Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame^] | 1 | /* |
| 2 | * Copyright (c) 2017 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 | |
| 25 | layout(local_size_x = LOCAL_SIZE_X, local_size_y = LOCAL_SIZE_Y, local_size_z = LOCAL_SIZE_Z) in; |
| 26 | |
| 27 | #include "helpers.h" |
| 28 | |
| 29 | layout(std140) uniform shader_params |
| 30 | { |
| 31 | TENSOR3D_PARAM_DECLARATION(src); |
| 32 | TENSOR3D_PARAM_DECLARATION(dst); |
| 33 | TENSOR3D_PARAM_DECLARATION(weights); |
| 34 | #ifdef BIAS |
| 35 | VECTOR_PARAM_DECLARATION(biases); |
| 36 | #endif /* BIAS */ |
| 37 | uint weights_stride_w; |
| 38 | uint weights_depth; |
| 39 | }; |
| 40 | |
| 41 | #if defined(DATA_TYPE_FP32) |
| 42 | precision highp float; |
| 43 | |
| 44 | BUFFER_DECLARATION(src, 1, float, readonly); |
| 45 | BUFFER_DECLARATION(dst, 2, float, writeonly); |
| 46 | BUFFER_DECLARATION(weights, 3, float, readonly); |
| 47 | #ifdef BIAS |
| 48 | BUFFER_DECLARATION(biases, 4, float, readonly); |
| 49 | #endif /* BIAS */ |
| 50 | |
| 51 | /** This kernel performs a direct convolution to convolve the low three dimensions. |
| 52 | * |
| 53 | * @note The data type must be passed at compile time using "#define DATA_TYPE_FP32" |
| 54 | * @note The convolution stride x must be passed at compile time using "#define STRIDE_X" e.g. "#define STRIDE_X 1" |
| 55 | * @note In case biases will be added to the convolution "#define HAS_BIAS" has to be passed to append the final matrix with 1 in each row. |
| 56 | * |
| 57 | * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 |
| 58 | * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) |
| 59 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 60 | * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 61 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 62 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 63 | * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 64 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor |
| 65 | * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr |
| 66 | * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 67 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 68 | * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 69 | * @param[in] dst_step_y dst_stride_y * number of elements along Z processed per workitem(in bytes) |
| 70 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 71 | * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) |
| 72 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 73 | * @param[out] weights_ptr Pointer to the weights tensor. Supported data types: same as @p src_ptr |
| 74 | * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes) |
| 75 | * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes) |
| 76 | * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes) |
| 77 | * @param[in] weights_step_y weights_stride_y * number of elements along y processed per workitem(in bytes) |
| 78 | * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes) |
| 79 | * @param[in] weights_step_z weights_stride_z * number of elements along Z processed per workitem(in bytes) |
| 80 | * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor |
| 81 | * @param[in] biases_ptr Pointer to the biases tensor. Same as @p src_ptr |
| 82 | * @param[in] biases_stride_x Stride of the biases tensor in X dimension (in bytes) |
| 83 | * @param[in] biases_step_x biases_stride_x * number of elements along X processed per workitem(in bytes) |
| 84 | * @param[in] biases_offset_first_element_in_bytes The offset of the first element in the biases tensor |
| 85 | * @param[in] weights_stride_w Stride of the weights tensor in the 4th dimension |
| 86 | * @param[in] weights_depth The third dimensions of the weights tensors |
| 87 | */ |
| 88 | void main() |
| 89 | { |
| 90 | Image src = CONVERT_TO_IMAGE_STRUCT(src); |
| 91 | Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(weights); |
| 92 | Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst); |
| 93 | |
| 94 | #ifdef BIAS |
| 95 | Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases); |
| 96 | #endif /* BIAS */ |
| 97 | |
| 98 | float pixels = CONVERT(0, float); |
| 99 | uint z_index = gl_GlobalInvocationID.z; |
| 100 | weights.current_offset += z_index * weights_stride_w >> 2; |
| 101 | float temp; |
| 102 | float temp_weight; |
| 103 | |
| 104 | for(int d = 0; d < int(weights_depth); ++d) |
| 105 | { |
| 106 | temp = LOAD4(src, CURRENT_OFFSET(src)); |
| 107 | temp_weight = LOAD4(weights, CURRENT_OFFSET(weights)); |
| 108 | pixels += temp * temp_weight; |
| 109 | |
| 110 | src.current_offset += (src_stride_z >> 2); |
| 111 | weights.current_offset += (weights_stride_z >> 2); |
| 112 | } |
| 113 | |
| 114 | #ifdef BIAS |
| 115 | pixels += LOAD4(biases, vector_offset(biases, int(z_index))); |
| 116 | #endif /* BIAS */ |
| 117 | |
| 118 | STORE4(dst, CURRENT_OFFSET(dst), pixels); |
| 119 | } |
| 120 | #elif defined(DATA_TYPE_FP16) |
| 121 | precision mediump float; |
| 122 | |
| 123 | BUFFER_DECLARATION(src, 1, uvec4, readonly); |
| 124 | BUFFER_DECLARATION(dst, 2, uvec4, writeonly); |
| 125 | BUFFER_DECLARATION(weights, 3, uint, readonly); |
| 126 | #ifdef BIAS |
| 127 | BUFFER_DECLARATION(biases, 4, uint, readonly); |
| 128 | #endif /* BIAS */ |
| 129 | |
| 130 | #if STRIDE_X == 2 |
| 131 | #define CONVOLVE(s, w) convolve_stride2(s, w) |
| 132 | #elif STRIDE_X == 1 /* STRIDE_X == 1 */ |
| 133 | #define CONVOLVE(s, w) convolve_stride1(s, w) |
| 134 | #else /* STRIDE_X not equals 1 or 2 */ |
| 135 | #error STRIDE_X larger than 2 is not supported |
| 136 | #endif /* STRIDE_X == 2 */ |
| 137 | |
| 138 | vec4[2] convolve_stride1(Image src, float w) |
| 139 | { |
| 140 | uvec4 packed_s; |
| 141 | vec4 s[2]; |
| 142 | |
| 143 | GC_LOAD1_2D_OFFSET(packed_s, src, 0, 0); |
| 144 | |
| 145 | s[0] = vec4(unpackHalf2x16(packed_s.x), unpackHalf2x16(packed_s.y)); |
| 146 | s[1] = vec4(unpackHalf2x16(packed_s.z), unpackHalf2x16(packed_s.w)); |
| 147 | |
| 148 | s[0] *= w; |
| 149 | s[1] *= w; |
| 150 | |
| 151 | return s; |
| 152 | } |
| 153 | |
| 154 | vec4[2] convolve_stride2(Image src, float w) |
| 155 | { |
| 156 | uvec4 packed_s; |
| 157 | vec4 s[2]; |
| 158 | vec4 r[2]; |
| 159 | |
| 160 | GC_LOAD1_2D_OFFSET(packed_s, src, 0, 0); |
| 161 | s[0] = vec4(unpackHalf2x16(packed_s.x), unpackHalf2x16(packed_s.y)); |
| 162 | s[1] = vec4(unpackHalf2x16(packed_s.z), unpackHalf2x16(packed_s.w)); |
| 163 | |
| 164 | r[0] = vec4(s[0].xz, s[1].xz); |
| 165 | |
| 166 | GC_LOAD1_2D_OFFSET(packed_s, src, 8, 0); |
| 167 | s[0] = vec4(unpackHalf2x16(packed_s.x), unpackHalf2x16(packed_s.y)); |
| 168 | s[1] = vec4(unpackHalf2x16(packed_s.z), unpackHalf2x16(packed_s.w)); |
| 169 | |
| 170 | r[1] = vec4(s[0].xz, s[1].xz); |
| 171 | |
| 172 | r[0] *= w; |
| 173 | r[1] *= w; |
| 174 | |
| 175 | return r; |
| 176 | } |
| 177 | |
| 178 | /** This kernel performs a direct convolution to convolve the low three dimensions. |
| 179 | * |
| 180 | * @note The data type must be passed at compile time using "#define DATA_TYPE_FP16" |
| 181 | * @note The convolution stride x must be passed at compile time using "#define STRIDE_X" e.g. "#define STRIDE_X 1" |
| 182 | * @note In case biases will be added to the convolution "#define HAS_BIAS" has to be passed to append the final matrix with 1 in each row. |
| 183 | * |
| 184 | * @param[in] src_ptr Pointer to the source tensor. Supported data types: F16 |
| 185 | * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) |
| 186 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 187 | * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 188 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 189 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 190 | * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 191 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor |
| 192 | * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr |
| 193 | * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 194 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 195 | * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 196 | * @param[in] dst_step_y dst_stride_y * number of elements along Z processed per workitem(in bytes) |
| 197 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 198 | * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) |
| 199 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 200 | * @param[out] weights_ptr Pointer to the weights tensor. Supported data types: same as @p src_ptr |
| 201 | * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes) |
| 202 | * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes) |
| 203 | * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes) |
| 204 | * @param[in] weights_step_y weights_stride_y * number of elements along y processed per workitem(in bytes) |
| 205 | * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes) |
| 206 | * @param[in] weights_step_z weights_stride_z * number of elements along Z processed per workitem(in bytes) |
| 207 | * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor |
| 208 | * @param[in] biases_ptr Pointer to the biases tensor. Same as @p src_ptr |
| 209 | * @param[in] biases_stride_x Stride of the biases tensor in X dimension (in bytes) |
| 210 | * @param[in] biases_step_x biases_stride_x * number of elements along X processed per workitem(in bytes) |
| 211 | * @param[in] biases_offset_first_element_in_bytes The offset of the first element in the biases tensor |
| 212 | * @param[in] weights_stride_w Stride of the weights tensor in the 4th dimension |
| 213 | * @param[in] weights_depth The third dimensions of the weights tensors |
| 214 | */ |
| 215 | void main() |
| 216 | { |
| 217 | Image src = GC_CONVERT_TO_IMAGE_STRUCT(src); |
| 218 | Tensor3D weights = GC_CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(weights); |
| 219 | Tensor3D dst = GC_CONVERT_TO_TENSOR3D_STRUCT(dst); |
| 220 | |
| 221 | #ifdef BIAS |
| 222 | Vector biases = GC_CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases); |
| 223 | #endif /* BIAS */ |
| 224 | |
| 225 | vec4 pixels[2]; |
| 226 | pixels[0] = vec4(0.f); |
| 227 | pixels[1] = vec4(0.f); |
| 228 | |
| 229 | uint z_index = gl_GlobalInvocationID.z; |
| 230 | |
| 231 | weights.current_offset += z_index * weights_stride_w; |
| 232 | |
| 233 | uint packed_w; |
| 234 | float w; |
| 235 | |
| 236 | for(int d = 0; d < int(weights_depth); ++d) |
| 237 | { |
| 238 | GC_LOAD1_3D_OFFSET(packed_w, weights, 0, 0, 0); |
| 239 | w = unpackHalf2x16(packed_w).x; |
| 240 | |
| 241 | vec4 r[2] = CONVOLVE(src, w); |
| 242 | pixels[0] += r[0]; |
| 243 | pixels[1] += r[1]; |
| 244 | |
| 245 | src.current_offset += src_stride_z; |
| 246 | weights.current_offset += weights_stride_z; |
| 247 | } |
| 248 | |
| 249 | #ifdef BIAS |
| 250 | uint packed_b; |
| 251 | float b; |
| 252 | |
| 253 | GC_LOAD1_1D_OFFSET(packed_b, biases, z_index); |
| 254 | |
| 255 | if(z_index % uint(2) == uint(0)) |
| 256 | { |
| 257 | b = unpackHalf2x16(packed_b).x; |
| 258 | } |
| 259 | else |
| 260 | { |
| 261 | b = unpackHalf2x16(packed_b).y; |
| 262 | } |
| 263 | |
| 264 | pixels[0] += vec4(b); |
| 265 | pixels[1] += vec4(b); |
| 266 | #endif /* BIAS */ |
| 267 | |
| 268 | uvec4 packed_d; |
| 269 | packed_d = uvec4(packHalf2x16(pixels[0].xy), packHalf2x16(pixels[0].zw), |
| 270 | packHalf2x16(pixels[1].xy), packHalf2x16(pixels[1].zw)); |
| 271 | GC_STORE1_3D_OFFSET(packed_d, dst, 0, 0, 0); |
| 272 | } |
| 273 | #else /* DATA_TYPE_FP32 */ |
| 274 | #error Data type not supported |
| 275 | #endif /* DATA_TYPE_FP32 */ |