From b30dcc5ab8eb2bd37f0ab742af1ec45113d54296 Mon Sep 17 00:00:00 2001 From: Gian Marco Iodice Date: Tue, 20 Jun 2017 09:07:21 +0100 Subject: COMPMID-345 - In-place computation for Activation Layer Change-Id: I25ebfccc3d3e758cc8164e0b33805c0bb303891a Reviewed-on: http://mpd-gerrit.cambridge.arm.com/78226 Tested-by: Kaizen Reviewed-by: Pablo Tello Reviewed-by: Georgios Pinitas --- src/core/CL/cl_kernels/activation_layer.cl | 18 +++++-- src/core/CL/kernels/CLActivationLayerKernel.cpp | 67 ++++++++++++++++++++++--- 2 files changed, 73 insertions(+), 12 deletions(-) (limited to 'src/core/CL') diff --git a/src/core/CL/cl_kernels/activation_layer.cl b/src/core/CL/cl_kernels/activation_layer.cl index e3cbb6c801..136191aa22 100644 --- a/src/core/CL/cl_kernels/activation_layer.cl +++ b/src/core/CL/cl_kernels/activation_layer.cl @@ -24,6 +24,8 @@ #include "helpers.h" /** This performs an activation function floating point inputs. + * + * @note In order to perform the activation function "in-place", the pre-processor -DIN_PLACE must be passed at compile time * * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short * @note Activation function should be given as a preprocessor argument using -DNAME. e.g. -DTANH @@ -48,12 +50,20 @@ * @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination image */ __kernel void activation_layer( - TENSOR3D_DECLARATION(input), - TENSOR3D_DECLARATION(output)) + TENSOR3D_DECLARATION(input) +#if !defined IN_PLACE + , + TENSOR3D_DECLARATION(output) +#endif +) { // Get pixels pointer - Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input); + Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input); +#if defined IN_PLACE + Tensor3D output = input; +#else Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output); +#endif // Load data VEC_DATA_TYPE(DATA_TYPE, 16) @@ -63,7 +73,7 @@ __kernel void activation_layer( #if defined LOGISTIC data = 1 / (1 + exp(-data)); #elif defined TANH - data = (VEC_DATA_TYPE(DATA_TYPE, 16))A * tanh((VEC_DATA_TYPE(DATA_TYPE, 16))B * data); + data = (VEC_DATA_TYPE(DATA_TYPE, 16))A * tanh((VEC_DATA_TYPE(DATA_TYPE, 16))B * data); #elif defined RELU data = max(0, data); #elif defined BRELU diff --git a/src/core/CL/kernels/CLActivationLayerKernel.cpp b/src/core/CL/kernels/CLActivationLayerKernel.cpp index 83bbe6a3be..6439426e83 100644 --- a/src/core/CL/kernels/CLActivationLayerKernel.cpp +++ b/src/core/CL/kernels/CLActivationLayerKernel.cpp @@ -35,17 +35,24 @@ using namespace arm_compute; -void CLActivationLayerKernel::configure(const ICLTensor *input, ICLTensor *output, ActivationLayerInfo act_info) +CLActivationLayerKernel::CLActivationLayerKernel() + : _input(nullptr), _output(nullptr) +{ +} + +void CLActivationLayerKernel::configure(ICLTensor *input, ICLTensor *output, ActivationLayerInfo act_info) { ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32); - ARM_COMPUTE_ERROR_ON_NULLPTR(output); - // Output auto inizialitation if not yet initialized - auto_init_if_empty(*output->info(), input->info()->tensor_shape(), 1, input->info()->data_type(), input->info()->fixed_point_position()); + if(output != nullptr) + { + // Output auto inizialitation 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_SHAPES(input, output); - ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); + ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(input, output); + ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); + ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); + } // Set build options std::set build_opts; @@ -54,11 +61,55 @@ void CLActivationLayerKernel::configure(const ICLTensor *input, ICLTensor *outpu build_opts.insert(("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()))); build_opts.insert(("-DA=" + val_to_string(act_info.a()))); build_opts.insert(("-DB=" + val_to_string(act_info.b()))); + build_opts.insert(output == nullptr ? "-DIN_PLACE" : ""); // Create kernel _kernel = static_cast(CLKernelLibrary::get().create_kernel("activation_layer", build_opts)); // Make sure _kernel is initialized before calling the parent's configure constexpr unsigned int num_elems_processed_per_iteration = 16; - ICLSimple3DKernel::configure(input, output, num_elems_processed_per_iteration); + + _input = input; + _output = output; + + // Configure kernel window + Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration)); + + if(output != nullptr) + { + AccessWindowHorizontal output_access(output->info(), 0, num_elems_processed_per_iteration); + + update_window_and_padding(win, + AccessWindowHorizontal(input->info(), 0, num_elems_processed_per_iteration), + output_access); + + output_access.set_valid_region(win, input->info()->valid_region()); + } + else + { + update_window_and_padding(win, + AccessWindowHorizontal(input->info(), 0, num_elems_processed_per_iteration)); + } + + ICLKernel::configure(win); +} + +void CLActivationLayerKernel::run(const Window &window, cl::CommandQueue &queue) +{ + ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); + ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); + + Window slice = window.first_slice_window_3D(); + + do + { + unsigned int idx = 0; + add_3D_tensor_argument(idx, _input, slice); + if(_output != nullptr) + { + add_3D_tensor_argument(idx, _output, slice); + } + enqueue(queue, *this, slice); + } + while(window.slide_window_slice_3D(slice)); } -- cgit v1.2.1