aboutsummaryrefslogtreecommitdiff
path: root/src/runtime/CL/functions/CLDeconvolutionLayer.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/runtime/CL/functions/CLDeconvolutionLayer.cpp')
-rw-r--r--src/runtime/CL/functions/CLDeconvolutionLayer.cpp129
1 files changed, 109 insertions, 20 deletions
diff --git a/src/runtime/CL/functions/CLDeconvolutionLayer.cpp b/src/runtime/CL/functions/CLDeconvolutionLayer.cpp
index 918848745e..4e0d1501ba 100644
--- a/src/runtime/CL/functions/CLDeconvolutionLayer.cpp
+++ b/src/runtime/CL/functions/CLDeconvolutionLayer.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2017-2021 Arm Limited.
+ * Copyright (c) 2017-2023 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -23,12 +23,18 @@
*/
#include "arm_compute/runtime/CL/functions/CLDeconvolutionLayer.h"
+#include "arm_compute/core/Types.h"
#include "arm_compute/core/Utils.h"
-#include "arm_compute/core/Validate.h"
#include "arm_compute/core/utils/misc/ShapeCalculator.h"
#include "arm_compute/core/utils/quantization/AsymmHelpers.h"
+#include "arm_compute/core/Validate.h"
#include "arm_compute/runtime/CL/CLScheduler.h"
+#include "src/common/utils/Log.h"
+#include "src/core/CL/ICLKernel.h"
+#include "src/gpu/cl/IClOperator.h"
+#include "src/gpu/cl/operators/ClTransposedConvolution.h"
+
#include <cmath>
#include <memory>
#include <tuple>
@@ -36,26 +42,62 @@
using namespace arm_compute;
using namespace arm_compute::misc::shape_calculator;
+struct CLDeconvolutionLayer::Impl
+{
+ const ICLTensor *src{nullptr};
+ const ICLTensor *weights{nullptr};
+ const ICLTensor *biases{nullptr};
+ ICLTensor *dst{nullptr};
+ std::unique_ptr<opencl::IClOperator> op{nullptr};
+};
+
+CLDeconvolutionLayer::~CLDeconvolutionLayer() = default;
+
CLDeconvolutionLayer::CLDeconvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager)
- : _memory_manager(std::move(memory_manager)), _function()
+ : _memory_manager(std::move(memory_manager)), _function(), _impl(std::make_unique<Impl>())
{
}
-void CLDeconvolutionLayer::configure(ICLTensor *input, ICLTensor *weights, const ICLTensor *bias, ICLTensor *output, const PadStrideInfo &deconv_info,
- const WeightsInfo &weights_info)
+void CLDeconvolutionLayer::configure(ICLTensor *input,
+ ICLTensor *weights,
+ const ICLTensor *bias,
+ ICLTensor *output,
+ const PadStrideInfo &deconv_info,
+ const WeightsInfo &weights_info)
{
configure(CLKernelLibrary::get().get_compile_context(), input, weights, bias, output, deconv_info, weights_info);
}
-void CLDeconvolutionLayer::configure(const CLCompileContext &compile_context, ICLTensor *input, ICLTensor *weights, const ICLTensor *bias, ICLTensor *output, const PadStrideInfo &deconv_info,
- const WeightsInfo &weights_info)
+void CLDeconvolutionLayer::configure(const CLCompileContext &compile_context,
+ ICLTensor *input,
+ ICLTensor *weights,
+ const ICLTensor *bias,
+ ICLTensor *output,
+ const PadStrideInfo &deconv_info,
+ const WeightsInfo &weights_info)
{
ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
+ ARM_COMPUTE_LOG_PARAMS(input, weights, bias, output, deconv_info, weights_info);
- switch(CLDeconvolutionLayer::get_deconvolution_method(input->info(), weights->info(), nullptr, output->info(), deconv_info, weights_info))
+ switch (CLDeconvolutionLayer::get_deconvolution_method(input->info(), weights->info(), nullptr, output->info(),
+ deconv_info, weights_info))
{
case DeconvolutionMethod::DIRECT:
{
+ auto op = std::make_unique<opencl::ClTransposedConvolution>();
+ op->configure(compile_context, input->info(), weights->info(), bias != nullptr ? bias->info() : nullptr,
+ output->info(), deconv_info);
+
+ _impl->src = input;
+ _impl->weights = weights;
+ _impl->biases = bias;
+ _impl->dst = output;
+
+ _impl->op = std::move(op);
+ break;
+ }
+ case DeconvolutionMethod::UPSCALE_CONV2D:
+ {
auto f = std::make_unique<CLDirectDeconvolutionLayer>();
f->configure(compile_context, input, weights, bias, output, deconv_info, weights_info);
_function = std::move(f);
@@ -74,16 +116,28 @@ void CLDeconvolutionLayer::configure(const CLCompileContext &compile_context, IC
}
}
-Status CLDeconvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *bias, ITensorInfo *output, const PadStrideInfo &deconv_info,
- const WeightsInfo &weights_info)
+Status CLDeconvolutionLayer::validate(const ITensorInfo *input,
+ const ITensorInfo *weights,
+ const ITensorInfo *bias,
+ ITensorInfo *output,
+ const PadStrideInfo &deconv_info,
+ const WeightsInfo &weights_info)
{
ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output);
- switch(CLDeconvolutionLayer::get_deconvolution_method(input, weights, bias, output, deconv_info, weights_info))
+ switch (CLDeconvolutionLayer::get_deconvolution_method(input, weights, bias, output, deconv_info, weights_info))
{
case DeconvolutionMethod::DIRECT:
{
+ // Validate transposed convolution operator
+ ARM_COMPUTE_RETURN_ON_ERROR(
+ opencl::ClTransposedConvolution::validate(input, weights, bias, output, deconv_info));
+ break;
+ }
+ case DeconvolutionMethod::UPSCALE_CONV2D:
+ {
// Validate direct convolution layer
- ARM_COMPUTE_RETURN_ON_ERROR(CLDirectDeconvolutionLayer::validate(input, weights, bias, output, deconv_info, weights_info));
+ ARM_COMPUTE_RETURN_ON_ERROR(
+ CLDirectDeconvolutionLayer::validate(input, weights, bias, output, deconv_info, weights_info));
break;
}
case DeconvolutionMethod::GEMM:
@@ -100,24 +154,40 @@ Status CLDeconvolutionLayer::validate(const ITensorInfo *input, const ITensorInf
return Status{};
}
-DeconvolutionMethod CLDeconvolutionLayer::get_deconvolution_method(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *bias, ITensorInfo *output, const PadStrideInfo &deconv_info,
- const WeightsInfo &weights_info)
+DeconvolutionMethod CLDeconvolutionLayer::get_deconvolution_method(const ITensorInfo *input,
+ const ITensorInfo *weights,
+ const ITensorInfo *bias,
+ ITensorInfo *output,
+ const PadStrideInfo &deconv_info,
+ const WeightsInfo &weights_info)
{
ARM_COMPUTE_UNUSED(output, bias, weights_info);
- if(is_data_type_quantized_per_channel(weights->data_type()))
+ if (is_data_type_quantized_per_channel(weights->data_type()))
{
- return DeconvolutionMethod::DIRECT;
+ return DeconvolutionMethod::UPSCALE_CONV2D;
}
const DataLayout data_layout = input->data_layout();
const size_t idx_w = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
const size_t idx_h = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
+ const size_t idx_n = get_data_layout_dimension_index(data_layout, DataLayoutDimension::BATCHES);
+ const size_t ofm = weights->tensor_shape()[idx_n];
- if(weights->dimension(idx_w) != deconv_info.stride().first || weights->dimension(idx_h) != deconv_info.stride().second)
+ if (weights->dimension(idx_w) != deconv_info.stride().first ||
+ weights->dimension(idx_h) != deconv_info.stride().second)
{
- return DeconvolutionMethod::DIRECT;
+ // We observe better performance for FP32 types only when ofm <= 16, and for FP16 only when ofm <= 32.
+ if (input->data_layout() == DataLayout::NHWC && !((input->data_type() == DataType::F32) && (ofm > 16)) &&
+ !((input->data_type() == DataType::F16) && (ofm > 32)))
+ {
+ return DeconvolutionMethod::DIRECT;
+ }
+ else
+ {
+ return DeconvolutionMethod::UPSCALE_CONV2D;
+ }
}
return DeconvolutionMethod::GEMM;
@@ -126,10 +196,29 @@ DeconvolutionMethod CLDeconvolutionLayer::get_deconvolution_method(const ITensor
void CLDeconvolutionLayer::run()
{
prepare();
- _function->run();
+
+ if (_impl->op != nullptr)
+ {
+ // Optimized Operator will be used
+ ITensorPack pack;
+
+ pack.add_tensor(TensorType::ACL_SRC_0, _impl->src);
+ pack.add_tensor(TensorType::ACL_SRC_1, _impl->weights);
+ pack.add_tensor(TensorType::ACL_SRC_2, _impl->biases);
+ pack.add_tensor(TensorType::ACL_DST, _impl->dst);
+
+ _impl->op->run(pack);
+ }
+ else
+ {
+ _function->run();
+ }
}
void CLDeconvolutionLayer::prepare()
{
- _function->prepare();
+ if (_impl->op == nullptr)
+ {
+ _function->prepare();
+ }
}
generated by cgit v1.2.1 (git 2.23.0) at 2024-05-04 01:57:14 +0000