1/* 2 * Copyright (c) 2016-2020 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 "warp_helpers.h" 26 27/** Transforms four 2D coordinates. This is used to map the output coordinates to the input coordinates. 28 * 29 * @param[in] coord 2D coordinates to transform. 30 * @param[in] scale input/output scale ratio 31 * 32 * @return a float8 containing 4 2D transformed values in the input image. 33 */ 34inline const float8 transform_nearest(const float2 coord, const float2 scale) 35{ 36#ifdef SAMPLING_POLICY_TOP_LEFT 37 const float4 in_x_coords = (float4)(coord.s0, 1 + coord.s0, 2 + coord.s0, 3 + coord.s0); 38 const float4 new_x = in_x_coords * (float4)(scale.s0); 39 const float4 new_y = (float4)(coord.s1 * scale.s1); 40 return (float8)(new_x.s0, new_y.s0, new_x.s1, new_y.s1, new_x.s2, new_y.s2, new_x.s3, new_y.s3); 41#elif SAMPLING_POLICY_CENTER 42 const float4 in_x_coords = (float4)(coord.s0, 1 + coord.s0, 2 + coord.s0, 3 + coord.s0); 43 const float4 new_x = (in_x_coords + ((float4)(0.5f))) * (float4)(scale.s0); 44 const float4 new_y = (float4)((coord.s1 + 0.5f) * scale.s1); 45 return (float8)(new_x.s0, new_y.s0, new_x.s1, new_y.s1, new_x.s2, new_y.s2, new_x.s3, new_y.s3); 46#else /* SAMPLING_POLICY */ 47#error("Unsupported sampling policy"); 48#endif /* SAMPLING_POLICY */ 49} 50 51/** Transforms four 2D coordinates. This is used to map the output coordinates to the input coordinates. 52 * 53 * @param[in] coord 2D coordinates to transform. 54 * @param[in] scale input/output scale ratio 55 * 56 * @return a float8 containing 4 2D transformed values in the input image. 57 */ 58inline const float8 transform_bilinear(const float2 coord, const float2 scale) 59{ 60 const float4 in_x_coords = (float4)(coord.s0, 1 + coord.s0, 2 + coord.s0, 3 + coord.s0); 61#ifdef SAMPLING_POLICY_TOP_LEFT 62 const float4 new_x = in_x_coords * (float4)(scale.s0); 63 const float4 new_y = (float4)(coord.s1 * scale.s1); 64 return (float8)(new_x.s0, new_y.s0, new_x.s1, new_y.s1, new_x.s2, new_y.s2, new_x.s3, new_y.s3); 65#elif SAMPLING_POLICY_CENTER 66 const float4 new_x = (in_x_coords + ((float4)(0.5f))) * (float4)(scale.s0) - (float4)(0.5f); 67 const float4 new_y = (float4)((coord.s1 + 0.5f) * scale.s1 - 0.5f); 68 return (float8)(new_x.s0, new_y.s0, new_x.s1, new_y.s1, new_x.s2, new_y.s2, new_x.s3, new_y.s3); 69#else /* SAMPLING_POLICY */ 70#error("Unsupported sampling policy"); 71#endif /* SAMPLING_POLICY */ 72} 73 74/** Performs an affine transformation on an image interpolating with the NEAREAST NEIGHBOUR method. Input and output are single channel U8 or S16. 75 * 76 * @note Sampling policy to used is passed as -DSAMPLING_POLICY_(TYPE) e.g. -DSAMPLING_POLICY_TOP_LEFT 77 * 78 * @param[in] in_ptr Pointer to the source image. Supported data types: U8, S16. 79 * @param[in] in_stride_x Stride of the source image in X dimension (in bytes) 80 * @param[in] in_step_x src_stride_x * number of elements along X processed per workitem(in bytes) 81 * @param[in] in_stride_y Stride of the source image in Y dimension (in bytes) 82 * @param[in] in_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) 83 * @param[in] in_offset_first_element_in_bytes The offset of the first element in the source image 84 * @param[out] out_ptr Pointer to the destination image. Supported data types: U8, S16. (Must be the same as the input) 85 * @param[in] out_stride_x Stride of the destination image in X dimension (in bytes) 86 * @param[in] out_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) 87 * @param[in] out_stride_y Stride of the destination image in Y dimension (in bytes) 88 * @param[in] out_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) 89 * @param[in] out_offset_first_element_in_bytes The offset of the first element in the destination image 90 * @param[in] input_width Input image width 91 * @param[in] input_height Input image height 92 * @param[in] scale_x The scale factor along x dimension 93 * @param[in] scale_y The scale factor along y dimension 94 */ 95__kernel void scale_nearest_neighbour_nchw( 96 IMAGE_DECLARATION(in), 97 IMAGE_DECLARATION(out), 98 const float input_width, 99 const float input_height, 100 const float scale_x, 101 const float scale_y) 102{ 103 Image in = CONVERT_TO_IMAGE_STRUCT_NO_STEP(in); 104 Image out = CONVERT_TO_IMAGE_STRUCT(out); 105 const float2 r = (float2)(scale_x, scale_y); 106 float8 transformed = transform_nearest(get_current_coords(), r); 107#ifdef ALIGN_CORNERS 108 transformed = round(transformed); 109#endif // ALIGN_CORNERS 110 const float8 tc = clamp_to_border_with_size(transformed, input_width, input_height, BORDER_SIZE); 111 vstore4(read_texels4(&in, convert_int8(tc)), 0, (__global DATA_TYPE *)out.ptr); 112} 113 114/** Performs an affine transformation on an image interpolating with the BILINEAR method. 115 * 116 * @note Sampling policy to used is passed as -DSAMPLING_POLICY_(TYPE) e.g. -DSAMPLING_POLICY_TOP_LEFT 117 * 118 * @param[in] in_ptr Pointer to the source image. Supported data types: U8, S16. 119 * @param[in] in_stride_x Stride of the source image in X dimension (in bytes) 120 * @param[in] in_step_x src_stride_x * number of elements along X processed per workitem(in bytes) 121 * @param[in] in_stride_y Stride of the source image in Y dimension (in bytes) 122 * @param[in] in_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) 123 * @param[in] in_offset_first_element_in_bytes The offset of the first element in the source image 124 * @param[out] out_ptr Pointer to the destination image. Supported data types: U8, S16. (Must be the same as the input) 125 * @param[in] out_stride_x Stride of the destination image in X dimension (in bytes) 126 * @param[in] out_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) 127 * @param[in] out_stride_y Stride of the destination image in Y dimension (in bytes) 128 * @param[in] out_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) 129 * @param[in] out_offset_first_element_in_bytes The offset of the first element in the destination image 130 * @param[in] input_width Input image width 131 * @param[in] input_height Input image height 132 * @param[in] scale_x The scale factor along x dimension 133 * @param[in] scale_y The scale factor along y dimension 134 */ 135__kernel void scale_bilinear_nchw( 136 IMAGE_DECLARATION(in), 137 IMAGE_DECLARATION(out), 138 const float input_width, 139 const float input_height, 140 const float scale_x, 141 const float scale_y) 142{ 143 Image in = CONVERT_TO_IMAGE_STRUCT_NO_STEP(in); 144 Image out = CONVERT_TO_IMAGE_STRUCT(out); 145 const float2 r = (float2)(scale_x, scale_y); 146 const float8 tc = transform_bilinear(get_current_coords(), r); 147 vstore4(bilinear_interpolate_with_border(&in, tc, input_width, input_height, BORDER_SIZE), 0, (__global DATA_TYPE *)out.ptr); 148} 149 150#if defined(DEPTH_OUT) 151/** Performs scale on an image interpolating with the NEAREAST NEIGHBOUR method. Input and output are single channel F32. (NHWC) 152 * 153 * @note Sampling policy to used is passed as -DSAMPLING_POLICY_(TYPE) e.g. -DSAMPLING_POLICY_TOP_LEFT 154 * @note Output tensor's depth should be given as a preprocessor argument using -DDEPTH_OUT=size. e.g. -DDEPTH=16 155 * 156 * @param[in] in_ptr Pointer to the source image. Supported data types: U8/S16/F16/F32. 157 * @param[in] in_stride_x Stride of the source image in X dimension (in bytes) 158 * @param[in] in_step_x src_stride_x * number of elements along X processed per workitem(in bytes) 159 * @param[in] in_stride_y Stride of the source image in Y dimension (in bytes) 160 * @param[in] in_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) 161 * @param[in] in_stride_z Stride of the source image in Z dimension (in bytes) 162 * @param[in] in_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) 163 * @param[in] in_offset_first_element_in_bytes The offset of the first element in the source image 164 * @param[out] out_ptr Pointer to the destination image. Supported data types: same as @p in_ptr 165 * @param[in] out_stride_x Stride of the destination image in X dimension (in bytes) 166 * @param[in] out_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) 167 * @param[in] out_stride_y Stride of the destination image in Y dimension (in bytes) 168 * @param[in] out_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) 169 * @param[in] out_stride_z Stride of the destination image in Z dimension (in bytes) 170 * @param[in] out_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) 171 * @param[in] out_offset_first_element_in_bytes The offset of the first element in the destination image 172 * @param[in] input_width Input image width 173 * @param[in] input_height Input image height 174 * @param[in] scale_x The scale factor along x dimension 175 * @param[in] scale_y The scale factor along y dimension 176 */ 177__kernel void scale_nearest_neighbour_nhwc( 178 TENSOR4D_DECLARATION(in), 179 TENSOR4D_DECLARATION(out), 180 const float input_width, 181 const float input_height, 182 const float scale_x, 183 const float scale_y) 184{ 185 Tensor4D in = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(in, 0); 186 Tensor4D out = CONVERT_TO_TENSOR4D_STRUCT(out, DEPTH_OUT); 187 188#ifdef SAMPLING_POLICY_TOP_LEFT 189 float new_x = get_global_id(1) * scale_x; 190 float new_y = (get_global_id(2) % DEPTH_OUT) * scale_y; 191#elif SAMPLING_POLICY_CENTER 192 float new_x = (get_global_id(1) + 0.5f) * scale_x; 193 float new_y = ((get_global_id(2) % DEPTH_OUT) + 0.5f) * scale_y; 194#else /* SAMPLING_POLICY */ 195#error("Unsupported sampling policy"); 196#endif /* SAMPLING_POLICY */ 197#ifdef ALIGN_CORNERS 198 new_x = round(new_x); 199 new_y = round(new_y); 200#endif /* ALIGN_CORNERS */ 201 const float clamped_x = clamp(new_x, 0.0f, input_width - 1); 202 const float clamped_y = clamp(new_y, 0.0f, input_height - 1); 203 204 *((__global DATA_TYPE *)out.ptr) = *((__global DATA_TYPE *)tensor4D_offset(&in, get_global_id(0), convert_int(clamped_x), convert_int(clamped_y), (get_global_id(2) / DEPTH_OUT))); 205} 206 207/** Performs scale on an image interpolating with the BILINEAR method. (NHWC) 208 * 209 * @note Sampling policy to be used is passed as -DSAMPLING_POLICY_(TYPE) e.g. -DSAMPLING_POLICY_TOP_LEFT 210 * @note If border mode replicate is used, is should be passed as -DBORDER_MODE_REPLICATE 211 * @note Output tensor's depth should be given as a preprocessor argument using -DDEPTH_OUT=size. e.g. -DDEPTH=16 212 * 213 * @param[in] in_ptr Pointer to the source image. Supported data types: U8/S16/F16/F32. 214 * @param[in] in_stride_x Stride of the source image in X dimension (in bytes) 215 * @param[in] in_step_x src_stride_x * number of elements along X processed per workitem(in bytes) 216 * @param[in] in_stride_y Stride of the source image in Y dimension (in bytes) 217 * @param[in] in_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) 218 * @param[in] in_stride_z Stride of the source image in Z dimension (in bytes) 219 * @param[in] in_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) 220 * @param[in] in_offset_first_element_in_bytes The offset of the first element in the source image 221 * @param[out] out_ptr Pointer to the destination image. Supported data types: same as @p in_ptr 222 * @param[in] out_stride_x Stride of the destination image in X dimension (in bytes) 223 * @param[in] out_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) 224 * @param[in] out_stride_y Stride of the destination image in Y dimension (in bytes) 225 * @param[in] out_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) 226 * @param[in] out_stride_z Stride of the destination image in Z dimension (in bytes) 227 * @param[in] out_step_z dst_stride_y * number of elements along Z processed per workitem(in bytes) 228 * @param[in] out_offset_first_element_in_bytes The offset of the first element in the destination image 229 * @param[in] input_width Input image width 230 * @param[in] input_height Input image height 231 * @param[in] scale_x The scale factor along x dimension 232 * @param[in] scale_y The scale factor along y dimension 233 */ 234__kernel void scale_bilinear_nhwc( 235 TENSOR4D_DECLARATION(in), 236 TENSOR4D_DECLARATION(out), 237 const float input_width, 238 const float input_height, 239 const float scale_x, 240 const float scale_y) 241{ 242 Tensor4D in = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(in, 0); 243 Tensor4D out = CONVERT_TO_TENSOR4D_STRUCT(out, DEPTH_OUT); 244 245#ifdef SAMPLING_POLICY_TOP_LEFT 246 const float new_x = get_global_id(1) * scale_x; 247 const float new_y = (get_global_id(2) % DEPTH_OUT) * scale_y; 248#elif SAMPLING_POLICY_CENTER 249 const float new_x = (get_global_id(1) + 0.5f) * scale_x - 0.5f; 250 const float new_y = ((get_global_id(2) % DEPTH_OUT) + 0.5f) * scale_y - 0.5f; 251#else /* SAMPLING_POLICY */ 252#error("Unsupported sampling policy"); 253#endif /* SAMPLING_POLICY */ 254 255 const float new_xf = floor(new_x); 256 const float new_yf = floor(new_y); 257 float clamped_x = clamp(new_xf, 0.0f, input_width - 1); 258 float clamped_x1 = clamp(new_xf + 1, 0.0f, input_width - 1); 259 float clamped_x_ = clamped_x; 260 float clamped_x1_ = clamped_x1; 261 const float clamped_y = clamp(new_yf, 0.0f, input_height - 1); 262 const float clamped_y1 = clamp(new_yf + 1, 0.0f, input_height - 1); 263 264#ifndef BORDER_MODE_REPLICATE 265 clamped_x1 = select(clamped_x1, 0.0f - BORDER_SIZE, new_yf + 1 < 0.f || new_yf + 1 > input_height - 1 || new_xf + 1 < 0.f || new_xf + 1 > input_width - 1); 266 clamped_x_ = select(clamped_x_, 0.0f - BORDER_SIZE, new_yf + 1 > input_height - 1 || new_xf < 0.f || new_xf > input_width - 1); 267 clamped_x = select(clamped_x, 0.0f - BORDER_SIZE, new_yf < 0.f || new_yf > input_height - 1 || new_xf < 0.f || new_xf > input_width - 1); 268 clamped_x1_ = select(clamped_x1_, 0.0f - BORDER_SIZE, new_xf + 1 < 0.f || new_xf + 1 > input_width - 1 || new_yf < 0.f || new_yf > input_height - 1); 269#endif /* BORDER_MODE_REPLICATE */ 270 271 float4 ins = (float4)(*((__global DATA_TYPE *)tensor4D_offset(&in, get_global_id(0), convert_int(clamped_x), convert_int(clamped_y), (get_global_id(2) / DEPTH_OUT))), 272 *((__global DATA_TYPE *)tensor4D_offset(&in, get_global_id(0), convert_int(clamped_x1_), convert_int(clamped_y), (get_global_id(2) / DEPTH_OUT))), 273 *((__global DATA_TYPE *)tensor4D_offset(&in, get_global_id(0), convert_int(clamped_x_), convert_int(clamped_y1), (get_global_id(2) / DEPTH_OUT))), 274 *((__global DATA_TYPE *)tensor4D_offset(&in, get_global_id(0), convert_int(clamped_x1), convert_int(clamped_y1), (get_global_id(2) / DEPTH_OUT)))); 275 276 const float a = new_x - new_xf; 277 const float b = 1.f - a; 278 const float a1 = new_y - new_yf; 279 const float b1 = 1.f - a1; 280 const float fr = ((ins.s0 * b * b1) + (ins.s1 * a * b1) + (ins.s2 * b * a1) + (ins.s3 * a * a1)); 281 282 *((__global DATA_TYPE *)out.ptr) = CONVERT(fr, DATA_TYPE); 283} 284#endif /* defined(DEPTH_OUT) */