aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp
diff options
context:
space:
mode:
authorMichael Tyler <michael.tyler@arm.com>2022-12-15 12:39:29 +0000
committermichael.tyler <michael.tyler@arm.com>2023-01-16 09:31:00 +0000
commitba209750abc1ac7e42bab9fef5db284384d70fb3 (patch)
tree1065f242db9a9a5e48bd4a9f2fd68aef1924827a /src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp
parent8094f9dd5307c55f545b2cb41ec80a739a9b4d6f (diff)
downloadComputeLibrary-ba209750abc1ac7e42bab9fef5db284384d70fb3.tar.gz
Update CPU kernels to remove x19
Resolves: COMPMID-5805 Signed-off-by: Michael Tyler <michael.tyler@arm.com> Change-Id: I250f64531e209625e4ff176dd5a552c1c34bc484 Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8909 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> Benchmark: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp')
-rw-r--r--src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp652
1 files changed, 326 insertions, 326 deletions
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp
index c485b7dde0..cda34358f5 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2021 Arm Limited.
+ * Copyright (c) 2021, 2023 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -88,369 +88,369 @@ void sve_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst_direct_impl(
__asm__ __volatile__(
"ptrue p3.b\n"
- "mov x6, #0x0\n"
- "mov x7, #0x0\n"
+ "mov x13, #0x0\n"
+ "mov x8, #0x0\n"
"1:" // Tile loop
- "str x6, [%x[params_struct], %[offsetof_args_tile_i]]\n"
+ "str x13, [%x[params_struct], %[offsetof_args_tile_i]]\n"
+ "mov x25, #0x3\n"
"mov x24, #0x3\n"
- "str x7, [%x[params_struct], %[offsetof_args_tile_j]]\n"
- "mov x23, #0x3\n"
- "ldr x8, [%x[params_struct], %[offsetof_args_params]]\n"
- "mov x17, #0x0\n"
- "ldr x22, [%x[params_struct], %[offsetof_args_ld_input_row]]\n"
- "cntw x16\n"
- "ldr x15, [%x[params_struct], %[offsetof_args_ld_input_col]]\n"
- "sub x21, XZR, x16\n"
+ "str x8, [%x[params_struct], %[offsetof_args_tile_j]]\n"
+ "ldr x23, [%x[params_struct], %[offsetof_args_ld_input_row]]\n"
+ "ldr x17, [%x[params_struct], %[offsetof_args_ld_input_col]]\n"
+ "mul x22, x13, x23\n" // offset = tile_i * ld_input_row
+ "ldr x21, [%x[params_struct], %[offsetof_args_ld_output_row]]\n"
+ "madd x22, x8, x17, x22\n" // offset += tile_j * ld_input_col
+ "ldr x16, [%x[params_struct], %[offsetof_args_ld_output_col]]\n"
+ "cntw x15\n"
+ "mul x20, x13, x21\n" // offset = tile_i * ld_output_row
"ldr x14, [%x[params_struct], %[offsetof_args_inptr]]\n"
- "mul x19, x6, x22\n" // offset = tile_i * ld_input_row
- "ldr x20, [%x[params_struct], %[offsetof_args_ld_output_row]]\n"
- "madd x19, x7, x15, x19\n" // offset += tile_j * ld_input_col
- "ldr x13, [%x[params_struct], %[offsetof_args_ld_output_col]]\n"
- "mul x19, x19, x24\n" // offset *= kernel_stride * output_size
- "ldr x12, [%x[params_struct], %[offsetof_args_outptr]]\n"
- "add x14, x14, x19, LSL #2\n" // inptr[0] += offset * sizeof(float)
- "ld1rw { z18.s }, p3/Z, [%x[params_struct], %[offsetof_args_min]]\n"
- "add x11, x14, x22, LSL #2\n"
- "ld1rw { z17.s }, p3/Z, [%x[params_struct], %[offsetof_args_max]]\n"
- "add x10, x11, x22, LSL #2\n"
- "ld1w { z16.s }, p3/Z, [x8]\n"
- "add x9, x10, x22, LSL #2\n"
- "ld1w { z0.s }, p3/Z, [x8, #1, MUL VL]\n"
- "add x28, x9, x22, LSL #2\n"
- "ld1w { z1.s }, p3/Z, [x8, #2, MUL VL]\n"
- "add x27, x15, x15\n"
- "ld1w { z2.s }, p3/Z, [x8, #3, MUL VL]\n"
- "add x26, x27, x15\n"
- "ld1w { z3.s }, p3/Z, [x8, #4, MUL VL]\n"
- "add x25, x26, x15\n"
- "ld1w { z4.s }, p3/Z, [x8, #5, MUL VL]\n"
- "mul x19, x6, x20\n" // offset = tile_i * ld_output_row
- "ld1w { z5.s }, p3/Z, [x8, #6, MUL VL]\n"
- "madd x19, x7, x13, x19\n" // offset += tile_j * ld_output_col
- "ld1w { z6.s }, p3/Z, [x8, #7, MUL VL]\n"
- "mul x19, x19, x23\n" // offset *= output_tile_size
- "add x24, x13, x13\n"
- "add x12, x12, x19, LSL #2\n" // outptrs[0] += offset * sizeof(float)
- "add x23, x12, x20, LSL #2\n"
- "add x22, x23, x20, LSL #2\n"
+ "ldr x13, [%x[params_struct], %[offsetof_args_params]]\n"
+ "add x12, x17, x17\n"
+ "mul x22, x22, x25\n" // offset *= kernel_stride * output_size
+ "add x14, x14, x22, LSL #2\n" // inptr[0] += offset * sizeof(float)
+ "ldr x11, [%x[params_struct], %[offsetof_args_outptr]]\n"
+ "add x10, x14, x23, LSL #2\n"
+ "madd x20, x8, x16, x20\n" // offset += tile_j * ld_output_col
+ "add x9, x10, x23, LSL #2\n"
"whilelt p2.s, XZR, %x[n_channels]\n"
- "ld1w { z9.s }, p2/Z, [x10, x27, LSL #2]\n"
+ "ld1w { z18.s }, p3/Z, [x13]\n"
+ "mul x20, x20, x24\n" // offset *= output_tile_size
+ "ld1w { z0.s }, p3/Z, [x13, #1, MUL VL]\n"
+ "ld1w { z1.s }, p3/Z, [x13, #2, MUL VL]\n"
+ "add x28, x9, x23, LSL #2\n"
+ "ld1w { z2.s }, p3/Z, [x13, #3, MUL VL]\n"
+ "ld1w { z3.s }, p3/Z, [x13, #4, MUL VL]\n"
+ "add x27, x12, x17\n"
+ "add x11, x11, x20, LSL #2\n" // outptrs[0] += offset * sizeof(float)
+ "ld1w { z4.s }, p3/Z, [x13, #5, MUL VL]\n"
+ "ld1w { z5.s }, p3/Z, [x13, #6, MUL VL]\n"
+ "add x26, x28, x23, LSL #2\n"
+ "add x25, x27, x17\n"
+ "ld1w { z6.s }, p3/Z, [x13, #7, MUL VL]\n"
+ "addvl x13, x13, #16\n"
+ "add x24, x11, x21, LSL #2\n"
+ "ld1rw { z17.s }, p3/Z, [%x[params_struct], %[offsetof_args_min]]\n"
+ "cmp x15, %x[n_channels]\n"
+ "add x23, x24, x21, LSL #2\n"
+ "ld1rw { z16.s }, p3/Z, [%x[params_struct], %[offsetof_args_max]]\n"
+ "ld1w { z7.s }, p3/Z, [x13, #-8, MUL VL]\n"
+ "add x22, x16, x16\n"
+ "mov x21, #0x0\n"
+ "ld1w { z8.s }, p3/Z, [x13, #-7, MUL VL]\n"
+ "ld1w { z9.s }, p2/Z, [x9, x12, LSL #2]\n"
+ "sub x20, XZR, x15\n"
"ld1w { z10.s }, p2/Z, [x14]\n"
- "addvl x8, x8, #16\n"
"ld1w { z11.s }, p2/Z, [x14, x25, LSL #2]\n"
- "cmp x16, %x[n_channels]\n"
- "ld1w { z7.s }, p3/Z, [x8, #-8, MUL VL]\n"
- "ld1w { z8.s }, p3/Z, [x8, #-7, MUL VL]\n"
- "addvl x8, x8, #-6\n"
- "ld1w { z12.s }, p2/Z, [x28]\n"
- "ld1w { z13.s }, p2/Z, [x11, x27, LSL #2]\n"
+ "addvl x13, x13, #-6\n"
+ "ld1w { z12.s }, p2/Z, [x26]\n"
+ "ld1w { z13.s }, p2/Z, [x10, x12, LSL #2]\n"
"bge 3f\n"
"2:" // Tile loop: Channel loop
- "movprfx z31, z16\n fmla z31.s, p3/M, z8.s, z9.s\n"
- "whilelt p1.s, x16, %x[n_channels]\n"
- "movprfx z30, z16\n fmla z30.s, p3/M, z7.s, z9.s\n"
+ "movprfx z24, z18\n fmla z24.s, p3/M, z7.s, z9.s\n"
+ "movprfx z23, z18\n fmla z23.s, p3/M, z8.s, z9.s\n"
+ "whilelt p1.s, x15, %x[n_channels]\n"
"incw x21\n"
- "movprfx z29, z16\n fmla z29.s, p3/M, z6.s, z9.s\n"
+ "movprfx z25, z18\n fmla z25.s, p3/M, z6.s, z9.s\n"
+ "fmla z24.s, p3/M, z4.s, z13.s\n"
+ "incw x15\n"
"mov p0.b, p2.b\n"
- "movprfx z28, z16\n fmla z28.s, p3/M, z5.s, z9.s\n"
- "incw x17\n"
- "movprfx z27, z16\n fmla z27.s, p3/M, z4.s, z9.s\n"
- "incw x16\n"
- "movprfx z26, z16\n fmla z26.s, p3/M, z3.s, z9.s\n"
- "movprfx z25, z16\n fmla z25.s, p3/M, z2.s, z9.s\n"
- "movprfx z24, z16\n fmla z24.s, p3/M, z1.s, z9.s\n"
- "movprfx z23, z16\n fmla z23.s, p3/M, z0.s, z9.s\n"
- "ld1w { z16.s }, p3/Z, [x8]\n"
- "fmla z31.s, p3/M, z0.s, z10.s\n"
- "ld1w { z10.s }, p2/Z, [x10, x26, LSL #2]\n"
- "fmla z29.s, p3/M, z2.s, z11.s\n"
- "ld1w { z11.s }, p2/Z, [x10, x15, LSL #2]\n"
- "fmla z25.s, p3/M, z6.s, z12.s\n"
- "ld1w { z12.s }, p2/Z, [x28, x25, LSL #2]\n"
- "fmla z30.s, p3/M, z4.s, z13.s\n"
- "fmla z31.s, p3/M, z5.s, z13.s\n"
- "fmla z29.s, p3/M, z3.s, z13.s\n"
- "fmla z28.s, p3/M, z2.s, z13.s\n"
+ "movprfx z26, z18\n fmla z26.s, p3/M, z5.s, z9.s\n"
+ "movprfx z27, z18\n fmla z27.s, p3/M, z4.s, z9.s\n"
+ "incw x20\n"
+ "movprfx z28, z18\n fmla z28.s, p3/M, z3.s, z9.s\n"
+ "fmla z23.s, p3/M, z0.s, z10.s\n"
+ "ld1w { z10.s }, p2/Z, [x9, x27, LSL #2]\n"
+ "fmla z25.s, p3/M, z2.s, z11.s\n"
+ "ld1w { z11.s }, p2/Z, [x9, x17, LSL #2]\n"
+ "movprfx z29, z18\n fmla z29.s, p3/M, z2.s, z9.s\n"
+ "fmla z24.s, p3/M, z6.s, z11.s\n"
+ "movprfx z31, z18\n fmla z31.s, p3/M, z0.s, z9.s\n"
+ "fmla z23.s, p3/M, z5.s, z13.s\n"
+ "fmla z25.s, p3/M, z3.s, z13.s\n"
+ "fmla z26.s, p3/M, z2.s, z13.s\n"
"fmla z27.s, p3/M, z1.s, z13.s\n"
- "fmla z26.s, p3/M, z0.s, z13.s\n"
- "ld1w { z13.s }, p2/Z, [x14, x15, LSL #2]\n"
- "fmla z23.s, p3/M, z8.s, z12.s\n"
- "ld1w { z12.s }, p2/Z, [x14, x26, LSL #2]\n"
- "fmla z31.s, p3/M, z7.s, z11.s\n"
- "fmla z30.s, p3/M, z6.s, z11.s\n"
- "fmla z28.s, p3/M, z4.s, z11.s\n"
+ "fmla z28.s, p3/M, z0.s, z13.s\n"
+ "ld1w { z13.s }, p2/Z, [x14, x17, LSL #2]\n"
+ "fmla z29.s, p3/M, z6.s, z12.s\n"
+ "ld1w { z12.s }, p2/Z, [x26, x25, LSL #2]\n"
+ "movprfx z30, z18\n fmla z30.s, p3/M, z1.s, z9.s\n"
+ "fmla z24.s, p3/M, z0.s, z13.s\n"
+ "ld1w { z18.s }, p3/Z, [x13]\n"
+ "fmla z31.s, p3/M, z8.s, z12.s\n"
+ "ld1w { z12.s }, p2/Z, [x14, x27, LSL #2]\n"
+ "fmla z23.s, p3/M, z7.s, z11.s\n"
+ "fmla z30.s, p3/M, z0.s, z11.s\n"
+ "fmla z26.s, p3/M, z4.s, z11.s\n"
"fmla z27.s, p3/M, z3.s, z11.s\n"
- "fmla z25.s, p3/M, z1.s, z11.s\n"
- "fmla z24.s, p3/M, z0.s, z11.s\n"
- "ld1w { z11.s }, p2/Z, [x11]\n"
- "fmla z31.s, p3/M, z1.s, z13.s\n"
- "fmla z30.s, p3/M, z0.s, z13.s\n"
- "ld1w { z13.s }, p2/Z, [x11, x25, LSL #2]\n"
- "fmla z29.s, p3/M, z1.s, z12.s\n"
+ "fmla z29.s, p3/M, z1.s, z11.s\n"
+ "ld1w { z11.s }, p2/Z, [x10]\n"
+ "fmla z24.s, p3/M, z2.s, z12.s\n"
+ "fmla z25.s, p3/M, z1.s, z12.s\n"
+ "ld1w { z12.s }, p2/Z, [x28]\n"
+ "fmla z28.s, p3/M, z4.s, z10.s\n"
+ "fmla z23.s, p3/M, z1.s, z13.s\n"
+ "ld1w { z13.s }, p2/Z, [x10, x25, LSL #2]\n"
+ "fmla z30.s, p3/M, z2.s, z10.s\n"
+ "fmla z31.s, p3/M, z1.s, z10.s\n"
+ "fmla z24.s, p3/M, z8.s, z10.s\n"
+ "fmla z25.s, p3/M, z7.s, z10.s\n"
"fmla z27.s, p3/M, z5.s, z10.s\n"
- "fmla z26.s, p3/M, z4.s, z10.s\n"
- "fmla z30.s, p3/M, z2.s, z12.s\n"
- "ld1w { z12.s }, p2/Z, [x9]\n"
- "fmla z29.s, p3/M, z7.s, z10.s\n"
- "fmla z24.s, p3/M, z2.s, z10.s\n"
- "fmla z23.s, p3/M, z1.s, z10.s\n"
- "fmla z30.s, p3/M, z8.s, z10.s\n"
- "ld1w { z10.s }, p2/Z, [x9, x27, LSL #2]\n"
- "fmla z31.s, p3/M, z3.s, z11.s\n"
- "fmla z28.s, p3/M, z0.s, z11.s\n"
- "ld1w { z11.s }, p2/Z, [x9, x25, LSL #2]\n"
- "fmla z29.s, p3/M, z5.s, z13.s\n"
- "fmla z26.s, p3/M, z2.s, z13.s\n"
- "ld1w { z13.s }, p2/Z, [x28, x15, LSL #2]\n"
- "fmla z25.s, p3/M, z3.s, z12.s\n"
- "fmla z28.s, p3/M, z6.s, z12.s\n"
- "ld1w { z12.s }, p2/Z, [x11, x15, LSL #2]\n"
+ "fmla z26.s, p3/M, z0.s, z11.s\n"
+ "ld1w { z10.s }, p2/Z, [x28, x12, LSL #2]\n"
+ "fmla z29.s, p3/M, z3.s, z12.s\n"
+ "fmla z28.s, p3/M, z2.s, z13.s\n"
+ "fmla z30.s, p3/M, z4.s, z10.s\n"
+ "fmla z31.s, p3/M, z3.s, z10.s\n"
+ "fmla z23.s, p3/M, z3.s, z11.s\n"
+ "fmla z25.s, p3/M, z5.s, z13.s\n"
+ "ld1w { z11.s }, p2/Z, [x28, x25, LSL #2]\n"
+ "ld1w { z13.s }, p2/Z, [x26, x17, LSL #2]\n"
+ "fmla z26.s, p3/M, z6.s, z12.s\n"
"fmla z27.s, p3/M, z7.s, z10.s\n"
- "fmla z26.s, p3/M, z6.s, z10.s\n"
- "fmla z25.s, p3/M, z5.s, z10.s\n"
- "fmla z28.s, p3/M, z8.s, z10.s\n"
- "fmla z24.s, p3/M, z4.s, z10.s\n"
- "fmla z23.s, p3/M, z3.s, z10.s\n"
- "fmla z26.s, p3/M, z8.s, z11.s\n"
- "fmla z25.s, p3/M, z7.s, z13.s\n"
- "fmla z24.s, p3/M, z6.s, z13.s\n"
- "ld1w { z13.s }, p2/Z, [x28, x26, LSL #2]\n"
- "fmla z23.s, p3/M, z5.s, z11.s\n"
- "ld1w { z11.s }, p2/Z, [x11, x26, LSL #2]\n"
- "addvl x11, x11, #1\n"
- "fmla z31.s, p3/M, z4.s, z12.s\n"
- "fmla z30.s, p3/M, z3.s, z12.s\n"
- "fmla z28.s, p3/M, z1.s, z12.s\n"
+ "ld1w { z12.s }, p2/Z, [x10, x17, LSL #2]\n"
+ "fmla z29.s, p3/M, z5.s, z10.s\n"
+ "fmla z28.s, p3/M, z6.s, z10.s\n"
+ "fmla z31.s, p3/M, z5.s, z11.s\n"
+ "fmla z30.s, p3/M, z6.s, z13.s\n"
+ "fmla z26.s, p3/M, z8.s, z10.s\n"
+ "fmla z29.s, p3/M, z7.s, z13.s\n"
+ "ld1w { z13.s }, p2/Z, [x26, x27, LSL #2]\n"
+ "fmla z24.s, p3/M, z3.s, z12.s\n"
"fmla z27.s, p3/M, z0.s, z12.s\n"
- "ld1w { z12.s }, p2/Z, [x9, x15, LSL #2]\n"
- "fmla z29.s, p3/M, z4.s, z11.s\n"
- "fmla z30.s, p3/M, z5.s, z11.s\n"
- "fmla z26.s, p3/M, z1.s, z11.s\n"
+ "fmla z28.s, p3/M, z8.s, z11.s\n"
+ "ld1w { z11.s }, p2/Z, [x10, x27, LSL #2]\n"
+ "fmla z30.s, p3/M, z8.s, z13.s\n"
+ "addvl x10, x10, #1\n"
+ "fmla z31.s, p3/M, z7.s, z13.s\n"
+ "fmla z23.s, p3/M, z4.s, z12.s\n"
+ "ld1w { z13.s }, p2/Z, [x28, x27, LSL #2]\n"
+ "fmla z26.s, p3/M, z1.s, z12.s\n"
+ "fmla z24.s, p3/M, z5.s, z11.s\n"
+ "ld1w { z12.s }, p2/Z, [x28, x17, LSL #2]\n"
+ "addvl x28, x28, #1\n"
+ "fmla z25.s, p3/M, z4.s, z11.s\n"
"fmla z27.s, p3/M, z2.s, z11.s\n"
- "ld1w { z11.s }, p2/Z, [x14, x27, LSL #2]\n"
+ "fmla z28.s, p3/M, z1.s, z11.s\n"
+ "ld1w { z11.s }, p2/Z, [x14, x12, LSL #2]\n"
+ "fmla z29.s, p3/M, z4.s, z12.s\n"
"addvl x14, x14, #1\n"
- "fmla z24.s, p3/M, z8.s, z13.s\n"
+ "fmla z30.s, p3/M, z3.s, z12.s\n"
+ "fmla z31.s, p3/M, z4.s, z13.s\n"
+ "ld1w { z4.s }, p3/Z, [x13, #5, MUL VL]\n"
"ld1w { z10.s }, p1/Z, [x14]\n"
- "fmla z23.s, p3/M, z7.s, z13.s\n"
- "ld1w { z13.s }, p2/Z, [x9, x26, LSL #2]\n"
- "addvl x9, x9, #1\n"
- "fmla z28.s, p3/M, z7.s, z12.s\n"
+ "fmla z26.s, p3/M, z7.s, z12.s\n"
"fmla z27.s, p3/M, z6.s, z12.s\n"
- "fmla z25.s, p3/M, z4.s, z12.s\n"
- "fmla z24.s, p3/M, z3.s, z12.s\n"
- "ld1w { z12.s }, p2/Z, [x10]\n"
+ "ld1w { z12.s }, p2/Z, [x9]\n"
+ "fmla z23.s, p3/M, z2.s, z11.s\n"
+ "fmla z24.s, p3/M, z1.s, z11.s\n"
+ "fmax z24.s, p3/M, z24.s, z17.s\n"
+ "ld1w { z1.s }, p3/Z, [x13, #2, MUL VL]\n"
+ "fmla z25.s, p3/M, z0.s, z11.s\n"
+ "ld1w { z11.s }, p2/Z, [x9, x25, LSL #2]\n"
+ "fmla z28.s, p3/M, z7.s, z13.s\n"
+ "addvl x9, x9, #1\n"
+ "fmla z30.s, p3/M, z5.s, z13.s\n"
+ "fmla z29.s, p3/M, z0.s, z12.s\n"
+ "ld1w { z0.s }, p3/Z, [x13, #1, MUL VL]\n"
+ "fmin z24.s, p3/M, z24.s, z16.s\n"
"fmla z31.s, p3/M, z2.s, z11.s\n"
- "fmla z30.s, p3/M, z1.s, z11.s\n"
- "ld1w { z1.s }, p3/Z, [x8, #2, MUL VL]\n"
- "fmla z29.s, p3/M, z0.s, z11.s\n"
- "ld1w { z11.s }, p2/Z, [x10, x25, LSL #2]\n"
- "addvl x10, x10, #1\n"
"fmla z27.s, p3/M, z8.s, z13.s\n"
- "ld1w { z9.s }, p1/Z, [x10, x27, LSL #2]\n"
- "fmla z26.s, p3/M, z7.s, z13.s\n"
- "fmla z24.s, p3/M, z5.s, z13.s\n"
- "fmla z23.s, p3/M, z4.s, z13.s\n"
- "ld1w { z13.s }, p2/Z, [x28, x27, LSL #2]\n"
- "whilelt p2.s, x17, %x[n_channels]\n"
- "fmla z31.s, p3/M, z6.s, z12.s\n"
- "ld1w { z4.s }, p3/Z, [x8, #5, MUL VL]\n"
- "addvl x28, x28, #1\n"
- "fmla z28.s, p3/M, z3.s, z12.s\n"
- "ld1w { z3.s }, p3/Z, [x8, #4, MUL VL]\n"
- "cmp x16, %x[n_channels]\n"
- "fmla z25.s, p3/M, z0.s, z12.s\n"
- "ld1w { z12.s }, p1/Z, [x28]\n"
- "fmla z29.s, p3/M, z8.s, z11.s\n"
- "ld1w { z0.s }, p3/Z, [x8, #1, MUL VL]\n"
- "fmla z26.s, p3/M, z5.s, z11.s\n"
- "ld1w { z5.s }, p3/Z, [x8, #6, MUL VL]\n"
- "fmla z23.s, p3/M, z2.s, z11.s\n"
+ "ld1w { z13.s }, p2/Z, [x26, x12, LSL #2]\n"
+ "fmax z27.s, p3/M, z27.s, z17.s\n"
+ "fmla z23.s, p3/M, z6.s, z12.s\n"
+ "fmla z26.s, p3/M, z3.s, z12.s\n"
+ "fmax z23.s, p3/M, z23.s, z17.s\n"
+ "fmax z26.s, p3/M, z26.s, z17.s\n"
+ "fmla z25.s, p3/M, z8.s, z11.s\n"
+ "fmla z28.s, p3/M, z5.s, z11.s\n"
+ "fmax z25.s, p3/M, z25.s, z17.s\n"
+ "fmax z28.s, p3/M, z28.s, z17.s\n"
+ "fmla z29.s, p3/M, z8.s, z13.s\n"
+ "fmla z30.s, p3/M, z7.s, z13.s\n"
+ "fmax z29.s, p3/M, z29.s, z17.s\n"
+ "fmax z30.s, p3/M, z30.s, z17.s\n"
+ "fmla z31.s, p3/M, z6.s, z13.s\n"
+ "fmax z31.s, p3/M, z31.s, z17.s\n"
+ "addvl x26, x26, #1\n"
+ "ld1w { z2.s }, p3/Z, [x13, #3, MUL VL]\n"
+ "ld1w { z3.s }, p3/Z, [x13, #4, MUL VL]\n"
+ "ld1w { z5.s }, p3/Z, [x13, #6, MUL VL]\n"
+ "whilelt p2.s, x21, %x[n_channels]\n"
+ "cmp x15, %x[n_channels]\n"
+ "ld1w { z6.s }, p3/Z, [x13, #7, MUL VL]\n"
+ "addvl x13, x13, #16\n"
+ "fmin z23.s, p3/M, z23.s, z16.s\n"
+ "ld1w { z9.s }, p1/Z, [x9, x12, LSL #2]\n"
+ "fmin z25.s, p3/M, z25.s, z16.s\n"
+ "fmin z26.s, p3/M, z26.s, z16.s\n"
"ld1w { z11.s }, p1/Z, [x14, x25, LSL #2]\n"
- "fmla z24.s, p3/M, z7.s, z13.s\n"
- "ld1w { z2.s }, p3/Z, [x8, #3, MUL VL]\n"
- "fmla z25.s, p3/M, z8.s, z13.s\n"
- "fmax z31.s, p3/M, z31.s, z18.s\n"
- "fmla z23.s, p3/M, z6.s, z13.s\n"
- "ld1w { z13.s }, p1/Z, [x11, x27, LSL #2]\n"
- "fmax z30.s, p3/M, z30.s, z18.s\n"
- "ld1w { z6.s }, p3/Z, [x8, #7, MUL VL]\n"
- "addvl x8, x8, #16\n"
- "fmin z31.s, p3/M, z31.s, z17.s\n"
- "ld1w { z7.s }, p3/Z, [x8, #-8, MUL VL]\n"
- "fmax z29.s, p3/M, z29.s, z18.s\n"
- "ld1w { z8.s }, p3/Z, [x8, #-7, MUL VL]\n"
- "addvl x8, x8, #-6\n"
- "fmin z30.s, p3/M, z30.s, z17.s\n"
- "st1w { z31.s }, p0, [x12]\n"
- "fmin z29.s, p3/M, z29.s, z17.s\n"
- "fmax z28.s, p3/M, z28.s, z18.s\n"
- "st1w { z30.s }, p0, [x12, x13, LSL #2]\n"
- "fmax z27.s, p3/M, z27.s, z18.s\n"
- "fmax z26.s, p3/M, z26.s, z18.s\n"
- "st1w { z29.s }, p0, [x12, x24, LSL #2]\n"
- "fmin z28.s, p3/M, z28.s, z17.s\n"
- "addvl x12, x12, #1\n"
- "fmax z25.s, p3/M, z25.s, z18.s\n"
- "st1w { z28.s }, p0, [x23]\n"
- "fmin z27.s, p3/M, z27.s, z17.s\n"
- "fmin z26.s, p3/M, z26.s, z17.s\n"
- "st1w { z27.s }, p0, [x23, x13, LSL #2]\n"
- "fmin z25.s, p3/M, z25.s, z17.s\n"
- "fmax z24.s, p3/M, z24.s, z18.s\n"
- "st1w { z26.s }, p0, [x23, x24, LSL #2]\n"
+ "ld1w { z12.s }, p1/Z, [x26]\n"
+ "fmin z27.s, p3/M, z27.s, z16.s\n"
+ "fmin z28.s, p3/M, z28.s, z16.s\n"
+ "ld1w { z13.s }, p1/Z, [x10, x12, LSL #2]\n"
+ "st1w { z23.s }, p0, [x11]\n"
+ "fmin z29.s, p3/M, z29.s, z16.s\n"
+ "fmin z30.s, p3/M, z30.s, z16.s\n"
+ "st1w { z24.s }, p0, [x11, x16, LSL #2]\n"
+ "ld1w { z7.s }, p3/Z, [x13, #-8, MUL VL]\n"
+ "fmin z31.s, p3/M, z31.s, z16.s\n"
+ "st1w { z25.s }, p0, [x11, x22, LSL #2]\n"
+ "addvl x11, x11, #1\n"
+ "ld1w { z8.s }, p3/Z, [x13, #-7, MUL VL]\n"
+ "st1w { z26.s }, p0, [x24]\n"
+ "addvl x13, x13, #-6\n"
+ "st1w { z27.s }, p0, [x24, x16, LSL #2]\n"
+ "st1w { z28.s }, p0, [x24, x22, LSL #2]\n"
+ "addvl x24, x24, #1\n"
+ "st1w { z29.s }, p0, [x23]\n"
+ "st1w { z30.s }, p0, [x23, x16, LSL #2]\n"
+ "st1w { z31.s }, p0, [x23, x22, LSL #2]\n"
"addvl x23, x23, #1\n"
- "fmax z23.s, p3/M, z23.s, z18.s\n"
- "st1w { z25.s }, p0, [x22]\n"
- "fmin z24.s, p3/M, z24.s, z17.s\n"
- "fmin z23.s, p3/M, z23.s, z17.s\n"
- "st1w { z24.s }, p0, [x22, x13, LSL #2]\n"
- "st1w { z23.s }, p0, [x22, x24, LSL #2]\n"
- "addvl x22, x22, #1\n"
"blt 2b\n"
"3:" // Tile loop: Channel tail
- "movprfx z31, z16\n fmla z31.s, p3/M, z8.s, z9.s\n"
- "ldr x6, [%x[params_struct], %[offsetof_args_tile_i]]\n"
- "mov p0.b, p2.b\n"
- "movprfx z30, z16\n fmla z30.s, p3/M, z7.s, z9.s\n"
- "ldr x7, [%x[params_struct], %[offsetof_args_tile_j]]\n"
- "add x21, x6, #0x1\n"
- "movprfx z29, z16\n fmla z29.s, p3/M, z6.s, z9.s\n"
+ "movprfx z24, z18\n fmla z24.s, p3/M, z7.s, z9.s\n"
+ "movprfx z23, z18\n fmla z23.s, p3/M, z8.s, z9.s\n"
+ "ldr x8, [%x[params_struct], %[offsetof_args_tile_j]]\n"
+ "ldr x13, [%x[params_struct], %[offsetof_args_tile_i]]\n"
+ "movprfx z25, z18\n fmla z25.s, p3/M, z6.s, z9.s\n"
+ "fmla z24.s, p3/M, z4.s, z13.s\n"
+ "ldr x20, [%x[params_struct], %[offsetof_args_n_tile_cols]]\n"
+ "add x8, x8, #0x1\n"
+ "movprfx z26, z18\n fmla z26.s, p3/M, z5.s, z9.s\n"
+ "movprfx z27, z18\n fmla z27.s, p3/M, z4.s, z9.s\n"
+ "cmp x8, x20\n"
+ "add x21, x13, #0x1\n"
+ "movprfx z28, z18\n fmla z28.s, p3/M, z3.s, z9.s\n"
+ "fmla z23.s, p3/M, z0.s, z10.s\n"
+ "ld1w { z10.s }, p2/Z, [x9, x27, LSL #2]\n"
"ldr x20, [%x[params_struct], %[offsetof_args_n_tile_rows]]\n"
- "movprfx z28, z16\n fmla z28.s, p3/M, z5.s, z9.s\n"
- "ldr x19, [%x[params_struct], %[offsetof_args_n_tile_cols]]\n"
- "add x7, x7, #0x1\n"
- "movprfx z27, z16\n fmla z27.s, p3/M, z4.s, z9.s\n"
- "cmp x7, x19\n"
- "movprfx z26, z16\n fmla z26.s, p3/M, z3.s, z9.s\n"
- "movprfx z25, z16\n fmla z25.s, p3/M, z2.s, z9.s\n"
- "csel x7, x7, XZR, LT\n"
- "movprfx z24, z16\n fmla z24.s, p3/M, z1.s, z9.s\n"
- "csel x6, x6, x21, LT\n"
- "movprfx z23, z16\n fmla z23.s, p3/M, z0.s, z9.s\n"
- "cmp x6, x20\n"
- "fmla z31.s, p3/M, z0.s, z10.s\n"
- "ld1w { z10.s }, p2/Z, [x10, x26, LSL #2]\n"
- "fmla z29.s, p3/M, z2.s, z11.s\n"
- "ld1w { z11.s }, p2/Z, [x10, x15, LSL #2]\n"
- "fmla z25.s, p3/M, z6.s, z12.s\n"
- "ld1w { z12.s }, p2/Z, [x28, x25, LSL #2]\n"
- "fmla z30.s, p3/M, z4.s, z13.s\n"
- "fmla z31.s, p3/M, z5.s, z13.s\n"
- "fmla z29.s, p3/M, z3.s, z13.s\n"
- "fmla z28.s, p3/M, z2.s, z13.s\n"
+ "fmla z25.s, p3/M, z2.s, z11.s\n"
+ "ld1w { z11.s }, p2/Z, [x9, x17, LSL #2]\n"
+ "movprfx z29, z18\n fmla z29.s, p3/M, z2.s, z9.s\n"
+ "csel x13, x13, x21, LT\n"
+ "fmla z24.s, p3/M, z6.s, z11.s\n"
+ "movprfx z31, z18\n fmla z31.s, p3/M, z0.s, z9.s\n"
+ "mov p0.b, p2.b\n"
+ "csel x8, x8, XZR, LT\n"
+ "fmla z23.s, p3/M, z5.s, z13.s\n"
+ "fmla z25.s, p3/M, z3.s, z13.s\n"
+ "cmp x13, x20\n"
+ "fmla z26.s, p3/M, z2.s, z13.s\n"
"fmla z27.s, p3/M, z1.s, z13.s\n"
- "fmla z26.s, p3/M, z0.s, z13.s\n"
- "ld1w { z13.s }, p2/Z, [x14, x15, LSL #2]\n"
- "fmla z23.s, p3/M, z8.s, z12.s\n"
- "ld1w { z12.s }, p2/Z, [x14, x26, LSL #2]\n"
- "fmla z31.s, p3/M, z7.s, z11.s\n"
- "fmla z30.s, p3/M, z6.s, z11.s\n"
- "fmla z28.s, p3/M, z4.s, z11.s\n"
+ "fmla z28.s, p3/M, z0.s, z13.s\n"
+ "ld1w { z13.s }, p2/Z, [x14, x17, LSL #2]\n"
+ "fmla z29.s, p3/M, z6.s, z12.s\n"
+ "ld1w { z12.s }, p2/Z, [x26, x25, LSL #2]\n"
+ "movprfx z30, z18\n fmla z30.s, p3/M, z1.s, z9.s\n"
+ "fmla z24.s, p3/M, z0.s, z13.s\n"
+ "fmla z31.s, p3/M, z8.s, z12.s\n"
+ "ld1w { z12.s }, p2/Z, [x14, x27, LSL #2]\n"
+ "fmla z23.s, p3/M, z7.s, z11.s\n"
+ "fmla z30.s, p3/M, z0.s, z11.s\n"
+ "fmla z26.s, p3/M, z4.s, z11.s\n"
"fmla z27.s, p3/M, z3.s, z11.s\n"
- "fmla z25.s, p3/M, z1.s, z11.s\n"
- "fmla z24.s, p3/M, z0.s, z11.s\n"
- "ld1w { z11.s }, p2/Z, [x11]\n"
- "fmla z31.s, p3/M, z1.s, z13.s\n"
- "fmla z30.s, p3/M, z0.s, z13.s\n"
- "ld1w { z13.s }, p2/Z, [x11, x25, LSL #2]\n"
- "fmla z29.s, p3/M, z1.s, z12.s\n"
+ "fmla z29.s, p3/M, z1.s, z11.s\n"
+ "ld1w { z11.s }, p2/Z, [x10]\n"
+ "fmla z24.s, p3/M, z2.s, z12.s\n"
+ "fmla z25.s, p3/M, z1.s, z12.s\n"
+ "ld1w { z12.s }, p2/Z, [x28]\n"
+ "fmla z28.s, p3/M, z4.s, z10.s\n"
+ "fmla z23.s, p3/M, z1.s, z13.s\n"
+ "ld1w { z13.s }, p2/Z, [x10, x25, LSL #2]\n"
+ "fmla z30.s, p3/M, z2.s, z10.s\n"
+ "fmla z31.s, p3/M, z1.s, z10.s\n"
+ "fmla z24.s, p3/M, z8.s, z10.s\n"
+ "fmla z25.s, p3/M, z7.s, z10.s\n"
"fmla z27.s, p3/M, z5.s, z10.s\n"
- "fmla z26.s, p3/M, z4.s, z10.s\n"
- "fmla z30.s, p3/M, z2.s, z12.s\n"
- "ld1w { z12.s }, p2/Z, [x9]\n"
- "fmla z29.s, p3/M, z7.s, z10.s\n"
- "fmla z24.s, p3/M, z2.s, z10.s\n"
- "fmla z23.s, p3/M, z1.s, z10.s\n"
- "fmla z30.s, p3/M, z8.s, z10.s\n"
- "ld1w { z10.s }, p2/Z, [x9, x27, LSL #2]\n"
- "fmla z31.s, p3/M, z3.s, z11.s\n"
- "fmla z28.s, p3/M, z0.s, z11.s\n"
- "ld1w { z11.s }, p2/Z, [x9, x25, LSL #2]\n"
- "fmla z29.s, p3/M, z5.s, z13.s\n"
- "fmla z26.s, p3/M, z2.s, z13.s\n"
- "ld1w { z13.s }, p2/Z, [x28, x15, LSL #2]\n"
- "fmla z25.s, p3/M, z3.s, z12.s\n"
- "fmla z28.s, p3/M, z6.s, z12.s\n"
- "ld1w { z12.s }, p2/Z, [x11, x15, LSL #2]\n"
+ "fmla z26.s, p3/M, z0.s, z11.s\n"
+ "ld1w { z10.s }, p2/Z, [x28, x12, LSL #2]\n"
+ "fmla z29.s, p3/M, z3.s, z12.s\n"
+ "fmla z28.s, p3/M, z2.s, z13.s\n"
+ "fmla z30.s, p3/M, z4.s, z10.s\n"
+ "fmla z31.s, p3/M, z3.s, z10.s\n"
+ "fmla z23.s, p3/M, z3.s, z11.s\n"
+ "fmla z25.s, p3/M, z5.s, z13.s\n"
+ "ld1w { z11.s }, p2/Z, [x28, x25, LSL #2]\n"
+ "ld1w { z13.s }, p2/Z, [x26, x17, LSL #2]\n"
+ "fmla z26.s, p3/M, z6.s, z12.s\n"
"fmla z27.s, p3/M, z7.s, z10.s\n"
- "fmla z26.s, p3/M, z6.s, z10.s\n"
- "fmla z25.s, p3/M, z5.s, z10.s\n"
- "fmla z28.s, p3/M, z8.s, z10.s\n"
- "fmla z24.s, p3/M, z4.s, z10.s\n"
- "fmla z23.s, p3/M, z3.s, z10.s\n"
- "fmla z26.s, p3/M, z8.s, z11.s\n"
- "fmla z25.s, p3/M, z7.s, z13.s\n"
- "fmla z24.s, p3/M, z6.s, z13.s\n"
- "ld1w { z13.s }, p2/Z, [x28, x26, LSL #2]\n"
- "fmla z23.s, p3/M, z5.s, z11.s\n"
- "ld1w { z11.s }, p2/Z, [x11, x26, LSL #2]\n"
- "fmla z31.s, p3/M, z4.s, z12.s\n"
- "fmla z30.s, p3/M, z3.s, z12.s\n"
- "fmla z28.s, p3/M, z1.s, z12.s\n"
+ "ld1w { z12.s }, p2/Z, [x10, x17, LSL #2]\n"
+ "fmla z29.s, p3/M, z5.s, z10.s\n"
+ "fmla z28.s, p3/M, z6.s, z10.s\n"
+ "fmla z31.s, p3/M, z5.s, z11.s\n"
+ "fmla z30.s, p3/M, z6.s, z13.s\n"
+ "fmla z26.s, p3/M, z8.s, z10.s\n"
+ "fmla z29.s, p3/M, z7.s, z13.s\n"
+ "ld1w { z13.s }, p2/Z, [x26, x27, LSL #2]\n"
+ "fmla z24.s, p3/M, z3.s, z12.s\n"
"fmla z27.s, p3/M, z0.s, z12.s\n"
- "ld1w { z12.s }, p2/Z, [x9, x15, LSL #2]\n"
- "fmla z29.s, p3/M, z4.s, z11.s\n"
- "fmla z30.s, p3/M, z5.s, z11.s\n"
- "fmla z26.s, p3/M, z1.s, z11.s\n"
+ "fmla z28.s, p3/M, z8.s, z11.s\n"
+ "ld1w { z11.s }, p2/Z, [x10, x27, LSL #2]\n"
+ "fmla z30.s, p3/M, z8.s, z13.s\n"
+ "fmla z31.s, p3/M, z7.s, z13.s\n"
+ "fmla z23.s, p3/M, z4.s, z12.s\n"
+ "ld1w { z13.s }, p2/Z, [x28, x27, LSL #2]\n"
+ "fmla z26.s, p3/M, z1.s, z12.s\n"
+ "fmla z24.s, p3/M, z5.s, z11.s\n"
+ "ld1w { z12.s }, p2/Z, [x28, x17, LSL #2]\n"
+ "fmla z25.s, p3/M, z4.s, z11.s\n"
"fmla z27.s, p3/M, z2.s, z11.s\n"
- "ld1w { z11.s }, p2/Z, [x14, x27, LSL #2]\n"
- "fmla z24.s, p3/M, z8.s, z13.s\n"
- "fmla z23.s, p3/M, z7.s, z13.s\n"
- "ld1w { z13.s }, p2/Z, [x9, x26, LSL #2]\n"
- "fmla z28.s, p3/M, z7.s, z12.s\n"
+ "fmla z28.s, p3/M, z1.s, z11.s\n"
+ "ld1w { z11.s }, p2/Z, [x14, x12, LSL #2]\n"
+ "fmla z29.s, p3/M, z4.s, z12.s\n"
+ "fmla z30.s, p3/M, z3.s, z12.s\n"
+ "fmla z31.s, p3/M, z4.s, z13.s\n"
+ "fmla z26.s, p3/M, z7.s, z12.s\n"
"fmla z27.s, p3/M, z6.s, z12.s\n"
- "fmla z25.s, p3/M, z4.s, z12.s\n"
- "fmla z24.s, p3/M, z3.s, z12.s\n"
- "ld1w { z12.s }, p2/Z, [x10]\n"
+ "ld1w { z12.s }, p2/Z, [x9]\n"
+ "fmla z23.s, p3/M, z2.s, z11.s\n"
+ "fmla z24.s, p3/M, z1.s, z11.s\n"
+ "fmax z24.s, p3/M, z24.s, z17.s\n"
+ "fmin z24.s, p3/M, z24.s, z16.s\n"
+ "fmla z25.s, p3/M, z0.s, z11.s\n"
+ "ld1w { z11.s }, p2/Z, [x9, x25, LSL #2]\n"
+ "fmla z28.s, p3/M, z7.s, z13.s\n"
+ "fmla z30.s, p3/M, z5.s, z13.s\n"
+ "fmla z29.s, p3/M, z0.s, z12.s\n"
"fmla z31.s, p3/M, z2.s, z11.s\n"
- "fmla z30.s, p3/M, z1.s, z11.s\n"
- "fmla z29.s, p3/M, z0.s, z11.s\n"
- "ld1w { z11.s }, p2/Z, [x10, x25, LSL #2]\n"
"fmla z27.s, p3/M, z8.s, z13.s\n"
- "fmla z26.s, p3/M, z7.s, z13.s\n"
- "fmla z24.s, p3/M, z5.s, z13.s\n"
- "fmla z23.s, p3/M, z4.s, z13.s\n"
- "ld1w { z13.s }, p2/Z, [x28, x27, LSL #2]\n"
- "fmla z31.s, p3/M, z6.s, z12.s\n"
- "fmla z28.s, p3/M, z3.s, z12.s\n"
- "fmla z25.s, p3/M, z0.s, z12.s\n"
- "fmla z29.s, p3/M, z8.s, z11.s\n"
- "fmla z26.s, p3/M, z5.s, z11.s\n"
- "fmla z23.s, p3/M, z2.s, z11.s\n"
- "fmla z25.s, p3/M, z8.s, z13.s\n"
- "fmla z24.s, p3/M, z7.s, z13.s\n"
- "fmax z31.s, p3/M, z31.s, z18.s\n"
- "fmla z23.s, p3/M, z6.s, z13.s\n"
- "fmax z30.s, p3/M, z30.s, z18.s\n"
- "fmax z29.s, p3/M, z29.s, z18.s\n"
- "fmin z31.s, p3/M, z31.s, z17.s\n"
- "st1w { z31.s }, p0, [x12]\n"
- "fmin z30.s, p3/M, z30.s, z17.s\n"
- "fmin z29.s, p3/M, z29.s, z17.s\n"
- "st1w { z30.s }, p0, [x12, x13, LSL #2]\n"
- "fmax z28.s, p3/M, z28.s, z18.s\n"
- "fmax z27.s, p3/M, z27.s, z18.s\n"
- "st1w { z29.s }, p0, [x12, x24, LSL #2]\n"
- "fmax z26.s, p3/M, z26.s, z18.s\n"
- "fmax z25.s, p3/M, z25.s, z18.s\n"
- "fmax z24.s, p3/M, z24.s, z18.s\n"
- "fmin z28.s, p3/M, z28.s, z17.s\n"
- "st1w { z28.s }, p0, [x23]\n"
- "fmin z27.s, p3/M, z27.s, z17.s\n"
- "fmin z26.s, p3/M, z26.s, z17.s\n"
- "st1w { z27.s }, p0, [x23, x13, LSL #2]\n"
- "fmin z25.s, p3/M, z25.s, z17.s\n"
- "fmin z24.s, p3/M, z24.s, z17.s\n"
- "st1w { z26.s }, p0, [x23, x24, LSL #2]\n"
- "fmax z23.s, p3/M, z23.s, z18.s\n"
- "st1w { z25.s }, p0, [x22]\n"
- "fmin z23.s, p3/M, z23.s, z17.s\n"
- "st1w { z24.s }, p0, [x22, x13, LSL #2]\n"
- "st1w { z23.s }, p0, [x22, x24, LSL #2]\n"
+ "ld1w { z13.s }, p2/Z, [x26, x12, LSL #2]\n"
+ "fmax z27.s, p3/M, z27.s, z17.s\n"
+ "fmla z23.s, p3/M, z6.s, z12.s\n"
+ "fmla z26.s, p3/M, z3.s, z12.s\n"
+ "fmax z23.s, p3/M, z23.s, z17.s\n"
+ "fmax z26.s, p3/M, z26.s, z17.s\n"
+ "fmla z25.s, p3/M, z8.s, z11.s\n"
+ "fmla z28.s, p3/M, z5.s, z11.s\n"
+ "fmax z25.s, p3/M, z25.s, z17.s\n"
+ "fmax z28.s, p3/M, z28.s, z17.s\n"
+ "fmla z29.s, p3/M, z8.s, z13.s\n"
+ "fmla z30.s, p3/M, z7.s, z13.s\n"
+ "fmax z29.s, p3/M, z29.s, z17.s\n"
+ "fmax z30.s, p3/M, z30.s, z17.s\n"
+ "fmla z31.s, p3/M, z6.s, z13.s\n"
+ "fmax z31.s, p3/M, z31.s, z17.s\n"
+ "fmin z23.s, p3/M, z23.s, z16.s\n"
+ "st1w { z23.s }, p0, [x11]\n"
+ "fmin z25.s, p3/M, z25.s, z16.s\n"
+ "fmin z26.s, p3/M, z26.s, z16.s\n"
+ "st1w { z24.s }, p0, [x11, x16, LSL #2]\n"
+ "fmin z27.s, p3/M, z27.s, z16.s\n"
+ "fmin z28.s, p3/M, z28.s, z16.s\n"
+ "st1w { z25.s }, p0, [x11, x22, LSL #2]\n"
+ "fmin z29.s, p3/M, z29.s, z16.s\n"
+ "fmin z30.s, p3/M, z30.s, z16.s\n"
+ "st1w { z26.s }, p0, [x24]\n"
+ "fmin z31.s, p3/M, z31.s, z16.s\n"
+ "st1w { z27.s }, p0, [x24, x16, LSL #2]\n"
+ "st1w { z28.s }, p0, [x24, x22, LSL #2]\n"
+ "st1w { z29.s }, p0, [x23]\n"
+ "st1w { z30.s }, p0, [x23, x16, 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", "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", "z16", "z17", "z18", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+ : "cc", "memory", "p0", "p1", "p2", "p3", "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", "z16", "z17", "z18", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
);
}