diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp | 360 |
1 files changed, 180 insertions, 180 deletions
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp index 96cfd5e497..e0c7d71e61 100644 --- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp +++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2022-2023 Arm Limited. + * Copyright (c) 2022-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -88,240 +88,240 @@ void sme2_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst_direct_impl( __asm__ __volatile__( ".inst 0xd503477f // SMSTART ZA\n" - "ptrue p3.b\n" - ".inst 0x25207810 // ptrue pn8.b\n" "mov x4, #0x0\n" "mov x5, #0x0\n" + "ptrue p3.b\n" + ".inst 0x25207810 // ptrue pn8.b\n" "1:" // Tile loop "str x4, [%x[params_struct], %[offsetof_args_tile_i]]\n" "mov x22, #0x2\n" "str x5, [%x[params_struct], %[offsetof_args_tile_j]]\n" "ldr x21, [%x[params_struct], %[offsetof_args_ld_input_row]]\n" - "mul x20, x4, x21\n" // offset = tile_i * ld_input_row "ldr x6, [%x[params_struct], %[offsetof_args_ld_input_col]]\n" + "ldr x7, [%x[params_struct], %[offsetof_args_inptr]]\n" + "mul x20, x4, x21\n" // offset = tile_i * ld_input_row + "ldr x8, [%x[params_struct], %[offsetof_args_params]]\n" "madd x20, x5, x6, x20\n" // offset += tile_j * ld_input_col + "add x17, x6, x6\n" "mul x20, x20, x22\n" // offset *= kernel_stride * output_size - "ldr x7, [%x[params_struct], %[offsetof_args_inptr]]\n" + "add x16, x17, x6\n" "add x7, x7, x20, LSL #2\n" // inptr[0] += offset * sizeof(float) - "add x8, x7, x21, LSL #2\n" - "add x17, x8, x21, LSL #2\n" - "add x16, x6, x6\n" - "ldr x15, [%x[params_struct], %[offsetof_args_params]]\n" - "add x14, x17, x21, LSL #2\n" - "add x13, x16, x6\n" + "add x15, x7, x21, LSL #2\n" + "add x14, x15, x21, LSL #2\n" + "add x13, x14, x21, LSL #2\n" "cbnz x5, 2f\n" "ldr x20, [%x[params_struct], %[offsetof_args_n_tile_cols]]\n" - "sub x21, x20, x5\n" - "sub x21, x21, #0x1\n" "lsl x12, %x[n_channels], #0x2\n" - "mov x20, #0x8\n" - "and x21, x21, #0x3fffff\n" - "mul x20, x20, x6\n" - "orr x12, x12, x21, LSL #22\n" - "orr x12, x12, x20, LSL #38\n" - "add x11, x8, x6, LSL #2\n" - "add x10, x7, x13, LSL #2\n" - "add x9, x8, x16, LSL #2\n" - "add x28, x17, x6, LSL #2\n" - "add x27, x14, x13, LSL #2\n" + "mov x21, #0x8\n" + "mul x21, x21, x6\n" + "add x11, x15, x6, LSL #2\n" + "add x10, x7, x16, LSL #2\n" + "add x9, x15, x17, LSL #2\n" + "sub x20, x20, x5\n" + "add x28, x14, x6, LSL #2\n" + "sub x20, x20, #0x1\n" + "add x27, x13, x16, LSL #2\n" + "and x20, x20, #0x3fffff\n" "add x26, x7, x6, LSL #2\n" - "add x25, x7, x16, LSL #2\n" - "add x24, x17, x16, LSL #2\n" - "add x23, x8, x13, LSL #2\n" - "add x22, x17, x13, LSL #2\n" - "add x21, x14, x6, LSL #2\n" - "add x20, x14, x16, LSL #2\n" + "orr x12, x12, x20, LSL #22\n" + "add x25, x7, x17, LSL #2\n" + "orr x12, x12, x21, LSL #38\n" + "add x24, x14, x17, LSL #2\n" + "add x23, x15, x16, LSL #2\n" + "add x22, x14, x16, LSL #2\n" + "add x21, x13, x6, LSL #2\n" + "add x20, x13, x17, LSL #2\n" ".inst 0xf8ac497a // rprfm pldonce, x12, [x11]\n" ".inst 0xf8ac48fa // rprfm pldonce, x12, [x7]\n" ".inst 0xf8ac495a // rprfm pldonce, x12, [x10]\n" ".inst 0xf8ac493a // rprfm pldonce, x12, [x9]\n" ".inst 0xf8ac4b9a // rprfm pldonce, x12, [x28]\n" - ".inst 0xf8ac49da // rprfm pldonce, x12, [x14]\n" + ".inst 0xf8ac49ba // rprfm pldonce, x12, [x13]\n" ".inst 0xf8ac4b7a // rprfm pldonce, x12, [x27]\n" ".inst 0xf8ac4b5a // rprfm pldonce, x12, [x26]\n" ".inst 0xf8ac4b3a // rprfm pldonce, x12, [x25]\n" ".inst 0xf8ac4b1a // rprfm pldonce, x12, [x24]\n" - ".inst 0xf8ac491a // rprfm pldonce, x12, [x8]\n" + ".inst 0xf8ac49fa // rprfm pldonce, x12, [x15]\n" ".inst 0xf8ac4afa // rprfm pldonce, x12, [x23]\n" - ".inst 0xf8ac4a3a // rprfm pldonce, x12, [x17]\n" + ".inst 0xf8ac49da // rprfm pldonce, x12, [x14]\n" ".inst 0xf8ac4ada // rprfm pldonce, x12, [x22]\n" ".inst 0xf8ac4aba // rprfm pldonce, x12, [x21]\n" ".inst 0xf8ac4a9a // rprfm pldonce, x12, [x20]\n" "2:" // Tile loop: Prefetch input rows: End - "ldr x22, [%x[params_struct], %[offsetof_args_ld_output_row]]\n" - "mul x21, x4, x22\n" // offset = tile_i * ld_output_row + "ldr x23, [%x[params_struct], %[offsetof_args_ld_output_row]]\n" "mov x20, #0x2\n" - "ld1w { z22.s }, p3/Z, [x15]\n" - "ldr x25, [%x[params_struct], %[offsetof_args_ld_output_col]]\n" - "madd x21, x5, x25, x21\n" // offset += tile_j * ld_output_col - "addvl x15, x15, #1\n" - ".inst 0xa040c1e0 // ld1w { z0.s-z3.s }, pn8.b/Z, [x15]\n" + "ld1w { z22.s }, p3/Z, [x8]\n" + "addvl x8, x8, #1\n" + "ldr x26, [%x[params_struct], %[offsetof_args_ld_output_col]]\n" + "cntw x25\n" + ".inst 0xa040c100 // ld1w { z0.s-z3.s }, pn8.b/Z, [x8]\n" + "addvl x8, x8, #4\n" "ldr x24, [%x[params_struct], %[offsetof_args_outptr]]\n" - "mul x21, x21, x20\n" // offset *= output_tile_size - "cntw x23\n" - "ld1rw { z21.s }, p3/Z, [%x[params_struct], %[offsetof_args_min]]\n" - "addvl x15, x15, #4\n" - "add x24, x24, x21, LSL #2\n" // outptrs[0] += offset * sizeof(float) - ".inst 0xa040c1e4 // ld1w { z4.s-z7.s }, pn8.b/Z, [x15]\n" "whilelt p2.s, XZR, %x[n_channels]\n" - "addvl x15, x15, #4\n" - "ld1rw { z14.s }, p3/Z, [%x[params_struct], %[offsetof_args_max]]\n" - "cmp x23, %x[n_channels]\n" - "add x22, x24, x22, LSL #2\n" - "ld1w { z8.s }, p3/Z, [x15]\n" + ".inst 0xa040c104 // ld1w { z4.s-z7.s }, pn8.b/Z, [x8]\n" + "addvl x8, x8, #4\n" + "mul x22, x4, x23\n" // offset = tile_i * ld_output_row + "cmp x25, %x[n_channels]\n" + "ld1rw { z15.s }, p3/Z, [%x[params_struct], %[offsetof_args_min]]\n" + "madd x22, x5, x26, x22\n" // offset += tile_j * ld_output_col + "ld1rw { z21.s }, p3/Z, [%x[params_struct], %[offsetof_args_max]]\n" "mov x21, #0x0\n" - "sub x20, XZR, x23\n" - "ld1w { z9.s }, p2/Z, [x8, x6, LSL #2]\n" + "mul x22, x22, x20\n" // offset *= output_tile_size + "sub x20, XZR, x25\n" + "ld1w { z8.s }, p3/Z, [x8]\n" + "add x24, x24, x22, LSL #2\n" // outptrs[0] += offset * sizeof(float) + "ld1w { z9.s }, p2/Z, [x15, x6, LSL #2]\n" + "addvl x8, x8, #1\n" + "add x23, x24, x23, LSL #2\n" "ld1w { z10.s }, p2/Z, [x7]\n" - "addvl x15, x15, #1\n" - "ld1w { z11.s }, p2/Z, [x7, x13, LSL #2]\n" - "ld1w { z12.s }, p2/Z, [x8, x16, LSL #2]\n" - "ld1w { z13.s }, p2/Z, [x17, x6, LSL #2]\n" + "ld1w { z11.s }, p2/Z, [x7, x16, LSL #2]\n" + "ld1w { z12.s }, p2/Z, [x15, x17, LSL #2]\n" + "ld1w { z13.s }, p2/Z, [x14, x6, LSL #2]\n" "bge 4f\n" "3:" // Tile loop: Channel loop - "movprfx z28, z22\n fmla z28.s, p3/M, z4.s, z9.s\n" - "movprfx z29, z22\n fmla z29.s, p3/M, z3.s, z9.s\n" - "whilelt p1.s, x23, %x[n_channels]\n" - "incw x21\n" - "movprfx z30, z22\n fmla z30.s, p3/M, z1.s, z9.s\n" - "movprfx z31, z22\n fmla z31.s, p3/M, z0.s, z9.s\n" - "ld1w { z18.s }, p2/Z, [x14]\n" - "incw x23\n" - "fmla z28.s, p3/M, z0.s, z10.s\n" - "fmla z29.s, p3/M, z2.s, z11.s\n" - "ld1w { z17.s }, p2/Z, [x14, x13, LSL #2]\n" - "mov p0.b, p2.b\n" - "fmla z30.s, p3/M, z2.s, z12.s\n" - "fmla z31.s, p3/M, z1.s, z12.s\n" - "ld1w { z16.s }, p2/Z, [x17, x16, LSL #2]\n" - "incw x20\n" - "fmla z28.s, p3/M, z5.s, z12.s\n" - "fmla z29.s, p3/M, z4.s, z12.s\n" - "ld1w { z11.s }, p2/Z, [x7, x6, LSL #2]\n" - "fmla z30.s, p3/M, z6.s, z18.s\n" - "fmla z31.s, p3/M, z3.s, z13.s\n" - "ld1w { z10.s }, p2/Z, [x7, x16, LSL #2]\n" - "addvl x7, x7, #1\n" - "fmla z28.s, p3/M, z7.s, z13.s\n" - "fmla z29.s, p3/M, z6.s, z13.s\n" - "ld1w { z22.s }, p3/Z, [x15]\n" - "addvl x15, x15, #1\n" - "fmla z30.s, p3/M, z4.s, z13.s\n" - "fmla z31.s, p3/M, z8.s, z17.s\n" - "ld1w { z9.s }, p2/Z, [x8]\n" - "fmla z28.s, p3/M, z1.s, z11.s\n" - "fmla z29.s, p3/M, z0.s, z11.s\n" - "ld1w { z19.s }, p2/Z, [x8, x13, LSL #2]\n" - "addvl x8, x8, #1\n" - "fmla z30.s, p3/M, z5.s, z16.s\n" - "fmla z31.s, p3/M, z4.s, z16.s\n" - "fmla z28.s, p3/M, z2.s, z10.s\n" - "fmla z29.s, p3/M, z1.s, z10.s\n" - "ld1w { z18.s }, p2/Z, [x17]\n" - "fmla z30.s, p3/M, z0.s, z9.s\n" - "fmla z31.s, p3/M, z2.s, z19.s\n" - "fmla z28.s, p3/M, z8.s, z16.s\n" - "fmla z29.s, p3/M, z7.s, z16.s\n" - "ld1w { z17.s }, p2/Z, [x17, x13, LSL #2]\n" - "addvl x17, x17, #1\n" - "fmla z30.s, p3/M, z3.s, z18.s\n" - "fmla z31.s, p3/M, z5.s, z17.s\n" - "ld1w { z13.s }, p1/Z, [x17, x6, LSL #2]\n" - "fmla z28.s, p3/M, z3.s, z9.s\n" - "ld1w { z16.s }, p2/Z, [x14, x6, LSL #2]\n" - "fmla z29.s, p3/M, z5.s, z19.s\n" - "fmla z30.s, p3/M, z7.s, z16.s\n" - "fmla z31.s, p3/M, z6.s, z16.s\n" - "ld1w { z16.s }, p2/Z, [x14, x16, LSL #2]\n" - "whilelt p2.s, x21, %x[n_channels]\n" - "fmla z28.s, p3/M, z6.s, z18.s\n" - "fmla z29.s, p3/M, z8.s, z17.s\n" - ".inst 0xa040c1e0 // ld1w { z0.s-z3.s }, pn8.b/Z, [x15]\n" - "addvl x15, x15, #4\n" - "fmla z30.s, p3/M, z8.s, z16.s\n" - "fmla z31.s, p3/M, z7.s, z16.s\n" - ".inst 0xa040c1e4 // ld1w { z4.s-z7.s }, pn8.b/Z, [x15]\n" - "addvl x15, x15, #4\n" - "cmp x23, %x[n_channels]\n" - ".inst 0xc1aecabc // fclamp { z28.s-z31.s }, z21.s, z14.s\n" - "addvl x14, x14, #1\n" - "ld1w { z9.s }, p1/Z, [x8, x6, LSL #2]\n" - "ld1w { z10.s }, p1/Z, [x7]\n" - "st1w { z28.s }, p0, [x24]\n" - "ld1w { z11.s }, p1/Z, [x7, x13, LSL #2]\n" - "st1w { z29.s }, p0, [x24, x25, LSL #2]\n" - "addvl x24, x24, #1\n" - "ld1w { z12.s }, p1/Z, [x8, x16, LSL #2]\n" - "st1w { z30.s }, p0, [x22]\n" - "st1w { z31.s }, p0, [x22, x25, LSL #2]\n" - "addvl x22, x22, #1\n" - "ld1w { z8.s }, p3/Z, [x15]\n" - "addvl x15, x15, #1\n" - "blt 3b\n" - "4:" // Tile loop: Channel tail "movprfx z24, z22\n fmla z24.s, p3/M, z4.s, z9.s\n" "movprfx z25, z22\n fmla z25.s, p3/M, z3.s, z9.s\n" - "ldr x5, [%x[params_struct], %[offsetof_args_tile_j]]\n" - "add x5, x5, #0x1\n" + "whilelt p1.s, x25, %x[n_channels]\n" + "incw x21\n" "movprfx z26, z22\n fmla z26.s, p3/M, z1.s, z9.s\n" "movprfx z27, z22\n fmla z27.s, p3/M, z0.s, z9.s\n" - "ld1w { z17.s }, p2/Z, [x14]\n" - "ldr x4, [%x[params_struct], %[offsetof_args_tile_i]]\n" + "ld1w { z17.s }, p2/Z, [x13]\n" + "incw x25\n" + "ld1w { z22.s }, p3/Z, [x8]\n" + "addvl x8, x8, #1\n" + "mov p0.b, p2.b\n" + "incw x20\n" "fmla z24.s, p3/M, z0.s, z10.s\n" "fmla z25.s, p3/M, z2.s, z11.s\n" - "ld1w { z16.s }, p2/Z, [x14, x13, LSL #2]\n" - "ldr x20, [%x[params_struct], %[offsetof_args_n_tile_cols]]\n" + "ld1w { z16.s }, p2/Z, [x13, x16, LSL #2]\n" "fmla z26.s, p3/M, z2.s, z12.s\n" "fmla z27.s, p3/M, z1.s, z12.s\n" - "ld1w { z20.s }, p2/Z, [x17, x16, LSL #2]\n" - "ldr x21, [%x[params_struct], %[offsetof_args_n_tile_rows]]\n" + "ld1w { z18.s }, p2/Z, [x14, x17, LSL #2]\n" "fmla z24.s, p3/M, z5.s, z12.s\n" "fmla z25.s, p3/M, z4.s, z12.s\n" - "ld1w { z18.s }, p2/Z, [x7, x6, LSL #2]\n" - "cmp x5, x20\n" + "ld1w { z28.s }, p2/Z, [x7, x6, LSL #2]\n" "fmla z26.s, p3/M, z6.s, z17.s\n" "fmla z27.s, p3/M, z3.s, z13.s\n" - "ld1w { z17.s }, p2/Z, [x7, x16, LSL #2]\n" - "add x20, x4, #0x1\n" + "ld1w { z14.s }, p2/Z, [x7, x17, LSL #2]\n" + "addvl x7, x7, #1\n" "fmla z24.s, p3/M, z7.s, z13.s\n" "fmla z25.s, p3/M, z6.s, z13.s\n" - "csel x4, x4, x20, LT\n" - "mov p0.b, p2.b\n" "fmla z26.s, p3/M, z4.s, z13.s\n" "fmla z27.s, p3/M, z8.s, z16.s\n" - "ld1w { z16.s }, p2/Z, [x8]\n" - "csel x5, x5, XZR, LT\n" - "fmla z24.s, p3/M, z1.s, z18.s\n" - "fmla z25.s, p3/M, z0.s, z18.s\n" - "ld1w { z19.s }, p2/Z, [x8, x13, LSL #2]\n" - "cmp x4, x21\n" - "fmla z26.s, p3/M, z5.s, z20.s\n" - "fmla z27.s, p3/M, z4.s, z20.s\n" - "fmla z24.s, p3/M, z2.s, z17.s\n" - "fmla z25.s, p3/M, z1.s, z17.s\n" - "ld1w { z18.s }, p2/Z, [x17]\n" - "fmla z26.s, p3/M, z0.s, z16.s\n" - "fmla z27.s, p3/M, z2.s, z19.s\n" - "fmla z24.s, p3/M, z8.s, z20.s\n" - "fmla z25.s, p3/M, z7.s, z20.s\n" - "ld1w { z17.s }, p2/Z, [x17, x13, LSL #2]\n" - "fmla z26.s, p3/M, z3.s, z18.s\n" - "fmla z27.s, p3/M, z5.s, z17.s\n" - "fmla z24.s, p3/M, z3.s, z16.s\n" - "ld1w { z16.s }, p2/Z, [x14, x6, LSL #2]\n" - "fmla z25.s, p3/M, z5.s, z19.s\n" - "fmla z26.s, p3/M, z7.s, z16.s\n" - "fmla z27.s, p3/M, z6.s, z16.s\n" - "ld1w { z16.s }, p2/Z, [x14, x16, LSL #2]\n" - "fmla z24.s, p3/M, z6.s, z18.s\n" - "fmla z25.s, p3/M, z8.s, z17.s\n" + "ld1w { z17.s }, p2/Z, [x15]\n" + "fmla z24.s, p3/M, z1.s, z28.s\n" + "fmla z25.s, p3/M, z0.s, z28.s\n" + "ld1w { z16.s }, p2/Z, [x15, x16, LSL #2]\n" + "addvl x15, x15, #1\n" + "fmla z26.s, p3/M, z5.s, z18.s\n" + "fmla z27.s, p3/M, z4.s, z18.s\n" + "fmla z24.s, p3/M, z2.s, z14.s\n" + "fmla z25.s, p3/M, z1.s, z14.s\n" + "ld1w { z19.s }, p2/Z, [x14]\n" + "fmla z26.s, p3/M, z0.s, z17.s\n" + "fmla z27.s, p3/M, z2.s, z16.s\n" + "fmla z24.s, p3/M, z8.s, z18.s\n" + "fmla z25.s, p3/M, z7.s, z18.s\n" + "ld1w { z18.s }, p2/Z, [x14, x16, LSL #2]\n" + "addvl x14, x14, #1\n" + "fmla z26.s, p3/M, z3.s, z19.s\n" + "ld1w { z13.s }, p1/Z, [x14, x6, LSL #2]\n" + "fmla z27.s, p3/M, z5.s, z18.s\n" + "fmla z24.s, p3/M, z3.s, z17.s\n" + "ld1w { z17.s }, p2/Z, [x13, x6, LSL #2]\n" + "fmla z25.s, p3/M, z5.s, z16.s\n" + "ld1w { z16.s }, p2/Z, [x13, x17, LSL #2]\n" + "whilelt p2.s, x21, %x[n_channels]\n" + "cmp x25, %x[n_channels]\n" + ".inst 0xa040c100 // ld1w { z0.s-z3.s }, pn8.b/Z, [x8]\n" + "addvl x8, x8, #4\n" + "addvl x13, x13, #1\n" + "fmla z26.s, p3/M, z7.s, z17.s\n" + "fmla z27.s, p3/M, z6.s, z17.s\n" + "ld1w { z11.s }, p1/Z, [x7, x16, LSL #2]\n" + "fmla z24.s, p3/M, z6.s, z19.s\n" + "fmla z25.s, p3/M, z8.s, z18.s\n" + "ld1w { z9.s }, p1/Z, [x15, x6, LSL #2]\n" + "ld1w { z10.s }, p1/Z, [x7]\n" "fmla z26.s, p3/M, z8.s, z16.s\n" "fmla z27.s, p3/M, z7.s, z16.s\n" - ".inst 0xc1aecab8 // fclamp { z24.s-z27.s }, z21.s, z14.s\n" + ".inst 0xa040c104 // ld1w { z4.s-z7.s }, pn8.b/Z, [x8]\n" + "addvl x8, x8, #4\n" + "ld1w { z12.s }, p1/Z, [x15, x17, LSL #2]\n" + "ld1w { z8.s }, p3/Z, [x8]\n" + "addvl x8, x8, #1\n" + ".inst 0xc1b5c9f8 // fclamp { z24.s-z27.s }, z15.s, z21.s\n" "st1w { z24.s }, p0, [x24]\n" - "st1w { z25.s }, p0, [x24, x25, LSL #2]\n" - "st1w { z26.s }, p0, [x22]\n" - "st1w { z27.s }, p0, [x22, x25, LSL #2]\n" + "st1w { z25.s }, p0, [x24, x26, LSL #2]\n" + "addvl x24, x24, #1\n" + "st1w { z26.s }, p0, [x23]\n" + "st1w { z27.s }, p0, [x23, x26, LSL #2]\n" + "addvl x23, x23, #1\n" + "blt 3b\n" + "4:" // Tile loop: Channel tail + "movprfx z28, z22\n fmla z28.s, p3/M, z4.s, z9.s\n" + "movprfx z29, z22\n fmla z29.s, p3/M, z3.s, z9.s\n" + "ldr x5, [%x[params_struct], %[offsetof_args_tile_j]]\n" + "mov p0.b, p2.b\n" + "movprfx z30, z22\n fmla z30.s, p3/M, z1.s, z9.s\n" + "movprfx z31, z22\n fmla z31.s, p3/M, z0.s, z9.s\n" + "ld1w { z18.s }, p2/Z, [x13]\n" + "ldr x4, [%x[params_struct], %[offsetof_args_tile_i]]\n" + "ldr x22, [%x[params_struct], %[offsetof_args_n_tile_cols]]\n" + "ldr x21, [%x[params_struct], %[offsetof_args_n_tile_rows]]\n" + "add x5, x5, #0x1\n" + "fmla z28.s, p3/M, z0.s, z10.s\n" + "fmla z29.s, p3/M, z2.s, z11.s\n" + "ld1w { z17.s }, p2/Z, [x13, x16, LSL #2]\n" + "add x20, x4, #0x1\n" + "fmla z30.s, p3/M, z2.s, z12.s\n" + "fmla z31.s, p3/M, z1.s, z12.s\n" + "ld1w { z20.s }, p2/Z, [x14, x17, LSL #2]\n" + "cmp x5, x22\n" + "csel x4, x4, x20, LT\n" + "csel x5, x5, XZR, LT\n" + "cmp x4, x21\n" + "fmla z28.s, p3/M, z5.s, z12.s\n" + "fmla z29.s, p3/M, z4.s, z12.s\n" + "ld1w { z16.s }, p2/Z, [x7, x6, LSL #2]\n" + "fmla z30.s, p3/M, z6.s, z18.s\n" + "fmla z31.s, p3/M, z3.s, z13.s\n" + "ld1w { z18.s }, p2/Z, [x7, x17, LSL #2]\n" + "fmla z28.s, p3/M, z7.s, z13.s\n" + "fmla z29.s, p3/M, z6.s, z13.s\n" + "fmla z30.s, p3/M, z4.s, z13.s\n" + "fmla z31.s, p3/M, z8.s, z17.s\n" + "ld1w { z17.s }, p2/Z, [x15]\n" + "fmla z28.s, p3/M, z1.s, z16.s\n" + "fmla z29.s, p3/M, z0.s, z16.s\n" + "ld1w { z16.s }, p2/Z, [x15, x16, LSL #2]\n" + "fmla z30.s, p3/M, z5.s, z20.s\n" + "fmla z31.s, p3/M, z4.s, z20.s\n" + "fmla z28.s, p3/M, z2.s, z18.s\n" + "fmla z29.s, p3/M, z1.s, z18.s\n" + "ld1w { z19.s }, p2/Z, [x14]\n" + "fmla z30.s, p3/M, z0.s, z17.s\n" + "fmla z31.s, p3/M, z2.s, z16.s\n" + "fmla z28.s, p3/M, z8.s, z20.s\n" + "fmla z29.s, p3/M, z7.s, z20.s\n" + "ld1w { z18.s }, p2/Z, [x14, x16, LSL #2]\n" + "fmla z30.s, p3/M, z3.s, z19.s\n" + "fmla z31.s, p3/M, z5.s, z18.s\n" + "fmla z28.s, p3/M, z3.s, z17.s\n" + "ld1w { z17.s }, p2/Z, [x13, x6, LSL #2]\n" + "fmla z29.s, p3/M, z5.s, z16.s\n" + "ld1w { z16.s }, p2/Z, [x13, x17, LSL #2]\n" + "fmla z30.s, p3/M, z7.s, z17.s\n" + "fmla z31.s, p3/M, z6.s, z17.s\n" + "fmla z28.s, p3/M, z6.s, z19.s\n" + "fmla z29.s, p3/M, z8.s, z18.s\n" + "fmla z30.s, p3/M, z8.s, z16.s\n" + "fmla z31.s, p3/M, z7.s, z16.s\n" + ".inst 0xc1b5c9fc // fclamp { z28.s-z31.s }, z15.s, z21.s\n" + "st1w { z28.s }, p0, [x24]\n" + "st1w { z29.s }, p0, [x24, x26, LSL #2]\n" + "st1w { z30.s }, p0, [x23]\n" + "st1w { z31.s }, p0, [x23, x26, LSL #2]\n" "blt 1b\n" ".inst 0xd503467f // SMSTOP\n" : |