aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/NEROIAlignLayerKernel.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/core/NEON/kernels/NEROIAlignLayerKernel.cpp')
-rw-r--r--src/core/NEON/kernels/NEROIAlignLayerKernel.cpp79
1 files changed, 42 insertions, 37 deletions
diff --git a/src/core/NEON/kernels/NEROIAlignLayerKernel.cpp b/src/core/NEON/kernels/NEROIAlignLayerKernel.cpp
index 802aebb526..486cd6d331 100644
--- a/src/core/NEON/kernels/NEROIAlignLayerKernel.cpp
+++ b/src/core/NEON/kernels/NEROIAlignLayerKernel.cpp
@@ -26,11 +26,12 @@
#include "arm_compute/core/Helpers.h"
#include "arm_compute/core/TensorInfo.h"
#include "arm_compute/core/Utils.h"
-#include "arm_compute/core/Window.h"
#include "arm_compute/core/utils/misc/ShapeCalculator.h"
#include "arm_compute/core/utils/misc/Utility.h"
-#include "src/core/CPP/Validate.h"
+#include "arm_compute/core/Window.h"
+
#include "src/core/common/Registrars.h"
+#include "src/core/CPP/Validate.h"
#include "src/core/helpers/AutoConfiguration.h"
#include "src/core/helpers/WindowHelpers.h"
#include "src/cpu/kernels/roialign/list.h"
@@ -49,7 +50,12 @@ struct ROIAlignSelectorData
};
using ROIAlignSelctorPtr = std::add_pointer<bool(const ROIAlignSelectorData &data)>::type;
-using ROIAlignUKernelPtr = std::add_pointer<void(const ITensor *input, ITensor *output, const ITensor *rois, ROIPoolingLayerInfo pool_info, const Window &window, const ThreadInfo &info)>::type;
+using ROIAlignUKernelPtr = std::add_pointer<void(const ITensor *input,
+ ITensor *output,
+ const ITensor *rois,
+ ROIPoolingLayerInfo pool_info,
+ const Window &window,
+ const ThreadInfo &info)>::type;
struct ROIAlignKernel
{
@@ -58,31 +64,18 @@ struct ROIAlignKernel
ROIAlignUKernelPtr ukernel;
};
-static const ROIAlignKernel available_kernels[] =
-{
- {
- "fp32_neon_roialign",
- [](const ROIAlignSelectorData & data) { return data.dt == DataType::F32; },
- REGISTER_FP32_NEON(arm_compute::cpu::neon_fp32_roialign)
- },
+static const ROIAlignKernel available_kernels[] = {
+ {"fp32_neon_roialign", [](const ROIAlignSelectorData &data) { return data.dt == DataType::F32; },
+ REGISTER_FP32_NEON(arm_compute::cpu::neon_fp32_roialign)},
#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
- {
- "fp16_neon_roialign",
- [](const ROIAlignSelectorData & data) { return data.dt == DataType::F16; },
- REGISTER_FP16_NEON(arm_compute::cpu::neon_fp16_roialign)
- },
+ {"fp16_neon_roialign", [](const ROIAlignSelectorData &data) { return data.dt == DataType::F16; },
+ REGISTER_FP16_NEON(arm_compute::cpu::neon_fp16_roialign)},
#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
#if defined(ARM_COMPUTE_ENABLE_NEON)
- {
- "qu8_neon_roialign",
- [](const ROIAlignSelectorData & data) { return data.dt == DataType::QASYMM8; },
- REGISTER_QASYMM8_NEON(arm_compute::cpu::neon_qu8_roialign)
- },
- {
- "qs8_neon_roialign",
- [](const ROIAlignSelectorData & data) { return data.dt == DataType::QASYMM8_SIGNED; },
- REGISTER_QASYMM8_SIGNED_NEON(arm_compute::cpu::neon_qs8_roialign)
- },
+ {"qu8_neon_roialign", [](const ROIAlignSelectorData &data) { return data.dt == DataType::QASYMM8; },
+ REGISTER_QASYMM8_NEON(arm_compute::cpu::neon_qu8_roialign)},
+ {"qs8_neon_roialign", [](const ROIAlignSelectorData &data) { return data.dt == DataType::QASYMM8_SIGNED; },
+ REGISTER_QASYMM8_SIGNED_NEON(arm_compute::cpu::neon_qs8_roialign)},
#endif //defined(ARM_COMPUTE_ENABLE_NEON)
};
@@ -94,9 +87,9 @@ static const ROIAlignKernel available_kernels[] =
*/
const ROIAlignKernel *get_implementation(const ROIAlignSelectorData &data)
{
- for(const auto &uk : available_kernels)
+ for (const auto &uk : available_kernels)
{
- if(uk.is_selected(data))
+ if (uk.is_selected(data))
{
return &uk;
}
@@ -104,24 +97,29 @@ const ROIAlignKernel *get_implementation(const ROIAlignSelectorData &data)
return nullptr;
}
-Status validate_arguments(const ITensorInfo *input, const ITensorInfo *rois, ITensorInfo *output, const ROIPoolingLayerInfo &pool_info)
+Status validate_arguments(const ITensorInfo *input,
+ const ITensorInfo *rois,
+ ITensorInfo *output,
+ const ROIPoolingLayerInfo &pool_info)
{
ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, rois, output);
ARM_COMPUTE_RETURN_ERROR_ON(rois->dimension(0) != 5);
ARM_COMPUTE_RETURN_ERROR_ON(rois->num_dimensions() > 2);
- ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::F32, DataType::F16);
+ ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED,
+ DataType::F32, DataType::F16);
ARM_COMPUTE_RETURN_ERROR_ON_DATA_LAYOUT_NOT_IN(input, DataLayout::NHWC, DataLayout::NCHW);
ARM_COMPUTE_RETURN_ERROR_ON((pool_info.pooled_width() == 0) || (pool_info.pooled_height() == 0));
ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input);
- if(output->total_size() != 0)
+ if (output->total_size() != 0)
{
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input, output);
- ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(compute_roi_align_shape(*input, *rois, pool_info), output->tensor_shape());
+ ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(compute_roi_align_shape(*input, *rois, pool_info),
+ output->tensor_shape());
}
- if(input->data_type() == DataType::QASYMM8 || input->data_type() == DataType::QASYMM8_SIGNED)
+ if (input->data_type() == DataType::QASYMM8 || input->data_type() == DataType::QASYMM8_SIGNED)
{
ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(rois, 1, DataType::QASYMM16);
@@ -143,13 +141,17 @@ NEROIAlignLayerKernel::NEROIAlignLayerKernel()
{
}
-void NEROIAlignLayerKernel::configure(const ITensor *input, const ITensor *rois, ITensor *output, const ROIPoolingLayerInfo &pool_info)
+void NEROIAlignLayerKernel::configure(const ITensor *input,
+ const ITensor *rois,
+ ITensor *output,
+ const ROIPoolingLayerInfo &pool_info)
{
ARM_COMPUTE_ERROR_ON_NULLPTR(input, output, rois);
ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), rois->info(), output->info(), pool_info));
// Output auto inizialitation if not yet initialized
const TensorShape output_shape = compute_roi_align_shape(*input->info(), *rois->info(), pool_info);
- auto_init_if_empty((*output->info()), output_shape, 1, input->info()->data_type(), input->info()->quantization_info());
+ auto_init_if_empty((*output->info()), output_shape, 1, input->info()->data_type(),
+ input->info()->quantization_info());
output->info()->set_data_layout(input->info()->data_layout());
// Configure kernel window
@@ -167,7 +169,10 @@ void NEROIAlignLayerKernel::configure(const ITensor *input, const ITensor *rois,
INEKernel::configure(window);
}
-Status NEROIAlignLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *rois, ITensorInfo *output, const ROIPoolingLayerInfo &pool_info)
+Status NEROIAlignLayerKernel::validate(const ITensorInfo *input,
+ const ITensorInfo *rois,
+ ITensorInfo *output,
+ const ROIPoolingLayerInfo &pool_info)
{
ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, rois, output, pool_info));
return Status{};
@@ -176,9 +181,9 @@ Status NEROIAlignLayerKernel::validate(const ITensorInfo *input, const ITensorIn
void NEROIAlignLayerKernel::run(const Window &window, const ThreadInfo &info)
{
const DataLayout data_layout = _input->info()->data_layout();
- if(data_layout == DataLayout::NCHW || data_layout == DataLayout::NHWC)
+ if (data_layout == DataLayout::NCHW || data_layout == DataLayout::NHWC)
{
- const auto *uk = get_implementation(ROIAlignSelectorData{ _input->info()->data_type() });
+ const auto *uk = get_implementation(ROIAlignSelectorData{_input->info()->data_type()});
ARM_COMPUTE_ERROR_ON(uk == nullptr || uk->ukernel == nullptr);
uk->ukernel(_input, _output, _rois, _pool_info, window, info);
generated by cgit v1.2.1 (git 2.23.0) at 2024-05-09 22:19:25 +0000