| Giorgio Arena | a50e5e0 | 2018-07-02 13:42:23 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2018 ARM Limited. |
| 3 | * |
| 4 | * SPDX-License-Identifier: MIT |
| 5 | * |
| 6 | * Permission is hereby granted, free of charge, to any person obtaining a copy |
| 7 | * of this software and associated documentation files (the "Software"), to |
| 8 | * deal in the Software without restriction, including without limitation the |
| 9 | * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or |
| 10 | * sell copies of the Software, and to permit persons to whom the Software is |
| 11 | * furnished to do so, subject to the following conditions: |
| 12 | * |
| 13 | * The above copyright notice and this permission notice shall be included in all |
| 14 | * copies or substantial portions of the Software. |
| 15 | * |
| 16 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 17 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 18 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
| 19 | * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
| 20 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
| 21 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE |
| 22 | * SOFTWARE. |
| 23 | */ |
| 24 | #include "helpers.h" |
| 25 | |
| 26 | #if defined(SRC_DIM_Z) |
| 27 | |
| 28 | /** This OpenCL kernel performs Winograd filter transform 3x3/3x1/1x3 when the data layout is NCHW and the output tile is 2x2/2x1/1x2 |
| 29 | * |
| 30 | * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64 |
| 31 | * @note If this kernel is used to perform Winograd filter transform 3x1, -DWINOGRAD_FILTER_TRANSFORM_HORIZONTAL has to be passed at compile time |
| 32 | * @note If this kernel is used to perform Winograd filter transform 1x3, -DWINOGRAD_FILTER_TRANSFORM_VERTICAL has to be passed at compile time |
| 33 | * |
| 34 | * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 |
| 35 | * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) |
| 36 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 37 | * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 38 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 39 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 40 | * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 41 | * @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes) |
| 42 | * @param[in] src_step_w src_stride_w * number of elements along W processed per workitem(in bytes) |
| 43 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor |
| 44 | * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr |
| 45 | * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 46 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 47 | * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 48 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 49 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 50 | * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 51 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 52 | */ |
| 53 | __kernel void winograd_filter_transform_2x2_3x3_nchw( |
| 54 | TENSOR4D_DECLARATION(src), |
| 55 | TENSOR3D_DECLARATION(dst)) |
| 56 | { |
| 57 | Tensor4D src = CONVERT_TO_TENSOR4D_STRUCT(src, SRC_DIM_Z); |
| 58 | |
| 59 | const __global uchar *src_addr = tensor4D_offset(&src, 0, 0, 0, 0); |
| 60 | |
| 61 | // Load the values from the input tensor |
| 62 | #if defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) |
| 63 | float3 w0 = vload3(0, (__global float *)(src_addr)); |
| 64 | #elif defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) |
| 65 | float3 w0 = (float3)(*((__global float *)(src_addr + 0 * src_stride_y)), |
| 66 | *((__global float *)(src_addr + 1 * src_stride_y)), |
| 67 | *((__global float *)(src_addr + 2 * src_stride_y))); |
| 68 | #else // defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) |
| Gian Marco Iodice | 876be2a | 2018-07-03 12:22:09 +0100 | [diff] [blame] | 69 | float3 w0 = vload3(0, (__global float *)(src_addr + 0 * src_stride_y)); |
| 70 | float3 w1 = vload3(0, (__global float *)(src_addr + 1 * src_stride_y)); |
| 71 | float3 w2 = vload3(0, (__global float *)(src_addr + 2 * src_stride_y)); |
| Giorgio Arena | a50e5e0 | 2018-07-02 13:42:23 +0100 | [diff] [blame] | 72 | #endif // defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) |
| 73 | |
| 74 | // Row 0 |
| 75 | float4 out0 = 0.0f; |
| 76 | out0.s0 = (w0.s0); |
| 77 | out0.s1 = (w0.s0 + w0.s1 + w0.s2) * 0.5f; |
| 78 | out0.s2 = (w0.s0 + w0.s2 - w0.s1) * 0.5f; |
| 79 | out0.s3 = (w0.s2); |
| 80 | |
| 81 | #if !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) |
| 82 | // Row 1 |
| 83 | float4 out1 = 0.0f; |
| 84 | out1.s0 = (w0.s0 + w1.s0 + w2.s0) * 0.5f; |
| 85 | out1.s1 = (w0.s0 + w1.s0 + w2.s0 + w0.s1 + w1.s1 + w2.s1 + w0.s2 + w1.s2 + w2.s2) * 0.25f; |
| 86 | out1.s2 = (w0.s0 + w1.s0 + w2.s0 + w0.s2 + w1.s2 + w2.s2 - w0.s1 - w1.s1 - w2.s1) * 0.25f; |
| 87 | out1.s3 = (w0.s2 + w1.s2 + w2.s2) * 0.5f; |
| 88 | |
| 89 | // Row 2 |
| 90 | float4 out2 = 0.0f; |
| 91 | out2.s0 = (w0.s0 + w2.s0 - w1.s0) * 0.5f; |
| 92 | out2.s1 = (w0.s0 + w2.s0 + w0.s1 + w2.s1 + w0.s2 + w2.s2 - w1.s0 - w1.s1 - w1.s2) * 0.25f; |
| 93 | out2.s2 = (w0.s0 + w2.s0 + w1.s1 + w0.s2 + w2.s2 - w1.s0 - w0.s1 - w2.s1 - w1.s2) * 0.25f; |
| 94 | out2.s3 = (w0.s2 + w2.s2 - w1.s2) * 0.5f; |
| 95 | |
| 96 | // Row 3 |
| 97 | float4 out3 = 0.0f; |
| 98 | out3.s0 = (w2.s0); |
| 99 | out3.s1 = (w2.s0 + w2.s1 + w2.s2) * 0.5f; |
| 100 | out3.s2 = (w2.s0 + w2.s2 - w2.s1) * 0.5f; |
| 101 | out3.s3 = (w2.s2); |
| 102 | #endif // !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) |
| 103 | |
| 104 | int z = get_global_id(2); |
| 105 | int x0 = z / SRC_DIM_Z; // idx filter |
| 106 | int y0 = z % SRC_DIM_Z; // idx channel |
| 107 | |
| 108 | // Get output address |
| 109 | __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x0 * dst_stride_x + y0 * dst_stride_y; |
| 110 | |
| 111 | // Store the values across the channels |
| 112 | // 16 channels for 3x3 kernels |
| 113 | // 4 channels for 3x1 or 1x3 kernels |
| 114 | *(__global float *)(dst_addr + 0 * dst_stride_z) = out0.s0; |
| 115 | *(__global float *)(dst_addr + 1 * dst_stride_z) = out0.s1; |
| 116 | *(__global float *)(dst_addr + 2 * dst_stride_z) = out0.s2; |
| 117 | *(__global float *)(dst_addr + 3 * dst_stride_z) = out0.s3; |
| 118 | |
| 119 | #if !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) |
| 120 | *(__global float *)(dst_addr + 4 * dst_stride_z) = out1.s0; |
| 121 | *(__global float *)(dst_addr + 5 * dst_stride_z) = out1.s1; |
| 122 | *(__global float *)(dst_addr + 6 * dst_stride_z) = out1.s2; |
| 123 | *(__global float *)(dst_addr + 7 * dst_stride_z) = out1.s3; |
| 124 | *(__global float *)(dst_addr + 8 * dst_stride_z) = out2.s0; |
| 125 | *(__global float *)(dst_addr + 9 * dst_stride_z) = out2.s1; |
| 126 | *(__global float *)(dst_addr + 10 * dst_stride_z) = out2.s2; |
| 127 | *(__global float *)(dst_addr + 11 * dst_stride_z) = out2.s3; |
| 128 | *(__global float *)(dst_addr + 12 * dst_stride_z) = out3.s0; |
| 129 | *(__global float *)(dst_addr + 13 * dst_stride_z) = out3.s1; |
| 130 | *(__global float *)(dst_addr + 14 * dst_stride_z) = out3.s2; |
| 131 | *(__global float *)(dst_addr + 15 * dst_stride_z) = out3.s3; |
| 132 | #endif // !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) |
| 133 | } |
| 134 | |
| 135 | /** This OpenCL kernel performs Winograd filter transform 3x3/3x1/1x3 when the data layout is NCHW and the output tile is 4x4/4x1/1x4 |
| 136 | * |
| 137 | * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64 |
| 138 | * @note If this kernel is used to perform Winograd filter transform 3x1, -DWINOGRAD_FILTER_TRANSFORM_HORIZONTAL has to be passed at compile time |
| 139 | * @note If this kernel is used to perform Winograd filter transform 1x3, -DWINOGRAD_FILTER_TRANSFORM_VERTICAL has to be passed at compile time |
| 140 | * |
| 141 | * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 |
| 142 | * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) |
| 143 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 144 | * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 145 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 146 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 147 | * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 148 | * @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes) |
| 149 | * @param[in] src_step_w src_stride_w * number of elements along W processed per workitem(in bytes) |
| 150 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor |
| 151 | * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr |
| 152 | * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 153 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 154 | * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 155 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 156 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 157 | * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 158 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 159 | */ |
| 160 | __kernel void winograd_filter_transform_4x4_3x3_nchw( |
| 161 | TENSOR4D_DECLARATION(src), |
| 162 | TENSOR3D_DECLARATION(dst)) |
| 163 | { |
| 164 | Tensor4D src = CONVERT_TO_TENSOR4D_STRUCT(src, SRC_DIM_Z); |
| 165 | |
| 166 | const __global uchar *src_addr = tensor4D_offset(&src, 0, 0, 0, 0); |
| 167 | |
| 168 | // Load the values from the input tensor |
| 169 | #if defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) |
| 170 | float3 w0 = vload3(0, (__global float *)(src_addr)); |
| 171 | #elif defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) |
| 172 | float3 w0 = (float3)(*((__global float *)(src_addr + 0 * src_stride_y)), |
| 173 | *((__global float *)(src_addr + 1 * src_stride_y)), |
| 174 | *((__global float *)(src_addr + 2 * src_stride_y))); |
| 175 | #else // defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) |
| Gian Marco Iodice | 876be2a | 2018-07-03 12:22:09 +0100 | [diff] [blame] | 176 | float3 w0 = vload3(0, (__global float *)(src_addr + 0 * src_stride_y)); |
| 177 | float3 w1 = vload3(0, (__global float *)(src_addr + 1 * src_stride_y)); |
| 178 | float3 w2 = vload3(0, (__global float *)(src_addr + 2 * src_stride_y)); |
| Giorgio Arena | a50e5e0 | 2018-07-02 13:42:23 +0100 | [diff] [blame] | 179 | #endif // defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) |
| 180 | |
| 181 | // Row 0 |
| 182 | float8 out0 = 0.0f; |
| 183 | out0.s0 = (w0.s0) / 16.f; |
| 184 | out0.s1 = (-w0.s0 - w0.s1 - w0.s2) / 24.f; |
| 185 | out0.s2 = (-w0.s0 + w0.s1 - w0.s2) / 24.f; |
| 186 | out0.s3 = (w0.s0 + 2.f * w0.s1 + 4.f * w0.s2) / 96.f; |
| 187 | out0.s4 = (w0.s0 - 2.f * w0.s1 + 4.f * w0.s2) / 96.f; |
| 188 | out0.s5 = (w0.s2) / 4.f; |
| 189 | |
| 190 | #if !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) |
| 191 | // Row 1 |
| 192 | float8 out1 = 0.0f; |
| 193 | out1.s0 = (-w0.s0 - w1.s0 - w2.s0) / 24.f; |
| 194 | out1.s1 = (w0.s0 + w1.s0 + w2.s0 + w0.s1 + w1.s1 + w2.s1 + w0.s2 + w1.s2 + w2.s2) / 36.f; |
| 195 | out1.s2 = (w0.s0 + w1.s0 + w2.s0 - w0.s1 - w1.s1 - w2.s1 + w0.s2 + w1.s2 + w2.s2) / 36.f; |
| 196 | out1.s3 = (-w0.s0 - w1.s0 - w2.s0 + 2.f * (-w0.s1 - w1.s1 - w2.s1) + 4.f * (-w0.s2 - w1.s2 - w2.s2)) / 144.f; |
| 197 | out1.s4 = (-w0.s0 - w1.s0 - w2.s0 + 2.f * (w0.s1 + w1.s1 + w2.s1) + 4.f * (-w0.s2 - w1.s2 - w2.s2)) / 144.f; |
| 198 | out1.s5 = (-w0.s2 - w1.s2 - w2.s2) / 6.f; |
| 199 | |
| 200 | // Row 2 |
| 201 | float8 out2 = 0.0f; |
| 202 | out2.s0 = (-w0.s0 + w1.s0 - w2.s0) / 24.f; |
| 203 | out2.s1 = (w0.s0 - w1.s0 + w2.s0 + w0.s1 - w1.s1 + w2.s1 + w0.s2 - w1.s2 + w2.s2) / 36.f; |
| 204 | out2.s2 = (w0.s0 - w1.s0 + w2.s0 - w0.s1 + w1.s1 - w2.s1 + w0.s2 - w1.s2 + w2.s2) / 36.f; |
| 205 | out2.s3 = (-w0.s0 + w1.s0 - w2.s0 + 2.f * (-w0.s1 + w1.s1 - w2.s1) + 4.f * (-w0.s2 + w1.s2 - w2.s2)) / 144.f; |
| 206 | out2.s4 = (-w0.s0 + w1.s0 - w2.s0 + 2.f * (w0.s1 - w1.s1 + w2.s1) + 4.f * (-w0.s2 + w1.s2 - w2.s2)) / 144.f; |
| 207 | out2.s5 = (-w0.s2 + w1.s2 - w2.s2) / 6.f; |
| 208 | |
| 209 | // Row 3 |
| 210 | float8 out3 = 0.0f; |
| 211 | out3.s0 = (w0.s0 + 2.f * w1.s0 + 4.f * w2.s0) / 96.f; |
| 212 | out3.s1 = (-w0.s0 - 2.f * w1.s0 - 4.f * w2.s0 - w0.s1 - 2.f * w1.s1 - 4.f * w2.s1 - w0.s2 - 2.f * w1.s2 - 4.f * w2.s2) / 144.f; |
| 213 | out3.s2 = (-w0.s0 - 2.f * w1.s0 - 4.f * w2.s0 + w0.s1 + 2.f * w1.s1 + 4.f * w2.s1 - w0.s2 - 2.f * w1.s2 - 4.f * w2.s2) / 144.f; |
| 214 | out3.s3 = ((w0.s0 + 2.f * w1.s0 + 4.f * w2.s0) + 2.f * (w0.s1 + 2.f * w1.s1 + 4.f * w2.s1) + 4.f * (w0.s2 + 2.f * w1.s2 + 4.f * w2.s2)) / 576.f; |
| 215 | out3.s4 = ((w0.s0 + 2.f * w1.s0 + 4.f * w2.s0) + 2.f * (-w0.s1 - 2.f * w1.s1 - 4.f * w2.s1) + 4.f * (w0.s2 + 2.f * w1.s2 + 4.f * w2.s2)) / 576.f; |
| 216 | out3.s5 = (w0.s2 + 2.f * w1.s2 + 4.f * w2.s2) / 24.f; |
| 217 | |
| 218 | // Row 4 |
| 219 | float8 out4 = 0.0f; |
| 220 | out4.s0 = (w0.s0 - 2.f * w1.s0 + 4.f * w2.s0) / 96.f; |
| 221 | out4.s1 = (-w0.s0 + 2.f * w1.s0 - 4.f * w2.s0 - w0.s1 + 2.f * w1.s1 - 4.f * w2.s1 - w0.s2 + 2.f * w1.s2 - 4.f * w2.s2) / 144.f; |
| 222 | out4.s2 = (-w0.s0 + 2.f * w1.s0 - 4.f * w2.s0 + w0.s1 - 2.f * w1.s1 + 4.f * w2.s1 - w0.s2 + 2.f * w1.s2 - 4.f * w2.s2) / 144.f; |
| 223 | out4.s3 = ((w0.s0 - 2.f * w1.s0 + 4.f * w2.s0) + 2.f * (w0.s1 - 2.f * w1.s1 + 4.f * w2.s1) + 4.f * (w0.s2 - 2.f * w1.s2 + 4.f * w2.s2)) / 576.f; |
| 224 | out4.s4 = ((w0.s0 - 2.f * w1.s0 + 4.f * w2.s0) + 2.f * (-w0.s1 + 2.f * w1.s1 - 4.f * w2.s1) + 4.f * (w0.s2 - 2.f * w1.s2 + 4.f * w2.s2)) / 576.f; |
| 225 | out4.s5 = (w0.s2 - 2.f * w1.s2 + 4.f * w2.s2) / 24.f; |
| 226 | |
| 227 | // Row 5 |
| 228 | float8 out5 = 0.0f; |
| 229 | out5.s0 = (w2.s0) / 4.f; |
| 230 | out5.s1 = (-w2.s0 - w2.s1 - w2.s2) / 6.f; |
| 231 | out5.s2 = (-w2.s0 + w2.s1 - w2.s2) / 6.f; |
| 232 | out5.s3 = (w2.s0 + 2.f * w2.s1 + 4.f * w2.s2) / 24.f; |
| 233 | out5.s4 = (w2.s0 - 2.f * w2.s1 + 4.f * w2.s2) / 24.f; |
| 234 | out5.s5 = (w2.s2); |
| 235 | #endif // !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) |
| 236 | |
| 237 | int z = get_global_id(2); |
| 238 | int x0 = z / SRC_DIM_Z; // idx filter |
| 239 | int y0 = z % SRC_DIM_Z; // idx channel |
| 240 | |
| 241 | // Get output address |
| 242 | __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x0 * dst_stride_x + y0 * dst_stride_y; |
| 243 | |
| 244 | // Store the values across the channels |
| 245 | // 36 channels for 3x3 kernels |
| 246 | // 6 channels for 3x1 or 1x3 kernels |
| 247 | *(__global float *)(dst_addr + 0 * dst_stride_z) = out0.s0; |
| 248 | *(__global float *)(dst_addr + 1 * dst_stride_z) = out0.s1; |
| 249 | *(__global float *)(dst_addr + 2 * dst_stride_z) = out0.s2; |
| 250 | *(__global float *)(dst_addr + 3 * dst_stride_z) = out0.s3; |
| 251 | *(__global float *)(dst_addr + 4 * dst_stride_z) = out0.s4; |
| 252 | *(__global float *)(dst_addr + 5 * dst_stride_z) = out0.s5; |
| 253 | |
| 254 | #if !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) |
| 255 | *(__global float *)(dst_addr + 6 * dst_stride_z) = out1.s0; |
| 256 | *(__global float *)(dst_addr + 7 * dst_stride_z) = out1.s1; |
| 257 | *(__global float *)(dst_addr + 8 * dst_stride_z) = out1.s2; |
| 258 | *(__global float *)(dst_addr + 9 * dst_stride_z) = out1.s3; |
| 259 | *(__global float *)(dst_addr + 10 * dst_stride_z) = out1.s4; |
| 260 | *(__global float *)(dst_addr + 11 * dst_stride_z) = out1.s5; |
| 261 | *(__global float *)(dst_addr + 12 * dst_stride_z) = out2.s0; |
| 262 | *(__global float *)(dst_addr + 13 * dst_stride_z) = out2.s1; |
| 263 | *(__global float *)(dst_addr + 14 * dst_stride_z) = out2.s2; |
| 264 | *(__global float *)(dst_addr + 15 * dst_stride_z) = out2.s3; |
| 265 | *(__global float *)(dst_addr + 16 * dst_stride_z) = out2.s4; |
| 266 | *(__global float *)(dst_addr + 17 * dst_stride_z) = out2.s5; |
| 267 | *(__global float *)(dst_addr + 18 * dst_stride_z) = out3.s0; |
| 268 | *(__global float *)(dst_addr + 19 * dst_stride_z) = out3.s1; |
| 269 | *(__global float *)(dst_addr + 20 * dst_stride_z) = out3.s2; |
| 270 | *(__global float *)(dst_addr + 21 * dst_stride_z) = out3.s3; |
| 271 | *(__global float *)(dst_addr + 22 * dst_stride_z) = out3.s4; |
| 272 | *(__global float *)(dst_addr + 23 * dst_stride_z) = out3.s5; |
| 273 | *(__global float *)(dst_addr + 24 * dst_stride_z) = out4.s0; |
| 274 | *(__global float *)(dst_addr + 25 * dst_stride_z) = out4.s1; |
| 275 | *(__global float *)(dst_addr + 26 * dst_stride_z) = out4.s2; |
| 276 | *(__global float *)(dst_addr + 27 * dst_stride_z) = out4.s3; |
| 277 | *(__global float *)(dst_addr + 28 * dst_stride_z) = out4.s4; |
| 278 | *(__global float *)(dst_addr + 29 * dst_stride_z) = out4.s5; |
| 279 | *(__global float *)(dst_addr + 30 * dst_stride_z) = out5.s0; |
| 280 | *(__global float *)(dst_addr + 31 * dst_stride_z) = out5.s1; |
| 281 | *(__global float *)(dst_addr + 32 * dst_stride_z) = out5.s2; |
| 282 | *(__global float *)(dst_addr + 33 * dst_stride_z) = out5.s3; |
| 283 | *(__global float *)(dst_addr + 34 * dst_stride_z) = out5.s4; |
| 284 | *(__global float *)(dst_addr + 35 * dst_stride_z) = out5.s5; |
| 285 | #endif // !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) |
| 286 | } |
| 287 | |
| Giorgio Arena | 149fdf3 | 2018-07-04 17:03:33 +0100 | [diff] [blame] | 288 | /** This OpenCL kernel performs Winograd filter transform 3x3/3x1/1x3 when the data layout is NHWC and the output tile is 4x4/4x1/1x4 |
| Giorgio Arena | a50e5e0 | 2018-07-02 13:42:23 +0100 | [diff] [blame] | 289 | * |
| 290 | * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64 |
| Giorgio Arena | 149fdf3 | 2018-07-04 17:03:33 +0100 | [diff] [blame] | 291 | * @note If this kernel is used to perform Winograd filter transform 3x1, -DWINOGRAD_FILTER_TRANSFORM_HORIZONTAL has to be passed at compile time |
| 292 | * @note If this kernel is used to perform Winograd filter transform 1x3, -DWINOGRAD_FILTER_TRANSFORM_VERTICAL has to be passed at compile time |
| Giorgio Arena | a50e5e0 | 2018-07-02 13:42:23 +0100 | [diff] [blame] | 293 | * |
| 294 | * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 |
| 295 | * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) |
| 296 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 297 | * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 298 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 299 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 300 | * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 301 | * @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes) |
| 302 | * @param[in] src_step_w src_stride_w * number of elements along W processed per workitem(in bytes) |
| 303 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor |
| 304 | * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr |
| 305 | * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 306 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 307 | * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 308 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 309 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 310 | * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 311 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 312 | */ |
| 313 | __kernel void winograd_filter_transform_4x4_3x3_nhwc( |
| 314 | TENSOR4D_DECLARATION(src), |
| 315 | TENSOR3D_DECLARATION(dst)) |
| 316 | { |
| 317 | Tensor4D src = CONVERT_TO_TENSOR4D_STRUCT(src, SRC_DIM_Z); |
| 318 | |
| 319 | const __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + get_global_id(0) * src_step_x + get_global_id(1) * src_step_y + get_global_id(2) * src_step_w; |
| 320 | |
| 321 | // Load the values from the input tensor |
| Giorgio Arena | 149fdf3 | 2018-07-04 17:03:33 +0100 | [diff] [blame] | 322 | #if defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) |
| 323 | float w00 = *((__global float *)(src_addr + 0 * src_stride_z)); |
| 324 | float w01 = *((__global float *)(src_addr + 1 * src_stride_z)); |
| 325 | float w02 = *((__global float *)(src_addr + 2 * src_stride_z)); |
| 326 | #else // defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) |
| 327 | float w00 = *((__global float *)(src_addr + 0 * src_stride_z + 0 * src_stride_y)); |
| 328 | float w01 = *((__global float *)(src_addr + 0 * src_stride_z + 1 * src_stride_y)); |
| 329 | float w02 = *((__global float *)(src_addr + 0 * src_stride_z + 2 * src_stride_y)); |
| 330 | #if !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) |
| 331 | float w10 = *((__global float *)(src_addr + 1 * src_stride_z + 0 * src_stride_y)); |
| 332 | float w11 = *((__global float *)(src_addr + 1 * src_stride_z + 1 * src_stride_y)); |
| 333 | float w12 = *((__global float *)(src_addr + 1 * src_stride_z + 2 * src_stride_y)); |
| 334 | float w20 = *((__global float *)(src_addr + 2 * src_stride_z + 0 * src_stride_y)); |
| 335 | float w21 = *((__global float *)(src_addr + 2 * src_stride_z + 1 * src_stride_y)); |
| 336 | float w22 = *((__global float *)(src_addr + 2 * src_stride_z + 2 * src_stride_y)); |
| 337 | #endif // !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) |
| 338 | #endif // defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) |
| Giorgio Arena | a50e5e0 | 2018-07-02 13:42:23 +0100 | [diff] [blame] | 339 | |
| 340 | // Row 0 |
| Giorgio Arena | 149fdf3 | 2018-07-04 17:03:33 +0100 | [diff] [blame] | 341 | float out00, out01, out02, out03, out04, out05; |
| Giorgio Arena | a50e5e0 | 2018-07-02 13:42:23 +0100 | [diff] [blame] | 342 | out00 = (w00) / 16.f; |
| 343 | out01 = (-w00 - w01 - w02) / 24.f; |
| 344 | out02 = (-w00 + w01 - w02) / 24.f; |
| 345 | out03 = (w00 + 2.f * w01 + 4.f * w02) / 96.f; |
| 346 | out04 = (w00 - 2.f * w01 + 4.f * w02) / 96.f; |
| 347 | out05 = (w02) / 4.f; |
| 348 | |
| Giorgio Arena | 149fdf3 | 2018-07-04 17:03:33 +0100 | [diff] [blame] | 349 | #if !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) |
| Giorgio Arena | a50e5e0 | 2018-07-02 13:42:23 +0100 | [diff] [blame] | 350 | // Row 1 |
| Giorgio Arena | 149fdf3 | 2018-07-04 17:03:33 +0100 | [diff] [blame] | 351 | float out10, out11, out12, out13, out14, out15; |
| Giorgio Arena | a50e5e0 | 2018-07-02 13:42:23 +0100 | [diff] [blame] | 352 | out10 = (-w00 - w10 - w20) / 24.f; |
| 353 | out11 = (w00 + w10 + w20 + w01 + w11 + w21 + w02 + w12 + w22) / 36.f; |
| 354 | out12 = (w00 + w10 + w20 - w01 - w11 - w21 + w02 + w12 + w22) / 36.f; |
| 355 | out13 = (-w00 - w10 - w20 + 2.f * (-w01 - w11 - w21) + 4.f * (-w02 - w12 - w22)) / 144.f; |
| 356 | out14 = (-w00 - w10 - w20 + 2.f * (w01 + w11 + w21) + 4.f * (-w02 - w12 - w22)) / 144.f; |
| 357 | out15 = (-w02 - w12 - w22) / 6.f; |
| 358 | |
| 359 | // Row 2 |
| Giorgio Arena | 149fdf3 | 2018-07-04 17:03:33 +0100 | [diff] [blame] | 360 | float out20, out21, out22, out23, out24, out25; |
| Giorgio Arena | a50e5e0 | 2018-07-02 13:42:23 +0100 | [diff] [blame] | 361 | out20 = (-w00 + w10 - w20) / 24.f; |
| 362 | out21 = (w00 - w10 + w20 + w01 - w11 + w21 + w02 - w12 + w22) / 36.f; |
| 363 | out22 = (w00 - w10 + w20 - w01 + w11 - w21 + w02 - w12 + w22) / 36.f; |
| 364 | out23 = (-w00 + w10 - w20 + 2.f * (-w01 + w11 - w21) + 4.f * (-w02 + w12 - w22)) / 144.f; |
| 365 | out24 = (-w00 + w10 - w20 + 2.f * (w01 - w11 + w21) + 4.f * (-w02 + w12 - w22)) / 144.f; |
| 366 | out25 = (-w02 + w12 - w22) / 6.f; |
| 367 | |
| 368 | // Row 3 |
| Giorgio Arena | 149fdf3 | 2018-07-04 17:03:33 +0100 | [diff] [blame] | 369 | float out30, out31, out32, out33, out34, out35; |
| Giorgio Arena | a50e5e0 | 2018-07-02 13:42:23 +0100 | [diff] [blame] | 370 | out30 = (w00 + 2.f * w10 + 4.f * w20) / 96.f; |
| 371 | out31 = (-w00 - 2.f * w10 - 4.f * w20 - w01 - 2.f * w11 - 4.f * w21 - w02 - 2.f * w12 - 4.f * w22) / 144.f; |
| 372 | out32 = (-w00 - 2.f * w10 - 4.f * w20 + w01 + 2.f * w11 + 4.f * w21 - w02 - 2.f * w12 - 4.f * w22) / 144.f; |
| 373 | out33 = ((w00 + 2.f * w10 + 4.f * w20) + 2.f * (w01 + 2.f * w11 + 4.f * w21) + 4.f * (w02 + 2.f * w12 + 4.f * w22)) / 576.f; |
| 374 | out34 = ((w00 + 2.f * w10 + 4.f * w20) + 2.f * (-w01 - 2.f * w11 - 4.f * w21) + 4.f * (w02 + 2.f * w12 + 4.f * w22)) / 576.f; |
| 375 | out35 = (w02 + 2.f * w12 + 4.f * w22) / 24.f; |
| 376 | |
| 377 | // Row 4 |
| Giorgio Arena | 149fdf3 | 2018-07-04 17:03:33 +0100 | [diff] [blame] | 378 | float out40, out41, out42, out43, out44, out45; |
| Giorgio Arena | a50e5e0 | 2018-07-02 13:42:23 +0100 | [diff] [blame] | 379 | out40 = (w00 - 2.f * w10 + 4.f * w20) / 96.f; |
| 380 | out41 = (-w00 + 2.f * w10 - 4.f * w20 - w01 + 2.f * w11 - 4.f * w21 - w02 + 2.f * w12 - 4.f * w22) / 144.f; |
| 381 | out42 = (-w00 + 2.f * w10 - 4.f * w20 + w01 - 2.f * w11 + 4.f * w21 - w02 + 2.f * w12 - 4.f * w22) / 144.f; |
| 382 | out43 = ((w00 - 2.f * w10 + 4.f * w20) + 2.f * (w01 - 2.f * w11 + 4.f * w21) + 4.f * (w02 - 2.f * w12 + 4.f * w22)) / 576.f; |
| 383 | out44 = ((w00 - 2.f * w10 + 4.f * w20) + 2.f * (-w01 + 2.f * w11 - 4.f * w21) + 4.f * (w02 - 2.f * w12 + 4.f * w22)) / 576.f; |
| 384 | out45 = (w02 - 2.f * w12 + 4.f * w22) / 24.f; |
| 385 | |
| 386 | // Row 5 |
| Giorgio Arena | 149fdf3 | 2018-07-04 17:03:33 +0100 | [diff] [blame] | 387 | float out50, out51, out52, out53, out54, out55; |
| Giorgio Arena | a50e5e0 | 2018-07-02 13:42:23 +0100 | [diff] [blame] | 388 | out50 = (w20) / 4.f; |
| 389 | out51 = (-w20 - w21 - w22) / 6.f; |
| 390 | out52 = (-w20 + w21 - w22) / 6.f; |
| 391 | out53 = (w20 + 2.f * w21 + 4.f * w22) / 24.f; |
| 392 | out54 = (w20 - 2.f * w21 + 4.f * w22) / 24.f; |
| 393 | out55 = (w22); |
| Giorgio Arena | 149fdf3 | 2018-07-04 17:03:33 +0100 | [diff] [blame] | 394 | #endif // !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) |
| Giorgio Arena | a50e5e0 | 2018-07-02 13:42:23 +0100 | [diff] [blame] | 395 | |
| 396 | int x0 = get_global_id(2); // idx filter |
| 397 | int y0 = get_global_id(0); // idx channel |
| 398 | |
| 399 | // Get output address |
| Giorgio Arena | 149fdf3 | 2018-07-04 17:03:33 +0100 | [diff] [blame] | 400 | __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x0 * sizeof(float) + y0 * dst_stride_y; |
| Giorgio Arena | a50e5e0 | 2018-07-02 13:42:23 +0100 | [diff] [blame] | 401 | |
| 402 | // Store the values across the channels |
| Giorgio Arena | 149fdf3 | 2018-07-04 17:03:33 +0100 | [diff] [blame] | 403 | // 36 channels for 3x3 kernels |
| 404 | // 6 channels for 3x1 or 1x3 kernels |
| 405 | *(__global float *)(dst_addr + 0 * dst_stride_z) = out00; |
| 406 | *(__global float *)(dst_addr + 1 * dst_stride_z) = out01; |
| 407 | *(__global float *)(dst_addr + 2 * dst_stride_z) = out02; |
| 408 | *(__global float *)(dst_addr + 3 * dst_stride_z) = out03; |
| 409 | *(__global float *)(dst_addr + 4 * dst_stride_z) = out04; |
| 410 | *(__global float *)(dst_addr + 5 * dst_stride_z) = out05; |
| 411 | #if !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) |
| Giorgio Arena | a50e5e0 | 2018-07-02 13:42:23 +0100 | [diff] [blame] | 412 | *(__global float *)(dst_addr + 6 * dst_stride_z) = out10; |
| 413 | *(__global float *)(dst_addr + 7 * dst_stride_z) = out11; |
| 414 | *(__global float *)(dst_addr + 8 * dst_stride_z) = out12; |
| 415 | *(__global float *)(dst_addr + 9 * dst_stride_z) = out13; |
| 416 | *(__global float *)(dst_addr + 10 * dst_stride_z) = out14; |
| 417 | *(__global float *)(dst_addr + 11 * dst_stride_z) = out15; |
| 418 | *(__global float *)(dst_addr + 12 * dst_stride_z) = out20; |
| 419 | *(__global float *)(dst_addr + 13 * dst_stride_z) = out21; |
| 420 | *(__global float *)(dst_addr + 14 * dst_stride_z) = out22; |
| 421 | *(__global float *)(dst_addr + 15 * dst_stride_z) = out23; |
| 422 | *(__global float *)(dst_addr + 16 * dst_stride_z) = out24; |
| 423 | *(__global float *)(dst_addr + 17 * dst_stride_z) = out25; |
| 424 | *(__global float *)(dst_addr + 18 * dst_stride_z) = out30; |
| 425 | *(__global float *)(dst_addr + 19 * dst_stride_z) = out31; |
| 426 | *(__global float *)(dst_addr + 20 * dst_stride_z) = out32; |
| 427 | *(__global float *)(dst_addr + 21 * dst_stride_z) = out33; |
| 428 | *(__global float *)(dst_addr + 22 * dst_stride_z) = out34; |
| 429 | *(__global float *)(dst_addr + 23 * dst_stride_z) = out35; |
| 430 | *(__global float *)(dst_addr + 24 * dst_stride_z) = out40; |
| 431 | *(__global float *)(dst_addr + 25 * dst_stride_z) = out41; |
| 432 | *(__global float *)(dst_addr + 26 * dst_stride_z) = out42; |
| 433 | *(__global float *)(dst_addr + 27 * dst_stride_z) = out43; |
| 434 | *(__global float *)(dst_addr + 28 * dst_stride_z) = out44; |
| 435 | *(__global float *)(dst_addr + 29 * dst_stride_z) = out45; |
| 436 | *(__global float *)(dst_addr + 30 * dst_stride_z) = out50; |
| 437 | *(__global float *)(dst_addr + 31 * dst_stride_z) = out51; |
| 438 | *(__global float *)(dst_addr + 32 * dst_stride_z) = out52; |
| 439 | *(__global float *)(dst_addr + 33 * dst_stride_z) = out53; |
| 440 | *(__global float *)(dst_addr + 34 * dst_stride_z) = out54; |
| 441 | *(__global float *)(dst_addr + 35 * dst_stride_z) = out55; |
| Giorgio Arena | 149fdf3 | 2018-07-04 17:03:33 +0100 | [diff] [blame] | 442 | #endif // !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) |
| Giorgio Arena | a50e5e0 | 2018-07-02 13:42:23 +0100 | [diff] [blame] | 443 | } |
| Giorgio Arena | 149fdf3 | 2018-07-04 17:03:33 +0100 | [diff] [blame] | 444 | |
| Gian Marco Iodice | 876be2a | 2018-07-03 12:22:09 +0100 | [diff] [blame] | 445 | /** This OpenCL kernel performs Winograd filter transform 5x5/5x1 or 1x5 when the data layout is NCHW and the output tile is 4x4/4x1 or 1x4 |
| Giorgio Arena | a50e5e0 | 2018-07-02 13:42:23 +0100 | [diff] [blame] | 446 | * |
| 447 | * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64 |
| 448 | * |
| Gian Marco Iodice | 876be2a | 2018-07-03 12:22:09 +0100 | [diff] [blame] | 449 | * @note If this kernel is used to perform Winograd filter transform 5x1, -DWINOGRAD_FILTER_TRANSFORM_HORIZONTAL has to be passed at compile time |
| 450 | * @note If this kernel is used to perform Winograd filter transform 1x5, -DWINOGRAD_FILTER_TRANSFORM_VERTICAL has to be passed at compile time |
| 451 | * |
| Giorgio Arena | a50e5e0 | 2018-07-02 13:42:23 +0100 | [diff] [blame] | 452 | * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 |
| 453 | * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) |
| 454 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 455 | * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 456 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 457 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 458 | * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 459 | * @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes) |
| 460 | * @param[in] src_step_w src_stride_w * number of elements along W processed per workitem(in bytes) |
| 461 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor |
| 462 | * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr |
| 463 | * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 464 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 465 | * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 466 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 467 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 468 | * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 469 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 470 | */ |
| 471 | __kernel void winograd_filter_transform_4x4_5x5_nchw( |
| 472 | TENSOR4D_DECLARATION(src), |
| 473 | TENSOR3D_DECLARATION(dst)) |
| 474 | { |
| 475 | Tensor4D src = CONVERT_TO_TENSOR4D_STRUCT(src, SRC_DIM_Z); |
| 476 | |
| 477 | const __global uchar *src_addr = tensor4D_offset(&src, 0, 0, 0, 0); |
| 478 | |
| 479 | // Load the values from the input tensor |
| Gian Marco Iodice | 876be2a | 2018-07-03 12:22:09 +0100 | [diff] [blame] | 480 | #if defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) |
| 481 | float4 w00 = vload4(0, (__global float *)(src_addr + 0 * src_stride_y)); |
| 482 | float w01 = *((__global float *)(src_addr + 0 * src_stride_y) + 4); |
| 483 | #elif defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) |
| 484 | float4 w00 = (float4)(*((__global float *)(src_addr + 0 * src_stride_y)), |
| 485 | *((__global float *)(src_addr + 1 * src_stride_y)), |
| 486 | *((__global float *)(src_addr + 2 * src_stride_y)), |
| 487 | *((__global float *)(src_addr + 3 * src_stride_y))); |
| 488 | float w01 = *((__global float *)(src_addr + 4 * src_stride_y)); |
| 489 | #else // defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) |
| 490 | float4 w00 = vload4(0, (__global float *)(src_addr + 0 * src_stride_y)); |
| 491 | float w01 = *((__global float *)(src_addr + 0 * src_stride_y) + 4); |
| 492 | float4 w10 = vload4(0, (__global float *)(src_addr + 1 * src_stride_y)); |
| 493 | float w11 = *((__global float *)(src_addr + 1 * src_stride_y) + 4); |
| 494 | float4 w20 = vload4(0, (__global float *)(src_addr + 2 * src_stride_y)); |
| 495 | float w21 = *((__global float *)(src_addr + 2 * src_stride_y) + 4); |
| 496 | float4 w30 = vload4(0, (__global float *)(src_addr + 3 * src_stride_y)); |
| 497 | float w31 = *((__global float *)(src_addr + 3 * src_stride_y) + 4); |
| 498 | float4 w40 = vload4(0, (__global float *)(src_addr + 4 * src_stride_y)); |
| 499 | float w41 = *((__global float *)(src_addr + 4 * src_stride_y) + 4); |
| 500 | #endif // defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) |
| Giorgio Arena | a50e5e0 | 2018-07-02 13:42:23 +0100 | [diff] [blame] | 501 | |
| Gian Marco Iodice | 876be2a | 2018-07-03 12:22:09 +0100 | [diff] [blame] | 502 | // Transform the input tile |
| Giorgio Arena | a50e5e0 | 2018-07-02 13:42:23 +0100 | [diff] [blame] | 503 | |
| 504 | // Row 0 |
| Gian Marco Iodice | 876be2a | 2018-07-03 12:22:09 +0100 | [diff] [blame] | 505 | float8 out0 = 0.0f; |
| 506 | out0.s0 = w00.s0; |
| 507 | out0.s1 = -2.f * (w00.s0 + w00.s1 + w00.s2 + w00.s3 + w01) / 9.f; |
| 508 | out0.s2 = -2.f * (w00.s0 - w00.s1 + w00.s2 - w00.s3 + w01) / 9.f; |
| 509 | out0.s3 = (w00.s0 + 2.f * w00.s1 + 4.f * w00.s2 + 8.f * w00.s3 + 16.f * w01) / 90.f; |
| 510 | out0.s4 = (w00.s0 - 2.f * w00.s1 + 4.f * w00.s2 - 8.f * w00.s3 + 16.f * w01) / 90.f; |
| 511 | out0.s5 = (16.f * w00.s0 + 8.f * w00.s1 + 4.f * w00.s2 + 2.f * w00.s3 + w01) / 180.f; |
| 512 | out0.s6 = (16.f * w00.s0 - 8.f * w00.s1 + 4.f * w00.s2 - 2.f * w00.s3 + w01) / 180.f; |
| 513 | out0.s7 = w01; |
| Giorgio Arena | a50e5e0 | 2018-07-02 13:42:23 +0100 | [diff] [blame] | 514 | |
| Gian Marco Iodice | 876be2a | 2018-07-03 12:22:09 +0100 | [diff] [blame] | 515 | #if !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) |
| Giorgio Arena | a50e5e0 | 2018-07-02 13:42:23 +0100 | [diff] [blame] | 516 | // Row 1 |
| Gian Marco Iodice | 876be2a | 2018-07-03 12:22:09 +0100 | [diff] [blame] | 517 | float8 out1 = 0.0f; |
| 518 | out1.s0 = -2.f * (w00.s0 + w10.s0 + w20.s0 + w30.s0 + w40.s0) / 9.f; |
| 519 | out1.s1 = 4.f * ((w00.s0 + w10.s0 + w20.s0 + w30.s0 + w40.s0) + (w00.s1 + w10.s1 + w20.s1 + w30.s1 + w40.s1) + (w00.s2 + w10.s2 + w20.s2 + w30.s2 + w40.s2) + |
| 520 | (w00.s3 + w10.s3 + w20.s3 + w30.s3 + w40.s3) + (w01 + w11 + w21 + w31 + w41)) / 81.f; |
| 521 | out1.s2 = 4.f * ((w00.s0 + w10.s0 + w20.s0 + w30.s0 + w40.s0) - (w00.s1 + w10.s1 + w20.s1 + w30.s1 + w40.s1) + (w00.s2 + w10.s2 + w20.s2 + w30.s2 + w40.s2) - |
| 522 | (w00.s3 + w10.s3 + w20.s3 + w30.s3 + w40.s3) + (w01 + w11 + w21 + w31 + w41)) / 81.f; |
| 523 | out1.s3 = -((w00.s0 + w10.s0 + w20.s0 + w30.s0 + w40.s0) + 2.f * (w00.s1 + w10.s1 + w20.s1 + w30.s1 + w40.s1) + 4.f * (w00.s2 + w10.s2 + w20.s2 + w30.s2 + w40.s2) + 8.f * |
| 524 | (w00.s3 + w10.s3 + w20.s3 + w30.s3 + w40.s3) + 16.f * (w01 + w11 + w21 + w31 + w41)) / 405.f; |
| 525 | out1.s4 = -((w00.s0 + w10.s0 + w20.s0 + w30.s0 + w40.s0) - 2.f * (w00.s1 + w10.s1 + w20.s1 + w30.s1 + w40.s1) + 4.f * (w00.s2 + w10.s2 + w20.s2 + w30.s2 + w40.s2) - 8.f * |
| 526 | (w00.s3 + w10.s3 + w20.s3 + w30.s3 + w40.s3) + 16.f * (w01 + w11 + w21 + w31 + w41)) / 405.f; |
| 527 | out1.s5 = -(16.f * (w00.s0 + w10.s0 + w20.s0 + w30.s0 + w40.s0) + 8.f * (w00.s1 + w10.s1 + w20.s1 + w30.s1 + w40.s1) + 4.f * (w00.s2 + w10.s2 + w20.s2 + w30.s2 + w40.s2) + 2.f * |
| 528 | (w00.s3 + w10.s3 + w20.s3 + w30.s3 + w40.s3) + (w01 + w11 + w21 + w31 + w41)) / 810.f; |
| 529 | out1.s6 = -(16.f * (w00.s0 + w10.s0 + w20.s0 + w30.s0 + w40.s0) - 8.f * (w00.s1 + w10.s1 + w20.s1 + w30.s1 + w40.s1) + 4.f * (w00.s2 + w10.s2 + w20.s2 + w30.s2 + w40.s2) - 2.f * |
| 530 | (w00.s3 + w10.s3 + w20.s3 + w30.s3 + w40.s3) + (w01 + w11 + w21 + w31 + w41)) / 810.f; |
| 531 | out1.s7 = -2.f * (w01 + w11 + w21 + w31 + w41) / 9.f; |
| Giorgio Arena | a50e5e0 | 2018-07-02 13:42:23 +0100 | [diff] [blame] | 532 | |
| 533 | // Row 2 |
| Gian Marco Iodice | 876be2a | 2018-07-03 12:22:09 +0100 | [diff] [blame] | 534 | float8 out2 = 0.0f; |
| 535 | out2.s0 = -2.f * (w00.s0 - w10.s0 + w20.s0 - w30.s0 + w40.s0) / 9.f; |
| 536 | out2.s1 = 4.f * ((w00.s0 - w10.s0 + w20.s0 - w30.s0 + w40.s0) + (w00.s1 - w10.s1 + w20.s1 - w30.s1 + w40.s1) + (w00.s2 - w10.s2 + w20.s2 - w30.s2 + w40.s2) + |
| 537 | (w00.s3 - w10.s3 + w20.s3 - w30.s3 + w40.s3) + (w01 - w11 + w21 - w31 + w41)) / 81.f; |
| 538 | out2.s2 = 4.f * ((w00.s0 - w10.s0 + w20.s0 - w30.s0 + w40.s0) - (w00.s1 - w10.s1 + w20.s1 - w30.s1 + w40.s1) + (w00.s2 - w10.s2 + w20.s2 - w30.s2 + w40.s2) - |
| 539 | (w00.s3 - w10.s3 + w20.s3 - w30.s3 + w40.s3) + (w01 - w11 + w21 - w31 + w41)) / 81.f; |
| 540 | out2.s3 = -((w00.s0 - w10.s0 + w20.s0 - w30.s0 + w40.s0) + 2.f * (w00.s1 - w10.s1 + w20.s1 - w30.s1 + w40.s1) + 4.f * (w00.s2 - w10.s2 + w20.s2 - w30.s2 + w40.s2) + 8.f * |
| 541 | (w00.s3 - w10.s3 + w20.s3 - w30.s3 + w40.s3) + 16.f * (w01 - w11 + w21 - w31 + w41)) / 405.f; |
| 542 | out2.s4 = -((w00.s0 - w10.s0 + w20.s0 - w30.s0 + w40.s0) - 2.f * (w00.s1 - w10.s1 + w20.s1 - w30.s1 + w40.s1) + 4.f * (w00.s2 - w10.s2 + w20.s2 - w30.s2 + w40.s2) - 8.f * |
| 543 | (w00.s3 - w10.s3 + w20.s3 - w30.s3 + w40.s3) + 16.f * (w01 - w11 + w21 - w31 + w41)) / 405.f; |
| 544 | out2.s5 = -(16.f * (w00.s0 - w10.s0 + w20.s0 - w30.s0 + w40.s0) + 8.f * (w00.s1 - w10.s1 + w20.s1 - w30.s1 + w40.s1) + 4.f * (w00.s2 - w10.s2 + w20.s2 - w30.s2 + w40.s2) + 2.f * |
| 545 | (w00.s3 - w10.s3 + w20.s3 - w30.s3 + w40.s3) + (w01 - w11 + w21 - w31 + w41)) / 810.f; |
| 546 | out2.s6 = -(16.f * (w00.s0 - w10.s0 + w20.s0 - w30.s0 + w40.s0) - 8.f * (w00.s1 - w10.s1 + w20.s1 - w30.s1 + w40.s1) + 4.f * (w00.s2 - w10.s2 + w20.s2 - w30.s2 + w40.s2) - 2.f * |
| 547 | (w00.s3 - w10.s3 + w20.s3 - w30.s3 + w40.s3) + (w01 - w11 + w21 - w31 + w41)) / 810.f; |
| 548 | out2.s7 = -2.f * (w01 - w11 + w21 - w31 + w41) / 9.f; |
| Giorgio Arena | a50e5e0 | 2018-07-02 13:42:23 +0100 | [diff] [blame] | 549 | |
| 550 | // Row 3 |
| Gian Marco Iodice | 876be2a | 2018-07-03 12:22:09 +0100 | [diff] [blame] | 551 | float8 out3 = 0.0f; |
| 552 | out3.s0 = (w00.s0 + 2.f * w10.s0 + 4.f * w20.s0 + 8.f * w30.s0 + 16.f * w40.s0) / 90.f; |
| 553 | out3.s1 = -((w00.s0 + 2.f * w10.s0 + 4.f * w20.s0 + 8.f * w30.s0 + 16.f * w40.s0) + (w00.s1 + 2.f * w10.s1 + 4.f * w20.s1 + 8.f * w30.s1 + 16.f * w40.s1) + |
| 554 | (w00.s2 + 2.f * w10.s2 + 4.f * w20.s2 + 8.f * w30.s2 + 16.f * w40.s2) + (w00.s3 + 2.f * w10.s3 + 4.f * w20.s3 + 8.f * w30.s3 + 16.f * w40.s3) + |
| 555 | (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41)) / 405.f; |
| 556 | out3.s2 = -((w00.s0 + 2.f * w10.s0 + 4.f * w20.s0 + 8.f * w30.s0 + 16.f * w40.s0) - (w00.s1 + 2.f * w10.s1 + 4.f * w20.s1 + 8.f * w30.s1 + 16.f * w40.s1) + |
| 557 | (w00.s2 + 2.f * w10.s2 + 4.f * w20.s2 + 8.f * w30.s2 + 16.f * w40.s2) - (w00.s3 + 2.f * w10.s3 + 4.f * w20.s3 + 8.f * w30.s3 + 16.f * w40.s3) + |
| 558 | (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41)) / 405.f; |
| 559 | out3.s3 = ((w00.s0 + 2.f * w10.s0 + 4.f * w20.s0 + 8.f * w30.s0 + 16.f * w40.s0) + 2.f * (w00.s1 + 2.f * w10.s1 + 4.f * w20.s1 + 8.f * w30.s1 + 16.f * w40.s1) + 4.f * |
| 560 | (w00.s2 + 2.f * w10.s2 + 4.f * w20.s2 + 8.f * w30.s2 + 16.f * w40.s2) + 8.f * (w00.s3 + 2.f * w10.s3 + 4.f * w20.s3 + 8.f * w30.s3 + 16.f * w40.s3) + 16.f * |
| 561 | (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41)) / 8100.f; |
| 562 | out3.s4 = ((w00.s0 + 2.f * w10.s0 + 4.f * w20.s0 + 8.f * w30.s0 + 16.f * w40.s0) - 2.f * (w00.s1 + 2.f * w10.s1 + 4.f * w20.s1 + 8.f * w30.s1 + 16.f * w40.s1) + 4.f * |
| 563 | (w00.s2 + 2.f * w10.s2 + 4.f * w20.s2 + 8.f * w30.s2 + 16.f * w40.s2) - 8.f * (w00.s3 + 2.f * w10.s3 + 4.f * w20.s3 + 8.f * w30.s3 + 16.f * w40.s3) + 16.f * |
| 564 | (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41)) / 8100.f; |
| 565 | out3.s5 = (16.f * (w00.s0 + 2.f * w10.s0 + 4.f * w20.s0 + 8.f * w30.s0 + 16.f * w40.s0) + 8.f * (w00.s1 + 2.f * w10.s1 + 4.f * w20.s1 + 8.f * w30.s1 + 16.f * w40.s1) + 4.f * |
| 566 | (w00.s2 + 2.f * w10.s2 + 4.f * w20.s2 + 8.f * w30.s2 + 16.f * w40.s2) + 2.f * (w00.s3 + 2.f * w10.s3 + 4.f * w20.s3 + 8.f * w30.s3 + 16.f * w40.s3) + |
| 567 | (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41)) / 16200.f; |
| 568 | out3.s6 = (16.f * (w00.s0 + 2.f * w10.s0 + 4.f * w20.s0 + 8.f * w30.s0 + 16.f * w40.s0) - 8.f * (w00.s1 + 2.f * w10.s1 + 4.f * w20.s1 + 8.f * w30.s1 + 16.f * w40.s1) + 4.f * |
| 569 | (w00.s2 + 2.f * w10.s2 + 4.f * w20.s2 + 8.f * w30.s2 + 16.f * w40.s2) - 2.f * (w00.s3 + 2.f * w10.s3 + 4.f * w20.s3 + 8.f * w30.s3 + 16.f * w40.s3) + |
| 570 | (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41)) / 16200.f; |
| 571 | out3.s7 = (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41) / 90.f; |
| Giorgio Arena | a50e5e0 | 2018-07-02 13:42:23 +0100 | [diff] [blame] | 572 | |
| 573 | // Row 4 |
| Gian Marco Iodice | 876be2a | 2018-07-03 12:22:09 +0100 | [diff] [blame] | 574 | float8 out4 = 0.0f; |
| 575 | out4.s0 = (w00.s0 - 2.f * w10.s0 + 4.f * w20.s0 - 8.f * w30.s0 + 16.f * w40.s0) / 90.f; |
| 576 | out4.s1 = -((w00.s0 - 2.f * w10.s0 + 4.f * w20.s0 - 8.f * w30.s0 + 16.f * w40.s0) + (w00.s1 - 2.f * w10.s1 + 4.f * w20.s1 - 8.f * w30.s1 + 16.f * w40.s1) + |
| 577 | (w00.s2 - 2.f * w10.s2 + 4.f * w20.s2 - 8.f * w30.s2 + 16.f * w40.s2) + (w00.s3 - 2.f * w10.s3 + 4.f * w20.s3 - 8.f * w30.s3 + 16.f * w40.s3) + |
| 578 | (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41)) / 405.f; |
| 579 | out4.s2 = -((w00.s0 - 2.f * w10.s0 + 4.f * w20.s0 - 8.f * w30.s0 + 16.f * w40.s0) - (w00.s1 - 2.f * w10.s1 + 4.f * w20.s1 - 8.f * w30.s1 + 16.f * w40.s1) + |
| 580 | (w00.s2 - 2.f * w10.s2 + 4.f * w20.s2 - 8.f * w30.s2 + 16.f * w40.s2) - (w00.s3 - 2.f * w10.s3 + 4.f * w20.s3 - 8.f * w30.s3 + 16.f * w40.s3) + |
| 581 | (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41)) / 405.f; |
| 582 | out4.s3 = ((w00.s0 - 2.f * w10.s0 + 4.f * w20.s0 - 8.f * w30.s0 + 16.f * w40.s0) + 2.f * (w00.s1 - 2.f * w10.s1 + 4.f * w20.s1 - 8.f * w30.s1 + 16.f * w40.s1) + 4.f * |
| 583 | (w00.s2 - 2.f * w10.s2 + 4.f * w20.s2 - 8.f * w30.s2 + 16.f * w40.s2) + 8.f * (w00.s3 - 2.f * w10.s3 + 4.f * w20.s3 - 8.f * w30.s3 + 16.f * w40.s3) + 16.f * |
| 584 | (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41)) / 8100.f; |
| 585 | out4.s4 = ((w00.s0 - 2.f * w10.s0 + 4.f * w20.s0 - 8.f * w30.s0 + 16.f * w40.s0) - 2.f * (w00.s1 - 2.f * w10.s1 + 4.f * w20.s1 - 8.f * w30.s1 + 16.f * w40.s1) + 4.f * |
| 586 | (w00.s2 - 2.f * w10.s2 + 4.f * w20.s2 - 8.f * w30.s2 + 16.f * w40.s2) - 8.f * (w00.s3 - 2.f * w10.s3 + 4.f * w20.s3 - 8.f * w30.s3 + 16.f * w40.s3) + 16.f * |
| 587 | (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41)) / 8100.f; |
| 588 | out4.s5 = (16.f * (w00.s0 - 2.f * w10.s0 + 4.f * w20.s0 - 8.f * w30.s0 + 16.f * w40.s0) + 8.f * (w00.s1 - 2.f * w10.s1 + 4.f * w20.s1 - 8.f * w30.s1 + 16.f * w40.s1) + 4.f * |
| 589 | (w00.s2 - 2.f * w10.s2 + 4.f * w20.s2 - 8.f * w30.s2 + 16.f * w40.s2) + 2.f * (w00.s3 - 2.f * w10.s3 + 4.f * w20.s3 - 8.f * w30.s3 + 16.f * w40.s3) + |
| 590 | (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41)) / 16200.f; |
| 591 | out4.s6 = (16.f * (w00.s0 - 2.f * w10.s0 + 4.f * w20.s0 - 8.f * w30.s0 + 16.f * w40.s0) - 8.f * (w00.s1 - 2.f * w10.s1 + 4.f * w20.s1 - 8.f * w30.s1 + 16.f * w40.s1) + 4.f * |
| 592 | (w00.s2 - 2.f * w10.s2 + 4.f * w20.s2 - 8.f * w30.s2 + 16.f * w40.s2) - 2.f * (w00.s3 - 2.f * w10.s3 + 4.f * w20.s3 - 8.f * w30.s3 + 16.f * w40.s3) + |
| 593 | (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41)) / 16200.f; |
| 594 | out4.s7 = (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41) / 90.f; |
| Giorgio Arena | a50e5e0 | 2018-07-02 13:42:23 +0100 | [diff] [blame] | 595 | |
| 596 | // Row 5 |
| Gian Marco Iodice | 876be2a | 2018-07-03 12:22:09 +0100 | [diff] [blame] | 597 | float8 out5 = 0.0f; |
| 598 | out5.s0 = (16.f * w00.s0 + 8.f * w10.s0 + 4.f * w20.s0 + 2.f * w30.s0 + w40.s0) / 180.f; |
| 599 | out5.s1 = -((16.f * w00.s0 + 8.f * w10.s0 + 4.f * w20.s0 + 2.f * w30.s0 + w40.s0) + (16.f * w00.s1 + 8.f * w10.s1 + 4.f * w20.s1 + 2.f * w30.s1 + w40.s1) + |
| 600 | (16.f * w00.s2 + 8.f * w10.s2 + 4.f * w20.s2 + 2.f * w30.s2 + w40.s2) + (16.f * w00.s3 + 8.f * w10.s3 + 4.f * w20.s3 + 2.f * w30.s3 + w40.s3) + |
| 601 | (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41)) / 810.f; |
| 602 | out5.s2 = -((16.f * w00.s0 + 8.f * w10.s0 + 4.f * w20.s0 + 2.f * w30.s0 + w40.s0) - (16.f * w00.s1 + 8.f * w10.s1 + 4.f * w20.s1 + 2.f * w30.s1 + w40.s1) + |
| 603 | (16.f * w00.s2 + 8.f * w10.s2 + 4.f * w20.s2 + 2.f * w30.s2 + w40.s2) - (16.f * w00.s3 + 8.f * w10.s3 + 4.f * w20.s3 + 2.f * w30.s3 + w40.s3) + |
| 604 | (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41)) / 810.f; |
| 605 | out5.s3 = ((16.f * w00.s0 + 8.f * w10.s0 + 4.f * w20.s0 + 2.f * w30.s0 + w40.s0) + 2.f * (16.f * w00.s1 + 8.f * w10.s1 + 4.f * w20.s1 + 2.f * w30.s1 + w40.s1) + 4.f * |
| 606 | (16.f * w00.s2 + 8.f * w10.s2 + 4.f * w20.s2 + 2.f * w30.s2 + w40.s2) + 8.f * (16.f * w00.s3 + 8.f * w10.s3 + 4.f * w20.s3 + 2.f * w30.s3 + w40.s3) + 16.f * |
| 607 | (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41)) / 16200.f; |
| 608 | out5.s4 = ((16.f * w00.s0 + 8.f * w10.s0 + 4.f * w20.s0 + 2.f * w30.s0 + w40.s0) - 2.f * (16.f * w00.s1 + 8.f * w10.s1 + 4.f * w20.s1 + 2.f * w30.s1 + w40.s1) + 4.f * |
| 609 | (16.f * w00.s2 + 8.f * w10.s2 + 4.f * w20.s2 + 2.f * w30.s2 + w40.s2) - 8.f * (16.f * w00.s3 + 8.f * w10.s3 + 4.f * w20.s3 + 2.f * w30.s3 + w40.s3) + 16.f * |
| 610 | (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41)) / 16200.f; |
| 611 | out5.s5 = (16.f * (16.f * w00.s0 + 8.f * w10.s0 + 4.f * w20.s0 + 2.f * w30.s0 + w40.s0) + 8.f * (16.f * w00.s1 + 8.f * w10.s1 + 4.f * w20.s1 + 2.f * w30.s1 + w40.s1) + 4.f * |
| 612 | (16.f * w00.s2 + 8.f * w10.s2 + 4.f * w20.s2 + 2.f * w30.s2 + w40.s2) + 2.f * (16.f * w00.s3 + 8.f * w10.s3 + 4.f * w20.s3 + 2.f * w30.s3 + w40.s3) + |
| 613 | (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41)) / 32400.f; |
| 614 | out5.s6 = (16.f * (16.f * w00.s0 + 8.f * w10.s0 + 4.f * w20.s0 + 2.f * w30.s0 + w40.s0) - 8.f * (16.f * w00.s1 + 8.f * w10.s1 + 4.f * w20.s1 + 2.f * w30.s1 + w40.s1) + 4.f * |
| 615 | (16.f * w00.s2 + 8.f * w10.s2 + 4.f * w20.s2 + 2.f * w30.s2 + w40.s2) - 2.f * (16.f * w00.s3 + 8.f * w10.s3 + 4.f * w20.s3 + 2.f * w30.s3 + w40.s3) + |
| 616 | (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41)) / 32400.f; |
| 617 | out5.s7 = (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41) / 180.f; |
| Giorgio Arena | a50e5e0 | 2018-07-02 13:42:23 +0100 | [diff] [blame] | 618 | |
| 619 | // Row 6 |
| Gian Marco Iodice | 876be2a | 2018-07-03 12:22:09 +0100 | [diff] [blame] | 620 | float8 out6 = 0.0f; |
| 621 | out6.s0 = (16.f * w00.s0 - 8.f * w10.s0 + 4.f * w20.s0 - 2.f * w30.s0 + w40.s0) / 180.f; |
| 622 | out6.s1 = -((16.f * w00.s0 - 8.f * w10.s0 + 4.f * w20.s0 - 2.f * w30.s0 + w40.s0) + (16.f * w00.s1 - 8.f * w10.s1 + 4.f * w20.s1 - 2.f * w30.s1 + w40.s1) + |
| 623 | (16.f * w00.s2 - 8.f * w10.s2 + 4.f * w20.s2 - 2.f * w30.s2 + w40.s2) + (16.f * w00.s3 - 8.f * w10.s3 + 4.f * w20.s3 - 2.f * w30.s3 + w40.s3) + |
| 624 | (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41)) / 810.f; |
| 625 | out6.s2 = -((16.f * w00.s0 - 8.f * w10.s0 + 4.f * w20.s0 - 2.f * w30.s0 + w40.s0) - (16.f * w00.s1 - 8.f * w10.s1 + 4.f * w20.s1 - 2.f * w30.s1 + w40.s1) + |
| 626 | (16.f * w00.s2 - 8.f * w10.s2 + 4.f * w20.s2 - 2.f * w30.s2 + w40.s2) - (16.f * w00.s3 - 8.f * w10.s3 + 4.f * w20.s3 - 2.f * w30.s3 + w40.s3) + |
| 627 | (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41)) / 810.f; |
| 628 | out6.s3 = ((16.f * w00.s0 - 8.f * w10.s0 + 4.f * w20.s0 - 2.f * w30.s0 + w40.s0) + 2.f * (16.f * w00.s1 - 8.f * w10.s1 + 4.f * w20.s1 - 2.f * w30.s1 + w40.s1) + 4.f * |
| 629 | (16.f * w00.s2 - 8.f * w10.s2 + 4.f * w20.s2 - 2.f * w30.s2 + w40.s2) + 8.f * (16.f * w00.s3 - 8.f * w10.s3 + 4.f * w20.s3 - 2.f * w30.s3 + w40.s3) + 16.f * |
| 630 | (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41)) / 16200.f; |
| 631 | out6.s4 = ((16.f * w00.s0 - 8.f * w10.s0 + 4.f * w20.s0 - 2.f * w30.s0 + w40.s0) - 2.f * (16.f * w00.s1 - 8.f * w10.s1 + 4.f * w20.s1 - 2.f * w30.s1 + w40.s1) + 4.f * |
| 632 | (16.f * w00.s2 - 8.f * w10.s2 + 4.f * w20.s2 - 2.f * w30.s2 + w40.s2) - 8.f * (16.f * w00.s3 - 8.f * w10.s3 + 4.f * w20.s3 - 2.f * w30.s3 + w40.s3) + 16.f * |
| 633 | (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41)) / 16200.f; |
| 634 | out6.s5 = (16.f * (16.f * w00.s0 - 8.f * w10.s0 + 4.f * w20.s0 - 2.f * w30.s0 + w40.s0) + 8.f * (16.f * w00.s1 - 8.f * w10.s1 + 4.f * w20.s1 - 2.f * w30.s1 + w40.s1) + 4.f * |
| 635 | (16.f * w00.s2 - 8.f * w10.s2 + 4.f * w20.s2 - 2.f * w30.s2 + w40.s2) + 2.f * (16.f * w00.s3 - 8.f * w10.s3 + 4.f * w20.s3 - 2.f * w30.s3 + w40.s3) + |
| 636 | (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41)) / 32400.f; |
| 637 | out6.s6 = (16.f * (16.f * w00.s0 - 8.f * w10.s0 + 4.f * w20.s0 - 2.f * w30.s0 + w40.s0) - 8.f * (16.f * w00.s1 - 8.f * w10.s1 + 4.f * w20.s1 - 2.f * w30.s1 + w40.s1) + 4.f * |
| 638 | (16.f * w00.s2 - 8.f * w10.s2 + 4.f * w20.s2 - 2.f * w30.s2 + w40.s2) - 2.f * (16.f * w00.s3 - 8.f * w10.s3 + 4.f * w20.s3 - 2.f * w30.s3 + w40.s3) + |
| 639 | (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41)) / 32400.f; |
| 640 | out6.s7 = (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41) / 180.f; |
| Giorgio Arena | a50e5e0 | 2018-07-02 13:42:23 +0100 | [diff] [blame] | 641 | |
| 642 | // Row 7 |
| Gian Marco Iodice | 876be2a | 2018-07-03 12:22:09 +0100 | [diff] [blame] | 643 | float8 out7 = 0.0f; |
| 644 | out7.s0 = w40.s0; |
| 645 | out7.s1 = -2.f * (w40.s0 + w40.s1 + w40.s2 + w40.s3 + w41) / 9.f; |
| 646 | out7.s2 = -2.f * (w40.s0 - w40.s1 + w40.s2 - w40.s3 + w41) / 9.f; |
| 647 | out7.s3 = (w40.s0 + 2.f * w40.s1 + 4.f * w40.s2 + 8.f * w40.s3 + 16.f * w41) / 90.f; |
| 648 | out7.s4 = (w40.s0 - 2.f * w40.s1 + 4.f * w40.s2 - 8.f * w40.s3 + 16.f * w41) / 90.f; |
| 649 | out7.s5 = (16.f * w40.s0 + 8.f * w40.s1 + 4.f * w40.s2 + 2.f * w40.s3 + w41) / 180.f; |
| 650 | out7.s6 = (16.f * w40.s0 - 8.f * w40.s1 + 4.f * w40.s2 - 2.f * w40.s3 + w41) / 180.f; |
| 651 | out7.s7 = w41; |
| 652 | #endif // !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) |
| Giorgio Arena | a50e5e0 | 2018-07-02 13:42:23 +0100 | [diff] [blame] | 653 | |
| 654 | int z = get_global_id(2); |
| 655 | int x0 = z / SRC_DIM_Z; // idx filter |
| 656 | int y0 = z % SRC_DIM_Z; // idx channel |
| 657 | |
| 658 | // Get output address |
| Gian Marco Iodice | 876be2a | 2018-07-03 12:22:09 +0100 | [diff] [blame] | 659 | __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x0 * sizeof(float) + y0 * dst_stride_y; |
| Giorgio Arena | a50e5e0 | 2018-07-02 13:42:23 +0100 | [diff] [blame] | 660 | |
| Gian Marco Iodice | 876be2a | 2018-07-03 12:22:09 +0100 | [diff] [blame] | 661 | // Store the values across the channels |
| 662 | *(__global float *)(dst_addr + 0 * dst_stride_z) = out0.s0; |
| 663 | *(__global float *)(dst_addr + 1 * dst_stride_z) = out0.s1; |
| 664 | *(__global float *)(dst_addr + 2 * dst_stride_z) = out0.s2; |
| 665 | *(__global float *)(dst_addr + 3 * dst_stride_z) = out0.s3; |
| 666 | *(__global float *)(dst_addr + 4 * dst_stride_z) = out0.s4; |
| 667 | *(__global float *)(dst_addr + 5 * dst_stride_z) = out0.s5; |
| 668 | *(__global float *)(dst_addr + 6 * dst_stride_z) = out0.s6; |
| 669 | *(__global float *)(dst_addr + 7 * dst_stride_z) = out0.s7; |
| 670 | |
| 671 | #if !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) |
| Giorgio Arena | a50e5e0 | 2018-07-02 13:42:23 +0100 | [diff] [blame] | 672 | *(__global float *)(dst_addr + 8 * dst_stride_z) = out1.s0; |
| 673 | *(__global float *)(dst_addr + 9 * dst_stride_z) = out1.s1; |
| 674 | *(__global float *)(dst_addr + 10 * dst_stride_z) = out1.s2; |
| 675 | *(__global float *)(dst_addr + 11 * dst_stride_z) = out1.s3; |
| 676 | *(__global float *)(dst_addr + 12 * dst_stride_z) = out1.s4; |
| 677 | *(__global float *)(dst_addr + 13 * dst_stride_z) = out1.s5; |
| 678 | *(__global float *)(dst_addr + 14 * dst_stride_z) = out1.s6; |
| 679 | *(__global float *)(dst_addr + 15 * dst_stride_z) = out1.s7; |
| 680 | *(__global float *)(dst_addr + 16 * dst_stride_z) = out2.s0; |
| 681 | *(__global float *)(dst_addr + 17 * dst_stride_z) = out2.s1; |
| 682 | *(__global float *)(dst_addr + 18 * dst_stride_z) = out2.s2; |
| 683 | *(__global float *)(dst_addr + 19 * dst_stride_z) = out2.s3; |
| 684 | *(__global float *)(dst_addr + 20 * dst_stride_z) = out2.s4; |
| 685 | *(__global float *)(dst_addr + 21 * dst_stride_z) = out2.s5; |
| 686 | *(__global float *)(dst_addr + 22 * dst_stride_z) = out2.s6; |
| 687 | *(__global float *)(dst_addr + 23 * dst_stride_z) = out2.s7; |
| 688 | *(__global float *)(dst_addr + 24 * dst_stride_z) = out3.s0; |
| 689 | *(__global float *)(dst_addr + 25 * dst_stride_z) = out3.s1; |
| 690 | *(__global float *)(dst_addr + 26 * dst_stride_z) = out3.s2; |
| 691 | *(__global float *)(dst_addr + 27 * dst_stride_z) = out3.s3; |
| 692 | *(__global float *)(dst_addr + 28 * dst_stride_z) = out3.s4; |
| 693 | *(__global float *)(dst_addr + 29 * dst_stride_z) = out3.s5; |
| 694 | *(__global float *)(dst_addr + 30 * dst_stride_z) = out3.s6; |
| 695 | *(__global float *)(dst_addr + 31 * dst_stride_z) = out3.s7; |
| 696 | *(__global float *)(dst_addr + 32 * dst_stride_z) = out4.s0; |
| 697 | *(__global float *)(dst_addr + 33 * dst_stride_z) = out4.s1; |
| 698 | *(__global float *)(dst_addr + 34 * dst_stride_z) = out4.s2; |
| 699 | *(__global float *)(dst_addr + 35 * dst_stride_z) = out4.s3; |
| 700 | *(__global float *)(dst_addr + 36 * dst_stride_z) = out4.s4; |
| 701 | *(__global float *)(dst_addr + 37 * dst_stride_z) = out4.s5; |
| 702 | *(__global float *)(dst_addr + 38 * dst_stride_z) = out4.s6; |
| 703 | *(__global float *)(dst_addr + 39 * dst_stride_z) = out4.s7; |
| 704 | *(__global float *)(dst_addr + 40 * dst_stride_z) = out5.s0; |
| 705 | *(__global float *)(dst_addr + 41 * dst_stride_z) = out5.s1; |
| 706 | *(__global float *)(dst_addr + 42 * dst_stride_z) = out5.s2; |
| 707 | *(__global float *)(dst_addr + 43 * dst_stride_z) = out5.s3; |
| 708 | *(__global float *)(dst_addr + 44 * dst_stride_z) = out5.s4; |
| 709 | *(__global float *)(dst_addr + 45 * dst_stride_z) = out5.s5; |
| 710 | *(__global float *)(dst_addr + 46 * dst_stride_z) = out5.s6; |
| 711 | *(__global float *)(dst_addr + 47 * dst_stride_z) = out5.s7; |
| 712 | *(__global float *)(dst_addr + 48 * dst_stride_z) = out6.s0; |
| 713 | *(__global float *)(dst_addr + 49 * dst_stride_z) = out6.s1; |
| 714 | *(__global float *)(dst_addr + 50 * dst_stride_z) = out6.s2; |
| 715 | *(__global float *)(dst_addr + 51 * dst_stride_z) = out6.s3; |
| 716 | *(__global float *)(dst_addr + 52 * dst_stride_z) = out6.s4; |
| 717 | *(__global float *)(dst_addr + 53 * dst_stride_z) = out6.s5; |
| 718 | *(__global float *)(dst_addr + 54 * dst_stride_z) = out6.s6; |
| 719 | *(__global float *)(dst_addr + 55 * dst_stride_z) = out6.s7; |
| 720 | *(__global float *)(dst_addr + 56 * dst_stride_z) = out7.s0; |
| 721 | *(__global float *)(dst_addr + 57 * dst_stride_z) = out7.s1; |
| 722 | *(__global float *)(dst_addr + 58 * dst_stride_z) = out7.s2; |
| 723 | *(__global float *)(dst_addr + 59 * dst_stride_z) = out7.s3; |
| 724 | *(__global float *)(dst_addr + 60 * dst_stride_z) = out7.s4; |
| 725 | *(__global float *)(dst_addr + 61 * dst_stride_z) = out7.s5; |
| 726 | *(__global float *)(dst_addr + 62 * dst_stride_z) = out7.s6; |
| 727 | *(__global float *)(dst_addr + 63 * dst_stride_z) = out7.s7; |
| Gian Marco Iodice | 876be2a | 2018-07-03 12:22:09 +0100 | [diff] [blame] | 728 | #endif // !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) |
| Giorgio Arena | a50e5e0 | 2018-07-02 13:42:23 +0100 | [diff] [blame] | 729 | } |
| 730 | |
| Gian Marco Iodice | d28b751 | 2018-07-06 12:59:28 +0100 | [diff] [blame] | 731 | /** This OpenCL kernel performs Winograd filter transform 5x5/5x1 or 1x5 when the data layout is NHWC and the output tile is 4x4/4x1 or 1x4 |
| Giorgio Arena | a50e5e0 | 2018-07-02 13:42:23 +0100 | [diff] [blame] | 732 | * |
| 733 | * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64 |
| Gian Marco Iodice | d28b751 | 2018-07-06 12:59:28 +0100 | [diff] [blame] | 734 | * @note If this kernel is used to perform Winograd filter transform 5x1, -DWINOGRAD_FILTER_TRANSFORM_HORIZONTAL has to be passed at compile time |
| 735 | * @note If this kernel is used to perform Winograd filter transform 1x5, -DWINOGRAD_FILTER_TRANSFORM_VERTICAL has to be passed at compile time |
| Giorgio Arena | a50e5e0 | 2018-07-02 13:42:23 +0100 | [diff] [blame] | 736 | * |
| 737 | * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 |
| 738 | * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) |
| 739 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 740 | * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 741 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 742 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 743 | * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 744 | * @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes) |
| 745 | * @param[in] src_step_w src_stride_w * number of elements along W processed per workitem(in bytes) |
| 746 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor |
| 747 | * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr |
| 748 | * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 749 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 750 | * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 751 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 752 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 753 | * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 754 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 755 | */ |
| 756 | __kernel void winograd_filter_transform_4x4_5x5_nhwc( |
| 757 | TENSOR4D_DECLARATION(src), |
| 758 | TENSOR3D_DECLARATION(dst)) |
| 759 | { |
| 760 | Tensor4D src = CONVERT_TO_TENSOR4D_STRUCT(src, SRC_DIM_Z); |
| 761 | |
| 762 | const __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + get_global_id(0) * sizeof(float) + get_global_id(1) * src_step_y + get_global_id(2) * src_step_w; |
| 763 | |
| Gian Marco Iodice | d28b751 | 2018-07-06 12:59:28 +0100 | [diff] [blame] | 764 | #if defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) |
| Giorgio Arena | a50e5e0 | 2018-07-02 13:42:23 +0100 | [diff] [blame] | 765 | // Load the values from the input tensor |
| Gian Marco Iodice | d28b751 | 2018-07-06 12:59:28 +0100 | [diff] [blame] | 766 | float w00 = *((__global float *)(src_addr + 0 * src_stride_z)); |
| 767 | float w01 = *((__global float *)(src_addr + 1 * src_stride_z)); |
| 768 | float w02 = *((__global float *)(src_addr + 2 * src_stride_z)); |
| 769 | float w03 = *((__global float *)(src_addr + 3 * src_stride_z)); |
| 770 | float w04 = *((__global float *)(src_addr + 4 * src_stride_z)); |
| 771 | #else // defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) |
| 772 | // Load the values from the input tensor |
| 773 | float w00 = *((__global float *)(src_addr + 0 * src_stride_y)); |
| 774 | float w01 = *((__global float *)(src_addr + 1 * src_stride_y)); |
| 775 | float w02 = *((__global float *)(src_addr + 2 * src_stride_y)); |
| 776 | float w03 = *((__global float *)(src_addr + 3 * src_stride_y)); |
| 777 | float w04 = *((__global float *)(src_addr + 4 * src_stride_y)); |
| 778 | #endif // defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) |
| 779 | |
| 780 | #if !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) |
| Giorgio Arena | a50e5e0 | 2018-07-02 13:42:23 +0100 | [diff] [blame] | 781 | float w10 = *((__global float *)(src_addr + 1 * src_stride_z + 0 * src_stride_y)); |
| 782 | float w11 = *((__global float *)(src_addr + 1 * src_stride_z + 1 * src_stride_y)); |
| 783 | float w12 = *((__global float *)(src_addr + 1 * src_stride_z + 2 * src_stride_y)); |
| 784 | float w13 = *((__global float *)(src_addr + 1 * src_stride_z + 3 * src_stride_y)); |
| 785 | float w14 = *((__global float *)(src_addr + 1 * src_stride_z + 4 * src_stride_y)); |
| 786 | float w20 = *((__global float *)(src_addr + 2 * src_stride_z + 0 * src_stride_y)); |
| 787 | float w21 = *((__global float *)(src_addr + 2 * src_stride_z + 1 * src_stride_y)); |
| 788 | float w22 = *((__global float *)(src_addr + 2 * src_stride_z + 2 * src_stride_y)); |
| 789 | float w23 = *((__global float *)(src_addr + 2 * src_stride_z + 3 * src_stride_y)); |
| 790 | float w24 = *((__global float *)(src_addr + 2 * src_stride_z + 4 * src_stride_y)); |
| 791 | float w30 = *((__global float *)(src_addr + 3 * src_stride_z + 0 * src_stride_y)); |
| 792 | float w31 = *((__global float *)(src_addr + 3 * src_stride_z + 1 * src_stride_y)); |
| 793 | float w32 = *((__global float *)(src_addr + 3 * src_stride_z + 2 * src_stride_y)); |
| 794 | float w33 = *((__global float *)(src_addr + 3 * src_stride_z + 3 * src_stride_y)); |
| 795 | float w34 = *((__global float *)(src_addr + 3 * src_stride_z + 4 * src_stride_y)); |
| 796 | float w40 = *((__global float *)(src_addr + 4 * src_stride_z + 0 * src_stride_y)); |
| 797 | float w41 = *((__global float *)(src_addr + 4 * src_stride_z + 1 * src_stride_y)); |
| 798 | float w42 = *((__global float *)(src_addr + 4 * src_stride_z + 2 * src_stride_y)); |
| 799 | float w43 = *((__global float *)(src_addr + 4 * src_stride_z + 3 * src_stride_y)); |
| 800 | float w44 = *((__global float *)(src_addr + 4 * src_stride_z + 4 * src_stride_y)); |
| Gian Marco Iodice | d28b751 | 2018-07-06 12:59:28 +0100 | [diff] [blame] | 801 | #endif // !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) |
| Giorgio Arena | a50e5e0 | 2018-07-02 13:42:23 +0100 | [diff] [blame] | 802 | |
| 803 | // Row 0 |
| Gian Marco Iodice | d28b751 | 2018-07-06 12:59:28 +0100 | [diff] [blame] | 804 | float8 out0 = 0.0f; |
| 805 | out0.s0 = w00; |
| 806 | out0.s1 = -2.f * (w00 + w01 + w02 + w03 + w04) / 9.f; |
| 807 | out0.s2 = -2.f * (w00 - w01 + w02 - w03 + w04) / 9.f; |
| 808 | out0.s3 = (w00 + 2.f * w01 + 4.f * w02 + 8.f * w03 + 16.f * w04) / 90.f; |
| 809 | out0.s4 = (w00 - 2.f * w01 + 4.f * w02 - 8.f * w03 + 16.f * w04) / 90.f; |
| 810 | out0.s5 = (16.f * w00 + 8.f * w01 + 4.f * w02 + 2.f * w03 + w04) / 180.f; |
| 811 | out0.s6 = (16.f * w00 - 8.f * w01 + 4.f * w02 - 2.f * w03 + w04) / 180.f; |
| 812 | out0.s7 = w04; |
| Giorgio Arena | a50e5e0 | 2018-07-02 13:42:23 +0100 | [diff] [blame] | 813 | |
| Gian Marco Iodice | d28b751 | 2018-07-06 12:59:28 +0100 | [diff] [blame] | 814 | #if !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) |
| Giorgio Arena | a50e5e0 | 2018-07-02 13:42:23 +0100 | [diff] [blame] | 815 | // Row 1 |
| Gian Marco Iodice | d28b751 | 2018-07-06 12:59:28 +0100 | [diff] [blame] | 816 | float8 out1 = 0.0f; |
| 817 | out1.s0 = -2.f * (w00 + w10 + w20 + w30 + w40) / 9.f; |
| 818 | out1.s1 = 4.f * ((w00 + w10 + w20 + w30 + w40) + (w01 + w11 + w21 + w31 + w41) + (w02 + w12 + w22 + w32 + w42) + (w03 + w13 + w23 + w33 + w43) + (w04 + w14 + w24 + w34 + w44)) / 81.f; |
| 819 | out1.s2 = 4.f * ((w00 + w10 + w20 + w30 + w40) - (w01 + w11 + w21 + w31 + w41) + (w02 + w12 + w22 + w32 + w42) - (w03 + w13 + w23 + w33 + w43) + (w04 + w14 + w24 + w34 + w44)) / 81.f; |
| 820 | out1.s3 = -((w00 + w10 + w20 + w30 + w40) + 2.f * (w01 + w11 + w21 + w31 + w41) + 4.f * (w02 + w12 + w22 + w32 + w42) + 8.f * (w03 + w13 + w23 + w33 + w43) + 16.f * |
| 821 | (w04 + w14 + w24 + w34 + w44)) / 405.f; |
| 822 | out1.s4 = -((w00 + w10 + w20 + w30 + w40) - 2.f * (w01 + w11 + w21 + w31 + w41) + 4.f * (w02 + w12 + w22 + w32 + w42) - 8.f * (w03 + w13 + w23 + w33 + w43) + 16.f * |
| 823 | (w04 + w14 + w24 + w34 + w44)) / 405.f; |
| 824 | out1.s5 = -(16.f * (w00 + w10 + w20 + w30 + w40) + 8.f * (w01 + w11 + w21 + w31 + w41) + 4.f * (w02 + w12 + w22 + w32 + w42) + 2.f * (w03 + w13 + w23 + w33 + w43) + |
| 825 | (w04 + w14 + w24 + w34 + w44)) / 810.f; |
| 826 | out1.s6 = -(16.f * (w00 + w10 + w20 + w30 + w40) - 8.f * (w01 + w11 + w21 + w31 + w41) + 4.f * (w02 + w12 + w22 + w32 + w42) - 2.f * (w03 + w13 + w23 + w33 + w43) + |
| 827 | (w04 + w14 + w24 + w34 + w44)) / 810.f; |
| 828 | out1.s7 = -2.f * (w04 + w14 + w24 + w34 + w44) / 9.f; |
| Giorgio Arena | a50e5e0 | 2018-07-02 13:42:23 +0100 | [diff] [blame] | 829 | |
| 830 | // Row 2 |
| Gian Marco Iodice | d28b751 | 2018-07-06 12:59:28 +0100 | [diff] [blame] | 831 | float8 out2 = 0.0f; |
| 832 | out2.s0 = -2.f * (w00 - w10 + w20 - w30 + w40) / 9.f; |
| 833 | out2.s1 = 4.f * ((w00 - w10 + w20 - w30 + w40) + (w01 - w11 + w21 - w31 + w41) + (w02 - w12 + w22 - w32 + w42) + (w03 - w13 + w23 - w33 + w43) + (w04 - w14 + w24 - w34 + w44)) / 81.f; |
| 834 | out2.s2 = 4.f * ((w00 - w10 + w20 - w30 + w40) - (w01 - w11 + w21 - w31 + w41) + (w02 - w12 + w22 - w32 + w42) - (w03 - w13 + w23 - w33 + w43) + (w04 - w14 + w24 - w34 + w44)) / 81.f; |
| 835 | out2.s3 = -((w00 - w10 + w20 - w30 + w40) + 2.f * (w01 - w11 + w21 - w31 + w41) + 4.f * (w02 - w12 + w22 - w32 + w42) + 8.f * (w03 - w13 + w23 - w33 + w43) + 16.f * |
| 836 | (w04 - w14 + w24 - w34 + w44)) / 405.f; |
| 837 | out2.s4 = -((w00 - w10 + w20 - w30 + w40) - 2.f * (w01 - w11 + w21 - w31 + w41) + 4.f * (w02 - w12 + w22 - w32 + w42) - 8.f * (w03 - w13 + w23 - w33 + w43) + 16.f * |
| 838 | (w04 - w14 + w24 - w34 + w44)) / 405.f; |
| 839 | out2.s5 = -(16.f * (w00 - w10 + w20 - w30 + w40) + 8.f * (w01 - w11 + w21 - w31 + w41) + 4.f * (w02 - w12 + w22 - w32 + w42) + 2.f * (w03 - w13 + w23 - w33 + w43) + |
| 840 | (w04 - w14 + w24 - w34 + w44)) / 810.f; |
| 841 | out2.s6 = -(16.f * (w00 - w10 + w20 - w30 + w40) - 8.f * (w01 - w11 + w21 - w31 + w41) + 4.f * (w02 - w12 + w22 - w32 + w42) - 2.f * (w03 - w13 + w23 - w33 + w43) + |
| 842 | (w04 - w14 + w24 - w34 + w44)) / 810.f; |
| 843 | out2.s7 = -2.f * (w04 - w14 + w24 - w34 + w44) / 9.f; |
| Giorgio Arena | a50e5e0 | 2018-07-02 13:42:23 +0100 | [diff] [blame] | 844 | |
| 845 | // Row 3 |
| Gian Marco Iodice | d28b751 | 2018-07-06 12:59:28 +0100 | [diff] [blame] | 846 | float8 out3 = 0.0f; |
| 847 | out3.s0 = (w00 + 2.f * w10 + 4.f * w20 + 8.f * w30 + 16.f * w40) / 90.f; |
| 848 | out3.s1 = -((w00 + 2.f * w10 + 4.f * w20 + 8.f * w30 + 16.f * w40) + (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41) + (w02 + 2.f * w12 + 4.f * w22 + 8.f * w32 + 16.f * w42) + |
| 849 | (w03 + 2.f * w13 + 4.f * w23 + 8.f * w33 + 16.f * w43) + (w04 + 2.f * w14 + 4.f * w24 + 8.f * w34 + 16.f * w44)) / 405.f; |
| 850 | out3.s2 = -((w00 + 2.f * w10 + 4.f * w20 + 8.f * w30 + 16.f * w40) - (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41) + (w02 + 2.f * w12 + 4.f * w22 + 8.f * w32 + 16.f * w42) - |
| 851 | (w03 + 2.f * w13 + 4.f * w23 + 8.f * w33 + 16.f * w43) + (w04 + 2.f * w14 + 4.f * w24 + 8.f * w34 + 16.f * w44)) / 405.f; |
| 852 | out3.s3 = ((w00 + 2.f * w10 + 4.f * w20 + 8.f * w30 + 16.f * w40) + 2.f * (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41) + 4.f * |
| 853 | (w02 + 2.f * w12 + 4.f * w22 + 8.f * w32 + 16.f * w42) + 8.f * (w03 + 2.f * w13 + 4.f * w23 + 8.f * w33 + 16.f * w43) + 16.f * (w04 + 2.f * w14 + 4.f * w24 + 8.f * w34 + 16.f * w44)) / 8100.f; |
| 854 | out3.s4 = ((w00 + 2.f * w10 + 4.f * w20 + 8.f * w30 + 16.f * w40) - 2.f * (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41) + 4.f * |
| 855 | (w02 + 2.f * w12 + 4.f * w22 + 8.f * w32 + 16.f * w42) - 8.f * (w03 + 2.f * w13 + 4.f * w23 + 8.f * w33 + 16.f * w43) + 16.f * (w04 + 2.f * w14 + 4.f * w24 + 8.f * w34 + 16.f * w44)) / 8100.f; |
| 856 | out3.s5 = (16.f * (w00 + 2.f * w10 + 4.f * w20 + 8.f * w30 + 16.f * w40) + 8.f * (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41) + 4.f * |
| 857 | (w02 + 2.f * w12 + 4.f * w22 + 8.f * w32 + 16.f * w42) + 2.f * (w03 + 2.f * w13 + 4.f * w23 + 8.f * w33 + 16.f * w43) + (w04 + 2.f * w14 + 4.f * w24 + 8.f * w34 + 16.f * w44)) / 16200.f; |
| 858 | out3.s6 = (16.f * (w00 + 2.f * w10 + 4.f * w20 + 8.f * w30 + 16.f * w40) - 8.f * (w01 + 2.f * w11 + 4.f * w21 + 8.f * w31 + 16.f * w41) + 4.f * |
| 859 | (w02 + 2.f * w12 + 4.f * w22 + 8.f * w32 + 16.f * w42) - 2.f * (w03 + 2.f * w13 + 4.f * w23 + 8.f * w33 + 16.f * w43) + (w04 + 2.f * w14 + 4.f * w24 + 8.f * w34 + 16.f * w44)) / 16200.f; |
| 860 | out3.s7 = (w04 + 2.f * w14 + 4.f * w24 + 8.f * w34 + 16.f * w44) / 90.f; |
| Giorgio Arena | a50e5e0 | 2018-07-02 13:42:23 +0100 | [diff] [blame] | 861 | |
| 862 | // Row 4 |
| Gian Marco Iodice | d28b751 | 2018-07-06 12:59:28 +0100 | [diff] [blame] | 863 | float8 out4 = 0.0f; |
| 864 | out4.s0 = (w00 - 2.f * w10 + 4.f * w20 - 8.f * w30 + 16.f * w40) / 90.f; |
| 865 | out4.s1 = -((w00 - 2.f * w10 + 4.f * w20 - 8.f * w30 + 16.f * w40) + (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41) + (w02 - 2.f * w12 + 4.f * w22 - 8.f * w32 + 16.f * w42) + |
| 866 | (w03 - 2.f * w13 + 4.f * w23 - 8.f * w33 + 16.f * w43) + (w04 - 2.f * w14 + 4.f * w24 - 8.f * w34 + 16.f * w44)) / 405.f; |
| 867 | out4.s2 = -((w00 - 2.f * w10 + 4.f * w20 - 8.f * w30 + 16.f * w40) - (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41) + (w02 - 2.f * w12 + 4.f * w22 - 8.f * w32 + 16.f * w42) - |
| 868 | (w03 - 2.f * w13 + 4.f * w23 - 8.f * w33 + 16.f * w43) + (w04 - 2.f * w14 + 4.f * w24 - 8.f * w34 + 16.f * w44)) / 405.f; |
| 869 | out4.s3 = ((w00 - 2.f * w10 + 4.f * w20 - 8.f * w30 + 16.f * w40) + 2.f * (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41) + 4.f * |
| 870 | (w02 - 2.f * w12 + 4.f * w22 - 8.f * w32 + 16.f * w42) + 8.f * (w03 - 2.f * w13 + 4.f * w23 - 8.f * w33 + 16.f * w43) + 16.f * (w04 - 2.f * w14 + 4.f * w24 - 8.f * w34 + 16.f * w44)) / 8100.f; |
| 871 | out4.s4 = ((w00 - 2.f * w10 + 4.f * w20 - 8.f * w30 + 16.f * w40) - 2.f * (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41) + 4.f * |
| 872 | (w02 - 2.f * w12 + 4.f * w22 - 8.f * w32 + 16.f * w42) - 8.f * (w03 - 2.f * w13 + 4.f * w23 - 8.f * w33 + 16.f * w43) + 16.f * (w04 - 2.f * w14 + 4.f * w24 - 8.f * w34 + 16.f * w44)) / 8100.f; |
| 873 | out4.s5 = (16.f * (w00 - 2.f * w10 + 4.f * w20 - 8.f * w30 + 16.f * w40) + 8.f * (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41) + 4.f * |
| 874 | (w02 - 2.f * w12 + 4.f * w22 - 8.f * w32 + 16.f * w42) + 2.f * (w03 - 2.f * w13 + 4.f * w23 - 8.f * w33 + 16.f * w43) + (w04 - 2.f * w14 + 4.f * w24 - 8.f * w34 + 16.f * w44)) / 16200.f; |
| 875 | out4.s6 = (16.f * (w00 - 2.f * w10 + 4.f * w20 - 8.f * w30 + 16.f * w40) - 8.f * (w01 - 2.f * w11 + 4.f * w21 - 8.f * w31 + 16.f * w41) + 4.f * |
| 876 | (w02 - 2.f * w12 + 4.f * w22 - 8.f * w32 + 16.f * w42) - 2.f * (w03 - 2.f * w13 + 4.f * w23 - 8.f * w33 + 16.f * w43) + (w04 - 2.f * w14 + 4.f * w24 - 8.f * w34 + 16.f * w44)) / 16200.f; |
| 877 | out4.s7 = (w04 - 2.f * w14 + 4.f * w24 - 8.f * w34 + 16.f * w44) / 90.f; |
| Giorgio Arena | a50e5e0 | 2018-07-02 13:42:23 +0100 | [diff] [blame] | 878 | |
| 879 | // Row 5 |
| Gian Marco Iodice | d28b751 | 2018-07-06 12:59:28 +0100 | [diff] [blame] | 880 | float8 out5 = 0.0f; |
| 881 | out5.s0 = (16.f * w00 + 8.f * w10 + 4.f * w20 + 2.f * w30 + w40) / 180.f; |
| 882 | out5.s1 = -((16.f * w00 + 8.f * w10 + 4.f * w20 + 2.f * w30 + w40) + (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41) + (16.f * w02 + 8.f * w12 + 4.f * w22 + 2.f * w32 + w42) + |
| 883 | (16.f * w03 + 8.f * w13 + 4.f * w23 + 2.f * w33 + w43) + (16.f * w04 + 8.f * w14 + 4.f * w24 + 2.f * w34 + w44)) / 810.f; |
| 884 | out5.s2 = -((16.f * w00 + 8.f * w10 + 4.f * w20 + 2.f * w30 + w40) - (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41) + (16.f * w02 + 8.f * w12 + 4.f * w22 + 2.f * w32 + w42) - |
| 885 | (16.f * w03 + 8.f * w13 + 4.f * w23 + 2.f * w33 + w43) + (16.f * w04 + 8.f * w14 + 4.f * w24 + 2.f * w34 + w44)) / 810.f; |
| 886 | out5.s3 = ((16.f * w00 + 8.f * w10 + 4.f * w20 + 2.f * w30 + w40) + 2.f * (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41) + 4.f * |
| 887 | (16.f * w02 + 8.f * w12 + 4.f * w22 + 2.f * w32 + w42) + 8.f * (16.f * w03 + 8.f * w13 + 4.f * w23 + 2.f * w33 + w43) + 16.f * (16.f * w04 + 8.f * w14 + 4.f * w24 + 2.f * w34 + w44)) / 16200.f; |
| 888 | out5.s4 = ((16.f * w00 + 8.f * w10 + 4.f * w20 + 2.f * w30 + w40) - 2.f * (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41) + 4.f * |
| 889 | (16.f * w02 + 8.f * w12 + 4.f * w22 + 2.f * w32 + w42) - 8.f * (16.f * w03 + 8.f * w13 + 4.f * w23 + 2.f * w33 + w43) + 16.f * (16.f * w04 + 8.f * w14 + 4.f * w24 + 2.f * w34 + w44)) / 16200.f; |
| 890 | out5.s5 = (16.f * (16.f * w00 + 8.f * w10 + 4.f * w20 + 2.f * w30 + w40) + 8.f * (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41) + 4.f * |
| 891 | (16.f * w02 + 8.f * w12 + 4.f * w22 + 2.f * w32 + w42) + 2.f * (16.f * w03 + 8.f * w13 + 4.f * w23 + 2.f * w33 + w43) + (16.f * w04 + 8.f * w14 + 4.f * w24 + 2.f * w34 + w44)) / 32400.f; |
| 892 | out5.s6 = (16.f * (16.f * w00 + 8.f * w10 + 4.f * w20 + 2.f * w30 + w40) - 8.f * (16.f * w01 + 8.f * w11 + 4.f * w21 + 2.f * w31 + w41) + 4.f * |
| 893 | (16.f * w02 + 8.f * w12 + 4.f * w22 + 2.f * w32 + w42) - 2.f * (16.f * w03 + 8.f * w13 + 4.f * w23 + 2.f * w33 + w43) + (16.f * w04 + 8.f * w14 + 4.f * w24 + 2.f * w34 + w44)) / 32400.f; |
| 894 | out5.s7 = (16.f * w04 + 8.f * w14 + 4.f * w24 + 2.f * w34 + w44) / 180.f; |
| Giorgio Arena | a50e5e0 | 2018-07-02 13:42:23 +0100 | [diff] [blame] | 895 | |
| 896 | // Row 6 |
| Gian Marco Iodice | d28b751 | 2018-07-06 12:59:28 +0100 | [diff] [blame] | 897 | float8 out6 = 0.0f; |
| 898 | out6.s0 = (16.f * w00 - 8.f * w10 + 4.f * w20 - 2.f * w30 + w40) / 180.f; |
| 899 | out6.s1 = -((16.f * w00 - 8.f * w10 + 4.f * w20 - 2.f * w30 + w40) + (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41) + (16.f * w02 - 8.f * w12 + 4.f * w22 - 2.f * w32 + w42) + |
| 900 | (16.f * w03 - 8.f * w13 + 4.f * w23 - 2.f * w33 + w43) + (16.f * w04 - 8.f * w14 + 4.f * w24 - 2.f * w34 + w44)) / 810.f; |
| 901 | out6.s2 = -((16.f * w00 - 8.f * w10 + 4.f * w20 - 2.f * w30 + w40) - (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41) + (16.f * w02 - 8.f * w12 + 4.f * w22 - 2.f * w32 + w42) - |
| 902 | (16.f * w03 - 8.f * w13 + 4.f * w23 - 2.f * w33 + w43) + (16.f * w04 - 8.f * w14 + 4.f * w24 - 2.f * w34 + w44)) / 810.f; |
| 903 | out6.s3 = ((16.f * w00 - 8.f * w10 + 4.f * w20 - 2.f * w30 + w40) + 2.f * (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41) + 4.f * |
| 904 | (16.f * w02 - 8.f * w12 + 4.f * w22 - 2.f * w32 + w42) + 8.f * (16.f * w03 - 8.f * w13 + 4.f * w23 - 2.f * w33 + w43) + 16.f * (16.f * w04 - 8.f * w14 + 4.f * w24 - 2.f * w34 + w44)) / 16200.f; |
| 905 | out6.s4 = ((16.f * w00 - 8.f * w10 + 4.f * w20 - 2.f * w30 + w40) - 2.f * (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41) + 4.f * |
| 906 | (16.f * w02 - 8.f * w12 + 4.f * w22 - 2.f * w32 + w42) - 8.f * (16.f * w03 - 8.f * w13 + 4.f * w23 - 2.f * w33 + w43) + 16.f * (16.f * w04 - 8.f * w14 + 4.f * w24 - 2.f * w34 + w44)) / 16200.f; |
| 907 | out6.s5 = (16.f * (16.f * w00 - 8.f * w10 + 4.f * w20 - 2.f * w30 + w40) + 8.f * (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41) + 4.f * |
| 908 | (16.f * w02 - 8.f * w12 + 4.f * w22 - 2.f * w32 + w42) + 2.f * (16.f * w03 - 8.f * w13 + 4.f * w23 - 2.f * w33 + w43) + (16.f * w04 - 8.f * w14 + 4.f * w24 - 2.f * w34 + w44)) / 32400.f; |
| 909 | out6.s6 = (16.f * (16.f * w00 - 8.f * w10 + 4.f * w20 - 2.f * w30 + w40) - 8.f * (16.f * w01 - 8.f * w11 + 4.f * w21 - 2.f * w31 + w41) + 4.f * |
| 910 | (16.f * w02 - 8.f * w12 + 4.f * w22 - 2.f * w32 + w42) - 2.f * (16.f * w03 - 8.f * w13 + 4.f * w23 - 2.f * w33 + w43) + (16.f * w04 - 8.f * w14 + 4.f * w24 - 2.f * w34 + w44)) / 32400.f; |
| 911 | out6.s7 = (16.f * w04 - 8.f * w14 + 4.f * w24 - 2.f * w34 + w44) / 180.f; |
| Giorgio Arena | a50e5e0 | 2018-07-02 13:42:23 +0100 | [diff] [blame] | 912 | |
| 913 | // Row 7 |
| Gian Marco Iodice | d28b751 | 2018-07-06 12:59:28 +0100 | [diff] [blame] | 914 | float8 out7 = 0.0f; |
| 915 | out7.s0 = w40; |
| 916 | out7.s1 = -2.f * (w40 + w41 + w42 + w43 + w44) / 9.f; |
| 917 | out7.s2 = -2.f * (w40 - w41 + w42 - w43 + w44) / 9.f; |
| 918 | out7.s3 = (w40 + 2.f * w41 + 4.f * w42 + 8.f * w43 + 16.f * w44) / 90.f; |
| 919 | out7.s4 = (w40 - 2.f * w41 + 4.f * w42 - 8.f * w43 + 16.f * w44) / 90.f; |
| 920 | out7.s5 = (16.f * w40 + 8.f * w41 + 4.f * w42 + 2.f * w43 + w44) / 180.f; |
| 921 | out7.s6 = (16.f * w40 - 8.f * w41 + 4.f * w42 - 2.f * w43 + w44) / 180.f; |
| 922 | out7.s7 = w44; |
| 923 | #endif // !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) |
| Giorgio Arena | a50e5e0 | 2018-07-02 13:42:23 +0100 | [diff] [blame] | 924 | |
| 925 | int x0 = get_global_id(2); // idx filter |
| 926 | int y0 = get_global_id(0); // idx channel |
| 927 | |
| 928 | // Get output address |
| 929 | __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x0 * sizeof(float) + y0 * dst_stride_y; |
| 930 | |
| Gian Marco Iodice | d28b751 | 2018-07-06 12:59:28 +0100 | [diff] [blame] | 931 | // Store the values across the channels |
| 932 | *(__global float *)(dst_addr + 0 * dst_stride_z) = out0.s0; |
| 933 | *(__global float *)(dst_addr + 1 * dst_stride_z) = out0.s1; |
| 934 | *(__global float *)(dst_addr + 2 * dst_stride_z) = out0.s2; |
| 935 | *(__global float *)(dst_addr + 3 * dst_stride_z) = out0.s3; |
| 936 | *(__global float *)(dst_addr + 4 * dst_stride_z) = out0.s4; |
| 937 | *(__global float *)(dst_addr + 5 * dst_stride_z) = out0.s5; |
| 938 | *(__global float *)(dst_addr + 6 * dst_stride_z) = out0.s6; |
| 939 | *(__global float *)(dst_addr + 7 * dst_stride_z) = out0.s7; |
| 940 | |
| 941 | #if !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) |
| Giorgio Arena | a50e5e0 | 2018-07-02 13:42:23 +0100 | [diff] [blame] | 942 | *(__global float *)(dst_addr + 8 * dst_stride_z) = out1.s0; |
| 943 | *(__global float *)(dst_addr + 9 * dst_stride_z) = out1.s1; |
| 944 | *(__global float *)(dst_addr + 10 * dst_stride_z) = out1.s2; |
| 945 | *(__global float *)(dst_addr + 11 * dst_stride_z) = out1.s3; |
| 946 | *(__global float *)(dst_addr + 12 * dst_stride_z) = out1.s4; |
| 947 | *(__global float *)(dst_addr + 13 * dst_stride_z) = out1.s5; |
| 948 | *(__global float *)(dst_addr + 14 * dst_stride_z) = out1.s6; |
| 949 | *(__global float *)(dst_addr + 15 * dst_stride_z) = out1.s7; |
| 950 | *(__global float *)(dst_addr + 16 * dst_stride_z) = out2.s0; |
| 951 | *(__global float *)(dst_addr + 17 * dst_stride_z) = out2.s1; |
| 952 | *(__global float *)(dst_addr + 18 * dst_stride_z) = out2.s2; |
| 953 | *(__global float *)(dst_addr + 19 * dst_stride_z) = out2.s3; |
| 954 | *(__global float *)(dst_addr + 20 * dst_stride_z) = out2.s4; |
| 955 | *(__global float *)(dst_addr + 21 * dst_stride_z) = out2.s5; |
| 956 | *(__global float *)(dst_addr + 22 * dst_stride_z) = out2.s6; |
| 957 | *(__global float *)(dst_addr + 23 * dst_stride_z) = out2.s7; |
| 958 | *(__global float *)(dst_addr + 24 * dst_stride_z) = out3.s0; |
| 959 | *(__global float *)(dst_addr + 25 * dst_stride_z) = out3.s1; |
| 960 | *(__global float *)(dst_addr + 26 * dst_stride_z) = out3.s2; |
| 961 | *(__global float *)(dst_addr + 27 * dst_stride_z) = out3.s3; |
| 962 | *(__global float *)(dst_addr + 28 * dst_stride_z) = out3.s4; |
| 963 | *(__global float *)(dst_addr + 29 * dst_stride_z) = out3.s5; |
| 964 | *(__global float *)(dst_addr + 30 * dst_stride_z) = out3.s6; |
| 965 | *(__global float *)(dst_addr + 31 * dst_stride_z) = out3.s7; |
| 966 | *(__global float *)(dst_addr + 32 * dst_stride_z) = out4.s0; |
| 967 | *(__global float *)(dst_addr + 33 * dst_stride_z) = out4.s1; |
| 968 | *(__global float *)(dst_addr + 34 * dst_stride_z) = out4.s2; |
| 969 | *(__global float *)(dst_addr + 35 * dst_stride_z) = out4.s3; |
| 970 | *(__global float *)(dst_addr + 36 * dst_stride_z) = out4.s4; |
| 971 | *(__global float *)(dst_addr + 37 * dst_stride_z) = out4.s5; |
| 972 | *(__global float *)(dst_addr + 38 * dst_stride_z) = out4.s6; |
| 973 | *(__global float *)(dst_addr + 39 * dst_stride_z) = out4.s7; |
| 974 | *(__global float *)(dst_addr + 40 * dst_stride_z) = out5.s0; |
| 975 | *(__global float *)(dst_addr + 41 * dst_stride_z) = out5.s1; |
| 976 | *(__global float *)(dst_addr + 42 * dst_stride_z) = out5.s2; |
| 977 | *(__global float *)(dst_addr + 43 * dst_stride_z) = out5.s3; |
| 978 | *(__global float *)(dst_addr + 44 * dst_stride_z) = out5.s4; |
| 979 | *(__global float *)(dst_addr + 45 * dst_stride_z) = out5.s5; |
| 980 | *(__global float *)(dst_addr + 46 * dst_stride_z) = out5.s6; |
| 981 | *(__global float *)(dst_addr + 47 * dst_stride_z) = out5.s7; |
| 982 | *(__global float *)(dst_addr + 48 * dst_stride_z) = out6.s0; |
| 983 | *(__global float *)(dst_addr + 49 * dst_stride_z) = out6.s1; |
| 984 | *(__global float *)(dst_addr + 50 * dst_stride_z) = out6.s2; |
| 985 | *(__global float *)(dst_addr + 51 * dst_stride_z) = out6.s3; |
| 986 | *(__global float *)(dst_addr + 52 * dst_stride_z) = out6.s4; |
| 987 | *(__global float *)(dst_addr + 53 * dst_stride_z) = out6.s5; |
| 988 | *(__global float *)(dst_addr + 54 * dst_stride_z) = out6.s6; |
| 989 | *(__global float *)(dst_addr + 55 * dst_stride_z) = out6.s7; |
| 990 | *(__global float *)(dst_addr + 56 * dst_stride_z) = out7.s0; |
| 991 | *(__global float *)(dst_addr + 57 * dst_stride_z) = out7.s1; |
| 992 | *(__global float *)(dst_addr + 58 * dst_stride_z) = out7.s2; |
| 993 | *(__global float *)(dst_addr + 59 * dst_stride_z) = out7.s3; |
| 994 | *(__global float *)(dst_addr + 60 * dst_stride_z) = out7.s4; |
| 995 | *(__global float *)(dst_addr + 61 * dst_stride_z) = out7.s5; |
| 996 | *(__global float *)(dst_addr + 62 * dst_stride_z) = out7.s6; |
| 997 | *(__global float *)(dst_addr + 63 * dst_stride_z) = out7.s7; |
| Gian Marco Iodice | d28b751 | 2018-07-06 12:59:28 +0100 | [diff] [blame] | 998 | #endif // !defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) |
| Giorgio Arena | a50e5e0 | 2018-07-02 13:42:23 +0100 | [diff] [blame] | 999 | } |
| Gian Marco Iodice | 876be2a | 2018-07-03 12:22:09 +0100 | [diff] [blame] | 1000 | #endif // defined(SRC_DIM_Z) |
| 1001 | |
| 1002 | #if defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) |
| 1003 | /** This OpenCL kernel performs Winograd filter transform 3x1 when the data layout is NCHW and the output tile is 2x1 |
| 1004 | * |
| 1005 | * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64 |
| 1006 | * @note -DWINOGRAD_FILTER_TRANSFORM_HORIZONTAL has to be passed at compile time to perform Winograd Filter Transform |
| 1007 | * |
| 1008 | * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 |
| 1009 | * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) |
| 1010 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 1011 | * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 1012 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 1013 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 1014 | * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 1015 | * @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes) |
| 1016 | * @param[in] src_step_w src_stride_w * number of elements along W processed per workitem(in bytes) |
| 1017 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor |
| 1018 | * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr |
| 1019 | * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 1020 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 1021 | * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 1022 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 1023 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 1024 | * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 1025 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 1026 | */ |
| 1027 | __kernel void winograd_filter_transform_2x1_3x1_nchw( |
| 1028 | TENSOR4D_DECLARATION(src), |
| 1029 | TENSOR3D_DECLARATION(dst)) |
| 1030 | { |
| 1031 | winograd_filter_transform_2x2_3x3_nchw(src_ptr, |
| 1032 | src_stride_x, |
| 1033 | src_step_x, |
| 1034 | src_stride_y, |
| 1035 | src_step_y, |
| 1036 | src_stride_z, |
| 1037 | src_step_z, |
| 1038 | src_stride_w, |
| 1039 | src_step_w, |
| 1040 | src_offset_first_element_in_bytes, |
| 1041 | dst_ptr, |
| 1042 | dst_stride_x, |
| 1043 | dst_step_x, |
| 1044 | dst_stride_y, |
| 1045 | dst_step_y, |
| 1046 | dst_stride_z, |
| 1047 | dst_step_z, |
| 1048 | dst_offset_first_element_in_bytes); |
| 1049 | } |
| 1050 | |
| 1051 | /** This OpenCL kernel performs Winograd filter transform 3x1 when the data layout is NCHW and the output tile is 4x1 |
| 1052 | * |
| 1053 | * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64 |
| 1054 | * @note -DWINOGRAD_FILTER_TRANSFORM_HORIZONTAL has to be passed at compile time to perform Winograd Filter Transform |
| 1055 | * |
| 1056 | * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 |
| 1057 | * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) |
| 1058 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 1059 | * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 1060 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 1061 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 1062 | * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 1063 | * @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes) |
| 1064 | * @param[in] src_step_w src_stride_w * number of elements along W processed per workitem(in bytes) |
| 1065 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor |
| 1066 | * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr |
| 1067 | * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 1068 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 1069 | * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 1070 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 1071 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 1072 | * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 1073 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 1074 | */ |
| 1075 | __kernel void winograd_filter_transform_4x1_3x1_nchw( |
| 1076 | TENSOR4D_DECLARATION(src), |
| 1077 | TENSOR3D_DECLARATION(dst)) |
| 1078 | { |
| 1079 | winograd_filter_transform_4x4_3x3_nchw(src_ptr, |
| 1080 | src_stride_x, |
| 1081 | src_step_x, |
| 1082 | src_stride_y, |
| 1083 | src_step_y, |
| 1084 | src_stride_z, |
| 1085 | src_step_z, |
| 1086 | src_stride_w, |
| 1087 | src_step_w, |
| 1088 | src_offset_first_element_in_bytes, |
| 1089 | dst_ptr, |
| 1090 | dst_stride_x, |
| 1091 | dst_step_x, |
| 1092 | dst_stride_y, |
| 1093 | dst_step_y, |
| 1094 | dst_stride_z, |
| 1095 | dst_step_z, |
| 1096 | dst_offset_first_element_in_bytes); |
| 1097 | } |
| 1098 | |
| 1099 | /** This OpenCL kernel performs Winograd filter transform 5x1 when the data layout is NCHW and the output tile is 4x1 |
| 1100 | * |
| 1101 | * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64 |
| 1102 | * @note -DWINOGRAD_FILTER_TRANSFORM_HORIZONTAL has to be passed at compile time to perform Winograd Filter Transform |
| 1103 | * |
| 1104 | * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 |
| 1105 | * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) |
| 1106 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 1107 | * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 1108 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 1109 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 1110 | * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 1111 | * @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes) |
| 1112 | * @param[in] src_step_w src_stride_w * number of elements along W processed per workitem(in bytes) |
| 1113 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor |
| 1114 | * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr |
| 1115 | * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 1116 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 1117 | * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 1118 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 1119 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 1120 | * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 1121 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 1122 | */ |
| 1123 | __kernel void winograd_filter_transform_4x1_5x1_nchw( |
| 1124 | TENSOR4D_DECLARATION(src), |
| 1125 | TENSOR3D_DECLARATION(dst)) |
| 1126 | { |
| 1127 | winograd_filter_transform_4x4_5x5_nchw(src_ptr, |
| 1128 | src_stride_x, |
| 1129 | src_step_x, |
| 1130 | src_stride_y, |
| 1131 | src_step_y, |
| 1132 | src_stride_z, |
| 1133 | src_step_z, |
| 1134 | src_stride_w, |
| 1135 | src_step_w, |
| 1136 | src_offset_first_element_in_bytes, |
| 1137 | dst_ptr, |
| 1138 | dst_stride_x, |
| 1139 | dst_step_x, |
| 1140 | dst_stride_y, |
| 1141 | dst_step_y, |
| 1142 | dst_stride_z, |
| 1143 | dst_step_z, |
| 1144 | dst_offset_first_element_in_bytes); |
| 1145 | } |
| Gian Marco Iodice | d28b751 | 2018-07-06 12:59:28 +0100 | [diff] [blame] | 1146 | |
| 1147 | /** This OpenCL kernel performs Winograd filter transform 3x1 when the data layout is NHWC and the output tile is 4x1 |
| 1148 | * |
| 1149 | * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64 |
| 1150 | * @note -DWINOGRAD_FILTER_TRANSFORM_HORIZONTAL has to be passed at compile time to perform Winograd Filter Transform |
| 1151 | * |
| 1152 | * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 |
| 1153 | * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) |
| 1154 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 1155 | * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 1156 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 1157 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 1158 | * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 1159 | * @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes) |
| 1160 | * @param[in] src_step_w src_stride_w * number of elements along W processed per workitem(in bytes) |
| 1161 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor |
| 1162 | * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr |
| 1163 | * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 1164 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 1165 | * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 1166 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 1167 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 1168 | * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 1169 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 1170 | */ |
| 1171 | __kernel void winograd_filter_transform_4x1_3x1_nhwc( |
| 1172 | TENSOR4D_DECLARATION(src), |
| 1173 | TENSOR3D_DECLARATION(dst)) |
| 1174 | { |
| 1175 | winograd_filter_transform_4x4_3x3_nhwc(src_ptr, |
| 1176 | src_stride_x, |
| 1177 | src_step_x, |
| 1178 | src_stride_y, |
| 1179 | src_step_y, |
| 1180 | src_stride_z, |
| 1181 | src_step_z, |
| 1182 | src_stride_w, |
| 1183 | src_step_w, |
| 1184 | src_offset_first_element_in_bytes, |
| 1185 | dst_ptr, |
| 1186 | dst_stride_x, |
| 1187 | dst_step_x, |
| 1188 | dst_stride_y, |
| 1189 | dst_step_y, |
| 1190 | dst_stride_z, |
| 1191 | dst_step_z, |
| 1192 | dst_offset_first_element_in_bytes); |
| 1193 | } |
| 1194 | |
| 1195 | /** This OpenCL kernel performs Winograd filter transform 5x1 when the data layout is NHWC and the output tile is 4x1 |
| 1196 | * |
| 1197 | * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64 |
| 1198 | * @note -DWINOGRAD_FILTER_TRANSFORM_HORIZONTAL has to be passed at compile time to perform Winograd Filter Transform |
| 1199 | * |
| 1200 | * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 |
| 1201 | * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) |
| 1202 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 1203 | * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 1204 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 1205 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 1206 | * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 1207 | * @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes) |
| 1208 | * @param[in] src_step_w src_stride_w * number of elements along W processed per workitem(in bytes) |
| 1209 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor |
| 1210 | * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr |
| 1211 | * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 1212 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 1213 | * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 1214 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 1215 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 1216 | * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 1217 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 1218 | */ |
| 1219 | __kernel void winograd_filter_transform_4x1_5x1_nhwc( |
| 1220 | TENSOR4D_DECLARATION(src), |
| 1221 | TENSOR3D_DECLARATION(dst)) |
| 1222 | { |
| 1223 | winograd_filter_transform_4x4_5x5_nhwc(src_ptr, |
| 1224 | src_stride_x, |
| 1225 | src_step_x, |
| 1226 | src_stride_y, |
| 1227 | src_step_y, |
| 1228 | src_stride_z, |
| 1229 | src_step_z, |
| 1230 | src_stride_w, |
| 1231 | src_step_w, |
| 1232 | src_offset_first_element_in_bytes, |
| 1233 | dst_ptr, |
| 1234 | dst_stride_x, |
| 1235 | dst_step_x, |
| 1236 | dst_stride_y, |
| 1237 | dst_step_y, |
| 1238 | dst_stride_z, |
| 1239 | dst_step_z, |
| 1240 | dst_offset_first_element_in_bytes); |
| 1241 | } |
| Gian Marco Iodice | 876be2a | 2018-07-03 12:22:09 +0100 | [diff] [blame] | 1242 | #endif // defined(WINOGRAD_FILTER_TRANSFORM_HORIZONTAL) |
| 1243 | |
| 1244 | #if defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) |
| 1245 | /** This OpenCL kernel performs Winograd filter transform 1x3 when the data layout is NCHW and the output tile is 1x2 |
| 1246 | * |
| 1247 | * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64 |
| 1248 | * @note -DWINOGRAD_FILTER_TRANSFORM_VERTICAL has to be passed at compile time to perform Winograd Filter Transform |
| 1249 | * |
| 1250 | * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 |
| 1251 | * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) |
| 1252 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 1253 | * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 1254 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 1255 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 1256 | * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 1257 | * @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes) |
| 1258 | * @param[in] src_step_w src_stride_w * number of elements along W processed per workitem(in bytes) |
| 1259 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor |
| 1260 | * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr |
| 1261 | * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 1262 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 1263 | * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 1264 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 1265 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 1266 | * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 1267 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 1268 | */ |
| 1269 | __kernel void winograd_filter_transform_1x2_1x3_nchw( |
| 1270 | TENSOR4D_DECLARATION(src), |
| 1271 | TENSOR3D_DECLARATION(dst)) |
| 1272 | { |
| 1273 | winograd_filter_transform_2x2_3x3_nchw(src_ptr, |
| 1274 | src_stride_x, |
| 1275 | src_step_x, |
| 1276 | src_stride_y, |
| 1277 | src_step_y, |
| 1278 | src_stride_z, |
| 1279 | src_step_z, |
| 1280 | src_stride_w, |
| 1281 | src_step_w, |
| 1282 | src_offset_first_element_in_bytes, |
| 1283 | dst_ptr, |
| 1284 | dst_stride_x, |
| 1285 | dst_step_x, |
| 1286 | dst_stride_y, |
| 1287 | dst_step_y, |
| 1288 | dst_stride_z, |
| 1289 | dst_step_z, |
| 1290 | dst_offset_first_element_in_bytes); |
| 1291 | } |
| 1292 | |
| 1293 | /** This OpenCL kernel performs Winograd filter transform 1x3 when the data layout is NCHW and the output tile is 1x4 |
| 1294 | * |
| 1295 | * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64 |
| 1296 | * @note -DWINOGRAD_FILTER_TRANSFORM_VERTICAL has to be passed at compile time to perform Winograd Filter Transform |
| 1297 | * |
| 1298 | * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 |
| 1299 | * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) |
| 1300 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 1301 | * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 1302 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 1303 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 1304 | * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 1305 | * @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes) |
| 1306 | * @param[in] src_step_w src_stride_w * number of elements along W processed per workitem(in bytes) |
| 1307 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor |
| 1308 | * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr |
| 1309 | * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 1310 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 1311 | * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 1312 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 1313 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 1314 | * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 1315 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 1316 | */ |
| 1317 | __kernel void winograd_filter_transform_1x4_1x3_nchw( |
| 1318 | TENSOR4D_DECLARATION(src), |
| 1319 | TENSOR3D_DECLARATION(dst)) |
| 1320 | { |
| 1321 | winograd_filter_transform_4x4_3x3_nchw(src_ptr, |
| 1322 | src_stride_x, |
| 1323 | src_step_x, |
| 1324 | src_stride_y, |
| 1325 | src_step_y, |
| 1326 | src_stride_z, |
| 1327 | src_step_z, |
| 1328 | src_stride_w, |
| 1329 | src_step_w, |
| 1330 | src_offset_first_element_in_bytes, |
| 1331 | dst_ptr, |
| 1332 | dst_stride_x, |
| 1333 | dst_step_x, |
| 1334 | dst_stride_y, |
| 1335 | dst_step_y, |
| 1336 | dst_stride_z, |
| 1337 | dst_step_z, |
| 1338 | dst_offset_first_element_in_bytes); |
| 1339 | } |
| 1340 | |
| 1341 | /** This OpenCL kernel performs Winograd filter transform 1x5 when the data layout is NCHW and the output tile is 1x4 |
| 1342 | * |
| 1343 | * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64 |
| 1344 | * @note -DWINOGRAD_FILTER_TRANSFORM_VERTICAL has to be passed at compile time to perform Winograd Filter Transform |
| 1345 | * |
| 1346 | * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 |
| 1347 | * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) |
| 1348 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 1349 | * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 1350 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 1351 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 1352 | * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 1353 | * @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes) |
| 1354 | * @param[in] src_step_w src_stride_w * number of elements along W processed per workitem(in bytes) |
| 1355 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor |
| 1356 | * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr |
| 1357 | * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 1358 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 1359 | * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 1360 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 1361 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 1362 | * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 1363 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 1364 | */ |
| 1365 | __kernel void winograd_filter_transform_1x4_1x5_nchw( |
| 1366 | TENSOR4D_DECLARATION(src), |
| 1367 | TENSOR3D_DECLARATION(dst)) |
| 1368 | { |
| 1369 | winograd_filter_transform_4x4_5x5_nchw(src_ptr, |
| 1370 | src_stride_x, |
| 1371 | src_step_x, |
| 1372 | src_stride_y, |
| 1373 | src_step_y, |
| 1374 | src_stride_z, |
| 1375 | src_step_z, |
| 1376 | src_stride_w, |
| 1377 | src_step_w, |
| 1378 | src_offset_first_element_in_bytes, |
| 1379 | dst_ptr, |
| 1380 | dst_stride_x, |
| 1381 | dst_step_x, |
| 1382 | dst_stride_y, |
| 1383 | dst_step_y, |
| 1384 | dst_stride_z, |
| 1385 | dst_step_z, |
| 1386 | dst_offset_first_element_in_bytes); |
| 1387 | } |
| Gian Marco Iodice | d28b751 | 2018-07-06 12:59:28 +0100 | [diff] [blame] | 1388 | |
| 1389 | /** This OpenCL kernel performs Winograd filter transform 1x3 when the data layout is NHWC and the output tile is 1x4 |
| 1390 | * |
| 1391 | * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64 |
| 1392 | * @note -DWINOGRAD_FILTER_TRANSFORM_VERTICAL has to be passed at compile time to perform Winograd Filter Transform |
| 1393 | * |
| 1394 | * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 |
| 1395 | * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) |
| 1396 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 1397 | * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 1398 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 1399 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 1400 | * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 1401 | * @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes) |
| 1402 | * @param[in] src_step_w src_stride_w * number of elements along W processed per workitem(in bytes) |
| 1403 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor |
| 1404 | * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr |
| 1405 | * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 1406 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 1407 | * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 1408 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 1409 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 1410 | * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 1411 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 1412 | */ |
| 1413 | __kernel void winograd_filter_transform_1x4_1x3_nhwc( |
| 1414 | TENSOR4D_DECLARATION(src), |
| 1415 | TENSOR3D_DECLARATION(dst)) |
| 1416 | { |
| 1417 | winograd_filter_transform_4x4_3x3_nhwc(src_ptr, |
| 1418 | src_stride_x, |
| 1419 | src_step_x, |
| 1420 | src_stride_y, |
| 1421 | src_step_y, |
| 1422 | src_stride_z, |
| 1423 | src_step_z, |
| 1424 | src_stride_w, |
| 1425 | src_step_w, |
| 1426 | src_offset_first_element_in_bytes, |
| 1427 | dst_ptr, |
| 1428 | dst_stride_x, |
| 1429 | dst_step_x, |
| 1430 | dst_stride_y, |
| 1431 | dst_step_y, |
| 1432 | dst_stride_z, |
| 1433 | dst_step_z, |
| 1434 | dst_offset_first_element_in_bytes); |
| 1435 | } |
| 1436 | |
| 1437 | /** This OpenCL kernel performs Winograd filter transform 1x5 when the data layout is NHWC and the output tile is 1x4 |
| 1438 | * |
| 1439 | * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64 |
| 1440 | * @note -DWINOGRAD_FILTER_TRANSFORM_VERTICAL has to be passed at compile time to perform Winograd Filter Transform |
| 1441 | * |
| 1442 | * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 |
| 1443 | * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) |
| 1444 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 1445 | * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 1446 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 1447 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 1448 | * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 1449 | * @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes) |
| 1450 | * @param[in] src_step_w src_stride_w * number of elements along W processed per workitem(in bytes) |
| 1451 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor |
| 1452 | * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr |
| 1453 | * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 1454 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 1455 | * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 1456 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 1457 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 1458 | * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 1459 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 1460 | */ |
| 1461 | __kernel void winograd_filter_transform_1x4_1x5_nhwc( |
| 1462 | TENSOR4D_DECLARATION(src), |
| 1463 | TENSOR3D_DECLARATION(dst)) |
| 1464 | { |
| 1465 | winograd_filter_transform_4x4_5x5_nhwc(src_ptr, |
| 1466 | src_stride_x, |
| 1467 | src_step_x, |
| 1468 | src_stride_y, |
| 1469 | src_step_y, |
| 1470 | src_stride_z, |
| 1471 | src_step_z, |
| 1472 | src_stride_w, |
| 1473 | src_step_w, |
| 1474 | src_offset_first_element_in_bytes, |
| 1475 | dst_ptr, |
| 1476 | dst_stride_x, |
| 1477 | dst_step_x, |
| 1478 | dst_stride_y, |
| 1479 | dst_step_y, |
| 1480 | dst_stride_z, |
| 1481 | dst_step_z, |
| 1482 | dst_offset_first_element_in_bytes); |
| 1483 | } |
| Giorgio Arena | 149fdf3 | 2018-07-04 17:03:33 +0100 | [diff] [blame] | 1484 | #endif // defined(WINOGRAD_FILTER_TRANSFORM_VERTICAL) |