• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (c) 2017-2021 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  */
24 #include "src/gpu/cl/operators/ClGemmConv2d.h"
25 
26 #include "arm_compute/core/CL/ICLTensor.h"
27 #include "arm_compute/core/PixelValue.h"
28 #include "arm_compute/core/Size2D.h"
29 #include "arm_compute/core/TensorInfo.h"
30 #include "arm_compute/core/Utils.h"
31 #include "arm_compute/core/Validate.h"
32 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
33 #include "arm_compute/core/utils/quantization/AsymmHelpers.h"
34 #include "arm_compute/runtime/CL/CLScheduler.h"
35 #include "src/core/helpers/AutoConfiguration.h"
36 #include "src/core/helpers/MemoryHelpers.h"
37 #include "src/gpu/cl/kernels/ClActivationKernel.h"
38 #include "src/gpu/cl/kernels/ClCol2ImKernel.h"
39 #include "src/gpu/cl/kernels/ClIm2ColKernel.h"
40 #include "src/gpu/cl/kernels/ClWeightsReshapeKernel.h"
41 #include "src/gpu/cl/operators/ClGemm.h"
42 #include "src/gpu/cl/operators/ClGemmLowpMatrixMultiplyCore.h"
43 #include "src/gpu/cl/utils/ClAuxTensorHandler.h"
44 
45 #include "src/common/utils/Log.h"
46 #include "support/Cast.h"
47 
48 namespace arm_compute
49 {
50 using namespace experimental;
51 using namespace misc::shape_calculator;
52 using namespace utils::cast;
53 namespace opencl
54 {
ClGemmConv2d()55 ClGemmConv2d::ClGemmConv2d()
56     : _weights_reshape_kernel(nullptr), _im2col_kernel(nullptr), _mm_gemm(nullptr), _mm_gemmlowp(nullptr), _col2im_kernel(nullptr), _activation_kernel(nullptr), _im2col_output(), _weights_reshaped(),
57       _gemm_output(), _skip_im2col(false), _skip_col2im(false), _is_quantized(false), _fuse_activation(true), _append_bias(false), _is_prepared(false), _use_post_ops(false), _aux_mem(AuxTensorIdx::Count)
58 {
59 }
60 ClGemmConv2d::~ClGemmConv2d() = default;
61 
configure_mm(const ClCompileContext & compile_context,const ITensorInfo * src,ITensorInfo * weights,ITensorInfo * biases,ITensorInfo * dst,const GEMMLowpOutputStageInfo & gemmlowp_output_stage,int gemm_3d_depth,const ActivationLayerInfo & act_info,const experimental::PostOpList<ITensorInfo * > & post_ops)62 void ClGemmConv2d::configure_mm(const ClCompileContext &compile_context, const ITensorInfo *src, ITensorInfo *weights, ITensorInfo *biases, ITensorInfo *dst,
63                                 const GEMMLowpOutputStageInfo &gemmlowp_output_stage,
64                                 int gemm_3d_depth, const ActivationLayerInfo &act_info, const experimental::PostOpList<ITensorInfo *> &post_ops)
65 {
66     ARM_COMPUTE_ERROR_ON_NULLPTR(src, weights);
67     ARM_COMPUTE_ERROR_THROW_ON(validate_mm(src, weights, biases, dst, gemmlowp_output_stage, gemm_3d_depth, _skip_im2col, act_info));
68 
69     const GEMMInfo &gemm_info = GEMMInfo(false,                 // is_a_reshaped
70                                          false,                 // is_b_reshaped
71                                          true,                  // reshape_b_only_on_first_run
72                                          gemm_3d_depth,         // depth_output_gemm3d
73                                          _skip_im2col,          // reinterpret_input_as_3d
74                                          false,                 // retain_internal_weights
75                                          gemmlowp_output_stage, // gemmlowp_output_stage
76                                          false,                 // fast_math
77                                          false,                 // fp_mixed_precision
78                                          true,                  // broadcast_bias
79                                          act_info,              // activation_info
80                                          post_ops               // post ops
81                                         );
82 
83     TensorInfo tmp_src{ *src };
84     if(_is_quantized)
85     {
86         ARM_COMPUTE_ERROR_ON_MSG(post_ops.size() > 0, "ClGemmConv2d quantized types do not support post ops");
87         // Since we need negative offsets for computing convolution, we need to change QuantizationInfo()
88         // Extract and negate input and weights offset
89         const QuantizationInfo input_quantization_info   = src->quantization_info();
90         const QuantizationInfo weights_quantization_info = weights->quantization_info();
91 
92         tmp_src.set_quantization_info(QuantizationInfo(input_quantization_info.uniform().scale, -input_quantization_info.uniform().offset));
93         weights->set_quantization_info(QuantizationInfo(weights_quantization_info.uniform().scale, -weights_quantization_info.uniform().offset));
94 
95         _mm_gemmlowp = std::make_unique<ClGemmLowpMatrixMultiplyCore>();
96         _mm_gemmlowp->configure(compile_context, &tmp_src, weights, biases, dst, gemm_info);
97 
98         // Revert back QuantizatioInfo as weights could be used in other convolution layers
99         weights->set_quantization_info(weights_quantization_info);
100 
101         auto mm_mem_req = _mm_gemmlowp->workspace();
102         for(unsigned int cont = 0; cont < mm_mem_req.size(); ++cont)
103         {
104             _aux_mem[cont] = mm_mem_req[cont];
105         }
106     }
107     else
108     {
109         // Configure matrix multiply function
110         _mm_gemm = std::make_unique<ClGemm>();
111         _mm_gemm->configure(compile_context, &tmp_src, weights, biases, dst, 1.0f, 1.0f, gemm_info);
112         auto mm_mem_req = _mm_gemm->workspace();
113         for(unsigned int cont = 0; cont < mm_mem_req.size(); ++cont)
114         {
115             _aux_mem[cont] = mm_mem_req[cont];
116         }
117     }
118 }
119 
validate_mm(const ITensorInfo * src,const ITensorInfo * weights,const ITensorInfo * biases,const ITensorInfo * dst,const GEMMLowpOutputStageInfo & gemmlowp_output_stage,int gemm_3d_depth,bool skip_im2col,const ActivationLayerInfo & act_info,const experimental::PostOpList<ITensorInfo * > & post_ops)120 Status ClGemmConv2d::validate_mm(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst,
121                                  const GEMMLowpOutputStageInfo &gemmlowp_output_stage, int gemm_3d_depth, bool skip_im2col, const ActivationLayerInfo &act_info, const experimental::PostOpList<ITensorInfo *> &post_ops)
122 {
123     const bool is_quantized = is_data_type_quantized_asymmetric(src->data_type());
124 
125     const GEMMInfo &gemm_info = GEMMInfo(false,                 // is_a_reshaped
126                                          false,                 // is_b_reshaped
127                                          true,                  // reshape_b_only_on_first_run
128                                          gemm_3d_depth,         // depth_output_gemm3d
129                                          skip_im2col,           // reinterpret_input_as_3d
130                                          false,                 // retain_internal_weights
131                                          gemmlowp_output_stage, // gemmlowp_output_stage
132                                          false,                 // fast_math
133                                          false,                 // fp_mixed_precision
134                                          true,                  // broadcast_bias
135                                          act_info,              // activation_info
136                                          post_ops               // post ops
137                                         );
138 
139     if(is_quantized)
140     {
141         ARM_COMPUTE_RETURN_ERROR_ON_MSG(post_ops.size() > 0, "ClGemmConv2d quantized types do not support post ops");
142         // Since we need negative offsets for computing convolution, we need to change QuantizationInfo()
143         // Extract and negate input and weights offset
144         const QuantizationInfo input_quantization_info   = src->quantization_info();
145         const QuantizationInfo weights_quantization_info = weights->quantization_info();
146 
147         std::unique_ptr<ITensorInfo> src_qa     = src->clone();
148         std::unique_ptr<ITensorInfo> weights_qa = weights->clone();
149         src_qa->set_quantization_info(QuantizationInfo(input_quantization_info.uniform().scale, -input_quantization_info.uniform().offset));
150         weights_qa->set_quantization_info(QuantizationInfo(weights_quantization_info.uniform().scale, -weights_quantization_info.uniform().offset));
151 
152         // Perform validation step on GEMMLowp
153         return ClGemmLowpMatrixMultiplyCore::validate(src_qa.get(), weights_qa.get(), biases, dst, gemm_info);
154     }
155     else
156     {
157         // Perform validation step on Matrix multiply function
158         return ClGemm::validate(src, weights, biases, dst, 1.0f, 1.0f, gemm_info);
159     }
160 }
161 
configure(const CLCompileContext & compile_context,ITensorInfo * src,ITensorInfo * weights,ITensorInfo * biases,ITensorInfo * dst,const Conv2dInfo & conv2d_info,const WeightsInfo & weights_info)162 void ClGemmConv2d::configure(const CLCompileContext &compile_context, ITensorInfo *src, ITensorInfo *weights, ITensorInfo *biases, ITensorInfo *dst,
163                              const Conv2dInfo &conv2d_info, const WeightsInfo &weights_info)
164 {
165     ARM_COMPUTE_ERROR_ON_NULLPTR(src, weights, dst);
166 
167     ARM_COMPUTE_ERROR_THROW_ON(ClGemmConv2d::validate(src, weights, biases, dst,
168                                                       conv2d_info,
169                                                       weights_info));
170     ARM_COMPUTE_LOG_PARAMS(src, weights, biases, dst, conv2d_info, weights_info);
171 
172     const DataType   data_type   = src->data_type();
173     const DataLayout data_layout = src->data_layout();
174     const int        idx_width   = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
175     const int        idx_height  = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
176     const int        idx_kernels = get_data_layout_dimension_index(data_layout, DataLayoutDimension::BATCHES);
177 
178     const unsigned int kernel_width  = weights->dimension(idx_width);
179     const unsigned int kernel_height = weights->dimension(idx_height);
180     const unsigned int num_kernels   = weights->dimension(idx_kernels);
181 
182     const UniformQuantizationInfo iq_info = src->quantization_info().uniform();
183     const UniformQuantizationInfo oq_info = dst->quantization_info().uniform();
184 
185     _is_prepared  = weights_info.retain_internal_weights();
186     _is_quantized = is_data_type_quantized_asymmetric(src->data_type());
187     _skip_im2col  = (data_layout == DataLayout::NHWC && kernel_width == 1 && kernel_height == 1 && conv2d_info.conv_info.stride().first == 1 && conv2d_info.conv_info.stride().second == 1);
188     _skip_col2im  = data_layout == DataLayout::NHWC;
189 
190     // Only for quantize there are few cases where we cannot fuse the activation function in GEMM
191     _fuse_activation = true;
192     _use_post_ops    = conv2d_info.post_ops.size() > 0;
193 
194     const ITensorInfo *gemm_input_to_use  = src;
195     ITensorInfo       *gemm_output_to_use = dst;
196 
197     // Get parameters from conv_info
198     unsigned int stride_x = 0;
199     unsigned int stride_y = 0;
200     std::tie(stride_x, stride_y) = conv2d_info.conv_info.stride();
201 
202     // Get convolved dimensions
203     unsigned int conv_w = 0;
204     unsigned int conv_h = 0;
205     std::tie(conv_w, conv_h) = scaled_dimensions(src->dimension(idx_width),
206                                                  src->dimension(idx_height),
207                                                  kernel_width,
208                                                  kernel_height,
209                                                  conv2d_info.conv_info,
210                                                  conv2d_info.dilation);
211 
212     unsigned int mat_weights_cols = num_kernels / conv2d_info.num_groups;
213 
214     ITensorInfo *biases_to_use = biases;
215     _append_bias               = false;
216 
217     _weights_reshape_kernel = std::make_unique<kernels::ClWeightsReshapeKernel>();
218     if(conv2d_info.num_groups != 1 && biases != nullptr)
219     {
220         // num_groups != 1 can only be for NCHW
221         // Since it is missing an utility function to reshape the biases, we append the biases into the weights tensor
222         biases_to_use = nullptr;
223         _append_bias  = true;
224         _weights_reshape_kernel->configure(compile_context, weights, biases, &_weights_reshaped, conv2d_info.num_groups);
225     }
226     else
227     {
228         _weights_reshape_kernel->configure(compile_context, weights, nullptr, &_weights_reshaped, conv2d_info.num_groups);
229     }
230 
231     // Create tensor to store im2col reshaped inputs
232     if(!_skip_im2col)
233     {
234         // Configure and tune im2col. im2col output shape is auto-initialized
235         _im2col_kernel = std::make_unique<opencl::kernels::ClIm2ColKernel>();
236 
237         // Set the GPU target for im2col
238         _im2col_kernel->set_target(CLScheduler::get().target());
239         _im2col_kernel->configure(compile_context, src, &_im2col_output, Size2D(kernel_width, kernel_height), conv2d_info.conv_info, _append_bias, conv2d_info.dilation, conv2d_info.num_groups);
240 
241         // Set quantization info
242         _im2col_output.set_quantization_info(src->quantization_info());
243         CLScheduler::get().tune_kernel_static(*_im2col_kernel);
244 
245         // Update GEMM input
246         gemm_input_to_use = &_im2col_output;
247     }
248 
249     // Create GEMM output tensor
250     if(!_skip_col2im)
251     {
252         TensorShape shape_gemm;
253 
254         // If we cannot skip col2im it means we run im2col as well
255         shape_gemm = _im2col_output.tensor_shape();
256         shape_gemm.set(0, mat_weights_cols);
257         shape_gemm.set(1, conv_w * conv_h);
258 
259         _gemm_output = TensorInfo(shape_gemm, 1, data_type);
260         _gemm_output.set_quantization_info(dst->quantization_info()).set_data_layout(src->data_layout());
261 
262         // Update GEMM output
263         gemm_output_to_use = &_gemm_output;
264     }
265 
266     GEMMLowpOutputStageInfo gemmlowp_output_stage;
267     gemmlowp_output_stage.type            = GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT;
268     gemmlowp_output_stage.gemmlowp_offset = 0;
269 
270     // Configure output stage for quantized case
271     if(_is_quantized)
272     {
273         const auto         output_quant_info        = (dst->total_size() == 0) ? iq_info : oq_info;
274         const bool         is_quantized_per_channel = is_data_type_quantized_per_channel(weights->data_type());
275         const unsigned int num_filters              = (is_quantized_per_channel) ? num_kernels : 1;
276 
277         gemmlowp_output_stage.is_quantized_per_channel = is_quantized_per_channel;
278 
279         gemmlowp_output_stage.gemmlowp_multipliers.resize(num_filters);
280         gemmlowp_output_stage.gemmlowp_shifts.resize(num_filters);
281         quantization::compute_quantized_multipliers_and_shifts(src, weights, dst,
282                                                                gemmlowp_output_stage.gemmlowp_multipliers.data(),
283                                                                gemmlowp_output_stage.gemmlowp_shifts.data());
284         gemmlowp_output_stage.gemmlowp_multiplier = gemmlowp_output_stage.gemmlowp_multipliers[0];
285         gemmlowp_output_stage.gemmlowp_shift      = gemmlowp_output_stage.gemmlowp_shifts[0];
286 
287         PixelValue min_val{};
288         PixelValue max_val{};
289         std::tie(min_val, max_val) = get_min_max(dst->data_type());
290 
291         auto min_activation = min_val.get<int32_t>();
292         auto max_activation = max_val.get<int32_t>();
293 
294         const std::set<ActivationLayerInfo::ActivationFunction> supported_acts = { ActivationLayerInfo::ActivationFunction::RELU,
295                                                                                    ActivationLayerInfo::ActivationFunction::BOUNDED_RELU,
296                                                                                    ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU
297                                                                                  };
298 
299         if(conv2d_info.act_info.enabled())
300         {
301             if(supported_acts.count(conv2d_info.act_info.activation()) != 0)
302             {
303                 std::tie(min_activation, max_activation) = get_quantized_activation_min_max(conv2d_info.act_info, data_type, output_quant_info);
304             }
305             else
306             {
307                 _fuse_activation = false;
308             }
309         }
310 
311         // Set the GEMMLowp output stage info
312         gemmlowp_output_stage.gemmlowp_offset    = output_quant_info.offset;
313         gemmlowp_output_stage.gemmlowp_min_bound = min_activation;
314         gemmlowp_output_stage.gemmlowp_max_bound = max_activation;
315     }
316 
317     // Configure and tune GEMM
318     // In case of NHWC, we need to run GEMM3D (gemm_3d_depth != 0) in order to avoid reshaping the output matrix
319     const unsigned int gemm_3d_depth = (data_layout == DataLayout::NHWC) ? conv_h : 0;
320 
321     configure_mm(compile_context, gemm_input_to_use, &_weights_reshaped, biases_to_use, gemm_output_to_use, gemmlowp_output_stage, gemm_3d_depth, conv2d_info.act_info, conv2d_info.post_ops);
322 
323     if(!_skip_col2im)
324     {
325         ARM_COMPUTE_ERROR_ON_MSG(conv2d_info.post_ops.size() > 0, "ClGemmConv2d does not support post ops with col2im operation"); // Post ops must be performed after every other op
326         // Set the GPU target for col2im
327         _col2im_kernel = std::make_unique<opencl::kernels::ClCol2ImKernel>();
328         _col2im_kernel->set_target(CLScheduler::get().target());
329         // Configure and tune Col2Im
330         _col2im_kernel->configure(compile_context, gemm_output_to_use, dst, Size2D(conv_w, conv_h), conv2d_info.num_groups);
331         CLScheduler::get().tune_kernel_static(*_col2im_kernel.get());
332     }
333 
334     ARM_COMPUTE_ERROR_ON_MSG((dst->dimension(idx_width) != conv_w) || (dst->dimension(idx_height) != conv_h),
335                              "Output shape does not match the expected one");
336 
337     // Disable running of activation kernel if post ops are used
338     if(!_fuse_activation && !_use_post_ops)
339     {
340         _activation_kernel = std::make_unique<opencl::kernels::ClActivationKernel>();
341         _activation_kernel->configure(compile_context, dst, nullptr, conv2d_info.act_info);
342     }
343 
344     _aux_mem[Im2ColOutput]    = MemoryInfo(offset_int_vec(Im2ColOutput), MemoryLifetime::Temporary, _im2col_output.total_size());
345     _aux_mem[WeightsReshaped] = MemoryInfo(offset_int_vec(WeightsReshaped), MemoryLifetime::Persistent, _weights_reshaped.total_size());
346     _aux_mem[GemmOutput]      = MemoryInfo(offset_int_vec(GemmOutput), MemoryLifetime::Temporary, _gemm_output.total_size());
347 }
348 
validate(const ITensorInfo * src,const ITensorInfo * weights,const ITensorInfo * biases,const ITensorInfo * dst,const Conv2dInfo & conv2d_info,const WeightsInfo & weights_info)349 Status ClGemmConv2d::validate(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, const Conv2dInfo &conv2d_info,
350                               const WeightsInfo &weights_info)
351 {
352     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, weights, dst);
353     ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights_info.are_reshaped(), "Weights already reshaped are not supported!");
354     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::F16, DataType::F32);
355     const bool is_quantized_per_channel = is_data_type_quantized_per_channel(weights->data_type());
356 
357     if(!is_quantized_per_channel)
358     {
359         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, weights);
360     }
361     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(src, weights);
362     ARM_COMPUTE_RETURN_ERROR_ON_MSG((conv2d_info.num_groups != 1) && (src->data_layout() != DataLayout::NCHW), "Grouping (num_groups != 1) with NHWC data layout is not supported");
363     ARM_COMPUTE_RETURN_ERROR_ON_MSG((conv2d_info.num_groups != 1) && (src->data_type() == DataType::QASYMM8), "Grouping (num_groups != 1) is not supported with QASYMM8");
364     ARM_COMPUTE_RETURN_ERROR_ON(((src->dimension(2) / weights->dimension(2)) != conv2d_info.num_groups) && (src->data_layout() == DataLayout::NCHW));
365 
366     const DataLayout data_layout = src->data_layout();
367     const DataType   data_type   = src->data_type();
368     const int        idx_width   = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
369     const int        idx_height  = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
370     const int        idx_channel = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL);
371     const int        idx_kernels = get_data_layout_dimension_index(data_layout, DataLayoutDimension::BATCHES);
372 
373     const unsigned int kernel_width  = weights->dimension(idx_width);
374     const unsigned int kernel_height = weights->dimension(idx_height);
375     const unsigned int num_kernels   = weights->dimension(idx_kernels);
376 
377     TensorInfo         im2col_reshaped_info{};
378     TensorInfo         info_gemm{};
379     TensorInfo         weights_reshaped_info{};
380     const ITensorInfo *gemm_input_to_use  = src;
381     const ITensorInfo *gemm_output_to_use = dst;
382     const ITensorInfo *weights_to_use     = weights;
383     const bool         is_quantized       = is_data_type_quantized_asymmetric(data_type);
384     const bool         skip_im2col        = (data_layout == DataLayout::NHWC && kernel_width == 1 && kernel_height == 1 && conv2d_info.conv_info.stride().first == 1
385                                              && conv2d_info.conv_info.stride().second == 1);
386     const bool skip_col2im     = data_layout == DataLayout::NHWC;
387     bool       fuse_activation = true;
388     bool       use_post_ops    = conv2d_info.post_ops.size() > 0;
389 
390     ARM_COMPUTE_RETURN_ERROR_ON((weights->dimension(idx_channel) * conv2d_info.num_groups) != src->dimension(idx_channel));
391     ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 4);
392     ARM_COMPUTE_RETURN_ERROR_ON_MSG(!skip_im2col
393                                     && conv2d_info.post_ops.size() > 0,
394                                     "ClGemmConv2d does not support post ops with col2im or im2col operation"); // Post ops must be performed after every other op
395 
396     // Validate biases
397     if(biases != nullptr)
398     {
399         if(is_quantized)
400         {
401             ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::S32);
402         }
403         else
404         {
405             ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, biases);
406         }
407         ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(idx_kernels));
408         ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1);
409     }
410 
411     if(conv2d_info.act_info.enabled())
412     {
413         ARM_COMPUTE_ERROR_ON(conv2d_info.act_info.b() > conv2d_info.act_info.a());
414     }
415 
416     // Get convolved dimensions
417     unsigned int conv_w = 0;
418     unsigned int conv_h = 0;
419 
420     std::tie(conv_w, conv_h) = scaled_dimensions(src->dimension(idx_width),
421                                                  src->dimension(idx_height),
422                                                  kernel_width,
423                                                  kernel_height,
424                                                  conv2d_info.conv_info,
425                                                  conv2d_info.dilation);
426 
427     unsigned int mat_weights_cols = num_kernels / conv2d_info.num_groups;
428 
429     const ITensorInfo *biases_to_use = biases;
430     bool               append_bias   = false;
431 
432     if(conv2d_info.num_groups != 1 && biases != nullptr)
433     {
434         // num_groups != 1 can only be for NCHW
435         // Since it is missing an utility function to reshape the biases, we append the biases into the weights tensor
436         biases_to_use         = nullptr;
437         append_bias           = true;
438         weights_reshaped_info = TensorInfo(compute_weights_reshaped_shape(*weights, true, conv2d_info.num_groups), 1, data_type);
439     }
440     else
441     {
442         weights_reshaped_info = TensorInfo(compute_weights_reshaped_shape(*weights, false, conv2d_info.num_groups), 1, data_type);
443     }
444 
445     weights_to_use = &weights_reshaped_info;
446 
447     if(!skip_im2col)
448     {
449         const Size2D kernel_dims(kernel_width, kernel_height);
450 
451         // Output tensor auto initialization if not yet initialized
452         TensorShape expected_output_shape = compute_im2col_conv_shape(src, kernel_dims, conv2d_info.conv_info, append_bias, conv2d_info.dilation, conv2d_info.num_groups == 1, conv2d_info.num_groups);
453 
454         auto_init_if_empty(im2col_reshaped_info, src->clone()->set_tensor_shape(expected_output_shape));
455 
456         ARM_COMPUTE_RETURN_ON_ERROR(opencl::kernels::ClIm2ColKernel::validate(src, &im2col_reshaped_info, kernel_dims, conv2d_info.conv_info, append_bias, conv2d_info.dilation, conv2d_info.num_groups));
457         gemm_input_to_use = &im2col_reshaped_info;
458     }
459 
460     // Create GEMM output tensor
461     if(!skip_col2im)
462     {
463         TensorShape shape_gemm;
464 
465         shape_gemm = gemm_input_to_use->tensor_shape();
466         shape_gemm.set(0, mat_weights_cols);
467         shape_gemm.set(1, conv_w * conv_h);
468 
469         info_gemm = TensorInfo(shape_gemm, 1, data_type);
470         info_gemm.set_quantization_info(dst->quantization_info()).set_data_layout(src->data_layout());
471         gemm_output_to_use = &info_gemm;
472     }
473 
474     GEMMLowpOutputStageInfo gemmlowp_output_stage;
475     gemmlowp_output_stage.type                     = GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT;
476     gemmlowp_output_stage.gemmlowp_offset          = 0;
477     gemmlowp_output_stage.is_quantized_per_channel = is_quantized_per_channel;
478 
479     if(is_quantized)
480     {
481         const UniformQuantizationInfo iq_info           = src->quantization_info().uniform();
482         const UniformQuantizationInfo oq_info           = dst->quantization_info().uniform();
483         const auto                    output_quant_info = (dst->total_size() == 0) ? iq_info : oq_info;
484         const unsigned int            num_filters       = (is_quantized_per_channel) ? num_kernels : 1;
485 
486         gemmlowp_output_stage.gemmlowp_multipliers.resize(num_filters);
487         gemmlowp_output_stage.gemmlowp_shifts.resize(num_filters);
488         quantization::compute_quantized_multipliers_and_shifts(src, weights, dst,
489                                                                gemmlowp_output_stage.gemmlowp_multipliers.data(),
490                                                                gemmlowp_output_stage.gemmlowp_shifts.data());
491         gemmlowp_output_stage.gemmlowp_multiplier = gemmlowp_output_stage.gemmlowp_multipliers[0];
492         gemmlowp_output_stage.gemmlowp_shift      = gemmlowp_output_stage.gemmlowp_shifts[0];
493 
494         int min_activation = 0;
495         int max_activation = 0;
496 
497         const std::set<ActivationLayerInfo::ActivationFunction> supported_acts = { ActivationLayerInfo::ActivationFunction::RELU,
498                                                                                    ActivationLayerInfo::ActivationFunction::BOUNDED_RELU,
499                                                                                    ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU
500                                                                                  };
501 
502         if(conv2d_info.act_info.enabled())
503         {
504             if(supported_acts.count(conv2d_info.act_info.activation()) != 0)
505             {
506                 std::tie(min_activation, max_activation) = get_quantized_activation_min_max(conv2d_info.act_info, data_type, output_quant_info);
507             }
508             else
509             {
510                 fuse_activation = false;
511             }
512         }
513 
514         // Set the GEMMLowp output stage info
515         gemmlowp_output_stage.gemmlowp_offset    = output_quant_info.offset;
516         gemmlowp_output_stage.gemmlowp_min_bound = min_activation;
517         gemmlowp_output_stage.gemmlowp_max_bound = max_activation;
518     }
519 
520     // In case of NHWC, we need to run GEMM3D (gemm_3d_depth != 0) in order to avoid reshaping the output matrix
521     const unsigned int gemm_3d_depth = (data_layout == DataLayout::NHWC) ? conv_h : 0;
522 
523     ARM_COMPUTE_RETURN_ON_ERROR(validate_mm(gemm_input_to_use, weights_to_use, biases_to_use, gemm_output_to_use, gemmlowp_output_stage, gemm_3d_depth, skip_im2col, conv2d_info.act_info,
524                                             conv2d_info.post_ops));
525 
526     // Validate Col2Im
527     if(!skip_col2im)
528     {
529         ARM_COMPUTE_RETURN_ON_ERROR(kernels::ClCol2ImKernel::validate(gemm_output_to_use, dst, Size2D(conv_w, conv_h), conv2d_info.num_groups));
530     }
531 
532     // Validate Activation Layer
533     // Disable running (thus validation) of activation kernel if post ops are used
534     if(!fuse_activation && !use_post_ops)
535     {
536         ARM_COMPUTE_RETURN_ON_ERROR(kernels::ClActivationKernel::validate(dst, nullptr, conv2d_info.act_info));
537     }
538 
539     return Status{};
540 }
541 
run(ITensorPack & tensors)542 void ClGemmConv2d::run(ITensorPack &tensors)
543 {
544     prepare(tensors);
545 
546     auto src                = tensors.get_const_tensor(ACL_SRC_0);
547     auto biases             = tensors.get_const_tensor(ACL_SRC_2);
548     auto dst                = tensors.get_tensor(ACL_DST);
549     auto gemm_input_to_use  = src;
550     auto gemm_output_to_use = dst;
551 
552     CLAuxTensorHandler im2col_output(offset_int_vec(Im2ColOutput), _im2col_output, tensors, false);
553     CLAuxTensorHandler gemm_output(offset_int_vec(GemmOutput), _gemm_output, tensors, false);
554     CLAuxTensorHandler weights_reshaped(offset_int_vec(WeightsReshaped), _weights_reshaped, tensors, false);
555 
556     // Run im2col
557     if(!_skip_im2col)
558     {
559         ITensorPack pack =
560         {
561             { TensorType::ACL_SRC, src },
562             { TensorType::ACL_DST, im2col_output.get() }
563         };
564         CLScheduler::get().enqueue_op(*_im2col_kernel, pack, false);
565         gemm_input_to_use = im2col_output.get();
566     }
567     if(!_skip_col2im)
568     {
569         gemm_output_to_use = gemm_output.get();
570     }
571     ITensorPack pack_mm = tensors;
572     pack_mm.add_const_tensor(TensorType::ACL_SRC_0, gemm_input_to_use);
573     pack_mm.add_const_tensor(TensorType::ACL_SRC_1, weights_reshaped.get());
574     if(!_append_bias)
575     {
576         pack_mm.add_const_tensor(TensorType::ACL_SRC_2, biases);
577     }
578     pack_mm.add_tensor(TensorType::ACL_DST, gemm_output_to_use);
579     // Runs ClGemm or ClGemmLowpMatrixMultiplyCore functions
580     if(_is_quantized)
581     {
582         // Run gemmlowp
583         _mm_gemmlowp->run(pack_mm);
584     }
585     else
586     {
587         // Run gemm
588         _mm_gemm->run(pack_mm);
589     }
590 
591     // Reshape output matrix
592     if(!_skip_col2im)
593     {
594         ITensorPack pack =
595         {
596             { TensorType::ACL_SRC, gemm_output_to_use },
597             { TensorType::ACL_DST, dst }
598         };
599         CLScheduler::get().enqueue_op(*_col2im_kernel.get(), pack, false);
600     }
601 
602     //Run Activation Layer if we cannot fuse in GEMM
603     // Disable running of activation kernel if post ops are used
604     if(!_fuse_activation && !_use_post_ops)
605     {
606         ITensorPack pack =
607         {
608             { TensorType::ACL_SRC, dst },
609             { TensorType::ACL_DST, dst }
610         };
611         CLScheduler::get().enqueue_op(*_activation_kernel.get(), pack, false);
612     }
613 }
614 
prepare(ITensorPack & tensors)615 void ClGemmConv2d::prepare(ITensorPack &tensors)
616 {
617     if(!_is_prepared)
618     {
619         // Run weights reshaping and mark original weights tensor as unused
620         ICLTensor         *weights_reshaped_p = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(offset_int_vec(WeightsReshaped)));
621         CLAuxTensorHandler weights_reshaped(_weights_reshaped, *weights_reshaped_p);
622         auto               weights = tensors.get_const_tensor(TensorType::ACL_SRC_1);
623         ITensorPack        pack =
624         {
625             { TensorType::ACL_SRC, weights },
626             { TensorType::ACL_DST, weights_reshaped.get() }
627         };
628 
629         if(_append_bias)
630         {
631             const auto biases = tensors.get_const_tensor(TensorType::ACL_SRC_2);
632             pack.add_const_tensor(TensorType::ACL_BIAS, biases);
633         }
634         CLScheduler::get().enqueue_op(*_weights_reshape_kernel.get(), pack, true);
635         tensors.add_const_tensor(TensorType::ACL_SRC_1, weights_reshaped.get());
636 
637         // Prepare GEMM
638         _is_quantized ? _mm_gemmlowp->prepare(tensors) : _mm_gemm->prepare(tensors);
639         _is_prepared = true;
640     }
641 }
workspace() const642 experimental::MemoryRequirements ClGemmConv2d::workspace() const
643 {
644     return _aux_mem;
645 }
646 } // namespace opencl
647 } // namespace arm_compute
648