diff options
author | Sheri Zhang <sheri.zhang@arm.com> | 2020-10-14 13:34:54 +0100 |
---|---|---|
committer | Sheri Zhang <sheri.zhang@arm.com> | 2020-10-28 14:12:33 +0000 |
commit | 5a4284dc7d98a382d0fa492b64fabe430d5afdc6 (patch) | |
tree | c331acbf1bab0220b8609dfa2375ce9e995740fd /src/core/CL/kernels | |
parent | 1b50bd4d74d479f80c48ee1cd7b223245162aac3 (diff) | |
download | ComputeLibrary-5a4284dc7d98a382d0fa492b64fabe430d5afdc6.tar.gz |
COMPMID-3710: Remove OpenCL padding: CLDepthConvertLayerKernel
Signed-off-by: Sheri Zhang <sheri.zhang@arm.com>
Change-Id: Iee1d4655012ce4cb699535697aeefec673f0bc63
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4157
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/core/CL/kernels')
-rw-r--r-- | src/core/CL/kernels/CLDepthConvertLayerKernel.cpp | 22 |
1 files changed, 20 insertions, 2 deletions
diff --git a/src/core/CL/kernels/CLDepthConvertLayerKernel.cpp b/src/core/CL/kernels/CLDepthConvertLayerKernel.cpp index 24f638f8c4..452a14bd29 100644 --- a/src/core/CL/kernels/CLDepthConvertLayerKernel.cpp +++ b/src/core/CL/kernels/CLDepthConvertLayerKernel.cpp @@ -32,6 +32,7 @@ #include "arm_compute/core/Validate.h" #include "src/core/CL/CLValidate.h" #include "src/core/helpers/AutoConfiguration.h" +#include "src/core/helpers/WindowHelpers.h" #include "support/StringSupport.h" #include <cstddef> @@ -81,21 +82,27 @@ void CLDepthConvertLayerKernel::configure(const CLCompileContext &compile_contex { ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); + _input = input; + _output = output; + // Auto initialize output shape if not initialized (We can only auto-configure the shape, datatype must be given) set_shape_if_empty(*output->info(), input->info()->tensor_shape()); ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), policy, shift)); + auto padding_info = get_padding_info({ input, output }); + // Get data sizes const size_t input_size = data_size_from_type(input->info()->data_type()); const size_t output_size = data_size_from_type(output->info()->data_type()); // Get number of elements to process per iterations - constexpr unsigned int num_elems_processed_per_iteration = 16; + const unsigned int num_elems_processed_per_iteration = adjust_vec_size(16 / input->info()->element_size(), input->info()->dimension(0)); // Set build options 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(input->info()->dimension(0) % num_elems_processed_per_iteration)); build_opts.add_option("-DDATA_TYPE_IN=" + get_cl_type_from_data_type(input->info()->data_type())); build_opts.add_option("-DDATA_TYPE_OUT=" + get_cl_type_from_data_type(output->info()->data_type())); // Conversions from float always SATURATE as out-of-bounds conversion from float->integer is implementation defined @@ -111,14 +118,25 @@ void CLDepthConvertLayerKernel::configure(const CLCompileContext &compile_contex unsigned int idx = 2 * num_arguments_per_3D_tensor(); // Skip the input and output parameters _kernel.setArg(idx++, shift); + // Since we have a leftover vector size calculated using the input tensor shape, it is required to + // have the input region equal to the tensor shape + ValidRegion input_valid_region = input->info()->valid_region(); + input->info()->set_valid_region(ValidRegion(Coordinates(0, 0), input->info()->tensor_shape())); + // Configure kernel - ICLSimple2DKernel::configure(input, output, num_elems_processed_per_iteration); + Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration)); + ICLKernel::configure_internal(win); // Collapse window const Window &full_window = window(); Window collapsed_window = full_window.collapse_if_possible(full_window, Window::DimZ); ICLKernel::configure_internal(collapsed_window); + // Restore the valid region + input->info()->set_valid_region(input_valid_region); + + ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info)); + // Set config_id for enabling LWS tuning _config_id = kernel_name; _config_id += "_"; |