ramelg01 | 3751569 | 2022-02-26 22:06:20 +0000 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2022 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" // Needed for GET_SPATIAL_IDX() |
| 26 | |
| 27 | #if defined(POOL_AVG) || defined(POOL_L2) |
| 28 | #define POOL_OP(x, y) ((x) + (y)) |
| 29 | #else /* defined(POOL_AVG) || defined(POOL_L2) */ |
| 30 | #define POOL_OP(x, y) (fmax((x), (y))) |
| 31 | #endif /* defined(POOL_AVG) || defined(POOL_L2) */ |
| 32 | |
| 33 | #define SQRT_OP(x) sqrt((x)) |
| 34 | |
| 35 | #if defined(VEC_SIZE) && defined(VEC_SIZE_LEFTOVER) && defined(SRC_WIDTH) && defined(SRC_HEIGHT) && defined(SRC_DEPTH) && defined(DST_CHANNELS) && defined(DST_HEIGHT) && defined(DST_DEPTH) && defined(DST_BATCH_SIZE) && defined(ACC_DATA_TYPE) |
| 36 | |
| 37 | #if defined(POOL_SIZE_X) && defined(POOL_SIZE_Y) && defined(POOL_SIZE_Z) |
| 38 | |
| 39 | /** Performs 3d pooling layer of size equal to MxNXD. This OpenCL kernel can perform the following pooling types: |
| 40 | * -# max, -DPOOL_MAX must be passed at compile time |
| 41 | * -# average, -DPOOL_AVG must be passed at compile time. If padding has to be excluded, -DEXCLUDE_PADDING should be passed at compile time |
| 42 | * -# l2 normalisation, -DPOOL_L2 must be passed at compile time |
| 43 | * |
| 44 | * @note Datatype must be passed at compile type using -DDATA_TYPE e.g. -DDATA_TYPE=half. Supported data types are F32/F16 |
| 45 | * @note Accumulation data type must be passed at compile time using -DACC_DATA_TYPE e.g. -DACC_DATA_TYPE=float |
| 46 | * @note If -DFP_MIXED_PRECISION is passed at compile time, the kernel will use F32 for the partial result |
| 47 | * @note Pool size must be passed at compile time using -DPOOL_SIZE_X, -DPOOL_SIZE_Y, and -DPOOL_SIZE_Z. e.g. -DPOOL_SIZE_X=4, -DPOOL_SIZE_Y=4, -DPOOL_SIZE_Z=2 |
| 48 | * @note Input tensor width, height and depth must be passed at compile time using -DSRC_WIDTH, -DSRC_HEIGHT, and -DSRC_DEPTH |
| 49 | * @note Output tensor height, channels, depth, and batch size must be passed at compile time using -DDST_HEIGHT, -DDST_CHANNELS, -DDST_DEPTH, and -DDST_BATCH_SIZE |
| 50 | * @note Pool strides must be passed at compile time using -DSTRIDE_X, -DSTRIDE_Y and -DSTRIDE_Z which are the steps of the window along the x, y and z directions |
| 51 | * @note Pool pads must be passed at compile time using -DPAD_X, -DPAD_Y, -DPAD_Z |
| 52 | * @note Vector size must be passed at compile time using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16 |
| 53 | * @note Leftover vector size must be passed at compile time using -DVEC_SIZE_LEFTOVER. e.g. -DVEC_SIZE_LEFTOVER=3. It is defined as the remainder between the input's first dimension and VEC_SIZE |
| 54 | * @note The initial value for the pooling operation must be passed at compile time using -DINITIAL_VALUE e.g. -DINITIAL_VALUE=0 |
| 55 | * |
| 56 | * @param[in] input_ptr Pointer to the source tensor. Supported data types: F32/F16 |
| 57 | * @param[in] input_stride_x Stride of the source tensor in X dimension (in bytes) |
| 58 | * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) |
| 59 | * @param[in] input_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 60 | * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) |
| 61 | * @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 62 | * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) |
| 63 | * @param[in] input_stride_w Stride of the source tensor in W dimension (in bytes) |
| 64 | * @param[in] input_step_w input_stride_w * number of elements along W processed per workitem(in bytes) |
| 65 | * @param[in] input_stride_v Stride of the source tensor in V dimension (in bytes) |
| 66 | * @param[in] input_step_v input_stride_v * number of elements along V processed per workitem(in bytes) |
| 67 | * @param[in] input_offset_first_element_in_bytes The offset of the first element in the source tensor |
| 68 | * @param[out] output_ptr Pointer to the destination tensor. Supported data types: same as @p input_ptr |
| 69 | * @param[in] output_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 70 | * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes) |
| 71 | * @param[in] output_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 72 | * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes) |
| 73 | * @param[in] output_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 74 | * @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes) |
| 75 | * @param[in] output_stride_w Stride of the destination tensor in W dimension (in bytes) |
| 76 | * @param[in] output_step_w output_stride_w * number of elements along W processed per workitem(in bytes) |
| 77 | * @param[in] output_stride_v Stride of the destination tensor in V dimension (in bytes) |
| 78 | * @param[in] output_step_v output_stride_v * number of elements along V processed per workitem(in bytes) |
| 79 | * @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 80 | */ |
| 81 | __kernel void pooling_3d_layer_MxN_ndhwc( |
| 82 | TENSOR5D_DECLARATION(input), |
| 83 | TENSOR5D_DECLARATION(output)) |
| 84 | { |
| 85 | // Note: If C is not multiple of VEC_SIZE, we shift back of VEC_SIZE_LEFTOVER elements to compute the leftover elements for get_global_id(0) == 0 |
| 86 | // Note: If C is less than VEC_SIZE, VEC_SIZE should be SHRINKED to the closest smaller VEC_SIZE. This operation is performed on the host side |
| 87 | int idx_out_c = GET_SPATIAL_IDX(0, VEC_SIZE, VEC_SIZE_LEFTOVER); |
| 88 | int idx_out_w = GET_SPATIAL_IDX(1, 1, 0); |
| 89 | |
| 90 | // The depth size dimension and the batch size dimension are collapsed over the height dimension |
| 91 | int idx_out_h = GET_SPATIAL_IDX(2, 1, 0) % DST_HEIGHT; |
| 92 | int idx_out_d = (GET_SPATIAL_IDX(2, 1, 0) / DST_HEIGHT) % DST_DEPTH; |
| 93 | int idx_out_n = (GET_SPATIAL_IDX(2, 1, 0) / DST_HEIGHT) / DST_DEPTH; |
| 94 | |
| 95 | __global unsigned char *in_base_ptr = input_ptr + input_offset_first_element_in_bytes + idx_out_c * sizeof(DATA_TYPE) + idx_out_n * input_stride_v; |
| 96 | |
| 97 | __global unsigned char *out_base_ptr = output_ptr + output_offset_first_element_in_bytes + idx_out_c * sizeof(DATA_TYPE) + idx_out_w * output_stride_y + idx_out_h * output_stride_z + idx_out_d * |
| 98 | output_stride_w + idx_out_n * output_stride_v; |
| 99 | |
| 100 | VEC_DATA_TYPE(ACC_DATA_TYPE, VEC_SIZE) |
| 101 | res0 = INITIAL_VALUE; |
| 102 | |
| 103 | int idx_in_w = idx_out_w * STRIDE_X - (int)PAD_X; |
| 104 | int idx_in_h = idx_out_h * STRIDE_Y - (int)PAD_Y; |
| 105 | int idx_in_d = idx_out_d * STRIDE_Z - (int)PAD_Z; |
| 106 | |
| 107 | // The start of width to consider in calculation should exclude padding |
| 108 | int pool_x_s = max((int)0, -idx_in_w); |
| 109 | // Assumed Symmetric Padding (left padding = right padding = PAD_X), the filter end should be either the pool width or what is remaining from current pos to the (src width + pad right) |
| 110 | int pool_x_e = min((int)POOL_SIZE_X, (int)SRC_WIDTH + PAD_X - idx_in_w); |
| 111 | int pool_y_s = max((int)0, -idx_in_h); |
| 112 | int pool_y_e = min((int)POOL_SIZE_Y, (int)SRC_HEIGHT + PAD_Y - idx_in_h); |
| 113 | int pool_z_s = max((int)0, -idx_in_d); |
| 114 | int pool_z_e = min((int)POOL_SIZE_Z, (int)SRC_DEPTH + PAD_Z - idx_in_d); |
| 115 | |
| 116 | // The filter size with all padding in all directions considered. |
| 117 | int filter_size = pool_z_e * pool_y_e * pool_x_e; |
| 118 | |
| 119 | // The end of width to consider in calculation should exclude PAD_X |
| 120 | pool_x_e = min(pool_x_e, SRC_WIDTH - idx_in_w); |
| 121 | pool_y_e = min(pool_y_e, SRC_HEIGHT - idx_in_h); |
| 122 | pool_z_e = min(pool_z_e, SRC_DEPTH - idx_in_d); |
| 123 | |
| 124 | #if defined(EXCLUDE_PADDING) |
| 125 | filter_size = (pool_z_e - pool_z_s) * (pool_y_e - pool_y_s) * (pool_x_e - pool_x_s); |
| 126 | #endif // defined(EXCLUDE_PADDING) |
| 127 | |
| 128 | #if POOL_SIZE_X == SRC_WIDTH && POOL_SIZE_Y == SRC_HEIGHT && POOL_SIZE_Z == SRC_DEPTH && PAD_X == 0 && PAD_Y == 0 && PAD_Z == 0 |
| 129 | // Global pooling path |
| 130 | for(int z = 0; z < POOL_SIZE_Z; ++z) |
| 131 | { |
| 132 | int depth_offset_src = (z + idx_in_d) * input_stride_w; |
| 133 | for(int y = 0; y < POOL_SIZE_Y; ++y) |
| 134 | { |
| 135 | int height_offset_src = (y + idx_in_h) * input_stride_z; |
| 136 | #pragma unroll 8 |
| 137 | for(int x = 0; x < POOL_SIZE_X; ++x) |
| 138 | { |
| 139 | int width_offset_src = (x + idx_in_w) * input_stride_y; |
| 140 | #else // POOL_SIZE_X == SRC_WIDTH && POOL_SIZE_Y == SRC_HEIGHT && POOL_SIZE_Z == SRC_DEPTH && PAD_X == 0 && PAD_Y == 0 && PAD_Z == 0 |
| 141 | for(int z = pool_z_s; z < pool_z_e; ++z) |
| 142 | { |
| 143 | int depth_offset_src = (z + idx_in_d) * input_stride_w; |
| 144 | for(int y = pool_y_s; y < pool_y_e; ++y) |
| 145 | { |
| 146 | int height_offset_src = (y + idx_in_h) * input_stride_z; |
| 147 | #pragma unroll 8 |
| 148 | for(int x = pool_x_s; x < pool_x_e; ++x) |
| 149 | { |
| 150 | int width_offset_src = (x + idx_in_w) * input_stride_y; |
| 151 | #endif // POOL_SIZE_X == SRC_WIDTH && POOL_SIZE_Y == SRC_HEIGHT && POOL_SIZE_Z == SRC_DEPTH && PAD_X == 0 && PAD_Y == 0 && PAD_Z == 0 |
| 152 | VEC_DATA_TYPE(ACC_DATA_TYPE, VEC_SIZE) |
| 153 | data0; |
| 154 | #if defined(FP_MIXED_PRECISION) |
| 155 | // In case of FP_MIXED_PRECISION, ACC_DATA_TYPE is != DATA_TYPE |
| 156 | data0 = CONVERT(VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(in_base_ptr + width_offset_src + height_offset_src + depth_offset_src)), |
| 157 | VEC_DATA_TYPE(ACC_DATA_TYPE, VEC_SIZE)); |
| 158 | #else // defined(FP_MIXED_PRECISION) |
| 159 | data0 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(in_base_ptr + width_offset_src + height_offset_src + depth_offset_src)); |
| 160 | #endif // defined(FP_MIXED_PRECISION) |
| 161 | |
| 162 | #if defined(POOL_L2) |
| 163 | // Raise to power of 2 for L2 Pooling |
| 164 | data0 *= data0; |
| 165 | #endif // defined(POOL_L2) |
| 166 | res0 = POOL_OP(res0, data0); |
| 167 | } |
| 168 | } |
| 169 | } |
| 170 | |
| 171 | #if defined(POOL_AVG) || defined(POOL_L2) |
| 172 | res0 /= (VEC_DATA_TYPE(ACC_DATA_TYPE, VEC_SIZE))filter_size; |
| 173 | #endif // defined(POOL_AVG) || defined(POOL_L2) |
| 174 | |
| 175 | #if defined(POOL_L2) |
| 176 | // Take square root of the result in L2 pooling |
| 177 | res0 = SQRT_OP(res0); |
| 178 | #endif // defined(POOL_L2) |
| 179 | |
Mohammed Suhail Munshi | 5e549fa | 2022-03-16 11:14:06 +0000 | [diff] [blame] | 180 | VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| 181 | out_q0 = CONVERT(res0, VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)); |
| 182 | |
| 183 | |
| 184 | |
| 185 | // Store result |
| 186 | #if defined(QUANTIZED) |
| 187 | STORE_VECTOR_SELECT(out_q, DATA_TYPE, out_base_ptr, VEC_SIZE, VEC_SIZE_LEFTOVER, (VEC_SIZE_LEFTOVER != 0) && get_global_id(0) == 0); |
| 188 | #elif defined(FP_MIXED_PRECISION) |
ramelg01 | 3751569 | 2022-02-26 22:06:20 +0000 | [diff] [blame] | 189 | VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| 190 | res_converted0 = CONVERT(res0, VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)); |
| 191 | STORE_VECTOR_SELECT(res_converted, DATA_TYPE, out_base_ptr, VEC_SIZE, VEC_SIZE_LEFTOVER, (VEC_SIZE_LEFTOVER != 0) && get_global_id(0) == 0); |
| 192 | #else // defined(FP_MIXED_PRECISION) |
| 193 | STORE_VECTOR_SELECT(res, DATA_TYPE, out_base_ptr, VEC_SIZE, VEC_SIZE_LEFTOVER, (VEC_SIZE_LEFTOVER != 0) && get_global_id(0) == 0); |
| 194 | #endif // defined(FP_MIXED_PRECISION) |
| 195 | } |
| 196 | #endif // defined(POOL_SIZE_X) && defined(POOL_SIZE_Y) && defined(POOL_SIZE_Z) |
| 197 | #endif // defined(VEC_SIZE) && defined(VEC_SIZE_LEFTOVER) && defined(SRC_WIDTH) && defined(SRC_HEIGHT) && defined(SRC_DEPTH) && defined(DST_CHANNELS) && defined(DST_HEIGHT) && defined(DST_DEPTH) && defined(DST_BATCH_SIZE) && defined(ACC_DATA_TYPE) |