diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32bf16fp32_mmla_6x4VL/generic.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp32bf16fp32_mmla_6x4VL/generic.cpp | 1023 |
1 files changed, 513 insertions, 510 deletions
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 0d2b47ec39..11526b6f65 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 @@ -1,5 +1,5 @@ /* - * Copyright (c) 2021, 2023 Arm Limited. + * Copyright (c) 2021, 2023-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -48,18 +48,19 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( size_t output_offset = {}; size_t input_initial_col = {}; size_t input_offset = {}; + void *output_ptr = {}; + const float *bias = {}; } ka; unsigned long flags=0; - void *output_ptr; void *input_ptr; if (output_arg.is_indirect) { - output_ptr=(void *)(output_arg.indirect.ptr); + ka.output_ptr=(void *)(output_arg.indirect.ptr); ka.output_offset=output_arg.indirect.offset; flags |= 0x4; } else { - output_ptr=(void *)(output_arg.direct.base); + ka.output_ptr=(void *)(output_arg.direct.base); ka.output_offset=output_arg.direct.stride; } @@ -80,6 +81,7 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( ka.string_lengths = string_lengths; ka.N = N; ka.B_ptr = B_ptr; + ka.bias = bias; switch(act.type) { default: case Activation::Type::None: @@ -103,10 +105,10 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "cmp %x[M], #0x2\n" "bgt 27f\n" "beq 14f\n" - "mov x12, %x[bias]\n" + "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n" "ldr x11, [%x[args_ptr], %[offsetof_N]]\n" "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x9, %x[output_ptr]\n" + "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n" "2:" // Height 1: Column loop "mov x20, #0x0\n" "whilelt p4.s, x20, x11\n" @@ -119,26 +121,26 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "cbz x12, 3f\n" "ld1w { z8.s }, p5/Z, [x12]\n" "ld1w { z9.s }, p5/Z, [x12, #1, MUL VL]\n" - "zip2 z12.d, z8.d, z8.d\n" - "zip1 z8.d, z8.d, z8.d\n" "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n" "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n" + "addvl x12, x12, #4\n" + "zip2 z12.d, z8.d, z8.d\n" + "zip1 z8.d, z8.d, z8.d\n" "zip2 z13.d, z9.d, z9.d\n" "zip1 z9.d, z9.d, z9.d\n" "zip2 z14.d, z10.d, z10.d\n" "zip1 z10.d, z10.d, z10.d\n" - "addvl x12, x12, #4\n" "zip2 z15.d, z11.d, z11.d\n" "zip1 z11.d, z11.d, z11.d\n" "b 5f\n" "3:" // Height 1: no bias "tbz %x[flags], #0, 4f\n" - "ld1w { z16.s }, p4/Z, [x9]\n" + "ld1w { z19.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 z8.d, z19.d, z12.d\n" + "zip2 z12.d, z19.d, z12.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" @@ -159,8 +161,8 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "mov x28, #0x0\n" "6:" // Height 1: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr w27, [x20, x28, LSL #0x2]\n" "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr w27, [x20, x28, LSL #0x2]\n" "tbz %x[flags], #3, 7f\n" "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -176,52 +178,52 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "ble 10f\n" "9:" // Height 1: Multiply loop: Main loop head "whilelt p0.s, XZR, x27\n" + "ld1h { z17.h }, p5/Z, [x10]\n" + "ld1h { z16.h }, p5/Z, [x10, #1, MUL VL]\n" + "sub x27, x27, #0x4\n" + "cmp x27, #0x4\n" "ld1rqw { z18.s }, p0/Z, [x26]\n" + "add x26, x26, #0x10\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" + ".inst 0x6470e64c // bfmmla z12.s, z18.h, z16.h\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" + ".inst 0x6470e64d // bfmmla z13.s, z18.h, z16.h\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" + ".inst 0x6470e64e // bfmmla z14.s, z18.h, z16.h\n" "ld1h { z16.h }, p5/Z, [x10, #7, MUL VL]\n" - "sub x27, x27, #0x4\n" - "cmp x27, #0x4\n" + "addvl x10, x10, #8\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" + "ld1h { z17.h }, p5/Z, [x10]\n" + "ld1h { z16.h }, p5/Z, [x10, #1, 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" + ".inst 0x6470e64c // bfmmla z12.s, z18.h, z16.h\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" + ".inst 0x6470e64d // bfmmla z13.s, z18.h, z16.h\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" + ".inst 0x6470e64e // bfmmla z14.s, z18.h, z16.h\n" "ld1h { z16.h }, p5/Z, [x10, #7, MUL VL]\n" + "addvl x10, x10, #8\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" "add x28, x28, #0x1\n" @@ -232,9 +234,9 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "uzp1 z10.d, z10.d, z14.d\n" "uzp1 z11.d, z11.d, z15.d\n" "tbz %x[flags], #1, 12f\n" - "add x20, %x[args_ptr], %[offset_max]\n" - "ld1rw { z17.s }, p5/Z, [x20]\n" + "add x21, %x[args_ptr], %[offset_max]\n" "add x20, %x[args_ptr], %[offset_min]\n" + "ld1rw { z17.s }, p5/Z, [x21]\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" @@ -256,10 +258,10 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "bgt 2b\n" "b 80f\n" "14:" // Height 2 - "mov x12, %x[bias]\n" + "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n" "ldr x11, [%x[args_ptr], %[offsetof_N]]\n" "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x9, %x[output_ptr]\n" + "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n" "15:" // Height 2: Column loop "mov x20, #0x0\n" "whilelt p4.s, x20, x11\n" @@ -272,34 +274,34 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "cbz x12, 16f\n" "ld1w { z8.s }, p5/Z, [x12]\n" "ld1w { z9.s }, p5/Z, [x12, #1, MUL VL]\n" - "zip2 z12.d, z8.d, z8.d\n" - "zip1 z8.d, z8.d, z8.d\n" "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n" "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n" + "addvl x12, x12, #4\n" + "zip2 z12.d, z8.d, z8.d\n" + "zip1 z8.d, z8.d, z8.d\n" "zip2 z13.d, z9.d, z9.d\n" "zip1 z9.d, z9.d, z9.d\n" "zip2 z14.d, z10.d, z10.d\n" "zip1 z10.d, z10.d, z10.d\n" - "addvl x12, x12, #4\n" "zip2 z15.d, z11.d, z11.d\n" "zip1 z11.d, z11.d, z11.d\n" "b 18f\n" "16:" // Height 2: no bias "tbz %x[flags], #0, 17f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\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" + "add x20, x9, x20, LSL #2\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" + "ld1w { z15.s }, p1/Z, [x20, #3, MUL VL]\n" + "zip1 z8.d, z19.d, z12.d\n" + "zip2 z12.d, z19.d, z12.d\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" @@ -318,8 +320,8 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "mov x28, #0x0\n" "19:" // Height 2: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr w27, [x20, x28, LSL #0x2]\n" "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr w27, [x20, x28, LSL #0x2]\n" "tbz %x[flags], #3, 20f\n" "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -338,61 +340,61 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "ble 23f\n" "22:" // Height 2: Multiply loop: Main loop head "whilelt p0.s, XZR, x27\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" + "ld1h { z19.h }, p5/Z, [x10, #1, MUL VL]\n" + "sub x27, x27, #0x4\n" + "cmp x27, #0x4\n" + "ld1rqw { z18.s }, p0/Z, [x26]\n" + "add x26, x26, #0x10\n" + "ld1rqw { z16.s }, p0/Z, [x25]\n" + "add x25, x25, #0x10\n" + ".inst 0x658ab652 // bfcvt z18.h, p5/M, z18.s\n" + ".inst 0x658ab610 // bfcvt z16.h, p5/M, z16.s\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" + "uzp1 z16.h, z16.h, z16.h\n" + "trn1 z18.d, z18.d, z16.d\n" + ".inst 0x6471e648 // bfmmla z8.s, z18.h, z17.h\n" "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n" - ".inst 0x6470e66c // bfmmla z12.s, z19.h, z16.h\n" + ".inst 0x6473e64c // bfmmla z12.s, z18.h, z19.h\n" "ld1h { z16.h }, p5/Z, [x10, #3, MUL VL]\n" - ".inst 0x6471e669 // bfmmla z9.s, z19.h, z17.h\n" + ".inst 0x6471e649 // bfmmla z9.s, z18.h, z17.h\n" "ld1h { z17.h }, p5/Z, [x10, #4, MUL VL]\n" - ".inst 0x6470e66d // bfmmla z13.s, z19.h, z16.h\n" + ".inst 0x6470e64d // bfmmla z13.s, z18.h, z16.h\n" "ld1h { z16.h }, p5/Z, [x10, #5, MUL VL]\n" - ".inst 0x6471e66a // bfmmla z10.s, z19.h, z17.h\n" + ".inst 0x6471e64a // bfmmla z10.s, z18.h, z17.h\n" "ld1h { z17.h }, p5/Z, [x10, #6, MUL VL]\n" - ".inst 0x6470e66e // bfmmla z14.s, z19.h, z16.h\n" + ".inst 0x6470e64e // bfmmla z14.s, z18.h, z16.h\n" "ld1h { z16.h }, p5/Z, [x10, #7, MUL VL]\n" - "sub x27, x27, #0x4\n" - "cmp x27, #0x4\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" + ".inst 0x6471e64b // bfmmla z11.s, z18.h, z17.h\n" + ".inst 0x6470e64f // bfmmla z15.s, z18.h, z16.h\n" "bgt 22b\n" "23:" // Height 2: Multiply loop: Single iteration only "whilelt p0.s, XZR, x27\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" + "ld1h { z19.h }, p5/Z, [x10, #1, MUL VL]\n" + "ld1rqw { z18.s }, p0/Z, [x26]\n" + "ld1rqw { z16.s }, p0/Z, [x25]\n" + ".inst 0x658ab652 // bfcvt z18.h, p5/M, z18.s\n" + ".inst 0x658ab610 // bfcvt z16.h, p5/M, z16.s\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" + "uzp1 z16.h, z16.h, z16.h\n" + "trn1 z18.d, z18.d, z16.d\n" + ".inst 0x6471e648 // bfmmla z8.s, z18.h, z17.h\n" "ld1h { z17.h }, p5/Z, [x10, #2, MUL VL]\n" - ".inst 0x6470e66c // bfmmla z12.s, z19.h, z16.h\n" + ".inst 0x6473e64c // bfmmla z12.s, z18.h, z19.h\n" "ld1h { z16.h }, p5/Z, [x10, #3, MUL VL]\n" - ".inst 0x6471e669 // bfmmla z9.s, z19.h, z17.h\n" + ".inst 0x6471e649 // bfmmla z9.s, z18.h, z17.h\n" "ld1h { z17.h }, p5/Z, [x10, #4, MUL VL]\n" - ".inst 0x6470e66d // bfmmla z13.s, z19.h, z16.h\n" + ".inst 0x6470e64d // bfmmla z13.s, z18.h, z16.h\n" "ld1h { z16.h }, p5/Z, [x10, #5, MUL VL]\n" - ".inst 0x6471e66a // bfmmla z10.s, z19.h, z17.h\n" + ".inst 0x6471e64a // bfmmla z10.s, z18.h, z17.h\n" "ld1h { z17.h }, p5/Z, [x10, #6, MUL VL]\n" - ".inst 0x6470e66e // bfmmla z14.s, z19.h, z16.h\n" + ".inst 0x6470e64e // bfmmla z14.s, z18.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 0x6470e66f // bfmmla z15.s, z19.h, z16.h\n" + ".inst 0x6471e64b // bfmmla z11.s, z18.h, z17.h\n" + ".inst 0x6470e64f // bfmmla z15.s, z18.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" @@ -401,17 +403,17 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" "uzp1 z6.d, z8.d, z12.d\n" "uzp2 z8.d, z8.d, z12.d\n" - "add x25, x9, x20, LSL #2\n" "uzp1 z12.d, z9.d, z13.d\n" "uzp2 z9.d, z9.d, z13.d\n" "uzp1 z13.d, z10.d, z14.d\n" "uzp2 z10.d, z10.d, z14.d\n" "uzp1 z14.d, z11.d, z15.d\n" "uzp2 z11.d, z11.d, z15.d\n" + "add x26, x9, x20, LSL #2\n" "tbz %x[flags], #1, 25f\n" - "add x20, %x[args_ptr], %[offset_max]\n" - "ld1rw { z17.s }, p5/Z, [x20]\n" + "add x21, %x[args_ptr], %[offset_max]\n" "add x20, %x[args_ptr], %[offset_min]\n" + "ld1rw { z17.s }, p5/Z, [x21]\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" @@ -435,20 +437,20 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "st1w { z13.s }, p2, [x9, #2, MUL VL]\n" "st1w { z14.s }, p1, [x9, #3, MUL VL]\n" "addvl x9, x9, #4\n" - "st1w { z8.s }, p4, [x25]\n" - "st1w { z9.s }, p3, [x25, #1, MUL VL]\n" - "st1w { z10.s }, p2, [x25, #2, MUL VL]\n" - "st1w { z11.s }, p1, [x25, #3, MUL VL]\n" + "st1w { z8.s }, p4, [x26]\n" + "st1w { z9.s }, p3, [x26, #1, MUL VL]\n" + "st1w { z10.s }, p2, [x26, #2, MUL VL]\n" + "st1w { z11.s }, p1, [x26, #3, MUL VL]\n" "26:" // Height 2: Writeback done "decw x11, ALL, MUL #4\n" "cmp x11, XZR\n" "bgt 15b\n" "b 80f\n" "27:" // Height 3 - "mov x12, %x[bias]\n" + "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n" "ldr x11, [%x[args_ptr], %[offsetof_N]]\n" "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x9, %x[output_ptr]\n" + "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n" "28:" // Height 3: Column loop "mov x20, #0x0\n" "whilelt p4.s, x20, x11\n" @@ -461,15 +463,15 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "cbz x12, 29f\n" "ld1w { z8.s }, p5/Z, [x12]\n" "ld1w { z9.s }, p5/Z, [x12, #1, MUL VL]\n" - "zip2 z12.d, z8.d, z8.d\n" - "zip1 z8.d, z8.d, z8.d\n" "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n" "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n" + "addvl x12, x12, #4\n" + "zip2 z12.d, z8.d, z8.d\n" + "zip1 z8.d, z8.d, z8.d\n" "zip2 z13.d, z9.d, z9.d\n" "zip1 z9.d, z9.d, z9.d\n" "zip2 z14.d, z10.d, z10.d\n" "zip1 z10.d, z10.d, z10.d\n" - "addvl x12, x12, #4\n" "zip2 z15.d, z11.d, z11.d\n" "zip1 z11.d, z11.d, z11.d\n" "mov z16.d, z8.d\n" @@ -484,28 +486,28 @@ 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" + "ld1w { z24.s }, p4/Z, [x9]\n" + "ld1w { z26.s }, p3/Z, [x9, #1, MUL VL]\n" + "ld1w { z25.s }, p2/Z, [x9, #2, MUL VL]\n" + "ld1w { z16.s }, p1/Z, [x9, #3, 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, [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 z8.d, z24.d, z12.d\n" + "zip2 z12.d, z24.d, z12.d\n" + "ld1w { z24.s }, p1/Z, [x20, #3, MUL VL]\n" + "zip1 z9.d, z26.d, z13.d\n" + "zip2 z13.d, z26.d, z13.d\n" + "zip1 z10.d, z25.d, z14.d\n" + "zip2 z14.d, z25.d, z14.d\n" "zip1 z11.d, z16.d, z15.d\n" "zip2 z15.d, z16.d, z15.d\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" @@ -536,8 +538,8 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "mov x28, #0x0\n" "32:" // Height 3: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr w27, [x20, x28, LSL #0x2]\n" "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr w27, [x20, x28, LSL #0x2]\n" "tbz %x[flags], #3, 33f\n" "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -559,83 +561,83 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "ble 36f\n" "35:" // Height 3: Multiply loop: Main loop head "whilelt p0.s, XZR, x27\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" + "ld1h { z25.h }, p5/Z, [x10]\n" + "ld1h { z28.h }, p5/Z, [x10, #1, MUL VL]\n" + "sub x27, x27, #0x4\n" + "cmp x27, #0x4\n" + "ld1rqw { z27.s }, p0/Z, [x26]\n" + "add x26, x26, #0x10\n" + "ld1rqw { z24.s }, p0/Z, [x25]\n" + "add x25, x25, #0x10\n" "ld1rqw { z26.s }, p0/Z, [x24]\n" + "add x24, x24, #0x10\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" + ".inst 0x658ab718 // bfcvt z24.h, p5/M, z24.s\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 z28.d, z28.d, z27.d\n" + "uzp1 z24.h, z24.h, z24.h\n" + "trn1 z27.d, z27.d, z24.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 0x6479e768 // bfmmla z8.s, z27.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" - "add x26, x26, #0x10\n" - ".inst 0x6478e754 // bfmmla z20.s, z26.h, z24.h\n" + ".inst 0x647ce76c // bfmmla z12.s, z27.h, z28.h\n" + ".inst 0x647ce754 // bfmmla z20.s, z26.h, z28.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 0x6479e769 // bfmmla z9.s, z27.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" - "add x24, x24, #0x10\n" + ".inst 0x6478e76d // bfmmla z13.s, z27.h, z24.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 0x6479e76a // bfmmla z10.s, z27.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 0x6478e76e // bfmmla z14.s, z27.h, z24.h\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 0x6479e76b // bfmmla z11.s, z27.h, z25.h\n" ".inst 0x6479e753 // bfmmla z19.s, z26.h, z25.h\n" - ".inst 0x6478e78f // bfmmla z15.s, z28.h, z24.h\n" + ".inst 0x6478e76f // bfmmla z15.s, z27.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 { z28.s }, p0/Z, [x26]\n" - "ld1rqw { z27.s }, p0/Z, [x25]\n" - ".inst 0x658ab79c // bfcvt z28.h, p5/M, z28.s\n" + "ld1h { z25.h }, p5/Z, [x10]\n" + "ld1h { z28.h }, p5/Z, [x10, #1, MUL VL]\n" + "ld1rqw { z27.s }, p0/Z, [x26]\n" + "ld1rqw { z24.s }, p0/Z, [x25]\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" + ".inst 0x658ab718 // bfcvt z24.h, p5/M, z24.s\n" "uzp1 z27.h, z27.h, z27.h\n" + "uzp1 z24.h, z24.h, z24.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" + "trn1 z27.d, z27.d, z24.d\n" "uzp1 z26.h, z26.h, z26.h\n" - ".inst 0x6479e788 // bfmmla z8.s, z28.h, z25.h\n" + ".inst 0x6479e768 // bfmmla z8.s, z27.h, z25.h\n" + ".inst 0x647ce76c // bfmmla z12.s, z27.h, z28.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" + ".inst 0x647ce754 // bfmmla z20.s, z26.h, z28.h\n" "ld1h { z24.h }, p5/Z, [x10, #3, MUL VL]\n" - ".inst 0x6479e789 // bfmmla z9.s, z28.h, z25.h\n" + ".inst 0x6479e769 // bfmmla z9.s, z27.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 0x6478e76d // bfmmla z13.s, z27.h, z24.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 0x6479e76a // bfmmla z10.s, z27.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 0x6478e76e // bfmmla z14.s, z27.h, z24.h\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 0x6479e76b // bfmmla z11.s, z27.h, z25.h\n" ".inst 0x6479e753 // bfmmla z19.s, z26.h, z25.h\n" - ".inst 0x6478e78f // bfmmla z15.s, z28.h, z24.h\n" + ".inst 0x6478e76f // bfmmla z15.s, z27.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" @@ -643,24 +645,24 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "cmp x28, x20\n" "bne 32b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x9, x20, LSL #2\n" "uzp1 z6.d, z8.d, z12.d\n" "uzp2 z8.d, z8.d, z12.d\n" "uzp1 z12.d, z9.d, z13.d\n" "uzp2 z9.d, z9.d, z13.d\n" - "add x24, x25, x20, LSL #2\n" "uzp1 z13.d, z10.d, z14.d\n" "uzp2 z10.d, z10.d, z14.d\n" + "add x26, x9, x20, LSL #2\n" "uzp1 z14.d, z11.d, z15.d\n" "uzp2 z11.d, z11.d, z15.d\n" + "add x25, x26, x20, LSL #2\n" "uzp1 z16.d, z16.d, z20.d\n" "uzp1 z17.d, z17.d, z21.d\n" "uzp1 z18.d, z18.d, z22.d\n" "uzp1 z19.d, z19.d, z23.d\n" "tbz %x[flags], #1, 38f\n" - "add x20, %x[args_ptr], %[offset_max]\n" - "ld1rw { z25.s }, p5/Z, [x20]\n" + "add x21, %x[args_ptr], %[offset_max]\n" "add x20, %x[args_ptr], %[offset_min]\n" + "ld1rw { z25.s }, p5/Z, [x21]\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" @@ -692,24 +694,24 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "st1w { z13.s }, p2, [x9, #2, MUL VL]\n" "st1w { z14.s }, p1, [x9, #3, MUL VL]\n" "addvl x9, x9, #4\n" - "st1w { z8.s }, p4, [x25]\n" - "st1w { z9.s }, p3, [x25, #1, MUL VL]\n" - "st1w { z10.s }, p2, [x25, #2, MUL VL]\n" - "st1w { z11.s }, p1, [x25, #3, MUL VL]\n" - "st1w { z16.s }, p4, [x24]\n" - "st1w { z17.s }, p3, [x24, #1, MUL VL]\n" - "st1w { z18.s }, p2, [x24, #2, MUL VL]\n" - "st1w { z19.s }, p1, [x24, #3, MUL VL]\n" + "st1w { z8.s }, p4, [x26]\n" + "st1w { z9.s }, p3, [x26, #1, MUL VL]\n" + "st1w { z10.s }, p2, [x26, #2, MUL VL]\n" + "st1w { z11.s }, p1, [x26, #3, MUL VL]\n" + "st1w { z16.s }, p4, [x25]\n" + "st1w { z17.s }, p3, [x25, #1, MUL VL]\n" + "st1w { z18.s }, p2, [x25, #2, MUL VL]\n" + "st1w { z19.s }, p1, [x25, #3, MUL VL]\n" "39:" // Height 3: Writeback done "decw x11, ALL, MUL #4\n" "cmp x11, XZR\n" "bgt 28b\n" "b 80f\n" "40:" // Height 4 - "mov x12, %x[bias]\n" + "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n" "ldr x11, [%x[args_ptr], %[offsetof_N]]\n" "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x9, %x[output_ptr]\n" + "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n" "41:" // Height 4: Column loop "mov x20, #0x0\n" "whilelt p4.s, x20, x11\n" @@ -722,15 +724,15 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "cbz x12, 42f\n" "ld1w { z8.s }, p5/Z, [x12]\n" "ld1w { z9.s }, p5/Z, [x12, #1, MUL VL]\n" - "zip2 z12.d, z8.d, z8.d\n" - "zip1 z8.d, z8.d, z8.d\n" "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n" "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n" + "addvl x12, x12, #4\n" + "zip2 z12.d, z8.d, z8.d\n" + "zip1 z8.d, z8.d, z8.d\n" "zip2 z13.d, z9.d, z9.d\n" "zip1 z9.d, z9.d, z9.d\n" "zip2 z14.d, z10.d, z10.d\n" "zip1 z10.d, z10.d, z10.d\n" - "addvl x12, x12, #4\n" "zip2 z15.d, z11.d, z11.d\n" "zip1 z11.d, z11.d, z11.d\n" "mov z16.d, z8.d\n" @@ -745,37 +747,37 @@ 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" + "ld1w { z20.s }, p4/Z, [x9]\n" + "ld1w { z22.s }, p3/Z, [x9, #1, MUL VL]\n" + "ld1w { z25.s }, p2/Z, [x9, #2, MUL VL]\n" + "ld1w { z16.s }, p1/Z, [x9, #3, 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, [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, [x21, #3, MUL VL]\n" + "zip1 z8.d, z20.d, z12.d\n" + "zip2 z12.d, z20.d, z12.d\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, [x20, #1, MUL VL]\n" + "zip1 z9.d, z22.d, z13.d\n" + "zip2 z13.d, z22.d, z13.d\n" "ld1w { z22.s }, p2/Z, [x20, #2, MUL VL]\n" + "ld1w { z23.s }, p1/Z, [x20, #3, MUL VL]\n" + "zip1 z10.d, z25.d, z14.d\n" + "zip2 z14.d, z25.d, z14.d\n" + "zip1 z11.d, z16.d, z15.d\n" + "zip2 z15.d, z16.d, z15.d\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" "zip2 z21.d, z18.d, z21.d\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" @@ -802,8 +804,8 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "mov x28, #0x0\n" "45:" // Height 4: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr w27, [x20, x28, LSL #0x2]\n" "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr w27, [x20, x28, LSL #0x2]\n" "tbz %x[flags], #3, 46f\n" "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -828,110 +830,110 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "ble 49f\n" "48:" // Height 4: Multiply loop: Main loop head "whilelt p0.s, XZR, x27\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" + "ld1h { z29.h }, p5/Z, [x10]\n" + "ld1h { z28.h }, p5/Z, [x10, #1, MUL VL]\n" + "sub x27, x27, #0x4\n" + "cmp x27, #0x4\n" + "ld1rqw { z27.s }, p0/Z, [x26]\n" + "add x26, x26, #0x10\n" + "ld1rqw { z25.s }, p0/Z, [x25]\n" + "add x25, x25, #0x10\n" + "ld1rqw { z26.s }, p0/Z, [x24]\n" + "add x24, x24, #0x10\n" + "ld1rqw { z24.s }, p0/Z, [x23]\n" + "add x23, x23, #0x10\n" ".inst 0x658ab77b // bfcvt z27.h, p5/M, z27.s\n" + ".inst 0x658ab739 // bfcvt z25.h, p5/M, z25.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" + ".inst 0x658ab718 // bfcvt z24.h, p5/M, z24.s\n" "uzp1 z27.h, z27.h, z27.h\n" - "sub x27, x27, #0x4\n" - "cmp x27, #0x4\n" + "uzp1 z25.h, z25.h, z25.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 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" + "uzp1 z24.h, z24.h, z24.h\n" + "trn1 z27.d, z27.d, z25.d\n" + "trn1 z26.d, z26.d, z24.d\n" + ".inst 0x647de768 // bfmmla z8.s, z27.h, z29.h\n" + ".inst 0x647de750 // bfmmla z16.s, z26.h, z29.h\n" "ld1h { z25.h }, p5/Z, [x10, #2, MUL VL]\n" - ".inst 0x6478e774 // bfmmla z20.s, z27.h, z24.h\n" + ".inst 0x647ce76c // bfmmla z12.s, z27.h, z28.h\n" + ".inst 0x647ce754 // bfmmla z20.s, z26.h, z28.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 0x6479e771 // bfmmla z17.s, z27.h, z25.h\n" - ".inst 0x6478e7ad // bfmmla z13.s, z29.h, z24.h\n" + ".inst 0x6479e769 // bfmmla z9.s, z27.h, z25.h\n" + ".inst 0x6479e751 // bfmmla z17.s, z26.h, z25.h\n" "ld1h { z25.h }, p5/Z, [x10, #4, MUL VL]\n" - "add x24, x24, #0x10\n" - ".inst 0x6478e775 // bfmmla z21.s, z27.h, z24.h\n" + ".inst 0x6478e76d // bfmmla z13.s, z27.h, z24.h\n" + ".inst 0x6478e755 // bfmmla z21.s, z26.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 0x6479e772 // bfmmla z18.s, z27.h, z25.h\n" - ".inst 0x6478e7ae // bfmmla z14.s, z29.h, z24.h\n" + ".inst 0x6479e76a // bfmmla z10.s, z27.h, z25.h\n" + ".inst 0x6479e752 // bfmmla z18.s, z26.h, z25.h\n" "ld1h { z25.h }, p5/Z, [x10, #6, MUL VL]\n" - ".inst 0x6478e776 // bfmmla z22.s, z27.h, z24.h\n" + ".inst 0x6478e76e // bfmmla z14.s, z27.h, z24.h\n" + ".inst 0x6478e756 // bfmmla z22.s, z26.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 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" + ".inst 0x6479e76b // bfmmla z11.s, z27.h, z25.h\n" + ".inst 0x6479e753 // bfmmla z19.s, z26.h, z25.h\n" + ".inst 0x6478e76f // bfmmla z15.s, z27.h, z24.h\n" + ".inst 0x6478e757 // bfmmla z23.s, z26.h, z24.h\n" "bgt 48b\n" "49:" // Height 4: Multiply loop: Single iteration only "whilelt p0.s, XZR, x27\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" + "ld1h { z29.h }, p5/Z, [x10]\n" + "ld1h { z28.h }, p5/Z, [x10, #1, MUL VL]\n" + "ld1rqw { z27.s }, p0/Z, [x26]\n" + "ld1rqw { z25.s }, p0/Z, [x25]\n" + "ld1rqw { z26.s }, p0/Z, [x24]\n" + "ld1rqw { z24.s }, p0/Z, [x23]\n" ".inst 0x658ab77b // bfcvt z27.h, p5/M, z27.s\n" + ".inst 0x658ab739 // bfcvt z25.h, p5/M, z25.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" + ".inst 0x658ab718 // bfcvt z24.h, p5/M, z24.s\n" "uzp1 z27.h, z27.h, z27.h\n" + "uzp1 z25.h, z25.h, z25.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" + "uzp1 z24.h, z24.h, z24.h\n" + "trn1 z27.d, z27.d, z25.d\n" + "trn1 z26.d, z26.d, z24.d\n" + ".inst 0x647de768 // bfmmla z8.s, z27.h, z29.h\n" + ".inst 0x647de750 // bfmmla z16.s, z26.h, z29.h\n" "ld1h { z25.h }, p5/Z, [x10, #2, MUL VL]\n" - ".inst 0x6478e774 // bfmmla z20.s, z27.h, z24.h\n" + ".inst 0x647ce76c // bfmmla z12.s, z27.h, z28.h\n" + ".inst 0x647ce754 // bfmmla z20.s, z26.h, z28.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" + ".inst 0x6479e769 // bfmmla z9.s, z27.h, z25.h\n" + ".inst 0x6479e751 // bfmmla z17.s, z26.h, z25.h\n" "ld1h { z25.h }, p5/Z, [x10, #4, MUL VL]\n" - ".inst 0x6478e775 // bfmmla z21.s, z27.h, z24.h\n" + ".inst 0x6478e76d // bfmmla z13.s, z27.h, z24.h\n" + ".inst 0x6478e755 // bfmmla z21.s, z26.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" + ".inst 0x6479e76a // bfmmla z10.s, z27.h, z25.h\n" + ".inst 0x6479e752 // bfmmla z18.s, z26.h, z25.h\n" "ld1h { z25.h }, p5/Z, [x10, #6, MUL VL]\n" - ".inst 0x6478e776 // bfmmla z22.s, z27.h, z24.h\n" + ".inst 0x6478e76e // bfmmla z14.s, z27.h, z24.h\n" + ".inst 0x6478e756 // bfmmla z22.s, z26.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 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" + ".inst 0x6479e76b // bfmmla z11.s, z27.h, z25.h\n" + ".inst 0x6479e753 // bfmmla z19.s, z26.h, z25.h\n" + ".inst 0x6478e76f // bfmmla z15.s, z27.h, z24.h\n" + ".inst 0x6478e757 // bfmmla z23.s, z26.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" "cmp x28, x20\n" "bne 45b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x9, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" "uzp1 z6.d, z8.d, z12.d\n" "uzp2 z8.d, z8.d, z12.d\n" "uzp1 z12.d, z9.d, z13.d\n" - "add x23, x24, x20, LSL #2\n" "uzp2 z9.d, z9.d, z13.d\n" "uzp1 z13.d, z10.d, z14.d\n" "uzp2 z10.d, z10.d, z14.d\n" + "add x26, x9, x20, LSL #2\n" + "add x25, x26, x20, LSL #2\n" "uzp1 z14.d, z11.d, z15.d\n" "uzp2 z11.d, z11.d, z15.d\n" + "add x24, x25, x20, LSL #2\n" "uzp1 z15.d, z16.d, z20.d\n" "uzp2 z16.d, z16.d, z20.d\n" "uzp1 z20.d, z17.d, z21.d\n" @@ -941,9 +943,9 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "uzp1 z22.d, z19.d, z23.d\n" "uzp2 z19.d, z19.d, z23.d\n" "tbz %x[flags], #1, 51f\n" - "add x20, %x[args_ptr], %[offset_max]\n" - "ld1rw { z24.s }, p5/Z, [x20]\n" + "add x21, %x[args_ptr], %[offset_max]\n" "add x20, %x[args_ptr], %[offset_min]\n" + "ld1rw { z24.s }, p5/Z, [x21]\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" @@ -983,28 +985,28 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "st1w { z13.s }, p2, [x9, #2, MUL VL]\n" "st1w { z14.s }, p1, [x9, #3, MUL VL]\n" "addvl x9, x9, #4\n" - "st1w { z8.s }, p4, [x25]\n" - "st1w { z9.s }, p3, [x25, #1, MUL VL]\n" - "st1w { z10.s }, p2, [x25, #2, MUL VL]\n" - "st1w { z11.s }, p1, [x25, #3, MUL VL]\n" - "st1w { z15.s }, p4, [x24]\n" - "st1w { z20.s }, p3, [x24, #1, MUL VL]\n" - "st1w { z21.s }, p2, [x24, #2, MUL VL]\n" - "st1w { z22.s }, p1, [x24, #3, MUL VL]\n" - "st1w { z16.s }, p4, [x23]\n" - "st1w { z17.s }, p3, [x23, #1, MUL VL]\n" - "st1w { z18.s }, p2, [x23, #2, MUL VL]\n" - "st1w { z19.s }, p1, [x23, #3, MUL VL]\n" + "st1w { z8.s }, p4, [x26]\n" + "st1w { z9.s }, p3, [x26, #1, MUL VL]\n" + "st1w { z10.s }, p2, [x26, #2, MUL VL]\n" + "st1w { z11.s }, p1, [x26, #3, MUL VL]\n" + "st1w { z15.s }, p4, [x25]\n" + "st1w { z20.s }, p3, [x25, #1, MUL VL]\n" + "st1w { z21.s }, p2, [x25, #2, MUL VL]\n" + "st1w { z22.s }, p1, [x25, #3, MUL VL]\n" + "st1w { z16.s }, p4, [x24]\n" + "st1w { z17.s }, p3, [x24, #1, MUL VL]\n" + "st1w { z18.s }, p2, [x24, #2, MUL VL]\n" + "st1w { z19.s }, p1, [x24, #3, MUL VL]\n" "52:" // Height 4: Writeback done "decw x11, ALL, MUL #4\n" "cmp x11, XZR\n" "bgt 41b\n" "b 80f\n" "53:" // Height 5 - "mov x12, %x[bias]\n" + "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n" "ldr x11, [%x[args_ptr], %[offsetof_N]]\n" "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x9, %x[output_ptr]\n" + "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n" "54:" // Height 5: Column loop "mov x20, #0x0\n" "whilelt p4.s, x20, x11\n" @@ -1017,15 +1019,15 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "cbz x12, 55f\n" "ld1w { z8.s }, p5/Z, [x12]\n" "ld1w { z9.s }, p5/Z, [x12, #1, MUL VL]\n" - "zip2 z12.d, z8.d, z8.d\n" - "zip1 z8.d, z8.d, z8.d\n" "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n" "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n" + "addvl x12, x12, #4\n" + "zip2 z12.d, z8.d, z8.d\n" + "zip1 z8.d, z8.d, z8.d\n" "zip2 z13.d, z9.d, z9.d\n" "zip1 z9.d, z9.d, z9.d\n" "zip2 z14.d, z10.d, z10.d\n" "zip1 z10.d, z10.d, z10.d\n" - "addvl x12, x12, #4\n" "zip2 z15.d, z11.d, z11.d\n" "zip1 z11.d, z11.d, z11.d\n" "mov z16.d, z8.d\n" @@ -1048,46 +1050,46 @@ 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" + "ld1w { z20.s }, p4/Z, [x9]\n" + "ld1w { z22.s }, p3/Z, [x9, #1, MUL VL]\n" + "ld1w { z25.s }, p2/Z, [x9, #2, MUL VL]\n" + "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n" "add x23, x9, x20, LSL #2\n" "add x22, x23, x20, LSL #2\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, [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" + "add x20, x21, x20, LSL #2\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, [x22, #3, MUL VL]\n" + "zip1 z8.d, z20.d, z12.d\n" + "zip2 z12.d, z20.d, z12.d\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, [x21, #1, MUL VL]\n" + "zip1 z9.d, z22.d, z13.d\n" + "zip2 z13.d, z22.d, z13.d\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, [x21, #3, MUL VL]\n" + "zip1 z10.d, z25.d, z14.d\n" + "zip2 z14.d, z25.d, z14.d\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, [x20, #1, MUL VL]\n" + "zip1 z11.d, z16.d, z15.d\n" + "zip2 z15.d, z16.d, z15.d\n" "ld1w { z27.s }, p2/Z, [x20, #2, MUL VL]\n" + "ld1w { z0.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" + "zip2 z21.d, z18.d, z21.d\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" "zip2 z23.d, z24.d, z23.d\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" @@ -1126,8 +1128,8 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "mov x28, #0x0\n" "58:" // Height 5: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr w27, [x20, x28, LSL #0x2]\n" "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr w27, [x20, x28, LSL #0x2]\n" "tbz %x[flags], #3, 59f\n" "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -1155,115 +1157,115 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "ble 62f\n" "61:" // Height 5: Multiply loop: Main loop head "whilelt p0.s, XZR, x27\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 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 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" + "ld1h { z6.h }, p5/Z, [x10]\n" + "ld1h { z5.h }, p5/Z, [x10, #1, MUL VL]\n" "sub x27, x27, #0x4\n" - "uzp1 z3.h, z3.h, z3.h\n" - ".inst 0x658ab442 // bfcvt z2.h, p5/M, z2.s\n" "cmp x27, #0x4\n" + "ld1rqw { z4.s }, p0/Z, [x26]\n" "add x26, x26, #0x10\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" + "ld1rqw { z1.s }, p0/Z, [x25]\n" "add x25, x25, #0x10\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" + "ld1rqw { z3.s }, p0/Z, [x24]\n" "add x24, x24, #0x10\n" + "ld1rqw { z0.s }, p0/Z, [x23]\n" + "ld1rqw { z2.s }, p0/Z, [x22]\n" "add x23, x23, #0x10\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 0x6461e491 // bfmmla z17.s, z4.h, z1.h\n" + ".inst 0x658ab484 // bfcvt z4.h, p5/M, z4.s\n" + ".inst 0x658ab421 // bfcvt z1.h, p5/M, z1.s\n" + ".inst 0x658ab463 // bfcvt z3.h, p5/M, z3.s\n" + ".inst 0x658ab400 // bfcvt z0.h, p5/M, z0.s\n" + "uzp1 z4.h, z4.h, z4.h\n" + ".inst 0x658ab442 // bfcvt z2.h, p5/M, z2.s\n" + "uzp1 z1.h, z1.h, z1.h\n" + "uzp1 z3.h, z3.h, z3.h\n" + "uzp1 z0.h, z0.h, z0.h\n" + "trn1 z4.d, z4.d, z1.d\n" + "uzp1 z2.h, z2.h, z2.h\n" + "trn1 z3.d, z3.d, z0.d\n" + ".inst 0x6466e488 // bfmmla z8.s, z4.h, z6.h\n" + ".inst 0x6466e458 // bfmmla z24.s, z2.h, z6.h\n" + ".inst 0x6465e48c // bfmmla z12.s, z4.h, z5.h\n" + ".inst 0x6465e45c // bfmmla z28.s, z2.h, z5.h\n" + ".inst 0x6466e470 // bfmmla z16.s, z3.h, z6.h\n" + "ld1h { z1.h }, p5/Z, [x10, #2, MUL VL]\n" + ".inst 0x6465e474 // bfmmla z20.s, z3.h, z5.h\n" + "ld1h { z0.h }, p5/Z, [x10, #3, MUL VL]\n" + ".inst 0x6461e489 // bfmmla z9.s, z4.h, z1.h\n" + ".inst 0x6461e471 // bfmmla z17.s, z3.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 0x6460e48d // bfmmla z13.s, z4.h, z0.h\n" + ".inst 0x6460e475 // bfmmla z21.s, z3.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 0x6461e48a // bfmmla z10.s, z4.h, z1.h\n" + ".inst 0x6461e472 // bfmmla z18.s, z3.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 0x6460e48e // bfmmla z14.s, z4.h, z0.h\n" + ".inst 0x6460e476 // bfmmla z22.s, z3.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 0x6461e493 // bfmmla z19.s, z4.h, z1.h\n" + ".inst 0x6461e48b // bfmmla z11.s, z4.h, z1.h\n" + ".inst 0x6461e473 // bfmmla z19.s, z3.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 0x6460e48f // bfmmla z15.s, z4.h, z0.h\n" + ".inst 0x6460e477 // bfmmla z23.s, z3.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 { 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 0x658ab4a5 // bfcvt z5.h, p5/M, z5.s\n" - ".inst 0x658ab484 // bfcvt z4.h, p5/M, z4.s\n" + "ld1h { z6.h }, p5/Z, [x10]\n" + "ld1h { z5.h }, p5/Z, [x10, #1, MUL VL]\n" + "ld1rqw { z4.s }, p0/Z, [x26]\n" + "ld1rqw { z1.s }, p0/Z, [x25]\n" + "ld1rqw { z3.s }, p0/Z, [x24]\n" + "ld1rqw { z0.s }, p0/Z, [x23]\n" "ld1rqw { z2.s }, p0/Z, [x22]\n" + ".inst 0x658ab484 // bfcvt z4.h, p5/M, z4.s\n" + ".inst 0x658ab421 // bfcvt z1.h, p5/M, z1.s\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" + ".inst 0x658ab400 // bfcvt z0.h, p5/M, z0.s\n" "uzp1 z4.h, z4.h, z4.h\n" - "ld1h { z0.h }, p5/Z, [x10, #1, MUL VL]\n" - "uzp1 z3.h, z3.h, z3.h\n" + "uzp1 z1.h, z1.h, z1.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 z3.h, z3.h, z3.h\n" + "uzp1 z0.h, z0.h, z0.h\n" + "trn1 z4.d, z4.d, z1.d\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" + "trn1 z3.d, z3.d, z0.d\n" + ".inst 0x6466e488 // bfmmla z8.s, z4.h, z6.h\n" + ".inst 0x6465e48c // bfmmla z12.s, z4.h, z5.h\n" + ".inst 0x6466e470 // bfmmla z16.s, z3.h, z6.h\n" + ".inst 0x6466e458 // bfmmla z24.s, z2.h, z6.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" + ".inst 0x6465e474 // bfmmla z20.s, z3.h, z5.h\n" + ".inst 0x6465e45c // bfmmla z28.s, z2.h, z5.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 0x6461e489 // bfmmla z9.s, z4.h, z1.h\n" + ".inst 0x6461e471 // bfmmla z17.s, z3.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 0x6460e48d // bfmmla z13.s, z4.h, z0.h\n" + ".inst 0x6460e475 // bfmmla z21.s, z3.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 0x6461e48a // bfmmla z10.s, z4.h, z1.h\n" + ".inst 0x6461e472 // bfmmla z18.s, z3.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 0x6460e48e // bfmmla z14.s, z4.h, z0.h\n" + ".inst 0x6460e476 // bfmmla z22.s, z3.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 0x6461e493 // bfmmla z19.s, z4.h, z1.h\n" + ".inst 0x6461e48b // bfmmla z11.s, z4.h, z1.h\n" + ".inst 0x6461e473 // bfmmla z19.s, z3.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 0x6460e48f // bfmmla z15.s, z4.h, z0.h\n" + ".inst 0x6460e477 // bfmmla z23.s, z3.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" @@ -1271,20 +1273,20 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "cmp x28, x20\n" "bne 58b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x9, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" "uzp1 z6.d, z8.d, z12.d\n" - "add x23, x24, x20, LSL #2\n" "uzp2 z8.d, z8.d, z12.d\n" "uzp1 z12.d, z9.d, z13.d\n" - "add x22, x23, x20, LSL #2\n" "uzp2 z9.d, z9.d, z13.d\n" "uzp1 z13.d, z10.d, z14.d\n" "uzp2 z10.d, z10.d, z14.d\n" + "add x26, x9, x20, LSL #2\n" + "add x25, x26, x20, LSL #2\n" + "add x24, x25, x20, LSL #2\n" "uzp1 z14.d, z11.d, z15.d\n" "uzp2 z11.d, z11.d, z15.d\n" "uzp1 z15.d, z16.d, z20.d\n" "uzp2 z16.d, z16.d, z20.d\n" + "add x23, x24, x20, LSL #2\n" "uzp1 z20.d, z17.d, z21.d\n" "uzp2 z17.d, z17.d, z21.d\n" "uzp1 z21.d, z18.d, z22.d\n" @@ -1296,9 +1298,9 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "uzp1 z26.d, z26.d, z30.d\n" "uzp1 z27.d, z27.d, z31.d\n" "tbz %x[flags], #1, 64f\n" - "add x20, %x[args_ptr], %[offset_max]\n" - "ld1rw { z0.s }, p5/Z, [x20]\n" + "add x21, %x[args_ptr], %[offset_max]\n" "add x20, %x[args_ptr], %[offset_min]\n" + "ld1rw { z0.s }, p5/Z, [x21]\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" @@ -1346,22 +1348,22 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "st1w { z13.s }, p2, [x9, #2, MUL VL]\n" "st1w { z14.s }, p1, [x9, #3, MUL VL]\n" "addvl x9, x9, #4\n" - "st1w { z8.s }, p4, [x25]\n" - "st1w { z9.s }, p3, [x25, #1, MUL VL]\n" - "st1w { z10.s }, p2, [x25, #2, MUL VL]\n" - "st1w { z11.s }, p1, [x25, #3, MUL VL]\n" - "st1w { z15.s }, p4, [x24]\n" - "st1w { z20.s }, p3, [x24, #1, MUL VL]\n" - "st1w { z21.s }, p2, [x24, #2, MUL VL]\n" - "st1w { z22.s }, p1, [x24, #3, MUL VL]\n" - "st1w { z16.s }, p4, [x23]\n" - "st1w { z17.s }, p3, [x23, #1, MUL VL]\n" - "st1w { z18.s }, p2, [x23, #2, MUL VL]\n" - "st1w { z19.s }, p1, [x23, #3, MUL VL]\n" - "st1w { z24.s }, p4, [x22]\n" - "st1w { z25.s }, p3, [x22, #1, MUL VL]\n" - "st1w { z26.s }, p2, [x22, #2, MUL VL]\n" - "st1w { z27.s }, p1, [x22, #3, MUL VL]\n" + "st1w { z8.s }, p4, [x26]\n" + "st1w { z9.s }, p3, [x26, #1, MUL VL]\n" + "st1w { z10.s }, p2, [x26, #2, MUL VL]\n" + "st1w { z11.s }, p1, [x26, #3, MUL VL]\n" + "st1w { z15.s }, p4, [x25]\n" + "st1w { z20.s }, p3, [x25, #1, MUL VL]\n" + "st1w { z21.s }, p2, [x25, #2, MUL VL]\n" + "st1w { z22.s }, p1, [x25, #3, MUL VL]\n" + "st1w { z16.s }, p4, [x24]\n" + "st1w { z17.s }, p3, [x24, #1, MUL VL]\n" + "st1w { z18.s }, p2, [x24, #2, MUL VL]\n" + "st1w { z19.s }, p1, [x24, #3, MUL VL]\n" + "st1w { z24.s }, p4, [x23]\n" + "st1w { z25.s }, p3, [x23, #1, MUL VL]\n" + "st1w { z26.s }, p2, [x23, #2, MUL VL]\n" + "st1w { z27.s }, p1, [x23, #3, MUL VL]\n" "65:" // Height 5: Writeback done "decw x11, ALL, MUL #4\n" "cmp x11, XZR\n" @@ -1369,12 +1371,13 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "b 80f\n" "66:" // Height 6 "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n" + "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n" "mov x20, #0x18\n" - "mov x12, %x[bias]\n" + "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n" "ldr x11, [%x[args_ptr], %[offsetof_N]]\n" "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x9, %x[output_ptr]\n" - "madd %x[output_ptr], x21, x20, %x[output_ptr]\n" + "madd x20, x21, x20, x9\n" + "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n" "67:" // Height 6: Column loop "mov x20, #0x0\n" "whilelt p4.s, x20, x11\n" @@ -1387,15 +1390,15 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "cbz x12, 68f\n" "ld1w { z8.s }, p5/Z, [x12]\n" "ld1w { z9.s }, p5/Z, [x12, #1, MUL VL]\n" - "zip2 z12.d, z8.d, z8.d\n" - "zip1 z8.d, z8.d, z8.d\n" "ld1w { z10.s }, p5/Z, [x12, #2, MUL VL]\n" "ld1w { z11.s }, p5/Z, [x12, #3, MUL VL]\n" + "addvl x12, x12, #4\n" + "zip2 z12.d, z8.d, z8.d\n" + "zip1 z8.d, z8.d, z8.d\n" "zip2 z13.d, z9.d, z9.d\n" "zip1 z9.d, z9.d, z9.d\n" "zip2 z14.d, z10.d, z10.d\n" "zip1 z10.d, z10.d, z10.d\n" - "addvl x12, x12, #4\n" "zip2 z15.d, z11.d, z11.d\n" "zip1 z11.d, z11.d, z11.d\n" "mov z16.d, z8.d\n" @@ -1418,54 +1421,54 @@ 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" + "ld1w { z20.s }, p4/Z, [x9]\n" + "ld1w { z22.s }, p3/Z, [x9, #1, MUL VL]\n" + "ld1w { z25.s }, p2/Z, [x9, #2, MUL VL]\n" + "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\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" - "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, [x24]\n" - "zip1 z8.d, z17.d, z12.d\n" "ld1w { z13.s }, p3/Z, [x24, #1, MUL VL]\n" + "add x21, x22, x20, LSL #2\n" + "add x20, x21, x20, LSL #2\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, [x23, #3, MUL VL]\n" + "zip1 z8.d, z20.d, z12.d\n" + "zip2 z12.d, z20.d, z12.d\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, [x22, #1, MUL VL]\n" + "zip1 z9.d, z22.d, z13.d\n" + "zip2 z13.d, z22.d, z13.d\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, [x22, #3, MUL VL]\n" + "zip1 z10.d, z25.d, z14.d\n" + "zip2 z14.d, z25.d, z14.d\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, [x21, #1, MUL VL]\n" + "zip1 z11.d, z16.d, z15.d\n" + "zip2 z15.d, z16.d, z15.d\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 { z0.s }, p1/Z, [x21, #3, MUL VL]\n" + "zip1 z16.d, z17.d, z20.d\n" + "zip2 z20.d, z17.d, z20.d\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, [x20, #1, MUL VL]\n" + "zip1 z17.d, z18.d, z21.d\n" + "zip2 z21.d, z18.d, z21.d\n" "ld1w { z30.s }, p2/Z, [x20, #2, MUL VL]\n" + "ld1w { z31.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" + "zip2 z23.d, z24.d, z23.d\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" - "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" @@ -1501,8 +1504,8 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "mov x28, #0x0\n" "71:" // Height 6: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr w27, [x20, x28, LSL #0x2]\n" "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr w27, [x20, x28, LSL #0x2]\n" "tbz %x[flags], #3, 72f\n" "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -1533,146 +1536,146 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "ble 75f\n" "74:" // Height 6: Multiply loop: Main loop head "whilelt p0.s, XZR, x27\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" + "ld1h { z7.h }, p5/Z, [x10]\n" + "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n" + "sub x27, x27, #0x4\n" + "cmp x27, #0x4\n" + "ld1rqw { z5.s }, p0/Z, [x26]\n" + "add x26, x26, #0x10\n" + "ld1rqw { z4.s }, p0/Z, [x25]\n" + "add x25, x25, #0x10\n" + "ld1rqw { z3.s }, p0/Z, [x24]\n" + "add x24, x24, #0x10\n" + "ld1rqw { z1.s }, p0/Z, [x23]\n" + "ld1rqw { z2.s }, p0/Z, [x22]\n" + "ld1rqw { z0.s }, p0/Z, [x21]\n" + "add x23, x23, #0x10\n" + "add x22, x22, #0x10\n" ".inst 0x658ab4a5 // bfcvt z5.h, p5/M, z5.s\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" + "add x21, x21, #0x10\n" ".inst 0x658ab463 // bfcvt z3.h, p5/M, z3.s\n" + ".inst 0x658ab421 // bfcvt z1.h, p5/M, z1.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" + ".inst 0x658ab400 // bfcvt z0.h, p5/M, z0.s\n" "uzp1 z5.h, z5.h, z5.h\n" - "sub x27, x27, #0x4\n" - "cmp x27, #0x4\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 z1.h, z1.h, z1.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" + "uzp1 z0.h, z0.h, z0.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 0x6461e478 // bfmmla z24.s, z3.h, z1.h\n" - ".inst 0x6460e4ec // bfmmla z12.s, z7.h, z0.h\n" + "trn1 z3.d, z3.d, z1.d\n" + "trn1 z2.d, z2.d, z0.d\n" + ".inst 0x6467e4a8 // bfmmla z8.s, z5.h, z7.h\n" + ".inst 0x6467e470 // bfmmla z16.s, z3.h, z7.h\n" + ".inst 0x6466e4ac // bfmmla z12.s, z5.h, z6.h\n" + ".inst 0x6466e474 // bfmmla z20.s, z3.h, z6.h\n" + ".inst 0x6467e458 // bfmmla z24.s, z2.h, z7.h\n" "ld1h { z1.h }, p5/Z, [x10, #2, MUL VL]\n" - "add x22, x22, #0x10\n" - ".inst 0x6460e4b4 // bfmmla z20.s, z5.h, z0.h\n" - ".inst 0x6460e47c // bfmmla z28.s, z3.h, z0.h\n" + ".inst 0x6466e45c // bfmmla z28.s, z2.h, z6.h\n" "ld1h { z0.h }, p5/Z, [x10, #3, MUL VL]\n" - "add x21, x21, #0x10\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" + ".inst 0x6461e4a9 // bfmmla z9.s, z5.h, z1.h\n" + ".inst 0x6461e471 // bfmmla z17.s, z3.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 0x6460e4b5 // bfmmla z21.s, z5.h, z0.h\n" - ".inst 0x6460e47d // bfmmla z29.s, z3.h, z0.h\n" + ".inst 0x6460e4ad // bfmmla z13.s, z5.h, z0.h\n" + ".inst 0x6460e475 // bfmmla z21.s, z3.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 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" + ".inst 0x6461e4aa // bfmmla z10.s, z5.h, z1.h\n" + ".inst 0x6461e472 // bfmmla z18.s, z3.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 0x6460e4b6 // bfmmla z22.s, z5.h, z0.h\n" - ".inst 0x6460e47e // bfmmla z30.s, z3.h, z0.h\n" + ".inst 0x6460e4ae // bfmmla z14.s, z5.h, z0.h\n" + ".inst 0x6460e476 // bfmmla z22.s, z3.h, z0.h\n" + ".inst 0x6460e45e // bfmmla z30.s, z2.h, z0.h\n" "ld1h { z0.h }, p5/Z, [x10, #7, MUL VL]\n" "addvl x10, x10, #8\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" + ".inst 0x6461e4ab // bfmmla z11.s, z5.h, z1.h\n" + ".inst 0x6461e473 // bfmmla z19.s, z3.h, z1.h\n" + ".inst 0x6461e45b // bfmmla z27.s, z2.h, z1.h\n" + ".inst 0x6460e4af // bfmmla z15.s, z5.h, z0.h\n" + ".inst 0x6460e477 // bfmmla z23.s, z3.h, z0.h\n" + ".inst 0x6460e45f // bfmmla z31.s, z2.h, z0.h\n" "bgt 74b\n" "75:" // Height 6: Multiply loop: Single iteration only "whilelt p0.s, XZR, x27\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" + "ld1h { z7.h }, p5/Z, [x10]\n" + "ld1h { z6.h }, p5/Z, [x10, #1, MUL VL]\n" + "ld1rqw { z5.s }, p0/Z, [x26]\n" + "ld1rqw { z4.s }, p0/Z, [x25]\n" + "ld1rqw { z3.s }, p0/Z, [x24]\n" + "ld1rqw { z1.s }, p0/Z, [x23]\n" + "ld1rqw { z2.s }, p0/Z, [x22]\n" + "ld1rqw { z0.s }, p0/Z, [x21]\n" ".inst 0x658ab4a5 // bfcvt z5.h, p5/M, z5.s\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 0x658ab421 // bfcvt z1.h, p5/M, z1.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" + ".inst 0x658ab400 // bfcvt z0.h, p5/M, z0.s\n" "uzp1 z5.h, z5.h, z5.h\n" "uzp1 z4.h, z4.h, z4.h\n" "uzp1 z3.h, z3.h, z3.h\n" + "uzp1 z1.h, z1.h, z1.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" + "uzp1 z0.h, z0.h, z0.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" + "trn1 z3.d, z3.d, z1.d\n" + "trn1 z2.d, z2.d, z0.d\n" + ".inst 0x6467e4a8 // bfmmla z8.s, z5.h, z7.h\n" + ".inst 0x6467e470 // bfmmla z16.s, z3.h, z7.h\n" + ".inst 0x6466e4ac // bfmmla z12.s, z5.h, z6.h\n" + ".inst 0x6466e474 // bfmmla z20.s, z3.h, z6.h\n" + ".inst 0x6467e458 // bfmmla z24.s, z2.h, z7.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" + ".inst 0x6466e45c // bfmmla z28.s, z2.h, z6.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" + ".inst 0x6461e4a9 // bfmmla z9.s, z5.h, z1.h\n" + ".inst 0x6461e471 // bfmmla z17.s, z3.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 0x6460e4b5 // bfmmla z21.s, z5.h, z0.h\n" - ".inst 0x6460e47d // bfmmla z29.s, z3.h, z0.h\n" + ".inst 0x6460e4ad // bfmmla z13.s, z5.h, z0.h\n" + ".inst 0x6460e475 // bfmmla z21.s, z3.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 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" + ".inst 0x6461e4aa // bfmmla z10.s, z5.h, z1.h\n" + ".inst 0x6461e472 // bfmmla z18.s, z3.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 0x6460e4b6 // bfmmla z22.s, z5.h, z0.h\n" - ".inst 0x6460e47e // bfmmla z30.s, z3.h, z0.h\n" + ".inst 0x6460e4ae // bfmmla z14.s, z5.h, z0.h\n" + ".inst 0x6460e476 // bfmmla z22.s, z3.h, z0.h\n" + ".inst 0x6460e45e // bfmmla z30.s, z2.h, z0.h\n" "ld1h { z0.h }, p5/Z, [x10, #7, MUL VL]\n" "addvl x10, x10, #8\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" + ".inst 0x6461e4ab // bfmmla z11.s, z5.h, z1.h\n" + ".inst 0x6461e473 // bfmmla z19.s, z3.h, z1.h\n" + ".inst 0x6461e45b // bfmmla z27.s, z2.h, z1.h\n" + ".inst 0x6460e4af // bfmmla z15.s, z5.h, z0.h\n" + ".inst 0x6460e477 // bfmmla z23.s, z3.h, z0.h\n" + ".inst 0x6460e45f // bfmmla z31.s, z2.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" "cmp x28, x20\n" "bne 71b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x9, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" "uzp1 z6.d, z8.d, z12.d\n" - "add x23, x24, x20, LSL #2\n" - "add x22, x23, x20, LSL #2\n" "uzp2 z8.d, z8.d, z12.d\n" "uzp1 z12.d, z9.d, z13.d\n" "uzp2 z9.d, z9.d, z13.d\n" "uzp1 z13.d, z10.d, z14.d\n" - "add x21, x22, x20, LSL #2\n" "uzp2 z10.d, z10.d, z14.d\n" + "add x26, x9, x20, LSL #2\n" + "add x25, x26, x20, LSL #2\n" + "add x24, x25, x20, LSL #2\n" "uzp1 z14.d, z11.d, z15.d\n" "uzp2 z11.d, z11.d, z15.d\n" + "add x23, x24, x20, LSL #2\n" "uzp1 z15.d, z16.d, z20.d\n" "uzp2 z16.d, z16.d, z20.d\n" + "add x22, x23, x20, LSL #2\n" "uzp1 z20.d, z17.d, z21.d\n" "uzp2 z17.d, z17.d, z21.d\n" "uzp1 z21.d, z18.d, z22.d\n" @@ -1688,9 +1691,9 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "uzp1 z30.d, z27.d, z31.d\n" "uzp2 z27.d, z27.d, z31.d\n" "tbz %x[flags], #1, 77f\n" - "add x20, %x[args_ptr], %[offset_max]\n" - "ld1rw { z1.s }, p5/Z, [x20]\n" + "add x21, %x[args_ptr], %[offset_max]\n" "add x20, %x[args_ptr], %[offset_min]\n" + "ld1rw { z1.s }, p5/Z, [x21]\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" @@ -1746,26 +1749,26 @@ void sve_hybrid_fp32bf16fp32_mmla_6x4VL ( "st1w { z13.s }, p2, [x9, #2, MUL VL]\n" "st1w { z14.s }, p1, [x9, #3, MUL VL]\n" "addvl x9, x9, #4\n" - "st1w { z8.s }, p4, [x25]\n" - "st1w { z9.s }, p3, [x25, #1, MUL VL]\n" - "st1w { z10.s }, p2, [x25, #2, MUL VL]\n" - "st1w { z11.s }, p1, [x25, #3, MUL VL]\n" - "st1w { z15.s }, p4, [x24]\n" - "st1w { z20.s }, p3, [x24, #1, MUL VL]\n" - "st1w { z21.s }, p2, [x24, #2, MUL VL]\n" - "st1w { z22.s }, p1, [x24, #3, MUL VL]\n" - "st1w { z16.s }, p4, [x23]\n" - "st1w { z17.s }, p3, [x23, #1, MUL VL]\n" - "st1w { z18.s }, p2, [x23, #2, MUL VL]\n" - "st1w { z19.s }, p1, [x23, #3, MUL VL]\n" - "st1w { z23.s }, p4, [x22]\n" - "st1w { z28.s }, p3, [x22, #1, MUL VL]\n" - "st1w { z29.s }, p2, [x22, #2, MUL VL]\n" - "st1w { z30.s }, p1, [x22, #3, MUL VL]\n" - "st1w { z24.s }, p4, [x21]\n" - "st1w { z25.s }, p3, [x21, #1, MUL VL]\n" - "st1w { z26.s }, p2, [x21, #2, MUL VL]\n" - "st1w { z27.s }, p1, [x21, #3, MUL VL]\n" + "st1w { z8.s }, p4, [x26]\n" + "st1w { z9.s }, p3, [x26, #1, MUL VL]\n" + "st1w { z10.s }, p2, [x26, #2, MUL VL]\n" + "st1w { z11.s }, p1, [x26, #3, MUL VL]\n" + "st1w { z15.s }, p4, [x25]\n" + "st1w { z20.s }, p3, [x25, #1, MUL VL]\n" + "st1w { z21.s }, p2, [x25, #2, MUL VL]\n" + "st1w { z22.s }, p1, [x25, #3, MUL VL]\n" + "st1w { z16.s }, p4, [x24]\n" + "st1w { z17.s }, p3, [x24, #1, MUL VL]\n" + "st1w { z18.s }, p2, [x24, #2, MUL VL]\n" + "st1w { z19.s }, p1, [x24, #3, MUL VL]\n" + "st1w { z23.s }, p4, [x23]\n" + "st1w { z28.s }, p3, [x23, #1, MUL VL]\n" + "st1w { z29.s }, p2, [x23, #2, MUL VL]\n" + "st1w { z30.s }, p1, [x23, #3, MUL VL]\n" + "st1w { z24.s }, p4, [x22]\n" + "st1w { z25.s }, p3, [x22, #1, MUL VL]\n" + "st1w { z26.s }, p2, [x22, #2, MUL VL]\n" + "st1w { z27.s }, p1, [x22, #3, MUL VL]\n" "78:" // Height 6: Writeback done "decw x11, ALL, MUL #4\n" "cmp x11, XZR\n" @@ -1782,8 +1785,8 @@ 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)) + : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr) + : [args_ptr] "r" (&ka), [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_bias] "I" (offsetof(KernelArgs, bias)), [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_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [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" ); } |