Adnan AlSinan | 7075fe2 | 2021-07-05 13:12:52 +0100 | [diff] [blame^] | 1 | /* |
| 2 | * Copyright (c) 2017-2021 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 | #include "tile_helpers.h" |
| 26 | |
| 27 | #define MUL_OP(x, y) ((x) * (y)) |
| 28 | #define ADD_OP(x, y) ((x) + (y)) |
| 29 | #define DIV_OP(x, y) ((x) / (y)) |
| 30 | #define POW_OP(x, y) pow((x), (y)) |
| 31 | #define SQCVT_SAT(a) (a) |
| 32 | |
| 33 | #if defined(WIDTH_SIZE) |
| 34 | /** Apply cross-map normalization. |
| 35 | * |
| 36 | * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short |
| 37 | * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size, e.g. -DVEC_SIZE=16 |
| 38 | * @note The radius should be given as a preprocessor argument using -DRADIUS=size. e.g. -DRADIUS=5 |
| 39 | * @note The number of slices should be given as a preprocessor argument using -DNUM_SLICES=size. e.g. -DNUM_SLICES=192 |
| 40 | * @note Scaling coefficient (= alpha/norm_size), beta and kappa need to be passed at compile time using -DCOEFF, -DALPHA and -DKAPPA |
| 41 | * |
| 42 | * @param[in] input_ptr Pointer to the first source tensor. Supported data types: F16/F32 |
| 43 | * @param[in] input_stride_x Stride of the first source tensor in X dimension (in bytes) |
| 44 | * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) |
| 45 | * @param[in] input_stride_y Stride of the first source tensor in Y dimension (in bytes) |
| 46 | * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) |
| 47 | * @param[in] input_stride_z Stride of the first source tensor in Z dimension (in bytes) |
| 48 | * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) |
| 49 | * @param[in] input_offset_first_element_in_bytes The offset of the first element in the first source tensor |
| 50 | * @param[out] output_ptr Pointer to the destination tensor. Supported data types: same as @p input_ptr |
| 51 | * @param[in] output_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 52 | * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes) |
| 53 | * @param[in] output_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 54 | * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes) |
| 55 | * @param[in] output_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 56 | * @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes) |
| 57 | * @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 58 | */ |
| 59 | __kernel void normalization_layer_cross_map_nhwc(TENSOR3D_DECLARATION(input), |
| 60 | TENSOR3D_DECLARATION(output)) |
| 61 | { |
| 62 | // Offset computation |
| 63 | const uint x_offs = GET_SPATIAL_IDX(0, VEC_SIZE, VEC_SIZE_LEFTOVER); |
| 64 | |
| 65 | // Address computation |
| 66 | __global uchar *input_addr = input_ptr + input_offset_first_element_in_bytes + get_global_id(1) * input_stride_y + get_global_id(2) * input_stride_z; |
| 67 | __global uchar *output_addr = output_ptr + output_offset_first_element_in_bytes + x_offs * sizeof(DATA_TYPE) + get_global_id(1) * output_stride_y + get_global_id(2) * output_stride_z; |
| 68 | |
| 69 | VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| 70 | acc = 0; |
| 71 | const VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| 72 | coeff_v = SQCVT_SAT(COEFF); |
| 73 | const VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| 74 | beta_v = SQCVT_SAT(BETA); |
| 75 | const VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| 76 | kappa_v = SQCVT_SAT(KAPPA); |
| 77 | |
| 78 | const int left_slice = max((int)0, (int)x_offs - (int)RADIUS); |
| 79 | const int right_slice = min((int)WIDTH_SIZE - 1, (int)x_offs + (int)RADIUS); |
| 80 | |
| 81 | for(int i = left_slice; i <= right_slice; ++i) |
| 82 | { |
| 83 | VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| 84 | values = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(input_addr + i * sizeof(DATA_TYPE))); |
| 85 | acc = ADD_OP(acc, MUL_OP(values, values)); |
| 86 | } |
| 87 | |
| 88 | acc = ADD_OP(MUL_OP(acc, coeff_v), kappa_v); |
| 89 | const VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| 90 | normalized = POW_OP(acc, beta_v); |
| 91 | const VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| 92 | normalized_pixel0 = DIV_OP(VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(input_addr + x_offs * sizeof(DATA_TYPE))), normalized); |
| 93 | |
| 94 | STORE_VECTOR_SELECT(normalized_pixel, DATA_TYPE, output_addr, VEC_SIZE, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0); |
| 95 | } |
| 96 | #endif // defined(WIDTH_SIZE) |
| 97 | |
| 98 | #if defined(NUM_SLICES) && defined(DIM1_SIZE) |
| 99 | /** Apply in-map normalization when tensors are in the NHWC data layout format. |
| 100 | * |
| 101 | * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short |
| 102 | * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size, e.g. -DVEC_SIZE=16 |
| 103 | * @note The radius should be given as a preprocessor argument using -DRADIUS=size. e.g. -DRADIUS=5 |
| 104 | * @note The number of slices should be given as a preprocessor argument using -DNUM_SLICES=size. e.g. -DNUM_SLICES=192 |
| 105 | * @note Scaling coefficient (= alpha/norm_size), beta and kappa need to be passed at compile time using -DCOEFF, -DALPHA and -DKAPPA |
| 106 | * |
| 107 | * @param[in] input_ptr Pointer to the first source tensor. Supported data types: F16/F32 |
| 108 | * @param[in] input_stride_x Stride of the first source tensor in X dimension (in bytes) |
| 109 | * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) |
| 110 | * @param[in] input_stride_y Stride of the first source tensor in Y dimension (in bytes) |
| 111 | * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) |
| 112 | * @param[in] input_stride_z Stride of the first source tensor in Z dimension (in bytes) |
| 113 | * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) |
| 114 | * @param[in] input_offset_first_element_in_bytes The offset of the first element in the first source tensor |
| 115 | * @param[out] output_ptr Pointer to the destination tensor. Supported data types: same as @p input_ptr |
| 116 | * @param[in] output_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 117 | * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes) |
| 118 | * @param[in] output_stride_y Stride of the first destination tensor in Y dimension (in bytes) |
| 119 | * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes) |
| 120 | * @param[in] output_stride_z Stride of the first source tensor in Z dimension (in bytes) |
| 121 | * @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes) |
| 122 | * @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 123 | */ |
| 124 | __kernel void normalization_layer_in_map_nhwc(TENSOR3D_DECLARATION(input), |
| 125 | TENSOR3D_DECLARATION(output)) |
| 126 | { |
| 127 | // Offset computation |
| 128 | const uint x_offs = GET_SPATIAL_IDX(0, VEC_SIZE, VEC_SIZE_LEFTOVER); |
| 129 | const int current_cols = get_global_id(1); |
| 130 | const int current_rows = get_global_id(2); |
| 131 | |
| 132 | // Address computation |
| 133 | __global uchar *input_addr = input_ptr + input_offset_first_element_in_bytes + x_offs * sizeof(DATA_TYPE); |
| 134 | __global uchar *output_addr = output_ptr + output_offset_first_element_in_bytes + x_offs * sizeof(DATA_TYPE) + current_cols * output_stride_y + current_rows * output_stride_z; |
| 135 | |
| 136 | VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| 137 | acc = 0; |
| 138 | const VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| 139 | coeff_v = SQCVT_SAT(COEFF); |
| 140 | const VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| 141 | beta_v = SQCVT_SAT(BETA); |
| 142 | const VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| 143 | kappa_v = SQCVT_SAT(KAPPA); |
| 144 | |
| 145 | const int first_col = max(0, current_cols - (int)RADIUS); |
| 146 | const int last_col = min((int)DIM1_SIZE - 1, current_cols + (int)RADIUS); |
| 147 | |
| 148 | #if defined(IN_MAP_2D) |
| 149 | const int first_row = max(0, current_rows - (int)RADIUS); |
| 150 | const int last_row = min((int)NUM_SLICES - 1, current_rows + (int)RADIUS); |
| 151 | #endif /* defined(IN_MAP_2D) */ |
| 152 | |
| 153 | #if defined(IN_MAP_2D) |
| 154 | for(int j = first_row; j <= last_row; ++j) |
| 155 | { |
| 156 | #else // defined(IN_MAP_2D) |
| 157 | const int j = current_rows; |
| 158 | #endif /* defined(IN_MAP_2D) */ |
| 159 | for(int i = first_col; i <= last_col; ++i) |
| 160 | { |
| 161 | VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| 162 | values = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(input_addr + i * input_stride_y + j * input_stride_z)); |
| 163 | acc = ADD_OP(acc, MUL_OP(values, values)); |
| 164 | } |
| 165 | #if defined(IN_MAP_2D) |
| 166 | } |
| 167 | #endif /* defined(IN_MAP_2D) */ |
| 168 | |
| 169 | acc = ADD_OP(MUL_OP(acc, coeff_v), kappa_v); |
| 170 | const VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| 171 | normalized = POW_OP(acc, beta_v); |
| 172 | const VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| 173 | normalized_pixel0 = DIV_OP(VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(input_addr + current_cols * output_stride_y + current_rows *output_stride_z)), normalized); |
| 174 | |
| 175 | STORE_VECTOR_SELECT(normalized_pixel, DATA_TYPE, output_addr, VEC_SIZE, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0); |
| 176 | } |
| 177 | #endif // defined(NUM_SLICES) && defined(DIM1_SIZE) |