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 --- .../sve_hybrid_fp32bf16fp32_mmla_6x4VL/generic.cpp | 1473 ++++++++++---------- 1 file changed, 736 insertions(+), 737 deletions(-) (limited to 'src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32bf16fp32_mmla_6x4VL/generic.cpp') diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32bf16fp32_mmla_6x4VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32bf16fp32_mmla_6x4VL/generic.cpp index 23d7ff9c3b..0d2b47ec39 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32bf16fp32_mmla_6x4VL/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32bf16fp32_mmla_6x4VL/generic.cpp @@ -133,16 +133,16 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "b 5f\n" "3:" // Height 1: no bias "tbz %x[flags], #0, 4f\n" - "ld1w { z9.s }, p4/Z, [x9]\n" - "ld1w { z10.s }, p3/Z, [x9, #1, MUL VL]\n" - "zip1 z8.d, z9.d, z12.d\n" - "zip2 z12.d, z9.d, z12.d\n" - "ld1w { z11.s }, p2/Z, [x9, #2, MUL VL]\n" + "ld1w { z16.s }, p4/Z, [x9]\n" + "ld1w { z18.s }, p3/Z, [x9, #1, MUL VL]\n" + "zip1 z8.d, z16.d, z12.d\n" + "zip2 z12.d, z16.d, z12.d\n" + "ld1w { z17.s }, p2/Z, [x9, #2, MUL VL]\n" "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n" - "zip1 z9.d, z10.d, z13.d\n" - "zip2 z13.d, z10.d, z13.d\n" - "zip1 z10.d, z11.d, z14.d\n" - "zip2 z14.d, z11.d, z14.d\n" + "zip1 z9.d, z18.d, z13.d\n" + "zip2 z13.d, z18.d, z13.d\n" + "zip1 z10.d, z17.d, z14.d\n" + "zip2 z14.d, z17.d, z14.d\n" "zip1 z11.d, z16.d, z15.d\n" "zip2 z15.d, z16.d, z15.d\n" "b 5f\n" @@ -160,11 +160,11 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "6:" // Height 1: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w27, [x20, x28, 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, 7f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" "cbnz x28, 8f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #2\n" @@ -176,51 +176,51 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "ble 10f\n" "9:" // Height 1: Multiply loop: Main loop head "whilelt p0.s, XZR, x27\n" - "ld1rqw { z0.s }, p0/Z, [x26]\n" - ".inst 0x658ab400 // bfcvt z0.h, p5/M, z0.s\n" - "uzp1 z0.h, z0.h, z0.h\n" - "ld1h { z6.h }, p5/Z, [x10]\n" - "ld1h { z7.h }, p5/Z, [x10, #1, MUL VL]\n" - ".inst 0x6466e408 // bfmmla z8.s, z0.h, z6.h\n" - ".inst 0x6467e40c // bfmmla z12.s, z0.h, z7.h\n" - "ld1h { z6.h }, p5/Z, [x10, #2, MUL VL]\n" - "ld1h { z7.h }, p5/Z, [x10, #3, MUL VL]\n" - ".inst 0x6466e409 // bfmmla z9.s, z0.h, z6.h\n" - ".inst 0x6467e40d // bfmmla z13.s, z0.h, z7.h\n" - "ld1h { z6.h }, p5/Z, [x10, #4, MUL VL]\n" - "ld1h { z7.h }, p5/Z, [x10, #5, MUL VL]\n" - ".inst 0x6466e40a // bfmmla z10.s, z0.h, z6.h\n" - ".inst 0x6467e40e // bfmmla z14.s, z0.h, z7.h\n" - "ld1h { z6.h }, p5/Z, [x10, #6, MUL VL]\n" - "ld1h { z7.h }, p5/Z, [x10, #7, MUL VL]\n" + "ld1rqw { z18.s }, p0/Z, [x26]\n" + ".inst 0x658ab652 // bfcvt z18.h, p5/M, z18.s\n" + "uzp1 z18.h, z18.h, z18.h\n" + "ld1h { z17.h }, p5/Z, [x10]\n" + "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n" + ".inst 0x6471e648 // bfmmla z8.s, z18.h, z17.h\n" + ".inst 0x6470e64c // bfmmla z12.s, z18.h, z16.h\n" + "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n" + "ld1h { z16.h }, p5/Z, [x10, #3, MUL VL]\n" + ".inst 0x6471e649 // bfmmla z9.s, z18.h, z17.h\n" + ".inst 0x6470e64d // bfmmla z13.s, z18.h, z16.h\n" + "ld1h { z17.h }, p5/Z, [x10, #4, MUL VL]\n" + "ld1h { z16.h }, p5/Z, [x10, #5, MUL VL]\n" + ".inst 0x6471e64a // bfmmla z10.s, z18.h, z17.h\n" + ".inst 0x6470e64e // bfmmla z14.s, z18.h, z16.h\n" + "ld1h { z17.h }, p5/Z, [x10, #6, MUL VL]\n" + "ld1h { z16.h }, p5/Z, [x10, #7, MUL VL]\n" "sub x27, x27, #0x4\n" "cmp x27, #0x4\n" - ".inst 0x6466e40b // bfmmla z11.s, z0.h, z6.h\n" - ".inst 0x6467e40f // bfmmla z15.s, z0.h, z7.h\n" + ".inst 0x6471e64b // bfmmla z11.s, z18.h, z17.h\n" + ".inst 0x6470e64f // bfmmla z15.s, z18.h, z16.h\n" "add x26, x26, #0x10\n" "addvl x10, x10, #8\n" "bgt 9b\n" "10:" // Height 1: Multiply loop: Single iteration only "whilelt p0.s, XZR, x27\n" - "ld1rqw { z0.s }, p0/Z, [x26]\n" - ".inst 0x658ab400 // bfcvt z0.h, p5/M, z0.s\n" - "uzp1 z0.h, z0.h, z0.h\n" - "ld1h { z6.h }, p5/Z, [x10]\n" - "ld1h { z7.h }, p5/Z, [x10, #1, MUL VL]\n" - ".inst 0x6466e408 // bfmmla z8.s, z0.h, z6.h\n" - ".inst 0x6467e40c // bfmmla z12.s, z0.h, z7.h\n" - "ld1h { z6.h }, p5/Z, [x10, #2, MUL VL]\n" - "ld1h { z7.h }, p5/Z, [x10, #3, MUL VL]\n" - ".inst 0x6466e409 // bfmmla z9.s, z0.h, z6.h\n" - ".inst 0x6467e40d // bfmmla z13.s, z0.h, z7.h\n" - "ld1h { z6.h }, p5/Z, [x10, #4, MUL VL]\n" - "ld1h { z7.h }, p5/Z, [x10, #5, MUL VL]\n" - ".inst 0x6466e40a // bfmmla z10.s, z0.h, z6.h\n" - ".inst 0x6467e40e // bfmmla z14.s, z0.h, z7.h\n" - "ld1h { z6.h }, p5/Z, [x10, #6, MUL VL]\n" - "ld1h { z7.h }, p5/Z, [x10, #7, MUL VL]\n" - ".inst 0x6466e40b // bfmmla z11.s, z0.h, z6.h\n" - ".inst 0x6467e40f // bfmmla z15.s, z0.h, z7.h\n" + "ld1rqw { z18.s }, p0/Z, [x26]\n" + ".inst 0x658ab652 // bfcvt z18.h, p5/M, z18.s\n" + "uzp1 z18.h, z18.h, z18.h\n" + "ld1h { z17.h }, p5/Z, [x10]\n" + "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n" + ".inst 0x6471e648 // bfmmla z8.s, z18.h, z17.h\n" + ".inst 0x6470e64c // bfmmla z12.s, z18.h, z16.h\n" + "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n" + "ld1h { z16.h }, p5/Z, [x10, #3, MUL VL]\n" + ".inst 0x6471e649 // bfmmla z9.s, z18.h, z17.h\n" + ".inst 0x6470e64d // bfmmla z13.s, z18.h, z16.h\n" + "ld1h { z17.h }, p5/Z, [x10, #4, MUL VL]\n" + "ld1h { z16.h }, p5/Z, [x10, #5, MUL VL]\n" + ".inst 0x6471e64a // bfmmla z10.s, z18.h, z17.h\n" + ".inst 0x6470e64e // bfmmla z14.s, z18.h, z16.h\n" + "ld1h { z17.h }, p5/Z, [x10, #6, MUL VL]\n" + "ld1h { z16.h }, p5/Z, [x10, #7, MUL VL]\n" + ".inst 0x6471e64b // bfmmla z11.s, z18.h, z17.h\n" + ".inst 0x6470e64f // bfmmla z15.s, z18.h, z16.h\n" "addvl x10, x10, #8\n" "11:" // Height 1: Multiply loop: multiply skip "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" @@ -233,17 +233,17 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "uzp1 z11.d, z11.d, z15.d\n" "tbz %x[flags], #1, 12f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1rw { z1.s }, p5/Z, [x20]\n" + "ld1rw { z17.s }, p5/Z, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1rw { z0.s }, p5/Z, [x20]\n" - "fmin z8.s, p5/M, z8.s, z1.s\n" - "fmin z9.s, p5/M, z9.s, z1.s\n" - "fmin z10.s, p5/M, z10.s, z1.s\n" - "fmin z11.s, p5/M, z11.s, z1.s\n" - "fmax z8.s, p5/M, z8.s, z0.s\n" - "fmax z9.s, p5/M, z9.s, z0.s\n" - "fmax z10.s, p5/M, z10.s, z0.s\n" - "fmax z11.s, p5/M, z11.s, z0.s\n" + "ld1rw { z16.s }, p5/Z, [x20]\n" + "fmin z8.s, p5/M, z8.s, z17.s\n" + "fmin z9.s, p5/M, z9.s, z17.s\n" + "fmin z10.s, p5/M, z10.s, z17.s\n" + "fmin z11.s, p5/M, z11.s, z17.s\n" + "fmax z8.s, p5/M, z8.s, z16.s\n" + "fmax z9.s, p5/M, z9.s, z16.s\n" + "fmax z10.s, p5/M, z10.s, z16.s\n" + "fmax z11.s, p5/M, z11.s, z16.s\n" "12:" // Height 1: No activation "st1w { z8.s }, p4, [x9]\n" "st1w { z9.s }, p3, [x9, #1, MUL VL]\n" @@ -287,21 +287,21 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "16:" // Height 2: no bias "tbz %x[flags], #0, 17f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x9, x20, LSL #2\n" - "ld1w { z9.s }, p4/Z, [x9]\n" - "ld1w { z10.s }, p3/Z, [x9, #1, MUL VL]\n" - "ld1w { z11.s }, p2/Z, [x9, #2, MUL VL]\n" + "add x20, x9, x20, LSL #2\n" + "ld1w { z19.s }, p4/Z, [x9]\n" + "ld1w { z18.s }, p3/Z, [x9, #1, MUL VL]\n" + "ld1w { z17.s }, p2/Z, [x9, #2, MUL VL]\n" "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n" - "ld1w { z12.s }, p4/Z, [x25]\n" - "zip1 z8.d, z9.d, z12.d\n" - "zip2 z12.d, z9.d, z12.d\n" - "ld1w { z13.s }, p3/Z, [x25, #1, MUL VL]\n" - "ld1w { z14.s }, p2/Z, [x25, #2, MUL VL]\n" - "zip1 z9.d, z10.d, z13.d\n" - "zip2 z13.d, z10.d, z13.d\n" - "ld1w { z15.s }, p1/Z, [x25, #3, MUL VL]\n" - "zip1 z10.d, z11.d, z14.d\n" - "zip2 z14.d, z11.d, z14.d\n" + "ld1w { z12.s }, p4/Z, [x20]\n" + "zip1 z8.d, z19.d, z12.d\n" + "zip2 z12.d, z19.d, z12.d\n" + "ld1w { z13.s }, p3/Z, [x20, #1, MUL VL]\n" + "ld1w { z14.s }, p2/Z, [x20, #2, MUL VL]\n" + "zip1 z9.d, z18.d, z13.d\n" + "zip2 z13.d, z18.d, z13.d\n" + "ld1w { z15.s }, p1/Z, [x20, #3, MUL VL]\n" + "zip1 z10.d, z17.d, z14.d\n" + "zip2 z14.d, z17.d, z14.d\n" "zip1 z11.d, z16.d, z15.d\n" "zip2 z15.d, z16.d, z15.d\n" "b 18f\n" @@ -319,12 +319,12 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "19:" // Height 2: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w27, [x20, x28, 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, 20f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" - "ldr x25, [x21, #0x8]\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" + "ldr x25, [x20, #0x8]\n" "cbnz x28, 21f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #2\n" @@ -332,67 +332,67 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "b 21f\n" "20:" // Height 2: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20, LSL #2\n" + "add x25, x26, x21, LSL #2\n" "21:" // Height 2: input setup done "cmp x27, #0x4\n" "ble 23f\n" "22:" // Height 2: Multiply loop: Main loop head "whilelt p0.s, XZR, x27\n" - "ld1rqw { z0.s }, p0/Z, [x26]\n" - "ld1rqw { z1.s }, p0/Z, [x25]\n" - ".inst 0x658ab400 // bfcvt z0.h, p5/M, z0.s\n" - ".inst 0x658ab421 // bfcvt z1.h, p5/M, z1.s\n" - "uzp1 z0.h, z0.h, z0.h\n" - "ld1h { z6.h }, p5/Z, [x10]\n" - "ld1h { z7.h }, p5/Z, [x10, #1, MUL VL]\n" - "uzp1 z1.h, z1.h, z1.h\n" - "trn1 z0.d, z0.d, z1.d\n" - ".inst 0x6466e408 // bfmmla z8.s, z0.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x10, #2, MUL VL]\n" - ".inst 0x6467e40c // bfmmla z12.s, z0.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #3, MUL VL]\n" - ".inst 0x6466e409 // bfmmla z9.s, z0.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x10, #4, MUL VL]\n" - ".inst 0x6467e40d // bfmmla z13.s, z0.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #5, MUL VL]\n" - ".inst 0x6466e40a // bfmmla z10.s, z0.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x10, #6, MUL VL]\n" - ".inst 0x6467e40e // bfmmla z14.s, z0.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #7, MUL VL]\n" + "ld1rqw { z19.s }, p0/Z, [x26]\n" + "ld1rqw { z18.s }, p0/Z, [x25]\n" + ".inst 0x658ab673 // bfcvt z19.h, p5/M, z19.s\n" + ".inst 0x658ab652 // bfcvt z18.h, p5/M, z18.s\n" + "uzp1 z19.h, z19.h, z19.h\n" + "ld1h { z17.h }, p5/Z, [x10]\n" + "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n" + "uzp1 z18.h, z18.h, z18.h\n" + "trn1 z19.d, z19.d, z18.d\n" + ".inst 0x6471e668 // bfmmla z8.s, z19.h, z17.h\n" + "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n" + ".inst 0x6470e66c // bfmmla z12.s, z19.h, z16.h\n" + "ld1h { z16.h }, p5/Z, [x10, #3, MUL VL]\n" + ".inst 0x6471e669 // bfmmla z9.s, z19.h, z17.h\n" + "ld1h { z17.h }, p5/Z, [x10, #4, MUL VL]\n" + ".inst 0x6470e66d // bfmmla z13.s, z19.h, z16.h\n" + "ld1h { z16.h }, p5/Z, [x10, #5, MUL VL]\n" + ".inst 0x6471e66a // bfmmla z10.s, z19.h, z17.h\n" + "ld1h { z17.h }, p5/Z, [x10, #6, MUL VL]\n" + ".inst 0x6470e66e // bfmmla z14.s, z19.h, z16.h\n" + "ld1h { z16.h }, p5/Z, [x10, #7, MUL VL]\n" "sub x27, x27, #0x4\n" "cmp x27, #0x4\n" - ".inst 0x6466e40b // bfmmla z11.s, z0.h, z6.h\n" - ".inst 0x6467e40f // bfmmla z15.s, z0.h, z7.h\n" + ".inst 0x6471e66b // bfmmla z11.s, z19.h, z17.h\n" + ".inst 0x6470e66f // bfmmla z15.s, z19.h, z16.h\n" "add x26, x26, #0x10\n" "add x25, x25, #0x10\n" "addvl x10, x10, #8\n" "bgt 22b\n" "23:" // Height 2: Multiply loop: Single iteration only "whilelt p0.s, XZR, x27\n" - "ld1rqw { z0.s }, p0/Z, [x26]\n" - "ld1rqw { z1.s }, p0/Z, [x25]\n" - ".inst 0x658ab400 // bfcvt z0.h, p5/M, z0.s\n" - ".inst 0x658ab421 // bfcvt z1.h, p5/M, z1.s\n" - "uzp1 z0.h, z0.h, z0.h\n" - "ld1h { z6.h }, p5/Z, [x10]\n" - "ld1h { z7.h }, p5/Z, [x10, #1, MUL VL]\n" - "uzp1 z1.h, z1.h, z1.h\n" - "trn1 z0.d, z0.d, z1.d\n" - ".inst 0x6466e408 // bfmmla z8.s, z0.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x10, #2, MUL VL]\n" - ".inst 0x6467e40c // bfmmla z12.s, z0.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #3, MUL VL]\n" - ".inst 0x6466e409 // bfmmla z9.s, z0.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x10, #4, MUL VL]\n" - ".inst 0x6467e40d // bfmmla z13.s, z0.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #5, MUL VL]\n" - ".inst 0x6466e40a // bfmmla z10.s, z0.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x10, #6, MUL VL]\n" - ".inst 0x6467e40e // bfmmla z14.s, z0.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #7, MUL VL]\n" - ".inst 0x6466e40b // bfmmla z11.s, z0.h, z6.h\n" + "ld1rqw { z19.s }, p0/Z, [x26]\n" + "ld1rqw { z18.s }, p0/Z, [x25]\n" + ".inst 0x658ab673 // bfcvt z19.h, p5/M, z19.s\n" + ".inst 0x658ab652 // bfcvt z18.h, p5/M, z18.s\n" + "uzp1 z19.h, z19.h, z19.h\n" + "ld1h { z17.h }, p5/Z, [x10]\n" + "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n" + "uzp1 z18.h, z18.h, z18.h\n" + "trn1 z19.d, z19.d, z18.d\n" + ".inst 0x6471e668 // bfmmla z8.s, z19.h, z17.h\n" + "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n" + ".inst 0x6470e66c // bfmmla z12.s, z19.h, z16.h\n" + "ld1h { z16.h }, p5/Z, [x10, #3, MUL VL]\n" + ".inst 0x6471e669 // bfmmla z9.s, z19.h, z17.h\n" + "ld1h { z17.h }, p5/Z, [x10, #4, MUL VL]\n" + ".inst 0x6470e66d // bfmmla z13.s, z19.h, z16.h\n" + "ld1h { z16.h }, p5/Z, [x10, #5, MUL VL]\n" + ".inst 0x6471e66a // bfmmla z10.s, z19.h, z17.h\n" + "ld1h { z17.h }, p5/Z, [x10, #6, MUL VL]\n" + ".inst 0x6470e66e // bfmmla z14.s, z19.h, z16.h\n" + "ld1h { z16.h }, p5/Z, [x10, #7, MUL VL]\n" + ".inst 0x6471e66b // bfmmla z11.s, z19.h, z17.h\n" "addvl x10, x10, #8\n" - ".inst 0x6467e40f // bfmmla z15.s, z0.h, z7.h\n" + ".inst 0x6470e66f // bfmmla z15.s, z19.h, z16.h\n" "24:" // Height 2: Multiply loop: multiply skip "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x28, x28, #0x1\n" @@ -410,25 +410,25 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "uzp2 z11.d, z11.d, z15.d\n" "tbz %x[flags], #1, 25f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1rw { z1.s }, p5/Z, [x20]\n" + "ld1rw { z17.s }, p5/Z, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1rw { z0.s }, p5/Z, [x20]\n" - "fmin z6.s, p5/M, z6.s, z1.s\n" - "fmin z12.s, p5/M, z12.s, z1.s\n" - "fmin z13.s, p5/M, z13.s, z1.s\n" - "fmin z14.s, p5/M, z14.s, z1.s\n" - "fmin z8.s, p5/M, z8.s, z1.s\n" - "fmin z9.s, p5/M, z9.s, z1.s\n" - "fmin z10.s, p5/M, z10.s, z1.s\n" - "fmin z11.s, p5/M, z11.s, z1.s\n" - "fmax z6.s, p5/M, z6.s, z0.s\n" - "fmax z12.s, p5/M, z12.s, z0.s\n" - "fmax z13.s, p5/M, z13.s, z0.s\n" - "fmax z14.s, p5/M, z14.s, z0.s\n" - "fmax z8.s, p5/M, z8.s, z0.s\n" - "fmax z9.s, p5/M, z9.s, z0.s\n" - "fmax z10.s, p5/M, z10.s, z0.s\n" - "fmax z11.s, p5/M, z11.s, z0.s\n" + "ld1rw { z16.s }, p5/Z, [x20]\n" + "fmin z6.s, p5/M, z6.s, z17.s\n" + "fmin z12.s, p5/M, z12.s, z17.s\n" + "fmin z13.s, p5/M, z13.s, z17.s\n" + "fmin z14.s, p5/M, z14.s, z17.s\n" + "fmin z8.s, p5/M, z8.s, z17.s\n" + "fmin z9.s, p5/M, z9.s, z17.s\n" + "fmin z10.s, p5/M, z10.s, z17.s\n" + "fmin z11.s, p5/M, z11.s, z17.s\n" + "fmax z6.s, p5/M, z6.s, z16.s\n" + "fmax z12.s, p5/M, z12.s, z16.s\n" + "fmax z13.s, p5/M, z13.s, z16.s\n" + "fmax z14.s, p5/M, z14.s, z16.s\n" + "fmax z8.s, p5/M, z8.s, z16.s\n" + "fmax z9.s, p5/M, z9.s, z16.s\n" + "fmax z10.s, p5/M, z10.s, z16.s\n" + "fmax z11.s, p5/M, z11.s, z16.s\n" "25:" // Height 2: No activation "st1w { z6.s }, p4, [x9]\n" "st1w { z12.s }, p3, [x9, #1, MUL VL]\n" @@ -484,28 +484,28 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "29:" // Height 3: no bias "tbz %x[flags], #0, 30f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x9, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" - "ld1w { z9.s }, p4/Z, [x9]\n" - "ld1w { z10.s }, p3/Z, [x9, #1, MUL VL]\n" - "ld1w { z11.s }, p2/Z, [x9, #2, MUL VL]\n" + "add x21, x9, x20, LSL #2\n" + "add x20, x21, x20, LSL #2\n" + "ld1w { z19.s }, p4/Z, [x9]\n" + "ld1w { z17.s }, p3/Z, [x9, #1, MUL VL]\n" + "ld1w { z18.s }, p2/Z, [x9, #2, MUL VL]\n" "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n" - "ld1w { z12.s }, p4/Z, [x25]\n" - "zip1 z8.d, z9.d, z12.d\n" - "zip2 z12.d, z9.d, z12.d\n" - "ld1w { z13.s }, p3/Z, [x25, #1, MUL VL]\n" - "ld1w { z14.s }, p2/Z, [x25, #2, MUL VL]\n" - "zip1 z9.d, z10.d, z13.d\n" - "zip2 z13.d, z10.d, z13.d\n" - "ld1w { z15.s }, p1/Z, [x25, #3, MUL VL]\n" - "ld1w { z17.s }, p4/Z, [x24]\n" - "zip1 z10.d, z11.d, z14.d\n" - "zip2 z14.d, z11.d, z14.d\n" - "ld1w { z18.s }, p3/Z, [x24, #1, MUL VL]\n" - "ld1w { z19.s }, p2/Z, [x24, #2, MUL VL]\n" + "ld1w { z12.s }, p4/Z, [x21]\n" + "zip1 z8.d, z19.d, z12.d\n" + "zip2 z12.d, z19.d, z12.d\n" + "ld1w { z13.s }, p3/Z, [x21, #1, MUL VL]\n" + "ld1w { z14.s }, p2/Z, [x21, #2, MUL VL]\n" + "zip1 z9.d, z17.d, z13.d\n" + "zip2 z13.d, z17.d, z13.d\n" + "ld1w { z15.s }, p1/Z, [x21, #3, MUL VL]\n" + "ld1w { z17.s }, p4/Z, [x20]\n" + "zip1 z10.d, z18.d, z14.d\n" + "zip2 z14.d, z18.d, z14.d\n" + "ld1w { z18.s }, p3/Z, [x20, #1, MUL VL]\n" + "ld1w { z19.s }, p2/Z, [x20, #2, MUL VL]\n" "zip1 z11.d, z16.d, z15.d\n" "zip2 z15.d, z16.d, z15.d\n" - "ld1w { z24.s }, p1/Z, [x24, #3, MUL VL]\n" + "ld1w { z24.s }, p1/Z, [x20, #3, MUL VL]\n" "zip1 z16.d, z17.d, z20.d\n" "zip2 z20.d, z17.d, z20.d\n" "zip1 z17.d, z18.d, z21.d\n" @@ -537,13 +537,13 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "32:" // Height 3: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w27, [x20, x28, 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, 33f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" - "ldr x25, [x21, #0x8]\n" - "ldr x24, [x21, #0x10]\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" + "ldr x25, [x20, #0x8]\n" + "ldr x24, [x20, #0x10]\n" "cbnz x28, 34f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #2\n" @@ -552,91 +552,91 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "b 34f\n" "33:" // Height 3: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" + "add x25, x26, x21, LSL #2\n" + "add x24, x25, x21, LSL #2\n" "34:" // Height 3: input setup done "cmp x27, #0x4\n" "ble 36f\n" "35:" // Height 3: Multiply loop: Main loop head "whilelt p0.s, XZR, x27\n" - "ld1rqw { z0.s }, p0/Z, [x26]\n" - "ld1rqw { z1.s }, p0/Z, [x25]\n" - ".inst 0x658ab400 // bfcvt z0.h, p5/M, z0.s\n" - "ld1rqw { z2.s }, p0/Z, [x24]\n" - ".inst 0x658ab421 // bfcvt z1.h, p5/M, z1.s\n" - "uzp1 z0.h, z0.h, z0.h\n" - "ld1h { z6.h }, p5/Z, [x10]\n" - "uzp1 z1.h, z1.h, z1.h\n" - ".inst 0x658ab442 // bfcvt z2.h, p5/M, z2.s\n" - "ld1h { z7.h }, p5/Z, [x10, #1, MUL VL]\n" + "ld1rqw { z28.s }, p0/Z, [x26]\n" + "ld1rqw { z27.s }, p0/Z, [x25]\n" + ".inst 0x658ab79c // bfcvt z28.h, p5/M, z28.s\n" + "ld1rqw { z26.s }, p0/Z, [x24]\n" + ".inst 0x658ab77b // bfcvt z27.h, p5/M, z27.s\n" + "uzp1 z28.h, z28.h, z28.h\n" + "ld1h { z25.h }, p5/Z, [x10]\n" + "uzp1 z27.h, z27.h, z27.h\n" + ".inst 0x658ab75a // bfcvt z26.h, p5/M, z26.s\n" + "ld1h { z24.h }, p5/Z, [x10, #1, MUL VL]\n" "sub x27, x27, #0x4\n" - "trn1 z0.d, z0.d, z1.d\n" - "uzp1 z2.h, z2.h, z2.h\n" - ".inst 0x6466e408 // bfmmla z8.s, z0.h, z6.h\n" + "trn1 z28.d, z28.d, z27.d\n" + "uzp1 z26.h, z26.h, z26.h\n" + ".inst 0x6479e788 // bfmmla z8.s, z28.h, z25.h\n" "cmp x27, #0x4\n" - ".inst 0x6466e450 // bfmmla z16.s, z2.h, z6.h\n" - ".inst 0x6467e40c // bfmmla z12.s, z0.h, z7.h\n" - "ld1h { z6.h }, p5/Z, [x10, #2, MUL VL]\n" + ".inst 0x6479e750 // bfmmla z16.s, z26.h, z25.h\n" + ".inst 0x6478e78c // bfmmla z12.s, z28.h, z24.h\n" + "ld1h { z25.h }, p5/Z, [x10, #2, MUL VL]\n" "add x26, x26, #0x10\n" - ".inst 0x6467e454 // bfmmla z20.s, z2.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #3, MUL VL]\n" - ".inst 0x6466e409 // bfmmla z9.s, z0.h, z6.h\n" + ".inst 0x6478e754 // bfmmla z20.s, z26.h, z24.h\n" + "ld1h { z24.h }, p5/Z, [x10, #3, MUL VL]\n" + ".inst 0x6479e789 // bfmmla z9.s, z28.h, z25.h\n" "add x25, x25, #0x10\n" - ".inst 0x6466e451 // bfmmla z17.s, z2.h, z6.h\n" - ".inst 0x6467e40d // bfmmla z13.s, z0.h, z7.h\n" - "ld1h { z6.h }, p5/Z, [x10, #4, MUL VL]\n" + ".inst 0x6479e751 // bfmmla z17.s, z26.h, z25.h\n" + ".inst 0x6478e78d // bfmmla z13.s, z28.h, z24.h\n" + "ld1h { z25.h }, p5/Z, [x10, #4, MUL VL]\n" "add x24, x24, #0x10\n" - ".inst 0x6467e455 // bfmmla z21.s, z2.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #5, MUL VL]\n" - ".inst 0x6466e40a // bfmmla z10.s, z0.h, z6.h\n" - ".inst 0x6466e452 // bfmmla z18.s, z2.h, z6.h\n" - ".inst 0x6467e40e // bfmmla z14.s, z0.h, z7.h\n" - "ld1h { z6.h }, p5/Z, [x10, #6, MUL VL]\n" - ".inst 0x6467e456 // bfmmla z22.s, z2.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #7, MUL VL]\n" - ".inst 0x6466e40b // bfmmla z11.s, z0.h, z6.h\n" + ".inst 0x6478e755 // bfmmla z21.s, z26.h, z24.h\n" + "ld1h { z24.h }, p5/Z, [x10, #5, MUL VL]\n" + ".inst 0x6479e78a // bfmmla z10.s, z28.h, z25.h\n" + ".inst 0x6479e752 // bfmmla z18.s, z26.h, z25.h\n" + ".inst 0x6478e78e // bfmmla z14.s, z28.h, z24.h\n" + "ld1h { z25.h }, p5/Z, [x10, #6, MUL VL]\n" + ".inst 0x6478e756 // bfmmla z22.s, z26.h, z24.h\n" + "ld1h { z24.h }, p5/Z, [x10, #7, MUL VL]\n" + ".inst 0x6479e78b // bfmmla z11.s, z28.h, z25.h\n" "addvl x10, x10, #8\n" - ".inst 0x6466e453 // bfmmla z19.s, z2.h, z6.h\n" - ".inst 0x6467e40f // bfmmla z15.s, z0.h, z7.h\n" - ".inst 0x6467e457 // bfmmla z23.s, z2.h, z7.h\n" + ".inst 0x6479e753 // bfmmla z19.s, z26.h, z25.h\n" + ".inst 0x6478e78f // bfmmla z15.s, z28.h, z24.h\n" + ".inst 0x6478e757 // bfmmla z23.s, z26.h, z24.h\n" "bgt 35b\n" "36:" // Height 3: Multiply loop: Single iteration only "whilelt p0.s, XZR, x27\n" - "ld1rqw { z0.s }, p0/Z, [x26]\n" - "ld1rqw { z1.s }, p0/Z, [x25]\n" - ".inst 0x658ab400 // bfcvt z0.h, p5/M, z0.s\n" - "ld1rqw { z2.s }, p0/Z, [x24]\n" - ".inst 0x658ab421 // bfcvt z1.h, p5/M, z1.s\n" - "uzp1 z0.h, z0.h, z0.h\n" - "ld1h { z6.h }, p5/Z, [x10]\n" - "uzp1 z1.h, z1.h, z1.h\n" - ".inst 0x658ab442 // bfcvt z2.h, p5/M, z2.s\n" - "ld1h { z7.h }, p5/Z, [x10, #1, MUL VL]\n" - "trn1 z0.d, z0.d, z1.d\n" - "uzp1 z2.h, z2.h, z2.h\n" - ".inst 0x6466e408 // bfmmla z8.s, z0.h, z6.h\n" - ".inst 0x6466e450 // bfmmla z16.s, z2.h, z6.h\n" - ".inst 0x6467e40c // bfmmla z12.s, z0.h, z7.h\n" - "ld1h { z6.h }, p5/Z, [x10, #2, MUL VL]\n" - ".inst 0x6467e454 // bfmmla z20.s, z2.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #3, MUL VL]\n" - ".inst 0x6466e409 // bfmmla z9.s, z0.h, z6.h\n" - ".inst 0x6466e451 // bfmmla z17.s, z2.h, z6.h\n" - ".inst 0x6467e40d // bfmmla z13.s, z0.h, z7.h\n" - "ld1h { z6.h }, p5/Z, [x10, #4, MUL VL]\n" - ".inst 0x6467e455 // bfmmla z21.s, z2.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #5, MUL VL]\n" - ".inst 0x6466e40a // bfmmla z10.s, z0.h, z6.h\n" - ".inst 0x6466e452 // bfmmla z18.s, z2.h, z6.h\n" - ".inst 0x6467e40e // bfmmla z14.s, z0.h, z7.h\n" - "ld1h { z6.h }, p5/Z, [x10, #6, MUL VL]\n" - ".inst 0x6467e456 // bfmmla z22.s, z2.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #7, MUL VL]\n" - ".inst 0x6466e40b // bfmmla z11.s, z0.h, z6.h\n" + "ld1rqw { z28.s }, p0/Z, [x26]\n" + "ld1rqw { z27.s }, p0/Z, [x25]\n" + ".inst 0x658ab79c // bfcvt z28.h, p5/M, z28.s\n" + "ld1rqw { z26.s }, p0/Z, [x24]\n" + ".inst 0x658ab77b // bfcvt z27.h, p5/M, z27.s\n" + "uzp1 z28.h, z28.h, z28.h\n" + "ld1h { z25.h }, p5/Z, [x10]\n" + "uzp1 z27.h, z27.h, z27.h\n" + ".inst 0x658ab75a // bfcvt z26.h, p5/M, z26.s\n" + "ld1h { z24.h }, p5/Z, [x10, #1, MUL VL]\n" + "trn1 z28.d, z28.d, z27.d\n" + "uzp1 z26.h, z26.h, z26.h\n" + ".inst 0x6479e788 // bfmmla z8.s, z28.h, z25.h\n" + ".inst 0x6479e750 // bfmmla z16.s, z26.h, z25.h\n" + ".inst 0x6478e78c // bfmmla z12.s, z28.h, z24.h\n" + "ld1h { z25.h }, p5/Z, [x10, #2, MUL VL]\n" + ".inst 0x6478e754 // bfmmla z20.s, z26.h, z24.h\n" + "ld1h { z24.h }, p5/Z, [x10, #3, MUL VL]\n" + ".inst 0x6479e789 // bfmmla z9.s, z28.h, z25.h\n" + ".inst 0x6479e751 // bfmmla z17.s, z26.h, z25.h\n" + ".inst 0x6478e78d // bfmmla z13.s, z28.h, z24.h\n" + "ld1h { z25.h }, p5/Z, [x10, #4, MUL VL]\n" + ".inst 0x6478e755 // bfmmla z21.s, z26.h, z24.h\n" + "ld1h { z24.h }, p5/Z, [x10, #5, MUL VL]\n" + ".inst 0x6479e78a // bfmmla z10.s, z28.h, z25.h\n" + ".inst 0x6479e752 // bfmmla z18.s, z26.h, z25.h\n" + ".inst 0x6478e78e // bfmmla z14.s, z28.h, z24.h\n" + "ld1h { z25.h }, p5/Z, [x10, #6, MUL VL]\n" + ".inst 0x6478e756 // bfmmla z22.s, z26.h, z24.h\n" + "ld1h { z24.h }, p5/Z, [x10, #7, MUL VL]\n" + ".inst 0x6479e78b // bfmmla z11.s, z28.h, z25.h\n" "addvl x10, x10, #8\n" - ".inst 0x6466e453 // bfmmla z19.s, z2.h, z6.h\n" - ".inst 0x6467e40f // bfmmla z15.s, z0.h, z7.h\n" - ".inst 0x6467e457 // bfmmla z23.s, z2.h, z7.h\n" + ".inst 0x6479e753 // bfmmla z19.s, z26.h, z25.h\n" + ".inst 0x6478e78f // bfmmla z15.s, z28.h, z24.h\n" + ".inst 0x6478e757 // bfmmla z23.s, z26.h, z24.h\n" "37:" // Height 3: Multiply loop: multiply skip "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x28, x28, #0x1\n" @@ -659,33 +659,33 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "uzp1 z19.d, z19.d, z23.d\n" "tbz %x[flags], #1, 38f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1rw { z1.s }, p5/Z, [x20]\n" + "ld1rw { z25.s }, p5/Z, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1rw { z0.s }, p5/Z, [x20]\n" - "fmin z6.s, p5/M, z6.s, z1.s\n" - "fmin z12.s, p5/M, z12.s, z1.s\n" - "fmin z13.s, p5/M, z13.s, z1.s\n" - "fmin z14.s, p5/M, z14.s, z1.s\n" - "fmin z8.s, p5/M, z8.s, z1.s\n" - "fmin z9.s, p5/M, z9.s, z1.s\n" - "fmin z10.s, p5/M, z10.s, z1.s\n" - "fmin z11.s, p5/M, z11.s, z1.s\n" - "fmin z16.s, p5/M, z16.s, z1.s\n" - "fmin z17.s, p5/M, z17.s, z1.s\n" - "fmin z18.s, p5/M, z18.s, z1.s\n" - "fmin z19.s, p5/M, z19.s, z1.s\n" - "fmax z6.s, p5/M, z6.s, z0.s\n" - "fmax z12.s, p5/M, z12.s, z0.s\n" - "fmax z13.s, p5/M, z13.s, z0.s\n" - "fmax z14.s, p5/M, z14.s, z0.s\n" - "fmax z8.s, p5/M, z8.s, z0.s\n" - "fmax z9.s, p5/M, z9.s, z0.s\n" - "fmax z10.s, p5/M, z10.s, z0.s\n" - "fmax z11.s, p5/M, z11.s, z0.s\n" - "fmax z16.s, p5/M, z16.s, z0.s\n" - "fmax z17.s, p5/M, z17.s, z0.s\n" - "fmax z18.s, p5/M, z18.s, z0.s\n" - "fmax z19.s, p5/M, z19.s, z0.s\n" + "ld1rw { z24.s }, p5/Z, [x20]\n" + "fmin z6.s, p5/M, z6.s, z25.s\n" + "fmin z12.s, p5/M, z12.s, z25.s\n" + "fmin z13.s, p5/M, z13.s, z25.s\n" + "fmin z14.s, p5/M, z14.s, z25.s\n" + "fmin z8.s, p5/M, z8.s, z25.s\n" + "fmin z9.s, p5/M, z9.s, z25.s\n" + "fmin z10.s, p5/M, z10.s, z25.s\n" + "fmin z11.s, p5/M, z11.s, z25.s\n" + "fmin z16.s, p5/M, z16.s, z25.s\n" + "fmin z17.s, p5/M, z17.s, z25.s\n" + "fmin z18.s, p5/M, z18.s, z25.s\n" + "fmin z19.s, p5/M, z19.s, z25.s\n" + "fmax z6.s, p5/M, z6.s, z24.s\n" + "fmax z12.s, p5/M, z12.s, z24.s\n" + "fmax z13.s, p5/M, z13.s, z24.s\n" + "fmax z14.s, p5/M, z14.s, z24.s\n" + "fmax z8.s, p5/M, z8.s, z24.s\n" + "fmax z9.s, p5/M, z9.s, z24.s\n" + "fmax z10.s, p5/M, z10.s, z24.s\n" + "fmax z11.s, p5/M, z11.s, z24.s\n" + "fmax z16.s, p5/M, z16.s, z24.s\n" + "fmax z17.s, p5/M, z17.s, z24.s\n" + "fmax z18.s, p5/M, z18.s, z24.s\n" + "fmax z19.s, p5/M, z19.s, z24.s\n" "38:" // Height 3: No activation "st1w { z6.s }, p4, [x9]\n" "st1w { z12.s }, p3, [x9, #1, MUL VL]\n" @@ -745,37 +745,37 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "42:" // Height 4: no bias "tbz %x[flags], #0, 43f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x9, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" - "ld1w { z9.s }, p4/Z, [x9]\n" - "add x23, x24, x20, LSL #2\n" - "ld1w { z10.s }, p3/Z, [x9, #1, MUL VL]\n" - "ld1w { z11.s }, p2/Z, [x9, #2, MUL VL]\n" + "add x22, x9, x20, LSL #2\n" + "add x21, x22, x20, LSL #2\n" + "ld1w { z19.s }, p4/Z, [x9]\n" + "add x20, x21, x20, LSL #2\n" + "ld1w { z17.s }, p3/Z, [x9, #1, MUL VL]\n" + "ld1w { z18.s }, p2/Z, [x9, #2, MUL VL]\n" "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n" - "ld1w { z12.s }, p4/Z, [x25]\n" - "zip1 z8.d, z9.d, z12.d\n" - "zip2 z12.d, z9.d, z12.d\n" - "ld1w { z13.s }, p3/Z, [x25, #1, MUL VL]\n" - "ld1w { z14.s }, p2/Z, [x25, #2, MUL VL]\n" - "zip1 z9.d, z10.d, z13.d\n" - "zip2 z13.d, z10.d, z13.d\n" - "ld1w { z15.s }, p1/Z, [x25, #3, MUL VL]\n" - "ld1w { z17.s }, p4/Z, [x24]\n" - "zip1 z10.d, z11.d, z14.d\n" - "zip2 z14.d, z11.d, z14.d\n" - "ld1w { z18.s }, p3/Z, [x24, #1, MUL VL]\n" - "ld1w { z19.s }, p2/Z, [x24, #2, MUL VL]\n" + "ld1w { z12.s }, p4/Z, [x22]\n" + "zip1 z8.d, z19.d, z12.d\n" + "zip2 z12.d, z19.d, z12.d\n" + "ld1w { z13.s }, p3/Z, [x22, #1, MUL VL]\n" + "ld1w { z14.s }, p2/Z, [x22, #2, MUL VL]\n" + "zip1 z9.d, z17.d, z13.d\n" + "zip2 z13.d, z17.d, z13.d\n" + "ld1w { z15.s }, p1/Z, [x22, #3, MUL VL]\n" + "ld1w { z17.s }, p4/Z, [x21]\n" + "zip1 z10.d, z18.d, z14.d\n" + "zip2 z14.d, z18.d, z14.d\n" + "ld1w { z18.s }, p3/Z, [x21, #1, MUL VL]\n" + "ld1w { z19.s }, p2/Z, [x21, #2, MUL VL]\n" "zip1 z11.d, z16.d, z15.d\n" "zip2 z15.d, z16.d, z15.d\n" - "ld1w { z24.s }, p1/Z, [x24, #3, MUL VL]\n" - "ld1w { z20.s }, p4/Z, [x23]\n" + "ld1w { z24.s }, p1/Z, [x21, #3, MUL VL]\n" + "ld1w { z20.s }, p4/Z, [x20]\n" "zip1 z16.d, z17.d, z20.d\n" "zip2 z20.d, z17.d, z20.d\n" - "ld1w { z21.s }, p3/Z, [x23, #1, MUL VL]\n" - "ld1w { z22.s }, p2/Z, [x23, #2, MUL VL]\n" + "ld1w { z21.s }, p3/Z, [x20, #1, MUL VL]\n" + "ld1w { z22.s }, p2/Z, [x20, #2, MUL VL]\n" "zip1 z17.d, z18.d, z21.d\n" "zip2 z21.d, z18.d, z21.d\n" - "ld1w { z23.s }, p1/Z, [x23, #3, MUL VL]\n" + "ld1w { z23.s }, p1/Z, [x20, #3, MUL VL]\n" "zip1 z18.d, z19.d, z22.d\n" "zip2 z22.d, z19.d, z22.d\n" "zip1 z19.d, z24.d, z23.d\n" @@ -803,14 +803,14 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "45:" // Height 4: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w27, [x20, x28, 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, 46f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" - "ldr x25, [x21, #0x8]\n" - "ldr x24, [x21, #0x10]\n" - "ldr x23, [x21, #0x18]\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" + "ldr x25, [x20, #0x8]\n" + "ldr x24, [x20, #0x10]\n" + "ldr x23, [x20, #0x18]\n" "cbnz x28, 47f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #2\n" @@ -820,101 +820,101 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "b 47f\n" "46:" // Height 4: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" - "add x23, x24, x20, LSL #2\n" + "add x25, x26, x21, LSL #2\n" + "add x24, x25, x21, LSL #2\n" + "add x23, x24, x21, LSL #2\n" "47:" // Height 4: input setup done "cmp x27, #0x4\n" "ble 49f\n" "48:" // Height 4: Multiply loop: Main loop head "whilelt p0.s, XZR, x27\n" - "ld1rqw { z0.s }, p0/Z, [x26]\n" - "ld1rqw { z1.s }, p0/Z, [x25]\n" - ".inst 0x658ab400 // bfcvt z0.h, p5/M, z0.s\n" - "ld1rqw { z2.s }, p0/Z, [x24]\n" - "ld1rqw { z3.s }, p0/Z, [x23]\n" - ".inst 0x658ab421 // bfcvt z1.h, p5/M, z1.s\n" - ".inst 0x658ab442 // bfcvt z2.h, p5/M, z2.s\n" - ".inst 0x658ab463 // bfcvt z3.h, p5/M, z3.s\n" - "uzp1 z0.h, z0.h, z0.h\n" - "ld1h { z6.h }, p5/Z, [x10]\n" - "ld1h { z7.h }, p5/Z, [x10, #1, MUL VL]\n" - "uzp1 z1.h, z1.h, z1.h\n" - "uzp1 z2.h, z2.h, z2.h\n" + "ld1rqw { z29.s }, p0/Z, [x26]\n" + "ld1rqw { z28.s }, p0/Z, [x25]\n" + ".inst 0x658ab7bd // bfcvt z29.h, p5/M, z29.s\n" + "ld1rqw { z27.s }, p0/Z, [x24]\n" + "ld1rqw { z26.s }, p0/Z, [x23]\n" + ".inst 0x658ab79c // bfcvt z28.h, p5/M, z28.s\n" + ".inst 0x658ab77b // bfcvt z27.h, p5/M, z27.s\n" + ".inst 0x658ab75a // bfcvt z26.h, p5/M, z26.s\n" + "uzp1 z29.h, z29.h, z29.h\n" + "ld1h { z25.h }, p5/Z, [x10]\n" + "ld1h { z24.h }, p5/Z, [x10, #1, MUL VL]\n" + "uzp1 z28.h, z28.h, z28.h\n" + "uzp1 z27.h, z27.h, z27.h\n" "sub x27, x27, #0x4\n" "cmp x27, #0x4\n" - "uzp1 z3.h, z3.h, z3.h\n" - "trn1 z0.d, z0.d, z1.d\n" - ".inst 0x6466e408 // bfmmla z8.s, z0.h, z6.h\n" + "uzp1 z26.h, z26.h, z26.h\n" + "trn1 z29.d, z29.d, z28.d\n" + ".inst 0x6479e7a8 // bfmmla z8.s, z29.h, z25.h\n" "add x26, x26, #0x10\n" - "trn1 z2.d, z2.d, z3.d\n" - ".inst 0x6466e450 // bfmmla z16.s, z2.h, z6.h\n" - ".inst 0x6467e40c // bfmmla z12.s, z0.h, z7.h\n" - "ld1h { z6.h }, p5/Z, [x10, #2, MUL VL]\n" - ".inst 0x6467e454 // bfmmla z20.s, z2.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #3, MUL VL]\n" - ".inst 0x6466e409 // bfmmla z9.s, z0.h, z6.h\n" + "trn1 z27.d, z27.d, z26.d\n" + ".inst 0x6479e770 // bfmmla z16.s, z27.h, z25.h\n" + ".inst 0x6478e7ac // bfmmla z12.s, z29.h, z24.h\n" + "ld1h { z25.h }, p5/Z, [x10, #2, MUL VL]\n" + ".inst 0x6478e774 // bfmmla z20.s, z27.h, z24.h\n" + "ld1h { z24.h }, p5/Z, [x10, #3, MUL VL]\n" + ".inst 0x6479e7a9 // bfmmla z9.s, z29.h, z25.h\n" "add x25, x25, #0x10\n" - ".inst 0x6466e451 // bfmmla z17.s, z2.h, z6.h\n" - ".inst 0x6467e40d // bfmmla z13.s, z0.h, z7.h\n" - "ld1h { z6.h }, p5/Z, [x10, #4, MUL VL]\n" + ".inst 0x6479e771 // bfmmla z17.s, z27.h, z25.h\n" + ".inst 0x6478e7ad // bfmmla z13.s, z29.h, z24.h\n" + "ld1h { z25.h }, p5/Z, [x10, #4, MUL VL]\n" "add x24, x24, #0x10\n" - ".inst 0x6467e455 // bfmmla z21.s, z2.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #5, MUL VL]\n" - ".inst 0x6466e40a // bfmmla z10.s, z0.h, z6.h\n" + ".inst 0x6478e775 // bfmmla z21.s, z27.h, z24.h\n" + "ld1h { z24.h }, p5/Z, [x10, #5, MUL VL]\n" + ".inst 0x6479e7aa // bfmmla z10.s, z29.h, z25.h\n" "add x23, x23, #0x10\n" - ".inst 0x6466e452 // bfmmla z18.s, z2.h, z6.h\n" - ".inst 0x6467e40e // bfmmla z14.s, z0.h, z7.h\n" - "ld1h { z6.h }, p5/Z, [x10, #6, MUL VL]\n" - ".inst 0x6467e456 // bfmmla z22.s, z2.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #7, MUL VL]\n" - ".inst 0x6466e40b // bfmmla z11.s, z0.h, z6.h\n" + ".inst 0x6479e772 // bfmmla z18.s, z27.h, z25.h\n" + ".inst 0x6478e7ae // bfmmla z14.s, z29.h, z24.h\n" + "ld1h { z25.h }, p5/Z, [x10, #6, MUL VL]\n" + ".inst 0x6478e776 // bfmmla z22.s, z27.h, z24.h\n" + "ld1h { z24.h }, p5/Z, [x10, #7, MUL VL]\n" + ".inst 0x6479e7ab // bfmmla z11.s, z29.h, z25.h\n" "addvl x10, x10, #8\n" - ".inst 0x6466e453 // bfmmla z19.s, z2.h, z6.h\n" - ".inst 0x6467e40f // bfmmla z15.s, z0.h, z7.h\n" - ".inst 0x6467e457 // bfmmla z23.s, z2.h, z7.h\n" + ".inst 0x6479e773 // bfmmla z19.s, z27.h, z25.h\n" + ".inst 0x6478e7af // bfmmla z15.s, z29.h, z24.h\n" + ".inst 0x6478e777 // bfmmla z23.s, z27.h, z24.h\n" "bgt 48b\n" "49:" // Height 4: Multiply loop: Single iteration only "whilelt p0.s, XZR, x27\n" - "ld1rqw { z0.s }, p0/Z, [x26]\n" - "ld1rqw { z1.s }, p0/Z, [x25]\n" - ".inst 0x658ab400 // bfcvt z0.h, p5/M, z0.s\n" - "ld1rqw { z2.s }, p0/Z, [x24]\n" - "ld1rqw { z3.s }, p0/Z, [x23]\n" - ".inst 0x658ab421 // bfcvt z1.h, p5/M, z1.s\n" - ".inst 0x658ab442 // bfcvt z2.h, p5/M, z2.s\n" - ".inst 0x658ab463 // bfcvt z3.h, p5/M, z3.s\n" - "uzp1 z0.h, z0.h, z0.h\n" - "ld1h { z6.h }, p5/Z, [x10]\n" - "ld1h { z7.h }, p5/Z, [x10, #1, MUL VL]\n" - "uzp1 z1.h, z1.h, z1.h\n" - "uzp1 z2.h, z2.h, z2.h\n" - "uzp1 z3.h, z3.h, z3.h\n" - "trn1 z0.d, z0.d, z1.d\n" - ".inst 0x6466e408 // bfmmla z8.s, z0.h, z6.h\n" - "trn1 z2.d, z2.d, z3.d\n" - ".inst 0x6466e450 // bfmmla z16.s, z2.h, z6.h\n" - ".inst 0x6467e40c // bfmmla z12.s, z0.h, z7.h\n" - "ld1h { z6.h }, p5/Z, [x10, #2, MUL VL]\n" - ".inst 0x6467e454 // bfmmla z20.s, z2.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #3, MUL VL]\n" - ".inst 0x6466e409 // bfmmla z9.s, z0.h, z6.h\n" - ".inst 0x6466e451 // bfmmla z17.s, z2.h, z6.h\n" - ".inst 0x6467e40d // bfmmla z13.s, z0.h, z7.h\n" - "ld1h { z6.h }, p5/Z, [x10, #4, MUL VL]\n" - ".inst 0x6467e455 // bfmmla z21.s, z2.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #5, MUL VL]\n" - ".inst 0x6466e40a // bfmmla z10.s, z0.h, z6.h\n" - ".inst 0x6466e452 // bfmmla z18.s, z2.h, z6.h\n" - ".inst 0x6467e40e // bfmmla z14.s, z0.h, z7.h\n" - "ld1h { z6.h }, p5/Z, [x10, #6, MUL VL]\n" - ".inst 0x6467e456 // bfmmla z22.s, z2.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #7, MUL VL]\n" - ".inst 0x6466e40b // bfmmla z11.s, z0.h, z6.h\n" + "ld1rqw { z29.s }, p0/Z, [x26]\n" + "ld1rqw { z28.s }, p0/Z, [x25]\n" + ".inst 0x658ab7bd // bfcvt z29.h, p5/M, z29.s\n" + "ld1rqw { z27.s }, p0/Z, [x24]\n" + "ld1rqw { z26.s }, p0/Z, [x23]\n" + ".inst 0x658ab79c // bfcvt z28.h, p5/M, z28.s\n" + ".inst 0x658ab77b // bfcvt z27.h, p5/M, z27.s\n" + ".inst 0x658ab75a // bfcvt z26.h, p5/M, z26.s\n" + "uzp1 z29.h, z29.h, z29.h\n" + "ld1h { z25.h }, p5/Z, [x10]\n" + "ld1h { z24.h }, p5/Z, [x10, #1, MUL VL]\n" + "uzp1 z28.h, z28.h, z28.h\n" + "uzp1 z27.h, z27.h, z27.h\n" + "uzp1 z26.h, z26.h, z26.h\n" + "trn1 z29.d, z29.d, z28.d\n" + ".inst 0x6479e7a8 // bfmmla z8.s, z29.h, z25.h\n" + "trn1 z27.d, z27.d, z26.d\n" + ".inst 0x6479e770 // bfmmla z16.s, z27.h, z25.h\n" + ".inst 0x6478e7ac // bfmmla z12.s, z29.h, z24.h\n" + "ld1h { z25.h }, p5/Z, [x10, #2, MUL VL]\n" + ".inst 0x6478e774 // bfmmla z20.s, z27.h, z24.h\n" + "ld1h { z24.h }, p5/Z, [x10, #3, MUL VL]\n" + ".inst 0x6479e7a9 // bfmmla z9.s, z29.h, z25.h\n" + ".inst 0x6479e771 // bfmmla z17.s, z27.h, z25.h\n" + ".inst 0x6478e7ad // bfmmla z13.s, z29.h, z24.h\n" + "ld1h { z25.h }, p5/Z, [x10, #4, MUL VL]\n" + ".inst 0x6478e775 // bfmmla z21.s, z27.h, z24.h\n" + "ld1h { z24.h }, p5/Z, [x10, #5, MUL VL]\n" + ".inst 0x6479e7aa // bfmmla z10.s, z29.h, z25.h\n" + ".inst 0x6479e772 // bfmmla z18.s, z27.h, z25.h\n" + ".inst 0x6478e7ae // bfmmla z14.s, z29.h, z24.h\n" + "ld1h { z25.h }, p5/Z, [x10, #6, MUL VL]\n" + ".inst 0x6478e776 // bfmmla z22.s, z27.h, z24.h\n" + "ld1h { z24.h }, p5/Z, [x10, #7, MUL VL]\n" + ".inst 0x6479e7ab // bfmmla z11.s, z29.h, z25.h\n" "addvl x10, x10, #8\n" - ".inst 0x6466e453 // bfmmla z19.s, z2.h, z6.h\n" - ".inst 0x6467e40f // bfmmla z15.s, z0.h, z7.h\n" - ".inst 0x6467e457 // bfmmla z23.s, z2.h, z7.h\n" + ".inst 0x6479e773 // bfmmla z19.s, z27.h, z25.h\n" + ".inst 0x6478e7af // bfmmla z15.s, z29.h, z24.h\n" + ".inst 0x6478e777 // bfmmla z23.s, z27.h, z24.h\n" "50:" // Height 4: Multiply loop: multiply skip "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x28, x28, #0x1\n" @@ -942,41 +942,41 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "uzp2 z19.d, z19.d, z23.d\n" "tbz %x[flags], #1, 51f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1rw { z1.s }, p5/Z, [x20]\n" + "ld1rw { z24.s }, p5/Z, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1rw { z0.s }, p5/Z, [x20]\n" - "fmin z6.s, p5/M, z6.s, z1.s\n" - "fmin z12.s, p5/M, z12.s, z1.s\n" - "fmin z13.s, p5/M, z13.s, z1.s\n" - "fmin z14.s, p5/M, z14.s, z1.s\n" - "fmin z8.s, p5/M, z8.s, z1.s\n" - "fmin z9.s, p5/M, z9.s, z1.s\n" - "fmin z10.s, p5/M, z10.s, z1.s\n" - "fmin z11.s, p5/M, z11.s, z1.s\n" - "fmin z15.s, p5/M, z15.s, z1.s\n" - "fmin z20.s, p5/M, z20.s, z1.s\n" - "fmin z21.s, p5/M, z21.s, z1.s\n" - "fmin z22.s, p5/M, z22.s, z1.s\n" - "fmin z16.s, p5/M, z16.s, z1.s\n" - "fmin z17.s, p5/M, z17.s, z1.s\n" - "fmin z18.s, p5/M, z18.s, z1.s\n" - "fmin z19.s, p5/M, z19.s, z1.s\n" - "fmax z6.s, p5/M, z6.s, z0.s\n" - "fmax z12.s, p5/M, z12.s, z0.s\n" - "fmax z13.s, p5/M, z13.s, z0.s\n" - "fmax z14.s, p5/M, z14.s, z0.s\n" - "fmax z8.s, p5/M, z8.s, z0.s\n" - "fmax z9.s, p5/M, z9.s, z0.s\n" - "fmax z10.s, p5/M, z10.s, z0.s\n" - "fmax z11.s, p5/M, z11.s, z0.s\n" - "fmax z15.s, p5/M, z15.s, z0.s\n" - "fmax z20.s, p5/M, z20.s, z0.s\n" - "fmax z21.s, p5/M, z21.s, z0.s\n" - "fmax z22.s, p5/M, z22.s, z0.s\n" - "fmax z16.s, p5/M, z16.s, z0.s\n" - "fmax z17.s, p5/M, z17.s, z0.s\n" - "fmax z18.s, p5/M, z18.s, z0.s\n" - "fmax z19.s, p5/M, z19.s, z0.s\n" + "ld1rw { z23.s }, p5/Z, [x20]\n" + "fmin z6.s, p5/M, z6.s, z24.s\n" + "fmin z12.s, p5/M, z12.s, z24.s\n" + "fmin z13.s, p5/M, z13.s, z24.s\n" + "fmin z14.s, p5/M, z14.s, z24.s\n" + "fmin z8.s, p5/M, z8.s, z24.s\n" + "fmin z9.s, p5/M, z9.s, z24.s\n" + "fmin z10.s, p5/M, z10.s, z24.s\n" + "fmin z11.s, p5/M, z11.s, z24.s\n" + "fmin z15.s, p5/M, z15.s, z24.s\n" + "fmin z20.s, p5/M, z20.s, z24.s\n" + "fmin z21.s, p5/M, z21.s, z24.s\n" + "fmin z22.s, p5/M, z22.s, z24.s\n" + "fmin z16.s, p5/M, z16.s, z24.s\n" + "fmin z17.s, p5/M, z17.s, z24.s\n" + "fmin z18.s, p5/M, z18.s, z24.s\n" + "fmin z19.s, p5/M, z19.s, z24.s\n" + "fmax z6.s, p5/M, z6.s, z23.s\n" + "fmax z12.s, p5/M, z12.s, z23.s\n" + "fmax z13.s, p5/M, z13.s, z23.s\n" + "fmax z14.s, p5/M, z14.s, z23.s\n" + "fmax z8.s, p5/M, z8.s, z23.s\n" + "fmax z9.s, p5/M, z9.s, z23.s\n" + "fmax z10.s, p5/M, z10.s, z23.s\n" + "fmax z11.s, p5/M, z11.s, z23.s\n" + "fmax z15.s, p5/M, z15.s, z23.s\n" + "fmax z20.s, p5/M, z20.s, z23.s\n" + "fmax z21.s, p5/M, z21.s, z23.s\n" + "fmax z22.s, p5/M, z22.s, z23.s\n" + "fmax z16.s, p5/M, z16.s, z23.s\n" + "fmax z17.s, p5/M, z17.s, z23.s\n" + "fmax z18.s, p5/M, z18.s, z23.s\n" + "fmax z19.s, p5/M, z19.s, z23.s\n" "51:" // Height 4: No activation "st1w { z6.s }, p4, [x9]\n" "st1w { z12.s }, p3, [x9, #1, MUL VL]\n" @@ -1048,54 +1048,54 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "55:" // Height 5: no bias "tbz %x[flags], #0, 56f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x9, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" - "ld1w { z9.s }, p4/Z, [x9]\n" - "add x23, x24, x20, LSL #2\n" + "add x23, x9, x20, LSL #2\n" "add x22, x23, x20, LSL #2\n" - "ld1w { z10.s }, p3/Z, [x9, #1, MUL VL]\n" - "ld1w { z11.s }, p2/Z, [x9, #2, MUL VL]\n" + "ld1w { z19.s }, p4/Z, [x9]\n" + "add x21, x22, x20, LSL #2\n" + "add x20, x21, x20, LSL #2\n" + "ld1w { z17.s }, p3/Z, [x9, #1, MUL VL]\n" + "ld1w { z18.s }, p2/Z, [x9, #2, MUL VL]\n" "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n" - "ld1w { z12.s }, p4/Z, [x25]\n" - "zip1 z8.d, z9.d, z12.d\n" - "zip2 z12.d, z9.d, z12.d\n" - "ld1w { z13.s }, p3/Z, [x25, #1, MUL VL]\n" - "ld1w { z14.s }, p2/Z, [x25, #2, MUL VL]\n" - "zip1 z9.d, z10.d, z13.d\n" - "zip2 z13.d, z10.d, z13.d\n" - "ld1w { z15.s }, p1/Z, [x25, #3, MUL VL]\n" - "ld1w { z17.s }, p4/Z, [x24]\n" - "zip1 z10.d, z11.d, z14.d\n" - "zip2 z14.d, z11.d, z14.d\n" - "ld1w { z18.s }, p3/Z, [x24, #1, MUL VL]\n" - "ld1w { z19.s }, p2/Z, [x24, #2, MUL VL]\n" + "ld1w { z12.s }, p4/Z, [x23]\n" + "zip1 z8.d, z19.d, z12.d\n" + "zip2 z12.d, z19.d, z12.d\n" + "ld1w { z13.s }, p3/Z, [x23, #1, MUL VL]\n" + "ld1w { z14.s }, p2/Z, [x23, #2, MUL VL]\n" + "zip1 z9.d, z17.d, z13.d\n" + "zip2 z13.d, z17.d, z13.d\n" + "ld1w { z15.s }, p1/Z, [x23, #3, MUL VL]\n" + "ld1w { z17.s }, p4/Z, [x22]\n" + "zip1 z10.d, z18.d, z14.d\n" + "zip2 z14.d, z18.d, z14.d\n" + "ld1w { z18.s }, p3/Z, [x22, #1, MUL VL]\n" + "ld1w { z19.s }, p2/Z, [x22, #2, MUL VL]\n" "zip1 z11.d, z16.d, z15.d\n" "zip2 z15.d, z16.d, z15.d\n" - "ld1w { z24.s }, p1/Z, [x24, #3, MUL VL]\n" - "ld1w { z20.s }, p4/Z, [x23]\n" + "ld1w { z24.s }, p1/Z, [x22, #3, MUL VL]\n" + "ld1w { z20.s }, p4/Z, [x21]\n" "zip1 z16.d, z17.d, z20.d\n" "zip2 z20.d, z17.d, z20.d\n" - "ld1w { z21.s }, p3/Z, [x23, #1, MUL VL]\n" - "ld1w { z22.s }, p2/Z, [x23, #2, MUL VL]\n" + "ld1w { z21.s }, p3/Z, [x21, #1, MUL VL]\n" + "ld1w { z22.s }, p2/Z, [x21, #2, MUL VL]\n" "zip1 z17.d, z18.d, z21.d\n" "zip2 z21.d, z18.d, z21.d\n" - "ld1w { z23.s }, p1/Z, [x23, #3, MUL VL]\n" - "ld1w { z25.s }, p4/Z, [x22]\n" + "ld1w { z23.s }, p1/Z, [x21, #3, MUL VL]\n" + "ld1w { z25.s }, p4/Z, [x20]\n" "zip1 z18.d, z19.d, z22.d\n" "zip2 z22.d, z19.d, z22.d\n" - "ld1w { z26.s }, p3/Z, [x22, #1, MUL VL]\n" - "ld1w { z27.s }, p2/Z, [x22, #2, MUL VL]\n" + "ld1w { z26.s }, p3/Z, [x20, #1, MUL VL]\n" + "ld1w { z27.s }, p2/Z, [x20, #2, MUL VL]\n" "zip1 z19.d, z24.d, z23.d\n" "zip2 z23.d, z24.d, z23.d\n" - "ld1w { z6.s }, p1/Z, [x22, #3, MUL VL]\n" + "ld1w { z0.s }, p1/Z, [x20, #3, MUL VL]\n" "zip1 z24.d, z25.d, z28.d\n" "zip2 z28.d, z25.d, z28.d\n" "zip1 z25.d, z26.d, z29.d\n" "zip2 z29.d, z26.d, z29.d\n" "zip1 z26.d, z27.d, z30.d\n" "zip2 z30.d, z27.d, z30.d\n" - "zip1 z27.d, z6.d, z31.d\n" - "zip2 z31.d, z6.d, z31.d\n" + "zip1 z27.d, z0.d, z31.d\n" + "zip2 z31.d, z0.d, z31.d\n" "b 57f\n" "56:" // Height 5: no accumulate "mov z8.b, #0x0\n" @@ -1127,15 +1127,15 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "58:" // Height 5: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w27, [x20, x28, 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, 59f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" - "ldr x25, [x21, #0x8]\n" - "ldr x24, [x21, #0x10]\n" - "ldr x23, [x21, #0x18]\n" - "ldr x22, [x21, #0x20]\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" + "ldr x25, [x20, #0x8]\n" + "ldr x24, [x20, #0x10]\n" + "ldr x23, [x20, #0x18]\n" + "ldr x22, [x20, #0x20]\n" "cbnz x28, 60f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #2\n" @@ -1146,125 +1146,125 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "b 60f\n" "59:" // Height 5: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" - "add x23, x24, x20, LSL #2\n" - "add x22, x23, x20, LSL #2\n" + "add x25, x26, x21, LSL #2\n" + "add x24, x25, x21, LSL #2\n" + "add x23, x24, x21, LSL #2\n" + "add x22, x23, x21, LSL #2\n" "60:" // Height 5: input setup done "cmp x27, #0x4\n" "ble 62f\n" "61:" // Height 5: Multiply loop: Main loop head "whilelt p0.s, XZR, x27\n" - "ld1rqw { z0.s }, p0/Z, [x26]\n" - "ld1rqw { z1.s }, p0/Z, [x25]\n" - ".inst 0x658ab400 // bfcvt z0.h, p5/M, z0.s\n" - "ld1rqw { z2.s }, p0/Z, [x24]\n" + "ld1rqw { z6.s }, p0/Z, [x26]\n" + "ld1rqw { z5.s }, p0/Z, [x25]\n" + ".inst 0x658ab4c6 // bfcvt z6.h, p5/M, z6.s\n" + "ld1rqw { z4.s }, p0/Z, [x24]\n" "ld1rqw { z3.s }, p0/Z, [x23]\n" - ".inst 0x658ab421 // bfcvt z1.h, p5/M, z1.s\n" - ".inst 0x658ab442 // bfcvt z2.h, p5/M, z2.s\n" - "ld1rqw { z4.s }, p0/Z, [x22]\n" + ".inst 0x658ab4a5 // bfcvt z5.h, p5/M, z5.s\n" + ".inst 0x658ab484 // bfcvt z4.h, p5/M, z4.s\n" + "ld1rqw { z2.s }, p0/Z, [x22]\n" ".inst 0x658ab463 // bfcvt z3.h, p5/M, z3.s\n" - "uzp1 z0.h, z0.h, z0.h\n" - "ld1h { z6.h }, p5/Z, [x10]\n" - "uzp1 z1.h, z1.h, z1.h\n" - "uzp1 z2.h, z2.h, z2.h\n" - "ld1h { z7.h }, p5/Z, [x10, #1, MUL VL]\n" + "uzp1 z6.h, z6.h, z6.h\n" + "ld1h { z1.h }, p5/Z, [x10]\n" + "uzp1 z5.h, z5.h, z5.h\n" + "uzp1 z4.h, z4.h, z4.h\n" + "ld1h { z0.h }, p5/Z, [x10, #1, MUL VL]\n" "sub x27, x27, #0x4\n" "uzp1 z3.h, z3.h, z3.h\n" - ".inst 0x658ab484 // bfcvt z4.h, p5/M, z4.s\n" + ".inst 0x658ab442 // bfcvt z2.h, p5/M, z2.s\n" "cmp x27, #0x4\n" "add x26, x26, #0x10\n" - "trn1 z0.d, z0.d, z1.d\n" - "trn1 z2.d, z2.d, z3.d\n" - ".inst 0x6466e408 // bfmmla z8.s, z0.h, z6.h\n" + "trn1 z6.d, z6.d, z5.d\n" + "trn1 z4.d, z4.d, z3.d\n" + ".inst 0x6461e4c8 // bfmmla z8.s, z6.h, z1.h\n" "add x25, x25, #0x10\n" - "uzp1 z4.h, z4.h, z4.h\n" - ".inst 0x6466e450 // bfmmla z16.s, z2.h, z6.h\n" - ".inst 0x6466e498 // bfmmla z24.s, z4.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x10, #2, MUL VL]\n" - ".inst 0x6467e40c // bfmmla z12.s, z0.h, z7.h\n" - ".inst 0x6467e454 // bfmmla z20.s, z2.h, z7.h\n" + "uzp1 z2.h, z2.h, z2.h\n" + ".inst 0x6461e490 // bfmmla z16.s, z4.h, z1.h\n" + ".inst 0x6461e458 // bfmmla z24.s, z2.h, z1.h\n" + "ld1h { z1.h }, p5/Z, [x10, #2, MUL VL]\n" + ".inst 0x6460e4cc // bfmmla z12.s, z6.h, z0.h\n" + ".inst 0x6460e494 // bfmmla z20.s, z4.h, z0.h\n" "add x24, x24, #0x10\n" "add x23, x23, #0x10\n" - ".inst 0x6467e49c // bfmmla z28.s, z4.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #3, MUL VL]\n" - ".inst 0x6466e409 // bfmmla z9.s, z0.h, z6.h\n" + ".inst 0x6460e45c // bfmmla z28.s, z2.h, z0.h\n" + "ld1h { z0.h }, p5/Z, [x10, #3, MUL VL]\n" + ".inst 0x6461e4c9 // bfmmla z9.s, z6.h, z1.h\n" "add x22, x22, #0x10\n" - ".inst 0x6466e451 // bfmmla z17.s, z2.h, z6.h\n" - ".inst 0x6466e499 // bfmmla z25.s, z4.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x10, #4, MUL VL]\n" - ".inst 0x6467e40d // bfmmla z13.s, z0.h, z7.h\n" - ".inst 0x6467e455 // bfmmla z21.s, z2.h, z7.h\n" - ".inst 0x6467e49d // bfmmla z29.s, z4.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #5, MUL VL]\n" - ".inst 0x6466e40a // bfmmla z10.s, z0.h, z6.h\n" - ".inst 0x6466e452 // bfmmla z18.s, z2.h, z6.h\n" - ".inst 0x6466e49a // bfmmla z26.s, z4.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x10, #6, MUL VL]\n" - ".inst 0x6467e40e // bfmmla z14.s, z0.h, z7.h\n" - ".inst 0x6467e456 // bfmmla z22.s, z2.h, z7.h\n" - ".inst 0x6467e49e // bfmmla z30.s, z4.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #7, MUL VL]\n" - ".inst 0x6466e40b // bfmmla z11.s, z0.h, z6.h\n" + ".inst 0x6461e491 // bfmmla z17.s, z4.h, z1.h\n" + ".inst 0x6461e459 // bfmmla z25.s, z2.h, z1.h\n" + "ld1h { z1.h }, p5/Z, [x10, #4, MUL VL]\n" + ".inst 0x6460e4cd // bfmmla z13.s, z6.h, z0.h\n" + ".inst 0x6460e495 // bfmmla z21.s, z4.h, z0.h\n" + ".inst 0x6460e45d // bfmmla z29.s, z2.h, z0.h\n" + "ld1h { z0.h }, p5/Z, [x10, #5, MUL VL]\n" + ".inst 0x6461e4ca // bfmmla z10.s, z6.h, z1.h\n" + ".inst 0x6461e492 // bfmmla z18.s, z4.h, z1.h\n" + ".inst 0x6461e45a // bfmmla z26.s, z2.h, z1.h\n" + "ld1h { z1.h }, p5/Z, [x10, #6, MUL VL]\n" + ".inst 0x6460e4ce // bfmmla z14.s, z6.h, z0.h\n" + ".inst 0x6460e496 // bfmmla z22.s, z4.h, z0.h\n" + ".inst 0x6460e45e // bfmmla z30.s, z2.h, z0.h\n" + "ld1h { z0.h }, p5/Z, [x10, #7, MUL VL]\n" + ".inst 0x6461e4cb // bfmmla z11.s, z6.h, z1.h\n" "addvl x10, x10, #8\n" - ".inst 0x6466e453 // bfmmla z19.s, z2.h, z6.h\n" - ".inst 0x6466e49b // bfmmla z27.s, z4.h, z6.h\n" - ".inst 0x6467e40f // bfmmla z15.s, z0.h, z7.h\n" - ".inst 0x6467e457 // bfmmla z23.s, z2.h, z7.h\n" - ".inst 0x6467e49f // bfmmla z31.s, z4.h, z7.h\n" + ".inst 0x6461e493 // bfmmla z19.s, z4.h, z1.h\n" + ".inst 0x6461e45b // bfmmla z27.s, z2.h, z1.h\n" + ".inst 0x6460e4cf // bfmmla z15.s, z6.h, z0.h\n" + ".inst 0x6460e497 // bfmmla z23.s, z4.h, z0.h\n" + ".inst 0x6460e45f // bfmmla z31.s, z2.h, z0.h\n" "bgt 61b\n" "62:" // Height 5: Multiply loop: Single iteration only "whilelt p0.s, XZR, x27\n" - "ld1rqw { z0.s }, p0/Z, [x26]\n" - "ld1rqw { z1.s }, p0/Z, [x25]\n" - ".inst 0x658ab400 // bfcvt z0.h, p5/M, z0.s\n" - "ld1rqw { z2.s }, p0/Z, [x24]\n" + "ld1rqw { z6.s }, p0/Z, [x26]\n" + "ld1rqw { z5.s }, p0/Z, [x25]\n" + ".inst 0x658ab4c6 // bfcvt z6.h, p5/M, z6.s\n" + "ld1rqw { z4.s }, p0/Z, [x24]\n" "ld1rqw { z3.s }, p0/Z, [x23]\n" - ".inst 0x658ab421 // bfcvt z1.h, p5/M, z1.s\n" - ".inst 0x658ab442 // bfcvt z2.h, p5/M, z2.s\n" - "ld1rqw { z4.s }, p0/Z, [x22]\n" - ".inst 0x658ab463 // bfcvt z3.h, p5/M, z3.s\n" - "uzp1 z0.h, z0.h, z0.h\n" - "ld1h { z6.h }, p5/Z, [x10]\n" - "uzp1 z1.h, z1.h, z1.h\n" - "uzp1 z2.h, z2.h, z2.h\n" - "ld1h { z7.h }, p5/Z, [x10, #1, MUL VL]\n" - "uzp1 z3.h, z3.h, z3.h\n" + ".inst 0x658ab4a5 // bfcvt z5.h, p5/M, z5.s\n" ".inst 0x658ab484 // bfcvt z4.h, p5/M, z4.s\n" - "trn1 z0.d, z0.d, z1.d\n" - "trn1 z2.d, z2.d, z3.d\n" - ".inst 0x6466e408 // bfmmla z8.s, z0.h, z6.h\n" + "ld1rqw { z2.s }, p0/Z, [x22]\n" + ".inst 0x658ab463 // bfcvt z3.h, p5/M, z3.s\n" + "uzp1 z6.h, z6.h, z6.h\n" + "ld1h { z1.h }, p5/Z, [x10]\n" + "uzp1 z5.h, z5.h, z5.h\n" "uzp1 z4.h, z4.h, z4.h\n" - ".inst 0x6466e450 // bfmmla z16.s, z2.h, z6.h\n" - ".inst 0x6466e498 // bfmmla z24.s, z4.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x10, #2, MUL VL]\n" - ".inst 0x6467e40c // bfmmla z12.s, z0.h, z7.h\n" - ".inst 0x6467e454 // bfmmla z20.s, z2.h, z7.h\n" - ".inst 0x6467e49c // bfmmla z28.s, z4.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #3, MUL VL]\n" - ".inst 0x6466e409 // bfmmla z9.s, z0.h, z6.h\n" - ".inst 0x6466e451 // bfmmla z17.s, z2.h, z6.h\n" - ".inst 0x6466e499 // bfmmla z25.s, z4.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x10, #4, MUL VL]\n" - ".inst 0x6467e40d // bfmmla z13.s, z0.h, z7.h\n" - ".inst 0x6467e455 // bfmmla z21.s, z2.h, z7.h\n" - ".inst 0x6467e49d // bfmmla z29.s, z4.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #5, MUL VL]\n" - ".inst 0x6466e40a // bfmmla z10.s, z0.h, z6.h\n" - ".inst 0x6466e452 // bfmmla z18.s, z2.h, z6.h\n" - ".inst 0x6466e49a // bfmmla z26.s, z4.h, z6.h\n" - "ld1h { z6.h }, p5/Z, [x10, #6, MUL VL]\n" - ".inst 0x6467e40e // bfmmla z14.s, z0.h, z7.h\n" - ".inst 0x6467e456 // bfmmla z22.s, z2.h, z7.h\n" - ".inst 0x6467e49e // bfmmla z30.s, z4.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #7, MUL VL]\n" - ".inst 0x6466e40b // bfmmla z11.s, z0.h, z6.h\n" + "ld1h { z0.h }, p5/Z, [x10, #1, MUL VL]\n" + "uzp1 z3.h, z3.h, z3.h\n" + ".inst 0x658ab442 // bfcvt z2.h, p5/M, z2.s\n" + "trn1 z6.d, z6.d, z5.d\n" + "trn1 z4.d, z4.d, z3.d\n" + ".inst 0x6461e4c8 // bfmmla z8.s, z6.h, z1.h\n" + "uzp1 z2.h, z2.h, z2.h\n" + ".inst 0x6461e490 // bfmmla z16.s, z4.h, z1.h\n" + ".inst 0x6461e458 // bfmmla z24.s, z2.h, z1.h\n" + "ld1h { z1.h }, p5/Z, [x10, #2, MUL VL]\n" + ".inst 0x6460e4cc // bfmmla z12.s, z6.h, z0.h\n" + ".inst 0x6460e494 // bfmmla z20.s, z4.h, z0.h\n" + ".inst 0x6460e45c // bfmmla z28.s, z2.h, z0.h\n" + "ld1h { z0.h }, p5/Z, [x10, #3, MUL VL]\n" + ".inst 0x6461e4c9 // bfmmla z9.s, z6.h, z1.h\n" + ".inst 0x6461e491 // bfmmla z17.s, z4.h, z1.h\n" + ".inst 0x6461e459 // bfmmla z25.s, z2.h, z1.h\n" + "ld1h { z1.h }, p5/Z, [x10, #4, MUL VL]\n" + ".inst 0x6460e4cd // bfmmla z13.s, z6.h, z0.h\n" + ".inst 0x6460e495 // bfmmla z21.s, z4.h, z0.h\n" + ".inst 0x6460e45d // bfmmla z29.s, z2.h, z0.h\n" + "ld1h { z0.h }, p5/Z, [x10, #5, MUL VL]\n" + ".inst 0x6461e4ca // bfmmla z10.s, z6.h, z1.h\n" + ".inst 0x6461e492 // bfmmla z18.s, z4.h, z1.h\n" + ".inst 0x6461e45a // bfmmla z26.s, z2.h, z1.h\n" + "ld1h { z1.h }, p5/Z, [x10, #6, MUL VL]\n" + ".inst 0x6460e4ce // bfmmla z14.s, z6.h, z0.h\n" + ".inst 0x6460e496 // bfmmla z22.s, z4.h, z0.h\n" + ".inst 0x6460e45e // bfmmla z30.s, z2.h, z0.h\n" + "ld1h { z0.h }, p5/Z, [x10, #7, MUL VL]\n" + ".inst 0x6461e4cb // bfmmla z11.s, z6.h, z1.h\n" "addvl x10, x10, #8\n" - ".inst 0x6466e453 // bfmmla z19.s, z2.h, z6.h\n" - ".inst 0x6466e49b // bfmmla z27.s, z4.h, z6.h\n" - ".inst 0x6467e40f // bfmmla z15.s, z0.h, z7.h\n" - ".inst 0x6467e457 // bfmmla z23.s, z2.h, z7.h\n" - ".inst 0x6467e49f // bfmmla z31.s, z4.h, z7.h\n" + ".inst 0x6461e493 // bfmmla z19.s, z4.h, z1.h\n" + ".inst 0x6461e45b // bfmmla z27.s, z2.h, z1.h\n" + ".inst 0x6460e4cf // bfmmla z15.s, z6.h, z0.h\n" + ".inst 0x6460e497 // bfmmla z23.s, z4.h, z0.h\n" + ".inst 0x6460e45f // bfmmla z31.s, z2.h, z0.h\n" "63:" // Height 5: Multiply loop: multiply skip "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x28, x28, #0x1\n" @@ -1297,49 +1297,49 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "uzp1 z27.d, z27.d, z31.d\n" "tbz %x[flags], #1, 64f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1rw { z1.s }, p5/Z, [x20]\n" - "add x20, %x[args_ptr], %[offset_min]\n" "ld1rw { z0.s }, p5/Z, [x20]\n" - "fmin z6.s, p5/M, z6.s, z1.s\n" - "fmin z12.s, p5/M, z12.s, z1.s\n" - "fmin z13.s, p5/M, z13.s, z1.s\n" - "fmin z14.s, p5/M, z14.s, z1.s\n" - "fmin z8.s, p5/M, z8.s, z1.s\n" - "fmin z9.s, p5/M, z9.s, z1.s\n" - "fmin z10.s, p5/M, z10.s, z1.s\n" - "fmin z11.s, p5/M, z11.s, z1.s\n" - "fmin z15.s, p5/M, z15.s, z1.s\n" - "fmin z20.s, p5/M, z20.s, z1.s\n" - "fmin z21.s, p5/M, z21.s, z1.s\n" - "fmin z22.s, p5/M, z22.s, z1.s\n" - "fmin z16.s, p5/M, z16.s, z1.s\n" - "fmin z17.s, p5/M, z17.s, z1.s\n" - "fmin z18.s, p5/M, z18.s, z1.s\n" - "fmin z19.s, p5/M, z19.s, z1.s\n" - "fmin z24.s, p5/M, z24.s, z1.s\n" - "fmin z25.s, p5/M, z25.s, z1.s\n" - "fmin z26.s, p5/M, z26.s, z1.s\n" - "fmin z27.s, p5/M, z27.s, z1.s\n" - "fmax z6.s, p5/M, z6.s, z0.s\n" - "fmax z12.s, p5/M, z12.s, z0.s\n" - "fmax z13.s, p5/M, z13.s, z0.s\n" - "fmax z14.s, p5/M, z14.s, z0.s\n" - "fmax z8.s, p5/M, z8.s, z0.s\n" - "fmax z9.s, p5/M, z9.s, z0.s\n" - "fmax z10.s, p5/M, z10.s, z0.s\n" - "fmax z11.s, p5/M, z11.s, z0.s\n" - "fmax z15.s, p5/M, z15.s, z0.s\n" - "fmax z20.s, p5/M, z20.s, z0.s\n" - "fmax z21.s, p5/M, z21.s, z0.s\n" - "fmax z22.s, p5/M, z22.s, z0.s\n" - "fmax z16.s, p5/M, z16.s, z0.s\n" - "fmax z17.s, p5/M, z17.s, z0.s\n" - "fmax z18.s, p5/M, z18.s, z0.s\n" - "fmax z19.s, p5/M, z19.s, z0.s\n" - "fmax z24.s, p5/M, z24.s, z0.s\n" - "fmax z25.s, p5/M, z25.s, z0.s\n" - "fmax z26.s, p5/M, z26.s, z0.s\n" - "fmax z27.s, p5/M, z27.s, z0.s\n" + "add x20, %x[args_ptr], %[offset_min]\n" + "ld1rw { z23.s }, p5/Z, [x20]\n" + "fmin z6.s, p5/M, z6.s, z0.s\n" + "fmin z12.s, p5/M, z12.s, z0.s\n" + "fmin z13.s, p5/M, z13.s, z0.s\n" + "fmin z14.s, p5/M, z14.s, z0.s\n" + "fmin z8.s, p5/M, z8.s, z0.s\n" + "fmin z9.s, p5/M, z9.s, z0.s\n" + "fmin z10.s, p5/M, z10.s, z0.s\n" + "fmin z11.s, p5/M, z11.s, z0.s\n" + "fmin z15.s, p5/M, z15.s, z0.s\n" + "fmin z20.s, p5/M, z20.s, z0.s\n" + "fmin z21.s, p5/M, z21.s, z0.s\n" + "fmin z22.s, p5/M, z22.s, z0.s\n" + "fmin z16.s, p5/M, z16.s, z0.s\n" + "fmin z17.s, p5/M, z17.s, z0.s\n" + "fmin z18.s, p5/M, z18.s, z0.s\n" + "fmin z19.s, p5/M, z19.s, z0.s\n" + "fmin z24.s, p5/M, z24.s, z0.s\n" + "fmin z25.s, p5/M, z25.s, z0.s\n" + "fmin z26.s, p5/M, z26.s, z0.s\n" + "fmin z27.s, p5/M, z27.s, z0.s\n" + "fmax z6.s, p5/M, z6.s, z23.s\n" + "fmax z12.s, p5/M, z12.s, z23.s\n" + "fmax z13.s, p5/M, z13.s, z23.s\n" + "fmax z14.s, p5/M, z14.s, z23.s\n" + "fmax z8.s, p5/M, z8.s, z23.s\n" + "fmax z9.s, p5/M, z9.s, z23.s\n" + "fmax z10.s, p5/M, z10.s, z23.s\n" + "fmax z11.s, p5/M, z11.s, z23.s\n" + "fmax z15.s, p5/M, z15.s, z23.s\n" + "fmax z20.s, p5/M, z20.s, z23.s\n" + "fmax z21.s, p5/M, z21.s, z23.s\n" + "fmax z22.s, p5/M, z22.s, z23.s\n" + "fmax z16.s, p5/M, z16.s, z23.s\n" + "fmax z17.s, p5/M, z17.s, z23.s\n" + "fmax z18.s, p5/M, z18.s, z23.s\n" + "fmax z19.s, p5/M, z19.s, z23.s\n" + "fmax z24.s, p5/M, z24.s, z23.s\n" + "fmax z25.s, p5/M, z25.s, z23.s\n" + "fmax z26.s, p5/M, z26.s, z23.s\n" + "fmax z27.s, p5/M, z27.s, z23.s\n" "64:" // Height 5: No activation "st1w { z6.s }, p4, [x9]\n" "st1w { z12.s }, p3, [x9, #1, MUL VL]\n" @@ -1418,59 +1418,59 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "68:" // Height 6: no bias "tbz %x[flags], #0, 69f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x9, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" - "ld1w { z9.s }, p4/Z, [x9]\n" + "add x24, x9, x20, LSL #2\n" "add x23, x24, x20, LSL #2\n" + "ld1w { z17.s }, p4/Z, [x9]\n" "add x22, x23, x20, LSL #2\n" - "ld1w { z10.s }, p3/Z, [x9, #1, MUL VL]\n" - "ld1w { z11.s }, p2/Z, [x9, #2, MUL VL]\n" "add x21, x22, x20, LSL #2\n" + "ld1w { z18.s }, p3/Z, [x9, #1, MUL VL]\n" + "ld1w { z20.s }, p2/Z, [x9, #2, MUL VL]\n" + "add x20, x21, x20, LSL #2\n" "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n" - "ld1w { z12.s }, p4/Z, [x25]\n" - "zip1 z8.d, z9.d, z12.d\n" - "ld1w { z13.s }, p3/Z, [x25, #1, MUL VL]\n" - "ld1w { z14.s }, p2/Z, [x25, #2, MUL VL]\n" - "zip2 z12.d, z9.d, z12.d\n" - "zip1 z9.d, z10.d, z13.d\n" - "ld1w { z15.s }, p1/Z, [x25, #3, MUL VL]\n" - "ld1w { z17.s }, p4/Z, [x24]\n" - "zip2 z13.d, z10.d, z13.d\n" - "zip1 z10.d, z11.d, z14.d\n" - "ld1w { z18.s }, p3/Z, [x24, #1, MUL VL]\n" - "ld1w { z19.s }, p2/Z, [x24, #2, MUL VL]\n" - "zip2 z14.d, z11.d, z14.d\n" + "ld1w { z12.s }, p4/Z, [x24]\n" + "zip1 z8.d, z17.d, z12.d\n" + "ld1w { z13.s }, p3/Z, [x24, #1, MUL VL]\n" + "ld1w { z14.s }, p2/Z, [x24, #2, MUL VL]\n" + "zip2 z12.d, z17.d, z12.d\n" + "zip1 z9.d, z18.d, z13.d\n" + "ld1w { z15.s }, p1/Z, [x24, #3, MUL VL]\n" + "ld1w { z17.s }, p4/Z, [x23]\n" + "zip2 z13.d, z18.d, z13.d\n" + "zip1 z10.d, z20.d, z14.d\n" + "ld1w { z18.s }, p3/Z, [x23, #1, MUL VL]\n" + "ld1w { z19.s }, p2/Z, [x23, #2, MUL VL]\n" + "zip2 z14.d, z20.d, z14.d\n" "zip1 z11.d, z16.d, z15.d\n" - "ld1w { z24.s }, p1/Z, [x24, #3, MUL VL]\n" - "ld1w { z20.s }, p4/Z, [x23]\n" + "ld1w { z24.s }, p1/Z, [x23, #3, MUL VL]\n" + "ld1w { z20.s }, p4/Z, [x22]\n" "zip2 z15.d, z16.d, z15.d\n" "zip1 z16.d, z17.d, z20.d\n" - "ld1w { z21.s }, p3/Z, [x23, #1, MUL VL]\n" - "ld1w { z22.s }, p2/Z, [x23, #2, MUL VL]\n" + "ld1w { z21.s }, p3/Z, [x22, #1, MUL VL]\n" + "ld1w { z22.s }, p2/Z, [x22, #2, MUL VL]\n" "zip2 z20.d, z17.d, z20.d\n" "zip1 z17.d, z18.d, z21.d\n" - "ld1w { z23.s }, p1/Z, [x23, #3, MUL VL]\n" - "ld1w { z25.s }, p4/Z, [x22]\n" + "ld1w { z23.s }, p1/Z, [x22, #3, MUL VL]\n" + "ld1w { z25.s }, p4/Z, [x21]\n" "zip2 z21.d, z18.d, z21.d\n" "zip1 z18.d, z19.d, z22.d\n" - "ld1w { z26.s }, p3/Z, [x22, #1, MUL VL]\n" - "ld1w { z27.s }, p2/Z, [x22, #2, MUL VL]\n" + "ld1w { z26.s }, p3/Z, [x21, #1, MUL VL]\n" + "ld1w { z27.s }, p2/Z, [x21, #2, MUL VL]\n" "zip2 z22.d, z19.d, z22.d\n" "zip1 z19.d, z24.d, z23.d\n" - "ld1w { z6.s }, p1/Z, [x22, #3, MUL VL]\n" - "ld1w { z28.s }, p4/Z, [x21]\n" + "ld1w { z0.s }, p1/Z, [x21, #3, MUL VL]\n" + "ld1w { z28.s }, p4/Z, [x20]\n" "zip2 z23.d, z24.d, z23.d\n" "zip1 z24.d, z25.d, z28.d\n" - "ld1w { z29.s }, p3/Z, [x21, #1, MUL VL]\n" - "ld1w { z30.s }, p2/Z, [x21, #2, MUL VL]\n" + "ld1w { z29.s }, p3/Z, [x20, #1, MUL VL]\n" + "ld1w { z30.s }, p2/Z, [x20, #2, MUL VL]\n" "zip2 z28.d, z25.d, z28.d\n" "zip1 z25.d, z26.d, z29.d\n" - "ld1w { z31.s }, p1/Z, [x21, #3, MUL VL]\n" + "ld1w { z31.s }, p1/Z, [x20, #3, MUL VL]\n" "zip2 z29.d, z26.d, z29.d\n" "zip1 z26.d, z27.d, z30.d\n" "zip2 z30.d, z27.d, z30.d\n" - "zip1 z27.d, z6.d, z31.d\n" - "zip2 z31.d, z6.d, z31.d\n" + "zip1 z27.d, z0.d, z31.d\n" + "zip2 z31.d, z0.d, z31.d\n" "b 70f\n" "69:" // Height 6: no accumulate "mov z8.b, #0x0\n" @@ -1502,16 +1502,16 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "71:" // Height 6: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w27, [x20, x28, 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, 72f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" - "ldr x25, [x21, #0x8]\n" - "ldr x24, [x21, #0x10]\n" - "ldr x23, [x21, #0x18]\n" - "ldr x22, [x21, #0x20]\n" - "ldr x21, [x21, #0x28]\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" + "ldr x25, [x20, #0x8]\n" + "ldr x24, [x20, #0x10]\n" + "ldr x23, [x20, #0x18]\n" + "ldr x22, [x20, #0x20]\n" + "ldr x21, [x20, #0x28]\n" "cbnz x28, 73f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #2\n" @@ -1523,135 +1523,135 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "b 73f\n" "72:" // Height 6: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" - "add x23, x24, x20, LSL #2\n" - "add x22, x23, x20, LSL #2\n" - "add x21, x22, x20, LSL #2\n" + "add x25, x26, x21, LSL #2\n" + "add x24, x25, x21, LSL #2\n" + "add x23, x24, x21, LSL #2\n" + "add x22, x23, x21, LSL #2\n" + "add x21, x22, x21, LSL #2\n" "73:" // Height 6: input setup done "cmp x27, #0x4\n" "ble 75f\n" "74:" // Height 6: Multiply loop: Main loop head "whilelt p0.s, XZR, x27\n" - "ld1rqw { z0.s }, p0/Z, [x26]\n" - "ld1rqw { z1.s }, p0/Z, [x25]\n" - ".inst 0x658ab400 // bfcvt z0.h, p5/M, z0.s\n" - "ld1rqw { z2.s }, p0/Z, [x24]\n" - "ld1rqw { z3.s }, p0/Z, [x23]\n" - ".inst 0x658ab421 // bfcvt z1.h, p5/M, z1.s\n" - ".inst 0x658ab442 // bfcvt z2.h, p5/M, z2.s\n" - "ld1rqw { z4.s }, p0/Z, [x22]\n" - "ld1rqw { z5.s }, p0/Z, [x21]\n" - ".inst 0x658ab463 // bfcvt z3.h, p5/M, z3.s\n" - ".inst 0x658ab484 // bfcvt z4.h, p5/M, z4.s\n" + "ld1rqw { z7.s }, p0/Z, [x26]\n" + "ld1rqw { z6.s }, p0/Z, [x25]\n" + ".inst 0x658ab4e7 // bfcvt z7.h, p5/M, z7.s\n" + "ld1rqw { z5.s }, p0/Z, [x24]\n" + "ld1rqw { z4.s }, p0/Z, [x23]\n" + ".inst 0x658ab4c6 // bfcvt z6.h, p5/M, z6.s\n" ".inst 0x658ab4a5 // bfcvt z5.h, p5/M, z5.s\n" - "uzp1 z0.h, z0.h, z0.h\n" - "ld1h { z6.h }, p5/Z, [x10]\n" - "ld1h { z7.h }, p5/Z, [x10, #1, MUL VL]\n" - "uzp1 z1.h, z1.h, z1.h\n" - "uzp1 z2.h, z2.h, z2.h\n" + "ld1rqw { z3.s }, p0/Z, [x22]\n" + "ld1rqw { z2.s }, p0/Z, [x21]\n" + ".inst 0x658ab484 // bfcvt z4.h, p5/M, z4.s\n" + ".inst 0x658ab463 // bfcvt z3.h, p5/M, z3.s\n" + ".inst 0x658ab442 // bfcvt z2.h, p5/M, z2.s\n" + "uzp1 z7.h, z7.h, z7.h\n" + "ld1h { z1.h }, p5/Z, [x10]\n" + "ld1h { z0.h }, p5/Z, [x10, #1, MUL VL]\n" + "uzp1 z6.h, z6.h, z6.h\n" + "uzp1 z5.h, z5.h, z5.h\n" "sub x27, x27, #0x4\n" "cmp x27, #0x4\n" - "uzp1 z3.h, z3.h, z3.h\n" "uzp1 z4.h, z4.h, z4.h\n" + "uzp1 z3.h, z3.h, z3.h\n" "add x26, x26, #0x10\n" "add x25, x25, #0x10\n" - "uzp1 z5.h, z5.h, z5.h\n" - "trn1 z0.d, z0.d, z1.d\n" - ".inst 0x6466e408 // bfmmla z8.s, z0.h, z6.h\n" + "uzp1 z2.h, z2.h, z2.h\n" + "trn1 z7.d, z7.d, z6.d\n" + ".inst 0x6461e4e8 // bfmmla z8.s, z7.h, z1.h\n" "add x24, x24, #0x10\n" - "trn1 z2.d, z2.d, z3.d\n" - "trn1 z4.d, z4.d, z5.d\n" - ".inst 0x6466e450 // bfmmla z16.s, z2.h, z6.h\n" + "trn1 z5.d, z5.d, z4.d\n" + "trn1 z3.d, z3.d, z2.d\n" + ".inst 0x6461e4b0 // bfmmla z16.s, z5.h, z1.h\n" "add x23, x23, #0x10\n" - ".inst 0x6466e498 // bfmmla z24.s, z4.h, z6.h\n" - ".inst 0x6467e40c // bfmmla z12.s, z0.h, z7.h\n" - "ld1h { z6.h }, p5/Z, [x10, #2, MUL VL]\n" + ".inst 0x6461e478 // bfmmla z24.s, z3.h, z1.h\n" + ".inst 0x6460e4ec // bfmmla z12.s, z7.h, z0.h\n" + "ld1h { z1.h }, p5/Z, [x10, #2, MUL VL]\n" "add x22, x22, #0x10\n" - ".inst 0x6467e454 // bfmmla z20.s, z2.h, z7.h\n" - ".inst 0x6467e49c // bfmmla z28.s, z4.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #3, MUL VL]\n" + ".inst 0x6460e4b4 // bfmmla z20.s, z5.h, z0.h\n" + ".inst 0x6460e47c // bfmmla z28.s, z3.h, z0.h\n" + "ld1h { z0.h }, p5/Z, [x10, #3, MUL VL]\n" "add x21, x21, #0x10\n" - ".inst 0x6466e409 // bfmmla z9.s, z0.h, z6.h\n" - ".inst 0x6466e451 // bfmmla z17.s, z2.h, z6.h\n" - ".inst 0x6466e499 // bfmmla z25.s, z4.h, z6.h\n" - ".inst 0x6467e40d // bfmmla z13.s, z0.h, z7.h\n" - "ld1h { z6.h }, p5/Z, [x10, #4, MUL VL]\n" - ".inst 0x6467e455 // bfmmla z21.s, z2.h, z7.h\n" - ".inst 0x6467e49d // bfmmla z29.s, z4.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #5, MUL VL]\n" - ".inst 0x6466e40a // bfmmla z10.s, z0.h, z6.h\n" - ".inst 0x6466e452 // bfmmla z18.s, z2.h, z6.h\n" - ".inst 0x6466e49a // bfmmla z26.s, z4.h, z6.h\n" - ".inst 0x6467e40e // bfmmla z14.s, z0.h, z7.h\n" - "ld1h { z6.h }, p5/Z, [x10, #6, MUL VL]\n" - ".inst 0x6467e456 // bfmmla z22.s, z2.h, z7.h\n" - ".inst 0x6467e49e // bfmmla z30.s, z4.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #7, MUL VL]\n" + ".inst 0x6461e4e9 // bfmmla z9.s, z7.h, z1.h\n" + ".inst 0x6461e4b1 // bfmmla z17.s, z5.h, z1.h\n" + ".inst 0x6461e479 // bfmmla z25.s, z3.h, z1.h\n" + ".inst 0x6460e4ed // bfmmla z13.s, z7.h, z0.h\n" + "ld1h { z1.h }, p5/Z, [x10, #4, MUL VL]\n" + ".inst 0x6460e4b5 // bfmmla z21.s, z5.h, z0.h\n" + ".inst 0x6460e47d // bfmmla z29.s, z3.h, z0.h\n" + "ld1h { z0.h }, p5/Z, [x10, #5, MUL VL]\n" + ".inst 0x6461e4ea // bfmmla z10.s, z7.h, z1.h\n" + ".inst 0x6461e4b2 // bfmmla z18.s, z5.h, z1.h\n" + ".inst 0x6461e47a // bfmmla z26.s, z3.h, z1.h\n" + ".inst 0x6460e4ee // bfmmla z14.s, z7.h, z0.h\n" + "ld1h { z1.h }, p5/Z, [x10, #6, MUL VL]\n" + ".inst 0x6460e4b6 // bfmmla z22.s, z5.h, z0.h\n" + ".inst 0x6460e47e // bfmmla z30.s, z3.h, z0.h\n" + "ld1h { z0.h }, p5/Z, [x10, #7, MUL VL]\n" "addvl x10, x10, #8\n" - ".inst 0x6466e40b // bfmmla z11.s, z0.h, z6.h\n" - ".inst 0x6466e453 // bfmmla z19.s, z2.h, z6.h\n" - ".inst 0x6466e49b // bfmmla z27.s, z4.h, z6.h\n" - ".inst 0x6467e40f // bfmmla z15.s, z0.h, z7.h\n" - ".inst 0x6467e457 // bfmmla z23.s, z2.h, z7.h\n" - ".inst 0x6467e49f // bfmmla z31.s, z4.h, z7.h\n" + ".inst 0x6461e4eb // bfmmla z11.s, z7.h, z1.h\n" + ".inst 0x6461e4b3 // bfmmla z19.s, z5.h, z1.h\n" + ".inst 0x6461e47b // bfmmla z27.s, z3.h, z1.h\n" + ".inst 0x6460e4ef // bfmmla z15.s, z7.h, z0.h\n" + ".inst 0x6460e4b7 // bfmmla z23.s, z5.h, z0.h\n" + ".inst 0x6460e47f // bfmmla z31.s, z3.h, z0.h\n" "bgt 74b\n" "75:" // Height 6: Multiply loop: Single iteration only "whilelt p0.s, XZR, x27\n" - "ld1rqw { z0.s }, p0/Z, [x26]\n" - "ld1rqw { z1.s }, p0/Z, [x25]\n" - ".inst 0x658ab400 // bfcvt z0.h, p5/M, z0.s\n" - "ld1rqw { z2.s }, p0/Z, [x24]\n" - "ld1rqw { z3.s }, p0/Z, [x23]\n" - ".inst 0x658ab421 // bfcvt z1.h, p5/M, z1.s\n" - ".inst 0x658ab442 // bfcvt z2.h, p5/M, z2.s\n" - "ld1rqw { z4.s }, p0/Z, [x22]\n" - "ld1rqw { z5.s }, p0/Z, [x21]\n" - ".inst 0x658ab463 // bfcvt z3.h, p5/M, z3.s\n" - ".inst 0x658ab484 // bfcvt z4.h, p5/M, z4.s\n" + "ld1rqw { z7.s }, p0/Z, [x26]\n" + "ld1rqw { z6.s }, p0/Z, [x25]\n" + ".inst 0x658ab4e7 // bfcvt z7.h, p5/M, z7.s\n" + "ld1rqw { z5.s }, p0/Z, [x24]\n" + "ld1rqw { z4.s }, p0/Z, [x23]\n" + ".inst 0x658ab4c6 // bfcvt z6.h, p5/M, z6.s\n" ".inst 0x658ab4a5 // bfcvt z5.h, p5/M, z5.s\n" - "uzp1 z0.h, z0.h, z0.h\n" - "ld1h { z6.h }, p5/Z, [x10]\n" - "ld1h { z7.h }, p5/Z, [x10, #1, MUL VL]\n" - "uzp1 z1.h, z1.h, z1.h\n" - "uzp1 z2.h, z2.h, z2.h\n" - "uzp1 z3.h, z3.h, z3.h\n" - "uzp1 z4.h, z4.h, z4.h\n" + "ld1rqw { z3.s }, p0/Z, [x22]\n" + "ld1rqw { z2.s }, p0/Z, [x21]\n" + ".inst 0x658ab484 // bfcvt z4.h, p5/M, z4.s\n" + ".inst 0x658ab463 // bfcvt z3.h, p5/M, z3.s\n" + ".inst 0x658ab442 // bfcvt z2.h, p5/M, z2.s\n" + "uzp1 z7.h, z7.h, z7.h\n" + "ld1h { z1.h }, p5/Z, [x10]\n" + "ld1h { z0.h }, p5/Z, [x10, #1, MUL VL]\n" + "uzp1 z6.h, z6.h, z6.h\n" "uzp1 z5.h, z5.h, z5.h\n" - "trn1 z0.d, z0.d, z1.d\n" - ".inst 0x6466e408 // bfmmla z8.s, z0.h, z6.h\n" - "trn1 z2.d, z2.d, z3.d\n" - "trn1 z4.d, z4.d, z5.d\n" - ".inst 0x6466e450 // bfmmla z16.s, z2.h, z6.h\n" - ".inst 0x6466e498 // bfmmla z24.s, z4.h, z6.h\n" - ".inst 0x6467e40c // bfmmla z12.s, z0.h, z7.h\n" - "ld1h { z6.h }, p5/Z, [x10, #2, MUL VL]\n" - ".inst 0x6467e454 // bfmmla z20.s, z2.h, z7.h\n" - ".inst 0x6467e49c // bfmmla z28.s, z4.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #3, MUL VL]\n" - ".inst 0x6466e409 // bfmmla z9.s, z0.h, z6.h\n" - ".inst 0x6466e451 // bfmmla z17.s, z2.h, z6.h\n" - ".inst 0x6466e499 // bfmmla z25.s, z4.h, z6.h\n" - ".inst 0x6467e40d // bfmmla z13.s, z0.h, z7.h\n" - "ld1h { z6.h }, p5/Z, [x10, #4, MUL VL]\n" - ".inst 0x6467e455 // bfmmla z21.s, z2.h, z7.h\n" - ".inst 0x6467e49d // bfmmla z29.s, z4.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #5, MUL VL]\n" - ".inst 0x6466e40a // bfmmla z10.s, z0.h, z6.h\n" - ".inst 0x6466e452 // bfmmla z18.s, z2.h, z6.h\n" - ".inst 0x6466e49a // bfmmla z26.s, z4.h, z6.h\n" - ".inst 0x6467e40e // bfmmla z14.s, z0.h, z7.h\n" - "ld1h { z6.h }, p5/Z, [x10, #6, MUL VL]\n" - ".inst 0x6467e456 // bfmmla z22.s, z2.h, z7.h\n" - ".inst 0x6467e49e // bfmmla z30.s, z4.h, z7.h\n" - "ld1h { z7.h }, p5/Z, [x10, #7, MUL VL]\n" + "uzp1 z4.h, z4.h, z4.h\n" + "uzp1 z3.h, z3.h, z3.h\n" + "uzp1 z2.h, z2.h, z2.h\n" + "trn1 z7.d, z7.d, z6.d\n" + ".inst 0x6461e4e8 // bfmmla z8.s, z7.h, z1.h\n" + "trn1 z5.d, z5.d, z4.d\n" + "trn1 z3.d, z3.d, z2.d\n" + ".inst 0x6461e4b0 // bfmmla z16.s, z5.h, z1.h\n" + ".inst 0x6461e478 // bfmmla z24.s, z3.h, z1.h\n" + ".inst 0x6460e4ec // bfmmla z12.s, z7.h, z0.h\n" + "ld1h { z1.h }, p5/Z, [x10, #2, MUL VL]\n" + ".inst 0x6460e4b4 // bfmmla z20.s, z5.h, z0.h\n" + ".inst 0x6460e47c // bfmmla z28.s, z3.h, z0.h\n" + "ld1h { z0.h }, p5/Z, [x10, #3, MUL VL]\n" + ".inst 0x6461e4e9 // bfmmla z9.s, z7.h, z1.h\n" + ".inst 0x6461e4b1 // bfmmla z17.s, z5.h, z1.h\n" + ".inst 0x6461e479 // bfmmla z25.s, z3.h, z1.h\n" + ".inst 0x6460e4ed // bfmmla z13.s, z7.h, z0.h\n" + "ld1h { z1.h }, p5/Z, [x10, #4, MUL VL]\n" + ".inst 0x6460e4b5 // bfmmla z21.s, z5.h, z0.h\n" + ".inst 0x6460e47d // bfmmla z29.s, z3.h, z0.h\n" + "ld1h { z0.h }, p5/Z, [x10, #5, MUL VL]\n" + ".inst 0x6461e4ea // bfmmla z10.s, z7.h, z1.h\n" + ".inst 0x6461e4b2 // bfmmla z18.s, z5.h, z1.h\n" + ".inst 0x6461e47a // bfmmla z26.s, z3.h, z1.h\n" + ".inst 0x6460e4ee // bfmmla z14.s, z7.h, z0.h\n" + "ld1h { z1.h }, p5/Z, [x10, #6, MUL VL]\n" + ".inst 0x6460e4b6 // bfmmla z22.s, z5.h, z0.h\n" + ".inst 0x6460e47e // bfmmla z30.s, z3.h, z0.h\n" + "ld1h { z0.h }, p5/Z, [x10, #7, MUL VL]\n" "addvl x10, x10, #8\n" - ".inst 0x6466e40b // bfmmla z11.s, z0.h, z6.h\n" - ".inst 0x6466e453 // bfmmla z19.s, z2.h, z6.h\n" - ".inst 0x6466e49b // bfmmla z27.s, z4.h, z6.h\n" - ".inst 0x6467e40f // bfmmla z15.s, z0.h, z7.h\n" - ".inst 0x6467e457 // bfmmla z23.s, z2.h, z7.h\n" - ".inst 0x6467e49f // bfmmla z31.s, z4.h, z7.h\n" + ".inst 0x6461e4eb // bfmmla z11.s, z7.h, z1.h\n" + ".inst 0x6461e4b3 // bfmmla z19.s, z5.h, z1.h\n" + ".inst 0x6461e47b // bfmmla z27.s, z3.h, z1.h\n" + ".inst 0x6460e4ef // bfmmla z15.s, z7.h, z0.h\n" + ".inst 0x6460e4b7 // bfmmla z23.s, z5.h, z0.h\n" + ".inst 0x6460e47f // bfmmla z31.s, z3.h, z0.h\n" "76:" // Height 6: Multiply loop: multiply skip "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x28, x28, #0x1\n" @@ -1782,7 +1782,6 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "madd %x[input_ptr], x20, x21, %x[input_ptr]\n" "b 1b\n" "80:" // Exit - : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr) : [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, 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_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", "p0", "p1", "p2", "p3", "p4", "p5", "x9", "x10", "x11", "x12", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31" @@ -1790,4 +1789,4 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( } } // namespace arm_gemm -#endif // __ARM_FEATURE_SVE +#endif // ARM_COMPUTE_ENABLE_SVE -- cgit v1.2.1