diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_direct.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_direct.cpp | 424 |
1 files changed, 212 insertions, 212 deletions
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_direct.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_direct.cpp index 066b935486..131a8eec01 100644 --- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_direct.cpp +++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s2_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 * @@ -88,78 +88,78 @@ void sme2_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst_direct_impl( __asm__ __volatile__( ".inst 0xd503477f // SMSTART ZA\n" - "ptrue p3.b\n" - ".inst 0x25207810 // ptrue pn8.b\n" "mov x2, #0x0\n" "mov x3, #0x0\n" + "ptrue p3.b\n" + ".inst 0x25207810 // ptrue pn8.b\n" "1:" // Tile loop "str x2, [%x[params_struct], %[offsetof_args_tile_i]]\n" "mov x22, #0x4\n" "str x3, [%x[params_struct], %[offsetof_args_tile_j]]\n" "ldr x21, [%x[params_struct], %[offsetof_args_ld_input_row]]\n" - "mul x20, x2, x21\n" // offset = tile_i * ld_input_row "ldr x4, [%x[params_struct], %[offsetof_args_ld_input_col]]\n" + "ldr x5, [%x[params_struct], %[offsetof_args_inptr]]\n" + "mul x20, x2, x21\n" // offset = tile_i * ld_input_row + "ldr x6, [%x[params_struct], %[offsetof_args_params]]\n" "madd x20, x3, x4, x20\n" // offset += tile_j * ld_input_col + "add x7, x4, x4\n" "mul x20, x20, x22\n" // offset *= kernel_stride * output_size - "ldr x5, [%x[params_struct], %[offsetof_args_inptr]]\n" + "add x8, x7, x4\n" "add x5, x5, x20, LSL #2\n" // inptr[0] += offset * sizeof(float) - "add x6, x5, x21, LSL #2\n" - "add x7, x6, x21, LSL #2\n" - "add x8, x4, x4\n" - "ldr x17, [%x[params_struct], %[offsetof_args_params]]\n" - "add x16, x7, x21, LSL #2\n" - "add x15, x8, x4\n" - "add x14, x16, x21, LSL #2\n" - "add x13, x15, x4\n" + "add x17, x8, x4\n" + "add x16, x5, x21, LSL #2\n" + "add x15, x16, x21, LSL #2\n" + "add x14, x15, x21, LSL #2\n" + "add x13, x14, x21, LSL #2\n" "cbnz x3, 2f\n" "ldr x20, [%x[params_struct], %[offsetof_args_n_tile_cols]]\n" - "sub x21, x20, x3\n" - "sub x21, x21, #0x1\n" "lsl x12, %x[n_channels], #0x2\n" - "mov x20, #0x10\n" - "and x21, x21, #0x3fffff\n" - "mul x20, x20, x4\n" - "orr x12, x12, x21, LSL #22\n" - "orr x12, x12, x20, LSL #38\n" - "add x27, x7, x8, LSL #2\n" + "mov x28, #0x10\n" + "mul x28, x28, x4\n" + "add x27, x15, x7, LSL #2\n" "add x26, x5, x4, LSL #2\n" - "add x25, x5, x15, LSL #2\n" - "add x24, x5, x13, LSL #2\n" - "add x23, x6, x4, LSL #2\n" - "add x22, x5, x8, LSL #2\n" - "add x21, x6, x15, LSL #2\n" - "add x20, x6, x13, LSL #2\n" - "add x11, x6, x8, LSL #2\n" - "add x10, x16, x4, LSL #2\n" - "add x9, x7, x4, LSL #2\n" - "add x28, x16, x15, LSL #2\n" + "add x25, x5, x8, LSL #2\n" + "sub x20, x20, x3\n" + "add x24, x5, x17, LSL #2\n" + "sub x20, x20, #0x1\n" + "add x23, x16, x4, LSL #2\n" + "and x20, x20, #0x3fffff\n" + "add x22, x5, x7, LSL #2\n" + "orr x12, x12, x20, LSL #22\n" + "add x21, x16, x8, LSL #2\n" + "orr x12, x12, x28, LSL #38\n" + "add x20, x16, x17, LSL #2\n" + "add x11, x16, x7, LSL #2\n" + "add x10, x14, x4, LSL #2\n" + "add x9, x15, x4, LSL #2\n" + "add x28, x14, x8, LSL #2\n" ".inst 0xf8ac4b7a // rprfm pldonce, x12, [x27]\n" - "add x27, x7, x15, LSL #2\n" + "add x27, x15, x8, LSL #2\n" ".inst 0xf8ac48ba // rprfm pldonce, x12, [x5]\n" ".inst 0xf8ac4b5a // rprfm pldonce, x12, [x26]\n" - "add x26, x16, x13, LSL #2\n" + "add x26, x14, x17, LSL #2\n" ".inst 0xf8ac4b3a // rprfm pldonce, x12, [x25]\n" - "add x25, x7, x13, LSL #2\n" + "add x25, x15, x17, LSL #2\n" ".inst 0xf8ac4b1a // rprfm pldonce, x12, [x24]\n" - "add x24, x14, x4, LSL #2\n" - ".inst 0xf8ac48da // rprfm pldonce, x12, [x6]\n" + "add x24, x13, x4, LSL #2\n" + ".inst 0xf8ac4a1a // rprfm pldonce, x12, [x16]\n" ".inst 0xf8ac4afa // rprfm pldonce, x12, [x23]\n" - "add x23, x16, x8, LSL #2\n" + "add x23, x14, x7, LSL #2\n" ".inst 0xf8ac4ada // rprfm pldonce, x12, [x22]\n" - "add x22, x14, x15, LSL #2\n" + "add x22, x13, x8, LSL #2\n" ".inst 0xf8ac4aba // rprfm pldonce, x12, [x21]\n" - "add x21, x14, x8, LSL #2\n" + "add x21, x13, x7, LSL #2\n" ".inst 0xf8ac4a9a // rprfm pldonce, x12, [x20]\n" - "add x20, x14, x13, LSL #2\n" + "add x20, x13, x17, LSL #2\n" ".inst 0xf8ac497a // rprfm pldonce, x12, [x11]\n" - ".inst 0xf8ac4a1a // rprfm pldonce, x12, [x16]\n" - ".inst 0xf8ac48fa // rprfm pldonce, x12, [x7]\n" + ".inst 0xf8ac49da // rprfm pldonce, x12, [x14]\n" + ".inst 0xf8ac49fa // rprfm pldonce, x12, [x15]\n" ".inst 0xf8ac495a // rprfm pldonce, x12, [x10]\n" ".inst 0xf8ac493a // rprfm pldonce, x12, [x9]\n" ".inst 0xf8ac4b9a // rprfm pldonce, x12, [x28]\n" ".inst 0xf8ac4b7a // rprfm pldonce, x12, [x27]\n" ".inst 0xf8ac4b5a // rprfm pldonce, x12, [x26]\n" - ".inst 0xf8ac49da // rprfm pldonce, x12, [x14]\n" + ".inst 0xf8ac49ba // rprfm pldonce, x12, [x13]\n" ".inst 0xf8ac4b3a // rprfm pldonce, x12, [x25]\n" ".inst 0xf8ac4b1a // rprfm pldonce, x12, [x24]\n" ".inst 0xf8ac4afa // rprfm pldonce, x12, [x23]\n" @@ -167,199 +167,199 @@ void sme2_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst_direct_impl( ".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, x2, 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, [x17]\n" - "ldr x25, [%x[params_struct], %[offsetof_args_ld_output_col]]\n" - "madd x21, x3, x25, x21\n" // offset += tile_j * ld_output_col - "addvl x17, x17, #1\n" - ".inst 0xa040c220 // ld1w { z0.s-z3.s }, pn8.b/Z, [x17]\n" + "ld1w { z28.s }, p3/Z, [x6]\n" + "addvl x6, x6, #1\n" + "ldr x26, [%x[params_struct], %[offsetof_args_ld_output_col]]\n" + "cntw x25\n" + ".inst 0xa040c0c0 // ld1w { z0.s-z3.s }, pn8.b/Z, [x6]\n" + "addvl x6, x6, #4\n" "ldr x24, [%x[params_struct], %[offsetof_args_outptr]]\n" - "mul x21, x21, x20\n" // offset *= output_tile_size - "cntw x23\n" - "ld1rw { z26.s }, p3/Z, [%x[params_struct], %[offsetof_args_min]]\n" - "addvl x17, x17, #4\n" - "add x24, x24, x21, LSL #2\n" // outptrs[0] += offset * sizeof(float) - ".inst 0xa040c224 // ld1w { z4.s-z7.s }, pn8.b/Z, [x17]\n" "whilelt p2.s, XZR, %x[n_channels]\n" - "addvl x17, x17, #4\n" - "ld1rw { z24.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, [x17]\n" + ".inst 0xa040c0c4 // ld1w { z4.s-z7.s }, pn8.b/Z, [x6]\n" + "addvl x6, x6, #4\n" + "mul x22, x2, x23\n" // offset = tile_i * ld_output_row + "cmp x25, %x[n_channels]\n" + "ld1rw { z30.s }, p3/Z, [%x[params_struct], %[offsetof_args_min]]\n" + "madd x22, x3, x26, x22\n" // offset += tile_j * ld_output_col + "ld1rw { z31.s }, p3/Z, [%x[params_struct], %[offsetof_args_max]]\n" "mov x21, #0x0\n" - "sub x20, XZR, x23\n" - "ld1w { z9.s }, p2/Z, [x7, x8, LSL #2]\n" + "mul x22, x22, x20\n" // offset *= output_tile_size + "sub x20, XZR, x25\n" + "ld1w { z8.s }, p3/Z, [x6]\n" + "add x24, x24, x22, LSL #2\n" // outptrs[0] += offset * sizeof(float) + "ld1w { z9.s }, p2/Z, [x15, x7, LSL #2]\n" + "addvl x6, x6, #1\n" + "add x23, x24, x23, LSL #2\n" "ld1w { z10.s }, p2/Z, [x5]\n" - "addvl x17, x17, #1\n" "ld1w { z11.s }, p2/Z, [x5, x4, LSL #2]\n" - "ld1w { z12.s }, p2/Z, [x5, x15, LSL #2]\n" - "ld1w { z13.s }, p2/Z, [x5, x13, LSL #2]\n" - "ld1w { z14.s }, p2/Z, [x6]\n" - "ld1w { z15.s }, p2/Z, [x6, x4, LSL #2]\n" - "ld1w { z16.s }, p2/Z, [x5, x8, LSL #2]\n" + "ld1w { z12.s }, p2/Z, [x5, x8, LSL #2]\n" + "ld1w { z13.s }, p2/Z, [x5, x17, LSL #2]\n" + "ld1w { z14.s }, p2/Z, [x16]\n" + "ld1w { z15.s }, p2/Z, [x16, x4, LSL #2]\n" + "ld1w { z16.s }, p2/Z, [x5, x7, LSL #2]\n" "bge 4f\n" "3:" // Tile loop: Channel loop - "movprfx z28, z22\n fmla z28.s, p3/M, z8.s, z9.s\n" - "movprfx z29, z22\n fmla z29.s, p3/M, z6.s, z9.s\n" - "whilelt p1.s, x23, %x[n_channels]\n" + "movprfx z24, z28\n fmla z24.s, p3/M, z8.s, z9.s\n" + "movprfx z25, z28\n fmla z25.s, p3/M, z6.s, z9.s\n" + "whilelt p1.s, x25, %x[n_channels]\n" "incw x21\n" - "fmla z28.s, p3/M, z0.s, z10.s\n" - "fmla z29.s, p3/M, z1.s, z12.s\n" - "ld1w { z18.s }, p2/Z, [x6, x13, LSL #2]\n" - "incw x23\n" - "fmla z28.s, p3/M, z1.s, z11.s\n" - "fmla z29.s, p3/M, z2.s, z13.s\n" - "ld1w { z27.s }, p2/Z, [x6, x15, LSL #2]\n" + "movprfx z26, z28\n fmla z26.s, p3/M, z2.s, z9.s\n" + "movprfx z27, z28\n fmla z27.s, p3/M, z0.s, z9.s\n" + "ld1w { z28.s }, p3/Z, [x6]\n" + "addvl x6, x6, #1\n" + "incw x25\n" "mov p0.b, p2.b\n" - "fmla z28.s, p3/M, z3.s, z14.s\n" - "fmla z29.s, p3/M, z0.s, z16.s\n" - "ld1w { z17.s }, p2/Z, [x6, x8, LSL #2]\n" "addvl x5, x5, #1\n" - "fmla z28.s, p3/M, z4.s, z15.s\n" - "fmla z29.s, p3/M, z4.s, z27.s\n" - "ld1w { z25.s }, p2/Z, [x16]\n" - "addvl x6, x6, #1\n" - "fmla z28.s, p3/M, z2.s, z16.s\n" - "fmla z29.s, p3/M, z5.s, z18.s\n" - "ld1w { z12.s }, p2/Z, [x7]\n" "incw x20\n" - "movprfx z30, z22\n fmla z30.s, p3/M, z2.s, z9.s\n" - "movprfx z31, z22\n fmla z31.s, p3/M, z0.s, z9.s\n" - "ld1w { z18.s }, p2/Z, [x7, x15, LSL #2]\n" - "fmla z28.s, p3/M, z5.s, z17.s\n" - "fmla z29.s, p3/M, z3.s, z17.s\n" - "ld1w { z16.s }, p2/Z, [x16, x15, LSL #2]\n" - "fmla z30.s, p3/M, z3.s, z25.s\n" - "fmla z31.s, p3/M, z4.s, z16.s\n" - "ld1w { z10.s }, p2/Z, [x16, x4, LSL #2]\n" - "fmla z30.s, p3/M, z0.s, z12.s\n" - "fmla z31.s, p3/M, z1.s, z18.s\n" - "ld1w { z16.s }, p2/Z, [x16, x13, LSL #2]\n" - "fmla z30.s, p3/M, z4.s, z10.s\n" - "fmla z31.s, p3/M, z5.s, z16.s\n" - "ld1w { z16.s }, p2/Z, [x7, x4, LSL #2]\n" - "fmla z28.s, p3/M, z6.s, z12.s\n" - "ld1w { z22.s }, p2/Z, [x7, x13, LSL #2]\n" - "fmla z30.s, p3/M, z1.s, z16.s\n" - "addvl x7, x7, #1\n" - "fmla z31.s, p3/M, z2.s, z22.s\n" - "fmla z28.s, p3/M, z7.s, z16.s\n" - "ld1w { z16.s }, p2/Z, [x14]\n" - "ld1w { z17.s }, p2/Z, [x16, x8, LSL #2]\n" - "fmla z30.s, p3/M, z6.s, z16.s\n" - "fmla z31.s, p3/M, z3.s, z17.s\n" + "fmla z24.s, p3/M, z0.s, z10.s\n" + "fmla z25.s, p3/M, z1.s, z12.s\n" + "ld1w { z18.s }, p2/Z, [x16, x17, LSL #2]\n" + "ld1w { z10.s }, p1/Z, [x5]\n" + "fmla z24.s, p3/M, z1.s, z11.s\n" + "fmla z25.s, p3/M, z2.s, z13.s\n" + "ld1w { z9.s }, p2/Z, [x16, x8, LSL #2]\n" + "ld1w { z22.s }, p2/Z, [x16, x7, LSL #2]\n" "addvl x16, x16, #1\n" - "ld1w { z16.s }, p2/Z, [x14, x4, LSL #2]\n" - "fmla z30.s, p3/M, z7.s, z16.s\n" - "fmla z29.s, p3/M, z7.s, z18.s\n" - "ld1w { z16.s }, p2/Z, [x14, x15, LSL #2]\n" - "fmla z31.s, p3/M, z7.s, z16.s\n" - "fmla z30.s, p3/M, z5.s, z17.s\n" - "ld1w { z17.s }, p2/Z, [x14, x8, LSL #2]\n" - "fmla z31.s, p3/M, z6.s, z17.s\n" - "fmla z29.s, p3/M, z8.s, z22.s\n" - "ld1w { z16.s }, p2/Z, [x14, x13, LSL #2]\n" - "fmla z30.s, p3/M, z8.s, z17.s\n" - "fmla z31.s, p3/M, z8.s, z16.s\n" - "whilelt p2.s, x21, %x[n_channels]\n" - "ld1w { z22.s }, p3/Z, [x17]\n" - "addvl x17, x17, #1\n" - "cmp x23, %x[n_channels]\n" - ".inst 0xc1b8cb5c // fclamp { z28.s-z31.s }, z26.s, z24.s\n" - ".inst 0xa040c220 // ld1w { z0.s-z3.s }, pn8.b/Z, [x17]\n" - "addvl x17, x17, #4\n" + "fmla z24.s, p3/M, z3.s, z14.s\n" + "fmla z25.s, p3/M, z0.s, z16.s\n" + "ld1w { z17.s }, p2/Z, [x14]\n" + "fmla z26.s, p3/M, z3.s, z17.s\n" + "ld1w { z21.s }, p2/Z, [x14, x17, LSL #2]\n" + "fmla z24.s, p3/M, z4.s, z15.s\n" + "fmla z25.s, p3/M, z4.s, z9.s\n" + "ld1w { z19.s }, p2/Z, [x15]\n" + "ld1w { z17.s }, p2/Z, [x14, x4, LSL #2]\n" + "fmla z26.s, p3/M, z0.s, z19.s\n" + "fmla z24.s, p3/M, z2.s, z16.s\n" + "fmla z25.s, p3/M, z5.s, z18.s\n" + "ld1w { z18.s }, p2/Z, [x15, x8, LSL #2]\n" + "ld1w { z0.s }, p2/Z, [x15, x4, LSL #2]\n" + "fmla z26.s, p3/M, z4.s, z17.s\n" + "ld1w { z20.s }, p2/Z, [x15, x17, LSL #2]\n" + "addvl x15, x15, #1\n" + "fmla z24.s, p3/M, z5.s, z22.s\n" + "fmla z25.s, p3/M, z3.s, z22.s\n" + "ld1w { z16.s }, p2/Z, [x14, x8, LSL #2]\n" + "ld1w { z9.s }, p1/Z, [x15, x7, LSL #2]\n" + "fmla z27.s, p3/M, z4.s, z16.s\n" + "fmla z26.s, p3/M, z1.s, z0.s\n" + "ld1w { z17.s }, p2/Z, [x13, x4, LSL #2]\n" + "fmla z24.s, p3/M, z6.s, z19.s\n" + "ld1w { z16.s }, p2/Z, [x13]\n" + "fmla z25.s, p3/M, z7.s, z18.s\n" + "fmla z27.s, p3/M, z1.s, z18.s\n" + "fmla z26.s, p3/M, z6.s, z16.s\n" + "ld1w { z19.s }, p2/Z, [x13, x7, LSL #2]\n" + "fmla z24.s, p3/M, z7.s, z0.s\n" + "ld1w { z16.s }, p2/Z, [x14, x7, LSL #2]\n" + "fmla z25.s, p3/M, z8.s, z20.s\n" "addvl x14, x14, #1\n" - "st1w { z28.s }, p0, [x24]\n" - ".inst 0xa040c224 // ld1w { z4.s-z7.s }, pn8.b/Z, [x17]\n" - "addvl x17, x17, #4\n" - "st1w { z29.s }, p0, [x24, x25, LSL #2]\n" - "addvl x24, x24, #1\n" - "ld1w { z9.s }, p1/Z, [x7, x8, LSL #2]\n" - "st1w { z30.s }, p0, [x22]\n" - "ld1w { z10.s }, p1/Z, [x5]\n" - "st1w { z31.s }, p0, [x22, x25, LSL #2]\n" - "addvl x22, x22, #1\n" + "ld1w { z12.s }, p1/Z, [x5, x8, LSL #2]\n" + "fmla z27.s, p3/M, z5.s, z21.s\n" + "ld1w { z18.s }, p2/Z, [x13, x8, LSL #2]\n" + "fmla z26.s, p3/M, z7.s, z17.s\n" + "ld1w { z13.s }, p1/Z, [x5, x17, LSL #2]\n" + "fmla z27.s, p3/M, z2.s, z20.s\n" + "fmla z26.s, p3/M, z5.s, z16.s\n" + "ld1w { z17.s }, p2/Z, [x13, x17, LSL #2]\n" + "whilelt p2.s, x21, %x[n_channels]\n" + "cmp x25, %x[n_channels]\n" + "addvl x13, x13, #1\n" + "fmla z27.s, p3/M, z3.s, z16.s\n" + "fmla z26.s, p3/M, z8.s, z19.s\n" + ".inst 0xa040c0c0 // ld1w { z0.s-z3.s }, pn8.b/Z, [x6]\n" + "addvl x6, x6, #4\n" + "ld1w { z16.s }, p1/Z, [x5, x7, LSL #2]\n" + "fmla z27.s, p3/M, z7.s, z18.s\n" + "ld1w { z14.s }, p1/Z, [x16]\n" + "fmla z27.s, p3/M, z6.s, z19.s\n" + ".inst 0xa040c0c4 // ld1w { z4.s-z7.s }, pn8.b/Z, [x6]\n" + "addvl x6, x6, #4\n" + "ld1w { z15.s }, p1/Z, [x16, x4, LSL #2]\n" + "fmla z27.s, p3/M, z8.s, z17.s\n" "ld1w { z11.s }, p1/Z, [x5, x4, LSL #2]\n" - "ld1w { z12.s }, p1/Z, [x5, x15, LSL #2]\n" - "ld1w { z13.s }, p1/Z, [x5, x13, LSL #2]\n" - "ld1w { z14.s }, p1/Z, [x6]\n" - "ld1w { z15.s }, p1/Z, [x6, x4, LSL #2]\n" - "ld1w { z16.s }, p1/Z, [x5, x8, LSL #2]\n" - "ld1w { z8.s }, p3/Z, [x17]\n" - "addvl x17, x17, #1\n" + "ld1w { z8.s }, p3/Z, [x6]\n" + "addvl x6, x6, #1\n" + ".inst 0xc1bfcbd8 // fclamp { z24.s-z27.s }, z30.s, z31.s\n" + "st1w { z24.s }, p0, [x24]\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, z8.s, z9.s\n" - "movprfx z29, z22\n fmla z29.s, p3/M, z6.s, z9.s\n" + "movprfx z24, z28\n fmla z24.s, p3/M, z8.s, z9.s\n" + "movprfx z25, z28\n fmla z25.s, p3/M, z6.s, z9.s\n" "ldr x3, [%x[params_struct], %[offsetof_args_tile_j]]\n" - "add x3, x3, #0x1\n" - "fmla z28.s, p3/M, z0.s, z10.s\n" - "fmla z29.s, p3/M, z1.s, z12.s\n" - "ld1w { z18.s }, p2/Z, [x6, x13, LSL #2]\n" + "mov p0.b, p2.b\n" + "movprfx z26, z28\n fmla z26.s, p3/M, z2.s, z9.s\n" + "movprfx z27, z28\n fmla z27.s, p3/M, z0.s, z9.s\n" "ldr x2, [%x[params_struct], %[offsetof_args_tile_i]]\n" - "fmla z28.s, p3/M, z1.s, z11.s\n" - "fmla z29.s, p3/M, z2.s, z13.s\n" - "ld1w { z17.s }, p2/Z, [x6, x15, LSL #2]\n" - "ldr x20, [%x[params_struct], %[offsetof_args_n_tile_cols]]\n" - "fmla z28.s, p3/M, z3.s, z14.s\n" - "fmla z29.s, p3/M, z0.s, z16.s\n" - "ld1w { z20.s }, p2/Z, [x6, x8, LSL #2]\n" + "ldr x22, [%x[params_struct], %[offsetof_args_n_tile_cols]]\n" "ldr x21, [%x[params_struct], %[offsetof_args_n_tile_rows]]\n" - "fmla z28.s, p3/M, z4.s, z15.s\n" - "fmla z29.s, p3/M, z4.s, z17.s\n" - "ld1w { z17.s }, p2/Z, [x16]\n" - "cmp x3, x20\n" - "fmla z28.s, p3/M, z2.s, z16.s\n" - "fmla z29.s, p3/M, z5.s, z18.s\n" - "ld1w { z18.s }, p2/Z, [x7]\n" + "add x3, x3, #0x1\n" + "fmla z24.s, p3/M, z0.s, z10.s\n" + "fmla z25.s, p3/M, z1.s, z12.s\n" + "ld1w { z21.s }, p2/Z, [x16, x17, LSL #2]\n" "add x20, x2, #0x1\n" - "movprfx z30, z22\n fmla z30.s, p3/M, z2.s, z9.s\n" - "movprfx z31, z22\n fmla z31.s, p3/M, z0.s, z9.s\n" - "ld1w { z19.s }, p2/Z, [x7, x15, LSL #2]\n" + "cmp x3, x22\n" "csel x2, x2, x20, LT\n" - "fmla z28.s, p3/M, z5.s, z20.s\n" - "fmla z29.s, p3/M, z3.s, z20.s\n" - "ld1w { z16.s }, p2/Z, [x16, x15, LSL #2]\n" - "mov p0.b, p2.b\n" - "fmla z30.s, p3/M, z3.s, z17.s\n" - "fmla z31.s, p3/M, z4.s, z16.s\n" - "ld1w { z17.s }, p2/Z, [x16, x4, LSL #2]\n" "csel x3, x3, XZR, LT\n" - "fmla z30.s, p3/M, z0.s, z18.s\n" - "fmla z31.s, p3/M, z1.s, z19.s\n" - "ld1w { z16.s }, p2/Z, [x16, x13, LSL #2]\n" "cmp x2, x21\n" - "fmla z30.s, p3/M, z4.s, z17.s\n" - "fmla z31.s, p3/M, z5.s, z16.s\n" - "ld1w { z16.s }, p2/Z, [x7, x4, LSL #2]\n" - "fmla z28.s, p3/M, z6.s, z18.s\n" - "ld1w { z18.s }, p2/Z, [x7, x13, LSL #2]\n" - "fmla z30.s, p3/M, z1.s, z16.s\n" - "fmla z31.s, p3/M, z2.s, z18.s\n" - "fmla z28.s, p3/M, z7.s, z16.s\n" - "ld1w { z16.s }, p2/Z, [x14]\n" - "ld1w { z17.s }, p2/Z, [x16, x8, LSL #2]\n" - "fmla z30.s, p3/M, z6.s, z16.s\n" - "fmla z31.s, p3/M, z3.s, z17.s\n" - "ld1w { z16.s }, p2/Z, [x14, x4, LSL #2]\n" - "fmla z30.s, p3/M, z7.s, z16.s\n" - "fmla z29.s, p3/M, z7.s, z19.s\n" - "ld1w { z16.s }, p2/Z, [x14, x15, LSL #2]\n" - "fmla z31.s, p3/M, z7.s, z16.s\n" - "fmla z30.s, p3/M, z5.s, z17.s\n" - "ld1w { z17.s }, p2/Z, [x14, x8, LSL #2]\n" - "fmla z31.s, p3/M, z6.s, z17.s\n" - "fmla z29.s, p3/M, z8.s, z18.s\n" - "ld1w { z16.s }, p2/Z, [x14, x13, LSL #2]\n" - "fmla z30.s, p3/M, z8.s, z17.s\n" - "fmla z31.s, p3/M, z8.s, z16.s\n" - ".inst 0xc1b8cb5c // fclamp { z28.s-z31.s }, z26.s, z24.s\n" - "st1w { z28.s }, p0, [x24]\n" - "st1w { z29.s }, p0, [x24, x25, LSL #2]\n" - "st1w { z30.s }, p0, [x22]\n" - "st1w { z31.s }, p0, [x22, x25, LSL #2]\n" + "fmla z24.s, p3/M, z1.s, z11.s\n" + "fmla z25.s, p3/M, z2.s, z13.s\n" + "ld1w { z18.s }, p2/Z, [x16, x8, LSL #2]\n" + "ld1w { z20.s }, p2/Z, [x16, x7, LSL #2]\n" + "fmla z24.s, p3/M, z3.s, z14.s\n" + "fmla z25.s, p3/M, z0.s, z16.s\n" + "ld1w { z17.s }, p2/Z, [x14]\n" + "fmla z26.s, p3/M, z3.s, z17.s\n" + "ld1w { z23.s }, p2/Z, [x14, x17, LSL #2]\n" + "fmla z24.s, p3/M, z4.s, z15.s\n" + "fmla z25.s, p3/M, z4.s, z18.s\n" + "ld1w { z22.s }, p2/Z, [x15]\n" + "ld1w { z19.s }, p2/Z, [x14, x4, LSL #2]\n" + "fmla z26.s, p3/M, z0.s, z22.s\n" + "fmla z24.s, p3/M, z2.s, z16.s\n" + "fmla z25.s, p3/M, z5.s, z21.s\n" + "ld1w { z18.s }, p2/Z, [x15, x8, LSL #2]\n" + "ld1w { z17.s }, p2/Z, [x15, x4, LSL #2]\n" + "fmla z26.s, p3/M, z4.s, z19.s\n" + "ld1w { z21.s }, p2/Z, [x15, x17, LSL #2]\n" + "fmla z24.s, p3/M, z5.s, z20.s\n" + "fmla z25.s, p3/M, z3.s, z20.s\n" + "ld1w { z16.s }, p2/Z, [x14, x8, LSL #2]\n" + "fmla z27.s, p3/M, z4.s, z16.s\n" + "fmla z26.s, p3/M, z1.s, z17.s\n" + "ld1w { z20.s }, p2/Z, [x13, x4, LSL #2]\n" + "fmla z24.s, p3/M, z6.s, z22.s\n" + "ld1w { z16.s }, p2/Z, [x13]\n" + "fmla z25.s, p3/M, z7.s, z18.s\n" + "fmla z27.s, p3/M, z1.s, z18.s\n" + "fmla z26.s, p3/M, z6.s, z16.s\n" + "ld1w { z19.s }, p2/Z, [x13, x7, LSL #2]\n" + "fmla z24.s, p3/M, z7.s, z17.s\n" + "ld1w { z18.s }, p2/Z, [x14, x7, LSL #2]\n" + "fmla z25.s, p3/M, z8.s, z21.s\n" + "fmla z27.s, p3/M, z5.s, z23.s\n" + "ld1w { z17.s }, p2/Z, [x13, x8, LSL #2]\n" + "fmla z26.s, p3/M, z7.s, z20.s\n" + "fmla z27.s, p3/M, z2.s, z21.s\n" + "fmla z26.s, p3/M, z5.s, z18.s\n" + "ld1w { z16.s }, p2/Z, [x13, x17, LSL #2]\n" + "fmla z27.s, p3/M, z3.s, z18.s\n" + "fmla z26.s, p3/M, z8.s, z19.s\n" + "fmla z27.s, p3/M, z7.s, z17.s\n" + "fmla z27.s, p3/M, z6.s, z19.s\n" + "fmla z27.s, p3/M, z8.s, z16.s\n" + ".inst 0xc1bfcbd8 // fclamp { z24.s-z27.s }, z30.s, z31.s\n" + "st1w { z24.s }, p0, [x24]\n" + "st1w { z25.s }, p0, [x24, x26, LSL #2]\n" + "st1w { z26.s }, p0, [x23]\n" + "st1w { z27.s }, p0, [x23, x26, LSL #2]\n" "blt 1b\n" ".inst 0xd503467f // SMSTOP\n" : |