From be13cead34e566bdd561ad3ffc3f645b460e482e Mon Sep 17 00:00:00 2001 From: Michael Tyler Date: Tue, 17 Jan 2023 11:04:14 +0000 Subject: Revert "Update CPU kernels to remove x19" This reverts commit 3c59f01c209d2732a15d97d65565ead964787a8b. Resolves: COMPMID-5817 Change-Id: Ie2443a21854a95db1e3d0cafa2121c0187a5e237 Signed-off-by: Michael Tyler Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8974 Comments-Addressed: Arm Jenkins Tested-by: Arm Jenkins Reviewed-by: Gian Marco Iodice Benchmark: Arm Jenkins --- .../generic.cpp | 1306 ++++++++++---------- 1 file changed, 653 insertions(+), 653 deletions(-) (limited to 'src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8s8u8q_planar_5x5_s2_4rows_dot_za/generic.cpp') diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8s8u8q_planar_5x5_s2_4rows_dot_za/generic.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8s8u8q_planar_5x5_s2_4rows_dot_za/generic.cpp index 33bb4eb8ec..8cdc94d0e9 100644 --- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8s8u8q_planar_5x5_s2_4rows_dot_za/generic.cpp +++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_u8s8u8q_planar_5x5_s2_4rows_dot_za/generic.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2022-2023 Arm Limited. + * Copyright (c) 2022 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -69,20 +69,20 @@ void sme2_u8s8u8q_planar_5x5_s2_4rows_dot_za_impl( __asm__ __volatile__( ".inst 0xd503477f // SMSTART ZA\n" - "ldr x3, [%x[args], %[offsetof_Args_pad_bottom]]\n" + "ldr x4, [%x[args], %[offsetof_Args_pad_bottom]]\n" "ptrue p2.b\n" - "mov x20, #0xb\n" - "ldr x4, [%x[args], %[offsetof_Args_pad_top]]\n" + "mov x19, #0xb\n" + "ldr x5, [%x[args], %[offsetof_Args_pad_top]]\n" "ld1rh { z9.h }, p2/Z, [%x[qp], %[offsetof_Requantize32_a_offset]]\n" - "sub x20, x20, x3\n" + "sub x19, x19, x4\n" ".inst 0x25207812 // ptrue pn10.b\n" - "ldr x5, [%x[args], %[offsetof_Args_n_channels]]\n" - "whilelt p1.s, XZR, x5\n" - "whilelt p9.s, XZR, x20\n" + "ldr x6, [%x[args], %[offsetof_Args_n_channels]]\n" + "whilelt p1.s, XZR, x6\n" + "whilelt p9.s, XZR, x19\n" "ld1rw { z8.s }, p2/Z, [%x[qp], %[offsetof_Requantize32_c_offset]]\n" - "whilelt p8.s, XZR, x4\n" + "whilelt p8.s, XZR, x5\n" "addvl SP, SP, #-15\n" - "ldr x6, [%x[args], %[offsetof_Args_current_channel]]\n" + "ldr x7, [%x[args], %[offsetof_Args_current_channel]]\n" "neg z9.h, p2/M, z9.h\n" "eor p8.b, p2/Z, p8.b, p9.b\n" "ld1rw { z3.s }, p2/Z, [%x[qp], %[offsetof_Requantize32_per_layer_mul]]\n" @@ -90,227 +90,310 @@ void sme2_u8s8u8q_planar_5x5_s2_4rows_dot_za_impl( "ld1rw { z26.s }, p2/Z, [%x[qp], %[offsetof_Requantize32_minval]]\n" "ld1rw { z23.s }, p2/Z, [%x[qp], %[offsetof_Requantize32_maxval]]\n" "1:" // Channel loop - "ldr x20, [%x[qp], %[offsetof_Requantize32_bias]]\n" + "ldr x19, [%x[qp], %[offsetof_Requantize32_bias]]\n" "mov z28.s, #0x0\n" - "cbz x20, 2f\n" - "ld1w { z28.s }, p1/Z, [x20, x6, LSL #2]\n" + "cbz x19, 2f\n" + "ld1w { z28.s }, p1/Z, [x19, x7, LSL #2]\n" "2:" // Load bias: Done - "ldr x22, [%x[args], %[offsetof_Args_weights]]\n" - "mov x20, x22\n" - "ld1sb { z12.s }, p2/Z, [x20]\n" - "incw x20, ALL, MUL #5\n" + "ldr x21, [%x[args], %[offsetof_Args_weights]]\n" + "mov x19, x21\n" + "ld1sb { z12.s }, p2/Z, [x19]\n" + "incw x19, ALL, MUL #5\n" "ld1rh { z18.h }, p2/Z, [%x[qp], %[offsetof_Requantize32_b_offset]]\n" "sub z12.h, z12.h, z18.h\n" - "incw x22\n" + "incw x21\n" "mov z14.h, #0x0\n" - "ld1sb { z25.s }, p2/Z, [x20]\n" - "incw x20, ALL, MUL #5\n" + "ld1sb { z25.s }, p2/Z, [x19]\n" + "incw x19, ALL, MUL #5\n" "sub z25.h, z25.h, z18.h\n" "trn1 z2.h, z12.h, z25.h\n" - "ld1sb { z24.s }, p2/Z, [x20]\n" - "incw x20, ALL, MUL #5\n" + "ld1sb { z24.s }, p2/Z, [x19]\n" + "incw x19, ALL, MUL #5\n" "sub z24.h, z24.h, z18.h\n" - "addvl x21, SP, #15\n" - "ld1sb { z17.s }, p2/Z, [x20]\n" - "incw x20, ALL, MUL #5\n" + "addvl x20, SP, #15\n" + "ld1sb { z17.s }, p2/Z, [x19]\n" + "incw x19, ALL, MUL #5\n" "sub z17.h, z17.h, z18.h\n" "trn1 z10.h, z24.h, z17.h\n" - "ld1sb { z16.s }, p2/Z, [x20]\n" - "mov x20, x22\n" + "ld1sb { z16.s }, p2/Z, [x19]\n" + "mov x19, x21\n" "sub z16.h, z16.h, z18.h\n" - "incw x22\n" - "ld1sb { z12.s }, p2/Z, [x20]\n" - "incw x20, ALL, MUL #5\n" + "incw x21\n" + "ld1sb { z12.s }, p2/Z, [x19]\n" + "incw x19, ALL, MUL #5\n" "sub z12.h, z12.h, z18.h\n" - "addvl x21, x21, #-3\n" - "ld1sb { z25.s }, p2/Z, [x20]\n" - "incw x20, ALL, MUL #5\n" + "addvl x20, x20, #-3\n" + "ld1sb { z25.s }, p2/Z, [x19]\n" + "incw x19, ALL, MUL #5\n" "sub z25.h, z25.h, z18.h\n" "trn1 z0.h, z16.h, z14.h\n" - "ld1sb { z24.s }, p2/Z, [x20]\n" - "incw x20, ALL, MUL #5\n" + "ld1sb { z24.s }, p2/Z, [x19]\n" + "incw x19, ALL, MUL #5\n" "sub z24.h, z24.h, z18.h\n" - "st1h { z2.h }, p2, [x21]\n" - "ld1sb { z17.s }, p2/Z, [x20]\n" - "incw x20, ALL, MUL #5\n" + "st1h { z2.h }, p2, [x20]\n" + "ld1sb { z17.s }, p2/Z, [x19]\n" + "incw x19, ALL, MUL #5\n" "sub z17.h, z17.h, z18.h\n" "trn1 z2.h, z12.h, z25.h\n" - "ld1sb { z16.s }, p2/Z, [x20]\n" - "mov x20, x22\n" - "st1h { z10.h }, p2, [x21, #1, MUL VL]\n" + "ld1sb { z16.s }, p2/Z, [x19]\n" + "mov x19, x21\n" + "st1h { z10.h }, p2, [x20, #1, MUL VL]\n" "sub z16.h, z16.h, z18.h\n" - "ld1sb { z12.s }, p2/Z, [x20]\n" - "incw x20, ALL, MUL #5\n" + "ld1sb { z12.s }, p2/Z, [x19]\n" + "incw x19, ALL, MUL #5\n" "trn1 z10.h, z24.h, z17.h\n" "sub z12.h, z12.h, z18.h\n" - "ld1sb { z25.s }, p2/Z, [x20]\n" - "incw x20, ALL, MUL #5\n" + "ld1sb { z25.s }, p2/Z, [x19]\n" + "incw x19, ALL, MUL #5\n" "sub z25.h, z25.h, z18.h\n" - "st1h { z0.h }, p2, [x21, #2, MUL VL]\n" - "ld1sb { z24.s }, p2/Z, [x20]\n" - "incw x20, ALL, MUL #5\n" + "st1h { z0.h }, p2, [x20, #2, MUL VL]\n" + "ld1sb { z24.s }, p2/Z, [x19]\n" + "incw x19, ALL, MUL #5\n" "trn1 z0.h, z16.h, z14.h\n" - "incw x22\n" - "ld1sb { z17.s }, p2/Z, [x20]\n" - "incw x20, ALL, MUL #5\n" + "incw x21\n" + "ld1sb { z17.s }, p2/Z, [x19]\n" + "incw x19, ALL, MUL #5\n" "sub z24.h, z24.h, z18.h\n" "sub z17.h, z17.h, z18.h\n" - "ld1sb { z16.s }, p2/Z, [x20]\n" - "addvl x21, x21, #-3\n" - "mov x20, x22\n" - "st1h { z2.h }, p2, [x21]\n" + "ld1sb { z16.s }, p2/Z, [x19]\n" + "addvl x20, x20, #-3\n" + "mov x19, x21\n" + "st1h { z2.h }, p2, [x20]\n" "trn1 z2.h, z12.h, z25.h\n" - "ld1sb { z12.s }, p2/Z, [x20]\n" - "incw x20, ALL, MUL #5\n" + "ld1sb { z12.s }, p2/Z, [x19]\n" + "incw x19, ALL, MUL #5\n" "sub z16.h, z16.h, z18.h\n" - "ld1sb { z25.s }, p2/Z, [x20]\n" - "incw x20, ALL, MUL #5\n" - "st1h { z10.h }, p2, [x21, #1, MUL VL]\n" + "ld1sb { z25.s }, p2/Z, [x19]\n" + "incw x19, ALL, MUL #5\n" + "st1h { z10.h }, p2, [x20, #1, MUL VL]\n" "trn1 z10.h, z24.h, z17.h\n" - "ld1sb { z24.s }, p2/Z, [x20]\n" - "incw x20, ALL, MUL #5\n" + "ld1sb { z24.s }, p2/Z, [x19]\n" + "incw x19, ALL, MUL #5\n" "sub z12.h, z12.h, z18.h\n" "sub z25.h, z25.h, z18.h\n" - "ld1sb { z17.s }, p2/Z, [x20]\n" - "incw x20, ALL, MUL #5\n" - "st1h { z0.h }, p2, [x21, #2, MUL VL]\n" + "ld1sb { z17.s }, p2/Z, [x19]\n" + "incw x19, ALL, MUL #5\n" + "st1h { z0.h }, p2, [x20, #2, MUL VL]\n" "trn1 z0.h, z16.h, z14.h\n" - "ld1sb { z16.s }, p2/Z, [x20]\n" - "incw x22\n" + "ld1sb { z16.s }, p2/Z, [x19]\n" + "incw x21\n" "sub z24.h, z24.h, z18.h\n" "sub z17.h, z17.h, z18.h\n" - "addvl x21, x21, #-3\n" - "mov x20, x22\n" - "st1h { z2.h }, p2, [x21]\n" + "addvl x20, x20, #-3\n" + "mov x19, x21\n" + "st1h { z2.h }, p2, [x20]\n" "sub z16.h, z16.h, z18.h\n" "trn1 z2.h, z12.h, z25.h\n" - "ld1sb { z12.s }, p2/Z, [x20]\n" - "incw x20, ALL, MUL #5\n" - "st1h { z10.h }, p2, [x21, #1, MUL VL]\n" - "ld1sb { z25.s }, p2/Z, [x20]\n" - "incw x20, ALL, MUL #5\n" + "ld1sb { z12.s }, p2/Z, [x19]\n" + "incw x19, ALL, MUL #5\n" + "st1h { z10.h }, p2, [x20, #1, MUL VL]\n" + "ld1sb { z25.s }, p2/Z, [x19]\n" + "incw x19, ALL, MUL #5\n" "trn1 z10.h, z24.h, z17.h\n" - "st1h { z0.h }, p2, [x21, #2, MUL VL]\n" - "ld1sb { z24.s }, p2/Z, [x20]\n" - "incw x20, ALL, MUL #5\n" + "st1h { z0.h }, p2, [x20, #2, MUL VL]\n" + "ld1sb { z24.s }, p2/Z, [x19]\n" + "incw x19, ALL, MUL #5\n" "trn1 z0.h, z16.h, z14.h\n" "sub z12.h, z12.h, z18.h\n" - "ld1sb { z17.s }, p2/Z, [x20]\n" - "incw x20, ALL, MUL #5\n" + "ld1sb { z17.s }, p2/Z, [x19]\n" + "incw x19, ALL, MUL #5\n" "sub z25.h, z25.h, z18.h\n" "sub z24.h, z24.h, z18.h\n" - "ld1sb { z16.s }, p2/Z, [x20]\n" + "ld1sb { z16.s }, p2/Z, [x19]\n" "sub z17.h, z17.h, z18.h\n" "sub z16.h, z16.h, z18.h\n" - "ldr x20, [%x[qp], %[offsetof_Requantize32_per_channel_muls]]\n" - "addvl x21, x21, #-3\n" - "st1h { z2.h }, p2, [x21]\n" + "ldr x19, [%x[qp], %[offsetof_Requantize32_per_channel_muls]]\n" + "addvl x20, x20, #-3\n" + "st1h { z2.h }, p2, [x20]\n" "mov z29.d, z28.d\n" "mov z30.d, z28.d\n" - "st1h { z10.h }, p2, [x21, #1, MUL VL]\n" + "st1h { z10.h }, p2, [x20, #1, MUL VL]\n" "mov z31.d, z28.d\n" "trn1 z2.h, z12.h, z25.h\n" - "st1h { z0.h }, p2, [x21, #2, MUL VL]\n" - "addvl x21, x21, #-3\n" + "st1h { z0.h }, p2, [x20, #2, MUL VL]\n" + "addvl x20, x20, #-3\n" "trn1 z10.h, z24.h, z17.h\n" "trn1 z0.h, z16.h, z14.h\n" - "st1h { z2.h }, p2, [x21]\n" - "st1h { z10.h }, p2, [x21, #1, MUL VL]\n" - "st1h { z0.h }, p2, [x21, #2, MUL VL]\n" - "cbz x20, 3f\n" - "ld1w { z3.s }, p1/Z, [x20, x6, LSL #2]\n" + "st1h { z2.h }, p2, [x20]\n" + "st1h { z10.h }, p2, [x20, #1, MUL VL]\n" + "st1h { z0.h }, p2, [x20, #2, MUL VL]\n" + "cbz x19, 3f\n" + "ld1w { z3.s }, p1/Z, [x19, x7, LSL #2]\n" "3:" // Load mul: End - "ldr x20, [%x[qp], %[offsetof_Requantize32_per_channel_right_shifts]]\n" - "cbz x20, 4f\n" - "ld1w { z1.s }, p1/Z, [x20, x6, LSL #2]\n" + "ldr x19, [%x[qp], %[offsetof_Requantize32_per_channel_right_shifts]]\n" + "cbz x19, 4f\n" + "ld1w { z1.s }, p1/Z, [x19, x7, LSL #2]\n" "4:" // Load right_shift: End - "ldr x7, [%x[args], %[offsetof_Args_input_cols]]\n" - "sub x20, x7, #0x1\n" - "orr x23, x20, %x[ld_in_col], LSL #16\n" - "ldr x17, [%x[args], %[offsetof_Args_inptr]]\n" - "orr x23, x5, x23, LSL #22\n" - "mov x22, #0xb\n" - "add x21, x4, x3\n" - "lsl x20, %x[ld_in_row], #0x0\n" - "ldr x16, [%x[args], %[offsetof_Args_output_cols]]\n" + "ldr x17, [%x[args], %[offsetof_Args_input_cols]]\n" + "sub x19, x17, #0x1\n" + "orr x22, x19, %x[ld_in_col], LSL #16\n" + "ldr x16, [%x[args], %[offsetof_Args_inptr]]\n" + "orr x22, x6, x22, LSL #22\n" + "mov x21, #0xb\n" + "add x20, x5, x4\n" + "lsl x19, %x[ld_in_row], #0x0\n" + "ldr x15, [%x[args], %[offsetof_Args_output_cols]]\n" "mov x8, #0x0\n" - "lsl x23, x23, #0x0\n" - "sub x22, x22, x21\n" - "madd x20, x20, x4, x17\n" + "lsl x22, x22, #0x0\n" + "sub x21, x21, x20\n" + "madd x19, x19, x5, x16\n" "5:" // Issue prefetches - "subs x22, x22, #0x1\n" - ".inst 0xf8b74a9c // rprfm pldstrm, x23, [x20]\n" - "add x20, x20, %x[ld_in_col]\n" + "subs x21, x21, #0x1\n" + ".inst 0xf8b64a7c // rprfm pldstrm, x22, [x19]\n" + "add x19, x19, %x[ld_in_col]\n" "bgt 5b\n" - "ldr x25, [%x[args], %[offsetof_Args_outptrs]]\n" - "lsl x20, %x[ld_in_row], #0x0\n" - "msub x17, x4, x20, x17\n" + "ldr x24, [%x[args], %[offsetof_Args_outptrs]]\n" + "lsl x19, %x[ld_in_row], #0x0\n" + "msub x16, x5, x19, x16\n" ".inst 0xc0040f80 // mova za.d[x8, #0], { z28.d-z31.d }\n" - "ldr x20, [%x[args], %[offsetof_Args_ld_out_cols]]\n" + "ldr x19, [%x[args], %[offsetof_Args_ld_out_cols]]\n" ".inst 0xc0040f81 // mova za.d[x8, #1], { z28.d-z31.d }\n" - "mov x22, #0x4\n" - "ldp x15, x14, [x25], #0x10\n" + "mov x21, #0x4\n" + "ldp x14, x13, [x24], #0x10\n" ".inst 0xc0040f82 // mova za.d[x8, #2], { z28.d-z31.d }\n" - "ldp x13, x11, [x20], #0x10\n" + "ldp x11, x10, [x19], #0x10\n" ".inst 0xc0040f83 // mova za.d[x8, #3], { z28.d-z31.d }\n" - "ldr x21, [%x[args], %[offsetof_Args_pad_left]]\n" + "ldr x20, [%x[args], %[offsetof_Args_pad_left]]\n" ".inst 0xc0040f84 // mova za.d[x8, #4], { z28.d-z31.d }\n" - "ldp x10, x9, [x25], #0x10\n" - "ldp x28, x27, [x20], #0x10\n" - "cbz x21, 7f\n" - "cmp x21, x22\n" - "csel x20, x21, x22, LT\n" - "sub x21, x21, x20\n" - "sub x22, x22, x20\n" - "cbz x21, 7f\n" + "ldp x9, x28, [x24], #0x10\n" + "ldp x27, x26, [x19], #0x10\n" + "cbz x20, 7f\n" + "cmp x20, x21\n" + "csel x19, x20, x21, LT\n" + "sub x20, x20, x19\n" + "sub x21, x21, x19\n" + "cbz x20, 7f\n" ".inst 0xc0060c04 // mova { z4.d-z7.d }, za.d[x8, #0]\n" ".inst 0xc1a3ac04 // sqdmulh { z4.s-z7.s }, { z4.s-z7.s }, z3.s\n" - "and x22, x21, #0x1\n" + "and x21, x20, #0x1\n" ".inst 0xc1a1aa24 // srshl { z4.s-z7.s }, { z4.s-z7.s }, z1.s\n" - "add x21, x21, #0x1\n" - "lsr x21, x21, #0x1\n" + "add x20, x20, #0x1\n" + "lsr x20, x20, #0x1\n" ".inst 0xc1a8ab04 // add { z4.s-z7.s }, { z4.s-z7.s }, z8.s\n" - "sub x16, x16, x21\n" + "sub x15, x15, x20\n" ".inst 0xc1b7cf44 // sclamp { z4.s-z7.s }, z26.s, z23.s\n" "6:" // Left padding - "subs x21, x21, #0x1\n" - "st1b { z4.s }, p1, [x15]\n" - "add x15, x15, x13\n" - "st1b { z5.s }, p1, [x14]\n" + "subs x20, x20, #0x1\n" + "st1b { z4.s }, p1, [x14]\n" "add x14, x14, x11\n" - "st1b { z6.s }, p1, [x10]\n" - "add x10, x10, x28\n" - "st1b { z7.s }, p1, [x9]\n" + "st1b { z5.s }, p1, [x13]\n" + "add x13, x13, x10\n" + "st1b { z6.s }, p1, [x9]\n" "add x9, x9, x27\n" + "st1b { z7.s }, p1, [x28]\n" + "add x28, x28, x26\n" "bgt 6b\n" "7:" // Left padding: End - "adds XZR, x4, x3\n" + "adds XZR, x5, x4\n" "bne 14f\n" - "cbz x22, 12f\n" - "cmp x22, #0x1\n" - "sub x7, x7, x22\n" + "cbz x21, 12f\n" + "cmp x21, #0x1\n" + "sub x17, x17, x21\n" "beq 11f\n" - "cmp x22, #0x2\n" + "cmp x21, #0x2\n" "beq 10f\n" - "cmp x22, #0x3\n" + "cmp x21, #0x3\n" "beq 9f\n" "8:" // Unpadded: 4 priming loads - "add x21, x17, %x[ld_in_row]\n" - "ld1b { z11.s }, p1/Z, [x17]\n" - "addvl x20, SP, #12\n" + "add x20, x16, %x[ld_in_row]\n" + "ld1b { z11.s }, p1/Z, [x16]\n" + "addvl x19, SP, #12\n" + "ld1b { z21.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row]\n" + "trn1 z11.h, z11.h, z21.h\n" + "add z11.h, z11.h, z9.h\n" + "ld1b { z12.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row]\n" + "add x16, x16, %x[ld_in_col]\n" + "ld1b { z20.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row]\n" + "trn1 z12.h, z12.h, z20.h\n" + "add z12.h, z12.h, z9.h\n" + "ld1b { z13.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row]\n" + "ld1b { z19.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row]\n" + "trn1 z13.h, z13.h, z19.h\n" + "add z13.h, z13.h, z9.h\n" + "ld1b { z14.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row]\n" + "ld1b { z18.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row]\n" + "trn1 z14.h, z14.h, z18.h\n" + "add z14.h, z14.h, z9.h\n" + "ld1b { z15.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row]\n" + "ld1b { z17.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row]\n" + "trn1 z15.h, z15.h, z17.h\n" + "add z15.h, z15.h, z9.h\n" + ".inst 0xa1402a62 // ld1h { z2.h, z10.h }, pn10.b/Z, [x19]\n" + ".inst 0xc1721568 // sdot za.s[x8, 0], { z11.h-z14.h }, z2.h\n" + "ld1b { z16.s }, p1/Z, [x20]\n" + "mov z16.d, z16.d\n" + "add z16.h, z16.h, z9.h\n" + ".inst 0xc17a1588 // sdot za.s[x8, 0], { z12.h-z15.h }, z10.h\n" + "ld1h { z0.h }, p2/Z, [x19, #2, MUL VL]\n" + ".inst 0xc17015a8 // sdot za.s[x8, 0], { z13.h-z16.h }, z0.h\n" + "9:" // Unpadded: 3 priming loads + "add x20, x16, %x[ld_in_row]\n" + "ld1b { z11.s }, p1/Z, [x16]\n" + "addvl x19, SP, #9\n" + "ld1b { z21.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row]\n" + "trn1 z11.h, z11.h, z21.h\n" + "add z11.h, z11.h, z9.h\n" + "ld1b { z12.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row]\n" + "add x16, x16, %x[ld_in_col]\n" + "ld1b { z20.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row]\n" + "trn1 z12.h, z12.h, z20.h\n" + "add z12.h, z12.h, z9.h\n" + "ld1b { z13.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row]\n" + "ld1b { z19.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row]\n" + "trn1 z13.h, z13.h, z19.h\n" + "add z13.h, z13.h, z9.h\n" + "ld1b { z14.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row]\n" + "ld1b { z18.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row]\n" + "trn1 z14.h, z14.h, z18.h\n" + "add z14.h, z14.h, z9.h\n" + "ld1b { z15.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row]\n" + "ld1b { z17.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row]\n" + "trn1 z15.h, z15.h, z17.h\n" + "add z15.h, z15.h, z9.h\n" + ".inst 0xa1402a62 // ld1h { z2.h, z10.h }, pn10.b/Z, [x19]\n" + ".inst 0xc1721568 // sdot za.s[x8, 0], { z11.h-z14.h }, z2.h\n" + "ld1b { z16.s }, p1/Z, [x20]\n" + "mov z16.d, z16.d\n" + "add z16.h, z16.h, z9.h\n" + ".inst 0xc17a1588 // sdot za.s[x8, 0], { z12.h-z15.h }, z10.h\n" + "ld1h { z0.h }, p2/Z, [x19, #2, MUL VL]\n" + ".inst 0xc17015a8 // sdot za.s[x8, 0], { z13.h-z16.h }, z0.h\n" + "10:" // Unpadded: 2 priming loads + "add x21, x16, %x[ld_in_row]\n" + "ld1b { z11.s }, p1/Z, [x16]\n" + "addvl x20, SP, #6\n" "ld1b { z21.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row]\n" "trn1 z11.h, z11.h, z21.h\n" "add z11.h, z11.h, z9.h\n" "ld1b { z12.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row]\n" - "add x17, x17, %x[ld_in_col]\n" + "addvl x19, SP, #12\n" "ld1b { z20.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row]\n" "trn1 z12.h, z12.h, z20.h\n" "add z12.h, z12.h, z9.h\n" "ld1b { z13.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row]\n" + "add x16, x16, %x[ld_in_col]\n" "ld1b { z19.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row]\n" "trn1 z13.h, z13.h, z19.h\n" @@ -324,34 +407,40 @@ void sme2_u8s8u8q_planar_5x5_s2_4rows_dot_za_impl( "ld1b { z15.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row]\n" "ld1b { z17.s }, p1/Z, [x21]\n" - "add x21, x21, %x[ld_in_row]\n" "trn1 z15.h, z15.h, z17.h\n" + "add x21, x21, %x[ld_in_row]\n" "add z15.h, z15.h, z9.h\n" ".inst 0xa1402a82 // ld1h { z2.h, z10.h }, pn10.b/Z, [x20]\n" ".inst 0xc1721568 // sdot za.s[x8, 0], { z11.h-z14.h }, z2.h\n" "ld1b { z16.s }, p1/Z, [x21]\n" "mov z16.d, z16.d\n" - "add z16.h, z16.h, z9.h\n" ".inst 0xc17a1588 // sdot za.s[x8, 0], { z12.h-z15.h }, z10.h\n" + ".inst 0xa1402a62 // ld1h { z2.h, z10.h }, pn10.b/Z, [x19]\n" + ".inst 0xc1721569 // sdot za.s[x8, 1], { z11.h-z14.h }, z2.h\n" + "add z16.h, z16.h, z9.h\n" "ld1h { z0.h }, p2/Z, [x20, #2, MUL VL]\n" + ".inst 0xc17a1589 // sdot za.s[x8, 1], { z12.h-z15.h }, z10.h\n" ".inst 0xc17015a8 // sdot za.s[x8, 0], { z13.h-z16.h }, z0.h\n" - "9:" // Unpadded: 3 priming loads - "add x21, x17, %x[ld_in_row]\n" - "ld1b { z11.s }, p1/Z, [x17]\n" - "addvl x20, SP, #9\n" + "ld1h { z0.h }, p2/Z, [x19, #2, MUL VL]\n" + ".inst 0xc17015a9 // sdot za.s[x8, 1], { z13.h-z16.h }, z0.h\n" + "11:" // Unpadded: 1 priming loads + "add x21, x16, %x[ld_in_row]\n" + "ld1b { z11.s }, p1/Z, [x16]\n" + "addvl x20, SP, #3\n" "ld1b { z21.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row]\n" "trn1 z11.h, z11.h, z21.h\n" "add z11.h, z11.h, z9.h\n" "ld1b { z12.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row]\n" - "add x17, x17, %x[ld_in_col]\n" + "addvl x19, SP, #12\n" "ld1b { z20.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row]\n" "trn1 z12.h, z12.h, z20.h\n" "add z12.h, z12.h, z9.h\n" "ld1b { z13.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row]\n" + "add x16, x16, %x[ld_in_col]\n" "ld1b { z19.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row]\n" "trn1 z13.h, z13.h, z19.h\n" @@ -365,100 +454,127 @@ void sme2_u8s8u8q_planar_5x5_s2_4rows_dot_za_impl( "ld1b { z15.s }, p1/Z, [x21]\n" "add x21, x21, %x[ld_in_row]\n" "ld1b { z17.s }, p1/Z, [x21]\n" - "add x21, x21, %x[ld_in_row]\n" "trn1 z15.h, z15.h, z17.h\n" + "add x21, x21, %x[ld_in_row]\n" "add z15.h, z15.h, z9.h\n" ".inst 0xa1402a82 // ld1h { z2.h, z10.h }, pn10.b/Z, [x20]\n" ".inst 0xc1721568 // sdot za.s[x8, 0], { z11.h-z14.h }, z2.h\n" "ld1b { z16.s }, p1/Z, [x21]\n" "mov z16.d, z16.d\n" - "add z16.h, z16.h, z9.h\n" ".inst 0xc17a1588 // sdot za.s[x8, 0], { z12.h-z15.h }, z10.h\n" + ".inst 0xa1402a62 // ld1h { z2.h, z10.h }, pn10.b/Z, [x19]\n" + ".inst 0xc1721569 // sdot za.s[x8, 1], { z11.h-z14.h }, z2.h\n" + "add z16.h, z16.h, z9.h\n" "ld1h { z0.h }, p2/Z, [x20, #2, MUL VL]\n" + ".inst 0xc17a1589 // sdot za.s[x8, 1], { z12.h-z15.h }, z10.h\n" ".inst 0xc17015a8 // sdot za.s[x8, 0], { z13.h-z16.h }, z0.h\n" - "10:" // Unpadded: 2 priming loads - "add x22, x17, %x[ld_in_row]\n" - "ld1b { z11.s }, p1/Z, [x17]\n" - "addvl x21, SP, #6\n" - "ld1b { z21.s }, p1/Z, [x22]\n" - "add x22, x22, %x[ld_in_row]\n" + "ld1h { z0.h }, p2/Z, [x19, #2, MUL VL]\n" + ".inst 0xc17015a9 // sdot za.s[x8, 1], { z13.h-z16.h }, z0.h\n" + "12:" // Unpadded: 0 priming loads + "cmp x17, #0x2\n" + ".inst 0xa1402be2 // ld1h { z2.h, z10.h }, pn10.b/Z, [SP]\n" + "ld1h { z0.h }, p2/Z, [SP, #2, MUL VL]\n" + "blt 22f\n" + "add x20, x16, %x[ld_in_row]\n" + "ld1b { z11.s }, p1/Z, [x16]\n" + "sub x17, x17, #0x2\n" + "ld1b { z21.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row]\n" "trn1 z11.h, z11.h, z21.h\n" + "sub x15, x15, #0x1\n" + "ld1b { z12.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row]\n" + "lsr x19, x17, #0x1\n" "add z11.h, z11.h, z9.h\n" - "ld1b { z12.s }, p1/Z, [x22]\n" - "add x22, x22, %x[ld_in_row]\n" - "addvl x20, SP, #12\n" - "ld1b { z20.s }, p1/Z, [x22]\n" - "add x22, x22, %x[ld_in_row]\n" + "ld1b { z20.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row]\n" "trn1 z12.h, z12.h, z20.h\n" + "cmp x19, x15\n" + "ld1b { z13.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row]\n" + "csel x25, x19, x15, LT\n" "add z12.h, z12.h, z9.h\n" - "ld1b { z13.s }, p1/Z, [x22]\n" - "add x22, x22, %x[ld_in_row]\n" - "add x17, x17, %x[ld_in_col]\n" - "ld1b { z19.s }, p1/Z, [x22]\n" - "add x22, x22, %x[ld_in_row]\n" + "ld1b { z19.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row]\n" "trn1 z13.h, z13.h, z19.h\n" "add z13.h, z13.h, z9.h\n" - "ld1b { z14.s }, p1/Z, [x22]\n" - "add x22, x22, %x[ld_in_row]\n" - "ld1b { z18.s }, p1/Z, [x22]\n" - "add x22, x22, %x[ld_in_row]\n" + "ld1b { z14.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row]\n" + "add x16, x16, %x[ld_in_col]\n" + "ld1b { z18.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row]\n" "trn1 z14.h, z14.h, z18.h\n" "add z14.h, z14.h, z9.h\n" - "ld1b { z15.s }, p1/Z, [x22]\n" - "add x22, x22, %x[ld_in_row]\n" - "ld1b { z17.s }, p1/Z, [x22]\n" + "ld1b { z15.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row]\n" + "and x17, x17, #0x1\n" + "ld1b { z17.s }, p1/Z, [x20]\n" + "add x20, x20, %x[ld_in_row]\n" "trn1 z15.h, z15.h, z17.h\n" - "add x22, x22, %x[ld_in_row]\n" "add z15.h, z15.h, z9.h\n" - ".inst 0xa1402aa2 // ld1h { z2.h, z10.h }, pn10.b/Z, [x21]\n" - ".inst 0xc1721568 // sdot za.s[x8, 0], { z11.h-z14.h }, z2.h\n" - "ld1b { z16.s }, p1/Z, [x22]\n" + "ld1b { z16.s }, p1/Z, [x20]\n" "mov z16.d, z16.d\n" + "add z16.h, z16.h, z9.h\n" + "sub x15, x15, x25\n" + "cbz x25, 21f\n" + "13:" // Unpadded: Main loop + ".inst 0xc1721568 // sdot za.s[x8, 0], { z11.h-z14.h }, z2.h\n" + "addvl x24, SP, #6\n" + "addvl x23, SP, #12\n" ".inst 0xc17a1588 // sdot za.s[x8, 0], { z12.h-z15.h }, z10.h\n" - ".inst 0xa1402a82 // ld1h { z2.h, z10.h }, pn10.b/Z, [x20]\n" + ".inst 0xa1402b02 // ld1h { z2.h, z10.h }, pn10.b/Z, [x24]\n" + "add x22, x16, %x[ld_in_row]\n" + "addvl x21, SP, #3\n" ".inst 0xc1721569 // sdot za.s[x8, 1], { z11.h-z14.h }, z2.h\n" - "add z16.h, z16.h, z9.h\n" - "ld1h { z0.h }, p2/Z, [x21, #2, MUL VL]\n" + "addvl x20, SP, #9\n" + "subs x25, x25, #0x1\n" ".inst 0xc17a1589 // sdot za.s[x8, 1], { z12.h-z15.h }, z10.h\n" + ".inst 0xa1402ae2 // ld1h { z2.h, z10.h }, pn10.b/Z, [x23]\n" + ".inst 0xc172156a // sdot za.s[x8, 2], { z11.h-z14.h }, z2.h\n" + "ld1b { z11.s }, p1/Z, [x16]\n" + "add x16, x16, %x[ld_in_col]\n" + "add x19, x16, %x[ld_in_row]\n" ".inst 0xc17015a8 // sdot za.s[x8, 0], { z13.h-z16.h }, z0.h\n" - "ld1h { z0.h }, p2/Z, [x20, #2, MUL VL]\n" - ".inst 0xc17015a9 // sdot za.s[x8, 1], { z13.h-z16.h }, z0.h\n" - "11:" // Unpadded: 1 priming loads - "add x22, x17, %x[ld_in_row]\n" - "ld1b { z11.s }, p1/Z, [x17]\n" - "addvl x21, SP, #3\n" + "ld1h { z0.h }, p2/Z, [x24, #2, MUL VL]\n" + ".inst 0xc17a158a // sdot za.s[x8, 2], { z12.h-z15.h }, z10.h\n" "ld1b { z21.s }, p1/Z, [x22]\n" "add x22, x22, %x[ld_in_row]\n" "trn1 z11.h, z11.h, z21.h\n" + ".inst 0xc17015a9 // sdot za.s[x8, 1], { z13.h-z16.h }, z0.h\n" + "ld1h { z0.h }, p2/Z, [x23, #2, MUL VL]\n" "add z11.h, z11.h, z9.h\n" "ld1b { z12.s }, p1/Z, [x22]\n" "add x22, x22, %x[ld_in_row]\n" - "addvl x20, SP, #9\n" + ".inst 0xc17015aa // sdot za.s[x8, 2], { z13.h-z16.h }, z0.h\n" "ld1b { z20.s }, p1/Z, [x22]\n" "add x22, x22, %x[ld_in_row]\n" "trn1 z12.h, z12.h, z20.h\n" "add z12.h, z12.h, z9.h\n" "ld1b { z13.s }, p1/Z, [x22]\n" "add x22, x22, %x[ld_in_row]\n" - "add x17, x17, %x[ld_in_col]\n" + ".inst 0xc0060c04 // mova { z4.d-z7.d }, za.d[x8, #0]\n" + "add x8, x8, #0x1\n" "ld1b { z19.s }, p1/Z, [x22]\n" "add x22, x22, %x[ld_in_row]\n" "trn1 z13.h, z13.h, z19.h\n" "add z13.h, z13.h, z9.h\n" "ld1b { z14.s }, p1/Z, [x22]\n" "add x22, x22, %x[ld_in_row]\n" + ".inst 0xc1a3ac04 // sqdmulh { z4.s-z7.s }, { z4.s-z7.s }, z3.s\n" "ld1b { z18.s }, p1/Z, [x22]\n" "add x22, x22, %x[ld_in_row]\n" "trn1 z14.h, z14.h, z18.h\n" "add z14.h, z14.h, z9.h\n" "ld1b { z15.s }, p1/Z, [x22]\n" "add x22, x22, %x[ld_in_row]\n" + ".inst 0xc1a1aa24 // srshl { z4.s-z7.s }, { z4.s-z7.s }, z1.s\n" "ld1b { z17.s }, p1/Z, [x22]\n" "trn1 z15.h, z15.h, z17.h\n" "add x22, x22, %x[ld_in_row]\n" "add z15.h, z15.h, z9.h\n" ".inst 0xa1402aa2 // ld1h { z2.h, z10.h }, pn10.b/Z, [x21]\n" ".inst 0xc1721568 // sdot za.s[x8, 0], { z11.h-z14.h }, z2.h\n" + ".inst 0xc1a8ab04 // add { z4.s-z7.s }, { z4.s-z7.s }, z8.s\n" "ld1b { z16.s }, p1/Z, [x22]\n" "mov z16.d, z16.d\n" ".inst 0xc17a1588 // sdot za.s[x8, 0], { z12.h-z15.h }, z10.h\n" @@ -467,166 +583,50 @@ void sme2_u8s8u8q_planar_5x5_s2_4rows_dot_za_impl( "add z16.h, z16.h, z9.h\n" "ld1h { z0.h }, p2/Z, [x21, #2, MUL VL]\n" ".inst 0xc17a1589 // sdot za.s[x8, 1], { z12.h-z15.h }, z10.h\n" - ".inst 0xc17015a8 // sdot za.s[x8, 0], { z13.h-z16.h }, z0.h\n" - "ld1h { z0.h }, p2/Z, [x20, #2, MUL VL]\n" - ".inst 0xc17015a9 // sdot za.s[x8, 1], { z13.h-z16.h }, z0.h\n" - "12:" // Unpadded: 0 priming loads - "cmp x7, #0x2\n" - ".inst 0xa1402be2 // ld1h { z2.h, z10.h }, pn10.b/Z, [SP]\n" - "ld1h { z0.h }, p2/Z, [SP, #2, MUL VL]\n" - "blt 22f\n" - "add x21, x17, %x[ld_in_row]\n" - "ld1b { z11.s }, p1/Z, [x17]\n" - "sub x7, x7, #0x2\n" - "ld1b { z21.s }, p1/Z, [x21]\n" - "add x21, x21, %x[ld_in_row]\n" - "trn1 z11.h, z11.h, z21.h\n" - "sub x16, x16, #0x1\n" - "ld1b { z12.s }, p1/Z, [x21]\n" - "add x21, x21, %x[ld_in_row]\n" - "lsr x20, x7, #0x1\n" - "add z11.h, z11.h, z9.h\n" - "ld1b { z20.s }, p1/Z, [x21]\n" - "add x21, x21, %x[ld_in_row]\n" - "trn1 z12.h, z12.h, z20.h\n" - "cmp x20, x16\n" - "ld1b { z13.s }, p1/Z, [x21]\n" - "add x21, x21, %x[ld_in_row]\n" - "csel x26, x20, x16, LT\n" - "add z12.h, z12.h, z9.h\n" - "ld1b { z19.s }, p1/Z, [x21]\n" - "add x21, x21, %x[ld_in_row]\n" - "trn1 z13.h, z13.h, z19.h\n" - "add z13.h, z13.h, z9.h\n" - "ld1b { z14.s }, p1/Z, [x21]\n" - "add x21, x21, %x[ld_in_row]\n" - "add x17, x17, %x[ld_in_col]\n" - "ld1b { z18.s }, p1/Z, [x21]\n" - "add x21, x21, %x[ld_in_row]\n" - "trn1 z14.h, z14.h, z18.h\n" - "add z14.h, z14.h, z9.h\n" - "ld1b { z15.s }, p1/Z, [x21]\n" - "add x21, x21, %x[ld_in_row]\n" - "and x7, x7, #0x1\n" - "ld1b { z17.s }, p1/Z, [x21]\n" - "add x21, x21, %x[ld_in_row]\n" - "trn1 z15.h, z15.h, z17.h\n" - "add z15.h, z15.h, z9.h\n" - "ld1b { z16.s }, p1/Z, [x21]\n" - "mov z16.d, z16.d\n" - "add z16.h, z16.h, z9.h\n" - "sub x16, x16, x26\n" - "cbz x26, 21f\n" - "13:" // Unpadded: Main loop - ".inst 0xc1721568 // sdot za.s[x8, 0], { z11.h-z14.h }, z2.h\n" - "addvl x25, SP, #6\n" - "addvl x24, SP, #12\n" - ".inst 0xc17a1588 // sdot za.s[x8, 0], { z12.h-z15.h }, z10.h\n" - ".inst 0xa1402b22 // ld1h { z2.h, z10.h }, pn10.b/Z, [x25]\n" - "add x23, x17, %x[ld_in_row]\n" - "addvl x22, SP, #3\n" - ".inst 0xc1721569 // sdot za.s[x8, 1], { z11.h-z14.h }, z2.h\n" - "addvl x21, SP, #9\n" - "subs x26, x26, #0x1\n" - ".inst 0xc17a1589 // sdot za.s[x8, 1], { z12.h-z15.h }, z10.h\n" - ".inst 0xa1402b02 // ld1h { z2.h, z10.h }, pn10.b/Z, [x24]\n" - ".inst 0xc172156a // sdot za.s[x8, 2], { z11.h-z14.h }, z2.h\n" - "ld1b { z11.s }, p1/Z, [x17]\n" - "add x17, x17, %x[ld_in_col]\n" - "add x20, x17, %x[ld_in_row]\n" - ".inst 0xc17015a8 // sdot za.s[x8, 0], { z13.h-z16.h }, z0.h\n" - "ld1h { z0.h }, p2/Z, [x25, #2, MUL VL]\n" - ".inst 0xc17a158a // sdot za.s[x8, 2], { z12.h-z15.h }, z10.h\n" - "ld1b { z21.s }, p1/Z, [x23]\n" - "add x23, x23, %x[ld_in_row]\n" - "trn1 z11.h, z11.h, z21.h\n" - ".inst 0xc17015a9 // sdot za.s[x8, 1], { z13.h-z16.h }, z0.h\n" - "ld1h { z0.h }, p2/Z, [x24, #2, MUL VL]\n" - "add z11.h, z11.h, z9.h\n" - "ld1b { z12.s }, p1/Z, [x23]\n" - "add x23, x23, %x[ld_in_row]\n" - ".inst 0xc17015aa // sdot za.s[x8, 2], { z13.h-z16.h }, z0.h\n" - "ld1b { z20.s }, p1/Z, [x23]\n" - "add x23, x23, %x[ld_in_row]\n" - "trn1 z12.h, z12.h, z20.h\n" - "add z12.h, z12.h, z9.h\n" - "ld1b { z13.s }, p1/Z, [x23]\n" - "add x23, x23, %x[ld_in_row]\n" - ".inst 0xc0060c04 // mova { z4.d-z7.d }, za.d[x8, #0]\n" - "add x8, x8, #0x1\n" - "ld1b { z19.s }, p1/Z, [x23]\n" - "add x23, x23, %x[ld_in_row]\n" - "trn1 z13.h, z13.h, z19.h\n" - "add z13.h, z13.h, z9.h\n" - "ld1b { z14.s }, p1/Z, [x23]\n" - "add x23, x23, %x[ld_in_row]\n" - ".inst 0xc1a3ac04 // sqdmulh { z4.s-z7.s }, { z4.s-z7.s }, z3.s\n" - "ld1b { z18.s }, p1/Z, [x23]\n" - "add x23, x23, %x[ld_in_row]\n" - "trn1 z14.h, z14.h, z18.h\n" - "add z14.h, z14.h, z9.h\n" - "ld1b { z15.s }, p1/Z, [x23]\n" - "add x23, x23, %x[ld_in_row]\n" - ".inst 0xc1a1aa24 // srshl { z4.s-z7.s }, { z4.s-z7.s }, z1.s\n" - "ld1b { z17.s }, p1/Z, [x23]\n" - "trn1 z15.h, z15.h, z17.h\n" - "add x23, x23, %x[ld_in_row]\n" - "add z15.h, z15.h, z9.h\n" - ".inst 0xa1402ac2 // ld1h { z2.h, z10.h }, pn10.b/Z, [x22]\n" - ".inst 0xc1721568 // sdot za.s[x8, 0], { z11.h-z14.h }, z2.h\n" - ".inst 0xc1a8ab04 // add { z4.s-z7.s }, { z4.s-z7.s }, z8.s\n" - "ld1b { z16.s }, p1/Z, [x23]\n" - "mov z16.d, z16.d\n" - ".inst 0xc17a1588 // sdot za.s[x8, 0], { z12.h-z15.h }, z10.h\n" - ".inst 0xa1402aa2 // ld1h { z2.h, z10.h }, pn10.b/Z, [x21]\n" - ".inst 0xc1721569 // sdot za.s[x8, 1], { z11.h-z14.h }, z2.h\n" - "add z16.h, z16.h, z9.h\n" - "ld1h { z0.h }, p2/Z, [x22, #2, MUL VL]\n" - ".inst 0xc17a1589 // sdot za.s[x8, 1], { z12.h-z15.h }, z10.h\n" ".inst 0xc1b7cf44 // sclamp { z4.s-z7.s }, z26.s, z23.s\n" ".inst 0xc17015a8 // sdot za.s[x8, 0], { z13.h-z16.h }, z0.h\n" - "ld1h { z0.h }, p2/Z, [x21, #2, MUL VL]\n" - "st1b { z4.s }, p1, [x15]\n" - "add x15, x15, x13\n" - "ld1b { z11.s }, p1/Z, [x17]\n" - ".inst 0xc17015a9 // sdot za.s[x8, 1], { z13.h-z16.h }, z0.h\n" - "st1b { z5.s }, p1, [x14]\n" + "ld1h { z0.h }, p2/Z, [x20, #2, MUL VL]\n" + "st1b { z4.s }, p1, [x14]\n" "add x14, x14, x11\n" - "ld1b { z21.s }, p1/Z, [x20]\n" - "add x20, x20, %x[ld_in_row]\n" + "ld1b { z11.s }, p1/Z, [x16]\n" + ".inst 0xc17015a9 // sdot za.s[x8, 1], { z13.h-z16.h }, z0.h\n" + "st1b { z5.s }, p1, [x13]\n" + "add x13, x13, x10\n" + "ld1b { z21.s }, p1/Z, [x19]\n" + "add x19, x19, %x[ld_in_row]\n" "trn1 z11.h, z11.h, z21.h\n" - "st1b { z6.s }, p1, [x10]\n" - "ld1b { z12.s }, p1/Z, [x20]\n" - "add x20, x20, %x[ld_in_row]\n" - "add x10, x10, x28\n" - "st1b { z7.s }, p1, [x9]\n" - "ld1b { z20.s }, p1/Z, [x20]\n" - "add x20, x20, %x[ld_in_row]\n" - "trn1 z12.h, z12.h, z20.h\n" + "st1b { z6.s }, p1, [x9]\n" + "ld1b { z12.s }, p1/Z, [x19]\n" + "add x19, x19, %x[ld_in_row]\n" "add x9, x9, x27\n" - "ld1b { z13.s }, p1/Z, [x20]\n" - "add x20, x20, %x[ld_in_row]\n" + "st1b { z7.s }, p1, [x28]\n" + "ld1b { z20.s }, p1/Z, [x19]\n" + "add x19, x19, %x[ld_in_row]\n" + "trn1 z12.h, z12.h, z20.h\n" + "add x28, x28, x26\n" + "ld1b { z13.s }, p1/Z, [x19]\n" + "add x19, x19, %x[ld_in_row]\n" ".inst 0xc0040f84 // mova za.d[x8, #4], { z28.d-z31.d }\n" "add z11.h, z11.h, z9.h\n" - "ld1b { z19.s }, p1/Z, [x20]\n" - "add x20, x20, %x[ld_in_row]\n" + "ld1b { z19.s }, p1/Z, [x19]\n" + "add x19, x19, %x[ld_in_row]\n" "trn1 z13.h, z13.h, z19.h\n" "add z12.h, z12.h, z9.h\n" - "ld1b { z14.s }, p1/Z, [x20]\n" - "add x20, x20, %x[ld_in_row]\n" + "ld1b { z14.s }, p1/Z, [x19]\n" + "add x19, x19, %x[ld_in_row]\n" "add z13.h, z13.h, z9.h\n" - "add x17, x17, %x[ld_in_col]\n" - "ld1b { z18.s }, p1/Z, [x20]\n" - "add x20, x20, %x[ld_in_row]\n" + "add x16, x16, %x[ld_in_col]\n" + "ld1b { z18.s }, p1/Z, [x19]\n" + "add x19, x19, %x[ld_in_row]\n" "trn1 z14.h, z14.h, z18.h\n" "add z14.h, z14.h, z9.h\n" - "ld1b { z15.s }, p1/Z, [x20]\n" - "add x20, x20, %x[ld_in_row]\n" - "ld1b { z17.s }, p1/Z, [x20]\n" - "add x20, x20, %x[ld_in_row]\n" + "ld1b { z15.s }, p1/Z, [x19]\n" + "add x19, x19, %x[ld_in_row]\n" + "ld1b { z17.s }, p1/Z, [x19]\n" + "add x19, x19, %x[ld_in_row]\n" "trn1 z15.h, z15.h, z17.h\n" "add z15.h, z15.h, z9.h\n" - "ld1b { z16.s }, p1/Z, [x20]\n" + "ld1b { z16.s }, p1/Z, [x19]\n" "mov z16.d, z16.d\n" "add z16.h, z16.h, z9.h\n" ".inst 0xa1402be2 // ld1h { z2.h, z10.h }, pn10.b/Z, [SP]\n" @@ -634,717 +634,717 @@ void sme2_u8s8u8q_planar_5x5_s2_4rows_dot_za_impl( "bgt 13b\n" "b 21f\n" "14:" // Padded - "cbz x22, 19f\n" - "cmp x22, #0x1\n" - "sub x7, x7, x22\n" + "cbz x21, 19f\n" + "cmp x21, #0x1\n" + "sub x17, x17, x21\n" "beq 18f\n" - "cmp x22, #0x2\n" + "cmp x21, #0x2\n" "beq 17f\n" - "cmp x22, #0x3\n" + "cmp x21, #0x3\n" "beq 16f\n" "15:" // Padded: 4 priming loads "mov x12, #0x0\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1b { z11.s }, p0/Z, [x17]\n" + "ld1b { z11.s }, p0/Z, [x16]\n" "add z11.h, p0/M, z11.h, z9.h\n" - "add x21, x17, %x[ld_in_row]\n" + "add x20, x16, %x[ld_in_row]\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1b { z21.s }, p0/Z, [x21]\n" + "ld1b { z21.s }, p0/Z, [x20]\n" "add z21.h, p0/M, z21.h, z9.h\n" - "add x21, x21, %x[ld_in_row]\n" + "add x20, x20, %x[ld_in_row]\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "ld1b { z12.s }, p0/Z, [x21]\n" + "ld1b { z12.s }, p0/Z, [x20]\n" "add z12.h, p0/M, z12.h, z9.h\n" - "add x21, x21, %x[ld_in_row]\n" + "add x20, x20, %x[ld_in_row]\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" - "ld1b { z20.s }, p0/Z, [x21]\n" + "ld1b { z20.s }, p0/Z, [x20]\n" "add z20.h, p0/M, z20.h, z9.h\n" "mov x12, #0x4\n" - "add x21, x21, %x[ld_in_row]\n" + "add x20, x20, %x[ld_in_row]\n" "trn1 z11.h, z11.h, z21.h\n" "trn1 z12.h, z12.h, z20.h\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1b { z13.s }, p0/Z, [x21]\n" + "ld1b { z13.s }, p0/Z, [x20]\n" "add z13.h, p0/M, z13.h, z9.h\n" - "add x21, x21, %x[ld_in_row]\n" + "add x20, x20, %x[ld_in_row]\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1b { z19.s }, p0/Z, [x21]\n" + "ld1b { z19.s }, p0/Z, [x20]\n" "add z19.h, p0/M, z19.h, z9.h\n" - "add x21, x21, %x[ld_in_row]\n" + "add x20, x20, %x[ld_in_row]\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "ld1b { z14.s }, p0/Z, [x21]\n" + "ld1b { z14.s }, p0/Z, [x20]\n" "add z14.h, p0/M, z14.h, z9.h\n" - "add x21, x21, %x[ld_in_row]\n" + "add x20, x20, %x[ld_in_row]\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" - "ld1b { z18.s }, p0/Z, [x21]\n" + "ld1b { z18.s }, p0/Z, [x20]\n" "mov x12, #0x8\n" "add z18.h, p0/M, z18.h, z9.h\n" - "add x21, x21, %x[ld_in_row]\n" + "add x20, x20, %x[ld_in_row]\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1b { z15.s }, p0/Z, [x21]\n" + "ld1b { z15.s }, p0/Z, [x20]\n" "add z15.h, p0/M, z15.h, z9.h\n" - "add x21, x21, %x[ld_in_row]\n" + "add x20, x20, %x[ld_in_row]\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1b { z17.s }, p0/Z, [x21]\n" + "ld1b { z17.s }, p0/Z, [x20]\n" "add z17.h, p0/M, z17.h, z9.h\n" - "addvl x20, SP, #12\n" - "add x21, x21, %x[ld_in_row]\n" + "addvl x19, SP, #12\n" + "add x20, x20, %x[ld_in_row]\n" "trn1 z13.h, z13.h, z19.h\n" "trn1 z14.h, z14.h, z18.h\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - ".inst 0xa1402a82 // ld1h { z2.h, z10.h }, pn10.b/Z, [x20]\n" + ".inst 0xa1402a62 // ld1h { z2.h, z10.h }, pn10.b/Z, [x19]\n" "trn1 z15.h, z15.h, z17.h\n" ".inst 0xc1721568 // sdot za.s[x8, 0], { z11.h-z14.h }, z2.h\n" - "ld1b { z16.s }, p0/Z, [x21]\n" + "ld1b { z16.s }, p0/Z, [x20]\n" "add z16.h, p0/M, z16.h, z9.h\n" "mov z16.d, z16.d\n" - "add x17, x17, %x[ld_in_col]\n" + "add x16, x16, %x[ld_in_col]\n" ".inst 0xc17a1588 // sdot za.s[x8, 0], { z12.h-z15.h }, z10.h\n" - "ld1h { z0.h }, p2/Z, [x20, #2, MUL VL]\n" + "ld1h { z0.h }, p2/Z, [x19, #2, MUL VL]\n" ".inst 0xc17015a8 // sdot za.s[x8, 0], { z13.h-z16.h }, z0.h\n" "16:" // Padded: 3 priming loads "mov x12, #0x0\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1b { z11.s }, p0/Z, [x17]\n" + "ld1b { z11.s }, p0/Z, [x16]\n" "add z11.h, p0/M, z11.h, z9.h\n" - "add x21, x17, %x[ld_in_row]\n" + "add x20, x16, %x[ld_in_row]\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1b { z21.s }, p0/Z, [x21]\n" + "ld1b { z21.s }, p0/Z, [x20]\n" "add z21.h, p0/M, z21.h, z9.h\n" - "add x21, x21, %x[ld_in_row]\n" + "add x20, x20, %x[ld_in_row]\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "ld1b { z12.s }, p0/Z, [x21]\n" + "ld1b { z12.s }, p0/Z, [x20]\n" "add z12.h, p0/M, z12.h, z9.h\n" - "add x21, x21, %x[ld_in_row]\n" + "add x20, x20, %x[ld_in_row]\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" - "ld1b { z20.s }, p0/Z, [x21]\n" + "ld1b { z20.s }, p0/Z, [x20]\n" "add z20.h, p0/M, z20.h, z9.h\n" "mov x12, #0x4\n" - "add x21, x21, %x[ld_in_row]\n" + "add x20, x20, %x[ld_in_row]\n" "trn1 z11.h, z11.h, z21.h\n" "trn1 z12.h, z12.h, z20.h\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1b { z13.s }, p0/Z, [x21]\n" + "ld1b { z13.s }, p0/Z, [x20]\n" "add z13.h, p0/M, z13.h, z9.h\n" - "add x21, x21, %x[ld_in_row]\n" + "add x20, x20, %x[ld_in_row]\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1b { z19.s }, p0/Z, [x21]\n" + "ld1b { z19.s }, p0/Z, [x20]\n" "add z19.h, p0/M, z19.h, z9.h\n" - "add x21, x21, %x[ld_in_row]\n" + "add x20, x20, %x[ld_in_row]\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "ld1b { z14.s }, p0/Z, [x21]\n" + "ld1b { z14.s }, p0/Z, [x20]\n" "add z14.h, p0/M, z14.h, z9.h\n" - "add x21, x21, %x[ld_in_row]\n" + "add x20, x20, %x[ld_in_row]\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" - "ld1b { z18.s }, p0/Z, [x21]\n" + "ld1b { z18.s }, p0/Z, [x20]\n" "mov x12, #0x8\n" "add z18.h, p0/M, z18.h, z9.h\n" - "add x21, x21, %x[ld_in_row]\n" + "add x20, x20, %x[ld_in_row]\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1b { z15.s }, p0/Z, [x21]\n" + "ld1b { z15.s }, p0/Z, [x20]\n" "add z15.h, p0/M, z15.h, z9.h\n" - "add x21, x21, %x[ld_in_row]\n" + "add x20, x20, %x[ld_in_row]\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1b { z17.s }, p0/Z, [x21]\n" + "ld1b { z17.s }, p0/Z, [x20]\n" "add z17.h, p0/M, z17.h, z9.h\n" - "addvl x20, SP, #9\n" - "add x21, x21, %x[ld_in_row]\n" + "addvl x19, SP, #9\n" + "add x20, x20, %x[ld_in_row]\n" "trn1 z13.h, z13.h, z19.h\n" "trn1 z14.h, z14.h, z18.h\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - ".inst 0xa1402a82 // ld1h { z2.h, z10.h }, pn10.b/Z, [x20]\n" + ".inst 0xa1402a62 // ld1h { z2.h, z10.h }, pn10.b/Z, [x19]\n" "trn1 z15.h, z15.h, z17.h\n" ".inst 0xc1721568 // sdot za.s[x8, 0], { z11.h-z14.h }, z2.h\n" - "ld1b { z16.s }, p0/Z, [x21]\n" + "ld1b { z16.s }, p0/Z, [x20]\n" "add z16.h, p0/M, z16.h, z9.h\n" "mov z16.d, z16.d\n" - "add x17, x17, %x[ld_in_col]\n" + "add x16, x16, %x[ld_in_col]\n" ".inst 0xc17a1588 // sdot za.s[x8, 0], { z12.h-z15.h }, z10.h\n" - "ld1h { z0.h }, p2/Z, [x20, #2, MUL VL]\n" + "ld1h { z0.h }, p2/Z, [x19, #2, MUL VL]\n" ".inst 0xc17015a8 // sdot za.s[x8, 0], { z13.h-z16.h }, z0.h\n" "17:" // Padded: 2 priming loads "mov x12, #0x0\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1b { z11.s }, p0/Z, [x17]\n" + "ld1b { z11.s }, p0/Z, [x16]\n" "add z11.h, p0/M, z11.h, z9.h\n" - "add x20, x17, %x[ld_in_row]\n" + "add x19, x16, %x[ld_in_row]\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1b { z21.s }, p0/Z, [x20]\n" + "ld1b { z21.s }, p0/Z, [x19]\n" "add z21.h, p0/M, z21.h, z9.h\n" - "add x20, x20, %x[ld_in_row]\n" + "add x19, x19, %x[ld_in_row]\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "ld1b { z12.s }, p0/Z, [x20]\n" + "ld1b { z12.s }, p0/Z, [x19]\n" "add z12.h, p0/M, z12.h, z9.h\n" - "add x20, x20, %x[ld_in_row]\n" + "add x19, x19, %x[ld_in_row]\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" - "ld1b { z20.s }, p0/Z, [x20]\n" + "ld1b { z20.s }, p0/Z, [x19]\n" "add z20.h, p0/M, z20.h, z9.h\n" "mov x12, #0x4\n" - "add x20, x20, %x[ld_in_row]\n" + "add x19, x19, %x[ld_in_row]\n" "trn1 z11.h, z11.h, z21.h\n" "trn1 z12.h, z12.h, z20.h\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1b { z13.s }, p0/Z, [x20]\n" + "ld1b { z13.s }, p0/Z, [x19]\n" "add z13.h, p0/M, z13.h, z9.h\n" - "add x20, x20, %x[ld_in_row]\n" + "add x19, x19, %x[ld_in_row]\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1b { z19.s }, p0/Z, [x20]\n" + "ld1b { z19.s }, p0/Z, [x19]\n" "add z19.h, p0/M, z19.h, z9.h\n" - "add x20, x20, %x[ld_in_row]\n" + "add x19, x19, %x[ld_in_row]\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "ld1b { z14.s }, p0/Z, [x20]\n" + "ld1b { z14.s }, p0/Z, [x19]\n" "add z14.h, p0/M, z14.h, z9.h\n" - "add x20, x20, %x[ld_in_row]\n" + "add x19, x19, %x[ld_in_row]\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" - "ld1b { z18.s }, p0/Z, [x20]\n" + "ld1b { z18.s }, p0/Z, [x19]\n" "mov x12, #0x8\n" "add z18.h, p0/M, z18.h, z9.h\n" - "add x20, x20, %x[ld_in_row]\n" + "add x19, x19, %x[ld_in_row]\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1b { z15.s }, p0/Z, [x20]\n" + "ld1b { z15.s }, p0/Z, [x19]\n" "add z15.h, p0/M, z15.h, z9.h\n" - "add x20, x20, %x[ld_in_row]\n" + "add x19, x19, %x[ld_in_row]\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1b { z17.s }, p0/Z, [x20]\n" + "ld1b { z17.s }, p0/Z, [x19]\n" "add z17.h, p0/M, z17.h, z9.h\n" - "addvl x21, SP, #6\n" + "addvl x20, SP, #6\n" "trn1 z13.h, z13.h, z19.h\n" "trn1 z14.h, z14.h, z18.h\n" - ".inst 0xa1402aa2 // ld1h { z2.h, z10.h }, pn10.b/Z, [x21]\n" - "add x20, x20, %x[ld_in_row]\n" + ".inst 0xa1402a82 // ld1h { z2.h, z10.h }, pn10.b/Z, [x20]\n" + "add x19, x19, %x[ld_in_row]\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" "trn1 z15.h, z15.h, z17.h\n" ".inst 0xc1721568 // sdot za.s[x8, 0], { z11.h-z14.h }, z2.h\n" - "ld1b { z16.s }, p0/Z, [x20]\n" - "addvl x20, SP, #12\n" + "ld1b { z16.s }, p0/Z, [x19]\n" + "addvl x19, SP, #12\n" "add z16.h, p0/M, z16.h, z9.h\n" ".inst 0xc17a1588 // sdot za.s[x8, 0], { z12.h-z15.h }, z10.h\n" - ".inst 0xa1402a82 // ld1h { z2.h, z10.h }, pn10.b/Z, [x20]\n" + ".inst 0xa1402a62 // ld1h { z2.h, z10.h }, pn10.b/Z, [x19]\n" ".inst 0xc1721569 // sdot za.s[x8, 1], { z11.h-z14.h }, z2.h\n" "mov z16.d, z16.d\n" - "add x17, x17, %x[ld_in_col]\n" - "ld1h { z0.h }, p2/Z, [x21, #2, MUL VL]\n" + "add x16, x16, %x[ld_in_col]\n" + "ld1h { z0.h }, p2/Z, [x20, #2, MUL VL]\n" ".inst 0xc17a1589 // sdot za.s[x8, 1], { z12.h-z15.h }, z10.h\n" ".inst 0xc17015a8 // sdot za.s[x8, 0], { z13.h-z16.h }, z0.h\n" - "ld1h { z0.h }, p2/Z, [x20, #2, MUL VL]\n" + "ld1h { z0.h }, p2/Z, [x19, #2, MUL VL]\n" ".inst 0xc17015a9 // sdot za.s[x8, 1], { z13.h-z16.h }, z0.h\n" "18:" // Padded: 1 priming loads "mov x12, #0x0\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1b { z11.s }, p0/Z, [x17]\n" + "ld1b { z11.s }, p0/Z, [x16]\n" "add z11.h, p0/M, z11.h, z9.h\n" - "add x20, x17, %x[ld_in_row]\n" + "add x19, x16, %x[ld_in_row]\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1b { z21.s }, p0/Z, [x20]\n" + "ld1b { z21.s }, p0/Z, [x19]\n" "add z21.h, p0/M, z21.h, z9.h\n" - "add x20, x20, %x[ld_in_row]\n" + "add x19, x19, %x[ld_in_row]\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "ld1b { z12.s }, p0/Z, [x20]\n" + "ld1b { z12.s }, p0/Z, [x19]\n" "add z12.h, p0/M, z12.h, z9.h\n" - "add x20, x20, %x[ld_in_row]\n" + "add x19, x19, %x[ld_in_row]\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" - "ld1b { z20.s }, p0/Z, [x20]\n" + "ld1b { z20.s }, p0/Z, [x19]\n" "add z20.h, p0/M, z20.h, z9.h\n" "mov x12, #0x4\n" - "add x20, x20, %x[ld_in_row]\n" + "add x19, x19, %x[ld_in_row]\n" "trn1 z11.h, z11.h, z21.h\n" "trn1 z12.h, z12.h, z20.h\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1b { z13.s }, p0/Z, [x20]\n" + "ld1b { z13.s }, p0/Z, [x19]\n" "add z13.h, p0/M, z13.h, z9.h\n" - "add x20, x20, %x[ld_in_row]\n" + "add x19, x19, %x[ld_in_row]\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1b { z19.s }, p0/Z, [x20]\n" + "ld1b { z19.s }, p0/Z, [x19]\n" "add z19.h, p0/M, z19.h, z9.h\n" - "add x20, x20, %x[ld_in_row]\n" + "add x19, x19, %x[ld_in_row]\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "ld1b { z14.s }, p0/Z, [x20]\n" + "ld1b { z14.s }, p0/Z, [x19]\n" "add z14.h, p0/M, z14.h, z9.h\n" - "add x20, x20, %x[ld_in_row]\n" + "add x19, x19, %x[ld_in_row]\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" - "ld1b { z18.s }, p0/Z, [x20]\n" + "ld1b { z18.s }, p0/Z, [x19]\n" "mov x12, #0x8\n" "add z18.h, p0/M, z18.h, z9.h\n" - "add x20, x20, %x[ld_in_row]\n" + "add x19, x19, %x[ld_in_row]\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1b { z15.s }, p0/Z, [x20]\n" + "ld1b { z15.s }, p0/Z, [x19]\n" "add z15.h, p0/M, z15.h, z9.h\n" - "add x20, x20, %x[ld_in_row]\n" + "add x19, x19, %x[ld_in_row]\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1b { z17.s }, p0/Z, [x20]\n" + "ld1b { z17.s }, p0/Z, [x19]\n" "add z17.h, p0/M, z17.h, z9.h\n" - "addvl x21, SP, #3\n" + "addvl x20, SP, #3\n" "trn1 z13.h, z13.h, z19.h\n" "trn1 z14.h, z14.h, z18.h\n" - ".inst 0xa1402aa2 // ld1h { z2.h, z10.h }, pn10.b/Z, [x21]\n" - "add x20, x20, %x[ld_in_row]\n" + ".inst 0xa1402a82 // ld1h { z2.h, z10.h }, pn10.b/Z, [x20]\n" + "add x19, x19, %x[ld_in_row]\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" "trn1 z15.h, z15.h, z17.h\n" ".inst 0xc1721568 // sdot za.s[x8, 0], { z11.h-z14.h }, z2.h\n" - "ld1b { z16.s }, p0/Z, [x20]\n" - "addvl x20, SP, #9\n" + "ld1b { z16.s }, p0/Z, [x19]\n" + "addvl x19, SP, #12\n" "add z16.h, p0/M, z16.h, z9.h\n" ".inst 0xc17a1588 // sdot za.s[x8, 0], { z12.h-z15.h }, z10.h\n" - ".inst 0xa1402a82 // ld1h { z2.h, z10.h }, pn10.b/Z, [x20]\n" + ".inst 0xa1402a62 // ld1h { z2.h, z10.h }, pn10.b/Z, [x19]\n" ".inst 0xc1721569 // sdot za.s[x8, 1], { z11.h-z14.h }, z2.h\n" "mov z16.d, z16.d\n" - "add x17, x17, %x[ld_in_col]\n" - "ld1h { z0.h }, p2/Z, [x21, #2, MUL VL]\n" + "add x16, x16, %x[ld_in_col]\n" + "ld1h { z0.h }, p2/Z, [x20, #2, MUL VL]\n" ".inst 0xc17a1589 // sdot za.s[x8, 1], { z12.h-z15.h }, z10.h\n" ".inst 0xc17015a8 // sdot za.s[x8, 0], { z13.h-z16.h }, z0.h\n" - "ld1h { z0.h }, p2/Z, [x20, #2, MUL VL]\n" + "ld1h { z0.h }, p2/Z, [x19, #2, MUL VL]\n" ".inst 0xc17015a9 // sdot za.s[x8, 1], { z13.h-z16.h }, z0.h\n" "19:" // Padded: 0 priming loads - "cmp x7, #0x2\n" + "cmp x17, #0x2\n" ".inst 0xa1402be2 // ld1h { z2.h, z10.h }, pn10.b/Z, [SP]\n" "ld1h { z0.h }, p2/Z, [SP, #2, MUL VL]\n" "blt 22f\n" "mov x12, #0x0\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1b { z11.s }, p0/Z, [x17]\n" + "ld1b { z11.s }, p0/Z, [x16]\n" "add z11.h, p0/M, z11.h, z9.h\n" - "add x20, x17, %x[ld_in_row]\n" + "add x19, x16, %x[ld_in_row]\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1b { z21.s }, p0/Z, [x20]\n" + "ld1b { z21.s }, p0/Z, [x19]\n" "add z21.h, p0/M, z21.h, z9.h\n" - "add x20, x20, %x[ld_in_row]\n" + "add x19, x19, %x[ld_in_row]\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "ld1b { z12.s }, p0/Z, [x20]\n" + "ld1b { z12.s }, p0/Z, [x19]\n" "add z12.h, p0/M, z12.h, z9.h\n" - "add x20, x20, %x[ld_in_row]\n" + "add x19, x19, %x[ld_in_row]\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" - "ld1b { z20.s }, p0/Z, [x20]\n" + "ld1b { z20.s }, p0/Z, [x19]\n" "add z20.h, p0/M, z20.h, z9.h\n" "mov x12, #0x4\n" - "add x20, x20, %x[ld_in_row]\n" + "add x19, x19, %x[ld_in_row]\n" "trn1 z11.h, z11.h, z21.h\n" "trn1 z12.h, z12.h, z20.h\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1b { z13.s }, p0/Z, [x20]\n" + "ld1b { z13.s }, p0/Z, [x19]\n" "add z13.h, p0/M, z13.h, z9.h\n" - "add x20, x20, %x[ld_in_row]\n" + "add x19, x19, %x[ld_in_row]\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1b { z19.s }, p0/Z, [x20]\n" + "ld1b { z19.s }, p0/Z, [x19]\n" "add z19.h, p0/M, z19.h, z9.h\n" - "add x20, x20, %x[ld_in_row]\n" + "add x19, x19, %x[ld_in_row]\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "ld1b { z14.s }, p0/Z, [x20]\n" + "ld1b { z14.s }, p0/Z, [x19]\n" "add z14.h, p0/M, z14.h, z9.h\n" - "add x20, x20, %x[ld_in_row]\n" + "add x19, x19, %x[ld_in_row]\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" - "ld1b { z18.s }, p0/Z, [x20]\n" + "ld1b { z18.s }, p0/Z, [x19]\n" "mov x12, #0x8\n" "add z18.h, p0/M, z18.h, z9.h\n" - "add x20, x20, %x[ld_in_row]\n" + "add x19, x19, %x[ld_in_row]\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1b { z15.s }, p0/Z, [x20]\n" + "ld1b { z15.s }, p0/Z, [x19]\n" "add z15.h, p0/M, z15.h, z9.h\n" - "add x20, x20, %x[ld_in_row]\n" + "add x19, x19, %x[ld_in_row]\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1b { z17.s }, p0/Z, [x20]\n" + "ld1b { z17.s }, p0/Z, [x19]\n" "add z17.h, p0/M, z17.h, z9.h\n" - "add x20, x20, %x[ld_in_row]\n" + "add x19, x19, %x[ld_in_row]\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "ld1b { z16.s }, p0/Z, [x20]\n" + "ld1b { z16.s }, p0/Z, [x19]\n" "add z16.h, p0/M, z16.h, z9.h\n" - "sub x7, x7, #0x2\n" - "sub x16, x16, #0x1\n" + "sub x17, x17, #0x2\n" + "sub x15, x15, #0x1\n" "trn1 z13.h, z13.h, z19.h\n" "trn1 z14.h, z14.h, z18.h\n" - "lsr x20, x7, #0x1\n" - "cmp x20, x16\n" + "lsr x19, x17, #0x1\n" + "cmp x19, x15\n" "trn1 z15.h, z15.h, z17.h\n" "mov z16.d, z16.d\n" - "csel x25, x20, x16, LT\n" - "add x17, x17, %x[ld_in_col]\n" - "and x7, x7, #0x1\n" - "sub x16, x16, x25\n" - "cbz x25, 21f\n" + "csel x24, x19, x15, LT\n" + "add x16, x16, %x[ld_in_col]\n" + "and x17, x17, #0x1\n" + "sub x15, x15, x24\n" + "cbz x24, 21f\n" "20:" // Padded: Main loop ".inst 0xc1721568 // sdot za.s[x8, 0], { z11.h-z14.h }, z2.h\n" - "addvl x24, SP, #6\n" - "addvl x23, SP, #12\n" + "addvl x23, SP, #6\n" + "addvl x22, SP, #12\n" ".inst 0xc17a1588 // sdot za.s[x8, 0], { z12.h-z15.h }, z10.h\n" - ".inst 0xa1402b02 // ld1h { z2.h, z10.h }, pn10.b/Z, [x24]\n" + ".inst 0xa1402ae2 // ld1h { z2.h, z10.h }, pn10.b/Z, [x23]\n" "mov x12, #0x0\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" ".inst 0xc1721569 // sdot za.s[x8, 1], { z11.h-z14.h }, z2.h\n" - "add x20, x17, %x[ld_in_row]\n" - "addvl x22, SP, #3\n" + "add x19, x16, %x[ld_in_row]\n" + "addvl x21, SP, #3\n" ".inst 0xc17a1589 // sdot za.s[x8, 1], { z12.h-z15.h }, z10.h\n" - ".inst 0xa1402ae2 // ld1h { z2.h, z10.h }, pn10.b/Z, [x23]\n" - "addvl x21, SP, #9\n" - "subs x25, x25, #0x1\n" + ".inst 0xa1402ac2 // ld1h { z2.h, z10.h }, pn10.b/Z, [x22]\n" + "addvl x20, SP, #9\n" + "subs x24, x24, #0x1\n" ".inst 0xc172156a // sdot za.s[x8, 2], { z11.h-z14.h }, z2.h\n" - "ld1b { z11.s }, p0/Z, [x17]\n" + "ld1b { z11.s }, p0/Z, [x16]\n" "add z11.h, p0/M, z11.h, z9.h\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1b { z21.s }, p0/Z, [x20]\n" + "ld1b { z21.s }, p0/Z, [x19]\n" "add z21.h, p0/M, z21.h, z9.h\n" - "add x20, x20, %x[ld_in_row]\n" + "add x19, x19, %x[ld_in_row]\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" ".inst 0xc17a158a // sdot za.s[x8, 2], { z12.h-z15.h }, z10.h\n" - "ld1b { z12.s }, p0/Z, [x20]\n" + "ld1b { z12.s }, p0/Z, [x19]\n" "add z12.h, p0/M, z12.h, z9.h\n" - "add x20, x20, %x[ld_in_row]\n" + "add x19, x19, %x[ld_in_row]\n" ".inst 0xc17015a8 // sdot za.s[x8, 0], { z13.h-z16.h }, z0.h\n" - "ld1h { z0.h }, p2/Z, [x24, #2, MUL VL]\n" + "ld1h { z0.h }, p2/Z, [x23, #2, MUL VL]\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" "mov x12, #0x4\n" - "ld1b { z20.s }, p0/Z, [x20]\n" + "ld1b { z20.s }, p0/Z, [x19]\n" ".inst 0xc17015a9 // sdot za.s[x8, 1], { z13.h-z16.h }, z0.h\n" "add z20.h, p0/M, z20.h, z9.h\n" - "add x20, x20, %x[ld_in_row]\n" - "ld1h { z0.h }, p2/Z, [x23, #2, MUL VL]\n" + "add x19, x19, %x[ld_in_row]\n" + "ld1h { z0.h }, p2/Z, [x22, #2, MUL VL]\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" ".inst 0xc17015aa // sdot za.s[x8, 2], { z13.h-z16.h }, z0.h\n" "trn1 z11.h, z11.h, z21.h\n" - "ld1b { z13.s }, p0/Z, [x20]\n" + "ld1b { z13.s }, p0/Z, [x19]\n" "add z13.h, p0/M, z13.h, z9.h\n" - "add x20, x20, %x[ld_in_row]\n" + "add x19, x19, %x[ld_in_row]\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1b { z19.s }, p0/Z, [x20]\n" + "ld1b { z19.s }, p0/Z, [x19]\n" "add z19.h, p0/M, z19.h, z9.h\n" - "add x20, x20, %x[ld_in_row]\n" + "add x19, x19, %x[ld_in_row]\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "ld1b { z14.s }, p0/Z, [x20]\n" + "ld1b { z14.s }, p0/Z, [x19]\n" "add z14.h, p0/M, z14.h, z9.h\n" - "add x20, x20, %x[ld_in_row]\n" + "add x19, x19, %x[ld_in_row]\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" - "ld1b { z18.s }, p0/Z, [x20]\n" + "ld1b { z18.s }, p0/Z, [x19]\n" "mov x12, #0x8\n" "add z18.h, p0/M, z18.h, z9.h\n" - "add x20, x20, %x[ld_in_row]\n" + "add x19, x19, %x[ld_in_row]\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1b { z15.s }, p0/Z, [x20]\n" + "ld1b { z15.s }, p0/Z, [x19]\n" "add z15.h, p0/M, z15.h, z9.h\n" - "add x20, x20, %x[ld_in_row]\n" + "add x19, x19, %x[ld_in_row]\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1b { z17.s }, p0/Z, [x20]\n" + "ld1b { z17.s }, p0/Z, [x19]\n" "add z17.h, p0/M, z17.h, z9.h\n" "trn1 z12.h, z12.h, z20.h\n" "trn1 z13.h, z13.h, z19.h\n" "trn1 z14.h, z14.h, z18.h\n" - ".inst 0xa1402ac2 // ld1h { z2.h, z10.h }, pn10.b/Z, [x22]\n" - "add x20, x20, %x[ld_in_row]\n" + ".inst 0xa1402aa2 // ld1h { z2.h, z10.h }, pn10.b/Z, [x21]\n" + "add x19, x19, %x[ld_in_row]\n" ".inst 0xc0060c04 // mova { z4.d-z7.d }, za.d[x8, #0]\n" "add x8, x8, #0x1\n" "trn1 z15.h, z15.h, z17.h\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" ".inst 0xc1721568 // sdot za.s[x8, 0], { z11.h-z14.h }, z2.h\n" - "ld1b { z16.s }, p0/Z, [x20]\n" + "ld1b { z16.s }, p0/Z, [x19]\n" "mov x12, #0x0\n" "add z16.h, p0/M, z16.h, z9.h\n" ".inst 0xc17a1588 // sdot za.s[x8, 0], { z12.h-z15.h }, z10.h\n" - ".inst 0xa1402aa2 // ld1h { z2.h, z10.h }, pn10.b/Z, [x21]\n" - "add x17, x17, %x[ld_in_col]\n" + ".inst 0xa1402a82 // ld1h { z2.h, z10.h }, pn10.b/Z, [x20]\n" + "add x16, x16, %x[ld_in_col]\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" ".inst 0xc1721569 // sdot za.s[x8, 1], { z11.h-z14.h }, z2.h\n" - "ld1b { z11.s }, p0/Z, [x17]\n" + "ld1b { z11.s }, p0/Z, [x16]\n" "add z11.h, p0/M, z11.h, z9.h\n" - "add x20, x17, %x[ld_in_row]\n" + "add x19, x16, %x[ld_in_row]\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1b { z21.s }, p0/Z, [x20]\n" + "ld1b { z21.s }, p0/Z, [x19]\n" "add z21.h, p0/M, z21.h, z9.h\n" - "add x20, x20, %x[ld_in_row]\n" + "add x19, x19, %x[ld_in_row]\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" ".inst 0xc17a1589 // sdot za.s[x8, 1], { z12.h-z15.h }, z10.h\n" - "ld1b { z12.s }, p0/Z, [x20]\n" + "ld1b { z12.s }, p0/Z, [x19]\n" "mov z16.d, z16.d\n" - "ld1h { z0.h }, p2/Z, [x22, #2, MUL VL]\n" + "ld1h { z0.h }, p2/Z, [x21, #2, MUL VL]\n" "add z12.h, p0/M, z12.h, z9.h\n" - "add x20, x20, %x[ld_in_row]\n" + "add x19, x19, %x[ld_in_row]\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" - "ld1b { z20.s }, p0/Z, [x20]\n" + "ld1b { z20.s }, p0/Z, [x19]\n" ".inst 0xc17015a8 // sdot za.s[x8, 0], { z13.h-z16.h }, z0.h\n" "mov x12, #0x4\n" "add z20.h, p0/M, z20.h, z9.h\n" - "ld1h { z0.h }, p2/Z, [x21, #2, MUL VL]\n" - "add x20, x20, %x[ld_in_row]\n" + "ld1h { z0.h }, p2/Z, [x20, #2, MUL VL]\n" + "add x19, x19, %x[ld_in_row]\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" ".inst 0xc17015a9 // sdot za.s[x8, 1], { z13.h-z16.h }, z0.h\n" - "ld1b { z13.s }, p0/Z, [x20]\n" + "ld1b { z13.s }, p0/Z, [x19]\n" "add z13.h, p0/M, z13.h, z9.h\n" - "add x20, x20, %x[ld_in_row]\n" + "add x19, x19, %x[ld_in_row]\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1b { z19.s }, p0/Z, [x20]\n" + "ld1b { z19.s }, p0/Z, [x19]\n" "add z19.h, p0/M, z19.h, z9.h\n" - "add x20, x20, %x[ld_in_row]\n" + "add x19, x19, %x[ld_in_row]\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "ld1b { z14.s }, p0/Z, [x20]\n" + "ld1b { z14.s }, p0/Z, [x19]\n" "add z14.h, p0/M, z14.h, z9.h\n" - "add x20, x20, %x[ld_in_row]\n" + "add x19, x19, %x[ld_in_row]\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" - "ld1b { z18.s }, p0/Z, [x20]\n" + "ld1b { z18.s }, p0/Z, [x19]\n" "mov x12, #0x8\n" "add z18.h, p0/M, z18.h, z9.h\n" - "add x20, x20, %x[ld_in_row]\n" + "add x19, x19, %x[ld_in_row]\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1b { z15.s }, p0/Z, [x20]\n" + "ld1b { z15.s }, p0/Z, [x19]\n" ".inst 0xc1a3ac04 // sqdmulh { z4.s-z7.s }, { z4.s-z7.s }, z3.s\n" "add z15.h, p0/M, z15.h, z9.h\n" - "add x20, x20, %x[ld_in_row]\n" + "add x19, x19, %x[ld_in_row]\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1b { z17.s }, p0/Z, [x20]\n" + "ld1b { z17.s }, p0/Z, [x19]\n" ".inst 0xc1a1aa24 // srshl { z4.s-z7.s }, { z4.s-z7.s }, z1.s\n" - "add x20, x20, %x[ld_in_row]\n" + "add x19, x19, %x[ld_in_row]\n" ".inst 0xc0040f84 // mova za.d[x8, #4], { z28.d-z31.d }\n" "add z17.h, p0/M, z17.h, z9.h\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "ld1b { z16.s }, p0/Z, [x20]\n" + "ld1b { z16.s }, p0/Z, [x19]\n" "add z16.h, p0/M, z16.h, z9.h\n" ".inst 0xc1a8ab04 // add { z4.s-z7.s }, { z4.s-z7.s }, z8.s\n" ".inst 0xa1402be2 // ld1h { z2.h, z10.h }, pn10.b/Z, [SP]\n" - "add x17, x17, %x[ld_in_col]\n" + "add x16, x16, %x[ld_in_col]\n" ".inst 0xc1b7cf44 // sclamp { z4.s-z7.s }, z26.s, z23.s\n" - "st1b { z4.s }, p1, [x15]\n" - "add x15, x15, x13\n" - "ld1h { z0.h }, p2/Z, [SP, #2, MUL VL]\n" - "st1b { z5.s }, p1, [x14]\n" + "st1b { z4.s }, p1, [x14]\n" "add x14, x14, x11\n" + "ld1h { z0.h }, p2/Z, [SP, #2, MUL VL]\n" + "st1b { z5.s }, p1, [x13]\n" + "add x13, x13, x10\n" "trn1 z11.h, z11.h, z21.h\n" "trn1 z12.h, z12.h, z20.h\n" - "st1b { z6.s }, p1, [x10]\n" - "add x10, x10, x28\n" + "st1b { z6.s }, p1, [x9]\n" + "add x9, x9, x27\n" "trn1 z13.h, z13.h, z19.h\n" "trn1 z14.h, z14.h, z18.h\n" - "st1b { z7.s }, p1, [x9]\n" - "add x9, x9, x27\n" + "st1b { z7.s }, p1, [x28]\n" + "add x28, x28, x26\n" "trn1 z15.h, z15.h, z17.h\n" "mov z16.d, z16.d\n" "bgt 20b\n" "21:" // Main loop tail ".inst 0xc1721568 // sdot za.s[x8, 0], { z11.h-z14.h }, z2.h\n" - "addvl x24, SP, #6\n" - "addvl x23, SP, #12\n" + "addvl x23, SP, #6\n" + "addvl x22, SP, #12\n" ".inst 0xc17a1588 // sdot za.s[x8, 0], { z12.h-z15.h }, z10.h\n" - ".inst 0xa1402b02 // ld1h { z2.h, z10.h }, pn10.b/Z, [x24]\n" + ".inst 0xa1402ae2 // ld1h { z2.h, z10.h }, pn10.b/Z, [x23]\n" "mov x12, #0x0\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" ".inst 0xc1721569 // sdot za.s[x8, 1], { z11.h-z14.h }, z2.h\n" - "add x22, x17, %x[ld_in_row]\n" - "addvl x21, SP, #3\n" + "add x21, x16, %x[ld_in_row]\n" + "addvl x20, SP, #3\n" ".inst 0xc17a1589 // sdot za.s[x8, 1], { z12.h-z15.h }, z10.h\n" - ".inst 0xa1402ae2 // ld1h { z2.h, z10.h }, pn10.b/Z, [x23]\n" - "addvl x20, SP, #9\n" + ".inst 0xa1402ac2 // ld1h { z2.h, z10.h }, pn10.b/Z, [x22]\n" + "addvl x19, SP, #9\n" ".inst 0xc172156a // sdot za.s[x8, 2], { z11.h-z14.h }, z2.h\n" - "ld1b { z11.s }, p0/Z, [x17]\n" + "ld1b { z11.s }, p0/Z, [x16]\n" "add z11.h, p0/M, z11.h, z9.h\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1b { z21.s }, p0/Z, [x22]\n" + "ld1b { z21.s }, p0/Z, [x21]\n" "add z21.h, p0/M, z21.h, z9.h\n" - "add x22, x22, %x[ld_in_row]\n" + "add x21, x21, %x[ld_in_row]\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" ".inst 0xc17a158a // sdot za.s[x8, 2], { z12.h-z15.h }, z10.h\n" - "ld1b { z12.s }, p0/Z, [x22]\n" + "ld1b { z12.s }, p0/Z, [x21]\n" "add z12.h, p0/M, z12.h, z9.h\n" - "add x22, x22, %x[ld_in_row]\n" + "add x21, x21, %x[ld_in_row]\n" ".inst 0xc17015a8 // sdot za.s[x8, 0], { z13.h-z16.h }, z0.h\n" - "ld1h { z0.h }, p2/Z, [x24, #2, MUL VL]\n" + "ld1h { z0.h }, p2/Z, [x23, #2, MUL VL]\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" "mov x12, #0x4\n" - "ld1b { z20.s }, p0/Z, [x22]\n" + "ld1b { z20.s }, p0/Z, [x21]\n" ".inst 0xc17015a9 // sdot za.s[x8, 1], { z13.h-z16.h }, z0.h\n" "add z20.h, p0/M, z20.h, z9.h\n" - "add x22, x22, %x[ld_in_row]\n" - "ld1h { z0.h }, p2/Z, [x23, #2, MUL VL]\n" + "add x21, x21, %x[ld_in_row]\n" + "ld1h { z0.h }, p2/Z, [x22, #2, MUL VL]\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" ".inst 0xc17015aa // sdot za.s[x8, 2], { z13.h-z16.h }, z0.h\n" "trn1 z11.h, z11.h, z21.h\n" - "ld1b { z13.s }, p0/Z, [x22]\n" + "ld1b { z13.s }, p0/Z, [x21]\n" "add z13.h, p0/M, z13.h, z9.h\n" - "add x22, x22, %x[ld_in_row]\n" + "add x21, x21, %x[ld_in_row]\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1b { z19.s }, p0/Z, [x22]\n" + "ld1b { z19.s }, p0/Z, [x21]\n" "add z19.h, p0/M, z19.h, z9.h\n" - "add x22, x22, %x[ld_in_row]\n" + "add x21, x21, %x[ld_in_row]\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "ld1b { z14.s }, p0/Z, [x22]\n" + "ld1b { z14.s }, p0/Z, [x21]\n" "add z14.h, p0/M, z14.h, z9.h\n" - "add x22, x22, %x[ld_in_row]\n" + "add x21, x21, %x[ld_in_row]\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" - "ld1b { z18.s }, p0/Z, [x22]\n" + "ld1b { z18.s }, p0/Z, [x21]\n" "mov x12, #0x8\n" "add z18.h, p0/M, z18.h, z9.h\n" - "add x22, x22, %x[ld_in_row]\n" + "add x21, x21, %x[ld_in_row]\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1b { z15.s }, p0/Z, [x22]\n" + "ld1b { z15.s }, p0/Z, [x21]\n" "add z15.h, p0/M, z15.h, z9.h\n" - "add x22, x22, %x[ld_in_row]\n" + "add x21, x21, %x[ld_in_row]\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1b { z17.s }, p0/Z, [x22]\n" + "ld1b { z17.s }, p0/Z, [x21]\n" "add z17.h, p0/M, z17.h, z9.h\n" "trn1 z12.h, z12.h, z20.h\n" "trn1 z13.h, z13.h, z19.h\n" "trn1 z14.h, z14.h, z18.h\n" - ".inst 0xa1402aa2 // ld1h { z2.h, z10.h }, pn10.b/Z, [x21]\n" - "add x22, x22, %x[ld_in_row]\n" + ".inst 0xa1402a82 // ld1h { z2.h, z10.h }, pn10.b/Z, [x20]\n" + "add x21, x21, %x[ld_in_row]\n" ".inst 0xc0060c04 // mova { z4.d-z7.d }, za.d[x8, #0]\n" "add x8, x8, #0x1\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" "trn1 z15.h, z15.h, z17.h\n" ".inst 0xc1721568 // sdot za.s[x8, 0], { z11.h-z14.h }, z2.h\n" - "ld1b { z16.s }, p0/Z, [x22]\n" + "ld1b { z16.s }, p0/Z, [x21]\n" ".inst 0xc1a3ac04 // sqdmulh { z4.s-z7.s }, { z4.s-z7.s }, z3.s\n" "add z16.h, p0/M, z16.h, z9.h\n" ".inst 0xc17a1588 // sdot za.s[x8, 0], { z12.h-z15.h }, z10.h\n" - ".inst 0xa1402a82 // ld1h { z2.h, z10.h }, pn10.b/Z, [x20]\n" - "add x17, x17, %x[ld_in_col]\n" + ".inst 0xa1402a62 // ld1h { z2.h, z10.h }, pn10.b/Z, [x19]\n" + "add x16, x16, %x[ld_in_col]\n" ".inst 0xc1a1aa24 // srshl { z4.s-z7.s }, { z4.s-z7.s }, z1.s\n" ".inst 0xc1721569 // sdot za.s[x8, 1], { z11.h-z14.h }, z2.h\n" "mov z16.d, z16.d\n" - "ld1h { z0.h }, p2/Z, [x21, #2, MUL VL]\n" + "ld1h { z0.h }, p2/Z, [x20, #2, MUL VL]\n" ".inst 0xc17a1589 // sdot za.s[x8, 1], { z12.h-z15.h }, z10.h\n" ".inst 0xc1a8ab04 // add { z4.s-z7.s }, { z4.s-z7.s }, z8.s\n" ".inst 0xc17015a8 // sdot za.s[x8, 0], { z13.h-z16.h }, z0.h\n" - "ld1h { z0.h }, p2/Z, [x20, #2, MUL VL]\n" + "ld1h { z0.h }, p2/Z, [x19, #2, MUL VL]\n" ".inst 0xc1b7cf44 // sclamp { z4.s-z7.s }, z26.s, z23.s\n" - "st1b { z4.s }, p1, [x15]\n" - "add x15, x15, x13\n" - "st1b { z5.s }, p1, [x14]\n" + "st1b { z4.s }, p1, [x14]\n" "add x14, x14, x11\n" + "st1b { z5.s }, p1, [x13]\n" + "add x13, x13, x10\n" ".inst 0xc0040f84 // mova za.d[x8, #4], { z28.d-z31.d }\n" ".inst 0xa1402be2 // ld1h { z2.h, z10.h }, pn10.b/Z, [SP]\n" - "st1b { z6.s }, p1, [x10]\n" - "add x10, x10, x28\n" + "st1b { z6.s }, p1, [x9]\n" + "add x9, x9, x27\n" ".inst 0xc17015a9 // sdot za.s[x8, 1], { z13.h-z16.h }, z0.h\n" "ld1h { z0.h }, p2/Z, [SP, #2, MUL VL]\n" - "st1b { z7.s }, p1, [x9]\n" - "add x9, x9, x27\n" + "st1b { z7.s }, p1, [x28]\n" + "add x28, x28, x26\n" "22:" // Main loop skip tail - "cbz x7, 23f\n" // Skip remainder inputs + "cbz x17, 23f\n" // Skip remainder inputs "mov x12, #0x0\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1b { z11.s }, p0/Z, [x17]\n" + "ld1b { z11.s }, p0/Z, [x16]\n" "add z11.h, p0/M, z11.h, z9.h\n" - "add x20, x17, %x[ld_in_row]\n" + "add x19, x16, %x[ld_in_row]\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1b { z21.s }, p0/Z, [x20]\n" + "ld1b { z21.s }, p0/Z, [x19]\n" "add z21.h, p0/M, z21.h, z9.h\n" - "add x20, x20, %x[ld_in_row]\n" + "add x19, x19, %x[ld_in_row]\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "ld1b { z12.s }, p0/Z, [x20]\n" + "ld1b { z12.s }, p0/Z, [x19]\n" "add z12.h, p0/M, z12.h, z9.h\n" - "add x20, x20, %x[ld_in_row]\n" + "add x19, x19, %x[ld_in_row]\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" - "ld1b { z20.s }, p0/Z, [x20]\n" + "ld1b { z20.s }, p0/Z, [x19]\n" "add z20.h, p0/M, z20.h, z9.h\n" "mov x12, #0x4\n" - "add x20, x20, %x[ld_in_row]\n" + "add x19, x19, %x[ld_in_row]\n" "trn1 z11.h, z11.h, z21.h\n" "trn1 z12.h, z12.h, z20.h\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1b { z13.s }, p0/Z, [x20]\n" + "ld1b { z13.s }, p0/Z, [x19]\n" "add z13.h, p0/M, z13.h, z9.h\n" - "add x20, x20, %x[ld_in_row]\n" + "add x19, x19, %x[ld_in_row]\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1b { z19.s }, p0/Z, [x20]\n" + "ld1b { z19.s }, p0/Z, [x19]\n" "add z19.h, p0/M, z19.h, z9.h\n" - "add x20, x20, %x[ld_in_row]\n" + "add x19, x19, %x[ld_in_row]\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" - "ld1b { z14.s }, p0/Z, [x20]\n" + "ld1b { z14.s }, p0/Z, [x19]\n" "add z14.h, p0/M, z14.h, z9.h\n" - "add x20, x20, %x[ld_in_row]\n" + "add x19, x19, %x[ld_in_row]\n" ".inst 0x25f04500 // psel p0.s, p1.s/Z, p8.s[w12, #3]\n" - "ld1b { z18.s }, p0/Z, [x20]\n" + "ld1b { z18.s }, p0/Z, [x19]\n" "mov x12, #0x8\n" "add z18.h, p0/M, z18.h, z9.h\n" - "add x20, x20, %x[ld_in_row]\n" + "add x19, x19, %x[ld_in_row]\n" ".inst 0x25304500 // psel p0.s, p1.s/Z, p8.s[w12]\n" - "ld1b { z15.s }, p0/Z, [x20]\n" + "ld1b { z15.s }, p0/Z, [x19]\n" "add z15.h, p0/M, z15.h, z9.h\n" - "add x20, x20, %x[ld_in_row]\n" + "add x19, x19, %x[ld_in_row]\n" ".inst 0x25704500 // psel p0.s, p1.s/Z, p8.s[w12, #1]\n" - "ld1b { z17.s }, p0/Z, [x20]\n" + "ld1b { z17.s }, p0/Z, [x19]\n" "add z17.h, p0/M, z17.h, z9.h\n" - "add x20, x20, %x[ld_in_row]\n" + "add x19, x19, %x[ld_in_row]\n" ".inst 0x25b04500 // psel p0.s, p1.s/Z, p8.s[w12, #2]\n" "trn1 z13.h, z13.h, z19.h\n" "trn1 z14.h, z14.h, z18.h\n" - "ld1b { z16.s }, p0/Z, [x20]\n" + "ld1b { z16.s }, p0/Z, [x19]\n" "add z16.h, p0/M, z16.h, z9.h\n" "trn1 z15.h, z15.h, z17.h\n" - "addvl x21, SP, #6\n" + "addvl x20, SP, #6\n" ".inst 0xc1721568 // sdot za.s[x8, 0], { z11.h-z14.h }, z2.h\n" "mov z16.d, z16.d\n" - "addvl x20, SP, #12\n" - "sub x16, x16, #0x1\n" + "addvl x19, SP, #12\n" + "sub x15, x15, #0x1\n" ".inst 0xc17a1588 // sdot za.s[x8, 0], { z12.h-z15.h }, z10.h\n" - ".inst 0xa1402aa2 // ld1h { z2.h, z10.h }, pn10.b/Z, [x21]\n" + ".inst 0xa1402a82 // ld1h { z2.h, z10.h }, pn10.b/Z, [x20]\n" ".inst 0xc17015a8 // sdot za.s[x8, 0], { z13.h-z16.h }, z0.h\n" - "ld1h { z0.h }, p2/Z, [x21, #2, MUL VL]\n" + "ld1h { z0.h }, p2/Z, [x20, #2, MUL VL]\n" ".inst 0xc1721569 // sdot za.s[x8, 1], { z11.h-z14.h }, z2.h\n" ".inst 0xc0060c04 // mova { z4.d-z7.d }, za.d[x8, #0]\n" ".inst 0xc1a3ac04 // sqdmulh { z4.s-z7.s }, { z4.s-z7.s }, z3.s\n" ".inst 0xc17a1589 // sdot za.s[x8, 1], { z12.h-z15.h }, z10.h\n" - ".inst 0xa1402a82 // ld1h { z2.h, z10.h }, pn10.b/Z, [x20]\n" + ".inst 0xa1402a62 // ld1h { z2.h, z10.h }, pn10.b/Z, [x19]\n" ".inst 0xc1a1aa24 // srshl { z4.s-z7.s }, { z4.s-z7.s }, z1.s\n" ".inst 0xc172156a // sdot za.s[x8, 2], { z11.h-z14.h }, z2.h\n" ".inst 0xc1a8ab04 // add { z4.s-z7.s }, { z4.s-z7.s }, z8.s\n" ".inst 0xc17a158a // sdot za.s[x8, 2], { z12.h-z15.h }, z10.h\n" ".inst 0xc1b7cf44 // sclamp { z4.s-z7.s }, z26.s, z23.s\n" - "st1b { z4.s }, p1, [x15]\n" - "add x15, x15, x13\n" - ".inst 0xc17015a9 // sdot za.s[x8, 1], { z13.h-z16.h }, z0.h\n" - "ld1h { z0.h }, p2/Z, [x20, #2, MUL VL]\n" - "st1b { z5.s }, p1, [x14]\n" + "st1b { z4.s }, p1, [x14]\n" "add x14, x14, x11\n" + ".inst 0xc17015a9 // sdot za.s[x8, 1], { z13.h-z16.h }, z0.h\n" + "ld1h { z0.h }, p2/Z, [x19, #2, MUL VL]\n" + "st1b { z5.s }, p1, [x13]\n" + "add x13, x13, x10\n" ".inst 0xc17015aa // sdot za.s[x8, 2], { z13.h-z16.h }, z0.h\n" "add x8, x8, #0x1\n" - "st1b { z6.s }, p1, [x10]\n" - "add x10, x10, x28\n" - "st1b { z7.s }, p1, [x9]\n" + "st1b { z6.s }, p1, [x9]\n" "add x9, x9, x27\n" + "st1b { z7.s }, p1, [x28]\n" + "add x28, x28, x26\n" ".inst 0xc0040f84 // mova za.d[x8, #4], { z28.d-z31.d }\n" "23:" // Tail input: End - "cbz x16, 25f\n" + "cbz x15, 25f\n" "24:" // Right padding loop ".inst 0xc0060c04 // mova { z4.d-z7.d }, za.d[x8, #0]\n" ".inst 0xc1a3ac04 // sqdmulh { z4.s-z7.s }, { z4.s-z7.s }, z3.s\n" "add x8, x8, #0x1\n" ".inst 0xc1a1aa24 // srshl { z4.s-z7.s }, { z4.s-z7.s }, z1.s\n" - "subs x16, x16, #0x1\n" + "subs x15, x15, #0x1\n" ".inst 0xc0040f84 // mova za.d[x8, #4], { z28.d-z31.d }\n" ".inst 0xc1a8ab04 // add { z4.s-z7.s }, { z4.s-z7.s }, z8.s\n" ".inst 0xc1b7cf44 // sclamp { z4.s-z7.s }, z26.s, z23.s\n" - "st1b { z4.s }, p1, [x15]\n" - "add x15, x15, x13\n" - "st1b { z5.s }, p1, [x14]\n" + "st1b { z4.s }, p1, [x14]\n" "add x14, x14, x11\n" - "st1b { z6.s }, p1, [x10]\n" - "add x10, x10, x28\n" - "st1b { z7.s }, p1, [x9]\n" + "st1b { z5.s }, p1, [x13]\n" + "add x13, x13, x10\n" + "st1b { z6.s }, p1, [x9]\n" "add x9, x9, x27\n" + "st1b { z7.s }, p1, [x28]\n" + "add x28, x28, x26\n" "bgt 24b\n" "25:" // End - "ldr x22, [%x[args], %[offsetof_Args_weights]]\n" - "incw x22, ALL, MUL #16\n" - "incw x22, ALL, MUL #9\n" - "str x22, [%x[args], %[offsetof_Args_weights]]\n" - "ldr x20, [%x[args], %[offsetof_Args_ld_in_vl]]\n" - "incw x6\n" - "whilelt p1.s, x6, x5\n" - "ldr x17, [%x[args], %[offsetof_Args_inptr]]\n" - "add x17, x17, x20\n" - "str x17, [%x[args], %[offsetof_Args_inptr]]\n" - "ldr x25, [%x[args], %[offsetof_Args_outptrs]]\n" - "ldr x24, [%x[args], %[offsetof_Args_ld_out_vls]]\n" - "ldp x23, x22, [x25, #0x0]\n" - "ldp x21, x20, [x24, #0x0]\n" - "add x23, x23, x21\n" + "ldr x21, [%x[args], %[offsetof_Args_weights]]\n" + "incw x21, ALL, MUL #16\n" + "incw x21, ALL, MUL #9\n" + "str x21, [%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\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\n" - "stp x23, x22, [x25, #0x0]\n" - "ldp x23, x22, [x25, #0x10]\n" - "ldp x21, x20, [x24, #0x10]\n" - "add x23, x23, x21\n" + "add x21, x21, x19\n" + "stp x22, x21, [x24, #0x0]\n" + "ldp x22, x21, [x24, #0x10]\n" + "ldp x20, x19, [x23, #0x10]\n" "add x22, x22, x20\n" - "stp x23, x22, [x25, #0x10]\n" + "add x21, x21, x19\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_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)), [offsetof_Requantize32_a_offset] "I" (offsetof(arm_gemm::Requantize32, a_offset)), [offsetof_Requantize32_b_offset] "I" (offsetof(arm_gemm::Requantize32, b_offset)), [offsetof_Requantize32_bias] "I" (offsetof(arm_gemm::Requantize32, bias)), [offsetof_Requantize32_c_offset] "I" (offsetof(arm_gemm::Requantize32, c_offset)), [offsetof_Requantize32_maxval] "I" (offsetof(arm_gemm::Requantize32, maxval)), [offsetof_Requantize32_minval] "I" (offsetof(arm_gemm::Requantize32, minval)), [offsetof_Requantize32_per_channel_muls] "I" (offsetof(arm_gemm::Requantize32, per_channel_muls)), [offsetof_Requantize32_per_channel_right_shifts] "I" (offsetof(arm_gemm::Requantize32, per_channel_right_shifts)), [offsetof_Requantize32_per_layer_mul] "I" (offsetof(arm_gemm::Requantize32, per_layer_mul)), [offsetof_Requantize32_per_layer_right_shift] "I" (offsetof(arm_gemm::Requantize32, per_layer_right_shift)), [qp] "r" (&qp) - : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x3", "x4", "x5", "x6", "x7", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "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" + : "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" ); } -- cgit v1.2.1