From 5264b7d5555ec980f9c52c719122479d0d676af8 Mon Sep 17 00:00:00 2001 From: Pablo Tello Date: Mon, 21 Oct 2019 14:25:41 +0100 Subject: COMPMID-2576: Fuse activation in Winograd output transform. Change-Id: I26dd1307847adeaaefae0a7374b9858c07d71372 Signed-off-by: Pablo Tello Reviewed-on: https://review.mlplatform.org/c/2172 Comments-Addressed: Arm Jenkins Tested-by: Arm Jenkins Reviewed-by: Gian Marco Iodice --- .../output_4x4_3x3_fp32_fp32_integers.cpp | 1769 +------------------- 1 file changed, 84 insertions(+), 1685 deletions(-) (limited to 'src/core/NEON/kernels/convolution/winograd/winograd_transforms/output_4x4_3x3_fp32_fp32_integers.cpp') diff --git a/src/core/NEON/kernels/convolution/winograd/winograd_transforms/output_4x4_3x3_fp32_fp32_integers.cpp b/src/core/NEON/kernels/convolution/winograd/winograd_transforms/output_4x4_3x3_fp32_fp32_integers.cpp index fd16a4df1c..292999c8bc 100644 --- a/src/core/NEON/kernels/convolution/winograd/winograd_transforms/output_4x4_3x3_fp32_fp32_integers.cpp +++ b/src/core/NEON/kernels/convolution/winograd/winograd_transforms/output_4x4_3x3_fp32_fp32_integers.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019 ARM Limited. + * Copyright (c) 2017-2019 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -28,1683 +28,6 @@ namespace winograd { -#ifdef __aarch64__ - -template <> -void OutputTransform<3, 3, 6, 6, float, float, winograd::WinogradRoots::Integers>::transform_tile( - int n_channels, - const float* inptr, - const int matrix_stride, - const float* bptr, - float* output, - const int output_row_stride, - const int output_col_stride -) -{ - const float coeffs[2] = {2.0f, 4.0f}; - if (bptr != nullptr) - { - __asm__ __volatile__ ( - "ldr d0, [%[pcoeffs]]\n" - "add x21, %[in_col_stride1], %[in_col_stride1]\n" - "add x22, x21, %[in_col_stride1]\n" - "add x25, %[inptr0], %[in_row_stride]\n" - "add x15, %[output_col_stride1], %[output_col_stride1]\n" - "add x23, x22, %[in_col_stride1]\n" - "add x13, x25, %[in_row_stride]\n" - "add x16, x15, %[output_col_stride1]\n" - "add x24, x23, %[in_col_stride1]\n" - "add x26, x13, %[in_row_stride]\n" - "add x17, %[outptr0], %[output_row_stride]\n" - "add x14, x26, %[in_row_stride]\n" - "add x28, x17, %[output_row_stride]\n" - "lsr x19, %[n_channels], #2\n" - "add x27, x14, %[in_row_stride]\n" - "add x18, x28, %[output_row_stride]\n" - "and x20, %[n_channels], #3\n" - "cbz x19, 4f\n" - "1:\n" - "ldr q19, [%[inptr0]]\n" - "subs x19, x19, #1\n" - "ldr q20, [%[inptr0], %[in_col_stride1]]\n" - "ldr q4, [%[inptr0], x21]\n" - "fadd v1.4s, v20.4s, v4.4s\n" - "ldr q17, [%[inptr0], x22]\n" - "fsub v7.4s, v20.4s, v4.4s\n" - "ldr q22, [%[inptr0], x23]\n" - "fadd v5.4s, v17.4s, v22.4s\n" - "ldr q18, [%[inptr0], x24]\n" - "fsub v10.4s, v17.4s, v22.4s\n" - "ldr q25, [x25]\n" - "fadd v8.4s, v19.4s, v1.4s\n" - "ldr q12, [x25, %[in_col_stride1]]\n" - "mov v4.16b, v1.16b\n" - "ldr q23, [x25, x21]\n" - "mov v1.16b, v7.16b\n" - "ldr q9, [x25, x22]\n" - "fmul v10.4s, v10.4s, v0.s[0]\n" - "ldr q11, [x25, x23]\n" - "fadd v8.4s, v8.4s, v5.4s\n" - "ldr q6, [x25, x24]\n" - "fmla v4.4s, v5.4s, v0.s[1]\n" - "fadd v7.4s, v7.4s, v10.4s\n" - "fmla v1.4s, v10.4s, v0.s[1]\n" - "fadd v1.4s, v1.4s, v18.4s\n" - "beq 3f\n" - "2:\n" - "fadd v3.4s, v12.4s, v23.4s\n" - "ldr q2, [x13]\n" - "fadd v27.4s, v9.4s, v11.4s\n" - "ldr q21, [x13, %[in_col_stride1]]\n" - "fsub v16.4s, v12.4s, v23.4s\n" - "ldr q26, [x13, x21]\n" - "fsub v9.4s, v9.4s, v11.4s\n" - "ldr q17, [x13, x22]\n" - "fadd v14.4s, v25.4s, v3.4s\n" - "ldr q19, [x13, x23]\n" - "mov v11.16b, v3.16b\n" - "ldr q10, [x13, x24]\n" - "mov v3.16b, v16.16b\n" - "ldr q15, [x26]\n" - "fmul v9.4s, v9.4s, v0.s[0]\n" - "ldr q12, [x26, %[in_col_stride1]]\n" - "fadd v14.4s, v14.4s, v27.4s\n" - "ldr q20, [x26, x21]\n" - "fmla v11.4s, v27.4s, v0.s[1]\n" - "ldr q24, [x26, x22]\n" - "fadd v23.4s, v21.4s, v26.4s\n" - "ldr q29, [x26, x23]\n" - "fadd v13.4s, v16.4s, v9.4s\n" - "ldr q5, [x26, x24]\n" - "fmla v3.4s, v9.4s, v0.s[1]\n" - "ldr q18, [x14]\n" - "fadd v30.4s, v17.4s, v19.4s\n" - "add %[inptr0], %[inptr0], #16\n" - "fadd v16.4s, v2.4s, v23.4s\n" - "add x25, x25, #16\n" - "fsub v21.4s, v21.4s, v26.4s\n" - "ldr q22, [x14, %[in_col_stride1]]\n" - "fadd v3.4s, v3.4s, v6.4s\n" - "ldr q28, [x14, x21]\n" - "fsub v19.4s, v17.4s, v19.4s\n" - "add x13, x13, #16\n" - "fadd v16.4s, v16.4s, v30.4s\n" - "add x26, x26, #16\n" - "mov v17.16b, v23.16b\n" - "subs x19, x19, #1\n" - "fadd v26.4s, v12.4s, v20.4s\n" - "fsub v9.4s, v12.4s, v20.4s\n" - "fmul v19.4s, v19.4s, v0.s[0]\n" - "ldr q20, [x14, x22]\n" - "fmla v17.4s, v30.4s, v0.s[1]\n" - "fadd v25.4s, v24.4s, v29.4s\n" - "fsub v12.4s, v24.4s, v29.4s\n" - "fadd v24.4s, v22.4s, v28.4s\n" - "fadd v23.4s, v15.4s, v26.4s\n" - "mov v15.16b, v26.16b\n" - "fsub v22.4s, v22.4s, v28.4s\n" - "fadd v29.4s, v14.4s, v16.4s\n" - "fsub v16.4s, v14.4s, v16.4s\n" - "ldr q28, [x14, x23]\n" - "fmul v12.4s, v12.4s, v0.s[0]\n" - "fmla v15.4s, v25.4s, v0.s[1]\n" - "fadd v23.4s, v23.4s, v25.4s\n" - "mov v6.16b, v21.16b\n" - "fadd v30.4s, v21.4s, v19.4s\n" - "fadd v26.4s, v18.4s, v24.4s\n" - "mov v25.16b, v24.16b\n" - "fadd v18.4s, v8.4s, v29.4s\n" - "fmla v6.4s, v19.4s, v0.s[1]\n" - "fadd v27.4s, v20.4s, v28.4s\n" - "fsub v21.4s, v20.4s, v28.4s\n" - "mov v19.16b, v29.16b\n" - "fadd v29.4s, v13.4s, v30.4s\n" - "fsub v8.4s, v13.4s, v30.4s\n" - "fadd v14.4s, v9.4s, v12.4s\n" - "fadd v6.4s, v6.4s, v10.4s\n" - "ldr q20, [x14, x24]\n" - "fadd v26.4s, v26.4s, v27.4s\n" - "add x14, x14, #16\n" - "fmla v9.4s, v12.4s, v0.s[1]\n" - "ldr q24, [x27]\n" - "fmul v21.4s, v21.4s, v0.s[0]\n" - "fmla v25.4s, v27.4s, v0.s[1]\n" - "fadd v10.4s, v7.4s, v29.4s\n" - "ldr q2, [%[bptr]]\n" - "mov v7.16b, v29.16b\n" - "add %[bptr], %[bptr], #16\n" - "fadd v9.4s, v9.4s, v5.4s\n" - "fadd v13.4s, v23.4s, v26.4s\n" - "fsub v23.4s, v23.4s, v26.4s\n" - "fadd v27.4s, v11.4s, v17.4s\n" - "fsub v11.4s, v11.4s, v17.4s\n" - "fadd v30.4s, v15.4s, v25.4s\n" - "fsub v15.4s, v15.4s, v25.4s\n" - "ldr q28, [x27, %[in_col_stride1]]\n" - "fadd v18.4s, v18.4s, v13.4s\n" - "fmla v19.4s, v13.4s, v0.s[1]\n" - "fadd v26.4s, v22.4s, v21.4s\n" - "mov v12.16b, v22.16b\n" - "fmul v23.4s, v23.4s, v0.s[0]\n" - "fadd v17.4s, v4.4s, v27.4s\n" - "fmul v15.4s, v15.4s, v0.s[0]\n" - "mov v4.16b, v27.16b\n" - "fmla v12.4s, v21.4s, v0.s[1]\n" - "ldr q22, [x27, x21]\n" - "fadd v18.4s, v18.4s, v2.4s\n" - "fadd v19.4s, v19.4s, v2.4s\n" - "fadd v17.4s, v17.4s, v30.4s\n" - "fmla v4.4s, v30.4s, v0.s[1]\n" - "fadd v25.4s, v28.4s, v22.4s\n" - "fsub v27.4s, v28.4s, v22.4s\n" - "fadd v12.4s, v12.4s, v20.4s\n" - "ldr q29, [x27, x22]\n" - "str q18, [%[outptr0]]\n" - "fadd v22.4s, v16.4s, v23.4s\n" - "str q19, [x28]\n" - "fadd v28.4s, v24.4s, v25.4s\n" - "ldr q30, [x27, x23]\n" - "fadd v20.4s, v29.4s, v30.4s\n" - "fsub v18.4s, v29.4s, v30.4s\n" - "mov v21.16b, v25.16b\n" - "ldr q25, [x27, x24]\n" - "fmla v16.4s, v23.4s, v0.s[1]\n" - "ldr q19, [%[inptr0]]\n" - "fadd v17.4s, v17.4s, v2.4s\n" - "add x27, x27, #16\n" - "fadd v28.4s, v28.4s, v20.4s\n" - "fmul v18.4s, v18.4s, v0.s[0]\n" - "fmla v21.4s, v20.4s, v0.s[1]\n" - "ldr q20, [%[inptr0], %[in_col_stride1]]\n" - "fadd v22.4s, v22.4s, v2.4s\n" - "fadd v4.4s, v4.4s, v2.4s\n" - "str q17, [%[outptr0], x15]\n" - "mov v24.16b, v27.16b\n" - "fadd v23.4s, v27.4s, v18.4s\n" - "fadd v16.4s, v16.4s, v28.4s\n" - "fadd v13.4s, v14.4s, v26.4s\n" - "fsub v30.4s, v14.4s, v26.4s\n" - "str q22, [x17]\n" - "fmla v24.4s, v18.4s, v0.s[1]\n" - "str q4, [x28, x15]\n" - "mov v14.16b, v8.16b\n" - "fadd v29.4s, v11.4s, v15.4s\n" - "ldr q4, [%[inptr0], x21]\n" - "fadd v10.4s, v10.4s, v13.4s\n" - "ldr q17, [%[inptr0], x22]\n" - "fadd v24.4s, v24.4s, v25.4s\n" - "ldr q22, [%[inptr0], x23]\n" - "fmul v30.4s, v30.4s, v0.s[0]\n" - "fmla v7.4s, v13.4s, v0.s[1]\n" - "mov v26.16b, v11.16b\n" - "fadd v13.4s, v3.4s, v6.4s\n" - "fsub v3.4s, v3.4s, v6.4s\n" - "ldr q18, [%[inptr0], x24]\n" - "fadd v10.4s, v10.4s, v2.4s\n" - "fadd v29.4s, v29.4s, v2.4s\n" - "fadd v8.4s, v8.4s, v30.4s\n" - "fmla v14.4s, v30.4s, v0.s[1]\n" - "fmla v26.4s, v15.4s, v0.s[1]\n" - "ldr q25, [x25]\n" - "fadd v27.4s, v9.4s, v12.4s\n" - "fadd v1.4s, v1.4s, v13.4s\n" - "str q10, [%[outptr0], %[output_col_stride1]]\n" - "fsub v6.4s, v9.4s, v12.4s\n" - "str q29, [x17, x15]\n" - "fadd v14.4s, v14.4s, v23.4s\n" - "fadd v26.4s, v26.4s, v21.4s\n" - "ldr q12, [x25, %[in_col_stride1]]\n" - "fadd v1.4s, v1.4s, v27.4s\n" - "ldr q23, [x25, x21]\n" - "fmul v6.4s, v6.4s, v0.s[0]\n" - "ldr q9, [x25, x22]\n" - "mov v5.16b, v13.16b\n" - "ldr q11, [x25, x23]\n" - "mov v13.16b, v3.16b\n" - "fadd v8.4s, v8.4s, v2.4s\n" - "fadd v1.4s, v1.4s, v2.4s\n" - "fadd v7.4s, v7.4s, v2.4s\n" - "fadd v10.4s, v3.4s, v6.4s\n" - "fmla v5.4s, v27.4s, v0.s[1]\n" - "fmla v13.4s, v6.4s, v0.s[1]\n" - "ldr q6, [x25, x24]\n" - "str q8, [x17, %[output_col_stride1]]\n" - "fadd v16.4s, v16.4s, v2.4s\n" - "str q1, [%[outptr0], x16]\n" - "fadd v14.4s, v14.4s, v2.4s\n" - "str q7, [x28, %[output_col_stride1]]\n" - "fadd v10.4s, v10.4s, v2.4s\n" - "fadd v13.4s, v13.4s, v24.4s\n" - "add %[outptr0], %[outptr0], #16\n" - "str q16, [x18]\n" - "fadd v5.4s, v5.4s, v2.4s\n" - "str q14, [x18, %[output_col_stride1]]\n" - "fadd v26.4s, v26.4s, v2.4s\n" - "str q10, [x17, x16]\n" - "fadd v1.4s, v20.4s, v4.4s\n" - "fadd v13.4s, v13.4s, v2.4s\n" - "add x17, x17, #16\n" - "str q5, [x28, x16]\n" - "fadd v5.4s, v17.4s, v22.4s\n" - "str q26, [x18, x15]\n" - "fsub v7.4s, v20.4s, v4.4s\n" - "fadd v8.4s, v19.4s, v1.4s\n" - "add x28, x28, #16\n" - "str q13, [x18, x16]\n" - "mov v4.16b, v1.16b\n" - "fsub v10.4s, v17.4s, v22.4s\n" - "add x18, x18, #16\n" - "mov v1.16b, v7.16b\n" - "fadd v8.4s, v8.4s, v5.4s\n" - "fmla v4.4s, v5.4s, v0.s[1]\n" - "fmul v10.4s, v10.4s, v0.s[0]\n" - "fadd v7.4s, v7.4s, v10.4s\n" - "fmla v1.4s, v10.4s, v0.s[1]\n" - "fadd v1.4s, v1.4s, v18.4s\n" - "bne 2b\n" - "3:\n" - "fadd v3.4s, v12.4s, v23.4s\n" - "ldr q2, [x13]\n" - "fadd v27.4s, v9.4s, v11.4s\n" - "ldr q21, [x13, %[in_col_stride1]]\n" - "fsub v16.4s, v12.4s, v23.4s\n" - "ldr q26, [x13, x21]\n" - "fsub v9.4s, v9.4s, v11.4s\n" - "ldr q17, [x13, x22]\n" - "fadd v14.4s, v25.4s, v3.4s\n" - "ldr q19, [x13, x23]\n" - "mov v11.16b, v3.16b\n" - "ldr q10, [x13, x24]\n" - "mov v3.16b, v16.16b\n" - "ldr q15, [x26]\n" - "fmul v9.4s, v9.4s, v0.s[0]\n" - "ldr q12, [x26, %[in_col_stride1]]\n" - "fadd v14.4s, v14.4s, v27.4s\n" - "ldr q20, [x26, x21]\n" - "fmla v11.4s, v27.4s, v0.s[1]\n" - "ldr q24, [x26, x22]\n" - "fadd v23.4s, v21.4s, v26.4s\n" - "ldr q29, [x26, x23]\n" - "fadd v13.4s, v16.4s, v9.4s\n" - "ldr q5, [x26, x24]\n" - "fmla v3.4s, v9.4s, v0.s[1]\n" - "ldr q18, [x14]\n" - "fadd v30.4s, v17.4s, v19.4s\n" - "add %[inptr0], %[inptr0], #16\n" - "fadd v16.4s, v2.4s, v23.4s\n" - "add x25, x25, #16\n" - "fsub v21.4s, v21.4s, v26.4s\n" - "ldr q22, [x14, %[in_col_stride1]]\n" - "fadd v3.4s, v3.4s, v6.4s\n" - "ldr q28, [x14, x21]\n" - "fsub v19.4s, v17.4s, v19.4s\n" - "add x13, x13, #16\n" - "fadd v16.4s, v16.4s, v30.4s\n" - "add x26, x26, #16\n" - "mov v17.16b, v23.16b\n" - "fadd v26.4s, v12.4s, v20.4s\n" - "fsub v9.4s, v12.4s, v20.4s\n" - "ldr q2, [%[bptr]]\n" - "fmul v19.4s, v19.4s, v0.s[0]\n" - "add %[bptr], %[bptr], #16\n" - "fmla v17.4s, v30.4s, v0.s[1]\n" - "fadd v25.4s, v24.4s, v29.4s\n" - "fadd v23.4s, v15.4s, v26.4s\n" - "fsub v12.4s, v24.4s, v29.4s\n" - "mov v15.16b, v26.16b\n" - "fadd v24.4s, v22.4s, v28.4s\n" - "fsub v22.4s, v22.4s, v28.4s\n" - "fadd v29.4s, v14.4s, v16.4s\n" - "fsub v16.4s, v14.4s, v16.4s\n" - "ldr q20, [x14, x22]\n" - "fadd v23.4s, v23.4s, v25.4s\n" - "fmul v12.4s, v12.4s, v0.s[0]\n" - "fmla v15.4s, v25.4s, v0.s[1]\n" - "mov v6.16b, v21.16b\n" - "fadd v30.4s, v21.4s, v19.4s\n" - "fadd v26.4s, v18.4s, v24.4s\n" - "mov v25.16b, v24.16b\n" - "fadd v18.4s, v8.4s, v29.4s\n" - "fmla v6.4s, v19.4s, v0.s[1]\n" - "mov v19.16b, v29.16b\n" - "fadd v27.4s, v11.4s, v17.4s\n" - "fsub v11.4s, v11.4s, v17.4s\n" - "fadd v29.4s, v13.4s, v30.4s\n" - "fsub v8.4s, v13.4s, v30.4s\n" - "fadd v14.4s, v9.4s, v12.4s\n" - "fadd v6.4s, v6.4s, v10.4s\n" - "ldr q28, [x14, x23]\n" - "fadd v17.4s, v4.4s, v27.4s\n" - "mov v4.16b, v27.16b\n" - "fmla v9.4s, v12.4s, v0.s[1]\n" - "fadd v27.4s, v20.4s, v28.4s\n" - "fsub v21.4s, v20.4s, v28.4s\n" - "fadd v10.4s, v7.4s, v29.4s\n" - "mov v7.16b, v29.16b\n" - "fadd v13.4s, v3.4s, v6.4s\n" - "fsub v3.4s, v3.4s, v6.4s\n" - "ldr q20, [x14, x24]\n" - "fadd v9.4s, v9.4s, v5.4s\n" - "fadd v26.4s, v26.4s, v27.4s\n" - "fmul v21.4s, v21.4s, v0.s[0]\n" - "add x14, x14, #16\n" - "fmla v25.4s, v27.4s, v0.s[1]\n" - "mov v12.16b, v22.16b\n" - "fadd v1.4s, v1.4s, v13.4s\n" - "mov v5.16b, v13.16b\n" - "fadd v13.4s, v23.4s, v26.4s\n" - "fsub v23.4s, v23.4s, v26.4s\n" - "fadd v26.4s, v22.4s, v21.4s\n" - "ldr q24, [x27]\n" - "fmla v12.4s, v21.4s, v0.s[1]\n" - "fadd v30.4s, v15.4s, v25.4s\n" - "fsub v15.4s, v15.4s, v25.4s\n" - "ldr q28, [x27, %[in_col_stride1]]\n" - "fadd v18.4s, v18.4s, v13.4s\n" - "fmul v23.4s, v23.4s, v0.s[0]\n" - "fmla v19.4s, v13.4s, v0.s[1]\n" - "ldr q22, [x27, x21]\n" - "fadd v12.4s, v12.4s, v20.4s\n" - "ldr q29, [x27, x22]\n" - "fadd v17.4s, v17.4s, v30.4s\n" - "fmul v15.4s, v15.4s, v0.s[0]\n" - "fmla v4.4s, v30.4s, v0.s[1]\n" - "fadd v25.4s, v28.4s, v22.4s\n" - "fsub v27.4s, v28.4s, v22.4s\n" - "fadd v22.4s, v16.4s, v23.4s\n" - "fadd v18.4s, v18.4s, v2.4s\n" - "fadd v17.4s, v17.4s, v2.4s\n" - "fadd v19.4s, v19.4s, v2.4s\n" - "fadd v28.4s, v24.4s, v25.4s\n" - "mov v21.16b, v25.16b\n" - "fmla v16.4s, v23.4s, v0.s[1]\n" - "ldr q30, [x27, x23]\n" - "str q18, [%[outptr0]]\n" - "fadd v20.4s, v29.4s, v30.4s\n" - "str q17, [%[outptr0], x15]\n" - "fsub v18.4s, v29.4s, v30.4s\n" - "str q19, [x28]\n" - "mov v24.16b, v27.16b\n" - "fadd v13.4s, v14.4s, v26.4s\n" - "ldr q25, [x27, x24]\n" - "fadd v28.4s, v28.4s, v20.4s\n" - "add x27, x27, #16\n" - "fmul v18.4s, v18.4s, v0.s[0]\n" - "fmla v21.4s, v20.4s, v0.s[1]\n" - "fsub v30.4s, v14.4s, v26.4s\n" - "mov v14.16b, v8.16b\n" - "fadd v10.4s, v10.4s, v13.4s\n" - "fmla v7.4s, v13.4s, v0.s[1]\n" - "fadd v16.4s, v16.4s, v28.4s\n" - "fadd v29.4s, v11.4s, v15.4s\n" - "fadd v23.4s, v27.4s, v18.4s\n" - "fmla v24.4s, v18.4s, v0.s[1]\n" - "fmul v30.4s, v30.4s, v0.s[0]\n" - "mov v26.16b, v11.16b\n" - "fadd v27.4s, v9.4s, v12.4s\n" - "fsub v6.4s, v9.4s, v12.4s\n" - "mov v13.16b, v3.16b\n" - "fadd v10.4s, v10.4s, v2.4s\n" - "fadd v24.4s, v24.4s, v25.4s\n" - "fmla v26.4s, v15.4s, v0.s[1]\n" - "fadd v8.4s, v8.4s, v30.4s\n" - "fmla v14.4s, v30.4s, v0.s[1]\n" - "fadd v1.4s, v1.4s, v27.4s\n" - "fmul v6.4s, v6.4s, v0.s[0]\n" - "str q10, [%[outptr0], %[output_col_stride1]]\n" - "fmla v5.4s, v27.4s, v0.s[1]\n" - "fadd v26.4s, v26.4s, v21.4s\n" - "fadd v22.4s, v22.4s, v2.4s\n" - "fadd v14.4s, v14.4s, v23.4s\n" - "fadd v8.4s, v8.4s, v2.4s\n" - "fadd v10.4s, v3.4s, v6.4s\n" - "fmla v13.4s, v6.4s, v0.s[1]\n" - "fadd v1.4s, v1.4s, v2.4s\n" - "fadd v29.4s, v29.4s, v2.4s\n" - "str q22, [x17]\n" - "fadd v7.4s, v7.4s, v2.4s\n" - "str q8, [x17, %[output_col_stride1]]\n" - "fadd v4.4s, v4.4s, v2.4s\n" - "fadd v13.4s, v13.4s, v24.4s\n" - "fadd v10.4s, v10.4s, v2.4s\n" - "str q1, [%[outptr0], x16]\n" - "fadd v5.4s, v5.4s, v2.4s\n" - "str q29, [x17, x15]\n" - "fadd v16.4s, v16.4s, v2.4s\n" - "str q7, [x28, %[output_col_stride1]]\n" - "fadd v14.4s, v14.4s, v2.4s\n" - "str q10, [x17, x16]\n" - "fadd v26.4s, v26.4s, v2.4s\n" - "str q4, [x28, x15]\n" - "fadd v13.4s, v13.4s, v2.4s\n" - "str q5, [x28, x16]\n" - "add %[outptr0], %[outptr0], #16\n" - "str q16, [x18]\n" - "add x17, x17, #16\n" - "str q14, [x18, %[output_col_stride1]]\n" - "add x28, x28, #16\n" - "str q26, [x18, x15]\n" - "str q13, [x18, x16]\n" - "add x18, x18, #16\n" - "4:\n" - "cmp x20, #2\n" - "blt 5f\n" - "ldr d19, [%[inptr0]]\n" - "ldr d20, [%[inptr0], %[in_col_stride1]]\n" - "sub x20, x20, #2\n" - "ldr d4, [%[inptr0], x21]\n" - "ldr d17, [%[inptr0], x22]\n" - "fadd v1.4s, v20.4s, v4.4s\n" - "ldr d22, [%[inptr0], x23]\n" - "fadd v5.4s, v17.4s, v22.4s\n" - "ldr d18, [%[inptr0], x24]\n" - "fsub v7.4s, v20.4s, v4.4s\n" - "ldr d25, [x25]\n" - "fsub v10.4s, v17.4s, v22.4s\n" - "ldr d12, [x25, %[in_col_stride1]]\n" - "fadd v8.4s, v19.4s, v1.4s\n" - "ldr d23, [x25, x21]\n" - "mov v4.16b, v1.16b\n" - "ldr d9, [x25, x22]\n" - "mov v1.16b, v7.16b\n" - "ldr d11, [x25, x23]\n" - "fmul v10.4s, v10.4s, v0.s[0]\n" - "ldr d6, [x25, x24]\n" - "fadd v8.4s, v8.4s, v5.4s\n" - "ldr d2, [x13]\n" - "fmla v4.4s, v5.4s, v0.s[1]\n" - "ldr d21, [x13, %[in_col_stride1]]\n" - "fadd v3.4s, v12.4s, v23.4s\n" - "ldr d26, [x13, x21]\n" - "fadd v7.4s, v7.4s, v10.4s\n" - "ldr d17, [x13, x22]\n" - "fmla v1.4s, v10.4s, v0.s[1]\n" - "ldr d19, [x13, x23]\n" - "fadd v27.4s, v9.4s, v11.4s\n" - "ldr d10, [x13, x24]\n" - "fadd v14.4s, v25.4s, v3.4s\n" - "ldr d15, [x26]\n" - "fsub v16.4s, v12.4s, v23.4s\n" - "ldr d12, [x26, %[in_col_stride1]]\n" - "fadd v1.4s, v1.4s, v18.4s\n" - "ldr d20, [x26, x21]\n" - "fsub v9.4s, v9.4s, v11.4s\n" - "ldr d24, [x26, x22]\n" - "fadd v14.4s, v14.4s, v27.4s\n" - "ldr d29, [x26, x23]\n" - "mov v11.16b, v3.16b\n" - "ldr d5, [x26, x24]\n" - "mov v3.16b, v16.16b\n" - "ldr d18, [x14]\n" - "fmul v9.4s, v9.4s, v0.s[0]\n" - "add %[inptr0], %[inptr0], #8\n" - "fmla v11.4s, v27.4s, v0.s[1]\n" - "add x25, x25, #8\n" - "fadd v23.4s, v21.4s, v26.4s\n" - "add x13, x13, #8\n" - "fsub v21.4s, v21.4s, v26.4s\n" - "ldr d22, [x14, %[in_col_stride1]]\n" - "fadd v13.4s, v16.4s, v9.4s\n" - "add x26, x26, #8\n" - "fmla v3.4s, v9.4s, v0.s[1]\n" - "fadd v30.4s, v17.4s, v19.4s\n" - "fadd v16.4s, v2.4s, v23.4s\n" - "fsub v19.4s, v17.4s, v19.4s\n" - "mov v17.16b, v23.16b\n" - "fadd v26.4s, v12.4s, v20.4s\n" - "fsub v9.4s, v12.4s, v20.4s\n" - "ldr d28, [x14, x21]\n" - "fadd v3.4s, v3.4s, v6.4s\n" - "ldr d20, [x14, x22]\n" - "fadd v16.4s, v16.4s, v30.4s\n" - "fmul v19.4s, v19.4s, v0.s[0]\n" - "fmla v17.4s, v30.4s, v0.s[1]\n" - "fadd v25.4s, v24.4s, v29.4s\n" - "fadd v23.4s, v15.4s, v26.4s\n" - "fsub v12.4s, v24.4s, v29.4s\n" - "mov v15.16b, v26.16b\n" - "fadd v24.4s, v22.4s, v28.4s\n" - "fsub v22.4s, v22.4s, v28.4s\n" - "fadd v29.4s, v14.4s, v16.4s\n" - "fsub v16.4s, v14.4s, v16.4s\n" - "ldr d28, [x14, x23]\n" - "fadd v23.4s, v23.4s, v25.4s\n" - "fmul v12.4s, v12.4s, v0.s[0]\n" - "fmla v15.4s, v25.4s, v0.s[1]\n" - "mov v6.16b, v21.16b\n" - "fadd v30.4s, v21.4s, v19.4s\n" - "fadd v26.4s, v18.4s, v24.4s\n" - "mov v25.16b, v24.16b\n" - "fadd v18.4s, v8.4s, v29.4s\n" - "fmla v6.4s, v19.4s, v0.s[1]\n" - "fadd v27.4s, v20.4s, v28.4s\n" - "fsub v21.4s, v20.4s, v28.4s\n" - "mov v19.16b, v29.16b\n" - "fadd v29.4s, v13.4s, v30.4s\n" - "fsub v8.4s, v13.4s, v30.4s\n" - "fadd v14.4s, v9.4s, v12.4s\n" - "fadd v6.4s, v6.4s, v10.4s\n" - "ldr d20, [x14, x24]\n" - "fadd v26.4s, v26.4s, v27.4s\n" - "add x14, x14, #8\n" - "fmla v9.4s, v12.4s, v0.s[1]\n" - "ldr d24, [x27]\n" - "fmul v21.4s, v21.4s, v0.s[0]\n" - "fmla v25.4s, v27.4s, v0.s[1]\n" - "fadd v10.4s, v7.4s, v29.4s\n" - "ldr d2, [%[bptr]]\n" - "mov v7.16b, v29.16b\n" - "add %[bptr], %[bptr], #8\n" - "fadd v9.4s, v9.4s, v5.4s\n" - "fadd v13.4s, v23.4s, v26.4s\n" - "fsub v23.4s, v23.4s, v26.4s\n" - "fadd v27.4s, v11.4s, v17.4s\n" - "fsub v11.4s, v11.4s, v17.4s\n" - "fadd v30.4s, v15.4s, v25.4s\n" - "fsub v15.4s, v15.4s, v25.4s\n" - "ldr d28, [x27, %[in_col_stride1]]\n" - "fadd v18.4s, v18.4s, v13.4s\n" - "fmla v19.4s, v13.4s, v0.s[1]\n" - "fadd v26.4s, v22.4s, v21.4s\n" - "mov v12.16b, v22.16b\n" - "fmul v23.4s, v23.4s, v0.s[0]\n" - "fadd v17.4s, v4.4s, v27.4s\n" - "fmul v15.4s, v15.4s, v0.s[0]\n" - "mov v4.16b, v27.16b\n" - "fmla v12.4s, v21.4s, v0.s[1]\n" - "ldr d22, [x27, x21]\n" - "fadd v18.4s, v18.4s, v2.4s\n" - "fadd v19.4s, v19.4s, v2.4s\n" - "fadd v17.4s, v17.4s, v30.4s\n" - "fmla v4.4s, v30.4s, v0.s[1]\n" - "fadd v25.4s, v28.4s, v22.4s\n" - "fsub v27.4s, v28.4s, v22.4s\n" - "fadd v12.4s, v12.4s, v20.4s\n" - "ldr d29, [x27, x22]\n" - "str d18, [%[outptr0]]\n" - "fadd v22.4s, v16.4s, v23.4s\n" - "str d19, [x28]\n" - "fadd v28.4s, v24.4s, v25.4s\n" - "ldr d30, [x27, x23]\n" - "fadd v20.4s, v29.4s, v30.4s\n" - "fsub v18.4s, v29.4s, v30.4s\n" - "mov v21.16b, v25.16b\n" - "ldr d25, [x27, x24]\n" - "fmla v16.4s, v23.4s, v0.s[1]\n" - "add x27, x27, #8\n" - "mov v24.16b, v27.16b\n" - "fadd v17.4s, v17.4s, v2.4s\n" - "fadd v28.4s, v28.4s, v20.4s\n" - "fmul v18.4s, v18.4s, v0.s[0]\n" - "fmla v21.4s, v20.4s, v0.s[1]\n" - "fadd v13.4s, v14.4s, v26.4s\n" - "fsub v30.4s, v14.4s, v26.4s\n" - "mov v14.16b, v8.16b\n" - "str d17, [%[outptr0], x15]\n" - "fadd v29.4s, v11.4s, v15.4s\n" - "fadd v23.4s, v27.4s, v18.4s\n" - "fmla v24.4s, v18.4s, v0.s[1]\n" - "fadd v16.4s, v16.4s, v28.4s\n" - "fadd v10.4s, v10.4s, v13.4s\n" - "fmul v30.4s, v30.4s, v0.s[0]\n" - "fmla v7.4s, v13.4s, v0.s[1]\n" - "mov v26.16b, v11.16b\n" - "fadd v13.4s, v3.4s, v6.4s\n" - "fadd v24.4s, v24.4s, v25.4s\n" - "fadd v27.4s, v9.4s, v12.4s\n" - "fsub v3.4s, v3.4s, v6.4s\n" - "fsub v6.4s, v9.4s, v12.4s\n" - "fadd v8.4s, v8.4s, v30.4s\n" - "fmla v14.4s, v30.4s, v0.s[1]\n" - "fmla v26.4s, v15.4s, v0.s[1]\n" - "fadd v1.4s, v1.4s, v13.4s\n" - "mov v5.16b, v13.16b\n" - "fadd v10.4s, v10.4s, v2.4s\n" - "fmul v6.4s, v6.4s, v0.s[0]\n" - "mov v13.16b, v3.16b\n" - "fadd v14.4s, v14.4s, v23.4s\n" - "fadd v22.4s, v22.4s, v2.4s\n" - "fadd v26.4s, v26.4s, v21.4s\n" - "fadd v1.4s, v1.4s, v27.4s\n" - "str d10, [%[outptr0], %[output_col_stride1]]\n" - "fmla v5.4s, v27.4s, v0.s[1]\n" - "fadd v10.4s, v3.4s, v6.4s\n" - "fmla v13.4s, v6.4s, v0.s[1]\n" - "str d22, [x17]\n" - "fadd v8.4s, v8.4s, v2.4s\n" - "fadd v1.4s, v1.4s, v2.4s\n" - "fadd v29.4s, v29.4s, v2.4s\n" - "fadd v7.4s, v7.4s, v2.4s\n" - "fadd v4.4s, v4.4s, v2.4s\n" - "fadd v13.4s, v13.4s, v24.4s\n" - "fadd v10.4s, v10.4s, v2.4s\n" - "str d8, [x17, %[output_col_stride1]]\n" - "fadd v5.4s, v5.4s, v2.4s\n" - "str d1, [%[outptr0], x16]\n" - "fadd v16.4s, v16.4s, v2.4s\n" - "str d29, [x17, x15]\n" - "fadd v14.4s, v14.4s, v2.4s\n" - "str d10, [x17, x16]\n" - "fadd v26.4s, v26.4s, v2.4s\n" - "str d7, [x28, %[output_col_stride1]]\n" - "fadd v13.4s, v13.4s, v2.4s\n" - "str d4, [x28, x15]\n" - "add %[outptr0], %[outptr0], #8\n" - "str d5, [x28, x16]\n" - "add x17, x17, #8\n" - "str d16, [x18]\n" - "add x28, x28, #8\n" - "str d14, [x18, %[output_col_stride1]]\n" - "str d26, [x18, x15]\n" - "str d13, [x18, x16]\n" - "add x18, x18, #8\n" - "5:\n" - "cbz x20, 6f\n" - "ldr s19, [%[inptr0]]\n" - "ldr s20, [%[inptr0], %[in_col_stride1]]\n" - "ldr s4, [%[inptr0], x21]\n" - "fadd v1.4s, v20.4s, v4.4s\n" - "ldr s17, [%[inptr0], x22]\n" - "fsub v7.4s, v20.4s, v4.4s\n" - "ldr s22, [%[inptr0], x23]\n" - "fadd v5.4s, v17.4s, v22.4s\n" - "ldr s18, [%[inptr0], x24]\n" - "fsub v10.4s, v17.4s, v22.4s\n" - "ldr s25, [x25]\n" - "fadd v8.4s, v19.4s, v1.4s\n" - "ldr s12, [x25, %[in_col_stride1]]\n" - "mov v4.16b, v1.16b\n" - "ldr s23, [x25, x21]\n" - "mov v1.16b, v7.16b\n" - "ldr s9, [x25, x22]\n" - "fmul v10.4s, v10.4s, v0.s[0]\n" - "ldr s11, [x25, x23]\n" - "fadd v8.4s, v8.4s, v5.4s\n" - "ldr s6, [x25, x24]\n" - "fmla v4.4s, v5.4s, v0.s[1]\n" - "ldr s2, [x13]\n" - "fadd v3.4s, v12.4s, v23.4s\n" - "ldr s21, [x13, %[in_col_stride1]]\n" - "fadd v7.4s, v7.4s, v10.4s\n" - "ldr s26, [x13, x21]\n" - "fmla v1.4s, v10.4s, v0.s[1]\n" - "ldr s17, [x13, x22]\n" - "fadd v27.4s, v9.4s, v11.4s\n" - "ldr s19, [x13, x23]\n" - "fadd v14.4s, v25.4s, v3.4s\n" - "ldr s10, [x13, x24]\n" - "fsub v16.4s, v12.4s, v23.4s\n" - "ldr s15, [x26]\n" - "fadd v1.4s, v1.4s, v18.4s\n" - "ldr s12, [x26, %[in_col_stride1]]\n" - "fsub v9.4s, v9.4s, v11.4s\n" - "ldr s20, [x26, x21]\n" - "fadd v14.4s, v14.4s, v27.4s\n" - "ldr s24, [x26, x22]\n" - "mov v11.16b, v3.16b\n" - "ldr s29, [x26, x23]\n" - "mov v3.16b, v16.16b\n" - "ldr s5, [x26, x24]\n" - "fmul v9.4s, v9.4s, v0.s[0]\n" - "ldr s18, [x14]\n" - "fmla v11.4s, v27.4s, v0.s[1]\n" - "fadd v23.4s, v21.4s, v26.4s\n" - "fsub v21.4s, v21.4s, v26.4s\n" - "fadd v30.4s, v17.4s, v19.4s\n" - "fsub v19.4s, v17.4s, v19.4s\n" - "ldr s22, [x14, %[in_col_stride1]]\n" - "fadd v13.4s, v16.4s, v9.4s\n" - "fmla v3.4s, v9.4s, v0.s[1]\n" - "fadd v16.4s, v2.4s, v23.4s\n" - "mov v17.16b, v23.16b\n" - "fadd v26.4s, v12.4s, v20.4s\n" - "fsub v9.4s, v12.4s, v20.4s\n" - "fmul v19.4s, v19.4s, v0.s[0]\n" - "ldr s28, [x14, x21]\n" - "fadd v3.4s, v3.4s, v6.4s\n" - "ldr s20, [x14, x22]\n" - "fadd v16.4s, v16.4s, v30.4s\n" - "fmla v17.4s, v30.4s, v0.s[1]\n" - "fadd v25.4s, v24.4s, v29.4s\n" - "fadd v23.4s, v15.4s, v26.4s\n" - "fsub v12.4s, v24.4s, v29.4s\n" - "mov v15.16b, v26.16b\n" - "fadd v24.4s, v22.4s, v28.4s\n" - "fsub v22.4s, v22.4s, v28.4s\n" - "fadd v30.4s, v21.4s, v19.4s\n" - "mov v6.16b, v21.16b\n" - "fadd v23.4s, v23.4s, v25.4s\n" - "fmla v15.4s, v25.4s, v0.s[1]\n" - "fmul v12.4s, v12.4s, v0.s[0]\n" - "ldr s28, [x14, x23]\n" - "fmla v6.4s, v19.4s, v0.s[1]\n" - "fadd v27.4s, v20.4s, v28.4s\n" - "fadd v26.4s, v18.4s, v24.4s\n" - "fsub v21.4s, v20.4s, v28.4s\n" - "mov v25.16b, v24.16b\n" - "fadd v29.4s, v14.4s, v16.4s\n" - "fsub v16.4s, v14.4s, v16.4s\n" - "ldr s20, [x14, x24]\n" - "fadd v6.4s, v6.4s, v10.4s\n" - "ldr s24, [x27]\n" - "fadd v26.4s, v26.4s, v27.4s\n" - "fmul v21.4s, v21.4s, v0.s[0]\n" - "fmla v25.4s, v27.4s, v0.s[1]\n" - "fadd v18.4s, v8.4s, v29.4s\n" - "mov v19.16b, v29.16b\n" - "fadd v29.4s, v13.4s, v30.4s\n" - "fsub v8.4s, v13.4s, v30.4s\n" - "fadd v27.4s, v11.4s, v17.4s\n" - "fsub v11.4s, v11.4s, v17.4s\n" - "fadd v13.4s, v23.4s, v26.4s\n" - "fsub v23.4s, v23.4s, v26.4s\n" - "ldr s28, [x27, %[in_col_stride1]]\n" - "fadd v10.4s, v7.4s, v29.4s\n" - "mov v7.16b, v29.16b\n" - "fadd v17.4s, v4.4s, v27.4s\n" - "mov v4.16b, v27.16b\n" - "fadd v18.4s, v18.4s, v13.4s\n" - "fmla v19.4s, v13.4s, v0.s[1]\n" - "fmul v23.4s, v23.4s, v0.s[0]\n" - "fadd v30.4s, v15.4s, v25.4s\n" - "fsub v15.4s, v15.4s, v25.4s\n" - "fadd v13.4s, v3.4s, v6.4s\n" - "fsub v3.4s, v3.4s, v6.4s\n" - "ldr s2, [%[bptr]]\n" - "fadd v18.4s, v18.4s, v2.4s\n" - "fadd v19.4s, v19.4s, v2.4s\n" - "fadd v17.4s, v17.4s, v30.4s\n" - "fmla v4.4s, v30.4s, v0.s[1]\n" - "fadd v14.4s, v9.4s, v12.4s\n" - "fmul v15.4s, v15.4s, v0.s[0]\n" - "fadd v1.4s, v1.4s, v13.4s\n" - "str s18, [%[outptr0]]\n" - "fadd v26.4s, v22.4s, v21.4s\n" - "str s19, [x28]\n" - "fmla v9.4s, v12.4s, v0.s[1]\n" - "mov v12.16b, v22.16b\n" - "ldr s22, [x27, x21]\n" - "fadd v25.4s, v28.4s, v22.4s\n" - "fsub v27.4s, v28.4s, v22.4s\n" - "fadd v22.4s, v16.4s, v23.4s\n" - "fadd v9.4s, v9.4s, v5.4s\n" - "ldr s29, [x27, x22]\n" - "fmla v12.4s, v21.4s, v0.s[1]\n" - "ldr s30, [x27, x23]\n" - "fadd v28.4s, v24.4s, v25.4s\n" - "mov v21.16b, v25.16b\n" - "fmla v16.4s, v23.4s, v0.s[1]\n" - "ldr s25, [x27, x24]\n" - "mov v5.16b, v13.16b\n" - "fadd v17.4s, v17.4s, v2.4s\n" - "fadd v12.4s, v12.4s, v20.4s\n" - "fadd v20.4s, v29.4s, v30.4s\n" - "fsub v18.4s, v29.4s, v30.4s\n" - "mov v24.16b, v27.16b\n" - "fadd v22.4s, v22.4s, v2.4s\n" - "fadd v4.4s, v4.4s, v2.4s\n" - "str s17, [%[outptr0], x15]\n" - "fadd v13.4s, v14.4s, v26.4s\n" - "fadd v28.4s, v28.4s, v20.4s\n" - "fmla v21.4s, v20.4s, v0.s[1]\n" - "fmul v18.4s, v18.4s, v0.s[0]\n" - "fsub v30.4s, v14.4s, v26.4s\n" - "str s22, [x17]\n" - "mov v14.16b, v8.16b\n" - "str s4, [x28, x15]\n" - "fadd v10.4s, v10.4s, v13.4s\n" - "fadd v16.4s, v16.4s, v28.4s\n" - "fmla v7.4s, v13.4s, v0.s[1]\n" - "fadd v23.4s, v27.4s, v18.4s\n" - "fmla v24.4s, v18.4s, v0.s[1]\n" - "fmul v30.4s, v30.4s, v0.s[0]\n" - "fadd v29.4s, v11.4s, v15.4s\n" - "mov v26.16b, v11.16b\n" - "fadd v27.4s, v9.4s, v12.4s\n" - "fsub v6.4s, v9.4s, v12.4s\n" - "mov v13.16b, v3.16b\n" - "fadd v24.4s, v24.4s, v25.4s\n" - "fadd v10.4s, v10.4s, v2.4s\n" - "fadd v8.4s, v8.4s, v30.4s\n" - "fmla v14.4s, v30.4s, v0.s[1]\n" - "fmla v26.4s, v15.4s, v0.s[1]\n" - "fadd v1.4s, v1.4s, v27.4s\n" - "fmul v6.4s, v6.4s, v0.s[0]\n" - "fmla v5.4s, v27.4s, v0.s[1]\n" - "str s10, [%[outptr0], %[output_col_stride1]]\n" - "fadd v29.4s, v29.4s, v2.4s\n" - "fadd v14.4s, v14.4s, v23.4s\n" - "fadd v8.4s, v8.4s, v2.4s\n" - "fadd v26.4s, v26.4s, v21.4s\n" - "fadd v1.4s, v1.4s, v2.4s\n" - "fadd v10.4s, v3.4s, v6.4s\n" - "fmla v13.4s, v6.4s, v0.s[1]\n" - "str s29, [x17, x15]\n" - "fadd v7.4s, v7.4s, v2.4s\n" - "str s8, [x17, %[output_col_stride1]]\n" - "fadd v5.4s, v5.4s, v2.4s\n" - "str s1, [%[outptr0], x16]\n" - "fadd v16.4s, v16.4s, v2.4s\n" - "fadd v13.4s, v13.4s, v24.4s\n" - "fadd v10.4s, v10.4s, v2.4s\n" - "str s7, [x28, %[output_col_stride1]]\n" - "fadd v14.4s, v14.4s, v2.4s\n" - "str s5, [x28, x16]\n" - "fadd v26.4s, v26.4s, v2.4s\n" - "str s16, [x18]\n" - "fadd v13.4s, v13.4s, v2.4s\n" - "str s10, [x17, x16]\n" - "str s14, [x18, %[output_col_stride1]]\n" - "str s26, [x18, x15]\n" - "str s13, [x18, x16]\n" - "6:\n" - : [bptr] "+r" (bptr), [outptr0] "+r" (output), [inptr0] "+r" (inptr) - : [output_row_stride] "r" (output_row_stride * sizeof(float)), [output_col_stride1] "r" (output_col_stride * sizeof(float)), [pcoeffs] "r" (coeffs), [n_channels] "r" ((long) n_channels), [in_row_stride] "r" (6 * matrix_stride * sizeof(float)), [in_col_stride1] "r" (matrix_stride * sizeof(float)) - : "cc", "v0", "v1", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v2", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v3", "v30", "v4", "v5", "v6", "v7", "v8", "v9", "x13", "x14", "x15", "x16", "x17", "x18", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "memory" - ); - } - else - { - __asm__ __volatile__ ( - "ldr d0, [%[pcoeffs]]\n" - "add x21, %[in_col_stride1], %[in_col_stride1]\n" // Compute input column stride 2 - "add x22, x21, %[in_col_stride1]\n" // Compute input column stride 3 - "add x25, %[inptr0], %[in_row_stride]\n" // Compute input row pointers - "add x15, %[output_col_stride1], %[output_col_stride1]\n" // Compute output column stride 2 - "add x23, x22, %[in_col_stride1]\n" // Compute input column stride 4 - "add x13, x25, %[in_row_stride]\n" // Compute input row pointers - "add x16, x15, %[output_col_stride1]\n" // Compute output column stride 3 - "add x24, x23, %[in_col_stride1]\n" // Compute input column stride 5 - "add x26, x13, %[in_row_stride]\n" // Compute input row pointers - "add x17, %[outptr0], %[output_row_stride]\n" // Compute output row pointer 1 - "add x14, x26, %[in_row_stride]\n" // Compute input row pointers - "add x28, x17, %[output_row_stride]\n" // Compute output row pointer 2 - "lsr x19, %[n_channels], #2\n" - "add x27, x14, %[in_row_stride]\n" // Compute input row pointers - "add x18, x28, %[output_row_stride]\n" // Compute output row pointer 3 - "and x20, %[n_channels], #3\n" - "cbz x19, 4f\n" - "1:\n" // Quad head - "ldr q17, [%[inptr0]]\n" - "subs x19, x19, #1\n" - "ldr q23, [%[inptr0], %[in_col_stride1]]\n" - "ldr q27, [%[inptr0], x21]\n" - "fadd v4.4s, v23.4s, v27.4s\n" - "ldr q24, [%[inptr0], x22]\n" - "fsub v13.4s, v23.4s, v27.4s\n" - "ldr q11, [%[inptr0], x23]\n" - "fadd v10.4s, v24.4s, v11.4s\n" - "ldr q12, [%[inptr0], x24]\n" - "fsub v11.4s, v24.4s, v11.4s\n" - "ldr q20, [x25]\n" - "fadd v7.4s, v17.4s, v4.4s\n" - "ldr q19, [x25, %[in_col_stride1]]\n" - "mov v4.16b, v4.16b\n" - "ldr q22, [x25, x21]\n" - "mov v1.16b, v13.16b\n" - "ldr q14, [x25, x22]\n" - "fmul v11.4s, v11.4s, v0.s[0]\n" - "ldr q18, [x25, x23]\n" - "fadd v7.4s, v7.4s, v10.4s\n" - "ldr q3, [x25, x24]\n" - "fmla v4.4s, v10.4s, v0.s[1]\n" - "fadd v8.4s, v13.4s, v11.4s\n" - "fmla v1.4s, v11.4s, v0.s[1]\n" - "fadd v1.4s, v1.4s, v12.4s\n" - "beq 3f\n" - "2:\n" // Quad loop - "fadd v2.4s, v19.4s, v22.4s\n" - "ldr q16, [x13]\n" - "fadd v23.4s, v14.4s, v18.4s\n" - "ldr q21, [x13, %[in_col_stride1]]\n" - "fsub v15.4s, v19.4s, v22.4s\n" - "ldr q24, [x13, x21]\n" - "fsub v31.4s, v14.4s, v18.4s\n" - "ldr q25, [x13, x22]\n" - "fadd v11.4s, v20.4s, v2.4s\n" - "ldr q17, [x13, x23]\n" - "mov v13.16b, v2.16b\n" - "ldr q9, [x13, x24]\n" - "mov v2.16b, v15.16b\n" - "ldr q6, [x26]\n" - "fmul v31.4s, v31.4s, v0.s[0]\n" - "ldr q19, [x26, %[in_col_stride1]]\n" - "fadd v11.4s, v11.4s, v23.4s\n" - "ldr q22, [x26, x21]\n" - "fmla v13.4s, v23.4s, v0.s[1]\n" - "ldr q12, [x26, x22]\n" - "fadd v29.4s, v21.4s, v24.4s\n" - "ldr q26, [x26, x23]\n" - "fadd v15.4s, v15.4s, v31.4s\n" - "ldr q5, [x26, x24]\n" - "fmla v2.4s, v31.4s, v0.s[1]\n" - "ldr q10, [x14]\n" - "fadd v18.4s, v25.4s, v17.4s\n" - "add %[inptr0], %[inptr0], #16\n" - "fadd v27.4s, v16.4s, v29.4s\n" - "add x25, x25, #16\n" - "fsub v14.4s, v21.4s, v24.4s\n" - "ldr q30, [x14, %[in_col_stride1]]\n" - "fadd v2.4s, v2.4s, v3.4s\n" - "ldr q31, [x14, x21]\n" - "fsub v28.4s, v25.4s, v17.4s\n" - "add x13, x13, #16\n" - "fadd v27.4s, v27.4s, v18.4s\n" - "add x26, x26, #16\n" - "mov v21.16b, v29.16b\n" - "subs x19, x19, #1\n" - "fadd v20.4s, v19.4s, v22.4s\n" - "fsub v17.4s, v19.4s, v22.4s\n" - "fmul v28.4s, v28.4s, v0.s[0]\n" - "ldr q23, [x14, x22]\n" - "fmla v21.4s, v18.4s, v0.s[1]\n" - "fadd v29.4s, v12.4s, v26.4s\n" - "fsub v16.4s, v12.4s, v26.4s\n" - "fadd v25.4s, v30.4s, v31.4s\n" - "fadd v24.4s, v6.4s, v20.4s\n" - "mov v6.16b, v20.16b\n" - "fsub v22.4s, v30.4s, v31.4s\n" - "fadd v31.4s, v11.4s, v27.4s\n" - "fsub v12.4s, v11.4s, v27.4s\n" - "ldr q26, [x14, x23]\n" - "fmul v16.4s, v16.4s, v0.s[0]\n" - "fmla v6.4s, v29.4s, v0.s[1]\n" - "fadd v24.4s, v24.4s, v29.4s\n" - "mov v3.16b, v14.16b\n" - "fadd v20.4s, v14.4s, v28.4s\n" - "fadd v29.4s, v10.4s, v25.4s\n" - "mov v10.16b, v25.16b\n" - "fadd v25.4s, v7.4s, v31.4s\n" - "fmla v3.4s, v28.4s, v0.s[1]\n" - "fadd v14.4s, v23.4s, v26.4s\n" - "fsub v23.4s, v23.4s, v26.4s\n" - "mov v26.16b, v31.16b\n" - "fadd v31.4s, v15.4s, v20.4s\n" - "fsub v11.4s, v15.4s, v20.4s\n" - "fadd v20.4s, v17.4s, v16.4s\n" - "mov v7.16b, v17.16b\n" - "fadd v3.4s, v3.4s, v9.4s\n" - "ldr q18, [x14, x24]\n" - "fadd v29.4s, v29.4s, v14.4s\n" - "add x14, x14, #16\n" - "fmla v7.4s, v16.4s, v0.s[1]\n" - "ldr q19, [x27]\n" - "fmul v23.4s, v23.4s, v0.s[0]\n" - "fmla v10.4s, v14.4s, v0.s[1]\n" - "fadd v15.4s, v8.4s, v31.4s\n" - "mov v14.16b, v31.16b\n" - "fadd v28.4s, v24.4s, v29.4s\n" - "fsub v24.4s, v24.4s, v29.4s\n" - "fadd v7.4s, v7.4s, v5.4s\n" - "ldr q27, [x27, %[in_col_stride1]]\n" - "fadd v30.4s, v13.4s, v21.4s\n" - "fsub v9.4s, v13.4s, v21.4s\n" - "fadd v17.4s, v22.4s, v23.4s\n" - "mov v8.16b, v22.16b\n" - "fadd v25.4s, v25.4s, v28.4s\n" - "fmul v24.4s, v24.4s, v0.s[0]\n" - "fmla v26.4s, v28.4s, v0.s[1]\n" - "ldr q29, [x27, x21]\n" - "fmla v8.4s, v23.4s, v0.s[1]\n" - "ldr q28, [x27, x22]\n" - "fadd v13.4s, v4.4s, v30.4s\n" - "mov v4.16b, v30.16b\n" - "str q25, [%[outptr0]]\n" // Store output (0, 0) - "fadd v16.4s, v27.4s, v29.4s\n" - "str q26, [x28]\n" // Store output (2, 0) - "fsub v29.4s, v27.4s, v29.4s\n" - "fadd v8.4s, v8.4s, v18.4s\n" - "ldr q23, [x27, x23]\n" - "fadd v30.4s, v28.4s, v23.4s\n" - "ldr q25, [x27, x24]\n" - "fadd v19.4s, v19.4s, v16.4s\n" - "add x27, x27, #16\n" - "fsub v27.4s, v28.4s, v23.4s\n" - "mov v16.16b, v16.16b\n" - "fadd v22.4s, v20.4s, v17.4s\n" - "fsub v20.4s, v20.4s, v17.4s\n" - "fadd v21.4s, v12.4s, v24.4s\n" - "mov v26.16b, v12.16b\n" - "fadd v19.4s, v19.4s, v30.4s\n" - "fmla v16.4s, v30.4s, v0.s[1]\n" - "fmul v27.4s, v27.4s, v0.s[0]\n" - "ldr q17, [%[inptr0]]\n" - "fmla v26.4s, v24.4s, v0.s[1]\n" - "ldr q23, [%[inptr0], %[in_col_stride1]]\n" - "str q21, [x17]\n" // Store output (1, 0) - "mov v5.16b, v29.16b\n" - "fadd v15.4s, v15.4s, v22.4s\n" - "fmul v20.4s, v20.4s, v0.s[0]\n" - "fadd v18.4s, v29.4s, v27.4s\n" - "fmla v14.4s, v22.4s, v0.s[1]\n" - "fmla v5.4s, v27.4s, v0.s[1]\n" - "ldr q27, [%[inptr0], x21]\n" - "fadd v26.4s, v26.4s, v19.4s\n" - "ldr q24, [%[inptr0], x22]\n" - "str q15, [%[outptr0], %[output_col_stride1]]\n" // Store output (0, 1) - "fadd v12.4s, v11.4s, v20.4s\n" - "str q14, [x28, %[output_col_stride1]]\n" // Store output (2, 1) - "mov v28.16b, v11.16b\n" - "fadd v5.4s, v5.4s, v25.4s\n" - "ldr q11, [%[inptr0], x23]\n" - "str q26, [x18]\n" // Store output (3, 0) - "fadd v21.4s, v6.4s, v10.4s\n" - "str q12, [x17, %[output_col_stride1]]\n" // Store output (1, 1) - "fmla v28.4s, v20.4s, v0.s[1]\n" - "fsub v10.4s, v6.4s, v10.4s\n" - "ldr q12, [%[inptr0], x24]\n" - "mov v15.16b, v9.16b\n" - "ldr q20, [x25]\n" - "fadd v13.4s, v13.4s, v21.4s\n" - "ldr q19, [x25, %[in_col_stride1]]\n" - "fadd v28.4s, v28.4s, v18.4s\n" - "ldr q22, [x25, x21]\n" - "fmul v10.4s, v10.4s, v0.s[0]\n" - "ldr q14, [x25, x22]\n" - "fmla v4.4s, v21.4s, v0.s[1]\n" - "ldr q18, [x25, x23]\n" - "str q13, [%[outptr0], x15]\n" // Store output (0, 2) - "fadd v6.4s, v2.4s, v3.4s\n" - "str q28, [x18, %[output_col_stride1]]\n" // Store output (3, 1) - "fadd v30.4s, v7.4s, v8.4s\n" - "fadd v13.4s, v9.4s, v10.4s\n" - "fmla v15.4s, v10.4s, v0.s[1]\n" - "str q4, [x28, x15]\n" // Store output (2, 2) - "fsub v2.4s, v2.4s, v3.4s\n" - "fadd v1.4s, v1.4s, v6.4s\n" - "ldr q3, [x25, x24]\n" - "fsub v8.4s, v7.4s, v8.4s\n" - "mov v6.16b, v6.16b\n" - "str q13, [x17, x15]\n" // Store output (1, 2) - "fadd v15.4s, v15.4s, v16.4s\n" - "mov v9.16b, v2.16b\n" - "fadd v4.4s, v23.4s, v27.4s\n" - "fadd v1.4s, v1.4s, v30.4s\n" - "fmla v6.4s, v30.4s, v0.s[1]\n" - "fmul v8.4s, v8.4s, v0.s[0]\n" - "fadd v10.4s, v24.4s, v11.4s\n" - "str q15, [x18, x15]\n" // Store output (3, 2) - "fsub v13.4s, v23.4s, v27.4s\n" - "fadd v7.4s, v17.4s, v4.4s\n" - "fsub v11.4s, v24.4s, v11.4s\n" - "str q1, [%[outptr0], x16]\n" // Store output (0, 3) - "mov v4.16b, v4.16b\n" - "str q6, [x28, x16]\n" // Store output (2, 3) - "fadd v2.4s, v2.4s, v8.4s\n" - "fmla v9.4s, v8.4s, v0.s[1]\n" - "add %[outptr0], %[outptr0], #16\n" - "fadd v7.4s, v7.4s, v10.4s\n" - "add x28, x28, #16\n" - "fmul v11.4s, v11.4s, v0.s[0]\n" - "fmla v4.4s, v10.4s, v0.s[1]\n" - "str q2, [x17, x16]\n" // Store output (1, 3) - "mov v1.16b, v13.16b\n" - "fadd v9.4s, v9.4s, v5.4s\n" - "add x17, x17, #16\n" - "fadd v8.4s, v13.4s, v11.4s\n" - "fmla v1.4s, v11.4s, v0.s[1]\n" - "str q9, [x18, x16]\n" // Store output (3, 3) - "add x18, x18, #16\n" - "fadd v1.4s, v1.4s, v12.4s\n" - "bne 2b\n" - "3:\n" // Quad tail - "fadd v2.4s, v19.4s, v22.4s\n" - "ldr q16, [x13]\n" - "fadd v23.4s, v14.4s, v18.4s\n" - "ldr q21, [x13, %[in_col_stride1]]\n" - "fsub v15.4s, v19.4s, v22.4s\n" - "ldr q24, [x13, x21]\n" - "fsub v31.4s, v14.4s, v18.4s\n" - "ldr q25, [x13, x22]\n" - "fadd v11.4s, v20.4s, v2.4s\n" - "ldr q17, [x13, x23]\n" - "mov v13.16b, v2.16b\n" - "ldr q9, [x13, x24]\n" - "mov v2.16b, v15.16b\n" - "ldr q6, [x26]\n" - "fmul v31.4s, v31.4s, v0.s[0]\n" - "ldr q19, [x26, %[in_col_stride1]]\n" - "fadd v11.4s, v11.4s, v23.4s\n" - "ldr q22, [x26, x21]\n" - "fmla v13.4s, v23.4s, v0.s[1]\n" - "ldr q12, [x26, x22]\n" - "fadd v29.4s, v21.4s, v24.4s\n" - "ldr q26, [x26, x23]\n" - "fadd v15.4s, v15.4s, v31.4s\n" - "ldr q5, [x26, x24]\n" - "fmla v2.4s, v31.4s, v0.s[1]\n" - "ldr q10, [x14]\n" - "fadd v18.4s, v25.4s, v17.4s\n" - "add %[inptr0], %[inptr0], #16\n" - "fadd v27.4s, v16.4s, v29.4s\n" - "add x25, x25, #16\n" - "fsub v14.4s, v21.4s, v24.4s\n" - "ldr q30, [x14, %[in_col_stride1]]\n" - "fadd v2.4s, v2.4s, v3.4s\n" - "ldr q31, [x14, x21]\n" - "fsub v28.4s, v25.4s, v17.4s\n" - "add x13, x13, #16\n" - "fadd v27.4s, v27.4s, v18.4s\n" - "add x26, x26, #16\n" - "mov v21.16b, v29.16b\n" - "fadd v20.4s, v19.4s, v22.4s\n" - "fsub v17.4s, v19.4s, v22.4s\n" - "fadd v29.4s, v12.4s, v26.4s\n" - "fmul v28.4s, v28.4s, v0.s[0]\n" - "fsub v16.4s, v12.4s, v26.4s\n" - "fmla v21.4s, v18.4s, v0.s[1]\n" - "ldr q23, [x14, x22]\n" - "fadd v24.4s, v6.4s, v20.4s\n" - "mov v6.16b, v20.16b\n" - "fadd v25.4s, v30.4s, v31.4s\n" - "fsub v22.4s, v30.4s, v31.4s\n" - "fadd v20.4s, v14.4s, v28.4s\n" - "mov v3.16b, v14.16b\n" - "fmul v16.4s, v16.4s, v0.s[0]\n" - "fmla v6.4s, v29.4s, v0.s[1]\n" - "fadd v24.4s, v24.4s, v29.4s\n" - "ldr q26, [x14, x23]\n" - "fmla v3.4s, v28.4s, v0.s[1]\n" - "fadd v14.4s, v23.4s, v26.4s\n" - "fadd v29.4s, v10.4s, v25.4s\n" - "fsub v23.4s, v23.4s, v26.4s\n" - "mov v10.16b, v25.16b\n" - "fadd v31.4s, v11.4s, v27.4s\n" - "fsub v12.4s, v11.4s, v27.4s\n" - "ldr q18, [x14, x24]\n" - "fadd v3.4s, v3.4s, v9.4s\n" - "ldr q19, [x27]\n" - "fadd v29.4s, v29.4s, v14.4s\n" - "add x14, x14, #16\n" - "fmul v23.4s, v23.4s, v0.s[0]\n" - "fmla v10.4s, v14.4s, v0.s[1]\n" - "fadd v25.4s, v7.4s, v31.4s\n" - "mov v26.16b, v31.16b\n" - "fadd v31.4s, v15.4s, v20.4s\n" - "fsub v11.4s, v15.4s, v20.4s\n" - "fadd v28.4s, v24.4s, v29.4s\n" - "fsub v24.4s, v24.4s, v29.4s\n" - "fadd v30.4s, v13.4s, v21.4s\n" - "fsub v9.4s, v13.4s, v21.4s\n" - "fadd v20.4s, v17.4s, v16.4s\n" - "mov v7.16b, v17.16b\n" - "fadd v15.4s, v8.4s, v31.4s\n" - "mov v14.16b, v31.16b\n" - "fadd v25.4s, v25.4s, v28.4s\n" - "fmul v24.4s, v24.4s, v0.s[0]\n" - "fmla v7.4s, v16.4s, v0.s[1]\n" - "ldr q27, [x27, %[in_col_stride1]]\n" - "fmla v26.4s, v28.4s, v0.s[1]\n" - "ldr q29, [x27, x21]\n" - "fadd v13.4s, v4.4s, v30.4s\n" - "mov v4.16b, v30.16b\n" - "str q25, [%[outptr0]]\n" // Store output (0, 0) - "fadd v17.4s, v22.4s, v23.4s\n" - "fadd v7.4s, v7.4s, v5.4s\n" - "ldr q28, [x27, x22]\n" - "str q26, [x28]\n" // Store output (2, 0) - "mov v8.16b, v22.16b\n" - "fadd v16.4s, v27.4s, v29.4s\n" - "fsub v29.4s, v27.4s, v29.4s\n" - "fadd v21.4s, v12.4s, v24.4s\n" - "mov v26.16b, v12.16b\n" - "fmla v8.4s, v23.4s, v0.s[1]\n" - "fadd v22.4s, v20.4s, v17.4s\n" - "fsub v20.4s, v20.4s, v17.4s\n" - "ldr q23, [x27, x23]\n" - "fadd v19.4s, v19.4s, v16.4s\n" - "mov v16.16b, v16.16b\n" - "str q21, [x17]\n" // Store output (1, 0) - "fadd v30.4s, v28.4s, v23.4s\n" - "fadd v8.4s, v8.4s, v18.4s\n" - "ldr q25, [x27, x24]\n" - "fsub v27.4s, v28.4s, v23.4s\n" - "add x27, x27, #16\n" - "mov v5.16b, v29.16b\n" - "fmla v26.4s, v24.4s, v0.s[1]\n" - "fadd v19.4s, v19.4s, v30.4s\n" - "fmla v16.4s, v30.4s, v0.s[1]\n" - "fadd v15.4s, v15.4s, v22.4s\n" - "fmul v20.4s, v20.4s, v0.s[0]\n" - "fmul v27.4s, v27.4s, v0.s[0]\n" - "fmla v14.4s, v22.4s, v0.s[1]\n" - "mov v28.16b, v11.16b\n" - "fadd v21.4s, v6.4s, v10.4s\n" - "fadd v26.4s, v26.4s, v19.4s\n" - "fsub v10.4s, v6.4s, v10.4s\n" - "str q15, [%[outptr0], %[output_col_stride1]]\n" // Store output (0, 1) - "fadd v12.4s, v11.4s, v20.4s\n" - "str q14, [x28, %[output_col_stride1]]\n" // Store output (2, 1) - "fadd v18.4s, v29.4s, v27.4s\n" - "fmla v5.4s, v27.4s, v0.s[1]\n" - "fmla v28.4s, v20.4s, v0.s[1]\n" - "str q26, [x18]\n" // Store output (3, 0) - "fadd v13.4s, v13.4s, v21.4s\n" - "str q12, [x17, %[output_col_stride1]]\n" // Store output (1, 1) - "fmul v10.4s, v10.4s, v0.s[0]\n" - "fmla v4.4s, v21.4s, v0.s[1]\n" - "mov v15.16b, v9.16b\n" - "fadd v5.4s, v5.4s, v25.4s\n" - "fadd v28.4s, v28.4s, v18.4s\n" - "str q13, [%[outptr0], x15]\n" // Store output (0, 2) - "fadd v6.4s, v2.4s, v3.4s\n" - "fadd v13.4s, v9.4s, v10.4s\n" - "fmla v15.4s, v10.4s, v0.s[1]\n" - "str q4, [x28, x15]\n" // Store output (2, 2) - "fadd v30.4s, v7.4s, v8.4s\n" - "str q28, [x18, %[output_col_stride1]]\n" // Store output (3, 1) - "fsub v2.4s, v2.4s, v3.4s\n" - "fadd v1.4s, v1.4s, v6.4s\n" - "fsub v8.4s, v7.4s, v8.4s\n" - "str q13, [x17, x15]\n" // Store output (1, 2) - "fadd v15.4s, v15.4s, v16.4s\n" - "mov v6.16b, v6.16b\n" - "mov v9.16b, v2.16b\n" - "fadd v1.4s, v1.4s, v30.4s\n" - "fmul v8.4s, v8.4s, v0.s[0]\n" - "str q15, [x18, x15]\n" // Store output (3, 2) - "fmla v6.4s, v30.4s, v0.s[1]\n" - "str q1, [%[outptr0], x16]\n" // Store output (0, 3) - "fadd v2.4s, v2.4s, v8.4s\n" - "str q6, [x28, x16]\n" // Store output (2, 3) - "fmla v9.4s, v8.4s, v0.s[1]\n" - "add %[outptr0], %[outptr0], #16\n" - "add x28, x28, #16\n" - "str q2, [x17, x16]\n" // Store output (1, 3) - "fadd v9.4s, v9.4s, v5.4s\n" - "add x17, x17, #16\n" - "str q9, [x18, x16]\n" // Store output (3, 3) - "add x18, x18, #16\n" - "4:\n" // Double - "cmp x20, #2\n" - "blt 5f\n" - "ldr d17, [%[inptr0]]\n" - "ldr d23, [%[inptr0], %[in_col_stride1]]\n" - "sub x20, x20, #2\n" - "ldr d27, [%[inptr0], x21]\n" - "ldr d24, [%[inptr0], x22]\n" - "fadd v4.4s, v23.4s, v27.4s\n" - "ldr d11, [%[inptr0], x23]\n" - "fadd v10.4s, v24.4s, v11.4s\n" - "ldr d12, [%[inptr0], x24]\n" - "fsub v13.4s, v23.4s, v27.4s\n" - "ldr d20, [x25]\n" - "fsub v11.4s, v24.4s, v11.4s\n" - "ldr d19, [x25, %[in_col_stride1]]\n" - "fadd v7.4s, v17.4s, v4.4s\n" - "ldr d22, [x25, x21]\n" - "mov v4.16b, v4.16b\n" - "ldr d14, [x25, x22]\n" - "mov v1.16b, v13.16b\n" - "ldr d18, [x25, x23]\n" - "fmul v11.4s, v11.4s, v0.s[0]\n" - "ldr d3, [x25, x24]\n" - "fadd v7.4s, v7.4s, v10.4s\n" - "ldr d16, [x13]\n" - "fmla v4.4s, v10.4s, v0.s[1]\n" - "ldr d21, [x13, %[in_col_stride1]]\n" - "fadd v2.4s, v19.4s, v22.4s\n" - "ldr d24, [x13, x21]\n" - "fadd v8.4s, v13.4s, v11.4s\n" - "ldr d25, [x13, x22]\n" - "fmla v1.4s, v11.4s, v0.s[1]\n" - "ldr d17, [x13, x23]\n" - "fadd v23.4s, v14.4s, v18.4s\n" - "ldr d9, [x13, x24]\n" - "fadd v11.4s, v20.4s, v2.4s\n" - "ldr d6, [x26]\n" - "fsub v15.4s, v19.4s, v22.4s\n" - "ldr d19, [x26, %[in_col_stride1]]\n" - "fadd v1.4s, v1.4s, v12.4s\n" - "ldr d22, [x26, x21]\n" - "fsub v31.4s, v14.4s, v18.4s\n" - "ldr d12, [x26, x22]\n" - "fadd v11.4s, v11.4s, v23.4s\n" - "ldr d26, [x26, x23]\n" - "mov v13.16b, v2.16b\n" - "ldr d5, [x26, x24]\n" - "mov v2.16b, v15.16b\n" - "ldr d10, [x14]\n" - "fmul v31.4s, v31.4s, v0.s[0]\n" - "add %[inptr0], %[inptr0], #8\n" - "fmla v13.4s, v23.4s, v0.s[1]\n" - "add x25, x25, #8\n" - "fadd v29.4s, v21.4s, v24.4s\n" - "add x13, x13, #8\n" - "fsub v14.4s, v21.4s, v24.4s\n" - "ldr d30, [x14, %[in_col_stride1]]\n" - "fadd v15.4s, v15.4s, v31.4s\n" - "add x26, x26, #8\n" - "fmla v2.4s, v31.4s, v0.s[1]\n" - "fadd v18.4s, v25.4s, v17.4s\n" - "fadd v27.4s, v16.4s, v29.4s\n" - "fsub v28.4s, v25.4s, v17.4s\n" - "mov v21.16b, v29.16b\n" - "fadd v20.4s, v19.4s, v22.4s\n" - "fsub v17.4s, v19.4s, v22.4s\n" - "ldr d31, [x14, x21]\n" - "fadd v2.4s, v2.4s, v3.4s\n" - "ldr d23, [x14, x22]\n" - "fadd v27.4s, v27.4s, v18.4s\n" - "fmul v28.4s, v28.4s, v0.s[0]\n" - "fmla v21.4s, v18.4s, v0.s[1]\n" - "fadd v29.4s, v12.4s, v26.4s\n" - "fadd v24.4s, v6.4s, v20.4s\n" - "fsub v16.4s, v12.4s, v26.4s\n" - "mov v6.16b, v20.16b\n" - "fadd v25.4s, v30.4s, v31.4s\n" - "fsub v22.4s, v30.4s, v31.4s\n" - "fadd v31.4s, v11.4s, v27.4s\n" - "fsub v12.4s, v11.4s, v27.4s\n" - "ldr d26, [x14, x23]\n" - "fadd v24.4s, v24.4s, v29.4s\n" - "fmul v16.4s, v16.4s, v0.s[0]\n" - "fmla v6.4s, v29.4s, v0.s[1]\n" - "mov v3.16b, v14.16b\n" - "fadd v20.4s, v14.4s, v28.4s\n" - "fadd v29.4s, v10.4s, v25.4s\n" - "mov v10.16b, v25.16b\n" - "fadd v25.4s, v7.4s, v31.4s\n" - "fmla v3.4s, v28.4s, v0.s[1]\n" - "fadd v14.4s, v23.4s, v26.4s\n" - "fsub v23.4s, v23.4s, v26.4s\n" - "mov v26.16b, v31.16b\n" - "fadd v31.4s, v15.4s, v20.4s\n" - "fsub v11.4s, v15.4s, v20.4s\n" - "fadd v20.4s, v17.4s, v16.4s\n" - "mov v7.16b, v17.16b\n" - "fadd v3.4s, v3.4s, v9.4s\n" - "ldr d18, [x14, x24]\n" - "fadd v29.4s, v29.4s, v14.4s\n" - "add x14, x14, #8\n" - "fmla v7.4s, v16.4s, v0.s[1]\n" - "ldr d19, [x27]\n" - "fmul v23.4s, v23.4s, v0.s[0]\n" - "fmla v10.4s, v14.4s, v0.s[1]\n" - "fadd v15.4s, v8.4s, v31.4s\n" - "mov v14.16b, v31.16b\n" - "fadd v28.4s, v24.4s, v29.4s\n" - "fsub v24.4s, v24.4s, v29.4s\n" - "fadd v7.4s, v7.4s, v5.4s\n" - "ldr d27, [x27, %[in_col_stride1]]\n" - "fadd v30.4s, v13.4s, v21.4s\n" - "fsub v9.4s, v13.4s, v21.4s\n" - "fadd v17.4s, v22.4s, v23.4s\n" - "mov v8.16b, v22.16b\n" - "fadd v25.4s, v25.4s, v28.4s\n" - "fmul v24.4s, v24.4s, v0.s[0]\n" - "fmla v26.4s, v28.4s, v0.s[1]\n" - "ldr d29, [x27, x21]\n" - "fmla v8.4s, v23.4s, v0.s[1]\n" - "ldr d28, [x27, x22]\n" - "fadd v13.4s, v4.4s, v30.4s\n" - "mov v4.16b, v30.16b\n" - "str d25, [%[outptr0]]\n" // Store output (0, 0) - "fadd v16.4s, v27.4s, v29.4s\n" - "str d26, [x28]\n" // Store output (2, 0) - "fsub v29.4s, v27.4s, v29.4s\n" - "fadd v8.4s, v8.4s, v18.4s\n" - "ldr d23, [x27, x23]\n" - "fadd v30.4s, v28.4s, v23.4s\n" - "ldr d25, [x27, x24]\n" - "fadd v19.4s, v19.4s, v16.4s\n" - "add x27, x27, #8\n" - "fsub v27.4s, v28.4s, v23.4s\n" - "mov v16.16b, v16.16b\n" - "fadd v22.4s, v20.4s, v17.4s\n" - "fsub v20.4s, v20.4s, v17.4s\n" - "fadd v21.4s, v12.4s, v24.4s\n" - "mov v26.16b, v12.16b\n" - "fadd v19.4s, v19.4s, v30.4s\n" - "fmla v16.4s, v30.4s, v0.s[1]\n" - "fmul v27.4s, v27.4s, v0.s[0]\n" - "mov v5.16b, v29.16b\n" - "fmla v26.4s, v24.4s, v0.s[1]\n" - "fadd v15.4s, v15.4s, v22.4s\n" - "str d21, [x17]\n" // Store output (1, 0) - "fmul v20.4s, v20.4s, v0.s[0]\n" - "fmla v14.4s, v22.4s, v0.s[1]\n" - "mov v28.16b, v11.16b\n" - "fadd v18.4s, v29.4s, v27.4s\n" - "fmla v5.4s, v27.4s, v0.s[1]\n" - "str d15, [%[outptr0], %[output_col_stride1]]\n" // Store output (0, 1) - "fadd v26.4s, v26.4s, v19.4s\n" - "fadd v12.4s, v11.4s, v20.4s\n" - "fmla v28.4s, v20.4s, v0.s[1]\n" - "str d14, [x28, %[output_col_stride1]]\n" // Store output (2, 1) - "fadd v21.4s, v6.4s, v10.4s\n" - "fadd v5.4s, v5.4s, v25.4s\n" - "fsub v10.4s, v6.4s, v10.4s\n" - "str d26, [x18]\n" // Store output (3, 0) - "mov v15.16b, v9.16b\n" - "str d12, [x17, %[output_col_stride1]]\n" // Store output (1, 1) - "fadd v28.4s, v28.4s, v18.4s\n" - "fadd v13.4s, v13.4s, v21.4s\n" - "fmla v4.4s, v21.4s, v0.s[1]\n" - "fmul v10.4s, v10.4s, v0.s[0]\n" - "fadd v6.4s, v2.4s, v3.4s\n" - "fadd v30.4s, v7.4s, v8.4s\n" - "fsub v2.4s, v2.4s, v3.4s\n" - "str d28, [x18, %[output_col_stride1]]\n" // Store output (3, 1) - "fsub v8.4s, v7.4s, v8.4s\n" - "str d13, [%[outptr0], x15]\n" // Store output (0, 2) - "str d4, [x28, x15]\n" // Store output (2, 2) - "fadd v13.4s, v9.4s, v10.4s\n" - "fmla v15.4s, v10.4s, v0.s[1]\n" - "fadd v1.4s, v1.4s, v6.4s\n" - "mov v6.16b, v6.16b\n" - "fmul v8.4s, v8.4s, v0.s[0]\n" - "mov v9.16b, v2.16b\n" - "str d13, [x17, x15]\n" // Store output (1, 2) - "fadd v15.4s, v15.4s, v16.4s\n" - "fadd v1.4s, v1.4s, v30.4s\n" - "fmla v6.4s, v30.4s, v0.s[1]\n" - "fadd v2.4s, v2.4s, v8.4s\n" - "fmla v9.4s, v8.4s, v0.s[1]\n" - "str d15, [x18, x15]\n" // Store output (3, 2) - "str d1, [%[outptr0], x16]\n" // Store output (0, 3) - "str d2, [x17, x16]\n" // Store output (1, 3) - "fadd v9.4s, v9.4s, v5.4s\n" - "str d6, [x28, x16]\n" // Store output (2, 3) - "add %[outptr0], %[outptr0], #8\n" - "add x17, x17, #8\n" - "add x28, x28, #8\n" - "str d9, [x18, x16]\n" // Store output (3, 3) - "add x18, x18, #8\n" - "5:\n" // Scalar - "cbz x20, 6f\n" - "ldr s17, [%[inptr0]]\n" - "ldr s23, [%[inptr0], %[in_col_stride1]]\n" - "ldr s27, [%[inptr0], x21]\n" - "fadd v4.4s, v23.4s, v27.4s\n" - "ldr s24, [%[inptr0], x22]\n" - "fsub v13.4s, v23.4s, v27.4s\n" - "ldr s11, [%[inptr0], x23]\n" - "fadd v10.4s, v24.4s, v11.4s\n" - "ldr s12, [%[inptr0], x24]\n" - "fsub v11.4s, v24.4s, v11.4s\n" - "ldr s20, [x25]\n" - "fadd v7.4s, v17.4s, v4.4s\n" - "ldr s19, [x25, %[in_col_stride1]]\n" - "mov v4.16b, v4.16b\n" - "ldr s22, [x25, x21]\n" - "mov v1.16b, v13.16b\n" - "ldr s14, [x25, x22]\n" - "fmul v11.4s, v11.4s, v0.s[0]\n" - "ldr s18, [x25, x23]\n" - "fadd v7.4s, v7.4s, v10.4s\n" - "ldr s3, [x25, x24]\n" - "fmla v4.4s, v10.4s, v0.s[1]\n" - "ldr s16, [x13]\n" - "fadd v2.4s, v19.4s, v22.4s\n" - "ldr s21, [x13, %[in_col_stride1]]\n" - "fadd v8.4s, v13.4s, v11.4s\n" - "ldr s24, [x13, x21]\n" - "fmla v1.4s, v11.4s, v0.s[1]\n" - "ldr s25, [x13, x22]\n" - "fadd v23.4s, v14.4s, v18.4s\n" - "ldr s17, [x13, x23]\n" - "fadd v11.4s, v20.4s, v2.4s\n" - "ldr s9, [x13, x24]\n" - "fsub v15.4s, v19.4s, v22.4s\n" - "ldr s6, [x26]\n" - "fadd v1.4s, v1.4s, v12.4s\n" - "ldr s19, [x26, %[in_col_stride1]]\n" - "fsub v31.4s, v14.4s, v18.4s\n" - "ldr s22, [x26, x21]\n" - "fadd v11.4s, v11.4s, v23.4s\n" - "ldr s12, [x26, x22]\n" - "mov v13.16b, v2.16b\n" - "ldr s26, [x26, x23]\n" - "mov v2.16b, v15.16b\n" - "ldr s5, [x26, x24]\n" - "fmul v31.4s, v31.4s, v0.s[0]\n" - "ldr s10, [x14]\n" - "fmla v13.4s, v23.4s, v0.s[1]\n" - "fadd v29.4s, v21.4s, v24.4s\n" - "fsub v14.4s, v21.4s, v24.4s\n" - "fadd v18.4s, v25.4s, v17.4s\n" - "fsub v28.4s, v25.4s, v17.4s\n" - "ldr s30, [x14, %[in_col_stride1]]\n" - "fadd v15.4s, v15.4s, v31.4s\n" - "fmla v2.4s, v31.4s, v0.s[1]\n" - "fadd v27.4s, v16.4s, v29.4s\n" - "mov v21.16b, v29.16b\n" - "fadd v20.4s, v19.4s, v22.4s\n" - "fsub v17.4s, v19.4s, v22.4s\n" - "fmul v28.4s, v28.4s, v0.s[0]\n" - "ldr s31, [x14, x21]\n" - "fadd v2.4s, v2.4s, v3.4s\n" - "ldr s23, [x14, x22]\n" - "fadd v27.4s, v27.4s, v18.4s\n" - "fmla v21.4s, v18.4s, v0.s[1]\n" - "fadd v29.4s, v12.4s, v26.4s\n" - "fadd v24.4s, v6.4s, v20.4s\n" - "fsub v16.4s, v12.4s, v26.4s\n" - "mov v6.16b, v20.16b\n" - "fadd v25.4s, v30.4s, v31.4s\n" - "fsub v22.4s, v30.4s, v31.4s\n" - "fadd v20.4s, v14.4s, v28.4s\n" - "mov v3.16b, v14.16b\n" - "fadd v24.4s, v24.4s, v29.4s\n" - "fmla v6.4s, v29.4s, v0.s[1]\n" - "fmul v16.4s, v16.4s, v0.s[0]\n" - "ldr s26, [x14, x23]\n" - "fmla v3.4s, v28.4s, v0.s[1]\n" - "fadd v14.4s, v23.4s, v26.4s\n" - "fadd v29.4s, v10.4s, v25.4s\n" - "fsub v23.4s, v23.4s, v26.4s\n" - "mov v10.16b, v25.16b\n" - "fadd v31.4s, v11.4s, v27.4s\n" - "fsub v12.4s, v11.4s, v27.4s\n" - "ldr s18, [x14, x24]\n" - "fadd v3.4s, v3.4s, v9.4s\n" - "ldr s19, [x27]\n" - "fadd v29.4s, v29.4s, v14.4s\n" - "fmul v23.4s, v23.4s, v0.s[0]\n" - "fmla v10.4s, v14.4s, v0.s[1]\n" - "fadd v25.4s, v7.4s, v31.4s\n" - "mov v26.16b, v31.16b\n" - "fadd v31.4s, v15.4s, v20.4s\n" - "fsub v11.4s, v15.4s, v20.4s\n" - "fadd v30.4s, v13.4s, v21.4s\n" - "fsub v9.4s, v13.4s, v21.4s\n" - "fadd v28.4s, v24.4s, v29.4s\n" - "fsub v24.4s, v24.4s, v29.4s\n" - "ldr s27, [x27, %[in_col_stride1]]\n" - "fadd v15.4s, v8.4s, v31.4s\n" - "mov v14.16b, v31.16b\n" - "fadd v13.4s, v4.4s, v30.4s\n" - "mov v4.16b, v30.16b\n" - "fadd v25.4s, v25.4s, v28.4s\n" - "fmla v26.4s, v28.4s, v0.s[1]\n" - "fmul v24.4s, v24.4s, v0.s[0]\n" - "fadd v21.4s, v6.4s, v10.4s\n" - "fsub v10.4s, v6.4s, v10.4s\n" - "fadd v6.4s, v2.4s, v3.4s\n" - "fsub v2.4s, v2.4s, v3.4s\n" - "ldr s29, [x27, x21]\n" - "str s25, [%[outptr0]]\n" // Store output (0, 0) - "fadd v20.4s, v17.4s, v16.4s\n" - "str s26, [x28]\n" // Store output (2, 0) - "mov v7.16b, v17.16b\n" - "fadd v17.4s, v22.4s, v23.4s\n" - "mov v8.16b, v22.16b\n" - "fadd v13.4s, v13.4s, v21.4s\n" - "fmul v10.4s, v10.4s, v0.s[0]\n" - "fmla v7.4s, v16.4s, v0.s[1]\n" - "ldr s28, [x27, x22]\n" - "fmla v8.4s, v23.4s, v0.s[1]\n" - "ldr s23, [x27, x23]\n" - "fmla v4.4s, v21.4s, v0.s[1]\n" - "ldr s25, [x27, x24]\n" - "str s13, [%[outptr0], x15]\n" // Store output (0, 2) - "fadd v16.4s, v27.4s, v29.4s\n" - "fadd v7.4s, v7.4s, v5.4s\n" - "fadd v30.4s, v28.4s, v23.4s\n" - "fadd v8.4s, v8.4s, v18.4s\n" - "fsub v29.4s, v27.4s, v29.4s\n" - "str s4, [x28, x15]\n" // Store output (2, 2) - "fsub v27.4s, v28.4s, v23.4s\n" - "fadd v19.4s, v19.4s, v16.4s\n" - "mov v16.16b, v16.16b\n" - "fadd v21.4s, v12.4s, v24.4s\n" - "mov v26.16b, v12.16b\n" - "mov v5.16b, v29.16b\n" - "fadd v22.4s, v20.4s, v17.4s\n" - "fmul v27.4s, v27.4s, v0.s[0]\n" - "fmla v16.4s, v30.4s, v0.s[1]\n" - "fadd v19.4s, v19.4s, v30.4s\n" - "fmla v26.4s, v24.4s, v0.s[1]\n" - "str s21, [x17]\n" // Store output (1, 0) - "fsub v20.4s, v20.4s, v17.4s\n" - "fadd v15.4s, v15.4s, v22.4s\n" - "fmla v14.4s, v22.4s, v0.s[1]\n" - "fadd v18.4s, v29.4s, v27.4s\n" - "fmla v5.4s, v27.4s, v0.s[1]\n" - "fadd v26.4s, v26.4s, v19.4s\n" - "mov v28.16b, v11.16b\n" - "fmul v20.4s, v20.4s, v0.s[0]\n" - "fadd v13.4s, v9.4s, v10.4s\n" - "str s15, [%[outptr0], %[output_col_stride1]]\n" // Store output (0, 1) - "mov v15.16b, v9.16b\n" - "str s14, [x28, %[output_col_stride1]]\n" // Store output (2, 1) - "fadd v5.4s, v5.4s, v25.4s\n" - "str s26, [x18]\n" // Store output (3, 0) - "fadd v30.4s, v7.4s, v8.4s\n" - "str s13, [x17, x15]\n" // Store output (1, 2) - "fadd v12.4s, v11.4s, v20.4s\n" - "fmla v28.4s, v20.4s, v0.s[1]\n" - "fmla v15.4s, v10.4s, v0.s[1]\n" - "fadd v1.4s, v1.4s, v6.4s\n" - "fsub v8.4s, v7.4s, v8.4s\n" - "mov v6.16b, v6.16b\n" - "mov v9.16b, v2.16b\n" - "str s12, [x17, %[output_col_stride1]]\n" // Store output (1, 1) - "fadd v28.4s, v28.4s, v18.4s\n" - "fadd v15.4s, v15.4s, v16.4s\n" - "fadd v1.4s, v1.4s, v30.4s\n" - "fmul v8.4s, v8.4s, v0.s[0]\n" - "fmla v6.4s, v30.4s, v0.s[1]\n" - "str s28, [x18, %[output_col_stride1]]\n" // Store output (3, 1) - "str s1, [%[outptr0], x16]\n" // Store output (0, 3) - "str s6, [x28, x16]\n" // Store output (2, 3) - "fadd v2.4s, v2.4s, v8.4s\n" - "str s15, [x18, x15]\n" // Store output (3, 2) - "fmla v9.4s, v8.4s, v0.s[1]\n" - "str s2, [x17, x16]\n" // Store output (1, 3) - "fadd v9.4s, v9.4s, v5.4s\n" - "str s9, [x18, x16]\n" // Store output (3, 3) - "6:\n" // End - : [outptr0] "+r" (output), [inptr0] "+r" (inptr) - : [output_col_stride1] "r" (output_col_stride * sizeof(float)), [pcoeffs] "r" (coeffs), [n_channels] "r" ((long) n_channels), [in_row_stride] "r" (6 * matrix_stride * sizeof(float)), [in_col_stride1] "r" (matrix_stride * sizeof(float)), [output_row_stride] "r" (output_row_stride * sizeof(float)) - : "cc", "v0", "v1", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v2", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v3", "v30", "v31", "v4", "v5", "v6", "v7", "v8", "v9", "x13", "x14", "x15", "x16", "x17", "x18", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "memory" - ); - } -} - -#else - template <> void winograd::OutputTransform<3, 3, 6, 6, float, float, winograd::WinogradRoots::Integers>::transform_tile( const int n_channels, @@ -1713,7 +36,9 @@ void winograd::OutputTransform<3, 3, 6, 6, float, float, winograd::WinogradRoots const float* bptr, float* const output, const int output_row_stride, - const int output_col_stride + const int output_col_stride, + const float output_min, + const float output_max ) { // Construct a map to the output cells @@ -1728,7 +53,79 @@ void winograd::OutputTransform<3, 3, 6, 6, float, float, winograd::WinogradRoots // For each channel of the output int channels_remaining = n_channels; -#ifdef __arm__ + +#ifdef __aarch64__ + for (; channels_remaining >= 4; channels_remaining -= 4) + { + // Matrices used and computed during this transform + float32x4_t F[6][6], FZ[6][4], f[4][4], b; + + // Read a 6x6 tile in the Winograd domain + for (int i = 0, m = 0; i < 6; i++) + { + for (int j = 0; j < 6; j++, m++) + { + F[i][j] = vld1q_f32(inptr + m*matrix_stride); + } + } + inptr += 4; + + // Compute the matrix F Z + for (int i = 0; i < 6; i++) + { + // FZ[i][0] = 1*F[i][0] + 1*F[i][1] + 1*F[i][2] + 1*F[i][3] + 1*F[i][4]; + FZ[i][0] = vaddq_f32(vaddq_f32(vaddq_f32(F[i][0], F[i][1]), vaddq_f32(F[i][2], F[i][3])), F[i][4]); + + // FZ[i][1] = 1*F[i][1] + -1*F[i][2] + 2*F[i][3] + -2*F[i][4]; + FZ[i][1] = vmlaq_n_f32(vsubq_f32(F[i][1], F[i][2]), vsubq_f32(F[i][3], F[i][4]), 2.0f); + + // FZ[i][2] = 1*F[i][1] + 1*F[i][2] + 4*F[i][3] + 4*F[i][4]; + FZ[i][2] = vmlaq_n_f32(vaddq_f32(F[i][1], F[i][2]), vaddq_f32(F[i][3], F[i][4]), 4.0f); + + // FZ[i][3] = 1*F[i][1] + -1*F[i][2] + 8*F[i][3] + -8*F[i][4] + 1*F[i][5]; + FZ[i][3] = vaddq_f32(vmlaq_n_f32(vsubq_f32(F[i][1], F[i][2]), vsubq_f32(F[i][3], F[i][4]), 8.0f), F[i][5]); + } + + // Compute the output tile f = ZT F Z + for (int j = 0; j < 4; j++) + { + // f[0][j] = 1*FZ[0][j] + 1*FZ[1][j] + 1*FZ[2][j] + 1*FZ[3][j] + 1*FZ[4][j]; + f[0][j] = vaddq_f32(vaddq_f32(vaddq_f32(FZ[0][j], FZ[1][j]), vaddq_f32(FZ[2][j], FZ[3][j])), FZ[4][j]); + + // f[1][j] = 1*FZ[1][j] + -1*FZ[2][j] + 2*FZ[3][j] + -2*FZ[4][j]; + f[1][j] = vmlaq_n_f32(vsubq_f32(FZ[1][j], FZ[2][j]), vsubq_f32(FZ[3][j], FZ[4][j]), 2.0f); + + // f[2][j] = 1*FZ[1][j] + 1*FZ[2][j] + 4*FZ[3][j] + 4*FZ[4][j]; + f[2][j] = vmlaq_n_f32(vaddq_f32(FZ[1][j], FZ[2][j]), vaddq_f32(FZ[3][j], FZ[4][j]), 4.0f); + + // f[3][j] = 1*FZ[1][j] + -1*FZ[2][j] + 8*FZ[3][j] + -8*FZ[4][j] + 1*FZ[5][j]; + f[3][j] = vaddq_f32(vmlaq_n_f32(vsubq_f32(FZ[1][j], FZ[2][j]), vsubq_f32(FZ[3][j], FZ[4][j]), 8.0f), FZ[5][j]); + } + + // Write out the output tile + if (bptr != nullptr) + { + b = vld1q_f32(bptr); + bptr += 4; + } + else + { + b = vdupq_n_f32(0.0f); + } + for (int i = 0; i < output_tile_rows; i++) + { + for (int j = 0; j < output_tile_cols; j++) + { + const auto y = + vmaxq_f32(vminq_f32(vaddq_f32(f[i][j], b), vdupq_n_f32(output_max)), + vdupq_n_f32(output_min)); + vst1q_f32(outptrs[i][j], y); + outptrs[i][j] += 4; + } + } + } +#endif // __aarch64__ +#ifdef __arm_any__ for (; channels_remaining >= 2; channels_remaining -= 2) { // Matrices used and computed during this transform @@ -1790,12 +187,15 @@ void winograd::OutputTransform<3, 3, 6, 6, float, float, winograd::WinogradRoots { for (int j = 0; j < output_tile_cols; j++) { - vst1_f32(outptrs[i][j], vadd_f32(f[i][j], b)); + const auto y = + vmax_f32(vmin_f32(vadd_f32(f[i][j], b), vdup_n_f32(output_max)), + vdup_n_f32(output_min)); + vst1_f32(outptrs[i][j], y); outptrs[i][j] += 2; } } } -#endif // __arm__ +#endif // __arm_any__ for (; channels_remaining; channels_remaining--) { // Matrices used and computed during this transform @@ -1842,14 +242,13 @@ void winograd::OutputTransform<3, 3, 6, 6, float, float, winograd::WinogradRoots { for (int j = 0; j < output_tile_cols; j++) { - *(outptrs[i][j]++) = f[i][j] + b; + const auto y = std::max(std::min(f[i][j] + b, output_max), output_min); + *(outptrs[i][j]++) = y; } } } } -#endif - template class OutputTransform<3, 3, 6, 6, float, float, winograd::WinogradRoots::Integers>; } // namespace winograd -- cgit v1.2.1