From f932d2c8409831cb9cb97a2eb65be93ad4709cd6 Mon Sep 17 00:00:00 2001 From: Michele Di Giorgio Date: Mon, 6 Jul 2020 11:27:21 +0100 Subject: COMPMID-3386: Support memory injection in CLConcatenate functions/kernels Signed-off-by: Georgios Pinitas Change-Id: I611adf4f506d406540e920b0bd6befb4b5108918 Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/3601 Tested-by: Arm Jenkins Reviewed-by: Michalis Spyrou Comments-Addressed: Arm Jenkins --- .../kernels/CLWidthConcatenate4TensorsKernel.cpp | 94 +++++++++++----------- 1 file changed, 46 insertions(+), 48 deletions(-) (limited to 'src/core/CL/kernels/CLWidthConcatenate4TensorsKernel.cpp') diff --git a/src/core/CL/kernels/CLWidthConcatenate4TensorsKernel.cpp b/src/core/CL/kernels/CLWidthConcatenate4TensorsKernel.cpp index 84744a68aa..e49bb1c8ce 100644 --- a/src/core/CL/kernels/CLWidthConcatenate4TensorsKernel.cpp +++ b/src/core/CL/kernels/CLWidthConcatenate4TensorsKernel.cpp @@ -32,6 +32,7 @@ #include "arm_compute/core/Utils.h" #include "arm_compute/core/Window.h" #include "arm_compute/core/utils/helpers/tensor_info.h" +#include "arm_compute/core/utils/misc/Cast.h" #include "arm_compute/core/utils/misc/ShapeCalculator.h" #include "support/StringSupport.h" @@ -97,7 +98,6 @@ Status validate_arguments(const ITensorInfo *input1, const ITensorInfo *input2, } // namespace CLWidthConcatenate4TensorsKernel::CLWidthConcatenate4TensorsKernel() - : _input1(nullptr), _input2(nullptr), _input3(nullptr), _input4(nullptr), _output(nullptr) { } @@ -108,42 +108,33 @@ Status CLWidthConcatenate4TensorsKernel::validate(const ITensorInfo *input1, con return Status{}; } -void CLWidthConcatenate4TensorsKernel::configure(const ICLTensor *input1, const ICLTensor *input2, const ICLTensor *input3, const ICLTensor *input4, ICLTensor *output) -{ - configure(CLKernelLibrary::get().get_compile_context(), input1, input2, input3, input4, output); -} - -void CLWidthConcatenate4TensorsKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input1, const ICLTensor *input2, const ICLTensor *input3, const ICLTensor *input4, - ICLTensor *output) +void CLWidthConcatenate4TensorsKernel::configure(const CLCompileContext &compile_context, + ITensorInfo *input1, ITensorInfo *input2, + ITensorInfo *input3, ITensorInfo *input4, + ITensorInfo *output) { ARM_COMPUTE_ERROR_ON_NULLPTR(input1, input2, input3, input4, output); - ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input1->info(), input2->info(), input3->info(), input4->info(), output->info())); - - _input1 = input1; - _input2 = input2; - _input3 = input3; - _input4 = input4; - _output = output; + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input1, input2, input3, input4, output)); // Add build options CLBuildOptions build_opts; - build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input1->info()->data_type())); + build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input1->data_type())); build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration)); - build_opts.add_option("-DDEPTH=" + support::cpp11::to_string(input1->info()->dimension(2))); - build_opts.add_option("-DINPUT1_WIDTH=" + support::cpp11::to_string(input1->info()->dimension(0))); - build_opts.add_option("-DINPUT2_WIDTH=" + support::cpp11::to_string(input2->info()->dimension(0))); - build_opts.add_option("-DINPUT3_WIDTH=" + support::cpp11::to_string(input3->info()->dimension(0))); - build_opts.add_option("-DELEMENT_SIZE=" + support::cpp11::to_string(input1->info()->element_size())); + build_opts.add_option("-DDEPTH=" + support::cpp11::to_string(input1->dimension(2))); + build_opts.add_option("-DINPUT1_WIDTH=" + support::cpp11::to_string(input1->dimension(0))); + build_opts.add_option("-DINPUT2_WIDTH=" + support::cpp11::to_string(input2->dimension(0))); + build_opts.add_option("-DINPUT3_WIDTH=" + support::cpp11::to_string(input3->dimension(0))); + build_opts.add_option("-DELEMENT_SIZE=" + support::cpp11::to_string(input1->element_size())); // If input have different quantization info set quantization parameters needed for the re-quantization process - const bool have_different_qinfo = helpers::tensor_info::tensors_have_different_quantization_info(output->info(), input1->info(), input2->info(), input3->info(), input4->info()); - if(is_data_type_quantized_asymmetric(input1->info()->data_type()) && have_different_qinfo) + const bool have_different_qinfo = helpers::tensor_info::tensors_have_different_quantization_info(output, input1, input2, input3, input4); + if(is_data_type_quantized_asymmetric(input1->data_type()) && have_different_qinfo) { - const UniformQuantizationInfo iq1_info = input1->info()->quantization_info().uniform(); - const UniformQuantizationInfo iq2_info = input2->info()->quantization_info().uniform(); - const UniformQuantizationInfo iq3_info = input3->info()->quantization_info().uniform(); - const UniformQuantizationInfo iq4_info = input4->info()->quantization_info().uniform(); - const UniformQuantizationInfo oq_info = output->info()->quantization_info().uniform(); + const UniformQuantizationInfo iq1_info = input1->quantization_info().uniform(); + const UniformQuantizationInfo iq2_info = input2->quantization_info().uniform(); + const UniformQuantizationInfo iq3_info = input3->quantization_info().uniform(); + const UniformQuantizationInfo iq4_info = input4->quantization_info().uniform(); + const UniformQuantizationInfo oq_info = output->quantization_info().uniform(); build_opts.add_option("-DOFFSET_IN1=" + float_to_string_with_full_precision(iq1_info.offset)); build_opts.add_option("-DSCALE_IN1=" + float_to_string_with_full_precision(iq1_info.scale)); @@ -161,18 +152,18 @@ void CLWidthConcatenate4TensorsKernel::configure(const CLCompileContext &compile _kernel = create_kernel(compile_context, "concatenate_width_x4", build_opts.options()); // Configure kernel window - auto win_config = validate_and_configure_window(input1->info(), input2->info(), input3->info(), input4->info(), output->info()); + auto win_config = validate_and_configure_window(input1, input2, input3, input4, output); ARM_COMPUTE_ERROR_THROW_ON(std::get<0>(win_config)); ICLKernel::configure_internal(std::get<1>(win_config)); // Set output valid region - output->info()->set_valid_region(ValidRegion(Coordinates(), output->info()->tensor_shape())); + output->set_valid_region(ValidRegion(Coordinates(), output->tensor_shape())); // Pass paddings as arguments to the kernel - const unsigned int input1_width = input1->info()->dimension(0); - const unsigned int input2_width = input2->info()->dimension(0); - const unsigned int input3_width = input3->info()->dimension(0); + const unsigned int input1_width = input1->dimension(0); + const unsigned int input2_width = input2->dimension(0); + const unsigned int input3_width = input3->dimension(0); const unsigned int input1_right_padding = ceil_to_multiple(input1_width, num_elems_processed_per_iteration) - input1_width; const unsigned int input2_left_padding = input1_width % num_elems_processed_per_iteration; @@ -192,40 +183,47 @@ void CLWidthConcatenate4TensorsKernel::configure(const CLCompileContext &compile // Set config_id for enabling LWS tuning _config_id = "concatenate_width_x4_"; - _config_id += lower_string(string_from_data_type(input1->info()->data_type())); + _config_id += lower_string(string_from_data_type(input1->data_type())); _config_id += "_"; - _config_id += support::cpp11::to_string(input1->info()->dimension(0)); + _config_id += support::cpp11::to_string(input1->dimension(0)); _config_id += "_"; - _config_id += support::cpp11::to_string(input1->info()->dimension(1)); + _config_id += support::cpp11::to_string(input1->dimension(1)); _config_id += "_"; - _config_id += support::cpp11::to_string(input2->info()->dimension(0)); + _config_id += support::cpp11::to_string(input2->dimension(0)); _config_id += "_"; - _config_id += support::cpp11::to_string(input2->info()->dimension(1)); + _config_id += support::cpp11::to_string(input2->dimension(1)); _config_id += "_"; - _config_id += support::cpp11::to_string(input3->info()->dimension(0)); + _config_id += support::cpp11::to_string(input3->dimension(0)); _config_id += "_"; - _config_id += support::cpp11::to_string(input3->info()->dimension(1)); + _config_id += support::cpp11::to_string(input3->dimension(1)); _config_id += "_"; - _config_id += support::cpp11::to_string(input4->info()->dimension(0)); + _config_id += support::cpp11::to_string(input4->dimension(0)); _config_id += "_"; - _config_id += support::cpp11::to_string(input4->info()->dimension(1)); + _config_id += support::cpp11::to_string(input4->dimension(1)); } -void CLWidthConcatenate4TensorsKernel::run(const Window &window, cl::CommandQueue &queue) +void CLWidthConcatenate4TensorsKernel::run_op(const InputTensorMap &inputs, const OutputTensorMap &outputs, + const Window &window, cl::CommandQueue &queue) { ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); + const auto src0 = utils::cast::polymorphic_downcast(inputs.at(TensorType::ACL_SRC_VEC)); + const auto src1 = utils::cast::polymorphic_downcast(inputs.at(TensorType::ACL_SRC_VEC + 1)); + const auto src2 = utils::cast::polymorphic_downcast(inputs.at(TensorType::ACL_SRC_VEC + 2)); + const auto src3 = utils::cast::polymorphic_downcast(inputs.at(TensorType::ACL_SRC_VEC + 3)); + auto dst = utils::cast::polymorphic_downcast(outputs.at(TensorType::ACL_DST)); + Window slice = window.first_slice_window_4D(); do { unsigned int idx = 0; - add_4D_tensor_argument(idx, _input1, slice); - add_4D_tensor_argument(idx, _input2, slice); - add_4D_tensor_argument(idx, _input3, slice); - add_4D_tensor_argument(idx, _input4, slice); - add_4D_tensor_argument(idx, _output, slice); + add_4D_tensor_argument(idx, src0, slice); + add_4D_tensor_argument(idx, src1, slice); + add_4D_tensor_argument(idx, src2, slice); + add_4D_tensor_argument(idx, src3, slice); + add_4D_tensor_argument(idx, dst, slice); enqueue(queue, *this, window, lws_hint()); } while(window.slide_window_slice_4D(slice)); -- cgit v1.2.1