diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32_mla_6x4VL/generic.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32_mla_6x4VL/generic.cpp | 765 |
1 files changed, 384 insertions, 381 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32_mla_6x4VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32_mla_6x4VL/generic.cpp index b1ab31e618..299dec5b3c 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32_mla_6x4VL/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_ffhybrid_fp32_mla_6x4VL/generic.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2022-2023 Arm Limited. + * Copyright (c) 2022-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -49,18 +49,19 @@ void sve_ffhybrid_fp32_mla_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; } @@ -81,6 +82,7 @@ void sve_ffhybrid_fp32_mla_6x4VL ( ka.string_lengths = string_lengths; ka.N = N; ka.B_ptr = B_ptr; + ka.bias = bias; ka.B_stride = B_stride; switch(act.type) { default: @@ -106,15 +108,15 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "bgt 29f\n" "beq 15f\n" "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x15, %x[bias]\n" + "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n" "ldr x14, [%x[args_ptr], %[offsetof_N]]\n" "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" - "mov x13, %x[output_ptr]\n" + "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n" "2:" // Height 1: Column loop "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n" - "add x11, x12, x20, LSL #2\n" "cntw x21, ALL, MUL #3\n" + "add x11, x12, x20, LSL #2\n" "add x10, x11, x20, LSL #2\n" "add x9, x10, x20, LSL #2\n" "add x20, x9, x20, LSL #2\n" @@ -122,12 +124,12 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" "bgt 3f\n" "decw x21\n" - "cmp x14, x21\n" "mov x9, x12\n" + "cmp x14, x21\n" "bgt 3f\n" "decw x21\n" - "cmp x14, x21\n" "mov x10, x12\n" + "cmp x14, x21\n" "bgt 3f\n" "mov x11, x12\n" "3:" // Height 1: B setup done @@ -162,8 +164,8 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "mov x28, #0x0\n" "7:" // 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, 8f\n" "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -179,113 +181,113 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "ble 11f\n" "10:" // Height 1: Multiply loop: Main loop head "whilelt p0.s, XZR, x27\n" - "ld1rqw { z0.s }, p0/Z, [x26]\n" - "ld1w { z16.s }, p5/Z, [x12]\n" - "fmla z8.s, z16.s, z0.s[0]\n" + "ld1w { z17.s }, p5/Z, [x12]\n" "ld1w { z16.s }, p5/Z, [x11]\n" + "sub x27, x27, #0x4\n" + "cmp x27, #0x4\n" + "ld1rqw { z0.s }, p0/Z, [x26]\n" + "add x26, x26, #0x10\n" + "fmla z8.s, z17.s, z0.s[0]\n" + "ld1w { z17.s }, p5/Z, [x10]\n" "fmla z9.s, z16.s, z0.s[0]\n" - "ld1w { z16.s }, p5/Z, [x10]\n" - "fmla z10.s, z16.s, z0.s[0]\n" "ld1w { z16.s }, p5/Z, [x9]\n" + "fmla z10.s, z17.s, z0.s[0]\n" + "ld1w { z17.s }, p5/Z, [x12, #1, MUL VL]\n" "fmla z11.s, z16.s, z0.s[0]\n" - "ld1w { z16.s }, p5/Z, [x12, #1, MUL VL]\n" - "fmla z8.s, z16.s, z0.s[1]\n" "ld1w { z16.s }, p5/Z, [x11, #1, MUL VL]\n" + "fmla z8.s, z17.s, z0.s[1]\n" + "ld1w { z17.s }, p5/Z, [x10, #1, MUL VL]\n" "fmla z9.s, z16.s, z0.s[1]\n" - "ld1w { z16.s }, p5/Z, [x10, #1, MUL VL]\n" - "fmla z10.s, z16.s, z0.s[1]\n" "ld1w { z16.s }, p5/Z, [x9, #1, MUL VL]\n" + "fmla z10.s, z17.s, z0.s[1]\n" + "ld1w { z17.s }, p5/Z, [x12, #2, MUL VL]\n" "fmla z11.s, z16.s, z0.s[1]\n" - "ld1w { z16.s }, p5/Z, [x12, #2, MUL VL]\n" - "fmla z8.s, z16.s, z0.s[2]\n" "ld1w { z16.s }, p5/Z, [x11, #2, MUL VL]\n" + "fmla z8.s, z17.s, z0.s[2]\n" + "ld1w { z17.s }, p5/Z, [x10, #2, MUL VL]\n" "fmla z9.s, z16.s, z0.s[2]\n" - "ld1w { z16.s }, p5/Z, [x10, #2, MUL VL]\n" - "fmla z10.s, z16.s, z0.s[2]\n" "ld1w { z16.s }, p5/Z, [x9, #2, MUL VL]\n" + "fmla z10.s, z17.s, z0.s[2]\n" + "ld1w { z17.s }, p5/Z, [x12, #3, MUL VL]\n" + "addvl x12, x12, #4\n" "fmla z11.s, z16.s, z0.s[2]\n" - "ld1w { z16.s }, p5/Z, [x12, #3, MUL VL]\n" - "fmla z8.s, z16.s, z0.s[3]\n" "ld1w { z16.s }, p5/Z, [x11, #3, MUL VL]\n" - "fmla z9.s, z16.s, z0.s[3]\n" + "addvl x11, x11, #4\n" + "fmla z8.s, z17.s, z0.s[3]\n" "ld1w { z17.s }, p5/Z, [x10, #3, MUL VL]\n" - "sub x27, x27, #0x4\n" + "addvl x10, x10, #4\n" + "fmla z9.s, z16.s, z0.s[3]\n" "ld1w { z16.s }, p5/Z, [x9, #3, MUL VL]\n" - "cmp x27, #0x4\n" + "addvl x9, x9, #4\n" "fmla z10.s, z17.s, z0.s[3]\n" "fmla z11.s, z16.s, z0.s[3]\n" - "add x26, x26, #0x10\n" - "addvl x12, x12, #4\n" - "addvl x11, x11, #4\n" - "addvl x10, x10, #4\n" - "addvl x9, x9, #4\n" "bgt 10b\n" "11:" // Height 1: Multiply loop: Single iteration only "whilelt p0.s, XZR, x27\n" - "ld1rqw { z0.s }, p0/Z, [x26]\n" - "ld1w { z16.s }, p5/Z, [x12]\n" - "fmla z8.s, z16.s, z0.s[0]\n" + "ld1w { z17.s }, p5/Z, [x12]\n" "ld1w { z16.s }, p5/Z, [x11]\n" - "fmla z9.s, z16.s, z0.s[0]\n" - "ld1w { z17.s }, p5/Z, [x10]\n" "subs x27, x27, #0x1\n" - "ld1w { z16.s }, p5/Z, [x9]\n" - "fmla z10.s, z17.s, z0.s[0]\n" - "fmla z11.s, z16.s, z0.s[0]\n" "addvl x12, x12, #1\n" "addvl x11, x11, #1\n" + "ld1rqw { z0.s }, p0/Z, [x26]\n" + "fmla z8.s, z17.s, z0.s[0]\n" + "ld1w { z17.s }, p5/Z, [x10]\n" "addvl x10, x10, #1\n" + "fmla z9.s, z16.s, z0.s[0]\n" + "ld1w { z16.s }, p5/Z, [x9]\n" "addvl x9, x9, #1\n" + "fmla z10.s, z17.s, z0.s[0]\n" + "fmla z11.s, z16.s, z0.s[0]\n" "ble 12f\n" "ld1w { z17.s }, p5/Z, [x12]\n" "ld1w { z16.s }, p5/Z, [x11]\n" - "fmla z8.s, z17.s, z0.s[1]\n" - "fmla z9.s, z16.s, z0.s[1]\n" - "ld1w { z17.s }, p5/Z, [x10]\n" - "ld1w { z16.s }, p5/Z, [x9]\n" "subs x27, x27, #0x1\n" - "fmla z10.s, z17.s, z0.s[1]\n" - "fmla z11.s, z16.s, z0.s[1]\n" "addvl x12, x12, #1\n" "addvl x11, x11, #1\n" + "fmla z8.s, z17.s, z0.s[1]\n" + "ld1w { z17.s }, p5/Z, [x10]\n" "addvl x10, x10, #1\n" + "fmla z9.s, z16.s, z0.s[1]\n" + "ld1w { z16.s }, p5/Z, [x9]\n" "addvl x9, x9, #1\n" + "fmla z10.s, z17.s, z0.s[1]\n" + "fmla z11.s, z16.s, z0.s[1]\n" "ble 12f\n" "ld1w { z17.s }, p5/Z, [x12]\n" "ld1w { z16.s }, p5/Z, [x11]\n" - "fmla z8.s, z17.s, z0.s[2]\n" - "fmla z9.s, z16.s, z0.s[2]\n" - "ld1w { z17.s }, p5/Z, [x10]\n" - "ld1w { z16.s }, p5/Z, [x9]\n" "subs x27, x27, #0x1\n" - "fmla z10.s, z17.s, z0.s[2]\n" - "fmla z11.s, z16.s, z0.s[2]\n" "addvl x12, x12, #1\n" "addvl x11, x11, #1\n" + "fmla z8.s, z17.s, z0.s[2]\n" + "ld1w { z17.s }, p5/Z, [x10]\n" "addvl x10, x10, #1\n" + "fmla z9.s, z16.s, z0.s[2]\n" + "ld1w { z16.s }, p5/Z, [x9]\n" "addvl x9, x9, #1\n" + "fmla z10.s, z17.s, z0.s[2]\n" + "fmla z11.s, z16.s, z0.s[2]\n" "ble 12f\n" "ld1w { z17.s }, p5/Z, [x12]\n" "ld1w { z16.s }, p5/Z, [x11]\n" + "addvl x12, x12, #1\n" + "addvl x11, x11, #1\n" "fmla z8.s, z17.s, z0.s[3]\n" - "fmla z9.s, z16.s, z0.s[3]\n" "ld1w { z17.s }, p5/Z, [x10]\n" + "addvl x10, x10, #1\n" + "fmla z9.s, z16.s, z0.s[3]\n" "ld1w { z16.s }, p5/Z, [x9]\n" + "addvl x9, x9, #1\n" "fmla z10.s, z17.s, z0.s[3]\n" "fmla z11.s, z16.s, z0.s[3]\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" - "addvl x10, x10, #1\n" - "addvl x9, x9, #1\n" "12:" // Height 1: Multiply loop: multiply skip "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x28, x28, #0x1\n" "cmp x28, x20\n" "bne 7b\n" "tbz %x[flags], #1, 13f\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" @@ -308,15 +310,15 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "b 86f\n" "15:" // Height 2 "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x15, %x[bias]\n" + "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n" "ldr x14, [%x[args_ptr], %[offsetof_N]]\n" "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" - "mov x13, %x[output_ptr]\n" + "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n" "16:" // Height 2: Column loop "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n" - "add x11, x12, x20, LSL #2\n" "cntw x21, ALL, MUL #3\n" + "add x11, x12, x20, LSL #2\n" "add x10, x11, x20, LSL #2\n" "add x9, x10, x20, LSL #2\n" "add x20, x9, x20, LSL #2\n" @@ -324,12 +326,12 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" "bgt 17f\n" "decw x21\n" - "cmp x14, x21\n" "mov x9, x12\n" + "cmp x14, x21\n" "bgt 17f\n" "decw x21\n" - "cmp x14, x21\n" "mov x10, x12\n" + "cmp x14, x21\n" "bgt 17f\n" "mov x11, x12\n" "17:" // Height 2: B setup done @@ -344,22 +346,22 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "cbz x15, 18f\n" "ld1w { z8.s }, p5/Z, [x15]\n" "ld1w { z9.s }, p5/Z, [x15, #1, MUL VL]\n" - "mov z12.d, z8.d\n" - "mov z13.d, z9.d\n" "ld1w { z10.s }, p5/Z, [x15, #2, MUL VL]\n" "ld1w { z11.s }, p5/Z, [x15, #3, MUL VL]\n" + "addvl x15, x15, #4\n" + "mov z12.d, z8.d\n" + "mov z13.d, z9.d\n" "mov z14.d, z10.d\n" "mov z15.d, z11.d\n" - "addvl x15, x15, #4\n" "b 20f\n" "18:" // Height 2: no bias "tbz %x[flags], #0, 19f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x20, x13, x20, LSL #2\n" "ld1w { z8.s }, p4/Z, [x13]\n" "ld1w { z9.s }, p3/Z, [x13, #1, MUL VL]\n" "ld1w { z10.s }, p2/Z, [x13, #2, MUL VL]\n" "ld1w { z11.s }, p1/Z, [x13, #3, MUL VL]\n" + "add x20, x13, x20, LSL #2\n" "ld1w { z12.s }, p4/Z, [x20]\n" "ld1w { z13.s }, p3/Z, [x20, #1, MUL VL]\n" "ld1w { z14.s }, p2/Z, [x20, #2, MUL VL]\n" @@ -378,8 +380,8 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "mov x28, #0x0\n" "21:" // 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, 22f\n" "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -398,29 +400,29 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "ble 25f\n" "24:" // Height 2: Multiply loop: Main loop head "whilelt p0.s, XZR, x27\n" - "ld1rqw { z1.s }, p0/Z, [x26]\n" - "ld1rqw { z0.s }, p0/Z, [x25]\n" - "sub x27, x27, #0x4\n" "ld1w { z17.s }, p5/Z, [x12]\n" "ld1w { z16.s }, p5/Z, [x11]\n" + "sub x27, x27, #0x4\n" + "cmp x27, #0x4\n" + "ld1rqw { z1.s }, p0/Z, [x26]\n" + "add x26, x26, #0x10\n" + "ld1rqw { z0.s }, p0/Z, [x25]\n" + "add x25, x25, #0x10\n" "fmla z8.s, z17.s, z1.s[0]\n" "fmla z12.s, z17.s, z0.s[0]\n" + "ld1w { z17.s }, p5/Z, [x10]\n" "fmla z9.s, z16.s, z1.s[0]\n" "fmla z13.s, z16.s, z0.s[0]\n" - "ld1w { z17.s }, p5/Z, [x10]\n" "ld1w { z16.s }, p5/Z, [x9]\n" "fmla z10.s, z17.s, z1.s[0]\n" "fmla z14.s, z17.s, z0.s[0]\n" "ld1w { z17.s }, p5/Z, [x12, #1, MUL VL]\n" - "cmp x27, #0x4\n" "fmla z11.s, z16.s, z1.s[0]\n" "fmla z15.s, z16.s, z0.s[0]\n" "ld1w { z16.s }, p5/Z, [x11, #1, MUL VL]\n" - "add x26, x26, #0x10\n" "fmla z8.s, z17.s, z1.s[1]\n" "fmla z12.s, z17.s, z0.s[1]\n" "ld1w { z17.s }, p5/Z, [x10, #1, MUL VL]\n" - "add x25, x25, #0x10\n" "fmla z9.s, z16.s, z1.s[1]\n" "fmla z13.s, z16.s, z0.s[1]\n" "ld1w { z16.s }, p5/Z, [x9, #1, MUL VL]\n" @@ -459,89 +461,89 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "bgt 24b\n" "25:" // 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" - "subs x27, x27, #0x1\n" "ld1w { z17.s }, p5/Z, [x12]\n" "ld1w { z16.s }, p5/Z, [x11]\n" + "subs x27, x27, #0x1\n" + "addvl x12, x12, #1\n" + "addvl x11, x11, #1\n" + "ld1rqw { z0.s }, p0/Z, [x26]\n" + "ld1rqw { z1.s }, p0/Z, [x25]\n" "fmla z8.s, z17.s, z0.s[0]\n" "fmla z12.s, z17.s, z1.s[0]\n" + "ld1w { z17.s }, p5/Z, [x10]\n" + "addvl x10, x10, #1\n" "fmla z9.s, z16.s, z0.s[0]\n" "fmla z13.s, z16.s, z1.s[0]\n" - "ld1w { z17.s }, p5/Z, [x10]\n" "ld1w { z16.s }, p5/Z, [x9]\n" + "addvl x9, x9, #1\n" "fmla z10.s, z17.s, z0.s[0]\n" "fmla z14.s, z17.s, z1.s[0]\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "fmla z11.s, z16.s, z0.s[0]\n" "fmla z15.s, z16.s, z1.s[0]\n" - "addvl x10, x10, #1\n" - "addvl x9, x9, #1\n" "ble 26f\n" "ld1w { z17.s }, p5/Z, [x12]\n" "ld1w { z16.s }, p5/Z, [x11]\n" + "subs x27, x27, #0x1\n" + "addvl x12, x12, #1\n" + "addvl x11, x11, #1\n" "fmla z8.s, z17.s, z0.s[1]\n" "fmla z12.s, z17.s, z1.s[1]\n" + "ld1w { z17.s }, p5/Z, [x10]\n" + "addvl x10, x10, #1\n" "fmla z9.s, z16.s, z0.s[1]\n" "fmla z13.s, z16.s, z1.s[1]\n" - "ld1w { z17.s }, p5/Z, [x10]\n" "ld1w { z16.s }, p5/Z, [x9]\n" - "subs x27, x27, #0x1\n" + "addvl x9, x9, #1\n" "fmla z10.s, z17.s, z0.s[1]\n" "fmla z14.s, z17.s, z1.s[1]\n" - "addvl x12, x12, #1\n" "fmla z11.s, z16.s, z0.s[1]\n" "fmla z15.s, z16.s, z1.s[1]\n" - "addvl x11, x11, #1\n" - "addvl x10, x10, #1\n" - "addvl x9, x9, #1\n" "ble 26f\n" "ld1w { z17.s }, p5/Z, [x12]\n" "ld1w { z16.s }, p5/Z, [x11]\n" + "subs x27, x27, #0x1\n" + "addvl x12, x12, #1\n" + "addvl x11, x11, #1\n" "fmla z8.s, z17.s, z0.s[2]\n" "fmla z12.s, z17.s, z1.s[2]\n" + "ld1w { z17.s }, p5/Z, [x10]\n" + "addvl x10, x10, #1\n" "fmla z9.s, z16.s, z0.s[2]\n" "fmla z13.s, z16.s, z1.s[2]\n" - "ld1w { z17.s }, p5/Z, [x10]\n" "ld1w { z16.s }, p5/Z, [x9]\n" - "subs x27, x27, #0x1\n" + "addvl x9, x9, #1\n" "fmla z10.s, z17.s, z0.s[2]\n" "fmla z14.s, z17.s, z1.s[2]\n" - "addvl x12, x12, #1\n" "fmla z11.s, z16.s, z0.s[2]\n" "fmla z15.s, z16.s, z1.s[2]\n" - "addvl x11, x11, #1\n" - "addvl x10, x10, #1\n" - "addvl x9, x9, #1\n" "ble 26f\n" "ld1w { z17.s }, p5/Z, [x12]\n" "ld1w { z16.s }, p5/Z, [x11]\n" + "addvl x12, x12, #1\n" + "addvl x11, x11, #1\n" "fmla z8.s, z17.s, z0.s[3]\n" "fmla z12.s, z17.s, z1.s[3]\n" + "ld1w { z17.s }, p5/Z, [x10]\n" + "addvl x10, x10, #1\n" "fmla z9.s, z16.s, z0.s[3]\n" "fmla z13.s, z16.s, z1.s[3]\n" - "ld1w { z17.s }, p5/Z, [x10]\n" "ld1w { z16.s }, p5/Z, [x9]\n" + "addvl x9, x9, #1\n" "fmla z10.s, z17.s, z0.s[3]\n" "fmla z14.s, z17.s, z1.s[3]\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "fmla z11.s, z16.s, z0.s[3]\n" "fmla z15.s, z16.s, z1.s[3]\n" - "addvl x10, x10, #1\n" - "addvl x9, x9, #1\n" "26:" // Height 2: Multiply loop: multiply skip "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x28, x28, #0x1\n" "cmp x28, x20\n" "bne 21b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x13, x20, LSL #2\n" + "add x26, x13, x20, LSL #2\n" "tbz %x[flags], #1, 27f\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" @@ -565,10 +567,10 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "st1w { z10.s }, p2, [x13, #2, MUL VL]\n" "st1w { z11.s }, p1, [x13, #3, MUL VL]\n" "addvl x13, x13, #4\n" - "st1w { z12.s }, p4, [x25]\n" - "st1w { z13.s }, p3, [x25, #1, MUL VL]\n" - "st1w { z14.s }, p2, [x25, #2, MUL VL]\n" - "st1w { z15.s }, p1, [x25, #3, MUL VL]\n" + "st1w { z12.s }, p4, [x26]\n" + "st1w { z13.s }, p3, [x26, #1, MUL VL]\n" + "st1w { z14.s }, p2, [x26, #2, MUL VL]\n" + "st1w { z15.s }, p1, [x26, #3, MUL VL]\n" "28:" // Height 2: Writeback done "decw x14, ALL, MUL #4\n" "cmp x14, XZR\n" @@ -576,15 +578,15 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "b 86f\n" "29:" // Height 3 "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x15, %x[bias]\n" + "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n" "ldr x14, [%x[args_ptr], %[offsetof_N]]\n" "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" - "mov x13, %x[output_ptr]\n" + "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n" "30:" // Height 3: Column loop "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n" - "add x11, x12, x20, LSL #2\n" "cntw x21, ALL, MUL #3\n" + "add x11, x12, x20, LSL #2\n" "add x10, x11, x20, LSL #2\n" "add x9, x10, x20, LSL #2\n" "add x20, x9, x20, LSL #2\n" @@ -592,12 +594,12 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" "bgt 31f\n" "decw x21\n" - "cmp x14, x21\n" "mov x9, x12\n" + "cmp x14, x21\n" "bgt 31f\n" "decw x21\n" - "cmp x14, x21\n" "mov x10, x12\n" + "cmp x14, x21\n" "bgt 31f\n" "mov x11, x12\n" "31:" // Height 3: B setup done @@ -612,27 +614,27 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "cbz x15, 32f\n" "ld1w { z8.s }, p5/Z, [x15]\n" "ld1w { z9.s }, p5/Z, [x15, #1, MUL VL]\n" - "mov z12.d, z8.d\n" - "mov z13.d, z9.d\n" "ld1w { z10.s }, p5/Z, [x15, #2, MUL VL]\n" "ld1w { z11.s }, p5/Z, [x15, #3, MUL VL]\n" - "mov z14.d, z10.d\n" - "mov z15.d, z11.d\n" + "addvl x15, x15, #4\n" + "mov z12.d, z8.d\n" + "mov z13.d, z9.d\n" "mov z16.d, z8.d\n" "mov z17.d, z9.d\n" - "addvl x15, x15, #4\n" + "mov z14.d, z10.d\n" + "mov z15.d, z11.d\n" "mov z18.d, z10.d\n" "mov z19.d, z11.d\n" "b 34f\n" "32:" // Height 3: no bias "tbz %x[flags], #0, 33f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x21, x13, x20, LSL #2\n" - "add x20, x21, x20, LSL #2\n" "ld1w { z8.s }, p4/Z, [x13]\n" "ld1w { z9.s }, p3/Z, [x13, #1, MUL VL]\n" "ld1w { z10.s }, p2/Z, [x13, #2, MUL VL]\n" "ld1w { z11.s }, p1/Z, [x13, #3, MUL VL]\n" + "add x21, x13, x20, LSL #2\n" + "add x20, x21, x20, LSL #2\n" "ld1w { z12.s }, p4/Z, [x21]\n" "ld1w { z13.s }, p3/Z, [x21, #1, MUL VL]\n" "ld1w { z14.s }, p2/Z, [x21, #2, MUL VL]\n" @@ -659,8 +661,8 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "mov x28, #0x0\n" "35:" // 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, 36f\n" "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -682,62 +684,62 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "ble 39f\n" "38:" // Height 3: Multiply loop: Main loop head "whilelt p0.s, XZR, x27\n" + "ld1w { z21.s }, p5/Z, [x12]\n" + "ld1w { z20.s }, p5/Z, [x11]\n" + "sub x27, x27, #0x4\n" + "cmp x27, #0x4\n" "ld1rqw { z2.s }, p0/Z, [x26]\n" + "add x26, x26, #0x10\n" "ld1rqw { z1.s }, p0/Z, [x25]\n" - "sub x27, x27, #0x4\n" + "add x25, x25, #0x10\n" "ld1rqw { z0.s }, p0/Z, [x24]\n" - "ld1w { z21.s }, p5/Z, [x12]\n" + "add x24, x24, #0x10\n" "fmla z8.s, z21.s, z2.s[0]\n" "fmla z12.s, z21.s, z1.s[0]\n" - "ld1w { z20.s }, p5/Z, [x11]\n" - "fmla z16.s, z21.s, z0.s[0]\n" "fmla z9.s, z20.s, z2.s[0]\n" - "ld1w { z21.s }, p5/Z, [x10]\n" "fmla z13.s, z20.s, z1.s[0]\n" + "fmla z16.s, z21.s, z0.s[0]\n" + "ld1w { z21.s }, p5/Z, [x10]\n" "fmla z17.s, z20.s, z0.s[0]\n" "ld1w { z20.s }, p5/Z, [x9]\n" - "cmp x27, #0x4\n" "fmla z10.s, z21.s, z2.s[0]\n" "fmla z14.s, z21.s, z1.s[0]\n" - "add x26, x26, #0x10\n" - "add x25, x25, #0x10\n" "fmla z18.s, z21.s, z0.s[0]\n" - "fmla z11.s, z20.s, z2.s[0]\n" "ld1w { z21.s }, p5/Z, [x12, #1, MUL VL]\n" - "add x24, x24, #0x10\n" + "fmla z11.s, z20.s, z2.s[0]\n" "fmla z15.s, z20.s, z1.s[0]\n" "fmla z19.s, z20.s, z0.s[0]\n" "ld1w { z20.s }, p5/Z, [x11, #1, MUL VL]\n" "fmla z8.s, z21.s, z2.s[1]\n" "fmla z12.s, z21.s, z1.s[1]\n" "fmla z16.s, z21.s, z0.s[1]\n" - "fmla z9.s, z20.s, z2.s[1]\n" "ld1w { z21.s }, p5/Z, [x10, #1, MUL VL]\n" + "fmla z9.s, z20.s, z2.s[1]\n" "fmla z13.s, z20.s, z1.s[1]\n" "fmla z17.s, z20.s, z0.s[1]\n" "ld1w { z20.s }, p5/Z, [x9, #1, MUL VL]\n" "fmla z10.s, z21.s, z2.s[1]\n" "fmla z14.s, z21.s, z1.s[1]\n" "fmla z18.s, z21.s, z0.s[1]\n" - "fmla z11.s, z20.s, z2.s[1]\n" "ld1w { z21.s }, p5/Z, [x12, #2, MUL VL]\n" + "fmla z11.s, z20.s, z2.s[1]\n" "fmla z15.s, z20.s, z1.s[1]\n" "fmla z19.s, z20.s, z0.s[1]\n" "ld1w { z20.s }, p5/Z, [x11, #2, MUL VL]\n" "fmla z8.s, z21.s, z2.s[2]\n" "fmla z12.s, z21.s, z1.s[2]\n" "fmla z16.s, z21.s, z0.s[2]\n" - "fmla z9.s, z20.s, z2.s[2]\n" "ld1w { z21.s }, p5/Z, [x10, #2, MUL VL]\n" + "fmla z9.s, z20.s, z2.s[2]\n" "fmla z13.s, z20.s, z1.s[2]\n" "fmla z17.s, z20.s, z0.s[2]\n" "ld1w { z20.s }, p5/Z, [x9, #2, MUL VL]\n" "fmla z10.s, z21.s, z2.s[2]\n" "fmla z14.s, z21.s, z1.s[2]\n" "fmla z18.s, z21.s, z0.s[2]\n" - "fmla z11.s, z20.s, z2.s[2]\n" "ld1w { z21.s }, p5/Z, [x12, #3, MUL VL]\n" "addvl x12, x12, #4\n" + "fmla z11.s, z20.s, z2.s[2]\n" "fmla z15.s, z20.s, z1.s[2]\n" "fmla z19.s, z20.s, z0.s[2]\n" "ld1w { z20.s }, p5/Z, [x11, #3, MUL VL]\n" @@ -745,9 +747,9 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "fmla z8.s, z21.s, z2.s[3]\n" "fmla z12.s, z21.s, z1.s[3]\n" "fmla z16.s, z21.s, z0.s[3]\n" - "fmla z9.s, z20.s, z2.s[3]\n" "ld1w { z21.s }, p5/Z, [x10, #3, MUL VL]\n" "addvl x10, x10, #4\n" + "fmla z9.s, z20.s, z2.s[3]\n" "fmla z13.s, z20.s, z1.s[3]\n" "fmla z17.s, z20.s, z0.s[3]\n" "ld1w { z20.s }, p5/Z, [x9, #3, MUL VL]\n" @@ -761,91 +763,91 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "bgt 38b\n" "39:" // Height 3: Multiply loop: Single iteration only "whilelt p0.s, XZR, x27\n" + "ld1w { z21.s }, p5/Z, [x12]\n" + "ld1w { z20.s }, p5/Z, [x11]\n" + "subs x27, x27, #0x1\n" + "addvl x12, x12, #1\n" + "addvl x11, x11, #1\n" "ld1rqw { z0.s }, p0/Z, [x26]\n" "ld1rqw { z1.s }, p0/Z, [x25]\n" - "subs x27, x27, #0x1\n" "ld1rqw { z2.s }, p0/Z, [x24]\n" - "ld1w { z21.s }, p5/Z, [x12]\n" "fmla z8.s, z21.s, z0.s[0]\n" "fmla z12.s, z21.s, z1.s[0]\n" - "ld1w { z20.s }, p5/Z, [x11]\n" - "fmla z16.s, z21.s, z2.s[0]\n" "fmla z9.s, z20.s, z0.s[0]\n" - "ld1w { z21.s }, p5/Z, [x10]\n" "fmla z13.s, z20.s, z1.s[0]\n" + "fmla z16.s, z21.s, z2.s[0]\n" + "ld1w { z21.s }, p5/Z, [x10]\n" + "addvl x10, x10, #1\n" "fmla z17.s, z20.s, z2.s[0]\n" "ld1w { z20.s }, p5/Z, [x9]\n" - "addvl x12, x12, #1\n" + "addvl x9, x9, #1\n" "fmla z10.s, z21.s, z0.s[0]\n" "fmla z14.s, z21.s, z1.s[0]\n" - "addvl x11, x11, #1\n" - "addvl x10, x10, #1\n" "fmla z18.s, z21.s, z2.s[0]\n" "fmla z11.s, z20.s, z0.s[0]\n" - "addvl x9, x9, #1\n" "fmla z15.s, z20.s, z1.s[0]\n" "fmla z19.s, z20.s, z2.s[0]\n" "ble 40f\n" "ld1w { z21.s }, p5/Z, [x12]\n" "ld1w { z20.s }, p5/Z, [x11]\n" + "subs x27, x27, #0x1\n" + "addvl x12, x12, #1\n" + "addvl x11, x11, #1\n" "fmla z8.s, z21.s, z0.s[1]\n" "fmla z12.s, z21.s, z1.s[1]\n" "fmla z16.s, z21.s, z2.s[1]\n" - "fmla z9.s, z20.s, z0.s[1]\n" "ld1w { z21.s }, p5/Z, [x10]\n" - "subs x27, x27, #0x1\n" + "addvl x10, x10, #1\n" + "fmla z9.s, z20.s, z0.s[1]\n" "fmla z13.s, z20.s, z1.s[1]\n" "fmla z17.s, z20.s, z2.s[1]\n" "ld1w { z20.s }, p5/Z, [x9]\n" - "addvl x12, x12, #1\n" + "addvl x9, x9, #1\n" "fmla z10.s, z21.s, z0.s[1]\n" "fmla z14.s, z21.s, z1.s[1]\n" - "addvl x11, x11, #1\n" - "addvl x10, x10, #1\n" "fmla z18.s, z21.s, z2.s[1]\n" "fmla z11.s, z20.s, z0.s[1]\n" - "addvl x9, x9, #1\n" "fmla z15.s, z20.s, z1.s[1]\n" "fmla z19.s, z20.s, z2.s[1]\n" "ble 40f\n" "ld1w { z21.s }, p5/Z, [x12]\n" "ld1w { z20.s }, p5/Z, [x11]\n" + "subs x27, x27, #0x1\n" + "addvl x12, x12, #1\n" + "addvl x11, x11, #1\n" "fmla z8.s, z21.s, z0.s[2]\n" "fmla z12.s, z21.s, z1.s[2]\n" "fmla z16.s, z21.s, z2.s[2]\n" - "fmla z9.s, z20.s, z0.s[2]\n" "ld1w { z21.s }, p5/Z, [x10]\n" - "subs x27, x27, #0x1\n" + "addvl x10, x10, #1\n" + "fmla z9.s, z20.s, z0.s[2]\n" "fmla z13.s, z20.s, z1.s[2]\n" "fmla z17.s, z20.s, z2.s[2]\n" "ld1w { z20.s }, p5/Z, [x9]\n" - "addvl x12, x12, #1\n" + "addvl x9, x9, #1\n" "fmla z10.s, z21.s, z0.s[2]\n" "fmla z14.s, z21.s, z1.s[2]\n" - "addvl x11, x11, #1\n" - "addvl x10, x10, #1\n" "fmla z18.s, z21.s, z2.s[2]\n" "fmla z11.s, z20.s, z0.s[2]\n" - "addvl x9, x9, #1\n" "fmla z15.s, z20.s, z1.s[2]\n" "fmla z19.s, z20.s, z2.s[2]\n" "ble 40f\n" "ld1w { z21.s }, p5/Z, [x12]\n" "ld1w { z20.s }, p5/Z, [x11]\n" + "addvl x12, x12, #1\n" + "addvl x11, x11, #1\n" "fmla z8.s, z21.s, z0.s[3]\n" "fmla z12.s, z21.s, z1.s[3]\n" "fmla z16.s, z21.s, z2.s[3]\n" - "fmla z9.s, z20.s, z0.s[3]\n" "ld1w { z21.s }, p5/Z, [x10]\n" - "addvl x12, x12, #1\n" + "addvl x10, x10, #1\n" + "fmla z9.s, z20.s, z0.s[3]\n" "fmla z13.s, z20.s, z1.s[3]\n" "fmla z17.s, z20.s, z2.s[3]\n" "ld1w { z20.s }, p5/Z, [x9]\n" - "addvl x11, x11, #1\n" + "addvl x9, x9, #1\n" "fmla z10.s, z21.s, z0.s[3]\n" "fmla z14.s, z21.s, z1.s[3]\n" - "addvl x10, x10, #1\n" - "addvl x9, x9, #1\n" "fmla z18.s, z21.s, z2.s[3]\n" "fmla z11.s, z20.s, z0.s[3]\n" "fmla z15.s, z20.s, z1.s[3]\n" @@ -856,12 +858,12 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "cmp x28, x20\n" "bne 35b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x13, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" + "add x26, x13, x20, LSL #2\n" + "add x25, x26, x20, LSL #2\n" "tbz %x[flags], #1, 41f\n" - "add x20, %x[args_ptr], %[offset_max]\n" - "ld1rw { z21.s }, p5/Z, [x20]\n" + "add x21, %x[args_ptr], %[offset_max]\n" "add x20, %x[args_ptr], %[offset_min]\n" + "ld1rw { z21.s }, p5/Z, [x21]\n" "ld1rw { z20.s }, p5/Z, [x20]\n" "fmin z8.s, p5/M, z8.s, z21.s\n" "fmin z9.s, p5/M, z9.s, z21.s\n" @@ -893,14 +895,14 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "st1w { z10.s }, p2, [x13, #2, MUL VL]\n" "st1w { z11.s }, p1, [x13, #3, MUL VL]\n" "addvl x13, x13, #4\n" - "st1w { z12.s }, p4, [x25]\n" - "st1w { z13.s }, p3, [x25, #1, MUL VL]\n" - "st1w { z14.s }, p2, [x25, #2, MUL VL]\n" - "st1w { z15.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 { z12.s }, p4, [x26]\n" + "st1w { z13.s }, p3, [x26, #1, MUL VL]\n" + "st1w { z14.s }, p2, [x26, #2, MUL VL]\n" + "st1w { z15.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" "42:" // Height 3: Writeback done "decw x14, ALL, MUL #4\n" "cmp x14, XZR\n" @@ -908,15 +910,15 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "b 86f\n" "43:" // Height 4 "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x15, %x[bias]\n" + "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n" "ldr x14, [%x[args_ptr], %[offsetof_N]]\n" "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" - "mov x13, %x[output_ptr]\n" + "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n" "44:" // Height 4: Column loop "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n" - "add x11, x12, x20, LSL #2\n" "cntw x21, ALL, MUL #3\n" + "add x11, x12, x20, LSL #2\n" "add x10, x11, x20, LSL #2\n" "add x9, x10, x20, LSL #2\n" "add x20, x9, x20, LSL #2\n" @@ -924,12 +926,12 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" "bgt 45f\n" "decw x21\n" - "cmp x14, x21\n" "mov x9, x12\n" + "cmp x14, x21\n" "bgt 45f\n" "decw x21\n" - "cmp x14, x21\n" "mov x10, x12\n" + "cmp x14, x21\n" "bgt 45f\n" "mov x11, x12\n" "45:" // Height 4: B setup done @@ -944,18 +946,18 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "cbz x15, 46f\n" "ld1w { z8.s }, p5/Z, [x15]\n" "ld1w { z9.s }, p5/Z, [x15, #1, MUL VL]\n" - "mov z12.d, z8.d\n" - "mov z13.d, z9.d\n" "ld1w { z10.s }, p5/Z, [x15, #2, MUL VL]\n" "ld1w { z11.s }, p5/Z, [x15, #3, MUL VL]\n" - "mov z14.d, z10.d\n" - "mov z15.d, z11.d\n" + "addvl x15, x15, #4\n" + "mov z12.d, z8.d\n" + "mov z13.d, z9.d\n" "mov z16.d, z8.d\n" "mov z17.d, z9.d\n" - "addvl x15, x15, #4\n" + "mov z20.d, z8.d\n" + "mov z14.d, z10.d\n" + "mov z15.d, z11.d\n" "mov z18.d, z10.d\n" "mov z19.d, z11.d\n" - "mov z20.d, z8.d\n" "mov z21.d, z9.d\n" "mov z22.d, z10.d\n" "mov z23.d, z11.d\n" @@ -963,13 +965,13 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "46:" // Height 4: no bias "tbz %x[flags], #0, 47f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x22, x13, x20, LSL #2\n" - "add x21, x22, x20, LSL #2\n" "ld1w { z8.s }, p4/Z, [x13]\n" - "add x20, x21, x20, LSL #2\n" "ld1w { z9.s }, p3/Z, [x13, #1, MUL VL]\n" "ld1w { z10.s }, p2/Z, [x13, #2, MUL VL]\n" "ld1w { z11.s }, p1/Z, [x13, #3, MUL VL]\n" + "add x22, x13, x20, LSL #2\n" + "add x21, x22, x20, LSL #2\n" + "add x20, x21, x20, LSL #2\n" "ld1w { z12.s }, p4/Z, [x22]\n" "ld1w { z13.s }, p3/Z, [x22, #1, MUL VL]\n" "ld1w { z14.s }, p2/Z, [x22, #2, MUL VL]\n" @@ -1004,8 +1006,8 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "mov x28, #0x0\n" "49:" // 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, 50f\n" "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -1030,25 +1032,25 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "ble 53f\n" "52:" // Height 4: Multiply loop: Main loop head "whilelt p0.s, XZR, x27\n" + "ld1w { z25.s }, p5/Z, [x12]\n" + "ld1w { z24.s }, p5/Z, [x11]\n" + "sub x27, x27, #0x4\n" + "cmp x27, #0x4\n" "ld1rqw { z3.s }, p0/Z, [x26]\n" + "add x26, x26, #0x10\n" "ld1rqw { z2.s }, p0/Z, [x25]\n" - "sub x27, x27, #0x4\n" + "add x25, x25, #0x10\n" "ld1rqw { z1.s }, p0/Z, [x24]\n" + "add x24, x24, #0x10\n" "ld1rqw { z0.s }, p0/Z, [x23]\n" - "cmp x27, #0x4\n" - "add x26, x26, #0x10\n" - "ld1w { z25.s }, p5/Z, [x12]\n" - "ld1w { z24.s }, p5/Z, [x11]\n" + "add x23, x23, #0x10\n" "fmla z8.s, z25.s, z3.s[0]\n" "fmla z12.s, z25.s, z2.s[0]\n" + "fmla z9.s, z24.s, z3.s[0]\n" + "fmla z13.s, z24.s, z2.s[0]\n" "fmla z16.s, z25.s, z1.s[0]\n" "fmla z20.s, z25.s, z0.s[0]\n" "ld1w { z25.s }, p5/Z, [x10]\n" - "add x25, x25, #0x10\n" - "fmla z9.s, z24.s, z3.s[0]\n" - "fmla z13.s, z24.s, z2.s[0]\n" - "add x24, x24, #0x10\n" - "add x23, x23, #0x10\n" "fmla z17.s, z24.s, z1.s[0]\n" "fmla z21.s, z24.s, z0.s[0]\n" "ld1w { z24.s }, p5/Z, [x9]\n" @@ -1127,22 +1129,22 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "bgt 52b\n" "53:" // Height 4: Multiply loop: Single iteration only "whilelt p0.s, XZR, x27\n" + "ld1w { z25.s }, p5/Z, [x12]\n" + "ld1w { z24.s }, p5/Z, [x11]\n" + "subs x27, x27, #0x1\n" + "addvl x12, x12, #1\n" + "addvl x11, x11, #1\n" "ld1rqw { z0.s }, p0/Z, [x26]\n" "ld1rqw { z1.s }, p0/Z, [x25]\n" - "subs x27, x27, #0x1\n" "ld1rqw { z2.s }, p0/Z, [x24]\n" "ld1rqw { z3.s }, p0/Z, [x23]\n" - "ld1w { z25.s }, p5/Z, [x12]\n" - "ld1w { z24.s }, p5/Z, [x11]\n" "fmla z8.s, z25.s, z0.s[0]\n" "fmla z12.s, z25.s, z1.s[0]\n" + "fmla z9.s, z24.s, z0.s[0]\n" + "fmla z13.s, z24.s, z1.s[0]\n" "fmla z16.s, z25.s, z2.s[0]\n" "fmla z20.s, z25.s, z3.s[0]\n" "ld1w { z25.s }, p5/Z, [x10]\n" - "addvl x12, x12, #1\n" - "fmla z9.s, z24.s, z0.s[0]\n" - "fmla z13.s, z24.s, z1.s[0]\n" - "addvl x11, x11, #1\n" "addvl x10, x10, #1\n" "fmla z17.s, z24.s, z2.s[0]\n" "fmla z21.s, z24.s, z3.s[0]\n" @@ -1159,23 +1161,23 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "ble 54f\n" "ld1w { z25.s }, p5/Z, [x12]\n" "ld1w { z24.s }, p5/Z, [x11]\n" + "subs x27, x27, #0x1\n" + "addvl x12, x12, #1\n" + "addvl x11, x11, #1\n" "fmla z8.s, z25.s, z0.s[1]\n" "fmla z12.s, z25.s, z1.s[1]\n" "fmla z16.s, z25.s, z2.s[1]\n" "fmla z20.s, z25.s, z3.s[1]\n" "ld1w { z25.s }, p5/Z, [x10]\n" - "subs x27, x27, #0x1\n" + "addvl x10, x10, #1\n" "fmla z9.s, z24.s, z0.s[1]\n" "fmla z13.s, z24.s, z1.s[1]\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "fmla z17.s, z24.s, z2.s[1]\n" "fmla z21.s, z24.s, z3.s[1]\n" "ld1w { z24.s }, p5/Z, [x9]\n" - "addvl x10, x10, #1\n" + "addvl x9, x9, #1\n" "fmla z10.s, z25.s, z0.s[1]\n" "fmla z14.s, z25.s, z1.s[1]\n" - "addvl x9, x9, #1\n" "fmla z18.s, z25.s, z2.s[1]\n" "fmla z22.s, z25.s, z3.s[1]\n" "fmla z11.s, z24.s, z0.s[1]\n" @@ -1185,23 +1187,23 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "ble 54f\n" "ld1w { z25.s }, p5/Z, [x12]\n" "ld1w { z24.s }, p5/Z, [x11]\n" + "subs x27, x27, #0x1\n" + "addvl x12, x12, #1\n" + "addvl x11, x11, #1\n" "fmla z8.s, z25.s, z0.s[2]\n" "fmla z12.s, z25.s, z1.s[2]\n" "fmla z16.s, z25.s, z2.s[2]\n" "fmla z20.s, z25.s, z3.s[2]\n" "ld1w { z25.s }, p5/Z, [x10]\n" - "subs x27, x27, #0x1\n" + "addvl x10, x10, #1\n" "fmla z9.s, z24.s, z0.s[2]\n" "fmla z13.s, z24.s, z1.s[2]\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "fmla z17.s, z24.s, z2.s[2]\n" "fmla z21.s, z24.s, z3.s[2]\n" "ld1w { z24.s }, p5/Z, [x9]\n" - "addvl x10, x10, #1\n" + "addvl x9, x9, #1\n" "fmla z10.s, z25.s, z0.s[2]\n" "fmla z14.s, z25.s, z1.s[2]\n" - "addvl x9, x9, #1\n" "fmla z18.s, z25.s, z2.s[2]\n" "fmla z22.s, z25.s, z3.s[2]\n" "fmla z11.s, z24.s, z0.s[2]\n" @@ -1211,16 +1213,16 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "ble 54f\n" "ld1w { z25.s }, p5/Z, [x12]\n" "ld1w { z24.s }, p5/Z, [x11]\n" + "addvl x12, x12, #1\n" + "addvl x11, x11, #1\n" "fmla z8.s, z25.s, z0.s[3]\n" "fmla z12.s, z25.s, z1.s[3]\n" "fmla z16.s, z25.s, z2.s[3]\n" "fmla z20.s, z25.s, z3.s[3]\n" "ld1w { z25.s }, p5/Z, [x10]\n" - "addvl x12, x12, #1\n" + "addvl x10, x10, #1\n" "fmla z9.s, z24.s, z0.s[3]\n" "fmla z13.s, z24.s, z1.s[3]\n" - "addvl x11, x11, #1\n" - "addvl x10, x10, #1\n" "fmla z17.s, z24.s, z2.s[3]\n" "fmla z21.s, z24.s, z3.s[3]\n" "ld1w { z24.s }, p5/Z, [x9]\n" @@ -1239,13 +1241,13 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "cmp x28, x20\n" "bne 49b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x13, x20, LSL #2\n" + "add x26, x13, x20, LSL #2\n" + "add x25, x26, x20, LSL #2\n" "add x24, x25, x20, LSL #2\n" - "add x23, x24, x20, LSL #2\n" "tbz %x[flags], #1, 55f\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 z8.s, p5/M, z8.s, z25.s\n" "fmin z9.s, p5/M, z9.s, z25.s\n" @@ -1285,18 +1287,18 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "st1w { z10.s }, p2, [x13, #2, MUL VL]\n" "st1w { z11.s }, p1, [x13, #3, MUL VL]\n" "addvl x13, x13, #4\n" - "st1w { z12.s }, p4, [x25]\n" - "st1w { z13.s }, p3, [x25, #1, MUL VL]\n" - "st1w { z14.s }, p2, [x25, #2, MUL VL]\n" - "st1w { z15.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 { z20.s }, p4, [x23]\n" - "st1w { z21.s }, p3, [x23, #1, MUL VL]\n" - "st1w { z22.s }, p2, [x23, #2, MUL VL]\n" - "st1w { z23.s }, p1, [x23, #3, MUL VL]\n" + "st1w { z12.s }, p4, [x26]\n" + "st1w { z13.s }, p3, [x26, #1, MUL VL]\n" + "st1w { z14.s }, p2, [x26, #2, MUL VL]\n" + "st1w { z15.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" + "st1w { z20.s }, p4, [x24]\n" + "st1w { z21.s }, p3, [x24, #1, MUL VL]\n" + "st1w { z22.s }, p2, [x24, #2, MUL VL]\n" + "st1w { z23.s }, p1, [x24, #3, MUL VL]\n" "56:" // Height 4: Writeback done "decw x14, ALL, MUL #4\n" "cmp x14, XZR\n" @@ -1304,15 +1306,15 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "b 86f\n" "57:" // Height 5 "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x15, %x[bias]\n" + "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n" "ldr x14, [%x[args_ptr], %[offsetof_N]]\n" "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" - "mov x13, %x[output_ptr]\n" + "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n" "58:" // Height 5: Column loop "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n" - "add x11, x12, x20, LSL #2\n" "cntw x21, ALL, MUL #3\n" + "add x11, x12, x20, LSL #2\n" "add x10, x11, x20, LSL #2\n" "add x9, x10, x20, LSL #2\n" "add x20, x9, x20, LSL #2\n" @@ -1320,12 +1322,12 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" "bgt 59f\n" "decw x21\n" - "cmp x14, x21\n" "mov x9, x12\n" + "cmp x14, x21\n" "bgt 59f\n" "decw x21\n" - "cmp x14, x21\n" "mov x10, x12\n" + "cmp x14, x21\n" "bgt 59f\n" "mov x11, x12\n" "59:" // Height 5: B setup done @@ -1340,18 +1342,18 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "cbz x15, 60f\n" "ld1w { z8.s }, p5/Z, [x15]\n" "ld1w { z9.s }, p5/Z, [x15, #1, MUL VL]\n" - "mov z12.d, z8.d\n" - "mov z13.d, z9.d\n" "ld1w { z10.s }, p5/Z, [x15, #2, MUL VL]\n" "ld1w { z11.s }, p5/Z, [x15, #3, MUL VL]\n" - "mov z14.d, z10.d\n" - "mov z15.d, z11.d\n" + "addvl x15, x15, #4\n" + "mov z12.d, z8.d\n" + "mov z13.d, z9.d\n" "mov z16.d, z8.d\n" "mov z17.d, z9.d\n" - "addvl x15, x15, #4\n" + "mov z20.d, z8.d\n" + "mov z14.d, z10.d\n" + "mov z15.d, z11.d\n" "mov z18.d, z10.d\n" "mov z19.d, z11.d\n" - "mov z20.d, z8.d\n" "mov z21.d, z9.d\n" "mov z22.d, z10.d\n" "mov z23.d, z11.d\n" @@ -1363,16 +1365,16 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "60:" // Height 5: no bias "tbz %x[flags], #0, 61f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x23, x13, x20, LSL #2\n" - "add x22, x23, x20, LSL #2\n" "ld1w { z8.s }, p4/Z, [x13]\n" - "add x21, x22, x20, LSL #2\n" - "add x20, x21, x20, LSL #2\n" "ld1w { z9.s }, p3/Z, [x13, #1, MUL VL]\n" "ld1w { z10.s }, p2/Z, [x13, #2, MUL VL]\n" "ld1w { z11.s }, p1/Z, [x13, #3, MUL VL]\n" + "add x23, x13, x20, LSL #2\n" + "add x22, x23, x20, LSL #2\n" + "add x21, x22, x20, LSL #2\n" "ld1w { z12.s }, p4/Z, [x23]\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" "ld1w { z15.s }, p1/Z, [x23, #3, MUL VL]\n" "ld1w { z16.s }, p4/Z, [x22]\n" @@ -1413,8 +1415,8 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "mov x28, #0x0\n" "63:" // 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, 64f\n" "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -1442,29 +1444,29 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "ble 67f\n" "66:" // Height 5: Multiply loop: Main loop head "whilelt p0.s, XZR, x27\n" + "ld1w { z29.s }, p5/Z, [x12]\n" + "ld1w { z28.s }, p5/Z, [x11]\n" + "sub x27, x27, #0x4\n" + "cmp x27, #0x4\n" "ld1rqw { z4.s }, p0/Z, [x26]\n" + "add x26, x26, #0x10\n" "ld1rqw { z3.s }, p0/Z, [x25]\n" - "sub x27, x27, #0x4\n" + "add x25, x25, #0x10\n" "ld1rqw { z2.s }, p0/Z, [x24]\n" + "add x24, x24, #0x10\n" "ld1rqw { z1.s }, p0/Z, [x23]\n" - "cmp x27, #0x4\n" - "add x26, x26, #0x10\n" "ld1rqw { z0.s }, p0/Z, [x22]\n" - "ld1w { z29.s }, p5/Z, [x12]\n" + "add x23, x23, #0x10\n" + "add x22, x22, #0x10\n" "fmla z8.s, z29.s, z4.s[0]\n" "fmla z12.s, z29.s, z3.s[0]\n" - "ld1w { z28.s }, p5/Z, [x11]\n" + "fmla z9.s, z28.s, z4.s[0]\n" "fmla z16.s, z29.s, z2.s[0]\n" "fmla z20.s, z29.s, z1.s[0]\n" - "add x25, x25, #0x10\n" "fmla z24.s, z29.s, z0.s[0]\n" - "fmla z9.s, z28.s, z4.s[0]\n" - "ld1w { z29.s }, p5/Z, [x10]\n" - "add x24, x24, #0x10\n" "fmla z13.s, z28.s, z3.s[0]\n" + "ld1w { z29.s }, p5/Z, [x10]\n" "fmla z17.s, z28.s, z2.s[0]\n" - "add x23, x23, #0x10\n" - "add x22, x22, #0x10\n" "fmla z21.s, z28.s, z1.s[0]\n" "fmla z25.s, z28.s, z0.s[0]\n" "ld1w { z28.s }, p5/Z, [x9]\n" @@ -1473,8 +1475,8 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "fmla z18.s, z29.s, z2.s[0]\n" "fmla z22.s, z29.s, z1.s[0]\n" "fmla z26.s, z29.s, z0.s[0]\n" - "fmla z11.s, z28.s, z4.s[0]\n" "ld1w { z29.s }, p5/Z, [x12, #1, MUL VL]\n" + "fmla z11.s, z28.s, z4.s[0]\n" "fmla z15.s, z28.s, z3.s[0]\n" "fmla z19.s, z28.s, z2.s[0]\n" "fmla z23.s, z28.s, z1.s[0]\n" @@ -1485,8 +1487,8 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "fmla z16.s, z29.s, z2.s[1]\n" "fmla z20.s, z29.s, z1.s[1]\n" "fmla z24.s, z29.s, z0.s[1]\n" - "fmla z9.s, z28.s, z4.s[1]\n" "ld1w { z29.s }, p5/Z, [x10, #1, MUL VL]\n" + "fmla z9.s, z28.s, z4.s[1]\n" "fmla z13.s, z28.s, z3.s[1]\n" "fmla z17.s, z28.s, z2.s[1]\n" "fmla z21.s, z28.s, z1.s[1]\n" @@ -1497,8 +1499,8 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "fmla z18.s, z29.s, z2.s[1]\n" "fmla z22.s, z29.s, z1.s[1]\n" "fmla z26.s, z29.s, z0.s[1]\n" - "fmla z11.s, z28.s, z4.s[1]\n" "ld1w { z29.s }, p5/Z, [x12, #2, MUL VL]\n" + "fmla z11.s, z28.s, z4.s[1]\n" "fmla z15.s, z28.s, z3.s[1]\n" "fmla z19.s, z28.s, z2.s[1]\n" "fmla z23.s, z28.s, z1.s[1]\n" @@ -1509,8 +1511,8 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "fmla z16.s, z29.s, z2.s[2]\n" "fmla z20.s, z29.s, z1.s[2]\n" "fmla z24.s, z29.s, z0.s[2]\n" - "fmla z9.s, z28.s, z4.s[2]\n" "ld1w { z29.s }, p5/Z, [x10, #2, MUL VL]\n" + "fmla z9.s, z28.s, z4.s[2]\n" "fmla z13.s, z28.s, z3.s[2]\n" "fmla z17.s, z28.s, z2.s[2]\n" "fmla z21.s, z28.s, z1.s[2]\n" @@ -1521,30 +1523,30 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "fmla z18.s, z29.s, z2.s[2]\n" "fmla z22.s, z29.s, z1.s[2]\n" "fmla z26.s, z29.s, z0.s[2]\n" - "fmla z11.s, z28.s, z4.s[2]\n" "ld1w { z29.s }, p5/Z, [x12, #3, MUL VL]\n" "addvl x12, x12, #4\n" + "fmla z11.s, z28.s, z4.s[2]\n" "fmla z15.s, z28.s, z3.s[2]\n" "fmla z19.s, z28.s, z2.s[2]\n" "fmla z23.s, z28.s, z1.s[2]\n" "fmla z27.s, z28.s, z0.s[2]\n" "ld1w { z28.s }, p5/Z, [x11, #3, MUL VL]\n" - "addvl x11, x11, #4\n" "fmla z8.s, z29.s, z4.s[3]\n" + "addvl x11, x11, #4\n" "fmla z12.s, z29.s, z3.s[3]\n" "fmla z16.s, z29.s, z2.s[3]\n" "fmla z20.s, z29.s, z1.s[3]\n" "fmla z24.s, z29.s, z0.s[3]\n" - "fmla z9.s, z28.s, z4.s[3]\n" "ld1w { z29.s }, p5/Z, [x10, #3, MUL VL]\n" "addvl x10, x10, #4\n" + "fmla z9.s, z28.s, z4.s[3]\n" "fmla z13.s, z28.s, z3.s[3]\n" "fmla z17.s, z28.s, z2.s[3]\n" "fmla z21.s, z28.s, z1.s[3]\n" "fmla z25.s, z28.s, z0.s[3]\n" "ld1w { z28.s }, p5/Z, [x9, #3, MUL VL]\n" - "addvl x9, x9, #4\n" "fmla z10.s, z29.s, z4.s[3]\n" + "addvl x9, x9, #4\n" "fmla z14.s, z29.s, z3.s[3]\n" "fmla z18.s, z29.s, z2.s[3]\n" "fmla z22.s, z29.s, z1.s[3]\n" @@ -1557,25 +1559,25 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "bgt 66b\n" "67:" // Height 5: Multiply loop: Single iteration only "whilelt p0.s, XZR, x27\n" + "ld1w { z29.s }, p5/Z, [x12]\n" + "ld1w { z28.s }, p5/Z, [x11]\n" + "subs x27, x27, #0x1\n" + "addvl x12, x12, #1\n" + "addvl x11, x11, #1\n" "ld1rqw { z0.s }, p0/Z, [x26]\n" "ld1rqw { z1.s }, p0/Z, [x25]\n" - "subs x27, x27, #0x1\n" "ld1rqw { z2.s }, p0/Z, [x24]\n" "ld1rqw { z3.s }, p0/Z, [x23]\n" "ld1rqw { z4.s }, p0/Z, [x22]\n" - "ld1w { z29.s }, p5/Z, [x12]\n" "fmla z8.s, z29.s, z0.s[0]\n" "fmla z12.s, z29.s, z1.s[0]\n" - "ld1w { z28.s }, p5/Z, [x11]\n" + "fmla z9.s, z28.s, z0.s[0]\n" + "fmla z13.s, z28.s, z1.s[0]\n" "fmla z16.s, z29.s, z2.s[0]\n" "fmla z20.s, z29.s, z3.s[0]\n" - "addvl x12, x12, #1\n" "fmla z24.s, z29.s, z4.s[0]\n" - "fmla z9.s, z28.s, z0.s[0]\n" - "ld1w { z29.s }, p5/Z, [x10]\n" - "addvl x11, x11, #1\n" - "fmla z13.s, z28.s, z1.s[0]\n" "fmla z17.s, z28.s, z2.s[0]\n" + "ld1w { z29.s }, p5/Z, [x10]\n" "addvl x10, x10, #1\n" "fmla z21.s, z28.s, z3.s[0]\n" "fmla z25.s, z28.s, z4.s[0]\n" @@ -1594,19 +1596,19 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "ble 68f\n" "ld1w { z29.s }, p5/Z, [x12]\n" "ld1w { z28.s }, p5/Z, [x11]\n" + "subs x27, x27, #0x1\n" + "addvl x12, x12, #1\n" + "addvl x11, x11, #1\n" "fmla z8.s, z29.s, z0.s[1]\n" "fmla z12.s, z29.s, z1.s[1]\n" "fmla z16.s, z29.s, z2.s[1]\n" "fmla z20.s, z29.s, z3.s[1]\n" - "subs x27, x27, #0x1\n" - "addvl x12, x12, #1\n" "fmla z24.s, z29.s, z4.s[1]\n" "fmla z9.s, z28.s, z0.s[1]\n" "ld1w { z29.s }, p5/Z, [x10]\n" - "addvl x11, x11, #1\n" + "addvl x10, x10, #1\n" "fmla z13.s, z28.s, z1.s[1]\n" "fmla z17.s, z28.s, z2.s[1]\n" - "addvl x10, x10, #1\n" "fmla z21.s, z28.s, z3.s[1]\n" "fmla z25.s, z28.s, z4.s[1]\n" "ld1w { z28.s }, p5/Z, [x9]\n" @@ -1624,19 +1626,19 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "ble 68f\n" "ld1w { z29.s }, p5/Z, [x12]\n" "ld1w { z28.s }, p5/Z, [x11]\n" + "subs x27, x27, #0x1\n" + "addvl x12, x12, #1\n" + "addvl x11, x11, #1\n" "fmla z8.s, z29.s, z0.s[2]\n" "fmla z12.s, z29.s, z1.s[2]\n" "fmla z16.s, z29.s, z2.s[2]\n" "fmla z20.s, z29.s, z3.s[2]\n" - "subs x27, x27, #0x1\n" - "addvl x12, x12, #1\n" "fmla z24.s, z29.s, z4.s[2]\n" "fmla z9.s, z28.s, z0.s[2]\n" "ld1w { z29.s }, p5/Z, [x10]\n" - "addvl x11, x11, #1\n" + "addvl x10, x10, #1\n" "fmla z13.s, z28.s, z1.s[2]\n" "fmla z17.s, z28.s, z2.s[2]\n" - "addvl x10, x10, #1\n" "fmla z21.s, z28.s, z3.s[2]\n" "fmla z25.s, z28.s, z4.s[2]\n" "ld1w { z28.s }, p5/Z, [x9]\n" @@ -1654,12 +1656,12 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "ble 68f\n" "ld1w { z29.s }, p5/Z, [x12]\n" "ld1w { z28.s }, p5/Z, [x11]\n" + "addvl x12, x12, #1\n" + "addvl x11, x11, #1\n" "fmla z8.s, z29.s, z0.s[3]\n" "fmla z12.s, z29.s, z1.s[3]\n" "fmla z16.s, z29.s, z2.s[3]\n" "fmla z20.s, z29.s, z3.s[3]\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "fmla z24.s, z29.s, z4.s[3]\n" "fmla z9.s, z28.s, z0.s[3]\n" "ld1w { z29.s }, p5/Z, [x10]\n" @@ -1686,14 +1688,14 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "cmp x28, x20\n" "bne 63b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x13, x20, LSL #2\n" + "add x26, x13, x20, LSL #2\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" "tbz %x[flags], #1, 69f\n" - "add x20, %x[args_ptr], %[offset_max]\n" - "ld1rw { z29.s }, p5/Z, [x20]\n" + "add x21, %x[args_ptr], %[offset_max]\n" "add x20, %x[args_ptr], %[offset_min]\n" + "ld1rw { z29.s }, p5/Z, [x21]\n" "ld1rw { z28.s }, p5/Z, [x20]\n" "fmin z8.s, p5/M, z8.s, z29.s\n" "fmin z9.s, p5/M, z9.s, z29.s\n" @@ -1741,22 +1743,22 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "st1w { z10.s }, p2, [x13, #2, MUL VL]\n" "st1w { z11.s }, p1, [x13, #3, MUL VL]\n" "addvl x13, x13, #4\n" - "st1w { z12.s }, p4, [x25]\n" - "st1w { z13.s }, p3, [x25, #1, MUL VL]\n" - "st1w { z14.s }, p2, [x25, #2, MUL VL]\n" - "st1w { z15.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 { z20.s }, p4, [x23]\n" - "st1w { z21.s }, p3, [x23, #1, MUL VL]\n" - "st1w { z22.s }, p2, [x23, #2, MUL VL]\n" - "st1w { z23.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 { z12.s }, p4, [x26]\n" + "st1w { z13.s }, p3, [x26, #1, MUL VL]\n" + "st1w { z14.s }, p2, [x26, #2, MUL VL]\n" + "st1w { z15.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" + "st1w { z20.s }, p4, [x24]\n" + "st1w { z21.s }, p3, [x24, #1, MUL VL]\n" + "st1w { z22.s }, p2, [x24, #2, MUL VL]\n" + "st1w { z23.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" "70:" // Height 5: Writeback done "decw x14, ALL, MUL #4\n" "cmp x14, XZR\n" @@ -1764,18 +1766,19 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "b 86f\n" "71:" // Height 6 "ldr x20, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "ldr x15, [%x[args_ptr], %[offsetof_bias]]\n" + "mov x21, #0x18\n" "ldr x14, [%x[args_ptr], %[offsetof_N]]\n" "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" - "mov x21, #0x18\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "mov x15, %x[bias]\n" - "mov x13, %x[output_ptr]\n" - "madd %x[output_ptr], x20, x21, %x[output_ptr]\n" + "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n" + "madd x21, x20, x21, x13\n" + "str x21, [%x[args_ptr], %[offsetof_output_ptr]]\n" "72:" // Height 6: Column loop "ldr x12, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n" - "add x11, x12, x20, LSL #2\n" "cntw x21, ALL, MUL #3\n" + "add x11, x12, x20, LSL #2\n" "add x10, x11, x20, LSL #2\n" "add x9, x10, x20, LSL #2\n" "add x20, x9, x20, LSL #2\n" @@ -1783,12 +1786,12 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" "bgt 73f\n" "decw x21\n" - "cmp x14, x21\n" "mov x9, x12\n" + "cmp x14, x21\n" "bgt 73f\n" "decw x21\n" - "cmp x14, x21\n" "mov x10, x12\n" + "cmp x14, x21\n" "bgt 73f\n" "mov x11, x12\n" "73:" // Height 6: B setup done @@ -1803,18 +1806,18 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "cbz x15, 74f\n" "ld1w { z8.s }, p5/Z, [x15]\n" "ld1w { z9.s }, p5/Z, [x15, #1, MUL VL]\n" - "mov z12.d, z8.d\n" - "mov z13.d, z9.d\n" "ld1w { z10.s }, p5/Z, [x15, #2, MUL VL]\n" "ld1w { z11.s }, p5/Z, [x15, #3, MUL VL]\n" - "mov z14.d, z10.d\n" - "mov z15.d, z11.d\n" + "addvl x15, x15, #4\n" + "mov z12.d, z8.d\n" + "mov z13.d, z9.d\n" "mov z16.d, z8.d\n" "mov z17.d, z9.d\n" - "addvl x15, x15, #4\n" + "mov z20.d, z8.d\n" + "mov z14.d, z10.d\n" + "mov z15.d, z11.d\n" "mov z18.d, z10.d\n" "mov z19.d, z11.d\n" - "mov z20.d, z8.d\n" "mov z21.d, z9.d\n" "mov z22.d, z10.d\n" "mov z23.d, z11.d\n" @@ -1830,17 +1833,17 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "74:" // Height 6: no bias "tbz %x[flags], #0, 75f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x24, x13, x20, LSL #2\n" - "add x23, x24, x20, LSL #2\n" "ld1w { z8.s }, p4/Z, [x13]\n" - "add x22, x23, x20, LSL #2\n" - "add x21, x22, x20, LSL #2\n" "ld1w { z9.s }, p3/Z, [x13, #1, MUL VL]\n" "ld1w { z10.s }, p2/Z, [x13, #2, MUL VL]\n" - "add x20, x21, x20, LSL #2\n" "ld1w { z11.s }, p1/Z, [x13, #3, MUL VL]\n" + "add x24, x13, x20, LSL #2\n" + "add x23, x24, x20, LSL #2\n" + "add x22, x23, x20, LSL #2\n" "ld1w { z12.s }, p4/Z, [x24]\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" "ld1w { z15.s }, p1/Z, [x24, #3, MUL VL]\n" "ld1w { z16.s }, p4/Z, [x23]\n" @@ -1889,8 +1892,8 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "mov x28, #0x0\n" "77:" // 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, 78f\n" "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -1921,29 +1924,29 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "ble 81f\n" "80:" // Height 6: Multiply loop: Main loop head "whilelt p0.s, XZR, x27\n" + "ld1w { z1.s }, p5/Z, [x12]\n" + "ld1w { z0.s }, p5/Z, [x11]\n" + "sub x27, x27, #0x4\n" + "cmp x27, #0x4\n" "ld1rqw { z7.s }, p0/Z, [x26]\n" + "add x26, x26, #0x10\n" "ld1rqw { z6.s }, p0/Z, [x25]\n" - "sub x27, x27, #0x4\n" + "add x25, x25, #0x10\n" "ld1rqw { z5.s }, p0/Z, [x24]\n" + "add x24, x24, #0x10\n" "ld1rqw { z4.s }, p0/Z, [x23]\n" - "cmp x27, #0x4\n" - "add x26, x26, #0x10\n" "ld1rqw { z3.s }, p0/Z, [x22]\n" "ld1rqw { z2.s }, p0/Z, [x21]\n" - "add x25, x25, #0x10\n" - "add x24, x24, #0x10\n" - "ld1w { z1.s }, p5/Z, [x12]\n" - "ld1w { z0.s }, p5/Z, [x11]\n" + "add x23, x23, #0x10\n" + "add x22, x22, #0x10\n" "fmla z8.s, z1.s, z7.s[0]\n" "fmla z12.s, z1.s, z6.s[0]\n" + "add x21, x21, #0x10\n" "fmla z16.s, z1.s, z5.s[0]\n" "fmla z20.s, z1.s, z4.s[0]\n" - "add x23, x23, #0x10\n" - "add x22, x22, #0x10\n" "fmla z24.s, z1.s, z3.s[0]\n" "fmla z28.s, z1.s, z2.s[0]\n" "ld1w { z1.s }, p5/Z, [x10]\n" - "add x21, x21, #0x10\n" "fmla z9.s, z0.s, z7.s[0]\n" "fmla z13.s, z0.s, z6.s[0]\n" "fmla z17.s, z0.s, z5.s[0]\n" @@ -2054,27 +2057,27 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "bgt 80b\n" "81:" // Height 6: Multiply loop: Single iteration only "whilelt p0.s, XZR, x27\n" + "ld1w { z7.s }, p5/Z, [x12]\n" + "ld1w { z6.s }, p5/Z, [x11]\n" + "subs x27, x27, #0x1\n" + "addvl x12, x12, #1\n" + "addvl x11, x11, #1\n" "ld1rqw { z0.s }, p0/Z, [x26]\n" "ld1rqw { z1.s }, p0/Z, [x25]\n" - "subs x27, x27, #0x1\n" "ld1rqw { z2.s }, p0/Z, [x24]\n" "ld1rqw { z3.s }, p0/Z, [x23]\n" "ld1rqw { z4.s }, p0/Z, [x22]\n" "ld1rqw { z5.s }, p0/Z, [x21]\n" - "ld1w { z7.s }, p5/Z, [x12]\n" - "ld1w { z6.s }, p5/Z, [x11]\n" "fmla z8.s, z7.s, z0.s[0]\n" "fmla z12.s, z7.s, z1.s[0]\n" + "fmla z9.s, z6.s, z0.s[0]\n" + "fmla z13.s, z6.s, z1.s[0]\n" "fmla z16.s, z7.s, z2.s[0]\n" "fmla z20.s, z7.s, z3.s[0]\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "fmla z24.s, z7.s, z4.s[0]\n" "fmla z28.s, z7.s, z5.s[0]\n" "ld1w { z7.s }, p5/Z, [x10]\n" "addvl x10, x10, #1\n" - "fmla z9.s, z6.s, z0.s[0]\n" - "fmla z13.s, z6.s, z1.s[0]\n" "fmla z17.s, z6.s, z2.s[0]\n" "fmla z21.s, z6.s, z3.s[0]\n" "fmla z25.s, z6.s, z4.s[0]\n" @@ -2096,19 +2099,19 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "ble 82f\n" "ld1w { z7.s }, p5/Z, [x12]\n" "ld1w { z6.s }, p5/Z, [x11]\n" + "subs x27, x27, #0x1\n" + "addvl x12, x12, #1\n" + "addvl x11, x11, #1\n" "fmla z8.s, z7.s, z0.s[1]\n" "fmla z12.s, z7.s, z1.s[1]\n" "fmla z16.s, z7.s, z2.s[1]\n" "fmla z20.s, z7.s, z3.s[1]\n" - "subs x27, x27, #0x1\n" - "addvl x12, x12, #1\n" "fmla z24.s, z7.s, z4.s[1]\n" "fmla z28.s, z7.s, z5.s[1]\n" "ld1w { z7.s }, p5/Z, [x10]\n" - "addvl x11, x11, #1\n" + "addvl x10, x10, #1\n" "fmla z9.s, z6.s, z0.s[1]\n" "fmla z13.s, z6.s, z1.s[1]\n" - "addvl x10, x10, #1\n" "fmla z17.s, z6.s, z2.s[1]\n" "fmla z21.s, z6.s, z3.s[1]\n" "fmla z25.s, z6.s, z4.s[1]\n" @@ -2130,19 +2133,19 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "ble 82f\n" "ld1w { z7.s }, p5/Z, [x12]\n" "ld1w { z6.s }, p5/Z, [x11]\n" + "subs x27, x27, #0x1\n" + "addvl x12, x12, #1\n" + "addvl x11, x11, #1\n" "fmla z8.s, z7.s, z0.s[2]\n" "fmla z12.s, z7.s, z1.s[2]\n" "fmla z16.s, z7.s, z2.s[2]\n" "fmla z20.s, z7.s, z3.s[2]\n" - "subs x27, x27, #0x1\n" - "addvl x12, x12, #1\n" "fmla z24.s, z7.s, z4.s[2]\n" "fmla z28.s, z7.s, z5.s[2]\n" "ld1w { z7.s }, p5/Z, [x10]\n" - "addvl x11, x11, #1\n" + "addvl x10, x10, #1\n" "fmla z9.s, z6.s, z0.s[2]\n" "fmla z13.s, z6.s, z1.s[2]\n" - "addvl x10, x10, #1\n" "fmla z17.s, z6.s, z2.s[2]\n" "fmla z21.s, z6.s, z3.s[2]\n" "fmla z25.s, z6.s, z4.s[2]\n" @@ -2164,12 +2167,12 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "ble 82f\n" "ld1w { z7.s }, p5/Z, [x12]\n" "ld1w { z6.s }, p5/Z, [x11]\n" + "addvl x12, x12, #1\n" + "addvl x11, x11, #1\n" "fmla z8.s, z7.s, z0.s[3]\n" "fmla z12.s, z7.s, z1.s[3]\n" "fmla z16.s, z7.s, z2.s[3]\n" "fmla z20.s, z7.s, z3.s[3]\n" - "addvl x12, x12, #1\n" - "addvl x11, x11, #1\n" "fmla z24.s, z7.s, z4.s[3]\n" "fmla z28.s, z7.s, z5.s[3]\n" "ld1w { z7.s }, p5/Z, [x10]\n" @@ -2200,15 +2203,15 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "cmp x28, x20\n" "bne 77b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x13, x20, LSL #2\n" + "add x26, x13, x20, LSL #2\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" "tbz %x[flags], #1, 83f\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 z8.s, p5/M, z8.s, z1.s\n" "fmin z9.s, p5/M, z9.s, z1.s\n" @@ -2264,26 +2267,26 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "st1w { z10.s }, p2, [x13, #2, MUL VL]\n" "st1w { z11.s }, p1, [x13, #3, MUL VL]\n" "addvl x13, x13, #4\n" - "st1w { z12.s }, p4, [x25]\n" - "st1w { z13.s }, p3, [x25, #1, MUL VL]\n" - "st1w { z14.s }, p2, [x25, #2, MUL VL]\n" - "st1w { z15.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 { z20.s }, p4, [x23]\n" - "st1w { z21.s }, p3, [x23, #1, MUL VL]\n" - "st1w { z22.s }, p2, [x23, #2, MUL VL]\n" - "st1w { z23.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 { z28.s }, p4, [x21]\n" - "st1w { z29.s }, p3, [x21, #1, MUL VL]\n" - "st1w { z30.s }, p2, [x21, #2, MUL VL]\n" - "st1w { z31.s }, p1, [x21, #3, MUL VL]\n" + "st1w { z12.s }, p4, [x26]\n" + "st1w { z13.s }, p3, [x26, #1, MUL VL]\n" + "st1w { z14.s }, p2, [x26, #2, MUL VL]\n" + "st1w { z15.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" + "st1w { z20.s }, p4, [x24]\n" + "st1w { z21.s }, p3, [x24, #1, MUL VL]\n" + "st1w { z22.s }, p2, [x24, #2, MUL VL]\n" + "st1w { z23.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" + "st1w { z28.s }, p4, [x22]\n" + "st1w { z29.s }, p3, [x22, #1, MUL VL]\n" + "st1w { z30.s }, p2, [x22, #2, MUL VL]\n" + "st1w { z31.s }, p1, [x22, #3, MUL VL]\n" "84:" // Height 6: Writeback done "decw x14, ALL, MUL #4\n" "cmp x14, XZR\n" @@ -2300,8 +2303,8 @@ void sve_ffhybrid_fp32_mla_6x4VL ( "madd %x[input_ptr], x20, x21, %x[input_ptr]\n" "b 1b\n" "86:" // 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_B_stride] "I" (offsetof(KernelArgs, B_stride)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_cur_B_ptr] "I" (offsetof(KernelArgs, cur_B_ptr)), [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_B_stride] "I" (offsetof(KernelArgs, B_stride)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_bias] "I" (offsetof(KernelArgs, bias)), [offsetof_cur_B_ptr] "I" (offsetof(KernelArgs, cur_B_ptr)), [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", "x13", "x14", "x15", "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" ); } |