aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp
diff options
context:
space:
mode:
authorMichael Tyler <michael.tyler@arm.com>2023-02-01 16:37:07 +0000
committermichael.tyler <michael.tyler@arm.com>2023-02-08 15:33:26 +0000
commit7d9a626aaba9837cb82d189a9c4f0bcef58825bb (patch)
treee3d8cfeb7f8539cca3a8bf2f1a8f412d25d89041 /src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp
parent4e2bbbbb23e6f4bd452f7f865e51228e1f51efec (diff)
downloadComputeLibrary-7d9a626aaba9837cb82d189a9c4f0bcef58825bb.tar.gz
Update CPU kernels to remove x19 and w19
Resolves: COMPMID-5805 Change-Id: Idf720bbb136474810086f5089c5ed23b3f79835a Signed-off-by: Michael Tyler <michael.tyler@arm.com> Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9081 Benchmark: Arm Jenkins <bsgcomp@arm.com> Tested-by: Arm Jenkins <bsgcomp@arm.com> Comments-Addressed: Arm Jenkins <bsgcomp@arm.com> Reviewed-by: Gunes Bayir <gunes.bayir@arm.com> Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Diffstat (limited to 'src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp')
-rw-r--r--src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp1050
1 files changed, 525 insertions, 525 deletions
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp
index 84b4b3b72b..cdf77a1cf0 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2021 Arm Limited.
+ * Copyright (c) 2021, 2023 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -88,565 +88,565 @@ void sve_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst_direct_impl(
__asm__ __volatile__(
"ptrue p3.b\n"
- "mov x2, #0x0\n"
- "mov x3, #0x0\n"
+ "mov x16, #0x0\n"
+ "mov x4, #0x0\n"
"1:" // Tile loop
- "str x2, [%x[params_struct], %[offsetof_args_tile_i]]\n"
+ "str x16, [%x[params_struct], %[offsetof_args_tile_i]]\n"
+ "mov x25, #0x4\n"
"mov x24, #0x4\n"
- "str x3, [%x[params_struct], %[offsetof_args_tile_j]]\n"
- "mov x23, #0x4\n"
- "ldr x4, [%x[params_struct], %[offsetof_args_params]]\n"
- "mov x5, #0x0\n"
- "ldr x22, [%x[params_struct], %[offsetof_args_ld_input_row]]\n"
- "cntw x6\n"
- "ldr x7, [%x[params_struct], %[offsetof_args_ld_input_col]]\n"
- "sub x21, XZR, x6\n"
+ "str x4, [%x[params_struct], %[offsetof_args_tile_j]]\n"
+ "ldr x23, [%x[params_struct], %[offsetof_args_ld_input_row]]\n"
+ "ldr x22, [%x[params_struct], %[offsetof_args_ld_output_row]]\n"
+ "mul x21, x16, x23\n" // offset = tile_i * ld_input_row
+ "ldr x5, [%x[params_struct], %[offsetof_args_ld_input_col]]\n"
+ "ldr x6, [%x[params_struct], %[offsetof_args_ld_output_col]]\n"
+ "mul x20, x16, x22\n" // offset = tile_i * ld_output_row
+ "add x7, x5, x5\n"
+ "madd x21, x4, x5, x21\n" // offset += tile_j * ld_input_col
"ldr x8, [%x[params_struct], %[offsetof_args_inptr]]\n"
- "mul x19, x2, x22\n" // offset = tile_i * ld_input_row
- "ldr x20, [%x[params_struct], %[offsetof_args_ld_output_row]]\n"
- "madd x19, x3, x7, x19\n" // offset += tile_j * ld_input_col
- "ldr x17, [%x[params_struct], %[offsetof_args_ld_output_col]]\n"
- "mul x19, x19, x24\n" // offset *= kernel_stride * output_size
- "ldr x16, [%x[params_struct], %[offsetof_args_outptr]]\n"
- "add x8, x8, x19, LSL #2\n" // inptr[0] += offset * sizeof(float)
- "ld1rw { z15.s }, p3/Z, [%x[params_struct], %[offsetof_args_min]]\n"
- "add x15, x8, x22, LSL #2\n"
- "ld1rw { z14.s }, p3/Z, [%x[params_struct], %[offsetof_args_max]]\n"
- "add x14, x15, x22, LSL #2\n"
- "ld1w { z13.s }, p3/Z, [x4]\n"
- "add x13, x14, x22, LSL #2\n"
- "ld1w { z0.s }, p3/Z, [x4, #1, MUL VL]\n"
- "add x12, x13, x22, LSL #2\n"
- "ld1w { z1.s }, p3/Z, [x4, #2, MUL VL]\n"
- "add x11, x12, x22, LSL #2\n"
- "ld1w { z2.s }, p3/Z, [x4, #3, MUL VL]\n"
- "add x10, x7, x7\n"
- "ld1w { z3.s }, p3/Z, [x4, #4, MUL VL]\n"
- "add x9, x10, x7\n"
- "ld1w { z4.s }, p3/Z, [x4, #5, MUL VL]\n"
- "add x28, x9, x7\n"
- "ld1w { z5.s }, p3/Z, [x4, #6, MUL VL]\n"
- "add x27, x28, x7\n"
- "ld1w { z6.s }, p3/Z, [x4, #7, MUL VL]\n"
- "mul x19, x2, x20\n" // offset = tile_i * ld_output_row
- "add x26, x17, x17\n"
- "madd x19, x3, x17, x19\n" // offset += tile_j * ld_output_col
- "mul x19, x19, x23\n" // offset *= output_tile_size
- "add x16, x16, x19, LSL #2\n" // outptrs[0] += offset * sizeof(float)
- "add x25, x16, x20, LSL #2\n"
- "add x24, x25, x20, LSL #2\n"
- "add x23, x24, x20, LSL #2\n"
- "add x22, x26, x17\n"
+ "ldr x17, [%x[params_struct], %[offsetof_args_params]]\n"
+ "cntw x16\n"
+ "madd x20, x4, x6, x20\n" // offset += tile_j * ld_output_col
+ "ldr x15, [%x[params_struct], %[offsetof_args_outptr]]\n"
+ "add x14, x7, x5\n"
"whilelt p2.s, XZR, %x[n_channels]\n"
- "ld1w { z9.s }, p2/Z, [x14, x10, LSL #2]\n"
+ "mul x21, x21, x25\n" // offset *= kernel_stride * output_size
+ "add x8, x8, x21, LSL #2\n" // inptr[0] += offset * sizeof(float)
+ "add x13, x8, x23, LSL #2\n"
+ "ld1w { z15.s }, p3/Z, [x17]\n"
+ "mul x20, x20, x24\n" // offset *= output_tile_size
+ "add x12, x13, x23, LSL #2\n"
+ "add x15, x15, x20, LSL #2\n" // outptrs[0] += offset * sizeof(float)
+ "ld1w { z0.s }, p3/Z, [x17, #1, MUL VL]\n"
+ "ld1w { z1.s }, p3/Z, [x17, #2, MUL VL]\n"
+ "ld1w { z2.s }, p3/Z, [x17, #3, MUL VL]\n"
+ "add x11, x12, x23, LSL #2\n"
+ "add x10, x14, x5\n"
+ "ld1w { z3.s }, p3/Z, [x17, #4, MUL VL]\n"
+ "ld1w { z4.s }, p3/Z, [x17, #5, MUL VL]\n"
+ "add x9, x15, x22, LSL #2\n"
+ "add x28, x11, x23, LSL #2\n"
+ "ld1w { z5.s }, p3/Z, [x17, #6, MUL VL]\n"
+ "ld1w { z6.s }, p3/Z, [x17, #7, MUL VL]\n"
+ "addvl x17, x17, #16\n"
+ "add x27, x10, x5\n"
+ "add x26, x9, x22, LSL #2\n"
+ "add x25, x6, x6\n"
+ "ld1rw { z14.s }, p3/Z, [%x[params_struct], %[offsetof_args_min]]\n"
+ "ld1rw { z13.s }, p3/Z, [%x[params_struct], %[offsetof_args_max]]\n"
+ "cmp x16, %x[n_channels]\n"
+ "add x24, x28, x23, LSL #2\n"
+ "ld1w { z7.s }, p3/Z, [x17, #-8, MUL VL]\n"
+ "ld1w { z8.s }, p3/Z, [x17, #-7, MUL VL]\n"
+ "add x23, x26, x22, LSL #2\n"
+ "add x22, x25, x6\n"
+ "ld1w { z9.s }, p2/Z, [x12, x7, LSL #2]\n"
"ld1w { z10.s }, p2/Z, [x8]\n"
- "addvl x4, x4, #16\n"
+ "mov x21, #0x0\n"
+ "sub x20, XZR, x16\n"
"ld1w { z11.s }, p2/Z, [x8, x27, LSL #2]\n"
- "cmp x6, %x[n_channels]\n"
- "ld1w { z7.s }, p3/Z, [x4, #-8, MUL VL]\n"
- "ld1w { z8.s }, p3/Z, [x4, #-7, MUL VL]\n"
- "addvl x4, x4, #-6\n"
- "ld1w { z12.s }, p2/Z, [x14, x9, LSL #2]\n"
+ "ld1w { z12.s }, p2/Z, [x12, x14, LSL #2]\n"
+ "addvl x17, x17, #-6\n"
"bge 3f\n"
"2:" // Tile loop: Channel loop
- "movprfx z31, z13\n fmla z31.s, p3/M, z8.s, z9.s\n"
- "whilelt p1.s, x6, %x[n_channels]\n"
- "movprfx z30, z13\n fmla z30.s, p3/M, z7.s, z9.s\n"
+ "movprfx z21, z15\n fmla z21.s, p3/M, z4.s, z9.s\n"
+ "movprfx z16, z15\n fmla z16.s, p3/M, z8.s, z9.s\n"
+ "whilelt p1.s, x16, %x[n_channels]\n"
"incw x21\n"
- "movprfx z29, z13\n fmla z29.s, p3/M, z6.s, z9.s\n"
+ "movprfx z22, z15\n fmla z22.s, p3/M, z3.s, z9.s\n"
+ "movprfx z25, z15\n fmla z25.s, p3/M, z1.s, z9.s\n"
+ "incw x16\n"
"mov p0.b, p2.b\n"
- "movprfx z27, z13\n fmla z27.s, p3/M, z5.s, z9.s\n"
- "incw x5\n"
- "movprfx z26, z13\n fmla z26.s, p3/M, z4.s, z9.s\n"
- "incw x6\n"
- "movprfx z25, z13\n fmla z25.s, p3/M, z3.s, z9.s\n"
- "movprfx z23, z13\n fmla z23.s, p3/M, z2.s, z9.s\n"
- "movprfx z22, z13\n fmla z22.s, p3/M, z1.s, z9.s\n"
- "movprfx z21, z13\n fmla z21.s, p3/M, z0.s, z9.s\n"
- "ld1w { z9.s }, p2/Z, [x13, x10, LSL #2]\n"
- "fmla z31.s, p3/M, z0.s, z10.s\n"
- "ld1w { z10.s }, p2/Z, [x11]\n"
- "movprfx z28, z13\n fmla z28.s, p3/M, z2.s, z11.s\n"
- "ld1w { z11.s }, p2/Z, [x11, x27, LSL #2]\n"
- "fmla z30.s, p3/M, z8.s, z12.s\n"
- "fmla z29.s, p3/M, z7.s, z12.s\n"
- "fmla z26.s, p3/M, z5.s, z12.s\n"
- "fmla z28.s, p3/M, z6.s, z12.s\n"
- "fmla z25.s, p3/M, z4.s, z12.s\n"
- "movprfx z24, z13\n fmla z24.s, p3/M, z3.s, z12.s\n"
- "fmla z22.s, p3/M, z2.s, z12.s\n"
- "fmla z21.s, p3/M, z1.s, z12.s\n"
- "movprfx z20, z13\n fmla z20.s, p3/M, z0.s, z12.s\n"
- "ld1w { z12.s }, p2/Z, [x8, x7, LSL #2]\n"
- "movprfx z19, z13\n fmla z19.s, p3/M, z6.s, z10.s\n"
- "ld1w { z10.s }, p2/Z, [x13, x9, LSL #2]\n"
- "movprfx z16, z13\n fmla z16.s, p3/M, z8.s, z11.s\n"
- "ld1w { z11.s }, p2/Z, [x8, x28, LSL #2]\n"
- "fmla z27.s, p3/M, z8.s, z9.s\n"
- "fmla z26.s, p3/M, z7.s, z9.s\n"
- "fmla z25.s, p3/M, z6.s, z9.s\n"
- "fmla z23.s, p3/M, z5.s, z9.s\n"
- "fmla z22.s, p3/M, z4.s, z9.s\n"
- "fmla z21.s, p3/M, z3.s, z9.s\n"
- "fmla z19.s, p3/M, z2.s, z9.s\n"
- "movprfx z18, z13\n fmla z18.s, p3/M, z1.s, z9.s\n"
- "movprfx z17, z13\n fmla z17.s, p3/M, z0.s, z9.s\n"
- "ld1w { z9.s }, p2/Z, [x15]\n"
- "fmla z31.s, p3/M, z1.s, z12.s\n"
- "ld1w { z13.s }, p3/Z, [x4]\n"
- "fmla z30.s, p3/M, z0.s, z12.s\n"
- "ld1w { z12.s }, p2/Z, [x15, x27, LSL #2]\n"
- "fmla z29.s, p3/M, z2.s, z11.s\n"
- "fmla z28.s, p3/M, z1.s, z11.s\n"
- "ld1w { z11.s }, p2/Z, [x12]\n"
- "fmla z26.s, p3/M, z8.s, z10.s\n"
- "fmla z25.s, p3/M, z7.s, z10.s\n"
- "fmla z24.s, p3/M, z6.s, z10.s\n"
- "fmla z22.s, p3/M, z5.s, z10.s\n"
- "fmla z21.s, p3/M, z4.s, z10.s\n"
- "fmla z20.s, p3/M, z3.s, z10.s\n"
- "fmla z18.s, p3/M, z2.s, z10.s\n"
- "fmla z17.s, p3/M, z1.s, z10.s\n"
+ "movprfx z26, z15\n fmla z26.s, p3/M, z0.s, z9.s\n"
+ "fmla z21.s, p3/M, z5.s, z12.s\n"
+ "incw x20\n"
+ "movprfx z17, z15\n fmla z17.s, p3/M, z7.s, z9.s\n"
+ "movprfx z18, z15\n fmla z18.s, p3/M, z6.s, z9.s\n"
+ "movprfx z20, z15\n fmla z20.s, p3/M, z5.s, z9.s\n"
+ "movprfx z24, z15\n fmla z24.s, p3/M, z2.s, z9.s\n"
+ "ld1w { z9.s }, p2/Z, [x11, x7, LSL #2]\n"
"fmla z16.s, p3/M, z0.s, z10.s\n"
- "ld1w { z10.s }, p2/Z, [x15, x10, LSL #2]\n"
- "fmla z31.s, p3/M, z3.s, z9.s\n"
- "fmla z27.s, p3/M, z0.s, z9.s\n"
- "fmla z28.s, p3/M, z5.s, z12.s\n"
- "fmla z24.s, p3/M, z2.s, z12.s\n"
- "ld1w { z12.s }, p2/Z, [x15, x9, LSL #2]\n"
- "fmla z23.s, p3/M, z6.s, z11.s\n"
- "fmla z19.s, p3/M, z3.s, z11.s\n"
- "ld1w { z11.s }, p2/Z, [x12, x27, LSL #2]\n"
- "fmla z31.s, p3/M, z5.s, z10.s\n"
- "fmla z30.s, p3/M, z4.s, z10.s\n"
- "fmla z29.s, p3/M, z3.s, z10.s\n"
- "fmla z27.s, p3/M, z2.s, z10.s\n"
- "fmla z26.s, p3/M, z1.s, z10.s\n"
- "fmla z25.s, p3/M, z0.s, z10.s\n"
- "ld1w { z10.s }, p2/Z, [x14, x7, LSL #2]\n"
- "fmla z20.s, p3/M, z8.s, z11.s\n"
- "fmla z16.s, p3/M, z5.s, z11.s\n"
- "ld1w { z11.s }, p2/Z, [x11, x7, LSL #2]\n"
- "fmla z30.s, p3/M, z5.s, z12.s\n"
- "fmla z29.s, p3/M, z4.s, z12.s\n"
- "fmla z28.s, p3/M, z3.s, z12.s\n"
- "fmla z26.s, p3/M, z2.s, z12.s\n"
- "fmla z25.s, p3/M, z1.s, z12.s\n"
- "fmla z24.s, p3/M, z0.s, z12.s\n"
- "ld1w { z12.s }, p2/Z, [x14, x28, LSL #2]\n"
- "fmla z19.s, p3/M, z7.s, z11.s\n"
- "fmla z18.s, p3/M, z6.s, z11.s\n"
- "ld1w { z11.s }, p2/Z, [x11, x28, LSL #2]\n"
- "fmla z31.s, p3/M, z7.s, z10.s\n"
- "fmla z30.s, p3/M, z6.s, z10.s\n"
- "fmla z27.s, p3/M, z4.s, z10.s\n"
- "fmla z26.s, p3/M, z3.s, z10.s\n"
- "fmla z23.s, p3/M, z1.s, z10.s\n"
+ "movprfx z19, z15\n fmla z19.s, p3/M, z2.s, z11.s\n"
+ "ld1w { z10.s }, p2/Z, [x24]\n"
+ "ld1w { z11.s }, p2/Z, [x24, x27, LSL #2]\n"
+ "fmla z22.s, p3/M, z4.s, z12.s\n"
+ "fmla z25.s, p3/M, z2.s, z12.s\n"
+ "fmla z26.s, p3/M, z1.s, z12.s\n"
+ "movprfx z28, z15\n fmla z28.s, p3/M, z6.s, z10.s\n"
+ "ld1w { z10.s }, p2/Z, [x11, x14, LSL #2]\n"
+ "fmla z21.s, p3/M, z7.s, z9.s\n"
+ "fmla z17.s, p3/M, z8.s, z12.s\n"
+ "fmla z18.s, p3/M, z7.s, z12.s\n"
+ "fmla z19.s, p3/M, z6.s, z12.s\n"
+ "movprfx z23, z15\n fmla z23.s, p3/M, z3.s, z12.s\n"
+ "movprfx z27, z15\n fmla z27.s, p3/M, z0.s, z12.s\n"
+ "ld1w { z12.s }, p2/Z, [x8, x5, LSL #2]\n"
+ "movprfx z31, z15\n fmla z31.s, p3/M, z8.s, z11.s\n"
+ "fmla z22.s, p3/M, z6.s, z9.s\n"
+ "ld1w { z11.s }, p2/Z, [x8, x10, LSL #2]\n"
+ "fmla z25.s, p3/M, z4.s, z9.s\n"
+ "fmla z26.s, p3/M, z3.s, z9.s\n"
+ "fmla z20.s, p3/M, z8.s, z9.s\n"
+ "fmla z24.s, p3/M, z5.s, z9.s\n"
+ "fmla z28.s, p3/M, z2.s, z9.s\n"
+ "fmla z21.s, p3/M, z8.s, z10.s\n"
+ "fmla z16.s, p3/M, z1.s, z12.s\n"
+ "fmla z17.s, p3/M, z0.s, z12.s\n"
+ "ld1w { z12.s }, p2/Z, [x13, x27, LSL #2]\n"
+ "fmla z18.s, p3/M, z2.s, z11.s\n"
+ "fmla z19.s, p3/M, z1.s, z11.s\n"
+ "ld1w { z11.s }, p2/Z, [x28]\n"
+ "fmla z22.s, p3/M, z7.s, z10.s\n"
+ "fmla z23.s, p3/M, z6.s, z10.s\n"
+ "fmla z25.s, p3/M, z5.s, z10.s\n"
+ "fmla z26.s, p3/M, z4.s, z10.s\n"
+ "fmla z27.s, p3/M, z3.s, z10.s\n"
+ "fmla z31.s, p3/M, z0.s, z10.s\n"
+ "fmla z24.s, p3/M, z6.s, z11.s\n"
+ "fmla z28.s, p3/M, z3.s, z11.s\n"
+ "ld1w { z11.s }, p2/Z, [x28, x27, LSL #2]\n"
+ "fmla z19.s, p3/M, z5.s, z12.s\n"
+ "fmla z23.s, p3/M, z2.s, z12.s\n"
+ "ld1w { z12.s }, p2/Z, [x13, x14, LSL #2]\n"
+ "fmla z27.s, p3/M, z8.s, z11.s\n"
+ "fmla z31.s, p3/M, z5.s, z11.s\n"
+ "movprfx z29, z15\n fmla z29.s, p3/M, z1.s, z9.s\n"
+ "movprfx z30, z15\n fmla z30.s, p3/M, z0.s, z9.s\n"
+ "ld1w { z9.s }, p2/Z, [x13]\n"
+ "fmla z29.s, p3/M, z2.s, z10.s\n"
+ "fmla z30.s, p3/M, z1.s, z10.s\n"
+ "ld1w { z10.s }, p2/Z, [x13, x7, LSL #2]\n"
+ "fmla z20.s, p3/M, z0.s, z9.s\n"
+ "fmla z21.s, p3/M, z1.s, z10.s\n"
+ "fmla z16.s, p3/M, z3.s, z9.s\n"
+ "fmla z17.s, p3/M, z4.s, z10.s\n"
+ "ld1w { z11.s }, p2/Z, [x24, x5, LSL #2]\n"
+ "fmla z18.s, p3/M, z3.s, z10.s\n"
"fmla z22.s, p3/M, z0.s, z10.s\n"
- "ld1w { z10.s }, p2/Z, [x8, x10, LSL #2]\n"
- "fmla z17.s, p3/M, z8.s, z11.s\n"
- "fmla z16.s, p3/M, z7.s, z11.s\n"
- "ld1w { z11.s }, p2/Z, [x13, x7, LSL #2]\n"
- "fmla z29.s, p3/M, z8.s, z12.s\n"
- "fmla z28.s, p3/M, z7.s, z12.s\n"
- "fmla z25.s, p3/M, z5.s, z12.s\n"
- "fmla z24.s, p3/M, z4.s, z12.s\n"
+ "fmla z20.s, p3/M, z2.s, z10.s\n"
"fmla z21.s, p3/M, z2.s, z12.s\n"
- "fmla z20.s, p3/M, z1.s, z12.s\n"
- "ld1w { z12.s }, p2/Z, [x8, x9, LSL #2]\n"
+ "fmla z16.s, p3/M, z5.s, z10.s\n"
+ "fmla z17.s, p3/M, z5.s, z12.s\n"
+ "ld1w { z10.s }, p2/Z, [x12, x5, LSL #2]\n"
+ "fmla z18.s, p3/M, z4.s, z12.s\n"
+ "fmla z19.s, p3/M, z3.s, z12.s\n"
+ "fmla z22.s, p3/M, z1.s, z12.s\n"
+ "fmla z23.s, p3/M, z0.s, z12.s\n"
+ "ld1w { z12.s }, p2/Z, [x12, x10, LSL #2]\n"
+ "fmla z28.s, p3/M, z7.s, z11.s\n"
+ "fmla z29.s, p3/M, z6.s, z11.s\n"
+ "ld1w { z11.s }, p2/Z, [x24, x10, LSL #2]\n"
+ "fmla z20.s, p3/M, z4.s, z10.s\n"
+ "fmla z21.s, p3/M, z3.s, z10.s\n"
+ "fmla z24.s, p3/M, z1.s, z10.s\n"
+ "fmla z25.s, p3/M, z0.s, z10.s\n"
+ "fmla z16.s, p3/M, z7.s, z10.s\n"
+ "fmla z17.s, p3/M, z6.s, z10.s\n"
+ "ld1w { z10.s }, p2/Z, [x8, x7, LSL #2]\n"
+ "fmla z30.s, p3/M, z8.s, z11.s\n"
+ "fmla z31.s, p3/M, z7.s, z11.s\n"
+ "ld1w { z11.s }, p2/Z, [x11, x5, LSL #2]\n"
+ "fmla z18.s, p3/M, z8.s, z12.s\n"
+ "fmla z19.s, p3/M, z7.s, z12.s\n"
+ "fmla z22.s, p3/M, z5.s, z12.s\n"
+ "fmla z23.s, p3/M, z4.s, z12.s\n"
+ "fmla z26.s, p3/M, z2.s, z12.s\n"
+ "fmla z27.s, p3/M, z1.s, z12.s\n"
+ "ld1w { z12.s }, p2/Z, [x8, x14, LSL #2]\n"
"addvl x8, x8, #1\n"
- "fmla z31.s, p3/M, z2.s, z10.s\n"
- "fmla z30.s, p3/M, z1.s, z10.s\n"
- "fmla z29.s, p3/M, z0.s, z10.s\n"
- "ld1w { z10.s }, p2/Z, [x14]\n"
- "fmla z27.s, p3/M, z7.s, z11.s\n"
+ "fmla z20.s, p3/M, z7.s, z11.s\n"
+ "fmla z21.s, p3/M, z6.s, z11.s\n"
+ "fmla z24.s, p3/M, z4.s, z11.s\n"
+ "fmla z25.s, p3/M, z3.s, z11.s\n"
+ "fmla z28.s, p3/M, z1.s, z11.s\n"
+ "fmla z29.s, p3/M, z0.s, z11.s\n"
+ "ld1w { z11.s }, p2/Z, [x11, x10, LSL #2]\n"
+ "fmla z16.s, p3/M, z2.s, z10.s\n"
+ "fmla z17.s, p3/M, z1.s, z10.s\n"
+ "fmla z18.s, p3/M, z0.s, z10.s\n"
+ "ld1w { z10.s }, p2/Z, [x12]\n"
+ "fmla z30.s, p3/M, z2.s, z11.s\n"
+ "fmla z19.s, p3/M, z0.s, z12.s\n"
+ "fmla z20.s, p3/M, z3.s, z10.s\n"
+ "fmla z24.s, p3/M, z0.s, z10.s\n"
+ "fmla z22.s, p3/M, z8.s, z11.s\n"
+ "fmla z23.s, p3/M, z7.s, z11.s\n"
+ "fmla z26.s, p3/M, z5.s, z11.s\n"
+ "fmla z27.s, p3/M, z4.s, z11.s\n"
+ "fmla z31.s, p3/M, z1.s, z11.s\n"
+ "ld1w { z11.s }, p2/Z, [x28, x7, LSL #2]\n"
+ "fmla z17.s, p3/M, z2.s, z12.s\n"
+ "fmla z18.s, p3/M, z1.s, z12.s\n"
+ "ld1w { z12.s }, p2/Z, [x12, x27, LSL #2]\n"
+ "addvl x12, x12, #1\n"
+ "fmla z16.s, p3/M, z6.s, z10.s\n"
+ "ld1w { z10.s }, p2/Z, [x11]\n"
+ "fmla z29.s, p3/M, z4.s, z11.s\n"
+ "fmla z30.s, p3/M, z3.s, z11.s\n"
+ "fmla z19.s, p3/M, z8.s, z12.s\n"
+ "fmla z23.s, p3/M, z5.s, z12.s\n"
+ "fmla z27.s, p3/M, z2.s, z12.s\n"
+ "ld1w { z12.s }, p2/Z, [x11, x27, LSL #2]\n"
+ "addvl x11, x11, #1\n"
+ "fmla z20.s, p3/M, z6.s, z10.s\n"
+ "fmla z24.s, p3/M, z3.s, z10.s\n"
+ "fmla z28.s, p3/M, z0.s, z10.s\n"
+ "ld1w { z10.s }, p2/Z, [x24, x7, LSL #2]\n"
+ "fmla z31.s, p3/M, z2.s, z12.s\n"
+ "fmla z29.s, p3/M, z7.s, z10.s\n"
+ "fmla z30.s, p3/M, z6.s, z10.s\n"
+ "fmla z24.s, p3/M, z8.s, z11.s\n"
+ "fmla z25.s, p3/M, z7.s, z11.s\n"
"fmla z26.s, p3/M, z6.s, z11.s\n"
- "fmla z23.s, p3/M, z4.s, z11.s\n"
- "fmla z22.s, p3/M, z3.s, z11.s\n"
- "fmla z19.s, p3/M, z1.s, z11.s\n"
- "fmla z18.s, p3/M, z0.s, z11.s\n"
- "ld1w { z11.s }, p2/Z, [x13, x28, LSL #2]\n"
- "fmla z30.s, p3/M, z2.s, z12.s\n"
- "fmla z29.s, p3/M, z1.s, z12.s\n"
- "fmla z28.s, p3/M, z0.s, z12.s\n"
- "ld1w { z12.s }, p2/Z, [x14, x27, LSL #2]\n"
- "addvl x14, x14, #1\n"
- "fmla z31.s, p3/M, z6.s, z10.s\n"
- "ld1w { z9.s }, p1/Z, [x14, x10, LSL #2]\n"
- "fmla z27.s, p3/M, z3.s, z10.s\n"
- "fmla z23.s, p3/M, z0.s, z10.s\n"
- "ld1w { z10.s }, p2/Z, [x13]\n"
+ "fmla z28.s, p3/M, z5.s, z11.s\n"
+ "ld1w { z11.s }, p2/Z, [x28, x14, LSL #2]\n"
+ "fmla z27.s, p3/M, z5.s, z12.s\n"
+ "fmla z29.s, p3/M, z5.s, z11.s\n"
+ "fmla z30.s, p3/M, z4.s, z11.s\n"
+ "fmla z31.s, p3/M, z3.s, z11.s\n"
+ "fmla z23.s, p3/M, z8.s, z12.s\n"
+ "ld1w { z12.s }, p2/Z, [x24, x14, LSL #2]\n"
+ "fmla z28.s, p3/M, z8.s, z10.s\n"
+ "ld1w { z10.s }, p2/Z, [x13, x5, LSL #2]\n"
"fmla z25.s, p3/M, z8.s, z11.s\n"
- "fmla z24.s, p3/M, z7.s, z11.s\n"
- "fmla z21.s, p3/M, z5.s, z11.s\n"
- "fmla z20.s, p3/M, z4.s, z11.s\n"
- "fmla z17.s, p3/M, z2.s, z11.s\n"
- "fmla z16.s, p3/M, z1.s, z11.s\n"
- "ld1w { z11.s }, p2/Z, [x12, x10, LSL #2]\n"
- "fmla z28.s, p3/M, z8.s, z12.s\n"
- "fmla z24.s, p3/M, z5.s, z12.s\n"
- "fmla z20.s, p3/M, z2.s, z12.s\n"
- "ld1w { z12.s }, p2/Z, [x13, x27, LSL #2]\n"
+ "fmla z26.s, p3/M, z7.s, z11.s\n"
+ "addvl x24, x24, #1\n"
+ "fmla z27.s, p3/M, z6.s, z11.s\n"
+ "fmla z29.s, p3/M, z8.s, z12.s\n"
+ "ld1w { z11.s }, p2/Z, [x13, x10, LSL #2]\n"
"addvl x13, x13, #1\n"
- "fmla z27.s, p3/M, z6.s, z10.s\n"
- "fmla z23.s, p3/M, z3.s, z10.s\n"
- "fmla z19.s, p3/M, z0.s, z10.s\n"
- "ld1w { z10.s }, p2/Z, [x11, x10, LSL #2]\n"
- "fmla z22.s, p3/M, z7.s, z11.s\n"
- "fmla z21.s, p3/M, z6.s, z11.s\n"
- "fmla z23.s, p3/M, z8.s, z11.s\n"
- "fmla z19.s, p3/M, z5.s, z11.s\n"
- "fmla z18.s, p3/M, z4.s, z11.s\n"
- "fmla z17.s, p3/M, z3.s, z11.s\n"
- "ld1w { z11.s }, p2/Z, [x12, x9, LSL #2]\n"
- "fmla z24.s, p3/M, z8.s, z12.s\n"
- "fmla z20.s, p3/M, z5.s, z12.s\n"
- "fmla z16.s, p3/M, z2.s, z12.s\n"
- "ld1w { z12.s }, p2/Z, [x11, x9, LSL #2]\n"
- "addvl x11, x11, #1\n"
- "fmla z19.s, p3/M, z8.s, z10.s\n"
- "fmla z18.s, p3/M, z7.s, z10.s\n"
- "fmla z17.s, p3/M, z6.s, z10.s\n"
- "ld1w { z10.s }, p2/Z, [x15, x7, LSL #2]\n"
- "fmla z22.s, p3/M, z8.s, z11.s\n"
- "fmla z21.s, p3/M, z7.s, z11.s\n"
- "fmla z20.s, p3/M, z6.s, z11.s\n"
+ "fmla z30.s, p3/M, z7.s, z12.s\n"
+ "fmla z31.s, p3/M, z6.s, z12.s\n"
+ "ld1w { z12.s }, p2/Z, [x28, x5, LSL #2]\n"
+ "fmla z16.s, p3/M, z4.s, z10.s\n"
+ "fmla z17.s, p3/M, z3.s, z10.s\n"
+ "fmax z16.s, p3/M, z16.s, z14.s\n"
+ "fmax z17.s, p3/M, z17.s, z14.s\n"
+ "fmla z20.s, p3/M, z1.s, z10.s\n"
+ "fmla z21.s, p3/M, z0.s, z10.s\n"
+ "ld1w { z10.s }, p2/Z, [x28, x10, LSL #2]\n"
+ "fmax z20.s, p3/M, z20.s, z14.s\n"
"fmla z18.s, p3/M, z5.s, z11.s\n"
- "fmla z17.s, p3/M, z4.s, z11.s\n"
- "fmla z16.s, p3/M, z3.s, z11.s\n"
- "ld1w { z11.s }, p2/Z, [x15, x28, LSL #2]\n"
- "addvl x15, x15, #1\n"
- "fmla z18.s, p3/M, z8.s, z12.s\n"
+ "fmla z19.s, p3/M, z4.s, z11.s\n"
+ "fmax z18.s, p3/M, z18.s, z14.s\n"
+ "fmax z19.s, p3/M, z19.s, z14.s\n"
+ "fmla z22.s, p3/M, z2.s, z11.s\n"
+ "fmla z23.s, p3/M, z1.s, z11.s\n"
+ "fmax z21.s, p3/M, z21.s, z14.s\n"
+ "fmax z22.s, p3/M, z22.s, z14.s\n"
+ "fmla z24.s, p3/M, z7.s, z12.s\n"
+ "fmla z25.s, p3/M, z6.s, z12.s\n"
+ "fmax z23.s, p3/M, z23.s, z14.s\n"
+ "fmax z24.s, p3/M, z24.s, z14.s\n"
+ "fmla z28.s, p3/M, z4.s, z12.s\n"
+ "fmla z29.s, p3/M, z3.s, z12.s\n"
+ "fmax z25.s, p3/M, z25.s, z14.s\n"
+ "fmax z28.s, p3/M, z28.s, z14.s\n"
+ "fmla z26.s, p3/M, z8.s, z10.s\n"
+ "fmla z27.s, p3/M, z7.s, z10.s\n"
+ "fmax z26.s, p3/M, z26.s, z14.s\n"
+ "fmax z27.s, p3/M, z27.s, z14.s\n"
+ "fmla z30.s, p3/M, z5.s, z10.s\n"
"fmla z31.s, p3/M, z4.s, z10.s\n"
- "fmla z17.s, p3/M, z7.s, z12.s\n"
- "fmla z16.s, p3/M, z6.s, z12.s\n"
- "ld1w { z12.s }, p2/Z, [x12, x7, LSL #2]\n"
- "fmla z30.s, p3/M, z3.s, z10.s\n"
- "fmla z27.s, p3/M, z1.s, z10.s\n"
- "fmla z26.s, p3/M, z0.s, z10.s\n"
- "ld1w { z10.s }, p2/Z, [x12, x28, LSL #2]\n"
- "whilelt p2.s, x5, %x[n_channels]\n"
- "fmla z29.s, p3/M, z5.s, z11.s\n"
- "ld1w { z0.s }, p3/Z, [x4, #1, MUL VL]\n"
- "addvl x12, x12, #1\n"
- "fmla z28.s, p3/M, z4.s, z11.s\n"
- "cmp x6, %x[n_channels]\n"
- "fmla z25.s, p3/M, z2.s, z11.s\n"
- "ld1w { z2.s }, p3/Z, [x4, #3, MUL VL]\n"
- "fmla z24.s, p3/M, z1.s, z11.s\n"
- "ld1w { z11.s }, p1/Z, [x8, x27, LSL #2]\n"
- "fmla z23.s, p3/M, z7.s, z12.s\n"
- "ld1w { z1.s }, p3/Z, [x4, #2, MUL VL]\n"
- "fmla z22.s, p3/M, z6.s, z12.s\n"
- "ld1w { z6.s }, p3/Z, [x4, #7, MUL VL]\n"
- "fmla z19.s, p3/M, z4.s, z12.s\n"
- "fmla z18.s, p3/M, z3.s, z12.s\n"
- "ld1w { z12.s }, p1/Z, [x14, x9, LSL #2]\n"
- "fmla z21.s, p3/M, z8.s, z10.s\n"
- "ld1w { z3.s }, p3/Z, [x4, #4, MUL VL]\n"
- "fmla z20.s, p3/M, z7.s, z10.s\n"
- "fmla z17.s, p3/M, z5.s, z10.s\n"
- "ld1w { z5.s }, p3/Z, [x4, #6, MUL VL]\n"
- "fmla z16.s, p3/M, z4.s, z10.s\n"
+ "fmax z29.s, p3/M, z29.s, z14.s\n"
+ "fmax z30.s, p3/M, z30.s, z14.s\n"
+ "fmax z31.s, p3/M, z31.s, z14.s\n"
+ "ld1w { z15.s }, p3/Z, [x17]\n"
+ "ld1w { z0.s }, p3/Z, [x17, #1, MUL VL]\n"
+ "whilelt p2.s, x21, %x[n_channels]\n"
+ "ld1w { z1.s }, p3/Z, [x17, #2, MUL VL]\n"
+ "ld1w { z2.s }, p3/Z, [x17, #3, MUL VL]\n"
+ "cmp x16, %x[n_channels]\n"
+ "fmin z16.s, p3/M, z16.s, z13.s\n"
+ "ld1w { z3.s }, p3/Z, [x17, #4, MUL VL]\n"
+ "ld1w { z4.s }, p3/Z, [x17, #5, MUL VL]\n"
+ "fmin z17.s, p3/M, z17.s, z13.s\n"
+ "fmin z18.s, p3/M, z18.s, z13.s\n"
+ "ld1w { z5.s }, p3/Z, [x17, #6, MUL VL]\n"
+ "ld1w { z6.s }, p3/Z, [x17, #7, MUL VL]\n"
+ "addvl x17, x17, #16\n"
+ "fmin z19.s, p3/M, z19.s, z13.s\n"
+ "fmin z20.s, p3/M, z20.s, z13.s\n"
+ "fmin z21.s, p3/M, z21.s, z13.s\n"
+ "ld1w { z9.s }, p1/Z, [x12, x7, LSL #2]\n"
"ld1w { z10.s }, p1/Z, [x8]\n"
- "fmax z31.s, p3/M, z31.s, z15.s\n"
- "ld1w { z4.s }, p3/Z, [x4, #5, MUL VL]\n"
- "addvl x4, x4, #16\n"
- "fmax z30.s, p3/M, z30.s, z15.s\n"
- "ld1w { z7.s }, p3/Z, [x4, #-8, MUL VL]\n"
- "fmax z29.s, p3/M, z29.s, z15.s\n"
- "ld1w { z8.s }, p3/Z, [x4, #-7, MUL VL]\n"
- "addvl x4, x4, #-6\n"
- "fmin z31.s, p3/M, z31.s, z14.s\n"
- "st1w { z31.s }, p0, [x16]\n"
- "fmin z30.s, p3/M, z30.s, z14.s\n"
- "fmin z29.s, p3/M, z29.s, z14.s\n"
- "st1w { z30.s }, p0, [x16, x17, LSL #2]\n"
- "fmax z28.s, p3/M, z28.s, z15.s\n"
- "fmax z27.s, p3/M, z27.s, z15.s\n"
- "st1w { z29.s }, p0, [x16, x26, LSL #2]\n"
- "fmax z26.s, p3/M, z26.s, z15.s\n"
- "fmax z25.s, p3/M, z25.s, z15.s\n"
- "fmax z24.s, p3/M, z24.s, z15.s\n"
- "fmin z28.s, p3/M, z28.s, z14.s\n"
- "st1w { z28.s }, p0, [x16, x22, LSL #2]\n"
- "fmin z27.s, p3/M, z27.s, z14.s\n"
- "addvl x16, x16, #1\n"
- "fmin z26.s, p3/M, z26.s, z14.s\n"
- "st1w { z27.s }, p0, [x25]\n"
- "fmin z25.s, p3/M, z25.s, z14.s\n"
- "fmin z24.s, p3/M, z24.s, z14.s\n"
- "st1w { z26.s }, p0, [x25, x17, LSL #2]\n"
- "fmax z23.s, p3/M, z23.s, z15.s\n"
- "st1w { z25.s }, p0, [x25, x26, LSL #2]\n"
- "fmax z22.s, p3/M, z22.s, z15.s\n"
- "fmax z21.s, p3/M, z21.s, z15.s\n"
- "st1w { z24.s }, p0, [x25, x22, LSL #2]\n"
- "addvl x25, x25, #1\n"
- "fmin z23.s, p3/M, z23.s, z14.s\n"
- "st1w { z23.s }, p0, [x24]\n"
- "fmin z22.s, p3/M, z22.s, z14.s\n"
- "fmin z21.s, p3/M, z21.s, z14.s\n"
- "st1w { z22.s }, p0, [x24, x17, LSL #2]\n"
- "fmax z20.s, p3/M, z20.s, z15.s\n"
- "fmax z19.s, p3/M, z19.s, z15.s\n"
- "st1w { z21.s }, p0, [x24, x26, LSL #2]\n"
- "fmax z18.s, p3/M, z18.s, z15.s\n"
- "fmax z17.s, p3/M, z17.s, z15.s\n"
- "fmax z16.s, p3/M, z16.s, z15.s\n"
- "fmin z20.s, p3/M, z20.s, z14.s\n"
- "st1w { z20.s }, p0, [x24, x22, LSL #2]\n"
- "fmin z19.s, p3/M, z19.s, z14.s\n"
- "addvl x24, x24, #1\n"
- "fmin z18.s, p3/M, z18.s, z14.s\n"
- "st1w { z19.s }, p0, [x23]\n"
- "fmin z17.s, p3/M, z17.s, z14.s\n"
- "fmin z16.s, p3/M, z16.s, z14.s\n"
- "st1w { z18.s }, p0, [x23, x17, LSL #2]\n"
- "st1w { z17.s }, p0, [x23, x26, LSL #2]\n"
- "st1w { z16.s }, p0, [x23, x22, LSL #2]\n"
+ "fmin z22.s, p3/M, z22.s, z13.s\n"
+ "fmin z23.s, p3/M, z23.s, z13.s\n"
+ "ld1w { z11.s }, p1/Z, [x8, x27, LSL #2]\n"
+ "ld1w { z12.s }, p1/Z, [x12, x14, LSL #2]\n"
+ "fmin z24.s, p3/M, z24.s, z13.s\n"
+ "fmin z25.s, p3/M, z25.s, z13.s\n"
+ "st1w { z16.s }, p0, [x15]\n"
+ "ld1w { z7.s }, p3/Z, [x17, #-8, MUL VL]\n"
+ "fmin z26.s, p3/M, z26.s, z13.s\n"
+ "fmin z27.s, p3/M, z27.s, z13.s\n"
+ "st1w { z17.s }, p0, [x15, x6, LSL #2]\n"
+ "ld1w { z8.s }, p3/Z, [x17, #-7, MUL VL]\n"
+ "fmin z28.s, p3/M, z28.s, z13.s\n"
+ "fmin z29.s, p3/M, z29.s, z13.s\n"
+ "st1w { z18.s }, p0, [x15, x25, LSL #2]\n"
+ "fmin z30.s, p3/M, z30.s, z13.s\n"
+ "fmin z31.s, p3/M, z31.s, z13.s\n"
+ "st1w { z19.s }, p0, [x15, x22, LSL #2]\n"
+ "addvl x28, x28, #1\n"
+ "st1w { z20.s }, p0, [x9]\n"
+ "addvl x15, x15, #1\n"
+ "st1w { z21.s }, p0, [x9, x6, LSL #2]\n"
+ "addvl x17, x17, #-6\n"
+ "st1w { z22.s }, p0, [x9, x25, LSL #2]\n"
+ "st1w { z23.s }, p0, [x9, x22, LSL #2]\n"
+ "addvl x9, x9, #1\n"
+ "st1w { z24.s }, p0, [x26]\n"
+ "st1w { z25.s }, p0, [x26, x6, LSL #2]\n"
+ "st1w { z26.s }, p0, [x26, x25, LSL #2]\n"
+ "st1w { z27.s }, p0, [x26, x22, LSL #2]\n"
+ "addvl x26, x26, #1\n"
+ "st1w { z28.s }, p0, [x23]\n"
+ "st1w { z29.s }, p0, [x23, x6, LSL #2]\n"
+ "st1w { z30.s }, p0, [x23, x25, LSL #2]\n"
+ "st1w { z31.s }, p0, [x23, x22, LSL #2]\n"
"addvl x23, x23, #1\n"
"blt 2b\n"
"3:" // Tile loop: Channel tail
- "movprfx z31, z13\n fmla z31.s, p3/M, z8.s, z9.s\n"
- "ldr x2, [%x[params_struct], %[offsetof_args_tile_i]]\n"
- "mov p0.b, p2.b\n"
- "movprfx z30, z13\n fmla z30.s, p3/M, z7.s, z9.s\n"
- "ldr x3, [%x[params_struct], %[offsetof_args_tile_j]]\n"
- "add x21, x2, #0x1\n"
- "movprfx z29, z13\n fmla z29.s, p3/M, z6.s, z9.s\n"
+ "movprfx z21, z15\n fmla z21.s, p3/M, z4.s, z9.s\n"
+ "movprfx z16, z15\n fmla z16.s, p3/M, z8.s, z9.s\n"
+ "ldr x4, [%x[params_struct], %[offsetof_args_tile_j]]\n"
+ "ldr x16, [%x[params_struct], %[offsetof_args_tile_i]]\n"
+ "movprfx z22, z15\n fmla z22.s, p3/M, z3.s, z9.s\n"
+ "movprfx z25, z15\n fmla z25.s, p3/M, z1.s, z9.s\n"
+ "ldr x20, [%x[params_struct], %[offsetof_args_n_tile_cols]]\n"
+ "add x4, x4, #0x1\n"
+ "movprfx z26, z15\n fmla z26.s, p3/M, z0.s, z9.s\n"
+ "fmla z21.s, p3/M, z5.s, z12.s\n"
+ "cmp x4, x20\n"
+ "add x21, x16, #0x1\n"
+ "movprfx z17, z15\n fmla z17.s, p3/M, z7.s, z9.s\n"
+ "movprfx z18, z15\n fmla z18.s, p3/M, z6.s, z9.s\n"
"ldr x20, [%x[params_struct], %[offsetof_args_n_tile_rows]]\n"
- "movprfx z27, z13\n fmla z27.s, p3/M, z5.s, z9.s\n"
- "ldr x19, [%x[params_struct], %[offsetof_args_n_tile_cols]]\n"
- "add x3, x3, #0x1\n"
- "movprfx z26, z13\n fmla z26.s, p3/M, z4.s, z9.s\n"
- "cmp x3, x19\n"
- "movprfx z25, z13\n fmla z25.s, p3/M, z3.s, z9.s\n"
- "movprfx z23, z13\n fmla z23.s, p3/M, z2.s, z9.s\n"
- "csel x3, x3, XZR, LT\n"
- "movprfx z22, z13\n fmla z22.s, p3/M, z1.s, z9.s\n"
- "csel x2, x2, x21, LT\n"
- "movprfx z21, z13\n fmla z21.s, p3/M, z0.s, z9.s\n"
- "ld1w { z9.s }, p2/Z, [x13, x10, LSL #2]\n"
- "cmp x2, x20\n"
- "fmla z31.s, p3/M, z0.s, z10.s\n"
- "ld1w { z10.s }, p2/Z, [x11]\n"
- "movprfx z28, z13\n fmla z28.s, p3/M, z2.s, z11.s\n"
- "ld1w { z11.s }, p2/Z, [x11, x27, LSL #2]\n"
- "fmla z30.s, p3/M, z8.s, z12.s\n"
- "fmla z29.s, p3/M, z7.s, z12.s\n"
- "fmla z26.s, p3/M, z5.s, z12.s\n"
- "fmla z28.s, p3/M, z6.s, z12.s\n"
- "fmla z25.s, p3/M, z4.s, z12.s\n"
- "movprfx z24, z13\n fmla z24.s, p3/M, z3.s, z12.s\n"
- "fmla z22.s, p3/M, z2.s, z12.s\n"
- "fmla z21.s, p3/M, z1.s, z12.s\n"
- "movprfx z20, z13\n fmla z20.s, p3/M, z0.s, z12.s\n"
- "ld1w { z12.s }, p2/Z, [x8, x7, LSL #2]\n"
- "movprfx z19, z13\n fmla z19.s, p3/M, z6.s, z10.s\n"
- "ld1w { z10.s }, p2/Z, [x13, x9, LSL #2]\n"
- "movprfx z16, z13\n fmla z16.s, p3/M, z8.s, z11.s\n"
- "ld1w { z11.s }, p2/Z, [x8, x28, LSL #2]\n"
- "fmla z27.s, p3/M, z8.s, z9.s\n"
- "fmla z26.s, p3/M, z7.s, z9.s\n"
- "fmla z25.s, p3/M, z6.s, z9.s\n"
- "fmla z23.s, p3/M, z5.s, z9.s\n"
- "fmla z22.s, p3/M, z4.s, z9.s\n"
- "fmla z21.s, p3/M, z3.s, z9.s\n"
- "fmla z19.s, p3/M, z2.s, z9.s\n"
- "movprfx z18, z13\n fmla z18.s, p3/M, z1.s, z9.s\n"
- "movprfx z17, z13\n fmla z17.s, p3/M, z0.s, z9.s\n"
- "ld1w { z9.s }, p2/Z, [x15]\n"
- "fmla z31.s, p3/M, z1.s, z12.s\n"
- "fmla z30.s, p3/M, z0.s, z12.s\n"
- "ld1w { z12.s }, p2/Z, [x15, x27, LSL #2]\n"
- "fmla z29.s, p3/M, z2.s, z11.s\n"
- "fmla z28.s, p3/M, z1.s, z11.s\n"
- "ld1w { z11.s }, p2/Z, [x12]\n"
- "fmla z26.s, p3/M, z8.s, z10.s\n"
- "fmla z25.s, p3/M, z7.s, z10.s\n"
- "fmla z24.s, p3/M, z6.s, z10.s\n"
- "fmla z22.s, p3/M, z5.s, z10.s\n"
- "fmla z21.s, p3/M, z4.s, z10.s\n"
- "fmla z20.s, p3/M, z3.s, z10.s\n"
- "fmla z18.s, p3/M, z2.s, z10.s\n"
- "fmla z17.s, p3/M, z1.s, z10.s\n"
+ "csel x16, x16, x21, LT\n"
+ "movprfx z20, z15\n fmla z20.s, p3/M, z5.s, z9.s\n"
+ "movprfx z24, z15\n fmla z24.s, p3/M, z2.s, z9.s\n"
+ "ld1w { z9.s }, p2/Z, [x11, x7, LSL #2]\n"
+ "mov p0.b, p2.b\n"
"fmla z16.s, p3/M, z0.s, z10.s\n"
- "ld1w { z10.s }, p2/Z, [x15, x10, LSL #2]\n"
- "fmla z31.s, p3/M, z3.s, z9.s\n"
- "fmla z27.s, p3/M, z0.s, z9.s\n"
- "fmla z28.s, p3/M, z5.s, z12.s\n"
- "fmla z24.s, p3/M, z2.s, z12.s\n"
- "ld1w { z12.s }, p2/Z, [x15, x9, LSL #2]\n"
- "fmla z23.s, p3/M, z6.s, z11.s\n"
- "fmla z19.s, p3/M, z3.s, z11.s\n"
- "ld1w { z11.s }, p2/Z, [x12, x27, LSL #2]\n"
- "fmla z31.s, p3/M, z5.s, z10.s\n"
- "fmla z30.s, p3/M, z4.s, z10.s\n"
- "fmla z29.s, p3/M, z3.s, z10.s\n"
- "fmla z27.s, p3/M, z2.s, z10.s\n"
- "fmla z26.s, p3/M, z1.s, z10.s\n"
+ "movprfx z19, z15\n fmla z19.s, p3/M, z2.s, z11.s\n"
+ "ld1w { z10.s }, p2/Z, [x24]\n"
+ "ld1w { z11.s }, p2/Z, [x24, x27, LSL #2]\n"
+ "fmla z22.s, p3/M, z4.s, z12.s\n"
+ "fmla z25.s, p3/M, z2.s, z12.s\n"
+ "csel x4, x4, XZR, LT\n"
+ "cmp x16, x20\n"
+ "fmla z26.s, p3/M, z1.s, z12.s\n"
+ "movprfx z28, z15\n fmla z28.s, p3/M, z6.s, z10.s\n"
+ "ld1w { z10.s }, p2/Z, [x11, x14, LSL #2]\n"
+ "fmla z21.s, p3/M, z7.s, z9.s\n"
+ "fmla z17.s, p3/M, z8.s, z12.s\n"
+ "fmla z18.s, p3/M, z7.s, z12.s\n"
+ "fmla z19.s, p3/M, z6.s, z12.s\n"
+ "movprfx z23, z15\n fmla z23.s, p3/M, z3.s, z12.s\n"
+ "movprfx z27, z15\n fmla z27.s, p3/M, z0.s, z12.s\n"
+ "ld1w { z12.s }, p2/Z, [x8, x5, LSL #2]\n"
+ "movprfx z31, z15\n fmla z31.s, p3/M, z8.s, z11.s\n"
+ "fmla z22.s, p3/M, z6.s, z9.s\n"
+ "ld1w { z11.s }, p2/Z, [x8, x10, LSL #2]\n"
+ "fmla z25.s, p3/M, z4.s, z9.s\n"
+ "fmla z26.s, p3/M, z3.s, z9.s\n"
+ "fmla z20.s, p3/M, z8.s, z9.s\n"
+ "fmla z24.s, p3/M, z5.s, z9.s\n"
+ "fmla z28.s, p3/M, z2.s, z9.s\n"
+ "fmla z21.s, p3/M, z8.s, z10.s\n"
+ "fmla z16.s, p3/M, z1.s, z12.s\n"
+ "fmla z17.s, p3/M, z0.s, z12.s\n"
+ "ld1w { z12.s }, p2/Z, [x13, x27, LSL #2]\n"
+ "fmla z18.s, p3/M, z2.s, z11.s\n"
+ "fmla z19.s, p3/M, z1.s, z11.s\n"
+ "ld1w { z11.s }, p2/Z, [x28]\n"
+ "fmla z22.s, p3/M, z7.s, z10.s\n"
+ "fmla z23.s, p3/M, z6.s, z10.s\n"
+ "fmla z25.s, p3/M, z5.s, z10.s\n"
+ "fmla z26.s, p3/M, z4.s, z10.s\n"
+ "fmla z27.s, p3/M, z3.s, z10.s\n"
+ "fmla z31.s, p3/M, z0.s, z10.s\n"
+ "fmla z24.s, p3/M, z6.s, z11.s\n"
+ "fmla z28.s, p3/M, z3.s, z11.s\n"
+ "ld1w { z11.s }, p2/Z, [x28, x27, LSL #2]\n"
+ "fmla z19.s, p3/M, z5.s, z12.s\n"
+ "fmla z23.s, p3/M, z2.s, z12.s\n"
+ "ld1w { z12.s }, p2/Z, [x13, x14, LSL #2]\n"
+ "fmla z27.s, p3/M, z8.s, z11.s\n"
+ "fmla z31.s, p3/M, z5.s, z11.s\n"
+ "movprfx z29, z15\n fmla z29.s, p3/M, z1.s, z9.s\n"
+ "movprfx z30, z15\n fmla z30.s, p3/M, z0.s, z9.s\n"
+ "ld1w { z9.s }, p2/Z, [x13]\n"
+ "fmla z29.s, p3/M, z2.s, z10.s\n"
+ "fmla z30.s, p3/M, z1.s, z10.s\n"
+ "ld1w { z10.s }, p2/Z, [x13, x7, LSL #2]\n"
+ "fmla z20.s, p3/M, z0.s, z9.s\n"
+ "fmla z21.s, p3/M, z1.s, z10.s\n"
+ "fmla z16.s, p3/M, z3.s, z9.s\n"
+ "fmla z17.s, p3/M, z4.s, z10.s\n"
+ "ld1w { z11.s }, p2/Z, [x24, x5, LSL #2]\n"
+ "fmla z18.s, p3/M, z3.s, z10.s\n"
+ "fmla z22.s, p3/M, z0.s, z10.s\n"
+ "fmla z20.s, p3/M, z2.s, z10.s\n"
+ "fmla z21.s, p3/M, z2.s, z12.s\n"
+ "fmla z16.s, p3/M, z5.s, z10.s\n"
+ "fmla z17.s, p3/M, z5.s, z12.s\n"
+ "ld1w { z10.s }, p2/Z, [x12, x5, LSL #2]\n"
+ "fmla z18.s, p3/M, z4.s, z12.s\n"
+ "fmla z19.s, p3/M, z3.s, z12.s\n"
+ "fmla z22.s, p3/M, z1.s, z12.s\n"
+ "fmla z23.s, p3/M, z0.s, z12.s\n"
+ "ld1w { z12.s }, p2/Z, [x12, x10, LSL #2]\n"
+ "fmla z28.s, p3/M, z7.s, z11.s\n"
+ "fmla z29.s, p3/M, z6.s, z11.s\n"
+ "ld1w { z11.s }, p2/Z, [x24, x10, LSL #2]\n"
+ "fmla z20.s, p3/M, z4.s, z10.s\n"
+ "fmla z21.s, p3/M, z3.s, z10.s\n"
+ "fmla z24.s, p3/M, z1.s, z10.s\n"
"fmla z25.s, p3/M, z0.s, z10.s\n"
- "ld1w { z10.s }, p2/Z, [x14, x7, LSL #2]\n"
- "fmla z20.s, p3/M, z8.s, z11.s\n"
- "fmla z16.s, p3/M, z5.s, z11.s\n"
- "ld1w { z11.s }, p2/Z, [x11, x7, LSL #2]\n"
- "fmla z30.s, p3/M, z5.s, z12.s\n"
- "fmla z29.s, p3/M, z4.s, z12.s\n"
- "fmla z28.s, p3/M, z3.s, z12.s\n"
+ "fmla z16.s, p3/M, z7.s, z10.s\n"
+ "fmla z17.s, p3/M, z6.s, z10.s\n"
+ "ld1w { z10.s }, p2/Z, [x8, x7, LSL #2]\n"
+ "fmla z30.s, p3/M, z8.s, z11.s\n"
+ "fmla z31.s, p3/M, z7.s, z11.s\n"
+ "ld1w { z11.s }, p2/Z, [x11, x5, LSL #2]\n"
+ "fmla z18.s, p3/M, z8.s, z12.s\n"
+ "fmla z19.s, p3/M, z7.s, z12.s\n"
+ "fmla z22.s, p3/M, z5.s, z12.s\n"
+ "fmla z23.s, p3/M, z4.s, z12.s\n"
"fmla z26.s, p3/M, z2.s, z12.s\n"
- "fmla z25.s, p3/M, z1.s, z12.s\n"
- "fmla z24.s, p3/M, z0.s, z12.s\n"
- "ld1w { z12.s }, p2/Z, [x14, x28, LSL #2]\n"
- "fmla z19.s, p3/M, z7.s, z11.s\n"
- "fmla z18.s, p3/M, z6.s, z11.s\n"
- "ld1w { z11.s }, p2/Z, [x11, x28, LSL #2]\n"
- "fmla z31.s, p3/M, z7.s, z10.s\n"
+ "fmla z27.s, p3/M, z1.s, z12.s\n"
+ "ld1w { z12.s }, p2/Z, [x8, x14, LSL #2]\n"
+ "fmla z20.s, p3/M, z7.s, z11.s\n"
+ "fmla z21.s, p3/M, z6.s, z11.s\n"
+ "fmla z24.s, p3/M, z4.s, z11.s\n"
+ "fmla z25.s, p3/M, z3.s, z11.s\n"
+ "fmla z28.s, p3/M, z1.s, z11.s\n"
+ "fmla z29.s, p3/M, z0.s, z11.s\n"
+ "ld1w { z11.s }, p2/Z, [x11, x10, LSL #2]\n"
+ "fmla z16.s, p3/M, z2.s, z10.s\n"
+ "fmla z17.s, p3/M, z1.s, z10.s\n"
+ "fmla z18.s, p3/M, z0.s, z10.s\n"
+ "ld1w { z10.s }, p2/Z, [x12]\n"
+ "fmla z30.s, p3/M, z2.s, z11.s\n"
+ "fmla z19.s, p3/M, z0.s, z12.s\n"
+ "fmla z20.s, p3/M, z3.s, z10.s\n"
+ "fmla z24.s, p3/M, z0.s, z10.s\n"
+ "fmla z22.s, p3/M, z8.s, z11.s\n"
+ "fmla z23.s, p3/M, z7.s, z11.s\n"
+ "fmla z26.s, p3/M, z5.s, z11.s\n"
+ "fmla z27.s, p3/M, z4.s, z11.s\n"
+ "fmla z31.s, p3/M, z1.s, z11.s\n"
+ "ld1w { z11.s }, p2/Z, [x28, x7, LSL #2]\n"
+ "fmla z17.s, p3/M, z2.s, z12.s\n"
+ "fmla z18.s, p3/M, z1.s, z12.s\n"
+ "ld1w { z12.s }, p2/Z, [x12, x27, LSL #2]\n"
+ "fmla z16.s, p3/M, z6.s, z10.s\n"
+ "ld1w { z10.s }, p2/Z, [x11]\n"
+ "fmla z29.s, p3/M, z4.s, z11.s\n"
+ "fmla z30.s, p3/M, z3.s, z11.s\n"
+ "fmla z19.s, p3/M, z8.s, z12.s\n"
+ "fmla z23.s, p3/M, z5.s, z12.s\n"
+ "fmla z27.s, p3/M, z2.s, z12.s\n"
+ "ld1w { z12.s }, p2/Z, [x11, x27, LSL #2]\n"
+ "fmla z20.s, p3/M, z6.s, z10.s\n"
+ "fmla z24.s, p3/M, z3.s, z10.s\n"
+ "fmla z28.s, p3/M, z0.s, z10.s\n"
+ "ld1w { z10.s }, p2/Z, [x24, x7, LSL #2]\n"
+ "fmla z31.s, p3/M, z2.s, z12.s\n"
+ "fmla z29.s, p3/M, z7.s, z10.s\n"
"fmla z30.s, p3/M, z6.s, z10.s\n"
- "fmla z27.s, p3/M, z4.s, z10.s\n"
- "fmla z26.s, p3/M, z3.s, z10.s\n"
- "fmla z23.s, p3/M, z1.s, z10.s\n"
- "fmla z22.s, p3/M, z0.s, z10.s\n"
- "ld1w { z10.s }, p2/Z, [x8, x10, LSL #2]\n"
- "fmla z17.s, p3/M, z8.s, z11.s\n"
- "fmla z16.s, p3/M, z7.s, z11.s\n"
- "ld1w { z11.s }, p2/Z, [x13, x7, LSL #2]\n"
- "fmla z29.s, p3/M, z8.s, z12.s\n"
- "fmla z28.s, p3/M, z7.s, z12.s\n"
- "fmla z25.s, p3/M, z5.s, z12.s\n"
- "fmla z24.s, p3/M, z4.s, z12.s\n"
- "fmla z21.s, p3/M, z2.s, z12.s\n"
- "fmla z20.s, p3/M, z1.s, z12.s\n"
- "ld1w { z12.s }, p2/Z, [x8, x9, LSL #2]\n"
- "fmla z31.s, p3/M, z2.s, z10.s\n"
- "fmla z30.s, p3/M, z1.s, z10.s\n"
- "fmla z29.s, p3/M, z0.s, z10.s\n"
- "ld1w { z10.s }, p2/Z, [x14]\n"
- "fmla z27.s, p3/M, z7.s, z11.s\n"
+ "fmla z24.s, p3/M, z8.s, z11.s\n"
+ "fmla z25.s, p3/M, z7.s, z11.s\n"
"fmla z26.s, p3/M, z6.s, z11.s\n"
- "fmla z23.s, p3/M, z4.s, z11.s\n"
- "fmla z22.s, p3/M, z3.s, z11.s\n"
- "fmla z19.s, p3/M, z1.s, z11.s\n"
- "fmla z18.s, p3/M, z0.s, z11.s\n"
- "ld1w { z11.s }, p2/Z, [x13, x28, LSL #2]\n"
- "fmla z30.s, p3/M, z2.s, z12.s\n"
- "fmla z29.s, p3/M, z1.s, z12.s\n"
- "fmla z28.s, p3/M, z0.s, z12.s\n"
- "ld1w { z12.s }, p2/Z, [x14, x27, LSL #2]\n"
- "fmla z31.s, p3/M, z6.s, z10.s\n"
- "fmla z27.s, p3/M, z3.s, z10.s\n"
- "fmla z23.s, p3/M, z0.s, z10.s\n"
- "ld1w { z10.s }, p2/Z, [x13]\n"
+ "fmla z28.s, p3/M, z5.s, z11.s\n"
+ "ld1w { z11.s }, p2/Z, [x28, x14, LSL #2]\n"
+ "fmla z27.s, p3/M, z5.s, z12.s\n"
+ "fmla z29.s, p3/M, z5.s, z11.s\n"
+ "fmla z30.s, p3/M, z4.s, z11.s\n"
+ "fmla z31.s, p3/M, z3.s, z11.s\n"
+ "fmla z23.s, p3/M, z8.s, z12.s\n"
+ "ld1w { z12.s }, p2/Z, [x24, x14, LSL #2]\n"
+ "fmla z28.s, p3/M, z8.s, z10.s\n"
+ "ld1w { z10.s }, p2/Z, [x13, x5, LSL #2]\n"
"fmla z25.s, p3/M, z8.s, z11.s\n"
- "fmla z24.s, p3/M, z7.s, z11.s\n"
- "fmla z21.s, p3/M, z5.s, z11.s\n"
- "fmla z20.s, p3/M, z4.s, z11.s\n"
- "fmla z17.s, p3/M, z2.s, z11.s\n"
- "fmla z16.s, p3/M, z1.s, z11.s\n"
- "ld1w { z11.s }, p2/Z, [x12, x10, LSL #2]\n"
- "fmla z28.s, p3/M, z8.s, z12.s\n"
- "fmla z24.s, p3/M, z5.s, z12.s\n"
- "fmla z20.s, p3/M, z2.s, z12.s\n"
- "ld1w { z12.s }, p2/Z, [x13, x27, LSL #2]\n"
- "fmla z27.s, p3/M, z6.s, z10.s\n"
- "fmla z23.s, p3/M, z3.s, z10.s\n"
- "fmla z19.s, p3/M, z0.s, z10.s\n"
- "ld1w { z10.s }, p2/Z, [x11, x10, LSL #2]\n"
- "fmla z22.s, p3/M, z7.s, z11.s\n"
- "fmla z21.s, p3/M, z6.s, z11.s\n"
- "fmla z23.s, p3/M, z8.s, z11.s\n"
- "fmla z19.s, p3/M, z5.s, z11.s\n"
- "fmla z18.s, p3/M, z4.s, z11.s\n"
- "fmla z17.s, p3/M, z3.s, z11.s\n"
- "ld1w { z11.s }, p2/Z, [x12, x9, LSL #2]\n"
- "fmla z24.s, p3/M, z8.s, z12.s\n"
- "fmla z20.s, p3/M, z5.s, z12.s\n"
- "fmla z16.s, p3/M, z2.s, z12.s\n"
- "ld1w { z12.s }, p2/Z, [x11, x9, LSL #2]\n"
- "fmla z19.s, p3/M, z8.s, z10.s\n"
- "fmla z18.s, p3/M, z7.s, z10.s\n"
- "fmla z17.s, p3/M, z6.s, z10.s\n"
- "ld1w { z10.s }, p2/Z, [x15, x7, LSL #2]\n"
- "fmla z22.s, p3/M, z8.s, z11.s\n"
- "fmla z21.s, p3/M, z7.s, z11.s\n"
- "fmla z20.s, p3/M, z6.s, z11.s\n"
+ "fmla z26.s, p3/M, z7.s, z11.s\n"
+ "fmla z27.s, p3/M, z6.s, z11.s\n"
+ "fmla z29.s, p3/M, z8.s, z12.s\n"
+ "ld1w { z11.s }, p2/Z, [x13, x10, LSL #2]\n"
+ "fmla z30.s, p3/M, z7.s, z12.s\n"
+ "fmla z31.s, p3/M, z6.s, z12.s\n"
+ "ld1w { z12.s }, p2/Z, [x28, x5, LSL #2]\n"
+ "fmla z16.s, p3/M, z4.s, z10.s\n"
+ "fmla z17.s, p3/M, z3.s, z10.s\n"
+ "fmax z16.s, p3/M, z16.s, z14.s\n"
+ "fmax z17.s, p3/M, z17.s, z14.s\n"
+ "fmla z20.s, p3/M, z1.s, z10.s\n"
+ "fmla z21.s, p3/M, z0.s, z10.s\n"
+ "ld1w { z10.s }, p2/Z, [x28, x10, LSL #2]\n"
+ "fmax z20.s, p3/M, z20.s, z14.s\n"
"fmla z18.s, p3/M, z5.s, z11.s\n"
- "fmla z17.s, p3/M, z4.s, z11.s\n"
- "fmla z16.s, p3/M, z3.s, z11.s\n"
- "ld1w { z11.s }, p2/Z, [x15, x28, LSL #2]\n"
+ "fmla z19.s, p3/M, z4.s, z11.s\n"
+ "fmax z18.s, p3/M, z18.s, z14.s\n"
+ "fmax z19.s, p3/M, z19.s, z14.s\n"
+ "fmla z22.s, p3/M, z2.s, z11.s\n"
+ "fmla z23.s, p3/M, z1.s, z11.s\n"
+ "fmax z21.s, p3/M, z21.s, z14.s\n"
+ "fmax z22.s, p3/M, z22.s, z14.s\n"
+ "fmla z24.s, p3/M, z7.s, z12.s\n"
+ "fmla z25.s, p3/M, z6.s, z12.s\n"
+ "fmax z23.s, p3/M, z23.s, z14.s\n"
+ "fmax z24.s, p3/M, z24.s, z14.s\n"
+ "fmla z28.s, p3/M, z4.s, z12.s\n"
+ "fmla z29.s, p3/M, z3.s, z12.s\n"
+ "fmax z25.s, p3/M, z25.s, z14.s\n"
+ "fmax z28.s, p3/M, z28.s, z14.s\n"
+ "fmla z26.s, p3/M, z8.s, z10.s\n"
+ "fmla z27.s, p3/M, z7.s, z10.s\n"
+ "fmax z26.s, p3/M, z26.s, z14.s\n"
+ "fmax z27.s, p3/M, z27.s, z14.s\n"
+ "fmla z30.s, p3/M, z5.s, z10.s\n"
"fmla z31.s, p3/M, z4.s, z10.s\n"
- "fmla z18.s, p3/M, z8.s, z12.s\n"
- "fmla z17.s, p3/M, z7.s, z12.s\n"
- "fmla z16.s, p3/M, z6.s, z12.s\n"
- "ld1w { z12.s }, p2/Z, [x12, x7, LSL #2]\n"
- "fmla z30.s, p3/M, z3.s, z10.s\n"
- "fmla z27.s, p3/M, z1.s, z10.s\n"
- "fmla z26.s, p3/M, z0.s, z10.s\n"
- "ld1w { z10.s }, p2/Z, [x12, x28, LSL #2]\n"
- "fmla z29.s, p3/M, z5.s, z11.s\n"
- "fmla z28.s, p3/M, z4.s, z11.s\n"
- "fmla z25.s, p3/M, z2.s, z11.s\n"
- "fmla z24.s, p3/M, z1.s, z11.s\n"
- "fmla z23.s, p3/M, z7.s, z12.s\n"
- "fmla z22.s, p3/M, z6.s, z12.s\n"
- "fmla z19.s, p3/M, z4.s, z12.s\n"
- "fmla z18.s, p3/M, z3.s, z12.s\n"
- "fmla z21.s, p3/M, z8.s, z10.s\n"
- "fmla z20.s, p3/M, z7.s, z10.s\n"
- "fmla z17.s, p3/M, z5.s, z10.s\n"
- "fmla z16.s, p3/M, z4.s, z10.s\n"
- "fmax z31.s, p3/M, z31.s, z15.s\n"
- "fmax z30.s, p3/M, z30.s, z15.s\n"
- "fmax z29.s, p3/M, z29.s, z15.s\n"
- "fmax z28.s, p3/M, z28.s, z15.s\n"
- "fmin z31.s, p3/M, z31.s, z14.s\n"
- "st1w { z31.s }, p0, [x16]\n"
- "fmin z30.s, p3/M, z30.s, z14.s\n"
- "fmin z29.s, p3/M, z29.s, z14.s\n"
- "st1w { z30.s }, p0, [x16, x17, LSL #2]\n"
- "fmin z28.s, p3/M, z28.s, z14.s\n"
- "fmax z27.s, p3/M, z27.s, z15.s\n"
- "st1w { z29.s }, p0, [x16, x26, LSL #2]\n"
- "fmax z26.s, p3/M, z26.s, z15.s\n"
- "st1w { z28.s }, p0, [x16, x22, LSL #2]\n"
- "fmin z27.s, p3/M, z27.s, z14.s\n"
- "fmax z25.s, p3/M, z25.s, z15.s\n"
- "st1w { z27.s }, p0, [x25]\n"
- "fmin z26.s, p3/M, z26.s, z14.s\n"
- "fmin z25.s, p3/M, z25.s, z14.s\n"
- "st1w { z26.s }, p0, [x25, x17, LSL #2]\n"
- "fmax z24.s, p3/M, z24.s, z15.s\n"
- "fmax z23.s, p3/M, z23.s, z15.s\n"
- "st1w { z25.s }, p0, [x25, x26, LSL #2]\n"
- "fmax z22.s, p3/M, z22.s, z15.s\n"
- "fmax z21.s, p3/M, z21.s, z15.s\n"
- "fmax z20.s, p3/M, z20.s, z15.s\n"
- "fmin z24.s, p3/M, z24.s, z14.s\n"
- "st1w { z24.s }, p0, [x25, x22, LSL #2]\n"
- "fmin z23.s, p3/M, z23.s, z14.s\n"
- "fmin z22.s, p3/M, z22.s, z14.s\n"
- "st1w { z23.s }, p0, [x24]\n"
- "fmin z21.s, p3/M, z21.s, z14.s\n"
- "fmin z20.s, p3/M, z20.s, z14.s\n"
- "st1w { z22.s }, p0, [x24, x17, LSL #2]\n"
- "fmax z19.s, p3/M, z19.s, z15.s\n"
- "st1w { z21.s }, p0, [x24, x26, LSL #2]\n"
- "fmax z18.s, p3/M, z18.s, z15.s\n"
- "fmax z17.s, p3/M, z17.s, z15.s\n"
- "st1w { z20.s }, p0, [x24, x22, LSL #2]\n"
- "fmin z19.s, p3/M, z19.s, z14.s\n"
- "st1w { z19.s }, p0, [x23]\n"
- "fmin z18.s, p3/M, z18.s, z14.s\n"
- "fmin z17.s, p3/M, z17.s, z14.s\n"
- "st1w { z18.s }, p0, [x23, x17, LSL #2]\n"
- "fmax z16.s, p3/M, z16.s, z15.s\n"
- "st1w { z17.s }, p0, [x23, x26, LSL #2]\n"
- "fmin z16.s, p3/M, z16.s, z14.s\n"
- "st1w { z16.s }, p0, [x23, x22, LSL #2]\n"
+ "fmax z29.s, p3/M, z29.s, z14.s\n"
+ "fmax z30.s, p3/M, z30.s, z14.s\n"
+ "fmax z31.s, p3/M, z31.s, z14.s\n"
+ "fmin z16.s, p3/M, z16.s, z13.s\n"
+ "st1w { z16.s }, p0, [x15]\n"
+ "fmin z17.s, p3/M, z17.s, z13.s\n"
+ "fmin z18.s, p3/M, z18.s, z13.s\n"
+ "st1w { z17.s }, p0, [x15, x6, LSL #2]\n"
+ "fmin z19.s, p3/M, z19.s, z13.s\n"
+ "fmin z20.s, p3/M, z20.s, z13.s\n"
+ "st1w { z18.s }, p0, [x15, x25, LSL #2]\n"
+ "fmin z21.s, p3/M, z21.s, z13.s\n"
+ "fmin z22.s, p3/M, z22.s, z13.s\n"
+ "st1w { z19.s }, p0, [x15, x22, LSL #2]\n"
+ "fmin z23.s, p3/M, z23.s, z13.s\n"
+ "fmin z24.s, p3/M, z24.s, z13.s\n"
+ "st1w { z20.s }, p0, [x9]\n"
+ "fmin z25.s, p3/M, z25.s, z13.s\n"
+ "fmin z26.s, p3/M, z26.s, z13.s\n"
+ "st1w { z21.s }, p0, [x9, x6, LSL #2]\n"
+ "fmin z27.s, p3/M, z27.s, z13.s\n"
+ "fmin z28.s, p3/M, z28.s, z13.s\n"
+ "st1w { z22.s }, p0, [x9, x25, LSL #2]\n"
+ "fmin z29.s, p3/M, z29.s, z13.s\n"
+ "fmin z30.s, p3/M, z30.s, z13.s\n"
+ "st1w { z23.s }, p0, [x9, x22, LSL #2]\n"
+ "fmin z31.s, p3/M, z31.s, z13.s\n"
+ "st1w { z24.s }, p0, [x26]\n"
+ "st1w { z25.s }, p0, [x26, x6, LSL #2]\n"
+ "st1w { z26.s }, p0, [x26, x25, LSL #2]\n"
+ "st1w { z27.s }, p0, [x26, x22, LSL #2]\n"
+ "st1w { z28.s }, p0, [x23]\n"
+ "st1w { z29.s }, p0, [x23, x6, LSL #2]\n"
+ "st1w { z30.s }, p0, [x23, x25, LSL #2]\n"
+ "st1w { z31.s }, p0, [x23, x22, LSL #2]\n"
"blt 1b\n"
:
: [n_channels] "r" ((unsigned long) n_channels), [offsetof_args_inptr] "I" (offsetof(Args, inptr)), [offsetof_args_ld_input_col] "I" (offsetof(Args, ld_input_col)), [offsetof_args_ld_input_row] "I" (offsetof(Args, ld_input_row)), [offsetof_args_ld_output_col] "I" (offsetof(Args, ld_output_col)), [offsetof_args_ld_output_row] "I" (offsetof(Args, ld_output_row)), [offsetof_args_max] "I" (offsetof(Args, max)), [offsetof_args_min] "I" (offsetof(Args, min)), [offsetof_args_n_tile_cols] "I" (offsetof(Args, n_tile_cols)), [offsetof_args_n_tile_rows] "I" (offsetof(Args, n_tile_rows)), [offsetof_args_outptr] "I" (offsetof(Args, outptr)), [offsetof_args_params] "I" (offsetof(Args, params)), [offsetof_args_tile_i] "I" (offsetof(Args, tile_i)), [offsetof_args_tile_j] "I" (offsetof(Args, tile_j)), [params_struct] "r" (&params_struct)
- : "cc", "memory", "p0", "p1", "p2", "p3", "x2", "x3", "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"
+ : "cc", "memory", "p0", "p1", "p2", "p3", "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"
);
}