From 1167487ea8e54a76d0a3625e0aa84e2ad9ffd317 Mon Sep 17 00:00:00 2001 From: Giorgio Arena Date: Wed, 7 Feb 2018 15:38:12 +0000 Subject: COMPMID-897 Merge batch normalization with bounded relu Change-Id: I9a607fe620f795cdea1a99fdd3f5f8c2fc76f980 Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/119234 Tested-by: Jenkins Reviewed-by: Gian Marco Iodice Reviewed-by: Georgios Pinitas --- .../CL/kernels/CLBatchNormalizationLayerKernel.cpp | 49 +++++++++++++++------- 1 file changed, 33 insertions(+), 16 deletions(-) (limited to 'src/core/CL/kernels/CLBatchNormalizationLayerKernel.cpp') diff --git a/src/core/CL/kernels/CLBatchNormalizationLayerKernel.cpp b/src/core/CL/kernels/CLBatchNormalizationLayerKernel.cpp index 663b044b5d..95487a23db 100644 --- a/src/core/CL/kernels/CLBatchNormalizationLayerKernel.cpp +++ b/src/core/CL/kernels/CLBatchNormalizationLayerKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -42,7 +42,7 @@ namespace Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *mean, const ITensorInfo *var, const ITensorInfo *beta, const ITensorInfo *gamma, - float epsilon) + float epsilon, ActivationLayerInfo act_info) { ARM_COMPUTE_UNUSED(epsilon); ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32); @@ -50,6 +50,14 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, mean, var, beta, gamma); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, mean, var, beta, gamma); ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(2) != mean->dimension(0)); + if(act_info.enabled()) + { + ActivationLayerInfo::ActivationFunction act = act_info.activation(); + ARM_COMPUTE_RETURN_ERROR_ON(input->data_type() != DataType::F32 && input->data_type() != DataType::F16); + ARM_COMPUTE_RETURN_ERROR_ON(act != ActivationLayerInfo::ActivationLayerInfo::ActivationFunction::RELU && act != ActivationLayerInfo::ActivationLayerInfo::ActivationFunction::BOUNDED_RELU + && act != ActivationLayerInfo::ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU); + ARM_COMPUTE_RETURN_ERROR_ON(act_info.b() > act_info.a()); + } if(output != nullptr && output->total_size() != 0) { @@ -98,7 +106,7 @@ CLBatchNormalizationLayerKernel::CLBatchNormalizationLayerKernel() } void CLBatchNormalizationLayerKernel::configure(ICLTensor *input, ICLTensor *output, const ICLTensor *mean, const ICLTensor *var, const ICLTensor *beta, const ICLTensor *gamma, - float epsilon) + float epsilon, ActivationLayerInfo act_info) { ARM_COMPUTE_ERROR_ON_NULLPTR(input, mean, var, beta, gamma); @@ -118,22 +126,22 @@ void CLBatchNormalizationLayerKernel::configure(ICLTensor *input, ICLTensor *out } ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), (output != nullptr) ? output->info() : nullptr, - mean->info(), var->info(), beta->info(), gamma->info(), epsilon)); + mean->info(), var->info(), beta->info(), gamma->info(), epsilon, act_info)); const unsigned int num_elems_processed_per_iteration = 16 / input->info()->element_size(); // Set build options - std::set build_opts; - build_opts.emplace(("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()))); - build_opts.emplace(("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration))); - build_opts.emplace(output == nullptr ? "-DIN_PLACE" : ""); - if(is_data_type_fixed_point(input->info()->data_type())) - { - build_opts.emplace("-DFIXED_POINT_POSITION=" + support::cpp11::to_string(input->info()->fixed_point_position())); - } + CLBuildOptions build_opts; + build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type())); + build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration)); + build_opts.add_option_if(act_info.enabled(), "-D" + string_from_activation_func(act_info.activation())); + build_opts.add_option_if(act_info.enabled(), "-DA_VAL=" + float_to_string_with_full_precision(act_info.a())); + build_opts.add_option_if(act_info.enabled(), "-DB_VAL=" + float_to_string_with_full_precision(act_info.b())); + build_opts.add_option_if(output == nullptr, "-DIN_PLACE"); + build_opts.add_option_if(is_data_type_fixed_point(input->info()->data_type()), "-DFIXED_POINT_POSITION=" + support::cpp11::to_string(input->info()->fixed_point_position())); // Create kernel - _kernel = static_cast(CLKernelLibrary::get().create_kernel("batchnormalization_layer", build_opts)); + _kernel = static_cast(CLKernelLibrary::get().create_kernel("batchnormalization_layer", build_opts.options())); // Set kernel static arguments unsigned int include_output = (output != nullptr) ? 1 : 0; @@ -144,14 +152,23 @@ void CLBatchNormalizationLayerKernel::configure(ICLTensor *input, ICLTensor *out auto win_config = validate_and_configure_window(input->info(), (output == nullptr) ? nullptr : output->info()); ARM_COMPUTE_ERROR_THROW_ON(win_config.first); ICLKernel::configure(win_config.second); + + _config_id = "batch_normalization_layer_"; + _config_id += string_from_data_type(input->info()->data_type()); + _config_id += "_"; + _config_id += support::cpp11::to_string(input->info()->dimension(0)); + _config_id += "_"; + _config_id += support::cpp11::to_string(input->info()->dimension(1)); + _config_id += "_"; + _config_id += support::cpp11::to_string(input->info()->dimension(2)); } Status CLBatchNormalizationLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *mean, const ITensorInfo *var, const ITensorInfo *beta, const ITensorInfo *gamma, - float epsilon) + float epsilon, ActivationLayerInfo act_info) { - ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, mean, var, beta, gamma, epsilon)); + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, mean, var, beta, gamma, epsilon, act_info)); ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), (output == nullptr) ? nullptr : output->clone().get()).first); return Status{}; @@ -182,7 +199,7 @@ void CLBatchNormalizationLayerKernel::run(const Window &window, cl::CommandQueue { add_3D_tensor_argument(idx, _output, slice); } - enqueue(queue, *this, slice); + enqueue(queue, *this, slice, _lws_hint); } while(window.slide_window_slice_3D(slice)); } -- cgit v1.2.1