Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +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 | #include "helpers.h" |
| 25 | |
| 26 | #if defined USE_F16 |
| 27 | #define MINVAL HALF_MIN |
| 28 | #define SELECT_DATA_TYPE short |
| 29 | #define DATA_TYPE half |
| 30 | #else |
| 31 | #define MINVAL FLT_MIN |
| 32 | #define SELECT_DATA_TYPE int |
| 33 | #define DATA_TYPE float |
| 34 | #endif |
| 35 | |
| 36 | __constant VEC_DATA_TYPE(DATA_TYPE, 16) type_min = (VEC_DATA_TYPE(DATA_TYPE, 16))(MINVAL); |
| 37 | __constant uint16 idx16 = (uint16)(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15); |
| 38 | |
| 39 | /** Identifies the maximum value across the 1st dimension. |
| 40 | * |
| 41 | * @note Datatype must be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short |
| 42 | * @note In case F16 is used -DUSE_HALF must be passed otherwise the kernel will default to used F32. |
| 43 | * @note In case the input is not multiple of 16 -DNON_MULTIPLE_OF_16 must be passed. |
| 44 | * |
| 45 | * @param[in] src_ptr Pointer to the source tensor slice. Supported data types: F16, F32 |
| 46 | * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) |
| 47 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 48 | * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 49 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 50 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor |
| 51 | * @param[out] dst_ptr Pointer to the destination tensor slice. Supported data types: F16, F32 |
| 52 | * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 53 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 54 | * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 55 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 56 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 57 | * @param[in] width Input image width |
| 58 | */ |
| 59 | __kernel void softmax_layer_max( |
| 60 | IMAGE_DECLARATION(src), |
| 61 | IMAGE_DECLARATION(dst), |
| 62 | uint width) |
| 63 | { |
| 64 | Image src = CONVERT_TO_IMAGE_STRUCT(src); |
| 65 | Image dst = CONVERT_TO_IMAGE_STRUCT(dst); |
| 66 | |
| 67 | // Initialize local maximum |
| 68 | VEC_DATA_TYPE(DATA_TYPE, 16) |
| 69 | max_val = (VEC_DATA_TYPE(DATA_TYPE, 16))type_min; |
| 70 | |
| 71 | // Calculate max of row |
| 72 | const uint width4 = width >> 4; |
| 73 | for(uint i = 0; i < width4; i++) |
| 74 | { |
| 75 | VEC_DATA_TYPE(DATA_TYPE, 16) |
| 76 | data = vload16(0, (__global DATA_TYPE *)offset(&src, i << 4, 0)); |
| 77 | max_val = max(data, max_val); |
| 78 | } |
| 79 | |
| 80 | #if defined NON_MULTIPLE_OF_16 |
| 81 | // Handle non multiple of 16 |
| 82 | VEC_DATA_TYPE(DATA_TYPE, 16) |
| 83 | data = vload16(0, (__global DATA_TYPE *)offset(&src, width4 << 4, 0)); |
| 84 | VEC_DATA_TYPE(SELECT_DATA_TYPE, 16) |
| 85 | widx = CONVERT(((uint16)(width4 << 4) + idx16) < width, VEC_DATA_TYPE(SELECT_DATA_TYPE, 16)); |
| 86 | max_val = max(max_val, select(type_min, data, widx)); |
| 87 | #endif |
| 88 | |
| 89 | // Perform max reduction |
| 90 | max_val.s01234567 = max(max_val.s01234567, max_val.s89ABCDEF); |
| 91 | max_val.s0123 = max(max_val.s0123, max_val.s4567); |
| 92 | max_val.s01 = max(max_val.s01, max_val.s23); |
| 93 | max_val.s0 = max(max_val.s0, max_val.s1); |
| 94 | |
| 95 | // Store result |
| 96 | *((__global DATA_TYPE *)dst.ptr) = max_val.s0; |
| 97 | } |
| 98 | |
| 99 | /** Shifts the values of the input tensor by the max calculated in softmax_layer_max kernel, |
| 100 | * then gets the exponent of each element as sums all elements across each row. |
| 101 | * |
| 102 | * @note Datatype must be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short |
| 103 | * @note In case F16 is used -DUSE_HALF must be passed otherwise the kernel will default to used F32. |
| 104 | * @note In case the input is not multiple of 16 -DNON_MULTIPLE_OF_16 must be passed. |
| 105 | * |
| 106 | * @param[in] src_ptr Pointer to the source tensor slice. Supported data types: F16, F32 |
| 107 | * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) |
| 108 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 109 | * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 110 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 111 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor |
| 112 | * @param[in] max_ptr Pointer to the max values tensor slice. Supported data types: F16, F32 |
| 113 | * @param[in] max_stride_x Stride of the max values tensor in X dimension (in bytes) |
| 114 | * @param[in] max_step_x max_stride_x * number of elements along X processed per workitem(in bytes) |
| 115 | * @param[in] max_stride_y Stride of the max values tensor in Y dimension (in bytes) |
| 116 | * @param[in] max_step_y max_stride_y * number of elements along Y processed per workitem(in bytes) |
| 117 | * @param[in] max_offset_first_element_in_bytes The offset of the first element in the max values tensor |
| 118 | * @param[out] dst_ptr Pointer to the destination tensor slice. Supported data types: F16, F32 |
| 119 | * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 120 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 121 | * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 122 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 123 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 124 | * @param[out] sum_ptr Pointer to the sum values tensor slice. Supported data types: F16, F32 |
| 125 | * @param[in] sum_stride_x Stride of the sum values tensor in X dimension (in bytes) |
| 126 | * @param[in] sum_step_x sum_stride_x * number of elements along X processed per workitem(in bytes) |
| 127 | * @param[in] sum_stride_y Stride of the sum values tensor in Y dimension (in bytes) |
| 128 | * @param[in] sum_step_y sum_stride_y * number of elements along Y processed per workitem(in bytes) |
| 129 | * @param[in] sum_offset_first_element_in_bytes The offset of the first element in the sum values tensor |
| 130 | * @param[in] width Input image width |
| 131 | */ |
| 132 | __kernel void softmax_layer_shift_exp_sum( |
| 133 | IMAGE_DECLARATION(src), |
| 134 | IMAGE_DECLARATION(max), |
| 135 | IMAGE_DECLARATION(dst), |
| 136 | IMAGE_DECLARATION(sum), |
| 137 | uint width) |
| 138 | { |
| 139 | Image src = CONVERT_TO_IMAGE_STRUCT(src); |
| 140 | Image dst = CONVERT_TO_IMAGE_STRUCT(dst); |
| 141 | Image max = CONVERT_TO_IMAGE_STRUCT(max); |
| 142 | Image sum = CONVERT_TO_IMAGE_STRUCT(sum); |
| 143 | |
| 144 | // Load max value of 1D logits vector (row) |
| 145 | DATA_TYPE max_val = *((__global DATA_TYPE *)offset(&max, 0, 0)); |
| 146 | |
| 147 | // Set sum vector |
| 148 | VEC_DATA_TYPE(DATA_TYPE, 16) |
| 149 | sum1D = 0; |
| 150 | |
| 151 | // Shift values, exp and sum |
| 152 | const uint width4 = width >> 4; |
| 153 | for(uint i = 0; i < width4; i++) |
| 154 | { |
| 155 | VEC_DATA_TYPE(DATA_TYPE, 16) |
| 156 | data = vload16(0, (__global DATA_TYPE *)offset(&src, i << 4, 0)); |
| 157 | data = exp(data - max_val); |
| 158 | vstore16(data, 0, (__global DATA_TYPE *)offset(&dst, i << 4, 0)); |
| 159 | sum1D += data; |
| 160 | } |
| 161 | |
| 162 | #if defined NON_MULTIPLE_OF_16 |
| 163 | // Handle non multiple of 16 |
| 164 | VEC_DATA_TYPE(DATA_TYPE, 16) |
| 165 | data = vload16(0, (__global DATA_TYPE *)offset(&src, width4 << 4, 0)); |
| 166 | data = exp(data - max_val); |
| 167 | VEC_DATA_TYPE(SELECT_DATA_TYPE, 16) |
| 168 | widx = CONVERT(((uint16)(width4 << 4) + idx16) < width, VEC_DATA_TYPE(SELECT_DATA_TYPE, 16)); |
| 169 | data = select(0, data, widx); |
| 170 | vstore16(data, 0, (__global DATA_TYPE *)offset(&dst, width4 << 4, 0)); |
| 171 | sum1D += data; |
| 172 | #endif |
| 173 | |
| 174 | // Perform min/max reduction |
| 175 | sum1D.s01234567 = sum1D.s01234567 + sum1D.s89ABCDEF; |
| 176 | sum1D.s0123 = sum1D.s0123 + sum1D.s4567; |
| 177 | sum1D.s01 = sum1D.s01 + sum1D.s23; |
| 178 | sum1D.s0 = sum1D.s0 + sum1D.s1; |
| 179 | |
| 180 | // Calculate and store result |
| 181 | *((__global DATA_TYPE *)sum.ptr) = sum1D.s0; |
| 182 | } |
| 183 | |
| 184 | /** Divides all the values of the input tensor by the sum calculated from softmax_layer_shift_exp_sum kernel. |
| 185 | * |
| 186 | * @note Datatype must be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short |
| 187 | * |
| 188 | * @param[in] src_ptr Pointer to the source tensor slice. Supported data types: F16, F32 |
| 189 | * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) |
| 190 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 191 | * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 192 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 193 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor |
| 194 | * @param[in] sum_ptr Pointer to the sum values tensor slice. Supported data types: F16, F32 |
| 195 | * @param[in] sum_stride_x Stride of the sum values tensor in X dimension (in bytes) |
| 196 | * @param[in] sum_step_x sum_stride_x * number of elements along X processed per workitem(in bytes) |
| 197 | * @param[in] sum_stride_y Stride of the sum values tensor in Y dimension (in bytes) |
| 198 | * @param[in] sum_step_y sum_stride_y * number of elements along Y processed per workitem(in bytes) |
| 199 | * @param[in] sum_offset_first_element_in_bytes The offset of the first element in the sum values tensor |
| 200 | * @param[out] dst_ptr Pointer to the destination tensor slice. Supported data types: F16, F32 |
| 201 | * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 202 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 203 | * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 204 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 205 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 206 | */ |
| 207 | __kernel void softmax_layer_norm( |
| 208 | IMAGE_DECLARATION(src), |
| 209 | IMAGE_DECLARATION(sum), |
| 210 | IMAGE_DECLARATION(dst)) |
| 211 | { |
| 212 | Image src = CONVERT_TO_IMAGE_STRUCT(src); |
| 213 | Image dst = CONVERT_TO_IMAGE_STRUCT(dst); |
| 214 | Image sum = CONVERT_TO_IMAGE_STRUCT_NO_STEP(sum); |
| 215 | |
| 216 | // Load max value of 1D logits vector (row) |
| 217 | DATA_TYPE sum_val = *((__global DATA_TYPE *)offset(&sum, 0, get_global_id(1))); |
| 218 | VEC_DATA_TYPE(DATA_TYPE, 16) |
| 219 | data = vload16(0, (__global DATA_TYPE *)offset(&src, 0, 0)); |
| 220 | vstore16(data / sum_val, 0, (__global DATA_TYPE *)offset(&dst, 0, 0)); |
| 221 | } |