From d60a6b9d7977c6bd63ff7c523bed84d42363898b Mon Sep 17 00:00:00 2001 From: Gian Marco Iodice Date: Thu, 10 Aug 2017 10:43:40 +0100 Subject: COMPMID-477 - Optimized CLNormalizationLayer CLPixelWiseMultiplication has been removed within the function Change-Id: Ibe7edd7921d5cef6ff68fdeeca89771129a8eaea Reviewed-on: http://mpd-gerrit.cambridge.arm.com/84459 Reviewed-by: Anthony Barbier Tested-by: Kaizen Reviewed-by: Georgios Pinitas --- src/core/CL/cl_kernels/normalization_layer.cl | 120 +++++++++------------ src/core/CL/kernels/CLNormalizationLayerKernel.cpp | 25 ++--- 2 files changed, 61 insertions(+), 84 deletions(-) (limited to 'src/core') diff --git a/src/core/CL/cl_kernels/normalization_layer.cl b/src/core/CL/cl_kernels/normalization_layer.cl index e2a5c4079a..4e65560b95 100644 --- a/src/core/CL/cl_kernels/normalization_layer.cl +++ b/src/core/CL/cl_kernels/normalization_layer.cl @@ -54,43 +54,33 @@ * * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size, e.g. -DVEC_SIZE=16 + * @note The radius should be given as a preprocessor argument using -DRADIUS=size. e.g. -DRADIUS=5 + * @note The number of slices should be given as a preprocessor argument using -DNUM_SLICES=size. e.g. -DNUM_SLICES=192 * @note In case of fixed-point operation -DFIXED_POINT_POSITION=fixed_point_position must be provided: e.g. -DFIXED_POINT_POSITION=3 * @note Scaling coefficient (= alpha/norm_size), beta and kappa need to be passed at compile time using -DCOEFF, -DALPHA and -DKAPPA * - * @param[in] input_ptr Pointer to the first source tensor. Supported data types: QS8/QS16/F16/F32 - * @param[in] input_stride_x Stride of the first source tensor in X dimension (in bytes) - * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] input_stride_y Stride of the first source tensor in Y dimension (in bytes) - * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] input_stride_z Stride of the first source tensor in Z dimension (in bytes) - * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] input_offset_first_element_in_bytes The offset of the first element in the first source tensor - * @param[in] squared_input_ptr Pointer to the second source tensor. Supported data types: same as @p input_ptr - * @param[in] squared_input_stride_x Stride of the second source tensor in X dimension (in bytes) - * @param[in] squared_input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] squared_input_stride_y Stride of the second source tensor in Y dimension (in bytes) - * @param[in] squared_input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] squared_input_stride_z Stride of the second source tensor in Z dimension (in bytes) - * @param[in] squared_input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] squared_input_offset_first_element_in_bytes The offset of the second element in the second source tensor - * @param[out] output_ptr Pointer to the destination tensor. Supported data types: same as @p input_ptr - * @param[in] output_stride_x Stride of the destination tensor in X dimension (in bytes) - * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] output_stride_y Stride of the destination tensor in Y dimension (in bytes) - * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] output_stride_z Stride of the destination tensor in Z dimension (in bytes) - * @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination tensor - * @param[in] radius Number of elements on the right or left side to normalize across + * @param[in] input_ptr Pointer to the first source tensor. Supported data types: QS8/QS16/F16/F32 + * @param[in] input_stride_x Stride of the first source tensor in X dimension (in bytes) + * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] input_stride_y Stride of the first source tensor in Y dimension (in bytes) + * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] input_stride_z Stride of the first source tensor in Z dimension (in bytes) + * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] input_offset_first_element_in_bytes The offset of the first element in the first source tensor + * @param[out] output_ptr Pointer to the destination tensor. Supported data types: same as @p input_ptr + * @param[in] output_stride_x Stride of the destination tensor in X dimension (in bytes) + * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] output_stride_y Stride of the destination tensor in Y dimension (in bytes) + * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] output_stride_z Stride of the destination tensor in Z dimension (in bytes) + * @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination tensor */ __kernel void normalization_layer_cross_map(TENSOR3D_DECLARATION(input), - TENSOR3D_DECLARATION(squared_input), - TENSOR3D_DECLARATION(output), - uint radius) + TENSOR3D_DECLARATION(output)) { - Tensor3D in = CONVERT_TO_TENSOR3D_STRUCT(input); - Tensor3D squared_in = CONVERT_TO_TENSOR3D_STRUCT(squared_input); - Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(output); + Tensor3D in = CONVERT_TO_TENSOR3D_STRUCT(input); + Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(output); VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) acc = (VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))0; @@ -101,15 +91,16 @@ __kernel void normalization_layer_cross_map(TENSOR3D_DECLARATION(input), const VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) kappa_v = (VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))SQCVT_SAT(KAPPA); - const int num_of_slices = get_global_size(2); const int current_slice = get_global_id(2); - const int left_slice = max(current_slice - (int)radius, (int)0); - const int right_slice = min(current_slice + (int)radius, (int)(num_of_slices - 1)); + const int left_slice = max(current_slice - (int)RADIUS, (int)0); + const int right_slice = min(current_slice + (int)RADIUS, (int)(NUM_SLICES - 1)); for(int i = left_slice; i <= right_slice; i++) { - acc = ADD_OP(acc, LOAD_OP(0, (__global DATA_TYPE *)tensor3D_offset(&squared_in, 0, 0, i - current_slice))); + VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) + values = LOAD_OP(0, (__global DATA_TYPE *)tensor3D_offset(&in, 0, 0, i - current_slice)); + acc = ADD_OP(acc, MUL_OP(values, values)); } acc = ADD_OP(MUL_OP(acc, coeff_v), kappa_v); @@ -125,43 +116,32 @@ __kernel void normalization_layer_cross_map(TENSOR3D_DECLARATION(input), * * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size, e.g. -DVEC_SIZE=16 + * @note The radius should be given as a preprocessor argument using -DRADIUS=size. e.g. -DRADIUS=5 * @note In case of fixed-point operation -DFIXED_POINT_POSITION=fixed_point_position must be provided: e.g. -DFIXED_POINT_POSITION=3 * @note Scaling coefficient (= alpha/norm_size), beta and kappa need to be passed at compile time using -DCOEFF, -DALPHA and -DKAPPA * - * @param[in] input_ptr Pointer to the first source tensor. Supported data types: QS8/F16/F32 - * @param[in] input_stride_x Stride of the first source tensor in X dimension (in bytes) - * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] input_stride_y Stride of the first source tensor in Y dimension (in bytes) - * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] input_stride_z Stride of the first source tensor in Z dimension (in bytes) - * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] input_offset_first_element_in_bytes The offset of the first element in the first source tensor - * @param[in] squared_input_ptr Pointer to the second source tensor. Supported data types: same as @p input_ptr - * @param[in] squared_input_stride_x Stride of the second source tensor in X dimension (in bytes) - * @param[in] squared_input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] squared_input_stride_y Stride of the second source tensor in Y dimension (in bytes) - * @param[in] squared_input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] squared_input_stride_z Stride of the second source tensor in Z dimension (in bytes) - * @param[in] squared_input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] squared_input_offset_first_element_in_bytes The offset of the second element in the second source tensor - * @param[out] output_ptr Pointer to the destination tensor. Supported data types: same as @p input_ptr - * @param[in] output_stride_x Stride of the destination tensor in X dimension (in bytes) - * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] output_stride_y Stride of the first destination tensor in Y dimension (in bytes) - * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] output_stride_z Stride of the first source tensor in Z dimension (in bytes) - * @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination tensor - * @param[in] radius Number of elements on the right or left side to normalize across + * @param[in] input_ptr Pointer to the first source tensor. Supported data types: QS8/F16/F32 + * @param[in] input_stride_x Stride of the first source tensor in X dimension (in bytes) + * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] input_stride_y Stride of the first source tensor in Y dimension (in bytes) + * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] input_stride_z Stride of the first source tensor in Z dimension (in bytes) + * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] input_offset_first_element_in_bytes The offset of the first element in the first source tensor + * @param[out] output_ptr Pointer to the destination tensor. Supported data types: same as @p input_ptr + * @param[in] output_stride_x Stride of the destination tensor in X dimension (in bytes) + * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] output_stride_y Stride of the first destination tensor in Y dimension (in bytes) + * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] output_stride_z Stride of the first source tensor in Z dimension (in bytes) + * @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination tensor */ __kernel void normalization_layer_in_map_1D(TENSOR3D_DECLARATION(input), - TENSOR3D_DECLARATION(squared_input), - TENSOR3D_DECLARATION(output), - uint radius) + TENSOR3D_DECLARATION(output)) { - Tensor3D in = CONVERT_TO_TENSOR3D_STRUCT(input); - Tensor3D squared_in = CONVERT_TO_TENSOR3D_STRUCT(squared_input); - Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(output); + Tensor3D in = CONVERT_TO_TENSOR3D_STRUCT(input); + Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(output); VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) acc = (VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))0; @@ -174,12 +154,14 @@ __kernel void normalization_layer_in_map_1D(TENSOR3D_DECLARATION(input), const int current_pos = get_global_id(0) << 2; - const int left_pos = max(current_pos - (int)radius, -3); - const int right_pos = min(current_pos + (int)radius, (int)((get_global_size(0) << 2) + 3 - 1)); + const int left_pos = max(current_pos - (int)RADIUS, -3); + const int right_pos = min(current_pos + (int)RADIUS, (int)((get_global_size(0) << 2) + 3 - 1)); for(int i = left_pos; i <= right_pos; i += 1) { - acc = ADD_OP(acc, LOAD_OP(0, (__global DATA_TYPE *)tensor3D_offset(&squared_in, i - current_pos, 0, 0))); + VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) + values = LOAD_OP(0, (__global DATA_TYPE *)tensor3D_offset(&in, i - current_pos, 0, 0)); + acc = ADD_OP(acc, MUL_OP(values, values)); } acc = ADD_OP(MUL_OP(acc, coeff_v), kappa_v); diff --git a/src/core/CL/kernels/CLNormalizationLayerKernel.cpp b/src/core/CL/kernels/CLNormalizationLayerKernel.cpp index b382e9d510..a74473980b 100644 --- a/src/core/CL/kernels/CLNormalizationLayerKernel.cpp +++ b/src/core/CL/kernels/CLNormalizationLayerKernel.cpp @@ -36,7 +36,7 @@ using namespace arm_compute; CLNormalizationLayerKernel::CLNormalizationLayerKernel() - : _input(nullptr), _squared_input(nullptr), _output(nullptr), _border_size(0), _is_in_map(false) + : _input(nullptr), _output(nullptr), _border_size(0), _is_in_map(false) { } @@ -45,7 +45,7 @@ BorderSize CLNormalizationLayerKernel::border_size() const return _border_size; } -void CLNormalizationLayerKernel::configure(const ICLTensor *input, const ICLTensor *squared_input, ICLTensor *output, NormalizationLayerInfo norm_info) +void CLNormalizationLayerKernel::configure(const ICLTensor *input, ICLTensor *output, NormalizationLayerInfo norm_info) { ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32); ARM_COMPUTE_ERROR_ON_NULLPTR(output); @@ -53,21 +53,20 @@ void CLNormalizationLayerKernel::configure(const ICLTensor *input, const ICLTens // Output tensor auto initialization if not yet initialized auto_init_if_empty(*output->info(), input->info()->tensor_shape(), 1, input->info()->data_type(), input->info()->fixed_point_position()); - ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, squared_input, output); - ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(input, squared_input, output); + ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); + ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(input, output); ARM_COMPUTE_ERROR_ON_MSG(!(norm_info.norm_size() % 2), "Normalization size should be odd"); ARM_COMPUTE_ERROR_ON_MSG(norm_info.type() == NormType::IN_MAP_2D, "2D In-Map Normalization not implemented"); if(is_data_type_fixed_point(input->info()->data_type())) { - ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, squared_input, output); + ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); ARM_COMPUTE_ERROR_ON_VALUE_NOT_REPRESENTABLE_IN_FIXED_POINT(norm_info.beta(), input); ARM_COMPUTE_ERROR_ON_VALUE_NOT_REPRESENTABLE_IN_FIXED_POINT(norm_info.kappa(), input); ARM_COMPUTE_ERROR_ON_VALUE_NOT_REPRESENTABLE_IN_FIXED_POINT(norm_info.scale_coeff(), input); } - _input = input; - _squared_input = squared_input; - _output = output; + _input = input; + _output = output; _is_in_map = (norm_info.type() != NormType::CROSS_MAP); const unsigned int border_width = _is_in_map ? std::min(norm_info.norm_size() / 2, 3U) : 0; @@ -87,23 +86,20 @@ void CLNormalizationLayerKernel::configure(const ICLTensor *input, const ICLTens build_opts.emplace(("-DBETA=" + float_to_string_with_full_precision(norm_info.beta()))); build_opts.emplace(("-DKAPPA=" + float_to_string_with_full_precision(norm_info.kappa()))); build_opts.emplace(("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration))); + build_opts.emplace(("-DRADIUS=" + support::cpp11::to_string(norm_info.norm_size() / 2))); + build_opts.emplace(("-DNUM_SLICES=" + support::cpp11::to_string(input->info()->dimension(2)))); // Create kernel std::string kernel_name = (norm_info.type() == NormType::IN_MAP_1D) ? "normalization_layer_in_map_1D" : "normalization_layer_cross_map"; _kernel = static_cast(CLKernelLibrary::get().create_kernel(kernel_name, build_opts)); - // Set kernel static arguments - unsigned int idx = 3 * num_arguments_per_3D_tensor(); // Skip the input and output parameters - _kernel.setArg(idx++, norm_info.norm_size() / 2); - // Configure kernel window Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration)); AccessWindowHorizontal input_access(input->info(), -_border_size.left, num_elems_read_per_iteration); - AccessWindowHorizontal squared_input_access(squared_input->info(), -_border_size.left, num_elems_read_per_iteration); AccessWindowHorizontal output_access(output->info(), 0, num_elems_processed_per_iteration); - update_window_and_padding(win, input_access, squared_input_access, output_access); + update_window_and_padding(win, input_access, output_access); output_access.set_valid_region(win, input->info()->valid_region()); @@ -123,7 +119,6 @@ void CLNormalizationLayerKernel::run(const Window &window, cl::CommandQueue &que { unsigned int idx = 0; add_3D_tensor_argument(idx, _input, slice); - add_3D_tensor_argument(idx, _squared_input, slice); add_3D_tensor_argument(idx, _output, slice); enqueue(queue, *this, slice); } -- cgit v1.2.1