0x0]\n" + "ldr q7, [x16, #0x10]\n" ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n" ".inst 0x4f84e0d8 // sdot v24.4s, v6.16b, v4.4b[0]\n" ".inst 0x4f85e0dc // sdot v28.4s, v6.16b, v5.4b[0]\n" + "ldr q6, [x16, #0x20]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n" ".inst 0x4f84e0f9 // sdot v25.4s, v7.16b, v4.4b[0]\n" ".inst 0x4f85e0fd // sdot v29.4s, v7.16b, v5.4b[0]\n" - "ldr q6, [x13, #0x20]\n" - "ldr q7, [x13, #0x30]\n" - "add x13, x13, #0x40\n" + "ldr q7, [x16, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "add x16, x16, #0x40\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n" @@ -3304,13 +3304,13 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( ".inst 0x4f85e0ff // sdot v31.4s, v7.16b, v5.4b[0]\n" "194:" // Height 6: Multiply loop: No odd multiplies "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" - "add x11, x11, #0x1\n" - "cmp x11, x19\n" + "add x14, x14, #0x1\n" + "cmp x14, x19\n" "bne 184b\n" "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "prfm pstl1keep, [x12, #0x0]\n" - "cmp x14, #0x10\n" - "add x23, x12, x19, LSL #2\n" + "prfm pstl1keep, [x15, #0x0]\n" + "cmp x17, #0x10\n" + "add x23, x15, x19, LSL #2\n" "prfm pstl1keep, [x23, #0x0]\n" "add x22, x23, x19, LSL #2\n" "prfm pstl1keep, [x22, #0x0]\n" @@ -3321,9 +3321,9 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( "add x19, x20, x19, LSL #2\n" "prfm pstl1keep, [x19, #0x0]\n" "bge 203f\n" - "tbz x14, #3, 198f\n" - "st1 { v8.4s }, [x12], #0x10\n" - "st1 { v9.4s }, [x12], #0x10\n" + "tbz x17, #3, 198f\n" + "st1 { v8.4s }, [x15], #0x10\n" + "st1 { v9.4s }, [x15], #0x10\n" "st1 { v12.4s }, [x23], #0x10\n" "st1 { v13.4s }, [x23], #0x10\n" "st1 { v16.4s }, [x22], #0x10\n" @@ -3334,22 +3334,22 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( "st1 { v25.4s }, [x20], #0x10\n" "st1 { v28.4s }, [x19], #0x10\n" "st1 { v29.4s }, [x19], #0x10\n" - "tbz x14, #2, 196f\n" - "st1 { v10.4s }, [x12], #0x10\n" + "tbz x17, #2, 196f\n" + "st1 { v10.4s }, [x15], #0x10\n" "st1 { v14.4s }, [x23], #0x10\n" "st1 { v18.4s }, [x22], #0x10\n" "st1 { v22.4s }, [x21], #0x10\n" "st1 { v26.4s }, [x20], #0x10\n" "st1 { v30.4s }, [x19], #0x10\n" - "tbz x14, #1, 195f\n" - "str d11, [x12], #0x8\n" + "tbz x17, #1, 195f\n" + "str d11, [x15], #0x8\n" "str d15, [x23], #0x8\n" "str d19, [x22], #0x8\n" "str d23, [x21], #0x8\n" "str d27, [x20], #0x8\n" "str d31, [x19], #0x8\n" - "tbz x14, #0, 202f\n" - "st1 { v11.s }[2], [x12]\n" + "tbz x17, #0, 202f\n" + "st1 { v11.s }[2], [x15]\n" "st1 { v15.s }[2], [x23]\n" "st1 { v19.s }[2], [x22]\n" "st1 { v23.s }[2], [x21]\n" @@ -3357,8 +3357,8 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( "st1 { v31.s }[2], [x19]\n" "b 202f\n" "195:" // Height 6: Partial direct writeback: partial_1_12 - "tbz x14, #0, 202f\n" - "str s11, [x12, #0x0]\n" + "tbz x17, #0, 202f\n" + "str s11, [x15, #0x0]\n" "str s15, [x23, #0x0]\n" "str s19, [x22, #0x0]\n" "str s23, [x21, #0x0]\n" @@ -3366,15 +3366,15 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( "str s31, [x19, #0x0]\n" "b 202f\n" "196:" // Height 6: Partial direct writeback: partial_2_8 - "tbz x14, #1, 197f\n" - "str d10, [x12], #0x8\n" + "tbz x17, #1, 197f\n" + "str d10, [x15], #0x8\n" "str d14, [x23], #0x8\n" "str d18, [x22], #0x8\n" "str d22, [x21], #0x8\n" "str d26, [x20], #0x8\n" "str d30, [x19], #0x8\n" - "tbz x14, #0, 202f\n" - "st1 { v10.s }[2], [x12]\n" + "tbz x17, #0, 202f\n" + "st1 { v10.s }[2], [x15]\n" "st1 { v14.s }[2], [x23]\n" "st1 { v18.s }[2], [x22]\n" "st1 { v22.s }[2], [x21]\n" @@ -3382,8 +3382,8 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( "st1 { v30.s }[2], [x19]\n" "b 202f\n" "197:" // Height 6: Partial direct writeback: partial_1_8 - "tbz x14, #0, 202f\n" - "str s10, [x12, #0x0]\n" + "tbz x17, #0, 202f\n" + "str s10, [x15, #0x0]\n" "str s14, [x23, #0x0]\n" "str s18, [x22, #0x0]\n" "str s22, [x21, #0x0]\n" @@ -3391,22 +3391,22 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( "str s30, [x19, #0x0]\n" "b 202f\n" "198:" // Height 6: Partial direct writeback: partial_4_0 - "tbz x14, #2, 200f\n" - "st1 { v8.4s }, [x12], #0x10\n" + "tbz x17, #2, 200f\n" + "st1 { v8.4s }, [x15], #0x10\n" "st1 { v12.4s }, [x23], #0x10\n" "st1 { v16.4s }, [x22], #0x10\n" "st1 { v20.4s }, [x21], #0x10\n" "st1 { v24.4s }, [x20], #0x10\n" "st1 { v28.4s }, [x19], #0x10\n" - "tbz x14, #1, 199f\n" - "str d9, [x12], #0x8\n" + "tbz x17, #1, 199f\n" + "str d9, [x15], #0x8\n" "str d13, [x23], #0x8\n" "str d17, [x22], #0x8\n" "str d21, [x21], #0x8\n" "str d25, [x20], #0x8\n" "str d29, [x19], #0x8\n" - "tbz x14, #0, 202f\n" - "st1 { v9.s }[2], [x12]\n" + "tbz x17, #0, 202f\n" + "st1 { v9.s }[2], [x15]\n" "st1 { v13.s }[2], [x23]\n" "st1 { v17.s }[2], [x22]\n" "st1 { v21.s }[2], [x21]\n" @@ -3414,8 +3414,8 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( "st1 { v29.s }[2], [x19]\n" "b 202f\n" "199:" // Height 6: Partial direct writeback: partial_1_4 - "tbz x14, #0, 202f\n" - "str s9, [x12, #0x0]\n" + "tbz x17, #0, 202f\n" + "str s9, [x15, #0x0]\n" "str s13, [x23, #0x0]\n" "str s17, [x22, #0x0]\n" "str s21, [x21, #0x0]\n" @@ -3423,15 +3423,15 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( "str s29, [x19, #0x0]\n" "b 202f\n" "200:" // Height 6: Partial direct writeback: partial_2_0 - "tbz x14, #1, 201f\n" - "str d8, [x12], #0x8\n" + "tbz x17, #1, 201f\n" + "str d8, [x15], #0x8\n" "str d12, [x23], #0x8\n" "str d16, [x22], #0x8\n" "str d20, [x21], #0x8\n" "str d24, [x20], #0x8\n" "str d28, [x19], #0x8\n" - "tbz x14, #0, 202f\n" - "st1 { v8.s }[2], [x12]\n" + "tbz x17, #0, 202f\n" + "st1 { v8.s }[2], [x15]\n" "st1 { v12.s }[2], [x23]\n" "st1 { v16.s }[2], [x22]\n" "st1 { v20.s }[2], [x21]\n" @@ -3439,7 +3439,7 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( "st1 { v28.s }[2], [x19]\n" "b 202f\n" "201:" // Height 6: Partial direct writeback: partial_1_0 - "str s8, [x12, #0x0]\n" + "str s8, [x15, #0x0]\n" "str s12, [x23, #0x0]\n" "str s16, [x22, #0x0]\n" "str s20, [x21, #0x0]\n" @@ -3448,11 +3448,11 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( "202:" // Height 6: Partial direct writeback: Done "b 204f\n" "203:" // Height 6: Full writeback - "str q8, [x12, #0x0]\n" - "str q9, [x12, #0x10]\n" - "str q10, [x12, #0x20]\n" - "str q11, [x12, #0x30]\n" - "add x12, x12, #0x40\n" + "str q8, [x15, #0x0]\n" + "str q9, [x15, #0x10]\n" + "str q10, [x15, #0x20]\n" + "str q11, [x15, #0x30]\n" + "add x15, x15, #0x40\n" "str q12, [x23, #0x0]\n" "str q13, [x23, #0x10]\n" "str q14, [x23, #0x20]\n" @@ -3474,7 +3474,7 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( "str q30, [x19, #0x20]\n" "str q31, [x19, #0x30]\n" "204:" // Height 6: Writeback done - "subs x14, x14, #0x10\n" + "subs x17, x17, #0x10\n" "bgt 172b\n" "subs %x[M], %x[M], #0x6\n" "beq 206f\n" @@ -3491,7 +3491,7 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr) : [args_ptr] "r" (&ka), [flags] "r" (flags), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)) - : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x13", "x14", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" ); } diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_dot_4x16/a55.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_dot_4x16/a55.cpp index 690085cb3b..954e2891fb 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_dot_4x16/a55.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_dot_4x16/a55.cpp @@ -130,11 +130,11 @@ void a64_hybrid_u8qa_dot_4x16_a55 ( "mov v5.d[1], x24\n" "ldr d7, [x10, #0x30]\n" "ldr x19, [x10, #0x38]\n" - "mov v6.d[1], x23\n" ".inst 0x6f80e0b1 // udot v17.4s, v5.16b, v0.4b[0]\n" + "mov v6.d[1], x23\n" "ldr d8, [x10, #0x40]\n" - "mov v7.d[1], x19\n" ".inst 0x6f80e0d2 // udot v18.4s, v6.16b, v0.4b[0]\n" + "mov v7.d[1], x19\n" "ldr x23, [x10, #0x48]\n" ".inst 0x6f80e0f3 // udot v19.4s, v7.16b, v0.4b[0]\n" "ldr d9, [x10, #0x50]\n" @@ -142,11 +142,11 @@ void a64_hybrid_u8qa_dot_4x16_a55 ( "mov v8.d[1], x23\n" "ldr d10, [x10, #0x60]\n" "ldr x23, [x10, #0x68]\n" - "mov v9.d[1], x19\n" ".inst 0x6fa0e110 // udot v16.4s, v8.16b, v0.4b[1]\n" + "mov v9.d[1], x19\n" "ldr d4, [x10, #0x70]\n" - "mov v10.d[1], x23\n" ".inst 0x6fa0e131 // udot v17.4s, v9.16b, v0.4b[1]\n" + "mov v10.d[1], x23\n" "ldr x19, [x10, #0x78]\n" ".inst 0x6fa0e152 // udot v18.4s, v10.16b, v0.4b[1]\n" "ldr d5, [x10, #0x80]\n" @@ -154,11 +154,11 @@ void a64_hybrid_u8qa_dot_4x16_a55 ( "mov v4.d[1], x19\n" "ldr d6, [x10, #0x90]\n" "ldr x23, [x10, #0x98]\n" - "mov v5.d[1], x24\n" ".inst 0x6fa0e093 // udot v19.4s, v4.16b, v0.4b[1]\n" + "mov v5.d[1], x24\n" "ldr d7, [x10, #0xa0]\n" - "mov v6.d[1], x23\n" ".inst 0x6f80e8b0 // udot v16.4s, v5.16b, v0.4b[2]\n" + "mov v6.d[1], x23\n" "ldr x19, [x10, #0xa8]\n" ".inst 0x6f80e8d1 // udot v17.4s, v6.16b, v0.4b[2]\n" "ldr d8, [x10, #0xb0]\n" @@ -166,11 +166,11 @@ void a64_hybrid_u8qa_dot_4x16_a55 ( "mov v7.d[1], x19\n" "ldr d9, [x10, #0xc0]\n" "ldr x19, [x10, #0xc8]\n" - "mov v8.d[1], x23\n" ".inst 0x6f80e8f2 // udot v18.4s, v7.16b, v0.4b[2]\n" + "mov v8.d[1], x23\n" "ldr d10, [x10, #0xd0]\n" - "mov v9.d[1], x19\n" ".inst 0x6f80e913 // udot v19.4s, v8.16b, v0.4b[2]\n" + "mov v9.d[1], x19\n" "ldr x23, [x10, #0xd8]\n" ".inst 0x6fa0e930 // udot v16.4s, v9.16b, v0.4b[3]\n" "ldr d4, [x10, #0xe0]\n" @@ -179,10 +179,10 @@ void a64_hybrid_u8qa_dot_4x16_a55 ( "ldr d5, [x10, #0xf0]\n" "ldr x24, [x10, #0xf8]\n" "add x10, x10, #0x100\n" - "mov v4.d[1], x19\n" ".inst 0x6fa0e951 // udot v17.4s, v10.16b, v0.4b[3]\n" - "mov v5.d[1], x24\n" + "mov v4.d[1], x19\n" ".inst 0x6fa0e892 // udot v18.4s, v4.16b, v0.4b[3]\n" + "mov v5.d[1], x24\n" ".inst 0x6fa0e8b3 // udot v19.4s, v5.16b, v0.4b[3]\n" "tbnz %x[flags], #31, 8f\n" ".inst 0x6e8f940b // udot v11.4s, v0.16b, v15.16b\n" @@ -208,24 +208,24 @@ void a64_hybrid_u8qa_dot_4x16_a55 ( "ldr q10, [x10, #0x60]\n" ".inst 0x6fa0e110 // udot v16.4s, v8.16b, v0.4b[1]\n" "ldr q4, [x10, #0x70]\n" - "ldr q5, [x10, #0x80]\n" ".inst 0x6fa0e131 // udot v17.4s, v9.16b, v0.4b[1]\n" - "ldr q6, [x10, #0x90]\n" + "ldr q5, [x10, #0x80]\n" ".inst 0x6fa0e152 // udot v18.4s, v10.16b, v0.4b[1]\n" - "ldr q7, [x10, #0xa0]\n" + "ldr q6, [x10, #0x90]\n" ".inst 0x6fa0e093 // udot v19.4s, v4.16b, v0.4b[1]\n" - "ldr q8, [x10, #0xb0]\n" + "ldr q7, [x10, #0xa0]\n" ".inst 0x6f80e8b0 // udot v16.4s, v5.16b, v0.4b[2]\n" - "ldr q9, [x10, #0xc0]\n" - "ldr q10, [x10, #0xd0]\n" + "ldr q8, [x10, #0xb0]\n" ".inst 0x6f80e8d1 // udot v17.4s, v6.16b, v0.4b[2]\n" - "ldr q4, [x10, #0xe0]\n" + "ldr q9, [x10, #0xc0]\n" ".inst 0x6f80e8f2 // udot v18.4s, v7.16b, v0.4b[2]\n" - "ldr q5, [x10, #0xf0]\n" + "ldr q10, [x10, #0xd0]\n" ".inst 0x6f80e913 // udot v19.4s, v8.16b, v0.4b[2]\n" - "add x10, x10, #0x100\n" + "ldr q4, [x10, #0xe0]\n" ".inst 0x6fa0e930 // udot v16.4s, v9.16b, v0.4b[3]\n" + "ldr q5, [x10, #0xf0]\n" ".inst 0x6fa0e951 // udot v17.4s, v10.16b, v0.4b[3]\n" + "add x10, x10, #0x100\n" ".inst 0x6fa0e892 // udot v18.4s, v4.16b, v0.4b[3]\n" ".inst 0x6fa0e8b3 // udot v19.4s, v5.16b, v0.4b[3]\n" "tbnz %x[flags], #31, 10f\n" @@ -451,83 +451,83 @@ void a64_hybrid_u8qa_dot_4x16_a55 ( "blt 39f\n" "37:" // Height 2: Multiply loop: Main loop head ".inst 0x6f80e090 // udot v16.4s, v4.16b, v0.4b[0]\n" - ".inst 0x6f81e094 // udot v20.4s, v4.16b, v1.4b[0]\n" "ldr d5, [x10, #0x10]\n" - "add x25, x25, #0x10\n" + ".inst 0x6f81e094 // udot v20.4s, v4.16b, v1.4b[0]\n" "ldr x24, [x10, #0x18]\n" - "add x22, x22, #0x10\n" "ldr d6, [x10, #0x20]\n" + "add x25, x25, #0x10\n" "ldr x23, [x10, #0x28]\n" + "add x22, x22, #0x10\n" "mov v5.d[1], x24\n" "ldr d7, [x10, #0x30]\n" "ldr x19, [x10, #0x38]\n" - "mov v6.d[1], x23\n" ".inst 0x6f80e0b1 // udot v17.4s, v5.16b, v0.4b[0]\n" + "mov v6.d[1], x23\n" ".inst 0x6f81e0b5 // udot v21.4s, v5.16b, v1.4b[0]\n" "ldr d8, [x10, #0x40]\n" - "mov v7.d[1], x19\n" ".inst 0x6f80e0d2 // udot v18.4s, v6.16b, v0.4b[0]\n" + "mov v7.d[1], x19\n" ".inst 0x6f81e0d6 // udot v22.4s, v6.16b, v1.4b[0]\n" "ldr x23, [x10, #0x48]\n" ".inst 0x6f80e0f3 // udot v19.4s, v7.16b, v0.4b[0]\n" - ".inst 0x6f81e0f7 // udot v23.4s, v7.16b, v1.4b[0]\n" "ldr d9, [x10, #0x50]\n" - "mov v8.d[1], x23\n" + ".inst 0x6f81e0f7 // udot v23.4s, v7.16b, v1.4b[0]\n" "ldr x19, [x10, #0x58]\n" + "mov v8.d[1], x23\n" "ldr d10, [x10, #0x60]\n" + "ldr x23, [x10, #0x68]\n" ".inst 0x6fa0e110 // udot v16.4s, v8.16b, v0.4b[1]\n" - ".inst 0x6fa1e114 // udot v20.4s, v8.16b, v1.4b[1]\n" "mov v9.d[1], x19\n" - "ldr x23, [x10, #0x68]\n" + ".inst 0x6fa1e114 // udot v20.4s, v8.16b, v1.4b[1]\n" "ldr d4, [x10, #0x70]\n" ".inst 0x6fa0e131 // udot v17.4s, v9.16b, v0.4b[1]\n" - ".inst 0x6fa1e135 // udot v21.4s, v9.16b, v1.4b[1]\n" "mov v10.d[1], x23\n" + ".inst 0x6fa1e135 // udot v21.4s, v9.16b, v1.4b[1]\n" "ldr x19, [x10, #0x78]\n" - "ldr d5, [x10, #0x80]\n" ".inst 0x6fa0e152 // udot v18.4s, v10.16b, v0.4b[1]\n" + "ldr d5, [x10, #0x80]\n" ".inst 0x6fa1e156 // udot v22.4s, v10.16b, v1.4b[1]\n" - "mov v4.d[1], x19\n" "ldr x24, [x10, #0x88]\n" + "mov v4.d[1], x19\n" "ldr d6, [x10, #0x90]\n" + "ldr x23, [x10, #0x98]\n" ".inst 0x6fa0e093 // udot v19.4s, v4.16b, v0.4b[1]\n" - ".inst 0x6fa1e097 // udot v23.4s, v4.16b, v1.4b[1]\n" "mov v5.d[1], x24\n" - "ldr x23, [x10, #0x98]\n" + ".inst 0x6fa1e097 // udot v23.4s, v4.16b, v1.4b[1]\n" "ldr d7, [x10, #0xa0]\n" ".inst 0x6f80e8b0 // udot v16.4s, v5.16b, v0.4b[2]\n" - ".inst 0x6f81e8b4 // udot v20.4s, v5.16b, v1.4b[2]\n" "mov v6.d[1], x23\n" + ".inst 0x6f81e8b4 // udot v20.4s, v5.16b, v1.4b[2]\n" "ldr x19, [x10, #0xa8]\n" - "ldr d8, [x10, #0xb0]\n" ".inst 0x6f80e8d1 // udot v17.4s, v6.16b, v0.4b[2]\n" + "ldr d8, [x10, #0xb0]\n" ".inst 0x6f81e8d5 // udot v21.4s, v6.16b, v1.4b[2]\n" - "mov v7.d[1], x19\n" "ldr x23, [x10, #0xb8]\n" + "mov v7.d[1], x19\n" "ldr d9, [x10, #0xc0]\n" + "ldr x19, [x10, #0xc8]\n" ".inst 0x6f80e8f2 // udot v18.4s, v7.16b, v0.4b[2]\n" - ".inst 0x6f81e8f6 // udot v22.4s, v7.16b, v1.4b[2]\n" "mov v8.d[1], x23\n" - "ldr x19, [x10, #0xc8]\n" + ".inst 0x6f81e8f6 // udot v22.4s, v7.16b, v1.4b[2]\n" "ldr d10, [x10, #0xd0]\n" ".inst 0x6f80e913 // udot v19.4s, v8.16b, v0.4b[2]\n" - ".inst 0x6f81e917 // udot v23.4s, v8.16b, v1.4b[2]\n" "mov v9.d[1], x19\n" + ".inst 0x6f81e917 // udot v23.4s, v8.16b, v1.4b[2]\n" "ldr x23, [x10, #0xd8]\n" - "ldr d4, [x10, #0xe0]\n" ".inst 0x6fa0e930 // udot v16.4s, v9.16b, v0.4b[3]\n" + "ldr d4, [x10, #0xe0]\n" ".inst 0x6fa1e934 // udot v20.4s, v9.16b, v1.4b[3]\n" - "mov v10.d[1], x23\n" "ldr x19, [x10, #0xe8]\n" + "mov v10.d[1], x23\n" "ldr d5, [x10, #0xf0]\n" - ".inst 0x6fa0e951 // udot v17.4s, v10.16b, v0.4b[3]\n" - ".inst 0x6fa1e955 // udot v21.4s, v10.16b, v1.4b[3]\n" - "mov v4.d[1], x19\n" "ldr x24, [x10, #0xf8]\n" "add x10, x10, #0x100\n" + ".inst 0x6fa0e951 // udot v17.4s, v10.16b, v0.4b[3]\n" + "mov v4.d[1], x19\n" + ".inst 0x6fa1e955 // udot v21.4s, v10.16b, v1.4b[3]\n" ".inst 0x6fa0e892 // udot v18.4s, v4.16b, v0.4b[3]\n" - ".inst 0x6fa1e896 // udot v22.4s, v4.16b, v1.4b[3]\n" "mov v5.d[1], x24\n" + ".inst 0x6fa1e896 // udot v22.4s, v4.16b, v1.4b[3]\n" ".inst 0x6fa0e8b3 // udot v19.4s, v5.16b, v0.4b[3]\n" ".inst 0x6fa1e8b7 // udot v23.4s, v5.16b, v1.4b[3]\n" "tbnz %x[flags], #31, 38f\n" @@ -544,50 +544,50 @@ void a64_hybrid_u8qa_dot_4x16_a55 ( "bge 37b\n" "39:" // Height 2: Multiply loop: Single iteration only ".inst 0x6f80e090 // udot v16.4s, v4.16b, v0.4b[0]\n" - ".inst 0x6f81e094 // udot v20.4s, v4.16b, v1.4b[0]\n" "ldr q5, [x10, #0x10]\n" - "sub x26, x26, #0x10\n" + ".inst 0x6f81e094 // udot v20.4s, v4.16b, v1.4b[0]\n" "ldr q6, [x10, #0x20]\n" - "add x25, x25, #0x10\n" "ldr q7, [x10, #0x30]\n" - "add x22, x22, #0x10\n" + "sub x26, x26, #0x10\n" ".inst 0x6f80e0b1 // udot v17.4s, v5.16b, v0.4b[0]\n" + "ldr q8, [x10, #0x40]\n" ".inst 0x6f81e0b5 // udot v21.4s, v5.16b, v1.4b[0]\n" + "ldr q9, [x10, #0x50]\n" ".inst 0x6f80e0d2 // udot v18.4s, v6.16b, v0.4b[0]\n" + "ldr q10, [x10, #0x60]\n" ".inst 0x6f81e0d6 // udot v22.4s, v6.16b, v1.4b[0]\n" + "ldr q4, [x10, #0x70]\n" ".inst 0x6f80e0f3 // udot v19.4s, v7.16b, v0.4b[0]\n" + "ldr q5, [x10, #0x80]\n" ".inst 0x6f81e0f7 // udot v23.4s, v7.16b, v1.4b[0]\n" - "ldr q8, [x10, #0x40]\n" - "ldr q9, [x10, #0x50]\n" - "ldr q10, [x10, #0x60]\n" + "ldr q6, [x10, #0x90]\n" ".inst 0x6fa0e110 // udot v16.4s, v8.16b, v0.4b[1]\n" + "ldr q7, [x10, #0xa0]\n" ".inst 0x6fa1e114 // udot v20.4s, v8.16b, v1.4b[1]\n" + "ldr q8, [x10, #0xb0]\n" ".inst 0x6fa0e131 // udot v17.4s, v9.16b, v0.4b[1]\n" + "add x25, x25, #0x10\n" ".inst 0x6fa1e135 // udot v21.4s, v9.16b, v1.4b[1]\n" + "ldr q9, [x10, #0xc0]\n" ".inst 0x6fa0e152 // udot v18.4s, v10.16b, v0.4b[1]\n" + "add x22, x22, #0x10\n" ".inst 0x6fa1e156 // udot v22.4s, v10.16b, v1.4b[1]\n" - "ldr q4, [x10, #0x70]\n" - "ldr q5, [x10, #0x80]\n" - "ldr q6, [x10, #0x90]\n" + "ldr q10, [x10, #0xd0]\n" ".inst 0x6fa0e093 // udot v19.4s, v4.16b, v0.4b[1]\n" ".inst 0x6fa1e097 // udot v23.4s, v4.16b, v1.4b[1]\n" + "ldr q4, [x10, #0xe0]\n" ".inst 0x6f80e8b0 // udot v16.4s, v5.16b, v0.4b[2]\n" ".inst 0x6f81e8b4 // udot v20.4s, v5.16b, v1.4b[2]\n" + "ldr q5, [x10, #0xf0]\n" ".inst 0x6f80e8d1 // udot v17.4s, v6.16b, v0.4b[2]\n" + "add x10, x10, #0x100\n" ".inst 0x6f81e8d5 // udot v21.4s, v6.16b, v1.4b[2]\n" - "ldr q7, [x10, #0xa0]\n" - "ldr q8, [x10, #0xb0]\n" - "ldr q9, [x10, #0xc0]\n" ".inst 0x6f80e8f2 // udot v18.4s, v7.16b, v0.4b[2]\n" ".inst 0x6f81e8f6 // udot v22.4s, v7.16b, v1.4b[2]\n" ".inst 0x6f80e913 // udot v19.4s, v8.16b, v0.4b[2]\n" ".inst 0x6f81e917 // udot v23.4s, v8.16b, v1.4b[2]\n" ".inst 0x6fa0e930 // udot v16.4s, v9.16b, v0.4b[3]\n" ".inst 0x6fa1e934 // udot v20.4s, v9.16b, v1.4b[3]\n" - "ldr q10, [x10, #0xd0]\n" - "ldr q4, [x10, #0xe0]\n" - "ldr q5, [x10, #0xf0]\n" - "add x10, x10, #0x100\n" ".inst 0x6fa0e951 // udot v17.4s, v10.16b, v0.4b[3]\n" ".inst 0x6fa1e955 // udot v21.4s, v10.16b, v1.4b[3]\n" ".inst 0x6fa0e892 // udot v18.4s, v4.16b, v0.4b[3]\n" @@ -617,13 +617,13 @@ void a64_hybrid_u8qa_dot_4x16_a55 ( "cmp x26, #0x4\n" "ldr q8, [x10, #0x20]\n" ".inst 0x6f80e0d0 // udot v16.4s, v6.16b, v0.4b[0]\n" + "ldr q9, [x10, #0x30]\n" ".inst 0x6f81e0d4 // udot v20.4s, v6.16b, v1.4b[0]\n" + "add x10, x10, #0x40\n" ".inst 0x6f80e0f1 // udot v17.4s, v7.16b, v0.4b[0]\n" ".inst 0x6f81e0f5 // udot v21.4s, v7.16b, v1.4b[0]\n" ".inst 0x6f80e112 // udot v18.4s, v8.16b, v0.4b[0]\n" ".inst 0x6f81e116 // udot v22.4s, v8.16b, v1.4b[0]\n" - "ldr q9, [x10, #0x30]\n" - "add x10, x10, #0x40\n" ".inst 0x6f80e133 // udot v19.4s, v9.16b, v0.4b[0]\n" ".inst 0x6f81e137 // udot v23.4s, v9.16b, v1.4b[0]\n" "bge 42b\n" @@ -648,13 +648,13 @@ void a64_hybrid_u8qa_dot_4x16_a55 ( "ldr q4, [x10, #0x10]\n" "ldr q5, [x10, #0x20]\n" ".inst 0x6f80e150 // udot v16.4s, v10.16b, v0.4b[0]\n" + "ldr q6, [x10, #0x30]\n" ".inst 0x6f81e154 // udot v20.4s, v10.16b, v1.4b[0]\n" + "add x10, x10, #0x40\n" ".inst 0x6f80e091 // udot v17.4s, v4.16b, v0.4b[0]\n" ".inst 0x6f81e095 // udot v21.4s, v4.16b, v1.4b[0]\n" ".inst 0x6f80e0b2 // udot v18.4s, v5.16b, v0.4b[0]\n" ".inst 0x6f81e0b6 // udot v22.4s, v5.16b, v1.4b[0]\n" - "ldr q6, [x10, #0x30]\n" - "add x10, x10, #0x40\n" ".inst 0x6f80e0d3 // udot v19.4s, v6.16b, v0.4b[0]\n" ".inst 0x6f81e0d7 // udot v23.4s, v6.16b, v1.4b[0]\n" "48:" // Height 2: Multiply loop: No odd multiplies @@ -907,97 +907,97 @@ void a64_hybrid_u8qa_dot_4x16_a55 ( "blt 69f\n" "67:" // Height 3: Multiply loop: Main loop head ".inst 0x6f80e090 // udot v16.4s, v4.16b, v0.4b[0]\n" - ".inst 0x6f81e094 // udot v20.4s, v4.16b, v1.4b[0]\n" - ".inst 0x6f82e098 // udot v24.4s, v4.16b, v2.4b[0]\n" "ldr d5, [x10, #0x10]\n" + ".inst 0x6f81e094 // udot v20.4s, v4.16b, v1.4b[0]\n" "ldr x24, [x10, #0x18]\n" - "add x25, x25, #0x10\n" + ".inst 0x6f82e098 // udot v24.4s, v4.16b, v2.4b[0]\n" "ldr d6, [x10, #0x20]\n" - "add x22, x22, #0x10\n" "ldr x23, [x10, #0x28]\n" - "add x21, x21, #0x10\n" + "add x25, x25, #0x10\n" "mov v5.d[1], x24\n" "ldr d7, [x10, #0x30]\n" "ldr x19, [x10, #0x38]\n" - "mov v6.d[1], x23\n" + "add x22, x22, #0x10\n" ".inst 0x6f80e0b1 // udot v17.4s, v5.16b, v0.4b[0]\n" + "mov v6.d[1], x23\n" ".inst 0x6f81e0b5 // udot v21.4s, v5.16b, v1.4b[0]\n" + "ldr d8, [x10, #0x40]\n" ".inst 0x6f82e0b9 // udot v25.4s, v5.16b, v2.4b[0]\n" "mov v7.d[1], x19\n" ".inst 0x6f80e0d2 // udot v18.4s, v6.16b, v0.4b[0]\n" - ".inst 0x6f81e0d6 // udot v22.4s, v6.16b, v1.4b[0]\n" - ".inst 0x6f82e0da // udot v26.4s, v6.16b, v2.4b[0]\n" - ".inst 0x6f80e0f3 // udot v19.4s, v7.16b, v0.4b[0]\n" - ".inst 0x6f81e0f7 // udot v23.4s, v7.16b, v1.4b[0]\n" - ".inst 0x6f82e0fb // udot v27.4s, v7.16b, v2.4b[0]\n" - "ldr d8, [x10, #0x40]\n" "ldr x23, [x10, #0x48]\n" + ".inst 0x6f81e0d6 // udot v22.4s, v6.16b, v1.4b[0]\n" "ldr d9, [x10, #0x50]\n" + ".inst 0x6f82e0da // udot v26.4s, v6.16b, v2.4b[0]\n" "ldr x19, [x10, #0x58]\n" + ".inst 0x6f80e0f3 // udot v19.4s, v7.16b, v0.4b[0]\n" "mov v8.d[1], x23\n" + ".inst 0x6f81e0f7 // udot v23.4s, v7.16b, v1.4b[0]\n" "ldr d10, [x10, #0x60]\n" - "ldr x23, [x10, #0x68]\n" + ".inst 0x6f82e0fb // udot v27.4s, v7.16b, v2.4b[0]\n" "mov v9.d[1], x19\n" ".inst 0x6fa0e110 // udot v16.4s, v8.16b, v0.4b[1]\n" + "ldr x23, [x10, #0x68]\n" ".inst 0x6fa1e114 // udot v20.4s, v8.16b, v1.4b[1]\n" + "ldr d4, [x10, #0x70]\n" ".inst 0x6fa2e118 // udot v24.4s, v8.16b, v2.4b[1]\n" - "mov v10.d[1], x23\n" + "ldr x19, [x10, #0x78]\n" ".inst 0x6fa0e131 // udot v17.4s, v9.16b, v0.4b[1]\n" + "mov v10.d[1], x23\n" ".inst 0x6fa1e135 // udot v21.4s, v9.16b, v1.4b[1]\n" + "ldr d5, [x10, #0x80]\n" ".inst 0x6fa2e139 // udot v25.4s, v9.16b, v2.4b[1]\n" + "mov v4.d[1], x19\n" ".inst 0x6fa0e152 // udot v18.4s, v10.16b, v0.4b[1]\n" - ".inst 0x6fa1e156 // udot v22.4s, v10.16b, v1.4b[1]\n" - ".inst 0x6fa2e15a // udot v26.4s, v10.16b, v2.4b[1]\n" - "ldr d4, [x10, #0x70]\n" - "ldr x19, [x10, #0x78]\n" - "ldr d5, [x10, #0x80]\n" "ldr x24, [x10, #0x88]\n" - "mov v4.d[1], x19\n" + ".inst 0x6fa1e156 // udot v22.4s, v10.16b, v1.4b[1]\n" "ldr d6, [x10, #0x90]\n" + ".inst 0x6fa2e15a // udot v26.4s, v10.16b, v2.4b[1]\n" "ldr x23, [x10, #0x98]\n" - "mov v5.d[1], x24\n" ".inst 0x6fa0e093 // udot v19.4s, v4.16b, v0.4b[1]\n" + "mov v5.d[1], x24\n" ".inst 0x6fa1e097 // udot v23.4s, v4.16b, v1.4b[1]\n" + "ldr d7, [x10, #0xa0]\n" ".inst 0x6fa2e09b // udot v27.4s, v4.16b, v2.4b[1]\n" "mov v6.d[1], x23\n" ".inst 0x6f80e8b0 // udot v16.4s, v5.16b, v0.4b[2]\n" - ".inst 0x6f81e8b4 // udot v20.4s, v5.16b, v1.4b[2]\n" - ".inst 0x6f82e8b8 // udot v24.4s, v5.16b, v2.4b[2]\n" - ".inst 0x6f80e8d1 // udot v17.4s, v6.16b, v0.4b[2]\n" - ".inst 0x6f81e8d5 // udot v21.4s, v6.16b, v1.4b[2]\n" - ".inst 0x6f82e8d9 // udot v25.4s, v6.16b, v2.4b[2]\n" - "ldr d7, [x10, #0xa0]\n" "ldr x19, [x10, #0xa8]\n" + ".inst 0x6f81e8b4 // udot v20.4s, v5.16b, v1.4b[2]\n" "ldr d8, [x10, #0xb0]\n" + ".inst 0x6f82e8b8 // udot v24.4s, v5.16b, v2.4b[2]\n" "ldr x23, [x10, #0xb8]\n" + ".inst 0x6f80e8d1 // udot v17.4s, v6.16b, v0.4b[2]\n" "mov v7.d[1], x19\n" + ".inst 0x6f81e8d5 // udot v21.4s, v6.16b, v1.4b[2]\n" "ldr d9, [x10, #0xc0]\n" - "ldr x19, [x10, #0xc8]\n" + ".inst 0x6f82e8d9 // udot v25.4s, v6.16b, v2.4b[2]\n" "mov v8.d[1], x23\n" ".inst 0x6f80e8f2 // udot v18.4s, v7.16b, v0.4b[2]\n" + "ldr x19, [x10, #0xc8]\n" ".inst 0x6f81e8f6 // udot v22.4s, v7.16b, v1.4b[2]\n" + "ldr d10, [x10, #0xd0]\n" ".inst 0x6f82e8fa // udot v26.4s, v7.16b, v2.4b[2]\n" - "mov v9.d[1], x19\n" + "ldr x23, [x10, #0xd8]\n" ".inst 0x6f80e913 // udot v19.4s, v8.16b, v0.4b[2]\n" + "mov v9.d[1], x19\n" ".inst 0x6f81e917 // udot v23.4s, v8.16b, v1.4b[2]\n" + "ldr d4, [x10, #0xe0]\n" ".inst 0x6f82e91b // udot v27.4s, v8.16b, v2.4b[2]\n" + "mov v10.d[1], x23\n" ".inst 0x6fa0e930 // udot v16.4s, v9.16b, v0.4b[3]\n" - ".inst 0x6fa1e934 // udot v20.4s, v9.16b, v1.4b[3]\n" - ".inst 0x6fa2e938 // udot v24.4s, v9.16b, v2.4b[3]\n" - "ldr d10, [x10, #0xd0]\n" - "ldr x23, [x10, #0xd8]\n" - "ldr d4, [x10, #0xe0]\n" "ldr x19, [x10, #0xe8]\n" - "mov v10.d[1], x23\n" + ".inst 0x6fa1e934 // udot v20.4s, v9.16b, v1.4b[3]\n" "ldr d5, [x10, #0xf0]\n" + ".inst 0x6fa2e938 // udot v24.4s, v9.16b, v2.4b[3]\n" "ldr x24, [x10, #0xf8]\n" - "add x10, x10, #0x100\n" - "mov v4.d[1], x19\n" ".inst 0x6fa0e951 // udot v17.4s, v10.16b, v0.4b[3]\n" + "mov v4.d[1], x19\n" ".inst 0x6fa1e955 // udot v21.4s, v10.16b, v1.4b[3]\n" + "add x21, x21, #0x10\n" ".inst 0x6fa2e959 // udot v25.4s, v10.16b, v2.4b[3]\n" "mov v5.d[1], x24\n" ".inst 0x6fa0e892 // udot v18.4s, v4.16b, v0.4b[3]\n" + "add x10, x10, #0x100\n" ".inst 0x6fa1e896 // udot v22.4s, v4.16b, v1.4b[3]\n" ".inst 0x6fa2e89a // udot v26.4s, v4.16b, v2.4b[3]\n" ".inst 0x6fa0e8b3 // udot v19.4s, v5.16b, v0.4b[3]\n" @@ -1020,51 +1020,55 @@ void a64_hybrid_u8qa_dot_4x16_a55 ( "bge 67b\n" "69:" // Height 3: Multiply loop: Single iteration only ".inst 0x6f80e090 // udot v16.4s, v4.16b, v0.4b[0]\n" - ".inst 0x6f81e094 // udot v20.4s, v4.16b, v1.4b[0]\n" - ".inst 0x6f82e098 // udot v24.4s, v4.16b, v2.4b[0]\n" "ldr q5, [x10, #0x10]\n" + ".inst 0x6f81e094 // udot v20.4s, v4.16b, v1.4b[0]\n" "ldr q6, [x10, #0x20]\n" - "sub x26, x26, #0x10\n" + ".inst 0x6f82e098 // udot v24.4s, v4.16b, v2.4b[0]\n" "ldr q7, [x10, #0x30]\n" - "add x25, x25, #0x10\n" ".inst 0x6f80e0b1 // udot v17.4s, v5.16b, v0.4b[0]\n" + "ldr q8, [x10, #0x40]\n" ".inst 0x6f81e0b5 // udot v21.4s, v5.16b, v1.4b[0]\n" + "ldr q9, [x10, #0x50]\n" ".inst 0x6f82e0b9 // udot v25.4s, v5.16b, v2.4b[0]\n" + "ldr q10, [x10, #0x60]\n" ".inst 0x6f80e0d2 // udot v18.4s, v6.16b, v0.4b[0]\n" + "ldr q4, [x10, #0x70]\n" ".inst 0x6f81e0d6 // udot v22.4s, v6.16b, v1.4b[0]\n" + "ldr q5, [x10, #0x80]\n" ".inst 0x6f82e0da // udot v26.4s, v6.16b, v2.4b[0]\n" + "ldr q6, [x10, #0x90]\n" ".inst 0x6f80e0f3 // udot v19.4s, v7.16b, v0.4b[0]\n" + "sub x26, x26, #0x10\n" ".inst 0x6f81e0f7 // udot v23.4s, v7.16b, v1.4b[0]\n" + "add x25, x25, #0x10\n" ".inst 0x6f82e0fb // udot v27.4s, v7.16b, v2.4b[0]\n" - "ldr q8, [x10, #0x40]\n" - "ldr q9, [x10, #0x50]\n" - "add x22, x22, #0x10\n" - "ldr q10, [x10, #0x60]\n" - "add x21, x21, #0x10\n" + "ldr q7, [x10, #0xa0]\n" ".inst 0x6fa0e110 // udot v16.4s, v8.16b, v0.4b[1]\n" + "add x22, x22, #0x10\n" ".inst 0x6fa1e114 // udot v20.4s, v8.16b, v1.4b[1]\n" + "add x21, x21, #0x10\n" ".inst 0x6fa2e118 // udot v24.4s, v8.16b, v2.4b[1]\n" + "ldr q8, [x10, #0xb0]\n" ".inst 0x6fa0e131 // udot v17.4s, v9.16b, v0.4b[1]\n" ".inst 0x6fa1e135 // udot v21.4s, v9.16b, v1.4b[1]\n" ".inst 0x6fa2e139 // udot v25.4s, v9.16b, v2.4b[1]\n" + "ldr q9, [x10, #0xc0]\n" ".inst 0x6fa0e152 // udot v18.4s, v10.16b, v0.4b[1]\n" ".inst 0x6fa1e156 // udot v22.4s, v10.16b, v1.4b[1]\n" ".inst 0x6fa2e15a // udot v26.4s, v10.16b, v2.4b[1]\n" - "ldr q4, [x10, #0x70]\n" - "ldr q5, [x10, #0x80]\n" - "ldr q6, [x10, #0x90]\n" + "ldr q10, [x10, #0xd0]\n" ".inst 0x6fa0e093 // udot v19.4s, v4.16b, v0.4b[1]\n" ".inst 0x6fa1e097 // udot v23.4s, v4.16b, v1.4b[1]\n" ".inst 0x6fa2e09b // udot v27.4s, v4.16b, v2.4b[1]\n" + "ldr q4, [x10, #0xe0]\n" ".inst 0x6f80e8b0 // udot v16.4s, v5.16b, v0.4b[2]\n" ".inst 0x6f81e8b4 // udot v20.4s, v5.16b, v1.4b[2]\n" ".inst 0x6f82e8b8 // udot v24.4s, v5.16b, v2.4b[2]\n" + "ldr q5, [x10, #0xf0]\n" ".inst 0x6f80e8d1 // udot v17.4s, v6.16b, v0.4b[2]\n" + "add x10, x10, #0x100\n" ".inst 0x6f81e8d5 // udot v21.4s, v6.16b, v1.4b[2]\n" ".inst 0x6f82e8d9 // udot v25.4s, v6.16b, v2.4b[2]\n" - "ldr q7, [x10, #0xa0]\n" - "ldr q8, [x10, #0xb0]\n" - "ldr q9, [x10, #0xc0]\n" ".inst 0x6f80e8f2 // udot v18.4s, v7.16b, v0.4b[2]\n" ".inst 0x6f81e8f6 // udot v22.4s, v7.16b, v1.4b[2]\n" ".inst 0x6f82e8fa // udot v26.4s, v7.16b, v2.4b[2]\n" @@ -1074,10 +1078,6 @@ void a64_hybrid_u8qa_dot_4x16_a55 ( ".inst 0x6fa0e930 // udot v16.4s, v9.16b, v0.4b[3]\n" ".inst 0x6fa1e934 // udot v20.4s, v9.16b, v1.4b[3]\n" ".inst 0x6fa2e938 // udot v24.4s, v9.16b, v2.4b[3]\n" - "ldr q10, [x10, #0xd0]\n" - "ldr q4, [x10, #0xe0]\n" - "ldr q5, [x10, #0xf0]\n" - "add x10, x10, #0x100\n" ".inst 0x6fa0e951 // udot v17.4s, v10.16b, v0.4b[3]\n" ".inst 0x6fa1e955 // udot v21.4s, v10.16b, v1.4b[3]\n" ".inst 0x6fa2e959 // udot v25.4s, v10.16b, v2.4b[3]\n" @@ -1114,7 +1114,9 @@ void a64_hybrid_u8qa_dot_4x16_a55 ( "cmp x26, #0x4\n" "ldr q8, [x10, #0x20]\n" ".inst 0x6f80e0d0 // udot v16.4s, v6.16b, v0.4b[0]\n" + "ldr q9, [x10, #0x30]\n" ".inst 0x6f81e0d4 // udot v20.4s, v6.16b, v1.4b[0]\n" + "add x10, x10, #0x40\n" ".inst 0x6f82e0d8 // udot v24.4s, v6.16b, v2.4b[0]\n" ".inst 0x6f80e0f1 // udot v17.4s, v7.16b, v0.4b[0]\n" ".inst 0x6f81e0f5 // udot v21.4s, v7.16b, v1.4b[0]\n" @@ -1122,8 +1124,6 @@ void a64_hybrid_u8qa_dot_4x16_a55 ( ".inst 0x6f80e112 // udot v18.4s, v8.16b, v0.4b[0]\n" ".inst 0x6f81e116 // udot v22.4s, v8.16b, v1.4b[0]\n" ".inst 0x6f82e11a // udot v26.4s, v8.16b, v2.4b[0]\n" - "ldr q9, [x10, #0x30]\n" - "add x10, x10, #0x40\n" ".inst 0x6f80e133 // udot v19.4s, v9.16b, v0.4b[0]\n" ".inst 0x6f81e137 // udot v23.4s, v9.16b, v1.4b[0]\n" ".inst 0x6f82e13b // udot v27.4s, v9.16b, v2.4b[0]\n" @@ -1153,7 +1153,9 @@ void a64_hybrid_u8qa_dot_4x16_a55 ( "ldr q4, [x10, #0x10]\n" "ldr q5, [x10, #0x20]\n" ".inst 0x6f80e150 // udot v16.4s, v10.16b, v0.4b[0]\n" + "ldr q6, [x10, #0x30]\n" ".inst 0x6f81e154 // udot v20.4s, v10.16b, v1.4b[0]\n" + "add x10, x10, #0x40\n" ".inst 0x6f82e158 // udot v24.4s, v10.16b, v2.4b[0]\n" ".inst 0x6f80e091 // udot v17.4s, v4.16b, v0.4b[0]\n" ".inst 0x6f81e095 // udot v21.4s, v4.16b, v1.4b[0]\n" @@ -1161,8 +1163,6 @@ void a64_hybrid_u8qa_dot_4x16_a55 ( ".inst 0x6f80e0b2 // udot v18.4s, v5.16b, v0.4b[0]\n" ".inst 0x6f81e0b6 // udot v22.4s, v5.16b, v1.4b[0]\n" ".inst 0x6f82e0ba // udot v26.4s, v5.16b, v2.4b[0]\n" - "ldr q6, [x10, #0x30]\n" - "add x10, x10, #0x40\n" ".inst 0x6f80e0d3 // udot v19.4s, v6.16b, v0.4b[0]\n" ".inst 0x6f81e0d7 // udot v23.4s, v6.16b, v1.4b[0]\n" ".inst 0x6f82e0db // udot v27.4s, v6.16b, v2.4b[0]\n" @@ -1492,110 +1492,110 @@ void a64_hybrid_u8qa_dot_4x16_a55 ( "blt 99f\n" "97:" // Height 4: Multiply loop: Main loop head ".inst 0x6f80e090 // udot v16.4s, v4.16b, v0.4b[0]\n" - ".inst 0x6f81e094 // udot v20.4s, v4.16b, v1.4b[0]\n" - ".inst 0x6f82e098 // udot v24.4s, v4.16b, v2.4b[0]\n" - ".inst 0x6f83e09c // udot v28.4s, v4.16b, v3.4b[0]\n" "ldr d5, [x10, #0x10]\n" - "add x25, x25, #0x10\n" + ".inst 0x6f81e094 // udot v20.4s, v4.16b, v1.4b[0]\n" "ldr x24, [x10, #0x18]\n" - "add x22, x22, #0x10\n" + ".inst 0x6f82e098 // udot v24.4s, v4.16b, v2.4b[0]\n" "ldr d6, [x10, #0x20]\n" - "add x21, x21, #0x10\n" + ".inst 0x6f83e09c // udot v28.4s, v4.16b, v3.4b[0]\n" "ldr x23, [x10, #0x28]\n" - "add x20, x20, #0x10\n" "mov v5.d[1], x24\n" "ldr d7, [x10, #0x30]\n" "ldr x19, [x10, #0x38]\n" - "mov v6.d[1], x23\n" + "add x25, x25, #0x10\n" ".inst 0x6f80e0b1 // udot v17.4s, v5.16b, v0.4b[0]\n" + "mov v6.d[1], x23\n" ".inst 0x6f81e0b5 // udot v21.4s, v5.16b, v1.4b[0]\n" + "ldr d8, [x10, #0x40]\n" ".inst 0x6f82e0b9 // udot v25.4s, v5.16b, v2.4b[0]\n" "mov v7.d[1], x19\n" ".inst 0x6f83e0bd // udot v29.4s, v5.16b, v3.4b[0]\n" + "ldr x23, [x10, #0x48]\n" ".inst 0x6f80e0d2 // udot v18.4s, v6.16b, v0.4b[0]\n" + "ldr d9, [x10, #0x50]\n" ".inst 0x6f81e0d6 // udot v22.4s, v6.16b, v1.4b[0]\n" + "ldr x19, [x10, #0x58]\n" ".inst 0x6f82e0da // udot v26.4s, v6.16b, v2.4b[0]\n" + "mov v8.d[1], x23\n" ".inst 0x6f83e0de // udot v30.4s, v6.16b, v3.4b[0]\n" + "ldr d10, [x10, #0x60]\n" ".inst 0x6f80e0f3 // udot v19.4s, v7.16b, v0.4b[0]\n" + "mov v9.d[1], x19\n" ".inst 0x6f81e0f7 // udot v23.4s, v7.16b, v1.4b[0]\n" + "ldr x23, [x10, #0x68]\n" ".inst 0x6f82e0fb // udot v27.4s, v7.16b, v2.4b[0]\n" + "ldr d4, [x10, #0x70]\n" ".inst 0x6f83e0ff // udot v31.4s, v7.16b, v3.4b[0]\n" - "ldr d8, [x10, #0x40]\n" - "ldr x23, [x10, #0x48]\n" - "ldr d9, [x10, #0x50]\n" - "ldr x19, [x10, #0x58]\n" - "mov v8.d[1], x23\n" - "ldr d10, [x10, #0x60]\n" - "ldr x23, [x10, #0x68]\n" - "mov v9.d[1], x19\n" + "ldr x19, [x10, #0x78]\n" ".inst 0x6fa0e110 // udot v16.4s, v8.16b, v0.4b[1]\n" + "mov v10.d[1], x23\n" ".inst 0x6fa1e114 // udot v20.4s, v8.16b, v1.4b[1]\n" + "ldr d5, [x10, #0x80]\n" ".inst 0x6fa2e118 // udot v24.4s, v8.16b, v2.4b[1]\n" - "mov v10.d[1], x23\n" + "mov v4.d[1], x19\n" ".inst 0x6fa3e11c // udot v28.4s, v8.16b, v3.4b[1]\n" + "ldr x24, [x10, #0x88]\n" ".inst 0x6fa0e131 // udot v17.4s, v9.16b, v0.4b[1]\n" + "ldr d6, [x10, #0x90]\n" ".inst 0x6fa1e135 // udot v21.4s, v9.16b, v1.4b[1]\n" + "ldr x23, [x10, #0x98]\n" ".inst 0x6fa2e139 // udot v25.4s, v9.16b, v2.4b[1]\n" + "mov v5.d[1], x24\n" ".inst 0x6fa3e13d // udot v29.4s, v9.16b, v3.4b[1]\n" + "ldr d7, [x10, #0xa0]\n" ".inst 0x6fa0e152 // udot v18.4s, v10.16b, v0.4b[1]\n" + "mov v6.d[1], x23\n" ".inst 0x6fa1e156 // udot v22.4s, v10.16b, v1.4b[1]\n" + "ldr x19, [x10, #0xa8]\n" ".inst 0x6fa2e15a // udot v26.4s, v10.16b, v2.4b[1]\n" + "ldr d8, [x10, #0xb0]\n" ".inst 0x6fa3e15e // udot v30.4s, v10.16b, v3.4b[1]\n" - "ldr d4, [x10, #0x70]\n" - "ldr x19, [x10, #0x78]\n" - "ldr d5, [x10, #0x80]\n" - "ldr x24, [x10, #0x88]\n" - "mov v4.d[1], x19\n" - "ldr d6, [x10, #0x90]\n" - "ldr x23, [x10, #0x98]\n" - "mov v5.d[1], x24\n" + "ldr x23, [x10, #0xb8]\n" ".inst 0x6fa0e093 // udot v19.4s, v4.16b, v0.4b[1]\n" + "mov v7.d[1], x19\n" ".inst 0x6fa1e097 // udot v23.4s, v4.16b, v1.4b[1]\n" + "ldr d9, [x10, #0xc0]\n" ".inst 0x6fa2e09b // udot v27.4s, v4.16b, v2.4b[1]\n" - "mov v6.d[1], x23\n" + "mov v8.d[1], x23\n" ".inst 0x6fa3e09f // udot v31.4s, v4.16b, v3.4b[1]\n" + "ldr x19, [x10, #0xc8]\n" ".inst 0x6f80e8b0 // udot v16.4s, v5.16b, v0.4b[2]\n" + "ldr d10, [x10, #0xd0]\n" ".inst 0x6f81e8b4 // udot v20.4s, v5.16b, v1.4b[2]\n" + "ldr x23, [x10, #0xd8]\n" ".inst 0x6f82e8b8 // udot v24.4s, v5.16b, v2.4b[2]\n" + "mov v9.d[1], x19\n" ".inst 0x6f83e8bc // udot v28.4s, v5.16b, v3.4b[2]\n" + "ldr d4, [x10, #0xe0]\n" ".inst 0x6f80e8d1 // udot v17.4s, v6.16b, v0.4b[2]\n" + "mov v10.d[1], x23\n" ".inst 0x6f81e8d5 // udot v21.4s, v6.16b, v1.4b[2]\n" + "ldr x19, [x10, #0xe8]\n" ".inst 0x6f82e8d9 // udot v25.4s, v6.16b, v2.4b[2]\n" + "ldr d5, [x10, #0xf0]\n" ".inst 0x6f83e8dd // udot v29.4s, v6.16b, v3.4b[2]\n" - "ldr d7, [x10, #0xa0]\n" - "ldr x19, [x10, #0xa8]\n" - "ldr d8, [x10, #0xb0]\n" - "ldr x23, [x10, #0xb8]\n" - "mov v7.d[1], x19\n" - "ldr d9, [x10, #0xc0]\n" - "ldr x19, [x10, #0xc8]\n" - "mov v8.d[1], x23\n" + "ldr x24, [x10, #0xf8]\n" ".inst 0x6f80e8f2 // udot v18.4s, v7.16b, v0.4b[2]\n" + "mov v4.d[1], x19\n" ".inst 0x6f81e8f6 // udot v22.4s, v7.16b, v1.4b[2]\n" + "add x22, x22, #0x10\n" ".inst 0x6f82e8fa // udot v26.4s, v7.16b, v2.4b[2]\n" - "mov v9.d[1], x19\n" + "mov v5.d[1], x24\n" ".inst 0x6f83e8fe // udot v30.4s, v7.16b, v3.4b[2]\n" + "add x21, x21, #0x10\n" ".inst 0x6f80e913 // udot v19.4s, v8.16b, v0.4b[2]\n" + "add x20, x20, #0x10\n" ".inst 0x6f81e917 // udot v23.4s, v8.16b, v1.4b[2]\n" + "add x10, x10, #0x100\n" ".inst 0x6f82e91b // udot v27.4s, v8.16b, v2.4b[2]\n" ".inst 0x6f83e91f // udot v31.4s, v8.16b, v3.4b[2]\n" ".inst 0x6fa0e930 // udot v16.4s, v9.16b, v0.4b[3]\n" ".inst 0x6fa1e934 // udot v20.4s, v9.16b, v1.4b[3]\n" ".inst 0x6fa2e938 // udot v24.4s, v9.16b, v2.4b[3]\n" ".inst 0x6fa3e93c // udot v28.4s, v9.16b, v3.4b[3]\n" - "ldr d10, [x10, #0xd0]\n" - "ldr x23, [x10, #0xd8]\n" - "ldr d4, [x10, #0xe0]\n" - "ldr x19, [x10, #0xe8]\n" - "mov v10.d[1], x23\n" - "ldr d5, [x10, #0xf0]\n" - "ldr x24, [x10, #0xf8]\n" - "add x10, x10, #0x100\n" - "mov v4.d[1], x19\n" ".inst 0x6fa0e951 // udot v17.4s, v10.16b, v0.4b[3]\n" ".inst 0x6fa1e955 // udot v21.4s, v10.16b, v1.4b[3]\n" ".inst 0x6fa2e959 // udot v25.4s, v10.16b, v2.4b[3]\n" - "mov v5.d[1], x24\n" ".inst 0x6fa3e95d // udot v29.4s, v10.16b, v3.4b[3]\n" ".inst 0x6fa0e892 // udot v18.4s, v4.16b, v0.4b[3]\n" ".inst 0x6fa1e896 // udot v22.4s, v4.16b, v1.4b[3]\n" @@ -1625,62 +1625,66 @@ void a64_hybrid_u8qa_dot_4x16_a55 ( "bge 97b\n" "99:" // Height 4: Multiply loop: Single iteration only ".inst 0x6f80e090 // udot v16.4s, v4.16b, v0.4b[0]\n" - ".inst 0x6f81e094 // udot v20.4s, v4.16b, v1.4b[0]\n" - ".inst 0x6f82e098 // udot v24.4s, v4.16b, v2.4b[0]\n" - ".inst 0x6f83e09c // udot v28.4s, v4.16b, v3.4b[0]\n" "ldr q5, [x10, #0x10]\n" - "sub x26, x26, #0x10\n" + ".inst 0x6f81e094 // udot v20.4s, v4.16b, v1.4b[0]\n" "ldr q6, [x10, #0x20]\n" - "add x25, x25, #0x10\n" + ".inst 0x6f82e098 // udot v24.4s, v4.16b, v2.4b[0]\n" "ldr q7, [x10, #0x30]\n" - "add x22, x22, #0x10\n" + ".inst 0x6f83e09c // udot v28.4s, v4.16b, v3.4b[0]\n" + "ldr q8, [x10, #0x40]\n" ".inst 0x6f80e0b1 // udot v17.4s, v5.16b, v0.4b[0]\n" + "ldr q9, [x10, #0x50]\n" ".inst 0x6f81e0b5 // udot v21.4s, v5.16b, v1.4b[0]\n" + "ldr q10, [x10, #0x60]\n" ".inst 0x6f82e0b9 // udot v25.4s, v5.16b, v2.4b[0]\n" + "ldr q4, [x10, #0x70]\n" ".inst 0x6f83e0bd // udot v29.4s, v5.16b, v3.4b[0]\n" + "ldr q5, [x10, #0x80]\n" ".inst 0x6f80e0d2 // udot v18.4s, v6.16b, v0.4b[0]\n" + "sub x26, x26, #0x10\n" ".inst 0x6f81e0d6 // udot v22.4s, v6.16b, v1.4b[0]\n" + "add x25, x25, #0x10\n" ".inst 0x6f82e0da // udot v26.4s, v6.16b, v2.4b[0]\n" + "add x22, x22, #0x10\n" ".inst 0x6f83e0de // udot v30.4s, v6.16b, v3.4b[0]\n" + "ldr q6, [x10, #0x90]\n" ".inst 0x6f80e0f3 // udot v19.4s, v7.16b, v0.4b[0]\n" + "add x21, x21, #0x10\n" ".inst 0x6f81e0f7 // udot v23.4s, v7.16b, v1.4b[0]\n" + "add x20, x20, #0x10\n" ".inst 0x6f82e0fb // udot v27.4s, v7.16b, v2.4b[0]\n" ".inst 0x6f83e0ff // udot v31.4s, v7.16b, v3.4b[0]\n" - "ldr q8, [x10, #0x40]\n" - "add x21, x21, #0x10\n" - "ldr q9, [x10, #0x50]\n" - "add x20, x20, #0x10\n" - "ldr q10, [x10, #0x60]\n" + "ldr q7, [x10, #0xa0]\n" ".inst 0x6fa0e110 // udot v16.4s, v8.16b, v0.4b[1]\n" ".inst 0x6fa1e114 // udot v20.4s, v8.16b, v1.4b[1]\n" ".inst 0x6fa2e118 // udot v24.4s, v8.16b, v2.4b[1]\n" ".inst 0x6fa3e11c // udot v28.4s, v8.16b, v3.4b[1]\n" + "ldr q8, [x10, #0xb0]\n" ".inst 0x6fa0e131 // udot v17.4s, v9.16b, v0.4b[1]\n" ".inst 0x6fa1e135 // udot v21.4s, v9.16b, v1.4b[1]\n" ".inst 0x6fa2e139 // udot v25.4s, v9.16b, v2.4b[1]\n" ".inst 0x6fa3e13d // udot v29.4s, v9.16b, v3.4b[1]\n" + "ldr q9, [x10, #0xc0]\n" ".inst 0x6fa0e152 // udot v18.4s, v10.16b, v0.4b[1]\n" ".inst 0x6fa1e156 // udot v22.4s, v10.16b, v1.4b[1]\n" ".inst 0x6fa2e15a // udot v26.4s, v10.16b, v2.4b[1]\n" ".inst 0x6fa3e15e // udot v30.4s, v10.16b, v3.4b[1]\n" - "ldr q4, [x10, #0x70]\n" - "ldr q5, [x10, #0x80]\n" - "ldr q6, [x10, #0x90]\n" + "ldr q10, [x10, #0xd0]\n" ".inst 0x6fa0e093 // udot v19.4s, v4.16b, v0.4b[1]\n" ".inst 0x6fa1e097 // udot v23.4s, v4.16b, v1.4b[1]\n" ".inst 0x6fa2e09b // udot v27.4s, v4.16b, v2.4b[1]\n" ".inst 0x6fa3e09f // udot v31.4s, v4.16b, v3.4b[1]\n" + "ldr q4, [x10, #0xe0]\n" ".inst 0x6f80e8b0 // udot v16.4s, v5.16b, v0.4b[2]\n" ".inst 0x6f81e8b4 // udot v20.4s, v5.16b, v1.4b[2]\n" ".inst 0x6f82e8b8 // udot v24.4s, v5.16b, v2.4b[2]\n" ".inst 0x6f83e8bc // udot v28.4s, v5.16b, v3.4b[2]\n" + "ldr q5, [x10, #0xf0]\n" ".inst 0x6f80e8d1 // udot v17.4s, v6.16b, v0.4b[2]\n" + "add x10, x10, #0x100\n" ".inst 0x6f81e8d5 // udot v21.4s, v6.16b, v1.4b[2]\n" ".inst 0x6f82e8d9 // udot v25.4s, v6.16b, v2.4b[2]\n" ".inst 0x6f83e8dd // udot v29.4s, v6.16b, v3.4b[2]\n" - "ldr q7, [x10, #0xa0]\n" - "ldr q8, [x10, #0xb0]\n" - "ldr q9, [x10, #0xc0]\n" ".inst 0x6f80e8f2 // udot v18.4s, v7.16b, v0.4b[2]\n" ".inst 0x6f81e8f6 // udot v22.4s, v7.16b, v1.4b[2]\n" ".inst 0x6f82e8fa // udot v26.4s, v7.16b, v2.4b[2]\n" @@ -1693,10 +1697,6 @@ void a64_hybrid_u8qa_dot_4x16_a55 ( ".inst 0x6fa1e934 // udot v20.4s, v9.16b, v1.4b[3]\n" ".inst 0x6fa2e938 // udot v24.4s, v9.16b, v2.4b[3]\n" ".inst 0x6fa3e93c // udot v28.4s, v9.16b, v3.4b[3]\n" - "ldr q10, [x10, #0xd0]\n" - "ldr q4, [x10, #0xe0]\n" - "ldr q5, [x10, #0xf0]\n" - "add x10, x10, #0x100\n" ".inst 0x6fa0e951 // udot v17.4s, v10.16b, v0.4b[3]\n" ".inst 0x6fa1e955 // udot v21.4s, v10.16b, v1.4b[3]\n" ".inst 0x6fa2e959 // udot v25.4s, v10.16b, v2.4b[3]\n" @@ -1740,7 +1740,9 @@ void a64_hybrid_u8qa_dot_4x16_a55 ( "cmp x26, #0x4\n" "ldr q8, [x10, #0x20]\n" ".inst 0x6f80e0d0 // udot v16.4s, v6.16b, v0.4b[0]\n" + "ldr q9, [x10, #0x30]\n" ".inst 0x6f81e0d4 // udot v20.4s, v6.16b, v1.4b[0]\n" + "add x10, x10, #0x40\n" ".inst 0x6f82e0d8 // udot v24.4s, v6.16b, v2.4b[0]\n" ".inst 0x6f83e0dc // udot v28.4s, v6.16b, v3.4b[0]\n" ".inst 0x6f80e0f1 // udot v17.4s, v7.16b, v0.4b[0]\n" @@ -1751,8 +1753,6 @@ void a64_hybrid_u8qa_dot_4x16_a55 ( ".inst 0x6f81e116 // udot v22.4s, v8.16b, v1.4b[0]\n" ".inst 0x6f82e11a // udot v26.4s, v8.16b, v2.4b[0]\n" ".inst 0x6f83e11e // udot v30.4s, v8.16b, v3.4b[0]\n" - "ldr q9, [x10, #0x30]\n" - "add x10, x10, #0x40\n" ".inst 0x6f80e133 // udot v19.4s, v9.16b, v0.4b[0]\n" ".inst 0x6f81e137 // udot v23.4s, v9.16b, v1.4b[0]\n" ".inst 0x6f82e13b // udot v27.4s, v9.16b, v2.4b[0]\n" @@ -1787,7 +1787,9 @@ void a64_hybrid_u8qa_dot_4x16_a55 ( "ldr q4, [x10, #0x10]\n" "ldr q5, [x10, #0x20]\n" ".inst 0x6f80e150 // udot v16.4s, v10.16b, v0.4b[0]\n" + "ldr q6, [x10, #0x30]\n" ".inst 0x6f81e154 // udot v20.4s, v10.16b, v1.4b[0]\n" + "add x10, x10, #0x40\n" ".inst 0x6f82e158 // udot v24.4s, v10.16b, v2.4b[0]\n" ".inst 0x6f83e15c // udot v28.4s, v10.16b, v3.4b[0]\n" ".inst 0x6f80e091 // udot v17.4s, v4.16b, v0.4b[0]\n" @@ -1798,8 +1800,6 @@ void a64_hybrid_u8qa_dot_4x16_a55 ( ".inst 0x6f81e0b6 // udot v22.4s, v5.16b, v1.4b[0]\n" ".inst 0x6f82e0ba // udot v26.4s, v5.16b, v2.4b[0]\n" ".inst 0x6f83e0be // udot v30.4s, v5.16b, v3.4b[0]\n" - "ldr q6, [x10, #0x30]\n" - "add x10, x10, #0x40\n" ".inst 0x6f80e0d3 // udot v19.4s, v6.16b, v0.4b[0]\n" ".inst 0x6f81e0d7 // udot v23.4s, v6.16b, v1.4b[0]\n" ".inst 0x6f82e0db // udot v27.4s, v6.16b, v2.4b[0]\n" diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16/a55.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16/a55.cpp index f131eb2eb5..8833651768 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16/a55.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16/a55.cpp @@ -87,73 +87,73 @@ void a64_hybrid_u8u32_dot_6x16_a55 ( "cmp %x[M], #0x2\n" "bgt 69f\n" "beq 35f\n" - "ldr x14, [%x[args_ptr], %[offsetof_N]]\n" - "ldr x13, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x12, %x[output_ptr]\n" + "ldr x17, [%x[args_ptr], %[offsetof_N]]\n" + "ldr x16, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "mov x15, %x[output_ptr]\n" "2:" // Height 1: Column loop "tbz %x[flags], #0, 12f\n" - "cmp x14, #0x10\n" + "cmp x17, #0x10\n" "bge 11f\n" - "tbz x14, #3, 6f\n" - "ld1 { v8.4s }, [x12], #0x10\n" - "ld1 { v9.4s }, [x12], #0x10\n" - "tbz x14, #2, 4f\n" - "ld1 { v10.4s }, [x12], #0x10\n" - "tbz x14, #1, 3f\n" + "tbz x17, #3, 6f\n" + "ld1 { v8.4s }, [x15], #0x10\n" + "ld1 { v9.4s }, [x15], #0x10\n" + "tbz x17, #2, 4f\n" + "ld1 { v10.4s }, [x15], #0x10\n" + "tbz x17, #1, 3f\n" "mov x24, #0x38\n" - "ldr d11, [x12], #0x8\n" - "tbz x14, #0, 10f\n" - "ld1 { v11.s }[2], [x12]\n" + "ldr d11, [x15], #0x8\n" + "tbz x17, #0, 10f\n" + "ld1 { v11.s }[2], [x15]\n" "b 10f\n" "3:" // Height 1: Partial accumulate: partial_1_12 "mov x24, #0x30\n" - "tbz x14, #0, 10f\n" - "ldr s11, [x12, #0x0]\n" + "tbz x17, #0, 10f\n" + "ldr s11, [x15, #0x0]\n" "b 10f\n" "4:" // Height 1: Partial accumulate: partial_2_8 - "tbz x14, #1, 5f\n" - "ldr d10, [x12], #0x8\n" + "tbz x17, #1, 5f\n" + "ldr d10, [x15], #0x8\n" "mov x24, #0x28\n" - "tbz x14, #0, 10f\n" - "ld1 { v10.s }[2], [x12]\n" + "tbz x17, #0, 10f\n" + "ld1 { v10.s }[2], [x15]\n" "b 10f\n" "5:" // Height 1: Partial accumulate: partial_1_8 "mov x24, #0x20\n" - "tbz x14, #0, 10f\n" - "ldr s10, [x12, #0x0]\n" + "tbz x17, #0, 10f\n" + "ldr s10, [x15, #0x0]\n" "b 10f\n" "6:" // Height 1: Partial accumulate: partial_4_0 - "tbz x14, #2, 8f\n" - "ld1 { v8.4s }, [x12], #0x10\n" - "tbz x14, #1, 7f\n" + "tbz x17, #2, 8f\n" + "ld1 { v8.4s }, [x15], #0x10\n" + "tbz x17, #1, 7f\n" "mov x24, #0x18\n" - "ldr d9, [x12], #0x8\n" - "tbz x14, #0, 10f\n" - "ld1 { v9.s }[2], [x12]\n" + "ldr d9, [x15], #0x8\n" + "tbz x17, #0, 10f\n" + "ld1 { v9.s }[2], [x15]\n" "b 10f\n" "7:" // Height 1: Partial accumulate: partial_1_4 "mov x24, #0x10\n" - "tbz x14, #0, 10f\n" - "ldr s9, [x12, #0x0]\n" + "tbz x17, #0, 10f\n" + "ldr s9, [x15, #0x0]\n" "b 10f\n" "8:" // Height 1: Partial accumulate: partial_2_0 - "tbz x14, #1, 9f\n" - "ldr d8, [x12], #0x8\n" + "tbz x17, #1, 9f\n" + "ldr d8, [x15], #0x8\n" "mov x24, #0x8\n" - "tbz x14, #0, 10f\n" - "ld1 { v8.s }[2], [x12]\n" + "tbz x17, #0, 10f\n" + "ld1 { v8.s }[2], [x15]\n" "b 10f\n" "9:" // Height 1: Partial accumulate: partial_1_0 - "ldr s8, [x12, #0x0]\n" + "ldr s8, [x15, #0x0]\n" "mov x24, #0x0\n" "10:" // Height 1: Partial accumulate: Done - "sub x12, x12, x24\n" + "sub x15, x15, x24\n" "b 13f\n" "11:" // Height 1: full accumulate - "ldr q8, [x12, #0x0]\n" - "ldr q9, [x12, #0x10]\n" - "ldr q10, [x12, #0x20]\n" - "ldr q11, [x12, #0x30]\n" + "ldr q8, [x15, #0x0]\n" + "ldr q9, [x15, #0x10]\n" + "ldr q10, [x15, #0x20]\n" + "ldr q11, [x15, #0x30]\n" "b 13f\n" "12:" // Height 1: no accumulate "movi v8.4s, #0x0\n" @@ -161,325 +161,325 @@ void a64_hybrid_u8u32_dot_6x16_a55 ( "movi v10.4s, #0x0\n" "movi v11.4s, #0x0\n" "13:" // Height 1: setup done - "mov x11, #0x0\n" + "mov x14, #0x0\n" "14:" // Height 1: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w10, [x20, x11, LSL #0x2]\n" + "ldr w13, [x20, x14, LSL #0x2]\n" "tbz %x[flags], #3, 15f\n" - "ldr x20, [%x[input_ptr], x11, LSL #0x3]\n" + "ldr x20, [%x[input_ptr], x14, LSL #0x3]\n" "add x20, x20, x19, LSL #3\n" - "ldr x9, [x20, #0x0]\n" - "cbnz x11, 16f\n" + "ldr x12, [x20, #0x0]\n" + "cbnz x14, 16f\n" "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x9, x9, x19\n" + "add x12, x12, x19\n" "b 16f\n" "15:" // Height 1: setup direct input - "mov x9, %x[input_ptr]\n" + "mov x12, %x[input_ptr]\n" "16:" // Height 1: input setup done - "cmp x10, #0x10\n" + "cmp x13, #0x10\n" "blt 19f\n" - "ldr q0, [x9, #0x0]\n" - "ldr q6, [x13, #0x0]\n" - "cmp x10, #0x20\n" + "ldr q0, [x12, #0x0]\n" + "ldr q6, [x16, #0x0]\n" + "cmp x13, #0x20\n" "blt 18f\n" "17:" // Height 1: Multiply loop: Main loop head ".inst 0x6f80e0c8 // udot v8.4s, v6.16b, v0.4b[0]\n" - "ldr d7, [x13, #0x10]\n" - "ldr x19, [x13, #0x18]\n" - "add x9, x9, #0x10\n" - "ldr d6, [x13, #0x20]\n" - "sub x10, x10, #0x10\n" - "ldr x28, [x13, #0x28]\n" - "cmp x10, #0x20\n" - "mov v7.d[1], x19\n" - "prfm pldl1keep, [x9, #0x80]\n" - "ldr x19, [x13, #0x38]\n" - "mov v6.d[1], x28\n" + "ldr d7, [x16, #0x10]\n" + "ldr x11, [x16, #0x18]\n" + "add x12, x12, #0x10\n" + "ldr d6, [x16, #0x20]\n" + "sub x13, x13, #0x10\n" + "ldr x10, [x16, #0x28]\n" + "cmp x13, #0x20\n" + "mov v7.d[1], x11\n" + "prfm pldl1keep, [x12, #0x80]\n" + "ldr x11, [x16, #0x38]\n" ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" - "ldr d7, [x13, #0x30]\n" + "mov v6.d[1], x10\n" + "ldr d7, [x16, #0x30]\n" ".inst 0x6f80e0ca // udot v10.4s, v6.16b, v0.4b[0]\n" - "ldr d6, [x13, #0x40]\n" - "ldr x28, [x13, #0x48]\n" - "mov v7.d[1], x19\n" - "ldr x19, [x13, #0x58]\n" - "ldr x27, [x9, #0x8]\n" - "mov v6.d[1], x28\n" + "ldr d6, [x16, #0x40]\n" + "ldr x10, [x16, #0x48]\n" + "mov v7.d[1], x11\n" + "ldr x11, [x16, #0x58]\n" + "ldr x9, [x12, #0x8]\n" ".inst 0x6f80e0eb // udot v11.4s, v7.16b, v0.4b[0]\n" - "ldr d7, [x13, #0x50]\n" + "mov v6.d[1], x10\n" + "ldr d7, [x16, #0x50]\n" ".inst 0x6fa0e0c8 // udot v8.4s, v6.16b, v0.4b[1]\n" - "ldr d6, [x13, #0x60]\n" - "ldr x28, [x13, #0x68]\n" - "mov v7.d[1], x19\n" - "ldr x19, [x13, #0x78]\n" - "mov v6.d[1], x28\n" + "ldr d6, [x16, #0x60]\n" + "ldr x10, [x16, #0x68]\n" + "mov v7.d[1], x11\n" + "ldr x11, [x16, #0x78]\n" ".inst 0x6fa0e0e9 // udot v9.4s, v7.16b, v0.4b[1]\n" - "ldr d7, [x13, #0x70]\n" + "mov v6.d[1], x10\n" + "ldr d7, [x16, #0x70]\n" ".inst 0x6fa0e0ca // udot v10.4s, v6.16b, v0.4b[1]\n" - "ldr d6, [x13, #0x80]\n" - "ldr x28, [x13, #0x88]\n" - "mov v7.d[1], x19\n" - "ldr x19, [x13, #0x98]\n" - "mov v6.d[1], x28\n" + "ldr d6, [x16, #0x80]\n" + "ldr x10, [x16, #0x88]\n" + "mov v7.d[1], x11\n" + "ldr x11, [x16, #0x98]\n" ".inst 0x6fa0e0eb // udot v11.4s, v7.16b, v0.4b[1]\n" - "ldr d7, [x13, #0x90]\n" + "mov v6.d[1], x10\n" + "ldr d7, [x16, #0x90]\n" ".inst 0x6f80e8c8 // udot v8.4s, v6.16b, v0.4b[2]\n" - "ldr d6, [x13, #0xa0]\n" - "ldr x28, [x13, #0xa8]\n" - "mov v7.d[1], x19\n" - "ldr x19, [x13, #0xb8]\n" - "mov v6.d[1], x28\n" + "ldr d6, [x16, #0xa0]\n" + "ldr x10, [x16, #0xa8]\n" + "mov v7.d[1], x11\n" + "ldr x11, [x16, #0xb8]\n" ".inst 0x6f80e8e9 // udot v9.4s, v7.16b, v0.4b[2]\n" - "ldr d7, [x13, #0xb0]\n" + "mov v6.d[1], x10\n" + "ldr d7, [x16, #0xb0]\n" ".inst 0x6f80e8ca // udot v10.4s, v6.16b, v0.4b[2]\n" - "ldr d6, [x13, #0xc0]\n" - "ldr x28, [x13, #0xc8]\n" - "mov v7.d[1], x19\n" - "ldr x19, [x13, #0xd8]\n" - "mov v6.d[1], x28\n" + "ldr d6, [x16, #0xc0]\n" + "ldr x10, [x16, #0xc8]\n" + "mov v7.d[1], x11\n" + "ldr x11, [x16, #0xd8]\n" ".inst 0x6f80e8eb // udot v11.4s, v7.16b, v0.4b[2]\n" - "ldr d7, [x13, #0xd0]\n" + "mov v6.d[1], x10\n" + "ldr d7, [x16, #0xd0]\n" ".inst 0x6fa0e8c8 // udot v8.4s, v6.16b, v0.4b[3]\n" - "ldr d6, [x13, #0xe0]\n" - "ldr x28, [x13, #0xe8]\n" - "mov v7.d[1], x19\n" - "ldr x19, [x13, #0xf8]\n" - "mov v6.d[1], x28\n" + "ldr d6, [x16, #0xe0]\n" + "ldr x10, [x16, #0xe8]\n" + "mov v7.d[1], x11\n" + "ldr x11, [x16, #0xf8]\n" ".inst 0x6fa0e8e9 // udot v9.4s, v7.16b, v0.4b[3]\n" - "ldr d7, [x13, #0xf0]\n" - "add x13, x13, #0x100\n" + "mov v6.d[1], x10\n" + "ldr d7, [x16, #0xf0]\n" + "add x16, x16, #0x100\n" ".inst 0x6fa0e8ca // udot v10.4s, v6.16b, v0.4b[3]\n" - "ldr d6, [x13, #0x0]\n" - "ldr x28, [x13, #0x8]\n" - "mov v7.d[1], x19\n" - "mov v6.d[1], x28\n" + "ldr d6, [x16, #0x0]\n" + "ldr x10, [x16, #0x8]\n" + "mov v7.d[1], x11\n" ".inst 0x6fa0e8eb // udot v11.4s, v7.16b, v0.4b[3]\n" - "ldr d0, [x9, #0x0]\n" - "mov v0.d[1], x27\n" + "mov v6.d[1], x10\n" + "ldr d0, [x12, #0x0]\n" + "mov v0.d[1], x9\n" "bge 17b\n" "18:" // Height 1: Multiply loop: Single iteration only ".inst 0x6f80e0c8 // udot v8.4s, v6.16b, v0.4b[0]\n" - "ldr q7, [x13, #0x10]\n" - "ldr q6, [x13, #0x20]\n" - "sub x10, x10, #0x10\n" - "add x9, x9, #0x10\n" + "ldr q7, [x16, #0x10]\n" + "ldr q6, [x16, #0x20]\n" + "sub x13, x13, #0x10\n" + "add x12, x12, #0x10\n" ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" - "prfm pldl1keep, [x9, #0x80]\n" + "prfm pldl1keep, [x12, #0x80]\n" ".inst 0x6f80e0ca // udot v10.4s, v6.16b, v0.4b[0]\n" - "ldr q7, [x13, #0x30]\n" - "ldr q6, [x13, #0x40]\n" + "ldr q7, [x16, #0x30]\n" + "ldr q6, [x16, #0x40]\n" ".inst 0x6f80e0eb // udot v11.4s, v7.16b, v0.4b[0]\n" - "ldr q7, [x13, #0x50]\n" + "ldr q7, [x16, #0x50]\n" ".inst 0x6fa0e0c8 // udot v8.4s, v6.16b, v0.4b[1]\n" - "ldr q6, [x13, #0x60]\n" + "ldr q6, [x16, #0x60]\n" ".inst 0x6fa0e0e9 // udot v9.4s, v7.16b, v0.4b[1]\n" - "ldr q7, [x13, #0x70]\n" + "ldr q7, [x16, #0x70]\n" ".inst 0x6fa0e0ca // udot v10.4s, v6.16b, v0.4b[1]\n" - "ldr q6, [x13, #0x80]\n" + "ldr q6, [x16, #0x80]\n" ".inst 0x6fa0e0eb // udot v11.4s, v7.16b, v0.4b[1]\n" - "ldr q7, [x13, #0x90]\n" + "ldr q7, [x16, #0x90]\n" ".inst 0x6f80e8c8 // udot v8.4s, v6.16b, v0.4b[2]\n" - "ldr q6, [x13, #0xa0]\n" + "ldr q6, [x16, #0xa0]\n" ".inst 0x6f80e8e9 // udot v9.4s, v7.16b, v0.4b[2]\n" - "ldr q7, [x13, #0xb0]\n" + "ldr q7, [x16, #0xb0]\n" ".inst 0x6f80e8ca // udot v10.4s, v6.16b, v0.4b[2]\n" - "ldr q6, [x13, #0xc0]\n" + "ldr q6, [x16, #0xc0]\n" ".inst 0x6f80e8eb // udot v11.4s, v7.16b, v0.4b[2]\n" - "ldr q7, [x13, #0xd0]\n" + "ldr q7, [x16, #0xd0]\n" ".inst 0x6fa0e8c8 // udot v8.4s, v6.16b, v0.4b[3]\n" - "ldr q6, [x13, #0xe0]\n" + "ldr q6, [x16, #0xe0]\n" ".inst 0x6fa0e8e9 // udot v9.4s, v7.16b, v0.4b[3]\n" - "ldr q7, [x13, #0xf0]\n" + "ldr q7, [x16, #0xf0]\n" ".inst 0x6fa0e8ca // udot v10.4s, v6.16b, v0.4b[3]\n" - "add x13, x13, #0x100\n" + "add x16, x16, #0x100\n" ".inst 0x6fa0e8eb // udot v11.4s, v7.16b, v0.4b[3]\n" "19:" // Height 1: Multiply loop: Main loop skip - "cbz x10, 24f\n" - "cmp x10, #0x4\n" + "cbz x13, 24f\n" + "cmp x13, #0x4\n" "blt 21f\n" "20:" // Height 1: Multiply loop: Odd block loop - "ldr s0, [x9], #0x4\n" - "sub x10, x10, #0x4\n" - "ldr q6, [x13, #0x0]\n" - "cmp x10, #0x4\n" - "ldr q7, [x13, #0x10]\n" + "ldr s0, [x12], #0x4\n" + "sub x13, x13, #0x4\n" + "ldr q6, [x16, #0x0]\n" + "cmp x13, #0x4\n" + "ldr q7, [x16, #0x10]\n" ".inst 0x6f80e0c8 // udot v8.4s, v6.16b, v0.4b[0]\n" - "ldr q6, [x13, #0x20]\n" + "ldr q6, [x16, #0x20]\n" ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" - "ldr q7, [x13, #0x30]\n" - "add x13, x13, #0x40\n" + "ldr q7, [x16, #0x30]\n" + "add x16, x16, #0x40\n" ".inst 0x6f80e0ca // udot v10.4s, v6.16b, v0.4b[0]\n" ".inst 0x6f80e0eb // udot v11.4s, v7.16b, v0.4b[0]\n" "bge 20b\n" - "cbz x10, 24f\n" + "cbz x13, 24f\n" "21:" // Height 1: Multiply loop: Skip odd blocks - "tbz x10, #1, 22f\n" - "ldr h0, [x9], #0x2\n" - "tbz x10, #0, 23f\n" - "ld1 { v0.b }[2], [x9]\n" + "tbz x13, #1, 22f\n" + "ldr h0, [x12], #0x2\n" + "tbz x13, #0, 23f\n" + "ld1 { v0.b }[2], [x12]\n" "b 23f\n" "22:" // Height 1: Multiply loop: Ragged operand read: partial_1_0 - "ldr b0, [x9, #0x0]\n" + "ldr b0, [x12, #0x0]\n" "23:" // Height 1: Multiply loop: Ragged operand read: Done - "ldr q6, [x13, #0x0]\n" - "ldr q7, [x13, #0x10]\n" + "ldr q6, [x16, #0x0]\n" + "ldr q7, [x16, #0x10]\n" ".inst 0x6f80e0c8 // udot v8.4s, v6.16b, v0.4b[0]\n" - "ldr q6, [x13, #0x20]\n" + "ldr q6, [x16, #0x20]\n" ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" - "ldr q7, [x13, #0x30]\n" - "add x13, x13, #0x40\n" + "ldr q7, [x16, #0x30]\n" + "add x16, x16, #0x40\n" ".inst 0x6f80e0ca // udot v10.4s, v6.16b, v0.4b[0]\n" ".inst 0x6f80e0eb // udot v11.4s, v7.16b, v0.4b[0]\n" "24:" // Height 1: Multiply loop: No odd multiplies "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" - "add x11, x11, #0x1\n" - "cmp x11, x19\n" + "add x14, x14, #0x1\n" + "cmp x14, x19\n" "bne 14b\n" - "prfm pstl1keep, [x12, #0x0]\n" - "cmp x14, #0x10\n" + "prfm pstl1keep, [x15, #0x0]\n" + "cmp x17, #0x10\n" "bge 33f\n" - "tbz x14, #3, 28f\n" - "st1 { v8.4s }, [x12], #0x10\n" - "st1 { v9.4s }, [x12], #0x10\n" - "tbz x14, #2, 26f\n" - "st1 { v10.4s }, [x12], #0x10\n" - "tbz x14, #1, 25f\n" - "str d11, [x12], #0x8\n" - "tbz x14, #0, 32f\n" - "st1 { v11.s }[2], [x12]\n" + "tbz x17, #3, 28f\n" + "st1 { v8.4s }, [x15], #0x10\n" + "st1 { v9.4s }, [x15], #0x10\n" + "tbz x17, #2, 26f\n" + "st1 { v10.4s }, [x15], #0x10\n" + "tbz x17, #1, 25f\n" + "str d11, [x15], #0x8\n" + "tbz x17, #0, 32f\n" + "st1 { v11.s }[2], [x15]\n" "b 32f\n" "25:" // Height 1: Partial direct writeback: partial_1_12 - "tbz x14, #0, 32f\n" - "str s11, [x12, #0x0]\n" + "tbz x17, #0, 32f\n" + "str s11, [x15, #0x0]\n" "b 32f\n" "26:" // Height 1: Partial direct writeback: partial_2_8 - "tbz x14, #1, 27f\n" - "str d10, [x12], #0x8\n" - "tbz x14, #0, 32f\n" - "st1 { v10.s }[2], [x12]\n" + "tbz x17, #1, 27f\n" + "str d10, [x15], #0x8\n" + "tbz x17, #0, 32f\n" + "st1 { v10.s }[2], [x15]\n" "b 32f\n" "27:" // Height 1: Partial direct writeback: partial_1_8 - "tbz x14, #0, 32f\n" - "str s10, [x12, #0x0]\n" + "tbz x17, #0, 32f\n" + "str s10, [x15, #0x0]\n" "b 32f\n" "28:" // Height 1: Partial direct writeback: partial_4_0 - "tbz x14, #2, 30f\n" - "st1 { v8.4s }, [x12], #0x10\n" - "tbz x14, #1, 29f\n" - "str d9, [x12], #0x8\n" - "tbz x14, #0, 32f\n" - "st1 { v9.s }[2], [x12]\n" + "tbz x17, #2, 30f\n" + "st1 { v8.4s }, [x15], #0x10\n" + "tbz x17, #1, 29f\n" + "str d9, [x15], #0x8\n" + "tbz x17, #0, 32f\n" + "st1 { v9.s }[2], [x15]\n" "b 32f\n" "29:" // Height 1: Partial direct writeback: partial_1_4 - "tbz x14, #0, 32f\n" - "str s9, [x12, #0x0]\n" + "tbz x17, #0, 32f\n" + "str s9, [x15, #0x0]\n" "b 32f\n" "30:" // Height 1: Partial direct writeback: partial_2_0 - "tbz x14, #1, 31f\n" - "str d8, [x12], #0x8\n" - "tbz x14, #0, 32f\n" - "st1 { v8.s }[2], [x12]\n" + "tbz x17, #1, 31f\n" + "str d8, [x15], #0x8\n" + "tbz x17, #0, 32f\n" + "st1 { v8.s }[2], [x15]\n" "b 32f\n" "31:" // Height 1: Partial direct writeback: partial_1_0 - "str s8, [x12, #0x0]\n" + "str s8, [x15, #0x0]\n" "32:" // Height 1: Partial direct writeback: Done "b 34f\n" "33:" // Height 1: Full writeback - "str q8, [x12, #0x0]\n" - "str q9, [x12, #0x10]\n" - "str q10, [x12, #0x20]\n" - "str q11, [x12, #0x30]\n" - "add x12, x12, #0x40\n" + "str q8, [x15, #0x0]\n" + "str q9, [x15, #0x10]\n" + "str q10, [x15, #0x20]\n" + "str q11, [x15, #0x30]\n" + "add x15, x15, #0x40\n" "34:" // Height 1: Writeback done - "subs x14, x14, #0x10\n" + "subs x17, x17, #0x10\n" "bgt 2b\n" "b 206f\n" "35:" // Height 2 - "ldr x14, [%x[args_ptr], %[offsetof_N]]\n" - "mov x12, %x[output_ptr]\n" - "ldr x13, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "ldr x17, [%x[args_ptr], %[offsetof_N]]\n" + "mov x15, %x[output_ptr]\n" + "ldr x16, [%x[args_ptr], %[offsetof_B_ptr]]\n" "36:" // Height 2: Column loop "tbz %x[flags], #0, 46f\n" "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "cmp x14, #0x10\n" - "add x23, x12, x19, LSL #2\n" + "cmp x17, #0x10\n" + "add x23, x15, x19, LSL #2\n" "bge 45f\n" - "tbz x14, #3, 40f\n" - "ld1 { v8.4s }, [x12], #0x10\n" + "tbz x17, #3, 40f\n" + "ld1 { v8.4s }, [x15], #0x10\n" "ld1 { v12.4s }, [x23], #0x10\n" - "ld1 { v9.4s }, [x12], #0x10\n" + "ld1 { v9.4s }, [x15], #0x10\n" "ld1 { v13.4s }, [x23], #0x10\n" - "tbz x14, #2, 38f\n" - "ld1 { v10.4s }, [x12], #0x10\n" + "tbz x17, #2, 38f\n" + "ld1 { v10.4s }, [x15], #0x10\n" "ld1 { v14.4s }, [x23], #0x10\n" - "tbz x14, #1, 37f\n" + "tbz x17, #1, 37f\n" "mov x24, #0x38\n" - "ldr d11, [x12], #0x8\n" + "ldr d11, [x15], #0x8\n" "ldr d15, [x23], #0x8\n" - "tbz x14, #0, 44f\n" - "ld1 { v11.s }[2], [x12]\n" + "tbz x17, #0, 44f\n" + "ld1 { v11.s }[2], [x15]\n" "ld1 { v15.s }[2], [x23]\n" "b 44f\n" "37:" // Height 2: Partial accumulate: partial_1_12 "mov x24, #0x30\n" - "tbz x14, #0, 44f\n" - "ldr s11, [x12, #0x0]\n" + "tbz x17, #0, 44f\n" + "ldr s11, [x15, #0x0]\n" "ldr s15, [x23, #0x0]\n" "b 44f\n" "38:" // Height 2: Partial accumulate: partial_2_8 - "tbz x14, #1, 39f\n" - "ldr d10, [x12], #0x8\n" + "tbz x17, #1, 39f\n" + "ldr d10, [x15], #0x8\n" "ldr d14, [x23], #0x8\n" "mov x24, #0x28\n" - "tbz x14, #0, 44f\n" - "ld1 { v10.s }[2], [x12]\n" + "tbz x17, #0, 44f\n" + "ld1 { v10.s }[2], [x15]\n" "ld1 { v14.s }[2], [x23]\n" "b 44f\n" "39:" // Height 2: Partial accumulate: partial_1_8 "mov x24, #0x20\n" - "tbz x14, #0, 44f\n" - "ldr s10, [x12, #0x0]\n" + "tbz x17, #0, 44f\n" + "ldr s10, [x15, #0x0]\n" "ldr s14, [x23, #0x0]\n" "b 44f\n" "40:" // Height 2: Partial accumulate: partial_4_0 - "tbz x14, #2, 42f\n" - "ld1 { v8.4s }, [x12], #0x10\n" + "tbz x17, #2, 42f\n" + "ld1 { v8.4s }, [x15], #0x10\n" "ld1 { v12.4s }, [x23], #0x10\n" - "tbz x14, #1, 41f\n" + "tbz x17, #1, 41f\n" "mov x24, #0x18\n" - "ldr d9, [x12], #0x8\n" + "ldr d9, [x15], #0x8\n" "ldr d13, [x23], #0x8\n" - "tbz x14, #0, 44f\n" - "ld1 { v9.s }[2], [x12]\n" + "tbz x17, #0, 44f\n" + "ld1 { v9.s }[2], [x15]\n" "ld1 { v13.s }[2], [x23]\n" "b 44f\n" "41:" // Height 2: Partial accumulate: partial_1_4 "mov x24, #0x10\n" - "tbz x14, #0, 44f\n" - "ldr s9, [x12, #0x0]\n" + "tbz x17, #0, 44f\n" + "ldr s9, [x15, #0x0]\n" "ldr s13, [x23, #0x0]\n" "b 44f\n" "42:" // Height 2: Partial accumulate: partial_2_0 - "tbz x14, #1, 43f\n" - "ldr d8, [x12], #0x8\n" + "tbz x17, #1, 43f\n" + "ldr d8, [x15], #0x8\n" "ldr d12, [x23], #0x8\n" "mov x24, #0x8\n" - "tbz x14, #0, 44f\n" - "ld1 { v8.s }[2], [x12]\n" + "tbz x17, #0, 44f\n" + "ld1 { v8.s }[2], [x15]\n" "ld1 { v12.s }[2], [x23]\n" "b 44f\n" "43:" // Height 2: Partial accumulate: partial_1_0 - "ldr s8, [x12, #0x0]\n" + "ldr s8, [x15, #0x0]\n" "mov x24, #0x0\n" "ldr s12, [x23, #0x0]\n" "44:" // Height 2: Partial accumulate: Done - "sub x12, x12, x24\n" + "sub x15, x15, x24\n" "b 47f\n" "45:" // Height 2: full accumulate - "ldr q8, [x12, #0x0]\n" - "ldr q9, [x12, #0x10]\n" - "ldr q10, [x12, #0x20]\n" - "ldr q11, [x12, #0x30]\n" + "ldr q8, [x15, #0x0]\n" + "ldr q9, [x15, #0x10]\n" + "ldr q10, [x15, #0x20]\n" + "ldr q11, [x15, #0x30]\n" "ldr q12, [x23, #0x0]\n" "ldr q13, [x23, #0x10]\n" "ldr q14, [x23, #0x20]\n" @@ -495,420 +495,420 @@ void a64_hybrid_u8u32_dot_6x16_a55 ( "movi v14.4s, #0x0\n" "movi v15.4s, #0x0\n" "47:" // Height 2: setup done - "mov x11, #0x0\n" + "mov x14, #0x0\n" "48:" // Height 2: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w10, [x20, x11, LSL #0x2]\n" + "ldr w13, [x20, x14, LSL #0x2]\n" "tbz %x[flags], #3, 49f\n" - "ldr x20, [%x[input_ptr], x11, LSL #0x3]\n" + "ldr x20, [%x[input_ptr], x14, LSL #0x3]\n" "add x20, x20, x19, LSL #3\n" - "ldr x9, [x20, #0x0]\n" - "ldr x26, [x20, #0x8]\n" - "cbnz x11, 50f\n" + "ldr x12, [x20, #0x0]\n" + "ldr x28, [x20, #0x8]\n" + "cbnz x14, 50f\n" "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x9, x9, x19\n" - "add x26, x26, x19\n" + "add x12, x12, x19\n" + "add x28, x28, x19\n" "b 50f\n" "49:" // Height 2: setup direct input - "mov x9, %x[input_ptr]\n" - "add x26, x9, x19\n" + "mov x12, %x[input_ptr]\n" + "add x28, x12, x19\n" "50:" // Height 2: input setup done - "cmp x10, #0x10\n" + "cmp x13, #0x10\n" "blt 53f\n" - "ldr q0, [x9, #0x0]\n" - "ldr q1, [x26, #0x0]\n" - "cmp x10, #0x20\n" - "ldr q6, [x13, #0x0]\n" + "ldr q0, [x12, #0x0]\n" + "ldr q1, [x28, #0x0]\n" + "cmp x13, #0x20\n" + "ldr q6, [x16, #0x0]\n" "blt 52f\n" "51:" // Height 2: Multiply loop: Main loop head ".inst 0x6f80e0c8 // udot v8.4s, v6.16b, v0.4b[0]\n" + "ldr d7, [x16, #0x10]\n" ".inst 0x6f81e0cc // udot v12.4s, v6.16b, v1.4b[0]\n" - "ldr d7, [x13, #0x10]\n" - "add x9, x9, #0x10\n" - "ldr x19, [x13, #0x18]\n" - "add x26, x26, #0x10\n" - "prfm pldl1keep, [x9, #0x80]\n" - "sub x10, x10, #0x10\n" - "ldr d6, [x13, #0x20]\n" - "cmp x10, #0x20\n" - "mov v7.d[1], x19\n" - "ldr x28, [x13, #0x28]\n" - "ldr x19, [x13, #0x38]\n" + "ldr x11, [x16, #0x18]\n" + "ldr d6, [x16, #0x20]\n" + "add x12, x12, #0x10\n" + "ldr x10, [x16, #0x28]\n" + "add x28, x28, #0x10\n" + "mov v7.d[1], x11\n" + "prfm pldl1keep, [x12, #0x80]\n" + "prfm pldl1keep, [x28, #0x80]\n" + "sub x13, x13, #0x10\n" ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" + "mov v6.d[1], x10\n" ".inst 0x6f81e0ed // udot v13.4s, v7.16b, v1.4b[0]\n" - "mov v6.d[1], x28\n" - "ldr d7, [x13, #0x30]\n" - "prfm pldl1keep, [x26, #0x80]\n" + "ldr d7, [x16, #0x30]\n" ".inst 0x6f80e0ca // udot v10.4s, v6.16b, v0.4b[0]\n" + "ldr x11, [x16, #0x38]\n" ".inst 0x6f81e0ce // udot v14.4s, v6.16b, v1.4b[0]\n" - "mov v7.d[1], x19\n" - "ldr d6, [x13, #0x40]\n" - "ldr x28, [x13, #0x48]\n" + "ldr d6, [x16, #0x40]\n" + "ldr x10, [x16, #0x48]\n" + "cmp x13, #0x20\n" + "mov v7.d[1], x11\n" + "ldr x11, [x16, #0x58]\n" + "ldr x9, [x12, #0x8]\n" ".inst 0x6f80e0eb // udot v11.4s, v7.16b, v0.4b[0]\n" + "mov v6.d[1], x10\n" ".inst 0x6f81e0ef // udot v15.4s, v7.16b, v1.4b[0]\n" - "ldr d7, [x13, #0x50]\n" - "mov v6.d[1], x28\n" - "ldr x19, [x13, #0x58]\n" - "ldr x28, [x13, #0x68]\n" + "ldr d7, [x16, #0x50]\n" ".inst 0x6fa0e0c8 // udot v8.4s, v6.16b, v0.4b[1]\n" + "ldr x10, [x16, #0x68]\n" ".inst 0x6fa1e0cc // udot v12.4s, v6.16b, v1.4b[1]\n" - "mov v7.d[1], x19\n" - "ldr d6, [x13, #0x60]\n" - "ldr x19, [x13, #0x78]\n" + "ldr d6, [x16, #0x60]\n" + "mov v7.d[1], x11\n" + "ldr x11, [x16, #0x78]\n" + "ldr x27, [x28, #0x8]\n" ".inst 0x6fa0e0e9 // udot v9.4s, v7.16b, v0.4b[1]\n" + "mov v6.d[1], x10\n" ".inst 0x6fa1e0ed // udot v13.4s, v7.16b, v1.4b[1]\n" - "mov v6.d[1], x28\n" - "ldr d7, [x13, #0x70]\n" - "ldr x28, [x13, #0x88]\n" + "ldr d7, [x16, #0x70]\n" ".inst 0x6fa0e0ca // udot v10.4s, v6.16b, v0.4b[1]\n" + "ldr x10, [x16, #0x88]\n" ".inst 0x6fa1e0ce // udot v14.4s, v6.16b, v1.4b[1]\n" - "mov v7.d[1], x19\n" - "ldr d6, [x13, #0x80]\n" - "ldr x19, [x13, #0x98]\n" + "ldr d6, [x16, #0x80]\n" + "mov v7.d[1], x11\n" + "ldr x11, [x16, #0x98]\n" ".inst 0x6fa0e0eb // udot v11.4s, v7.16b, v0.4b[1]\n" + "mov v6.d[1], x10\n" ".inst 0x6fa1e0ef // udot v15.4s, v7.16b, v1.4b[1]\n" - "mov v6.d[1], x28\n" - "ldr d7, [x13, #0x90]\n" - "ldr x28, [x13, #0xa8]\n" + "ldr d7, [x16, #0x90]\n" ".inst 0x6f80e8c8 // udot v8.4s, v6.16b, v0.4b[2]\n" + "ldr x10, [x16, #0xa8]\n" ".inst 0x6f81e8cc // udot v12.4s, v6.16b, v1.4b[2]\n" - "mov v7.d[1], x19\n" - "ldr d6, [x13, #0xa0]\n" - "ldr x19, [x13, #0xb8]\n" + "ldr d6, [x16, #0xa0]\n" + "mov v7.d[1], x11\n" + "ldr x11, [x16, #0xb8]\n" ".inst 0x6f80e8e9 // udot v9.4s, v7.16b, v0.4b[2]\n" + "mov v6.d[1], x10\n" ".inst 0x6f81e8ed // udot v13.4s, v7.16b, v1.4b[2]\n" - "mov v6.d[1], x28\n" - "ldr d7, [x13, #0xb0]\n" - "ldr x28, [x13, #0xc8]\n" + "ldr d7, [x16, #0xb0]\n" ".inst 0x6f80e8ca // udot v10.4s, v6.16b, v0.4b[2]\n" + "ldr x10, [x16, #0xc8]\n" ".inst 0x6f81e8ce // udot v14.4s, v6.16b, v1.4b[2]\n" - "mov v7.d[1], x19\n" - "ldr d6, [x13, #0xc0]\n" - "ldr x19, [x13, #0xd8]\n" + "ldr d6, [x16, #0xc0]\n" + "mov v7.d[1], x11\n" + "ldr x11, [x16, #0xd8]\n" ".inst 0x6f80e8eb // udot v11.4s, v7.16b, v0.4b[2]\n" + "mov v6.d[1], x10\n" ".inst 0x6f81e8ef // udot v15.4s, v7.16b, v1.4b[2]\n" - "mov v6.d[1], x28\n" - "ldr d7, [x13, #0xd0]\n" - "ldr x28, [x13, #0xe8]\n" + "ldr d7, [x16, #0xd0]\n" ".inst 0x6fa0e8c8 // udot v8.4s, v6.16b, v0.4b[3]\n" + "ldr x10, [x16, #0xe8]\n" ".inst 0x6fa1e8cc // udot v12.4s, v6.16b, v1.4b[3]\n" - "mov v7.d[1], x19\n" - "ldr d6, [x13, #0xe0]\n" - "ldr x19, [x13, #0xf8]\n" + "ldr d6, [x16, #0xe0]\n" + "mov v7.d[1], x11\n" + "ldr x11, [x16, #0xf8]\n" ".inst 0x6fa0e8e9 // udot v9.4s, v7.16b, v0.4b[3]\n" + "mov v6.d[1], x10\n" ".inst 0x6fa1e8ed // udot v13.4s, v7.16b, v1.4b[3]\n" - "mov v6.d[1], x28\n" - "ldr d7, [x13, #0xf0]\n" - "ldr x27, [x9, #0x8]\n" - "add x13, x13, #0x100\n" + "ldr d7, [x16, #0xf0]\n" ".inst 0x6fa0e8ca // udot v10.4s, v6.16b, v0.4b[3]\n" + "add x16, x16, #0x100\n" ".inst 0x6fa1e8ce // udot v14.4s, v6.16b, v1.4b[3]\n" - "mov v7.d[1], x19\n" - "ldr x25, [x26, #0x8]\n" - "ldr d6, [x13, #0x0]\n" + "ldr d6, [x16, #0x0]\n" + "mov v7.d[1], x11\n" + "ldr x10, [x16, #0x8]\n" ".inst 0x6fa0e8eb // udot v11.4s, v7.16b, v0.4b[3]\n" + "ldr d0, [x12, #0x0]\n" ".inst 0x6fa1e8ef // udot v15.4s, v7.16b, v1.4b[3]\n" - "ldr d0, [x9, #0x0]\n" - "ldr d1, [x26, #0x0]\n" - "ldr x28, [x13, #0x8]\n" - "mov v0.d[1], x27\n" - "mov v1.d[1], x25\n" - "mov v6.d[1], x28\n" + "mov v6.d[1], x10\n" + "ldr d1, [x28, #0x0]\n" + "mov v0.d[1], x9\n" + "mov v1.d[1], x27\n" "bge 51b\n" "52:" // Height 2: Multiply loop: Single iteration only ".inst 0x6f80e0c8 // udot v8.4s, v6.16b, v0.4b[0]\n" + "ldr q7, [x16, #0x10]\n" ".inst 0x6f81e0cc // udot v12.4s, v6.16b, v1.4b[0]\n" - "ldr q7, [x13, #0x10]\n" - "sub x10, x10, #0x10\n" - "ldr q6, [x13, #0x20]\n" - "add x9, x9, #0x10\n" - "prfm pldl1keep, [x9, #0x80]\n" - "add x26, x26, #0x10\n" + "ldr q6, [x16, #0x20]\n" + "sub x13, x13, #0x10\n" + "add x12, x12, #0x10\n" ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" + "prfm pldl1keep, [x12, #0x80]\n" ".inst 0x6f81e0ed // udot v13.4s, v7.16b, v1.4b[0]\n" + "ldr q7, [x16, #0x30]\n" ".inst 0x6f80e0ca // udot v10.4s, v6.16b, v0.4b[0]\n" + "add x28, x28, #0x10\n" ".inst 0x6f81e0ce // udot v14.4s, v6.16b, v1.4b[0]\n" - "ldr q7, [x13, #0x30]\n" - "prfm pldl1keep, [x26, #0x80]\n" - "ldr q6, [x13, #0x40]\n" + "prfm pldl1keep, [x28, #0x80]\n" ".inst 0x6f80e0eb // udot v11.4s, v7.16b, v0.4b[0]\n" + "ldr q6, [x16, #0x40]\n" ".inst 0x6f81e0ef // udot v15.4s, v7.16b, v1.4b[0]\n" - "ldr q7, [x13, #0x50]\n" + "ldr q7, [x16, #0x50]\n" ".inst 0x6fa0e0c8 // udot v8.4s, v6.16b, v0.4b[1]\n" ".inst 0x6fa1e0cc // udot v12.4s, v6.16b, v1.4b[1]\n" - "ldr q6, [x13, #0x60]\n" + "ldr q6, [x16, #0x60]\n" ".inst 0x6fa0e0e9 // udot v9.4s, v7.16b, v0.4b[1]\n" ".inst 0x6fa1e0ed // udot v13.4s, v7.16b, v1.4b[1]\n" - "ldr q7, [x13, #0x70]\n" + "ldr q7, [x16, #0x70]\n" ".inst 0x6fa0e0ca // udot v10.4s, v6.16b, v0.4b[1]\n" ".inst 0x6fa1e0ce // udot v14.4s, v6.16b, v1.4b[1]\n" - "ldr q6, [x13, #0x80]\n" + "ldr q6, [x16, #0x80]\n" ".inst 0x6fa0e0eb // udot v11.4s, v7.16b, v0.4b[1]\n" ".inst 0x6fa1e0ef // udot v15.4s, v7.16b, v1.4b[1]\n" - "ldr q7, [x13, #0x90]\n" + "ldr q7, [x16, #0x90]\n" ".inst 0x6f80e8c8 // udot v8.4s, v6.16b, v0.4b[2]\n" ".inst 0x6f81e8cc // udot v12.4s, v6.16b, v1.4b[2]\n" - "ldr q6, [x13, #0xa0]\n" + "ldr q6, [x16, #0xa0]\n" ".inst 0x6f80e8e9 // udot v9.4s, v7.16b, v0.4b[2]\n" ".inst 0x6f81e8ed // udot v13.4s, v7.16b, v1.4b[2]\n" - "ldr q7, [x13, #0xb0]\n" + "ldr q7, [x16, #0xb0]\n" ".inst 0x6f80e8ca // udot v10.4s, v6.16b, v0.4b[2]\n" ".inst 0x6f81e8ce // udot v14.4s, v6.16b, v1.4b[2]\n" - "ldr q6, [x13, #0xc0]\n" + "ldr q6, [x16, #0xc0]\n" ".inst 0x6f80e8eb // udot v11.4s, v7.16b, v0.4b[2]\n" ".inst 0x6f81e8ef // udot v15.4s, v7.16b, v1.4b[2]\n" - "ldr q7, [x13, #0xd0]\n" + "ldr q7, [x16, #0xd0]\n" ".inst 0x6fa0e8c8 // udot v8.4s, v6.16b, v0.4b[3]\n" ".inst 0x6fa1e8cc // udot v12.4s, v6.16b, v1.4b[3]\n" - "ldr q6, [x13, #0xe0]\n" + "ldr q6, [x16, #0xe0]\n" ".inst 0x6fa0e8e9 // udot v9.4s, v7.16b, v0.4b[3]\n" ".inst 0x6fa1e8ed // udot v13.4s, v7.16b, v1.4b[3]\n" - "ldr q7, [x13, #0xf0]\n" - "add x13, x13, #0x100\n" + "ldr q7, [x16, #0xf0]\n" ".inst 0x6fa0e8ca // udot v10.4s, v6.16b, v0.4b[3]\n" + "add x16, x16, #0x100\n" ".inst 0x6fa1e8ce // udot v14.4s, v6.16b, v1.4b[3]\n" ".inst 0x6fa0e8eb // udot v11.4s, v7.16b, v0.4b[3]\n" ".inst 0x6fa1e8ef // udot v15.4s, v7.16b, v1.4b[3]\n" "53:" // Height 2: Multiply loop: Main loop skip - "cbz x10, 58f\n" - "cmp x10, #0x4\n" + "cbz x13, 58f\n" + "cmp x13, #0x4\n" "blt 55f\n" "54:" // Height 2: Multiply loop: Odd block loop - "ldr s0, [x9], #0x4\n" - "sub x10, x10, #0x4\n" - "ldr s1, [x26], #0x4\n" - "cmp x10, #0x4\n" - "ldr q6, [x13, #0x0]\n" - "ldr q7, [x13, #0x10]\n" + "ldr s0, [x12], #0x4\n" + "sub x13, x13, #0x4\n" + "ldr s1, [x28], #0x4\n" + "cmp x13, #0x4\n" + "ldr q6, [x16, #0x0]\n" + "ldr q7, [x16, #0x10]\n" ".inst 0x6f80e0c8 // udot v8.4s, v6.16b, v0.4b[0]\n" ".inst 0x6f81e0cc // udot v12.4s, v6.16b, v1.4b[0]\n" + "ldr q6, [x16, #0x20]\n" ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" ".inst 0x6f81e0ed // udot v13.4s, v7.16b, v1.4b[0]\n" - "ldr q6, [x13, #0x20]\n" - "ldr q7, [x13, #0x30]\n" - "add x13, x13, #0x40\n" + "ldr q7, [x16, #0x30]\n" ".inst 0x6f80e0ca // udot v10.4s, v6.16b, v0.4b[0]\n" + "add x16, x16, #0x40\n" ".inst 0x6f81e0ce // udot v14.4s, v6.16b, v1.4b[0]\n" ".inst 0x6f80e0eb // udot v11.4s, v7.16b, v0.4b[0]\n" ".inst 0x6f81e0ef // udot v15.4s, v7.16b, v1.4b[0]\n" "bge 54b\n" - "cbz x10, 58f\n" + "cbz x13, 58f\n" "55:" // Height 2: Multiply loop: Skip odd blocks - "tbz x10, #1, 56f\n" - "ldr h0, [x9], #0x2\n" - "ldr h1, [x26], #0x2\n" - "tbz x10, #0, 57f\n" - "ld1 { v0.b }[2], [x9]\n" - "ld1 { v1.b }[2], [x26]\n" + "tbz x13, #1, 56f\n" + "ldr h0, [x12], #0x2\n" + "ldr h1, [x28], #0x2\n" + "tbz x13, #0, 57f\n" + "ld1 { v0.b }[2], [x12]\n" + "ld1 { v1.b }[2], [x28]\n" "b 57f\n" "56:" // Height 2: Multiply loop: Ragged operand read: partial_1_0 - "ldr b0, [x9, #0x0]\n" - "ldr b1, [x26, #0x0]\n" + "ldr b0, [x12, #0x0]\n" + "ldr b1, [x28, #0x0]\n" "57:" // Height 2: Multiply loop: Ragged operand read: Done - "ldr q6, [x13, #0x0]\n" - "ldr q7, [x13, #0x10]\n" + "ldr q6, [x16, #0x0]\n" + "ldr q7, [x16, #0x10]\n" ".inst 0x6f80e0c8 // udot v8.4s, v6.16b, v0.4b[0]\n" ".inst 0x6f81e0cc // udot v12.4s, v6.16b, v1.4b[0]\n" + "ldr q6, [x16, #0x20]\n" ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" ".inst 0x6f81e0ed // udot v13.4s, v7.16b, v1.4b[0]\n" - "ldr q6, [x13, #0x20]\n" - "ldr q7, [x13, #0x30]\n" - "add x13, x13, #0x40\n" + "ldr q7, [x16, #0x30]\n" ".inst 0x6f80e0ca // udot v10.4s, v6.16b, v0.4b[0]\n" + "add x16, x16, #0x40\n" ".inst 0x6f81e0ce // udot v14.4s, v6.16b, v1.4b[0]\n" ".inst 0x6f80e0eb // udot v11.4s, v7.16b, v0.4b[0]\n" ".inst 0x6f81e0ef // udot v15.4s, v7.16b, v1.4b[0]\n" "58:" // Height 2: Multiply loop: No odd multiplies "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" - "add x11, x11, #0x1\n" - "cmp x11, x19\n" + "add x14, x14, #0x1\n" + "cmp x14, x19\n" "bne 48b\n" "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "prfm pstl1keep, [x12, #0x0]\n" - "cmp x14, #0x10\n" - "add x23, x12, x19, LSL #2\n" + "prfm pstl1keep, [x15, #0x0]\n" + "cmp x17, #0x10\n" + "add x23, x15, x19, LSL #2\n" "prfm pstl1keep, [x23, #0x0]\n" "bge 67f\n" - "tbz x14, #3, 62f\n" - "st1 { v8.4s }, [x12], #0x10\n" - "st1 { v9.4s }, [x12], #0x10\n" + "tbz x17, #3, 62f\n" + "st1 { v8.4s }, [x15], #0x10\n" + "st1 { v9.4s }, [x15], #0x10\n" "st1 { v12.4s }, [x23], #0x10\n" "st1 { v13.4s }, [x23], #0x10\n" - "tbz x14, #2, 60f\n" - "st1 { v10.4s }, [x12], #0x10\n" + "tbz x17, #2, 60f\n" + "st1 { v10.4s }, [x15], #0x10\n" "st1 { v14.4s }, [x23], #0x10\n" - "tbz x14, #1, 59f\n" - "str d11, [x12], #0x8\n" + "tbz x17, #1, 59f\n" + "str d11, [x15], #0x8\n" "str d15, [x23], #0x8\n" - "tbz x14, #0, 66f\n" - "st1 { v11.s }[2], [x12]\n" + "tbz x17, #0, 66f\n" + "st1 { v11.s }[2], [x15]\n" "st1 { v15.s }[2], [x23]\n" "b 66f\n" "59:" // Height 2: Partial direct writeback: partial_1_12 - "tbz x14, #0, 66f\n" - "str s11, [x12, #0x0]\n" + "tbz x17, #0, 66f\n" + "str s11, [x15, #0x0]\n" "str s15, [x23, #0x0]\n" "b 66f\n" "60:" // Height 2: Partial direct writeback: partial_2_8 - "tbz x14, #1, 61f\n" - "str d10, [x12], #0x8\n" + "tbz x17, #1, 61f\n" + "str d10, [x15], #0x8\n" "str d14, [x23], #0x8\n" - "tbz x14, #0, 66f\n" - "st1 { v10.s }[2], [x12]\n" + "tbz x17, #0, 66f\n" + "st1 { v10.s }[2], [x15]\n" "st1 { v14.s }[2], [x23]\n" "b 66f\n" "61:" // Height 2: Partial direct writeback: partial_1_8 - "tbz x14, #0, 66f\n" - "str s10, [x12, #0x0]\n" + "tbz x17, #0, 66f\n" + "str s10, [x15, #0x0]\n" "str s14, [x23, #0x0]\n" "b 66f\n" "62:" // Height 2: Partial direct writeback: partial_4_0 - "tbz x14, #2, 64f\n" - "st1 { v8.4s }, [x12], #0x10\n" + "tbz x17, #2, 64f\n" + "st1 { v8.4s }, [x15], #0x10\n" "st1 { v12.4s }, [x23], #0x10\n" - "tbz x14, #1, 63f\n" - "str d9, [x12], #0x8\n" + "tbz x17, #1, 63f\n" + "str d9, [x15], #0x8\n" "str d13, [x23], #0x8\n" - "tbz x14, #0, 66f\n" - "st1 { v9.s }[2], [x12]\n" + "tbz x17, #0, 66f\n" + "st1 { v9.s }[2], [x15]\n" "st1 { v13.s }[2], [x23]\n" "b 66f\n" "63:" // Height 2: Partial direct writeback: partial_1_4 - "tbz x14, #0, 66f\n" - "str s9, [x12, #0x0]\n" + "tbz x17, #0, 66f\n" + "str s9, [x15, #0x0]\n" "str s13, [x23, #0x0]\n" "b 66f\n" "64:" // Height 2: Partial direct writeback: partial_2_0 - "tbz x14, #1, 65f\n" - "str d8, [x12], #0x8\n" + "tbz x17, #1, 65f\n" + "str d8, [x15], #0x8\n" "str d12, [x23], #0x8\n" - "tbz x14, #0, 66f\n" - "st1 { v8.s }[2], [x12]\n" + "tbz x17, #0, 66f\n" + "st1 { v8.s }[2], [x15]\n" "st1 { v12.s }[2], [x23]\n" "b 66f\n" "65:" // Height 2: Partial direct writeback: partial_1_0 - "str s8, [x12, #0x0]\n" + "str s8, [x15, #0x0]\n" "str s12, [x23, #0x0]\n" "66:" // Height 2: Partial direct writeback: Done "b 68f\n" "67:" // Height 2: Full writeback - "str q8, [x12, #0x0]\n" - "str q9, [x12, #0x10]\n" - "str q10, [x12, #0x20]\n" - "str q11, [x12, #0x30]\n" - "add x12, x12, #0x40\n" + "str q8, [x15, #0x0]\n" + "str q9, [x15, #0x10]\n" + "str q10, [x15, #0x20]\n" + "str q11, [x15, #0x30]\n" + "add x15, x15, #0x40\n" "str q12, [x23, #0x0]\n" "str q13, [x23, #0x10]\n" "str q14, [x23, #0x20]\n" "str q15, [x23, #0x30]\n" "68:" // Height 2: Writeback done - "subs x14, x14, #0x10\n" + "subs x17, x17, #0x10\n" "bgt 36b\n" "b 206f\n" "69:" // Height 3 - "ldr x14, [%x[args_ptr], %[offsetof_N]]\n" - "mov x12, %x[output_ptr]\n" - "ldr x13, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "ldr x17, [%x[args_ptr], %[offsetof_N]]\n" + "mov x15, %x[output_ptr]\n" + "ldr x16, [%x[args_ptr], %[offsetof_B_ptr]]\n" "70:" // Height 3: Column loop "tbz %x[flags], #0, 80f\n" "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "cmp x14, #0x10\n" - "add x23, x12, x19, LSL #2\n" + "cmp x17, #0x10\n" + "add x23, x15, x19, LSL #2\n" "add x22, x23, x19, LSL #2\n" "bge 79f\n" - "tbz x14, #3, 74f\n" - "ld1 { v8.4s }, [x12], #0x10\n" + "tbz x17, #3, 74f\n" + "ld1 { v8.4s }, [x15], #0x10\n" "ld1 { v12.4s }, [x23], #0x10\n" "ld1 { v16.4s }, [x22], #0x10\n" - "ld1 { v9.4s }, [x12], #0x10\n" + "ld1 { v9.4s }, [x15], #0x10\n" "ld1 { v13.4s }, [x23], #0x10\n" "ld1 { v17.4s }, [x22], #0x10\n" - "tbz x14, #2, 72f\n" - "ld1 { v10.4s }, [x12], #0x10\n" + "tbz x17, #2, 72f\n" + "ld1 { v10.4s }, [x15], #0x10\n" "ld1 { v14.4s }, [x23], #0x10\n" "ld1 { v18.4s }, [x22], #0x10\n" - "tbz x14, #1, 71f\n" - "ldr d11, [x12], #0x8\n" + "tbz x17, #1, 71f\n" + "ldr d11, [x15], #0x8\n" "mov x24, #0x38\n" "ldr d15, [x23], #0x8\n" "ldr d19, [x22], #0x8\n" - "tbz x14, #0, 78f\n" - "ld1 { v11.s }[2], [x12]\n" + "tbz x17, #0, 78f\n" + "ld1 { v11.s }[2], [x15]\n" "ld1 { v15.s }[2], [x23]\n" "ld1 { v19.s }[2], [x22]\n" "b 78f\n" "71:" // Height 3: Partial accumulate: partial_1_12 "mov x24, #0x30\n" - "tbz x14, #0, 78f\n" - "ldr s11, [x12, #0x0]\n" + "tbz x17, #0, 78f\n" + "ldr s11, [x15, #0x0]\n" "ldr s15, [x23, #0x0]\n" "ldr s19, [x22, #0x0]\n" "b 78f\n" "72:" // Height 3: Partial accumulate: partial_2_8 - "tbz x14, #1, 73f\n" - "ldr d10, [x12], #0x8\n" + "tbz x17, #1, 73f\n" + "ldr d10, [x15], #0x8\n" "ldr d14, [x23], #0x8\n" "mov x24, #0x28\n" "ldr d18, [x22], #0x8\n" - "tbz x14, #0, 78f\n" - "ld1 { v10.s }[2], [x12]\n" + "tbz x17, #0, 78f\n" + "ld1 { v10.s }[2], [x15]\n" "ld1 { v14.s }[2], [x23]\n" "ld1 { v18.s }[2], [x22]\n" "b 78f\n" "73:" // Height 3: Partial accumulate: partial_1_8 "mov x24, #0x20\n" - "tbz x14, #0, 78f\n" - "ldr s10, [x12, #0x0]\n" + "tbz x17, #0, 78f\n" + "ldr s10, [x15, #0x0]\n" "ldr s14, [x23, #0x0]\n" "ldr s18, [x22, #0x0]\n" "b 78f\n" "74:" // Height 3: Partial accumulate: partial_4_0 - "tbz x14, #2, 76f\n" - "ld1 { v8.4s }, [x12], #0x10\n" + "tbz x17, #2, 76f\n" + "ld1 { v8.4s }, [x15], #0x10\n" "ld1 { v12.4s }, [x23], #0x10\n" "ld1 { v16.4s }, [x22], #0x10\n" - "tbz x14, #1, 75f\n" - "ldr d9, [x12], #0x8\n" + "tbz x17, #1, 75f\n" + "ldr d9, [x15], #0x8\n" "mov x24, #0x18\n" "ldr d13, [x23], #0x8\n" "ldr d17, [x22], #0x8\n" - "tbz x14, #0, 78f\n" - "ld1 { v9.s }[2], [x12]\n" + "tbz x17, #0, 78f\n" + "ld1 { v9.s }[2], [x15]\n" "ld1 { v13.s }[2], [x23]\n" "ld1 { v17.s }[2], [x22]\n" "b 78f\n" "75:" // Height 3: Partial accumulate: partial_1_4 "mov x24, #0x10\n" - "tbz x14, #0, 78f\n" - "ldr s9, [x12, #0x0]\n" + "tbz x17, #0, 78f\n" + "ldr s9, [x15, #0x0]\n" "ldr s13, [x23, #0x0]\n" "ldr s17, [x22, #0x0]\n" "b 78f\n" "76:" // Height 3: Partial accumulate: partial_2_0 - "tbz x14, #1, 77f\n" - "ldr d8, [x12], #0x8\n" + "tbz x17, #1, 77f\n" + "ldr d8, [x15], #0x8\n" "ldr d12, [x23], #0x8\n" "mov x24, #0x8\n" "ldr d16, [x22], #0x8\n" - "tbz x14, #0, 78f\n" - "ld1 { v8.s }[2], [x12]\n" + "tbz x17, #0, 78f\n" + "ld1 { v8.s }[2], [x15]\n" "ld1 { v12.s }[2], [x23]\n" "ld1 { v16.s }[2], [x22]\n" "b 78f\n" "77:" // Height 3: Partial accumulate: partial_1_0 - "ldr s8, [x12, #0x0]\n" + "ldr s8, [x15, #0x0]\n" "mov x24, #0x0\n" "ldr s12, [x23, #0x0]\n" "ldr s16, [x22, #0x0]\n" "78:" // Height 3: Partial accumulate: Done - "sub x12, x12, x24\n" + "sub x15, x15, x24\n" "b 81f\n" "79:" // Height 3: full accumulate - "ldr q8, [x12, #0x0]\n" - "ldr q9, [x12, #0x10]\n" - "ldr q10, [x12, #0x20]\n" - "ldr q11, [x12, #0x30]\n" + "ldr q8, [x15, #0x0]\n" + "ldr q9, [x15, #0x10]\n" + "ldr q10, [x15, #0x20]\n" + "ldr q11, [x15, #0x30]\n" "ldr q12, [x23, #0x0]\n" "ldr q13, [x23, #0x10]\n" "ldr q14, [x23, #0x20]\n" @@ -932,280 +932,280 @@ void a64_hybrid_u8u32_dot_6x16_a55 ( "movi v18.4s, #0x0\n" "movi v19.4s, #0x0\n" "81:" // Height 3: setup done - "mov x11, #0x0\n" + "mov x14, #0x0\n" "82:" // Height 3: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w10, [x20, x11, LSL #0x2]\n" + "ldr w13, [x20, x14, LSL #0x2]\n" "tbz %x[flags], #3, 83f\n" - "ldr x20, [%x[input_ptr], x11, LSL #0x3]\n" + "ldr x20, [%x[input_ptr], x14, LSL #0x3]\n" "add x20, x20, x19, LSL #3\n" - "ldr x9, [x20, #0x0]\n" - "ldr x26, [x20, #0x8]\n" - "ldr x24, [x20, #0x10]\n" - "cbnz x11, 84f\n" + "ldr x12, [x20, #0x0]\n" + "ldr x28, [x20, #0x8]\n" + "ldr x26, [x20, #0x10]\n" + "cbnz x14, 84f\n" "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x9, x9, x19\n" + "add x12, x12, x19\n" + "add x28, x28, x19\n" "add x26, x26, x19\n" - "add x24, x24, x19\n" "b 84f\n" "83:" // Height 3: setup direct input - "mov x9, %x[input_ptr]\n" - "add x26, x9, x19\n" - "add x24, x26, x19\n" + "mov x12, %x[input_ptr]\n" + "add x28, x12, x19\n" + "add x26, x28, x19\n" "84:" // Height 3: input setup done - "cmp x10, #0x10\n" + "cmp x13, #0x10\n" "blt 87f\n" - "ldr q0, [x9, #0x0]\n" - "ldr q1, [x26, #0x0]\n" - "cmp x10, #0x20\n" - "ldr q2, [x24, #0x0]\n" - "ldr q6, [x13, #0x0]\n" + "ldr q0, [x12, #0x0]\n" + "ldr q1, [x28, #0x0]\n" + "cmp x13, #0x20\n" + "ldr q2, [x26, #0x0]\n" + "ldr q6, [x16, #0x0]\n" "blt 86f\n" "85:" // Height 3: Multiply loop: Main loop head ".inst 0x6f80e0c8 // udot v8.4s, v6.16b, v0.4b[0]\n" + "ldr d7, [x16, #0x10]\n" ".inst 0x6f81e0cc // udot v12.4s, v6.16b, v1.4b[0]\n" + "ldr x11, [x16, #0x18]\n" ".inst 0x6f82e0d0 // udot v16.4s, v6.16b, v2.4b[0]\n" - "ldr d7, [x13, #0x10]\n" - "ldr x19, [x13, #0x18]\n" - "add x9, x9, #0x10\n" - "ldr d6, [x13, #0x20]\n" - "add x26, x26, #0x10\n" - "ldr x28, [x13, #0x28]\n" - "add x24, x24, #0x10\n" - "mov v7.d[1], x19\n" - "prfm pldl1keep, [x9, #0x80]\n" - "ldr x19, [x13, #0x38]\n" - "sub x10, x10, #0x10\n" - "mov v6.d[1], x28\n" + "ldr d6, [x16, #0x20]\n" + "ldr x10, [x16, #0x28]\n" + "add x12, x12, #0x10\n" + "mov v7.d[1], x11\n" + "prfm pldl1keep, [x12, #0x80]\n" + "ldr x11, [x16, #0x38]\n" + "add x28, x28, #0x10\n" ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" + "mov v6.d[1], x10\n" ".inst 0x6f81e0ed // udot v13.4s, v7.16b, v1.4b[0]\n" + "prfm pldl1keep, [x28, #0x80]\n" ".inst 0x6f82e0f1 // udot v17.4s, v7.16b, v2.4b[0]\n" + "ldr d7, [x16, #0x30]\n" ".inst 0x6f80e0ca // udot v10.4s, v6.16b, v0.4b[0]\n" + "ldr x10, [x16, #0x48]\n" ".inst 0x6f81e0ce // udot v14.4s, v6.16b, v1.4b[0]\n" + "ldr x9, [x12, #0x8]\n" ".inst 0x6f82e0d2 // udot v18.4s, v6.16b, v2.4b[0]\n" - "ldr d7, [x13, #0x30]\n" - "prfm pldl1keep, [x26, #0x80]\n" - "cmp x10, #0x20\n" - "ldr d6, [x13, #0x40]\n" - "mov v7.d[1], x19\n" - "ldr x28, [x13, #0x48]\n" - "ldr x19, [x13, #0x58]\n" + "mov v7.d[1], x11\n" + "ldr d6, [x16, #0x40]\n" + "add x26, x26, #0x10\n" ".inst 0x6f80e0eb // udot v11.4s, v7.16b, v0.4b[0]\n" + "prfm pldl1keep, [x26, #0x80]\n" ".inst 0x6f81e0ef // udot v15.4s, v7.16b, v1.4b[0]\n" - "mov v6.d[1], x28\n" + "ldr x11, [x16, #0x58]\n" ".inst 0x6f82e0f3 // udot v19.4s, v7.16b, v2.4b[0]\n" - "ldr d7, [x13, #0x50]\n" + "mov v6.d[1], x10\n" + "ldr d7, [x16, #0x50]\n" + "sub x13, x13, #0x10\n" ".inst 0x6fa0e0c8 // udot v8.4s, v6.16b, v0.4b[1]\n" + "ldr x10, [x16, #0x68]\n" ".inst 0x6fa1e0cc // udot v12.4s, v6.16b, v1.4b[1]\n" + "ldr x27, [x28, #0x8]\n" ".inst 0x6fa2e0d0 // udot v16.4s, v6.16b, v2.4b[1]\n" - "prfm pldl1keep, [x24, #0x80]\n" - "mov v7.d[1], x19\n" - "ldr d6, [x13, #0x60]\n" - "ldr x28, [x13, #0x68]\n" + "mov v7.d[1], x11\n" + "ldr d6, [x16, #0x60]\n" + "cmp x13, #0x20\n" ".inst 0x6fa0e0e9 // udot v9.4s, v7.16b, v0.4b[1]\n" + "ldr x11, [x16, #0x78]\n" ".inst 0x6fa1e0ed // udot v13.4s, v7.16b, v1.4b[1]\n" + "ldr x25, [x26, #0x8]\n" ".inst 0x6fa2e0f1 // udot v17.4s, v7.16b, v2.4b[1]\n" - "ldr d7, [x13, #0x70]\n" - "mov v6.d[1], x28\n" - "ldr x19, [x13, #0x78]\n" - "ldr x28, [x13, #0x88]\n" + "mov v6.d[1], x10\n" + "ldr d7, [x16, #0x70]\n" ".inst 0x6fa0e0ca // udot v10.4s, v6.16b, v0.4b[1]\n" + "ldr x10, [x16, #0x88]\n" ".inst 0x6fa1e0ce // udot v14.4s, v6.16b, v1.4b[1]\n" - "mov v7.d[1], x19\n" ".inst 0x6fa2e0d2 // udot v18.4s, v6.16b, v2.4b[1]\n" - "ldr d6, [x13, #0x80]\n" + "mov v7.d[1], x11\n" + "ldr d6, [x16, #0x80]\n" ".inst 0x6fa0e0eb // udot v11.4s, v7.16b, v0.4b[1]\n" + "ldr x11, [x16, #0x98]\n" ".inst 0x6fa1e0ef // udot v15.4s, v7.16b, v1.4b[1]\n" ".inst 0x6fa2e0f3 // udot v19.4s, v7.16b, v2.4b[1]\n" - "ldr d7, [x13, #0x90]\n" - "mov v6.d[1], x28\n" - "ldr x19, [x13, #0x98]\n" - "ldr x28, [x13, #0xa8]\n" + "mov v6.d[1], x10\n" + "ldr d7, [x16, #0x90]\n" ".inst 0x6f80e8c8 // udot v8.4s, v6.16b, v0.4b[2]\n" + "ldr x10, [x16, #0xa8]\n" ".inst 0x6f81e8cc // udot v12.4s, v6.16b, v1.4b[2]\n" - "mov v7.d[1], x19\n" ".inst 0x6f82e8d0 // udot v16.4s, v6.16b, v2.4b[2]\n" - "ldr d6, [x13, #0xa0]\n" + "mov v7.d[1], x11\n" + "ldr d6, [x16, #0xa0]\n" ".inst 0x6f80e8e9 // udot v9.4s, v7.16b, v0.4b[2]\n" + "ldr x11, [x16, #0xb8]\n" ".inst 0x6f81e8ed // udot v13.4s, v7.16b, v1.4b[2]\n" ".inst 0x6f82e8f1 // udot v17.4s, v7.16b, v2.4b[2]\n" - "ldr d7, [x13, #0xb0]\n" - "mov v6.d[1], x28\n" - "ldr x19, [x13, #0xb8]\n" - "ldr x28, [x13, #0xc8]\n" + "mov v6.d[1], x10\n" + "ldr d7, [x16, #0xb0]\n" ".inst 0x6f80e8ca // udot v10.4s, v6.16b, v0.4b[2]\n" + "ldr x10, [x16, #0xc8]\n" ".inst 0x6f81e8ce // udot v14.4s, v6.16b, v1.4b[2]\n" - "mov v7.d[1], x19\n" ".inst 0x6f82e8d2 // udot v18.4s, v6.16b, v2.4b[2]\n" - "ldr d6, [x13, #0xc0]\n" + "mov v7.d[1], x11\n" + "ldr d6, [x16, #0xc0]\n" ".inst 0x6f80e8eb // udot v11.4s, v7.16b, v0.4b[2]\n" + "ldr x11, [x16, #0xd8]\n" ".inst 0x6f81e8ef // udot v15.4s, v7.16b, v1.4b[2]\n" ".inst 0x6f82e8f3 // udot v19.4s, v7.16b, v2.4b[2]\n" - "ldr d7, [x13, #0xd0]\n" - "mov v6.d[1], x28\n" - "ldr x19, [x13, #0xd8]\n" - "ldr x28, [x13, #0xe8]\n" + "mov v6.d[1], x10\n" + "ldr d7, [x16, #0xd0]\n" ".inst 0x6fa0e8c8 // udot v8.4s, v6.16b, v0.4b[3]\n" + "ldr x10, [x16, #0xe8]\n" ".inst 0x6fa1e8cc // udot v12.4s, v6.16b, v1.4b[3]\n" - "mov v7.d[1], x19\n" ".inst 0x6fa2e8d0 // udot v16.4s, v6.16b, v2.4b[3]\n" - "ldr d6, [x13, #0xe0]\n" + "mov v7.d[1], x11\n" + "ldr d6, [x16, #0xe0]\n" ".inst 0x6fa0e8e9 // udot v9.4s, v7.16b, v0.4b[3]\n" + "ldr x11, [x16, #0xf8]\n" ".inst 0x6fa1e8ed // udot v13.4s, v7.16b, v1.4b[3]\n" ".inst 0x6fa2e8f1 // udot v17.4s, v7.16b, v2.4b[3]\n" - "ldr d7, [x13, #0xf0]\n" - "mov v6.d[1], x28\n" - "ldr x19, [x13, #0xf8]\n" - "ldr x27, [x9, #0x8]\n" - "add x13, x13, #0x100\n" + "mov v6.d[1], x10\n" + "ldr d7, [x16, #0xf0]\n" + "add x16, x16, #0x100\n" ".inst 0x6fa0e8ca // udot v10.4s, v6.16b, v0.4b[3]\n" + "ldr x10, [x16, #0x8]\n" ".inst 0x6fa1e8ce // udot v14.4s, v6.16b, v1.4b[3]\n" - "mov v7.d[1], x19\n" ".inst 0x6fa2e8d2 // udot v18.4s, v6.16b, v2.4b[3]\n" - "ldr x25, [x26, #0x8]\n" + "mov v7.d[1], x11\n" + "ldr d6, [x16, #0x0]\n" ".inst 0x6fa0e8eb // udot v11.4s, v7.16b, v0.4b[3]\n" + "ldr d0, [x12, #0x0]\n" ".inst 0x6fa1e8ef // udot v15.4s, v7.16b, v1.4b[3]\n" + "ldr d1, [x28, #0x0]\n" ".inst 0x6fa2e8f3 // udot v19.4s, v7.16b, v2.4b[3]\n" - "ldr d0, [x9, #0x0]\n" - "ldr d1, [x26, #0x0]\n" - "ldr d2, [x24, #0x0]\n" - "mov v0.d[1], x27\n" - "ldr x21, [x24, #0x8]\n" - "mov v1.d[1], x25\n" - "ldr d6, [x13, #0x0]\n" - "ldr x28, [x13, #0x8]\n" - "mov v2.d[1], x21\n" - "mov v6.d[1], x28\n" + "mov v6.d[1], x10\n" + "mov v0.d[1], x9\n" + "ldr d2, [x26, #0x0]\n" + "mov v1.d[1], x27\n" + "mov v2.d[1], x25\n" "bge 85b\n" "86:" // Height 3: Multiply loop: Single iteration only ".inst 0x6f80e0c8 // udot v8.4s, v6.16b, v0.4b[0]\n" + "ldr q7, [x16, #0x10]\n" ".inst 0x6f81e0cc // udot v12.4s, v6.16b, v1.4b[0]\n" + "sub x13, x13, #0x10\n" ".inst 0x6f82e0d0 // udot v16.4s, v6.16b, v2.4b[0]\n" - "ldr q7, [x13, #0x10]\n" - "ldr q6, [x13, #0x20]\n" - "sub x10, x10, #0x10\n" - "add x9, x9, #0x10\n" - "add x26, x26, #0x10\n" + "ldr q6, [x16, #0x20]\n" ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" + "add x12, x12, #0x10\n" ".inst 0x6f81e0ed // udot v13.4s, v7.16b, v1.4b[0]\n" + "prfm pldl1keep, [x12, #0x80]\n" ".inst 0x6f82e0f1 // udot v17.4s, v7.16b, v2.4b[0]\n" + "ldr q7, [x16, #0x30]\n" ".inst 0x6f80e0ca // udot v10.4s, v6.16b, v0.4b[0]\n" + "add x28, x28, #0x10\n" ".inst 0x6f81e0ce // udot v14.4s, v6.16b, v1.4b[0]\n" + "prfm pldl1keep, [x28, #0x80]\n" ".inst 0x6f82e0d2 // udot v18.4s, v6.16b, v2.4b[0]\n" - "prfm pldl1keep, [x9, #0x80]\n" - "add x24, x24, #0x10\n" - "ldr q7, [x13, #0x30]\n" - "prfm pldl1keep, [x26, #0x80]\n" - "ldr q6, [x13, #0x40]\n" + "ldr q6, [x16, #0x40]\n" ".inst 0x6f80e0eb // udot v11.4s, v7.16b, v0.4b[0]\n" + "add x26, x26, #0x10\n" ".inst 0x6f81e0ef // udot v15.4s, v7.16b, v1.4b[0]\n" + "prfm pldl1keep, [x26, #0x80]\n" ".inst 0x6f82e0f3 // udot v19.4s, v7.16b, v2.4b[0]\n" - "ldr q7, [x13, #0x50]\n" + "ldr q7, [x16, #0x50]\n" ".inst 0x6fa0e0c8 // udot v8.4s, v6.16b, v0.4b[1]\n" ".inst 0x6fa1e0cc // udot v12.4s, v6.16b, v1.4b[1]\n" ".inst 0x6fa2e0d0 // udot v16.4s, v6.16b, v2.4b[1]\n" - "prfm pldl1keep, [x24, #0x80]\n" + "ldr q6, [x16, #0x60]\n" ".inst 0x6fa0e0e9 // udot v9.4s, v7.16b, v0.4b[1]\n" ".inst 0x6fa1e0ed // udot v13.4s, v7.16b, v1.4b[1]\n" ".inst 0x6fa2e0f1 // udot v17.4s, v7.16b, v2.4b[1]\n" - "ldr q6, [x13, #0x60]\n" - "ldr q7, [x13, #0x70]\n" + "ldr q7, [x16, #0x70]\n" ".inst 0x6fa0e0ca // udot v10.4s, v6.16b, v0.4b[1]\n" ".inst 0x6fa1e0ce // udot v14.4s, v6.16b, v1.4b[1]\n" ".inst 0x6fa2e0d2 // udot v18.4s, v6.16b, v2.4b[1]\n" + "ldr q6, [x16, #0x80]\n" ".inst 0x6fa0e0eb // udot v11.4s, v7.16b, v0.4b[1]\n" ".inst 0x6fa1e0ef // udot v15.4s, v7.16b, v1.4b[1]\n" ".inst 0x6fa2e0f3 // udot v19.4s, v7.16b, v2.4b[1]\n" - "ldr q6, [x13, #0x80]\n" - "ldr q7, [x13, #0x90]\n" + "ldr q7, [x16, #0x90]\n" ".inst 0x6f80e8c8 // udot v8.4s, v6.16b, v0.4b[2]\n" ".inst 0x6f81e8cc // udot v12.4s, v6.16b, v1.4b[2]\n" ".inst 0x6f82e8d0 // udot v16.4s, v6.16b, v2.4b[2]\n" + "ldr q6, [x16, #0xa0]\n" ".inst 0x6f80e8e9 // udot v9.4s, v7.16b, v0.4b[2]\n" ".inst 0x6f81e8ed // udot v13.4s, v7.16b, v1.4b[2]\n" ".inst 0x6f82e8f1 // udot v17.4s, v7.16b, v2.4b[2]\n" - "ldr q6, [x13, #0xa0]\n" - "ldr q7, [x13, #0xb0]\n" + "ldr q7, [x16, #0xb0]\n" ".inst 0x6f80e8ca // udot v10.4s, v6.16b, v0.4b[2]\n" ".inst 0x6f81e8ce // udot v14.4s, v6.16b, v1.4b[2]\n" ".inst 0x6f82e8d2 // udot v18.4s, v6.16b, v2.4b[2]\n" + "ldr q6, [x16, #0xc0]\n" ".inst 0x6f80e8eb // udot v11.4s, v7.16b, v0.4b[2]\n" ".inst 0x6f81e8ef // udot v15.4s, v7.16b, v1.4b[2]\n" ".inst 0x6f82e8f3 // udot v19.4s, v7.16b, v2.4b[2]\n" - "ldr q6, [x13, #0xc0]\n" - "ldr q7, [x13, #0xd0]\n" + "ldr q7, [x16, #0xd0]\n" ".inst 0x6fa0e8c8 // udot v8.4s, v6.16b, v0.4b[3]\n" ".inst 0x6fa1e8cc // udot v12.4s, v6.16b, v1.4b[3]\n" ".inst 0x6fa2e8d0 // udot v16.4s, v6.16b, v2.4b[3]\n" + "ldr q6, [x16, #0xe0]\n" ".inst 0x6fa0e8e9 // udot v9.4s, v7.16b, v0.4b[3]\n" ".inst 0x6fa1e8ed // udot v13.4s, v7.16b, v1.4b[3]\n" ".inst 0x6fa2e8f1 // udot v17.4s, v7.16b, v2.4b[3]\n" - "ldr q6, [x13, #0xe0]\n" - "ldr q7, [x13, #0xf0]\n" - "add x13, x13, #0x100\n" + "ldr q7, [x16, #0xf0]\n" ".inst 0x6fa0e8ca // udot v10.4s, v6.16b, v0.4b[3]\n" + "add x16, x16, #0x100\n" ".inst 0x6fa1e8ce // udot v14.4s, v6.16b, v1.4b[3]\n" ".inst 0x6fa2e8d2 // udot v18.4s, v6.16b, v2.4b[3]\n" ".inst 0x6fa0e8eb // udot v11.4s, v7.16b, v0.4b[3]\n" ".inst 0x6fa1e8ef // udot v15.4s, v7.16b, v1.4b[3]\n" ".inst 0x6fa2e8f3 // udot v19.4s, v7.16b, v2.4b[3]\n" "87:" // Height 3: Multiply loop: Main loop skip - "cbz x10, 92f\n" - "cmp x10, #0x4\n" + "cbz x13, 92f\n" + "cmp x13, #0x4\n" "blt 89f\n" "88:" // Height 3: Multiply loop: Odd block loop - "ldr s0, [x9], #0x4\n" - "sub x10, x10, #0x4\n" - "ldr s1, [x26], #0x4\n" - "cmp x10, #0x4\n" - "ldr s2, [x24], #0x4\n" - "ldr q6, [x13, #0x0]\n" - "ldr q7, [x13, #0x10]\n" + "ldr s0, [x12], #0x4\n" + "sub x13, x13, #0x4\n" + "ldr s1, [x28], #0x4\n" + "cmp x13, #0x4\n" + "ldr s2, [x26], #0x4\n" + "ldr q6, [x16, #0x0]\n" + "ldr q7, [x16, #0x10]\n" ".inst 0x6f80e0c8 // udot v8.4s, v6.16b, v0.4b[0]\n" ".inst 0x6f81e0cc // udot v12.4s, v6.16b, v1.4b[0]\n" ".inst 0x6f82e0d0 // udot v16.4s, v6.16b, v2.4b[0]\n" + "ldr q6, [x16, #0x20]\n" ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" ".inst 0x6f81e0ed // udot v13.4s, v7.16b, v1.4b[0]\n" ".inst 0x6f82e0f1 // udot v17.4s, v7.16b, v2.4b[0]\n" - "ldr q6, [x13, #0x20]\n" - "ldr q7, [x13, #0x30]\n" - "add x13, x13, #0x40\n" + "ldr q7, [x16, #0x30]\n" ".inst 0x6f80e0ca // udot v10.4s, v6.16b, v0.4b[0]\n" + "add x16, x16, #0x40\n" ".inst 0x6f81e0ce // udot v14.4s, v6.16b, v1.4b[0]\n" ".inst 0x6f82e0d2 // udot v18.4s, v6.16b, v2.4b[0]\n" ".inst 0x6f80e0eb // udot v11.4s, v7.16b, v0.4b[0]\n" ".inst 0x6f81e0ef // udot v15.4s, v7.16b, v1.4b[0]\n" ".inst 0x6f82e0f3 // udot v19.4s, v7.16b, v2.4b[0]\n" "bge 88b\n" - "cbz x10, 92f\n" + "cbz x13, 92f\n" "89:" // Height 3: Multiply loop: Skip odd blocks - "tbz x10, #1, 90f\n" - "ldr h0, [x9], #0x2\n" - "ldr h1, [x26], #0x2\n" - "ldr h2, [x24], #0x2\n" - "tbz x10, #0, 91f\n" - "ld1 { v0.b }[2], [x9]\n" - "ld1 { v1.b }[2], [x26]\n" - "ld1 { v2.b }[2], [x24]\n" + "tbz x13, #1, 90f\n" + "ldr h0, [x12], #0x2\n" + "ldr h1, [x28], #0x2\n" + "ldr h2, [x26], #0x2\n" + "tbz x13, #0, 91f\n" + "ld1 { v0.b }[2], [x12]\n" + "ld1 { v1.b }[2], [x28]\n" + "ld1 { v2.b }[2], [x26]\n" "b 91f\n" "90:" // Height 3: Multiply loop: Ragged operand read: partial_1_0 - "ldr b0, [x9, #0x0]\n" - "ldr b1, [x26, #0x0]\n" - "ldr b2, [x24, #0x0]\n" + "ldr b0, [x12, #0x0]\n" + "ldr b1, [x28, #0x0]\n" + "ldr b2, [x26, #0x0]\n" "91:" // Height 3: Multiply loop: Ragged operand read: Done - "ldr q6, [x13, #0x0]\n" - "ldr q7, [x13, #0x10]\n" + "ldr q6, [x16, #0x0]\n" + "ldr q7, [x16, #0x10]\n" ".inst 0x6f80e0c8 // udot v8.4s, v6.16b, v0.4b[0]\n" ".inst 0x6f81e0cc // udot v12.4s, v6.16b, v1.4b[0]\n" ".inst 0x6f82e0d0 // udot v16.4s, v6.16b, v2.4b[0]\n" + "ldr q6, [x16, #0x20]\n" ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" ".inst 0x6f81e0ed // udot v13.4s, v7.16b, v1.4b[0]\n" ".inst 0x6f82e0f1 // udot v17.4s, v7.16b, v2.4b[0]\n" - "ldr q6, [x13, #0x20]\n" - "ldr q7, [x13, #0x30]\n" - "add x13, x13, #0x40\n" + "ldr q7, [x16, #0x30]\n" ".inst 0x6f80e0ca // udot v10.4s, v6.16b, v0.4b[0]\n" + "add x16, x16, #0x40\n" ".inst 0x6f81e0ce // udot v14.4s, v6.16b, v1.4b[0]\n" ".inst 0x6f82e0d2 // udot v18.4s, v6.16b, v2.4b[0]\n" ".inst 0x6f80e0eb // udot v11.4s, v7.16b, v0.4b[0]\n" @@ -1213,101 +1213,101 @@ void a64_hybrid_u8u32_dot_6x16_a55 ( ".inst 0x6f82e0f3 // udot v19.4s, v7.16b, v2.4b[0]\n" "92:" // Height 3: Multiply loop: No odd multiplies "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" - "add x11, x11, #0x1\n" - "cmp x11, x19\n" + "add x14, x14, #0x1\n" + "cmp x14, x19\n" "bne 82b\n" "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "prfm pstl1keep, [x12, #0x0]\n" - "cmp x14, #0x10\n" - "add x23, x12, x19, LSL #2\n" + "prfm pstl1keep, [x15, #0x0]\n" + "cmp x17, #0x10\n" + "add x23, x15, x19, LSL #2\n" "prfm pstl1keep, [x23, #0x0]\n" "add x22, x23, x19, LSL #2\n" "prfm pstl1keep, [x22, #0x0]\n" "bge 101f\n" - "tbz x14, #3, 96f\n" - "st1 { v8.4s }, [x12], #0x10\n" - "st1 { v9.4s }, [x12], #0x10\n" + "tbz x17, #3, 96f\n" + "st1 { v8.4s }, [x15], #0x10\n" + "st1 { v9.4s }, [x15], #0x10\n" "st1 { v12.4s }, [x23], #0x10\n" "st1 { v13.4s }, [x23], #0x10\n" "st1 { v16.4s }, [x22], #0x10\n" "st1 { v17.4s }, [x22], #0x10\n" - "tbz x14, #2, 94f\n" - "st1 { v10.4s }, [x12], #0x10\n" + "tbz x17, #2, 94f\n" + "st1 { v10.4s }, [x15], #0x10\n" "st1 { v14.4s }, [x23], #0x10\n" "st1 { v18.4s }, [x22], #0x10\n" - "tbz x14, #1, 93f\n" - "str d11, [x12], #0x8\n" + "tbz x17, #1, 93f\n" + "str d11, [x15], #0x8\n" "str d15, [x23], #0x8\n" "str d19, [x22], #0x8\n" - "tbz x14, #0, 100f\n" - "st1 { v11.s }[2], [x12]\n" + "tbz x17, #0, 100f\n" + "st1 { v11.s }[2], [x15]\n" "st1 { v15.s }[2], [x23]\n" "st1 { v19.s }[2], [x22]\n" "b 100f\n" "93:" // Height 3: Partial direct writeback: partial_1_12 - "tbz x14, #0, 100f\n" - "str s11, [x12, #0x0]\n" + "tbz x17, #0, 100f\n" + "str s11, [x15, #0x0]\n" "str s15, [x23, #0x0]\n" "str s19, [x22, #0x0]\n" "b 100f\n" "94:" // Height 3: Partial direct writeback: partial_2_8 - "tbz x14, #1, 95f\n" - "str d10, [x12], #0x8\n" + "tbz x17, #1, 95f\n" + "str d10, [x15], #0x8\n" "str d14, [x23], #0x8\n" "str d18, [x22], #0x8\n" - "tbz x14, #0, 100f\n" - "st1 { v10.s }[2], [x12]\n" + "tbz x17, #0, 100f\n" + "st1 { v10.s }[2], [x15]\n" "st1 { v14.s }[2], [x23]\n" "st1 { v18.s }[2], [x22]\n" "b 100f\n" "95:" // Height 3: Partial direct writeback: partial_1_8 - "tbz x14, #0, 100f\n" - "str s10, [x12, #0x0]\n" + "tbz x17, #0, 100f\n" + "str s10, [x15, #0x0]\n" "str s14, [x23, #0x0]\n" "str s18, [x22, #0x0]\n" "b 100f\n" "96:" // Height 3: Partial direct writeback: partial_4_0 - "tbz x14, #2, 98f\n" - "st1 { v8.4s }, [x12], #0x10\n" + "tbz x17, #2, 98f\n" + "st1 { v8.4s }, [x15], #0x10\n" "st1 { v12.4s }, [x23], #0x10\n" "st1 { v16.4s }, [x22], #0x10\n" - "tbz x14, #1, 97f\n" - "str d9, [x12], #0x8\n" + "tbz x17, #1, 97f\n" + "str d9, [x15], #0x8\n" "str d13, [x23], #0x8\n" "str d17, [x22], #0x8\n" - "tbz x14, #0, 100f\n" - "st1 { v9.s }[2], [x12]\n" + "tbz x17, #0, 100f\n" + "st1 { v9.s }[2], [x15]\n" "st1 { v13.s }[2], [x23]\n" "st1 { v17.s }[2], [x22]\n" "b 100f\n" "97:" // Height 3: Partial direct writeback: partial_1_4 - "tbz x14, #0, 100f\n" - "str s9, [x12, #0x0]\n" + "tbz x17, #0, 100f\n" + "str s9, [x15, #0x0]\n" "str s13, [x23, #0x0]\n" "str s17, [x22, #0x0]\n" "b 100f\n" "98:" // Height 3: Partial direct writeback: partial_2_0 - "tbz x14, #1, 99f\n" - "str d8, [x12], #0x8\n" + "tbz x17, #1, 99f\n" + "str d8, [x15], #0x8\n" "str d12, [x23], #0x8\n" "str d16, [x22], #0x8\n" - "tbz x14, #0, 100f\n" - "st1 { v8.s }[2], [x12]\n" + "tbz x17, #0, 100f\n" + "st1 { v8.s }[2], [x15]\n" "st1 { v12.s }[2], [x23]\n" "st1 { v16.s }[2], [x22]\n" "b 100f\n" "99:" // Height 3: Partial direct writeback: partial_1_0 - "str s8, [x12, #0x0]\n" + "str s8, [x15, #0x0]\n" "str s12, [x23, #0x0]\n" "str s16, [x22, #0x0]\n" "100:" // Height 3: Partial direct writeback: Done "b 102f\n" "101:" // Height 3: Full writeback - "str q8, [x12, #0x0]\n" - "str q9, [x12, #0x10]\n" - "str q10, [x12, #0x20]\n" - "str q11, [x12, #0x30]\n" - "add x12, x12, #0x40\n" + "str q8, [x15, #0x0]\n" + "str q9, [x15, #0x10]\n" + "str q10, [x15, #0x20]\n" + "str q11, [x15, #0x30]\n" + "add x15, x15, #0x40\n" "str q12, [x23, #0x0]\n" "str q13, [x23, #0x10]\n" "str q14, [x23, #0x20]\n" @@ -1317,129 +1317,129 @@ void a64_hybrid_u8u32_dot_6x16_a55 ( "str q18, [x22, #0x20]\n" "str q19, [x22, #0x30]\n" "102:" // Height 3: Writeback done - "subs x14, x14, #0x10\n" + "subs x17, x17, #0x10\n" "bgt 70b\n" "b 206f\n" "103:" // Height 4 - "ldr x14, [%x[args_ptr], %[offsetof_N]]\n" - "mov x12, %x[output_ptr]\n" - "ldr x13, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "ldr x17, [%x[args_ptr], %[offsetof_N]]\n" + "mov x15, %x[output_ptr]\n" + "ldr x16, [%x[args_ptr], %[offsetof_B_ptr]]\n" "104:" // Height 4: Column loop "tbz %x[flags], #0, 114f\n" "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "cmp x14, #0x10\n" - "add x23, x12, x19, LSL #2\n" + "cmp x17, #0x10\n" + "add x23, x15, x19, LSL #2\n" "add x22, x23, x19, LSL #2\n" "add x21, x22, x19, LSL #2\n" "bge 113f\n" - "tbz x14, #3, 108f\n" - "ld1 { v8.4s }, [x12], #0x10\n" + "tbz x17, #3, 108f\n" + "ld1 { v8.4s }, [x15], #0x10\n" "ld1 { v12.4s }, [x23], #0x10\n" "ld1 { v16.4s }, [x22], #0x10\n" - "ld1 { v9.4s }, [x12], #0x10\n" + "ld1 { v9.4s }, [x15], #0x10\n" "ld1 { v13.4s }, [x23], #0x10\n" "ld1 { v17.4s }, [x22], #0x10\n" "ld1 { v20.4s }, [x21], #0x10\n" "ld1 { v21.4s }, [x21], #0x10\n" - "tbz x14, #2, 106f\n" - "ld1 { v10.4s }, [x12], #0x10\n" + "tbz x17, #2, 106f\n" + "ld1 { v10.4s }, [x15], #0x10\n" "ld1 { v14.4s }, [x23], #0x10\n" "ld1 { v18.4s }, [x22], #0x10\n" "ld1 { v22.4s }, [x21], #0x10\n" - "tbz x14, #1, 105f\n" - "ldr d11, [x12], #0x8\n" + "tbz x17, #1, 105f\n" + "ldr d11, [x15], #0x8\n" "mov x24, #0x38\n" "ldr d15, [x23], #0x8\n" "ldr d19, [x22], #0x8\n" "ldr d23, [x21], #0x8\n" - "tbz x14, #0, 112f\n" - "ld1 { v11.s }[2], [x12]\n" + "tbz x17, #0, 112f\n" + "ld1 { v11.s }[2], [x15]\n" "ld1 { v15.s }[2], [x23]\n" "ld1 { v19.s }[2], [x22]\n" "ld1 { v23.s }[2], [x21]\n" "b 112f\n" "105:" // Height 4: Partial accumulate: partial_1_12 "mov x24, #0x30\n" - "tbz x14, #0, 112f\n" - "ldr s11, [x12, #0x0]\n" + "tbz x17, #0, 112f\n" + "ldr s11, [x15, #0x0]\n" "ldr s15, [x23, #0x0]\n" "ldr s19, [x22, #0x0]\n" "ldr s23, [x21, #0x0]\n" "b 112f\n" "106:" // Height 4: Partial accumulate: partial_2_8 - "tbz x14, #1, 107f\n" - "ldr d10, [x12], #0x8\n" + "tbz x17, #1, 107f\n" + "ldr d10, [x15], #0x8\n" "ldr d14, [x23], #0x8\n" "mov x24, #0x28\n" "ldr d18, [x22], #0x8\n" "ldr d22, [x21], #0x8\n" - "tbz x14, #0, 112f\n" - "ld1 { v10.s }[2], [x12]\n" + "tbz x17, #0, 112f\n" + "ld1 { v10.s }[2], [x15]\n" "ld1 { v14.s }[2], [x23]\n" "ld1 { v18.s }[2], [x22]\n" "ld1 { v22.s }[2], [x21]\n" "b 112f\n" "107:" // Height 4: Partial accumulate: partial_1_8 "mov x24, #0x20\n" - "tbz x14, #0, 112f\n" - "ldr s10, [x12, #0x0]\n" + "tbz x17, #0, 112f\n" + "ldr s10, [x15, #0x0]\n" "ldr s14, [x23, #0x0]\n" "ldr s18, [x22, #0x0]\n" "ldr s22, [x21, #0x0]\n" "b 112f\n" "108:" // Height 4: Partial accumulate: partial_4_0 - "tbz x14, #2, 110f\n" - "ld1 { v8.4s }, [x12], #0x10\n" + "tbz x17, #2, 110f\n" + "ld1 { v8.4s }, [x15], #0x10\n" "ld1 { v12.4s }, [x23], #0x10\n" "ld1 { v16.4s }, [x22], #0x10\n" "ld1 { v20.4s }, [x21], #0x10\n" - "tbz x14, #1, 109f\n" - "ldr d9, [x12], #0x8\n" + "tbz x17, #1, 109f\n" + "ldr d9, [x15], #0x8\n" "mov x24, #0x18\n" "ldr d13, [x23], #0x8\n" "ldr d17, [x22], #0x8\n" "ldr d21, [x21], #0x8\n" - "tbz x14, #0, 112f\n" - "ld1 { v9.s }[2], [x12]\n" + "tbz x17, #0, 112f\n" + "ld1 { v9.s }[2], [x15]\n" "ld1 { v13.s }[2], [x23]\n" "ld1 { v17.s }[2], [x22]\n" "ld1 { v21.s }[2], [x21]\n" "b 112f\n" "109:" // Height 4: Partial accumulate: partial_1_4 "mov x24, #0x10\n" - "tbz x14, #0, 112f\n" - "ldr s9, [x12, #0x0]\n" + "tbz x17, #0, 112f\n" + "ldr s9, [x15, #0x0]\n" "ldr s13, [x23, #0x0]\n" "ldr s17, [x22, #0x0]\n" "ldr s21, [x21, #0x0]\n" "b 112f\n" "110:" // Height 4: Partial accumulate: partial_2_0 - "tbz x14, #1, 111f\n" - "ldr d8, [x12], #0x8\n" + "tbz x17, #1, 111f\n" + "ldr d8, [x15], #0x8\n" "ldr d12, [x23], #0x8\n" "mov x24, #0x8\n" "ldr d16, [x22], #0x8\n" "ldr d20, [x21], #0x8\n" - "tbz x14, #0, 112f\n" - "ld1 { v8.s }[2], [x12]\n" + "tbz x17, #0, 112f\n" + "ld1 { v8.s }[2], [x15]\n" "ld1 { v12.s }[2], [x23]\n" "ld1 { v16.s }[2], [x22]\n" "ld1 { v20.s }[2], [x21]\n" "b 112f\n" "111:" // Height 4: Partial accumulate: partial_1_0 - "ldr s8, [x12, #0x0]\n" + "ldr s8, [x15, #0x0]\n" "mov x24, #0x0\n" "ldr s12, [x23, #0x0]\n" "ldr s16, [x22, #0x0]\n" "ldr s20, [x21, #0x0]\n" "112:" // Height 4: Partial accumulate: Done - "sub x12, x12, x24\n" + "sub x15, x15, x24\n" "b 115f\n" "113:" // Height 4: full accumulate - "ldr q8, [x12, #0x0]\n" - "ldr q9, [x12, #0x10]\n" - "ldr q10, [x12, #0x20]\n" - "ldr q11, [x12, #0x30]\n" + "ldr q8, [x15, #0x0]\n" + "ldr q9, [x15, #0x10]\n" + "ldr q10, [x15, #0x20]\n" + "ldr q11, [x15, #0x30]\n" "ldr q12, [x23, #0x0]\n" "ldr q13, [x23, #0x10]\n" "ldr q14, [x23, #0x20]\n" @@ -1471,292 +1471,292 @@ void a64_hybrid_u8u32_dot_6x16_a55 ( "movi v22.4s, #0x0\n" "movi v23.4s, #0x0\n" "115:" // Height 4: setup done - "mov x11, #0x0\n" + "mov x14, #0x0\n" "116:" // Height 4: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w10, [x20, x11, LSL #0x2]\n" + "ldr w13, [x20, x14, LSL #0x2]\n" "tbz %x[flags], #3, 117f\n" - "ldr x20, [%x[input_ptr], x11, LSL #0x3]\n" + "ldr x20, [%x[input_ptr], x14, LSL #0x3]\n" "add x20, x20, x19, LSL #3\n" - "ldr x9, [x20, #0x0]\n" - "ldr x26, [x20, #0x8]\n" - "ldr x24, [x20, #0x10]\n" - "ldr x23, [x20, #0x18]\n" - "cbnz x11, 118f\n" + "ldr x12, [x20, #0x0]\n" + "ldr x28, [x20, #0x8]\n" + "ldr x26, [x20, #0x10]\n" + "ldr x24, [x20, #0x18]\n" + "cbnz x14, 118f\n" "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x9, x9, x19\n" + "add x12, x12, x19\n" + "add x28, x28, x19\n" "add x26, x26, x19\n" "add x24, x24, x19\n" - "add x23, x23, x19\n" "b 118f\n" "117:" // Height 4: setup direct input - "mov x9, %x[input_ptr]\n" - "add x26, x9, x19\n" + "mov x12, %x[input_ptr]\n" + "add x28, x12, x19\n" + "add x26, x28, x19\n" "add x24, x26, x19\n" - "add x23, x24, x19\n" "118:" // Height 4: input setup done - "cmp x10, #0x10\n" + "cmp x13, #0x10\n" "blt 121f\n" - "ldr q0, [x9, #0x0]\n" - "ldr q1, [x26, #0x0]\n" - "cmp x10, #0x20\n" - "ldr q2, [x24, #0x0]\n" - "ldr q3, [x23, #0x0]\n" - "ldr q6, [x13, #0x0]\n" + "ldr q0, [x12, #0x0]\n" + "ldr q1, [x28, #0x0]\n" + "cmp x13, #0x20\n" + "ldr q2, [x26, #0x0]\n" + "ldr q3, [x24, #0x0]\n" + "ldr q6, [x16, #0x0]\n" "blt 120f\n" "119:" // Height 4: Multiply loop: Main loop head ".inst 0x6f80e0c8 // udot v8.4s, v6.16b, v0.4b[0]\n" + "ldr d7, [x16, #0x10]\n" ".inst 0x6f81e0cc // udot v12.4s, v6.16b, v1.4b[0]\n" + "ldr x11, [x16, #0x18]\n" ".inst 0x6f82e0d0 // udot v16.4s, v6.16b, v2.4b[0]\n" + "ldr x10, [x16, #0x28]\n" ".inst 0x6f83e0d4 // udot v20.4s, v6.16b, v3.4b[0]\n" - "ldr d7, [x13, #0x10]\n" - "add x9, x9, #0x10\n" - "ldr x19, [x13, #0x18]\n" - "add x26, x26, #0x10\n" - "prfm pldl1keep, [x9, #0x80]\n" - "add x24, x24, #0x10\n" - "ldr d6, [x13, #0x20]\n" - "add x23, x23, #0x10\n" - "mov v7.d[1], x19\n" - "ldr x28, [x13, #0x28]\n" - "ldr x19, [x13, #0x38]\n" - "sub x10, x10, #0x10\n" + "ldr d6, [x16, #0x20]\n" + "mov v7.d[1], x11\n" + "ldr x11, [x16, #0x38]\n" + "add x12, x12, #0x10\n" + "add x28, x28, #0x10\n" ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" + "mov v6.d[1], x10\n" ".inst 0x6f81e0ed // udot v13.4s, v7.16b, v1.4b[0]\n" - "mov v6.d[1], x28\n" + "prfm pldl1keep, [x12, #0x80]\n" ".inst 0x6f82e0f1 // udot v17.4s, v7.16b, v2.4b[0]\n" + "prfm pldl1keep, [x28, #0x80]\n" ".inst 0x6f83e0f5 // udot v21.4s, v7.16b, v3.4b[0]\n" - "ldr d7, [x13, #0x30]\n" + "ldr d7, [x16, #0x30]\n" ".inst 0x6f80e0ca // udot v10.4s, v6.16b, v0.4b[0]\n" + "ldr x10, [x16, #0x48]\n" ".inst 0x6f81e0ce // udot v14.4s, v6.16b, v1.4b[0]\n" + "ldr x9, [x12, #0x8]\n" ".inst 0x6f82e0d2 // udot v18.4s, v6.16b, v2.4b[0]\n" + "mov v7.d[1], x11\n" ".inst 0x6f83e0d6 // udot v22.4s, v6.16b, v3.4b[0]\n" - "mov v7.d[1], x19\n" - "prfm pldl1keep, [x26, #0x80]\n" - "ldr d6, [x13, #0x40]\n" - "cmp x10, #0x20\n" + "ldr d6, [x16, #0x40]\n" ".inst 0x6f80e0eb // udot v11.4s, v7.16b, v0.4b[0]\n" + "ldr x11, [x16, #0x58]\n" ".inst 0x6f81e0ef // udot v15.4s, v7.16b, v1.4b[0]\n" + "ldr x27, [x28, #0x8]\n" ".inst 0x6f82e0f3 // udot v19.4s, v7.16b, v2.4b[0]\n" + "mov v6.d[1], x10\n" ".inst 0x6f83e0f7 // udot v23.4s, v7.16b, v3.4b[0]\n" - "ldr x28, [x13, #0x48]\n" - "ldr d7, [x13, #0x50]\n" - "ldr x19, [x13, #0x58]\n" - "mov v6.d[1], x28\n" - "prfm pldl1keep, [x24, #0x80]\n" - "ldr x28, [x13, #0x68]\n" + "ldr d7, [x16, #0x50]\n" ".inst 0x6fa0e0c8 // udot v8.4s, v6.16b, v0.4b[1]\n" + "ldr x10, [x16, #0x68]\n" ".inst 0x6fa1e0cc // udot v12.4s, v6.16b, v1.4b[1]\n" + "add x26, x26, #0x10\n" ".inst 0x6fa2e0d0 // udot v16.4s, v6.16b, v2.4b[1]\n" + "mov v7.d[1], x11\n" ".inst 0x6fa3e0d4 // udot v20.4s, v6.16b, v3.4b[1]\n" - "mov v7.d[1], x19\n" - "ldr d6, [x13, #0x60]\n" - "ldr x19, [x13, #0x78]\n" + "prfm pldl1keep, [x26, #0x80]\n" ".inst 0x6fa0e0e9 // udot v9.4s, v7.16b, v0.4b[1]\n" + "ldr d6, [x16, #0x60]\n" ".inst 0x6fa1e0ed // udot v13.4s, v7.16b, v1.4b[1]\n" - "mov v6.d[1], x28\n" + "ldr x11, [x16, #0x78]\n" ".inst 0x6fa2e0f1 // udot v17.4s, v7.16b, v2.4b[1]\n" + "ldr x25, [x26, #0x8]\n" ".inst 0x6fa3e0f5 // udot v21.4s, v7.16b, v3.4b[1]\n" - "ldr d7, [x13, #0x70]\n" + "mov v6.d[1], x10\n" + "ldr d7, [x16, #0x70]\n" + "add x24, x24, #0x10\n" ".inst 0x6fa0e0ca // udot v10.4s, v6.16b, v0.4b[1]\n" + "prfm pldl1keep, [x24, #0x80]\n" ".inst 0x6fa1e0ce // udot v14.4s, v6.16b, v1.4b[1]\n" + "ldr x10, [x16, #0x88]\n" ".inst 0x6fa2e0d2 // udot v18.4s, v6.16b, v2.4b[1]\n" + "mov v7.d[1], x11\n" ".inst 0x6fa3e0d6 // udot v22.4s, v6.16b, v3.4b[1]\n" - "mov v7.d[1], x19\n" - "prfm pldl1keep, [x23, #0x80]\n" - "ldr d6, [x13, #0x80]\n" + "ldr d6, [x16, #0x80]\n" ".inst 0x6fa0e0eb // udot v11.4s, v7.16b, v0.4b[1]\n" + "ldr x11, [x16, #0x98]\n" ".inst 0x6fa1e0ef // udot v15.4s, v7.16b, v1.4b[1]\n" + "ldr x23, [x24, #0x8]\n" ".inst 0x6fa2e0f3 // udot v19.4s, v7.16b, v2.4b[1]\n" + "mov v6.d[1], x10\n" ".inst 0x6fa3e0f7 // udot v23.4s, v7.16b, v3.4b[1]\n" - "ldr x28, [x13, #0x88]\n" - "ldr d7, [x13, #0x90]\n" - "ldr x19, [x13, #0x98]\n" - "mov v6.d[1], x28\n" - "ldr x28, [x13, #0xa8]\n" - "ldr x27, [x9, #0x8]\n" + "ldr d7, [x16, #0x90]\n" ".inst 0x6f80e8c8 // udot v8.4s, v6.16b, v0.4b[2]\n" + "ldr x10, [x16, #0xa8]\n" ".inst 0x6f81e8cc // udot v12.4s, v6.16b, v1.4b[2]\n" + "sub x13, x13, #0x10\n" ".inst 0x6f82e8d0 // udot v16.4s, v6.16b, v2.4b[2]\n" + "mov v7.d[1], x11\n" ".inst 0x6f83e8d4 // udot v20.4s, v6.16b, v3.4b[2]\n" - "mov v7.d[1], x19\n" - "ldr d6, [x13, #0xa0]\n" - "ldr x19, [x13, #0xb8]\n" + "ldr d6, [x16, #0xa0]\n" ".inst 0x6f80e8e9 // udot v9.4s, v7.16b, v0.4b[2]\n" + "ldr x11, [x16, #0xb8]\n" ".inst 0x6f81e8ed // udot v13.4s, v7.16b, v1.4b[2]\n" - "mov v6.d[1], x28\n" + "cmp x13, #0x20\n" ".inst 0x6f82e8f1 // udot v17.4s, v7.16b, v2.4b[2]\n" + "mov v6.d[1], x10\n" ".inst 0x6f83e8f5 // udot v21.4s, v7.16b, v3.4b[2]\n" - "ldr d7, [x13, #0xb0]\n" + "ldr d7, [x16, #0xb0]\n" ".inst 0x6f80e8ca // udot v10.4s, v6.16b, v0.4b[2]\n" + "ldr x10, [x16, #0xc8]\n" ".inst 0x6f81e8ce // udot v14.4s, v6.16b, v1.4b[2]\n" ".inst 0x6f82e8d2 // udot v18.4s, v6.16b, v2.4b[2]\n" + "mov v7.d[1], x11\n" ".inst 0x6f83e8d6 // udot v22.4s, v6.16b, v3.4b[2]\n" - "mov v7.d[1], x19\n" - "ldr d6, [x13, #0xc0]\n" - "ldr x28, [x13, #0xc8]\n" + "ldr d6, [x16, #0xc0]\n" ".inst 0x6f80e8eb // udot v11.4s, v7.16b, v0.4b[2]\n" + "ldr x11, [x16, #0xd8]\n" ".inst 0x6f81e8ef // udot v15.4s, v7.16b, v1.4b[2]\n" ".inst 0x6f82e8f3 // udot v19.4s, v7.16b, v2.4b[2]\n" + "mov v6.d[1], x10\n" ".inst 0x6f83e8f7 // udot v23.4s, v7.16b, v3.4b[2]\n" - "mov v6.d[1], x28\n" - "ldr d7, [x13, #0xd0]\n" - "ldr x19, [x13, #0xd8]\n" + "ldr d7, [x16, #0xd0]\n" ".inst 0x6fa0e8c8 // udot v8.4s, v6.16b, v0.4b[3]\n" + "ldr x10, [x16, #0xe8]\n" ".inst 0x6fa1e8cc // udot v12.4s, v6.16b, v1.4b[3]\n" ".inst 0x6fa2e8d0 // udot v16.4s, v6.16b, v2.4b[3]\n" + "mov v7.d[1], x11\n" ".inst 0x6fa3e8d4 // udot v20.4s, v6.16b, v3.4b[3]\n" - "mov v7.d[1], x19\n" - "ldr d6, [x13, #0xe0]\n" - "ldr x28, [x13, #0xe8]\n" + "ldr d6, [x16, #0xe0]\n" ".inst 0x6fa0e8e9 // udot v9.4s, v7.16b, v0.4b[3]\n" + "ldr x11, [x16, #0xf8]\n" ".inst 0x6fa1e8ed // udot v13.4s, v7.16b, v1.4b[3]\n" ".inst 0x6fa2e8f1 // udot v17.4s, v7.16b, v2.4b[3]\n" + "mov v6.d[1], x10\n" ".inst 0x6fa3e8f5 // udot v21.4s, v7.16b, v3.4b[3]\n" - "mov v6.d[1], x28\n" - "ldr d7, [x13, #0xf0]\n" - "ldr x19, [x13, #0xf8]\n" - "add x13, x13, #0x100\n" + "ldr d7, [x16, #0xf0]\n" ".inst 0x6fa0e8ca // udot v10.4s, v6.16b, v0.4b[3]\n" + "add x16, x16, #0x100\n" ".inst 0x6fa1e8ce // udot v14.4s, v6.16b, v1.4b[3]\n" + "ldr x10, [x16, #0x8]\n" ".inst 0x6fa2e8d2 // udot v18.4s, v6.16b, v2.4b[3]\n" + "mov v7.d[1], x11\n" ".inst 0x6fa3e8d6 // udot v22.4s, v6.16b, v3.4b[3]\n" - "mov v7.d[1], x19\n" - "ldr d6, [x13, #0x0]\n" - "ldr x28, [x13, #0x8]\n" + "ldr d6, [x16, #0x0]\n" ".inst 0x6fa0e8eb // udot v11.4s, v7.16b, v0.4b[3]\n" + "ldr d0, [x12, #0x0]\n" ".inst 0x6fa1e8ef // udot v15.4s, v7.16b, v1.4b[3]\n" + "ldr d1, [x28, #0x0]\n" ".inst 0x6fa2e8f3 // udot v19.4s, v7.16b, v2.4b[3]\n" + "mov v6.d[1], x10\n" ".inst 0x6fa3e8f7 // udot v23.4s, v7.16b, v3.4b[3]\n" - "mov v6.d[1], x28\n" - "ldr d0, [x9, #0x0]\n" - "ldr d1, [x26, #0x0]\n" - "ldr x25, [x26, #0x8]\n" - "mov v0.d[1], x27\n" - "ldr d2, [x24, #0x0]\n" - "ldr x21, [x24, #0x8]\n" - "mov v1.d[1], x25\n" - "ldr d3, [x23, #0x0]\n" - "ldr x19, [x23, #0x8]\n" - "mov v2.d[1], x21\n" - "mov v3.d[1], x19\n" + "mov v0.d[1], x9\n" + "mov v1.d[1], x27\n" + "ldr d2, [x26, #0x0]\n" + "ldr d3, [x24, #0x0]\n" + "mov v2.d[1], x25\n" + "mov v3.d[1], x23\n" "bge 119b\n" "120:" // Height 4: Multiply loop: Single iteration only ".inst 0x6f80e0c8 // udot v8.4s, v6.16b, v0.4b[0]\n" + "ldr q7, [x16, #0x10]\n" ".inst 0x6f81e0cc // udot v12.4s, v6.16b, v1.4b[0]\n" + "sub x13, x13, #0x10\n" ".inst 0x6f82e0d0 // udot v16.4s, v6.16b, v2.4b[0]\n" + "add x12, x12, #0x10\n" ".inst 0x6f83e0d4 // udot v20.4s, v6.16b, v3.4b[0]\n" - "ldr q7, [x13, #0x10]\n" - "sub x10, x10, #0x10\n" - "ldr q6, [x13, #0x20]\n" - "add x9, x9, #0x10\n" - "prfm pldl1keep, [x9, #0x80]\n" - "add x26, x26, #0x10\n" + "prfm pldl1keep, [x12, #0x80]\n" ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" + "ldr q6, [x16, #0x20]\n" ".inst 0x6f81e0ed // udot v13.4s, v7.16b, v1.4b[0]\n" + "add x28, x28, #0x10\n" ".inst 0x6f82e0f1 // udot v17.4s, v7.16b, v2.4b[0]\n" + "prfm pldl1keep, [x28, #0x80]\n" ".inst 0x6f83e0f5 // udot v21.4s, v7.16b, v3.4b[0]\n" + "ldr q7, [x16, #0x30]\n" ".inst 0x6f80e0ca // udot v10.4s, v6.16b, v0.4b[0]\n" + "add x26, x26, #0x10\n" ".inst 0x6f81e0ce // udot v14.4s, v6.16b, v1.4b[0]\n" + "prfm pldl1keep, [x26, #0x80]\n" ".inst 0x6f82e0d2 // udot v18.4s, v6.16b, v2.4b[0]\n" - ".inst 0x6f83e0d6 // udot v22.4s, v6.16b, v3.4b[0]\n" - "ldr q7, [x13, #0x30]\n" "add x24, x24, #0x10\n" - "prfm pldl1keep, [x26, #0x80]\n" - "add x23, x23, #0x10\n" - "ldr q6, [x13, #0x40]\n" + ".inst 0x6f83e0d6 // udot v22.4s, v6.16b, v3.4b[0]\n" + "prfm pldl1keep, [x24, #0x80]\n" ".inst 0x6f80e0eb // udot v11.4s, v7.16b, v0.4b[0]\n" + "ldr q6, [x16, #0x40]\n" ".inst 0x6f81e0ef // udot v15.4s, v7.16b, v1.4b[0]\n" ".inst 0x6f82e0f3 // udot v19.4s, v7.16b, v2.4b[0]\n" ".inst 0x6f83e0f7 // udot v23.4s, v7.16b, v3.4b[0]\n" + "ldr q7, [x16, #0x50]\n" ".inst 0x6fa0e0c8 // udot v8.4s, v6.16b, v0.4b[1]\n" ".inst 0x6fa1e0cc // udot v12.4s, v6.16b, v1.4b[1]\n" ".inst 0x6fa2e0d0 // udot v16.4s, v6.16b, v2.4b[1]\n" ".inst 0x6fa3e0d4 // udot v20.4s, v6.16b, v3.4b[1]\n" - "ldr q7, [x13, #0x50]\n" - "prfm pldl1keep, [x24, #0x80]\n" - "ldr q6, [x13, #0x60]\n" + "ldr q6, [x16, #0x60]\n" ".inst 0x6fa0e0e9 // udot v9.4s, v7.16b, v0.4b[1]\n" ".inst 0x6fa1e0ed // udot v13.4s, v7.16b, v1.4b[1]\n" ".inst 0x6fa2e0f1 // udot v17.4s, v7.16b, v2.4b[1]\n" ".inst 0x6fa3e0f5 // udot v21.4s, v7.16b, v3.4b[1]\n" + "ldr q7, [x16, #0x70]\n" ".inst 0x6fa0e0ca // udot v10.4s, v6.16b, v0.4b[1]\n" ".inst 0x6fa1e0ce // udot v14.4s, v6.16b, v1.4b[1]\n" ".inst 0x6fa2e0d2 // udot v18.4s, v6.16b, v2.4b[1]\n" ".inst 0x6fa3e0d6 // udot v22.4s, v6.16b, v3.4b[1]\n" - "ldr q7, [x13, #0x70]\n" - "prfm pldl1keep, [x23, #0x80]\n" - "ldr q6, [x13, #0x80]\n" + "ldr q6, [x16, #0x80]\n" ".inst 0x6fa0e0eb // udot v11.4s, v7.16b, v0.4b[1]\n" ".inst 0x6fa1e0ef // udot v15.4s, v7.16b, v1.4b[1]\n" ".inst 0x6fa2e0f3 // udot v19.4s, v7.16b, v2.4b[1]\n" ".inst 0x6fa3e0f7 // udot v23.4s, v7.16b, v3.4b[1]\n" + "ldr q7, [x16, #0x90]\n" ".inst 0x6f80e8c8 // udot v8.4s, v6.16b, v0.4b[2]\n" ".inst 0x6f81e8cc // udot v12.4s, v6.16b, v1.4b[2]\n" ".inst 0x6f82e8d0 // udot v16.4s, v6.16b, v2.4b[2]\n" ".inst 0x6f83e8d4 // udot v20.4s, v6.16b, v3.4b[2]\n" - "ldr q7, [x13, #0x90]\n" - "ldr q6, [x13, #0xa0]\n" + "ldr q6, [x16, #0xa0]\n" ".inst 0x6f80e8e9 // udot v9.4s, v7.16b, v0.4b[2]\n" ".inst 0x6f81e8ed // udot v13.4s, v7.16b, v1.4b[2]\n" ".inst 0x6f82e8f1 // udot v17.4s, v7.16b, v2.4b[2]\n" ".inst 0x6f83e8f5 // udot v21.4s, v7.16b, v3.4b[2]\n" + "ldr q7, [x16, #0xb0]\n" ".inst 0x6f80e8ca // udot v10.4s, v6.16b, v0.4b[2]\n" ".inst 0x6f81e8ce // udot v14.4s, v6.16b, v1.4b[2]\n" ".inst 0x6f82e8d2 // udot v18.4s, v6.16b, v2.4b[2]\n" ".inst 0x6f83e8d6 // udot v22.4s, v6.16b, v3.4b[2]\n" - "ldr q7, [x13, #0xb0]\n" - "ldr q6, [x13, #0xc0]\n" + "ldr q6, [x16, #0xc0]\n" ".inst 0x6f80e8eb // udot v11.4s, v7.16b, v0.4b[2]\n" ".inst 0x6f81e8ef // udot v15.4s, v7.16b, v1.4b[2]\n" ".inst 0x6f82e8f3 // udot v19.4s, v7.16b, v2.4b[2]\n" ".inst 0x6f83e8f7 // udot v23.4s, v7.16b, v3.4b[2]\n" + "ldr q7, [x16, #0xd0]\n" ".inst 0x6fa0e8c8 // udot v8.4s, v6.16b, v0.4b[3]\n" ".inst 0x6fa1e8cc // udot v12.4s, v6.16b, v1.4b[3]\n" ".inst 0x6fa2e8d0 // udot v16.4s, v6.16b, v2.4b[3]\n" ".inst 0x6fa3e8d4 // udot v20.4s, v6.16b, v3.4b[3]\n" - "ldr q7, [x13, #0xd0]\n" - "ldr q6, [x13, #0xe0]\n" + "ldr q6, [x16, #0xe0]\n" ".inst 0x6fa0e8e9 // udot v9.4s, v7.16b, v0.4b[3]\n" ".inst 0x6fa1e8ed // udot v13.4s, v7.16b, v1.4b[3]\n" ".inst 0x6fa2e8f1 // udot v17.4s, v7.16b, v2.4b[3]\n" ".inst 0x6fa3e8f5 // udot v21.4s, v7.16b, v3.4b[3]\n" + "ldr q7, [x16, #0xf0]\n" ".inst 0x6fa0e8ca // udot v10.4s, v6.16b, v0.4b[3]\n" + "add x16, x16, #0x100\n" ".inst 0x6fa1e8ce // udot v14.4s, v6.16b, v1.4b[3]\n" ".inst 0x6fa2e8d2 // udot v18.4s, v6.16b, v2.4b[3]\n" ".inst 0x6fa3e8d6 // udot v22.4s, v6.16b, v3.4b[3]\n" - "ldr q7, [x13, #0xf0]\n" - "add x13, x13, #0x100\n" ".inst 0x6fa0e8eb // udot v11.4s, v7.16b, v0.4b[3]\n" ".inst 0x6fa1e8ef // udot v15.4s, v7.16b, v1.4b[3]\n" ".inst 0x6fa2e8f3 // udot v19.4s, v7.16b, v2.4b[3]\n" ".inst 0x6fa3e8f7 // udot v23.4s, v7.16b, v3.4b[3]\n" "121:" // Height 4: Multiply loop: Main loop skip - "cbz x10, 126f\n" - "cmp x10, #0x4\n" + "cbz x13, 126f\n" + "cmp x13, #0x4\n" "blt 123f\n" "122:" // Height 4: Multiply loop: Odd block loop - "ldr s0, [x9], #0x4\n" - "sub x10, x10, #0x4\n" - "ldr s1, [x26], #0x4\n" - "cmp x10, #0x4\n" - "ldr s2, [x24], #0x4\n" - "ldr s3, [x23], #0x4\n" - "ldr q6, [x13, #0x0]\n" - "ldr q7, [x13, #0x10]\n" + "ldr s0, [x12], #0x4\n" + "sub x13, x13, #0x4\n" + "ldr s1, [x28], #0x4\n" + "cmp x13, #0x4\n" + "ldr s2, [x26], #0x4\n" + "ldr s3, [x24], #0x4\n" + "ldr q6, [x16, #0x0]\n" + "ldr q7, [x16, #0x10]\n" ".inst 0x6f80e0c8 // udot v8.4s, v6.16b, v0.4b[0]\n" ".inst 0x6f81e0cc // udot v12.4s, v6.16b, v1.4b[0]\n" ".inst 0x6f82e0d0 // udot v16.4s, v6.16b, v2.4b[0]\n" ".inst 0x6f83e0d4 // udot v20.4s, v6.16b, v3.4b[0]\n" + "ldr q6, [x16, #0x20]\n" ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" ".inst 0x6f81e0ed // udot v13.4s, v7.16b, v1.4b[0]\n" ".inst 0x6f82e0f1 // udot v17.4s, v7.16b, v2.4b[0]\n" ".inst 0x6f83e0f5 // udot v21.4s, v7.16b, v3.4b[0]\n" - "ldr q6, [x13, #0x20]\n" - "ldr q7, [x13, #0x30]\n" - "add x13, x13, #0x40\n" + "ldr q7, [x16, #0x30]\n" ".inst 0x6f80e0ca // udot v10.4s, v6.16b, v0.4b[0]\n" + "add x16, x16, #0x40\n" ".inst 0x6f81e0ce // udot v14.4s, v6.16b, v1.4b[0]\n" ".inst 0x6f82e0d2 // udot v18.4s, v6.16b, v2.4b[0]\n" ".inst 0x6f83e0d6 // udot v22.4s, v6.16b, v3.4b[0]\n" @@ -1765,39 +1765,39 @@ void a64_hybrid_u8u32_dot_6x16_a55 ( ".inst 0x6f82e0f3 // udot v19.4s, v7.16b, v2.4b[0]\n" ".inst 0x6f83e0f7 // udot v23.4s, v7.16b, v3.4b[0]\n" "bge 122b\n" - "cbz x10, 126f\n" + "cbz x13, 126f\n" "123:" // Height 4: Multiply loop: Skip odd blocks - "tbz x10, #1, 124f\n" - "ldr h0, [x9], #0x2\n" - "ldr h1, [x26], #0x2\n" - "ldr h2, [x24], #0x2\n" - "ldr h3, [x23], #0x2\n" - "tbz x10, #0, 125f\n" - "ld1 { v0.b }[2], [x9]\n" - "ld1 { v1.b }[2], [x26]\n" - "ld1 { v2.b }[2], [x24]\n" - "ld1 { v3.b }[2], [x23]\n" + "tbz x13, #1, 124f\n" + "ldr h0, [x12], #0x2\n" + "ldr h1, [x28], #0x2\n" + "ldr h2, [x26], #0x2\n" + "ldr h3, [x24], #0x2\n" + "tbz x13, #0, 125f\n" + "ld1 { v0.b }[2], [x12]\n" + "ld1 { v1.b }[2], [x28]\n" + "ld1 { v2.b }[2], [x26]\n" + "ld1 { v3.b }[2], [x24]\n" "b 125f\n" "124:" // Height 4: Multiply loop: Ragged operand read: partial_1_0 - "ldr b0, [x9, #0x0]\n" - "ldr b1, [x26, #0x0]\n" - "ldr b2, [x24, #0x0]\n" - "ldr b3, [x23, #0x0]\n" + "ldr b0, [x12, #0x0]\n" + "ldr b1, [x28, #0x0]\n" + "ldr b2, [x26, #0x0]\n" + "ldr b3, [x24, #0x0]\n" "125:" // Height 4: Multiply loop: Ragged operand read: Done - "ldr q6, [x13, #0x0]\n" - "ldr q7, [x13, #0x10]\n" + "ldr q6, [x16, #0x0]\n" + "ldr q7, [x16, #0x10]\n" ".inst 0x6f80e0c8 // udot v8.4s, v6.16b, v0.4b[0]\n" ".inst 0x6f81e0cc // udot v12.4s, v6.16b, v1.4b[0]\n" ".inst 0x6f82e0d0 // udot v16.4s, v6.16b, v2.4b[0]\n" ".inst 0x6f83e0d4 // udot v20.4s, v6.16b, v3.4b[0]\n" + "ldr q6, [x16, #0x20]\n" ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" ".inst 0x6f81e0ed // udot v13.4s, v7.16b, v1.4b[0]\n" ".inst 0x6f82e0f1 // udot v17.4s, v7.16b, v2.4b[0]\n" ".inst 0x6f83e0f5 // udot v21.4s, v7.16b, v3.4b[0]\n" - "ldr q6, [x13, #0x20]\n" - "ldr q7, [x13, #0x30]\n" - "add x13, x13, #0x40\n" + "ldr q7, [x16, #0x30]\n" ".inst 0x6f80e0ca // udot v10.4s, v6.16b, v0.4b[0]\n" + "add x16, x16, #0x40\n" ".inst 0x6f81e0ce // udot v14.4s, v6.16b, v1.4b[0]\n" ".inst 0x6f82e0d2 // udot v18.4s, v6.16b, v2.4b[0]\n" ".inst 0x6f83e0d6 // udot v22.4s, v6.16b, v3.4b[0]\n" @@ -1807,119 +1807,119 @@ void a64_hybrid_u8u32_dot_6x16_a55 ( ".inst 0x6f83e0f7 // udot v23.4s, v7.16b, v3.4b[0]\n" "126:" // Height 4: Multiply loop: No odd multiplies "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" - "add x11, x11, #0x1\n" - "cmp x11, x19\n" + "add x14, x14, #0x1\n" + "cmp x14, x19\n" "bne 116b\n" "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "prfm pstl1keep, [x12, #0x0]\n" - "cmp x14, #0x10\n" - "add x23, x12, x19, LSL #2\n" + "prfm pstl1keep, [x15, #0x0]\n" + "cmp x17, #0x10\n" + "add x23, x15, x19, LSL #2\n" "prfm pstl1keep, [x23, #0x0]\n" "add x22, x23, x19, LSL #2\n" "prfm pstl1keep, [x22, #0x0]\n" "add x21, x22, x19, LSL #2\n" "prfm pstl1keep, [x21, #0x0]\n" "bge 135f\n" - "tbz x14, #3, 130f\n" - "st1 { v8.4s }, [x12], #0x10\n" - "st1 { v9.4s }, [x12], #0x10\n" + "tbz x17, #3, 130f\n" + "st1 { v8.4s }, [x15], #0x10\n" + "st1 { v9.4s }, [x15], #0x10\n" "st1 { v12.4s }, [x23], #0x10\n" "st1 { v13.4s }, [x23], #0x10\n" "st1 { v16.4s }, [x22], #0x10\n" "st1 { v17.4s }, [x22], #0x10\n" "st1 { v20.4s }, [x21], #0x10\n" "st1 { v21.4s }, [x21], #0x10\n" - "tbz x14, #2, 128f\n" - "st1 { v10.4s }, [x12], #0x10\n" + "tbz x17, #2, 128f\n" + "st1 { v10.4s }, [x15], #0x10\n" "st1 { v14.4s }, [x23], #0x10\n" "st1 { v18.4s }, [x22], #0x10\n" "st1 { v22.4s }, [x21], #0x10\n" - "tbz x14, #1, 127f\n" - "str d11, [x12], #0x8\n" + "tbz x17, #1, 127f\n" + "str d11, [x15], #0x8\n" "str d15, [x23], #0x8\n" "str d19, [x22], #0x8\n" "str d23, [x21], #0x8\n" - "tbz x14, #0, 134f\n" - "st1 { v11.s }[2], [x12]\n" + "tbz x17, #0, 134f\n" + "st1 { v11.s }[2], [x15]\n" "st1 { v15.s }[2], [x23]\n" "st1 { v19.s }[2], [x22]\n" "st1 { v23.s }[2], [x21]\n" "b 134f\n" "127:" // Height 4: Partial direct writeback: partial_1_12 - "tbz x14, #0, 134f\n" - "str s11, [x12, #0x0]\n" + "tbz x17, #0, 134f\n" + "str s11, [x15, #0x0]\n" "str s15, [x23, #0x0]\n" "str s19, [x22, #0x0]\n" "str s23, [x21, #0x0]\n" "b 134f\n" "128:" // Height 4: Partial direct writeback: partial_2_8 - "tbz x14, #1, 129f\n" - "str d10, [x12], #0x8\n" + "tbz x17, #1, 129f\n" + "str d10, [x15], #0x8\n" "str d14, [x23], #0x8\n" "str d18, [x22], #0x8\n" "str d22, [x21], #0x8\n" - "tbz x14, #0, 134f\n" - "st1 { v10.s }[2], [x12]\n" + "tbz x17, #0, 134f\n" + "st1 { v10.s }[2], [x15]\n" "st1 { v14.s }[2], [x23]\n" "st1 { v18.s }[2], [x22]\n" "st1 { v22.s }[2], [x21]\n" "b 134f\n" "129:" // Height 4: Partial direct writeback: partial_1_8 - "tbz x14, #0, 134f\n" - "str s10, [x12, #0x0]\n" + "tbz x17, #0, 134f\n" + "str s10, [x15, #0x0]\n" "str s14, [x23, #0x0]\n" "str s18, [x22, #0x0]\n" "str s22, [x21, #0x0]\n" "b 134f\n" "130:" // Height 4: Partial direct writeback: partial_4_0 - "tbz x14, #2, 132f\n" - "st1 { v8.4s }, [x12], #0x10\n" + "tbz x17, #2, 132f\n" + "st1 { v8.4s }, [x15], #0x10\n" "st1 { v12.4s }, [x23], #0x10\n" "st1 { v16.4s }, [x22], #0x10\n" "st1 { v20.4s }, [x21], #0x10\n" - "tbz x14, #1, 131f\n" - "str d9, [x12], #0x8\n" + "tbz x17, #1, 131f\n" + "str d9, [x15], #0x8\n" "str d13, [x23], #0x8\n" "str d17, [x22], #0x8\n" "str d21, [x21], #0x8\n" - "tbz x14, #0, 134f\n" - "st1 { v9.s }[2], [x12]\n" + "tbz x17, #0, 134f\n" + "st1 { v9.s }[2], [x15]\n" "st1 { v13.s }[2], [x23]\n" "st1 { v17.s }[2], [x22]\n" "st1 { v21.s }[2], [x21]\n" "b 134f\n" "131:" // Height 4: Partial direct writeback: partial_1_4 - "tbz x14, #0, 134f\n" - "str s9, [x12, #0x0]\n" + "tbz x17, #0, 134f\n" + "str s9, [x15, #0x0]\n" "str s13, [x23, #0x0]\n" "str s17, [x22, #0x0]\n" "str s21, [x21, #0x0]\n" "b 134f\n" "132:" // Height 4: Partial direct writeback: partial_2_0 - "tbz x14, #1, 133f\n" - "str d8, [x12], #0x8\n" + "tbz x17, #1, 133f\n" + "str d8, [x15], #0x8\n" "str d12, [x23], #0x8\n" "str d16, [x22], #0x8\n" "str d20, [x21], #0x8\n" - "tbz x14, #0, 134f\n" - "st1 { v8.s }[2], [x12]\n" + "tbz x17, #0, 134f\n" + "st1 { v8.s }[2], [x15]\n" "st1 { v12.s }[2], [x23]\n" "st1 { v16.s }[2], [x22]\n" "st1 { v20.s }[2], [x21]\n" "b 134f\n" "133:" // Height 4: Partial direct writeback: partial_1_0 - "str s8, [x12, #0x0]\n" + "str s8, [x15, #0x0]\n" "str s12, [x23, #0x0]\n" "str s16, [x22, #0x0]\n" "str s20, [x21, #0x0]\n" "134:" // Height 4: Partial direct writeback: Done "b 136f\n" "135:" // Height 4: Full writeback - "str q8, [x12, #0x0]\n" - "str q9, [x12, #0x10]\n" - "str q10, [x12, #0x20]\n" - "str q11, [x12, #0x30]\n" - "add x12, x12, #0x40\n" + "str q8, [x15, #0x0]\n" + "str q9, [x15, #0x10]\n" + "str q10, [x15, #0x20]\n" + "str q11, [x15, #0x30]\n" + "add x15, x15, #0x40\n" "str q12, [x23, #0x0]\n" "str q13, [x23, #0x10]\n" "str q14, [x23, #0x20]\n" @@ -1933,48 +1933,48 @@ void a64_hybrid_u8u32_dot_6x16_a55 ( "str q22, [x21, #0x20]\n" "str q23, [x21, #0x30]\n" "136:" // Height 4: Writeback done - "subs x14, x14, #0x10\n" + "subs x17, x17, #0x10\n" "bgt 104b\n" "b 206f\n" "137:" // Height 5 - "ldr x14, [%x[args_ptr], %[offsetof_N]]\n" - "mov x12, %x[output_ptr]\n" - "ldr x13, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "ldr x17, [%x[args_ptr], %[offsetof_N]]\n" + "mov x15, %x[output_ptr]\n" + "ldr x16, [%x[args_ptr], %[offsetof_B_ptr]]\n" "138:" // Height 5: Column loop "tbz %x[flags], #0, 148f\n" "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "cmp x14, #0x10\n" - "add x23, x12, x19, LSL #2\n" + "cmp x17, #0x10\n" + "add x23, x15, x19, LSL #2\n" "add x22, x23, x19, LSL #2\n" "add x21, x22, x19, LSL #2\n" "add x20, x21, x19, LSL #2\n" "bge 147f\n" - "tbz x14, #3, 142f\n" - "ld1 { v8.4s }, [x12], #0x10\n" + "tbz x17, #3, 142f\n" + "ld1 { v8.4s }, [x15], #0x10\n" "ld1 { v12.4s }, [x23], #0x10\n" "ld1 { v16.4s }, [x22], #0x10\n" - "ld1 { v9.4s }, [x12], #0x10\n" + "ld1 { v9.4s }, [x15], #0x10\n" "ld1 { v13.4s }, [x23], #0x10\n" "ld1 { v17.4s }, [x22], #0x10\n" "ld1 { v20.4s }, [x21], #0x10\n" "ld1 { v24.4s }, [x20], #0x10\n" "ld1 { v21.4s }, [x21], #0x10\n" "ld1 { v25.4s }, [x20], #0x10\n" - "tbz x14, #2, 140f\n" - "ld1 { v10.4s }, [x12], #0x10\n" + "tbz x17, #2, 140f\n" + "ld1 { v10.4s }, [x15], #0x10\n" "ld1 { v14.4s }, [x23], #0x10\n" "ld1 { v18.4s }, [x22], #0x10\n" "ld1 { v22.4s }, [x21], #0x10\n" "ld1 { v26.4s }, [x20], #0x10\n" - "tbz x14, #1, 139f\n" - "ldr d11, [x12], #0x8\n" + "tbz x17, #1, 139f\n" + "ldr d11, [x15], #0x8\n" "mov x24, #0x38\n" "ldr d15, [x23], #0x8\n" "ldr d19, [x22], #0x8\n" "ldr d23, [x21], #0x8\n" "ldr d27, [x20], #0x8\n" - "tbz x14, #0, 146f\n" - "ld1 { v11.s }[2], [x12]\n" + "tbz x17, #0, 146f\n" + "ld1 { v11.s }[2], [x15]\n" "ld1 { v15.s }[2], [x23]\n" "ld1 { v19.s }[2], [x22]\n" "ld1 { v23.s }[2], [x21]\n" @@ -1982,23 +1982,23 @@ void a64_hybrid_u8u32_dot_6x16_a55 ( "b 146f\n" "139:" // Height 5: Partial accumulate: partial_1_12 "mov x24, #0x30\n" - "tbz x14, #0, 146f\n" - "ldr s11, [x12, #0x0]\n" + "tbz x17, #0, 146f\n" + "ldr s11, [x15, #0x0]\n" "ldr s15, [x23, #0x0]\n" "ldr s19, [x22, #0x0]\n" "ldr s23, [x21, #0x0]\n" "ldr s27, [x20, #0x0]\n" "b 146f\n" "140:" // Height 5: Partial accumulate: partial_2_8 - "tbz x14, #1, 141f\n" - "ldr d10, [x12], #0x8\n" + "tbz x17, #1, 141f\n" + "ldr d10, [x15], #0x8\n" "ldr d14, [x23], #0x8\n" "mov x24, #0x28\n" "ldr d18, [x22], #0x8\n" "ldr d22, [x21], #0x8\n" "ldr d26, [x20], #0x8\n" - "tbz x14, #0, 146f\n" - "ld1 { v10.s }[2], [x12]\n" + "tbz x17, #0, 146f\n" + "ld1 { v10.s }[2], [x15]\n" "ld1 { v14.s }[2], [x23]\n" "ld1 { v18.s }[2], [x22]\n" "ld1 { v22.s }[2], [x21]\n" @@ -2006,29 +2006,29 @@ void a64_hybrid_u8u32_dot_6x16_a55 ( "b 146f\n" "141:" // Height 5: Partial accumulate: partial_1_8 "mov x24, #0x20\n" - "tbz x14, #0, 146f\n" - "ldr s10, [x12, #0x0]\n" + "tbz x17, #0, 146f\n" + "ldr s10, [x15, #0x0]\n" "ldr s14, [x23, #0x0]\n" "ldr s18, [x22, #0x0]\n" "ldr s22, [x21, #0x0]\n" "ldr s26, [x20, #0x0]\n" "b 146f\n" "142:" // Height 5: Partial accumulate: partial_4_0 - "tbz x14, #2, 144f\n" - "ld1 { v8.4s }, [x12], #0x10\n" + "tbz x17, #2, 144f\n" + "ld1 { v8.4s }, [x15], #0x10\n" "ld1 { v12.4s }, [x23], #0x10\n" "ld1 { v16.4s }, [x22], #0x10\n" "ld1 { v20.4s }, [x21], #0x10\n" "ld1 { v24.4s }, [x20], #0x10\n" - "tbz x14, #1, 143f\n" - "ldr d9, [x12], #0x8\n" + "tbz x17, #1, 143f\n" + "ldr d9, [x15], #0x8\n" "mov x24, #0x18\n" "ldr d13, [x23], #0x8\n" "ldr d17, [x22], #0x8\n" "ldr d21, [x21], #0x8\n" "ldr d25, [x20], #0x8\n" - "tbz x14, #0, 146f\n" - "ld1 { v9.s }[2], [x12]\n" + "tbz x17, #0, 146f\n" + "ld1 { v9.s }[2], [x15]\n" "ld1 { v13.s }[2], [x23]\n" "ld1 { v17.s }[2], [x22]\n" "ld1 { v21.s }[2], [x21]\n" @@ -2036,43 +2036,43 @@ void a64_hybrid_u8u32_dot_6x16_a55 ( "b 146f\n" "143:" // Height 5: Partial accumulate: partial_1_4 "mov x24, #0x10\n" - "tbz x14, #0, 146f\n" - "ldr s9, [x12, #0x0]\n" + "tbz x17, #0, 146f\n" + "ldr s9, [x15, #0x0]\n" "ldr s13, [x23, #0x0]\n" "ldr s17, [x22, #0x0]\n" "ldr s21, [x21, #0x0]\n" "ldr s25, [x20, #0x0]\n" "b 146f\n" "144:" // Height 5: Partial accumulate: partial_2_0 - "tbz x14, #1, 145f\n" - "ldr d8, [x12], #0x8\n" + "tbz x17, #1, 145f\n" + "ldr d8, [x15], #0x8\n" "ldr d12, [x23], #0x8\n" "mov x24, #0x8\n" "ldr d16, [x22], #0x8\n" "ldr d20, [x21], #0x8\n" "ldr d24, [x20], #0x8\n" - "tbz x14, #0, 146f\n" - "ld1 { v8.s }[2], [x12]\n" + "tbz x17, #0, 146f\n" + "ld1 { v8.s }[2], [x15]\n" "ld1 { v12.s }[2], [x23]\n" "ld1 { v16.s }[2], [x22]\n" "ld1 { v20.s }[2], [x21]\n" "ld1 { v24.s }[2], [x20]\n" "b 146f\n" "145:" // Height 5: Partial accumulate: partial_1_0 - "ldr s8, [x12, #0x0]\n" + "ldr s8, [x15, #0x0]\n" "mov x24, #0x0\n" "ldr s12, [x23, #0x0]\n" "ldr s16, [x22, #0x0]\n" "ldr s20, [x21, #0x0]\n" "ldr s24, [x20, #0x0]\n" "146:" // Height 5: Partial accumulate: Done - "sub x12, x12, x24\n" + "sub x15, x15, x24\n" "b 149f\n" "147:" // Height 5: full accumulate - "ldr q8, [x12, #0x0]\n" - "ldr q9, [x12, #0x10]\n" - "ldr q10, [x12, #0x20]\n" - "ldr q11, [x12, #0x30]\n" + "ldr q8, [x15, #0x0]\n" + "ldr q9, [x15, #0x10]\n" + "ldr q10, [x15, #0x20]\n" + "ldr q11, [x15, #0x30]\n" "ldr q12, [x23, #0x0]\n" "ldr q13, [x23, #0x10]\n" "ldr q14, [x23, #0x20]\n" @@ -2112,338 +2112,338 @@ void a64_hybrid_u8u32_dot_6x16_a55 ( "movi v26.4s, #0x0\n" "movi v27.4s, #0x0\n" "149:" // Height 5: setup done - "mov x11, #0x0\n" + "mov x14, #0x0\n" "150:" // Height 5: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w10, [x20, x11, LSL #0x2]\n" + "ldr w13, [x20, x14, LSL #0x2]\n" "tbz %x[flags], #3, 151f\n" - "ldr x20, [%x[input_ptr], x11, LSL #0x3]\n" + "ldr x20, [%x[input_ptr], x14, LSL #0x3]\n" "add x20, x20, x19, LSL #3\n" - "ldr x9, [x20, #0x0]\n" - "ldr x26, [x20, #0x8]\n" - "ldr x24, [x20, #0x10]\n" - "ldr x23, [x20, #0x18]\n" + "ldr x12, [x20, #0x0]\n" + "ldr x28, [x20, #0x8]\n" + "ldr x26, [x20, #0x10]\n" + "ldr x24, [x20, #0x18]\n" "ldr x22, [x20, #0x20]\n" - "cbnz x11, 152f\n" + "cbnz x14, 152f\n" "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x9, x9, x19\n" + "add x12, x12, x19\n" + "add x28, x28, x19\n" "add x26, x26, x19\n" "add x24, x24, x19\n" - "add x23, x23, x19\n" "add x22, x22, x19\n" "b 152f\n" "151:" // Height 5: setup direct input - "mov x9, %x[input_ptr]\n" - "add x26, x9, x19\n" + "mov x12, %x[input_ptr]\n" + "add x28, x12, x19\n" + "add x26, x28, x19\n" "add x24, x26, x19\n" - "add x23, x24, x19\n" - "add x22, x23, x19\n" + "add x22, x24, x19\n" "152:" // Height 5: input setup done - "cmp x10, #0x10\n" + "cmp x13, #0x10\n" "blt 155f\n" - "ldr q0, [x9, #0x0]\n" - "ldr q1, [x26, #0x0]\n" - "cmp x10, #0x20\n" - "ldr q2, [x24, #0x0]\n" - "ldr q3, [x23, #0x0]\n" + "ldr q0, [x12, #0x0]\n" + "ldr q1, [x28, #0x0]\n" + "cmp x13, #0x20\n" + "ldr q2, [x26, #0x0]\n" + "ldr q3, [x24, #0x0]\n" "ldr q4, [x22, #0x0]\n" - "ldr q6, [x13, #0x0]\n" + "ldr q6, [x16, #0x0]\n" "blt 154f\n" "153:" // Height 5: Multiply loop: Main loop head ".inst 0x6f80e0c8 // udot v8.4s, v6.16b, v0.4b[0]\n" + "ldr d7, [x16, #0x10]\n" ".inst 0x6f81e0cc // udot v12.4s, v6.16b, v1.4b[0]\n" + "ldr x11, [x16, #0x18]\n" ".inst 0x6f82e0d0 // udot v16.4s, v6.16b, v2.4b[0]\n" + "ldr x10, [x16, #0x28]\n" ".inst 0x6f83e0d4 // udot v20.4s, v6.16b, v3.4b[0]\n" + "add x12, x12, #0x10\n" ".inst 0x6f84e0d8 // udot v24.4s, v6.16b, v4.4b[0]\n" - "ldr d7, [x13, #0x10]\n" - "ldr x19, [x13, #0x18]\n" - "add x9, x9, #0x10\n" - "ldr d6, [x13, #0x20]\n" - "add x26, x26, #0x10\n" - "ldr x28, [x13, #0x28]\n" - "add x24, x24, #0x10\n" - "mov v7.d[1], x19\n" - "prfm pldl1keep, [x9, #0x80]\n" - "ldr x19, [x13, #0x38]\n" - "add x23, x23, #0x10\n" - "mov v6.d[1], x28\n" + "mov v7.d[1], x11\n" + "prfm pldl1keep, [x12, #0x80]\n" + "add x28, x28, #0x10\n" ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" + "prfm pldl1keep, [x28, #0x80]\n" ".inst 0x6f81e0ed // udot v13.4s, v7.16b, v1.4b[0]\n" + "ldr d6, [x16, #0x20]\n" ".inst 0x6f82e0f1 // udot v17.4s, v7.16b, v2.4b[0]\n" + "ldr x11, [x16, #0x38]\n" ".inst 0x6f83e0f5 // udot v21.4s, v7.16b, v3.4b[0]\n" + "ldr x9, [x12, #0x8]\n" ".inst 0x6f84e0f9 // udot v25.4s, v7.16b, v4.4b[0]\n" + "mov v6.d[1], x10\n" + "ldr d7, [x16, #0x30]\n" + "add x26, x26, #0x10\n" ".inst 0x6f80e0ca // udot v10.4s, v6.16b, v0.4b[0]\n" + "prfm pldl1keep, [x26, #0x80]\n" ".inst 0x6f81e0ce // udot v14.4s, v6.16b, v1.4b[0]\n" + "ldr x10, [x16, #0x48]\n" ".inst 0x6f82e0d2 // udot v18.4s, v6.16b, v2.4b[0]\n" + "mov v7.d[1], x11\n" ".inst 0x6f83e0d6 // udot v22.4s, v6.16b, v3.4b[0]\n" + "ldr x11, [x16, #0x58]\n" ".inst 0x6f84e0da // udot v26.4s, v6.16b, v4.4b[0]\n" - "ldr d7, [x13, #0x30]\n" - "prfm pldl1keep, [x26, #0x80]\n" - "add x22, x22, #0x10\n" - "ldr d6, [x13, #0x40]\n" - "sub x10, x10, #0x10\n" - "mov v7.d[1], x19\n" - "ldr x28, [x13, #0x48]\n" - "ldr x19, [x13, #0x58]\n" - "cmp x10, #0x20\n" + "ldr d6, [x16, #0x40]\n" ".inst 0x6f80e0eb // udot v11.4s, v7.16b, v0.4b[0]\n" + "ldr x27, [x28, #0x8]\n" ".inst 0x6f81e0ef // udot v15.4s, v7.16b, v1.4b[0]\n" - "mov v6.d[1], x28\n" + "ldr x25, [x26, #0x8]\n" ".inst 0x6f82e0f3 // udot v19.4s, v7.16b, v2.4b[0]\n" + "mov v6.d[1], x10\n" ".inst 0x6f83e0f7 // udot v23.4s, v7.16b, v3.4b[0]\n" + "ldr x10, [x16, #0x68]\n" ".inst 0x6f84e0fb // udot v27.4s, v7.16b, v4.4b[0]\n" + "ldr d7, [x16, #0x50]\n" ".inst 0x6fa0e0c8 // udot v8.4s, v6.16b, v0.4b[1]\n" + "add x24, x24, #0x10\n" ".inst 0x6fa1e0cc // udot v12.4s, v6.16b, v1.4b[1]\n" + "prfm pldl1keep, [x24, #0x80]\n" ".inst 0x6fa2e0d0 // udot v16.4s, v6.16b, v2.4b[1]\n" + "mov v7.d[1], x11\n" ".inst 0x6fa3e0d4 // udot v20.4s, v6.16b, v3.4b[1]\n" + "ldr x11, [x16, #0x78]\n" ".inst 0x6fa4e0d8 // udot v24.4s, v6.16b, v4.4b[1]\n" - "ldr d7, [x13, #0x50]\n" - "prfm pldl1keep, [x24, #0x80]\n" - "ldr d6, [x13, #0x60]\n" - "mov v7.d[1], x19\n" - "ldr x28, [x13, #0x68]\n" - "ldr x19, [x13, #0x78]\n" + "ldr d6, [x16, #0x60]\n" ".inst 0x6fa0e0e9 // udot v9.4s, v7.16b, v0.4b[1]\n" + "ldr x23, [x24, #0x8]\n" ".inst 0x6fa1e0ed // udot v13.4s, v7.16b, v1.4b[1]\n" - "mov v6.d[1], x28\n" + "add x22, x22, #0x10\n" ".inst 0x6fa2e0f1 // udot v17.4s, v7.16b, v2.4b[1]\n" + "mov v6.d[1], x10\n" ".inst 0x6fa3e0f5 // udot v21.4s, v7.16b, v3.4b[1]\n" + "prfm pldl1keep, [x22, #0x80]\n" ".inst 0x6fa4e0f9 // udot v25.4s, v7.16b, v4.4b[1]\n" + "ldr d7, [x16, #0x70]\n" ".inst 0x6fa0e0ca // udot v10.4s, v6.16b, v0.4b[1]\n" + "ldr x10, [x16, #0x88]\n" ".inst 0x6fa1e0ce // udot v14.4s, v6.16b, v1.4b[1]\n" + "ldr x21, [x22, #0x8]\n" ".inst 0x6fa2e0d2 // udot v18.4s, v6.16b, v2.4b[1]\n" + "mov v7.d[1], x11\n" ".inst 0x6fa3e0d6 // udot v22.4s, v6.16b, v3.4b[1]\n" + "ldr x11, [x16, #0x98]\n" ".inst 0x6fa4e0da // udot v26.4s, v6.16b, v4.4b[1]\n" - "ldr d7, [x13, #0x70]\n" - "prfm pldl1keep, [x23, #0x80]\n" - "ldr d6, [x13, #0x80]\n" - "mov v7.d[1], x19\n" - "ldr x28, [x13, #0x88]\n" - "ldr x19, [x13, #0x98]\n" + "ldr d6, [x16, #0x80]\n" ".inst 0x6fa0e0eb // udot v11.4s, v7.16b, v0.4b[1]\n" + "sub x13, x13, #0x10\n" ".inst 0x6fa1e0ef // udot v15.4s, v7.16b, v1.4b[1]\n" - "mov v6.d[1], x28\n" + "cmp x13, #0x20\n" ".inst 0x6fa2e0f3 // udot v19.4s, v7.16b, v2.4b[1]\n" + "mov v6.d[1], x10\n" ".inst 0x6fa3e0f7 // udot v23.4s, v7.16b, v3.4b[1]\n" + "ldr x10, [x16, #0xa8]\n" ".inst 0x6fa4e0fb // udot v27.4s, v7.16b, v4.4b[1]\n" + "ldr d7, [x16, #0x90]\n" ".inst 0x6f80e8c8 // udot v8.4s, v6.16b, v0.4b[2]\n" ".inst 0x6f81e8cc // udot v12.4s, v6.16b, v1.4b[2]\n" ".inst 0x6f82e8d0 // udot v16.4s, v6.16b, v2.4b[2]\n" + "mov v7.d[1], x11\n" ".inst 0x6f83e8d4 // udot v20.4s, v6.16b, v3.4b[2]\n" + "ldr x11, [x16, #0xb8]\n" ".inst 0x6f84e8d8 // udot v24.4s, v6.16b, v4.4b[2]\n" - "ldr d7, [x13, #0x90]\n" - "prfm pldl1keep, [x22, #0x80]\n" - "ldr d6, [x13, #0xa0]\n" - "mov v7.d[1], x19\n" - "ldr x28, [x13, #0xa8]\n" - "ldr x19, [x13, #0xb8]\n" + "ldr d6, [x16, #0xa0]\n" ".inst 0x6f80e8e9 // udot v9.4s, v7.16b, v0.4b[2]\n" ".inst 0x6f81e8ed // udot v13.4s, v7.16b, v1.4b[2]\n" - "mov v6.d[1], x28\n" ".inst 0x6f82e8f1 // udot v17.4s, v7.16b, v2.4b[2]\n" + "mov v6.d[1], x10\n" ".inst 0x6f83e8f5 // udot v21.4s, v7.16b, v3.4b[2]\n" + "ldr x10, [x16, #0xc8]\n" ".inst 0x6f84e8f9 // udot v25.4s, v7.16b, v4.4b[2]\n" + "ldr d7, [x16, #0xb0]\n" ".inst 0x6f80e8ca // udot v10.4s, v6.16b, v0.4b[2]\n" ".inst 0x6f81e8ce // udot v14.4s, v6.16b, v1.4b[2]\n" ".inst 0x6f82e8d2 // udot v18.4s, v6.16b, v2.4b[2]\n" + "mov v7.d[1], x11\n" ".inst 0x6f83e8d6 // udot v22.4s, v6.16b, v3.4b[2]\n" + "ldr x11, [x16, #0xd8]\n" ".inst 0x6f84e8da // udot v26.4s, v6.16b, v4.4b[2]\n" - "ldr d7, [x13, #0xb0]\n" - "ldr d6, [x13, #0xc0]\n" - "ldr x28, [x13, #0xc8]\n" - "mov v7.d[1], x19\n" - "ldr x19, [x13, #0xd8]\n" - "ldr x27, [x9, #0x8]\n" - "mov v6.d[1], x28\n" + "ldr d6, [x16, #0xc0]\n" ".inst 0x6f80e8eb // udot v11.4s, v7.16b, v0.4b[2]\n" ".inst 0x6f81e8ef // udot v15.4s, v7.16b, v1.4b[2]\n" ".inst 0x6f82e8f3 // udot v19.4s, v7.16b, v2.4b[2]\n" + "mov v6.d[1], x10\n" ".inst 0x6f83e8f7 // udot v23.4s, v7.16b, v3.4b[2]\n" + "ldr x10, [x16, #0xe8]\n" ".inst 0x6f84e8fb // udot v27.4s, v7.16b, v4.4b[2]\n" + "ldr d7, [x16, #0xd0]\n" ".inst 0x6fa0e8c8 // udot v8.4s, v6.16b, v0.4b[3]\n" ".inst 0x6fa1e8cc // udot v12.4s, v6.16b, v1.4b[3]\n" ".inst 0x6fa2e8d0 // udot v16.4s, v6.16b, v2.4b[3]\n" + "mov v7.d[1], x11\n" ".inst 0x6fa3e8d4 // udot v20.4s, v6.16b, v3.4b[3]\n" + "ldr x11, [x16, #0xf8]\n" ".inst 0x6fa4e8d8 // udot v24.4s, v6.16b, v4.4b[3]\n" - "ldr d7, [x13, #0xd0]\n" - "ldr d6, [x13, #0xe0]\n" - "ldr x28, [x13, #0xe8]\n" - "mov v7.d[1], x19\n" - "ldr x19, [x13, #0xf8]\n" - "ldr x25, [x26, #0x8]\n" - "mov v6.d[1], x28\n" + "ldr d6, [x16, #0xe0]\n" ".inst 0x6fa0e8e9 // udot v9.4s, v7.16b, v0.4b[3]\n" ".inst 0x6fa1e8ed // udot v13.4s, v7.16b, v1.4b[3]\n" ".inst 0x6fa2e8f1 // udot v17.4s, v7.16b, v2.4b[3]\n" + "mov v6.d[1], x10\n" ".inst 0x6fa3e8f5 // udot v21.4s, v7.16b, v3.4b[3]\n" ".inst 0x6fa4e8f9 // udot v25.4s, v7.16b, v4.4b[3]\n" + "ldr d7, [x16, #0xf0]\n" ".inst 0x6fa0e8ca // udot v10.4s, v6.16b, v0.4b[3]\n" + "add x16, x16, #0x100\n" ".inst 0x6fa1e8ce // udot v14.4s, v6.16b, v1.4b[3]\n" + "ldr x10, [x16, #0x8]\n" ".inst 0x6fa2e8d2 // udot v18.4s, v6.16b, v2.4b[3]\n" + "mov v7.d[1], x11\n" ".inst 0x6fa3e8d6 // udot v22.4s, v6.16b, v3.4b[3]\n" ".inst 0x6fa4e8da // udot v26.4s, v6.16b, v4.4b[3]\n" - "ldr d7, [x13, #0xf0]\n" - "ldr x21, [x24, #0x8]\n" - "add x13, x13, #0x100\n" - "ldr d6, [x13, #0x0]\n" - "mov v7.d[1], x19\n" - "ldr x28, [x13, #0x8]\n" - "ldr x19, [x23, #0x8]\n" + "ldr d6, [x16, #0x0]\n" ".inst 0x6fa0e8eb // udot v11.4s, v7.16b, v0.4b[3]\n" + "ldr d0, [x12, #0x0]\n" ".inst 0x6fa1e8ef // udot v15.4s, v7.16b, v1.4b[3]\n" - "mov v6.d[1], x28\n" + "ldr d1, [x28, #0x0]\n" ".inst 0x6fa2e8f3 // udot v19.4s, v7.16b, v2.4b[3]\n" + "mov v6.d[1], x10\n" ".inst 0x6fa3e8f7 // udot v23.4s, v7.16b, v3.4b[3]\n" + "mov v0.d[1], x9\n" ".inst 0x6fa4e8fb // udot v27.4s, v7.16b, v4.4b[3]\n" - "ldr d0, [x9, #0x0]\n" - "ldr d1, [x26, #0x0]\n" - "ldr d2, [x24, #0x0]\n" - "mov v0.d[1], x27\n" - "ldr d3, [x23, #0x0]\n" - "mov v1.d[1], x25\n" + "mov v1.d[1], x27\n" + "ldr d2, [x26, #0x0]\n" + "ldr d3, [x24, #0x0]\n" "ldr d4, [x22, #0x0]\n" - "mov v2.d[1], x21\n" - "ldr x21, [x22, #0x8]\n" - "mov v3.d[1], x19\n" + "mov v2.d[1], x25\n" + "mov v3.d[1], x23\n" "mov v4.d[1], x21\n" "bge 153b\n" "154:" // Height 5: Multiply loop: Single iteration only ".inst 0x6f80e0c8 // udot v8.4s, v6.16b, v0.4b[0]\n" + "ldr q7, [x16, #0x10]\n" ".inst 0x6f81e0cc // udot v12.4s, v6.16b, v1.4b[0]\n" + "sub x13, x13, #0x10\n" ".inst 0x6f82e0d0 // udot v16.4s, v6.16b, v2.4b[0]\n" + "add x12, x12, #0x10\n" ".inst 0x6f83e0d4 // udot v20.4s, v6.16b, v3.4b[0]\n" + "prfm pldl1keep, [x12, #0x80]\n" ".inst 0x6f84e0d8 // udot v24.4s, v6.16b, v4.4b[0]\n" - "ldr q7, [x13, #0x10]\n" - "ldr q6, [x13, #0x20]\n" - "sub x10, x10, #0x10\n" - "add x9, x9, #0x10\n" - "add x26, x26, #0x10\n" + "ldr q6, [x16, #0x20]\n" ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" + "add x28, x28, #0x10\n" ".inst 0x6f81e0ed // udot v13.4s, v7.16b, v1.4b[0]\n" + "prfm pldl1keep, [x28, #0x80]\n" ".inst 0x6f82e0f1 // udot v17.4s, v7.16b, v2.4b[0]\n" + "add x26, x26, #0x10\n" ".inst 0x6f83e0f5 // udot v21.4s, v7.16b, v3.4b[0]\n" + "prfm pldl1keep, [x26, #0x80]\n" ".inst 0x6f84e0f9 // udot v25.4s, v7.16b, v4.4b[0]\n" + "ldr q7, [x16, #0x30]\n" ".inst 0x6f80e0ca // udot v10.4s, v6.16b, v0.4b[0]\n" + "add x24, x24, #0x10\n" ".inst 0x6f81e0ce // udot v14.4s, v6.16b, v1.4b[0]\n" + "prfm pldl1keep, [x24, #0x80]\n" ".inst 0x6f82e0d2 // udot v18.4s, v6.16b, v2.4b[0]\n" + "add x22, x22, #0x10\n" ".inst 0x6f83e0d6 // udot v22.4s, v6.16b, v3.4b[0]\n" + "prfm pldl1keep, [x22, #0x80]\n" ".inst 0x6f84e0da // udot v26.4s, v6.16b, v4.4b[0]\n" - "prfm pldl1keep, [x9, #0x80]\n" - "add x24, x24, #0x10\n" - "ldr q7, [x13, #0x30]\n" - "add x23, x23, #0x10\n" - "prfm pldl1keep, [x26, #0x80]\n" - "add x22, x22, #0x10\n" - "ldr q6, [x13, #0x40]\n" + "ldr q6, [x16, #0x40]\n" ".inst 0x6f80e0eb // udot v11.4s, v7.16b, v0.4b[0]\n" ".inst 0x6f81e0ef // udot v15.4s, v7.16b, v1.4b[0]\n" ".inst 0x6f82e0f3 // udot v19.4s, v7.16b, v2.4b[0]\n" ".inst 0x6f83e0f7 // udot v23.4s, v7.16b, v3.4b[0]\n" ".inst 0x6f84e0fb // udot v27.4s, v7.16b, v4.4b[0]\n" + "ldr q7, [x16, #0x50]\n" ".inst 0x6fa0e0c8 // udot v8.4s, v6.16b, v0.4b[1]\n" ".inst 0x6fa1e0cc // udot v12.4s, v6.16b, v1.4b[1]\n" ".inst 0x6fa2e0d0 // udot v16.4s, v6.16b, v2.4b[1]\n" ".inst 0x6fa3e0d4 // udot v20.4s, v6.16b, v3.4b[1]\n" ".inst 0x6fa4e0d8 // udot v24.4s, v6.16b, v4.4b[1]\n" - "ldr q7, [x13, #0x50]\n" - "prfm pldl1keep, [x24, #0x80]\n" - "ldr q6, [x13, #0x60]\n" + "ldr q6, [x16, #0x60]\n" ".inst 0x6fa0e0e9 // udot v9.4s, v7.16b, v0.4b[1]\n" ".inst 0x6fa1e0ed // udot v13.4s, v7.16b, v1.4b[1]\n" ".inst 0x6fa2e0f1 // udot v17.4s, v7.16b, v2.4b[1]\n" ".inst 0x6fa3e0f5 // udot v21.4s, v7.16b, v3.4b[1]\n" ".inst 0x6fa4e0f9 // udot v25.4s, v7.16b, v4.4b[1]\n" + "ldr q7, [x16, #0x70]\n" ".inst 0x6fa0e0ca // udot v10.4s, v6.16b, v0.4b[1]\n" ".inst 0x6fa1e0ce // udot v14.4s, v6.16b, v1.4b[1]\n" ".inst 0x6fa2e0d2 // udot v18.4s, v6.16b, v2.4b[1]\n" ".inst 0x6fa3e0d6 // udot v22.4s, v6.16b, v3.4b[1]\n" ".inst 0x6fa4e0da // udot v26.4s, v6.16b, v4.4b[1]\n" - "ldr q7, [x13, #0x70]\n" - "prfm pldl1keep, [x23, #0x80]\n" - "ldr q6, [x13, #0x80]\n" + "ldr q6, [x16, #0x80]\n" ".inst 0x6fa0e0eb // udot v11.4s, v7.16b, v0.4b[1]\n" ".inst 0x6fa1e0ef // udot v15.4s, v7.16b, v1.4b[1]\n" ".inst 0x6fa2e0f3 // udot v19.4s, v7.16b, v2.4b[1]\n" ".inst 0x6fa3e0f7 // udot v23.4s, v7.16b, v3.4b[1]\n" ".inst 0x6fa4e0fb // udot v27.4s, v7.16b, v4.4b[1]\n" + "ldr q7, [x16, #0x90]\n" ".inst 0x6f80e8c8 // udot v8.4s, v6.16b, v0.4b[2]\n" ".inst 0x6f81e8cc // udot v12.4s, v6.16b, v1.4b[2]\n" ".inst 0x6f82e8d0 // udot v16.4s, v6.16b, v2.4b[2]\n" ".inst 0x6f83e8d4 // udot v20.4s, v6.16b, v3.4b[2]\n" ".inst 0x6f84e8d8 // udot v24.4s, v6.16b, v4.4b[2]\n" - "ldr q7, [x13, #0x90]\n" - "prfm pldl1keep, [x22, #0x80]\n" - "ldr q6, [x13, #0xa0]\n" + "ldr q6, [x16, #0xa0]\n" ".inst 0x6f80e8e9 // udot v9.4s, v7.16b, v0.4b[2]\n" ".inst 0x6f81e8ed // udot v13.4s, v7.16b, v1.4b[2]\n" ".inst 0x6f82e8f1 // udot v17.4s, v7.16b, v2.4b[2]\n" ".inst 0x6f83e8f5 // udot v21.4s, v7.16b, v3.4b[2]\n" ".inst 0x6f84e8f9 // udot v25.4s, v7.16b, v4.4b[2]\n" + "ldr q7, [x16, #0xb0]\n" ".inst 0x6f80e8ca // udot v10.4s, v6.16b, v0.4b[2]\n" ".inst 0x6f81e8ce // udot v14.4s, v6.16b, v1.4b[2]\n" ".inst 0x6f82e8d2 // udot v18.4s, v6.16b, v2.4b[2]\n" ".inst 0x6f83e8d6 // udot v22.4s, v6.16b, v3.4b[2]\n" ".inst 0x6f84e8da // udot v26.4s, v6.16b, v4.4b[2]\n" - "ldr q7, [x13, #0xb0]\n" - "ldr q6, [x13, #0xc0]\n" + "ldr q6, [x16, #0xc0]\n" ".inst 0x6f80e8eb // udot v11.4s, v7.16b, v0.4b[2]\n" ".inst 0x6f81e8ef // udot v15.4s, v7.16b, v1.4b[2]\n" ".inst 0x6f82e8f3 // udot v19.4s, v7.16b, v2.4b[2]\n" ".inst 0x6f83e8f7 // udot v23.4s, v7.16b, v3.4b[2]\n" ".inst 0x6f84e8fb // udot v27.4s, v7.16b, v4.4b[2]\n" + "ldr q7, [x16, #0xd0]\n" ".inst 0x6fa0e8c8 // udot v8.4s, v6.16b, v0.4b[3]\n" ".inst 0x6fa1e8cc // udot v12.4s, v6.16b, v1.4b[3]\n" ".inst 0x6fa2e8d0 // udot v16.4s, v6.16b, v2.4b[3]\n" ".inst 0x6fa3e8d4 // udot v20.4s, v6.16b, v3.4b[3]\n" ".inst 0x6fa4e8d8 // udot v24.4s, v6.16b, v4.4b[3]\n" - "ldr q7, [x13, #0xd0]\n" - "ldr q6, [x13, #0xe0]\n" + "ldr q6, [x16, #0xe0]\n" ".inst 0x6fa0e8e9 // udot v9.4s, v7.16b, v0.4b[3]\n" ".inst 0x6fa1e8ed // udot v13.4s, v7.16b, v1.4b[3]\n" ".inst 0x6fa2e8f1 // udot v17.4s, v7.16b, v2.4b[3]\n" ".inst 0x6fa3e8f5 // udot v21.4s, v7.16b, v3.4b[3]\n" ".inst 0x6fa4e8f9 // udot v25.4s, v7.16b, v4.4b[3]\n" + "ldr q7, [x16, #0xf0]\n" ".inst 0x6fa0e8ca // udot v10.4s, v6.16b, v0.4b[3]\n" + "add x16, x16, #0x100\n" ".inst 0x6fa1e8ce // udot v14.4s, v6.16b, v1.4b[3]\n" ".inst 0x6fa2e8d2 // udot v18.4s, v6.16b, v2.4b[3]\n" ".inst 0x6fa3e8d6 // udot v22.4s, v6.16b, v3.4b[3]\n" ".inst 0x6fa4e8da // udot v26.4s, v6.16b, v4.4b[3]\n" - "ldr q7, [x13, #0xf0]\n" - "add x13, x13, #0x100\n" ".inst 0x6fa0e8eb // udot v11.4s, v7.16b, v0.4b[3]\n" ".inst 0x6fa1e8ef // udot v15.4s, v7.16b, v1.4b[3]\n" ".inst 0x6fa2e8f3 // udot v19.4s, v7.16b, v2.4b[3]\n" ".inst 0x6fa3e8f7 // udot v23.4s, v7.16b, v3.4b[3]\n" ".inst 0x6fa4e8fb // udot v27.4s, v7.16b, v4.4b[3]\n" "155:" // Height 5: Multiply loop: Main loop skip - "cbz x10, 160f\n" - "cmp x10, #0x4\n" + "cbz x13, 160f\n" + "cmp x13, #0x4\n" "blt 157f\n" "156:" // Height 5: Multiply loop: Odd block loop - "ldr s0, [x9], #0x4\n" - "sub x10, x10, #0x4\n" - "ldr s1, [x26], #0x4\n" - "cmp x10, #0x4\n" - "ldr s2, [x24], #0x4\n" - "ldr s3, [x23], #0x4\n" + "ldr s0, [x12], #0x4\n" + "sub x13, x13, #0x4\n" + "ldr s1, [x28], #0x4\n" + "cmp x13, #0x4\n" + "ldr s2, [x26], #0x4\n" + "ldr s3, [x24], #0x4\n" "ldr s4, [x22], #0x4\n" - "ldr q6, [x13, #0x0]\n" - "ldr q7, [x13, #0x10]\n" + "ldr q6, [x16, #0x0]\n" + "ldr q7, [x16, #0x10]\n" ".inst 0x6f80e0c8 // udot v8.4s, v6.16b, v0.4b[0]\n" ".inst 0x6f81e0cc // udot v12.4s, v6.16b, v1.4b[0]\n" ".inst 0x6f82e0d0 // udot v16.4s, v6.16b, v2.4b[0]\n" ".inst 0x6f83e0d4 // udot v20.4s, v6.16b, v3.4b[0]\n" ".inst 0x6f84e0d8 // udot v24.4s, v6.16b, v4.4b[0]\n" + "ldr q6, [x16, #0x20]\n" ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" ".inst 0x6f81e0ed // udot v13.4s, v7.16b, v1.4b[0]\n" ".inst 0x6f82e0f1 // udot v17.4s, v7.16b, v2.4b[0]\n" ".inst 0x6f83e0f5 // udot v21.4s, v7.16b, v3.4b[0]\n" ".inst 0x6f84e0f9 // udot v25.4s, v7.16b, v4.4b[0]\n" - "ldr q6, [x13, #0x20]\n" - "ldr q7, [x13, #0x30]\n" - "add x13, x13, #0x40\n" + "ldr q7, [x16, #0x30]\n" ".inst 0x6f80e0ca // udot v10.4s, v6.16b, v0.4b[0]\n" + "add x16, x16, #0x40\n" ".inst 0x6f81e0ce // udot v14.4s, v6.16b, v1.4b[0]\n" ".inst 0x6f82e0d2 // udot v18.4s, v6.16b, v2.4b[0]\n" ".inst 0x6f83e0d6 // udot v22.4s, v6.16b, v3.4b[0]\n" @@ -2454,44 +2454,44 @@ void a64_hybrid_u8u32_dot_6x16_a55 ( ".inst 0x6f83e0f7 // udot v23.4s, v7.16b, v3.4b[0]\n" ".inst 0x6f84e0fb // udot v27.4s, v7.16b, v4.4b[0]\n" "bge 156b\n" - "cbz x10, 160f\n" + "cbz x13, 160f\n" "157:" // Height 5: Multiply loop: Skip odd blocks - "tbz x10, #1, 158f\n" - "ldr h0, [x9], #0x2\n" - "ldr h1, [x26], #0x2\n" - "ldr h2, [x24], #0x2\n" - "ldr h3, [x23], #0x2\n" + "tbz x13, #1, 158f\n" + "ldr h0, [x12], #0x2\n" + "ldr h1, [x28], #0x2\n" + "ldr h2, [x26], #0x2\n" + "ldr h3, [x24], #0x2\n" "ldr h4, [x22], #0x2\n" - "tbz x10, #0, 159f\n" - "ld1 { v0.b }[2], [x9]\n" - "ld1 { v1.b }[2], [x26]\n" - "ld1 { v2.b }[2], [x24]\n" - "ld1 { v3.b }[2], [x23]\n" + "tbz x13, #0, 159f\n" + "ld1 { v0.b }[2], [x12]\n" + "ld1 { v1.b }[2], [x28]\n" + "ld1 { v2.b }[2], [x26]\n" + "ld1 { v3.b }[2], [x24]\n" "ld1 { v4.b }[2], [x22]\n" "b 159f\n" "158:" // Height 5: Multiply loop: Ragged operand read: partial_1_0 - "ldr b0, [x9, #0x0]\n" - "ldr b1, [x26, #0x0]\n" - "ldr b2, [x24, #0x0]\n" - "ldr b3, [x23, #0x0]\n" + "ldr b0, [x12, #0x0]\n" + "ldr b1, [x28, #0x0]\n" + "ldr b2, [x26, #0x0]\n" + "ldr b3, [x24, #0x0]\n" "ldr b4, [x22, #0x0]\n" "159:" // Height 5: Multiply loop: Ragged operand read: Done - "ldr q6, [x13, #0x0]\n" - "ldr q7, [x13, #0x10]\n" + "ldr q6, [x16, #0x0]\n" + "ldr q7, [x16, #0x10]\n" ".inst 0x6f80e0c8 // udot v8.4s, v6.16b, v0.4b[0]\n" ".inst 0x6f81e0cc // udot v12.4s, v6.16b, v1.4b[0]\n" ".inst 0x6f82e0d0 // udot v16.4s, v6.16b, v2.4b[0]\n" ".inst 0x6f83e0d4 // udot v20.4s, v6.16b, v3.4b[0]\n" ".inst 0x6f84e0d8 // udot v24.4s, v6.16b, v4.4b[0]\n" + "ldr q6, [x16, #0x20]\n" ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" ".inst 0x6f81e0ed // udot v13.4s, v7.16b, v1.4b[0]\n" ".inst 0x6f82e0f1 // udot v17.4s, v7.16b, v2.4b[0]\n" ".inst 0x6f83e0f5 // udot v21.4s, v7.16b, v3.4b[0]\n" ".inst 0x6f84e0f9 // udot v25.4s, v7.16b, v4.4b[0]\n" - "ldr q6, [x13, #0x20]\n" - "ldr q7, [x13, #0x30]\n" - "add x13, x13, #0x40\n" + "ldr q7, [x16, #0x30]\n" ".inst 0x6f80e0ca // udot v10.4s, v6.16b, v0.4b[0]\n" + "add x16, x16, #0x40\n" ".inst 0x6f81e0ce // udot v14.4s, v6.16b, v1.4b[0]\n" ".inst 0x6f82e0d2 // udot v18.4s, v6.16b, v2.4b[0]\n" ".inst 0x6f83e0d6 // udot v22.4s, v6.16b, v3.4b[0]\n" @@ -2503,13 +2503,13 @@ void a64_hybrid_u8u32_dot_6x16_a55 ( ".inst 0x6f84e0fb // udot v27.4s, v7.16b, v4.4b[0]\n" "160:" // Height 5: Multiply loop: No odd multiplies "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" - "add x11, x11, #0x1\n" - "cmp x11, x19\n" + "add x14, x14, #0x1\n" + "cmp x14, x19\n" "bne 150b\n" "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "prfm pstl1keep, [x12, #0x0]\n" - "cmp x14, #0x10\n" - "add x23, x12, x19, LSL #2\n" + "prfm pstl1keep, [x15, #0x0]\n" + "cmp x17, #0x10\n" + "add x23, x15, x19, LSL #2\n" "prfm pstl1keep, [x23, #0x0]\n" "add x22, x23, x19, LSL #2\n" "prfm pstl1keep, [x22, #0x0]\n" @@ -2518,9 +2518,9 @@ void a64_hybrid_u8u32_dot_6x16_a55 ( "add x20, x21, x19, LSL #2\n" "prfm pstl1keep, [x20, #0x0]\n" "bge 169f\n" - "tbz x14, #3, 164f\n" - "st1 { v8.4s }, [x12], #0x10\n" - "st1 { v9.4s }, [x12], #0x10\n" + "tbz x17, #3, 164f\n" + "st1 { v8.4s }, [x15], #0x10\n" + "st1 { v9.4s }, [x15], #0x10\n" "st1 { v12.4s }, [x23], #0x10\n" "st1 { v13.4s }, [x23], #0x10\n" "st1 { v16.4s }, [x22], #0x10\n" @@ -2529,99 +2529,99 @@ void a64_hybrid_u8u32_dot_6x16_a55 ( "st1 { v21.4s }, [x21], #0x10\n" "st1 { v24.4s }, [x20], #0x10\n" "st1 { v25.4s }, [x20], #0x10\n" - "tbz x14, #2, 162f\n" - "st1 { v10.4s }, [x12], #0x10\n" + "tbz x17, #2, 162f\n" + "st1 { v10.4s }, [x15], #0x10\n" "st1 { v14.4s }, [x23], #0x10\n" "st1 { v18.4s }, [x22], #0x10\n" "st1 { v22.4s }, [x21], #0x10\n" "st1 { v26.4s }, [x20], #0x10\n" - "tbz x14, #1, 161f\n" - "str d11, [x12], #0x8\n" + "tbz x17, #1, 161f\n" + "str d11, [x15], #0x8\n" "str d15, [x23], #0x8\n" "str d19, [x22], #0x8\n" "str d23, [x21], #0x8\n" "str d27, [x20], #0x8\n" - "tbz x14, #0, 168f\n" - "st1 { v11.s }[2], [x12]\n" + "tbz x17, #0, 168f\n" + "st1 { v11.s }[2], [x15]\n" "st1 { v15.s }[2], [x23]\n" "st1 { v19.s }[2], [x22]\n" "st1 { v23.s }[2], [x21]\n" "st1 { v27.s }[2], [x20]\n" "b 168f\n" "161:" // Height 5: Partial direct writeback: partial_1_12 - "tbz x14, #0, 168f\n" - "str s11, [x12, #0x0]\n" + "tbz x17, #0, 168f\n" + "str s11, [x15, #0x0]\n" "str s15, [x23, #0x0]\n" "str s19, [x22, #0x0]\n" "str s23, [x21, #0x0]\n" "str s27, [x20, #0x0]\n" "b 168f\n" "162:" // Height 5: Partial direct writeback: partial_2_8 - "tbz x14, #1, 163f\n" - "str d10, [x12], #0x8\n" + "tbz x17, #1, 163f\n" + "str d10, [x15], #0x8\n" "str d14, [x23], #0x8\n" "str d18, [x22], #0x8\n" "str d22, [x21], #0x8\n" "str d26, [x20], #0x8\n" - "tbz x14, #0, 168f\n" - "st1 { v10.s }[2], [x12]\n" + "tbz x17, #0, 168f\n" + "st1 { v10.s }[2], [x15]\n" "st1 { v14.s }[2], [x23]\n" "st1 { v18.s }[2], [x22]\n" "st1 { v22.s }[2], [x21]\n" "st1 { v26.s }[2], [x20]\n" "b 168f\n" "163:" // Height 5: Partial direct writeback: partial_1_8 - "tbz x14, #0, 168f\n" - "str s10, [x12, #0x0]\n" + "tbz x17, #0, 168f\n" + "str s10, [x15, #0x0]\n" "str s14, [x23, #0x0]\n" "str s18, [x22, #0x0]\n" "str s22, [x21, #0x0]\n" "str s26, [x20, #0x0]\n" "b 168f\n" "164:" // Height 5: Partial direct writeback: partial_4_0 - "tbz x14, #2, 166f\n" - "st1 { v8.4s }, [x12], #0x10\n" + "tbz x17, #2, 166f\n" + "st1 { v8.4s }, [x15], #0x10\n" "st1 { v12.4s }, [x23], #0x10\n" "st1 { v16.4s }, [x22], #0x10\n" "st1 { v20.4s }, [x21], #0x10\n" "st1 { v24.4s }, [x20], #0x10\n" - "tbz x14, #1, 165f\n" - "str d9, [x12], #0x8\n" + "tbz x17, #1, 165f\n" + "str d9, [x15], #0x8\n" "str d13, [x23], #0x8\n" "str d17, [x22], #0x8\n" "str d21, [x21], #0x8\n" "str d25, [x20], #0x8\n" - "tbz x14, #0, 168f\n" - "st1 { v9.s }[2], [x12]\n" + "tbz x17, #0, 168f\n" + "st1 { v9.s }[2], [x15]\n" "st1 { v13.s }[2], [x23]\n" "st1 { v17.s }[2], [x22]\n" "st1 { v21.s }[2], [x21]\n" "st1 { v25.s }[2], [x20]\n" "b 168f\n" "165:" // Height 5: Partial direct writeback: partial_1_4 - "tbz x14, #0, 168f\n" - "str s9, [x12, #0x0]\n" + "tbz x17, #0, 168f\n" + "str s9, [x15, #0x0]\n" "str s13, [x23, #0x0]\n" "str s17, [x22, #0x0]\n" "str s21, [x21, #0x0]\n" "str s25, [x20, #0x0]\n" "b 168f\n" "166:" // Height 5: Partial direct writeback: partial_2_0 - "tbz x14, #1, 167f\n" - "str d8, [x12], #0x8\n" + "tbz x17, #1, 167f\n" + "str d8, [x15], #0x8\n" "str d12, [x23], #0x8\n" "str d16, [x22], #0x8\n" "str d20, [x21], #0x8\n" "str d24, [x20], #0x8\n" - "tbz x14, #0, 168f\n" - "st1 { v8.s }[2], [x12]\n" + "tbz x17, #0, 168f\n" + "st1 { v8.s }[2], [x15]\n" "st1 { v12.s }[2], [x23]\n" "st1 { v16.s }[2], [x22]\n" "st1 { v20.s }[2], [x21]\n" "st1 { v24.s }[2], [x20]\n" "b 168f\n" "167:" // Height 5: Partial direct writeback: partial_1_0 - "str s8, [x12, #0x0]\n" + "str s8, [x15, #0x0]\n" "str s12, [x23, #0x0]\n" "str s16, [x22, #0x0]\n" "str s20, [x21, #0x0]\n" @@ -2629,11 +2629,11 @@ void a64_hybrid_u8u32_dot_6x16_a55 ( "168:" // Height 5: Partial direct writeback: Done "b 170f\n" "169:" // Height 5: Full writeback - "str q8, [x12, #0x0]\n" - "str q9, [x12, #0x10]\n" - "str q10, [x12, #0x20]\n" - "str q11, [x12, #0x30]\n" - "add x12, x12, #0x40\n" + "str q8, [x15, #0x0]\n" + "str q9, [x15, #0x10]\n" + "str q10, [x15, #0x20]\n" + "str q11, [x15, #0x30]\n" + "add x15, x15, #0x40\n" "str q12, [x23, #0x0]\n" "str q13, [x23, #0x10]\n" "str q14, [x23, #0x20]\n" @@ -2651,31 +2651,31 @@ void a64_hybrid_u8u32_dot_6x16_a55 ( "str q26, [x20, #0x20]\n" "str q27, [x20, #0x30]\n" "170:" // Height 5: Writeback done - "subs x14, x14, #0x10\n" + "subs x17, x17, #0x10\n" "bgt 138b\n" "b 206f\n" "171:" // Height 6 - "ldr x14, [%x[args_ptr], %[offsetof_N]]\n" - "mov x12, %x[output_ptr]\n" - "ldr x13, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "ldr x17, [%x[args_ptr], %[offsetof_N]]\n" + "mov x15, %x[output_ptr]\n" + "ldr x16, [%x[args_ptr], %[offsetof_B_ptr]]\n" "mov x20, #0x18\n" "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" "madd %x[output_ptr], x19, x20, %x[output_ptr]\n" "172:" // Height 6: Column loop "tbz %x[flags], #0, 182f\n" "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "cmp x14, #0x10\n" - "add x23, x12, x19, LSL #2\n" + "cmp x17, #0x10\n" + "add x23, x15, x19, LSL #2\n" "add x22, x23, x19, LSL #2\n" "add x21, x22, x19, LSL #2\n" "add x20, x21, x19, LSL #2\n" "add x19, x20, x19, LSL #2\n" "bge 181f\n" - "tbz x14, #3, 176f\n" - "ld1 { v8.4s }, [x12], #0x10\n" + "tbz x17, #3, 176f\n" + "ld1 { v8.4s }, [x15], #0x10\n" "ld1 { v12.4s }, [x23], #0x10\n" "ld1 { v16.4s }, [x22], #0x10\n" - "ld1 { v9.4s }, [x12], #0x10\n" + "ld1 { v9.4s }, [x15], #0x10\n" "ld1 { v13.4s }, [x23], #0x10\n" "ld1 { v17.4s }, [x22], #0x10\n" "ld1 { v20.4s }, [x21], #0x10\n" @@ -2684,23 +2684,23 @@ void a64_hybrid_u8u32_dot_6x16_a55 ( "ld1 { v21.4s }, [x21], #0x10\n" "ld1 { v25.4s }, [x20], #0x10\n" "ld1 { v29.4s }, [x19], #0x10\n" - "tbz x14, #2, 174f\n" - "ld1 { v10.4s }, [x12], #0x10\n" + "tbz x17, #2, 174f\n" + "ld1 { v10.4s }, [x15], #0x10\n" "ld1 { v14.4s }, [x23], #0x10\n" "ld1 { v18.4s }, [x22], #0x10\n" "ld1 { v22.4s }, [x21], #0x10\n" "ld1 { v26.4s }, [x20], #0x10\n" "ld1 { v30.4s }, [x19], #0x10\n" - "tbz x14, #1, 173f\n" - "ldr d11, [x12], #0x8\n" + "tbz x17, #1, 173f\n" + "ldr d11, [x15], #0x8\n" "mov x24, #0x38\n" "ldr d15, [x23], #0x8\n" "ldr d19, [x22], #0x8\n" "ldr d23, [x21], #0x8\n" "ldr d27, [x20], #0x8\n" "ldr d31, [x19], #0x8\n" - "tbz x14, #0, 180f\n" - "ld1 { v11.s }[2], [x12]\n" + "tbz x17, #0, 180f\n" + "ld1 { v11.s }[2], [x15]\n" "ld1 { v15.s }[2], [x23]\n" "ld1 { v19.s }[2], [x22]\n" "ld1 { v23.s }[2], [x21]\n" @@ -2709,8 +2709,8 @@ void a64_hybrid_u8u32_dot_6x16_a55 ( "b 180f\n" "173:" // Height 6: Partial accumulate: partial_1_12 "mov x24, #0x30\n" - "tbz x14, #0, 180f\n" - "ldr s11, [x12, #0x0]\n" + "tbz x17, #0, 180f\n" + "ldr s11, [x15, #0x0]\n" "ldr s15, [x23, #0x0]\n" "ldr s19, [x22, #0x0]\n" "ldr s23, [x21, #0x0]\n" @@ -2718,16 +2718,16 @@ void a64_hybrid_u8u32_dot_6x16_a55 ( "ldr s31, [x19, #0x0]\n" "b 180f\n" "174:" // Height 6: Partial accumulate: partial_2_8 - "tbz x14, #1, 175f\n" - "ldr d10, [x12], #0x8\n" + "tbz x17, #1, 175f\n" + "ldr d10, [x15], #0x8\n" "ldr d14, [x23], #0x8\n" "mov x24, #0x28\n" "ldr d18, [x22], #0x8\n" "ldr d22, [x21], #0x8\n" "ldr d26, [x20], #0x8\n" "ldr d30, [x19], #0x8\n" - "tbz x14, #0, 180f\n" - "ld1 { v10.s }[2], [x12]\n" + "tbz x17, #0, 180f\n" + "ld1 { v10.s }[2], [x15]\n" "ld1 { v14.s }[2], [x23]\n" "ld1 { v18.s }[2], [x22]\n" "ld1 { v22.s }[2], [x21]\n" @@ -2736,8 +2736,8 @@ void a64_hybrid_u8u32_dot_6x16_a55 ( "b 180f\n" "175:" // Height 6: Partial accumulate: partial_1_8 "mov x24, #0x20\n" - "tbz x14, #0, 180f\n" - "ldr s10, [x12, #0x0]\n" + "tbz x17, #0, 180f\n" + "ldr s10, [x15, #0x0]\n" "ldr s14, [x23, #0x0]\n" "ldr s18, [x22, #0x0]\n" "ldr s22, [x21, #0x0]\n" @@ -2745,23 +2745,23 @@ void a64_hybrid_u8u32_dot_6x16_a55 ( "ldr s30, [x19, #0x0]\n" "b 180f\n" "176:" // Height 6: Partial accumulate: partial_4_0 - "tbz x14, #2, 178f\n" - "ld1 { v8.4s }, [x12], #0x10\n" + "tbz x17, #2, 178f\n" + "ld1 { v8.4s }, [x15], #0x10\n" "ld1 { v12.4s }, [x23], #0x10\n" "ld1 { v16.4s }, [x22], #0x10\n" "ld1 { v20.4s }, [x21], #0x10\n" "ld1 { v24.4s }, [x20], #0x10\n" "ld1 { v28.4s }, [x19], #0x10\n" - "tbz x14, #1, 177f\n" - "ldr d9, [x12], #0x8\n" + "tbz x17, #1, 177f\n" + "ldr d9, [x15], #0x8\n" "mov x24, #0x18\n" "ldr d13, [x23], #0x8\n" "ldr d17, [x22], #0x8\n" "ldr d21, [x21], #0x8\n" "ldr d25, [x20], #0x8\n" "ldr d29, [x19], #0x8\n" - "tbz x14, #0, 180f\n" - "ld1 { v9.s }[2], [x12]\n" + "tbz x17, #0, 180f\n" + "ld1 { v9.s }[2], [x15]\n" "ld1 { v13.s }[2], [x23]\n" "ld1 { v17.s }[2], [x22]\n" "ld1 { v21.s }[2], [x21]\n" @@ -2770,8 +2770,8 @@ void a64_hybrid_u8u32_dot_6x16_a55 ( "b 180f\n" "177:" // Height 6: Partial accumulate: partial_1_4 "mov x24, #0x10\n" - "tbz x14, #0, 180f\n" - "ldr s9, [x12, #0x0]\n" + "tbz x17, #0, 180f\n" + "ldr s9, [x15, #0x0]\n" "ldr s13, [x23, #0x0]\n" "ldr s17, [x22, #0x0]\n" "ldr s21, [x21, #0x0]\n" @@ -2779,16 +2779,16 @@ void a64_hybrid_u8u32_dot_6x16_a55 ( "ldr s29, [x19, #0x0]\n" "b 180f\n" "178:" // Height 6: Partial accumulate: partial_2_0 - "tbz x14, #1, 179f\n" - "ldr d8, [x12], #0x8\n" + "tbz x17, #1, 179f\n" + "ldr d8, [x15], #0x8\n" "ldr d12, [x23], #0x8\n" "mov x24, #0x8\n" "ldr d16, [x22], #0x8\n" "ldr d20, [x21], #0x8\n" "ldr d24, [x20], #0x8\n" "ldr d28, [x19], #0x8\n" - "tbz x14, #0, 180f\n" - "ld1 { v8.s }[2], [x12]\n" + "tbz x17, #0, 180f\n" + "ld1 { v8.s }[2], [x15]\n" "ld1 { v12.s }[2], [x23]\n" "ld1 { v16.s }[2], [x22]\n" "ld1 { v20.s }[2], [x21]\n" @@ -2796,7 +2796,7 @@ void a64_hybrid_u8u32_dot_6x16_a55 ( "ld1 { v28.s }[2], [x19]\n" "b 180f\n" "179:" // Height 6: Partial accumulate: partial_1_0 - "ldr s8, [x12, #0x0]\n" + "ldr s8, [x15, #0x0]\n" "mov x24, #0x0\n" "ldr s12, [x23, #0x0]\n" "ldr s16, [x22, #0x0]\n" @@ -2804,13 +2804,13 @@ void a64_hybrid_u8u32_dot_6x16_a55 ( "ldr s24, [x20, #0x0]\n" "ldr s28, [x19, #0x0]\n" "180:" // Height 6: Partial accumulate: Done - "sub x12, x12, x24\n" + "sub x15, x15, x24\n" "b 183f\n" "181:" // Height 6: full accumulate - "ldr q8, [x12, #0x0]\n" - "ldr q9, [x12, #0x10]\n" - "ldr q10, [x12, #0x20]\n" - "ldr q11, [x12, #0x30]\n" + "ldr q8, [x15, #0x0]\n" + "ldr q9, [x15, #0x10]\n" + "ldr q10, [x15, #0x20]\n" + "ldr q11, [x15, #0x30]\n" "ldr q12, [x23, #0x0]\n" "ldr q13, [x23, #0x10]\n" "ldr q14, [x23, #0x20]\n" @@ -2858,347 +2858,347 @@ void a64_hybrid_u8u32_dot_6x16_a55 ( "movi v30.4s, #0x0\n" "movi v31.4s, #0x0\n" "183:" // Height 6: setup done - "mov x11, #0x0\n" + "mov x14, #0x0\n" "184:" // Height 6: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w10, [x20, x11, LSL #0x2]\n" + "ldr w13, [x20, x14, LSL #0x2]\n" "tbz %x[flags], #3, 185f\n" - "ldr x20, [%x[input_ptr], x11, LSL #0x3]\n" + "ldr x20, [%x[input_ptr], x14, LSL #0x3]\n" "add x20, x20, x19, LSL #3\n" - "ldr x9, [x20, #0x0]\n" - "ldr x26, [x20, #0x8]\n" - "ldr x24, [x20, #0x10]\n" - "ldr x23, [x20, #0x18]\n" + "ldr x12, [x20, #0x0]\n" + "ldr x28, [x20, #0x8]\n" + "ldr x26, [x20, #0x10]\n" + "ldr x24, [x20, #0x18]\n" "ldr x22, [x20, #0x20]\n" "ldr x20, [x20, #0x28]\n" - "cbnz x11, 186f\n" + "cbnz x14, 186f\n" "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x9, x9, x19\n" + "add x12, x12, x19\n" + "add x28, x28, x19\n" "add x26, x26, x19\n" "add x24, x24, x19\n" - "add x23, x23, x19\n" "add x22, x22, x19\n" "add x20, x20, x19\n" "b 186f\n" "185:" // Height 6: setup direct input - "mov x9, %x[input_ptr]\n" - "add x26, x9, x19\n" + "mov x12, %x[input_ptr]\n" + "add x28, x12, x19\n" + "add x26, x28, x19\n" "add x24, x26, x19\n" - "add x23, x24, x19\n" - "add x22, x23, x19\n" + "add x22, x24, x19\n" "add x20, x22, x19\n" "186:" // Height 6: input setup done - "cmp x10, #0x10\n" + "cmp x13, #0x10\n" "blt 189f\n" - "ldr q0, [x9, #0x0]\n" - "ldr q1, [x26, #0x0]\n" - "cmp x10, #0x20\n" - "ldr q2, [x24, #0x0]\n" - "ldr q3, [x23, #0x0]\n" + "ldr q0, [x12, #0x0]\n" + "ldr q1, [x28, #0x0]\n" + "cmp x13, #0x20\n" + "ldr q2, [x26, #0x0]\n" + "ldr q3, [x24, #0x0]\n" "ldr q4, [x22, #0x0]\n" "ldr q5, [x20, #0x0]\n" - "ldr q6, [x13, #0x0]\n" + "ldr q6, [x16, #0x0]\n" "blt 188f\n" "187:" // Height 6: Multiply loop: Main loop head ".inst 0x6f80e0c8 // udot v8.4s, v6.16b, v0.4b[0]\n" + "ldr d7, [x16, #0x10]\n" ".inst 0x6f81e0cc // udot v12.4s, v6.16b, v1.4b[0]\n" + "ldr x11, [x16, #0x18]\n" ".inst 0x6f82e0d0 // udot v16.4s, v6.16b, v2.4b[0]\n" + "ldr x10, [x16, #0x28]\n" ".inst 0x6f83e0d4 // udot v20.4s, v6.16b, v3.4b[0]\n" + "add x12, x12, #0x10\n" ".inst 0x6f84e0d8 // udot v24.4s, v6.16b, v4.4b[0]\n" + "mov v7.d[1], x11\n" ".inst 0x6f85e0dc // udot v28.4s, v6.16b, v5.4b[0]\n" - "ldr d7, [x13, #0x10]\n" - "add x9, x9, #0x10\n" - "ldr x19, [x13, #0x18]\n" - "add x26, x26, #0x10\n" - "prfm pldl1keep, [x9, #0x80]\n" - "add x24, x24, #0x10\n" - "ldr d6, [x13, #0x20]\n" - "add x23, x23, #0x10\n" - "mov v7.d[1], x19\n" - "ldr x28, [x13, #0x28]\n" - "ldr x19, [x13, #0x38]\n" - "add x22, x22, #0x10\n" + "prfm pldl1keep, [x12, #0x80]\n" ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" + "ldr d6, [x16, #0x20]\n" ".inst 0x6f81e0ed // udot v13.4s, v7.16b, v1.4b[0]\n" - "mov v6.d[1], x28\n" + "ldr x11, [x16, #0x38]\n" ".inst 0x6f82e0f1 // udot v17.4s, v7.16b, v2.4b[0]\n" + "ldr x9, [x12, #0x8]\n" ".inst 0x6f83e0f5 // udot v21.4s, v7.16b, v3.4b[0]\n" + "mov v6.d[1], x10\n" ".inst 0x6f84e0f9 // udot v25.4s, v7.16b, v4.4b[0]\n" + "ldr x10, [x16, #0x48]\n" ".inst 0x6f85e0fd // udot v29.4s, v7.16b, v5.4b[0]\n" + "ldr d7, [x16, #0x30]\n" ".inst 0x6f80e0ca // udot v10.4s, v6.16b, v0.4b[0]\n" + "add x28, x28, #0x10\n" ".inst 0x6f81e0ce // udot v14.4s, v6.16b, v1.4b[0]\n" + "prfm pldl1keep, [x28, #0x80]\n" ".inst 0x6f82e0d2 // udot v18.4s, v6.16b, v2.4b[0]\n" + "mov v7.d[1], x11\n" ".inst 0x6f83e0d6 // udot v22.4s, v6.16b, v3.4b[0]\n" + "ldr x11, [x16, #0x58]\n" ".inst 0x6f84e0da // udot v26.4s, v6.16b, v4.4b[0]\n" + "ldr x27, [x28, #0x8]\n" ".inst 0x6f85e0de // udot v30.4s, v6.16b, v5.4b[0]\n" - "ldr d7, [x13, #0x30]\n" - "prfm pldl1keep, [x26, #0x80]\n" - "add x20, x20, #0x10\n" - "ldr d6, [x13, #0x40]\n" - "sub x10, x10, #0x10\n" - "mov v7.d[1], x19\n" - "ldr x28, [x13, #0x48]\n" - "ldr x19, [x13, #0x58]\n" - "cmp x10, #0x20\n" + "ldr d6, [x16, #0x40]\n" ".inst 0x6f80e0eb // udot v11.4s, v7.16b, v0.4b[0]\n" + "add x26, x26, #0x10\n" ".inst 0x6f81e0ef // udot v15.4s, v7.16b, v1.4b[0]\n" - "mov v6.d[1], x28\n" + "prfm pldl1keep, [x26, #0x80]\n" ".inst 0x6f82e0f3 // udot v19.4s, v7.16b, v2.4b[0]\n" + "mov v6.d[1], x10\n" ".inst 0x6f83e0f7 // udot v23.4s, v7.16b, v3.4b[0]\n" + "ldr x10, [x16, #0x68]\n" ".inst 0x6f84e0fb // udot v27.4s, v7.16b, v4.4b[0]\n" + "ldr x25, [x26, #0x8]\n" ".inst 0x6f85e0ff // udot v31.4s, v7.16b, v5.4b[0]\n" + "ldr d7, [x16, #0x50]\n" ".inst 0x6fa0e0c8 // udot v8.4s, v6.16b, v0.4b[1]\n" + "add x24, x24, #0x10\n" ".inst 0x6fa1e0cc // udot v12.4s, v6.16b, v1.4b[1]\n" + "prfm pldl1keep, [x24, #0x80]\n" ".inst 0x6fa2e0d0 // udot v16.4s, v6.16b, v2.4b[1]\n" + "mov v7.d[1], x11\n" ".inst 0x6fa3e0d4 // udot v20.4s, v6.16b, v3.4b[1]\n" + "ldr x11, [x16, #0x78]\n" ".inst 0x6fa4e0d8 // udot v24.4s, v6.16b, v4.4b[1]\n" + "ldr x23, [x24, #0x8]\n" ".inst 0x6fa5e0dc // udot v28.4s, v6.16b, v5.4b[1]\n" - "ldr d7, [x13, #0x50]\n" - "prfm pldl1keep, [x24, #0x80]\n" - "ldr d6, [x13, #0x60]\n" - "mov v7.d[1], x19\n" - "ldr x28, [x13, #0x68]\n" - "ldr x19, [x13, #0x78]\n" + "ldr d6, [x16, #0x60]\n" ".inst 0x6fa0e0e9 // udot v9.4s, v7.16b, v0.4b[1]\n" + "add x22, x22, #0x10\n" ".inst 0x6fa1e0ed // udot v13.4s, v7.16b, v1.4b[1]\n" - "mov v6.d[1], x28\n" + "prfm pldl1keep, [x22, #0x80]\n" ".inst 0x6fa2e0f1 // udot v17.4s, v7.16b, v2.4b[1]\n" + "mov v6.d[1], x10\n" ".inst 0x6fa3e0f5 // udot v21.4s, v7.16b, v3.4b[1]\n" + "ldr x10, [x16, #0x88]\n" ".inst 0x6fa4e0f9 // udot v25.4s, v7.16b, v4.4b[1]\n" + "ldr x21, [x22, #0x8]\n" ".inst 0x6fa5e0fd // udot v29.4s, v7.16b, v5.4b[1]\n" + "ldr d7, [x16, #0x70]\n" ".inst 0x6fa0e0ca // udot v10.4s, v6.16b, v0.4b[1]\n" + "add x20, x20, #0x10\n" ".inst 0x6fa1e0ce // udot v14.4s, v6.16b, v1.4b[1]\n" + "prfm pldl1keep, [x20, #0x80]\n" ".inst 0x6fa2e0d2 // udot v18.4s, v6.16b, v2.4b[1]\n" + "mov v7.d[1], x11\n" ".inst 0x6fa3e0d6 // udot v22.4s, v6.16b, v3.4b[1]\n" + "ldr x11, [x16, #0x98]\n" ".inst 0x6fa4e0da // udot v26.4s, v6.16b, v4.4b[1]\n" + "ldr x19, [x20, #0x8]\n" ".inst 0x6fa5e0de // udot v30.4s, v6.16b, v5.4b[1]\n" - "ldr d7, [x13, #0x70]\n" - "prfm pldl1keep, [x23, #0x80]\n" - "ldr d6, [x13, #0x80]\n" - "mov v7.d[1], x19\n" - "ldr x28, [x13, #0x88]\n" - "ldr x19, [x13, #0x98]\n" + "ldr d6, [x16, #0x80]\n" ".inst 0x6fa0e0eb // udot v11.4s, v7.16b, v0.4b[1]\n" + "sub x13, x13, #0x10\n" ".inst 0x6fa1e0ef // udot v15.4s, v7.16b, v1.4b[1]\n" - "mov v6.d[1], x28\n" + "cmp x13, #0x20\n" ".inst 0x6fa2e0f3 // udot v19.4s, v7.16b, v2.4b[1]\n" + "mov v6.d[1], x10\n" ".inst 0x6fa3e0f7 // udot v23.4s, v7.16b, v3.4b[1]\n" + "ldr x10, [x16, #0xa8]\n" ".inst 0x6fa4e0fb // udot v27.4s, v7.16b, v4.4b[1]\n" ".inst 0x6fa5e0ff // udot v31.4s, v7.16b, v5.4b[1]\n" + "ldr d7, [x16, #0x90]\n" ".inst 0x6f80e8c8 // udot v8.4s, v6.16b, v0.4b[2]\n" ".inst 0x6f81e8cc // udot v12.4s, v6.16b, v1.4b[2]\n" ".inst 0x6f82e8d0 // udot v16.4s, v6.16b, v2.4b[2]\n" + "mov v7.d[1], x11\n" ".inst 0x6f83e8d4 // udot v20.4s, v6.16b, v3.4b[2]\n" + "ldr x11, [x16, #0xb8]\n" ".inst 0x6f84e8d8 // udot v24.4s, v6.16b, v4.4b[2]\n" ".inst 0x6f85e8dc // udot v28.4s, v6.16b, v5.4b[2]\n" - "ldr d7, [x13, #0x90]\n" - "prfm pldl1keep, [x22, #0x80]\n" - "ldr d6, [x13, #0xa0]\n" - "mov v7.d[1], x19\n" - "ldr x28, [x13, #0xa8]\n" - "ldr x19, [x13, #0xb8]\n" + "ldr d6, [x16, #0xa0]\n" ".inst 0x6f80e8e9 // udot v9.4s, v7.16b, v0.4b[2]\n" ".inst 0x6f81e8ed // udot v13.4s, v7.16b, v1.4b[2]\n" - "mov v6.d[1], x28\n" ".inst 0x6f82e8f1 // udot v17.4s, v7.16b, v2.4b[2]\n" + "mov v6.d[1], x10\n" ".inst 0x6f83e8f5 // udot v21.4s, v7.16b, v3.4b[2]\n" + "ldr x10, [x16, #0xc8]\n" ".inst 0x6f84e8f9 // udot v25.4s, v7.16b, v4.4b[2]\n" ".inst 0x6f85e8fd // udot v29.4s, v7.16b, v5.4b[2]\n" + "ldr d7, [x16, #0xb0]\n" ".inst 0x6f80e8ca // udot v10.4s, v6.16b, v0.4b[2]\n" ".inst 0x6f81e8ce // udot v14.4s, v6.16b, v1.4b[2]\n" ".inst 0x6f82e8d2 // udot v18.4s, v6.16b, v2.4b[2]\n" + "mov v7.d[1], x11\n" ".inst 0x6f83e8d6 // udot v22.4s, v6.16b, v3.4b[2]\n" + "ldr x11, [x16, #0xd8]\n" ".inst 0x6f84e8da // udot v26.4s, v6.16b, v4.4b[2]\n" ".inst 0x6f85e8de // udot v30.4s, v6.16b, v5.4b[2]\n" - "ldr d7, [x13, #0xb0]\n" - "prfm pldl1keep, [x20, #0x80]\n" - "ldr d6, [x13, #0xc0]\n" - "mov v7.d[1], x19\n" - "ldr x28, [x13, #0xc8]\n" - "ldr x19, [x13, #0xd8]\n" + "ldr d6, [x16, #0xc0]\n" ".inst 0x6f80e8eb // udot v11.4s, v7.16b, v0.4b[2]\n" ".inst 0x6f81e8ef // udot v15.4s, v7.16b, v1.4b[2]\n" - "mov v6.d[1], x28\n" ".inst 0x6f82e8f3 // udot v19.4s, v7.16b, v2.4b[2]\n" + "mov v6.d[1], x10\n" ".inst 0x6f83e8f7 // udot v23.4s, v7.16b, v3.4b[2]\n" + "ldr x10, [x16, #0xe8]\n" ".inst 0x6f84e8fb // udot v27.4s, v7.16b, v4.4b[2]\n" ".inst 0x6f85e8ff // udot v31.4s, v7.16b, v5.4b[2]\n" + "ldr d7, [x16, #0xd0]\n" ".inst 0x6fa0e8c8 // udot v8.4s, v6.16b, v0.4b[3]\n" ".inst 0x6fa1e8cc // udot v12.4s, v6.16b, v1.4b[3]\n" ".inst 0x6fa2e8d0 // udot v16.4s, v6.16b, v2.4b[3]\n" + "mov v7.d[1], x11\n" ".inst 0x6fa3e8d4 // udot v20.4s, v6.16b, v3.4b[3]\n" + "ldr x11, [x16, #0xf8]\n" ".inst 0x6fa4e8d8 // udot v24.4s, v6.16b, v4.4b[3]\n" ".inst 0x6fa5e8dc // udot v28.4s, v6.16b, v5.4b[3]\n" - "ldr d7, [x13, #0xd0]\n" - "ldr d6, [x13, #0xe0]\n" - "ldr x28, [x13, #0xe8]\n" - "mov v7.d[1], x19\n" - "ldr x19, [x13, #0xf8]\n" - "ldr x27, [x9, #0x8]\n" - "mov v6.d[1], x28\n" + "ldr d6, [x16, #0xe0]\n" ".inst 0x6fa0e8e9 // udot v9.4s, v7.16b, v0.4b[3]\n" ".inst 0x6fa1e8ed // udot v13.4s, v7.16b, v1.4b[3]\n" ".inst 0x6fa2e8f1 // udot v17.4s, v7.16b, v2.4b[3]\n" + "mov v6.d[1], x10\n" ".inst 0x6fa3e8f5 // udot v21.4s, v7.16b, v3.4b[3]\n" ".inst 0x6fa4e8f9 // udot v25.4s, v7.16b, v4.4b[3]\n" ".inst 0x6fa5e8fd // udot v29.4s, v7.16b, v5.4b[3]\n" + "ldr d7, [x16, #0xf0]\n" ".inst 0x6fa0e8ca // udot v10.4s, v6.16b, v0.4b[3]\n" + "add x16, x16, #0x100\n" ".inst 0x6fa1e8ce // udot v14.4s, v6.16b, v1.4b[3]\n" + "ldr x10, [x16, #0x8]\n" ".inst 0x6fa2e8d2 // udot v18.4s, v6.16b, v2.4b[3]\n" + "mov v7.d[1], x11\n" ".inst 0x6fa3e8d6 // udot v22.4s, v6.16b, v3.4b[3]\n" ".inst 0x6fa4e8da // udot v26.4s, v6.16b, v4.4b[3]\n" ".inst 0x6fa5e8de // udot v30.4s, v6.16b, v5.4b[3]\n" - "ldr d7, [x13, #0xf0]\n" - "ldr x25, [x26, #0x8]\n" - "add x13, x13, #0x100\n" - "ldr d6, [x13, #0x0]\n" - "mov v7.d[1], x19\n" - "ldr x28, [x13, #0x8]\n" - "ldr x21, [x24, #0x8]\n" + "ldr d6, [x16, #0x0]\n" ".inst 0x6fa0e8eb // udot v11.4s, v7.16b, v0.4b[3]\n" + "ldr d0, [x12, #0x0]\n" ".inst 0x6fa1e8ef // udot v15.4s, v7.16b, v1.4b[3]\n" - "mov v6.d[1], x28\n" + "ldr d1, [x28, #0x0]\n" ".inst 0x6fa2e8f3 // udot v19.4s, v7.16b, v2.4b[3]\n" + "mov v6.d[1], x10\n" ".inst 0x6fa3e8f7 // udot v23.4s, v7.16b, v3.4b[3]\n" + "mov v0.d[1], x9\n" ".inst 0x6fa4e8fb // udot v27.4s, v7.16b, v4.4b[3]\n" + "mov v1.d[1], x27\n" ".inst 0x6fa5e8ff // udot v31.4s, v7.16b, v5.4b[3]\n" - "ldr d0, [x9, #0x0]\n" - "ldr d1, [x26, #0x0]\n" - "ldr d2, [x24, #0x0]\n" - "mov v0.d[1], x27\n" - "ldr d3, [x23, #0x0]\n" - "mov v1.d[1], x25\n" - "ldr x19, [x23, #0x8]\n" - "mov v2.d[1], x21\n" + "ldr d2, [x26, #0x0]\n" + "ldr d3, [x24, #0x0]\n" "ldr d4, [x22, #0x0]\n" - "ldr x21, [x22, #0x8]\n" - "mov v3.d[1], x19\n" + "mov v2.d[1], x25\n" "ldr d5, [x20, #0x0]\n" - "ldr x19, [x20, #0x8]\n" + "mov v3.d[1], x23\n" "mov v4.d[1], x21\n" "mov v5.d[1], x19\n" "bge 187b\n" "188:" // Height 6: Multiply loop: Single iteration only ".inst 0x6f80e0c8 // udot v8.4s, v6.16b, v0.4b[0]\n" + "ldr q7, [x16, #0x10]\n" ".inst 0x6f81e0cc // udot v12.4s, v6.16b, v1.4b[0]\n" + "sub x13, x13, #0x10\n" ".inst 0x6f82e0d0 // udot v16.4s, v6.16b, v2.4b[0]\n" + "add x12, x12, #0x10\n" ".inst 0x6f83e0d4 // udot v20.4s, v6.16b, v3.4b[0]\n" + "prfm pldl1keep, [x12, #0x80]\n" ".inst 0x6f84e0d8 // udot v24.4s, v6.16b, v4.4b[0]\n" + "add x28, x28, #0x10\n" ".inst 0x6f85e0dc // udot v28.4s, v6.16b, v5.4b[0]\n" - "ldr q7, [x13, #0x10]\n" - "sub x10, x10, #0x10\n" - "ldr q6, [x13, #0x20]\n" - "add x9, x9, #0x10\n" - "prfm pldl1keep, [x9, #0x80]\n" - "add x26, x26, #0x10\n" + "prfm pldl1keep, [x28, #0x80]\n" ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" + "ldr q6, [x16, #0x20]\n" ".inst 0x6f81e0ed // udot v13.4s, v7.16b, v1.4b[0]\n" + "add x26, x26, #0x10\n" ".inst 0x6f82e0f1 // udot v17.4s, v7.16b, v2.4b[0]\n" + "prfm pldl1keep, [x26, #0x80]\n" ".inst 0x6f83e0f5 // udot v21.4s, v7.16b, v3.4b[0]\n" + "add x24, x24, #0x10\n" ".inst 0x6f84e0f9 // udot v25.4s, v7.16b, v4.4b[0]\n" + "prfm pldl1keep, [x24, #0x80]\n" ".inst 0x6f85e0fd // udot v29.4s, v7.16b, v5.4b[0]\n" + "ldr q7, [x16, #0x30]\n" ".inst 0x6f80e0ca // udot v10.4s, v6.16b, v0.4b[0]\n" + "add x22, x22, #0x10\n" ".inst 0x6f81e0ce // udot v14.4s, v6.16b, v1.4b[0]\n" + "prfm pldl1keep, [x22, #0x80]\n" ".inst 0x6f82e0d2 // udot v18.4s, v6.16b, v2.4b[0]\n" + "add x20, x20, #0x10\n" ".inst 0x6f83e0d6 // udot v22.4s, v6.16b, v3.4b[0]\n" + "prfm pldl1keep, [x20, #0x80]\n" ".inst 0x6f84e0da // udot v26.4s, v6.16b, v4.4b[0]\n" ".inst 0x6f85e0de // udot v30.4s, v6.16b, v5.4b[0]\n" - "ldr q7, [x13, #0x30]\n" - "add x24, x24, #0x10\n" - "prfm pldl1keep, [x26, #0x80]\n" - "add x23, x23, #0x10\n" - "ldr q6, [x13, #0x40]\n" - "add x22, x22, #0x10\n" + "ldr q6, [x16, #0x40]\n" ".inst 0x6f80e0eb // udot v11.4s, v7.16b, v0.4b[0]\n" ".inst 0x6f81e0ef // udot v15.4s, v7.16b, v1.4b[0]\n" ".inst 0x6f82e0f3 // udot v19.4s, v7.16b, v2.4b[0]\n" ".inst 0x6f83e0f7 // udot v23.4s, v7.16b, v3.4b[0]\n" ".inst 0x6f84e0fb // udot v27.4s, v7.16b, v4.4b[0]\n" ".inst 0x6f85e0ff // udot v31.4s, v7.16b, v5.4b[0]\n" + "ldr q7, [x16, #0x50]\n" ".inst 0x6fa0e0c8 // udot v8.4s, v6.16b, v0.4b[1]\n" ".inst 0x6fa1e0cc // udot v12.4s, v6.16b, v1.4b[1]\n" ".inst 0x6fa2e0d0 // udot v16.4s, v6.16b, v2.4b[1]\n" ".inst 0x6fa3e0d4 // udot v20.4s, v6.16b, v3.4b[1]\n" ".inst 0x6fa4e0d8 // udot v24.4s, v6.16b, v4.4b[1]\n" ".inst 0x6fa5e0dc // udot v28.4s, v6.16b, v5.4b[1]\n" - "ldr q7, [x13, #0x50]\n" - "add x20, x20, #0x10\n" - "prfm pldl1keep, [x24, #0x80]\n" - "ldr q6, [x13, #0x60]\n" + "ldr q6, [x16, #0x60]\n" ".inst 0x6fa0e0e9 // udot v9.4s, v7.16b, v0.4b[1]\n" ".inst 0x6fa1e0ed // udot v13.4s, v7.16b, v1.4b[1]\n" ".inst 0x6fa2e0f1 // udot v17.4s, v7.16b, v2.4b[1]\n" ".inst 0x6fa3e0f5 // udot v21.4s, v7.16b, v3.4b[1]\n" ".inst 0x6fa4e0f9 // udot v25.4s, v7.16b, v4.4b[1]\n" ".inst 0x6fa5e0fd // udot v29.4s, v7.16b, v5.4b[1]\n" + "ldr q7, [x16, #0x70]\n" ".inst 0x6fa0e0ca // udot v10.4s, v6.16b, v0.4b[1]\n" ".inst 0x6fa1e0ce // udot v14.4s, v6.16b, v1.4b[1]\n" ".inst 0x6fa2e0d2 // udot v18.4s, v6.16b, v2.4b[1]\n" ".inst 0x6fa3e0d6 // udot v22.4s, v6.16b, v3.4b[1]\n" ".inst 0x6fa4e0da // udot v26.4s, v6.16b, v4.4b[1]\n" ".inst 0x6fa5e0de // udot v30.4s, v6.16b, v5.4b[1]\n" - "ldr q7, [x13, #0x70]\n" - "prfm pldl1keep, [x23, #0x80]\n" - "ldr q6, [x13, #0x80]\n" + "ldr q6, [x16, #0x80]\n" ".inst 0x6fa0e0eb // udot v11.4s, v7.16b, v0.4b[1]\n" ".inst 0x6fa1e0ef // udot v15.4s, v7.16b, v1.4b[1]\n" ".inst 0x6fa2e0f3 // udot v19.4s, v7.16b, v2.4b[1]\n" ".inst 0x6fa3e0f7 // udot v23.4s, v7.16b, v3.4b[1]\n" ".inst 0x6fa4e0fb // udot v27.4s, v7.16b, v4.4b[1]\n" ".inst 0x6fa5e0ff // udot v31.4s, v7.16b, v5.4b[1]\n" + "ldr q7, [x16, #0x90]\n" ".inst 0x6f80e8c8 // udot v8.4s, v6.16b, v0.4b[2]\n" ".inst 0x6f81e8cc // udot v12.4s, v6.16b, v1.4b[2]\n" ".inst 0x6f82e8d0 // udot v16.4s, v6.16b, v2.4b[2]\n" ".inst 0x6f83e8d4 // udot v20.4s, v6.16b, v3.4b[2]\n" ".inst 0x6f84e8d8 // udot v24.4s, v6.16b, v4.4b[2]\n" ".inst 0x6f85e8dc // udot v28.4s, v6.16b, v5.4b[2]\n" - "ldr q7, [x13, #0x90]\n" - "prfm pldl1keep, [x22, #0x80]\n" - "ldr q6, [x13, #0xa0]\n" + "ldr q6, [x16, #0xa0]\n" ".inst 0x6f80e8e9 // udot v9.4s, v7.16b, v0.4b[2]\n" ".inst 0x6f81e8ed // udot v13.4s, v7.16b, v1.4b[2]\n" ".inst 0x6f82e8f1 // udot v17.4s, v7.16b, v2.4b[2]\n" ".inst 0x6f83e8f5 // udot v21.4s, v7.16b, v3.4b[2]\n" ".inst 0x6f84e8f9 // udot v25.4s, v7.16b, v4.4b[2]\n" ".inst 0x6f85e8fd // udot v29.4s, v7.16b, v5.4b[2]\n" + "ldr q7, [x16, #0xb0]\n" ".inst 0x6f80e8ca // udot v10.4s, v6.16b, v0.4b[2]\n" ".inst 0x6f81e8ce // udot v14.4s, v6.16b, v1.4b[2]\n" ".inst 0x6f82e8d2 // udot v18.4s, v6.16b, v2.4b[2]\n" ".inst 0x6f83e8d6 // udot v22.4s, v6.16b, v3.4b[2]\n" ".inst 0x6f84e8da // udot v26.4s, v6.16b, v4.4b[2]\n" ".inst 0x6f85e8de // udot v30.4s, v6.16b, v5.4b[2]\n" - "ldr q7, [x13, #0xb0]\n" - "prfm pldl1keep, [x20, #0x80]\n" - "ldr q6, [x13, #0xc0]\n" + "ldr q6, [x16, #0xc0]\n" ".inst 0x6f80e8eb // udot v11.4s, v7.16b, v0.4b[2]\n" ".inst 0x6f81e8ef // udot v15.4s, v7.16b, v1.4b[2]\n" ".inst 0x6f82e8f3 // udot v19.4s, v7.16b, v2.4b[2]\n" ".inst 0x6f83e8f7 // udot v23.4s, v7.16b, v3.4b[2]\n" ".inst 0x6f84e8fb // udot v27.4s, v7.16b, v4.4b[2]\n" ".inst 0x6f85e8ff // udot v31.4s, v7.16b, v5.4b[2]\n" + "ldr q7, [x16, #0xd0]\n" ".inst 0x6fa0e8c8 // udot v8.4s, v6.16b, v0.4b[3]\n" ".inst 0x6fa1e8cc // udot v12.4s, v6.16b, v1.4b[3]\n" ".inst 0x6fa2e8d0 // udot v16.4s, v6.16b, v2.4b[3]\n" ".inst 0x6fa3e8d4 // udot v20.4s, v6.16b, v3.4b[3]\n" ".inst 0x6fa4e8d8 // udot v24.4s, v6.16b, v4.4b[3]\n" ".inst 0x6fa5e8dc // udot v28.4s, v6.16b, v5.4b[3]\n" - "ldr q7, [x13, #0xd0]\n" - "ldr q6, [x13, #0xe0]\n" + "ldr q6, [x16, #0xe0]\n" ".inst 0x6fa0e8e9 // udot v9.4s, v7.16b, v0.4b[3]\n" ".inst 0x6fa1e8ed // udot v13.4s, v7.16b, v1.4b[3]\n" ".inst 0x6fa2e8f1 // udot v17.4s, v7.16b, v2.4b[3]\n" ".inst 0x6fa3e8f5 // udot v21.4s, v7.16b, v3.4b[3]\n" ".inst 0x6fa4e8f9 // udot v25.4s, v7.16b, v4.4b[3]\n" ".inst 0x6fa5e8fd // udot v29.4s, v7.16b, v5.4b[3]\n" + "ldr q7, [x16, #0xf0]\n" ".inst 0x6fa0e8ca // udot v10.4s, v6.16b, v0.4b[3]\n" + "add x16, x16, #0x100\n" ".inst 0x6fa1e8ce // udot v14.4s, v6.16b, v1.4b[3]\n" ".inst 0x6fa2e8d2 // udot v18.4s, v6.16b, v2.4b[3]\n" ".inst 0x6fa3e8d6 // udot v22.4s, v6.16b, v3.4b[3]\n" ".inst 0x6fa4e8da // udot v26.4s, v6.16b, v4.4b[3]\n" ".inst 0x6fa5e8de // udot v30.4s, v6.16b, v5.4b[3]\n" - "ldr q7, [x13, #0xf0]\n" - "add x13, x13, #0x100\n" ".inst 0x6fa0e8eb // udot v11.4s, v7.16b, v0.4b[3]\n" ".inst 0x6fa1e8ef // udot v15.4s, v7.16b, v1.4b[3]\n" ".inst 0x6fa2e8f3 // udot v19.4s, v7.16b, v2.4b[3]\n" @@ -3206,36 +3206,36 @@ void a64_hybrid_u8u32_dot_6x16_a55 ( ".inst 0x6fa4e8fb // udot v27.4s, v7.16b, v4.4b[3]\n" ".inst 0x6fa5e8ff // udot v31.4s, v7.16b, v5.4b[3]\n" "189:" // Height 6: Multiply loop: Main loop skip - "cbz x10, 194f\n" - "cmp x10, #0x4\n" + "cbz x13, 194f\n" + "cmp x13, #0x4\n" "blt 191f\n" "190:" // Height 6: Multiply loop: Odd block loop - "ldr s0, [x9], #0x4\n" - "sub x10, x10, #0x4\n" - "ldr s1, [x26], #0x4\n" - "cmp x10, #0x4\n" - "ldr s2, [x24], #0x4\n" - "ldr s3, [x23], #0x4\n" + "ldr s0, [x12], #0x4\n" + "sub x13, x13, #0x4\n" + "ldr s1, [x28], #0x4\n" + "cmp x13, #0x4\n" + "ldr s2, [x26], #0x4\n" + "ldr s3, [x24], #0x4\n" "ldr s4, [x22], #0x4\n" "ldr s5, [x20], #0x4\n" - "ldr q6, [x13, #0x0]\n" - "ldr q7, [x13, #0x10]\n" + "ldr q6, [x16, #0x0]\n" + "ldr q7, [x16, #0x10]\n" ".inst 0x6f80e0c8 // udot v8.4s, v6.16b, v0.4b[0]\n" ".inst 0x6f81e0cc // udot v12.4s, v6.16b, v1.4b[0]\n" ".inst 0x6f82e0d0 // udot v16.4s, v6.16b, v2.4b[0]\n" ".inst 0x6f83e0d4 // udot v20.4s, v6.16b, v3.4b[0]\n" ".inst 0x6f84e0d8 // udot v24.4s, v6.16b, v4.4b[0]\n" ".inst 0x6f85e0dc // udot v28.4s, v6.16b, v5.4b[0]\n" + "ldr q6, [x16, #0x20]\n" ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" ".inst 0x6f81e0ed // udot v13.4s, v7.16b, v1.4b[0]\n" ".inst 0x6f82e0f1 // udot v17.4s, v7.16b, v2.4b[0]\n" ".inst 0x6f83e0f5 // udot v21.4s, v7.16b, v3.4b[0]\n" ".inst 0x6f84e0f9 // udot v25.4s, v7.16b, v4.4b[0]\n" ".inst 0x6f85e0fd // udot v29.4s, v7.16b, v5.4b[0]\n" - "ldr q6, [x13, #0x20]\n" - "ldr q7, [x13, #0x30]\n" - "add x13, x13, #0x40\n" + "ldr q7, [x16, #0x30]\n" ".inst 0x6f80e0ca // udot v10.4s, v6.16b, v0.4b[0]\n" + "add x16, x16, #0x40\n" ".inst 0x6f81e0ce // udot v14.4s, v6.16b, v1.4b[0]\n" ".inst 0x6f82e0d2 // udot v18.4s, v6.16b, v2.4b[0]\n" ".inst 0x6f83e0d6 // udot v22.4s, v6.16b, v3.4b[0]\n" @@ -3248,49 +3248,49 @@ void a64_hybrid_u8u32_dot_6x16_a55 ( ".inst 0x6f84e0fb // udot v27.4s, v7.16b, v4.4b[0]\n" ".inst 0x6f85e0ff // udot v31.4s, v7.16b, v5.4b[0]\n" "bge 190b\n" - "cbz x10, 194f\n" + "cbz x13, 194f\n" "191:" // Height 6: Multiply loop: Skip odd blocks - "tbz x10, #1, 192f\n" - "ldr h0, [x9], #0x2\n" - "ldr h1, [x26], #0x2\n" - "ldr h2, [x24], #0x2\n" - "ldr h3, [x23], #0x2\n" + "tbz x13, #1, 192f\n" + "ldr h0, [x12], #0x2\n" + "ldr h1, [x28], #0x2\n" + "ldr h2, [x26], #0x2\n" + "ldr h3, [x24], #0x2\n" "ldr h4, [x22], #0x2\n" "ldr h5, [x20], #0x2\n" - "tbz x10, #0, 193f\n" - "ld1 { v0.b }[2], [x9]\n" - "ld1 { v1.b }[2], [x26]\n" - "ld1 { v2.b }[2], [x24]\n" - "ld1 { v3.b }[2], [x23]\n" + "tbz x13, #0, 193f\n" + "ld1 { v0.b }[2], [x12]\n" + "ld1 { v1.b }[2], [x28]\n" + "ld1 { v2.b }[2], [x26]\n" + "ld1 { v3.b }[2], [x24]\n" "ld1 { v4.b }[2], [x22]\n" "ld1 { v5.b }[2], [x20]\n" "b 193f\n" "192:" // Height 6: Multiply loop: Ragged operand read: partial_1_0 - "ldr b0, [x9, #0x0]\n" - "ldr b1, [x26, #0x0]\n" - "ldr b2, [x24, #0x0]\n" - "ldr b3, [x23, #0x0]\n" + "ldr b0, [x12, #0x0]\n" + "ldr b1, [x28, #0x0]\n" + "ldr b2, [x26, #0x0]\n" + "ldr b3, [x24, #0x0]\n" "ldr b4, [x22, #0x0]\n" "ldr b5, [x20, #0x0]\n" "193:" // Height 6: Multiply loop: Ragged operand read: Done - "ldr q6, [x13, #0x0]\n" - "ldr q7, [x13, #0x10]\n" + "ldr q6, [x16, #0x0]\n" + "ldr q7, [x16, #0x10]\n" ".inst 0x6f80e0c8 // udot v8.4s, v6.16b, v0.4b[0]\n" ".inst 0x6f81e0cc // udot v12.4s, v6.16b, v1.4b[0]\n" ".inst 0x6f82e0d0 // udot v16.4s, v6.16b, v2.4b[0]\n" ".inst 0x6f83e0d4 // udot v20.4s, v6.16b, v3.4b[0]\n" ".inst 0x6f84e0d8 // udot v24.4s, v6.16b, v4.4b[0]\n" ".inst 0x6f85e0dc // udot v28.4s, v6.16b, v5.4b[0]\n" + "ldr q6, [x16, #0x20]\n" ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" ".inst 0x6f81e0ed // udot v13.4s, v7.16b, v1.4b[0]\n" ".inst 0x6f82e0f1 // udot v17.4s, v7.16b, v2.4b[0]\n" ".inst 0x6f83e0f5 // udot v21.4s, v7.16b, v3.4b[0]\n" ".inst 0x6f84e0f9 // udot v25.4s, v7.16b, v4.4b[0]\n" ".inst 0x6f85e0fd // udot v29.4s, v7.16b, v5.4b[0]\n" - "ldr q6, [x13, #0x20]\n" - "ldr q7, [x13, #0x30]\n" - "add x13, x13, #0x40\n" + "ldr q7, [x16, #0x30]\n" ".inst 0x6f80e0ca // udot v10.4s, v6.16b, v0.4b[0]\n" + "add x16, x16, #0x40\n" ".inst 0x6f81e0ce // udot v14.4s, v6.16b, v1.4b[0]\n" ".inst 0x6f82e0d2 // udot v18.4s, v6.16b, v2.4b[0]\n" ".inst 0x6f83e0d6 // udot v22.4s, v6.16b, v3.4b[0]\n" @@ -3304,13 +3304,13 @@ void a64_hybrid_u8u32_dot_6x16_a55 ( ".inst 0x6f85e0ff // udot v31.4s, v7.16b, v5.4b[0]\n" "194:" // Height 6: Multiply loop: No odd multiplies "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" - "add x11, x11, #0x1\n" - "cmp x11, x19\n" + "add x14, x14, #0x1\n" + "cmp x14, x19\n" "bne 184b\n" "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "prfm pstl1keep, [x12, #0x0]\n" - "cmp x14, #0x10\n" - "add x23, x12, x19, LSL #2\n" + "prfm pstl1keep, [x15, #0x0]\n" + "cmp x17, #0x10\n" + "add x23, x15, x19, LSL #2\n" "prfm pstl1keep, [x23, #0x0]\n" "add x22, x23, x19, LSL #2\n" "prfm pstl1keep, [x22, #0x0]\n" @@ -3321,9 +3321,9 @@ void a64_hybrid_u8u32_dot_6x16_a55 ( "add x19, x20, x19, LSL #2\n" "prfm pstl1keep, [x19, #0x0]\n" "bge 203f\n" - "tbz x14, #3, 198f\n" - "st1 { v8.4s }, [x12], #0x10\n" - "st1 { v9.4s }, [x12], #0x10\n" + "tbz x17, #3, 198f\n" + "st1 { v8.4s }, [x15], #0x10\n" + "st1 { v9.4s }, [x15], #0x10\n" "st1 { v12.4s }, [x23], #0x10\n" "st1 { v13.4s }, [x23], #0x10\n" "st1 { v16.4s }, [x22], #0x10\n" @@ -3334,22 +3334,22 @@ void a64_hybrid_u8u32_dot_6x16_a55 ( "st1 { v25.4s }, [x20], #0x10\n" "st1 { v28.4s }, [x19], #0x10\n" "st1 { v29.4s }, [x19], #0x10\n" - "tbz x14, #2, 196f\n" - "st1 { v10.4s }, [x12], #0x10\n" + "tbz x17, #2, 196f\n" + "st1 { v10.4s }, [x15], #0x10\n" "st1 { v14.4s }, [x23], #0x10\n" "st1 { v18.4s }, [x22], #0x10\n" "st1 { v22.4s }, [x21], #0x10\n" "st1 { v26.4s }, [x20], #0x10\n" "st1 { v30.4s }, [x19], #0x10\n" - "tbz x14, #1, 195f\n" - "str d11, [x12], #0x8\n" + "tbz x17, #1, 195f\n" + "str d11, [x15], #0x8\n" "str d15, [x23], #0x8\n" "str d19, [x22], #0x8\n" "str d23, [x21], #0x8\n" "str d27, [x20], #0x8\n" "str d31, [x19], #0x8\n" - "tbz x14, #0, 202f\n" - "st1 { v11.s }[2], [x12]\n" + "tbz x17, #0, 202f\n" + "st1 { v11.s }[2], [x15]\n" "st1 { v15.s }[2], [x23]\n" "st1 { v19.s }[2], [x22]\n" "st1 { v23.s }[2], [x21]\n" @@ -3357,8 +3357,8 @@ void a64_hybrid_u8u32_dot_6x16_a55 ( "st1 { v31.s }[2], [x19]\n" "b 202f\n" "195:" // Height 6: Partial direct writeback: partial_1_12 - "tbz x14, #0, 202f\n" - "str s11, [x12, #0x0]\n" + "tbz x17, #0, 202f\n" + "str s11, [x15, #0x0]\n" "str s15, [x23, #0x0]\n" "str s19, [x22, #0x0]\n" "str s23, [x21, #0x0]\n" @@ -3366,15 +3366,15 @@ void a64_hybrid_u8u32_dot_6x16_a55 ( "str s31, [x19, #0x0]\n" "b 202f\n" "196:" // Height 6: Partial direct writeback: partial_2_8 - "tbz x14, #1, 197f\n" - "str d10, [x12], #0x8\n" + "tbz x17, #1, 197f\n" + "str d10, [x15], #0x8\n" "str d14, [x23], #0x8\n" "str d18, [x22], #0x8\n" "str d22, [x21], #0x8\n" "str d26, [x20], #0x8\n" "str d30, [x19], #0x8\n" - "tbz x14, #0, 202f\n" - "st1 { v10.s }[2], [x12]\n" + "tbz x17, #0, 202f\n" + "st1 { v10.s }[2], [x15]\n" "st1 { v14.s }[2], [x23]\n" "st1 { v18.s }[2], [x22]\n" "st1 { v22.s }[2], [x21]\n" @@ -3382,8 +3382,8 @@ void a64_hybrid_u8u32_dot_6x16_a55 ( "st1 { v30.s }[2], [x19]\n" "b 202f\n" "197:" // Height 6: Partial direct writeback: partial_1_8 - "tbz x14, #0, 202f\n" - "str s10, [x12, #0x0]\n" + "tbz x17, #0, 202f\n" + "str s10, [x15, #0x0]\n" "str s14, [x23, #0x0]\n" "str s18, [x22, #0x0]\n" "str s22, [x21, #0x0]\n" @@ -3391,22 +3391,22 @@ void a64_hybrid_u8u32_dot_6x16_a55 ( "str s30, [x19, #0x0]\n" "b 202f\n" "198:" // Height 6: Partial direct writeback: partial_4_0 - "tbz x14, #2, 200f\n" - "st1 { v8.4s }, [x12], #0x10\n" + "tbz x17, #2, 200f\n" + "st1 { v8.4s }, [x15], #0x10\n" "st1 { v12.4s }, [x23], #0x10\n" "st1 { v16.4s }, [x22], #0x10\n" "st1 { v20.4s }, [x21], #0x10\n" "st1 { v24.4s }, [x20], #0x10\n" "st1 { v28.4s }, [x19], #0x10\n" - "tbz x14, #1, 199f\n" - "str d9, [x12], #0x8\n" + "tbz x17, #1, 199f\n" + "str d9, [x15], #0x8\n" "str d13, [x23], #0x8\n" "str d17, [x22], #0x8\n" "str d21, [x21], #0x8\n" "str d25, [x20], #0x8\n" "str d29, [x19], #0x8\n" - "tbz x14, #0, 202f\n" - "st1 { v9.s }[2], [x12]\n" + "tbz x17, #0, 202f\n" + "st1 { v9.s }[2], [x15]\n" "st1 { v13.s }[2], [x23]\n" "st1 { v17.s }[2], [x22]\n" "st1 { v21.s }[2], [x21]\n" @@ -3414,8 +3414,8 @@ void a64_hybrid_u8u32_dot_6x16_a55 ( "st1 { v29.s }[2], [x19]\n" "b 202f\n" "199:" // Height 6: Partial direct writeback: partial_1_4 - "tbz x14, #0, 202f\n" - "str s9, [x12, #0x0]\n" + "tbz x17, #0, 202f\n" + "str s9, [x15, #0x0]\n" "str s13, [x23, #0x0]\n" "str s17, [x22, #0x0]\n" "str s21, [x21, #0x0]\n" @@ -3423,15 +3423,15 @@ void a64_hybrid_u8u32_dot_6x16_a55 ( "str s29, [x19, #0x0]\n" "b 202f\n" "200:" // Height 6: Partial direct writeback: partial_2_0 - "tbz x14, #1, 201f\n" - "str d8, [x12], #0x8\n" + "tbz x17, #1, 201f\n" + "str d8, [x15], #0x8\n" "str d12, [x23], #0x8\n" "str d16, [x22], #0x8\n" "str d20, [x21], #0x8\n" "str d24, [x20], #0x8\n" "str d28, [x19], #0x8\n" - "tbz x14, #0, 202f\n" - "st1 { v8.s }[2], [x12]\n" + "tbz x17, #0, 202f\n" + "st1 { v8.s }[2], [x15]\n" "st1 { v12.s }[2], [x23]\n" "st1 { v16.s }[2], [x22]\n" "st1 { v20.s }[2], [x21]\n" @@ -3439,7 +3439,7 @@ void a64_hybrid_u8u32_dot_6x16_a55 ( "st1 { v28.s }[2], [x19]\n" "b 202f\n" "201:" // Height 6: Partial direct writeback: partial_1_0 - "str s8, [x12, #0x0]\n" + "str s8, [x15, #0x0]\n" "str s12, [x23, #0x0]\n" "str s16, [x22, #0x0]\n" "str s20, [x21, #0x0]\n" @@ -3448,11 +3448,11 @@ void a64_hybrid_u8u32_dot_6x16_a55 ( "202:" // Height 6: Partial direct writeback: Done "b 204f\n" "203:" // Height 6: Full writeback - "str q8, [x12, #0x0]\n" - "str q9, [x12, #0x10]\n" - "str q10, [x12, #0x20]\n" - "str q11, [x12, #0x30]\n" - "add x12, x12, #0x40\n" + "str q8, [x15, #0x0]\n" + "str q9, [x15, #0x10]\n" + "str q10, [x15, #0x20]\n" + "str q11, [x15, #0x30]\n" + "add x15, x15, #0x40\n" "str q12, [x23, #0x0]\n" "str q13, [x23, #0x10]\n" "str q14, [x23, #0x20]\n" @@ -3474,7 +3474,7 @@ void a64_hybrid_u8u32_dot_6x16_a55 ( "str q30, [x19, #0x20]\n" "str q31, [x19, #0x30]\n" "204:" // Height 6: Writeback done - "subs x14, x14, #0x10\n" + "subs x17, x17, #0x10\n" "bgt 172b\n" "subs %x[M], %x[M], #0x6\n" "beq 206f\n" @@ -3491,7 +3491,7 @@ void a64_hybrid_u8u32_dot_6x16_a55 ( : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr) : [args_ptr] "r" (&ka), [flags] "r" (flags), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)) - : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x13", "x14", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" ); } -- cgit v1.2.1