Compute Library
 19.08
CLKernelLibrary.cpp
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2016-2019 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  */
25 
27 #include "arm_compute/core/Error.h"
28 #include "arm_compute/core/Utils.h"
29 
30 #include <algorithm>
31 #include <fstream>
32 #include <iostream>
33 #include <utility>
34 #include <vector>
35 
36 using namespace arm_compute;
37 
39  : _build_opts()
40 {
41 }
42 
43 void CLBuildOptions::add_option(std::string option)
44 {
45  _build_opts.emplace(std::move(option));
46 }
47 
48 void CLBuildOptions::add_option_if(bool cond, std::string option)
49 {
50  if(cond)
51  {
52  add_option(std::move(option));
53  }
54 }
55 
56 void CLBuildOptions::add_option_if_else(bool cond, std::string option_true, std::string option_false)
57 {
58  (cond) ? add_option(std::move(option_true)) : add_option(std::move(option_false));
59 }
60 
61 void CLBuildOptions::add_options(const StringSet &options)
62 {
63  _build_opts.insert(options.begin(), options.end());
64 }
65 
66 void CLBuildOptions::add_options_if(bool cond, const StringSet &options)
67 {
68  if(cond)
69  {
71  }
72 }
73 
74 const CLBuildOptions::StringSet &CLBuildOptions::options() const
75 {
76  return _build_opts;
77 }
78 
80  : _context(), _device(), _is_binary(false), _name(), _source(), _binary()
81 {
82 }
83 
84 Program::Program(cl::Context context, std::string name, std::string source)
85  : _context(std::move(context)), _device(), _is_binary(false), _name(std::move(name)), _source(std::move(source)), _binary()
86 {
87 }
88 
89 Program::Program(cl::Context context, cl::Device device, std::string name, std::vector<unsigned char> binary)
90  : _context(std::move(context)), _device(std::move(device)), _is_binary(true), _name(std::move(name)), _source(), _binary(std::move(binary))
91 {
92 }
93 
94 Program::operator cl::Program() const
95 {
96  if(_is_binary)
97  {
98  return cl::Program(_context, { _device }, { _binary });
99  }
100  else
101  {
102  return cl::Program(_context, _source, false);
103  }
104 }
105 
106 bool Program::build(const cl::Program &program, const std::string &build_options)
107 {
108  try
109  {
110  return program.build(build_options.c_str()) == CL_SUCCESS;
111  }
112  catch(const cl::Error &e)
113  {
114  cl_int err = CL_SUCCESS;
115  const auto build_info = program.getBuildInfo<CL_PROGRAM_BUILD_LOG>(&err);
116 
117  for(auto &pair : build_info)
118  {
119  std::cerr << pair.second << std::endl;
120  }
121 
122  return false;
123  }
124 }
125 
126 cl::Program Program::build(const std::string &build_options) const
127 {
128  cl::Program cl_program = static_cast<cl::Program>(*this);
129  build(cl_program, build_options);
130  return cl_program;
131 }
132 
134  : _name(), _kernel()
135 {
136 }
137 
138 Kernel::Kernel(std::string name, const cl::Program &program)
139  : _name(std::move(name)),
140  _kernel(cl::Kernel(program, _name.c_str()))
141 {
142 }
143 
144 const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
145 {
146  { "absdiff", "absdiff.cl" },
147  { "accumulate", "accumulate.cl" },
148  { "accumulate_squared", "accumulate.cl" },
149  { "accumulate_weighted", "accumulate.cl" },
150  { "activation_layer", "activation_layer.cl" },
151  { "activation_layer_quant", "activation_layer_quant.cl" },
152  { "activation_layer_quant_f32", "activation_layer_quant.cl" },
153  { "batch_to_space_nchw", "batch_to_space.cl" },
154  { "batch_to_space_static_nchw", "batch_to_space.cl" },
155  { "batch_to_space_nhwc", "batch_to_space.cl" },
156  { "batch_to_space_static_nhwc", "batch_to_space.cl" },
157  { "batchnormalization_layer_nchw", "batchnormalization_layer.cl" },
158  { "batchnormalization_layer_nhwc", "batchnormalization_layer.cl" },
159  { "bitwise_or", "bitwise_op.cl" },
160  { "bitwise_and", "bitwise_op.cl" },
161  { "bitwise_xor", "bitwise_op.cl" },
162  { "bitwise_not", "bitwise_op.cl" },
163  { "bounding_box_transform", "bounding_box_transform.cl" },
164  { "channel_combine_NV", "channel_combine.cl" },
165  { "channel_combine_RGB888", "channel_combine.cl" },
166  { "channel_combine_RGBA8888", "channel_combine.cl" },
167  { "channel_combine_UYVY422", "channel_combine.cl" },
168  { "channel_combine_YUYV422", "channel_combine.cl" },
169  { "channel_shuffle_nchw", "channel_shuffle.cl" },
170  { "channel_shuffle_nhwc", "channel_shuffle.cl" },
171  { "channel_extract_NV12", "channel_extract.cl" },
172  { "channel_extract_NV21", "channel_extract.cl" },
173  { "channel_extract_RGB888", "channel_extract.cl" },
174  { "channel_extract_RGBA8888", "channel_extract.cl" },
175  { "channel_extract_UYVY422", "channel_extract.cl" },
176  { "channel_extract_YUYV422", "channel_extract.cl" },
177  { "combine_gradients_L1", "canny.cl" },
178  { "combine_gradients_L2", "canny.cl" },
179  { "compare_equal", "comparisons.cl" },
180  { "compare_equal_quantized", "comparisons.cl" },
181  { "compare_notequal", "comparisons.cl" },
182  { "compare_notequal_quantized", "comparisons.cl" },
183  { "compare_greater", "comparisons.cl" },
184  { "compare_greater_quantized", "comparisons.cl" },
185  { "compare_greaterequal", "comparisons.cl" },
186  { "compare_greaterequal_quantized", "comparisons.cl" },
187  { "compare_less", "comparisons.cl" },
188  { "compare_less_quantized", "comparisons.cl" },
189  { "compare_lessequal", "comparisons.cl" },
190  { "compare_lessequal_quantized", "comparisons.cl" },
191  { "concatenate", "concatenate.cl" },
192  { "concatenate_width", "concatenate.cl" },
193  { "concatenate_height", "concatenate.cl" },
194  { "concatenate_width_x2", "concatenate.cl" },
195  { "concatenate_width_x4", "concatenate.cl" },
196  { "convolution_rectangle", "convolution_rectangle.cl" },
197  { "col2im", "col2im.cl" },
198  { "convert_depth_down", "depth_convert.cl" },
199  { "convert_depth_up", "depth_convert.cl" },
200  { "convert_fc_weights", "convert_fc_weights.cl" },
201  { "convolution3x3_static", "convolution3x3.cl" },
202  { "convolution5x5_static", "convolution5x5.cl" },
203  { "convolution7x7_static", "convolution7x7.cl" },
204  { "convolution9x9_static", "convolution9x9.cl" },
205  { "convolution_separable1x5_static", "convolution5x5.cl" },
206  { "convolution_separable5x1_static", "convolution5x5.cl" },
207  { "convolution_separable1x7_static", "convolution7x7.cl" },
208  { "convolution_separable7x1_static", "convolution7x7.cl" },
209  { "convolution_separable1x9_static", "convolution9x9.cl" },
210  { "convolution_separable9x1_static", "convolution9x9.cl" },
211  { "copy_tensor", "copy_tensor.cl" },
212  { "copy_pad_tensor", "copy_tensor.cl" },
213  { "copy_plane", "channel_extract.cl" },
214  { "copy_planes_3p", "channel_combine.cl" },
215  { "copy_to_keypoint", "fast_corners.cl" },
216  { "crop_tensor", "crop_tensor.cl" },
217  { "deconvolution_reshape", "deconvolution_layer.cl" },
218  { "deconvolution_upsample", "deconvolution_layer.cl" },
219  { "depthwise_convolution_3x3", "depthwise_convolution.cl" },
220  { "depthwise_convolution_3x3_f16", "depthwise_convolution.cl" },
221  { "depthwise_convolution_3x3_nhwc", "depthwise_convolution.cl" },
222  { "depthwise_convolution_3x3_nhwc_stride1", "depthwise_convolution.cl" },
223  { "dwc_3x3_native_qasymm8_nchw", "depthwise_convolution_quantized.cl" },
224  { "dwc_3x3_native_qasymm8_dot8_nchw", "depthwise_convolution_quantized.cl" },
225  { "dwc_3x3_reshaped_qasymm8_nhwc", "depthwise_convolution_quantized.cl" },
226  { "dwc_3x3_reshaped_qasymm8_stride1_nhwc", "depthwise_convolution_quantized.cl" },
227  { "dwc_3x3_reshaped_qasymm8_dot8_stride1_nhwc", "depthwise_convolution_quantized.cl" },
228  { "depth_to_space_nchw", "depth_to_space.cl" },
229  { "depth_to_space_nhwc", "depth_to_space.cl" },
230  { "depthwise_convolution_3x3_stridex1_stridey1_bifrost_f16", "depthwise_convolution.cl" },
231  { "depthwise_convolution_3x3_stridex2_stridey2_bifrost_f16", "depthwise_convolution.cl" },
232  { "depthwise_convolution_3x3_stridex1_stridey1_bifrost_f32", "depthwise_convolution.cl" },
233  { "depthwise_convolution_3x3_stridex2_stridey2_bifrost_f32", "depthwise_convolution.cl" },
234  { "depthwise_convolution_reshape_weights", "depthwise_convolution.cl" },
235  { "depthwise_convolution_reshape_weights_generic", "depthwise_convolution.cl" },
236  { "depthwise_im2col", "depthwise_convolution.cl" },
237  { "depthwise_vector_to_tensor", "depthwise_convolution.cl" },
238  { "dequantization_layer", "dequantization_layer.cl" },
239  { "derivative", "derivative.cl" },
240  { "dilate", "dilate.cl" },
241  { "direct_convolution1x1", "direct_convolution1x1.cl" },
242  { "direct_convolution1x1_nhwc", "direct_convolution1x1.cl" },
243  { "direct_convolution1x1_f32_bifrost", "direct_convolution1x1.cl" },
244  { "direct_convolution3x3", "direct_convolution3x3.cl" },
245  { "direct_convolution3x3_nhwc", "direct_convolution3x3.cl" },
246  { "direct_convolution3x3_f32_bifrost", "direct_convolution3x3.cl" },
247  { "direct_convolution5x5", "direct_convolution5x5.cl" },
248  { "direct_convolution5x5_nhwc", "direct_convolution5x5.cl" },
249  { "direct_convolution5x5_f32_bifrost", "direct_convolution5x5.cl" },
250  { "direct_convolution_1x1_3x3_5x5_quantized", "direct_convolution_1x1_3x3_5x5_quantized.cl" },
251  { "direct_convolution9x9_nhwc", "direct_convolution9x9.cl" },
252  { "elementwise_operation_ADD", "elementwise_operation.cl" },
253  { "elementwise_operation_SUB", "elementwise_operation.cl" },
254  { "elementwise_operation_MAX", "elementwise_operation.cl" },
255  { "elementwise_operation_MIN", "elementwise_operation.cl" },
256  { "elementwise_operation_DIV", "elementwise_operation.cl" },
257  { "elementwise_operation_SQUARED_DIFF", "elementwise_operation.cl" },
258  { "elementwise_operation_POWER", "elementwise_operation.cl" },
259  { "elementwise_operation_PRELU", "elementwise_operation.cl" },
260  { "elementwise_operation_ADD_quantized", "elementwise_operation_quantized.cl" },
261  { "elementwise_operation_SUB_quantized", "elementwise_operation_quantized.cl" },
262  { "elementwise_operation_MAX_quantized", "elementwise_operation_quantized.cl" },
263  { "elementwise_operation_MIN_quantized", "elementwise_operation_quantized.cl" },
264  { "elementwise_operation_DIV_quantized", "elementwise_operation_quantized.cl" },
265  { "elementwise_operation_SQUARED_DIFF_quantized", "elementwise_operation_quantized.cl" },
266  { "elementwise_operation_PRELU_quantized", "elementwise_operation_quantized.cl" },
267  { "elementwise_unary", "elementwise_unary.cl" },
268  { "erode", "erode.cl" },
269  { "fast_corners", "fast_corners.cl" },
270  { "fft_digit_reverse_axis_0", "fft_digit_reverse.cl" },
271  { "fft_digit_reverse_axis_1", "fft_digit_reverse.cl" },
272  { "fft_radix_2_first_stage_axis_0", "fft.cl" },
273  { "fft_radix_2_first_stage_axis_1", "fft.cl" },
274  { "fft_radix_2_axis_0", "fft.cl" },
275  { "fft_radix_2_axis_1", "fft.cl" },
276  { "fft_radix_3_first_stage_axis_0", "fft.cl" },
277  { "fft_radix_3_first_stage_axis_1", "fft.cl" },
278  { "fft_radix_3_axis_0", "fft.cl" },
279  { "fft_radix_3_axis_1", "fft.cl" },
280  { "fft_radix_4_first_stage_axis_0", "fft.cl" },
281  { "fft_radix_4_first_stage_axis_1", "fft.cl" },
282  { "fft_radix_4_axis_0", "fft.cl" },
283  { "fft_radix_4_axis_1", "fft.cl" },
284  { "fft_radix_5_first_stage_axis_0", "fft.cl" },
285  { "fft_radix_5_first_stage_axis_1", "fft.cl" },
286  { "fft_radix_5_axis_0", "fft.cl" },
287  { "fft_radix_5_axis_1", "fft.cl" },
288  { "fft_radix_7_first_stage_axis_0", "fft.cl" },
289  { "fft_radix_7_first_stage_axis_1", "fft.cl" },
290  { "fft_radix_7_axis_0", "fft.cl" },
291  { "fft_radix_7_axis_1", "fft.cl" },
292  { "fft_radix_8_first_stage_axis_0", "fft.cl" },
293  { "fft_radix_8_first_stage_axis_1", "fft.cl" },
294  { "fft_radix_8_axis_0", "fft.cl" },
295  { "fft_radix_8_axis_1", "fft.cl" },
296  { "fft_scale_conj", "fft_scale.cl" },
297  { "fill_image_borders_constant", "fill_border.cl" },
298  { "fill_image_borders_replicate", "fill_border.cl" },
299  { "finalize", "optical_flow_pyramid_lk.cl" },
300  { "flatten", "flatten.cl" },
301  { "floor_layer", "floor.cl" },
302  { "fuse_batchnormalization_layer", "batchnormalization_layer.cl" },
303  { "gather", "gather.cl" },
304  { "gaussian1x5_sub_x", "gaussian_pyramid.cl" },
305  { "gaussian5x1_sub_y", "gaussian_pyramid.cl" },
306  { "gemm_accumulate_biases", "gemm.cl" },
307  { "gemm_ma_f16", "gemm.cl" },
308  { "gemm_ma_f32", "gemm.cl" },
309  { "gemm_mv", "gemv.cl" },
310  { "gemm_mv_quantized", "gemv.cl" },
311  { "gemm_mm_interleaved_transposed_f16", "gemm.cl" },
312  { "gemm_mm_interleaved_transposed_f16_acc32", "gemm.cl" },
313  { "gemm_mm_interleaved_transposed_f16_bifrost", "gemm.cl" },
314  { "gemm_mm_interleaved_transposed_f32", "gemm.cl" },
315  { "gemm_mm_interleaved_transposed_f32_bifrost", "gemm.cl" },
316  { "gemm_mm_floating_point", "gemm.cl" },
317  { "gemm_mm_floating_point_f16_bifrost", "gemm.cl" },
318  { "gemm_mm_floating_point_f16_bifrost_acc32", "gemm.cl" },
319  { "gemm_mm_floating_point_f32_bifrost", "gemm.cl" },
320  { "gemm_mm_floating_point_f32_bifrost_1000", "gemm.cl" },
321  { "gemm_mm_native", "gemm.cl" },
322  { "gemm_mm_reshaped_lhs_nt_rhs_t", "gemm.cl" },
323  { "gemm_mm_reshaped_only_rhs_nt", "gemm.cl" },
324  { "gemm_mm_reshaped_only_rhs_t", "gemm.cl" },
325  { "gemm_lc_vm_f32", "gemm.cl" },
326  { "gemm_reshape_lhs_matrix_nt", "gemm.cl" },
327  { "gemm_reshape_lhs_matrix_t", "gemm.cl" },
328  { "gemm_reshape_rhs_matrix_nt", "gemm.cl" },
329  { "gemm_reshape_rhs_matrix_t", "gemm.cl" },
330  { "gemmlowp_matrix_a_reduction", "gemmlowp.cl" },
331  { "gemmlowp_matrix_a_reduction_dot8", "gemmlowp.cl" },
332  { "gemmlowp_matrix_b_reduction", "gemmlowp.cl" },
333  { "gemmlowp_mm_midgard", "gemmlowp.cl" },
334  { "gemmlowp_mm_native", "gemmlowp.cl" },
335  { "gemmlowp_mm_reshaped_lhs_nt_rhs_t", "gemmlowp.cl" },
336  { "gemmlowp_mm_reshaped_only_rhs_t", "gemmlowp.cl" },
337  { "gemmlowp_offset_contribution", "gemmlowp.cl" },
338  { "gemmlowp_offset_contribution_quantize_down", "gemmlowp.cl" },
339  { "gemmlowp_offset_contribution_quantize_down_fixedpoint", "gemmlowp.cl" },
340  { "gemmlowp_output_stage_quantize_down", "gemmlowp.cl" },
341  { "gemmlowp_output_stage_quantize_down_fixedpoint", "gemmlowp.cl" },
342  { "gemmlowp_output_stage_quantize_down_fixedpoint_qsymm16", "gemmlowp.cl" },
343  { "gemmlowp_output_stage_quantize_down_float", "gemmlowp.cl" },
344  { "generate_proposals_compute_all_anchors", "generate_proposals.cl" },
345  { "harris_score_3x3", "harris_corners.cl" },
346  { "harris_score_5x5", "harris_corners.cl" },
347  { "harris_score_7x7", "harris_corners.cl" },
348  { "hist_border_kernel", "histogram.cl" },
349  { "hist_border_kernel_fixed", "histogram.cl" },
350  { "hist_local_kernel", "histogram.cl" },
351  { "hist_local_kernel_fixed", "histogram.cl" },
352  { "hog_block_normalization", "hog.cl" },
353  { "hog_detector", "hog.cl" },
354  { "hog_orientation_binning", "hog.cl" },
355  { "hysteresis", "canny.cl" },
356  { "im2col1x1_stridex1_nchw", "im2col.cl" },
357  { "im2col3x3_nchw", "im2col.cl" },
358  { "im2col5x5_nchw", "im2col.cl" },
359  { "im2col11x11_padx0_pady0_nchw", "im2col.cl" },
360  { "im2col_generic_nchw", "im2col.cl" },
361  { "im2col_generic_padx0_pady0_nchw", "im2col.cl" },
362  { "im2col3x3_nhwc", "im2col.cl" },
363  { "im2col9x9_nhwc", "im2col.cl" },
364  { "im2col_generic_nhwc", "im2col.cl" },
365  { "init_level", "optical_flow_pyramid_lk.cl" },
366  { "init_level_max", "optical_flow_pyramid_lk.cl" },
367  { "init_level_max_initial_estimate", "optical_flow_pyramid_lk.cl" },
368  { "integral_horizontal", "integral_image.cl" },
369  { "integral_vertical", "integral_image.cl" },
370  { "IYUV_to_NV12_bt709", "color_convert.cl" },
371  { "IYUV_to_RGB888_bt709", "color_convert.cl" },
372  { "IYUV_to_RGBA8888_bt709", "color_convert.cl" },
373  { "IYUV_to_YUV444_bt709", "color_convert.cl" },
374  { "l2_normalize_x", "l2_normalize.cl" },
375  { "l2_normalize_y", "l2_normalize.cl" },
376  { "l2_normalize_z", "l2_normalize.cl" },
377  { "lktracker_stage0", "optical_flow_pyramid_lk.cl" },
378  { "lktracker_stage1", "optical_flow_pyramid_lk.cl" },
379  { "magnitude_phase", "magnitude_phase.cl" },
380  { "mean_stddev_accumulate", "mean_stddev.cl" },
381  { "mean_stddev_normalization", "mean_stddev_normalization.cl" },
382  { "memset", "memset.cl" },
383  { "minmax", "minmaxloc.cl" },
384  { "minmax_border", "minmaxloc.cl" },
385  { "minmax_layer", "minmax_layer.cl" },
386  { "minmaxloc", "minmaxloc.cl" },
387  { "non_linear_filter_box3x3", "non_linear_filter3x3.cl" },
388  { "non_linear_filter_cross3x3", "non_linear_filter3x3.cl" },
389  { "non_linear_filter_disk3x3", "non_linear_filter3x3.cl" },
390  { "non_linear_filter_box5x5", "non_linear_filter5x5.cl" },
391  { "non_linear_filter_cross5x5", "non_linear_filter5x5.cl" },
392  { "non_linear_filter_disk5x5", "non_linear_filter5x5.cl" },
393  { "non_max_suppression", "nonmax.cl" },
394  { "normalization_layer_cross_map", "normalization_layer.cl" },
395  { "normalization_layer_in_map_nchw", "normalization_layer.cl" },
396  { "normalization_layer_in_map_nhwc", "normalization_layer.cl" },
397  { "normalize_planar_yuv_layer_nchw", "normalize_planar_yuv_layer.cl" },
398  { "normalize_planar_yuv_layer_nhwc", "normalize_planar_yuv_layer.cl" },
399  { "normalize_planar_yuv_layer_q8_nchw", "normalize_planar_yuv_layer_quantized.cl" },
400  { "normalize_planar_yuv_layer_q8_nhwc", "normalize_planar_yuv_layer_quantized.cl" },
401  { "NV12_to_IYUV_bt709", "color_convert.cl" },
402  { "NV12_to_RGB888_bt709", "color_convert.cl" },
403  { "NV12_to_RGBA8888_bt709", "color_convert.cl" },
404  { "NV12_to_YUV444_bt709", "color_convert.cl" },
405  { "NV21_to_IYUV_bt709", "color_convert.cl" },
406  { "NV21_to_RGB888_bt709", "color_convert.cl" },
407  { "NV21_to_RGBA8888_bt709", "color_convert.cl" },
408  { "NV21_to_YUV444_bt709", "color_convert.cl" },
409  { "output_stage_quantized", "direct_convolution_1x1_3x3_5x5_quantized.cl" },
410  { "permute", "permute.cl" },
411  { "pixelwise_mul_complex", "pixelwise_mul_float.cl" },
412  { "pixelwise_mul_float", "pixelwise_mul_float.cl" },
413  { "pixelwise_mul_int", "pixelwise_mul_int.cl" },
414  { "pixelwise_mul_quantized", "pixelwise_mul_int.cl" },
415  { "pooling_layer_2", "pooling_layer.cl" },
416  { "pooling_layer_3", "pooling_layer.cl" },
417  { "pooling_layer_optimized_3", "pooling_layer.cl" },
418  { "pooling_layer_7", "pooling_layer.cl" },
419  { "pooling_layer_MxN_nchw", "pooling_layer.cl" },
420  { "pooling_layer_MxN_nhwc", "pooling_layer.cl" },
421  { "pooling_layer_MxN_quantized_nhwc", "pooling_layer_quantized.cl" },
422  { "pooling_layer_MxN_quantized_nchw", "pooling_layer_quantized.cl" },
423  { "prior_box_layer_nchw", "prior_box_layer.cl" },
424  { "quantization_layer", "quantization_layer.cl" },
425  { "range", "range.cl" },
426  { "range_quantized", "range.cl" },
427  { "reduction_operation_x", "reduction_operation.cl" },
428  { "reduction_operation_non_parallel_x", "reduction_operation.cl" },
429  { "reduction_operation_y", "reduction_operation.cl" },
430  { "reduction_operation_z", "reduction_operation.cl" },
431  { "reduction_operation_w", "reduction_operation.cl" },
432  { "remap_nearest_neighbour", "remap.cl" },
433  { "remap_bilinear", "remap.cl" },
434  { "reorg_layer_nchw", "reorg_layer.cl" },
435  { "reorg_layer_nhwc", "reorg_layer.cl" },
436  { "reshape_layer", "reshape_layer.cl" },
437  { "reshape_to_columns", "convolution_layer.cl" },
438  { "reverse", "reverse.cl" },
439  { "RGB888_to_IYUV_bt709", "color_convert.cl" },
440  { "RGB888_to_NV12_bt709", "color_convert.cl" },
441  { "RGB888_to_RGBA8888_bt709", "color_convert.cl" },
442  { "RGB888_to_U8_bt709", "color_convert.cl" },
443  { "RGB888_to_YUV444_bt709", "color_convert.cl" },
444  { "RGBA8888_to_IYUV_bt709", "color_convert.cl" },
445  { "RGBA8888_to_NV12_bt709", "color_convert.cl" },
446  { "RGBA8888_to_RGB888_bt709", "color_convert.cl" },
447  { "RGBA8888_to_YUV444_bt709", "color_convert.cl" },
448  { "roi_align_layer", "roi_align_layer.cl" },
449  { "roi_pooling_layer", "roi_pooling_layer.cl" },
450  { "scale_nearest_neighbour_nchw", "scale.cl" },
451  { "scale_nearest_neighbour_nhwc", "scale.cl" },
452  { "scale_bilinear_nchw", "scale.cl" },
453  { "scale_bilinear_nhwc", "scale.cl" },
454  { "scale_bilinear_quantized_nchw", "scale_quantized.cl" },
455  { "scale_bilinear_quantized_nhwc", "scale_quantized.cl" },
456  { "scharr3x3", "scharr_filter.cl" },
457  { "select_same_rank", "select.cl" },
458  { "select_different_rank_2", "select.cl" },
459  { "select_different_rank_n", "select.cl" },
460  { "sobel3x3", "sobel_filter.cl" },
461  { "sobel_separable5x1", "sobel_filter.cl" },
462  { "sobel_separable1x5", "sobel_filter.cl" },
463  { "sobel_separable7x1", "sobel_filter.cl" },
464  { "sobel_separable1x7", "sobel_filter.cl" },
465  { "softmax_layer_norm", "softmax_layer.cl" },
466  { "softmax_layer_norm_quantized", "softmax_layer_quantized.cl" },
467  { "softmax_layer_max_shift_exp_sum_quantized_serial", "softmax_layer_quantized.cl" },
468  { "softmax_layer_max_shift_exp_sum_quantized_parallel", "softmax_layer_quantized.cl" },
469  { "softmax_layer_max_shift_exp_sum_serial", "softmax_layer.cl" },
470  { "space_to_batch_nchw", "space_to_batch.cl" },
471  { "space_to_batch_static_nchw", "space_to_batch.cl" },
472  { "space_to_batch_nhwc", "space_to_batch.cl" },
473  { "space_to_batch_static_nhwc", "space_to_batch.cl" },
474  { "space_to_depth_nchw", "space_to_depth.cl" },
475  { "space_to_depth_nhwc", "space_to_depth.cl" },
476  { "softmax_layer_max_shift_exp_sum_parallel", "softmax_layer.cl" },
477  { "stack_layer", "stack_layer.cl" },
478  { "strided_slice", "slice_ops.cl" },
479  { "suppress_non_maximum", "canny.cl" },
480  { "tablelookup_U8", "tablelookup.cl" },
481  { "tablelookup_S16", "tablelookup.cl" },
482  { "threshold_binary", "threshold.cl" },
483  { "threshold_range", "threshold.cl" },
484  { "tile", "tile.cl" },
485  { "transpose", "transpose.cl" },
486  { "UYVY422_to_IYUV_bt709", "color_convert.cl" },
487  { "UYVY422_to_NV12_bt709", "color_convert.cl" },
488  { "UYVY422_to_RGB888_bt709", "color_convert.cl" },
489  { "UYVY422_to_RGBA8888_bt709", "color_convert.cl" },
490  { "upsample_layer_nchw", "upsample_layer.cl" },
491  { "upsample_layer_nhwc", "upsample_layer.cl" },
492  { "warp_affine_nearest_neighbour", "warp_affine.cl" },
493  { "warp_affine_bilinear", "warp_affine.cl" },
494  { "warp_perspective_nearest_neighbour", "warp_perspective.cl" },
495  { "warp_perspective_bilinear", "warp_perspective.cl" },
496  { "winograd_filter_transform_2x2_3x3_nchw", "winograd_filter_transform.cl" },
497  { "winograd_filter_transform_2x1_3x1_nchw", "winograd_filter_transform.cl" },
498  { "winograd_filter_transform_1x2_1x3_nchw", "winograd_filter_transform.cl" },
499  { "winograd_filter_transform_4x4_3x3_nchw", "winograd_filter_transform.cl" },
500  { "winograd_filter_transform_4x1_3x1_nchw", "winograd_filter_transform.cl" },
501  { "winograd_filter_transform_1x4_1x3_nchw", "winograd_filter_transform.cl" },
502  { "winograd_filter_transform_4x4_5x5_nchw", "winograd_filter_transform.cl" },
503  { "winograd_filter_transform_4x1_5x1_nchw", "winograd_filter_transform.cl" },
504  { "winograd_filter_transform_1x4_1x5_nchw", "winograd_filter_transform.cl" },
505  { "winograd_filter_transform_4x1_3x1_nhwc", "winograd_filter_transform.cl" },
506  { "winograd_filter_transform_1x4_1x3_nhwc", "winograd_filter_transform.cl" },
507  { "winograd_filter_transform_4x4_3x3_nhwc", "winograd_filter_transform.cl" },
508  { "winograd_filter_transform_4x4_5x5_nhwc", "winograd_filter_transform.cl" },
509  { "winograd_filter_transform_4x1_5x1_nhwc", "winograd_filter_transform.cl" },
510  { "winograd_filter_transform_1x4_1x5_nhwc", "winograd_filter_transform.cl" },
511  { "winograd_filter_transform_2x2_7x7_nhwc", "winograd_filter_transform.cl" },
512  { "winograd_filter_transform_2x1_7x1_nhwc", "winograd_filter_transform.cl" },
513  { "winograd_filter_transform_1x2_1x7_nhwc", "winograd_filter_transform.cl" },
514  { "winograd_input_transform_2x2_3x3_stepz1_nchw", "winograd_input_transform.cl" },
515  { "winograd_input_transform_2x2_3x3_stepz2_nchw", "winograd_input_transform.cl" },
516  { "winograd_input_transform_2x1_3x1_stepz1_nchw", "winograd_input_transform.cl" },
517  { "winograd_input_transform_2x1_3x1_stepz2_nchw", "winograd_input_transform.cl" },
518  { "winograd_input_transform_1x2_1x3_stepz1_nchw", "winograd_input_transform.cl" },
519  { "winograd_input_transform_1x2_1x3_stepz2_nchw", "winograd_input_transform.cl" },
520  { "winograd_input_transform_4x4_3x3_stepz1_nchw", "winograd_input_transform.cl" },
521  { "winograd_input_transform_4x1_3x1_stepz1_nchw", "winograd_input_transform.cl" },
522  { "winograd_input_transform_1x4_1x3_stepz1_nchw", "winograd_input_transform.cl" },
523  { "winograd_input_transform_4x4_5x5_stepz1_nchw", "winograd_input_transform.cl" },
524  { "winograd_input_transform_4x1_5x1_stepz1_nchw", "winograd_input_transform.cl" },
525  { "winograd_input_transform_1x4_1x5_stepz1_nchw", "winograd_input_transform.cl" },
526  { "winograd_input_transform_4x1_3x1_stepz1_nhwc", "winograd_input_transform.cl" },
527  { "winograd_input_transform_1x4_1x3_stepz1_nhwc", "winograd_input_transform.cl" },
528  { "winograd_input_transform_4x4_3x3_stepz1_nhwc", "winograd_input_transform.cl" },
529  { "winograd_input_transform_4x4_5x5_stepz1_nhwc", "winograd_input_transform.cl" },
530  { "winograd_input_transform_4x1_5x1_stepz1_nhwc", "winograd_input_transform.cl" },
531  { "winograd_input_transform_1x4_1x5_stepz1_nhwc", "winograd_input_transform.cl" },
532  { "winograd_input_transform_2x2_7x7_stepz1_nhwc", "winograd_input_transform.cl" },
533  { "winograd_input_transform_2x1_7x1_stepz1_nhwc", "winograd_input_transform.cl" },
534  { "winograd_input_transform_1x2_1x7_stepz1_nhwc", "winograd_input_transform.cl" },
535  { "winograd_output_transform_2x2_3x3_nchw", "winograd_output_transform.cl" },
536  { "winograd_output_transform_2x1_3x1_nchw", "winograd_output_transform.cl" },
537  { "winograd_output_transform_1x2_1x3_nchw", "winograd_output_transform.cl" },
538  { "winograd_output_transform_4x4_3x3_nchw", "winograd_output_transform.cl" },
539  { "winograd_output_transform_4x1_3x1_nchw", "winograd_output_transform.cl" },
540  { "winograd_output_transform_1x4_1x3_nchw", "winograd_output_transform.cl" },
541  { "winograd_output_transform_4x4_5x5_nchw", "winograd_output_transform.cl" },
542  { "winograd_output_transform_4x1_5x1_nchw", "winograd_output_transform.cl" },
543  { "winograd_output_transform_1x4_1x5_nchw", "winograd_output_transform.cl" },
544  { "winograd_output_transform_4x1_3x1_nhwc", "winograd_output_transform.cl" },
545  { "winograd_output_transform_1x4_1x3_nhwc", "winograd_output_transform.cl" },
546  { "winograd_output_transform_4x4_3x3_nhwc", "winograd_output_transform.cl" },
547  { "winograd_output_transform_4x4_5x5_nhwc", "winograd_output_transform.cl" },
548  { "winograd_output_transform_4x1_5x1_nhwc", "winograd_output_transform.cl" },
549  { "winograd_output_transform_1x4_1x5_nhwc", "winograd_output_transform.cl" },
550  { "winograd_output_transform_2x2_7x7_nhwc", "winograd_output_transform.cl" },
551  { "winograd_output_transform_2x1_7x1_nhwc", "winograd_output_transform.cl" },
552  { "winograd_output_transform_1x2_1x7_nhwc", "winograd_output_transform.cl" },
553  { "yolo_layer_nchw", "yolo_layer.cl" },
554  { "yolo_layer_nhwc", "yolo_layer.cl" },
555  { "YUYV422_to_IYUV_bt709", "color_convert.cl" },
556  { "YUYV422_to_NV12_bt709", "color_convert.cl" },
557  { "YUYV422_to_RGB888_bt709", "color_convert.cl" },
558  { "YUYV422_to_RGBA8888_bt709", "color_convert.cl" },
559 };
560 
561 const std::map<std::string, std::string> CLKernelLibrary::_program_source_map =
562 {
563 #ifdef EMBEDDED_KERNELS
564  {
565  "absdiff.cl",
566 #include "./cl_kernels/absdiff.clembed"
567  },
568  {
569  "accumulate.cl",
570 #include "./cl_kernels/accumulate.clembed"
571  },
572  {
573  "activation_layer.cl",
574 #include "./cl_kernels/activation_layer.clembed"
575  },
576  {
577  "activation_layer_quant.cl",
578 #include "./cl_kernels/activation_layer_quant.clembed"
579  },
580  {
581  "batch_to_space.cl",
582 #include "./cl_kernels/batch_to_space.clembed"
583  },
584  {
585  "bitwise_op.cl",
586 #include "./cl_kernels/bitwise_op.clembed"
587  },
588  {
589  "bounding_box_transform.cl",
590 #include "./cl_kernels/bounding_box_transform.clembed"
591  },
592  {
593  "canny.cl",
594 #include "./cl_kernels/canny.clembed"
595  },
596  {
597  "channel_combine.cl",
598 #include "./cl_kernels/channel_combine.clembed"
599  },
600  {
601  "channel_extract.cl",
602 #include "./cl_kernels/channel_extract.clembed"
603  },
604  {
605  "channel_shuffle.cl",
606 #include "./cl_kernels/channel_shuffle.clembed"
607  },
608  {
609  "col2im.cl",
610 #include "./cl_kernels/col2im.clembed"
611  },
612  {
613  "comparisons.cl",
614 #include "./cl_kernels/comparisons.clembed"
615  },
616  {
617  "concatenate.cl",
618 #include "./cl_kernels/concatenate.clembed"
619  },
620  {
621  "color_convert.cl",
622 #include "./cl_kernels/color_convert.clembed"
623  },
624  {
625  "convert_fc_weights.cl",
626 #include "./cl_kernels/convert_fc_weights.clembed"
627  },
628  {
629  "convolution3x3.cl",
630 #include "./cl_kernels/convolution3x3.clembed"
631  },
632  {
633  "convolution5x5.cl",
634 #include "./cl_kernels/convolution5x5.clembed"
635  },
636  {
637  "convolution7x7.cl",
638 #include "./cl_kernels/convolution7x7.clembed"
639  },
640  {
641  "convolution9x9.cl",
642 #include "./cl_kernels/convolution9x9.clembed"
643  },
644  {
645  "convolution_layer.cl",
646 #include "./cl_kernels/convolution_layer.clembed"
647  },
648  {
649  "convolution_rectangle.cl",
650 #include "./cl_kernels/convolution_rectangle.clembed"
651  },
652  {
653  "copy_tensor.cl",
654 #include "./cl_kernels/copy_tensor.clembed"
655  },
656  {
657  "crop_tensor.cl",
658 #include "./cl_kernels/crop_tensor.clembed"
659  },
660  {
661  "upsample_layer.cl",
662 #include "./cl_kernels/upsample_layer.clembed"
663  },
664  {
665  "deconvolution_layer.cl",
666 #include "./cl_kernels/deconvolution_layer.clembed"
667  },
668  {
669  "depth_convert.cl",
670 #include "./cl_kernels/depth_convert.clembed"
671  },
672  {
673  "depth_to_space.cl",
674 #include "./cl_kernels/depth_to_space.clembed"
675  },
676  {
677  "depthwise_convolution.cl",
678 #include "./cl_kernels/depthwise_convolution.clembed"
679  },
680  {
681  "depthwise_convolution_quantized.cl",
682 #include "./cl_kernels/depthwise_convolution_quantized.clembed"
683  },
684  {
685  "dequantization_layer.cl",
686 #include "./cl_kernels/dequantization_layer.clembed"
687  },
688  {
689  "derivative.cl",
690 #include "./cl_kernels/derivative.clembed"
691  },
692  {
693  "dilate.cl",
694 #include "./cl_kernels/dilate.clembed"
695  },
696  {
697  "direct_convolution1x1.cl",
698 #include "./cl_kernels/direct_convolution1x1.clembed"
699  },
700  {
701  "direct_convolution3x3.cl",
702 #include "./cl_kernels/direct_convolution3x3.clembed"
703  },
704  {
705  "direct_convolution5x5.cl",
706 #include "./cl_kernels/direct_convolution5x5.clembed"
707  },
708  {
709  "direct_convolution_1x1_3x3_5x5_quantized.cl",
710 #include "./cl_kernels/direct_convolution_1x1_3x3_5x5_quantized.clembed"
711  },
712  {
713  "direct_convolution9x9.cl",
714 #include "./cl_kernels/direct_convolution9x9.clembed"
715  },
716  {
717  "elementwise_operation.cl",
718 #include "./cl_kernels/elementwise_operation.clembed"
719  },
720  {
721  "elementwise_operation_quantized.cl",
722 #include "./cl_kernels/elementwise_operation_quantized.clembed"
723  },
724  {
725  "elementwise_unary.cl",
726 #include "./cl_kernels/elementwise_unary.clembed"
727  },
728  {
729  "erode.cl",
730 #include "./cl_kernels/erode.clembed"
731  },
732  {
733  "fast_corners.cl",
734 #include "./cl_kernels/fast_corners.clembed"
735  },
736  {
737  "fft.cl",
738 #include "./cl_kernels/fft.clembed"
739  },
740  {
741  "fft_digit_reverse.cl",
742 #include "./cl_kernels/fft_digit_reverse.clembed"
743  },
744  {
745  "fft_scale.cl",
746 #include "./cl_kernels/fft_scale.clembed"
747  },
748  {
749  "fill_border.cl",
750 #include "./cl_kernels/fill_border.clembed"
751  },
752  {
753  "flatten.cl",
754 #include "./cl_kernels/flatten.clembed"
755  },
756  {
757  "floor.cl",
758 #include "./cl_kernels/floor.clembed"
759  },
760  {
761  "gather.cl",
762 #include "./cl_kernels/gather.clembed"
763  },
764  {
765  "gaussian_pyramid.cl",
766 #include "./cl_kernels/gaussian_pyramid.clembed"
767  },
768  {
769  "gemm.cl",
770 #include "./cl_kernels/gemm.clembed"
771  },
772  {
773  "gemmlowp.cl",
774 #include "./cl_kernels/gemmlowp.clembed"
775  },
776  {
777  "gemv.cl",
778 #include "./cl_kernels/gemv.clembed"
779  },
780  {
781  "generate_proposals.cl",
782 #include "./cl_kernels/generate_proposals.clembed"
783  },
784  {
785  "harris_corners.cl",
786 #include "./cl_kernels/harris_corners.clembed"
787  },
788  {
789  "helpers.h",
790 #include "./cl_kernels/helpers.hembed"
791  },
792  {
793  "helpers_asymm.h",
794 #include "./cl_kernels/helpers_asymm.hembed"
795  },
796  {
797  "histogram.cl",
798 #include "./cl_kernels/histogram.clembed"
799  },
800  {
801  "hog.cl",
802 #include "./cl_kernels/hog.clembed"
803  },
804  {
805  "im2col.cl",
806 #include "./cl_kernels/im2col.clembed"
807  },
808  {
809  "integral_image.cl",
810 #include "./cl_kernels/integral_image.clembed"
811  },
812  {
813  "l2_normalize.cl",
814 #include "./cl_kernels/l2_normalize.clembed"
815  },
816  {
817  "magnitude_phase.cl",
818 #include "./cl_kernels/magnitude_phase.clembed"
819  },
820  {
821  "mean_stddev.cl",
822 #include "./cl_kernels/mean_stddev.clembed"
823  },
824  {
825  "mean_stddev_normalization.cl",
826 #include "./cl_kernels/mean_stddev_normalization.clembed"
827  },
828  {
829  "memset.cl",
830 #include "./cl_kernels/memset.clembed"
831  },
832  {
833  "minmaxloc.cl",
834 #include "./cl_kernels/minmaxloc.clembed"
835  },
836  {
837  "minmax_layer.cl",
838 #include "./cl_kernels/minmax_layer.clembed"
839  },
840  {
841  "non_linear_filter3x3.cl",
842 #include "./cl_kernels/non_linear_filter3x3.clembed"
843  },
844  {
845  "non_linear_filter5x5.cl",
846 #include "./cl_kernels/non_linear_filter5x5.clembed"
847  },
848  {
849  "non_linear_filter_helpers.h",
850 #include "./cl_kernels/non_linear_filter_helpers.hembed"
851  },
852  {
853  "nonmax.cl",
854 #include "./cl_kernels/nonmax.clembed"
855  },
856  {
857  "normalization_layer.cl",
858 #include "./cl_kernels/normalization_layer.clembed"
859  },
860  {
861  "normalize_planar_yuv_layer.cl",
862 #include "./cl_kernels/normalize_planar_yuv_layer.clembed"
863  },
864  {
865  "normalize_planar_yuv_layer_quantized.cl",
866 #include "./cl_kernels/normalize_planar_yuv_layer_quantized.clembed"
867  },
868  {
869  "batchnormalization_layer.cl",
870 #include "./cl_kernels/batchnormalization_layer.clembed"
871  },
872  {
873  "optical_flow_pyramid_lk.cl",
874 #include "./cl_kernels/optical_flow_pyramid_lk.clembed"
875  },
876  {
877  "permute.cl",
878 #include "./cl_kernels/permute.clembed"
879  },
880  {
881  "pixelwise_mul_float.cl",
882 #include "./cl_kernels/pixelwise_mul_float.clembed"
883  },
884  {
885  "pixelwise_mul_int.cl",
886 #include "./cl_kernels/pixelwise_mul_int.clembed"
887  },
888  {
889  "pooling_layer.cl",
890 #include "./cl_kernels/pooling_layer.clembed"
891  },
892  {
893  "pooling_layer_quantized.cl",
894 #include "./cl_kernels/pooling_layer_quantized.clembed"
895  },
896  {
897  "prior_box_layer.cl",
898 #include "./cl_kernels/prior_box_layer.clembed"
899  },
900  {
901  "quantization_layer.cl",
902 #include "./cl_kernels/quantization_layer.clembed"
903  },
904  {
905  "range.cl",
906 #include "./cl_kernels/range.clembed"
907  },
908  {
909  "reduction_operation.cl",
910 #include "./cl_kernels/reduction_operation.clembed"
911  },
912  {
913  "remap.cl",
914 #include "./cl_kernels/remap.clembed"
915  },
916  {
917  "reorg_layer.cl",
918 #include "./cl_kernels/reorg_layer.clembed"
919  },
920  {
921  "reshape_layer.cl",
922 #include "./cl_kernels/reshape_layer.clembed"
923  },
924  {
925  "reverse.cl",
926 #include "./cl_kernels/reverse.clembed"
927  },
928  {
929  "roi_align_layer.cl",
930 #include "./cl_kernels/roi_align_layer.clembed"
931  },
932  {
933  "roi_pooling_layer.cl",
934 #include "./cl_kernels/roi_pooling_layer.clembed"
935  },
936  {
937  "scale.cl",
938 #include "./cl_kernels/scale.clembed"
939  },
940  {
941  "scale_quantized.cl",
942 #include "./cl_kernels/scale_quantized.clembed"
943  },
944  {
945  "scharr_filter.cl",
946 #include "./cl_kernels/scharr_filter.clembed"
947  },
948  {
949  "select.cl",
950 #include "./cl_kernels/select.clembed"
951  },
952  {
953  "sobel_filter.cl",
954 #include "./cl_kernels/sobel_filter.clembed"
955  },
956  {
957  "softmax_layer.cl",
958 #include "./cl_kernels/softmax_layer.clembed"
959  },
960  {
961  "softmax_layer_quantized.cl",
962 #include "./cl_kernels/softmax_layer_quantized.clembed"
963  },
964  {
965  "slice_ops.cl",
966 #include "./cl_kernels/slice_ops.clembed"
967  },
968  {
969  "space_to_batch.cl",
970 #include "./cl_kernels/space_to_batch.clembed"
971  },
972  {
973  "space_to_depth.cl",
974 #include "./cl_kernels/space_to_depth.clembed"
975  },
976  {
977  "stack_layer.cl",
978 #include "./cl_kernels/stack_layer.clembed"
979  },
980  {
981  "tablelookup.cl",
982 #include "./cl_kernels/tablelookup.clembed"
983  },
984  {
985  "threshold.cl",
986 #include "./cl_kernels/threshold.clembed"
987  },
988  {
989  "tile.cl",
990 #include "./cl_kernels/tile.clembed"
991  },
992  {
993  "transpose.cl",
994 #include "./cl_kernels/transpose.clembed"
995  },
996  {
997  "types.h",
998 #include "./cl_kernels/types.hembed"
999  },
1000  {
1001  "warp_affine.cl",
1002 #include "./cl_kernels/warp_affine.clembed"
1003  },
1004  {
1005  "warp_helpers.h",
1006 #include "./cl_kernels/warp_helpers.hembed"
1007  },
1008  {
1009  "warp_perspective.cl",
1010 #include "./cl_kernels/warp_perspective.clembed"
1011  },
1012  {
1013  "winograd_filter_transform.cl",
1014 #include "./cl_kernels/winograd_filter_transform.clembed"
1015  },
1016  {
1017  "winograd_input_transform.cl",
1018 #include "./cl_kernels/winograd_input_transform.clembed"
1019  },
1020  {
1021  "winograd_output_transform.cl",
1022 #include "./cl_kernels/winograd_output_transform.clembed"
1023  },
1024  {
1025  "yolo_layer.cl",
1026 #include "./cl_kernels/yolo_layer.clembed"
1027  },
1028 #endif /* EMBEDDED_KERNELS */
1029 };
1030 
1031 CLKernelLibrary::CLKernelLibrary()
1032  : _context(), _device(), _kernel_path("."), _programs_map(), _built_programs_map()
1033 {
1034  opencl_is_available(); // Make sure the OpenCL symbols are initialised *before* the CLKernelLibrary is built
1035 }
1036 
1038 {
1039  static CLKernelLibrary _kernel_library;
1040  return _kernel_library;
1041 }
1042 
1043 Kernel CLKernelLibrary::create_kernel(const std::string &kernel_name, const StringSet &build_options_set) const
1044 {
1045  // Find which program contains the kernel
1046  auto kernel_program_it = _kernel_program_map.find(kernel_name);
1047 
1048  if(_kernel_program_map.end() == kernel_program_it)
1049  {
1050  ARM_COMPUTE_ERROR("Kernel %s not found in the CLKernelLibrary", kernel_name.c_str());
1051  }
1052  std::string concat_str;
1053 
1054 #if defined(ARM_COMPUTE_DEBUG_ENABLED)
1055  // Enable debug properties in CL kernels
1056  concat_str += " -DARM_COMPUTE_DEBUG_ENABLED";
1057 #endif // defined(ARM_COMPUTE_DEBUG_ENABLED)
1058 
1060  concat_str += " -DGPU_ARCH=" + support::cpp11::to_string(
1061  static_cast<std::underlying_type<GPUTarget>::type>(gpu_arch));
1062  if(fp16_supported())
1063  {
1064  concat_str += " -DARM_COMPUTE_OPENCL_FP16_ENABLED=1 ";
1065  }
1066 
1067  if(dot8_supported(_device))
1068  {
1069  concat_str += " -DARM_COMPUTE_OPENCL_DOT8_ENABLED=1 ";
1070  }
1071 
1072  if(dot8_acc_supported(_device))
1073  {
1074  concat_str += " -DARM_COMPUTE_OPENCL_DOT8_ACC_ENABLED=1 ";
1075  }
1076 
1077  if(get_cl_version(_device) == CLVersion::CL20)
1078  {
1079  concat_str += " -cl-std=CL2.0 ";
1080  }
1081  else if(arm_non_uniform_workgroup_supported(_device))
1082  {
1083  concat_str += " -cl-arm-non-uniform-work-group-size ";
1084  }
1085  else
1086  {
1087  ARM_COMPUTE_ERROR("Non uniform workgroup size is not supported!!");
1088  }
1089 
1090  // Check if the program has been built before with same build options.
1091  const std::string program_name = kernel_program_it->second;
1092  const std::string build_options = stringify_set(build_options_set) + concat_str;
1093 
1094  const std::string built_program_name = program_name + "_" + build_options;
1095  auto built_program_it = _built_programs_map.find(built_program_name);
1096 
1097  cl::Program cl_program;
1098 
1099  if(_built_programs_map.end() != built_program_it)
1100  {
1101  // If program has been built, retrieve to create kernel from it
1102  cl_program = built_program_it->second;
1103  }
1104  else
1105  {
1106  // Get program
1107  Program program = load_program(program_name);
1108 
1109  // Build program
1110  cl_program = program.build(build_options);
1111 
1112  // Add built program to internal map
1113  _built_programs_map.emplace(built_program_name, cl_program);
1114  }
1115 
1116  // Create and return kernel
1117  return Kernel(kernel_name, cl_program);
1118 }
1119 
1120 void CLKernelLibrary::add_built_program(const std::string &built_program_name, const cl::Program &program)
1121 {
1122  _built_programs_map.emplace(built_program_name, program);
1123 }
1124 
1126 {
1127  return ::fp16_supported(_device);
1128 }
1129 
1131 {
1132  return device_supports_extension(_device, "cl_khr_int64_base_atomics");
1133 }
1134 
1135 const Program &CLKernelLibrary::load_program(const std::string &program_name) const
1136 {
1137  const auto program_it = _programs_map.find(program_name);
1138 
1139  if(program_it != _programs_map.end())
1140  {
1141  return program_it->second;
1142  }
1143 
1144  Program program;
1145 
1146 #ifdef EMBEDDED_KERNELS
1147  const auto program_source_it = _program_source_map.find(program_name);
1148 
1149  if(_program_source_map.end() == program_source_it)
1150  {
1151  ARM_COMPUTE_ERROR("Embedded program for %s does not exist.", program_name.c_str());
1152  }
1153 
1154  program = Program(_context, program_name, program_source_it->second);
1155 #else /* EMBEDDED_KERNELS */
1156  // Check for binary
1157  std::string source_name = _kernel_path + program_name;
1158  std::string binary_name = source_name + "bin";
1159 
1160  if(std::ifstream(binary_name).is_open())
1161  {
1162  const std::string program_binary = read_file(binary_name, true);
1163  program = Program(_context, _device, program_name, std::vector<unsigned char>(program_binary.begin(), program_binary.end()));
1164  }
1165  else if(std::ifstream(source_name).is_open())
1166  {
1167  program = Program(_context, program_name, read_file(source_name, false));
1168  }
1169  else
1170  {
1171  ARM_COMPUTE_ERROR("Kernel file %s does not exist.", source_name.c_str());
1172  }
1173 #endif /* EMBEDDED_KERNELS */
1174 
1175  // Insert program to program map
1176  const auto new_program = _programs_map.emplace(program_name, std::move(program));
1177 
1178  return new_program.first->second;
1179 }
1180 
1181 std::string CLKernelLibrary::stringify_set(const StringSet &s) const
1182 {
1183  std::string concat_set;
1184 
1185 #ifndef EMBEDDED_KERNELS
1186  concat_set += "-I" + _kernel_path + " ";
1187 #endif /* EMBEDDED_KERNELS */
1188 
1189  // Concatenate set
1190  for(const auto &el : s)
1191  {
1192  concat_set += " " + el;
1193  }
1194 
1195  return concat_set;
1196 }
1197 
1198 std::string CLKernelLibrary::get_program_source(const std::string &program_name)
1199 {
1200  const auto program_source_it = _program_source_map.find(program_name);
1201 
1202  if(program_source_it == _program_source_map.end())
1203  {
1204  ARM_COMPUTE_ERROR("Embedded program for %s does not exist.", program_name.c_str());
1205  }
1206 
1207  return program_source_it->second;
1208 }
1209 
1210 size_t CLKernelLibrary::max_local_workgroup_size(const cl::Kernel &kernel) const
1211 {
1212  size_t result;
1213 
1214  size_t err = kernel.getWorkGroupInfo(_device, CL_KERNEL_WORK_GROUP_SIZE, &result);
1215  ARM_COMPUTE_ERROR_ON_MSG(err != 0, "clGetKernelWorkGroupInfo failed to return the maximum workgroup size for the kernel");
1216  ARM_COMPUTE_UNUSED(err);
1217 
1218  return result;
1219 }
1220 
1222 {
1223  GPUTarget _target = get_target_from_device(_device);
1224  cl::NDRange default_range;
1225 
1226  switch(_target)
1227  {
1228  case GPUTarget::MIDGARD:
1229  case GPUTarget::T600:
1230  case GPUTarget::T700:
1231  case GPUTarget::T800:
1232  default_range = cl::NDRange(128u, 1);
1233  break;
1234  default:
1235  default_range = cl::NullRange;
1236  }
1237 
1238  return default_range;
1239 }
1240 
1242 {
1243  return _device.getInfo<CL_DEVICE_VERSION>();
1244 }
1245 
1247 {
1248  return _device.getInfo<CL_DEVICE_MAX_COMPUTE_UNITS>();
1249 }
#define ARM_COMPUTE_ERROR(...)
Print the given message then throw an std::runtime_error.
Definition: Error.h:261
bool dot8_acc_supported(const cl::Device &device)
Helper function to check whether the cl_arm_integer_dot_product_accumulate_int8 extension is supporte...
Definition: CLHelpers.cpp:159
bool dot8_supported(const cl::Device &device)
Helper function to check whether the cl_arm_integer_dot_product_int8 extension is supported.
Definition: CLHelpers.cpp:149
Kernel()
Default Constructor.
bool fp16_supported(const cl::Device &device)
Helper function to check whether the cl_khr_fp16 extension is supported.
Definition: CLHelpers.cpp:144
void add_built_program(const std::string &built_program_name, const cl::Program &program)
Add a new built program to the cache.
std::string get_device_version()
Return the device version.
const StringSet & options() const
Gets the current options list set.
std::string to_string(T &&value)
Convert integer and float values to string.
static CLKernelLibrary & get()
Access the KernelLibrary singleton.
CLBuildOptions()
Default constructor.
GPUTarget get_arch_from_target(GPUTarget target)
Helper function to get the GPU arch.
Definition: GPUTarget.cpp:189
Copyright (c) 2017-2018 ARM Limited.
cl_uint get_num_compute_units()
Return the maximum number of compute units in the device.
size_t max_local_workgroup_size(const cl::Kernel &kernel) const
Find the maximum number of local work items in a workgroup can be supported for the kernel.
void add_option(std::string option)
Adds option to the existing build option list.
void add_options(const StringSet &options)
Appends given build options to the current's objects options.
cl::NDRange default_ndrange() const
Return the default NDRange for the device.
#define ARM_COMPUTE_UNUSED(...)
To avoid unused variables warnings.
Definition: Error.h:160
std::string read_file(const std::string &filename, bool binary)
Load an entire file in memory.
Definition: Utils.cpp:47
std::string get_program_source(const std::string &program_name)
Gets the source of the selected program.
static bool build(const cl::Program &program, const std::string &build_options="")
Build the given CL program.
void add_option_if(bool cond, std::string option)
Adds option if a given condition is true;.
GPUTarget get_target_from_device(const cl::Device &device)
Helper function to get the GPU target from CL device.
Definition: CLHelpers.cpp:131
Program()
Default constructor.
CLVersion get_cl_version(const cl::Device &device)
Helper function to get the highest OpenCL version supported.
Definition: CLHelpers.cpp:164
bool int64_base_atomics_supported() const
Returns true if int64_base_atomics extension is supported by the CL device.
bool device_supports_extension(const cl::Device &device, const char *extension_name)
Helper function to check whether a given extension is supported.
Definition: CLHelpers.cpp:187
GPUTarget
Available GPU Targets.
Definition: GPUTarget.h:34
bool fp16_supported() const
Returns true if FP16 is supported by the CL device.
Kernel create_kernel(const std::string &kernel_name, const StringSet &build_options_set={}) const
Creates a kernel from the kernel library.
CLKernelLibrary class.
bool arm_non_uniform_workgroup_supported(const cl::Device &device)
Helper function to check whether the arm_non_uniform_work_group_size extension is supported.
Definition: CLHelpers.cpp:139
void add_options_if(bool cond, const StringSet &options)
Appends given build options to the current's objects options if a given condition is true.
bool opencl_is_available()
Check if OpenCL is available.
Definition: OpenCL.cpp:136
#define ARM_COMPUTE_ERROR_ON_MSG(cond,...)
Definition: Error.h:328
void add_option_if_else(bool cond, std::string option_true, std::string option_false)
Adds first option if condition is true else the second one.