Compute Library
 21.11
CpuGemmConv2d.cpp
Go to the documentation of this file.
1 /*
2  * Copyright (c) 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 
28 #include "arm_compute/core/Utils.h"
33 
34 #include "src/common/utils/Log.h"
44 
45 #include <set>
46 #include <tuple>
47 
49 using namespace arm_compute::experimental;
50 
51 namespace arm_compute
52 {
53 namespace cpu
54 {
55 CpuGemmConv2d::CpuGemmConv2d()
56  : _weights_reshape_kernel(nullptr), _im2col_kernel(), _mm_gemm(), _mm_gemmlowp(), _col2im_kernel(), _reshape_kernel(), _im2col_output(), _weights_reshaped(), _gemm_output(), _gemm_output_3d(),
57  _data_layout(DataLayout::NCHW), _skip_im2col(false), _skip_col2im(false), _is_quantized(false), _is_prepared(false), _aux_mem(AuxTensorIdx::Count)
58 {
59 }
61 
62 void CpuGemmConv2d::configure_mm(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, ITensorInfo *dst, const ActivationLayerInfo &act_info,
63  bool enable_fast_math, int gemm_3d_depth)
64 {
65  ARM_COMPUTE_ERROR_ON_NULLPTR(src, weights);
66  ARM_COMPUTE_ERROR_THROW_ON(validate_mm(src, weights, biases, dst, act_info, enable_fast_math, gemm_3d_depth, _skip_im2col));
67 
68  // Create GEMMInfo structure
69  const GEMMInfo &gemm_info = GEMMInfo(false, false, true /* Reshape weights only for the first run */,
70  gemm_3d_depth, _skip_im2col /* Reinterpret the input as 3D if im2col is skipped */,
71  false, GEMMLowpOutputStageInfo(), false, enable_fast_math, false, act_info);
72 
73  // Supported activations in GEMM
74  const std::set<ActivationLayerInfo::ActivationFunction> supported_acts = { ActivationLayerInfo::ActivationFunction::RELU,
77  };
78 
79  if(_is_quantized)
80  {
81  TensorInfo tmp_src{ *src };
82  TensorInfo tmp_weights{ *weights };
83  // Since we need negative offsets for computing convolution, we need to change QuantizationInfo()
84  // Extract and negate input and weights offset
85  const QuantizationInfo iqinfo = src->quantization_info();
86  const QuantizationInfo wqinfo = weights->quantization_info();
87  const QuantizationInfo oqinfo = (dst->total_size() == 0) ? iqinfo : dst->quantization_info();
88  const UniformQuantizationInfo uiqinfo = iqinfo.uniform();
89  const UniformQuantizationInfo uoqinfo = oqinfo.uniform();
90  const DataType data_type = src->data_type();
91 
92  tmp_src.set_quantization_info(QuantizationInfo(uiqinfo.scale, -uiqinfo.offset));
93  if(!is_data_type_quantized_per_channel(tmp_weights.data_type()))
94  {
95  const UniformQuantizationInfo uwqinfo = wqinfo.uniform();
96  tmp_weights.set_quantization_info(QuantizationInfo(uwqinfo.scale, -uwqinfo.offset));
97  }
98 
99  // Merge activation with output stage
100  PixelValue type_min{};
101  PixelValue type_max{};
102  std::tie(type_min, type_max) = get_min_max(data_type);
103  int32_t min_activation = type_min.get<int32_t>();
104  int32_t max_activation = type_max.get<int32_t>();
105 
106  if(supported_acts.count(act_info.activation()) != 0)
107  {
108  std::tie(min_activation, max_activation) = get_quantized_activation_min_max(act_info, data_type, uoqinfo);
109  }
110 
113  output_info.gemmlowp_offset = uoqinfo.offset;
114  output_info.gemmlowp_min_bound = min_activation;
115  output_info.gemmlowp_max_bound = max_activation;
116  output_info.is_quantized_per_channel = (tmp_weights.data_type() == DataType::QSYMM8_PER_CHANNEL);
117  quantization::calculate_quantized_multipliers(iqinfo, wqinfo, oqinfo, output_info);
118 
119  _mm_gemmlowp = std::make_unique<CpuGemmLowpMatrixMultiplyCore>();
120  _mm_gemmlowp->configure(&tmp_src, &tmp_weights, biases, dst, GEMMInfo(false, false, true, gemm_3d_depth, _skip_im2col, false, output_info, false, enable_fast_math, false, act_info));
121 
122  auto mm_mem_req = _mm_gemmlowp->workspace();
123  for(unsigned int cont = 0; cont < mm_mem_req.size(); ++cont)
124  {
125  _aux_mem[cont] = mm_mem_req[cont];
126  }
127  }
128  else
129  {
130  // Configure matrix multiply function
131  _mm_gemm = std::make_unique<CpuGemm>();
132  _mm_gemm->configure(src, weights, biases, dst, 1.0f, 0.0f, gemm_info);
133  auto mm_mem_req = _mm_gemm->workspace();
134  for(unsigned int cont = 0; cont < mm_mem_req.size(); ++cont)
135  {
136  _aux_mem[cont] = mm_mem_req[cont];
137  }
138  }
139 }
140 
141 Status CpuGemmConv2d::validate_mm(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst,
142  const ActivationLayerInfo &act_info, bool enable_fast_math, int gemm_3d_depth, bool skip_im2col)
143 {
144  const DataType data_type = src->data_type();
145  const bool is_quantized = is_data_type_quantized_asymmetric(data_type);
146  const bool is_activation_enabled = act_info.enabled();
147 
148  // Create GEMMInfo structure
149  const GEMMInfo gemm_info = GEMMInfo(false, false, true /* Reshape weights only for the first run */,
150  gemm_3d_depth, skip_im2col /* Reinterpret the input as 3D if im2col is skipped */,
151  false, GEMMLowpOutputStageInfo(), false, enable_fast_math, false, act_info);
152 
153  if(is_quantized)
154  {
155  // Since we need negative offsets for computing convolution, we need to change QuantizationInfo()
156  // Extract and negate input and weights offset
157  const QuantizationInfo &iqinfo = src->quantization_info();
158  const QuantizationInfo &wqinfo = weights->quantization_info();
159  const QuantizationInfo &oqinfo = (dst->total_size() == 0) ? iqinfo : dst->quantization_info();
160  const UniformQuantizationInfo uoqinfo = oqinfo.uniform();
161 
162  // Merge activation with output stage
163  PixelValue type_min{};
164  PixelValue type_max{};
165  std::tie(type_min, type_max) = get_min_max(data_type);
166  int32_t min_activation = type_min.get<int32_t>();
167  int32_t max_activation = type_max.get<int32_t>();
168 
169  const std::set<ActivationLayerInfo::ActivationFunction> supported_acts = { ActivationLayerInfo::ActivationFunction::RELU,
172  };
173  if(is_activation_enabled && supported_acts.count(act_info.activation()) != 0)
174  {
175  std::tie(min_activation, max_activation) = get_quantized_activation_min_max(act_info, data_type, uoqinfo);
176  }
177 
180  output_info.gemmlowp_offset = uoqinfo.offset;
181  output_info.gemmlowp_min_bound = min_activation;
182  output_info.gemmlowp_max_bound = max_activation;
185 
186  // Perform validation step on GEMMLowp
187  std::unique_ptr<ITensorInfo> input_qa = src->clone();
188  std::unique_ptr<ITensorInfo> weights_qa = weights->clone();
189  input_qa->set_quantization_info(QuantizationInfo(iqinfo.uniform().scale, -iqinfo.uniform().offset));
190  weights_qa->set_quantization_info(QuantizationInfo(wqinfo.uniform().scale, -wqinfo.uniform().offset));
191  return CpuGemmLowpMatrixMultiplyCore::validate(input_qa.get(), weights_qa.get(), biases, dst, GEMMInfo(false, false, true, gemm_3d_depth, skip_im2col, false, output_info,
192  false, enable_fast_math, false, act_info));
193  }
194  else
195  {
196  // Perform validation step on Matrix multiply function
197  return CpuGemm::validate(src, weights, nullptr, dst, 1.0f, 0.0f, gemm_info);
198  }
199 }
200 
201 Status CpuGemmConv2d::validate_gemm3d(const ITensorInfo *input_info, const ITensorInfo *weights_info, const ActivationLayerInfo &act_info, int gemm_3d_depth, bool skip_im2col)
202 {
203  const DataType data_type = input_info->data_type();
204  const unsigned int mult_y = skip_im2col ? 1U : gemm_3d_depth;
205  const unsigned int mult_z = skip_im2col ? gemm_3d_depth : 1U;
206 
207  // Set dummy tensor shapes for the validation
208  const TensorInfo dummy_input_info(TensorShape(4U, 4U * mult_y, 1U * mult_z), 1, data_type, input_info->quantization_info());
209  const TensorInfo dummy_weights_info(TensorShape(4U, 4U), 1, data_type, weights_info->quantization_info());
210  const TensorInfo dummy_output_info(TensorShape(4U, 4U, gemm_3d_depth), 1, data_type, input_info->quantization_info());
211 
212  return validate_mm(&dummy_input_info, &dummy_weights_info, nullptr, &dummy_output_info, act_info, false, gemm_3d_depth, skip_im2col);
213 }
214 
215 void CpuGemmConv2d::configure(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, ITensorInfo *dst, const PadStrideInfo &conv_info, const WeightsInfo &weights_info,
216  const Size2D &dilation, const ActivationLayerInfo &act_info, bool enable_fast_math, unsigned int num_groups)
217 {
218  ARM_COMPUTE_ERROR_ON_NULLPTR(src, weights, dst);
219  ARM_COMPUTE_UNUSED(num_groups, weights_info);
221  weights,
222  biases,
223  dst,
224  conv_info,
225  weights_info,
226  dilation,
227  act_info,
228  enable_fast_math,
229  num_groups));
230  ARM_COMPUTE_LOG_PARAMS(src, weights, biases, dst, conv_info, weights_info, dilation, act_info, enable_fast_math, num_groups);
231 
232  const DataType data_type = src->data_type();
233  const DataLayout data_layout = src->data_layout();
236  const int idx_kernels = get_data_layout_dimension_index(data_layout, DataLayoutDimension::BATCHES);
237 
238  const unsigned int kernel_width = weights->dimension(idx_width);
239  const unsigned int kernel_height = weights->dimension(idx_height);
240 
241  _is_prepared = weights_info.retain_internal_weights();
242  _is_quantized = is_data_type_quantized_asymmetric(src->data_type());
243  _data_layout = data_layout;
244  _skip_im2col = (data_layout == DataLayout::NHWC && kernel_width == 1 && kernel_height == 1 && conv_info.stride().first == 1 && conv_info.stride().second == 1);
245 
246  const ITensorInfo *gemm_input_to_use = src;
247  ITensorInfo *gemm_output_to_use = dst;
248 
249  // Get convolved dimensions
250  unsigned int conv_w = 0;
251  unsigned int conv_h = 0;
252  std::tie(conv_w, conv_h) = scaled_dimensions(src->dimension(idx_width),
253  src->dimension(idx_height),
254  kernel_width,
255  kernel_height,
256  conv_info,
257  dilation);
258  ARM_COMPUTE_ERROR_ON_MSG((dst->dimension(idx_width) != conv_w) || (dst->dimension(idx_height) != conv_h),
259  "Output shape does not match the expected one");
260 
261  // Check if GEMM3D is supported
262  if(data_layout == DataLayout::NHWC)
263  {
264  _skip_col2im = bool(validate_gemm3d(src, weights, act_info, conv_h, true));
265  // If not supported, we need to perform im2col and col2im (or reshape layer)
266  if(!_skip_col2im)
267  {
268  _skip_im2col = false;
269  }
270  }
271  else
272  {
273  _skip_col2im = false;
274  }
275 
276  // Get parameters from conv_info
277  unsigned int stride_x = 0;
278  unsigned int stride_y = 0;
279  std::tie(stride_x, stride_y) = conv_info.stride();
280 
281  unsigned int mat_weights_cols = weights->dimension(idx_kernels);
282 
283  // _weights_reshaped will be auto configured in the kernel.
284  // Just append biases and do not transpose 1xW as it will be reshaped in CpuGemm
285  _weights_reshape_kernel = std::make_unique<kernels::CpuWeightsReshapeKernel>();
286  _weights_reshape_kernel->configure(weights, nullptr, &_weights_reshaped);
287  _weights_reshaped.set_quantization_info(weights->quantization_info());
288 
289  // Create tensor to store im2col reshaped inputs
290  if(!_skip_im2col)
291  {
292  // Configure
293  _im2col_kernel = std::make_unique<kernels::CpuIm2ColKernel>();
294  _im2col_kernel->configure(src, &_im2col_output, Size2D(kernel_width, kernel_height), conv_info, false, dilation);
295 
296  // Update GEMM input
297  gemm_input_to_use = &_im2col_output;
298  }
299 
300  // Create temporary GEMM output tensor in case we cannot skip col2im
301  const DataType output_data_type = data_type == DataType::BFLOAT16 ? DataType::F32 : data_type;
302  if(!_skip_col2im)
303  {
304  TensorShape shape_gemm;
305 
306  // Calculate GEMM output shape
307  shape_gemm = _im2col_output.tensor_shape();
308  shape_gemm.set(0, mat_weights_cols);
309  shape_gemm.set(1, conv_w * conv_h);
310 
311  _gemm_output = TensorInfo(shape_gemm, 1, output_data_type);
313  _gemm_output_3d = TensorInfo(_gemm_output);
314 
315  // Update GEMM output
316  gemm_output_to_use = &_gemm_output;
317  }
318  else
319  {
320  _gemm_output_3d = TensorInfo(*dst);
321  _gemm_output_3d.set_data_type(output_data_type).set_data_layout(src->data_layout()).set_is_resizable(true);
322  _gemm_output = TensorInfo(_gemm_output_3d);
323 
324  // Update GEMM output
325  gemm_output_to_use = &_gemm_output_3d;
326  }
327 
328  // Configure GEMM
329  // In case we need to skip col2im, GEMM3D (gemm_3d_depth != 0) must be called in order to avoid reshaping the output matrix
330  const unsigned int gemm_3d_depth = _skip_col2im ? conv_h : 0;
331  configure_mm(gemm_input_to_use, &_weights_reshaped, biases, gemm_output_to_use, act_info, enable_fast_math, gemm_3d_depth);
332 
333  if(!_skip_col2im && _data_layout == DataLayout::NCHW)
334  {
335  // Configure col2im
336  _col2im_kernel = std::make_unique<kernels::CpuCol2ImKernel>();
337  _col2im_kernel->configure(gemm_output_to_use, dst, Size2D(conv_w, conv_h));
338  }
339  else
340  {
341  // Configure reshape layer
342  _reshape_kernel = std::make_unique<kernels::CpuReshapeKernel>();
343  _reshape_kernel->configure(gemm_output_to_use, dst);
344  }
345 
346  // Check if GEMM transforms weights
347  // Modernise through COMPMID-4535
348  bool gemm_trans_wei = _aux_mem[1].size > 0; // Asm Pretranspose
349  gemm_trans_wei = _mm_gemm != nullptr ? _aux_mem[3].size > 0 : gemm_trans_wei; // Tranpose RHS
350  gemm_trans_wei = _mm_gemmlowp != nullptr ? _aux_mem[5].size > 0 : gemm_trans_wei; // Transpose RHS
351 
352  // Check lifetime
353  _aux_mem[Im2ColOutput] = MemoryInfo(offset_int_vec(Im2ColOutput), MemoryLifetime::Temporary, _im2col_output.total_size());
354  _aux_mem[WeightsReshaped] = MemoryInfo(offset_int_vec(WeightsReshaped), gemm_trans_wei ? MemoryLifetime::Prepare : MemoryLifetime::Persistent, _weights_reshaped.total_size());
355  _aux_mem[GemmOutput] = MemoryInfo(offset_int_vec(GemmOutput), MemoryLifetime::Temporary, _gemm_output.total_size());
356 }
357 
358 Status CpuGemmConv2d::validate(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, const PadStrideInfo &conv_info,
359  const WeightsInfo &weights_info, const Size2D &dilation, const ActivationLayerInfo &act_info, bool enable_fast_math, unsigned int num_groups)
360 {
361  ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, weights, dst);
362  ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights_info.are_reshaped(), "Weights already reshaped are not supported!");
366  ARM_COMPUTE_RETURN_ERROR_ON_MSG(num_groups > 1, "Grouping (num_groups != 1) is not supported");
367 
368  const DataLayout data_layout = src->data_layout();
369  const DataType data_type = src->data_type();
372  const int idx_channel = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL);
373  const int idx_kernels = get_data_layout_dimension_index(data_layout, DataLayoutDimension::BATCHES);
374 
375  const unsigned int kernel_width = weights->dimension(idx_width);
376  const unsigned int kernel_height = weights->dimension(idx_height);
377 
378  TensorInfo im2col_reshaped_info{};
379  TensorInfo info_gemm{};
380  TensorInfo tmp_info{};
381  TensorInfo weights_reshaped_info{};
382  const ITensorInfo *gemm_input_to_use = src;
383  const ITensorInfo *gemm_output_to_use = dst;
384  const ITensorInfo *weights_to_use = weights;
385 
386  const bool append_bias = false;
387  const bool is_quantized = is_data_type_quantized_asymmetric(data_type);
388  const bool is_bf16 = data_type == DataType::BFLOAT16;
389  bool skip_im2col = (data_layout == DataLayout::NHWC && kernel_width == 1 && kernel_height == 1 && conv_info.stride().first == 1 && conv_info.stride().second == 1);
390 
391  // Get convolved dimensions
392  unsigned int conv_w = 0;
393  unsigned int conv_h = 0;
394 
395  std::tie(conv_w, conv_h) = scaled_dimensions(src->dimension(idx_width),
396  src->dimension(idx_height),
397  kernel_width,
398  kernel_height,
399  conv_info,
400  dilation);
401 
402  // Check if GEMM3D is supported
403  bool skip_col2im = false;
404  if(data_layout == DataLayout::NHWC)
405  {
406  skip_col2im = bool(validate_gemm3d(src, weights, act_info, conv_h, true));
407  // If not supported, we need to perform im2col and col2im (or reshape layer)
408  if(!skip_col2im)
409  {
410  skip_im2col = false;
411  }
412  }
413 
414  if(skip_col2im)
415  {
416  // If not supported, we need to perform im2col and col2im (or reshape layer)
417  if(!bool(validate_gemm3d(src, weights, act_info, conv_h, skip_im2col)))
418  {
419  skip_im2col = false;
420  skip_col2im = false;
421  }
422  }
423 
424  ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(idx_channel) != src->dimension(idx_channel));
426 
427  // Validate biases
428  if(biases != nullptr)
429  {
430  if(is_quantized)
431  {
433  }
434  else if(is_bf16)
435  {
437  }
438  else
439  {
441  }
442  ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(idx_kernels));
444  }
445 
446  unsigned int mat_weights_cols = weights->dimension(idx_kernels);
447  unsigned int mat_weights_rows = weights->dimension(idx_width) * weights->dimension(idx_height) * weights->dimension(idx_channel);
448 
449  weights_reshaped_info = TensorInfo(compute_weights_reshaped_shape(*weights, append_bias), 1, data_type);
450  weights_reshaped_info.set_quantization_info(weights->quantization_info());
451  weights_to_use = &weights_reshaped_info;
452 
453  if(!skip_im2col)
454  {
455  // Create tensor info for im2col reshaped inputs
456  // For CPU, the batch size is on the fourth dimension
457  TensorShape shape_im2col = src->tensor_shape();
458  shape_im2col.set(0, mat_weights_rows);
459  shape_im2col.set(1, conv_w * conv_h);
460  shape_im2col.set(2, 1);
461 
462  im2col_reshaped_info = TensorInfo(shape_im2col, 1, data_type);
463  im2col_reshaped_info.set_quantization_info(src->quantization_info());
464  ARM_COMPUTE_RETURN_ON_ERROR(kernels::CpuIm2ColKernel::validate(src, &im2col_reshaped_info, Size2D(kernel_width, kernel_height), conv_info, append_bias, dilation));
465  gemm_input_to_use = &im2col_reshaped_info;
466  }
467 
468  // Create temporary GEMM output tensor in case we cannot skip col2im
469  const DataType output_data_type = data_type == DataType::BFLOAT16 ? DataType::F32 : data_type;
470  if(!skip_col2im)
471  {
472  TensorShape shape_gemm = gemm_input_to_use->tensor_shape();
473  shape_gemm.set(0, mat_weights_cols);
474  shape_gemm.set(1, conv_w * conv_h);
475  info_gemm = TensorInfo(shape_gemm, 1, output_data_type);
476  }
477  else
478  {
479  info_gemm = TensorInfo(dst->tensor_shape(), 1, output_data_type);
480  }
482  gemm_output_to_use = &info_gemm;
483  ARM_COMPUTE_RETURN_ON_ERROR(validate_mm(gemm_input_to_use, weights_to_use, biases, gemm_output_to_use, act_info, enable_fast_math, skip_col2im ? conv_h : 0, skip_im2col));
484 
485  // Validate Col2Im/ReshapeLayer
486  if(!skip_col2im && (data_layout == DataLayout::NCHW))
487  {
488  ARM_COMPUTE_RETURN_ON_ERROR(kernels::CpuCol2ImKernel::validate(gemm_output_to_use, dst, Size2D(conv_w, conv_h)));
489  }
490 
491  return Status{};
492 }
493 
495 {
496  prepare(tensors);
497 
498  auto src = tensors.get_const_tensor(ACL_SRC_0);
499  auto dst = tensors.get_tensor(ACL_DST);
500  auto gemm_input_to_use = src;
501 
502  CpuAuxTensorHandler im2col_output(offset_int_vec(Im2ColOutput), _im2col_output, tensors, false);
503  CpuAuxTensorHandler gemm_output(offset_int_vec(GemmOutput), _gemm_output, tensors, false);
504  CpuAuxTensorHandler reshaped_wei(offset_int_vec(WeightsReshaped), _weights_reshaped, tensors, false);
505 
506  bool out_has_padding = _skip_col2im && (dst->info()->padding().bottom != 0 || dst->info()->padding().top != 0);
507  if(!_skip_im2col)
508  {
509  // Run input reshaping
510  unsigned int y_dim = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::HEIGHT);
511  ITensorPack pack =
512  {
513  { TensorType::ACL_SRC, src },
514  { TensorType::ACL_DST, im2col_output.get() }
515  };
516  NEScheduler::get().schedule_op(_im2col_kernel.get(), y_dim, _im2col_kernel->window(), pack);
517  gemm_input_to_use = im2col_output.get();
518  }
519 
520  // Handle the case where output has top/bottom padding
521  const ITensor *out_to_use = out_has_padding ? gemm_output.get() : dst;
522  Tensor gemm3d;
523  _gemm_output_3d.extend_padding(out_to_use->info()->padding());
524  gemm3d.allocator()->soft_init(_gemm_output_3d);
525  gemm3d.allocator()->import_memory(out_to_use->buffer());
526  auto gemm_output_to_use = gemm_output.get();
527 
528  if(_skip_im2col)
529  {
530  gemm_output_to_use = &gemm3d;
531  }
532  if(_skip_col2im && !out_has_padding)
533  {
534  gemm_output_to_use = dst;
535  }
536 
537  // Runs CpuGemm or CpuGemmLowpMatrixMultiplyCore functions
538  ITensorPack pack_mm = tensors;
539  pack_mm.add_const_tensor(TensorType::ACL_SRC_0, gemm_input_to_use);
540  pack_mm.add_const_tensor(TensorType::ACL_SRC_1, reshaped_wei.get());
541  pack_mm.add_tensor(TensorType::ACL_DST, gemm_output_to_use);
542  if(_is_quantized)
543  {
544  // Run gemmlowp
545  _mm_gemmlowp->run(pack_mm);
546  }
547  else
548  {
549  // Run gemm
550  _mm_gemm->run(pack_mm);
551  }
552 
553  // Reshape output matrix
554  if(!_skip_col2im)
555  {
556  if(_data_layout == DataLayout::NCHW)
557  {
558  ITensorPack pack =
559  {
560  { TensorType::ACL_SRC, gemm_output.get() },
561  { TensorType::ACL_DST, dst }
562  };
563  NEScheduler::get().schedule_op(_col2im_kernel.get(), Window::DimY, _col2im_kernel->window(), pack);
564  }
565  else
566  {
567  ITensorPack pack =
568  {
569  { TensorType::ACL_SRC, gemm_output_to_use },
570  { TensorType::ACL_DST, dst }
571  };
572  NEScheduler::get().schedule_op(_reshape_kernel.get(), Window::DimY, _reshape_kernel->window(), pack);
573  }
574  }
575  else if(out_has_padding)
576  {
577  ITensorPack pack =
578  {
579  { TensorType::ACL_SRC, gemm_output_to_use },
580  { TensorType::ACL_DST, dst }
581  };
582  NEScheduler::get().schedule_op(_reshape_kernel.get(), Window::DimY, _reshape_kernel->window(), pack);
583  }
584 }
585 
587 {
588  if(!_is_prepared)
589  {
590  // Run weights reshaping and mark original weights tensor as unused
591  CpuAuxTensorHandler weights_reshaped(offset_int_vec(WeightsReshaped), _weights_reshaped, tensors);
592  auto weights = tensors.get_const_tensor(TensorType::ACL_SRC_1);
593  ITensorPack pack =
594  {
595  { TensorType::ACL_SRC, weights },
596  { TensorType::ACL_DST, weights_reshaped.get() }
597  };
598  NEScheduler::get().schedule_op(_weights_reshape_kernel.get(), 3, _weights_reshape_kernel->window(), pack);
599  weights->mark_as_unused();
600 
601  // Prepare GEMM
602  ITensorPack gemm_pack = tensors;
603  gemm_pack.add_const_tensor(TensorType::ACL_SRC_1, weights_reshaped.get());
604  _is_quantized ? _mm_gemmlowp->prepare(gemm_pack) : _mm_gemm->prepare(gemm_pack);
605 
606  _is_prepared = true;
607  }
608 }
610 {
611  return _aux_mem;
612 }
613 } // namespace cpu
614 } // namespace arm_compute
unsigned int top
top of the border
Definition: Types.h:377
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
Shape of a tensor.
Definition: TensorShape.h:39
Quantize using a fixed point multiplication.
void soft_init(TensorInfo &input, size_t alignment=0)
Initialize a tensor based with a reference TensorInfo.
#define ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(...)
Definition: Validate.h:490
bool enabled() const
Check if initialised.
Definition: Types.h:1559
virtual size_t dimension(size_t index) const =0
Return the size of the requested dimension.
void add_const_tensor(int id, const ITensor *tensor)
Add const tensor to the pack.
Definition: ITensorPack.cpp:49
#define ARM_COMPUTE_RETURN_ON_ERROR(status)
Checks if a status contains an error and returns it.
Definition: Error.h:204
virtual DataType data_type() const =0
Data type used for each element of the tensor.
virtual void schedule_op(ICPPKernel *kernel, const Hints &hints, const Window &window, ITensorPack &tensors)=0
Runs the kernel in the same thread as the caller synchronously.
static Status validate(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *d, float alpha, float beta, const GEMMInfo &gemm_info=GEMMInfo())
Static function to check if given info will lead to a valid configuration of CpuGemm.
Definition: CpuGemm.cpp:152
bool extend_padding(const PaddingSize &padding) override
Update the offset to the first element, the strides and the total size.
Definition: TensorInfo.cpp:247
bool are_reshaped() const
Flag which specifies if the weights tensor has been reshaped.
Definition: Types.h:1752
1 channel, 1 F32 per channel
ITensorInfo & set_data_type(DataType data_type) override
Set the data type to the specified value.
Definition: TensorInfo.cpp:287
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.
unsigned int bottom
bottom of the border
Definition: Types.h:379
int32_t gemmlowp_offset
GEMMLowp output stage offset used for quantizing to QASYMM8.
Definition: Types.h:1925
Status class.
Definition: Error.h:52
int32_t gemmlowp_max_bound
GEMMLowp max value used to saturate down the output result before converting back to QASYMM8...
Definition: Types.h:1929
#define ARM_COMPUTE_RETURN_ERROR_ON(cond)
If the condition is true, an error is returned.
Definition: Error.h:296
Activation Layer Information class.
Definition: Types.h:1509
GEMMLowpOutputStageType type
GEMMLowp output stage type.
Definition: Types.h:1924
Interface for CPU tensor.
Definition: ITensor.h:36
void prepare(ITensorPack &tensors) override
Prepare the function for executing.
SimpleTensor< float > src
Definition: DFT.cpp:155
Copyright (c) 2017-2021 Arm Limited.
std::vector< MemoryInfo > MemoryRequirements
Definition: Types.h:132
1 channel, 1 F16 per channel
std::pair< unsigned int, unsigned int > scaled_dimensions(int width, int height, int kernel_width, int kernel_height, const PadStrideInfo &pad_stride_info, const Size2D &dilation=Size2D(1U, 1U))
Returns expected width and height of output scaled tensor depending on dimensions rounding mode...
Definition: Utils.cpp:395
ITensorInfo & set_quantization_info(const QuantizationInfo &quantization_info) override
Set the quantization settings (scale and offset) of the tensor.
Definition: TensorInfo.cpp:346
TensorAllocator * allocator()
Return a pointer to the tensor&#39;s allocator.
Definition: Tensor.cpp:48
bool is_quantized_per_channel
GEMMLowp quantized per-channel flag.
Definition: Types.h:1933
Convolution Layer Weights Information class.
Definition: Types.h:1728
#define ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(...)
Definition: Validate.h:159
void run(ITensorPack &tensors) override
Run the kernels contained in the function.
1 channel, 1 S32 per channel
16-bit brain floating-point number
const DataType data_type
Definition: Im2Col.cpp:150
const ITensor * get_const_tensor(int id) const
Get constant tensor of a given id.
Definition: ITensorPack.cpp:54
Quantization information.
static Status validate(const ITensorInfo *src, const ITensorInfo *dst, const Size2D &convolved_dims)
Static function to check if given info will lead to a valid configuration.
#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:1058
virtual const TensorShape & tensor_shape() const =0
Size for each dimension of the tensor.
virtual ITensorInfo & set_data_layout(const DataLayout &data_layout)=0
Set the data layout of the tensor.
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:488
Status calculate_quantized_multipliers(const QuantizationInfo &iq_info, const QuantizationInfo &wq_info, const QuantizationInfo &oq_info, GEMMLowpOutputStageInfo &stage_info)
Calculate quantized representation of per-channel multipliers.
static Status validate(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info=WeightsInfo(), const Size2D &dilation=Size2D(1U, 1U), const ActivationLayerInfo &act_info=ActivationLayerInfo(), bool enable_fast_math=false, unsigned int num_groups=1)
Static function to check if given info will lead to a valid configuration.
quantized, asymmetric fixed-point 8-bit number unsigned
#define ARM_COMPUTE_ERROR_ON_MSG(cond, msg)
Definition: Error.h:456
const unsigned int num_groups
Definition: Im2Col.cpp:153
size_t total_size() const override
Returns the total size of the tensor in bytes.
Definition: TensorInfo.h:250
virtual uint8_t * buffer() const =0
Interface to be implemented by the child class to return a pointer to CPU memory. ...
std::pair< unsigned int, unsigned int > stride() const
Get the stride.
Definition: Types.h:704
experimental::MemoryRequirements workspace() const override
Return the memory requirements required by the workspace.
UniformQuantizationInfo uniform() const
Return per layer quantization info.
virtual std::unique_ptr< T > clone() const =0
Provide a clone of the current object of class T.
GEMMLowp output stage info.
Definition: Types.h:1922
virtual ITensorInfo * info() const =0
Interface to be implemented by the child class to return the tensor&#39;s metadata.
Basic implementation of the tensor interface.
Definition: Tensor.h:37
Padding and stride information class.
Definition: Types.h:656
virtual PaddingSize padding() const =0
Padding of tensor.
virtual QuantizationInfo quantization_info() const =0
Get the quantization settings (scale and offset) of the tensor.
Num samples, channels, height, width.
src_info set_data_layout(data_layout)
bool is_data_type_quantized_asymmetric(DataType dt)
Check if a given data type is of asymmetric quantized type.
Definition: Utils.h:1003
quantized, symmetric per channel fixed-point 8-bit number
static constexpr size_t DimY
Alias for dimension 1 also known as Y dimension.
Definition: Window.h:45
ITensor * get_tensor(int id)
Get tensor of a given id from the pac.
Definition: ITensorPack.cpp:64
virtual size_t total_size() const =0
Returns the total size of the tensor in bytes.
TensorShape compute_weights_reshaped_shape(const ITensorInfo &weights, bool has_bias=false, unsigned int num_groups=1)
Calculate the reshaped shape of the weights.
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
Class for specifying the size of an image or rectangle.
Definition: Size2D.h:34
#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
Status import_memory(void *memory)
Import an existing memory as a tensor&#39;s backing memory.
#define ARM_COMPUTE_RETURN_ERROR_ON_MSG(cond, msg)
If the condition is true, an error is returned.
Definition: Error.h:244
Tensor packing service.
Definition: ITensorPack.h:39
#define ARM_COMPUTE_LOG_PARAMS(...)
#define ARM_COMPUTE_ERROR_ON_NULLPTR(...)
Definition: Validate.h:157
Store the tensor&#39;s metadata.
Definition: TensorInfo.h:43
int offset_int_vec(int offset)
Definition: MemoryHelpers.h:38
GEMM information class.
Definition: Types.h:1974
ActivationFunction activation() const
Get the type of activation function.
Definition: Types.h:1544
quantized, asymmetric fixed-point 8-bit number signed
int32_t gemmlowp_min_bound
GEMMLowp min value used to saturate down the output result before converting back to QASYMM8...
Definition: Types.h:1928
const TensorShape & tensor_shape() const override
Size for each dimension of the tensor.
Definition: TensorInfo.h:234
void configure(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, ITensorInfo *dst, const PadStrideInfo &conv_info, const WeightsInfo &weights_info=WeightsInfo(), const Size2D &dilation=Size2D(1U, 1U), const ActivationLayerInfo &act_info=ActivationLayerInfo(), bool enable_fast_math=false, unsigned int num_groups=1)
Set the input and output tensors.
DataType
Available data types.
Definition: Types.h:79
DataLayout
[DataLayout enum definition]
Definition: Types.h:113
std::tuple< PixelValue, PixelValue > get_min_max(DataType dt)
Compute the mininum and maximum values a data type can take.
Definition: Utils.h:564
TensorShape & set(size_t dimension, size_t value, bool apply_dim_correction=true, bool increase_dim_unit=true)
Accessor to set the value of one of the dimensions.
Definition: TensorShape.h:79
bool retain_internal_weights() const
Definition: Types.h:1772
void add_tensor(int id, ITensor *tensor)
Add tensor to the pack.
Definition: ITensorPack.cpp:39
static Status validate(const ITensorInfo *src, const ITensorInfo *dst, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, const Size2D &dilation=Size2D(1U, 1U), unsigned int num_groups=1)
Static function to check if given info will lead to a valid configuration.
virtual DataLayout data_layout() const =0
Get the data layout of the tensor.
static IScheduler & get()
Access the scheduler singleton.
Definition: Scheduler.cpp:94
static Status validate(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *dst, const GEMMInfo &gemm_info=GEMMInfo())
Static function to check if given info will lead to a valid configuration.