diff options
| author | Michael Tyler <michael.tyler@arm.com> | 2022-12-15 12:39:29 +0000 |
|---|---|---|
| committer | michael.tyler <michael.tyler@arm.com> | 2023-01-16 09:31:00 +0000 |
| commit | ba209750abc1ac7e42bab9fef5db284384d70fb3 (patch) | |
| tree | 1065f242db9a9a5e48bd4a9f2fd68aef1924827a /src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp | |
| parent | 8094f9dd5307c55f545b2cb41ec80a739a9b4d6f (diff) | |
| download | ComputeLibrary-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/a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp | 932 |
1 files changed, 466 insertions, 466 deletions
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp index 9b4b3ee50d..2b1dc3646d 100644 --- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp +++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_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 * @@ -87,83 +87,83 @@ void a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst_direct_impl( ); __asm__ __volatile__( + "mov x24, #0x0\n" "mov x23, #0x0\n" - "mov x22, #0x0\n" "1:" // Tile loop - "str x23, [%x[params_struct], %[offsetof_args_tile_i]]\n" + "str x24, [%x[params_struct], %[offsetof_args_tile_i]]\n" + "mov x27, #0x3\n" "mov x26, #0x3\n" - "mov x25, #0x3\n" - "str x22, [%x[params_struct], %[offsetof_args_tile_j]]\n" - "ldr x24, [%x[params_struct], %[offsetof_args_ld_input_row]]\n" - "ldr x21, [%x[params_struct], %[offsetof_args_ld_output_row]]\n" - "mul x20, x23, x24\n" // offset = tile_i * ld_input_row - "ldr x17, [%x[params_struct], %[offsetof_args_ld_input_col]]\n" - "ldr x16, [%x[params_struct], %[offsetof_args_ld_output_col]]\n" - "mul x19, x23, x21\n" // offset = tile_i * ld_output_row - "mov x23, #0x10\n" // cntb _, ALL, #1 - "madd x20, x22, x17, x20\n" // offset += tile_j * ld_input_col - "ldr x15, [%x[params_struct], %[offsetof_args_inptr]]\n" + "str x23, [%x[params_struct], %[offsetof_args_tile_j]]\n" + "ldr x25, [%x[params_struct], %[offsetof_args_ld_input_row]]\n" + "ldr x22, [%x[params_struct], %[offsetof_args_ld_output_row]]\n" + "mul x21, x24, x25\n" // offset = tile_i * ld_input_row + "ldr x8, [%x[params_struct], %[offsetof_args_ld_input_col]]\n" + "ldr x17, [%x[params_struct], %[offsetof_args_ld_output_col]]\n" + "mul x20, x24, x22\n" // offset = tile_i * ld_output_row + "mov x24, #0x10\n" // cntb _, ALL, #1 + "madd x21, x23, x8, x21\n" // offset += tile_j * ld_input_col + "ldr x16, [%x[params_struct], %[offsetof_args_inptr]]\n" + "lsl x8, x8, #0x1\n" + "ldr x15, [%x[params_struct], %[offsetof_args_outptr]]\n" + "madd x20, x23, x17, x20\n" // offset += tile_j * ld_output_col "lsl x17, x17, #0x1\n" - "ldr x14, [%x[params_struct], %[offsetof_args_outptr]]\n" - "madd x19, x22, x16, x19\n" // offset += tile_j * ld_output_col - "lsl x16, x16, #0x1\n" - "lsr x22, %x[n_channels], #0x3\n" - "ldr x13, [%x[params_struct], %[offsetof_args_params]]\n" - "mul x20, x20, x26\n" // offset *= kernel_stride * output_size - "add x15, x15, x20, LSL #1\n" // inptr[0] += offset * sizeof(__fp16) - "add x12, x15, x24, LSL #1\n" - "mul x19, x19, x25\n" // offset *= output_tile_size - "add x11, x12, x24, LSL #1\n" - "add x10, x17, x17\n" - "add x14, x14, x19, LSL #1\n" // outptrs[0] += offset * sizeof(__fp16) - "add x9, x11, x24, LSL #1\n" - "add x28, x10, x17\n" - "add x27, x14, x21, LSL #1\n" + "lsr x23, %x[n_channels], #0x3\n" + "ldr x14, [%x[params_struct], %[offsetof_args_params]]\n" + "mul x21, x21, x27\n" // offset *= kernel_stride * output_size + "add x16, x16, x21, LSL #1\n" // inptr[0] += offset * sizeof(__fp16) + "add x13, x16, x25, LSL #1\n" + "mul x20, x20, x26\n" // offset *= output_tile_size + "add x12, x13, x25, LSL #1\n" + "add x11, x8, x8\n" + "add x15, x15, x20, LSL #1\n" // outptrs[0] += offset * sizeof(__fp16) + "add x10, x12, x25, LSL #1\n" + "add x9, x11, x8\n" + "add x28, x15, x22, LSL #1\n" "add x20, %x[params_struct], %[offsetof_args_min]\n" - "add x19, %x[params_struct], %[offsetof_args_max]\n" "ld1r { v18.8h }, [x20]\n" - "ld1r { v17.8h }, [x19]\n" - "add x26, x9, x24, LSL #1\n" - "add x25, x28, x17\n" - "add x24, x27, x21, LSL #1\n" - "add x21, x16, x16\n" - "mov x20, #0x0\n" - "sub x19, XZR, x23\n" - "cbz x22, 4f\n" - "ldr q16, [x13, #0x0]\n" - "cmp x23, x22, LSL #4\n" - "ldr q0, [x13, #0x10]\n" - "ldr q1, [x13, #0x20]\n" - "ldr q2, [x13, #0x30]\n" - "ldr q3, [x13, #0x40]\n" - "ldr q4, [x13, #0x50]\n" - "ldr q5, [x13, #0x60]\n" - "ldr q6, [x13, #0x70]\n" - "ldr q7, [x13, #0x80]\n" - "ldr q8, [x13, #0x90]\n" - "ldr q9, [x11, x10]\n" - "add x13, x13, #0xa0\n" - "ld1 { v10.8h }, [x15]\n" - "ldr q11, [x15, x25]\n" - "ld1 { v12.8h }, [x26]\n" - "ldr q13, [x12, x10]\n" + "add x20, %x[params_struct], %[offsetof_args_max]\n" + "ld1r { v17.8h }, [x20]\n" + "add x27, x10, x25, LSL #1\n" + "add x26, x9, x8\n" + "add x25, x28, x22, LSL #1\n" + "add x22, x17, x17\n" + "mov x21, #0x0\n" + "sub x20, XZR, x24\n" + "cbz x23, 4f\n" + "ldr q16, [x14, #0x0]\n" + "ldr q0, [x14, #0x10]\n" + "cmp x24, x23, LSL #4\n" + "ldr q1, [x14, #0x20]\n" + "ldr q2, [x14, #0x30]\n" + "ldr q3, [x14, #0x40]\n" + "ldr q4, [x14, #0x50]\n" + "ldr q5, [x14, #0x60]\n" + "ldr q6, [x14, #0x70]\n" + "ldr q7, [x14, #0x80]\n" + "ldr q8, [x14, #0x90]\n" + "add x14, x14, #0xa0\n" + "ldr q9, [x12, x11]\n" + "ld1 { v10.8h }, [x16]\n" + "ldr q11, [x16, x26]\n" + "ld1 { v12.8h }, [x27]\n" + "ldr q13, [x13, x11]\n" "bge 3f\n" "2:" // Tile loop: Channel loop "mov v24.16b, v16.16b\n fmla v24.8h, v7.8h, v9.8h\n" "mov v23.16b, v16.16b\n fmla v23.8h, v8.8h, v9.8h\n" - "add x23, x23, #0x10\n" - "cmp x23, x22, LSL #4\n" + "add x24, x24, #0x10\n" + "cmp x24, x23, LSL #4\n" "mov v25.16b, v16.16b\n fmla v25.8h, v6.8h, v9.8h\n" "fmla v24.8h, v4.8h, v13.8h\n" - "add x19, x19, #0x10\n" "add x20, x20, #0x10\n" + "add x21, x21, #0x10\n" "mov v26.16b, v16.16b\n fmla v26.8h, v5.8h, v9.8h\n" "mov v27.16b, v16.16b\n fmla v27.8h, v4.8h, v9.8h\n" "mov v28.16b, v16.16b\n fmla v28.8h, v3.8h, v9.8h\n" "fmla v23.8h, v0.8h, v10.8h\n" - "ldr q10, [x11, x28]\n" + "ldr q10, [x12, x9]\n" "fmla v25.8h, v2.8h, v11.8h\n" - "ldr q11, [x11, x17]\n" + "ldr q11, [x12, x8]\n" "mov v29.16b, v16.16b\n fmla v29.8h, v2.8h, v9.8h\n" "fmla v24.8h, v6.8h, v11.8h\n" "mov v31.16b, v16.16b\n fmla v31.8h, v0.8h, v9.8h\n" @@ -172,139 +172,139 @@ void a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst_direct_impl( "fmla v26.8h, v2.8h, v13.8h\n" "fmla v27.8h, v1.8h, v13.8h\n" "fmla v28.8h, v0.8h, v13.8h\n" - "ldr q13, [x15, x17]\n" + "ldr q13, [x16, x8]\n" "fmla v29.8h, v6.8h, v12.8h\n" - "ldr q12, [x26, x25]\n" + "ldr q12, [x27, x26]\n" "mov v30.16b, v16.16b\n fmla v30.8h, v1.8h, v9.8h\n" + "ldr q16, [x14, #0x0]\n" "fmla v24.8h, v0.8h, v13.8h\n" - "ldr q16, [x13, #0x0]\n" "fmla v31.8h, v8.8h, v12.8h\n" - "ldr q12, [x15, x28]\n" + "ldr q12, [x16, x9]\n" "fmla v23.8h, v7.8h, v11.8h\n" "fmla v30.8h, v0.8h, v11.8h\n" "fmla v26.8h, v4.8h, v11.8h\n" "fmla v27.8h, v3.8h, v11.8h\n" "fmla v29.8h, v1.8h, v11.8h\n" - "ld1 { v11.8h }, [x12]\n" + "ld1 { v11.8h }, [x13]\n" "fmla v24.8h, v2.8h, v12.8h\n" "fmla v25.8h, v1.8h, v12.8h\n" - "ld1 { v12.8h }, [x9]\n" + "ld1 { v12.8h }, [x10]\n" "fmla v28.8h, v4.8h, v10.8h\n" "fmla v23.8h, v1.8h, v13.8h\n" - "ldr q13, [x12, x25]\n" + "ldr q13, [x13, x26]\n" "fmla v30.8h, v2.8h, v10.8h\n" "fmla v31.8h, v1.8h, v10.8h\n" "fmla v24.8h, v8.8h, v10.8h\n" "fmla v25.8h, v7.8h, v10.8h\n" "fmla v27.8h, v5.8h, v10.8h\n" + "ldr q10, [x10, x11]\n" "fmla v26.8h, v0.8h, v11.8h\n" - "ldr q10, [x9, x10]\n" "fmla v29.8h, v3.8h, v12.8h\n" "fmla v28.8h, v2.8h, v13.8h\n" "fmla v30.8h, v4.8h, v10.8h\n" "fmla v31.8h, v3.8h, v10.8h\n" "fmla v23.8h, v3.8h, v11.8h\n" + "ldr q11, [x10, x26]\n" "fmla v25.8h, v5.8h, v13.8h\n" - "ldr q11, [x9, x25]\n" - "ldr q13, [x26, x17]\n" + "ldr q13, [x27, x8]\n" "fmla v26.8h, v6.8h, v12.8h\n" + "ldr q12, [x13, x8]\n" "fmla v27.8h, v7.8h, v10.8h\n" - "ldr q12, [x12, x17]\n" "fmla v29.8h, v5.8h, v10.8h\n" "fmla v28.8h, v6.8h, v10.8h\n" "fmla v31.8h, v5.8h, v11.8h\n" "fmla v30.8h, v6.8h, v13.8h\n" "fmla v26.8h, v8.8h, v10.8h\n" "fmla v29.8h, v7.8h, v13.8h\n" - "ldr q13, [x26, x28]\n" + "ldr q13, [x27, x9]\n" "fmla v24.8h, v3.8h, v12.8h\n" "fmla v27.8h, v0.8h, v12.8h\n" "fmla v28.8h, v8.8h, v11.8h\n" - "ldr q11, [x12, x28]\n" + "ldr q11, [x13, x9]\n" "fmla v30.8h, v8.8h, v13.8h\n" - "add x12, x12, #0x10\n" + "add x13, x13, #0x10\n" "fmla v31.8h, v7.8h, v13.8h\n" + "ldr q13, [x10, x9]\n" "fmla v23.8h, v4.8h, v12.8h\n" - "ldr q13, [x9, x28]\n" "fmla v26.8h, v1.8h, v12.8h\n" + "ldr q12, [x10, x8]\n" "fmla v24.8h, v5.8h, v11.8h\n" - "ldr q12, [x9, x17]\n" - "add x9, x9, #0x10\n" + "add x10, x10, #0x10\n" "fmla v25.8h, v4.8h, v11.8h\n" "fmla v27.8h, v2.8h, v11.8h\n" "fmla v28.8h, v1.8h, v11.8h\n" - "ldr q11, [x15, x10]\n" + "ldr q11, [x16, x11]\n" "fmla v29.8h, v4.8h, v12.8h\n" - "add x15, x15, #0x10\n" + "add x16, x16, #0x10\n" + "ld1 { v10.8h }, [x16]\n" "fmla v30.8h, v3.8h, v12.8h\n" "fmla v31.8h, v4.8h, v13.8h\n" - "ld1 { v10.8h }, [x15]\n" - "ldr q4, [x13, #0x50]\n" + "ldr q4, [x14, #0x50]\n" "fmla v26.8h, v7.8h, v12.8h\n" "fmla v27.8h, v6.8h, v12.8h\n" - "ld1 { v12.8h }, [x11]\n" + "ld1 { v12.8h }, [x12]\n" "fmla v23.8h, v2.8h, v11.8h\n" "fmla v24.8h, v1.8h, v11.8h\n" + "ldr q1, [x14, #0x20]\n" "fmax v24.8h, v24.8h, v18.8h\n" - "ldr q1, [x13, #0x20]\n" "fmla v25.8h, v0.8h, v11.8h\n" - "ldr q11, [x11, x25]\n" + "ldr q11, [x12, x26]\n" "fmla v28.8h, v7.8h, v13.8h\n" - "add x11, x11, #0x10\n" + "add x12, x12, #0x10\n" + "ldr q9, [x12, x11]\n" "fmla v30.8h, v5.8h, v13.8h\n" "fmla v29.8h, v0.8h, v12.8h\n" - "fmin v24.8h, v24.8h, v17.8h\n" - "ldr q9, [x11, x10]\n" + "ldr q0, [x14, #0x10]\n" "fmla v31.8h, v2.8h, v11.8h\n" + "ldr q2, [x14, #0x30]\n" "fmla v27.8h, v8.8h, v13.8h\n" - "ldr q13, [x26, x10]\n" - "fmax v27.8h, v27.8h, v18.8h\n" + "ldr q13, [x27, x11]\n" "fmla v23.8h, v6.8h, v12.8h\n" "fmla v26.8h, v3.8h, v12.8h\n" + "ldr q3, [x14, #0x40]\n" "fmax v23.8h, v23.8h, v18.8h\n" - "add x26, x26, #0x10\n" "fmla v25.8h, v8.8h, v11.8h\n" "fmla v28.8h, v5.8h, v11.8h\n" - "fmax v25.8h, v25.8h, v18.8h\n" - "ldr q11, [x15, x25]\n" + "ldr q11, [x16, x26]\n" + "ldr q5, [x14, #0x60]\n" "fmla v29.8h, v8.8h, v13.8h\n" + "ldr q8, [x14, #0x90]\n" "fmla v30.8h, v7.8h, v13.8h\n" - "fmax v26.8h, v26.8h, v18.8h\n" - "ld1 { v12.8h }, [x26]\n" + "ldr q7, [x14, #0x80]\n" "fmla v31.8h, v6.8h, v13.8h\n" + "ldr q13, [x13, x11]\n" + "ldr q6, [x14, #0x70]\n" + "fmax v25.8h, v25.8h, v18.8h\n" + "fmax v26.8h, v26.8h, v18.8h\n" + "fmax v27.8h, v27.8h, v18.8h\n" + "add x27, x27, #0x10\n" + "ld1 { v12.8h }, [x27]\n" "fmax v28.8h, v28.8h, v18.8h\n" - "ldr q13, [x12, x10]\n" - "ldr q0, [x13, #0x10]\n" "fmax v29.8h, v29.8h, v18.8h\n" + "add x14, x14, #0xa0\n" "fmax v30.8h, v30.8h, v18.8h\n" - "ldr q2, [x13, #0x30]\n" - "ldr q3, [x13, #0x40]\n" "fmax v31.8h, v31.8h, v18.8h\n" "fmin v23.8h, v23.8h, v17.8h\n" - "st1 { v23.8h }, [x14]\n" - "ldr q5, [x13, #0x60]\n" + "fmin v24.8h, v24.8h, v17.8h\n" + "st1 { v23.8h }, [x15]\n" "fmin v25.8h, v25.8h, v17.8h\n" "fmin v26.8h, v26.8h, v17.8h\n" - "str q24, [x14, x16]\n" - "ldr q6, [x13, #0x70]\n" + "str q24, [x15, x17]\n" "fmin v27.8h, v27.8h, v17.8h\n" "fmin v28.8h, v28.8h, v17.8h\n" - "str q25, [x14, x21]\n" - "add x14, x14, #0x10\n" + "str q25, [x15, x22]\n" + "add x15, x15, #0x10\n" "fmin v29.8h, v29.8h, v17.8h\n" "fmin v30.8h, v30.8h, v17.8h\n" - "st1 { v26.8h }, [x27]\n" - "ldr q7, [x13, #0x80]\n" + "st1 { v26.8h }, [x28]\n" "fmin v31.8h, v31.8h, v17.8h\n" - "str q27, [x27, x16]\n" - "ldr q8, [x13, #0x90]\n" - "add x13, x13, #0xa0\n" - "str q28, [x27, x21]\n" - "add x27, x27, #0x10\n" - "st1 { v29.8h }, [x24]\n" - "str q30, [x24, x16]\n" - "str q31, [x24, x21]\n" - "add x24, x24, #0x10\n" + "str q27, [x28, x17]\n" + "str q28, [x28, x22]\n" + "add x28, x28, #0x10\n" + "st1 { v29.8h }, [x25]\n" + "str q30, [x25, x17]\n" + "str q31, [x25, x22]\n" + "add x25, x25, #0x10\n" "blt 2b\n" "3:" // Tile loop: Channel tail "mov v24.16b, v16.16b\n fmla v24.8h, v7.8h, v9.8h\n" @@ -315,9 +315,9 @@ void a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst_direct_impl( "mov v27.16b, v16.16b\n fmla v27.8h, v4.8h, v9.8h\n" "mov v28.16b, v16.16b\n fmla v28.8h, v3.8h, v9.8h\n" "fmla v23.8h, v0.8h, v10.8h\n" - "ldr q10, [x11, x28]\n" + "ldr q10, [x12, x9]\n" "fmla v25.8h, v2.8h, v11.8h\n" - "ldr q11, [x11, x17]\n" + "ldr q11, [x12, x8]\n" "mov v29.16b, v16.16b\n fmla v29.8h, v2.8h, v9.8h\n" "fmla v24.8h, v6.8h, v11.8h\n" "mov v31.16b, v16.16b\n fmla v31.8h, v0.8h, v9.8h\n" @@ -326,92 +326,92 @@ void a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst_direct_impl( "fmla v26.8h, v2.8h, v13.8h\n" "fmla v27.8h, v1.8h, v13.8h\n" "fmla v28.8h, v0.8h, v13.8h\n" - "ldr q13, [x15, x17]\n" + "ldr q13, [x16, x8]\n" "fmla v29.8h, v6.8h, v12.8h\n" - "ldr q12, [x26, x25]\n" + "ldr q12, [x27, x26]\n" "mov v30.16b, v16.16b\n fmla v30.8h, v1.8h, v9.8h\n" "fmla v24.8h, v0.8h, v13.8h\n" "fmla v31.8h, v8.8h, v12.8h\n" - "ldr q12, [x15, x28]\n" + "ldr q12, [x16, x9]\n" "fmla v23.8h, v7.8h, v11.8h\n" "fmla v30.8h, v0.8h, v11.8h\n" "fmla v26.8h, v4.8h, v11.8h\n" "fmla v27.8h, v3.8h, v11.8h\n" "fmla v29.8h, v1.8h, v11.8h\n" - "ld1 { v11.8h }, [x12]\n" + "ld1 { v11.8h }, [x13]\n" "fmla v24.8h, v2.8h, v12.8h\n" "fmla v25.8h, v1.8h, v12.8h\n" - "ld1 { v12.8h }, [x9]\n" + "ld1 { v12.8h }, [x10]\n" "fmla v28.8h, v4.8h, v10.8h\n" "fmla v23.8h, v1.8h, v13.8h\n" - "ldr q13, [x12, x25]\n" + "ldr q13, [x13, x26]\n" "fmla v30.8h, v2.8h, v10.8h\n" "fmla v31.8h, v1.8h, v10.8h\n" "fmla v24.8h, v8.8h, v10.8h\n" "fmla v25.8h, v7.8h, v10.8h\n" "fmla v27.8h, v5.8h, v10.8h\n" + "ldr q10, [x10, x11]\n" "fmla v26.8h, v0.8h, v11.8h\n" - "ldr q10, [x9, x10]\n" "fmla v29.8h, v3.8h, v12.8h\n" "fmla v28.8h, v2.8h, v13.8h\n" "fmla v30.8h, v4.8h, v10.8h\n" "fmla v31.8h, v3.8h, v10.8h\n" "fmla v23.8h, v3.8h, v11.8h\n" + "ldr q11, [x10, x26]\n" "fmla v25.8h, v5.8h, v13.8h\n" - "ldr q11, [x9, x25]\n" - "ldr q13, [x26, x17]\n" + "ldr q13, [x27, x8]\n" "fmla v26.8h, v6.8h, v12.8h\n" + "ldr q12, [x13, x8]\n" "fmla v27.8h, v7.8h, v10.8h\n" - "ldr q12, [x12, x17]\n" "fmla v29.8h, v5.8h, v10.8h\n" "fmla v28.8h, v6.8h, v10.8h\n" "fmla v31.8h, v5.8h, v11.8h\n" "fmla v30.8h, v6.8h, v13.8h\n" "fmla v26.8h, v8.8h, v10.8h\n" "fmla v29.8h, v7.8h, v13.8h\n" - "ldr q13, [x26, x28]\n" + "ldr q13, [x27, x9]\n" "fmla v24.8h, v3.8h, v12.8h\n" "fmla v27.8h, v0.8h, v12.8h\n" "fmla v28.8h, v8.8h, v11.8h\n" - "ldr q11, [x12, x28]\n" + "ldr q11, [x13, x9]\n" "fmla v30.8h, v8.8h, v13.8h\n" - "add x12, x12, #0x10\n" + "add x13, x13, #0x10\n" "fmla v31.8h, v7.8h, v13.8h\n" + "ldr q13, [x10, x9]\n" "fmla v23.8h, v4.8h, v12.8h\n" - "ldr q13, [x9, x28]\n" "fmla v26.8h, v1.8h, v12.8h\n" + "ldr q12, [x10, x8]\n" "fmla v24.8h, v5.8h, v11.8h\n" - "ldr q12, [x9, x17]\n" - "add x9, x9, #0x10\n" + "add x10, x10, #0x10\n" "fmla v25.8h, v4.8h, v11.8h\n" "fmla v27.8h, v2.8h, v11.8h\n" "fmla v28.8h, v1.8h, v11.8h\n" - "ldr q11, [x15, x10]\n" + "ldr q11, [x16, x11]\n" "fmla v29.8h, v4.8h, v12.8h\n" - "add x15, x15, #0x10\n" + "add x16, x16, #0x10\n" "fmla v30.8h, v3.8h, v12.8h\n" "fmla v31.8h, v4.8h, v13.8h\n" "fmla v26.8h, v7.8h, v12.8h\n" "fmla v27.8h, v6.8h, v12.8h\n" - "ld1 { v12.8h }, [x11]\n" + "ld1 { v12.8h }, [x12]\n" "fmla v23.8h, v2.8h, v11.8h\n" "fmla v24.8h, v1.8h, v11.8h\n" "fmax v24.8h, v24.8h, v18.8h\n" "fmla v25.8h, v0.8h, v11.8h\n" - "ldr q11, [x11, x25]\n" + "ldr q11, [x12, x26]\n" "fmla v28.8h, v7.8h, v13.8h\n" "fmin v24.8h, v24.8h, v17.8h\n" "fmla v30.8h, v5.8h, v13.8h\n" "fmla v29.8h, v0.8h, v12.8h\n" - "add x11, x11, #0x10\n" + "add x12, x12, #0x10\n" "fmla v31.8h, v2.8h, v11.8h\n" "fmla v27.8h, v8.8h, v13.8h\n" - "ldr q13, [x26, x10]\n" + "ldr q13, [x27, x11]\n" "fmax v27.8h, v27.8h, v18.8h\n" "fmla v23.8h, v6.8h, v12.8h\n" "fmla v26.8h, v3.8h, v12.8h\n" "fmax v23.8h, v23.8h, v18.8h\n" - "add x26, x26, #0x10\n" + "add x27, x27, #0x10\n" "fmla v25.8h, v8.8h, v11.8h\n" "fmla v28.8h, v5.8h, v11.8h\n" "fmax v25.8h, v25.8h, v18.8h\n" @@ -424,94 +424,94 @@ void a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst_direct_impl( "fmax v30.8h, v30.8h, v18.8h\n" "fmax v31.8h, v31.8h, v18.8h\n" "fmin v23.8h, v23.8h, v17.8h\n" - "st1 { v23.8h }, [x14]\n" + "st1 { v23.8h }, [x15]\n" "fmin v25.8h, v25.8h, v17.8h\n" "fmin v26.8h, v26.8h, v17.8h\n" - "str q24, [x14, x16]\n" + "str q24, [x15, x17]\n" "fmin v27.8h, v27.8h, v17.8h\n" "fmin v28.8h, v28.8h, v17.8h\n" - "str q25, [x14, x21]\n" - "add x14, x14, #0x10\n" + "str q25, [x15, x22]\n" + "add x15, x15, #0x10\n" "fmin v29.8h, v29.8h, v17.8h\n" "fmin v30.8h, v30.8h, v17.8h\n" - "st1 { v26.8h }, [x27]\n" + "st1 { v26.8h }, [x28]\n" "fmin v31.8h, v31.8h, v17.8h\n" - "str q27, [x27, x16]\n" - "str q28, [x27, x21]\n" - "add x27, x27, #0x10\n" - "st1 { v29.8h }, [x24]\n" - "str q30, [x24, x16]\n" - "str q31, [x24, x21]\n" - "add x24, x24, #0x10\n" + "str q27, [x28, x17]\n" + "str q28, [x28, x22]\n" + "add x28, x28, #0x10\n" + "st1 { v29.8h }, [x25]\n" + "str q30, [x25, x17]\n" + "str q31, [x25, x22]\n" + "add x25, x25, #0x10\n" "4:" // Tile loop: Oddments "tst %x[n_channels], #0x7\n" "beq 93f\n" - "ldr q16, [x13, #0x0]\n" - "ldr q0, [x13, #0x10]\n" - "ldr q1, [x13, #0x20]\n" - "ldr q2, [x13, #0x30]\n" - "add x23, x11, x10\n" - "add x22, x15, XZR\n" - "ldr q3, [x13, #0x40]\n" - "ldr q4, [x13, #0x50]\n" - "add x21, x15, x25\n" - "add x20, x26, XZR\n" - "ldr q5, [x13, #0x60]\n" - "ldr q6, [x13, #0x70]\n" - "add x19, x12, x10\n" - "ldr q7, [x13, #0x80]\n" - "ldr q8, [x13, #0x90]\n" + "ldr q16, [x14, #0x0]\n" + "ldr q0, [x14, #0x10]\n" + "add x24, x12, x11\n" + "add x23, x16, XZR\n" + "ldr q1, [x14, #0x20]\n" + "ldr q2, [x14, #0x30]\n" + "add x22, x16, x26\n" + "add x21, x27, XZR\n" + "ldr q3, [x14, #0x40]\n" + "ldr q4, [x14, #0x50]\n" + "add x20, x13, x11\n" + "ldr q5, [x14, #0x60]\n" + "ldr q6, [x14, #0x70]\n" + "ldr q7, [x14, #0x80]\n" + "ldr q8, [x14, #0x90]\n" "tbz %x[n_channels], #2, 6f\n" - "ldr d9, [x23], #0x8\n" - "ldr d10, [x22], #0x8\n" - "ldr d11, [x21], #0x8\n" - "ldr d12, [x20], #0x8\n" - "ldr d13, [x19], #0x8\n" + "ldr d9, [x24], #0x8\n" + "ldr d10, [x23], #0x8\n" + "ldr d11, [x22], #0x8\n" + "ldr d12, [x21], #0x8\n" + "ldr d13, [x20], #0x8\n" "tbz %x[n_channels], #1, 5f\n" - "ld1 { v9.s }[2], [x23], #0x4\n" - "ld1 { v10.s }[2], [x22], #0x4\n" - "ld1 { v11.s }[2], [x21], #0x4\n" - "ld1 { v12.s }[2], [x20], #0x4\n" - "ld1 { v13.s }[2], [x19], #0x4\n" + "ld1 { v9.s }[2], [x24], #0x4\n" + "ld1 { v10.s }[2], [x23], #0x4\n" + "ld1 { v11.s }[2], [x22], #0x4\n" + "ld1 { v12.s }[2], [x21], #0x4\n" + "ld1 { v13.s }[2], [x20], #0x4\n" "tbz %x[n_channels], #0, 8f\n" - "ld1 { v9.h }[6], [x23]\n" - "ld1 { v10.h }[6], [x22]\n" - "ld1 { v11.h }[6], [x21]\n" - "ld1 { v12.h }[6], [x20]\n" - "ld1 { v13.h }[6], [x19]\n" + "ld1 { v9.h }[6], [x24]\n" + "ld1 { v10.h }[6], [x23]\n" + "ld1 { v11.h }[6], [x22]\n" + "ld1 { v12.h }[6], [x21]\n" + "ld1 { v13.h }[6], [x20]\n" "b 8f\n" "5:" // Tile loop: Oddments: Load inputs: (2, 2), (0, 0), (0, 4), (4, 0), (1, 2): Bit 2: Bit 1: Unset "tbz %x[n_channels], #0, 8f\n" - "ld1 { v9.h }[4], [x23]\n" - "ld1 { v10.h }[4], [x22]\n" - "ld1 { v11.h }[4], [x21]\n" - "ld1 { v12.h }[4], [x20]\n" - "ld1 { v13.h }[4], [x19]\n" + "ld1 { v9.h }[4], [x24]\n" + "ld1 { v10.h }[4], [x23]\n" + "ld1 { v11.h }[4], [x22]\n" + "ld1 { v12.h }[4], [x21]\n" + "ld1 { v13.h }[4], [x20]\n" "b 8f\n" "6:" // Tile loop: Oddments: Load inputs: (2, 2), (0, 0), (0, 4), (4, 0), (1, 2): Bit 2: Unset "tbz %x[n_channels], #1, 7f\n" - "ldr s9, [x23], #0x4\n" - "ldr s10, [x22], #0x4\n" - "ldr s11, [x21], #0x4\n" - "ldr s12, [x20], #0x4\n" - "ldr s13, [x19], #0x4\n" + "ldr s9, [x24], #0x4\n" + "ldr s10, [x23], #0x4\n" + "ldr s11, [x22], #0x4\n" + "ldr s12, [x21], #0x4\n" + "ldr s13, [x20], #0x4\n" "tbz %x[n_channels], #0, 8f\n" - "ld1 { v9.h }[2], [x23]\n" - "ld1 { v10.h }[2], [x22]\n" - "ld1 { v11.h }[2], [x21]\n" - "ld1 { v12.h }[2], [x20]\n" - "ld1 { v13.h }[2], [x19]\n" + "ld1 { v9.h }[2], [x24]\n" + "ld1 { v10.h }[2], [x23]\n" + "ld1 { v11.h }[2], [x22]\n" + "ld1 { v12.h }[2], [x21]\n" + "ld1 { v13.h }[2], [x20]\n" "b 8f\n" "7:" // Tile loop: Oddments: Load inputs: (2, 2), (0, 0), (0, 4), (4, 0), (1, 2): Bit 2: Unset: Bit 1: Unset - "ldr h9, [x23, #0x0]\n" - "ldr h10, [x22, #0x0]\n" - "ldr h11, [x21, #0x0]\n" - "ldr h12, [x20, #0x0]\n" - "ldr h13, [x19, #0x0]\n" + "ldr h9, [x24, #0x0]\n" + "ldr h10, [x23, #0x0]\n" + "ldr h11, [x22, #0x0]\n" + "ldr h12, [x21, #0x0]\n" + "ldr h13, [x20, #0x0]\n" "8:" // Tile loop: Oddments: Load inputs: (2, 2), (0, 0), (0, 4), (4, 0), (1, 2): Bit 2: End "mov v23.16b, v16.16b\n fmla v23.8h, v8.8h, v9.8h\n" "mov v25.16b, v16.16b\n fmla v25.8h, v6.8h, v9.8h\n" - "add x19, x26, x25\n" + "add x20, x27, x26\n" "mov v24.16b, v16.16b\n fmla v24.8h, v7.8h, v9.8h\n" "mov v26.16b, v16.16b\n fmla v26.8h, v5.8h, v9.8h\n" "mov v27.16b, v16.16b\n fmla v27.8h, v4.8h, v9.8h\n" @@ -529,483 +529,483 @@ void a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst_direct_impl( "fmla v27.8h, v1.8h, v13.8h\n" "fmla v28.8h, v0.8h, v13.8h\n" "tbz %x[n_channels], #2, 10f\n" - "ldr d12, [x19], #0x8\n" + "ldr d12, [x20], #0x8\n" "tbz %x[n_channels], #1, 9f\n" - "ld1 { v12.s }[2], [x19], #0x4\n" + "ld1 { v12.s }[2], [x20], #0x4\n" "tbz %x[n_channels], #0, 12f\n" - "ld1 { v12.h }[6], [x19]\n" + "ld1 { v12.h }[6], [x20]\n" "b 12f\n" "9:" // Tile loop: Oddments: Load inputs: (4, 4): Bit 2: Bit 1: Unset "tbz %x[n_channels], #0, 12f\n" - "ld1 { v12.h }[4], [x19]\n" + "ld1 { v12.h }[4], [x20]\n" "b 12f\n" "10:" // Tile loop: Oddments: Load inputs: (4, 4): Bit 2: Unset "tbz %x[n_channels], #1, 11f\n" - "ldr s12, [x19], #0x4\n" + "ldr s12, [x20], #0x4\n" "tbz %x[n_channels], #0, 12f\n" - "ld1 { v12.h }[2], [x19]\n" + "ld1 { v12.h }[2], [x20]\n" "b 12f\n" "11:" // Tile loop: Oddments: Load inputs: (4, 4): Bit 2: Unset: Bit 1: Unset - "ldr h12, [x19, #0x0]\n" + "ldr h12, [x20, #0x0]\n" "12:" // Tile loop: Oddments: Load inputs: (4, 4): Bit 2: End "fmla v31.8h, v8.8h, v12.8h\n" - "add x19, x11, x17\n" + "add x20, x12, x8\n" "tbz %x[n_channels], #2, 14f\n" - "ldr d11, [x19], #0x8\n" + "ldr d11, [x20], #0x8\n" "tbz %x[n_channels], #1, 13f\n" - "ld1 { v11.s }[2], [x19], #0x4\n" + "ld1 { v11.s }[2], [x20], #0x4\n" "tbz %x[n_channels], #0, 16f\n" - "ld1 { v11.h }[6], [x19]\n" + "ld1 { v11.h }[6], [x20]\n" "b 16f\n" "13:" // Tile loop: Oddments: Load inputs: (2, 1): Bit 2: Bit 1: Unset "tbz %x[n_channels], #0, 16f\n" - "ld1 { v11.h }[4], [x19]\n" + "ld1 { v11.h }[4], [x20]\n" "b 16f\n" "14:" // Tile loop: Oddments: Load inputs: (2, 1): Bit 2: Unset "tbz %x[n_channels], #1, 15f\n" - "ldr s11, [x19], #0x4\n" + "ldr s11, [x20], #0x4\n" "tbz %x[n_channels], #0, 16f\n" - "ld1 { v11.h }[2], [x19]\n" + "ld1 { v11.h }[2], [x20]\n" "b 16f\n" "15:" // Tile loop: Oddments: Load inputs: (2, 1): Bit 2: Unset: Bit 1: Unset - "ldr h11, [x19, #0x0]\n" + "ldr h11, [x20, #0x0]\n" "16:" // Tile loop: Oddments: Load inputs: (2, 1): Bit 2: End "fmla v23.8h, v7.8h, v11.8h\n" "fmla v24.8h, v6.8h, v11.8h\n" - "add x19, x15, x17\n" + "add x20, x16, x8\n" "fmla v26.8h, v4.8h, v11.8h\n" "fmla v27.8h, v3.8h, v11.8h\n" "fmla v29.8h, v1.8h, v11.8h\n" "fmla v30.8h, v0.8h, v11.8h\n" "tbz %x[n_channels], #2, 18f\n" - "ldr d13, [x19], #0x8\n" + "ldr d13, [x20], #0x8\n" "tbz %x[n_channels], #1, 17f\n" - "ld1 { v13.s }[2], [x19], #0x4\n" + "ld1 { v13.s }[2], [x20], #0x4\n" "tbz %x[n_channels], #0, 20f\n" - "ld1 { v13.h }[6], [x19]\n" + "ld1 { v13.h }[6], [x20]\n" "b 20f\n" "17:" // Tile loop: Oddments: Load inputs: (0, 1): Bit 2: Bit 1: Unset "tbz %x[n_channels], #0, 20f\n" - "ld1 { v13.h }[4], [x19]\n" + "ld1 { v13.h }[4], [x20]\n" "b 20f\n" "18:" // Tile loop: Oddments: Load inputs: (0, 1): Bit 2: Unset "tbz %x[n_channels], #1, 19f\n" - "ldr s13, [x19], #0x4\n" + "ldr s13, [x20], #0x4\n" "tbz %x[n_channels], #0, 20f\n" - "ld1 { v13.h }[2], [x19]\n" + "ld1 { v13.h }[2], [x20]\n" "b 20f\n" "19:" // Tile loop: Oddments: Load inputs: (0, 1): Bit 2: Unset: Bit 1: Unset - "ldr h13, [x19, #0x0]\n" + "ldr h13, [x20, #0x0]\n" "20:" // Tile loop: Oddments: Load inputs: (0, 1): Bit 2: End "fmla v23.8h, v1.8h, v13.8h\n" "fmla v24.8h, v0.8h, v13.8h\n" - "add x19, x15, x28\n" + "add x20, x16, x9\n" "tbz %x[n_channels], #2, 22f\n" - "ldr d12, [x19], #0x8\n" + "ldr d12, [x20], #0x8\n" "tbz %x[n_channels], #1, 21f\n" - "ld1 { v12.s }[2], [x19], #0x4\n" + "ld1 { v12.s }[2], [x20], #0x4\n" "tbz %x[n_channels], #0, 24f\n" - "ld1 { v12.h }[6], [x19]\n" + "ld1 { v12.h }[6], [x20]\n" "b 24f\n" "21:" // Tile loop: Oddments: Load inputs: (0, 3): Bit 2: Bit 1: Unset "tbz %x[n_channels], #0, 24f\n" - "ld1 { v12.h }[4], [x19]\n" + "ld1 { v12.h }[4], [x20]\n" "b 24f\n" "22:" // Tile loop: Oddments: Load inputs: (0, 3): Bit 2: Unset "tbz %x[n_channels], #1, 23f\n" - "ldr s12, [x19], #0x4\n" + "ldr s12, [x20], #0x4\n" "tbz %x[n_channels], #0, 24f\n" - "ld1 { v12.h }[2], [x19]\n" + "ld1 { v12.h }[2], [x20]\n" "b 24f\n" "23:" // Tile loop: Oddments: Load inputs: (0, 3): Bit 2: Unset: Bit 1: Unset - "ldr h12, [x19, #0x0]\n" + "ldr h12, [x20, #0x0]\n" "24:" // Tile loop: Oddments: Load inputs: (0, 3): Bit 2: End "fmla v24.8h, v2.8h, v12.8h\n" "fmla v25.8h, v1.8h, v12.8h\n" - "add x19, x11, x28\n" + "add x20, x12, x9\n" "tbz %x[n_channels], #2, 26f\n" - "ldr d10, [x19], #0x8\n" + "ldr d10, [x20], #0x8\n" "tbz %x[n_channels], #1, 25f\n" - "ld1 { v10.s }[2], [x19], #0x4\n" + "ld1 { v10.s }[2], [x20], #0x4\n" "tbz %x[n_channels], #0, 28f\n" - "ld1 { v10.h }[6], [x19]\n" + "ld1 { v10.h }[6], [x20]\n" "b 28f\n" "25:" // Tile loop: Oddments: Load inputs: (2, 3): Bit 2: Bit 1: Unset "tbz %x[n_channels], #0, 28f\n" - "ld1 { v10.h }[4], [x19]\n" + "ld1 { v10.h }[4], [x20]\n" "b 28f\n" "26:" // Tile loop: Oddments: Load inputs: (2, 3): Bit 2: Unset "tbz %x[n_channels], #1, 27f\n" - "ldr s10, [x19], #0x4\n" + "ldr s10, [x20], #0x4\n" "tbz %x[n_channels], #0, 28f\n" - "ld1 { v10.h }[2], [x19]\n" + "ld1 { v10.h }[2], [x20]\n" "b 28f\n" "27:" // Tile loop: Oddments: Load inputs: (2, 3): Bit 2: Unset: Bit 1: Unset - "ldr h10, [x19, #0x0]\n" + "ldr h10, [x20, #0x0]\n" "28:" // Tile loop: Oddments: Load inputs: (2, 3): Bit 2: End "fmla v24.8h, v8.8h, v10.8h\n" "fmla v25.8h, v7.8h, v10.8h\n" - "add x19, x12, XZR\n" + "add x20, x13, XZR\n" "fmla v27.8h, v5.8h, v10.8h\n" "fmla v28.8h, v4.8h, v10.8h\n" "fmla v30.8h, v2.8h, v10.8h\n" "fmla v31.8h, v1.8h, v10.8h\n" "tbz %x[n_channels], #2, 30f\n" - "ldr d11, [x19], #0x8\n" + "ldr d11, [x20], #0x8\n" "tbz %x[n_channels], #1, 29f\n" - "ld1 { v11.s }[2], [x19], #0x4\n" + "ld1 { v11.s }[2], [x20], #0x4\n" "tbz %x[n_channels], #0, 32f\n" - "ld1 { v11.h }[6], [x19]\n" + "ld1 { v11.h }[6], [x20]\n" "b 32f\n" "29:" // Tile loop: Oddments: Load inputs: (1, 0): Bit 2: Bit 1: Unset "tbz %x[n_channels], #0, 32f\n" - "ld1 { v11.h }[4], [x19]\n" + "ld1 { v11.h }[4], [x20]\n" "b 32f\n" "30:" // Tile loop: Oddments: Load inputs: (1, 0): Bit 2: Unset "tbz %x[n_channels], #1, 31f\n" - "ldr s11, [x19], #0x4\n" + "ldr s11, [x20], #0x4\n" "tbz %x[n_channels], #0, 32f\n" - "ld1 { v11.h }[2], [x19]\n" + "ld1 { v11.h }[2], [x20]\n" "b 32f\n" "31:" // Tile loop: Oddments: Load inputs: (1, 0): Bit 2: Unset: Bit 1: Unset - "ldr h11, [x19, #0x0]\n" + "ldr h11, [x20, #0x0]\n" "32:" // Tile loop: Oddments: Load inputs: (1, 0): Bit 2: End "fmla v23.8h, v3.8h, v11.8h\n" "fmla v26.8h, v0.8h, v11.8h\n" - "add x19, x12, x25\n" + "add x20, x13, x26\n" "tbz %x[n_channels], #2, 34f\n" - "ldr d13, [x19], #0x8\n" + "ldr d13, [x20], #0x8\n" "tbz %x[n_channels], #1, 33f\n" - "ld1 { v13.s }[2], [x19], #0x4\n" + "ld1 { v13.s }[2], [x20], #0x4\n" "tbz %x[n_channels], #0, 36f\n" - "ld1 { v13.h }[6], [x19]\n" + "ld1 { v13.h }[6], [x20]\n" "b 36f\n" "33:" // Tile loop: Oddments: Load inputs: (1, 4): Bit 2: Bit 1: Unset "tbz %x[n_channels], #0, 36f\n" - "ld1 { v13.h }[4], [x19]\n" + "ld1 { v13.h }[4], [x20]\n" "b 36f\n" "34:" // Tile loop: Oddments: Load inputs: (1, 4): Bit 2: Unset "tbz %x[n_channels], #1, 35f\n" - "ldr s13, [x19], #0x4\n" + "ldr s13, [x20], #0x4\n" "tbz %x[n_channels], #0, 36f\n" - "ld1 { v13.h }[2], [x19]\n" + "ld1 { v13.h }[2], [x20]\n" "b 36f\n" "35:" // Tile loop: Oddments: Load inputs: (1, 4): Bit 2: Unset: Bit 1: Unset - "ldr h13, [x19, #0x0]\n" + "ldr h13, [x20, #0x0]\n" "36:" // Tile loop: Oddments: Load inputs: (1, 4): Bit 2: End "fmla v25.8h, v5.8h, v13.8h\n" "fmla v28.8h, v2.8h, v13.8h\n" - "add x19, x9, XZR\n" + "add x20, x10, XZR\n" "tbz %x[n_channels], #2, 38f\n" - "ldr d12, [x19], #0x8\n" + "ldr d12, [x20], #0x8\n" "tbz %x[n_channels], #1, 37f\n" - "ld1 { v12.s }[2], [x19], #0x4\n" + "ld1 { v12.s }[2], [x20], #0x4\n" "tbz %x[n_channels], #0, 40f\n" - "ld1 { v12.h }[6], [x19]\n" + "ld1 { v12.h }[6], [x20]\n" "b 40f\n" "37:" // Tile loop: Oddments: Load inputs: (3, 0): Bit 2: Bit 1: Unset "tbz %x[n_channels], #0, 40f\n" - "ld1 { v12.h }[4], [x19]\n" + "ld1 { v12.h }[4], [x20]\n" "b 40f\n" "38:" // Tile loop: Oddments: Load inputs: (3, 0): Bit 2: Unset "tbz %x[n_channels], #1, 39f\n" - "ldr s12, [x19], #0x4\n" + "ldr s12, [x20], #0x4\n" "tbz %x[n_channels], #0, 40f\n" - "ld1 { v12.h }[2], [x19]\n" + "ld1 { v12.h }[2], [x20]\n" "b 40f\n" "39:" // Tile loop: Oddments: Load inputs: (3, 0): Bit 2: Unset: Bit 1: Unset - "ldr h12, [x19, #0x0]\n" + "ldr h12, [x20, #0x0]\n" "40:" // Tile loop: Oddments: Load inputs: (3, 0): Bit 2: End "fmla v26.8h, v6.8h, v12.8h\n" "fmla v29.8h, v3.8h, v12.8h\n" - "add x19, x9, x10\n" + "add x20, x10, x11\n" "tbz %x[n_channels], #2, 42f\n" - "ldr d10, [x19], #0x8\n" + "ldr d10, [x20], #0x8\n" "tbz %x[n_channels], #1, 41f\n" - "ld1 { v10.s }[2], [x19], #0x4\n" + "ld1 { v10.s }[2], [x20], #0x4\n" "tbz %x[n_channels], #0, 44f\n" - "ld1 { v10.h }[6], [x19]\n" + "ld1 { v10.h }[6], [x20]\n" "b 44f\n" "41:" // Tile loop: Oddments: Load inputs: (3, 2): Bit 2: Bit 1: Unset "tbz %x[n_channels], #0, 44f\n" - "ld1 { v10.h }[4], [x19]\n" + "ld1 { v10.h }[4], [x20]\n" "b 44f\n" "42:" // Tile loop: Oddments: Load inputs: (3, 2): Bit 2: Unset "tbz %x[n_channels], #1, 43f\n" - "ldr s10, [x19], #0x4\n" + "ldr s10, [x20], #0x4\n" "tbz %x[n_channels], #0, 44f\n" - "ld1 { v10.h }[2], [x19]\n" + "ld1 { v10.h }[2], [x20]\n" "b 44f\n" "43:" // Tile loop: Oddments: Load inputs: (3, 2): Bit 2: Unset: Bit 1: Unset - "ldr h10, [x19, #0x0]\n" + "ldr h10, [x20, #0x0]\n" "44:" // Tile loop: Oddments: Load inputs: (3, 2): Bit 2: End "fmla v26.8h, v8.8h, v10.8h\n" "fmla v27.8h, v7.8h, v10.8h\n" - "add x19, x9, x25\n" + "add x20, x10, x26\n" "fmla v28.8h, v6.8h, v10.8h\n" "fmla v29.8h, v5.8h, v10.8h\n" "fmla v30.8h, v4.8h, v10.8h\n" "fmla v31.8h, v3.8h, v10.8h\n" "tbz %x[n_channels], #2, 46f\n" - "ldr d11, [x19], #0x8\n" + "ldr d11, [x20], #0x8\n" "tbz %x[n_channels], #1, 45f\n" - "ld1 { v11.s }[2], [x19], #0x4\n" + "ld1 { v11.s }[2], [x20], #0x4\n" "tbz %x[n_channels], #0, 48f\n" - "ld1 { v11.h }[6], [x19]\n" + "ld1 { v11.h }[6], [x20]\n" "b 48f\n" "45:" // Tile loop: Oddments: Load inputs: (3, 4): Bit 2: Bit 1: Unset "tbz %x[n_channels], #0, 48f\n" - "ld1 { v11.h }[4], [x19]\n" + "ld1 { v11.h }[4], [x20]\n" "b 48f\n" "46:" // Tile loop: Oddments: Load inputs: (3, 4): Bit 2: Unset "tbz %x[n_channels], #1, 47f\n" - "ldr s11, [x19], #0x4\n" + "ldr s11, [x20], #0x4\n" "tbz %x[n_channels], #0, 48f\n" - "ld1 { v11.h }[2], [x19]\n" + "ld1 { v11.h }[2], [x20]\n" "b 48f\n" "47:" // Tile loop: Oddments: Load inputs: (3, 4): Bit 2: Unset: Bit 1: Unset - "ldr h11, [x19, #0x0]\n" + "ldr h11, [x20, #0x0]\n" "48:" // Tile loop: Oddments: Load inputs: (3, 4): Bit 2: End "fmla v28.8h, v8.8h, v11.8h\n" "fmla v31.8h, v5.8h, v11.8h\n" - "add x19, x26, x17\n" + "add x20, x27, x8\n" "tbz %x[n_channels], #2, 50f\n" - "ldr d13, [x19], #0x8\n" + "ldr d13, [x20], #0x8\n" "tbz %x[n_channels], #1, 49f\n" - "ld1 { v13.s }[2], [x19], #0x4\n" + "ld1 { v13.s }[2], [x20], #0x4\n" "tbz %x[n_channels], #0, 52f\n" - "ld1 { v13.h }[6], [x19]\n" + "ld1 { v13.h }[6], [x20]\n" "b 52f\n" "49:" // Tile loop: Oddments: Load inputs: (4, 1): Bit 2: Bit 1: Unset "tbz %x[n_channels], #0, 52f\n" - "ld1 { v13.h }[4], [x19]\n" + "ld1 { v13.h }[4], [x20]\n" "b 52f\n" "50:" // Tile loop: Oddments: Load inputs: (4, 1): Bit 2: Unset "tbz %x[n_channels], #1, 51f\n" - "ldr s13, [x19], #0x4\n" + "ldr s13, [x20], #0x4\n" "tbz %x[n_channels], #0, 52f\n" - "ld1 { v13.h }[2], [x19]\n" + "ld1 { v13.h }[2], [x20]\n" "b 52f\n" "51:" // Tile loop: Oddments: Load inputs: (4, 1): Bit 2: Unset: Bit 1: Unset - "ldr h13, [x19, #0x0]\n" + "ldr h13, [x20, #0x0]\n" "52:" // Tile loop: Oddments: Load inputs: (4, 1): Bit 2: End "fmla v29.8h, v7.8h, v13.8h\n" "fmla v30.8h, v6.8h, v13.8h\n" - "add x19, x12, x17\n" + "add x20, x13, x8\n" "tbz %x[n_channels], #2, 54f\n" - "ldr d12, [x19], #0x8\n" + "ldr d12, [x20], #0x8\n" "tbz %x[n_channels], #1, 53f\n" - "ld1 { v12.s }[2], [x19], #0x4\n" + "ld1 { v12.s }[2], [x20], #0x4\n" "tbz %x[n_channels], #0, 56f\n" - "ld1 { v12.h }[6], [x19]\n" + "ld1 { v12.h }[6], [x20]\n" "b 56f\n" "53:" // Tile loop: Oddments: Load inputs: (1, 1): Bit 2: Bit 1: Unset "tbz %x[n_channels], #0, 56f\n" - "ld1 { v12.h }[4], [x19]\n" + "ld1 { v12.h }[4], [x20]\n" "b 56f\n" "54:" // Tile loop: Oddments: Load inputs: (1, 1): Bit 2: Unset "tbz %x[n_channels], #1, 55f\n" - "ldr s12, [x19], #0x4\n" + "ldr s12, [x20], #0x4\n" "tbz %x[n_channels], #0, 56f\n" - "ld1 { v12.h }[2], [x19]\n" + "ld1 { v12.h }[2], [x20]\n" "b 56f\n" "55:" // Tile loop: Oddments: Load inputs: (1, 1): Bit 2: Unset: Bit 1: Unset - "ldr h12, [x19, #0x0]\n" + "ldr h12, [x20, #0x0]\n" "56:" // Tile loop: Oddments: Load inputs: (1, 1): Bit 2: End "fmla v23.8h, v4.8h, v12.8h\n" "fmla v24.8h, v3.8h, v12.8h\n" - "add x19, x12, x28\n" + "add x20, x13, x9\n" "fmla v26.8h, v1.8h, v12.8h\n" "fmla v27.8h, v0.8h, v12.8h\n" "tbz %x[n_channels], #2, 58f\n" - "ldr d11, [x19], #0x8\n" + "ldr d11, [x20], #0x8\n" "tbz %x[n_channels], #1, 57f\n" - "ld1 { v11.s }[2], [x19], #0x4\n" + "ld1 { v11.s }[2], [x20], #0x4\n" "tbz %x[n_channels], #0, 60f\n" - "ld1 { v11.h }[6], [x19]\n" + "ld1 { v11.h }[6], [x20]\n" "b 60f\n" "57:" // Tile loop: Oddments: Load inputs: (1, 3): Bit 2: Bit 1: Unset "tbz %x[n_channels], #0, 60f\n" - "ld1 { v11.h }[4], [x19]\n" + "ld1 { v11.h }[4], [x20]\n" "b 60f\n" "58:" // Tile loop: Oddments: Load inputs: (1, 3): Bit 2: Unset "tbz %x[n_channels], #1, 59f\n" - "ldr s11, [x19], #0x4\n" + "ldr s11, [x20], #0x4\n" "tbz %x[n_channels], #0, 60f\n" - "ld1 { v11.h }[2], [x19]\n" + "ld1 { v11.h }[2], [x20]\n" "b 60f\n" "59:" // Tile loop: Oddments: Load inputs: (1, 3): Bit 2: Unset: Bit 1: Unset - "ldr h11, [x19, #0x0]\n" + "ldr h11, [x20, #0x0]\n" "60:" // Tile loop: Oddments: Load inputs: (1, 3): Bit 2: End "fmla v24.8h, v5.8h, v11.8h\n" "fmla v25.8h, v4.8h, v11.8h\n" - "add x19, x26, x28\n" + "add x20, x27, x9\n" "fmla v27.8h, v2.8h, v11.8h\n" "fmla v28.8h, v1.8h, v11.8h\n" "tbz %x[n_channels], #2, 62f\n" - "ldr d13, [x19], #0x8\n" + "ldr d13, [x20], #0x8\n" "tbz %x[n_channels], #1, 61f\n" - "ld1 { v13.s }[2], [x19], #0x4\n" + "ld1 { v13.s }[2], [x20], #0x4\n" "tbz %x[n_channels], #0, 64f\n" - "ld1 { v13.h }[6], [x19]\n" + "ld1 { v13.h }[6], [x20]\n" "b 64f\n" "61:" // Tile loop: Oddments: Load inputs: (4, 3): Bit 2: Bit 1: Unset "tbz %x[n_channels], #0, 64f\n" - "ld1 { v13.h }[4], [x19]\n" + "ld1 { v13.h }[4], [x20]\n" "b 64f\n" "62:" // Tile loop: Oddments: Load inputs: (4, 3): Bit 2: Unset "tbz %x[n_channels], #1, 63f\n" - "ldr s13, [x19], #0x4\n" + "ldr s13, [x20], #0x4\n" "tbz %x[n_channels], #0, 64f\n" - "ld1 { v13.h }[2], [x19]\n" + "ld1 { v13.h }[2], [x20]\n" "b 64f\n" "63:" // Tile loop: Oddments: Load inputs: (4, 3): Bit 2: Unset: Bit 1: Unset - "ldr h13, [x19, #0x0]\n" + "ldr h13, [x20, #0x0]\n" "64:" // Tile loop: Oddments: Load inputs: (4, 3): Bit 2: End "fmla v30.8h, v8.8h, v13.8h\n" "fmla v31.8h, v7.8h, v13.8h\n" - "add x19, x9, x17\n" + "add x20, x10, x8\n" "tbz %x[n_channels], #2, 66f\n" - "ldr d12, [x19], #0x8\n" + "ldr d12, [x20], #0x8\n" "tbz %x[n_channels], #1, 65f\n" - "ld1 { v12.s }[2], [x19], #0x4\n" + "ld1 { v12.s }[2], [x20], #0x4\n" "tbz %x[n_channels], #0, 68f\n" - "ld1 { v12.h }[6], [x19]\n" + "ld1 { v12.h }[6], [x20]\n" "b 68f\n" "65:" // Tile loop: Oddments: Load inputs: (3, 1): Bit 2: Bit 1: Unset "tbz %x[n_channels], #0, 68f\n" - "ld1 { v12.h }[4], [x19]\n" + "ld1 { v12.h }[4], [x20]\n" "b 68f\n" "66:" // Tile loop: Oddments: Load inputs: (3, 1): Bit 2: Unset "tbz %x[n_channels], #1, 67f\n" - "ldr s12, [x19], #0x4\n" + "ldr s12, [x20], #0x4\n" "tbz %x[n_channels], #0, 68f\n" - "ld1 { v12.h }[2], [x19]\n" + "ld1 { v12.h }[2], [x20]\n" "b 68f\n" "67:" // Tile loop: Oddments: Load inputs: (3, 1): Bit 2: Unset: Bit 1: Unset - "ldr h12, [x19, #0x0]\n" + "ldr h12, [x20, #0x0]\n" "68:" // Tile loop: Oddments: Load inputs: (3, 1): Bit 2: End "fmla v26.8h, v7.8h, v12.8h\n" "fmla v27.8h, v6.8h, v12.8h\n" - "add x19, x15, x10\n" + "add x20, x16, x11\n" "fmla v29.8h, v4.8h, v12.8h\n" "fmla v30.8h, v3.8h, v12.8h\n" "tbz %x[n_channels], #2, 70f\n" - "ldr d11, [x19], #0x8\n" + "ldr d11, [x20], #0x8\n" "tbz %x[n_channels], #1, 69f\n" - "ld1 { v11.s }[2], [x19], #0x4\n" + "ld1 { v11.s }[2], [x20], #0x4\n" "tbz %x[n_channels], #0, 72f\n" - "ld1 { v11.h }[6], [x19]\n" + "ld1 { v11.h }[6], [x20]\n" "b 72f\n" "69:" // Tile loop: Oddments: Load inputs: (0, 2): Bit 2: Bit 1: Unset "tbz %x[n_channels], #0, 72f\n" - "ld1 { v11.h }[4], [x19]\n" + "ld1 { v11.h }[4], [x20]\n" "b 72f\n" "70:" // Tile loop: Oddments: Load inputs: (0, 2): Bit 2: Unset "tbz %x[n_channels], #1, 71f\n" - "ldr s11, [x19], #0x4\n" + "ldr s11, [x20], #0x4\n" "tbz %x[n_channels], #0, 72f\n" - "ld1 { v11.h }[2], [x19]\n" + "ld1 { v11.h }[2], [x20]\n" "b 72f\n" "71:" // Tile loop: Oddments: Load inputs: (0, 2): Bit 2: Unset: Bit 1: Unset - "ldr h11, [x19, #0x0]\n" + "ldr h11, [x20, #0x0]\n" "72:" // Tile loop: Oddments: Load inputs: (0, 2): Bit 2: End "fmla v23.8h, v2.8h, v11.8h\n" "fmla v24.8h, v1.8h, v11.8h\n" - "add x19, x9, x28\n" + "add x20, x10, x9\n" "fmla v25.8h, v0.8h, v11.8h\n" "tbz %x[n_channels], #2, 74f\n" - "ldr d13, [x19], #0x8\n" + "ldr d13, [x20], #0x8\n" "tbz %x[n_channels], #1, 73f\n" - "ld1 { v13.s }[2], [x19], #0x4\n" + "ld1 { v13.s }[2], [x20], #0x4\n" "tbz %x[n_channels], #0, 76f\n" - "ld1 { v13.h }[6], [x19]\n" + "ld1 { v13.h }[6], [x20]\n" "b 76f\n" "73:" // Tile loop: Oddments: Load inputs: (3, 3): Bit 2: Bit 1: Unset "tbz %x[n_channels], #0, 76f\n" - "ld1 { v13.h }[4], [x19]\n" + "ld1 { v13.h }[4], [x20]\n" "b 76f\n" "74:" // Tile loop: Oddments: Load inputs: (3, 3): Bit 2: Unset "tbz %x[n_channels], #1, 75f\n" - "ldr s13, [x19], #0x4\n" + "ldr s13, [x20], #0x4\n" "tbz %x[n_channels], #0, 76f\n" - "ld1 { v13.h }[2], [x19]\n" + "ld1 { v13.h }[2], [x20]\n" "b 76f\n" "75:" // Tile loop: Oddments: Load inputs: (3, 3): Bit 2: Unset: Bit 1: Unset - "ldr h13, [x19, #0x0]\n" + "ldr h13, [x20, #0x0]\n" "76:" // Tile loop: Oddments: Load inputs: (3, 3): Bit 2: End "fmla v27.8h, v8.8h, v13.8h\n" "fmla v28.8h, v7.8h, v13.8h\n" - "add x19, x11, XZR\n" + "add x20, x12, XZR\n" "fmla v30.8h, v5.8h, v13.8h\n" "fmla v31.8h, v4.8h, v13.8h\n" "tbz %x[n_channels], #2, 78f\n" - "ldr d12, [x19], #0x8\n" + "ldr d12, [x20], #0x8\n" "tbz %x[n_channels], #1, 77f\n" - "ld1 { v12.s }[2], [x19], #0x4\n" + "ld1 { v12.s }[2], [x20], #0x4\n" "tbz %x[n_channels], #0, 80f\n" - "ld1 { v12.h }[6], [x19]\n" + "ld1 { v12.h }[6], [x20]\n" "b 80f\n" "77:" // Tile loop: Oddments: Load inputs: (2, 0): Bit 2: Bit 1: Unset "tbz %x[n_channels], #0, 80f\n" - "ld1 { v12.h }[4], [x19]\n" + "ld1 { v12.h }[4], [x20]\n" "b 80f\n" "78:" // Tile loop: Oddments: Load inputs: (2, 0): Bit 2: Unset "tbz %x[n_channels], #1, 79f\n" - "ldr s12, [x19], #0x4\n" + "ldr s12, [x20], #0x4\n" "tbz %x[n_channels], #0, 80f\n" - "ld1 { v12.h }[2], [x19]\n" + "ld1 { v12.h }[2], [x20]\n" "b 80f\n" "79:" // Tile loop: Oddments: Load inputs: (2, 0): Bit 2: Unset: Bit 1: Unset - "ldr h12, [x19, #0x0]\n" + "ldr h12, [x20, #0x0]\n" "80:" // Tile loop: Oddments: Load inputs: (2, 0): Bit 2: End "fmla v23.8h, v6.8h, v12.8h\n" "fmla v26.8h, v3.8h, v12.8h\n" - "add x19, x11, x25\n" + "add x20, x12, x26\n" "fmla v29.8h, v0.8h, v12.8h\n" "tbz %x[n_channels], #2, 82f\n" - "ldr d11, [x19], #0x8\n" + "ldr d11, [x20], #0x8\n" "tbz %x[n_channels], #1, 81f\n" - "ld1 { v11.s }[2], [x19], #0x4\n" + "ld1 { v11.s }[2], [x20], #0x4\n" "tbz %x[n_channels], #0, 84f\n" - "ld1 { v11.h }[6], [x19]\n" + "ld1 { v11.h }[6], [x20]\n" "b 84f\n" "81:" // Tile loop: Oddments: Load inputs: (2, 4): Bit 2: Bit 1: Unset "tbz %x[n_channels], #0, 84f\n" - "ld1 { v11.h }[4], [x19]\n" + "ld1 { v11.h }[4], [x20]\n" "b 84f\n" "82:" // Tile loop: Oddments: Load inputs: (2, 4): Bit 2: Unset "tbz %x[n_channels], #1, 83f\n" - "ldr s11, [x19], #0x4\n" + "ldr s11, [x20], #0x4\n" "tbz %x[n_channels], #0, 84f\n" - "ld1 { v11.h }[2], [x19]\n" + "ld1 { v11.h }[2], [x20]\n" "b 84f\n" "83:" // Tile loop: Oddments: Load inputs: (2, 4): Bit 2: Unset: Bit 1: Unset - "ldr h11, [x19, #0x0]\n" + "ldr h11, [x20, #0x0]\n" "84:" // Tile loop: Oddments: Load inputs: (2, 4): Bit 2: End "fmla v25.8h, v8.8h, v11.8h\n" "fmla v28.8h, v5.8h, v11.8h\n" - "add x19, x26, x10\n" + "add x20, x27, x11\n" "fmla v31.8h, v2.8h, v11.8h\n" "tbz %x[n_channels], #2, 86f\n" - "ldr d13, [x19], #0x8\n" + "ldr d13, [x20], #0x8\n" "tbz %x[n_channels], #1, 85f\n" - "ld1 { v13.s }[2], [x19], #0x4\n" + "ld1 { v13.s }[2], [x20], #0x4\n" "tbz %x[n_channels], #0, 88f\n" - "ld1 { v13.h }[6], [x19]\n" + "ld1 { v13.h }[6], [x20]\n" "b 88f\n" "85:" // Tile loop: Oddments: Load inputs: (4, 2): Bit 2: Bit 1: Unset "tbz %x[n_channels], #0, 88f\n" - "ld1 { v13.h }[4], [x19]\n" + "ld1 { v13.h }[4], [x20]\n" "b 88f\n" "86:" // Tile loop: Oddments: Load inputs: (4, 2): Bit 2: Unset "tbz %x[n_channels], #1, 87f\n" - "ldr s13, [x19], #0x4\n" + "ldr s13, [x20], #0x4\n" "tbz %x[n_channels], #0, 88f\n" - "ld1 { v13.h }[2], [x19]\n" + "ld1 { v13.h }[2], [x20]\n" "b 88f\n" "87:" // Tile loop: Oddments: Load inputs: (4, 2): Bit 2: Unset: Bit 1: Unset - "ldr h13, [x19, #0x0]\n" + "ldr h13, [x20, #0x0]\n" "88:" // Tile loop: Oddments: Load inputs: (4, 2): Bit 2: End "fmla v29.8h, v8.8h, v13.8h\n" "fmla v30.8h, v7.8h, v13.8h\n" @@ -1029,127 +1029,127 @@ void a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst_direct_impl( "fmin v30.8h, v30.8h, v17.8h\n" "fmin v31.8h, v31.8h, v17.8h\n" "tbz %x[n_channels], #2, 90f\n" - "mov x21, x14\n" - "mov x20, x27\n" - "mov x19, x24\n" - "st1 { v23.d }[0], [x21], x16\n" - "st1 { v26.d }[0], [x20], x16\n" - "add x14, x14, #0x8\n" - "add x27, x27, #0x8\n" - "st1 { v29.d }[0], [x19], x16\n" - "add x24, x24, #0x8\n" - "st1 { v24.d }[0], [x21], x16\n" - "st1 { v27.d }[0], [x20], x16\n" - "st1 { v30.d }[0], [x19], x16\n" - "st1 { v25.d }[0], [x21]\n" - "st1 { v28.d }[0], [x20]\n" - "st1 { v31.d }[0], [x19]\n" + "mov x22, x15\n" + "mov x21, x28\n" + "st1 { v23.d }[0], [x22], x17\n" + "mov x20, x25\n" + "st1 { v26.d }[0], [x21], x17\n" + "add x15, x15, #0x8\n" + "st1 { v29.d }[0], [x20], x17\n" + "add x28, x28, #0x8\n" + "add x25, x25, #0x8\n" + "st1 { v24.d }[0], [x22], x17\n" + "st1 { v27.d }[0], [x21], x17\n" + "st1 { v30.d }[0], [x20], x17\n" + "st1 { v25.d }[0], [x22]\n" + "st1 { v28.d }[0], [x21]\n" + "st1 { v31.d }[0], [x20]\n" "tbz %x[n_channels], #1, 89f\n" - "mov x21, x14\n" - "mov x20, x27\n" - "mov x19, x24\n" - "st1 { v23.s }[2], [x21], x16\n" - "add x14, x14, #0x4\n" - "st1 { v26.s }[2], [x20], x16\n" - "add x27, x27, #0x4\n" - "add x24, x24, #0x4\n" - "st1 { v29.s }[2], [x19], x16\n" - "st1 { v24.s }[2], [x21], x16\n" - "st1 { v27.s }[2], [x20], x16\n" - "st1 { v30.s }[2], [x19], x16\n" - "st1 { v25.s }[2], [x21]\n" - "st1 { v28.s }[2], [x20]\n" - "st1 { v31.s }[2], [x19]\n" + "mov x22, x15\n" + "mov x21, x28\n" + "st1 { v23.s }[2], [x22], x17\n" + "mov x20, x25\n" + "st1 { v26.s }[2], [x21], x17\n" + "add x15, x15, #0x4\n" + "st1 { v29.s }[2], [x20], x17\n" + "add x28, x28, #0x4\n" + "add x25, x25, #0x4\n" + "st1 { v24.s }[2], [x22], x17\n" + "st1 { v27.s }[2], [x21], x17\n" + "st1 { v30.s }[2], [x20], x17\n" + "st1 { v25.s }[2], [x22]\n" + "st1 { v28.s }[2], [x21]\n" + "st1 { v31.s }[2], [x20]\n" "tbz %x[n_channels], #0, 92f\n" - "mov x21, x14\n" - "mov x20, x27\n" - "mov x19, x24\n" - "st1 { v23.h }[6], [x21], x16\n" - "st1 { v26.h }[6], [x20], x16\n" - "st1 { v29.h }[6], [x19], x16\n" - "st1 { v24.h }[6], [x21], x16\n" - "st1 { v27.h }[6], [x20], x16\n" - "st1 { v30.h }[6], [x19], x16\n" - "st1 { v25.h }[6], [x21]\n" - "st1 { v28.h }[6], [x20]\n" - "st1 { v31.h }[6], [x19]\n" + "mov x22, x15\n" + "mov x21, x28\n" + "st1 { v23.h }[6], [x22], x17\n" + "mov x20, x25\n" + "st1 { v26.h }[6], [x21], x17\n" + "st1 { v29.h }[6], [x20], x17\n" + "st1 { v24.h }[6], [x22], x17\n" + "st1 { v27.h }[6], [x21], x17\n" + "st1 { v30.h }[6], [x20], x17\n" + "st1 { v25.h }[6], [x22]\n" + "st1 { v28.h }[6], [x21]\n" + "st1 { v31.h }[6], [x20]\n" "b 92f\n" "89:" // Tile loop: Oddments: Store: Bit 2: Bit 1: Unset "tbz %x[n_channels], #0, 92f\n" - "mov x21, x14\n" - "mov x20, x27\n" - "st1 { v23.h }[4], [x21], x16\n" - "mov x19, x24\n" - "st1 { v26.h }[4], [x20], x16\n" - "st1 { v29.h }[4], [x19], x16\n" - "st1 { v24.h }[4], [x21], x16\n" - "st1 { v27.h }[4], [x20], x16\n" - "st1 { v30.h }[4], [x19], x16\n" - "st1 { v25.h }[4], [x21]\n" - "st1 { v28.h }[4], [x20]\n" - "st1 { v31.h }[4], [x19]\n" + "mov x22, x15\n" + "mov x21, x28\n" + "st1 { v23.h }[4], [x22], x17\n" + "mov x20, x25\n" + "st1 { v26.h }[4], [x21], x17\n" + "st1 { v29.h }[4], [x20], x17\n" + "st1 { v24.h }[4], [x22], x17\n" + "st1 { v27.h }[4], [x21], x17\n" + "st1 { v30.h }[4], [x20], x17\n" + "st1 { v25.h }[4], [x22]\n" + "st1 { v28.h }[4], [x21]\n" + "st1 { v31.h }[4], [x20]\n" "b 92f\n" "90:" // Tile loop: Oddments: Store: Bit 2: Unset "tbz %x[n_channels], #1, 91f\n" - "mov x21, x14\n" - "mov x20, x27\n" - "st1 { v23.s }[0], [x21], x16\n" - "mov x19, x24\n" - "st1 { v26.s }[0], [x20], x16\n" - "add x14, x14, #0x4\n" - "st1 { v29.s }[0], [x19], x16\n" - "add x27, x27, #0x4\n" - "add x24, x24, #0x4\n" - "st1 { v24.s }[0], [x21], x16\n" - "st1 { v27.s }[0], [x20], x16\n" - "st1 { v30.s }[0], [x19], x16\n" - "st1 { v25.s }[0], [x21]\n" - "st1 { v28.s }[0], [x20]\n" - "st1 { v31.s }[0], [x19]\n" + "mov x22, x15\n" + "mov x21, x28\n" + "st1 { v23.s }[0], [x22], x17\n" + "mov x20, x25\n" + "st1 { v26.s }[0], [x21], x17\n" + "add x15, x15, #0x4\n" + "st1 { v29.s }[0], [x20], x17\n" + "add x28, x28, #0x4\n" + "add x25, x25, #0x4\n" + "st1 { v24.s }[0], [x22], x17\n" + "st1 { v27.s }[0], [x21], x17\n" + "st1 { v30.s }[0], [x20], x17\n" + "st1 { v25.s }[0], [x22]\n" + "st1 { v28.s }[0], [x21]\n" + "st1 { v31.s }[0], [x20]\n" "tbz %x[n_channels], #0, 92f\n" - "mov x21, x14\n" - "mov x20, x27\n" - "mov x19, x24\n" - "st1 { v23.h }[2], [x21], x16\n" - "st1 { v26.h }[2], [x20], x16\n" - "st1 { v29.h }[2], [x19], x16\n" - "st1 { v24.h }[2], [x21], x16\n" - "st1 { v27.h }[2], [x20], x16\n" - "st1 { v30.h }[2], [x19], x16\n" - "st1 { v25.h }[2], [x21]\n" - "st1 { v28.h }[2], [x20]\n" - "st1 { v31.h }[2], [x19]\n" + "mov x22, x15\n" + "mov x21, x28\n" + "st1 { v23.h }[2], [x22], x17\n" + "mov x20, x25\n" + "st1 { v26.h }[2], [x21], x17\n" + "st1 { v29.h }[2], [x20], x17\n" + "st1 { v24.h }[2], [x22], x17\n" + "st1 { v27.h }[2], [x21], x17\n" + "st1 { v30.h }[2], [x20], x17\n" + "st1 { v25.h }[2], [x22]\n" + "st1 { v28.h }[2], [x21]\n" + "st1 { v31.h }[2], [x20]\n" "b 92f\n" "91:" // Tile loop: Oddments: Store: Bit 2: Unset: Bit 1: Unset - "mov x21, x14\n" - "mov x20, x27\n" - "st1 { v23.h }[0], [x21], x16\n" - "mov x19, x24\n" - "st1 { v26.h }[0], [x20], x16\n" - "st1 { v29.h }[0], [x19], x16\n" - "st1 { v24.h }[0], [x21], x16\n" - "st1 { v27.h }[0], [x20], x16\n" - "st1 { v30.h }[0], [x19], x16\n" - "st1 { v25.h }[0], [x21]\n" - "st1 { v28.h }[0], [x20]\n" - "st1 { v31.h }[0], [x19]\n" + "mov x22, x15\n" + "mov x21, x28\n" + "st1 { v23.h }[0], [x22], x17\n" + "mov x20, x25\n" + "st1 { v26.h }[0], [x21], x17\n" + "st1 { v29.h }[0], [x20], x17\n" + "st1 { v24.h }[0], [x22], x17\n" + "st1 { v27.h }[0], [x21], x17\n" + "st1 { v30.h }[0], [x20], x17\n" + "st1 { v25.h }[0], [x22]\n" + "st1 { v28.h }[0], [x21]\n" + "st1 { v31.h }[0], [x20]\n" "92:" // Tile loop: Oddments: Store: Bit 2: End "93:" // Tile loop: End - "ldr x22, [%x[params_struct], %[offsetof_args_tile_j]]\n" - "ldr x23, [%x[params_struct], %[offsetof_args_tile_i]]\n" - "add x22, x22, #0x1\n" - "add x20, x23, #0x1\n" - "ldr x19, [%x[params_struct], %[offsetof_args_n_tile_cols]]\n" - "cmp x22, x19\n" - "ldr x19, [%x[params_struct], %[offsetof_args_n_tile_rows]]\n" - "csel x23, x23, x20, LT\n" - "csel x22, x22, XZR, LT\n" - "cmp x23, x19\n" + "ldr x23, [%x[params_struct], %[offsetof_args_tile_j]]\n" + "ldr x24, [%x[params_struct], %[offsetof_args_tile_i]]\n" + "add x23, x23, #0x1\n" + "add x21, x24, #0x1\n" + "ldr x20, [%x[params_struct], %[offsetof_args_n_tile_cols]]\n" + "cmp x23, x20\n" + "ldr x20, [%x[params_struct], %[offsetof_args_n_tile_rows]]\n" + "csel x24, x24, x21, LT\n" + "csel x23, x23, XZR, LT\n" + "cmp x24, x20\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" (¶ms_struct) - : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v16", "v17", "v18", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v16", "v17", "v18", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" ); } |