xref: /aosp_15_r20/external/ComputeLibrary/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp (revision c217d954acce2dbc11938adb493fc0abd69584f3)
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 "arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h"
25 
26 #include "arm_compute/core/CL/CLKernelLibrary.h"
27 #include "arm_compute/core/PixelValue.h"
28 #include "arm_compute/core/Size2D.h"
29 #include "arm_compute/core/Utils.h"
30 #include "arm_compute/core/Validate.h"
31 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
32 #include "arm_compute/core/utils/quantization/AsymmHelpers.h"
33 #include "arm_compute/runtime/CL/CLScheduler.h"
34 #include "src/core/experimental/PostOpUtils.h"
35 #include "src/core/helpers/MemoryHelpers.h"
36 #include "src/gpu/cl/operators/ClGemmConv2d.h"
37 #include "support/Cast.h"
38 
39 #include <cmath>
40 #include <memory>
41 #include <tuple>
42 
43 namespace arm_compute
44 {
45 using namespace arm_compute::misc::shape_calculator;
46 using namespace arm_compute::utils::cast;
47 using namespace arm_compute::experimental;
48 
49 struct CLGEMMConvolutionLayer::Impl
50 {
51     const ITensor                        *weights{ nullptr };
52     std::unique_ptr<opencl::ClGemmConv2d> op{ nullptr };
53     ITensorPack                           run_pack{};
54     ITensorPack                           prep_pack{};
55     MemoryGroup                           memory_group{};
56     IWeightsManager                      *weights_manager{ nullptr };
57     MemoryRequirements                    aux_mem_req{};
58     WorkspaceData<CLTensor>               workspace_tensors{};
59     bool                                  is_prepared{ false };
60 };
61 
CLGEMMConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager,IWeightsManager * weights_manager)62 CLGEMMConvolutionLayer::CLGEMMConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager, IWeightsManager *weights_manager)
63     : _impl(std::make_unique<Impl>())
64 {
65     _impl->memory_group    = MemoryGroup(memory_manager);
66     _impl->weights_manager = weights_manager;
67 }
68 
69 CLGEMMConvolutionLayer::~CLGEMMConvolutionLayer() = default;
70 
configure(const ICLTensor * input,const ICLTensor * weights,const ICLTensor * biases,ICLTensor * output,const PadStrideInfo & conv_info,const WeightsInfo & weights_info,const Size2D & dilation,const ActivationLayerInfo & act_info,unsigned int num_groups,const experimental::PostOpList<ICLTensor * > & post_ops)71 void CLGEMMConvolutionLayer::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info,
72                                        const Size2D &dilation, const ActivationLayerInfo &act_info, unsigned int num_groups, const experimental::PostOpList<ICLTensor *> &post_ops)
73 {
74     configure(CLKernelLibrary::get().get_compile_context(), input, weights, biases, output, conv_info, weights_info, dilation, act_info, num_groups, post_ops);
75 }
76 
configure(const CLCompileContext & compile_context,const ICLTensor * input,const ICLTensor * weights,const ICLTensor * biases,ICLTensor * output,const PadStrideInfo & conv_info,const WeightsInfo & weights_info,const Size2D & dilation,const ActivationLayerInfo & act_info,unsigned int num_groups,const experimental::PostOpList<ICLTensor * > & post_ops)77 void CLGEMMConvolutionLayer::configure(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output,
78                                        const PadStrideInfo &conv_info,
79                                        const WeightsInfo &weights_info, const Size2D &dilation, const ActivationLayerInfo &act_info, unsigned int num_groups, const experimental::PostOpList<ICLTensor *> &post_ops)
80 {
81     ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
82     _impl->weights = weights;
83     _impl->op      = std::make_unique<opencl::ClGemmConv2d>();
84     // Convert post op arguments to ITensorInfo
85     auto transformed_post_ops = experimental::transform_post_op_list_arguments<ICLTensor *, ITensorInfo *>(post_ops, [](auto tensor)
86     {
87         return tensor->info();
88     });
89     const Conv2dInfo conv2d_info = Conv2dInfo(conv_info, dilation, act_info, false, num_groups, transformed_post_ops);
90     _impl->op->configure(compile_context, input->info(), weights->info(), (biases != nullptr ? biases->info() : nullptr), output->info(), conv2d_info, weights_info);
91 
92     _impl->run_pack =
93     {
94         { TensorType::ACL_SRC_0, input },
95         { TensorType::ACL_SRC_1, weights },
96         { TensorType::ACL_SRC_2, biases },
97         { TensorType::ACL_DST, output }
98     };
99     // Add post op tensors
100     size_t post_op_tensor_index = 0;
101     for(const auto &op : post_ops.get_list())
102     {
103         for(auto &tensor : op->arguments())
104         {
105             _impl->run_pack.add_const_tensor(experimental::get_post_op_arg_type(post_op_tensor_index++), *tensor);
106         }
107     }
108     _impl->prep_pack =
109     {
110         { TensorType::ACL_SRC_1, weights },
111         { TensorType::ACL_SRC_2, biases },
112     };
113     _impl->aux_mem_req       = _impl->op->workspace();
114     _impl->workspace_tensors = manage_workspace<CLTensor>(_impl->aux_mem_req, _impl->memory_group, _impl->run_pack, _impl->prep_pack);
115 }
116 
validate(const ITensorInfo * input,const ITensorInfo * weights,const ITensorInfo * biases,const ITensorInfo * output,const PadStrideInfo & conv_info,const WeightsInfo & weights_info,const Size2D & dilation,const ActivationLayerInfo & act_info,unsigned int num_groups,const experimental::PostOpList<ITensorInfo * > & post_ops)117 Status CLGEMMConvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
118                                         const WeightsInfo &weights_info, const Size2D &dilation, const ActivationLayerInfo &act_info, unsigned int num_groups, const experimental::PostOpList<ITensorInfo *> &post_ops)
119 {
120     const Conv2dInfo conv2d_info = Conv2dInfo(conv_info, dilation, act_info, false, num_groups, post_ops);
121     return opencl::ClGemmConv2d::validate(input, weights, biases, output, conv2d_info, weights_info);
122 }
123 
run()124 void CLGEMMConvolutionLayer::run()
125 {
126     prepare();
127     MemoryGroupResourceScope scope_mg(_impl->memory_group);
128     _impl->op->run(_impl->run_pack);
129 }
130 
prepare()131 void CLGEMMConvolutionLayer::prepare()
132 {
133     if(!_impl->is_prepared)
134     {
135         _impl->op->prepare(_impl->prep_pack);
136         auto has_reshape = std::find_if(_impl->aux_mem_req.begin(),
137                                         _impl->aux_mem_req.end(),
138                                         [](const MemoryInfo & m) -> bool { return m.lifetime == MemoryLifetime::Persistent; });
139 
140         if(has_reshape != std::end(_impl->aux_mem_req))
141         {
142             _impl->weights->mark_as_unused();
143         }
144         else
145         {
146             // Pack the B matrix to be used as the underlying GEMM performs no reshapes
147             _impl->run_pack.add_const_tensor(ACL_SRC_1, _impl->weights);
148         }
149         release_temporaries(_impl->aux_mem_req, _impl->workspace_tensors);
150         _impl->is_prepared = true;
151     }
152 }
153 } // namespace arm_compute
154