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 ++++++------ src/runtime/CL/functions/CLLSTMLayer.cpp | 52 +++---- .../CL/functions/CLPixelWiseMultiplication.cpp | 152 +++++++++++++++++++-- src/runtime/CL/functions/CLQLSTMLayer.cpp | 40 +++--- 4 files changed, 233 insertions(+), 98 deletions(-) (limited to 'src') 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 }; } diff --git a/src/runtime/CL/functions/CLLSTMLayer.cpp b/src/runtime/CL/functions/CLLSTMLayer.cpp index a1c412401b..058b6027c2 100644 --- a/src/runtime/CL/functions/CLLSTMLayer.cpp +++ b/src/runtime/CL/functions/CLLSTMLayer.cpp @@ -55,7 +55,7 @@ void CLLSTMLayer::configure(const ICLTensor *input, const ICLTensor *input_to_forget_weights, const ICLTensor *input_to_cell_weights, const ICLTensor *input_to_output_weights, const ICLTensor *recurrent_to_forget_weights, const ICLTensor *recurrent_to_cell_weights, const ICLTensor *recurrent_to_output_weights, const ICLTensor *forget_gate_bias, const ICLTensor *cell_bias, const ICLTensor *output_gate_bias, - const ICLTensor *output_state_in, const ICLTensor *cell_state_in, + const ICLTensor *output_state_in, ICLTensor *cell_state_in, ICLTensor *scratch_buffer, ICLTensor *output_state_out, ICLTensor *cell_state_out, ICLTensor *output, const LSTMParams &lstm_params, const ActivationLayerInfo &activation_info, float cell_threshold, float projection_threshold) { @@ -68,7 +68,7 @@ void CLLSTMLayer::configure(const CLCompileContext &compile_context, const ICLTe const ICLTensor *input_to_forget_weights, const ICLTensor *input_to_cell_weights, const ICLTensor *input_to_output_weights, const ICLTensor *recurrent_to_forget_weights, const ICLTensor *recurrent_to_cell_weights, const ICLTensor *recurrent_to_output_weights, const ICLTensor *forget_gate_bias, const ICLTensor *cell_bias, const ICLTensor *output_gate_bias, - const ICLTensor *output_state_in, const ICLTensor *cell_state_in, + const ICLTensor *output_state_in, ICLTensor *cell_state_in, ICLTensor *scratch_buffer, ICLTensor *output_state_out, ICLTensor *cell_state_out, ICLTensor *output, const LSTMParams &lstm_params, const ActivationLayerInfo &activation_info, float cell_threshold, float projection_threshold) { @@ -489,14 +489,14 @@ Status CLLSTMLayer::validate(const ITensorInfo *input, if(lstm_params.has_peephole_opt()) { - ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplicationKernel::validate(cell_state_in, lstm_params.cell_to_forget_weights(), &forget_gate, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN)); + ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplication::validate(cell_state_in, lstm_params.cell_to_forget_weights(), &forget_gate, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN)); ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&forget_gate, &forget_gate, &forget_gate, ConvertPolicy::SATURATE)); } if(lstm_params.use_layer_norm()) { ARM_COMPUTE_RETURN_ON_ERROR(CLMeanStdDevNormalizationLayer::validate(&forget_gate)); - ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplicationKernel::validate(&forget_gate, lstm_params.forget_layer_norm_weights(), &forget_gate, 1, ConvertPolicy::SATURATE, - RoundingPolicy::TO_NEAREST_EVEN)); + ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplication::validate(&forget_gate, lstm_params.forget_layer_norm_weights(), &forget_gate, 1, ConvertPolicy::SATURATE, + RoundingPolicy::TO_NEAREST_EVEN)); ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&forget_gate, forget_gate_bias, &forget_gate, ConvertPolicy::SATURATE)); } ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(&forget_gate, &forget_gate, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC))); @@ -524,14 +524,14 @@ Status CLLSTMLayer::validate(const ITensorInfo *input, { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(lstm_params.cell_to_input_weights()); ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.cell_to_input_weights()->num_dimensions() > 1); - ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplicationKernel::validate(cell_state_in, lstm_params.cell_to_input_weights(), &input_gate, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN)); + ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplication::validate(cell_state_in, lstm_params.cell_to_input_weights(), &input_gate, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN)); ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&input_gate, &input_gate, &input_gate, ConvertPolicy::SATURATE)); } if(lstm_params.use_layer_norm()) { ARM_COMPUTE_RETURN_ON_ERROR(CLMeanStdDevNormalizationLayer::validate(&input_gate)); - ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplicationKernel::validate(&input_gate, lstm_params.input_layer_norm_weights(), &input_gate, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN)); + ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplication::validate(&input_gate, lstm_params.input_layer_norm_weights(), &input_gate, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN)); ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&input_gate, lstm_params.input_gate_bias(), &input_gate, ConvertPolicy::SATURATE)); } ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(&input_gate, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC))); @@ -548,13 +548,13 @@ Status CLLSTMLayer::validate(const ITensorInfo *input, if(lstm_params.use_layer_norm()) { ARM_COMPUTE_RETURN_ON_ERROR(CLMeanStdDevNormalizationLayer::validate(&cell_state_tmp)); - ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplicationKernel::validate(&cell_state_tmp, lstm_params.cell_layer_norm_weights(), &cell_state_tmp, 1, ConvertPolicy::SATURATE, - RoundingPolicy::TO_NEAREST_EVEN)); + ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplication::validate(&cell_state_tmp, lstm_params.cell_layer_norm_weights(), &cell_state_tmp, 1, ConvertPolicy::SATURATE, + RoundingPolicy::TO_NEAREST_EVEN)); ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&cell_state_tmp, cell_bias, &cell_state_tmp, ConvertPolicy::SATURATE)); } ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(&cell_state_tmp, nullptr, activation_info)); - ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplicationKernel::validate(&cell_state_tmp, &input_gate, &cell_state_tmp, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN)); - ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplicationKernel::validate(&cell_state_tmp, &forget_gate, &cell_state_tmp, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN)); + ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplication::validate(&cell_state_tmp, &input_gate, &cell_state_tmp, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN)); + ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplication::validate(&cell_state_tmp, &forget_gate, &cell_state_tmp, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN)); ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&cell_state_tmp, &cell_state_tmp, &cell_state_tmp, ConvertPolicy::SATURATE)); if(cell_threshold != 0.f) { @@ -573,22 +573,22 @@ Status CLLSTMLayer::validate(const ITensorInfo *input, if(lstm_params.has_peephole_opt()) { - ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplicationKernel::validate(&cell_state_tmp, lstm_params.cell_to_output_weights(), &output_gate_tmp, 1, ConvertPolicy::SATURATE, - RoundingPolicy::TO_NEAREST_EVEN)); + ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplication::validate(&cell_state_tmp, lstm_params.cell_to_output_weights(), &output_gate_tmp, 1, ConvertPolicy::SATURATE, + RoundingPolicy::TO_NEAREST_EVEN)); ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&output_gate_tmp, &output_gate_tmp, &output_gate_tmp, ConvertPolicy::SATURATE)); } if(lstm_params.use_layer_norm()) { ARM_COMPUTE_RETURN_ON_ERROR(CLMeanStdDevNormalizationLayer::validate(&output_gate_tmp)); - ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplicationKernel::validate(&output_gate_tmp, lstm_params.output_layer_norm_weights(), &output_gate_tmp, 1, ConvertPolicy::SATURATE, - RoundingPolicy::TO_NEAREST_EVEN)); + ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplication::validate(&output_gate_tmp, lstm_params.output_layer_norm_weights(), &output_gate_tmp, 1, ConvertPolicy::SATURATE, + RoundingPolicy::TO_NEAREST_EVEN)); ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&output_gate_tmp, output_gate_bias, &output_gate_tmp, ConvertPolicy::SATURATE)); } ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(&output_gate_tmp, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC))); // Validate output state ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(&cell_state_tmp, &cell_state_tmp, activation_info)); - ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplicationKernel::validate(&cell_state_tmp, &output_gate_tmp, &output_gate_tmp, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN)); + ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplication::validate(&cell_state_tmp, &output_gate_tmp, &output_gate_tmp, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN)); if(lstm_params.has_projection()) { ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(&output_gate_tmp, lstm_params.projection_weights(), lstm_params.projection_bias(), output_state_out)); @@ -629,13 +629,13 @@ void CLLSTMLayer::run() if(_run_peephole_opt) { - CLScheduler::get().enqueue(_pixelwise_mul_forget_gate); + _pixelwise_mul_forget_gate.run(); _accum_forget_gate1.run(); } if(_is_layer_norm_lstm) { _mean_std_norm_forget_gate.run(); - CLScheduler::get().enqueue(_pixelwise_mul_forget_gate_coeff); + _pixelwise_mul_forget_gate_coeff.run(); _accum_forget_gate_bias.run(); } _activation_forget_gate.run(); @@ -651,14 +651,14 @@ void CLLSTMLayer::run() if(_run_peephole_opt) { - CLScheduler::get().enqueue(_pixelwise_mul_input_gate); + _pixelwise_mul_input_gate.run(); _accum_input_gate1.run(); } if(_is_layer_norm_lstm) { _mean_std_norm_input_gate.run(); - CLScheduler::get().enqueue(_pixelwise_mul_input_gate_coeff); + _pixelwise_mul_input_gate_coeff.run(); _accum_input_gate_bias.run(); } _activation_input_gate.run(); @@ -671,12 +671,12 @@ void CLLSTMLayer::run() if(_is_layer_norm_lstm) { _mean_std_norm_cell_gate.run(); - CLScheduler::get().enqueue(_pixelwise_mul_cell_gate_coeff); + _pixelwise_mul_cell_gate_coeff.run(); _accum_cell_gate_bias.run(); } _activation_cell_state.run(); - CLScheduler::get().enqueue(_pixelwise_mul_cell_state1); - CLScheduler::get().enqueue(_pixelwise_mul_cell_state2); + _pixelwise_mul_cell_state1.run(); + _pixelwise_mul_cell_state2.run(); _accum_cell_state2.run(); if(_perform_cell_clipping) @@ -688,19 +688,19 @@ void CLLSTMLayer::run() if(_run_peephole_opt) { - CLScheduler::get().enqueue(_pixelwise_mul_output_state1); + _pixelwise_mul_output_state1.run(); _accum_output1.run(); } if(_is_layer_norm_lstm) { _mean_std_norm_output_gate.run(); - CLScheduler::get().enqueue(_pixelwise_mul_output_gate_coeff); + _pixelwise_mul_output_gate_coeff.run(); _accum_output_gate_bias.run(); } _activation_output.run(); _activation_output_state.run(); - CLScheduler::get().enqueue(_pixelwise_mul_output_state2); + _pixelwise_mul_output_state2.run(); if(_has_projection_weights) { diff --git a/src/runtime/CL/functions/CLPixelWiseMultiplication.cpp b/src/runtime/CL/functions/CLPixelWiseMultiplication.cpp index 139209c4e0..34e06a3d03 100644 --- a/src/runtime/CL/functions/CLPixelWiseMultiplication.cpp +++ b/src/runtime/CL/functions/CLPixelWiseMultiplication.cpp @@ -25,30 +25,50 @@ #include "arm_compute/core/CL/ICLTensor.h" #include "arm_compute/core/CL/kernels/CLPixelWiseMultiplicationKernel.h" +#include "arm_compute/runtime/CL/CLScheduler.h" #include "support/MemorySupport.h" #include namespace arm_compute { -void CLPixelWiseMultiplication::configure(ICLTensor *input1, ICLTensor *input2, ICLTensor *output, float scale, - ConvertPolicy overflow_policy, RoundingPolicy rounding_policy, const ActivationLayerInfo &act_info) +namespace { - configure(CLKernelLibrary::get().get_compile_context(), input1, input2, output, scale, overflow_policy, rounding_policy, act_info); +void select_border_input(InputTensorMap &tensor_map, InputTensorMap &inputs, OutputTensorMap &outputs) +{ + if(outputs.at(TensorType::ACL_DST)->info()->dimension(0) > 1) + { + if(inputs.at(TensorType::ACL_SRC_1)->info()->dimension(0) == 1) + { + tensor_map[TensorType::ACL_SRC] = inputs.at(TensorType::ACL_SRC_1); + } + else + { + tensor_map[TensorType::ACL_SRC] = inputs.at(TensorType::ACL_SRC_0); + } + } } +} // namespace -void CLPixelWiseMultiplication::configure(const CLCompileContext &compile_context, ICLTensor *input1, ICLTensor *input2, ICLTensor *output, float scale, +namespace experimental +{ +CLPixelWiseMultiplication::CLPixelWiseMultiplication() + : _border_handler() +{ +} + +void CLPixelWiseMultiplication::configure(const CLCompileContext &compile_context, ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output, float scale, ConvertPolicy overflow_policy, RoundingPolicy rounding_policy, const ActivationLayerInfo &act_info) { auto k = arm_compute::support::cpp14::make_unique(); k->configure(compile_context, input1, input2, output, scale, overflow_policy, rounding_policy, act_info); _kernel = std::move(k); - if(output->info()->dimension(0) > 1) + if(output->dimension(0) > 1) { - ICLTensor *broadcasted_info = (input1->info()->dimension(0) == 1) ? input1 : input2; + ITensorInfo *broadcasted_info = (input1->dimension(0) == 1) ? input1 : input2; - if(broadcasted_info->info()->dimension(0) == 1) + if(broadcasted_info->dimension(0) == 1) { _border_handler.configure(compile_context, broadcasted_info, _kernel->border_size(), BorderMode::REPLICATE); } @@ -61,22 +81,30 @@ Status CLPixelWiseMultiplication::validate(const ITensorInfo *input1, const ITen return CLPixelWiseMultiplicationKernel::validate(input1, input2, output, scale, overflow_policy, rounding_policy, act_info); } -void CLComplexPixelWiseMultiplication::configure(ICLTensor *input1, ICLTensor *input2, ICLTensor *output, const ActivationLayerInfo &act_info) +void CLPixelWiseMultiplication::run(InputTensorMap inputs, OutputTensorMap outputs, OperatorTensorMap workspace) { - configure(CLKernelLibrary::get().get_compile_context(), input1, input2, output, act_info); + InputTensorMap src; + select_border_input(src, inputs, outputs); + CLScheduler::get().enqueue_op(_border_handler, src, {}); + ICLOperator::run(inputs, outputs, workspace); } -void CLComplexPixelWiseMultiplication::configure(const CLCompileContext &compile_context, ICLTensor *input1, ICLTensor *input2, ICLTensor *output, const ActivationLayerInfo &act_info) +CLComplexPixelWiseMultiplication::CLComplexPixelWiseMultiplication() + : _border_handler() +{ +} + +void CLComplexPixelWiseMultiplication::configure(const CLCompileContext &compile_context, ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output, const ActivationLayerInfo &act_info) { auto k = arm_compute::support::cpp14::make_unique(); k->configure(compile_context, input1, input2, output, act_info); _kernel = std::move(k); - if(output->info()->dimension(0) > 1) + if(output->dimension(0) > 1) { - ICLTensor *broadcasted_info = (input1->info()->dimension(0) == 1) ? input1 : input2; + ITensorInfo *broadcasted_info = (input1->dimension(0) == 1) ? input1 : input2; - if(broadcasted_info->info()->dimension(0) == 1) + if(broadcasted_info->dimension(0) == 1) { _border_handler.configure(compile_context, broadcasted_info, _kernel->border_size(), BorderMode::REPLICATE); } @@ -87,4 +115,102 @@ Status CLComplexPixelWiseMultiplication::validate(const ITensorInfo *input1, con { return CLComplexPixelWiseMultiplicationKernel::validate(input1, input2, output, act_info); } + +void CLComplexPixelWiseMultiplication::run(InputTensorMap inputs, OutputTensorMap outputs, OperatorTensorMap workspace) +{ + InputTensorMap src; + select_border_input(src, inputs, outputs); + CLScheduler::get().enqueue_op(_border_handler, src, {}); + ICLOperator::run(inputs, outputs, workspace); +} +} // namespace experimental + +struct CLPixelWiseMultiplication::Impl +{ + const ICLTensor *src_0{ nullptr }; + const ICLTensor *src_1{ nullptr }; + ICLTensor *dst{ nullptr }; + std::unique_ptr op{ nullptr }; +}; + +CLPixelWiseMultiplication::CLPixelWiseMultiplication() + : _impl(support::cpp14::make_unique()) +{ +} +CLPixelWiseMultiplication::CLPixelWiseMultiplication(CLPixelWiseMultiplication &&) = default; +CLPixelWiseMultiplication &CLPixelWiseMultiplication::operator=(CLPixelWiseMultiplication &&) = default; +CLPixelWiseMultiplication::~CLPixelWiseMultiplication() = default; + +void CLPixelWiseMultiplication::configure(ICLTensor *input1, ICLTensor *input2, ICLTensor *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 CLPixelWiseMultiplication::configure(const CLCompileContext &compile_context, ICLTensor *input1, ICLTensor *input2, ICLTensor *output, float scale, + ConvertPolicy overflow_policy, RoundingPolicy rounding_policy, const ActivationLayerInfo &act_info) +{ + _impl->src_0 = input1; + _impl->src_1 = input2; + _impl->dst = output; + _impl->op = arm_compute::support::cpp14::make_unique(); + _impl->op->configure(compile_context, input1->info(), input2->info(), output->info(), scale, overflow_policy, rounding_policy, act_info); +} + +Status CLPixelWiseMultiplication::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, float scale, + ConvertPolicy overflow_policy, RoundingPolicy rounding_policy, const ActivationLayerInfo &act_info) +{ + return experimental::CLPixelWiseMultiplication::validate(input1, input2, output, scale, overflow_policy, rounding_policy, act_info); +} + +void CLPixelWiseMultiplication::run() +{ + const InputTensorMap src{ { TensorType::ACL_SRC_0, _impl->src_0 }, { TensorType::ACL_SRC_1, _impl->src_1 } }; + const OutputTensorMap dst{ { TensorType::ACL_DST, _impl->dst } }; + + _impl->op->run(src, dst, {}); +} + +struct CLComplexPixelWiseMultiplication::Impl +{ + const ICLTensor *src_0{ nullptr }; + const ICLTensor *src_1{ nullptr }; + ICLTensor *dst{ nullptr }; + std::unique_ptr op{ nullptr }; +}; + +CLComplexPixelWiseMultiplication::CLComplexPixelWiseMultiplication() + : _impl(support::cpp14::make_unique()) +{ +} +CLComplexPixelWiseMultiplication::CLComplexPixelWiseMultiplication(CLComplexPixelWiseMultiplication &&) = default; +CLComplexPixelWiseMultiplication &CLComplexPixelWiseMultiplication::operator=(CLComplexPixelWiseMultiplication &&) = default; +CLComplexPixelWiseMultiplication::~CLComplexPixelWiseMultiplication() = default; + +void CLComplexPixelWiseMultiplication::configure(ICLTensor *input1, ICLTensor *input2, ICLTensor *output, const ActivationLayerInfo &act_info) +{ + configure(CLKernelLibrary::get().get_compile_context(), input1, input2, output, act_info); +} + +void CLComplexPixelWiseMultiplication::configure(const CLCompileContext &compile_context, ICLTensor *input1, ICLTensor *input2, ICLTensor *output, const ActivationLayerInfo &act_info) +{ + _impl->src_0 = input1; + _impl->src_1 = input2; + _impl->dst = output; + _impl->op = arm_compute::support::cpp14::make_unique(); + _impl->op->configure(compile_context, input1->info(), input2->info(), output->info(), act_info); +} + +Status CLComplexPixelWiseMultiplication::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, const ActivationLayerInfo &act_info) +{ + return experimental::CLComplexPixelWiseMultiplication::validate(input1, input2, output, act_info); +} + +void CLComplexPixelWiseMultiplication::run() +{ + const InputTensorMap src{ { TensorType::ACL_SRC_0, _impl->src_0 }, { TensorType::ACL_SRC_1, _impl->src_1 } }; + const OutputTensorMap dst{ { TensorType::ACL_DST, _impl->dst } }; + + _impl->op->run(src, dst, {}); +} } // namespace arm_compute diff --git a/src/runtime/CL/functions/CLQLSTMLayer.cpp b/src/runtime/CL/functions/CLQLSTMLayer.cpp index c5c4aa3dfa..a40a5d068d 100644 --- a/src/runtime/CL/functions/CLQLSTMLayer.cpp +++ b/src/runtime/CL/functions/CLQLSTMLayer.cpp @@ -113,7 +113,7 @@ void CLQLSTMLayer::configure(const ICLTensor *input, const ICLTensor *input_to_forget_weights, const ICLTensor *input_to_cell_weights, const ICLTensor *input_to_output_weights, const ICLTensor *recurrent_to_forget_weights, const ICLTensor *recurrent_to_cell_weights, const ICLTensor *recurrent_to_output_weights, const ICLTensor *forget_gate_bias, const ICLTensor *cell_bias, const ICLTensor *output_gate_bias, - const ICLTensor *cell_state_in, const ICLTensor *output_state_in, + ICLTensor *cell_state_in, const ICLTensor *output_state_in, ICLTensor *cell_state_out, ICLTensor *output_state_out, ICLTensor *output, const LSTMParams &lstm_params) { @@ -126,7 +126,7 @@ void CLQLSTMLayer::configure(const CLCompileContext &compile_context, const ICLT const ICLTensor *input_to_forget_weights, const ICLTensor *input_to_cell_weights, const ICLTensor *input_to_output_weights, const ICLTensor *recurrent_to_forget_weights, const ICLTensor *recurrent_to_cell_weights, const ICLTensor *recurrent_to_output_weights, const ICLTensor *forget_gate_bias, const ICLTensor *cell_bias, const ICLTensor *output_gate_bias, - const ICLTensor *cell_state_in, const ICLTensor *output_state_in, + ICLTensor *cell_state_in, const ICLTensor *output_state_in, ICLTensor *cell_state_out, ICLTensor *output_state_out, ICLTensor *output, const LSTMParams &lstm_params) { @@ -382,7 +382,7 @@ void CLQLSTMLayer::configure(const CLCompileContext &compile_context, const ICLT input_activation_input->allocator()->allocate(); } // Cell. - // TODO(COMPMID-3396): Perform multiplication in the quantized domain in CLPixelWiseMultiplicationKernel + // TODO(COMPMID-3396): Perform multiplication in the quantized domain in CLPixelWiseMultiplication _pixelwise_mul_forget_cell.configure(compile_context, &_forget_gate, cell_state_in, &_forget_gate, 1.f, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO); const float cell_gate_scale = _cell_gate.info()->quantization_info().uniform().scale; const float mul_input_cell_scale = cell_gate_scale * std::pow(2, 15 + cell_shift); @@ -418,7 +418,7 @@ void CLQLSTMLayer::configure(const CLCompileContext &compile_context, const ICLT if(_has_peephole) { - // TODO(COMPMID-3396): Perform multiplication in the quantized domain in CLPixelWiseMultiplicationKernel + // TODO(COMPMID-3396): Perform multiplication in the quantized domain in CLPixelWiseMultiplication // Here we are not using the output stage because all operations are done in float _mul_cell_to_output_res.allocator()->init(TensorInfo(cell_state_out->info()->tensor_shape(), 1, DataType::S32)); _memory_group.manage(&_mul_cell_to_output_res); @@ -453,7 +453,7 @@ void CLQLSTMLayer::configure(const CLCompileContext &compile_context, const ICLT // Hidden. _hidden_tanh.configure(compile_context, cell_state_out, &_input_gate, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::TANH, 1.f, 1.f)); - // TODO(COMPMID-3396): Perform multiplication in the quantized domain in CLPixelWiseMultiplicationKernel + // TODO(COMPMID-3396): Perform multiplication in the quantized domain in CLPixelWiseMultiplication _memory_group.manage(&_hidden_mul_res); const TensorInfo hidden_mul_res(_input_gate.info()->tensor_shape(), 1, DataType::S32); _hidden_mul_res.allocator()->init(hidden_mul_res); @@ -696,8 +696,8 @@ Status CLQLSTMLayer::validate(const ITensorInfo *input, if(lstm_params.has_peephole_opt()) { ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(lstm_params.cell_to_forget_weights(), 1, DataType::QSYMM16); - ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplicationKernel::validate(cell_state_in, lstm_params.cell_to_forget_weights(), &mm_out_info, 1.f, ConvertPolicy::SATURATE, - RoundingPolicy::TO_ZERO)); + ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplication::validate(cell_state_in, lstm_params.cell_to_forget_weights(), &mm_out_info, 1.f, ConvertPolicy::SATURATE, + RoundingPolicy::TO_ZERO)); const float cell_to_forget_scale = std::pow(2, cell_shift) * lstm_params.cell_to_forget_weights()->quantization_info().uniform().scale / lstm_params.forget_intermediate_scale(); ARM_COMPUTE_RETURN_ON_ERROR(quantization::calculate_quantized_multiplier(cell_to_forget_scale, &gemmlowp_info.gemmlowp_multiplier, &gemmlowp_info.gemmlowp_shift)); ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpOutputStage::validate(&mm_out_info, nullptr, &forget_outstage_info, gemmlowp_info)); @@ -766,8 +766,8 @@ Status CLQLSTMLayer::validate(const ITensorInfo *input, if(lstm_params.has_peephole_opt()) { - ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplicationKernel::validate(cell_state_in, lstm_params.cell_to_input_weights(), &mm_out_info, 1.f, ConvertPolicy::SATURATE, - RoundingPolicy::TO_ZERO)); + ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplication::validate(cell_state_in, lstm_params.cell_to_input_weights(), &mm_out_info, 1.f, ConvertPolicy::SATURATE, + RoundingPolicy::TO_ZERO)); const float cell_to_input_scale = std::pow(2, cell_shift) * lstm_params.cell_to_input_weights()->quantization_info().uniform().scale / lstm_params.input_intermediate_scale(); ARM_COMPUTE_RETURN_ON_ERROR(quantization::calculate_quantized_multiplier(cell_to_input_scale, &gemmlowp_info.gemmlowp_multiplier, &gemmlowp_info.gemmlowp_shift)); ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpOutputStage::validate(&mm_out_info, &eff_bias_info, &input_outstage_info, gemmlowp_info)); @@ -784,8 +784,8 @@ Status CLQLSTMLayer::validate(const ITensorInfo *input, ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(&input_outstage_info, &input_gate_info, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC, 1.f, 1.f))); } // Cell. - ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplicationKernel::validate(&forget_gate_info, cell_state_in, &forget_gate_info, 1.f, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO)); - ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplicationKernel::validate(&input_gate_info, cell_state_in, &cell_gate_info, 1.f, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO)); + ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplication::validate(&forget_gate_info, cell_state_in, &forget_gate_info, 1.f, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO)); + ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplication::validate(&input_gate_info, cell_state_in, &cell_gate_info, 1.f, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO)); ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&forget_gate_info, &cell_gate_info, cell_state_out, ConvertPolicy::SATURATE)); if(quantized_cell_clip > 0) { @@ -809,8 +809,8 @@ Status CLQLSTMLayer::validate(const ITensorInfo *input, // Here we are not using the output stage because all operations are done in float // const float cell_to_output_scale = std::pow(2, cell_shift) * lstm_params.cell_to_output_weights()->quantization_info().uniform().scale / lstm_params.output_intermediate_scale(); // ARM_COMPUTE_RETURN_ON_ERROR(quantization::calculate_quantized_multiplier(cell_to_output_scale, &gemmlowp_info.gemmlowp_multiplier, &gemmlowp_info.gemmlowp_shift)); - ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplicationKernel::validate(cell_state_out, lstm_params.cell_to_output_weights(), &output_outstage_info, 1.f, ConvertPolicy::SATURATE, - RoundingPolicy::TO_ZERO)); + ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplication::validate(cell_state_out, lstm_params.cell_to_output_weights(), &output_outstage_info, 1.f, ConvertPolicy::SATURATE, + RoundingPolicy::TO_ZERO)); ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&output_outstage_info, &output_outstage_info, &output_outstage_info, ConvertPolicy::SATURATE)); } @@ -830,7 +830,7 @@ Status CLQLSTMLayer::validate(const ITensorInfo *input, const TensorInfo hidden_out_info(TensorShape(num_units, batch_size), 1, DataType::QASYMM8_SIGNED); ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.hidden_state_scale() == 0); - ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplicationKernel::validate(&output_gate_info, &input_gate_info, &hidden_mul_res, 1.f, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO)); + ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplication::validate(&output_gate_info, &input_gate_info, &hidden_mul_res, 1.f, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO)); const float hidden_state_scale = std::pow(2, -15) / lstm_params.hidden_state_scale() * std::pow(2, -15); ARM_COMPUTE_RETURN_ON_ERROR(quantization::calculate_quantized_multiplier(hidden_state_scale, &gemmlowp_info.gemmlowp_multiplier, &gemmlowp_info.gemmlowp_shift, /* ignore_epsilon */ true)); gemmlowp_info.gemmlowp_offset = lstm_params.hidden_state_zero(); @@ -926,7 +926,7 @@ void CLQLSTMLayer::run() if(_has_peephole) { - CLScheduler::get().enqueue(_pixelwise_mul_cell_to_forget); + _pixelwise_mul_cell_to_forget.run(); _cell_to_forget_outstage.run(); _accumulate_cell_forget.run(); } @@ -968,7 +968,7 @@ void CLQLSTMLayer::run() if(_has_peephole) { - CLScheduler::get().enqueue(_pixelwise_mul_cell_to_input); + _pixelwise_mul_cell_to_input.run(); _cell_to_input_outstage.run(); _accumulate_cell_input.run(); } @@ -982,8 +982,8 @@ void CLQLSTMLayer::run() } // Cell. - CLScheduler::get().enqueue(_pixelwise_mul_forget_cell); - CLScheduler::get().enqueue(_pixelwise_mul_input_cell); + _pixelwise_mul_forget_cell.run(); + _pixelwise_mul_input_cell.run(); _add_forget_cell.run(); if(_has_cell_clipping) { @@ -998,7 +998,7 @@ void CLQLSTMLayer::run() _accumulate_input_recurrent_output.run(); if(_has_peephole) { - CLScheduler::get().enqueue(_pixelwise_mul_cell_to_output); + _pixelwise_mul_cell_to_output.run(); _cell_to_output_outstage.run(); _accumulate_cell_to_output.run(); } @@ -1012,7 +1012,7 @@ void CLQLSTMLayer::run() // Hidden. _hidden_tanh.run(); - CLScheduler::get().enqueue(_pixelwise_mul_hidden); + _pixelwise_mul_hidden.run(); _hidden_outstage.run(); // Projection. -- cgit v1.2.1