diff options
Diffstat (limited to 'src/gpu/cl/kernels/ClGemmLowpQuantizeDownInt32ScaleKernel.cpp')
-rw-r--r-- | src/gpu/cl/kernels/ClGemmLowpQuantizeDownInt32ScaleKernel.cpp | 64 |
1 files changed, 42 insertions, 22 deletions
diff --git a/src/gpu/cl/kernels/ClGemmLowpQuantizeDownInt32ScaleKernel.cpp b/src/gpu/cl/kernels/ClGemmLowpQuantizeDownInt32ScaleKernel.cpp index 5d54db214a..31434ce61b 100644 --- a/src/gpu/cl/kernels/ClGemmLowpQuantizeDownInt32ScaleKernel.cpp +++ b/src/gpu/cl/kernels/ClGemmLowpQuantizeDownInt32ScaleKernel.cpp @@ -26,15 +26,14 @@ #include "arm_compute/core/CL/CLHelpers.h" #include "arm_compute/core/CL/ICLTensor.h" #include "arm_compute/core/Helpers.h" -#include "arm_compute/core/Validate.h" #include "arm_compute/core/Utils.h" #include "arm_compute/core/utils/helpers/AdjustVecSize.h" #include "arm_compute/core/utils/quantization/AsymmHelpers.h" #include "arm_compute/core/utils/StringUtils.h" +#include "arm_compute/core/Validate.h" #include "src/core/helpers/AutoConfiguration.h" #include "src/core/helpers/WindowHelpers.h" - #include "support/Cast.h" #include "support/StringSupport.h" @@ -46,25 +45,34 @@ namespace kernels { namespace { -Status validate_arguments(const ITensorInfo *src, const ITensorInfo *bias, const ITensorInfo *dst, const GEMMLowpOutputStageInfo *output_stage) +Status validate_arguments(const ITensorInfo *src, + const ITensorInfo *bias, + const ITensorInfo *dst, + const GEMMLowpOutputStageInfo *output_stage) { ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::S32); - ARM_COMPUTE_RETURN_ERROR_ON((output_stage->output_data_type != DataType::QASYMM8) && (output_stage->output_data_type != DataType::QASYMM8_SIGNED)); - ARM_COMPUTE_RETURN_ERROR_ON(output_stage->gemmlowp_max_bound > std::get<1>(quantization::get_min_max_values_from_quantized_data_type(output_stage->output_data_type))); - ARM_COMPUTE_RETURN_ERROR_ON(output_stage->gemmlowp_min_bound < std::get<0>(quantization::get_min_max_values_from_quantized_data_type(output_stage->output_data_type)) - || output_stage->gemmlowp_min_bound > output_stage->gemmlowp_max_bound); + ARM_COMPUTE_RETURN_ERROR_ON((output_stage->output_data_type != DataType::QASYMM8) && + (output_stage->output_data_type != DataType::QASYMM8_SIGNED)); + ARM_COMPUTE_RETURN_ERROR_ON( + output_stage->gemmlowp_max_bound > + std::get<1>(quantization::get_min_max_values_from_quantized_data_type(output_stage->output_data_type))); + ARM_COMPUTE_RETURN_ERROR_ON( + output_stage->gemmlowp_min_bound < + std::get<0>(quantization::get_min_max_values_from_quantized_data_type(output_stage->output_data_type)) || + output_stage->gemmlowp_min_bound > output_stage->gemmlowp_max_bound); // Check biases if exist - if(bias != nullptr) + if (bias != nullptr) { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, bias); ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() > 1); ARM_COMPUTE_RETURN_ERROR_ON(src->dimension(0) != bias->dimension(0)); } - if(dst->total_size() != 0) + if (dst->total_size() != 0) { - ARM_COMPUTE_RETURN_ERROR_ON_MSG(dst->data_type() != output_stage->output_data_type, "Mismatching output data type"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(dst->data_type() != output_stage->output_data_type, + "Mismatching output data type"); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(src, dst); } @@ -77,7 +85,10 @@ ClGemmLowpQuantizeDownInt32ScaleKernel::ClGemmLowpQuantizeDownInt32ScaleKernel() _type = CLKernelType::ELEMENTWISE; } -Status ClGemmLowpQuantizeDownInt32ScaleKernel::validate(const ITensorInfo *src, const ITensorInfo *bias, const ITensorInfo *dst, const GEMMLowpOutputStageInfo *output_stage) +Status ClGemmLowpQuantizeDownInt32ScaleKernel::validate(const ITensorInfo *src, + const ITensorInfo *bias, + const ITensorInfo *dst, + const GEMMLowpOutputStageInfo *output_stage) { ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst); ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, bias, dst, output_stage)); @@ -85,14 +96,17 @@ Status ClGemmLowpQuantizeDownInt32ScaleKernel::validate(const ITensorInfo *src, return Status{}; } -void ClGemmLowpQuantizeDownInt32ScaleKernel::configure(const CLCompileContext &compile_context, const ITensorInfo *src, const ITensorInfo *bias, ITensorInfo *dst, +void ClGemmLowpQuantizeDownInt32ScaleKernel::configure(const CLCompileContext &compile_context, + const ITensorInfo *src, + const ITensorInfo *bias, + ITensorInfo *dst, const GEMMLowpOutputStageInfo *output_stage) { // Perform validate step ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst); ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, bias, dst, output_stage)); - auto padding_info = get_padding_info({ src, bias, dst }); + auto padding_info = get_padding_info({src, bias, dst}); // Output auto inizialitation if not yet initialized auto_init_if_empty(*dst, src->clone()->set_data_type(output_stage->output_data_type)); @@ -104,13 +118,18 @@ void ClGemmLowpQuantizeDownInt32ScaleKernel::configure(const CLCompileContext &c auto max = output_stage->gemmlowp_max_bound; CLBuildOptions build_opts; build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration)); - build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(src->dimension(0) % num_elems_processed_per_iteration)); + build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + + support::cpp11::to_string(src->dimension(0) % num_elems_processed_per_iteration)); build_opts.add_option("-DRESULT_OFFSET=" + support::cpp11::to_string(output_stage->gemmlowp_offset)); build_opts.add_option("-DRESULT_MULT_INT=" + support::cpp11::to_string(output_stage->gemmlowp_multiplier)); build_opts.add_option("-DRESULT_SHIFT=" + support::cpp11::to_string(output_stage->gemmlowp_shift)); - build_opts.add_option_if((min > std::get<0>(quantization::get_min_max_values_from_quantized_data_type(output_stage->output_data_type))) && (min != max), + build_opts.add_option_if((min > std::get<0>(quantization::get_min_max_values_from_quantized_data_type( + output_stage->output_data_type))) && + (min != max), "-DMIN_BOUND=" + support::cpp11::to_string(min)); - build_opts.add_option_if((max < std::get<1>(quantization::get_min_max_values_from_quantized_data_type(output_stage->output_data_type))) && (min != max), + build_opts.add_option_if((max < std::get<1>(quantization::get_min_max_values_from_quantized_data_type( + output_stage->output_data_type))) && + (min != max), "-DMAX_BOUND=" + support::cpp11::to_string(max)); build_opts.add_option("-DOUTPUT_DATA_TYPE=" + get_cl_type_from_data_type(dst->data_type())); build_opts.add_option_if(bias != nullptr, "-DADD_BIAS"); @@ -135,15 +154,17 @@ void ClGemmLowpQuantizeDownInt32ScaleKernel::run_op(ITensorPack &tensors, const ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); - const auto src = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC)); - const auto bias = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_BIAS)); - auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST)); + const auto src = + utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC)); + const auto bias = + utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_BIAS)); + auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST)); Window collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ); Window slice = collapsed.first_slice_window_3D(); unsigned int idx1 = num_arguments_per_3D_tensor(); - if(bias != nullptr) + if (bias != nullptr) { Window biases_slice(slice); biases_slice.set(Window::DimY, Window::Dimension(0, 1, 1)); @@ -157,8 +178,7 @@ void ClGemmLowpQuantizeDownInt32ScaleKernel::run_op(ITensorPack &tensors, const add_3D_tensor_argument(idx, src, slice); add_3D_tensor_argument(idx1, dst, slice); enqueue(queue, *this, slice, lws_hint()); - } - while(collapsed.slide_window_slice_3D(slice)); + } while (collapsed.slide_window_slice_3D(slice)); } } // namespace kernels } // namespace opencl |