SiCong Li | c51b72f | 2017-07-28 14:46:20 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2016, 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 | #include "helpers.h" |
| 25 | |
| 26 | #if STRIDE_X == 3 |
| 27 | #define INPUT_PIXEL_STR(data_size) extract_input_stride3_##data_size |
| 28 | #define INPUT_PIXEL(data_size) INPUT_PIXEL_STR(data_size) |
| 29 | #elif STRIDE_X == 2 |
| 30 | #define INPUT_PIXEL(data_size) extract_input_stride2 |
| 31 | #elif STRIDE_X == 1 |
| 32 | #define INPUT_PIXEL(data_size) extract_input_stride1 |
| 33 | #else /* STRIDE_X not equals 1, 2 or 3 */ |
| 34 | #error "Only support strides 1, 2 and 3" |
| 35 | #endif /* STRIDE_X == 3 */ |
| 36 | |
| 37 | /** Extracts a 1D horizontal vector from the input tensor with stride as 1. |
| 38 | * |
| 39 | * @param[in] input_pixel Pointer to the first pixel. |
| 40 | * |
| 41 | * @return extracted input pixels. |
| 42 | */ |
| 43 | inline VEC_DATA_TYPE(DATA_TYPE, 8) extract_input_stride1(__global const DATA_TYPE *input_pixel) |
| 44 | { |
| 45 | return vload8(0, input_pixel); |
| 46 | } |
| 47 | |
| 48 | /** Extracts a 1D horizontal vector from the input tensor with stride as 2. |
| 49 | * |
| 50 | * @param[in] input_pixel Pointer to the first pixel. |
| 51 | * |
| 52 | * @return extracted input pixels. |
| 53 | */ |
| 54 | inline VEC_DATA_TYPE(DATA_TYPE, 8) extract_input_stride2(__global const DATA_TYPE *input_pixel) |
| 55 | { |
| 56 | VEC_DATA_TYPE(DATA_TYPE, 16) |
| 57 | temp = vload16(0, input_pixel); |
| 58 | return temp.s02468ace; |
| 59 | } |
| 60 | |
| 61 | /** Extracts a 1D horizontal vector from the input tensor with stride as 3 and 32-bit data size. |
| 62 | * |
| 63 | * @param[in] input_pixel Pointer to the first pixel. |
| 64 | * |
| 65 | * @return extracted input pixels. |
| 66 | */ |
| 67 | inline VEC_DATA_TYPE(DATA_TYPE, 8) extract_input_stride3_32(__global const DATA_TYPE *input_pixel) |
| 68 | { |
| 69 | VEC_DATA_TYPE(DATA_TYPE, 4) |
| 70 | temp1 = vload4(0, input_pixel); |
| 71 | VEC_DATA_TYPE(DATA_TYPE, 4) |
| 72 | temp2 = vload4(0, input_pixel + 6); |
| 73 | VEC_DATA_TYPE(DATA_TYPE, 4) |
| 74 | temp3 = vload4(0, input_pixel + 12); |
| 75 | VEC_DATA_TYPE(DATA_TYPE, 4) |
| 76 | temp4 = vload4(0, input_pixel + 18); |
| 77 | return (VEC_DATA_TYPE(DATA_TYPE, 8))(temp1.s03, temp2.s03, temp3.s03, temp4.s03); |
| 78 | } |
| 79 | |
| 80 | /** Extracts a 1D horizontal vector from the input tensor with stride as 3 and 16-bit data size. |
| 81 | * |
| 82 | * @param[in] input_pixel Pointer to the first pixel. |
| 83 | * |
| 84 | * @return extracted input pixels. |
| 85 | */ |
| 86 | inline VEC_DATA_TYPE(DATA_TYPE, 8) extract_input_stride3_16(__global const DATA_TYPE *input_pixel) |
| 87 | { |
| 88 | VEC_DATA_TYPE(DATA_TYPE, 8) |
| 89 | temp1 = vload8(0, input_pixel); |
| 90 | VEC_DATA_TYPE(DATA_TYPE, 8) |
| 91 | temp2 = vload8(0, input_pixel + 8); |
| 92 | VEC_DATA_TYPE(DATA_TYPE, 8) |
| 93 | temp3 = vload8(0, input_pixel + 16); |
| 94 | return (VEC_DATA_TYPE(DATA_TYPE, 8))(temp1.s036, temp2.s147, temp3.s25); |
| 95 | } |
| 96 | |
| 97 | /** Extracts a 1D horizontal vector from the input tensor with stride as 3 and 8-bit data size. |
| 98 | * |
| 99 | * @param[in] input_pixel Pointer to the first pixel. |
| 100 | * |
| 101 | * @return extracted input pixels. |
| 102 | */ |
| 103 | inline VEC_DATA_TYPE(DATA_TYPE, 8) extract_input_stride3_8(__global const DATA_TYPE *input_pixel) |
| 104 | { |
| 105 | VEC_DATA_TYPE(DATA_TYPE, 16) |
| 106 | temp1 = vload16(0, input_pixel); |
| 107 | VEC_DATA_TYPE(DATA_TYPE, 16) |
| 108 | temp2 = vload16(0, input_pixel + 12); |
| 109 | return (VEC_DATA_TYPE(DATA_TYPE, 8))(temp1.s0369, temp2.s0369); |
| 110 | } |
| 111 | |
| 112 | /** This kernel performs a direct convolution to convolve the low three dimensions. |
| 113 | * |
| 114 | * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float |
| 115 | * @note The data size must be passed at compile time using -DDATA_SIZE e.g. -DDATA_SIZE=32 |
Gian Marco Iodice | 5cb4d6a | 2017-08-08 10:53:00 +0100 | [diff] [blame^] | 116 | * @note The convolution stride x must be passed at compile time using -DSTRIDE_X e.g. -DSTRIDE_X=1 |
| 117 | * @note The third dimensions of the weights tensors must be passed at compile time using -DWEIGHTS_DEPTH |
SiCong Li | c51b72f | 2017-07-28 14:46:20 +0100 | [diff] [blame] | 118 | * @note In case biases will be added to the convolution -DHAS_BIAS has to be passed to append the final matrix with 1 in each row. |
| 119 | * |
Gian Marco Iodice | 5cb4d6a | 2017-08-08 10:53:00 +0100 | [diff] [blame^] | 120 | * @param[in] src_ptr Pointer to the source tensor. Supported data types: F16/F32 |
SiCong Li | c51b72f | 2017-07-28 14:46:20 +0100 | [diff] [blame] | 121 | * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) |
| 122 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 123 | * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 124 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 125 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 126 | * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) |
| 127 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor |
| 128 | * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr |
| 129 | * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 130 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 131 | * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 132 | * @param[in] dst_step_y dst_stride_y * number of elements along Z processed per workitem(in bytes) |
| 133 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 134 | * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) |
| 135 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 136 | * @param[out] weights_ptr Pointer to the weights tensor. Supported data types: same as @p weights_ptr |
| 137 | * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes) |
| 138 | * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes) |
| 139 | * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes) |
| 140 | * @param[in] weights_step_y weights_stride_y * number of elements along y processed per workitem(in bytes) |
| 141 | * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes) |
| 142 | * @param[in] weights_step_z weights_stride_z * number of elements along Z processed per workitem(in bytes) |
| 143 | * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor |
| 144 | * @param[in] biases_ptr Pointer to the biases tensor. Same as @p src_ptr |
| 145 | * @param[in] biases_stride_x Stride of the biases tensor in X dimension (in bytes) |
| 146 | * @param[in] biases_step_x biases_stride_x * number of elements along X processed per workitem(in bytes) |
| 147 | * @param[in] biases_offset_first_element_in_bytes The offset of the first element in the biases tensor |
Gian Marco Iodice | 5cb4d6a | 2017-08-08 10:53:00 +0100 | [diff] [blame^] | 148 | * @param[in] weights_stride_w Stride of the weights tensor in the 4th dimension |
SiCong Li | c51b72f | 2017-07-28 14:46:20 +0100 | [diff] [blame] | 149 | */ |
Gian Marco Iodice | 5cb4d6a | 2017-08-08 10:53:00 +0100 | [diff] [blame^] | 150 | #if defined(DATA_TYPE) && defined(DATA_SIZE) && defined(STRIDE_X) && defined(WEIGHTS_DEPTH) |
SiCong Li | c51b72f | 2017-07-28 14:46:20 +0100 | [diff] [blame] | 151 | __kernel void direct_convolution1x1( |
| 152 | TENSOR3D_DECLARATION(src), |
| 153 | TENSOR3D_DECLARATION(dst), |
| 154 | TENSOR3D_DECLARATION(weights), |
| 155 | #ifdef HAS_BIAS |
| 156 | VECTOR_DECLARATION(biases), |
| 157 | #endif /* defined(HAS_BIAS) */ |
Gian Marco Iodice | 5cb4d6a | 2017-08-08 10:53:00 +0100 | [diff] [blame^] | 158 | unsigned int weights_stride_w) |
SiCong Li | c51b72f | 2017-07-28 14:46:20 +0100 | [diff] [blame] | 159 | { |
| 160 | Image src = CONVERT_TO_IMAGE_STRUCT(src); |
| 161 | Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(weights); |
| 162 | Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst); |
| 163 | |
| 164 | #ifdef HAS_BIAS |
| 165 | Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases); |
| 166 | #endif /* defined(HAS_BIAS) */ |
| 167 | |
| 168 | VEC_DATA_TYPE(DATA_TYPE, 8) |
| 169 | pixels = 0; |
| 170 | |
| 171 | const uint z_index = get_global_id(2); |
| 172 | |
| 173 | weights.ptr += z_index * weights_stride_w; |
| 174 | |
Gian Marco Iodice | 5cb4d6a | 2017-08-08 10:53:00 +0100 | [diff] [blame^] | 175 | for(int d = 0; d < WEIGHTS_DEPTH; ++d) |
SiCong Li | c51b72f | 2017-07-28 14:46:20 +0100 | [diff] [blame] | 176 | { |
| 177 | DATA_TYPE weight = *(__global DATA_TYPE *)weights.ptr; |
| 178 | VEC_DATA_TYPE(DATA_TYPE, 8) |
| 179 | input_pixel = INPUT_PIXEL(DATA_SIZE)((__global DATA_TYPE *)src.ptr); |
| 180 | pixels += weight * input_pixel; |
| 181 | src.ptr += src_stride_z; |
| 182 | weights.ptr += weights_stride_z; |
| 183 | } |
| 184 | |
| 185 | #ifdef HAS_BIAS |
| 186 | pixels += (VEC_DATA_TYPE(DATA_TYPE, 8)) * ((__global DATA_TYPE *)(vector_offset(&biases, z_index))); |
| 187 | #endif /* defined(HAS_BIAS) */ |
| 188 | |
| 189 | vstore8(pixels, 0, (__global DATA_TYPE *)dst.ptr); |
| 190 | } |
Gian Marco Iodice | 5cb4d6a | 2017-08-08 10:53:00 +0100 | [diff] [blame^] | 191 | #endif // defined(DATA_TYPE) && defined(DATA_SIZE) && defined(STRIDE_X) && defined(WEIGHTS_DEPTH) |