diff options
author | Viet-Hoa Do <viet-hoa.do@arm.com> | 2022-06-01 11:47:14 +0100 |
---|---|---|
committer | Viet-Hoa Do <viet-hoa.do@arm.com> | 2022-11-28 16:57:42 +0000 |
commit | 03b2971ac69a86f10a1566938d1a25afee15746c (patch) | |
tree | aec7cfc047e1da278b4b71a706cda7b1b0faa158 /src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32bf16fp32_planar_5x5_s2_4rows_dot_za/generic.cpp | |
parent | 7dc0234331f2150a6b4ac5c2b49de419870f7cf5 (diff) | |
download | ComputeLibrary-03b2971ac69a86f10a1566938d1a25afee15746c.tar.gz |
Integrate SME2 kernels
* Add SME/SME2 detection.
* Integrate SME2 implementation for:
- Normal convolution
- Winograd
- Depthwise convolution
- Pooling
Resolves: COMPMID-5700
Signed-off-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Change-Id: I2f1ca1d05f8cfeee9309ed1c0a36096a4a6aad5c
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8692
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32bf16fp32_planar_5x5_s2_4rows_dot_za/generic.cpp')
-rw-r--r-- | src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32bf16fp32_planar_5x5_s2_4rows_dot_za/generic.cpp | 1246 |
1 files changed, 1246 insertions, 0 deletions
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32bf16fp32_planar_5x5_s2_4rows_dot_za/generic.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32bf16fp32_planar_5x5_s2_4rows_dot_za/generic.cpp new file mode 100644 index 0000000000..eae8994166 --- /dev/null +++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32bf16fp32_planar_5x5_s2_4rows_dot_za/generic.cpp @@ -0,0 +1,1246 @@ +/* + * Copyright (c) 2022 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + +#if defined(ARM_COMPUTE_ENABLE_SME2) + +#include <algorithm> +#include <cstddef> + +namespace arm_conv { +namespace depthwise { + +void sme2_fp32bf16fp32_planar_5x5_s2_4rows_dot_za_impl( + const float *inptr, + size_t ld_in_row, + size_t ld_in_col, + size_t ld_in_vl, + unsigned int pad_top, + unsigned int valid_input_rows, + unsigned int pad_left, + unsigned int valid_input_cols, + const float *weights, + const float *bias, + float **outptrs, + const size_t *outlds, + const size_t *outvllds, + unsigned int output_cols, + unsigned int start_channel, + unsigned int valid_channels, + float act_min, + float act_max +) +{ + struct Args + { + const float *inptr; + size_t ld_in_vl; + long unsigned int pad_top, pad_bottom, pad_left; + const float *weights; + const float *bias; + long unsigned int input_cols, output_cols; + float **outptrs; + const size_t *ld_out_cols; + const size_t *ld_out_vls; + long unsigned int current_channel, n_channels; + float clamp_min, clamp_max; + }; + + Args args = { inptr, ld_in_vl, pad_top, 11u - std::min(11u, pad_top + valid_input_rows), pad_left, weights, bias, valid_input_cols, output_cols, outptrs, outlds, outvllds, start_channel, valid_channels, act_min, act_max }; + + __asm__ __volatile__( + "ldr x4, [%x[args], %[offsetof_Args_pad_bottom]]\n" + "mov x19, #0xb\n" + ".inst 0xd503477f // SMSTART ZA\n" + "sub x19, x19, x4\n" + "ldr x5, [%x[args], %[offsetof_Args_pad_top]]\n" + "ptrue p2.b\n" + ".inst 0x25207812 // ptrue pn10.b\n" + "ld1rw { z30.s }, p2/Z, [%x[args], %[offsetof_Args_clamp_min]]\n" + "ldr x6, [%x[args], %[offsetof_Args_n_channels]]\n" + "whilelt p1.s, XZR, x6\n" + "whilelt p9.s, XZR, x19\n" + "ld1rw { z22.s }, p2/Z, [%x[args], %[offsetof_Args_clamp_max]]\n" + "whilelt p8.s, XZR, x5\n" + "addvl SP, SP, #-15\n" + "ldr x7, [%x[args], %[offsetof_Args_current_channel]]\n" + "eor p8.b, p2/Z, p8.b, p9.b\n" + "1:" // Channel loop + "ldr x19, [%x[args], %[offsetof_Args_bias]]\n" + "fmov z4.s, #0x0\n" + "cbz x19, 2f\n" + "ld1w { z4.s }, p1/Z, [x19, x7, LSL #2]\n" + "2:" // Load bias: Done + "ldr x19, [%x[args], %[offsetof_Args_weights]]\n" + "mov x20, x19\n" + "ld1w { z31.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" + "ld1w { z16.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" + ".inst 0x658aabe1 // bfcvt z1.h, p2/M, z31.s\n" + "incb x19\n" + "ld1w { z13.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" + ".inst 0x658aa9a9 // bfcvt z9.h, p2/M, z13.s\n" + "addvl x23, SP, #15\n" + "ld1w { z18.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" + ".inst 0x648aaa01 // bfcvtnt z1.h, p2/M, z16.s\n" + "addvl x23, x23, #-3\n" + "ld1w { z15.s }, p2/Z, [x20]\n" + "mov x20, x19\n" + "st1h { z1.h }, p2, [x23]\n" + ".inst 0x648aaa49 // bfcvtnt z9.h, p2/M, z18.s\n" + "ld1w { z31.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" + ".inst 0x658aabe1 // bfcvt z1.h, p2/M, z31.s\n" + "incb x19\n" + "ld1w { z16.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" + "st1h { z9.h }, p2, [x23, #1, MUL VL]\n" + ".inst 0x658aa9e2 // bfcvt z2.h, p2/M, z15.s\n" + "ld1w { z13.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" + ".inst 0x658aa9a9 // bfcvt z9.h, p2/M, z13.s\n" + ".inst 0x648aaa01 // bfcvtnt z1.h, p2/M, z16.s\n" + "ld1w { z18.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" + "st1h { z2.h }, p2, [x23, #2, MUL VL]\n" + "addvl x23, x23, #-3\n" + "ld1w { z15.s }, p2/Z, [x20]\n" + "mov x20, x19\n" + "st1h { z1.h }, p2, [x23]\n" + ".inst 0x648aaa49 // bfcvtnt z9.h, p2/M, z18.s\n" + "ld1w { z31.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" + "incb x19\n" + ".inst 0x658aabe1 // bfcvt z1.h, p2/M, z31.s\n" + "ld1w { z16.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" + ".inst 0x658aa9e2 // bfcvt z2.h, p2/M, z15.s\n" + "st1h { z9.h }, p2, [x23, #1, MUL VL]\n" + "ld1w { z13.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" + ".inst 0x658aa9a9 // bfcvt z9.h, p2/M, z13.s\n" + "st1h { z2.h }, p2, [x23, #2, MUL VL]\n" + "ld1w { z18.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" + "addvl x23, x23, #-3\n" + ".inst 0x648aaa01 // bfcvtnt z1.h, p2/M, z16.s\n" + "ld1w { z15.s }, p2/Z, [x20]\n" + "mov x20, x19\n" + "st1h { z1.h }, p2, [x23]\n" + ".inst 0x648aaa49 // bfcvtnt z9.h, p2/M, z18.s\n" + "ld1w { z31.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" + ".inst 0x658aabe1 // bfcvt z1.h, p2/M, z31.s\n" + "incb x19\n" + "ld1w { z16.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" + ".inst 0x658aa9e2 // bfcvt z2.h, p2/M, z15.s\n" + "st1h { z9.h }, p2, [x23, #1, MUL VL]\n" + "ld1w { z13.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" + ".inst 0x648aaa01 // bfcvtnt z1.h, p2/M, z16.s\n" + ".inst 0x658aa9a9 // bfcvt z9.h, p2/M, z13.s\n" + "ld1w { z18.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" + "ldr x17, [%x[args], %[offsetof_Args_input_cols]]\n" + "st1h { z2.h }, p2, [x23, #2, MUL VL]\n" + "ld1w { z15.s }, p2/Z, [x20]\n" + "mov x20, x19\n" + "addvl x23, x23, #-3\n" + "st1h { z1.h }, p2, [x23]\n" + "ld1w { z31.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" + ".inst 0x648aaa49 // bfcvtnt z9.h, p2/M, z18.s\n" + "st1h { z9.h }, p2, [x23, #1, MUL VL]\n" + "ld1w { z16.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" + ".inst 0x658aabe1 // bfcvt z1.h, p2/M, z31.s\n" + ".inst 0x658aa9e2 // bfcvt z2.h, p2/M, z15.s\n" + "ld1w { z13.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" + ".inst 0x658aa9a9 // bfcvt z9.h, p2/M, z13.s\n" + "ldr x16, [%x[args], %[offsetof_Args_inptr]]\n" + "ld1w { z18.s }, p2/Z, [x20]\n" + "incb x20, ALL, MUL #5\n" + "sub x19, x17, #0x1\n" + "st1h { z2.h }, p2, [x23, #2, MUL VL]\n" + "ld1w { z15.s }, p2/Z, [x20]\n" + "orr x22, x19, %x[ld_in_col], LSL #18\n" + "addvl x23, x23, #-3\n" + "mov z5.d, z4.d\n" + "orr x22, x6, x22, LSL #20\n" + "mov x21, #0xb\n" + "mov z6.d, z4.d\n" + "mov z7.d, z4.d\n" + "add x20, x5, x4\n" + "lsl x19, %x[ld_in_row], #0x2\n" + ".inst 0x648aaa01 // bfcvtnt z1.h, p2/M, z16.s\n" + "st1h { z1.h }, p2, [x23]\n" + ".inst 0x648aaa49 // bfcvtnt z9.h, p2/M, z18.s\n" + "st1h { z9.h }, p2, [x23, #1, MUL VL]\n" + ".inst 0x658aa9e2 // bfcvt z2.h, p2/M, z15.s\n" + "mov x8, #0x0\n" + "st1h { z2.h }, p2, [x23, #2, MUL VL]\n" + "ldr x15, [%x[args], %[offsetof_Args_output_cols]]\n" + "lsl x22, x22, #0x2\n" + "sub x21, x21, x20\n" + "madd x19, x19, x5, x16\n" + "3:" // Issue prefetches + "subs x21, x21, #0x1\n" + ".inst 0xf8b64a7c // rprfm pldstrm, x22, [x19]\n" + "add x19, x19, %x[ld_in_col], LSL #2\n" + "bgt 3b\n" + "ldr x24, [%x[args], %[offsetof_Args_outptrs]]\n" + "lsl x19, %x[ld_in_row], #0x2\n" + "msub x16, x5, x19, x16\n" + ".inst 0xc0040c80 // mova za.d[x8, #0], { z4.d-z7.d }\n" + "ldr x19, [%x[args], %[offsetof_Args_ld_out_cols]]\n" + ".inst 0xc0040c81 // mova za.d[x8, #1], { z4.d-z7.d }\n" + "mov x21, #0x4\n" + "ldp x14, x13, [x24], #0x10\n" + ".inst 0xc0040c82 // mova za.d[x8, #2], { z4.d-z7.d }\n" + "ldp x11, x10, [x19], #0x10\n" + ".inst 0xc0040c83 // mova za.d[x8, #3], { z4.d-z7.d }\n" + "ldr x20, [%x[args], %[offsetof_Args_pad_left]]\n" + ".inst 0xc0040c84 // mova za.d[x8, #4], { z4.d-z7.d }\n" + "ldp x9, x28, [x24], #0x10\n" + "ldp x27, x26, [x19], #0x10\n" + "cbz x20, 5f\n" + "cmp x20, x21\n" + "csel x19, x20, x21, LT\n" + "sub x20, x20, x19\n" + "sub x21, x21, x19\n" + "cbz x20, 5f\n" + ".inst 0xc0060c18 // mova { z24.d-z27.d }, za.d[x8, #0]\n" + "and x21, x20, #0x1\n" + "add x20, x20, #0x1\n" + ".inst 0xc1b6cbd8 // fclamp { z24.s-z27.s }, z30.s, z22.s\n" + "lsr x20, x20, #0x1\n" + "sub x15, x15, x20\n" + "4:" // Left padding + "subs x20, x20, #0x1\n" + "st1w { z24.s }, p1, [x14]\n" + "add x14, x14, x11, LSL #2\n" + "st1w { z25.s }, p1, [x13]\n" + "add x13, x13, x10, LSL #2\n" + "st1w { z26.s }, p1, [x9]\n" + "add x9, x9, x27, LSL #2\n" + "st1w { z27.s }, p1, [x28]\n" + "add x28, x28, x26, LSL #2\n" + "bgt 4b\n" + "5:" // Left padding: End + "adds XZR, x5, x4\n" + "bne 12f\n" + "cbz x21, 10f\n" + "cmp x21, #0x1\n" + "sub x17, x17, x21\n" + "beq 9f\n" + "cmp x21, #0x2\n" + "beq 8f\n" + "cmp x21, #0x3\n" + "beq 7f\n" + "6:" // Unpadded: 4 priming loads + "add x20, x16, %x[ld_in_row], LSL #2\n" + "ld1w { z23.s }, p1/Z, [x16]\n" + ".inst 0x658aaaea // bfcvt z10.h, p2/M, z23.s\n" + "addvl x19, SP, #12\n" + "ld1w { z16.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x648aaa0a // bfcvtnt z10.h, p2/M, z16.s\n" + "add x16, x16, %x[ld_in_col], LSL #2\n" + "ld1w { z16.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x658aaa0b // bfcvt z11.h, p2/M, z16.s\n" + "ld1w { z16.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x648aaa0b // bfcvtnt z11.h, p2/M, z16.s\n" + "ld1w { z16.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x658aaa0c // bfcvt z12.h, p2/M, z16.s\n" + "ld1w { z16.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x648aaa0c // bfcvtnt z12.h, p2/M, z16.s\n" + "ld1w { z16.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x658aaa0d // bfcvt z13.h, p2/M, z16.s\n" + "ld1w { z16.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x648aaa0d // bfcvtnt z13.h, p2/M, z16.s\n" + "ld1w { z16.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x658aaa0e // bfcvt z14.h, p2/M, z16.s\n" + ".inst 0xa1402a61 // ld1h { z1.h, z9.h }, pn10.b/Z, [x19]\n" + ".inst 0xc1311150 // bfdot za.s[x8, 0], { z10.h-z13.h }, z1.h\n" + "ld1w { z16.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x648aaa0e // bfcvtnt z14.h, p2/M, z16.s\n" + ".inst 0xc1391170 // bfdot za.s[x8, 0], { z11.h-z14.h }, z9.h\n" + "ld1w { z16.s }, p1/Z, [x20]\n" + ".inst 0x658aaa0f // bfcvt z15.h, p2/M, z16.s\n" + "ld1h { z2.h }, p2/Z, [x19, #2, MUL VL]\n" + ".inst 0xc1321190 // bfdot za.s[x8, 0], { z12.h-z15.h }, z2.h\n" + "7:" // Unpadded: 3 priming loads + "add x20, x16, %x[ld_in_row], LSL #2\n" + "ld1w { z16.s }, p1/Z, [x16]\n" + ".inst 0x658aaa0a // bfcvt z10.h, p2/M, z16.s\n" + "addvl x19, SP, #9\n" + "ld1w { z16.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x648aaa0a // bfcvtnt z10.h, p2/M, z16.s\n" + "add x16, x16, %x[ld_in_col], LSL #2\n" + "ld1w { z16.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x658aaa0b // bfcvt z11.h, p2/M, z16.s\n" + "ld1w { z16.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x648aaa0b // bfcvtnt z11.h, p2/M, z16.s\n" + "ld1w { z16.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x658aaa0c // bfcvt z12.h, p2/M, z16.s\n" + "ld1w { z16.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x648aaa0c // bfcvtnt z12.h, p2/M, z16.s\n" + "ld1w { z16.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x658aaa0d // bfcvt z13.h, p2/M, z16.s\n" + "ld1w { z16.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x648aaa0d // bfcvtnt z13.h, p2/M, z16.s\n" + "ld1w { z16.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x658aaa0e // bfcvt z14.h, p2/M, z16.s\n" + ".inst 0xa1402a61 // ld1h { z1.h, z9.h }, pn10.b/Z, [x19]\n" + ".inst 0xc1311150 // bfdot za.s[x8, 0], { z10.h-z13.h }, z1.h\n" + "ld1w { z16.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x648aaa0e // bfcvtnt z14.h, p2/M, z16.s\n" + ".inst 0xc1391170 // bfdot za.s[x8, 0], { z11.h-z14.h }, z9.h\n" + "ld1w { z16.s }, p1/Z, [x20]\n" + ".inst 0x658aaa0f // bfcvt z15.h, p2/M, z16.s\n" + "ld1h { z2.h }, p2/Z, [x19, #2, MUL VL]\n" + ".inst 0xc1321190 // bfdot za.s[x8, 0], { z12.h-z15.h }, z2.h\n" + "8:" // Unpadded: 2 priming loads + "add x21, x16, %x[ld_in_row], LSL #2\n" + "ld1w { z16.s }, p1/Z, [x16]\n" + ".inst 0x658aaa0a // bfcvt z10.h, p2/M, z16.s\n" + "addvl x20, SP, #6\n" + "ld1w { z16.s }, p1/Z, [x21]\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0x648aaa0a // bfcvtnt z10.h, p2/M, z16.s\n" + "addvl x19, SP, #12\n" + "ld1w { z16.s }, p1/Z, [x21]\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0x658aaa0b // bfcvt z11.h, p2/M, z16.s\n" + "add x16, x16, %x[ld_in_col], LSL #2\n" + "ld1w { z16.s }, p1/Z, [x21]\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0x648aaa0b // bfcvtnt z11.h, p2/M, z16.s\n" + "ld1w { z16.s }, p1/Z, [x21]\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0x658aaa0c // bfcvt z12.h, p2/M, z16.s\n" + "ld1w { z16.s }, p1/Z, [x21]\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0x648aaa0c // bfcvtnt z12.h, p2/M, z16.s\n" + "ld1w { z16.s }, p1/Z, [x21]\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0x658aaa0d // bfcvt z13.h, p2/M, z16.s\n" + "ld1w { z16.s }, p1/Z, [x21]\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0x648aaa0d // bfcvtnt z13.h, p2/M, z16.s\n" + "ld1w { z16.s }, p1/Z, [x21]\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0x658aaa0e // bfcvt z14.h, p2/M, z16.s\n" + ".inst 0xa1402a81 // ld1h { z1.h, z9.h }, pn10.b/Z, [x20]\n" + ".inst 0xc1311150 // bfdot za.s[x8, 0], { z10.h-z13.h }, z1.h\n" + "ld1w { z16.s }, p1/Z, [x21]\n" + ".inst 0x648aaa0e // bfcvtnt z14.h, p2/M, z16.s\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0xc1391170 // bfdot za.s[x8, 0], { z11.h-z14.h }, z9.h\n" + ".inst 0xa1402a61 // ld1h { z1.h, z9.h }, pn10.b/Z, [x19]\n" + ".inst 0xc1311151 // bfdot za.s[x8, 1], { z10.h-z13.h }, z1.h\n" + "ld1w { z16.s }, p1/Z, [x21]\n" + ".inst 0x658aaa0f // bfcvt z15.h, p2/M, z16.s\n" + ".inst 0xc1391171 // bfdot za.s[x8, 1], { z11.h-z14.h }, z9.h\n" + "ld1h { z2.h }, p2/Z, [x20, #2, MUL VL]\n" + ".inst 0xc1321190 // bfdot za.s[x8, 0], { z12.h-z15.h }, z2.h\n" + "ld1h { z2.h }, p2/Z, [x19, #2, MUL VL]\n" + ".inst 0xc1321191 // bfdot za.s[x8, 1], { z12.h-z15.h }, z2.h\n" + "9:" // Unpadded: 1 priming loads + "add x21, x16, %x[ld_in_row], LSL #2\n" + "ld1w { z16.s }, p1/Z, [x16]\n" + ".inst 0x658aaa0a // bfcvt z10.h, p2/M, z16.s\n" + "addvl x20, SP, #3\n" + "ld1w { z16.s }, p1/Z, [x21]\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0x648aaa0a // bfcvtnt z10.h, p2/M, z16.s\n" + "addvl x19, SP, #9\n" + "ld1w { z16.s }, p1/Z, [x21]\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0x658aaa0b // bfcvt z11.h, p2/M, z16.s\n" + "add x16, x16, %x[ld_in_col], LSL #2\n" + "ld1w { z16.s }, p1/Z, [x21]\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0x648aaa0b // bfcvtnt z11.h, p2/M, z16.s\n" + "ld1w { z16.s }, p1/Z, [x21]\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0x658aaa0c // bfcvt z12.h, p2/M, z16.s\n" + "ld1w { z16.s }, p1/Z, [x21]\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0x648aaa0c // bfcvtnt z12.h, p2/M, z16.s\n" + "ld1w { z16.s }, p1/Z, [x21]\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0x658aaa0d // bfcvt z13.h, p2/M, z16.s\n" + "ld1w { z16.s }, p1/Z, [x21]\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0x648aaa0d // bfcvtnt z13.h, p2/M, z16.s\n" + "ld1w { z16.s }, p1/Z, [x21]\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0x658aaa0e // bfcvt z14.h, p2/M, z16.s\n" + ".inst 0xa1402a81 // ld1h { z1.h, z9.h }, pn10.b/Z, [x20]\n" + ".inst 0xc1311150 // bfdot za.s[x8, 0], { z10.h-z13.h }, z1.h\n" + "ld1w { z16.s }, p1/Z, [x21]\n" + ".inst 0x648aaa0e // bfcvtnt z14.h, p2/M, z16.s\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0xc1391170 // bfdot za.s[x8, 0], { z11.h-z14.h }, z9.h\n" + ".inst 0xa1402a61 // ld1h { z1.h, z9.h }, pn10.b/Z, [x19]\n" + ".inst 0xc1311151 // bfdot za.s[x8, 1], { z10.h-z13.h }, z1.h\n" + "ld1w { z16.s }, p1/Z, [x21]\n" + ".inst 0x658aaa0f // bfcvt z15.h, p2/M, z16.s\n" + ".inst 0xc1391171 // bfdot za.s[x8, 1], { z11.h-z14.h }, z9.h\n" + "ld1h { z2.h }, p2/Z, [x20, #2, MUL VL]\n" + ".inst 0xc1321190 // bfdot za.s[x8, 0], { z12.h-z15.h }, z2.h\n" + "ld1h { z2.h }, p2/Z, [x19, #2, MUL VL]\n" + ".inst 0xc1321191 // bfdot za.s[x8, 1], { z12.h-z15.h }, z2.h\n" + "10:" // Unpadded: 0 priming loads + "cmp x17, #0x2\n" + ".inst 0xa1402be1 // ld1h { z1.h, z9.h }, pn10.b/Z, [SP]\n" + "ld1h { z2.h }, p2/Z, [SP, #2, MUL VL]\n" + "blt 20f\n" + "add x20, x16, %x[ld_in_row], LSL #2\n" + "ld1w { z16.s }, p1/Z, [x16]\n" + ".inst 0x658aaa0a // bfcvt z10.h, p2/M, z16.s\n" + "sub x17, x17, #0x2\n" + "ld1w { z16.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + "sub x15, x15, #0x1\n" + ".inst 0x648aaa0a // bfcvtnt z10.h, p2/M, z16.s\n" + "ld1w { z16.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x658aaa0b // bfcvt z11.h, p2/M, z16.s\n" + "lsr x19, x17, #0x1\n" + "ld1w { z16.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + "cmp x19, x15\n" + ".inst 0x648aaa0b // bfcvtnt z11.h, p2/M, z16.s\n" + "ld1w { z16.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x658aaa0c // bfcvt z12.h, p2/M, z16.s\n" + "csel x25, x19, x15, LT\n" + "ld1w { z16.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x648aaa0c // bfcvtnt z12.h, p2/M, z16.s\n" + "add x16, x16, %x[ld_in_col], LSL #2\n" + "ld1w { z16.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x658aaa0d // bfcvt z13.h, p2/M, z16.s\n" + "and x17, x17, #0x1\n" + "ld1w { z16.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x648aaa0d // bfcvtnt z13.h, p2/M, z16.s\n" + "sub x15, x15, x25\n" + "ld1w { z16.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x658aaa0e // bfcvt z14.h, p2/M, z16.s\n" + "ld1w { z16.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x648aaa0e // bfcvtnt z14.h, p2/M, z16.s\n" + "ld1w { z16.s }, p1/Z, [x20]\n" + ".inst 0x658aaa0f // bfcvt z15.h, p2/M, z16.s\n" + "cbz x25, 19f\n" + "11:" // Unpadded: Main loop + ".inst 0xc1311150 // bfdot za.s[x8, 0], { z10.h-z13.h }, z1.h\n" + "addvl x24, SP, #6\n" + "addvl x23, SP, #12\n" + "ld1w { z18.s }, p1/Z, [x16]\n" + ".inst 0xc1391170 // bfdot za.s[x8, 0], { z11.h-z14.h }, z9.h\n" + ".inst 0xa1402b01 // ld1h { z1.h, z9.h }, pn10.b/Z, [x24]\n" + "add x22, x16, %x[ld_in_row], LSL #2\n" + "addvl x21, SP, #3\n" + ".inst 0xc1311151 // bfdot za.s[x8, 1], { z10.h-z13.h }, z1.h\n" + "ld1w { z17.s }, p1/Z, [x22]\n" + "add x22, x22, %x[ld_in_row], LSL #2\n" + "add x16, x16, %x[ld_in_col], LSL #2\n" + ".inst 0xc1391171 // bfdot za.s[x8, 1], { z11.h-z14.h }, z9.h\n" + ".inst 0xa1402ae1 // ld1h { z1.h, z9.h }, pn10.b/Z, [x23]\n" + "addvl x20, SP, #9\n" + "add x19, x16, %x[ld_in_row], LSL #2\n" + ".inst 0xc1311152 // bfdot za.s[x8, 2], { z10.h-z13.h }, z1.h\n" + "ld1w { z16.s }, p1/Z, [x22]\n" + "add x22, x22, %x[ld_in_row], LSL #2\n" + ".inst 0x658aaa4a // bfcvt z10.h, p2/M, z18.s\n" + ".inst 0xc1321190 // bfdot za.s[x8, 0], { z12.h-z15.h }, z2.h\n" + "ld1h { z2.h }, p2/Z, [x24, #2, MUL VL]\n" + ".inst 0x648aaa2a // bfcvtnt z10.h, p2/M, z17.s\n" + "subs x25, x25, #0x1\n" + "ld1w { z17.s }, p1/Z, [x22]\n" + "add x22, x22, %x[ld_in_row], LSL #2\n" + ".inst 0xc1391172 // bfdot za.s[x8, 2], { z11.h-z14.h }, z9.h\n" + ".inst 0x658aaa0b // bfcvt z11.h, p2/M, z16.s\n" + "ld1w { z16.s }, p1/Z, [x22]\n" + "add x22, x22, %x[ld_in_row], LSL #2\n" + ".inst 0xc1321191 // bfdot za.s[x8, 1], { z12.h-z15.h }, z2.h\n" + ".inst 0x648aaa2b // bfcvtnt z11.h, p2/M, z17.s\n" + "ld1w { z17.s }, p1/Z, [x22]\n" + "add x22, x22, %x[ld_in_row], LSL #2\n" + ".inst 0xc0060c18 // mova { z24.d-z27.d }, za.d[x8, #0]\n" + ".inst 0xc1b6cbd8 // fclamp { z24.s-z27.s }, z30.s, z22.s\n" + "ld1h { z2.h }, p2/Z, [x23, #2, MUL VL]\n" + ".inst 0xc1321192 // bfdot za.s[x8, 2], { z12.h-z15.h }, z2.h\n" + ".inst 0x658aaa0c // bfcvt z12.h, p2/M, z16.s\n" + "add x8, x8, #0x1\n" + "ld1w { z16.s }, p1/Z, [x22]\n" + "add x22, x22, %x[ld_in_row], LSL #2\n" + ".inst 0x658aaa0d // bfcvt z13.h, p2/M, z16.s\n" + ".inst 0x648aaa2c // bfcvtnt z12.h, p2/M, z17.s\n" + "ld1w { z16.s }, p1/Z, [x22]\n" + "add x22, x22, %x[ld_in_row], LSL #2\n" + ".inst 0x648aaa0d // bfcvtnt z13.h, p2/M, z16.s\n" + "st1w { z24.s }, p1, [x14]\n" + "ld1w { z16.s }, p1/Z, [x22]\n" + "add x22, x22, %x[ld_in_row], LSL #2\n" + ".inst 0x658aaa0e // bfcvt z14.h, p2/M, z16.s\n" + "add x14, x14, x11, LSL #2\n" + ".inst 0xa1402aa1 // ld1h { z1.h, z9.h }, pn10.b/Z, [x21]\n" + ".inst 0xc1311150 // bfdot za.s[x8, 0], { z10.h-z13.h }, z1.h\n" + "st1w { z25.s }, p1, [x13]\n" + "add x13, x13, x10, LSL #2\n" + "ld1w { z16.s }, p1/Z, [x22]\n" + ".inst 0x648aaa0e // bfcvtnt z14.h, p2/M, z16.s\n" + "add x22, x22, %x[ld_in_row], LSL #2\n" + ".inst 0xc1391170 // bfdot za.s[x8, 0], { z11.h-z14.h }, z9.h\n" + ".inst 0xa1402a81 // ld1h { z1.h, z9.h }, pn10.b/Z, [x20]\n" + ".inst 0xc1311151 // bfdot za.s[x8, 1], { z10.h-z13.h }, z1.h\n" + "st1w { z26.s }, p1, [x9]\n" + "add x9, x9, x27, LSL #2\n" + "ld1w { z16.s }, p1/Z, [x22]\n" + ".inst 0x658aaa0f // bfcvt z15.h, p2/M, z16.s\n" + ".inst 0xc1391171 // bfdot za.s[x8, 1], { z11.h-z14.h }, z9.h\n" + "ld1w { z16.s }, p1/Z, [x16]\n" + ".inst 0x658aaa0a // bfcvt z10.h, p2/M, z16.s\n" + "st1w { z27.s }, p1, [x28]\n" + "add x28, x28, x26, LSL #2\n" + "ld1w { z16.s }, p1/Z, [x19]\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" + ".inst 0xc0040c84 // mova za.d[x8, #4], { z4.d-z7.d }\n" + ".inst 0x648aaa0a // bfcvtnt z10.h, p2/M, z16.s\n" + "ld1w { z16.s }, p1/Z, [x19]\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" + ".inst 0x658aaa0b // bfcvt z11.h, p2/M, z16.s\n" + "add x16, x16, %x[ld_in_col], LSL #2\n" + "ld1w { z16.s }, p1/Z, [x19]\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" + ".inst 0x648aaa0b // bfcvtnt z11.h, p2/M, z16.s\n" + "ld1w { z16.s }, p1/Z, [x19]\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" + "ld1h { z2.h }, p2/Z, [x21, #2, MUL VL]\n" + ".inst 0xc1321190 // bfdot za.s[x8, 0], { z12.h-z15.h }, z2.h\n" + "ld1w { z19.s }, p1/Z, [x19]\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" + "ld1w { z17.s }, p1/Z, [x19]\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" + "ld1w { z18.s }, p1/Z, [x19]\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" + "ld1h { z2.h }, p2/Z, [x20, #2, MUL VL]\n" + ".inst 0xc1321191 // bfdot za.s[x8, 1], { z12.h-z15.h }, z2.h\n" + ".inst 0x658aaa0c // bfcvt z12.h, p2/M, z16.s\n" + "ld1w { z16.s }, p1/Z, [x19]\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" + ".inst 0x658aaa2d // bfcvt z13.h, p2/M, z17.s\n" + ".inst 0x658aaa0e // bfcvt z14.h, p2/M, z16.s\n" + "ld1w { z17.s }, p1/Z, [x19]\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" + ".inst 0x648aaa6c // bfcvtnt z12.h, p2/M, z19.s\n" + ".inst 0x648aaa4d // bfcvtnt z13.h, p2/M, z18.s\n" + "ld1w { z16.s }, p1/Z, [x19]\n" + ".inst 0x648aaa2e // bfcvtnt z14.h, p2/M, z17.s\n" + ".inst 0x658aaa0f // bfcvt z15.h, p2/M, z16.s\n" + ".inst 0xa1402be1 // ld1h { z1.h, z9.h }, pn10.b/Z, [SP]\n" + "ld1h { z2.h }, p2/Z, [SP, #2, MUL VL]\n" + "bgt 11b\n" + "b 19f\n" + "12:" // Padded + "cbz x21, 17f\n" + "cmp x21, #0x1\n" + "sub x17, x17, x21\n" + "beq 16f\n" + "cmp x21, #0x2\n" + "beq 15f\n" + "cmp x21, #0x3\n" + "beq 14f\n" + "13:" // Padded: 4 priming loads + "mov x12, #0x0\n" + ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" + "ld1w { z16.s }, p0/Z, [x16]\n" + ".inst 0x658aaa0a // bfcvt z10.h, p2/M, z16.s\n" + "add x20, x16, %x[ld_in_row], LSL #2\n" + ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" + "ld1w { z16.s }, p0/Z, [x20]\n" + ".inst 0x648aaa0a // bfcvtnt z10.h, p2/M, z16.s\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" + "ld1w { z16.s }, p0/Z, [x20]\n" + ".inst 0x658aaa0b // bfcvt z11.h, p2/M, z16.s\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" + "ld1w { z16.s }, p0/Z, [x20]\n" + ".inst 0x648aaa0b // bfcvtnt z11.h, p2/M, z16.s\n" + "mov x12, #0x4\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" + "ld1w { z16.s }, p0/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x658aaa0c // bfcvt z12.h, p2/M, z16.s\n" + ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" + "ld1w { z16.s }, p0/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x648aaa0c // bfcvtnt z12.h, p2/M, z16.s\n" + ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" + "ld1w { z16.s }, p0/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x658aaa0d // bfcvt z13.h, p2/M, z16.s\n" + ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" + "mov x12, #0x8\n" + "ld1w { z16.s }, p0/Z, [x20]\n" + ".inst 0x648aaa0d // bfcvtnt z13.h, p2/M, z16.s\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" + "ld1w { z16.s }, p0/Z, [x20]\n" + ".inst 0x658aaa0e // bfcvt z14.h, p2/M, z16.s\n" + "addvl x19, SP, #12\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0xa1402a61 // ld1h { z1.h, z9.h }, pn10.b/Z, [x19]\n" + ".inst 0xc1311150 // bfdot za.s[x8, 0], { z10.h-z13.h }, z1.h\n" + ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" + "ld1w { z16.s }, p0/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x648aaa0e // bfcvtnt z14.h, p2/M, z16.s\n" + ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" + "ld1w { z16.s }, p0/Z, [x20]\n" + ".inst 0x658aaa0f // bfcvt z15.h, p2/M, z16.s\n" + ".inst 0xc1391170 // bfdot za.s[x8, 0], { z11.h-z14.h }, z9.h\n" + "ld1h { z2.h }, p2/Z, [x19, #2, MUL VL]\n" + "add x16, x16, %x[ld_in_col], LSL #2\n" + ".inst 0xc1321190 // bfdot za.s[x8, 0], { z12.h-z15.h }, z2.h\n" + "14:" // Padded: 3 priming loads + "mov x12, #0x0\n" + ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" + "ld1w { z16.s }, p0/Z, [x16]\n" + ".inst 0x658aaa0a // bfcvt z10.h, p2/M, z16.s\n" + "add x20, x16, %x[ld_in_row], LSL #2\n" + ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" + "ld1w { z16.s }, p0/Z, [x20]\n" + ".inst 0x648aaa0a // bfcvtnt z10.h, p2/M, z16.s\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" + "ld1w { z16.s }, p0/Z, [x20]\n" + ".inst 0x658aaa0b // bfcvt z11.h, p2/M, z16.s\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" + "ld1w { z16.s }, p0/Z, [x20]\n" + ".inst 0x648aaa0b // bfcvtnt z11.h, p2/M, z16.s\n" + "mov x12, #0x4\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" + "ld1w { z16.s }, p0/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x658aaa0c // bfcvt z12.h, p2/M, z16.s\n" + ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" + "ld1w { z16.s }, p0/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x648aaa0c // bfcvtnt z12.h, p2/M, z16.s\n" + ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" + "ld1w { z16.s }, p0/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x658aaa0d // bfcvt z13.h, p2/M, z16.s\n" + ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" + "mov x12, #0x8\n" + "ld1w { z16.s }, p0/Z, [x20]\n" + ".inst 0x648aaa0d // bfcvtnt z13.h, p2/M, z16.s\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" + "ld1w { z16.s }, p0/Z, [x20]\n" + ".inst 0x658aaa0e // bfcvt z14.h, p2/M, z16.s\n" + "addvl x19, SP, #9\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0xa1402a61 // ld1h { z1.h, z9.h }, pn10.b/Z, [x19]\n" + ".inst 0xc1311150 // bfdot za.s[x8, 0], { z10.h-z13.h }, z1.h\n" + ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" + "ld1w { z16.s }, p0/Z, [x20]\n" + "add x20, x20, %x[ld_in_row], LSL #2\n" + ".inst 0x648aaa0e // bfcvtnt z14.h, p2/M, z16.s\n" + ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" + "ld1w { z16.s }, p0/Z, [x20]\n" + ".inst 0x658aaa0f // bfcvt z15.h, p2/M, z16.s\n" + ".inst 0xc1391170 // bfdot za.s[x8, 0], { z11.h-z14.h }, z9.h\n" + "ld1h { z2.h }, p2/Z, [x19, #2, MUL VL]\n" + "add x16, x16, %x[ld_in_col], LSL #2\n" + ".inst 0xc1321190 // bfdot za.s[x8, 0], { z12.h-z15.h }, z2.h\n" + "15:" // Padded: 2 priming loads + "mov x12, #0x0\n" + ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" + "ld1w { z16.s }, p0/Z, [x16]\n" + ".inst 0x658aaa0a // bfcvt z10.h, p2/M, z16.s\n" + "add x21, x16, %x[ld_in_row], LSL #2\n" + ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" + "ld1w { z16.s }, p0/Z, [x21]\n" + ".inst 0x648aaa0a // bfcvtnt z10.h, p2/M, z16.s\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" + "ld1w { z16.s }, p0/Z, [x21]\n" + ".inst 0x658aaa0b // bfcvt z11.h, p2/M, z16.s\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" + "ld1w { z16.s }, p0/Z, [x21]\n" + ".inst 0x648aaa0b // bfcvtnt z11.h, p2/M, z16.s\n" + "mov x12, #0x4\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" + "ld1w { z16.s }, p0/Z, [x21]\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0x658aaa0c // bfcvt z12.h, p2/M, z16.s\n" + ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" + "ld1w { z16.s }, p0/Z, [x21]\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0x648aaa0c // bfcvtnt z12.h, p2/M, z16.s\n" + ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" + "ld1w { z16.s }, p0/Z, [x21]\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0x658aaa0d // bfcvt z13.h, p2/M, z16.s\n" + ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" + "mov x12, #0x8\n" + "ld1w { z16.s }, p0/Z, [x21]\n" + ".inst 0x648aaa0d // bfcvtnt z13.h, p2/M, z16.s\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" + "ld1w { z16.s }, p0/Z, [x21]\n" + ".inst 0x658aaa0e // bfcvt z14.h, p2/M, z16.s\n" + "addvl x20, SP, #6\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0xa1402a81 // ld1h { z1.h, z9.h }, pn10.b/Z, [x20]\n" + ".inst 0xc1311150 // bfdot za.s[x8, 0], { z10.h-z13.h }, z1.h\n" + ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" + "ld1w { z16.s }, p0/Z, [x21]\n" + ".inst 0x648aaa0e // bfcvtnt z14.h, p2/M, z16.s\n" + "addvl x19, SP, #12\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" + ".inst 0xc1391170 // bfdot za.s[x8, 0], { z11.h-z14.h }, z9.h\n" + ".inst 0xa1402a61 // ld1h { z1.h, z9.h }, pn10.b/Z, [x19]\n" + "ld1w { z16.s }, p0/Z, [x21]\n" + ".inst 0xc1311151 // bfdot za.s[x8, 1], { z10.h-z13.h }, z1.h\n" + ".inst 0x658aaa0f // bfcvt z15.h, p2/M, z16.s\n" + "add x16, x16, %x[ld_in_col], LSL #2\n" + "ld1h { z2.h }, p2/Z, [x20, #2, MUL VL]\n" + ".inst 0xc1391171 // bfdot za.s[x8, 1], { z11.h-z14.h }, z9.h\n" + ".inst 0xc1321190 // bfdot za.s[x8, 0], { z12.h-z15.h }, z2.h\n" + "ld1h { z2.h }, p2/Z, [x19, #2, MUL VL]\n" + ".inst 0xc1321191 // bfdot za.s[x8, 1], { z12.h-z15.h }, z2.h\n" + "16:" // Padded: 1 priming loads + "mov x12, #0x0\n" + ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" + "ld1w { z16.s }, p0/Z, [x16]\n" + ".inst 0x658aaa0a // bfcvt z10.h, p2/M, z16.s\n" + "add x21, x16, %x[ld_in_row], LSL #2\n" + ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" + "ld1w { z16.s }, p0/Z, [x21]\n" + ".inst 0x648aaa0a // bfcvtnt z10.h, p2/M, z16.s\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" + "ld1w { z16.s }, p0/Z, [x21]\n" + ".inst 0x658aaa0b // bfcvt z11.h, p2/M, z16.s\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" + "ld1w { z16.s }, p0/Z, [x21]\n" + ".inst 0x648aaa0b // bfcvtnt z11.h, p2/M, z16.s\n" + "mov x12, #0x4\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" + "ld1w { z16.s }, p0/Z, [x21]\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0x658aaa0c // bfcvt z12.h, p2/M, z16.s\n" + ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" + "ld1w { z16.s }, p0/Z, [x21]\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0x648aaa0c // bfcvtnt z12.h, p2/M, z16.s\n" + ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" + "ld1w { z16.s }, p0/Z, [x21]\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0x658aaa0d // bfcvt z13.h, p2/M, z16.s\n" + ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" + "mov x12, #0x8\n" + "ld1w { z16.s }, p0/Z, [x21]\n" + ".inst 0x648aaa0d // bfcvtnt z13.h, p2/M, z16.s\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" + "ld1w { z16.s }, p0/Z, [x21]\n" + ".inst 0x658aaa0e // bfcvt z14.h, p2/M, z16.s\n" + "addvl x20, SP, #3\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0xa1402a81 // ld1h { z1.h, z9.h }, pn10.b/Z, [x20]\n" + ".inst 0xc1311150 // bfdot za.s[x8, 0], { z10.h-z13.h }, z1.h\n" + ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" + "ld1w { z16.s }, p0/Z, [x21]\n" + ".inst 0x648aaa0e // bfcvtnt z14.h, p2/M, z16.s\n" + "addvl x19, SP, #9\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" + ".inst 0xc1391170 // bfdot za.s[x8, 0], { z11.h-z14.h }, z9.h\n" + ".inst 0xa1402a61 // ld1h { z1.h, z9.h }, pn10.b/Z, [x19]\n" + "ld1w { z16.s }, p0/Z, [x21]\n" + ".inst 0xc1311151 // bfdot za.s[x8, 1], { z10.h-z13.h }, z1.h\n" + ".inst 0x658aaa0f // bfcvt z15.h, p2/M, z16.s\n" + "add x16, x16, %x[ld_in_col], LSL #2\n" + "ld1h { z2.h }, p2/Z, [x20, #2, MUL VL]\n" + ".inst 0xc1391171 // bfdot za.s[x8, 1], { z11.h-z14.h }, z9.h\n" + ".inst 0xc1321190 // bfdot za.s[x8, 0], { z12.h-z15.h }, z2.h\n" + "ld1h { z2.h }, p2/Z, [x19, #2, MUL VL]\n" + ".inst 0xc1321191 // bfdot za.s[x8, 1], { z12.h-z15.h }, z2.h\n" + "17:" // Padded: 0 priming loads + "cmp x17, #0x2\n" + ".inst 0xa1402be1 // ld1h { z1.h, z9.h }, pn10.b/Z, [SP]\n" + "ld1h { z2.h }, p2/Z, [SP, #2, MUL VL]\n" + "blt 20f\n" + "mov x12, #0x0\n" + ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" + "ld1w { z16.s }, p0/Z, [x16]\n" + ".inst 0x658aaa0a // bfcvt z10.h, p2/M, z16.s\n" + "add x19, x16, %x[ld_in_row], LSL #2\n" + ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" + "ld1w { z16.s }, p0/Z, [x19]\n" + ".inst 0x648aaa0a // bfcvtnt z10.h, p2/M, z16.s\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" + ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" + "ld1w { z16.s }, p0/Z, [x19]\n" + ".inst 0x658aaa0b // bfcvt z11.h, p2/M, z16.s\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" + ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" + "ld1w { z16.s }, p0/Z, [x19]\n" + ".inst 0x648aaa0b // bfcvtnt z11.h, p2/M, z16.s\n" + "mov x12, #0x4\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" + ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" + "ld1w { z16.s }, p0/Z, [x19]\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" + ".inst 0x658aaa0c // bfcvt z12.h, p2/M, z16.s\n" + ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" + "ld1w { z16.s }, p0/Z, [x19]\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" + ".inst 0x648aaa0c // bfcvtnt z12.h, p2/M, z16.s\n" + ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" + "ld1w { z16.s }, p0/Z, [x19]\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" + ".inst 0x658aaa0d // bfcvt z13.h, p2/M, z16.s\n" + ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" + "mov x12, #0x8\n" + "ld1w { z16.s }, p0/Z, [x19]\n" + ".inst 0x648aaa0d // bfcvtnt z13.h, p2/M, z16.s\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" + ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" + "ld1w { z16.s }, p0/Z, [x19]\n" + ".inst 0x658aaa0e // bfcvt z14.h, p2/M, z16.s\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" + ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" + "ld1w { z16.s }, p0/Z, [x19]\n" + ".inst 0x648aaa0e // bfcvtnt z14.h, p2/M, z16.s\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" + ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" + "ld1w { z16.s }, p0/Z, [x19]\n" + ".inst 0x658aaa0f // bfcvt z15.h, p2/M, z16.s\n" + "sub x17, x17, #0x2\n" + "sub x15, x15, #0x1\n" + "lsr x19, x17, #0x1\n" + "cmp x19, x15\n" + "csel x23, x19, x15, LT\n" + "add x16, x16, %x[ld_in_col], LSL #2\n" + "and x17, x17, #0x1\n" + "sub x15, x15, x23\n" + "cbz x23, 19f\n" + "18:" // Padded: Main loop + ".inst 0xc1311150 // bfdot za.s[x8, 0], { z10.h-z13.h }, z1.h\n" + "addvl x22, SP, #6\n" + "addvl x20, SP, #12\n" + ".inst 0xc1391170 // bfdot za.s[x8, 0], { z11.h-z14.h }, z9.h\n" + ".inst 0xa1402ac1 // ld1h { z1.h, z9.h }, pn10.b/Z, [x22]\n" + "mov x12, #0x0\n" + ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" + ".inst 0xc1311151 // bfdot za.s[x8, 1], { z10.h-z13.h }, z1.h\n" + "ld1w { z16.s }, p0/Z, [x16]\n" + "add x19, x16, %x[ld_in_row], LSL #2\n" + ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" + ".inst 0xc1391171 // bfdot za.s[x8, 1], { z11.h-z14.h }, z9.h\n" + ".inst 0xa1402a81 // ld1h { z1.h, z9.h }, pn10.b/Z, [x20]\n" + "addvl x21, SP, #3\n" + "add x16, x16, %x[ld_in_col], LSL #2\n" + ".inst 0xc1311152 // bfdot za.s[x8, 2], { z10.h-z13.h }, z1.h\n" + "ld1w { z20.s }, p0/Z, [x19]\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" + ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" + "ld1w { z17.s }, p0/Z, [x19]\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" + ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" + ".inst 0xc1321190 // bfdot za.s[x8, 0], { z12.h-z15.h }, z2.h\n" + "mov x12, #0x4\n" + "ld1h { z2.h }, p2/Z, [x22, #2, MUL VL]\n" + ".inst 0xc1391172 // bfdot za.s[x8, 2], { z11.h-z14.h }, z9.h\n" + ".inst 0x658aaa0a // bfcvt z10.h, p2/M, z16.s\n" + "ld1w { z19.s }, p0/Z, [x19]\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" + ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" + ".inst 0xc1321191 // bfdot za.s[x8, 1], { z12.h-z15.h }, z2.h\n" + "ld1w { z16.s }, p0/Z, [x19]\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" + ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" + ".inst 0x658aaa2b // bfcvt z11.h, p2/M, z17.s\n" + "ld1w { z18.s }, p0/Z, [x19]\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" + ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" + ".inst 0x648aaa8a // bfcvtnt z10.h, p2/M, z20.s\n" + "ld1h { z2.h }, p2/Z, [x20, #2, MUL VL]\n" + ".inst 0xc1321192 // bfdot za.s[x8, 2], { z12.h-z15.h }, z2.h\n" + ".inst 0x658aaa0c // bfcvt z12.h, p2/M, z16.s\n" + "addvl x20, SP, #9\n" + "ld1w { z16.s }, p0/Z, [x19]\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" + ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" + ".inst 0x658aaa0d // bfcvt z13.h, p2/M, z16.s\n" + "mov x12, #0x8\n" + "ld1w { z17.s }, p0/Z, [x19]\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" + ".inst 0x648aaa6b // bfcvtnt z11.h, p2/M, z19.s\n" + ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" + "ld1w { z16.s }, p0/Z, [x19]\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" + ".inst 0x648aaa4c // bfcvtnt z12.h, p2/M, z18.s\n" + ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" + ".inst 0x648aaa2d // bfcvtnt z13.h, p2/M, z17.s\n" + ".inst 0x658aaa0e // bfcvt z14.h, p2/M, z16.s\n" + ".inst 0xa1402aa1 // ld1h { z1.h, z9.h }, pn10.b/Z, [x21]\n" + "ld1w { z16.s }, p0/Z, [x19]\n" + ".inst 0xc0060c18 // mova { z24.d-z27.d }, za.d[x8, #0]\n" + "add x8, x8, #0x1\n" + ".inst 0x648aaa0e // bfcvtnt z14.h, p2/M, z16.s\n" + ".inst 0xc1311150 // bfdot za.s[x8, 0], { z10.h-z13.h }, z1.h\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" + ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" + "ld1w { z17.s }, p0/Z, [x19]\n" + "mov x12, #0x0\n" + ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" + ".inst 0xc1391170 // bfdot za.s[x8, 0], { z11.h-z14.h }, z9.h\n" + ".inst 0xa1402a81 // ld1h { z1.h, z9.h }, pn10.b/Z, [x20]\n" + "ld1w { z16.s }, p0/Z, [x16]\n" + "add x19, x16, %x[ld_in_row], LSL #2\n" + ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" + ".inst 0xc1311151 // bfdot za.s[x8, 1], { z10.h-z13.h }, z1.h\n" + "ld1w { z21.s }, p0/Z, [x19]\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" + ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" + ".inst 0x658aaa2f // bfcvt z15.h, p2/M, z17.s\n" + "ld1w { z17.s }, p0/Z, [x19]\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" + ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" + ".inst 0xc1391171 // bfdot za.s[x8, 1], { z11.h-z14.h }, z9.h\n" + "mov x12, #0x4\n" + "ld1w { z20.s }, p0/Z, [x19]\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" + ".inst 0x658aaa0a // bfcvt z10.h, p2/M, z16.s\n" + ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" + "ld1w { z16.s }, p0/Z, [x19]\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" + ".inst 0x658aaa2b // bfcvt z11.h, p2/M, z17.s\n" + ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" + "ld1h { z2.h }, p2/Z, [x21, #2, MUL VL]\n" + ".inst 0xc1321190 // bfdot za.s[x8, 0], { z12.h-z15.h }, z2.h\n" + "subs x23, x23, #0x1\n" + "ld1w { z19.s }, p0/Z, [x19]\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" + ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" + ".inst 0xc1b6cbd8 // fclamp { z24.s-z27.s }, z30.s, z22.s\n" + "ld1w { z17.s }, p0/Z, [x19]\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" + ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" + "st1w { z24.s }, p1, [x14]\n" + "mov x12, #0x8\n" + "ld1w { z18.s }, p0/Z, [x19]\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" + "st1w { z25.s }, p1, [x13]\n" + ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" + "ld1h { z2.h }, p2/Z, [x20, #2, MUL VL]\n" + ".inst 0xc1321191 // bfdot za.s[x8, 1], { z12.h-z15.h }, z2.h\n" + ".inst 0x658aaa0c // bfcvt z12.h, p2/M, z16.s\n" + "ld1w { z16.s }, p0/Z, [x19]\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" + ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" + ".inst 0x658aaa2d // bfcvt z13.h, p2/M, z17.s\n" + "ld1w { z17.s }, p0/Z, [x19]\n" + "add x19, x19, %x[ld_in_row], LSL #2\n" + ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" + ".inst 0x658aaa0e // bfcvt z14.h, p2/M, z16.s\n" + "ld1w { z16.s }, p0/Z, [x19]\n" + "add x14, x14, x11, LSL #2\n" + "add x13, x13, x10, LSL #2\n" + "st1w { z26.s }, p1, [x9]\n" + "add x9, x9, x27, LSL #2\n" + "st1w { z27.s }, p1, [x28]\n" + "add x28, x28, x26, LSL #2\n" + ".inst 0xc0040c84 // mova za.d[x8, #4], { z4.d-z7.d }\n" + ".inst 0xa1402be1 // ld1h { z1.h, z9.h }, pn10.b/Z, [SP]\n" + ".inst 0x648aaaaa // bfcvtnt z10.h, p2/M, z21.s\n" + ".inst 0x648aaa8b // bfcvtnt z11.h, p2/M, z20.s\n" + "add x16, x16, %x[ld_in_col], LSL #2\n" + "ld1h { z2.h }, p2/Z, [SP, #2, MUL VL]\n" + ".inst 0x648aaa6c // bfcvtnt z12.h, p2/M, z19.s\n" + ".inst 0x648aaa4d // bfcvtnt z13.h, p2/M, z18.s\n" + ".inst 0x648aaa2e // bfcvtnt z14.h, p2/M, z17.s\n" + ".inst 0x658aaa0f // bfcvt z15.h, p2/M, z16.s\n" + "bgt 18b\n" + "19:" // Main loop tail + ".inst 0xc1311150 // bfdot za.s[x8, 0], { z10.h-z13.h }, z1.h\n" + "addvl x23, SP, #6\n" + "addvl x22, SP, #12\n" + ".inst 0xc1391170 // bfdot za.s[x8, 0], { z11.h-z14.h }, z9.h\n" + ".inst 0xa1402ae1 // ld1h { z1.h, z9.h }, pn10.b/Z, [x23]\n" + "mov x12, #0x0\n" + ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" + ".inst 0xc1311151 // bfdot za.s[x8, 1], { z10.h-z13.h }, z1.h\n" + "ld1w { z16.s }, p0/Z, [x16]\n" + "add x21, x16, %x[ld_in_row], LSL #2\n" + ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" + ".inst 0xc1391171 // bfdot za.s[x8, 1], { z11.h-z14.h }, z9.h\n" + ".inst 0xa1402ac1 // ld1h { z1.h, z9.h }, pn10.b/Z, [x22]\n" + "addvl x20, SP, #3\n" + "addvl x19, SP, #9\n" + ".inst 0xc1311152 // bfdot za.s[x8, 2], { z10.h-z13.h }, z1.h\n" + "ld1w { z20.s }, p0/Z, [x21]\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" + "ld1w { z17.s }, p0/Z, [x21]\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" + ".inst 0xc1321190 // bfdot za.s[x8, 0], { z12.h-z15.h }, z2.h\n" + "mov x12, #0x4\n" + "ld1h { z2.h }, p2/Z, [x23, #2, MUL VL]\n" + ".inst 0xc1391172 // bfdot za.s[x8, 2], { z11.h-z14.h }, z9.h\n" + ".inst 0x658aaa0a // bfcvt z10.h, p2/M, z16.s\n" + "ld1w { z19.s }, p0/Z, [x21]\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" + ".inst 0xc1321191 // bfdot za.s[x8, 1], { z12.h-z15.h }, z2.h\n" + "ld1w { z16.s }, p0/Z, [x21]\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" + ".inst 0x658aaa2b // bfcvt z11.h, p2/M, z17.s\n" + "ld1w { z18.s }, p0/Z, [x21]\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" + ".inst 0x648aaa8a // bfcvtnt z10.h, p2/M, z20.s\n" + "ld1h { z2.h }, p2/Z, [x22, #2, MUL VL]\n" + ".inst 0xc1321192 // bfdot za.s[x8, 2], { z12.h-z15.h }, z2.h\n" + ".inst 0x658aaa0c // bfcvt z12.h, p2/M, z16.s\n" + "add x16, x16, %x[ld_in_col], LSL #2\n" + "ld1w { z16.s }, p0/Z, [x21]\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" + ".inst 0x658aaa0d // bfcvt z13.h, p2/M, z16.s\n" + "mov x12, #0x8\n" + "ld1w { z17.s }, p0/Z, [x21]\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0x648aaa6b // bfcvtnt z11.h, p2/M, z19.s\n" + ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" + "ld1w { z16.s }, p0/Z, [x21]\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0x648aaa4c // bfcvtnt z12.h, p2/M, z18.s\n" + ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" + ".inst 0x648aaa2d // bfcvtnt z13.h, p2/M, z17.s\n" + ".inst 0x658aaa0e // bfcvt z14.h, p2/M, z16.s\n" + ".inst 0xa1402a81 // ld1h { z1.h, z9.h }, pn10.b/Z, [x20]\n" + "ld1w { z16.s }, p0/Z, [x21]\n" + ".inst 0xc0060c18 // mova { z24.d-z27.d }, za.d[x8, #0]\n" + "add x8, x8, #0x1\n" + ".inst 0x648aaa0e // bfcvtnt z14.h, p2/M, z16.s\n" + ".inst 0xc1311150 // bfdot za.s[x8, 0], { z10.h-z13.h }, z1.h\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" + "ld1w { z16.s }, p0/Z, [x21]\n" + ".inst 0xc1391170 // bfdot za.s[x8, 0], { z11.h-z14.h }, z9.h\n" + ".inst 0xa1402a61 // ld1h { z1.h, z9.h }, pn10.b/Z, [x19]\n" + ".inst 0x658aaa0f // bfcvt z15.h, p2/M, z16.s\n" + ".inst 0xc1311151 // bfdot za.s[x8, 1], { z10.h-z13.h }, z1.h\n" + "ld1h { z2.h }, p2/Z, [x20, #2, MUL VL]\n" + ".inst 0xc1b6cbd8 // fclamp { z24.s-z27.s }, z30.s, z22.s\n" + ".inst 0xc1391171 // bfdot za.s[x8, 1], { z11.h-z14.h }, z9.h\n" + "st1w { z24.s }, p1, [x14]\n" + "add x14, x14, x11, LSL #2\n" + ".inst 0xa1402be1 // ld1h { z1.h, z9.h }, pn10.b/Z, [SP]\n" + ".inst 0xc1321190 // bfdot za.s[x8, 0], { z12.h-z15.h }, z2.h\n" + "ld1h { z2.h }, p2/Z, [x19, #2, MUL VL]\n" + "st1w { z25.s }, p1, [x13]\n" + "add x13, x13, x10, LSL #2\n" + "st1w { z26.s }, p1, [x9]\n" + "add x9, x9, x27, LSL #2\n" + ".inst 0xc0040c84 // mova za.d[x8, #4], { z4.d-z7.d }\n" + "st1w { z27.s }, p1, [x28]\n" + "add x28, x28, x26, LSL #2\n" + ".inst 0xc1321191 // bfdot za.s[x8, 1], { z12.h-z15.h }, z2.h\n" + "ld1h { z2.h }, p2/Z, [SP, #2, MUL VL]\n" + "20:" // Main loop skip tail + "cbz x17, 21f\n" // Skip remainder inputs + "mov x12, #0x0\n" + ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" + "ld1w { z16.s }, p0/Z, [x16]\n" + ".inst 0x658aaa0a // bfcvt z10.h, p2/M, z16.s\n" + "add x21, x16, %x[ld_in_row], LSL #2\n" + ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" + "ld1w { z16.s }, p0/Z, [x21]\n" + ".inst 0x648aaa0a // bfcvtnt z10.h, p2/M, z16.s\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" + "ld1w { z16.s }, p0/Z, [x21]\n" + ".inst 0x658aaa0b // bfcvt z11.h, p2/M, z16.s\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" + "ld1w { z16.s }, p0/Z, [x21]\n" + ".inst 0x648aaa0b // bfcvtnt z11.h, p2/M, z16.s\n" + "mov x12, #0x4\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" + "ld1w { z16.s }, p0/Z, [x21]\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0x658aaa0c // bfcvt z12.h, p2/M, z16.s\n" + ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" + "ld1w { z16.s }, p0/Z, [x21]\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0x648aaa0c // bfcvtnt z12.h, p2/M, z16.s\n" + ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" + "ld1w { z16.s }, p0/Z, [x21]\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0x658aaa0d // bfcvt z13.h, p2/M, z16.s\n" + ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" + "mov x12, #0x8\n" + "ld1w { z16.s }, p0/Z, [x21]\n" + ".inst 0x648aaa0d // bfcvtnt z13.h, p2/M, z16.s\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" + "ld1w { z16.s }, p0/Z, [x21]\n" + ".inst 0x658aaa0e // bfcvt z14.h, p2/M, z16.s\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" + "ld1w { z16.s }, p0/Z, [x21]\n" + ".inst 0x648aaa0e // bfcvtnt z14.h, p2/M, z16.s\n" + ".inst 0xc1311150 // bfdot za.s[x8, 0], { z10.h-z13.h }, z1.h\n" + "addvl x20, SP, #6\n" + "add x21, x21, %x[ld_in_row], LSL #2\n" + ".inst 0xc1391170 // bfdot za.s[x8, 0], { z11.h-z14.h }, z9.h\n" + ".inst 0xa1402a81 // ld1h { z1.h, z9.h }, pn10.b/Z, [x20]\n" + ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" + "addvl x19, SP, #12\n" + ".inst 0xc1311151 // bfdot za.s[x8, 1], { z10.h-z13.h }, z1.h\n" + "ld1w { z16.s }, p0/Z, [x21]\n" + ".inst 0x658aaa0f // bfcvt z15.h, p2/M, z16.s\n" + "sub x15, x15, #0x1\n" + ".inst 0xc1391171 // bfdot za.s[x8, 1], { z11.h-z14.h }, z9.h\n" + ".inst 0xa1402a61 // ld1h { z1.h, z9.h }, pn10.b/Z, [x19]\n" + ".inst 0xc1311152 // bfdot za.s[x8, 2], { z10.h-z13.h }, z1.h\n" + ".inst 0xc1321190 // bfdot za.s[x8, 0], { z12.h-z15.h }, z2.h\n" + "ld1h { z2.h }, p2/Z, [x20, #2, MUL VL]\n" + ".inst 0xc1391172 // bfdot za.s[x8, 2], { z11.h-z14.h }, z9.h\n" + ".inst 0xc1321191 // bfdot za.s[x8, 1], { z12.h-z15.h }, z2.h\n" + "ld1h { z2.h }, p2/Z, [x19, #2, MUL VL]\n" + ".inst 0xc0060c18 // mova { z24.d-z27.d }, za.d[x8, #0]\n" + ".inst 0xc1b6cbd8 // fclamp { z24.s-z27.s }, z30.s, z22.s\n" + "st1w { z24.s }, p1, [x14]\n" + "add x14, x14, x11, LSL #2\n" + ".inst 0xc1321192 // bfdot za.s[x8, 2], { z12.h-z15.h }, z2.h\n" + "add x8, x8, #0x1\n" + "st1w { z25.s }, p1, [x13]\n" + "add x13, x13, x10, LSL #2\n" + "st1w { z26.s }, p1, [x9]\n" + "add x9, x9, x27, LSL #2\n" + ".inst 0xc0040c84 // mova za.d[x8, #4], { z4.d-z7.d }\n" + "st1w { z27.s }, p1, [x28]\n" + "add x28, x28, x26, LSL #2\n" + "21:" // Tail input: End + "cbz x15, 23f\n" + "22:" // Right padding loop + ".inst 0xc0060c18 // mova { z24.d-z27.d }, za.d[x8, #0]\n" + "add x8, x8, #0x1\n" + "subs x15, x15, #0x1\n" + ".inst 0xc1b6cbd8 // fclamp { z24.s-z27.s }, z30.s, z22.s\n" + "st1w { z24.s }, p1, [x14]\n" + "add x14, x14, x11, LSL #2\n" + ".inst 0xc0040c84 // mova za.d[x8, #4], { z4.d-z7.d }\n" + "st1w { z25.s }, p1, [x13]\n" + "add x13, x13, x10, LSL #2\n" + "st1w { z26.s }, p1, [x9]\n" + "add x9, x9, x27, LSL #2\n" + "st1w { z27.s }, p1, [x28]\n" + "add x28, x28, x26, LSL #2\n" + "bgt 22b\n" + "23:" // End + "ldr x19, [%x[args], %[offsetof_Args_weights]]\n" + "incb x19, ALL, MUL #16\n" + "incb x19, ALL, MUL #9\n" + "str x19, [%x[args], %[offsetof_Args_weights]]\n" + "ldr x19, [%x[args], %[offsetof_Args_ld_in_vl]]\n" + "incw x7\n" + "whilelt p1.s, x7, x6\n" + "ldr x16, [%x[args], %[offsetof_Args_inptr]]\n" + "add x16, x16, x19, LSL #2\n" + "str x16, [%x[args], %[offsetof_Args_inptr]]\n" + "ldr x24, [%x[args], %[offsetof_Args_outptrs]]\n" + "ldr x23, [%x[args], %[offsetof_Args_ld_out_vls]]\n" + "ldp x22, x21, [x24, #0x0]\n" + "ldp x20, x19, [x23, #0x0]\n" + "add x22, x22, x20, LSL #2\n" + "add x21, x21, x19, LSL #2\n" + "stp x22, x21, [x24, #0x0]\n" + "ldp x22, x21, [x24, #0x10]\n" + "ldp x20, x19, [x23, #0x10]\n" + "add x22, x22, x20, LSL #2\n" + "add x21, x21, x19, LSL #2\n" + "stp x22, x21, [x24, #0x10]\n" + "b.any 1b\n" + "addvl SP, SP, #15\n" + ".inst 0xd503467f // SMSTOP\n" + : + : [args] "r" (&args), [ld_in_col] "r" (ld_in_col), [ld_in_row] "r" (ld_in_row), [offsetof_Args_bias] "I" (offsetof(Args, bias)), [offsetof_Args_clamp_max] "I" (offsetof(Args, clamp_max)), [offsetof_Args_clamp_min] "I" (offsetof(Args, clamp_min)), [offsetof_Args_current_channel] "I" (offsetof(Args, current_channel)), [offsetof_Args_inptr] "I" (offsetof(Args, inptr)), [offsetof_Args_input_cols] "I" (offsetof(Args, input_cols)), [offsetof_Args_ld_in_vl] "I" (offsetof(Args, ld_in_vl)), [offsetof_Args_ld_out_cols] "I" (offsetof(Args, ld_out_cols)), [offsetof_Args_ld_out_vls] "I" (offsetof(Args, ld_out_vls)), [offsetof_Args_n_channels] "I" (offsetof(Args, n_channels)), [offsetof_Args_outptrs] "I" (offsetof(Args, outptrs)), [offsetof_Args_output_cols] "I" (offsetof(Args, output_cols)), [offsetof_Args_pad_bottom] "I" (offsetof(Args, pad_bottom)), [offsetof_Args_pad_left] "I" (offsetof(Args, pad_left)), [offsetof_Args_pad_top] "I" (offsetof(Args, pad_top)), [offsetof_Args_weights] "I" (offsetof(Args, weights)) + : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x4", "x5", "x6", "x7", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x19", "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" + ); +} + +} // namespace depthwise +} // namespace arm_conv + +#endif // defined(ARM_COMPUTE_ENABLE_SME2) |