From 1009e87127e08a58561373c3f83f22b5f0922641 Mon Sep 17 00:00:00 2001 From: Michalis Spyrou Date: Mon, 27 Jul 2020 12:48:34 +0100 Subject: COMPMID-3385: Async support to CLArithmetic* kernels/functions Pt.2 Signed-off-by: Michalis Spyrou Change-Id: Idc5ac2dd2ba5295c00c88b44a783645327a27e15 Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/3617 Tested-by: Arm Jenkins Reviewed-by: Georgios Pinitas Comments-Addressed: Arm Jenkins --- .../CL/kernels/CLPixelWiseMultiplicationKernel.cpp | 87 ++++++++++++---------- 1 file changed, 48 insertions(+), 39 deletions(-) (limited to 'src/core/CL/kernels/CLPixelWiseMultiplicationKernel.cpp') diff --git a/src/core/CL/kernels/CLPixelWiseMultiplicationKernel.cpp b/src/core/CL/kernels/CLPixelWiseMultiplicationKernel.cpp index 2289316447..95869f7e96 100644 --- a/src/core/CL/kernels/CLPixelWiseMultiplicationKernel.cpp +++ b/src/core/CL/kernels/CLPixelWiseMultiplicationKernel.cpp @@ -29,6 +29,7 @@ #include "arm_compute/core/CL/ICLTensor.h" #include "arm_compute/core/CL/OpenCL.h" #include "arm_compute/core/TensorInfo.h" +#include "arm_compute/core/utils/misc/Cast.h" #include "support/StringSupport.h" namespace arm_compute @@ -142,21 +143,21 @@ CLPixelWiseMultiplicationKernel::CLPixelWiseMultiplicationKernel() { } -void CLPixelWiseMultiplicationKernel::configure(const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output, float scale, +void CLPixelWiseMultiplicationKernel::configure(ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output, float scale, ConvertPolicy overflow_policy, RoundingPolicy rounding_policy, const ActivationLayerInfo &act_info) { configure(CLKernelLibrary::get().get_compile_context(), input1, input2, output, scale, overflow_policy, rounding_policy, act_info); } -void CLPixelWiseMultiplicationKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output, float scale, +void CLPixelWiseMultiplicationKernel::configure(const CLCompileContext &compile_context, ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output, float scale, ConvertPolicy overflow_policy, RoundingPolicy rounding_policy, const ActivationLayerInfo &act_info) { ARM_COMPUTE_ERROR_ON_NULLPTR(input1, input2, output); - ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input1->info(), input2->info(), output->info(), + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input1, input2, output, scale, overflow_policy, rounding_policy, act_info)); // Configure kernel window - auto win_config = validate_and_configure_window(input1->info(), input2->info(), output->info()); + auto win_config = validate_and_configure_window(input1, input2, output); ARM_COMPUTE_ERROR_THROW_ON(win_config.first); _input1 = input1; @@ -179,14 +180,14 @@ void CLPixelWiseMultiplicationKernel::configure(const CLCompileContext &compile_ std::string acc_type; // Check if it has float inputs and output - if(is_data_type_float(input1->info()->data_type()) || is_data_type_float(input2->info()->data_type())) + if(is_data_type_float(input1->data_type()) || is_data_type_float(input2->data_type())) { scale_int = -1; - acc_type = (input1->info()->data_type() == DataType::F32 || input2->info()->data_type() == DataType::F32) ? "float" : "half"; + acc_type = (input1->data_type() == DataType::F32 || input2->data_type() == DataType::F32) ? "float" : "half"; } else { - if(input1->info()->element_size() == 2 || input2->info()->element_size() == 2) + if(input1->element_size() == 2 || input2->element_size() == 2) { // Use 32-bit accumulator for 16-bit input acc_type = "int"; @@ -198,26 +199,26 @@ void CLPixelWiseMultiplicationKernel::configure(const CLCompileContext &compile_ } } - const bool is_quantized = is_data_type_quantized(input1->info()->data_type()); + const bool is_quantized = is_data_type_quantized(input1->data_type()); // Set kernel build options std::string kernel_name = "pixelwise_mul"; CLBuildOptions build_opts; - build_opts.add_option("-DDATA_TYPE_IN1=" + get_cl_type_from_data_type(input1->info()->data_type())); - build_opts.add_option("-DDATA_TYPE_IN2=" + get_cl_type_from_data_type(input2->info()->data_type())); - build_opts.add_option("-DDATA_TYPE_OUT=" + get_cl_type_from_data_type(output->info()->data_type())); + build_opts.add_option("-DDATA_TYPE_IN1=" + get_cl_type_from_data_type(input1->data_type())); + build_opts.add_option("-DDATA_TYPE_IN2=" + get_cl_type_from_data_type(input2->data_type())); + build_opts.add_option("-DDATA_TYPE_OUT=" + get_cl_type_from_data_type(output->data_type())); build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration)); - if(is_quantized && (output->info()->data_type() != DataType::S32)) + if(is_quantized && (output->data_type() != DataType::S32)) { - const UniformQuantizationInfo iq1_info = input1->info()->quantization_info().uniform(); - const UniformQuantizationInfo iq2_info = input2->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 oq_info = output->quantization_info().uniform(); - build_opts.add_option_if(is_data_type_quantized_asymmetric(input1->info()->data_type()), + build_opts.add_option_if(is_data_type_quantized_asymmetric(input1->data_type()), "-DOFFSET_IN1=" + support::cpp11::to_string(iq1_info.offset)); - build_opts.add_option_if(is_data_type_quantized_asymmetric(input2->info()->data_type()), + build_opts.add_option_if(is_data_type_quantized_asymmetric(input2->data_type()), "-DOFFSET_IN2=" + support::cpp11::to_string(iq2_info.offset)); - build_opts.add_option_if(is_data_type_quantized_asymmetric(output->info()->data_type()), + build_opts.add_option_if(is_data_type_quantized_asymmetric(output->data_type()), "-DOFFSET_OUT=" + support::cpp11::to_string(oq_info.offset)); build_opts.add_option("-DSCALE_IN1=" + float_to_string_with_full_precision(iq1_info.scale)); build_opts.add_option("-DSCALE_IN2=" + float_to_string_with_full_precision(iq2_info.scale)); @@ -227,7 +228,7 @@ void CLPixelWiseMultiplicationKernel::configure(const CLCompileContext &compile_ else { kernel_name += (scale_int >= 0) ? "_int" : "_float"; - build_opts.add_option_if_else(overflow_policy == ConvertPolicy::WRAP || is_data_type_float(output->info()->data_type()), "-DWRAP", "-DSATURATE"); + build_opts.add_option_if_else(overflow_policy == ConvertPolicy::WRAP || is_data_type_float(output->data_type()), "-DWRAP", "-DSATURATE"); build_opts.add_option_if_else(rounding_policy == RoundingPolicy::TO_ZERO, "-DROUND=_rtz", "-DROUND=_rte"); build_opts.add_option("-DACC_DATA_TYPE=" + acc_type); if(act_info.enabled()) @@ -266,14 +267,18 @@ Status CLPixelWiseMultiplicationKernel::validate(const ITensorInfo *input1, cons return Status{}; } -void CLPixelWiseMultiplicationKernel::run(const Window &window, cl::CommandQueue &queue) +void CLPixelWiseMultiplicationKernel::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 TensorShape &in_shape1 = _input1->info()->tensor_shape(); - const TensorShape &in_shape2 = _input2->info()->tensor_shape(); - const TensorShape &out_shape = _output->info()->tensor_shape(); + const auto src_0 = utils::cast::polymorphic_downcast(inputs.at(TensorType::ACL_SRC_0)); + const auto src_1 = utils::cast::polymorphic_downcast(inputs.at(TensorType::ACL_SRC_1)); + auto dst = utils::cast::polymorphic_downcast(outputs.at(TensorType::ACL_DST)); + + const TensorShape &in_shape1 = src_0->info()->tensor_shape(); + const TensorShape &in_shape2 = src_1->info()->tensor_shape(); + const TensorShape &out_shape = dst->info()->tensor_shape(); bool can_collapse = true; if(std::min(in_shape1.total_size(), in_shape2.total_size()) > 1) @@ -298,9 +303,9 @@ void CLPixelWiseMultiplicationKernel::run(const Window &window, cl::CommandQueue do { unsigned int idx = 0; - add_3D_tensor_argument(idx, _input1, slice_input1); - add_3D_tensor_argument(idx, _input2, slice_input2); - add_3D_tensor_argument(idx, _output, slice); + add_3D_tensor_argument(idx, src_0, slice_input1); + add_3D_tensor_argument(idx, src_1, slice_input2); + add_3D_tensor_argument(idx, dst, slice); enqueue(queue, *this, slice, lws_hint()); ARM_COMPUTE_UNUSED(collapsed.slide_window_slice_3D(slice_input1)); @@ -311,7 +316,7 @@ void CLPixelWiseMultiplicationKernel::run(const Window &window, cl::CommandQueue BorderSize CLPixelWiseMultiplicationKernel::border_size() const { - const unsigned int replicateSize = _output->info()->dimension(0) - std::min(_input1->info()->dimension(0), _input2->info()->dimension(0)); + const unsigned int replicateSize = _output->dimension(0) - std::min(_input1->dimension(0), _input2->dimension(0)); const unsigned int border = std::min(num_elems_processed_per_iteration - 1U, replicateSize); return BorderSize{ 0, border, 0, 0 }; } @@ -374,18 +379,18 @@ CLComplexPixelWiseMultiplicationKernel::CLComplexPixelWiseMultiplicationKernel() { } -void CLComplexPixelWiseMultiplicationKernel::configure(const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output, const ActivationLayerInfo &act_info) +void CLComplexPixelWiseMultiplicationKernel::configure(ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output, const ActivationLayerInfo &act_info) { configure(CLKernelLibrary::get().get_compile_context(), input1, input2, output, act_info); } -void CLComplexPixelWiseMultiplicationKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output, const ActivationLayerInfo &act_info) +void CLComplexPixelWiseMultiplicationKernel::configure(const CLCompileContext &compile_context, ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output, const ActivationLayerInfo &act_info) { ARM_COMPUTE_ERROR_ON_NULLPTR(input1, input2, output); - ARM_COMPUTE_ERROR_THROW_ON(validate_arguments_complex(input1->info(), input2->info(), output->info(), act_info)); + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments_complex(input1, input2, output, act_info)); // Configure kernel window - auto win_config = validate_and_configure_window_complex(input1->info(), input2->info(), output->info()); + auto win_config = validate_and_configure_window_complex(input1, input2, output); ARM_COMPUTE_ERROR_THROW_ON(win_config.first); _input1 = input1; @@ -415,14 +420,18 @@ Status CLComplexPixelWiseMultiplicationKernel::validate(const ITensorInfo *input return Status{}; } -void CLComplexPixelWiseMultiplicationKernel::run(const Window &window, cl::CommandQueue &queue) +void CLComplexPixelWiseMultiplicationKernel::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 TensorShape &in_shape1 = _input1->info()->tensor_shape(); - const TensorShape &in_shape2 = _input2->info()->tensor_shape(); - const TensorShape &out_shape = _output->info()->tensor_shape(); + const auto src_0 = utils::cast::polymorphic_downcast(inputs.at(TensorType::ACL_SRC_0)); + const auto src_1 = utils::cast::polymorphic_downcast(inputs.at(TensorType::ACL_SRC_1)); + auto dst = utils::cast::polymorphic_downcast(outputs.at(TensorType::ACL_DST)); + + const TensorShape &in_shape1 = src_0->info()->tensor_shape(); + const TensorShape &in_shape2 = src_1->info()->tensor_shape(); + const TensorShape &out_shape = dst->info()->tensor_shape(); bool can_collapse = true; if(std::min(in_shape1.total_size(), in_shape2.total_size()) > 1) @@ -447,9 +456,9 @@ void CLComplexPixelWiseMultiplicationKernel::run(const Window &window, cl::Comma do { unsigned int idx = 0; - add_3D_tensor_argument(idx, _input1, slice_input1); - add_3D_tensor_argument(idx, _input2, slice_input2); - add_3D_tensor_argument(idx, _output, slice); + add_3D_tensor_argument(idx, src_0, slice_input1); + add_3D_tensor_argument(idx, src_1, slice_input2); + add_3D_tensor_argument(idx, dst, slice); enqueue(queue, *this, slice, lws_hint()); ARM_COMPUTE_UNUSED(collapsed.slide_window_slice_3D(slice_input1)); @@ -460,7 +469,7 @@ void CLComplexPixelWiseMultiplicationKernel::run(const Window &window, cl::Comma BorderSize CLComplexPixelWiseMultiplicationKernel::border_size() const { - const unsigned int replicateSize = _output->info()->dimension(0) - std::min(_input1->info()->dimension(0), _input2->info()->dimension(0)); + const unsigned int replicateSize = _output->dimension(0) - std::min(_input1->dimension(0), _input2->dimension(0)); const unsigned int border = std::min(num_elems_processed_per_iteration_complex - 1U, replicateSize); return BorderSize{ 0, border, 0, 0 }; } -- cgit v1.2.1