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_u8qa_dot_4x16/a55.cpp | 392 ++++++++++----------- 1 file changed, 196 insertions(+), 196 deletions(-) (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_dot_4x16/a55.cpp') 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" -- cgit v1.2.1