xref: /aosp_15_r20/external/ComputeLibrary/src/cpu/operators/CpuPool2d.cpp (revision c217d954acce2dbc11938adb493fc0abd69584f3)
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  */
24 #include "src/cpu/operators/CpuPool2d.h"
25 
26 #include "arm_compute/core/ITensor.h"
27 #include "arm_compute/core/TensorInfo.h"
28 #include "arm_compute/runtime/NEON/NEScheduler.h"
29 #include "src/common/utils/Log.h"
30 #include "src/core/NEON/kernels/NEFillBorderKernel.h"
31 #include "src/cpu/kernels/CpuPool2dKernel.h"
32 #include "src/cpu/kernels/internal/CpuPool2dAssemblyWrapperKernel.h"
33 
34 using namespace arm_compute::experimental;
35 
36 namespace arm_compute
37 {
38 namespace cpu
39 {
CpuPool2d()40 CpuPool2d::CpuPool2d()
41     : _pooling_layer_kernel(),
42       _asm_glue(),
43       _is_global_pooling_layer(false),
44       _data_layout(DataLayout::NCHW),
45       _aux_mem(1)
46 {
47 }
48 
49 CpuPool2d::~CpuPool2d() = default;
50 
configure(ITensorInfo * src,ITensorInfo * dst,const PoolingLayerInfo & pool_info,ITensorInfo * indices)51 void CpuPool2d::configure(ITensorInfo *src, ITensorInfo *dst, const PoolingLayerInfo &pool_info, ITensorInfo *indices)
52 {
53     ARM_COMPUTE_LOG_PARAMS(src, dst, pool_info, indices);
54 
55     // Check if we can run assembly kernels. Currently, indices are not supported by those kernels
56     const bool run_optimised = bool(kernels::CpuPool2dAssemblyWrapperKernel::validate(src, dst, pool_info)) && (indices == nullptr);
57 
58     // Get data layout
59     _data_layout = pool_info.data_layout == DataLayout::UNKNOWN ? src->data_layout() : pool_info.data_layout;
60 
61     // Check if we have Global Pooling Layer
62     const unsigned int idx_width  = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::WIDTH);
63     const unsigned int idx_height = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::HEIGHT);
64     _is_global_pooling_layer      = (src->dimension(idx_width) == pool_info.pool_size.width) && (src->dimension(idx_height) == pool_info.pool_size.height);
65 
66     if(run_optimised)
67     {
68         const CPUInfo     &ci          = NEScheduler::get().cpu_info();
69         const unsigned int num_threads = NEScheduler::get().num_threads();
70 
71         auto pooling_wrapper = std::make_unique<kernels::CpuPool2dAssemblyWrapperKernel>();
72         ARM_COMPUTE_ERROR_ON(pooling_wrapper == nullptr);
73         pooling_wrapper->configure(src, dst, pool_info, ci);
74 
75         // Get kernel's memory requirements
76         constexpr size_t alignment      = 4096;
77         const size_t     workspace_size = pooling_wrapper->get_working_size(num_threads);
78         _aux_mem[0]                     = MemoryInfo(TensorType::ACL_INT_0, MemoryLifetime::Temporary, workspace_size, alignment);
79 
80         _asm_glue = std::move(pooling_wrapper);
81     }
82     else
83     {
84         // Configure pooling kernel
85         auto k = std::make_unique<kernels::CpuPool2dKernel>();
86         k->configure(src, dst, pool_info, indices);
87         _pooling_layer_kernel = std::move(k);
88     }
89 }
90 
validate(const ITensorInfo * src,const ITensorInfo * dst,const PoolingLayerInfo & pool_info,const ITensorInfo * indices)91 Status CpuPool2d::validate(const ITensorInfo *src, const ITensorInfo *dst, const PoolingLayerInfo &pool_info, const ITensorInfo *indices)
92 {
93     const bool run_optimised = bool(kernels::CpuPool2dAssemblyWrapperKernel::validate(src, dst, pool_info)) && (indices == nullptr);
94 
95     if(run_optimised)
96     {
97         return Status{};
98     }
99 
100     return kernels::CpuPool2dKernel::validate(src, dst, pool_info, indices);
101 }
102 
run(ITensorPack & tensors)103 void CpuPool2d::run(ITensorPack &tensors)
104 {
105     ARM_COMPUTE_ERROR_ON_MSG(tensors.empty(), "No tensors provided");
106 
107     if(_asm_glue)
108     {
109         const auto hints = (_is_global_pooling_layer) ? Window::DimX : Window::DimY;
110         NEScheduler::get().schedule_op(_asm_glue.get(), hints, _asm_glue->window(), tensors);
111     }
112     else
113     {
114         switch(_data_layout)
115         {
116             case DataLayout::NCHW:
117                 NEScheduler::get().schedule_op(_pooling_layer_kernel.get(), _is_global_pooling_layer ? Window::DimZ : Window::DimY, _pooling_layer_kernel->window(), tensors);
118                 break;
119             case DataLayout::NHWC:
120                 NEScheduler::get().schedule_op(_pooling_layer_kernel.get(), Window::DimX, _pooling_layer_kernel->window(), tensors);
121                 break;
122             default:
123                 ARM_COMPUTE_ERROR("Data layout not supported");
124         }
125     }
126 }
127 
workspace() const128 experimental::MemoryRequirements CpuPool2d::workspace() const
129 {
130     return _aux_mem;
131 }
132 } // namespace cpu
133 } // namespace arm_compute
134