This patch adds acl cl and neon TransposeConv kernel test sets.
It also tests CLDeconvolutionLayerEx.
Signed-off-by: Jiyoung Yun <jy910.yun@samsung.com>
#include "Operation.h"
#include "operations/Convolution.h"
+#include "operations/TransposeConv.h"
namespace kbenchmark
{
// Config Name Operation Name
OP("CONV_2D", Convolution)
+OP("TRANSPOSE_CONV", TransposeConv)
--- /dev/null
+/*
+ * Copyright (c) 2019 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef __KBENCHMARK_KERNELS_ACL_COMMON_UTILS_H__
+#define __KBENCHMARK_KERNELS_ACL_COMMON_UTILS_H__
+
+#include <algorithm>
+
+using namespace arm_compute;
+
+namespace kbenchmark
+{
+namespace kernels
+{
+namespace acl_common
+{
+
+struct PaddingInfo
+{
+ uint32_t top;
+ uint32_t bottom;
+ uint32_t left;
+ uint32_t right;
+};
+
+PaddingInfo calculatePadding(const std::string &padding_name, const uint32_t ifm_H,
+ const uint32_t ifm_W, const uint32_t ofm_H, const uint32_t ofm_W,
+ const uint32_t vertical_stride, const uint32_t horizontal_stride,
+ const uint32_t ker_H, const uint32_t ker_W)
+{
+ uint32_t top;
+ uint32_t bottom;
+ uint32_t left;
+ uint32_t right;
+
+ if (padding_name == "VALID")
+ {
+ top = bottom = left = right = 0;
+ }
+ else if (padding_name == "SAME")
+ {
+ const int32_t vertical_needed_input = (ofm_H - 1) * vertical_stride + ker_H;
+ const int32_t vertical_total_padding = std::max(0, vertical_needed_input - (int32_t)ifm_H);
+
+ const int32_t horizontal_needed_input = (ofm_W - 1) * horizontal_stride + ker_W;
+ const int32_t horizontal_total_padding = std::max(0, horizontal_needed_input - (int32_t)ifm_W);
+
+ top = vertical_total_padding / 2;
+ bottom = (vertical_total_padding + 1) / 2;
+ left = horizontal_total_padding / 2;
+ right = (horizontal_total_padding + 1) / 2;
+ }
+
+ return PaddingInfo{top, bottom, left, right};
+}
+
+PadStrideInfo asPadStrideInfo(const PaddingInfo &padding, uint32_t vertical_stride,
+ uint32_t horizontal_stride)
+{
+ return PadStrideInfo{horizontal_stride,
+ vertical_stride,
+ padding.left,
+ padding.right,
+ padding.top,
+ padding.bottom,
+ DimensionRoundingType::FLOOR};
+}
+
+ActivationLayerInfo asActivationLayerInfo(const std::string &act_name)
+{
+ if (act_name == "NONE")
+ {
+ return ActivationLayerInfo{};
+ }
+ else if (act_name == "RELU")
+ {
+ return ActivationLayerInfo{ActivationLayerInfo::ActivationFunction::RELU};
+ }
+ else
+ {
+ throw std::runtime_error{"Not support activation layer info"};
+ }
+}
+
+} // namespace acl_common
+} // namespace kernels
+} // namespace kbenchmark
+
+#endif // __KBENCHMARK_KERNELS_ACL_COMMON_UTILS_H__
cmake_parse_arguments(ARG "" "NAME" "SOURCES" ${ARGN})
add_library(${ARG_NAME} SHARED ${ARG_SOURCES})
+ target_include_directories(${ARG_NAME} PUBLIC ../acl_common)
target_link_libraries(${ARG_NAME} nonius)
target_link_libraries(${ARG_NAME} arm_compute)
+ target_link_libraries(${ARG_NAME} arm_compute_ex)
target_link_libraries(${ARG_NAME} pthread)
install(TARGETS ${ARG_NAME} DESTINATION lib/kben)
endfunction(add_kben_armcl_library)
-add_kben_armcl_library(NAME kben_armcl_conv SOURCES Convolution.cpp)
\ No newline at end of file
+add_kben_armcl_library(NAME kben_armcl_conv SOURCES Convolution.cpp)
+add_kben_armcl_library(NAME kben_armcl_transpose_conv SOURCES TransposeConv.cpp)
--- /dev/null
+/*
+ * Copyright (c) 2019 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+/**
+ * @file TransposeConv benchmark with various algorithms
+ */
+
+#include <nonius/nonius.h++>
+
+#include <arm_compute/core/Types.h>
+#include <arm_compute/runtime/CL/CLScheduler.h>
+#include <arm_compute/runtime/CL/CLFunctions.h>
+#include <arm_compute/runtime/CL/CLFunctionsEx.h>
+
+#include <cstdint>
+#include <cassert>
+#include <stdexcept>
+
+#include "Utils.h"
+
+using namespace arm_compute;
+using namespace kbenchmark::kernels::acl_common;
+
+//
+// Helpers
+//
+namespace
+{
+
+enum Layout
+{
+ NCHW,
+ NHWC
+};
+
+struct Initializer
+{
+ Initializer() { CLScheduler::get().default_init(); }
+};
+
+Initializer initializer;
+
+TensorInfo make_info(uint32_t N)
+{
+ TensorShape shape{N};
+ return TensorInfo{shape, 1, DataType::F32};
+}
+
+template <enum Layout> TensorInfo make_info(uint32_t N, uint32_t C, uint32_t H, uint32_t W);
+
+template <> TensorInfo make_info<NCHW>(uint32_t N, uint32_t C, uint32_t H, uint32_t W)
+{
+ TensorShape shape{W, H, C, N};
+ TensorInfo info{shape, 1, DataType::F32};
+ info.set_data_layout(DataLayout::NCHW);
+ return info;
+}
+
+template <> TensorInfo make_info<NHWC>(uint32_t N, uint32_t C, uint32_t H, uint32_t W)
+{
+ TensorShape shape{C, W, H, N};
+ TensorInfo info{shape, 1, DataType::F32};
+ info.set_data_layout(DataLayout::NHWC);
+ return info;
+}
+
+inline void check(const Status &status)
+{
+ if (!status)
+ {
+ std::cerr << status.error_description() << std::endl;
+ throw std::runtime_error{"ERROR"};
+ }
+}
+
+inline bool is_odd(uint32_t n) { return (n % 2 != 0) ? true : false; }
+
+} // namespace
+
+//
+// Benchmark Parameters
+//
+NONIUS_PARAM(BATCH, 1);
+
+NONIUS_PARAM(IFM_C, 3);
+NONIUS_PARAM(IFM_H, 244);
+NONIUS_PARAM(IFM_W, 244);
+
+NONIUS_PARAM(OFM_C, 3);
+NONIUS_PARAM(OFM_H, 244);
+NONIUS_PARAM(OFM_W, 244);
+
+NONIUS_PARAM(KER_H, 3);
+NONIUS_PARAM(KER_W, 3);
+
+NONIUS_PARAM(STRIDE_H, 1);
+NONIUS_PARAM(STRIDE_W, 1);
+
+NONIUS_PARAM(PADDING, std::string{"SAME"})
+
+//
+// Configuration Helpers
+//
+namespace
+{
+
+struct Configuration
+{
+ uint32_t ifm_N;
+ uint32_t ifm_C;
+ uint32_t ifm_H;
+ uint32_t ifm_W;
+
+ uint32_t ofm_N;
+ uint32_t ofm_C;
+ uint32_t ofm_H;
+ uint32_t ofm_W;
+
+ uint32_t ker_N;
+ uint32_t ker_C;
+ uint32_t ker_H;
+ uint32_t ker_W;
+
+ uint32_t vertical_stride;
+ uint32_t horizontal_stride;
+
+ PadStrideInfo deconv_info;
+
+ uint32_t inner_border_right;
+ uint32_t inner_border_top;
+
+ Configuration(nonius::chronometer meter)
+ {
+ ifm_N = meter.param<BATCH>();
+ ifm_C = meter.param<IFM_C>();
+ ifm_H = meter.param<IFM_H>();
+ ifm_W = meter.param<IFM_W>();
+
+ ofm_N = meter.param<BATCH>();
+ ofm_C = meter.param<OFM_C>();
+ ofm_H = meter.param<OFM_H>();
+ ofm_W = meter.param<OFM_W>();
+
+ ker_N = meter.param<OFM_C>();
+ ker_C = meter.param<IFM_C>();
+ ker_H = meter.param<KER_H>();
+ ker_W = meter.param<KER_W>();
+
+ vertical_stride = meter.param<STRIDE_H>();
+ horizontal_stride = meter.param<STRIDE_W>();
+
+ // NOTE The padding calculation formula of TransposeConv is opposite to Conv.
+ // So the location of ifm and ofm is changed.
+ auto padding_info = calculatePadding(meter.param<PADDING>(), ofm_H, ofm_W, ifm_H, ifm_W,
+ vertical_stride, horizontal_stride, ker_H, ker_W);
+
+ inner_border_right = padding_info.right - padding_info.left;
+ inner_border_top = padding_info.bottom - padding_info.top;
+
+ padding_info.left = padding_info.right;
+ padding_info.top = padding_info.bottom;
+
+ deconv_info = asPadStrideInfo(padding_info, vertical_stride, horizontal_stride);
+ }
+
+ template <Layout L> TensorInfo src_info() const
+ {
+ return make_info<L>(ifm_N, ifm_C, ifm_H, ifm_W);
+ }
+ template <Layout L> TensorInfo dst_info() const
+ {
+ return make_info<L>(ofm_N, ofm_C, ofm_H, ofm_W);
+ }
+ template <Layout L> TensorInfo ker_info() const
+ {
+ return make_info<L>(ker_N, ker_C, ker_H, ker_W);
+ }
+ TensorInfo bias_info(void) const { return make_info(ker_N); }
+};
+
+} // namespace
+
+//
+// Benchmark Implementations
+//
+namespace
+{
+
+inline nonius::benchmark_registry &local_benchmark_registry()
+{
+ static nonius::benchmark_registry registry;
+ return registry;
+}
+
+} // namespace
+
+#define NONIUS_LOCAL_BENCHMARK(name, ...) \
+ namespace \
+ { \
+ static ::nonius::benchmark_registrar \
+ NONIUS_DETAIL_UNIQUE_NAME(benchmark_registrar)(local_benchmark_registry(), name, \
+ __VA_ARGS__); \
+ }
+
+NONIUS_LOCAL_BENCHMARK("CLDeconvolutionLayer_NCHW", [](nonius::chronometer meter) {
+ CLDeconvolutionLayer deconv;
+
+ // Configure
+ Configuration p{meter};
+
+ CLTensor src_tensor{};
+ CLTensor dst_tensor{};
+ CLTensor ker_tensor{};
+
+ src_tensor.allocator()->init(p.src_info<NCHW>());
+ dst_tensor.allocator()->init(p.dst_info<NCHW>());
+ ker_tensor.allocator()->init(p.ker_info<NCHW>());
+
+ try
+ {
+ check(deconv.validate(src_tensor.info(), ker_tensor.info(), nullptr, dst_tensor.info(),
+ p.deconv_info, p.inner_border_right, p.inner_border_top));
+ }
+ catch (...)
+ {
+ meter.measure([&](int) {
+ // DO NOTHING
+ volatile int x = 0;
+ return x;
+ });
+ return;
+ }
+
+ deconv.configure(&src_tensor, &ker_tensor, nullptr, &dst_tensor, p.deconv_info,
+ p.inner_border_right, p.inner_border_top);
+
+ src_tensor.allocator()->allocate();
+ ker_tensor.allocator()->allocate();
+ dst_tensor.allocator()->allocate();
+
+ // Run!
+ meter.measure([&](int) {
+ deconv.run();
+ CLScheduler::get().sync();
+ });
+})
+
+NONIUS_LOCAL_BENCHMARK("CLDeconvolutionLayer_NHWC", [](nonius::chronometer meter) {
+ CLDeconvolutionLayer deconv;
+
+ // Configure
+ Configuration p{meter};
+
+ CLTensor src_tensor{};
+ CLTensor dst_tensor{};
+ CLTensor ker_tensor{};
+
+ src_tensor.allocator()->init(p.src_info<NHWC>());
+ dst_tensor.allocator()->init(p.dst_info<NHWC>());
+ ker_tensor.allocator()->init(p.ker_info<NHWC>());
+
+ try
+ {
+ check(deconv.validate(src_tensor.info(), ker_tensor.info(), nullptr, dst_tensor.info(),
+ p.deconv_info, p.inner_border_right, p.inner_border_top));
+ }
+ catch (...)
+ {
+ meter.measure([&](int) {
+ // DO NOTHING
+ volatile int x = 0;
+ return x;
+ });
+ return;
+ }
+
+ deconv.configure(&src_tensor, &ker_tensor, nullptr, &dst_tensor, p.deconv_info,
+ p.inner_border_right, p.inner_border_top);
+
+ src_tensor.allocator()->allocate();
+ ker_tensor.allocator()->allocate();
+ dst_tensor.allocator()->allocate();
+
+ // Run!
+ meter.measure([&](int) {
+ deconv.run();
+ CLScheduler::get().sync();
+ });
+})
+
+NONIUS_LOCAL_BENCHMARK("CLDeconvolutionLayerEx_NCHW", [](nonius::chronometer meter) {
+ CLDeconvolutionLayerEx deconv;
+
+ // Configure
+ Configuration p{meter};
+
+ CLTensor src_tensor{};
+ CLTensor dst_tensor{};
+ CLTensor ker_tensor{};
+
+ src_tensor.allocator()->init(p.src_info<NCHW>());
+ dst_tensor.allocator()->init(p.dst_info<NCHW>());
+ ker_tensor.allocator()->init(p.ker_info<NCHW>());
+
+ try
+ {
+ check(deconv.validate(src_tensor.info(), ker_tensor.info(), nullptr, dst_tensor.info(),
+ p.deconv_info, p.inner_border_right, p.inner_border_top));
+ }
+ catch (...)
+ {
+ meter.measure([&](int) {
+ // DO NOTHING
+ volatile int x = 0;
+ return x;
+ });
+ return;
+ }
+
+ deconv.configure(&src_tensor, &ker_tensor, nullptr, &dst_tensor, p.deconv_info,
+ p.inner_border_right, p.inner_border_top);
+
+ src_tensor.allocator()->allocate();
+ ker_tensor.allocator()->allocate();
+ dst_tensor.allocator()->allocate();
+
+ // Run!
+ meter.measure([&](int) {
+ deconv.run();
+ CLScheduler::get().sync();
+ });
+})
+
+NONIUS_LOCAL_BENCHMARK("CLDeconvolutionLayerEx_NHWC", [](nonius::chronometer meter) {
+ CLDeconvolutionLayerEx deconv;
+
+ // Configure
+ Configuration p{meter};
+
+ CLTensor src_tensor{};
+ CLTensor dst_tensor{};
+ CLTensor ker_tensor{};
+
+ src_tensor.allocator()->init(p.src_info<NHWC>());
+ dst_tensor.allocator()->init(p.dst_info<NHWC>());
+ ker_tensor.allocator()->init(p.ker_info<NHWC>());
+
+ try
+ {
+ check(deconv.validate(src_tensor.info(), ker_tensor.info(), nullptr, dst_tensor.info(),
+ p.deconv_info, p.inner_border_right, p.inner_border_top));
+ }
+ catch (...)
+ {
+ meter.measure([&](int) {
+ // DO NOTHING
+ volatile int x = 0;
+ return x;
+ });
+ return;
+ }
+
+ deconv.configure(&src_tensor, &ker_tensor, nullptr, &dst_tensor, p.deconv_info,
+ p.inner_border_right, p.inner_border_top);
+
+ src_tensor.allocator()->allocate();
+ ker_tensor.allocator()->allocate();
+ dst_tensor.allocator()->allocate();
+
+ // Run!
+ meter.measure([&](int) {
+ deconv.run();
+ CLScheduler::get().sync();
+ });
+})
+
+extern "C" nonius::benchmark_registry &benchmark_functions(void)
+{
+ return local_benchmark_registry();
+}
cmake_parse_arguments(ARG "" "NAME" "SOURCES" ${ARGN})
add_library(${ARG_NAME} SHARED ${ARG_SOURCES})
+ target_include_directories(${ARG_NAME} PUBLIC ../acl_common)
target_link_libraries(${ARG_NAME} nonius)
target_link_libraries(${ARG_NAME} arm_compute)
target_link_libraries(${ARG_NAME} pthread)
install(TARGETS ${ARG_NAME} DESTINATION lib/kben)
endfunction(add_kben_armne_library)
-add_kben_armne_library(NAME kben_armne_conv SOURCES Convolution.cpp)
\ No newline at end of file
+add_kben_armne_library(NAME kben_armne_conv SOURCES Convolution.cpp)
+add_kben_armne_library(NAME kben_armne_transpose_conv SOURCES TransposeConv.cpp)
--- /dev/null
+/*
+ * Copyright (c) 2019 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+/**
+ * @file TransposeConv benchmark with various algorithms
+ */
+
+#include <nonius/nonius.h++>
+
+#include <arm_compute/core/Types.h>
+#include <arm_compute/runtime/NEON/NEScheduler.h>
+#include <arm_compute/runtime/NEON/NEFunctions.h>
+
+#include <cstdint>
+#include <cassert>
+#include <stdexcept>
+
+#include "Utils.h"
+
+using namespace arm_compute;
+using namespace kbenchmark::kernels::acl_common;
+
+//
+// Helpers
+//
+namespace
+{
+
+enum Layout
+{
+ NCHW,
+ NHWC
+};
+
+TensorInfo make_info(uint32_t N)
+{
+ TensorShape shape{N};
+ return TensorInfo{shape, 1, DataType::F32};
+}
+
+template <enum Layout> TensorInfo make_info(uint32_t N, uint32_t C, uint32_t H, uint32_t W);
+
+template <> TensorInfo make_info<NCHW>(uint32_t N, uint32_t C, uint32_t H, uint32_t W)
+{
+ TensorShape shape{W, H, C, N};
+ TensorInfo info{shape, 1, DataType::F32};
+ info.set_data_layout(DataLayout::NCHW);
+ return info;
+}
+
+template <> TensorInfo make_info<NHWC>(uint32_t N, uint32_t C, uint32_t H, uint32_t W)
+{
+ TensorShape shape{C, W, H, N};
+ TensorInfo info{shape, 1, DataType::F32};
+ info.set_data_layout(DataLayout::NHWC);
+ return info;
+}
+
+inline void check(const Status &status)
+{
+ if (!status)
+ {
+ std::cerr << status.error_description() << std::endl;
+ throw std::runtime_error{"ERROR"};
+ }
+}
+
+inline bool is_odd(uint32_t n) { return (n % 2 != 0) ? true : false; }
+
+} // namespace
+
+//
+// Benchmark Parameters
+//
+NONIUS_PARAM(BATCH, 1);
+
+NONIUS_PARAM(IFM_C, 3);
+NONIUS_PARAM(IFM_H, 244);
+NONIUS_PARAM(IFM_W, 244);
+
+NONIUS_PARAM(OFM_C, 3);
+NONIUS_PARAM(OFM_H, 244);
+NONIUS_PARAM(OFM_W, 244);
+
+NONIUS_PARAM(KER_H, 3);
+NONIUS_PARAM(KER_W, 3);
+
+NONIUS_PARAM(STRIDE_H, 1);
+NONIUS_PARAM(STRIDE_W, 1);
+
+NONIUS_PARAM(PADDING, std::string{"SAME"})
+
+//
+// Configuration Helpers
+//
+namespace
+{
+
+struct Configuration
+{
+ uint32_t ifm_N;
+ uint32_t ifm_C;
+ uint32_t ifm_H;
+ uint32_t ifm_W;
+
+ uint32_t ofm_N;
+ uint32_t ofm_C;
+ uint32_t ofm_H;
+ uint32_t ofm_W;
+
+ uint32_t ker_N;
+ uint32_t ker_C;
+ uint32_t ker_H;
+ uint32_t ker_W;
+
+ uint32_t vertical_stride;
+ uint32_t horizontal_stride;
+
+ PadStrideInfo deconv_info;
+
+ uint32_t inner_border_right;
+ uint32_t inner_border_top;
+
+ Configuration(nonius::chronometer meter)
+ {
+ ifm_N = meter.param<BATCH>();
+ ifm_C = meter.param<IFM_C>();
+ ifm_H = meter.param<IFM_H>();
+ ifm_W = meter.param<IFM_W>();
+
+ ofm_N = meter.param<BATCH>();
+ ofm_C = meter.param<OFM_C>();
+ ofm_H = meter.param<OFM_H>();
+ ofm_W = meter.param<OFM_W>();
+
+ ker_N = meter.param<OFM_C>();
+ ker_C = meter.param<IFM_C>();
+ ker_H = meter.param<KER_H>();
+ ker_W = meter.param<KER_W>();
+
+ vertical_stride = meter.param<STRIDE_H>();
+ horizontal_stride = meter.param<STRIDE_W>();
+
+ // NOTE The padding calculation formula of TransposeConv is opposite to Conv.
+ // So the location of ifm and ofm is changed.
+ auto padding_info = calculatePadding(meter.param<PADDING>(), ofm_H, ofm_W, ifm_H, ifm_W,
+ vertical_stride, horizontal_stride, ker_H, ker_W);
+
+ inner_border_right = padding_info.right - padding_info.left;
+ inner_border_top = padding_info.bottom - padding_info.top;
+
+ padding_info.left = padding_info.right;
+ padding_info.top = padding_info.bottom;
+
+ deconv_info = asPadStrideInfo(padding_info, vertical_stride, horizontal_stride);
+ }
+
+ template <Layout L> TensorInfo src_info() const
+ {
+ return make_info<L>(ifm_N, ifm_C, ifm_H, ifm_W);
+ }
+ template <Layout L> TensorInfo dst_info() const
+ {
+ return make_info<L>(ofm_N, ofm_C, ofm_H, ofm_W);
+ }
+ template <Layout L> TensorInfo ker_info() const
+ {
+ return make_info<L>(ker_N, ker_C, ker_H, ker_W);
+ }
+ TensorInfo bias_info(void) const { return make_info(ker_N); }
+};
+
+} // namespace
+
+//
+// Benchmark Implementations
+//
+namespace
+{
+
+inline nonius::benchmark_registry &local_benchmark_registry()
+{
+ static nonius::benchmark_registry registry;
+ return registry;
+}
+
+} // namespace
+
+#define NONIUS_LOCAL_BENCHMARK(name, ...) \
+ namespace \
+ { \
+ static ::nonius::benchmark_registrar \
+ NONIUS_DETAIL_UNIQUE_NAME(benchmark_registrar)(local_benchmark_registry(), name, \
+ __VA_ARGS__); \
+ }
+
+NONIUS_LOCAL_BENCHMARK("NEDeconvolutionLayer_NCHW", [](nonius::chronometer meter) {
+ NEDeconvolutionLayer deconv;
+
+ // Configure
+ Configuration p{meter};
+
+ Tensor src_tensor{};
+ Tensor dst_tensor{};
+ Tensor ker_tensor{};
+
+ src_tensor.allocator()->init(p.src_info<NCHW>());
+ dst_tensor.allocator()->init(p.dst_info<NCHW>());
+ ker_tensor.allocator()->init(p.ker_info<NCHW>());
+
+ try
+ {
+ check(deconv.validate(src_tensor.info(), ker_tensor.info(), nullptr, dst_tensor.info(),
+ p.deconv_info, p.inner_border_right, p.inner_border_top));
+ }
+ catch (...)
+ {
+ meter.measure([&](int) {
+ // DO NOTHING
+ volatile int x = 0;
+ return x;
+ });
+ return;
+ }
+
+ deconv.configure(&src_tensor, &ker_tensor, nullptr, &dst_tensor, p.deconv_info,
+ p.inner_border_right, p.inner_border_top);
+
+ src_tensor.allocator()->allocate();
+ ker_tensor.allocator()->allocate();
+ dst_tensor.allocator()->allocate();
+
+ // Run!
+ meter.measure([&](int) { deconv.run(); });
+})
+
+NONIUS_LOCAL_BENCHMARK("NEDeconvolutionLayer_NHWC", [](nonius::chronometer meter) {
+ NEDeconvolutionLayer deconv;
+
+ // Configure
+ Configuration p{meter};
+
+ Tensor src_tensor{};
+ Tensor dst_tensor{};
+ Tensor ker_tensor{};
+
+ src_tensor.allocator()->init(p.src_info<NHWC>());
+ dst_tensor.allocator()->init(p.dst_info<NHWC>());
+ ker_tensor.allocator()->init(p.ker_info<NHWC>());
+
+ try
+ {
+ check(deconv.validate(src_tensor.info(), ker_tensor.info(), nullptr, dst_tensor.info(),
+ p.deconv_info, p.inner_border_right, p.inner_border_top));
+ }
+ catch (...)
+ {
+ meter.measure([&](int) {
+ // DO NOTHING
+ volatile int x = 0;
+ return x;
+ });
+ return;
+ }
+
+ deconv.configure(&src_tensor, &ker_tensor, nullptr, &dst_tensor, p.deconv_info,
+ p.inner_border_right, p.inner_border_top);
+
+ src_tensor.allocator()->allocate();
+ ker_tensor.allocator()->allocate();
+ dst_tensor.allocator()->allocate();
+
+ // Run!
+ meter.measure([&](int) { deconv.run(); });
+})
+
+extern "C" nonius::benchmark_registry &benchmark_functions(void)
+{
+ return local_benchmark_registry();
+}
--- /dev/null
+/*
+ * Copyright (c) 2019 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef __KBENCHMARK_OPERATIONS_TRANSPOSE_CONV_H__
+#define __KBENCHMARK_OPERATIONS_TRANSPOSE_CONV_H__
+
+#include "Operation.h"
+#include "Utils.h"
+
+namespace kbenchmark
+{
+namespace operation
+{
+
+class TransposeConv final : public Operation
+{
+public:
+ TransposeConv() = default;
+
+ nonius::parameters params(int layer_num, OperationInfo &info) override
+ {
+ nonius::parameters params;
+
+ params.insert({"LAYER", nonius::param{layer_num}});
+
+ params.insert({"BATCH", nonius::param{1}});
+
+ // TODO The output_shape will be used in Dynamic tensor case
+ auto _output_shape = get_key_dims({"input0"}, info);
+
+ auto _weights = get_key_dims({"input1"}, info);
+ params.insert({"KER_H", nonius::param{_weights[1]}});
+ params.insert({"KER_W", nonius::param{_weights[2]}});
+
+ auto _input = get_key_dims({"input2"}, info);
+ params.insert({"IFM_H", nonius::param{_input[1]}});
+ params.insert({"IFM_W", nonius::param{_input[2]}});
+ params.insert({"IFM_C", nonius::param{_input[3]}});
+
+ auto _output = get_key_dims({"output0"}, info);
+ params.insert({"OFM_H", nonius::param{_output[1]}});
+ params.insert({"OFM_W", nonius::param{_output[2]}});
+ params.insert({"OFM_C", nonius::param{_output[3]}});
+
+ auto _stride_h = get_key_int({"stride_h"}, info);
+ auto _stride_w = get_key_int({"stride_w"}, info);
+ params.insert({"STRIDE_H", nonius::param{_stride_h}});
+ params.insert({"STRIDE_W", nonius::param{_stride_w}});
+
+ auto _pad = get_key_string({"padding"}, info);
+ params.insert({"PADDING", nonius::param{_pad}});
+
+ return params;
+ }
+};
+
+} // namespace operation
+} // namespace kbenchmark
+
+#endif // __KBENCHMARK_OPERATIONS_TRANSPOSE_CONV_H__