| SiCong Li | 3e36369 | 2017-07-04 15:02:10 +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 DATA_SIZE == 32 |
| 27 | #define VEC_SIZE 4 |
| 28 | #define VEC_MAX vec4_max |
| 29 | #elif DATA_SIZE == 16 |
| 30 | #define VEC_SIZE 8 |
| 31 | #define VEC_MAX vec8_max |
| 32 | #else /* DATA_SIZE not equals 32 or 16 */ |
| 33 | #error "Unsupported data size" |
| 34 | #endif /* DATA_SIZE == 32 */ |
| 35 | |
| 36 | inline DATA_TYPE vec4_max(VEC_DATA_TYPE(DATA_TYPE, 4) vec) |
| 37 | { |
| 38 | VEC_DATA_TYPE(DATA_TYPE, 2) |
| 39 | temp = fmax(vec.lo, vec.hi); |
| 40 | return fmax(temp.x, temp.y); |
| 41 | } |
| 42 | |
| 43 | inline DATA_TYPE vec8_max(VEC_DATA_TYPE(DATA_TYPE, 8) vec) |
| 44 | { |
| 45 | VEC_DATA_TYPE(DATA_TYPE, 4) |
| 46 | temp = fmax(vec.lo, vec.hi); |
| 47 | return vec4_max(temp); |
| 48 | } |
| 49 | |
| 50 | /** Performs a roi pooling on a single output pixel. |
| 51 | * |
| 52 | * @param[in] input Pointer to input Tensor3D struct. |
| 53 | * @param[in] region_start_x Start x index projected onto the input tensor. |
| 54 | * @param[in] region_end_x End x index projected onto the input tensor. |
| 55 | * @param[in] region_start_y Start y index projected onto the input tensor. |
| 56 | * @param[in] region_end_y End y index projected onto the input tensor. |
| 57 | * @param[in] pz z index of the input tensor. |
| 58 | * |
| 59 | * @return A max pooled value from the region specified in the input tensor. |
| 60 | */ |
| 61 | inline DATA_TYPE roi_pool_1x1(const Tensor3D *input, int region_start_x, int region_end_x, int region_start_y, int region_end_y, int pz) |
| 62 | { |
| 63 | // Iterate through the pooling region |
| 64 | if((region_end_x <= region_start_x) || (region_end_y <= region_start_y)) |
| 65 | { |
| 66 | return (DATA_TYPE)0; |
| 67 | } |
| 68 | else |
| 69 | { |
| 70 | int num_iter = (int)((region_end_x - region_start_x) / VEC_SIZE); |
| 71 | VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| 72 | curr_max = (VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))(-FLT_MAX); |
| 73 | for(int j = region_start_y; j < region_end_y; ++j) |
| 74 | { |
| 75 | int i = region_start_x; |
| 76 | for(; i < region_start_x + num_iter * VEC_SIZE; i += VEC_SIZE) |
| 77 | { |
| 78 | VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| 79 | val = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)tensor3D_offset(input, i, j, pz)); |
| 80 | curr_max = fmax(val, curr_max); |
| 81 | } |
| 82 | for(; i < region_end_x; ++i) |
| 83 | { |
| 84 | DATA_TYPE val = *(__global DATA_TYPE *)tensor3D_offset(input, i, j, pz); |
| 85 | curr_max = fmax(curr_max, val); |
| 86 | } |
| 87 | } |
| 88 | return (DATA_TYPE)VEC_MAX(curr_max); |
| 89 | } |
| 90 | } |
| 91 | |
| 92 | /** Performs a roi pooling function. |
| 93 | * |
| 94 | * @note Datatype must be passed using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types are F16, F32; |
| 95 | * @note Datasize must be passed using -DDATA_SIZE e.g. -DDATA_SIZE=32; |
| 96 | * @note Input dimensions must be passed using -DMAX_DIM_X, -DMAX_DIM_Y and -DMAX_DIM_Z; |
| 97 | * @note Pooled region dimensions must be passed using -DPOOLED_DIM_X and -DPOOLED_DIM_Y; |
| 98 | * @note Spatial scale must be passed using -DSPATIAL_SCALE; |
| 99 | * |
| 100 | * @param[in] input_ptr Pointer to the source image. Supported data types: F16, F32 |
| 101 | * @param[in] input_stride_x Stride of the source image in X dimension (in bytes) |
| 102 | * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) |
| 103 | * @param[in] input_stride_y Stride of the source image in Y dimension (in bytes) |
| 104 | * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) |
| 105 | * @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 106 | * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) |
| 107 | * @param[in] input_offset_first_element_in_bytes The offset of the first element in the pooled region of the source image as specifed by ROI |
| 108 | * @param[in] rois_ptr Pointer to the rois array. Layout: {x, y, width, height, batch_indx} |
| 109 | * @param[in] rois_stride_x Stride of the rois array in X dimension (in bytes) |
| 110 | * @param[in] rois_step_x rois_stride_x * number of elements along X processed per workitem(in bytes) |
| 111 | * @param[in] rois_offset_first_element_in_bytes The offset of the first element in the rois array |
| 112 | * @param[out] output_ptr Pointer to the destination image. Supported data types: F16, F32 |
| 113 | * @param[in] output_stride_x Stride of the destination image in X dimension (in bytes) |
| 114 | * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes) |
| 115 | * @param[in] output_stride_y Stride of the destination image in Y dimension (in bytes) |
| 116 | * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes) |
| 117 | * @param[in] output_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 118 | * @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes) |
| 119 | * @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination image |
| 120 | * @param[in] input_stride_w Stride of the source image in W dimension (in bytes) |
| 121 | * @param[in] output_stride_w Stride of the destination image in W dimension (in bytes) |
| 122 | */ |
| 123 | __kernel void roi_pooling_layer( |
| 124 | TENSOR3D_DECLARATION(input), |
| 125 | VECTOR_DECLARATION(rois), |
| 126 | TENSOR3D_DECLARATION(output), |
| 127 | unsigned int input_stride_w, unsigned int output_stride_w) |
| 128 | { |
| 129 | // Get pixels pointer |
| 130 | Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(input); |
| 131 | Vector rois = CONVERT_TO_VECTOR_STRUCT_NO_STEP(rois); |
| 132 | Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(output); |
| 133 | |
| 134 | const int px = get_global_id(0); |
| 135 | const int py = get_global_id(1); |
| 136 | const int pw = get_global_id(2); |
| 137 | |
| 138 | // Load roi parameters |
| 139 | // roi is laid out as follows: |
| 140 | // { x, y, width, height, batch_index } |
| 141 | const ushort8 roi = vload8(0, (__global ushort *)vector_offset(&rois, pw)); |
| 142 | const int2 roi_anchor = convert_int2_sat(round(convert_float2(roi.s01) * (float)SPATIAL_SCALE)); |
| 143 | const int2 roi_dims = convert_int2_sat(fmax(round(convert_float2(roi.s23) * (float)SPATIAL_SCALE), 1.f)); |
| 144 | |
| 145 | // Determine pooled region in input image to pooled region in output image ratio |
| 146 | const float2 pool_region_ratio = convert_float2(roi_dims) / (float2)(POOLED_DIM_X, POOLED_DIM_Y); |
| 147 | |
| 148 | // Calculate pooled region start and end |
| 149 | const float2 spatial_indx = (float2)(px, py); |
| 150 | const int2 max_spatial_dims = (int2)(MAX_DIM_X, MAX_DIM_Y); |
| 151 | int2 region_start = convert_int2_sat(floor(spatial_indx * pool_region_ratio)) + roi_anchor; |
| 152 | int2 region_end = convert_int2_sat(ceil((spatial_indx + 1) * pool_region_ratio)) + roi_anchor; |
| 153 | |
| 154 | region_start = clamp(region_start, 0, max_spatial_dims); |
| 155 | region_end = clamp(region_end, 0, max_spatial_dims); |
| 156 | |
| 157 | // Move input and output pointer across the fourth dimension |
| 158 | input.ptr += roi.s4 * input_stride_w; |
| 159 | output.ptr += pw * output_stride_w; |
| 160 | |
| 161 | for(int pz = 0; pz < MAX_DIM_Z; ++pz) |
| 162 | { |
| 163 | *(__global DATA_TYPE *)tensor3D_offset(&output, px, py, pz) = (__global DATA_TYPE)roi_pool_1x1(&input, |
| 164 | region_start.x, |
| 165 | region_end.x, |
| 166 | region_start.y, |
| 167 | region_end.y, pz); |
| 168 | } |
| 169 | } |