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) 179 180 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) 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) 198