1 /*
2 * Copyright (c) 2021-2022 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/ClDirectConv2d.h"
25
26 #include "arm_compute/core/KernelDescriptors.h"
27 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
28 #include "arm_compute/runtime/CL/CLScheduler.h"
29 #include "src/core/CL/kernels/CLFillBorderKernel.h"
30 #include "src/core/helpers/AutoConfiguration.h"
31 #include "src/gpu/cl/kernels/ClActivationKernel.h"
32 #include "src/gpu/cl/kernels/ClDirectConv2dKernel.h"
33 #include "src/runtime/heuristics/direct_conv/ClDirectConvDefaultConfigBifrost.h"
34 #include "src/runtime/heuristics/direct_conv/ClDirectConvDefaultConfigValhall.h"
35 #include "src/runtime/heuristics/direct_conv/ClDirectConvKernelConfig.h"
36 #include "src/runtime/heuristics/direct_conv/IClDirectConvKernelConfig.h"
37
38 #include "src/common/utils/Log.h"
39
40 using namespace arm_compute::cl_direct_conv;
41
42 namespace arm_compute
43 {
44 namespace opencl
45 {
46 namespace
47 {
select_activation_src_dst(ITensorPack & tensors)48 ITensorPack select_activation_src_dst(ITensorPack &tensors)
49 {
50 ITensorPack pack;
51 pack.add_tensor(TensorType::ACL_SRC, tensors.get_tensor(TensorType::ACL_DST));
52 pack.add_tensor(TensorType::ACL_DST, tensors.get_tensor(TensorType::ACL_DST));
53 return pack;
54 }
55
config_direct_convolution_nhwc(const ITensorInfo * src,const ITensorInfo * weights,const PadStrideInfo & conv_info)56 DirectConvComputeKernelInfo config_direct_convolution_nhwc(const ITensorInfo *src, const ITensorInfo *weights, const PadStrideInfo &conv_info)
57 {
58 // Get GPU target
59 GPUTarget gpu_target = CLScheduler::get().target();
60
61 std::unique_ptr<IClDirectConvKernelConfig> t = ClDirectConvKernelConfigurationFactory::create(gpu_target);
62
63 return t->configure(src, weights, conv_info);
64 }
65
66 } // namespace
67
configure(const CLCompileContext & compile_context,ITensorInfo * src,ITensorInfo * weights,ITensorInfo * biases,ITensorInfo * dst,const PadStrideInfo & conv_info,const ActivationLayerInfo & act_info)68 void ClDirectConv2d::configure(const CLCompileContext &compile_context, ITensorInfo *src, ITensorInfo *weights, ITensorInfo *biases, ITensorInfo *dst,
69 const PadStrideInfo &conv_info, const ActivationLayerInfo &act_info)
70 {
71 ARM_COMPUTE_ERROR_ON_NULLPTR(src);
72 ARM_COMPUTE_LOG_PARAMS(src, weights, biases, dst, conv_info, act_info);
73
74 // Initialize the direct convolution descriptor
75 const DirectConvComputeKernelInfo desc = config_direct_convolution_nhwc(src, weights, conv_info);
76
77 // Configure direct convolution kernel
78 const ActivationLayerInfo conv2d_act_info = (src->data_layout() == DataLayout::NHWC && is_data_type_float(src->data_type())) ? act_info : ActivationLayerInfo();
79 auto k = std::make_unique<kernels::ClDirectConv2dKernel>();
80 k->set_target(CLScheduler::get().target());
81 k->configure(compile_context, src, weights, biases, dst, conv_info, conv2d_act_info, desc);
82 _direct_conv_kernel = std::move(k);
83
84 // Configure border handler
85 PixelValue zero_value(0.f);
86 if(is_data_type_quantized_asymmetric(src->data_type()))
87 {
88 zero_value = PixelValue(0, src->data_type(), src->quantization_info());
89 }
90 auto b = std::make_unique<CLFillBorderKernel>();
91 b->configure(compile_context, src, _direct_conv_kernel->border_size(), BorderMode::CONSTANT, zero_value);
92 _src_border_handler = std::move(b);
93
94 // Fused activation is currently supported for NHWC and floating point types
95 if(act_info.enabled() && !conv2d_act_info.enabled())
96 {
97 auto a = std::make_unique<kernels::ClActivationKernel>();
98 a->configure(compile_context, dst, dst, act_info);
99 _activation_kernel = std::move(a);
100 }
101
102 // Tune kernels
103 CLScheduler::get().tune_kernel_static(*_direct_conv_kernel);
104 }
105
validate(const ITensorInfo * src,const ITensorInfo * weights,const ITensorInfo * biases,const ITensorInfo * dst,const PadStrideInfo & conv_info,const ActivationLayerInfo & act_info)106 Status ClDirectConv2d::validate(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst,
107 const PadStrideInfo &conv_info, const ActivationLayerInfo &act_info)
108 {
109 // Initialize the direct convolution descriptor
110 const DirectConvComputeKernelInfo desc = config_direct_convolution_nhwc(src, weights, conv_info);
111
112 ARM_COMPUTE_RETURN_ON_ERROR(kernels::ClDirectConv2dKernel::validate(src, weights, biases, dst, conv_info, ActivationLayerInfo(), desc));
113 if(act_info.enabled())
114 {
115 ARM_COMPUTE_RETURN_ON_ERROR(kernels::ClActivationKernel::validate(dst, dst, act_info));
116 }
117 return Status{};
118 }
119
run(ITensorPack & tensors)120 void ClDirectConv2d::run(ITensorPack &tensors)
121 {
122 // Run border handler
123 CLScheduler::get().enqueue_op(*_src_border_handler.get(), tensors, false);
124 // Run direct convolution
125 CLScheduler::get().enqueue_op(*_direct_conv_kernel.get(), tensors, false);
126 // Run activation kernel
127 if(_activation_kernel)
128 {
129 auto act_pack = select_activation_src_dst(tensors);
130 CLScheduler::get().enqueue_op(*_activation_kernel.get(), act_pack, false);
131 }
132 }
133 } // namespace opencl
134 } // namespace arm_compute
135