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 "src/core/CL/kernels/CLPixelWiseMultiplicationKernel.h"
25
26 #include "arm_compute/core/CL/CLHelpers.h"
27 #include "arm_compute/core/CL/CLKernelLibrary.h"
28 #include "arm_compute/core/CL/ICLTensor.h"
29 #include "arm_compute/core/CL/OpenCL.h"
30 #include "arm_compute/core/TensorInfo.h"
31 #include "src/core/CL/CLValidate.h"
32 #include "src/core/helpers/AutoConfiguration.h"
33 #include "src/core/helpers/WindowHelpers.h"
34 #include "support/Cast.h"
35 #include "support/StringSupport.h"
36
37 namespace arm_compute
38 {
39 namespace
40 {
41 constexpr unsigned int num_elems_processed_per_iteration = 16;
42
validate_arguments(const ITensorInfo * input1,const ITensorInfo * input2,const ITensorInfo * output,float scale,ConvertPolicy overflow_policy,RoundingPolicy rounding_policy,const ActivationLayerInfo & act_info)43 Status validate_arguments(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, float scale,
44 ConvertPolicy overflow_policy, RoundingPolicy rounding_policy, const ActivationLayerInfo &act_info)
45 {
46 ARM_COMPUTE_UNUSED(overflow_policy);
47 ARM_COMPUTE_UNUSED(rounding_policy);
48
49 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input1, input2, output);
50 ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input1);
51 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1,
52 1,
53 DataType::U8, DataType::QASYMM8, DataType::QASYMM8_SIGNED,
54 DataType::S16, DataType::QSYMM16, DataType::F16,
55 DataType::F32);
56 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input2,
57 1,
58 DataType::U8, DataType::QASYMM8, DataType::QASYMM8_SIGNED,
59 DataType::S16, DataType::QSYMM16, DataType::F16,
60 DataType::F32);
61 ARM_COMPUTE_RETURN_ERROR_ON_MSG(scale < 0, "Scale cannot be negative.");
62 ARM_COMPUTE_RETURN_ERROR_ON(act_info.enabled() && !is_data_type_float(output->data_type()));
63
64 const TensorShape &out_shape = TensorShape::broadcast_shape(input1->tensor_shape(), input2->tensor_shape());
65
66 ARM_COMPUTE_RETURN_ERROR_ON_MSG(out_shape.total_size() == 0, "Inputs are not broadcast compatible");
67
68 // Validate in case of configured output
69 if(output->total_size() > 0)
70 {
71 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output,
72 1,
73 DataType::U8, DataType::QASYMM8, DataType::QASYMM8_SIGNED,
74 DataType::S16, DataType::QSYMM16, DataType::F16,
75 DataType::S32, DataType::F32);
76 ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->data_type() == DataType::U8 && (input1->data_type() != DataType::U8 || input2->data_type() != DataType::U8),
77 "Output can only be U8 if both inputs are U8");
78 ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->data_type() == DataType::QASYMM8 && (input1->data_type() != DataType::QASYMM8 || input2->data_type() != DataType::QASYMM8),
79 "Output can only be QASYMM8 if both inputs are QASYMM8");
80 ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->data_type() == DataType::QASYMM8_SIGNED && (input1->data_type() != DataType::QASYMM8_SIGNED || input2->data_type() != DataType::QASYMM8_SIGNED),
81 "Output can only be QASYMM8_SIGNED if both inputs are QASYMM8_SIGNED");
82 ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->data_type() == DataType::QSYMM16 && (input1->data_type() != DataType::QSYMM16 || input2->data_type() != DataType::QSYMM16),
83 "Output can only be QSYMM16 if both inputs are QSYMM16");
84 ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->data_type() == DataType::S32 && (input1->data_type() != DataType::QSYMM16 || input2->data_type() != DataType::QSYMM16),
85 "Output can only be S32 if both inputs are QSYMM16");
86 ARM_COMPUTE_RETURN_ERROR_ON_MSG(detail::have_different_dimensions(out_shape, output->tensor_shape(), 0), "Wrong shape for output");
87 }
88
89 return Status{};
90 }
91
validate_and_configure_window(ITensorInfo * input1,ITensorInfo * input2,ITensorInfo * output)92 std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output)
93 {
94 const std::pair<TensorShape, ValidRegion> broadcast_pair = ITensorInfo::broadcast_shape_and_valid_region(*input1, *input2);
95 const TensorShape &out_shape = broadcast_pair.first;
96 const ValidRegion &valid_region = broadcast_pair.second;
97
98 // Auto initialize output if not initialized
99 {
100 set_shape_if_empty(*output, out_shape);
101
102 if(input1->data_type() == DataType::S16 || input2->data_type() == DataType::S16)
103 {
104 set_format_if_unknown(*output, Format::S16);
105 }
106 else if(input1->data_type() == DataType::F32 || input2->data_type() == DataType::F32)
107 {
108 set_format_if_unknown(*output, Format::F32);
109 }
110 else if(input1->data_type() == DataType::QASYMM8)
111 {
112 set_data_type_if_unknown(*output, DataType::QASYMM8);
113 }
114 else if(input1->data_type() == DataType::QASYMM8_SIGNED)
115 {
116 set_data_type_if_unknown(*output, DataType::QASYMM8_SIGNED);
117 }
118 else if(input1->data_type() == DataType::QSYMM16)
119 {
120 set_data_type_if_unknown(*output, DataType::QSYMM16);
121 }
122 }
123
124 Window win = calculate_max_window(valid_region, Steps(num_elems_processed_per_iteration));
125 Window win_input1 = win.broadcast_if_dimension_le_one(*input1);
126 Window win_input2 = win.broadcast_if_dimension_le_one(*input2);
127
128 AccessWindowHorizontal input1_access(input1, 0, num_elems_processed_per_iteration);
129 AccessWindowHorizontal input2_access(input2, 0, num_elems_processed_per_iteration);
130 AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration);
131
132 bool window_changed = update_window_and_padding(win_input1, input1_access)
133 || update_window_and_padding(win_input2, input2_access)
134 || update_window_and_padding(win, output_access);
135
136 output_access.set_valid_region(win, valid_region);
137
138 Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
139 return std::make_pair(err, win);
140 }
141 } // namespace
142
CLPixelWiseMultiplicationKernel()143 CLPixelWiseMultiplicationKernel::CLPixelWiseMultiplicationKernel()
144 : _input1(nullptr), _input2(nullptr), _output(nullptr)
145 {
146 }
147
configure(ITensorInfo * input1,ITensorInfo * input2,ITensorInfo * output,float scale,ConvertPolicy overflow_policy,RoundingPolicy rounding_policy,const ActivationLayerInfo & act_info)148 void CLPixelWiseMultiplicationKernel::configure(ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output, float scale,
149 ConvertPolicy overflow_policy, RoundingPolicy rounding_policy, const ActivationLayerInfo &act_info)
150 {
151 configure(CLKernelLibrary::get().get_compile_context(), input1, input2, output, scale, overflow_policy, rounding_policy, act_info);
152 }
153
configure(const CLCompileContext & compile_context,ITensorInfo * input1,ITensorInfo * input2,ITensorInfo * output,float scale,ConvertPolicy overflow_policy,RoundingPolicy rounding_policy,const ActivationLayerInfo & act_info)154 void CLPixelWiseMultiplicationKernel::configure(const CLCompileContext &compile_context, ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output, float scale,
155 ConvertPolicy overflow_policy, RoundingPolicy rounding_policy, const ActivationLayerInfo &act_info)
156 {
157 ARM_COMPUTE_ERROR_ON_NULLPTR(input1, input2, output);
158 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input1, input2, output,
159 scale, overflow_policy, rounding_policy, act_info));
160
161 // Configure kernel window
162 auto win_config = validate_and_configure_window(input1, input2, output);
163 ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
164
165 _input1 = input1;
166 _input2 = input2;
167 _output = output;
168
169 int scale_int = -1;
170 // Extract sign, exponent and mantissa
171 int exponent = 0;
172 float normalized_mantissa = std::frexp(scale, &exponent);
173 // Use int scaling if factor is equal to 1/2^n for 0 <= n <= 15
174 // frexp returns 0.5 as mantissa which means that the exponent will be in the range of -1 <= e <= 14
175 // Moreover, it will be negative as we deal with 1/2^n
176 if((normalized_mantissa == 0.5f) && (-14 <= exponent) && (exponent <= 1))
177 {
178 // Store the positive exponent. We know that we compute 1/2^n
179 // Additionally we need to subtract 1 to compensate that frexp used a mantissa of 0.5
180 scale_int = std::abs(exponent - 1);
181 }
182
183 std::string acc_type;
184 // Check if it has float inputs and output
185 if(is_data_type_float(input1->data_type()) || is_data_type_float(input2->data_type()))
186 {
187 scale_int = -1;
188 acc_type = (input1->data_type() == DataType::F32 || input2->data_type() == DataType::F32) ? "float" : "half";
189 }
190 else
191 {
192 if(input1->element_size() == 2 || input2->element_size() == 2)
193 {
194 // Use 32-bit accumulator for 16-bit input
195 acc_type = "int";
196 }
197 else
198 {
199 // Use 16-bit accumulator for 8-bit input
200 acc_type = "ushort";
201 }
202 }
203
204 const bool is_quantized = is_data_type_quantized(input1->data_type());
205
206 // Set kernel build options
207 std::string kernel_name = "pixelwise_mul";
208 CLBuildOptions build_opts;
209 build_opts.add_option("-DDATA_TYPE_IN1=" + get_cl_type_from_data_type(input1->data_type()));
210 build_opts.add_option("-DDATA_TYPE_IN2=" + get_cl_type_from_data_type(input2->data_type()));
211 build_opts.add_option("-DDATA_TYPE_OUT=" + get_cl_type_from_data_type(output->data_type()));
212 build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration));
213 if(is_quantized && (output->data_type() != DataType::S32))
214 {
215 const UniformQuantizationInfo iq1_info = input1->quantization_info().uniform();
216 const UniformQuantizationInfo iq2_info = input2->quantization_info().uniform();
217 const UniformQuantizationInfo oq_info = output->quantization_info().uniform();
218
219 build_opts.add_option_if(is_data_type_quantized_asymmetric(input1->data_type()),
220 "-DOFFSET_IN1=" + support::cpp11::to_string(iq1_info.offset));
221 build_opts.add_option_if(is_data_type_quantized_asymmetric(input2->data_type()),
222 "-DOFFSET_IN2=" + support::cpp11::to_string(iq2_info.offset));
223 build_opts.add_option_if(is_data_type_quantized_asymmetric(output->data_type()),
224 "-DOFFSET_OUT=" + support::cpp11::to_string(oq_info.offset));
225 build_opts.add_option("-DSCALE_IN1=" + float_to_string_with_full_precision(iq1_info.scale));
226 build_opts.add_option("-DSCALE_IN2=" + float_to_string_with_full_precision(iq2_info.scale));
227 build_opts.add_option("-DSCALE_OUT=" + float_to_string_with_full_precision(oq_info.scale));
228 kernel_name += "_quantized";
229 }
230 else
231 {
232 kernel_name += (scale_int >= 0) ? "_int" : "_float";
233 build_opts.add_option_if_else(overflow_policy == ConvertPolicy::WRAP || is_data_type_float(output->data_type()), "-DWRAP", "-DSATURATE");
234 build_opts.add_option_if_else(rounding_policy == RoundingPolicy::TO_ZERO, "-DROUND=_rtz", "-DROUND=_rte");
235 build_opts.add_option("-DACC_DATA_TYPE=" + acc_type);
236 if(act_info.enabled())
237 {
238 build_opts.add_option("-DACTIVATION_TYPE=" + lower_string(string_from_activation_func(act_info.activation())));
239 build_opts.add_option("-DA_VAL=" + float_to_string_with_full_precision(act_info.a()));
240 build_opts.add_option("-DB_VAL=" + float_to_string_with_full_precision(act_info.b()));
241 }
242 }
243
244 // Create kernel
245 _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
246
247 // Set scale argument
248 unsigned int idx = 3 * num_arguments_per_3D_tensor(); // Skip the inputs and output parameters
249
250 if(scale_int >= 0 && !is_quantized)
251 {
252 _kernel.setArg(idx++, scale_int);
253 }
254 else
255 {
256 _kernel.setArg(idx++, scale);
257 }
258
259 ICLKernel::configure_internal(win_config.second);
260 }
261
validate(const ITensorInfo * input1,const ITensorInfo * input2,const ITensorInfo * output,float scale,ConvertPolicy overflow_policy,RoundingPolicy rounding_policy,const ActivationLayerInfo & act_info)262 Status CLPixelWiseMultiplicationKernel::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, float scale,
263 ConvertPolicy overflow_policy, RoundingPolicy rounding_policy, const ActivationLayerInfo &act_info)
264 {
265 ARM_COMPUTE_ERROR_ON_NULLPTR(input1, input2, output);
266 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input1, input2, output, scale, overflow_policy, rounding_policy, act_info));
267 ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input1->clone().get(), input2->clone().get(), output->clone().get()).first);
268
269 return Status{};
270 }
271
run_op(ITensorPack & tensors,const Window & window,cl::CommandQueue & queue)272 void CLPixelWiseMultiplicationKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
273 {
274 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
275 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
276
277 const auto src_0 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_0));
278 const auto src_1 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_1));
279 auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
280
281 const TensorShape &in_shape1 = src_0->info()->tensor_shape();
282 const TensorShape &in_shape2 = src_1->info()->tensor_shape();
283 const TensorShape &out_shape = dst->info()->tensor_shape();
284
285 bool can_collapse = true;
286 if(std::min(in_shape1.total_size(), in_shape2.total_size()) > 1)
287 {
288 can_collapse = (std::min(in_shape1.num_dimensions(), in_shape2.num_dimensions()) > Window::DimZ);
289 for(size_t d = Window::DimZ; can_collapse && (d < out_shape.num_dimensions()); ++d)
290 {
291 can_collapse = (in_shape1[d] == in_shape2[d]);
292 }
293 }
294
295 bool has_collapsed = false;
296 Window collapsed = can_collapse ? window.collapse_if_possible(ICLKernel::window(), Window::DimZ, &has_collapsed) : window;
297
298 const TensorShape &in_shape1_collapsed = has_collapsed ? in_shape1.collapsed_from(Window::DimZ) : in_shape1;
299 const TensorShape &in_shape2_collapsed = has_collapsed ? in_shape2.collapsed_from(Window::DimZ) : in_shape2;
300
301 Window slice = collapsed.first_slice_window_3D();
302 Window slice_input1 = slice.broadcast_if_dimension_le_one(in_shape1_collapsed);
303 Window slice_input2 = slice.broadcast_if_dimension_le_one(in_shape2_collapsed);
304
305 do
306 {
307 unsigned int idx = 0;
308 add_3D_tensor_argument(idx, src_0, slice_input1);
309 add_3D_tensor_argument(idx, src_1, slice_input2);
310 add_3D_tensor_argument(idx, dst, slice);
311 enqueue(queue, *this, slice, lws_hint());
312
313 ARM_COMPUTE_UNUSED(collapsed.slide_window_slice_3D(slice_input1));
314 ARM_COMPUTE_UNUSED(collapsed.slide_window_slice_3D(slice_input2));
315 }
316 while(collapsed.slide_window_slice_3D(slice));
317 }
318
border_size() const319 BorderSize CLPixelWiseMultiplicationKernel::border_size() const
320 {
321 const unsigned int replicateSize = _output->dimension(0) - std::min(_input1->dimension(0), _input2->dimension(0));
322 const unsigned int border = std::min<unsigned int>(num_elems_processed_per_iteration - 1U, replicateSize);
323 return BorderSize{ 0, border, 0, 0 };
324 }
325
326 namespace
327 {
328 constexpr unsigned int num_elems_processed_per_iteration_complex = 1;
329
validate_arguments_complex(const ITensorInfo * input1,const ITensorInfo * input2,const ITensorInfo * output,const ActivationLayerInfo & act_info)330 Status validate_arguments_complex(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, const ActivationLayerInfo &act_info)
331 {
332 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 2, DataType::F32);
333 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input2, 2, DataType::F32);
334
335 const TensorShape &out_shape = TensorShape::broadcast_shape(input1->tensor_shape(), input2->tensor_shape());
336
337 ARM_COMPUTE_RETURN_ERROR_ON_MSG(out_shape.total_size() == 0, "Inputs are not broadcast compatible");
338 ARM_COMPUTE_RETURN_ERROR_ON(act_info.enabled() && !is_data_type_float(output->data_type()));
339
340 // Validate in case of configured output
341 if(output->total_size() > 0)
342 {
343 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 2, DataType::F32);
344 ARM_COMPUTE_RETURN_ERROR_ON_MSG(detail::have_different_dimensions(out_shape, output->tensor_shape(), 0), "Wrong shape for output");
345 }
346
347 return Status{};
348 }
349
validate_and_configure_window_complex(ITensorInfo * input1,ITensorInfo * input2,ITensorInfo * output)350 std::pair<Status, Window> validate_and_configure_window_complex(ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output)
351 {
352 const std::pair<TensorShape, ValidRegion> broadcast_pair = ITensorInfo::broadcast_shape_and_valid_region(*input1, *input2);
353 const TensorShape &out_shape = broadcast_pair.first;
354 const ValidRegion &valid_region = broadcast_pair.second;
355
356 // Auto initialize output if not initialized
357 const TensorInfo out_info(out_shape, input1->num_channels(), input1->data_type());
358 auto_init_if_empty(*output, out_info);
359
360 Window win = calculate_max_window(valid_region, Steps(num_elems_processed_per_iteration_complex));
361 Window win_input1 = win.broadcast_if_dimension_le_one(*input1);
362 Window win_input2 = win.broadcast_if_dimension_le_one(*input2);
363
364 AccessWindowHorizontal input1_access(input1, 0, num_elems_processed_per_iteration_complex);
365 AccessWindowHorizontal input2_access(input2, 0, num_elems_processed_per_iteration_complex);
366 AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration_complex);
367
368 bool window_changed = update_window_and_padding(win_input1, input1_access)
369 || update_window_and_padding(win_input2, input2_access)
370 || update_window_and_padding(win, output_access);
371
372 output_access.set_valid_region(win, valid_region);
373
374 Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
375 return std::make_pair(err, win);
376 }
377 } // namespace
378
CLComplexPixelWiseMultiplicationKernel()379 CLComplexPixelWiseMultiplicationKernel::CLComplexPixelWiseMultiplicationKernel()
380 : _input1(nullptr), _input2(nullptr), _output(nullptr)
381 {
382 }
383
configure(ITensorInfo * input1,ITensorInfo * input2,ITensorInfo * output,const ActivationLayerInfo & act_info)384 void CLComplexPixelWiseMultiplicationKernel::configure(ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output, const ActivationLayerInfo &act_info)
385 {
386 configure(CLKernelLibrary::get().get_compile_context(), input1, input2, output, act_info);
387 }
388
configure(const CLCompileContext & compile_context,ITensorInfo * input1,ITensorInfo * input2,ITensorInfo * output,const ActivationLayerInfo & act_info)389 void CLComplexPixelWiseMultiplicationKernel::configure(const CLCompileContext &compile_context, ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output, const ActivationLayerInfo &act_info)
390 {
391 ARM_COMPUTE_ERROR_ON_NULLPTR(input1, input2, output);
392 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments_complex(input1, input2, output, act_info));
393
394 // Configure kernel window
395 auto win_config = validate_and_configure_window_complex(input1, input2, output);
396 ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
397
398 _input1 = input1;
399 _input2 = input2;
400 _output = output;
401
402 CLBuildOptions build_opts;
403 if(act_info.enabled())
404 {
405 build_opts.add_option("-DACTIVATION_TYPE=" + lower_string(string_from_activation_func(act_info.activation())));
406 build_opts.add_option("-DA_VAL=" + float_to_string_with_full_precision(act_info.a()));
407 build_opts.add_option("-DB_VAL=" + float_to_string_with_full_precision(act_info.b()));
408 }
409
410 // Create kernel
411 _kernel = create_kernel(compile_context, "pixelwise_mul_complex", build_opts.options());
412
413 ICLKernel::configure_internal(win_config.second);
414 }
415
validate(const ITensorInfo * input1,const ITensorInfo * input2,const ITensorInfo * output,const ActivationLayerInfo & act_info)416 Status CLComplexPixelWiseMultiplicationKernel::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, const ActivationLayerInfo &act_info)
417 {
418 ARM_COMPUTE_ERROR_ON_NULLPTR(input1, input2, output);
419 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_complex(input1, input2, output, act_info));
420 ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window_complex(input1->clone().get(), input2->clone().get(), output->clone().get()).first);
421
422 return Status{};
423 }
424
run_op(ITensorPack & tensors,const Window & window,cl::CommandQueue & queue)425 void CLComplexPixelWiseMultiplicationKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
426 {
427 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
428 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
429
430 const auto src_0 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_0));
431 const auto src_1 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_1));
432 auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
433
434 const TensorShape &in_shape1 = src_0->info()->tensor_shape();
435 const TensorShape &in_shape2 = src_1->info()->tensor_shape();
436 const TensorShape &out_shape = dst->info()->tensor_shape();
437
438 bool can_collapse = true;
439 if(std::min(in_shape1.total_size(), in_shape2.total_size()) > 1)
440 {
441 can_collapse = (std::min(in_shape1.num_dimensions(), in_shape2.num_dimensions()) > Window::DimZ);
442 for(size_t d = Window::DimZ; can_collapse && (d < out_shape.num_dimensions()); ++d)
443 {
444 can_collapse = (in_shape1[d] == in_shape2[d]);
445 }
446 }
447
448 bool has_collapsed = false;
449 Window collapsed = can_collapse ? window.collapse_if_possible(ICLKernel::window(), Window::DimZ, &has_collapsed) : window;
450
451 const TensorShape &in_shape1_collapsed = has_collapsed ? in_shape1.collapsed_from(Window::DimZ) : in_shape1;
452 const TensorShape &in_shape2_collapsed = has_collapsed ? in_shape2.collapsed_from(Window::DimZ) : in_shape2;
453
454 Window slice = collapsed.first_slice_window_3D();
455 Window slice_input1 = slice.broadcast_if_dimension_le_one(in_shape1_collapsed);
456 Window slice_input2 = slice.broadcast_if_dimension_le_one(in_shape2_collapsed);
457
458 do
459 {
460 unsigned int idx = 0;
461 add_3D_tensor_argument(idx, src_0, slice_input1);
462 add_3D_tensor_argument(idx, src_1, slice_input2);
463 add_3D_tensor_argument(idx, dst, slice);
464 enqueue(queue, *this, slice, lws_hint());
465
466 ARM_COMPUTE_UNUSED(collapsed.slide_window_slice_3D(slice_input1));
467 ARM_COMPUTE_UNUSED(collapsed.slide_window_slice_3D(slice_input2));
468 }
469 while(collapsed.slide_window_slice_3D(slice));
470 }
471
border_size() const472 BorderSize CLComplexPixelWiseMultiplicationKernel::border_size() const
473 {
474 const unsigned int replicateSize = _output->dimension(0) - std::min(_input1->dimension(0), _input2->dimension(0));
475 const unsigned int border = std::min<unsigned int>(num_elems_processed_per_iteration_complex - 1U, replicateSize);
476 return BorderSize{ 0, border, 0, 0 };
477 }
478 } // namespace arm_compute
479