Compute Library
 21.05
CLGEMMMatrixMultiplyNativeKernel.cpp
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2019-2021 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 
32 #include "arm_compute/core/Utils.h"
39 #include "support/StringSupport.h"
40 
41 #include <cstddef>
42 #include <cstdint>
43 #include <tuple>
44 
46 
47 namespace arm_compute
48 {
49 namespace
50 {
51 using ElementsProcessed = Steps;
52 
53 Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, float alpha, float beta, const GEMMLHSMatrixInfo &lhs_info,
54  const GEMMRHSMatrixInfo &rhs_info,
55  const GEMMKernelInfo &gemm_info)
56 {
57  ARM_COMPUTE_UNUSED(alpha);
58  ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input0, input1, output);
61  ARM_COMPUTE_RETURN_ERROR_ON_MSG(input0->num_dimensions() > 4, "The number of dimensions for the LHS matrix must be <= 4");
62  ARM_COMPUTE_RETURN_ERROR_ON_MSG(input1->num_dimensions() > 3, "The number of dimensions for the RHS matrix must be <= 3");
63  ARM_COMPUTE_RETURN_ERROR_ON_MSG(((rhs_info.k0 & (rhs_info.k0 - 1)) && rhs_info.k0 != 3), "Only 2,3,4,8,16 are supported for k0");
64  ARM_COMPUTE_RETURN_ERROR_ON(rhs_info.k0 > 16);
65  ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.m0 < 1 || lhs_info.m0 > 8);
66  ARM_COMPUTE_RETURN_ERROR_ON_MSG(((rhs_info.n0 & (rhs_info.n0 - 1)) && rhs_info.n0 != 3), "Only 2,3,4,8,16 are supported for n0");
67  ARM_COMPUTE_RETURN_ERROR_ON_MSG((gemm_info.reinterpret_input_as_3d || gemm_info.depth_output_gemm3d != 0) && (input2 != nullptr)
68  && (!gemm_info.broadcast_bias),
69  "Bias addition only supported with broadcast mode in case the input or output has to be reinterpreted as 3D");
70  ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.fp_mixed_precision, "Mixed precision not supported");
71  ARM_COMPUTE_RETURN_ERROR_ON_MSG(rhs_info.export_to_cl_image, "Export to CLImage not supported for GEMM native");
72 
73  const unsigned int m = gemm_info.m;
74  const unsigned int n = gemm_info.n;
75  const unsigned int k = gemm_info.k;
76 
80 
81  ARM_COMPUTE_RETURN_ERROR_ON(input0->dimension(0) != k);
82  ARM_COMPUTE_RETURN_ERROR_ON(input1->dimension(0) != n);
83  ARM_COMPUTE_RETURN_ERROR_ON(input1->dimension(1) != k);
84  if(gemm_info.reinterpret_input_as_3d)
85  {
86  ARM_COMPUTE_RETURN_ERROR_ON(input0->dimension(1) * input0->dimension(2) != m);
87  }
88  else
89  {
90  ARM_COMPUTE_RETURN_ERROR_ON(input0->dimension(1) != m);
91  }
92 
93  if(input2 != nullptr && !(helpers::float_ops::is_zero(beta)))
94  {
95  const unsigned int input2_dim0 = input2->dimension(0);
96  const unsigned int input2_dim1 = input2->dimension(1);
97 
99  if(gemm_info.broadcast_bias)
100  {
101  ARM_COMPUTE_RETURN_ERROR_ON_MSG((input2_dim1 != 1 || input2_dim0 != n), "Incorrect dimension of bias matrix which is to be broadcasted");
102  }
103  else
104  {
105  ARM_COMPUTE_RETURN_ERROR_ON_MSG((input2_dim0 != n || input2_dim1 != m), "Incorrect dimension of bias matrix");
106  }
107  }
108 
109  if(output->total_size() != 0)
110  {
111  const TensorInfo tensor_info_output = output->clone()->set_tensor_shape(compute_mm_shape(*input0, *input1, gemm_info));
112  ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &tensor_info_output);
114  }
115 
116  return Status{};
117 }
118 
119 std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input0, ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output, const GEMMLHSMatrixInfo &lhs_info,
120  const GEMMRHSMatrixInfo &rhs_info,
121  const GEMMKernelInfo &gemm_info, ElementsProcessed &num_elements_processed)
122 {
123  unsigned int &num_elems_processed_per_iteration_x = num_elements_processed[0];
124  unsigned int &num_elems_processed_per_iteration_y = num_elements_processed[1];
125  bool reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d;
126  bool reinterpret_output_as_3d = gemm_info.depth_output_gemm3d != 0;
127 
128  Window win{};
129  Window win_out{};
130  bool window_changed = false;
131 
132  // In case both input and output have to be reinterpreted as 3D tensors,
133  // force reinterpret_input_as_3d and reinterpret_output_as_3d to be false.
134  if(reinterpret_input_as_3d == reinterpret_output_as_3d)
135  {
136  reinterpret_output_as_3d = false;
137  }
138 
139  // Output tensor auto initialization if not yet initialized
140  auto_init_if_empty(*output, input0->clone()->set_tensor_shape(compute_mm_shape(*input0, *input1, gemm_info)));
141 
142  TensorInfo tmp_info(*output);
143 
144  if(reinterpret_output_as_3d)
145  {
146  // Since the output tensor has to be reinterpreted as 3D and the execute window is based on a 2D GEMM,
147  // the window needs to be constructed on the 2D collapsed version of the tensor
148  TensorShape tmp_shape(output->tensor_shape());
149  tmp_shape.collapse(2U, 1U);
150  tmp_info.set_tensor_shape(tmp_shape);
151  }
152 
153  // Configure kernel window
154  num_elems_processed_per_iteration_x = rhs_info.n0;
155  num_elems_processed_per_iteration_y = lhs_info.m0;
156 
157  win = calculate_max_window(tmp_info, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
158  win_out = calculate_max_window(*output, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
159 
160  AccessWindowStatic input0_access(input0, 0, 0,
161  input0->dimension(0),
162  input0->dimension(1));
163  AccessWindowStatic input1_access(input1, 0, 0,
164  ceil_to_multiple(input1->dimension(0), num_elems_processed_per_iteration_x),
165  input1->dimension(1));
166  AccessWindowStatic output_access(output, 0, 0,
167  output->dimension(0),
168  output->dimension(1));
169 
170  if(input2 != nullptr)
171  {
172  const int bias_processed_per_iteration_x = num_elems_processed_per_iteration_x;
173 
174  AccessWindowStatic input2_access(input2, 0, 0,
175  ceil_to_multiple(input2->dimension(0), bias_processed_per_iteration_x),
176  input2->dimension(1));
177 
178  window_changed = update_window_and_padding(win, input0_access, input1_access, input2_access) || // window used by the execute_window_loop
179  update_window_and_padding(win_out, output_access); // window used to update the padding requirements of output tensor
180  }
181  else
182  {
183  window_changed = update_window_and_padding(win, input0_access, input1_access) || // window used by the execute_window_loop
184  update_window_and_padding(win_out, output_access); // window used to update the padding requirements of output tensor
185  }
186 
187  // Collapse along the Z direction
188  // This collapse needs to be here in order to tune the Z dimension of LWS
189  Window collapsed = win;
190  const unsigned int dimension_to_collapse = std::min(static_cast<unsigned int>(output->num_dimensions()), 2u);
191  collapsed = win.collapse(win, dimension_to_collapse);
192 
193  Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
194  return std::make_pair(err, collapsed);
195 }
196 } // namespace
197 
199  : _input0(nullptr), _input1(nullptr), _input2(nullptr), _output(nullptr), _slide_matrix_b(true), _reinterpret_input_as_3d(false), _reinterpret_output_as_3d(false), _use_dummy_work_items(false),
200  _add_bias(false), _broadcast_bias(false)
201 {
202 }
203 
204 void CLGEMMMatrixMultiplyNativeKernel::configure(const ICLTensor *input0, const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output, float alpha, float beta,
205  const GEMMLHSMatrixInfo &lhs_info,
206  const GEMMRHSMatrixInfo &rhs_info, const GEMMKernelInfo &gemm_info)
207 {
208  configure(CLKernelLibrary::get().get_compile_context(), input0, input1, input2, output, alpha, beta, lhs_info, rhs_info, gemm_info);
209 }
210 
211 void CLGEMMMatrixMultiplyNativeKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input0, const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output, float alpha,
212  float beta,
213  const GEMMLHSMatrixInfo &lhs_info,
214  const GEMMRHSMatrixInfo &rhs_info, const GEMMKernelInfo &gemm_info)
215 {
216  ARM_COMPUTE_ERROR_ON_NULLPTR(input0, input1, output);
217 
218  ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input0->info(), input1->info(), (input2 != nullptr ? input2->info() : nullptr), output->info(), alpha, beta, lhs_info, rhs_info, gemm_info));
219 
220  auto padding_info = get_padding_info({ input0, output });
221  _input0 = input0;
222  _input1 = input1;
223  _input2 = helpers::float_ops::is_zero(beta) ? nullptr : input2;
224  _output = output;
225  _reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d;
226  _reinterpret_output_as_3d = gemm_info.depth_output_gemm3d != 0;
227  _use_dummy_work_items = preferred_dummy_work_items_support(CLKernelLibrary::get().get_device());
228  _add_bias = _input2 != nullptr;
229  _broadcast_bias = gemm_info.broadcast_bias;
230 
231  // In case both input and output have to be reinterpreted as 3D tensors,
232  // force reinterpret_input_as_3d and reinterpret_output_as_3d to be false.
233  if(_reinterpret_input_as_3d == _reinterpret_output_as_3d)
234  {
235  _reinterpret_input_as_3d = false;
236  _reinterpret_output_as_3d = false;
237  }
238 
239  // Check if we need to slide the matrix B
240  const unsigned int num_dimensions_input0 = _input0->info()->num_dimensions();
241  _slide_matrix_b = (_input1->info()->num_dimensions() >= num_dimensions_input0);
242 
243  ElementsProcessed num_elements_processed{};
244 
245  // Configure kernel window
246  auto win_config = validate_and_configure_window(input0->info(), input1->info(), input2 != nullptr ? input2->info() : nullptr, output->info(), lhs_info, rhs_info, gemm_info, num_elements_processed);
247  ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
248  ICLKernel::configure_internal(win_config.second);
249 
250  // If _reinterpret_input_as_3d = _reinterpret_output_as_3d = true,
251  // we will dispatch a batched-GEMM to reduce the complexity of the address calculation within the OpenCL kernel.
252  // This means that the actual m used by the kernel is given by output->info()->dimension(1) and not by gemm_info.m
253  const unsigned int internal_m = _reinterpret_output_as_3d ? gemm_info.m : output->info()->dimension(1);
254 
255  const unsigned int h_gemm_3d = _reinterpret_output_as_3d ? output->info()->dimension(1) : input0->info()->dimension(1);
256  const unsigned int d_gemm_3d = _reinterpret_output_as_3d ? output->info()->dimension(2) : input0->info()->dimension(2);
257 
258  // Calculate partial (store instead of load) M0 and partial N0 for the partial blocks at the end of a row/column if any. This is to avoid padding.
259  const unsigned int partial_store_m0 = internal_m % lhs_info.m0;
260  const unsigned int partial_store_n0 = gemm_info.n % rhs_info.n0;
261 
262  // Shrink M0 to be always <= M (internal_m) to prevent out-of-bounds reads.
263  // NOTE: This might have implications on heuristics and performance
264  const unsigned int internal_m0 = std::min(internal_m, lhs_info.m0);
265 
266  // Create build options
267  CLBuildOptions build_opts;
268  build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input0->info()->data_type()));
269  build_opts.add_option_if(!(helpers::float_ops::is_one(alpha)), "-DALPHA=" + float_to_string_with_full_precision(alpha));
270  build_opts.add_option_if(_input2 != nullptr, "-DBETA=" + float_to_string_with_full_precision(beta));
271  build_opts.add_option_if(helpers::float_ops::is_one(beta), "-DUNIT_BETA");
272  build_opts.add_option_if(gemm_info.broadcast_bias, "-DBROADCAST_BIAS");
273  build_opts.add_option_if(_reinterpret_input_as_3d, "-DREINTERPRET_INPUT_AS_3D");
274  build_opts.add_option_if(_reinterpret_output_as_3d, "-DREINTERPRET_OUTPUT_AS_3D");
275  build_opts.add_option_if(_reinterpret_input_as_3d || _reinterpret_output_as_3d, "-DHEIGHT_GEMM3D=" + support::cpp11::to_string(h_gemm_3d));
276  build_opts.add_option_if(_reinterpret_input_as_3d || _reinterpret_output_as_3d, "-DDEPTH_GEMM3D=" + support::cpp11::to_string(d_gemm_3d));
277  build_opts.add_option_if(!_slide_matrix_b, "-DMATRIX_B_DEPTH=" + support::cpp11::to_string(input1->info()->dimension(2)));
278  build_opts.add_option_if(_use_dummy_work_items, "-DDUMMY_WORK_ITEMS");
279  build_opts.add_option("-DM=" + support::cpp11::to_string(internal_m));
280  build_opts.add_option("-DN=" + support::cpp11::to_string(gemm_info.n));
281  build_opts.add_option("-DK=" + support::cpp11::to_string(gemm_info.k));
282  build_opts.add_option("-DM0=" + support::cpp11::to_string(internal_m0));
283  build_opts.add_option("-DN0=" + support::cpp11::to_string(rhs_info.n0));
284  build_opts.add_option("-DK0=" + support::cpp11::to_string(rhs_info.k0));
285  build_opts.add_option("-DPARTIAL_STORE_M0=" + support::cpp11::to_string(partial_store_m0));
286  build_opts.add_option("-DPARTIAL_STORE_N0=" + support::cpp11::to_string(partial_store_n0));
287  build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DACTIVATION_TYPE=" + lower_string(string_from_activation_func(gemm_info.activation_info.activation())));
288  build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DA_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.a()));
289  build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DB_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.b()));
290 
291  std::string kernel_name("gemm_mm_native");
292 
293  // Create kernel
294  _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
295 
296  // Set config_id for enabling LWS tuning
297  _config_id = kernel_name;
298  _config_id += "_";
299  _config_id += (_add_bias ? "add_bias_" : "");
300  _config_id += (_broadcast_bias ? "broadcast_bias_" : "");
301  _config_id += (_reinterpret_input_as_3d ? "3di_" : "");
302  _config_id += (_reinterpret_output_as_3d ? "3do_" : "");
303  _config_id += (gemm_info.activation_info.enabled() ? "fused_activation_" : "");
304  _config_id += lower_string(string_from_data_type(input0->info()->data_type()));
305  _config_id += "_";
306  _config_id += support::cpp11::to_string(output->info()->dimension(1));
307  _config_id += "_";
308  _config_id += support::cpp11::to_string(output->info()->dimension(0));
309  _config_id += "_";
310  _config_id += support::cpp11::to_string(gemm_info.k);
311  _config_id += "_";
312  _config_id += support::cpp11::to_string(output->info()->dimension(2));
313  _config_id += "_";
314  _config_id += support::cpp11::to_string(lhs_info.m0);
315  _config_id += "_";
316  _config_id += support::cpp11::to_string(rhs_info.n0);
317  _config_id += "_";
318  _config_id += support::cpp11::to_string(rhs_info.k0);
319 
321 }
322 
323 Status CLGEMMMatrixMultiplyNativeKernel::validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, float alpha, float beta,
324  const GEMMLHSMatrixInfo &lhs_info,
325  const GEMMRHSMatrixInfo &rhs_info, const GEMMKernelInfo &gemm_info)
326 {
327  ElementsProcessed num_elements_processed{};
328  ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input0, input1, input2, output, alpha, beta, lhs_info, rhs_info, gemm_info));
329  ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input0->clone().get(),
330  input1->clone().get(),
331  input2 != nullptr ? input2->clone().get() : nullptr,
332  output->clone().get(),
333  lhs_info,
334  rhs_info,
335  gemm_info,
336  num_elements_processed)
337  .first);
338 
339  return Status{};
340 }
341 
342 void CLGEMMMatrixMultiplyNativeKernel::run(const Window &window, cl::CommandQueue &queue)
343 {
346 
347  if(_input1->info()->num_dimensions() < 3)
348  {
349  // The stride_z for matrix B must be zero if we do not slice
350  ARM_COMPUTE_ERROR_ON(_input1->info()->strides_in_bytes()[3] != 0);
351  }
352 
354  Window slice_matrix_b = slice;
355 
356  slice_matrix_b.set(Window::DimX, Window::Dimension(0, 1, 1));
357  slice_matrix_b.set(Window::DimY, Window::Dimension(0, 1, 1));
358 
359  if(_reinterpret_input_as_3d)
360  {
361  // Pass bottom paddings to the kernel if the input has to be reinterpreted as 3D tensor
362  unsigned int idx0;
363  if(_add_bias)
364  {
365  idx0 = 4 * num_arguments_per_2D_tensor() + 4;
366  }
367  else
368  {
369  idx0 = 3 * num_arguments_per_2D_tensor() + 3;
370  }
371  const unsigned int total_cross_plane_pad = _input0->info()->padding().top + _input0->info()->padding().bottom;
372  _kernel.setArg<cl_uint>(idx0, static_cast<unsigned int>(total_cross_plane_pad));
373  }
374 
375  if(_reinterpret_output_as_3d)
376  {
377  // Pass bottom paddings to the kernel if the output has to be reinterpreted as 3D tensor
378  unsigned int idx0;
379  if(_add_bias)
380  {
381  idx0 = 4 * num_arguments_per_2D_tensor() + 4 + (_reinterpret_input_as_3d ? 1 : 0);
382  }
383  else
384  {
385  idx0 = 3 * num_arguments_per_2D_tensor() + 3 + (_reinterpret_input_as_3d ? 1 : 0);
386  }
387  const unsigned int total_cross_plane_pad = _output->info()->padding().top + _output->info()->padding().bottom;
388  _kernel.setArg<cl_uint>(idx0, static_cast<unsigned int>(total_cross_plane_pad));
389  }
390 
391  do
392  {
393  Window slice_b = slice;
394  // Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2
395  // This scenario can happen when the matrix multiplication is used to perform a convolution operation
396  if(!_slide_matrix_b)
397  {
398  slice_b = slice_matrix_b;
399  }
400 
401  unsigned int idx = 0;
402  add_2D_tensor_argument(idx, _input0, slice);
403  add_2D_tensor_argument(idx, _input1, slice_b);
404  if(_add_bias)
405  {
406  add_2D_tensor_argument(idx, _input2, slice);
407  }
408  add_2D_tensor_argument(idx, _output, slice);
409  _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input0->info()->strides_in_bytes()[2]));
410  _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input1->info()->strides_in_bytes()[2]));
411  if(_add_bias)
412  {
413  _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input2->info()->strides_in_bytes()[2]));
414  }
415  _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_output->info()->strides_in_bytes()[2]));
416  enqueue(queue, *this, slice, lws_hint(), _use_dummy_work_items);
417  }
419 }
420 } // namespace arm_compute
unsigned int top
top of the border
Definition: Types.h:375
virtual size_t num_dimensions() const =0
The number of dimensions of the tensor (rank)
bool is_one(float a, float epsilon=0.00001f)
Checks if the input floating point number is 1.0f checking if the difference is within a range define...
Definition: float_ops.h:97
Window calculate_max_window(const ValidRegion &valid_region, const Steps &steps, bool skip_border, BorderSize border_size)
bool broadcast_bias
Flag used to broadcast the bias addition.
const Window & window() const
The maximum window the kernel can be executed on.
Definition: IKernel.cpp:28
Descriptor used by the GEMM kernels.
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:489
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.
unsigned int depth_output_gemm3d
Depth of the output tensor in case is reinterpreted as 3D.
cl::NDRange lws_hint() const
Return the Local-Workgroup-Size hint.
Definition: ICLKernel.h:276
bool preferred_dummy_work_items_support(const cl::Device &device)
Helper function to check if "dummy work-items" are preferred to have a power of two NDRange In case d...
Definition: CLHelpers.cpp:361
#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.
TensorShape compute_mm_shape(const ITensorInfo &input0, const ITensorInfo &input1, bool is_interleaved_transposed, const GEMMReshapeInfo &reshape_info)
Calculate the matrix multiplication output shape of two tensors.
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.
GEMM LHS (Left Hand Side) matrix information.
Definition: Types.h:1903
Store the tensor's metadata.
Definition: ITensorInfo.h:40
#define ARM_COMPUTE_ERROR_THROW_ON(status)
Definition: Error.h:455
Describe one of the image's dimensions with a start, end and step.
Definition: Window.h:77
unsigned int bottom
bottom of the border
Definition: Types.h:377
Status class.
Definition: Error.h:52
std::string lower_string(const std::string &val)
Lower a given string.
Definition: Utils.cpp:326
#define ARM_COMPUTE_RETURN_ERROR_ON(cond)
If the condition is true, an error is returned.
Definition: Error.h:296
Copyright (c) 2017-2021 Arm Limited.
#define ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(...)
Definition: Validate.h:159
void add_option(std::string option)
Adds option to the existing build option list.
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:403
const std::string & string_from_data_type(DataType dt)
Convert a data type identity into a string.
Definition: Utils.cpp:135
static Status validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, float alpha, float beta, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, const GEMMKernelInfo &gemm_info)
Static function to check if given info will lead to a valid configuration of CLGEMMMatrixMultiplyNati...
static constexpr size_t DimX
Alias for dimension 0 also known as X dimension.
Definition: Window.h:43
bool update_window_and_padding(Window &win, Ts &&... patterns)
Update window and padding size for each of the access patterns.
Definition: WindowHelpers.h:46
#define ARM_COMPUTE_UNUSED(...)
To avoid unused variables warnings.
Definition: Error.h:152
GEMM RHS (Right Hand Side) matrix information.
Definition: Types.h:1918
std::string float_to_string_with_full_precision(float val)
Create a string with the float in full precision.
Definition: Utils.h:1061
auto ceil_to_multiple(S value, T divisor) -> decltype(((value+divisor - 1)/divisor) *divisor)
Computes the smallest number larger or equal to value that is a multiple of divisor.
Definition: Utils.h:71
std::string kernel_name
std::string get_cl_type_from_data_type(const DataType &dt)
Translates a tensor data type to the appropriate OpenCL type.
Definition: CLHelpers.cpp:37
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...
bool reinterpret_input_as_3d
Flag used to reinterpret the input as 3D.
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's metadata.
void add_option_if(bool cond, std::string option)
Adds option if a given condition is true;.
void set(size_t dimension, const Dimension &dim)
Set the values of a given dimension.
Definition: Window.inl:49
virtual PaddingSize padding() const =0
Padding of tensor.
static constexpr unsigned int num_arguments_per_2D_tensor()
Returns the number of arguments enqueued per 2D tensor object.
Definition: ICLKernel.h:206
bool slide_window_slice_3D(Window &slice) const
Slide the passed 3D window slice.
Definition: Window.h:335
#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:504
CLCompileContext class.
static constexpr size_t DimY
Alias for dimension 1 also known as Y dimension.
Definition: Window.h:45
void add_2D_tensor_argument(unsigned int &idx, const ICLTensor *tensor, const Window &window)
Add the passed 2D tensor's parameters to the object's kernel's arguments starting from the index idx.
Definition: ICLKernel.h:148
Interface for OpenCL tensor.
Definition: ICLTensor.h:42
#define ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(...)
Definition: Validate.h:439
#define ARM_COMPUTE_CREATE_ERROR(error_code, msg)
Creates an error with a given message.
Definition: Error.h:159
void configure(const ICLTensor *input0, const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output, float alpha, float beta, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, const GEMMKernelInfo &gemm_info)
Initialise the kernel's input and output.
Manages all the OpenCL kernels compilation and caching, provides accessors for the OpenCL Context.
#define ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(...)
Definition: Validate.h:541
#define ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(t, c,...)
Definition: Validate.h:788
Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, const GEMMLowpOutputStageInfo *output_stage)
Wrapper to configure the Khronos OpenCL C++ header.
bool is_zero(float a, float epsilon=0.00001f)
Checks if the input floating point number is 0.0f checking if the difference is within a range define...
Definition: float_ops.h:109
#define ARM_COMPUTE_RETURN_ERROR_ON_MSG(cond, msg)
If the condition is true, an error is returned.
Definition: Error.h:244
#define ARM_COMPUTE_ERROR_ON_NULLPTR(...)
Definition: Validate.h:157
virtual const Strides & strides_in_bytes() const =0
The strides in bytes for accessing each dimension of the tensor.
Window first_slice_window_3D() const
First 3D slice of the window.
Definition: Window.h:291
Describe a multidimensional execution window.
Definition: Window.h:39
#define ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(f, s)
Definition: Validate.h:201
void run(const Window &window, cl::CommandQueue &queue) override
Enqueue the OpenCL kernel to process the given window on the passed OpenCL command queue.
SimpleTensor< T > slice(const SimpleTensor< T > &src, Coordinates starts, Coordinates ends)