From 74921eee924625426429044decefe3673561b174 Mon Sep 17 00:00:00 2001 From: Michael Tyler Date: Wed, 12 Apr 2023 17:43:17 +0100 Subject: Update CPU kernel implementations and guard directives Resolves COMPMID-6023 Change-Id: I868975d14c4f98af6716726feda22405a6a4c891 Signed-off-by: Michael Tyler Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9686 Tested-by: Arm Jenkins Reviewed-by: Viet-Hoa Do Comments-Addressed: Arm Jenkins Benchmark: Arm Jenkins --- .../generic.cpp | 600 ++++++++++----------- 1 file changed, 300 insertions(+), 300 deletions(-) (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32bf16fp32_mmla_4x24/generic.cpp') diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32bf16fp32_mmla_4x24/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32bf16fp32_mmla_4x24/generic.cpp index 1f707fa962..8961e615d7 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32bf16fp32_mmla_4x24/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32bf16fp32_mmla_4x24/generic.cpp @@ -283,11 +283,11 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 ( "21:" // Height 1: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w25, [x20, x26, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 22f\n" - "ldr x21, [%x[input_ptr], x26, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x24, [x21, #0x0]\n" + "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x24, [x20, #0x0]\n" "cbnz x26, 23f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x24, x24, x20, LSL #2\n" @@ -307,32 +307,32 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 ( "24:" // Height 1: Multiply loop: Main loop head ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n" ".inst 0x6e44ec08 // bfmmla v8.4s, v0.8h, v4.8h\n" - "ldr q4, [x10, #0x0]\n" + "ldr q24, [x10, #0x0]\n" ".inst 0x6e45ec0e // bfmmla v14.4s, v0.8h, v5.8h\n" - "ldr q5, [x10, #0x10]\n" + "ldr q23, [x10, #0x10]\n" ".inst 0x6e46ec09 // bfmmla v9.4s, v0.8h, v6.8h\n" - "ldr q6, [x9, #0x0]\n" + "ldr q22, [x9, #0x0]\n" ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n" - "ldr q7, [x9, #0x10]\n" - ".inst 0x6e44ec0a // bfmmla v10.4s, v0.8h, v4.8h\n" - "ldr q4, [x28, #0x0]\n" - ".inst 0x6e45ec10 // bfmmla v16.4s, v0.8h, v5.8h\n" - "ldr q5, [x28, #0x10]\n" - ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n" - "ldr q6, [x27, #0x0]\n" - ".inst 0x6e47ec11 // bfmmla v17.4s, v0.8h, v7.8h\n" - "ldr q7, [x27, #0x10]\n" + "ldr q21, [x9, #0x10]\n" + ".inst 0x6e58ec0a // bfmmla v10.4s, v0.8h, v24.8h\n" + "ldr q24, [x28, #0x0]\n" + ".inst 0x6e57ec10 // bfmmla v16.4s, v0.8h, v23.8h\n" + "ldr q23, [x28, #0x10]\n" + ".inst 0x6e56ec0b // bfmmla v11.4s, v0.8h, v22.8h\n" + "ldr q22, [x27, #0x0]\n" + ".inst 0x6e55ec11 // bfmmla v17.4s, v0.8h, v21.8h\n" + "ldr q21, [x27, #0x10]\n" "sub x25, x25, #0x4\n" "cmp x25, #0x8\n" - ".inst 0x6e44ec0c // bfmmla v12.4s, v0.8h, v4.8h\n" + ".inst 0x6e58ec0c // bfmmla v12.4s, v0.8h, v24.8h\n" "add x12, x12, #0x20\n" "ldr q4, [x12, #0x0]\n" "add x11, x11, #0x20\n" - ".inst 0x6e45ec12 // bfmmla v18.4s, v0.8h, v5.8h\n" + ".inst 0x6e57ec12 // bfmmla v18.4s, v0.8h, v23.8h\n" "ldr q5, [x12, #0x10]\n" - ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n" + ".inst 0x6e56ec0d // bfmmla v13.4s, v0.8h, v22.8h\n" "ldr q6, [x11, #0x0]\n" - ".inst 0x6e47ec13 // bfmmla v19.4s, v0.8h, v7.8h\n" + ".inst 0x6e55ec13 // bfmmla v19.4s, v0.8h, v21.8h\n" "ld1 { v0.4s }, [x24], #0x10\n" "ldr q7, [x11, #0x10]\n" "add x10, x10, #0x20\n" @@ -343,28 +343,28 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 ( "25:" // Height 1: Multiply loop: Single iteration only ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n" ".inst 0x6e44ec08 // bfmmla v8.4s, v0.8h, v4.8h\n" - "ldr q4, [x10, #0x0]\n" + "ldr q22, [x10, #0x0]\n" ".inst 0x6e45ec0e // bfmmla v14.4s, v0.8h, v5.8h\n" - "ldr q5, [x10, #0x10]\n" + "ldr q25, [x10, #0x10]\n" ".inst 0x6e46ec09 // bfmmla v9.4s, v0.8h, v6.8h\n" - "ldr q6, [x9, #0x0]\n" + "ldr q21, [x9, #0x0]\n" ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n" - "ldr q7, [x9, #0x10]\n" - ".inst 0x6e44ec0a // bfmmla v10.4s, v0.8h, v4.8h\n" - "ldr q4, [x28, #0x0]\n" - ".inst 0x6e45ec10 // bfmmla v16.4s, v0.8h, v5.8h\n" - "ldr q5, [x28, #0x10]\n" - ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n" - "ldr q6, [x27, #0x0]\n" - ".inst 0x6e47ec11 // bfmmla v17.4s, v0.8h, v7.8h\n" - "ldr q7, [x27, #0x10]\n" + "ldr q24, [x9, #0x10]\n" + ".inst 0x6e56ec0a // bfmmla v10.4s, v0.8h, v22.8h\n" + "ldr q23, [x28, #0x0]\n" + ".inst 0x6e59ec10 // bfmmla v16.4s, v0.8h, v25.8h\n" + "ldr q22, [x28, #0x10]\n" + ".inst 0x6e55ec0b // bfmmla v11.4s, v0.8h, v21.8h\n" + "ldr q21, [x27, #0x0]\n" + ".inst 0x6e58ec11 // bfmmla v17.4s, v0.8h, v24.8h\n" + "ldr q3, [x27, #0x10]\n" "sub x25, x25, #0x4\n" - ".inst 0x6e44ec0c // bfmmla v12.4s, v0.8h, v4.8h\n" - ".inst 0x6e45ec12 // bfmmla v18.4s, v0.8h, v5.8h\n" + ".inst 0x6e57ec0c // bfmmla v12.4s, v0.8h, v23.8h\n" + ".inst 0x6e56ec12 // bfmmla v18.4s, v0.8h, v22.8h\n" "add x12, x12, #0x20\n" "add x11, x11, #0x20\n" - ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n" - ".inst 0x6e47ec13 // bfmmla v19.4s, v0.8h, v7.8h\n" + ".inst 0x6e55ec0d // bfmmla v13.4s, v0.8h, v21.8h\n" + ".inst 0x6e43ec13 // bfmmla v19.4s, v0.8h, v3.8h\n" "add x10, x10, #0x20\n" "add x9, x9, #0x20\n" "add x28, x28, #0x20\n" @@ -380,31 +380,31 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 ( "27:" // Height 1: Multiply loop: Ragged operand read: partial_1_0 "ldr s0, [x24, #0x0]\n" "28:" // Height 1: Multiply loop: Ragged operand read: Done - "ldr q4, [x12, #0x0]\n" - "ldr q5, [x12, #0x10]\n" + "ldr q21, [x12, #0x0]\n" + "ldr q30, [x12, #0x10]\n" ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n" - ".inst 0x6e44ec08 // bfmmla v8.4s, v0.8h, v4.8h\n" - "ldr q6, [x11, #0x0]\n" - "ldr q7, [x11, #0x10]\n" - ".inst 0x6e45ec0e // bfmmla v14.4s, v0.8h, v5.8h\n" - ".inst 0x6e46ec09 // bfmmla v9.4s, v0.8h, v6.8h\n" - "ldr q4, [x10, #0x0]\n" - "ldr q5, [x10, #0x10]\n" - ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n" - ".inst 0x6e44ec0a // bfmmla v10.4s, v0.8h, v4.8h\n" - "ldr q6, [x9, #0x0]\n" - "ldr q7, [x9, #0x10]\n" - ".inst 0x6e45ec10 // bfmmla v16.4s, v0.8h, v5.8h\n" - ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n" - "ldr q4, [x28, #0x0]\n" - "ldr q5, [x28, #0x10]\n" - ".inst 0x6e47ec11 // bfmmla v17.4s, v0.8h, v7.8h\n" - ".inst 0x6e44ec0c // bfmmla v12.4s, v0.8h, v4.8h\n" - "ldr q6, [x27, #0x0]\n" - "ldr q7, [x27, #0x10]\n" - ".inst 0x6e45ec12 // bfmmla v18.4s, v0.8h, v5.8h\n" - ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n" - ".inst 0x6e47ec13 // bfmmla v19.4s, v0.8h, v7.8h\n" + ".inst 0x6e55ec08 // bfmmla v8.4s, v0.8h, v21.8h\n" + "ldr q21, [x11, #0x0]\n" + "ldr q22, [x11, #0x10]\n" + ".inst 0x6e5eec0e // bfmmla v14.4s, v0.8h, v30.8h\n" + ".inst 0x6e55ec09 // bfmmla v9.4s, v0.8h, v21.8h\n" + "ldr q21, [x10, #0x0]\n" + "ldr q23, [x10, #0x10]\n" + ".inst 0x6e56ec0f // bfmmla v15.4s, v0.8h, v22.8h\n" + ".inst 0x6e55ec0a // bfmmla v10.4s, v0.8h, v21.8h\n" + "ldr q21, [x9, #0x0]\n" + "ldr q22, [x9, #0x10]\n" + ".inst 0x6e57ec10 // bfmmla v16.4s, v0.8h, v23.8h\n" + ".inst 0x6e55ec0b // bfmmla v11.4s, v0.8h, v21.8h\n" + "ldr q21, [x28, #0x0]\n" + "ldr q23, [x28, #0x10]\n" + ".inst 0x6e56ec11 // bfmmla v17.4s, v0.8h, v22.8h\n" + ".inst 0x6e55ec0c // bfmmla v12.4s, v0.8h, v21.8h\n" + "ldr q22, [x27, #0x0]\n" + "ldr q21, [x27, #0x10]\n" + ".inst 0x6e57ec12 // bfmmla v18.4s, v0.8h, v23.8h\n" + ".inst 0x6e56ec0d // bfmmla v13.4s, v0.8h, v22.8h\n" + ".inst 0x6e55ec13 // bfmmla v19.4s, v0.8h, v21.8h\n" "add x12, x12, #0x20\n" "add x11, x11, #0x20\n" "add x10, x10, #0x20\n" @@ -424,21 +424,21 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 ( "uzp1 v13.2d, v13.2d, v19.2d\n" "tbz %x[flags], #1, 30f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1r { v1.4s }, [x20]\n" + "ld1r { v22.4s }, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1r { v0.4s }, [x20]\n" - "fmin v8.4s, v8.4s, v1.4s\n" - "fmin v9.4s, v9.4s, v1.4s\n" - "fmin v10.4s, v10.4s, v1.4s\n" - "fmin v11.4s, v11.4s, v1.4s\n" - "fmin v12.4s, v12.4s, v1.4s\n" - "fmin v13.4s, v13.4s, v1.4s\n" - "fmax v8.4s, v8.4s, v0.4s\n" - "fmax v9.4s, v9.4s, v0.4s\n" - "fmax v10.4s, v10.4s, v0.4s\n" - "fmax v11.4s, v11.4s, v0.4s\n" - "fmax v12.4s, v12.4s, v0.4s\n" - "fmax v13.4s, v13.4s, v0.4s\n" + "ld1r { v21.4s }, [x20]\n" + "fmin v8.4s, v8.4s, v22.4s\n" + "fmin v9.4s, v9.4s, v22.4s\n" + "fmin v10.4s, v10.4s, v22.4s\n" + "fmin v11.4s, v11.4s, v22.4s\n" + "fmin v12.4s, v12.4s, v22.4s\n" + "fmin v13.4s, v13.4s, v22.4s\n" + "fmax v8.4s, v8.4s, v21.4s\n" + "fmax v9.4s, v9.4s, v21.4s\n" + "fmax v10.4s, v10.4s, v21.4s\n" + "fmax v11.4s, v11.4s, v21.4s\n" + "fmax v12.4s, v12.4s, v21.4s\n" + "fmax v13.4s, v13.4s, v21.4s\n" "30:" // Height 1: No activation "cmp x14, #0x18\n" "bge 43f\n" @@ -744,12 +744,12 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 ( "65:" // Height 2: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w25, [x20, x26, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 66f\n" - "ldr x21, [%x[input_ptr], x26, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x24, [x21, #0x0]\n" - "ldr x23, [x21, #0x8]\n" + "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x24, [x20, #0x0]\n" + "ldr x23, [x20, #0x8]\n" "cbnz x26, 67f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x24, x24, x20, LSL #2\n" @@ -757,7 +757,7 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 ( "b 67f\n" "66:" // Height 2: setup direct input "mov x24, %x[input_ptr]\n" - "add x23, x24, x20, LSL #2\n" + "add x23, x24, x21, LSL #2\n" "67:" // Height 2: input setup done "cmp x25, #0x4\n" "blt 70f\n" @@ -774,32 +774,32 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 ( ".inst 0x4ea16820 // bfcvtn2 v0.8h, v1.4s\n" "ld1 { v1.4s }, [x23], #0x10\n" ".inst 0x6e44ec08 // bfmmla v8.4s, v0.8h, v4.8h\n" - "ldr q4, [x10, #0x0]\n" + "ldr q30, [x10, #0x0]\n" ".inst 0x6e45ec0e // bfmmla v14.4s, v0.8h, v5.8h\n" - "ldr q5, [x10, #0x10]\n" + "ldr q23, [x10, #0x10]\n" ".inst 0x6e46ec09 // bfmmla v9.4s, v0.8h, v6.8h\n" - "ldr q6, [x9, #0x0]\n" + "ldr q22, [x9, #0x0]\n" ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n" - "ldr q7, [x9, #0x10]\n" - ".inst 0x6e44ec0a // bfmmla v10.4s, v0.8h, v4.8h\n" - "ldr q4, [x28, #0x0]\n" - ".inst 0x6e45ec10 // bfmmla v16.4s, v0.8h, v5.8h\n" - "ldr q5, [x28, #0x10]\n" - ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n" - "ldr q6, [x27, #0x0]\n" - ".inst 0x6e47ec11 // bfmmla v17.4s, v0.8h, v7.8h\n" - "ldr q7, [x27, #0x10]\n" + "ldr q21, [x9, #0x10]\n" + ".inst 0x6e5eec0a // bfmmla v10.4s, v0.8h, v30.8h\n" + "ldr q2, [x28, #0x0]\n" + ".inst 0x6e57ec10 // bfmmla v16.4s, v0.8h, v23.8h\n" + "ldr q23, [x28, #0x10]\n" + ".inst 0x6e56ec0b // bfmmla v11.4s, v0.8h, v22.8h\n" + "ldr q22, [x27, #0x0]\n" + ".inst 0x6e55ec11 // bfmmla v17.4s, v0.8h, v21.8h\n" + "ldr q21, [x27, #0x10]\n" "sub x25, x25, #0x4\n" "cmp x25, #0x8\n" "add x12, x12, #0x20\n" - ".inst 0x6e44ec0c // bfmmla v12.4s, v0.8h, v4.8h\n" + ".inst 0x6e42ec0c // bfmmla v12.4s, v0.8h, v2.8h\n" "ldr q4, [x12, #0x0]\n" "add x11, x11, #0x20\n" - ".inst 0x6e45ec12 // bfmmla v18.4s, v0.8h, v5.8h\n" + ".inst 0x6e57ec12 // bfmmla v18.4s, v0.8h, v23.8h\n" "ldr q5, [x12, #0x10]\n" - ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n" + ".inst 0x6e56ec0d // bfmmla v13.4s, v0.8h, v22.8h\n" "ldr q6, [x11, #0x0]\n" - ".inst 0x6e47ec13 // bfmmla v19.4s, v0.8h, v7.8h\n" + ".inst 0x6e55ec13 // bfmmla v19.4s, v0.8h, v21.8h\n" "ld1 { v0.4s }, [x24], #0x10\n" "add x10, x10, #0x20\n" "ldr q7, [x11, #0x10]\n" @@ -811,28 +811,28 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 ( ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n" ".inst 0x4ea16820 // bfcvtn2 v0.8h, v1.4s\n" ".inst 0x6e44ec08 // bfmmla v8.4s, v0.8h, v4.8h\n" - "ldr q4, [x10, #0x0]\n" + "ldr q24, [x10, #0x0]\n" ".inst 0x6e45ec0e // bfmmla v14.4s, v0.8h, v5.8h\n" - "ldr q5, [x10, #0x10]\n" + "ldr q23, [x10, #0x10]\n" ".inst 0x6e46ec09 // bfmmla v9.4s, v0.8h, v6.8h\n" - "ldr q6, [x9, #0x0]\n" + "ldr q22, [x9, #0x0]\n" ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n" - "ldr q7, [x9, #0x10]\n" - ".inst 0x6e44ec0a // bfmmla v10.4s, v0.8h, v4.8h\n" - "ldr q4, [x28, #0x0]\n" - ".inst 0x6e45ec10 // bfmmla v16.4s, v0.8h, v5.8h\n" - "ldr q5, [x28, #0x10]\n" - ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n" - "ldr q6, [x27, #0x0]\n" - ".inst 0x6e47ec11 // bfmmla v17.4s, v0.8h, v7.8h\n" - "ldr q7, [x27, #0x10]\n" + "ldr q21, [x9, #0x10]\n" + ".inst 0x6e58ec0a // bfmmla v10.4s, v0.8h, v24.8h\n" + "ldr q24, [x28, #0x0]\n" + ".inst 0x6e57ec10 // bfmmla v16.4s, v0.8h, v23.8h\n" + "ldr q23, [x28, #0x10]\n" + ".inst 0x6e56ec0b // bfmmla v11.4s, v0.8h, v22.8h\n" + "ldr q22, [x27, #0x0]\n" + ".inst 0x6e55ec11 // bfmmla v17.4s, v0.8h, v21.8h\n" + "ldr q21, [x27, #0x10]\n" "sub x25, x25, #0x4\n" - ".inst 0x6e44ec0c // bfmmla v12.4s, v0.8h, v4.8h\n" - ".inst 0x6e45ec12 // bfmmla v18.4s, v0.8h, v5.8h\n" + ".inst 0x6e58ec0c // bfmmla v12.4s, v0.8h, v24.8h\n" + ".inst 0x6e57ec12 // bfmmla v18.4s, v0.8h, v23.8h\n" "add x12, x12, #0x20\n" "add x11, x11, #0x20\n" - ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n" - ".inst 0x6e47ec13 // bfmmla v19.4s, v0.8h, v7.8h\n" + ".inst 0x6e56ec0d // bfmmla v13.4s, v0.8h, v22.8h\n" + ".inst 0x6e55ec13 // bfmmla v19.4s, v0.8h, v21.8h\n" "add x10, x10, #0x20\n" "add x9, x9, #0x20\n" "add x28, x28, #0x20\n" @@ -851,32 +851,32 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 ( "ldr s0, [x24, #0x0]\n" "ldr s1, [x23, #0x0]\n" "72:" // Height 2: Multiply loop: Ragged operand read: Done - "ldr q4, [x12, #0x0]\n" - "ldr q5, [x12, #0x10]\n" + "ldr q24, [x12, #0x0]\n" + "ldr q23, [x12, #0x10]\n" ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n" ".inst 0x4ea16820 // bfcvtn2 v0.8h, v1.4s\n" - "ldr q6, [x11, #0x0]\n" - "ldr q7, [x11, #0x10]\n" - ".inst 0x6e44ec08 // bfmmla v8.4s, v0.8h, v4.8h\n" - ".inst 0x6e45ec0e // bfmmla v14.4s, v0.8h, v5.8h\n" - "ldr q4, [x10, #0x0]\n" - "ldr q5, [x10, #0x10]\n" - ".inst 0x6e46ec09 // bfmmla v9.4s, v0.8h, v6.8h\n" - ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n" - "ldr q6, [x9, #0x0]\n" - "ldr q7, [x9, #0x10]\n" - ".inst 0x6e44ec0a // bfmmla v10.4s, v0.8h, v4.8h\n" - ".inst 0x6e45ec10 // bfmmla v16.4s, v0.8h, v5.8h\n" - "ldr q4, [x28, #0x0]\n" - "ldr q5, [x28, #0x10]\n" - ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n" - ".inst 0x6e47ec11 // bfmmla v17.4s, v0.8h, v7.8h\n" - "ldr q6, [x27, #0x0]\n" - "ldr q7, [x27, #0x10]\n" - ".inst 0x6e44ec0c // bfmmla v12.4s, v0.8h, v4.8h\n" - ".inst 0x6e45ec12 // bfmmla v18.4s, v0.8h, v5.8h\n" - ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n" - ".inst 0x6e47ec13 // bfmmla v19.4s, v0.8h, v7.8h\n" + "ldr q22, [x11, #0x0]\n" + "ldr q21, [x11, #0x10]\n" + ".inst 0x6e58ec08 // bfmmla v8.4s, v0.8h, v24.8h\n" + ".inst 0x6e57ec0e // bfmmla v14.4s, v0.8h, v23.8h\n" + "ldr q24, [x10, #0x0]\n" + "ldr q23, [x10, #0x10]\n" + ".inst 0x6e56ec09 // bfmmla v9.4s, v0.8h, v22.8h\n" + ".inst 0x6e55ec0f // bfmmla v15.4s, v0.8h, v21.8h\n" + "ldr q22, [x9, #0x0]\n" + "ldr q21, [x9, #0x10]\n" + ".inst 0x6e58ec0a // bfmmla v10.4s, v0.8h, v24.8h\n" + ".inst 0x6e57ec10 // bfmmla v16.4s, v0.8h, v23.8h\n" + "ldr q24, [x28, #0x0]\n" + "ldr q23, [x28, #0x10]\n" + ".inst 0x6e56ec0b // bfmmla v11.4s, v0.8h, v22.8h\n" + ".inst 0x6e55ec11 // bfmmla v17.4s, v0.8h, v21.8h\n" + "ldr q22, [x27, #0x0]\n" + "ldr q21, [x27, #0x10]\n" + ".inst 0x6e58ec0c // bfmmla v12.4s, v0.8h, v24.8h\n" + ".inst 0x6e57ec12 // bfmmla v18.4s, v0.8h, v23.8h\n" + ".inst 0x6e56ec0d // bfmmla v13.4s, v0.8h, v22.8h\n" + ".inst 0x6e55ec13 // bfmmla v19.4s, v0.8h, v21.8h\n" "add x12, x12, #0x20\n" "add x11, x11, #0x20\n" "add x10, x10, #0x20\n" @@ -904,33 +904,33 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 ( "uzp2 v13.2d, v13.2d, v19.2d\n" "tbz %x[flags], #1, 74f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1r { v1.4s }, [x20]\n" + "ld1r { v22.4s }, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1r { v0.4s }, [x20]\n" - "fmin v4.4s, v4.4s, v1.4s\n" - "fmin v14.4s, v14.4s, v1.4s\n" - "fmin v15.4s, v15.4s, v1.4s\n" - "fmin v16.4s, v16.4s, v1.4s\n" - "fmin v17.4s, v17.4s, v1.4s\n" - "fmin v18.4s, v18.4s, v1.4s\n" - "fmin v8.4s, v8.4s, v1.4s\n" - "fmin v9.4s, v9.4s, v1.4s\n" - "fmin v10.4s, v10.4s, v1.4s\n" - "fmin v11.4s, v11.4s, v1.4s\n" - "fmin v12.4s, v12.4s, v1.4s\n" - "fmin v13.4s, v13.4s, v1.4s\n" - "fmax v4.4s, v4.4s, v0.4s\n" - "fmax v14.4s, v14.4s, v0.4s\n" - "fmax v15.4s, v15.4s, v0.4s\n" - "fmax v16.4s, v16.4s, v0.4s\n" - "fmax v17.4s, v17.4s, v0.4s\n" - "fmax v18.4s, v18.4s, v0.4s\n" - "fmax v8.4s, v8.4s, v0.4s\n" - "fmax v9.4s, v9.4s, v0.4s\n" - "fmax v10.4s, v10.4s, v0.4s\n" - "fmax v11.4s, v11.4s, v0.4s\n" - "fmax v12.4s, v12.4s, v0.4s\n" - "fmax v13.4s, v13.4s, v0.4s\n" + "ld1r { v21.4s }, [x20]\n" + "fmin v4.4s, v4.4s, v22.4s\n" + "fmin v14.4s, v14.4s, v22.4s\n" + "fmin v15.4s, v15.4s, v22.4s\n" + "fmin v16.4s, v16.4s, v22.4s\n" + "fmin v17.4s, v17.4s, v22.4s\n" + "fmin v18.4s, v18.4s, v22.4s\n" + "fmin v8.4s, v8.4s, v22.4s\n" + "fmin v9.4s, v9.4s, v22.4s\n" + "fmin v10.4s, v10.4s, v22.4s\n" + "fmin v11.4s, v11.4s, v22.4s\n" + "fmin v12.4s, v12.4s, v22.4s\n" + "fmin v13.4s, v13.4s, v22.4s\n" + "fmax v4.4s, v4.4s, v21.4s\n" + "fmax v14.4s, v14.4s, v21.4s\n" + "fmax v15.4s, v15.4s, v21.4s\n" + "fmax v16.4s, v16.4s, v21.4s\n" + "fmax v17.4s, v17.4s, v21.4s\n" + "fmax v18.4s, v18.4s, v21.4s\n" + "fmax v8.4s, v8.4s, v21.4s\n" + "fmax v9.4s, v9.4s, v21.4s\n" + "fmax v10.4s, v10.4s, v21.4s\n" + "fmax v11.4s, v11.4s, v21.4s\n" + "fmax v12.4s, v12.4s, v21.4s\n" + "fmax v13.4s, v13.4s, v21.4s\n" "74:" // Height 2: No activation "cmp x14, #0x18\n" "bge 87f\n" @@ -1339,13 +1339,13 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 ( "109:" // Height 3: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w25, [x20, x26, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 110f\n" - "ldr x21, [%x[input_ptr], x26, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x24, [x21, #0x0]\n" - "ldr x23, [x21, #0x8]\n" - "ldr x22, [x21, #0x10]\n" + "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x24, [x20, #0x0]\n" + "ldr x23, [x20, #0x8]\n" + "ldr x22, [x20, #0x10]\n" "cbnz x26, 111f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x24, x24, x20, LSL #2\n" @@ -1354,8 +1354,8 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 ( "b 111f\n" "110:" // Height 3: setup direct input "mov x24, %x[input_ptr]\n" - "add x23, x24, x20, LSL #2\n" - "add x22, x23, x20, LSL #2\n" + "add x23, x24, x21, LSL #2\n" + "add x22, x23, x21, LSL #2\n" "111:" // Height 3: input setup done "cmp x25, #0x4\n" "blt 114f\n" @@ -1386,7 +1386,7 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 ( ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n" "cmp x25, #0x8\n" ".inst 0x6e47ec5b // bfmmla v27.4s, v2.8h, v7.8h\n" - "ldr q7, [x9, #0x10]\n" + "ldr q3, [x9, #0x10]\n" ".inst 0x6e44ec0a // bfmmla v10.4s, v0.8h, v4.8h\n" "add x12, x12, #0x20\n" ".inst 0x6e44ec56 // bfmmla v22.4s, v2.8h, v4.8h\n" @@ -1399,10 +1399,10 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 ( "add x10, x10, #0x20\n" ".inst 0x6e46ec57 // bfmmla v23.4s, v2.8h, v6.8h\n" "ldr q6, [x27, #0x0]\n" - ".inst 0x6e47ec11 // bfmmla v17.4s, v0.8h, v7.8h\n" + ".inst 0x6e43ec11 // bfmmla v17.4s, v0.8h, v3.8h\n" "add x9, x9, #0x20\n" - ".inst 0x6e47ec5d // bfmmla v29.4s, v2.8h, v7.8h\n" - "ldr q7, [x27, #0x10]\n" + ".inst 0x6e43ec5d // bfmmla v29.4s, v2.8h, v3.8h\n" + "ldr q3, [x27, #0x10]\n" "add x28, x28, #0x20\n" ".inst 0x6e44ec0c // bfmmla v12.4s, v0.8h, v4.8h\n" "add x27, x27, #0x20\n" @@ -1414,9 +1414,9 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 ( ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n" ".inst 0x6e46ec59 // bfmmla v25.4s, v2.8h, v6.8h\n" "ldr q6, [x11, #0x0]\n" - ".inst 0x6e47ec13 // bfmmla v19.4s, v0.8h, v7.8h\n" + ".inst 0x6e43ec13 // bfmmla v19.4s, v0.8h, v3.8h\n" "ld1 { v0.4s }, [x24], #0x10\n" - ".inst 0x6e47ec5f // bfmmla v31.4s, v2.8h, v7.8h\n" + ".inst 0x6e43ec5f // bfmmla v31.4s, v2.8h, v3.8h\n" "ld1 { v2.4s }, [x22], #0x10\n" "ldr q7, [x11, #0x10]\n" "bge 112b\n" @@ -1427,10 +1427,10 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 ( "sub x25, x25, #0x4\n" ".inst 0x0ea16842 // bfcvtn v2.4h, v2.4s\n" ".inst 0x6e44ec54 // bfmmla v20.4s, v2.8h, v4.8h\n" - "ldr q4, [x10, #0x0]\n" + "ldr q3, [x10, #0x0]\n" ".inst 0x6e45ec0e // bfmmla v14.4s, v0.8h, v5.8h\n" ".inst 0x6e45ec5a // bfmmla v26.4s, v2.8h, v5.8h\n" - "ldr q5, [x10, #0x10]\n" + "ldr q4, [x10, #0x10]\n" ".inst 0x6e46ec09 // bfmmla v9.4s, v0.8h, v6.8h\n" "add x12, x12, #0x20\n" ".inst 0x6e46ec55 // bfmmla v21.4s, v2.8h, v6.8h\n" @@ -1438,31 +1438,31 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 ( ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n" "add x11, x11, #0x20\n" ".inst 0x6e47ec5b // bfmmla v27.4s, v2.8h, v7.8h\n" - "ldr q7, [x9, #0x10]\n" - ".inst 0x6e44ec0a // bfmmla v10.4s, v0.8h, v4.8h\n" + "ldr q1, [x9, #0x10]\n" + ".inst 0x6e43ec0a // bfmmla v10.4s, v0.8h, v3.8h\n" "add x10, x10, #0x20\n" - ".inst 0x6e44ec56 // bfmmla v22.4s, v2.8h, v4.8h\n" - "ldr q4, [x28, #0x0]\n" - ".inst 0x6e45ec10 // bfmmla v16.4s, v0.8h, v5.8h\n" + ".inst 0x6e43ec56 // bfmmla v22.4s, v2.8h, v3.8h\n" + "ldr q5, [x28, #0x0]\n" + ".inst 0x6e44ec10 // bfmmla v16.4s, v0.8h, v4.8h\n" "add x9, x9, #0x20\n" - ".inst 0x6e45ec5c // bfmmla v28.4s, v2.8h, v5.8h\n" - "ldr q5, [x28, #0x10]\n" + ".inst 0x6e44ec5c // bfmmla v28.4s, v2.8h, v4.8h\n" + "ldr q4, [x28, #0x10]\n" ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n" "add x28, x28, #0x20\n" ".inst 0x6e46ec57 // bfmmla v23.4s, v2.8h, v6.8h\n" - "ldr q6, [x27, #0x0]\n" - ".inst 0x6e47ec11 // bfmmla v17.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec5d // bfmmla v29.4s, v2.8h, v7.8h\n" - "ldr q7, [x27, #0x10]\n" + "ldr q3, [x27, #0x0]\n" + ".inst 0x6e41ec11 // bfmmla v17.4s, v0.8h, v1.8h\n" + ".inst 0x6e41ec5d // bfmmla v29.4s, v2.8h, v1.8h\n" + "ldr q1, [x27, #0x10]\n" "add x27, x27, #0x20\n" - ".inst 0x6e44ec0c // bfmmla v12.4s, v0.8h, v4.8h\n" - ".inst 0x6e44ec58 // bfmmla v24.4s, v2.8h, v4.8h\n" - ".inst 0x6e45ec12 // bfmmla v18.4s, v0.8h, v5.8h\n" - ".inst 0x6e45ec5e // bfmmla v30.4s, v2.8h, v5.8h\n" - ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec59 // bfmmla v25.4s, v2.8h, v6.8h\n" - ".inst 0x6e47ec13 // bfmmla v19.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec5f // bfmmla v31.4s, v2.8h, v7.8h\n" + ".inst 0x6e45ec0c // bfmmla v12.4s, v0.8h, v5.8h\n" + ".inst 0x6e45ec58 // bfmmla v24.4s, v2.8h, v5.8h\n" + ".inst 0x6e44ec12 // bfmmla v18.4s, v0.8h, v4.8h\n" + ".inst 0x6e44ec5e // bfmmla v30.4s, v2.8h, v4.8h\n" + ".inst 0x6e43ec0d // bfmmla v13.4s, v0.8h, v3.8h\n" + ".inst 0x6e43ec59 // bfmmla v25.4s, v2.8h, v3.8h\n" + ".inst 0x6e41ec13 // bfmmla v19.4s, v0.8h, v1.8h\n" + ".inst 0x6e41ec5f // bfmmla v31.4s, v2.8h, v1.8h\n" "114:" // Height 3: Multiply loop: Main loop skip "cbz x25, 117f\n" "cbz x25, 117f\n" @@ -1480,51 +1480,51 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 ( "ldr s1, [x23, #0x0]\n" "ldr s2, [x22, #0x0]\n" "116:" // Height 3: Multiply loop: Ragged operand read: Done - "ldr q4, [x12, #0x0]\n" - "ldr q5, [x12, #0x10]\n" + "ldr q5, [x12, #0x0]\n" + "ldr q4, [x12, #0x10]\n" ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n" ".inst 0x4ea16820 // bfcvtn2 v0.8h, v1.4s\n" - "ldr q6, [x11, #0x0]\n" - "ldr q7, [x11, #0x10]\n" + "ldr q3, [x11, #0x0]\n" + "ldr q1, [x11, #0x10]\n" ".inst 0x0ea16842 // bfcvtn v2.4h, v2.4s\n" - ".inst 0x6e44ec08 // bfmmla v8.4s, v0.8h, v4.8h\n" - ".inst 0x6e44ec54 // bfmmla v20.4s, v2.8h, v4.8h\n" - "ldr q4, [x10, #0x0]\n" - ".inst 0x6e45ec0e // bfmmla v14.4s, v0.8h, v5.8h\n" + ".inst 0x6e45ec08 // bfmmla v8.4s, v0.8h, v5.8h\n" + ".inst 0x6e45ec54 // bfmmla v20.4s, v2.8h, v5.8h\n" + "ldr q5, [x10, #0x0]\n" + ".inst 0x6e44ec0e // bfmmla v14.4s, v0.8h, v4.8h\n" "add x12, x12, #0x20\n" - ".inst 0x6e45ec5a // bfmmla v26.4s, v2.8h, v5.8h\n" - "ldr q5, [x10, #0x10]\n" - ".inst 0x6e46ec09 // bfmmla v9.4s, v0.8h, v6.8h\n" + ".inst 0x6e44ec5a // bfmmla v26.4s, v2.8h, v4.8h\n" + "ldr q4, [x10, #0x10]\n" + ".inst 0x6e43ec09 // bfmmla v9.4s, v0.8h, v3.8h\n" "add x11, x11, #0x20\n" - ".inst 0x6e46ec55 // bfmmla v21.4s, v2.8h, v6.8h\n" - "ldr q6, [x9, #0x0]\n" - ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n" + ".inst 0x6e43ec55 // bfmmla v21.4s, v2.8h, v3.8h\n" + "ldr q3, [x9, #0x0]\n" + ".inst 0x6e41ec0f // bfmmla v15.4s, v0.8h, v1.8h\n" "add x10, x10, #0x20\n" - ".inst 0x6e47ec5b // bfmmla v27.4s, v2.8h, v7.8h\n" - "ldr q7, [x9, #0x10]\n" - ".inst 0x6e44ec0a // bfmmla v10.4s, v0.8h, v4.8h\n" + ".inst 0x6e41ec5b // bfmmla v27.4s, v2.8h, v1.8h\n" + "ldr q1, [x9, #0x10]\n" + ".inst 0x6e45ec0a // bfmmla v10.4s, v0.8h, v5.8h\n" "add x9, x9, #0x20\n" - ".inst 0x6e44ec56 // bfmmla v22.4s, v2.8h, v4.8h\n" - "ldr q4, [x28, #0x0]\n" - ".inst 0x6e45ec10 // bfmmla v16.4s, v0.8h, v5.8h\n" - ".inst 0x6e45ec5c // bfmmla v28.4s, v2.8h, v5.8h\n" - "ldr q5, [x28, #0x10]\n" - ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n" + ".inst 0x6e45ec56 // bfmmla v22.4s, v2.8h, v5.8h\n" + "ldr q5, [x28, #0x0]\n" + ".inst 0x6e44ec10 // bfmmla v16.4s, v0.8h, v4.8h\n" + ".inst 0x6e44ec5c // bfmmla v28.4s, v2.8h, v4.8h\n" + "ldr q4, [x28, #0x10]\n" + ".inst 0x6e43ec0b // bfmmla v11.4s, v0.8h, v3.8h\n" "add x28, x28, #0x20\n" - ".inst 0x6e46ec57 // bfmmla v23.4s, v2.8h, v6.8h\n" - "ldr q6, [x27, #0x0]\n" - ".inst 0x6e47ec11 // bfmmla v17.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec5d // bfmmla v29.4s, v2.8h, v7.8h\n" - "ldr q7, [x27, #0x10]\n" - ".inst 0x6e44ec0c // bfmmla v12.4s, v0.8h, v4.8h\n" + ".inst 0x6e43ec57 // bfmmla v23.4s, v2.8h, v3.8h\n" + "ldr q3, [x27, #0x0]\n" + ".inst 0x6e41ec11 // bfmmla v17.4s, v0.8h, v1.8h\n" + ".inst 0x6e41ec5d // bfmmla v29.4s, v2.8h, v1.8h\n" + "ldr q1, [x27, #0x10]\n" + ".inst 0x6e45ec0c // bfmmla v12.4s, v0.8h, v5.8h\n" "add x27, x27, #0x20\n" - ".inst 0x6e44ec58 // bfmmla v24.4s, v2.8h, v4.8h\n" - ".inst 0x6e45ec12 // bfmmla v18.4s, v0.8h, v5.8h\n" - ".inst 0x6e45ec5e // bfmmla v30.4s, v2.8h, v5.8h\n" - ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec59 // bfmmla v25.4s, v2.8h, v6.8h\n" - ".inst 0x6e47ec13 // bfmmla v19.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec5f // bfmmla v31.4s, v2.8h, v7.8h\n" + ".inst 0x6e45ec58 // bfmmla v24.4s, v2.8h, v5.8h\n" + ".inst 0x6e44ec12 // bfmmla v18.4s, v0.8h, v4.8h\n" + ".inst 0x6e44ec5e // bfmmla v30.4s, v2.8h, v4.8h\n" + ".inst 0x6e43ec0d // bfmmla v13.4s, v0.8h, v3.8h\n" + ".inst 0x6e43ec59 // bfmmla v25.4s, v2.8h, v3.8h\n" + ".inst 0x6e41ec13 // bfmmla v19.4s, v0.8h, v1.8h\n" + ".inst 0x6e41ec5f // bfmmla v31.4s, v2.8h, v1.8h\n" "117:" // Height 3: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x26, x26, #0x1\n" @@ -2070,14 +2070,14 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 ( "153:" // Height 4: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w25, [x20, x26, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 154f\n" - "ldr x21, [%x[input_ptr], x26, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x24, [x21, #0x0]\n" - "ldr x23, [x21, #0x8]\n" - "ldr x22, [x21, #0x10]\n" - "ldr x21, [x21, #0x18]\n" + "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x24, [x20, #0x0]\n" + "ldr x23, [x20, #0x8]\n" + "ldr x22, [x20, #0x10]\n" + "ldr x21, [x20, #0x18]\n" "cbnz x26, 155f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x24, x24, x20, LSL #2\n" @@ -2087,9 +2087,9 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 ( "b 155f\n" "154:" // Height 4: setup direct input "mov x24, %x[input_ptr]\n" - "add x23, x24, x20, LSL #2\n" - "add x22, x23, x20, LSL #2\n" - "add x21, x22, x20, LSL #2\n" + "add x23, x24, x21, LSL #2\n" + "add x22, x23, x21, LSL #2\n" + "add x21, x22, x21, LSL #2\n" "155:" // Height 4: input setup done "cmp x25, #0x4\n" "blt 158f\n" @@ -2167,40 +2167,40 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 ( ".inst 0x6e44ec08 // bfmmla v8.4s, v0.8h, v4.8h\n" "add x11, x11, #0x20\n" ".inst 0x6e44ec54 // bfmmla v20.4s, v2.8h, v4.8h\n" - "ldr q4, [x10, #0x0]\n" + "ldr q3, [x10, #0x0]\n" ".inst 0x6e45ec0e // bfmmla v14.4s, v0.8h, v5.8h\n" ".inst 0x6e45ec5a // bfmmla v26.4s, v2.8h, v5.8h\n" - "ldr q5, [x10, #0x10]\n" + "ldr q4, [x10, #0x10]\n" ".inst 0x6e46ec09 // bfmmla v9.4s, v0.8h, v6.8h\n" "add x10, x10, #0x20\n" ".inst 0x6e46ec55 // bfmmla v21.4s, v2.8h, v6.8h\n" "ldr q6, [x9, #0x0]\n" ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n" ".inst 0x6e47ec5b // bfmmla v27.4s, v2.8h, v7.8h\n" - "ldr q7, [x9, #0x10]\n" - ".inst 0x6e44ec0a // bfmmla v10.4s, v0.8h, v4.8h\n" + "ldr q1, [x9, #0x10]\n" + ".inst 0x6e43ec0a // bfmmla v10.4s, v0.8h, v3.8h\n" "add x9, x9, #0x20\n" - ".inst 0x6e44ec56 // bfmmla v22.4s, v2.8h, v4.8h\n" - "ldr q4, [x28, #0x0]\n" - ".inst 0x6e45ec10 // bfmmla v16.4s, v0.8h, v5.8h\n" - ".inst 0x6e45ec5c // bfmmla v28.4s, v2.8h, v5.8h\n" - "ldr q5, [x28, #0x10]\n" + ".inst 0x6e43ec56 // bfmmla v22.4s, v2.8h, v3.8h\n" + "ldr q5, [x28, #0x0]\n" + ".inst 0x6e44ec10 // bfmmla v16.4s, v0.8h, v4.8h\n" + ".inst 0x6e44ec5c // bfmmla v28.4s, v2.8h, v4.8h\n" + "ldr q4, [x28, #0x10]\n" ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n" "add x28, x28, #0x20\n" ".inst 0x6e46ec57 // bfmmla v23.4s, v2.8h, v6.8h\n" - "ldr q6, [x27, #0x0]\n" - ".inst 0x6e47ec11 // bfmmla v17.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec5d // bfmmla v29.4s, v2.8h, v7.8h\n" - "ldr q7, [x27, #0x10]\n" + "ldr q3, [x27, #0x0]\n" + ".inst 0x6e41ec11 // bfmmla v17.4s, v0.8h, v1.8h\n" + ".inst 0x6e41ec5d // bfmmla v29.4s, v2.8h, v1.8h\n" + "ldr q1, [x27, #0x10]\n" "add x27, x27, #0x20\n" - ".inst 0x6e44ec0c // bfmmla v12.4s, v0.8h, v4.8h\n" - ".inst 0x6e44ec58 // bfmmla v24.4s, v2.8h, v4.8h\n" - ".inst 0x6e45ec12 // bfmmla v18.4s, v0.8h, v5.8h\n" - ".inst 0x6e45ec5e // bfmmla v30.4s, v2.8h, v5.8h\n" - ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec59 // bfmmla v25.4s, v2.8h, v6.8h\n" - ".inst 0x6e47ec13 // bfmmla v19.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec5f // bfmmla v31.4s, v2.8h, v7.8h\n" + ".inst 0x6e45ec0c // bfmmla v12.4s, v0.8h, v5.8h\n" + ".inst 0x6e45ec58 // bfmmla v24.4s, v2.8h, v5.8h\n" + ".inst 0x6e44ec12 // bfmmla v18.4s, v0.8h, v4.8h\n" + ".inst 0x6e44ec5e // bfmmla v30.4s, v2.8h, v4.8h\n" + ".inst 0x6e43ec0d // bfmmla v13.4s, v0.8h, v3.8h\n" + ".inst 0x6e43ec59 // bfmmla v25.4s, v2.8h, v3.8h\n" + ".inst 0x6e41ec13 // bfmmla v19.4s, v0.8h, v1.8h\n" + ".inst 0x6e41ec5f // bfmmla v31.4s, v2.8h, v1.8h\n" "158:" // Height 4: Multiply loop: Main loop skip "cbz x25, 161f\n" "cbz x25, 161f\n" @@ -2221,52 +2221,52 @@ void a64_ffhybrid_fp32bf16fp32_mmla_4x24 ( "ldr s2, [x22, #0x0]\n" "ldr s3, [x21, #0x0]\n" "160:" // Height 4: Multiply loop: Ragged operand read: Done - "ldr q4, [x12, #0x0]\n" - "ldr q5, [x12, #0x10]\n" + "ldr q5, [x12, #0x0]\n" + "ldr q4, [x12, #0x10]\n" ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n" ".inst 0x0ea16842 // bfcvtn v2.4h, v2.4s\n" - "ldr q6, [x11, #0x0]\n" - "ldr q7, [x11, #0x10]\n" + "ldr q7, [x11, #0x0]\n" + "ldr q6, [x11, #0x10]\n" ".inst 0x4ea16820 // bfcvtn2 v0.8h, v1.4s\n" ".inst 0x4ea16862 // bfcvtn2 v2.8h, v3.4s\n" - ".inst 0x6e44ec08 // bfmmla v8.4s, v0.8h, v4.8h\n" - ".inst 0x6e44ec54 // bfmmla v20.4s, v2.8h, v4.8h\n" - "ldr q4, [x10, #0x0]\n" + ".inst 0x6e45ec08 // bfmmla v8.4s, v0.8h, v5.8h\n" + ".inst 0x6e45ec54 // bfmmla v20.4s, v2.8h, v5.8h\n" + "ldr q5, [x10, #0x0]\n" "add x12, x12, #0x20\n" - ".inst 0x6e45ec0e // bfmmla v14.4s, v0.8h, v5.8h\n" - ".inst 0x6e45ec5a // bfmmla v26.4s, v2.8h, v5.8h\n" - "ldr q5, [x10, #0x10]\n" + ".inst 0x6e44ec0e // bfmmla v14.4s, v0.8h, v4.8h\n" + ".inst 0x6e44ec5a // bfmmla v26.4s, v2.8h, v4.8h\n" + "ldr q4, [x10, #0x10]\n" "add x11, x11, #0x20\n" - ".inst 0x6e46ec09 // bfmmla v9.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec55 // bfmmla v21.4s, v2.8h, v6.8h\n" - "ldr q6, [x9, #0x0]\n" + ".inst 0x6e47ec09 // bfmmla v9.4s, v0.8h, v7.8h\n" + ".inst 0x6e47ec55 // bfmmla v21.4s, v2.8h, v7.8h\n" + "ldr q3, [x9, #0x0]\n" "add x10, x10, #0x20\n" - ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec5b // bfmmla v27.4s, v2.8h, v7.8h\n" - "ldr q7, [x9, #0x10]\n" + ".inst 0x6e46ec0f // bfmmla v15.4s, v0.8h, v6.8h\n" + ".inst 0x6e46ec5b // bfmmla v27.4s, v2.8h, v6.8h\n" + "ldr q1, [x9, #0x10]\n" "add x9, x9, #0x20\n" - ".inst 0x6e44ec0a // bfmmla v10.4s, v0.8h, v4.8h\n" - ".inst 0x6e44ec56 // bfmmla v22.4s, v2.8h, v4.8h\n" - "ldr q4, [x28, #0x0]\n" - ".inst 0x6e45ec10 // bfmmla v16.4s, v0.8h, v5.8h\n" - ".inst 0x6e45ec5c // bfmmla v28.4s, v2.8h, v5.8h\n" - "ldr q5, [x28, #0x10]\n" + ".inst 0x6e45ec0a // bfmmla v10.4s, v0.8h, v5.8h\n" + ".inst 0x6e45ec56 // bfmmla v22.4s, v2.8h, v5.8h\n" + "ldr q5, [x28, #0x0]\n" + ".inst 0x6e44ec10 // bfmmla v16.4s, v0.8h, v4.8h\n" + ".inst 0x6e44ec5c // bfmmla v28.4s, v2.8h, v4.8h\n" + "ldr q4, [x28, #0x10]\n" "add x28, x28, #0x20\n" - ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec57 // bfmmla v23.4s, v2.8h, v6.8h\n" - "ldr q6, [x27, #0x0]\n" - ".inst 0x6e47ec11 // bfmmla v17.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec5d // bfmmla v29.4s, v2.8h, v7.8h\n" - "ldr q7, [x27, #0x10]\n" + ".inst 0x6e43ec0b // bfmmla v11.4s, v0.8h, v3.8h\n" + ".inst 0x6e43ec57 // bfmmla v23.4s, v2.8h, v3.8h\n" + "ldr q3, [x27, #0x0]\n" + ".inst 0x6e41ec11 // bfmmla v17.4s, v0.8h, v1.8h\n" + ".inst 0x6e41ec5d // bfmmla v29.4s, v2.8h, v1.8h\n" + "ldr q1, [x27, #0x10]\n" "add x27, x27, #0x20\n" - ".inst 0x6e44ec0c // bfmmla v12.4s, v0.8h, v4.8h\n" - ".inst 0x6e44ec58 // bfmmla v24.4s, v2.8h, v4.8h\n" - ".inst 0x6e45ec12 // bfmmla v18.4s, v0.8h, v5.8h\n" - ".inst 0x6e45ec5e // bfmmla v30.4s, v2.8h, v5.8h\n" - ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec59 // bfmmla v25.4s, v2.8h, v6.8h\n" - ".inst 0x6e47ec13 // bfmmla v19.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec5f // bfmmla v31.4s, v2.8h, v7.8h\n" + ".inst 0x6e45ec0c // bfmmla v12.4s, v0.8h, v5.8h\n" + ".inst 0x6e45ec58 // bfmmla v24.4s, v2.8h, v5.8h\n" + ".inst 0x6e44ec12 // bfmmla v18.4s, v0.8h, v4.8h\n" + ".inst 0x6e44ec5e // bfmmla v30.4s, v2.8h, v4.8h\n" + ".inst 0x6e43ec0d // bfmmla v13.4s, v0.8h, v3.8h\n" + ".inst 0x6e43ec59 // bfmmla v25.4s, v2.8h, v3.8h\n" + ".inst 0x6e41ec13 // bfmmla v19.4s, v0.8h, v1.8h\n" + ".inst 0x6e41ec5f // bfmmla v31.4s, v2.8h, v1.8h\n" "161:" // Height 4: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x26, x26, #0x1\n" -- cgit v1.2.1