From a3b1b469276b10484cd45901ae3a4b48b506caa9 Mon Sep 17 00:00:00 2001 From: Georgios Pinitas Date: Thu, 16 Nov 2017 19:24:39 +0000 Subject: COMPMID-667: Add validation static method to NEON GEMMlowp Change-Id: I8a470cc1351593ad8eeaf4ec92e04865e83d4f3c Reviewed-on: http://mpd-gerrit.cambridge.arm.com/96147 Tested-by: Kaizen Reviewed-by: Anthony Barbier --- .../NEON/kernels/NEGEMMLowpReductionKernel.cpp | 107 +++++++++++++++------ 1 file changed, 79 insertions(+), 28 deletions(-) (limited to 'src/core/NEON/kernels/NEGEMMLowpReductionKernel.cpp') diff --git a/src/core/NEON/kernels/NEGEMMLowpReductionKernel.cpp b/src/core/NEON/kernels/NEGEMMLowpReductionKernel.cpp index 81d9b5bb81..a8395a15cb 100644 --- a/src/core/NEON/kernels/NEGEMMLowpReductionKernel.cpp +++ b/src/core/NEON/kernels/NEGEMMLowpReductionKernel.cpp @@ -44,6 +44,59 @@ namespace arm_compute class Coordinates; } // namespace arm_compute +namespace +{ +Error validate_arguments_matrix_a_reduction(const ITensorInfo *input, const ITensorInfo *output) +{ + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::S32); + + return Error{}; +} +std::pair validate_and_configure_window_matrix_a_reduction(ITensorInfo *input, ITensorInfo *output, bool is_reshaped) +{ + const unsigned int num_elems_processed_per_iteration = is_reshaped ? 4 : 1; + + Window win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration)); + + AccessWindowStatic input_access(input, 0, 0, ceil_to_multiple(input->dimension(0), 16), input->dimension(1)); + AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration); + + bool window_changed = update_window_and_padding(win, input_access, output_access); + + output_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), output->tensor_shape())); + + Error err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Error{}; + return std::make_pair(err, win); +} + +Error validate_arguments_matrix_b_reduction(const ITensorInfo *input, const ITensorInfo *output) +{ + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::S32); + + return Error{}; +} + +std::pair validate_and_configure_window_matrix_b_reduction(ITensorInfo *input, ITensorInfo *output) +{ + constexpr unsigned int num_elems_processed_per_iteration = 16; + + // Configure kernel window + Window win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration)); + + AccessWindowStatic input_access(input, 0, 0, ceil_to_multiple(input->dimension(0), 16), input->dimension(1)); + AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration); + + bool window_changed = update_window_and_padding(win, input_access, output_access); + + output_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), output->tensor_shape())); + + Error err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Error{}; + return std::make_pair(err, win); +} +} // namespace + INEGEMMLowpReductionKernel::INEGEMMLowpReductionKernel() : _input(), _output(), _k(0), _is_reshaped(false) { @@ -51,29 +104,28 @@ INEGEMMLowpReductionKernel::INEGEMMLowpReductionKernel() void NEGEMMLowpMatrixAReductionKernel::configure(const ITensor *mtx_a, ITensor *vector_sum_row, int32_t num_mtx_a_cols, bool is_interleaved4x4) { - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(mtx_a, 1, DataType::QASYMM8); - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(vector_sum_row, 1, DataType::S32); + // Perform validate step + ARM_COMPUTE_ERROR_ON_NULLPTR(mtx_a, vector_sum_row); + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments_matrix_a_reduction(mtx_a->info(), vector_sum_row->info())); _input = mtx_a; _output = vector_sum_row; _k = num_mtx_a_cols; _is_reshaped = is_interleaved4x4; - const unsigned int num_elems_processed_per_iteration = _is_reshaped ? 4 : 1; - // Configure kernel window - Window win = calculate_max_window(*_output->info(), Steps(num_elems_processed_per_iteration)); - - AccessWindowStatic input_access(_input->info(), 0, 0, ceil_to_multiple(_input->info()->dimension(0), 16), _input->info()->dimension(1)); - AccessWindowHorizontal output_access(_output->info(), 0, num_elems_processed_per_iteration); - - update_window_and_padding(win, - input_access, - output_access); + auto win_config = validate_and_configure_window_matrix_a_reduction(_input->info(), _output->info(), _is_reshaped); + ARM_COMPUTE_ERROR_THROW_ON(win_config.first); + INEKernel::configure(win_config.second); +} - output_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), _output->info()->tensor_shape())); +Error NEGEMMLowpMatrixAReductionKernel::validate(const ITensorInfo *mtx_a, const ITensorInfo *vector_sum_row, int32_t num_mtx_a_cols, bool is_interleaved4x4) +{ + ARM_COMPUTE_UNUSED(num_mtx_a_cols); + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_matrix_a_reduction(mtx_a, vector_sum_row)); + ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window_matrix_a_reduction(mtx_a->clone().get(), vector_sum_row->clone().get(), is_interleaved4x4).first); - INEKernel::configure(win); + return Error{}; } void NEGEMMLowpMatrixAReductionKernel::run(const Window &window, const ThreadInfo &info) @@ -200,29 +252,28 @@ void NEGEMMLowpMatrixAReductionKernel::run(const Window &window, const ThreadInf void NEGEMMLowpMatrixBReductionKernel::configure(const ITensor *mtx_b, ITensor *vector_sum_col, int32_t num_mtx_b_rows, bool is_transposed1xW) { - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(mtx_b, 1, DataType::QASYMM8); - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(vector_sum_col, 1, DataType::S32); + ARM_COMPUTE_ERROR_ON_NULLPTR(mtx_b, vector_sum_col); + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments_matrix_b_reduction(mtx_b->info(), vector_sum_col->info())); _input = mtx_b; _output = vector_sum_col; _k = num_mtx_b_rows; _is_reshaped = is_transposed1xW; - constexpr unsigned int num_elems_processed_per_iteration = 16; - // Configure kernel window - Window win = calculate_max_window(*vector_sum_col->info(), Steps(num_elems_processed_per_iteration)); - - AccessWindowStatic input_access(_input->info(), 0, 0, ceil_to_multiple(_input->info()->dimension(0), 16), _input->info()->dimension(1)); - AccessWindowHorizontal output_access(_output->info(), 0, num_elems_processed_per_iteration); - - update_window_and_padding(win, - input_access, - output_access); + auto win_config = validate_and_configure_window_matrix_b_reduction(_input->info(), _output->info()); + ARM_COMPUTE_ERROR_THROW_ON(win_config.first); + INEKernel::configure(win_config.second); +} - output_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), _output->info()->tensor_shape())); +Error NEGEMMLowpMatrixBReductionKernel::validate(const ITensorInfo *mtx_b, const ITensorInfo *vector_sum_col, int32_t num_mtx_b_rows, bool is_transposed1xW) +{ + ARM_COMPUTE_UNUSED(num_mtx_b_rows); + ARM_COMPUTE_UNUSED(is_transposed1xW); + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_matrix_b_reduction(mtx_b, vector_sum_col)); + ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window_matrix_b_reduction(mtx_b->clone().get(), vector_sum_col->clone().get()).first); - INEKernel::configure(win); + return Error{}; } void NEGEMMLowpMatrixBReductionKernel::run(const Window &window, const ThreadInfo &info) -- cgit v1.2.1