diff options
| -rw-r--r-- | arm_compute/core/CL/kernels/CLL2NormalizeLayerKernel.h | 12 | ||||
| -rw-r--r-- | arm_compute/core/NEON/kernels/NEL2NormalizeLayerKernel.h | 12 | ||||
| -rw-r--r-- | arm_compute/runtime/CL/functions/CLL2NormalizeLayer.h | 10 | ||||
| -rw-r--r-- | arm_compute/runtime/NEON/functions/NEL2NormalizeLayer.h | 10 | ||||
| -rw-r--r-- | src/core/CL/kernels/CLL2NormalizeLayerKernel.cpp | 30 | ||||
| -rw-r--r-- | src/core/NEON/kernels/NEL2NormalizeLayerKernel.cpp | 26 | ||||
| -rw-r--r-- | src/runtime/CL/functions/CLL2NormalizeLayer.cpp | 17 | ||||
| -rw-r--r-- | src/runtime/NEON/functions/NEL2NormalizeLayer.cpp | 17 | ||||
| -rw-r--r-- | tests/validation/CL/L2NormalizeLayer.cpp | 24 | ||||
| -rw-r--r-- | tests/validation/NEON/L2NormalizeLayer.cpp | 28 | ||||
| -rw-r--r-- | tests/validation/fixtures/L2NormalizeLayerFixture.h | 23 |
11 files changed, 128 insertions, 81 deletions
diff --git a/arm_compute/core/CL/kernels/CLL2NormalizeLayerKernel.h b/arm_compute/core/CL/kernels/CLL2NormalizeLayerKernel.h index 8dd4609250..ec192bed42 100644 --- a/arm_compute/core/CL/kernels/CLL2NormalizeLayerKernel.h +++ b/arm_compute/core/CL/kernels/CLL2NormalizeLayerKernel.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2018 ARM Limited. + * Copyright (c) 2017-2019 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -55,10 +55,10 @@ public: * Sum will have the same number of dimensions as input. * @param[out] output Destination tensor. Data types and data layouts supported: Same as @p input. * Output will have the same number of dimensions as input. - * @param[in] axis Axis along which to reduce. Supported reduction axis : 0, 1, 2 + * @param[in] axis Axis along which to reduce. Negative values wrap around. Maximum supported actual reduction axis : 2 * @param[in] epsilon Lower bound value for the normalization. */ - void configure(const ICLTensor *input, const ICLTensor *sum, ICLTensor *output, unsigned int axis, float epsilon); + void configure(const ICLTensor *input, const ICLTensor *sum, ICLTensor *output, int axis, float epsilon); /** Static function to check if given info will lead to a valid configuration of @ref CLL2NormalizeLayerKernel. * @@ -67,12 +67,12 @@ public: * Sum will have the same number of dimensions as input. * @param[in] output Destination tensor info. Data types and data layouts supported: Same as @p input. * Output will have the same number of dimensions as input. - * @param[in] axis Axis along which to reduce. Supported reduction axis : 0, 1, 2 + * @param[in] axis Axis along which to reduce. Negative values wrap around. Maximum supported actual reduction axis : 2 * @param[in] epsilon Lower bound value for the normalization. * * @return a status */ - static Status validate(const ITensorInfo *input, const ITensorInfo *sum, const ITensorInfo *output, unsigned int axis, float epsilon); + static Status validate(const ITensorInfo *input, const ITensorInfo *sum, const ITensorInfo *output, int axis, float epsilon); // Inherited methods overridden: void run(const Window &window, cl::CommandQueue &queue) override; @@ -81,7 +81,7 @@ private: const ICLTensor *_input; const ICLTensor *_sum; ICLTensor *_output; - unsigned int _axis; + unsigned int _actual_axis; float _epsilon; }; } // namespace arm_compute diff --git a/arm_compute/core/NEON/kernels/NEL2NormalizeLayerKernel.h b/arm_compute/core/NEON/kernels/NEL2NormalizeLayerKernel.h index f893c4ae6b..ab5e040885 100644 --- a/arm_compute/core/NEON/kernels/NEL2NormalizeLayerKernel.h +++ b/arm_compute/core/NEON/kernels/NEL2NormalizeLayerKernel.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2018 ARM Limited. + * Copyright (c) 2017-2019 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -57,10 +57,10 @@ public: * Sum will have the same number of dimensions as input. * @param[out] output Destination tensor. Data types and data layouts supported: same as @p input. * Output will have the same number of dimensions as input. - * @param[in] axis Dimension along which to reduce. Supported reduction axis : 0, 1, 2 + * @param[in] axis Axis along which to reduce. Negative values wrap around. Maximum supported actual reduction axis : 2 * @param[in] epsilon Lower bound value for the normalization. */ - void configure(const ITensor *input, const ITensor *sum, ITensor *output, unsigned int axis, float epsilon); + void configure(const ITensor *input, const ITensor *sum, ITensor *output, int axis, float epsilon); /** Static function to check if given info will lead to a valid configuration of @ref NEL2NormalizeLayerKernel. * @@ -69,12 +69,12 @@ public: * Sum will have the same number of dimensions as input. * @param[in] output Destination tensor info. Data types and data layouts supported: same as @p input. * Output will have the same number of dimensions as input. - * @param[in] axis Dimension along which to reduce. Supported reduction axis : 0, 1, 2 + * @param[in] axis Axis along which to reduce. Negative values wrap around. Maximum supported actual reduction axis : 2 * @param[in] epsilon Lower bound value for the normalization. * * @return a status */ - static Status validate(const ITensorInfo *input, const ITensorInfo *sum, const ITensorInfo *output, unsigned int axis, float epsilon); + static Status validate(const ITensorInfo *input, const ITensorInfo *sum, const ITensorInfo *output, int axis, float epsilon); // Inherited methods overridden: void run(const Window &window, const ThreadInfo &info) override; @@ -83,7 +83,7 @@ private: const ITensor *_input; const ITensor *_sum; ITensor *_output; - unsigned int _axis; + unsigned int _actual_axis; float _epsilon; }; } // namespace arm_compute diff --git a/arm_compute/runtime/CL/functions/CLL2NormalizeLayer.h b/arm_compute/runtime/CL/functions/CLL2NormalizeLayer.h index 2cabaee5de..15dcc58310 100644 --- a/arm_compute/runtime/CL/functions/CLL2NormalizeLayer.h +++ b/arm_compute/runtime/CL/functions/CLL2NormalizeLayer.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2018 ARM Limited. + * Copyright (c) 2017-2019 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -55,21 +55,21 @@ public: * * @param[in] input Source tensor. Data types supported: F16/F32. Data layouts supported: NCHW/NHWC. * @param[out] output Destination tensor. Data types and data layouts supported: Same as @p input. - * @param[in] axis Axis along which to reduce. Supported reduction axis : 0, 1, 2 + * @param[in] axis Axis along which to reduce. Negative values wrap around. Maximum supported actual reduction axis : 2 * @param[in] epsilon (Optional) Lower bound value for the normalization. */ - void configure(ICLTensor *input, ICLTensor *output, unsigned int axis, float epsilon = 1e-12f); + void configure(ICLTensor *input, ICLTensor *output, int axis, float epsilon = 1e-12f); /** Static function to check if given info will lead to a valid configuration of @ref CLL2NormalizeLayer. * * @param[in] input Source tensor info. Data types supported: F16/F32. Data layouts supported: NCHW/NHWC. * @param[in] output Destination tensor info. Data types and data layouts supported: Same as @p input. - * @param[in] axis Axis along which to reduce. Supported reduction axis : 0, 1, 2 + * @param[in] axis Axis along which to reduce. Negative values wrap around. Maximum supported actual reduction axis : 2 * @param[in] epsilon (Optional) Lower bound value for the normalization. * * @return a status */ - static Status validate(const ITensorInfo *input, const ITensorInfo *output, unsigned int axis, float epsilon = 1e-12f); + static Status validate(const ITensorInfo *input, const ITensorInfo *output, int axis, float epsilon = 1e-12f); // Inherited methods overridden: void run() override; diff --git a/arm_compute/runtime/NEON/functions/NEL2NormalizeLayer.h b/arm_compute/runtime/NEON/functions/NEL2NormalizeLayer.h index ba506fa9ab..e778f96e22 100644 --- a/arm_compute/runtime/NEON/functions/NEL2NormalizeLayer.h +++ b/arm_compute/runtime/NEON/functions/NEL2NormalizeLayer.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2018 ARM Limited. + * Copyright (c) 2017-2019 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -52,21 +52,21 @@ public: * * @param[in, out] input Source tensor. Data types supported: F16/F32. (Written to only for border_size != 0) * @param[out] output Destination tensor. Data types and data layouts supported: same as @p input. - * @param[in] axis Dimension along which to reduce. Supported reduction axis : 0, 1, 2 + * @param[in] axis Axis along which to reduce. Negative values wrap around. Maximum supported actual reduction axis : 2 * @param[in] epsilon (Optional) Lower bound value for the normalization. */ - void configure(ITensor *input, ITensor *output, unsigned int axis, float epsilon = 1e-12f); + void configure(ITensor *input, ITensor *output, int axis, float epsilon = 1e-12f); /** Static function to check if given info will lead to a valid configuration of @ref NEL2NormalizeLayer. * * @param[in] input Source tensor info. Data types supported: F16/F32. (Written to only for border_size != 0) * @param[in] output Destination tensor info. Data types and data layouts supported: same as @p input. - * @param[in] axis Dimension along which to reduce. Supported reduction axis : 0, 1, 2 + * @param[in] axis Axis along which to reduce. Negative values wrap around. Maximum supported actual reduction axis : 2 * @param[in] epsilon (Optional) Lower bound value for the normalization. * * @return a status */ - static Status validate(const ITensorInfo *input, const ITensorInfo *output, unsigned int axis, float epsilon = 1e-12f); + static Status validate(const ITensorInfo *input, const ITensorInfo *output, int axis, float epsilon = 1e-12f); // Inherited methods overridden: void run() override; diff --git a/src/core/CL/kernels/CLL2NormalizeLayerKernel.cpp b/src/core/CL/kernels/CLL2NormalizeLayerKernel.cpp index cb2e29449c..00af590104 100644 --- a/src/core/CL/kernels/CLL2NormalizeLayerKernel.cpp +++ b/src/core/CL/kernels/CLL2NormalizeLayerKernel.cpp @@ -35,29 +35,32 @@ #include "support/ToolchainSupport.h" -using namespace arm_compute; - +namespace arm_compute +{ CLL2NormalizeLayerKernel::CLL2NormalizeLayerKernel() - : _input(nullptr), _sum(nullptr), _output(nullptr), _axis(0), _epsilon(1e-12) + : _input(nullptr), _sum(nullptr), _output(nullptr), _actual_axis(0), _epsilon(1e-12) { } namespace { -Status validate_arguments(const ITensorInfo *input, const ITensorInfo *sum, const ITensorInfo *output, unsigned int axis, float epsilon) +constexpr int max_input_tensor_dim = 3; + +Status validate_arguments(const ITensorInfo *input, const ITensorInfo *sum, const ITensorInfo *output, int axis, float epsilon) { ARM_COMPUTE_UNUSED(epsilon); + const uint32_t actual_axis = wrap_around(axis, max_input_tensor_dim); ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, sum, output); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, sum); ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input); ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis > 2, "Axis greater than 2 is not supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis >= TensorShape::num_max_dimensions, "Reduction axis greater than max number of dimensions"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(actual_axis > 2, "Actual axis greater than 2 is not supported"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(actual_axis >= TensorShape::num_max_dimensions, "Actual normalization axis greater than max number of dimensions"); // Reduce shape on axis TensorShape sum_shape = input->tensor_shape(); - sum_shape.set(axis, 1); + sum_shape.set(actual_axis, 1); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(sum->tensor_shape(), sum_shape); if(output->total_size() != 0) @@ -92,7 +95,7 @@ std::tuple<Status, Window> validate_and_configure_window(ITensorInfo *input, ITe } } // namespace -void CLL2NormalizeLayerKernel::configure(const ICLTensor *input, const ICLTensor *sum, ICLTensor *output, unsigned int axis, float epsilon) +void CLL2NormalizeLayerKernel::configure(const ICLTensor *input, const ICLTensor *sum, ICLTensor *output, int axis, float epsilon) { ARM_COMPUTE_ERROR_ON_NULLPTR(input, sum, output); ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), sum->info(), output->info(), axis, epsilon)); @@ -100,7 +103,7 @@ void CLL2NormalizeLayerKernel::configure(const ICLTensor *input, const ICLTensor _input = input; _sum = sum; _output = output; - _axis = axis; + _actual_axis = wrap_around(axis, max_input_tensor_dim); _epsilon = epsilon; const unsigned int num_elems_processed_per_iteration = 16; @@ -113,7 +116,7 @@ void CLL2NormalizeLayerKernel::configure(const ICLTensor *input, const ICLTensor // Create kernel std::string kernel_name; unsigned int idx = 0; - switch(axis) + switch(_actual_axis) { case 0: kernel_name = "x"; @@ -128,7 +131,7 @@ void CLL2NormalizeLayerKernel::configure(const ICLTensor *input, const ICLTensor idx = num_arguments_per_3D_tensor() * 3; break; default: - ARM_COMPUTE_ERROR("Not supported"); + ARM_COMPUTE_ERROR("Axis not supported"); } _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("l2_normalize_" + kernel_name, build_opts)); @@ -149,7 +152,7 @@ void CLL2NormalizeLayerKernel::configure(const ICLTensor *input, const ICLTensor ICLKernel::configure_internal(std::get<1>(win_config)); } -Status CLL2NormalizeLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *sum, const ITensorInfo *output, unsigned int axis, float epsilon) +Status CLL2NormalizeLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *sum, const ITensorInfo *output, int axis, float epsilon) { ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, sum, output, axis, epsilon)); ARM_COMPUTE_RETURN_ON_ERROR(std::get<0>(validate_and_configure_window(input->clone().get(), output->clone().get()))); @@ -164,7 +167,7 @@ void CLL2NormalizeLayerKernel::run(const Window &window, cl::CommandQueue &queue Window window_sum(window); - switch(_axis) + switch(_actual_axis) { case 0: { @@ -218,3 +221,4 @@ void CLL2NormalizeLayerKernel::run(const Window &window, cl::CommandQueue &queue ARM_COMPUTE_ERROR("Not supported"); } } +} // namespace arm_compute
\ No newline at end of file diff --git a/src/core/NEON/kernels/NEL2NormalizeLayerKernel.cpp b/src/core/NEON/kernels/NEL2NormalizeLayerKernel.cpp index efdcc44e0e..9900446218 100644 --- a/src/core/NEON/kernels/NEL2NormalizeLayerKernel.cpp +++ b/src/core/NEON/kernels/NEL2NormalizeLayerKernel.cpp @@ -40,6 +40,8 @@ namespace arm_compute { namespace { +constexpr int max_input_tensor_dim = 3; + template <typename T, int S> void l2_normalize_X(const ITensor *in, const ITensor *sum, ITensor *out, float epsilon, const Window &window) { @@ -141,19 +143,20 @@ void l2_normalize_Z(const ITensor *in, const ITensor *sum, ITensor *out, float e while(window.slide_window_slice_3D(in_slice) && window.slide_window_slice_3D(sum_slice)); } -Status validate_arguments(const ITensorInfo *input, const ITensorInfo *sum, const ITensorInfo *output, unsigned int axis, float epsilon) +Status validate_arguments(const ITensorInfo *input, const ITensorInfo *sum, const ITensorInfo *output, int axis, float epsilon) { ARM_COMPUTE_UNUSED(epsilon); + const uint32_t actual_axis = wrap_around(axis, max_input_tensor_dim); ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, sum, output); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, sum); ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis > 2, "Axis greater than 2 is not supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis >= TensorShape::num_max_dimensions, "Normalization axis greater than max number of dimensions"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(actual_axis > 2, "Actual axis greater than 2 is not supported"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(actual_axis >= TensorShape::num_max_dimensions, "Actual normalization axis greater than max number of dimensions"); // Reduce shape on axis TensorShape sum_shape = input->tensor_shape(); - sum_shape.set(axis, 1); + sum_shape.set(actual_axis, 1); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(sum->tensor_shape(), sum_shape); if(output->total_size() != 0) @@ -167,10 +170,11 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *sum, cons return Status{}; } -std::tuple<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *sum, ITensorInfo *output, unsigned int axis) +std::tuple<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *sum, ITensorInfo *output, int axis) { + const uint32_t actual_axis = wrap_around(axis, max_input_tensor_dim); const unsigned int num_elems_processed_per_iteration = 16 / data_size_from_type(input->data_type()); - const unsigned int num_elems_processed_per_iteration_sum = (axis == 0) ? 1 : num_elems_processed_per_iteration; + const unsigned int num_elems_processed_per_iteration_sum = (actual_axis == 0) ? 1 : num_elems_processed_per_iteration; Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration)); @@ -191,11 +195,11 @@ std::tuple<Status, Window> validate_and_configure_window(ITensorInfo *input, ITe } // namespace NEL2NormalizeLayerKernel::NEL2NormalizeLayerKernel() - : _input(nullptr), _sum(nullptr), _output(nullptr), _axis(0), _epsilon(1e-12) + : _input(nullptr), _sum(nullptr), _output(nullptr), _actual_axis(0), _epsilon(1e-12) { } -void NEL2NormalizeLayerKernel::configure(const ITensor *input, const ITensor *sum, ITensor *output, unsigned int axis, float epsilon) +void NEL2NormalizeLayerKernel::configure(const ITensor *input, const ITensor *sum, ITensor *output, int axis, float epsilon) { ARM_COMPUTE_ERROR_ON_NULLPTR(input, sum, output); ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), sum->info(), output->info(), axis, epsilon)); @@ -203,7 +207,7 @@ void NEL2NormalizeLayerKernel::configure(const ITensor *input, const ITensor *su _input = input; _sum = sum; _output = output; - _axis = axis; + _actual_axis = wrap_around(axis, max_input_tensor_dim); _epsilon = epsilon; // Configure kernel window @@ -213,7 +217,7 @@ void NEL2NormalizeLayerKernel::configure(const ITensor *input, const ITensor *su INEKernel::configure(std::get<1>(win_config)); } -Status NEL2NormalizeLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *sum, const ITensorInfo *output, unsigned int axis, float epsilon) +Status NEL2NormalizeLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *sum, const ITensorInfo *output, int axis, float epsilon) { ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, sum, output, axis, epsilon)); ARM_COMPUTE_RETURN_ON_ERROR(std::get<0>(validate_and_configure_window(input->clone().get(), sum->clone().get(), output->clone().get(), axis))); @@ -227,7 +231,7 @@ void NEL2NormalizeLayerKernel::run(const Window &window, const ThreadInfo &info) ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); - switch(_axis) + switch(_actual_axis) { case 0: switch(_input->info()->data_type()) diff --git a/src/runtime/CL/functions/CLL2NormalizeLayer.cpp b/src/runtime/CL/functions/CLL2NormalizeLayer.cpp index 136cb5edef..e76e4f601e 100644 --- a/src/runtime/CL/functions/CLL2NormalizeLayer.cpp +++ b/src/runtime/CL/functions/CLL2NormalizeLayer.cpp @@ -34,25 +34,31 @@ namespace arm_compute { +namespace +{ +constexpr int max_input_tensor_dim = 3; +} // namespace + CLL2NormalizeLayer::CLL2NormalizeLayer(std::shared_ptr<IMemoryManager> memory_manager) : _memory_group(std::move(memory_manager)), _reduce_func(), _normalize_kernel(), _sumsq() { } -void CLL2NormalizeLayer::configure(ICLTensor *input, ICLTensor *output, unsigned int axis, float epsilon) +void CLL2NormalizeLayer::configure(ICLTensor *input, ICLTensor *output, int axis, float epsilon) { // Manage intermediate buffers _memory_group.manage(&_sumsq); // Configure kernels - _reduce_func.configure(input, &_sumsq, axis, ReductionOperation::SUM_SQUARE); + const uint32_t actual_axis = wrap_around(axis, max_input_tensor_dim); + _reduce_func.configure(input, &_sumsq, actual_axis, ReductionOperation::SUM_SQUARE); _normalize_kernel.configure(input, &_sumsq, output, axis, epsilon); // Allocate intermediate tensor _sumsq.allocator()->allocate(); } -Status CLL2NormalizeLayer::validate(const ITensorInfo *input, const ITensorInfo *output, unsigned int axis, float epsilon) +Status CLL2NormalizeLayer::validate(const ITensorInfo *input, const ITensorInfo *output, int axis, float epsilon) { TensorShape shape(input->tensor_shape()); @@ -61,10 +67,11 @@ Status CLL2NormalizeLayer::validate(const ITensorInfo *input, const ITensorInfo sum_sq.set_data_type(input->data_type()); sum_sq.set_tensor_shape(shape); - ARM_COMPUTE_RETURN_ON_ERROR(CLReductionOperation::validate(input, &sum_sq, axis, ReductionOperation::SUM_SQUARE)); + const uint32_t actual_axis = wrap_around(axis, max_input_tensor_dim); + ARM_COMPUTE_RETURN_ON_ERROR(CLReductionOperation::validate(input, &sum_sq, actual_axis, ReductionOperation::SUM_SQUARE)); // Reduce shape on axis - shape.set(axis, 1); + shape.set(actual_axis, 1); sum_sq.set_tensor_shape(shape); ARM_COMPUTE_RETURN_ON_ERROR(CLL2NormalizeLayerKernel::validate(input, &sum_sq, output, axis, epsilon)); diff --git a/src/runtime/NEON/functions/NEL2NormalizeLayer.cpp b/src/runtime/NEON/functions/NEL2NormalizeLayer.cpp index c9ab5c98e2..88ffdbfd08 100644 --- a/src/runtime/NEON/functions/NEL2NormalizeLayer.cpp +++ b/src/runtime/NEON/functions/NEL2NormalizeLayer.cpp @@ -28,25 +28,31 @@ namespace arm_compute { +namespace +{ +constexpr int max_input_tensor_dim = 3; +} // namespace + NEL2NormalizeLayer::NEL2NormalizeLayer(std::shared_ptr<IMemoryManager> memory_manager) : _memory_group(std::move(memory_manager)), _reduce_func(), _normalize_kernel(), _sumsq() { } -void NEL2NormalizeLayer::configure(ITensor *input, ITensor *output, unsigned int axis, float epsilon) +void NEL2NormalizeLayer::configure(ITensor *input, ITensor *output, int axis, float epsilon) { // Manage intermediate buffers _memory_group.manage(&_sumsq); // Configure Kernels - _reduce_func.configure(input, &_sumsq, axis, ReductionOperation::SUM_SQUARE); + const uint32_t actual_axis = wrap_around(axis, max_input_tensor_dim); + _reduce_func.configure(input, &_sumsq, actual_axis, ReductionOperation::SUM_SQUARE); _normalize_kernel.configure(input, &_sumsq, output, axis, epsilon); // Allocate intermediate tensors _sumsq.allocator()->allocate(); } -Status NEL2NormalizeLayer::validate(const ITensorInfo *input, const ITensorInfo *output, unsigned int axis, float epsilon) +Status NEL2NormalizeLayer::validate(const ITensorInfo *input, const ITensorInfo *output, int axis, float epsilon) { TensorShape shape(input->tensor_shape()); @@ -55,10 +61,11 @@ Status NEL2NormalizeLayer::validate(const ITensorInfo *input, const ITensorInfo sum_sq.set_data_type(input->data_type()); sum_sq.set_tensor_shape(shape); - ARM_COMPUTE_RETURN_ON_ERROR(NEReductionOperation::validate(input, &sum_sq, axis, ReductionOperation::SUM_SQUARE)); + const uint32_t actual_axis = wrap_around(axis, max_input_tensor_dim); + ARM_COMPUTE_RETURN_ON_ERROR(NEReductionOperation::validate(input, &sum_sq, actual_axis, ReductionOperation::SUM_SQUARE)); // Reduce shape on axis - shape.set(axis, 1); + shape.set(actual_axis, 1); sum_sq.set_tensor_shape(shape); ARM_COMPUTE_RETURN_ON_ERROR(NEL2NormalizeLayerKernel::validate(input, &sum_sq, output, axis, epsilon)); diff --git a/tests/validation/CL/L2NormalizeLayer.cpp b/tests/validation/CL/L2NormalizeLayer.cpp index fdbfa3ed4d..beedd81335 100644 --- a/tests/validation/CL/L2NormalizeLayer.cpp +++ b/tests/validation/CL/L2NormalizeLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2018 ARM Limited. + * Copyright (c) 2017-2019 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -46,8 +46,8 @@ namespace constexpr AbsoluteTolerance<float> tolerance_f32(0.00001f); constexpr AbsoluteTolerance<float> tolerance_f16(0.2f); -auto data = concat(combine(framework::dataset::make("DataLayout", { DataLayout::NCHW }), framework::dataset::make("Axis", { 0, 1, 2 })), combine(framework::dataset::make("DataLayout", { DataLayout::NHWC }), - framework::dataset::make("Axis", { 1, 2 }))); +auto data = concat(combine(framework::dataset::make("DataLayout", { DataLayout::NCHW }), framework::dataset::make("Axis", { -1, 0, 2 })), combine(framework::dataset::make("DataLayout", { DataLayout::NHWC }), + framework::dataset::make("Axis", { -2, 2 }))); } // namespace @@ -61,8 +61,9 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( TensorInfo(TensorShape(128U, 64U), 1, DataType::F32), // Mismatching shape input/output TensorInfo(TensorShape(128U, 64U), 2, DataType::F32), // Number of Input channels != 1 TensorInfo(TensorShape(128U, 64U), 1, DataType::S16), // DataType != F32 - TensorInfo(TensorShape(128U, 64U), 1, DataType::F32), // Axis >= num_max_dimensions - TensorInfo(TensorShape(128U, 64U), 1, DataType::F32), // Axis > 2 + TensorInfo(TensorShape(128U, 64U), 1, DataType::F32), + TensorInfo(TensorShape(128U, 64U), 1, DataType::F32), + TensorInfo(TensorShape(128U, 64U), 1, DataType::F32), TensorInfo(TensorShape(128U, 64U), 1, DataType::F32) }), framework::dataset::make("OutputInfo", { TensorInfo(TensorShape(128U, 64U), 1, DataType::F16), @@ -71,10 +72,19 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( TensorInfo(TensorShape(128U, 64U), 1, DataType::S16), TensorInfo(TensorShape(128U, 64U), 1, DataType::F32), TensorInfo(TensorShape(128U, 64U), 1, DataType::F32), + TensorInfo(TensorShape(128U, 64U), 1, DataType::F32), TensorInfo(TensorShape(128U, 64U), 1, DataType::F32) })), - framework::dataset::make("Axis", { 0U, 0U, 0U, 0U, static_cast<unsigned int>(TensorShape::num_max_dimensions), 3U, 0U })), - framework::dataset::make("Expected", { false, false, false, false, false, false, true })), + framework::dataset::make("Axis", { + 0, + 0, + 0, + 0, + static_cast<int>(TensorShape::num_max_dimensions), + 3, + -2, + 0 })), + framework::dataset::make("Expected", { false, false, false, false, true, true, true, true })), input_info, output_info, axis, expected) { bool is_valid = bool(CLL2NormalizeLayer::validate(&input_info.clone()->set_is_resizable(false), diff --git a/tests/validation/NEON/L2NormalizeLayer.cpp b/tests/validation/NEON/L2NormalizeLayer.cpp index 3164a65417..17147c1d50 100644 --- a/tests/validation/NEON/L2NormalizeLayer.cpp +++ b/tests/validation/NEON/L2NormalizeLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2018 ARM Limited. + * Copyright (c) 2017-2019 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -59,8 +59,9 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( TensorInfo(TensorShape(128U, 64U), 1, DataType::F32), // Mismatching shape input/output TensorInfo(TensorShape(128U, 64U), 2, DataType::F32), // Number of Input channels != 1 TensorInfo(TensorShape(128U, 64U), 1, DataType::S16), // DataType != F32 - TensorInfo(TensorShape(128U, 64U), 1, DataType::F32), // Axis >= num_max_dimensions - TensorInfo(TensorShape(128U, 64U), 1, DataType::F32), // Axis > 2 + TensorInfo(TensorShape(128U, 64U), 1, DataType::F32), + TensorInfo(TensorShape(128U, 64U), 1, DataType::F32), + TensorInfo(TensorShape(128U, 64U), 1, DataType::F32), TensorInfo(TensorShape(128U, 64U), 1, DataType::F32) }), framework::dataset::make("OutputInfo", { TensorInfo(TensorShape(128U, 64U), 1, DataType::F16), @@ -69,10 +70,19 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( TensorInfo(TensorShape(128U, 64U), 1, DataType::S16), TensorInfo(TensorShape(128U, 64U), 1, DataType::F32), TensorInfo(TensorShape(128U, 64U), 1, DataType::F32), + TensorInfo(TensorShape(128U, 64U), 1, DataType::F32), TensorInfo(TensorShape(128U, 64U), 1, DataType::F32) })), - framework::dataset::make("Axis", { 0U, 0U, 0U, 0U, static_cast<unsigned int>(TensorShape::num_max_dimensions), 3U, 0U })), - framework::dataset::make("Expected", { false, false, false, false, false, false, true })), + framework::dataset::make("Axis", { + 0, + 0, + 0, + 0, + static_cast<int>(TensorShape::num_max_dimensions), + 3, + -2, + 0 })), + framework::dataset::make("Expected", { false, false, false, false, true, true, true, true })), input_info, output_info, axis, expected) { bool is_valid = bool(NEL2NormalizeLayer::validate(&input_info.clone()->set_is_resizable(false), @@ -89,7 +99,7 @@ using NEL2NormalizeLayerFixture = L2NormalizeLayerValidationFixture<Tensor, Acce TEST_SUITE(FP32) FIXTURE_DATA_TEST_CASE(RunSmall, NEL2NormalizeLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", DataType::F32)), framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - framework::dataset::make("Axis", { 0, 1, 2 })), + framework::dataset::make("Axis", { -1, 0, 2 })), framework::dataset::make("Epsilon", { 1e-12 }))) { // Validate output @@ -98,7 +108,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEL2NormalizeLayerFixture<float>, framework::Da FIXTURE_DATA_TEST_CASE(RunLarge, NEL2NormalizeLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", DataType::F32)), framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - framework::dataset::make("Axis", { 0, 1, 2 })), + framework::dataset::make("Axis", { -1, 0, 2 })), framework::dataset::make("Epsilon", { 1e-12 }))) { // Validate output @@ -110,7 +120,7 @@ TEST_SUITE_END() // FP32 TEST_SUITE(FP16) FIXTURE_DATA_TEST_CASE(RunSmall, NEL2NormalizeLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", DataType::F16)), framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - framework::dataset::make("Axis", { 0, 1, 2 })), + framework::dataset::make("Axis", { -1, 0, 2 })), framework::dataset::make("Epsilon", { 1e-12 }))) { // Validate output @@ -119,7 +129,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEL2NormalizeLayerFixture<half>, framework::Dat FIXTURE_DATA_TEST_CASE(RunLarge, NEL2NormalizeLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", DataType::F16)), framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - framework::dataset::make("Axis", { 0, 1, 2 })), + framework::dataset::make("Axis", { -1, 0, 2 })), framework::dataset::make("Epsilon", { 1e-12 }))) { // Validate output diff --git a/tests/validation/fixtures/L2NormalizeLayerFixture.h b/tests/validation/fixtures/L2NormalizeLayerFixture.h index 574722bd88..e3e1510ff0 100644 --- a/tests/validation/fixtures/L2NormalizeLayerFixture.h +++ b/tests/validation/fixtures/L2NormalizeLayerFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2018 ARM Limited. + * Copyright (c) 2017-2019 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -40,12 +40,16 @@ namespace test { namespace validation { +namespace +{ +constexpr int max_input_tensor_dim = 3; +} // namespace template <typename TensorType, typename AccessorType, typename FunctionType, typename T> class L2NormalizeLayerValidationFixture : public framework::Fixture { public: template <typename...> - void setup(TensorShape shape, DataType data_type, DataLayout data_layout, unsigned int axis, float epsilon) + void setup(TensorShape shape, DataType data_type, DataLayout data_layout, int axis, float epsilon) { _target = compute_target(shape, data_type, data_layout, axis, epsilon); _reference = compute_reference(shape, data_type, data_layout, axis, epsilon); @@ -59,7 +63,7 @@ protected: library->fill(tensor, distribution, 0); } - TensorType compute_target(TensorShape shape, DataType data_type, DataLayout data_layout, unsigned int axis, float epsilon) + TensorType compute_target(TensorShape shape, DataType data_type, DataLayout data_layout, int axis, float epsilon) { if(data_layout == DataLayout::NHWC) { @@ -93,20 +97,21 @@ protected: return dst; } - SimpleTensor<T> compute_reference(const TensorShape &shape, DataType data_type, DataLayout data_layout, unsigned int axis, float epsilon) + SimpleTensor<T> compute_reference(const TensorShape &shape, DataType data_type, DataLayout data_layout, int axis, float epsilon) { + uint32_t actual_axis = wrap_around(axis, max_input_tensor_dim); if(data_layout == DataLayout::NHWC) { - switch(axis) + switch(actual_axis) { case 0: - axis = 2; + actual_axis = 2; break; case 1: - axis = 0; + actual_axis = 0; break; case 2: - axis = 1; + actual_axis = 1; break; default: break; @@ -118,7 +123,7 @@ protected: // Fill reference fill(src); - return reference::l2_normalize<T>(src, axis, epsilon); + return reference::l2_normalize<T>(src, actual_axis, epsilon); } TensorType _target{}; |