Compute Library
 22.11
CLDepthwiseConvolutionLayerNativeKernel.cpp
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2019-2022 Arm Limited.
3  *
4  * SPDX-License-Identifier: MIT
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a copy
7  * of this software and associated documentation files (the "Software"), to
8  * deal in the Software without restriction, including without limitation the
9  * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10  * sell copies of the Software, and to permit persons to whom the Software is
11  * furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included in all
14  * copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22  * SOFTWARE.
23  */
25 
31 #include "arm_compute/core/Utils.h"
34 #include "src/core/CL/CLUtils.h"
35 #include "src/core/CL/CLValidate.h"
36 #include "src/core/CL/ICLKernel.h"
40 #include "support/StringSupport.h"
41 
42 namespace arm_compute
43 {
44 namespace
45 {
46 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const DWCComputeKernelInfo &dwc_info,
47  const ConvolutionInfo &conv_info, const ITensorInfo *output_multipliers, const ITensorInfo *output_shifts)
48 {
49  ARM_COMPUTE_UNUSED(dwc_info);
50  bool in_place = false;
51  if(output == nullptr || output == input)
52  {
53  in_place = true;
54  output = input;
55  }
60  ARM_COMPUTE_RETURN_ERROR_ON(conv_info.pad_stride_info.stride().first > 1 && dwc_info.m0 != 1);
61  ARM_COMPUTE_RETURN_ERROR_ON(conv_info.dilation.x() > 1 && dwc_info.m0 != 1);
62  ARM_COMPUTE_RETURN_ERROR_ON((dwc_info.export_input_to_cl_image == true));
63  ARM_COMPUTE_RETURN_ERROR_ON_MSG((dwc_info.export_weights_to_cl_image == true) && (export_to_cl_image(weights) == false), "Weights cannot be exported to cl_image!");
64  ARM_COMPUTE_RETURN_ERROR_ON((dwc_info.export_weights_to_cl_image == true) && ((dwc_info.n0 % 4) != 0));
65  ARM_COMPUTE_RETURN_ERROR_ON(conv_info.pad_stride_info.stride().first < 1);
66  ARM_COMPUTE_RETURN_ERROR_ON(conv_info.pad_stride_info.stride().second < 1);
67  ARM_COMPUTE_RETURN_ERROR_ON((conv_info.dilation.x() < 1) || (conv_info.dilation.y() < 1));
68  const size_t idx_c = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::CHANNEL);
69  ARM_COMPUTE_UNUSED(idx_c);
70  ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(idx_c) != (input->dimension(idx_c) * conv_info.depth_multiplier));
71 
72  // In place restrictions
73  if(in_place)
74  {
75  const int weights_width_idx = get_data_layout_dimension_index(weights->data_layout(), DataLayoutDimension::WIDTH);
76  const int weights_height_idx = get_data_layout_dimension_index(weights->data_layout(), DataLayoutDimension::HEIGHT);
77  ARM_COMPUTE_RETURN_ERROR_ON(weights->tensor_shape()[weights_width_idx] != 1U || weights->tensor_shape()[weights_height_idx] != 1U);
78  ARM_COMPUTE_RETURN_ERROR_ON(conv_info.depth_multiplier != 1U);
79  ARM_COMPUTE_RETURN_ERROR_ON(conv_info.pad_stride_info.stride() != std::make_pair(1U, 1U));
80  ARM_COMPUTE_RETURN_ERROR_ON(conv_info.dilation != Size2D(1U, 1U));
81  ARM_COMPUTE_RETURN_ERROR_ON(conv_info.pad_stride_info.has_padding()); // Note that in princple padding can be supported with in_place but we choose not to support it
82  }
83 
84  const ConvolutionInfo info{ conv_info.pad_stride_info, conv_info.depth_multiplier, ActivationLayerInfo(), conv_info.dilation };
85  const TensorShape output_shape = arm_compute::misc::shape_calculator::compute_depthwise_convolution_shape(*input, *weights, conv_info);
86 
87  if(conv_info.depth_multiplier > 1 && dwc_info.n0 > 1)
88  {
89  ARM_COMPUTE_RETURN_ERROR_ON((conv_info.depth_multiplier % dwc_info.n0) != 0);
90  }
91 
92  const bool is_quantized = is_data_type_quantized(input->data_type());
93 
94  if(biases != nullptr)
95  {
96  ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != output_shape[idx_c]);
97  ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1);
98 
99  if(is_quantized)
100  {
102  }
103  else
104  {
106  }
107  }
108 
109  if(is_quantized)
110  {
111  ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output_multipliers, output_shifts);
114  ARM_COMPUTE_RETURN_ERROR_ON(output_multipliers->num_dimensions() > 1);
115  ARM_COMPUTE_RETURN_ERROR_ON(output_shifts->num_dimensions() > 1);
116 
117  if(is_data_type_quantized_per_channel(weights->data_type()))
118  {
120  ARM_COMPUTE_RETURN_ERROR_ON(output_shape[idx_c] != output_multipliers->dimension(0));
121  ARM_COMPUTE_RETURN_ERROR_ON(output_shape[idx_c] != output_shifts->dimension(0));
122  }
123  else
124  {
126  ARM_COMPUTE_RETURN_ERROR_ON(1 != output_multipliers->dimension(0));
127  ARM_COMPUTE_RETURN_ERROR_ON(1 != output_shifts->dimension(0));
128  }
129  }
130  else
131  {
133  }
134 
135  if(output->total_size() != 0)
136  {
139  }
140 
141  if(is_data_type_quantized(input->data_type()))
142  {
143  const UniformQuantizationInfo iq_info = input->quantization_info().uniform();
144  const UniformQuantizationInfo wq_info = weights->quantization_info().uniform();
145  const UniformQuantizationInfo oq_info = (output->total_size() != 0) ? output->quantization_info().uniform() : iq_info;
146 
147  float multiplier = iq_info.scale * wq_info.scale / oq_info.scale;
148  int output_multiplier = 0;
149  int output_shift = 0;
150  ARM_COMPUTE_RETURN_ON_ERROR(quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift));
151  }
152 
153  return Status{};
154 }
155 } // namespace
156 
158  : _input(nullptr),
159  _weights(nullptr),
160  _biases(nullptr),
161  _output(nullptr),
162  _depth_multiplier(1),
163  _output_multipliers(nullptr),
164  _output_shifts(nullptr),
165  _export_input_to_cl_image(false),
166  _export_weights_to_cl_image(false),
167  _is_quantized(false)
168 {
169  _type = CLKernelType::DEPTHWISE;
170 }
171 
173  const DWCComputeKernelInfo &dwc_info, const ConvolutionInfo &conv_info,
174  const ICLTensor *output_multipliers, const ICLTensor *output_shifts)
175 {
176  configure(CLKernelLibrary::get().get_compile_context(), input, weights, biases, output, dwc_info, conv_info, output_multipliers, output_shifts);
177 }
178 
179 void CLDepthwiseConvolutionLayerNativeKernel::configure(const CLCompileContext &compile_context, ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output,
180  const DWCComputeKernelInfo &dwc_info, const ConvolutionInfo &conv_info,
181  const ICLTensor *output_multipliers, const ICLTensor *output_shifts)
182 {
183  ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights);
184  if(output == nullptr)
185  {
186  // In-place
187  output = input;
188  }
189  ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), weights->info(), (biases != nullptr) ? biases->info() : nullptr, output->info(),
190  dwc_info, conv_info, (output_multipliers != nullptr) ? output_multipliers->info() : nullptr, (output_shifts != nullptr) ? output_shifts->info() : nullptr));
191 
192  auto padding_info = get_padding_info({ input, output });
193 
194  const TensorShape output_shape = arm_compute::misc::shape_calculator::compute_depthwise_convolution_shape(*(input->info()), *(weights->info()), conv_info);
195  auto_init_if_empty(*(output->info()), input->info()->clone()->set_tensor_shape(output_shape).set_quantization_info(output->info()->quantization_info()));
196 
197  _input = input;
198  _output = output;
199  _weights = weights;
200  _biases = biases;
201  _depth_multiplier = conv_info.depth_multiplier;
202  _output_multipliers = output_multipliers;
203  _output_shifts = output_shifts;
204  _export_input_to_cl_image = dwc_info.export_input_to_cl_image;
205  _export_weights_to_cl_image = dwc_info.export_weights_to_cl_image;
206  _is_quantized = is_data_type_quantized(input->info()->data_type());
207 
208  const unsigned int n0 = adjust_vec_size(dwc_info.n0, output->info()->dimension(0));
209  const unsigned int m0 = std::min(dwc_info.m0, (unsigned int)output->info()->dimension(1));
210  std::string kernel_name = "";
211 
212  CLBuildOptions build_opts;
213 
214  // Update the padding for the input/weights tensor if we can export to cl_image
215  if(_export_input_to_cl_image)
216  {
218  }
219 
220  if(_export_weights_to_cl_image)
221  {
223  }
224 
225  // Conditions of -cl-fast-relaxed-math causing accuracy issues can be traced from COMPMID-5324
226  const GPUTarget gpu_target = get_target();
227  const auto act_function = conv_info.act_info.activation();
228  const auto dst_data_type = _output->info()->data_type();
229 
230  if((gpu_target != GPUTarget::G71 && (gpu_target & GPUTarget::GPU_ARCH_MASK) == GPUTarget::BIFROST)
232  && (dst_data_type == DataType::F32 || dst_data_type == DataType::F16))
233  {
234  // -cl-fast-relaxed-math also sets -cl-finite-math-only and -cl-unsafe-math-optimizations
235  // to disable -cl-finite-math-only, we only include -cl-unsafe-math-optimizations
236  build_opts.add_option("-cl-unsafe-math-optimizations");
237  }
238  else
239  {
240  build_opts.add_option("-cl-fast-relaxed-math");
241  }
242 
243  build_opts.add_option("-DACTIVATION_TYPE=" + lower_string(string_from_activation_func(act_function)));
244  build_opts.add_option("-DDEPTH_MULTIPLIER=" + support::cpp11::to_string(conv_info.depth_multiplier));
245  build_opts.add_option_if_else(_export_input_to_cl_image, "-DSRC_TENSOR_TYPE=IMAGE", "-DSRC_TENSOR_TYPE=BUFFER");
246  // Note: SRC_DATA_TYPE must have the same data type of WEI_DATA_TYPE. In quantized, we could
247  // have a case where the data types for the activation and weights are different. However, since the implementation
248  // only works when both have same data type, we have to change the offset to take into account this aspect
249  build_opts.add_option("-DSRC_DATA_TYPE=" + get_cl_type_from_data_type(_input->info()->data_type()));
250  build_opts.add_option("-DDST_TENSOR_TYPE=BUFFER");
251  build_opts.add_option("-DDST_DATA_TYPE=" + get_cl_type_from_data_type(dst_data_type));
252  build_opts.add_option_if_else(_export_weights_to_cl_image, "-DWEI_TENSOR_TYPE=IMAGE", "-DWEI_TENSOR_TYPE=BUFFER");
253  build_opts.add_option("-DSRC_WIDTH=" + support::cpp11::to_string(_input->info()->dimension(1)));
254  build_opts.add_option("-DSRC_HEIGHT=" + support::cpp11::to_string(_input->info()->dimension(2)));
255  build_opts.add_option("-DDST_WIDTH=" + support::cpp11::to_string(_output->info()->dimension(1)));
256  build_opts.add_option("-DDST_HEIGHT=" + support::cpp11::to_string(_output->info()->dimension(2)));
257  build_opts.add_option("-DWEI_WIDTH=" + support::cpp11::to_string(_weights->info()->dimension(1)));
258  build_opts.add_option("-DWEI_HEIGHT=" + support::cpp11::to_string(_weights->info()->dimension(2)));
259  build_opts.add_option("-DWEI_DATA_TYPE=" + get_cl_type_from_data_type(_weights->info()->data_type()));
260  build_opts.add_option("-DPAD_TOP=" + support::cpp11::to_string(conv_info.pad_stride_info.pad_top()));
261  build_opts.add_option("-DPAD_LEFT=" + support::cpp11::to_string(conv_info.pad_stride_info.pad_left()));
262  build_opts.add_option("-DSTRIDE_X=" + support::cpp11::to_string(conv_info.pad_stride_info.stride().first));
263  build_opts.add_option("-DSTRIDE_Y=" + support::cpp11::to_string(conv_info.pad_stride_info.stride().second));
264  build_opts.add_option("-DDILATION_X=" + support::cpp11::to_string(conv_info.dilation.x()));
265  build_opts.add_option("-DDILATION_Y=" + support::cpp11::to_string(conv_info.dilation.y()));
266  build_opts.add_option("-DN0=" + support::cpp11::to_string(n0));
267  build_opts.add_option("-DM0=" + support::cpp11::to_string(m0));
268  build_opts.add_option("-DM0_A=" + support::cpp11::to_string(_weights->info()->dimension(1) + m0 - 1));
269  build_opts.add_option_if_else(conv_info.depth_multiplier > 1, "-DN0_A=1", "-DN0_A=" + support::cpp11::to_string(n0));
270  build_opts.add_option("-DPARTIAL_N0=" + support::cpp11::to_string(_output->info()->dimension(0) % n0));
271  build_opts.add_option_if(_input->info()->num_dimensions() > 3, "-DBATCHED_EXECUTION");
272 
273  // Force unroll with pragma when any of the following values exceed the maximum number of manual unroll
274  set_unroll_with_pragma(build_opts, { static_cast<int>(_weights->info()->dimension(1) + m0 - 1),
275  static_cast<int>(_weights->info()->dimension(1)),
276  static_cast<int>(_weights->info()->dimension(2))
277  });
278 
279  if(biases != nullptr)
280  {
281  build_opts.add_option(std::string("-DHAS_BIAS"));
282  build_opts.add_option(std::string("-DBIA_DATA_TYPE=" + get_cl_type_from_data_type(biases->info()->data_type())));
283  }
284 
285  if(_is_quantized)
286  {
287  kernel_name = "dwc_native_quantized_nhwc";
288  const UniformQuantizationInfo iqinfo = input->info()->quantization_info().uniform();
289  const UniformQuantizationInfo wqinfo = weights->info()->quantization_info().uniform();
290  const UniformQuantizationInfo oqinfo = output->info()->quantization_info().uniform();
291 
292  PixelValue zero_value = PixelValue(0, input->info()->data_type(), input->info()->quantization_info());
293  int zero_value_s32;
294  zero_value.get(zero_value_s32);
295 
296  float multiplier = iqinfo.scale * wqinfo.scale / oqinfo.scale;
297  int output_multiplier = 0;
298  int output_shift = 0;
299  quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift);
300  build_opts.add_option("-DDST_MULTIPLIER=" + support::cpp11::to_string(output_multiplier));
301  build_opts.add_option("-DDST_SHIFT=" + support::cpp11::to_string(output_shift));
302  build_opts.add_option("-DSRC_OFFSET=" + support::cpp11::to_string(-iqinfo.offset));
303  build_opts.add_option("-DWEI_OFFSET=" + support::cpp11::to_string(-wqinfo.offset));
304  build_opts.add_option("-DDST_OFFSET=" + support::cpp11::to_string(oqinfo.offset));
305  build_opts.add_option("-DZERO_VALUE=" + support::cpp11::to_string(zero_value_s32));
306  build_opts.add_option("-DACC_DATA_TYPE=" + get_cl_type_from_data_type(DataType::S32));
307  build_opts.add_option("-DDST_MULTIPLIERS_DATA_TYPE=" + get_cl_type_from_data_type(_output_multipliers->info()->data_type()));
308  build_opts.add_option("-DDST_SHIFTS_DATA_TYPE=" + get_cl_type_from_data_type(_output_shifts->info()->data_type()));
309  build_opts.add_option_if_else(weights->info()->data_type() == DataType::QSYMM8_PER_CHANNEL, "-DQUANTIZATION_TYPE=PER_CHANNEL", "-DQUANTIZATION_TYPE=PER_TENSOR");
310  // Note: We expect the input and output tensors to always adopt a per-tensor quantization approach
311  int a_val{};
312  int b_val{};
313  std::tie(b_val, a_val) = get_quantized_activation_min_max(conv_info.act_info, input->info()->data_type(), oqinfo);
314 
315  build_opts.add_option_if(conv_info.act_info.enabled(), "-DA_VAL=" + support::cpp11::to_string(a_val));
316  build_opts.add_option_if(conv_info.act_info.enabled(), "-DB_VAL=" + support::cpp11::to_string(b_val));
317  }
318  else
319  {
320  kernel_name = "dwc_native_fp_nhwc";
321  build_opts.add_option("-DACC_DATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
322  build_opts.add_option_if(conv_info.act_info.enabled(), "-DA_VAL=" + float_to_string_with_full_precision(conv_info.act_info.a()));
323  build_opts.add_option_if(conv_info.act_info.enabled(), "-DB_VAL=" + float_to_string_with_full_precision(conv_info.act_info.b()));
324  }
325 
326  Window win = calculate_max_window(*(output->info()), Steps(n0, m0));
327  ICLKernel::configure_internal(win);
328 
329  _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
330 
332 
333  // Set config_id for enabling LWS tuning
334  _config_id = kernel_name;
335  _config_id += "_";
336  _config_id += support::cpp11::to_string(input->info()->dimension(0));
337  _config_id += "_";
338  _config_id += support::cpp11::to_string(input->info()->dimension(1));
339  _config_id += "_";
340  _config_id += support::cpp11::to_string(input->info()->dimension(2));
341  _config_id += "_";
342  _config_id += support::cpp11::to_string(output->info()->dimension(0));
343  _config_id += "_";
344  _config_id += support::cpp11::to_string(output->info()->dimension(1));
345  _config_id += "_";
346  _config_id += support::cpp11::to_string(output->info()->dimension(2));
347  _config_id += "_";
348  _config_id += string_from_data_type(input->info()->data_type());
349 }
350 
352  const DWCComputeKernelInfo &dwc_info, const ConvolutionInfo &conv_info, const ITensorInfo *output_multipliers, const ITensorInfo *output_shifts)
353 {
354  ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, weights, biases, output, dwc_info, conv_info, output_multipliers, output_shifts));
355  return Status{};
356 }
357 
358 void CLDepthwiseConvolutionLayerNativeKernel::run(const Window &window, cl::CommandQueue &queue)
359 {
362 
363  // Collapse window
364  Window window_collapsed = window.collapse(ICLKernel::window(), Window::DimZ);
365 
366  Window slice = window_collapsed.first_slice_window_4D();
367 
368  cl::Image2D input_cl_image;
369  cl::Image2D weights_cl_image;
370 
371  if(_export_input_to_cl_image || _export_weights_to_cl_image)
372  {
373  // Export cl_buffer to cl_image
374  if(_export_input_to_cl_image)
375  {
376  const size_t image_w = _input->info()->dimension(0) / 4;
377  const size_t image_h = _input->info()->dimension(1) * _input->info()->dimension(2) * _input->info()->dimension(3);
378  const TensorShape shape2d(image_w, image_h);
379  const size_t image_row_pitch = _input->info()->strides_in_bytes()[1];
380  input_cl_image = create_image2d_from_buffer(CLKernelLibrary::get().context(), _input->cl_buffer(), shape2d, _input->info()->data_type(), image_row_pitch);
381  }
382 
383  if(_export_weights_to_cl_image)
384  {
385  const size_t image_w = _weights->info()->dimension(0) / 4;
386  const size_t image_h = _weights->info()->dimension(1) * _weights->info()->dimension(2) * _weights->info()->dimension(3);
387  const TensorShape shape2d(image_w, image_h);
388  const size_t image_row_pitch = _weights->info()->strides_in_bytes()[1];
389  weights_cl_image = create_image2d_from_buffer(CLKernelLibrary::get().context(), _weights->cl_buffer(), shape2d, _weights->info()->data_type(), image_row_pitch);
390  }
391  }
392 
393  unsigned int idx = 0;
394  if(_export_input_to_cl_image)
395  {
396  _kernel.setArg(idx++, input_cl_image);
397  }
398  add_4d_tensor_nhwc_argument(idx, _input);
399  add_4d_tensor_nhwc_argument(idx, _output);
400  if(_export_weights_to_cl_image)
401  {
402  _kernel.setArg(idx++, weights_cl_image);
403  }
404  add_4D_tensor_argument(idx, _weights, slice);
405  if(_is_quantized)
406  {
407  add_1D_tensor_argument(idx, _output_multipliers, slice);
408  add_1D_tensor_argument(idx, _output_shifts, slice);
409  }
410  if(_biases != nullptr)
411  {
412  add_1D_tensor_argument(idx, _biases, slice);
413  }
414  enqueue(queue, *this, slice, lws_hint());
415 }
416 } // namespace arm_compute
void add_4d_tensor_nhwc_argument(unsigned int &idx, const ICLTensor *tensor)
Add the passed NHWC 4D tensor&#39;s parameters to the object&#39;s kernel&#39;s arguments by passing strides...
Definition: ICLKernel.cpp:144
bool is_data_type_quantized(DataType dt)
Check if a given data type is of quantized type.
Definition: Utils.h:1030
virtual size_t num_dimensions() const =0
The number of dimensions of the tensor (rank)
Class describing the value of a pixel for any image format.
Definition: PixelValue.h:34
Window calculate_max_window(const ValidRegion &valid_region, const Steps &steps, bool skip_border, BorderSize border_size)
#define ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(tensor)
Definition: CLValidate.h:35
const Window & window() const
The maximum window the kernel can be executed on.
Definition: IKernel.cpp:28
Shape of a tensor.
Definition: TensorShape.h:39
TensorShape compute_depthwise_convolution_shape(const ITensorInfo &input, const ITensorInfo &weights, const ConvolutionInfo &info)
Calculate the depthwise convolution output shape of a tensor.
#define ARM_COMPUTE_RETURN_ERROR_ON_DATA_LAYOUT_NOT_IN(t,...)
Definition: Validate.h:742
void set_unroll_with_pragma(CLBuildOptions &built_opts, std::initializer_list< int > values)
Definition: CLHelpers.cpp:482
virtual size_t dimension(size_t index) const =0
Return the size of the requested dimension.
void enqueue(cl::CommandQueue &queue, ICLKernel &kernel, const Window &window, const cl::NDRange &lws_hint=CLKernelLibrary::get().default_ndrange(), bool use_dummy_work_items=false)
Add the kernel to the command queue with the given window.
Definition: ICLKernel.cpp:32
const StringSet & options() const
Gets the current options list set.
cl::NDRange lws_hint() const
Return the Local-Workgroup-Size hint.
Definition: ICLKernel.h:383
void get(uint8_t &v) const
Interpret the pixel value as a U8.
Definition: PixelValue.h:244
#define ARM_COMPUTE_RETURN_ON_ERROR(status)
Checks if a status contains an error and returns it.
Definition: Error.h:204
std::string to_string(T &&value)
Convert integer and float values to string.
virtual DataType data_type() const =0
Data type used for each element of the tensor.
1 channel, 1 F32 per channel
#define ARM_COMPUTE_ERROR_ON(cond)
If the condition is true then an error message is printed and an exception thrown.
Definition: Error.h:466
const std::string & string_from_activation_func(ActivationLayerInfo::ActivationFunction act)
Translates a given activation function to a string.
Definition: Utils.cpp:163
static CLKernelLibrary & get()
Access the KernelLibrary singleton.
Store the tensor&#39;s metadata.
Definition: ITensorInfo.h:40
#define ARM_COMPUTE_ERROR_THROW_ON(status)
Definition: Error.h:455
Quantization info when assuming per layer quantization.
void configure(const CLCompileContext &compile_context, ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const DWCComputeKernelInfo &dwc_info, const ConvolutionInfo &conv_info, const ICLTensor *output_multipliers=nullptr, const ICLTensor *output_shifts=nullptr)
Initialize the function&#39;s source, destination and parameters.
Manages all the OpenCL kernels compilation and caching, provides accessors for the OpenCL Context...
Status calculate_quantized_multiplier(float multiplier, int32_t *quant_multiplier, int32_t *shift, bool ignore_epsilon=false)
Calculate quantized representation of multiplier.
Status class.
Definition: Error.h:52
std::string lower_string(const std::string &val)
Lower a given string.
Definition: Utils.cpp:353
Status validate_arguments(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *dst, const PadStrideInfo &conv_info)
#define ARM_COMPUTE_RETURN_ERROR_ON(cond)
If the condition is true, an error is returned.
Definition: Error.h:296
void update_padding_for_cl_image(ITensorInfo *tensor)
Update padding required to export the OpenCL buffer to OpenCL image2d.
#define ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(...)
Definition: Validate.h:284
Copyright (c) 2017-2022 Arm Limited.
1 channel, 1 F16 per channel
#define ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(...)
Definition: Validate.h:159
1 channel, 1 S32 per channel
void run(const Window &window, cl::CommandQueue &queue) override
Enqueue the OpenCL kernel to process the given window on the passed OpenCL command queue...
void add_option(std::string option)
Adds option to the existing build option list.
Window collapse(const Window &full_window, size_t first, size_t last=Coordinates::num_max_dimensions) const
Collapse the dimensions between first and last.
Definition: Window.inl:111
cl::Kernel create_kernel(const CLCompileContext &ctx, const std::string &kernel_name, const std::set< std::string > &build_opts=std::set< std::string >())
Creates an opencl kernel using a compile context.
Definition: CLHelpers.cpp:404
const std::string & string_from_data_type(DataType dt)
Convert a data type identity into a string.
Definition: Utils.cpp:135
#define ARM_COMPUTE_UNUSED(...)
To avoid unused variables warnings.
Definition: Error.h:152
bool is_data_type_quantized_per_channel(DataType dt)
Check if a given data type is of per channel type.
Definition: Utils.h:1107
std::string float_to_string_with_full_precision(float val)
Create a string with the float in full precision.
Definition: Utils.h:1124
std::pair< int32_t, int32_t > get_quantized_activation_min_max(ActivationLayerInfo act_info, DataType data_type, UniformQuantizationInfo oq_info)
Returns a pair of minimum and maximum values for a quantized activation.
Definition: Utils.cpp:558
quantized, asymmetric fixed-point 8-bit number unsigned
Class to describe a number of elements in each dimension.
Definition: Steps.h:40
unsigned int n0
Number of columns processed by each thread.
GPUTarget get_target() const
Get the targeted GPU architecture.
Definition: ICLKernel.h:443
UniformQuantizationInfo uniform() const
Return per layer quantization info.
std::string get_cl_type_from_data_type(const DataType &dt)
Translates a tensor data type to the appropriate OpenCL type.
Definition: CLHelpers.cpp:39
bool auto_init_if_empty(ITensorInfo &info, const TensorShape &shape, int num_channels, DataType data_type, QuantizationInfo quantization_info=QuantizationInfo())
Auto initialize the tensor info (shape, number of channels and data type) if the current assignment i...
virtual std::unique_ptr< T > clone() const =0
Provide a clone of the current object of class T.
virtual ITensorInfo * info() const =0
Interface to be implemented by the child class to return the tensor&#39;s metadata.
void add_option_if(bool cond, std::string option)
Adds option if a given condition is true;.
bool export_weights_to_cl_image
Export the weights to cl_image.
virtual QuantizationInfo quantization_info() const =0
Get the quantization settings (scale and offset) of the tensor.
#define ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(k)
Definition: Validate.h:915
bool has_padding_changed(const std::unordered_map< const ITensorInfo *, PaddingSize > &padding_map)
Check if the previously stored padding info has changed after configuring a kernel.
Definition: Utils.cpp:603
CLCompileContext class.
Compute descriptor used by the depthwise convolution native kernel.
Depthwise CL kernel type.
Definition: CLTypes.h:83
quantized, symmetric per channel fixed-point 8-bit number
ScaleKernelInfo info(interpolation_policy, default_border_mode, PixelValue(), sampling_policy, false)
Interface for OpenCL tensor.
Definition: ICLTensor.h:42
static constexpr size_t DimZ
Alias for dimension 2 also known as Z dimension.
Definition: Window.h:47
GPUTarget
Available GPU Targets.
Definition: GPUTarget.h:34
size_t get_data_layout_dimension_index(const DataLayout &data_layout, const DataLayoutDimension &data_layout_dimension)
Get the index of the given dimension.
Definition: Helpers.inl:193
#define ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(...)
Definition: Validate.h:541
Num samples, height, width, channels.
#define ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(t, c,...)
Definition: Validate.h:788
static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const DWCComputeKernelInfo &dwc_info, const ConvolutionInfo &conv_info, const ITensorInfo *output_multipliers=nullptr, const ITensorInfo *output_shifts=nullptr)
Static function to check if given info will lead to a valid configuration of CLDepthwiseConvolutionLa...
std::unordered_map< const ITensorInfo *, PaddingSize > get_padding_info(std::initializer_list< const ITensorInfo *> infos)
Stores padding information before configuring a kernel.
Definition: Utils.cpp:588
unsigned int m0
Number of rows processed by each thread.
Window first_slice_window_4D() const
First 4D slice of the window.
Definition: Window.h:313
#define ARM_COMPUTE_RETURN_ERROR_ON_MSG(cond, msg)
If the condition is true, an error is returned.
Definition: Error.h:244
virtual const cl::Buffer & cl_buffer() const =0
Interface to be implemented by the child class to return a reference to the OpenCL buffer containing ...
#define ARM_COMPUTE_ERROR_ON_NULLPTR(...)
Definition: Validate.h:157
cl::Image2D create_image2d_from_buffer(const cl::Context &ctx, const cl::Buffer &buffer, const TensorShape &shape2d, DataType data_type, size_t image_row_pitch)
Create a cl::Image2D object from an OpenCL buffer.
Definition: CLUtils.cpp:35
unsigned int adjust_vec_size(unsigned int vec_size, size_t dim0)
Returns the adjusted vector size in case it is less than the input&#39;s first dimension, getting rounded down to its closest valid vector size.
Definition: Utils.h:1222
quantized, asymmetric fixed-point 8-bit number signed
virtual const Strides & strides_in_bytes() const =0
The strides in bytes for accessing each dimension of the tensor.
void add_1D_tensor_argument(unsigned int &idx, const ICLTensor *tensor, const Window &window)
Add the passed 1D tensor&#39;s parameters to the object&#39;s kernel&#39;s arguments starting from the index idx...
Definition: ICLKernel.h:178
std::string kernel_name
void add_4D_tensor_argument(unsigned int &idx, const ICLTensor *tensor, const Window &window)
Add the passed 4D tensor&#39;s parameters to the object&#39;s kernel&#39;s arguments starting from the index idx...
Definition: ICLKernel.h:236
bool export_to_cl_image(const ITensorInfo *tensor)
Definition: CLHelpers.cpp:444
Describe a multidimensional execution window.
Definition: Window.h:39
#define ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(f, s)
Definition: Validate.h:201
SimpleTensor< T > slice(const SimpleTensor< T > &src, Coordinates starts, Coordinates ends)
bool export_input_to_cl_image
Export input to cl_image.
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.