diff options
Diffstat (limited to 'src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.cpp')
-rw-r--r-- | src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.cpp | 100 |
1 files changed, 67 insertions, 33 deletions
diff --git a/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.cpp b/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.cpp index b5a007e832..343c31c73d 100644 --- a/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.cpp +++ b/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.cpp @@ -36,6 +36,53 @@ using namespace arm_compute; namespace arm_compute { +namespace +{ +Error validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min, int max) +{ + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::S32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QASYMM8); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output); + ARM_COMPUTE_RETURN_ERROR_ON(max > 255); + ARM_COMPUTE_RETURN_ERROR_ON(min < 0 || min > max); + + // Check biases if exist + if(bias != nullptr) + { + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, bias); + ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() > 1); + ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) != bias->dimension(0)); + } + return Error{}; +} + +std::pair<Error, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *bias, 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)); + + AccessWindowHorizontal input_access(input, 0, num_elems_processed_per_iteration); + AccessWindowHorizontal output_result_access(output, 0, num_elems_processed_per_iteration); + + bool window_changed = update_window_and_padding(win, + input_access, + output_result_access); + + if(bias != nullptr) + { + AccessWindowStatic bias_access(bias, 0, 0, ceil_to_multiple(bias->dimension(0), num_elems_processed_per_iteration), bias->tensor_shape()[1]); + window_changed = window_changed || update_window_and_padding(win, bias_access); + } + + output_result_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape())); + + Error err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Error{}; + return std::make_pair(err, win); +} +} // namespace + class Coordinates; } // namespace arm_compute @@ -43,25 +90,31 @@ CLGEMMLowpQuantizeDownInt32ToUint8ScaleKernel::CLGEMMLowpQuantizeDownInt32ToUint : _input(nullptr), _bias(nullptr), _output(nullptr) { } +Error CLGEMMLowpQuantizeDownInt32ToUint8ScaleKernel::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min, int max) +{ + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, bias, output, min, max)); + ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), + (bias != nullptr) ? bias->clone().get() : nullptr, + output->clone().get()) + .first); + + return Error{}; +} void CLGEMMLowpQuantizeDownInt32ToUint8ScaleKernel::configure(const ICLTensor *input, const ICLTensor *bias, ICLTensor *output, int result_offset, int result_mult_int, int result_shift, int min, int max) { - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::S32); - ARM_COMPUTE_ERROR_ON(max > 255); - ARM_COMPUTE_ERROR_ON(min < 0 || min > max); - - if(bias != nullptr) - { - ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, bias); - ARM_COMPUTE_ERROR_ON(bias->info()->num_dimensions() > 1); - ARM_COMPUTE_ERROR_ON(input->info()->dimension(0) != bias->info()->dimension(0)); - } + // Perform validate step + ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); // Output auto inizialitation if not yet initialized auto_init_if_empty(*output->info(), input->info()->clone()->set_data_type(DataType::QASYMM8)); - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QASYMM8); + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), + (bias != nullptr) ? bias->info() : nullptr, + output->info(), + min, + max)); _input = input; _bias = bias; @@ -79,29 +132,10 @@ void CLGEMMLowpQuantizeDownInt32ToUint8ScaleKernel::configure(const ICLTensor *i // Create kernel _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("gemmlowp_output_stage_quantize_down", build_opts.options())); - constexpr unsigned int num_elems_processed_per_iteration = 16; - // Configure kernel window - Window win = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration)); - - AccessWindowHorizontal input_access(input->info(), 0, num_elems_processed_per_iteration); - AccessWindowHorizontal output_result_access(output->info(), 0, num_elems_processed_per_iteration); - - update_window_and_padding(win, - input_access, - output_result_access); - - if(bias != nullptr) - { - AccessWindowStatic bias_access(bias->info(), 0, 0, ceil_to_multiple(bias->info()->dimension(0), num_elems_processed_per_iteration), bias->info()->tensor_shape()[1]); - - update_window_and_padding(win, - bias_access); - } - - output_result_access.set_valid_region(win, ValidRegion(Coordinates(), output->info()->tensor_shape())); - - ICLKernel::configure(win); + auto win_config = validate_and_configure_window(input->info(), (bias != nullptr) ? bias->info() : nullptr, output->info()); + ARM_COMPUTE_ERROR_THROW_ON(win_config.first); + ICLKernel::configure(win_config.second); } void CLGEMMLowpQuantizeDownInt32ToUint8ScaleKernel::run(const Window &window, cl::CommandQueue &queue) |