diff options
Diffstat (limited to 'src/core/gpu/cl')
34 files changed, 0 insertions, 4580 deletions
diff --git a/src/core/gpu/cl/ClCompileContext.h b/src/core/gpu/cl/ClCompileContext.h deleted file mode 100644 index e69cc0200f..0000000000 --- a/src/core/gpu/cl/ClCompileContext.h +++ /dev/null @@ -1,36 +0,0 @@ -/* - * Copyright (c) 2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#ifndef ARM_COMPUTE_CL_COMPILE_CONTEXT_H -#define ARM_COMPUTE_CL_COMPILE_CONTEXT_H - -#include "arm_compute/core/CL/CLCompileContext.h" - -namespace arm_compute -{ -namespace opencl -{ -using ClCompileContext = arm_compute::CLCompileContext; -} // namespace opencl -} // namespace arm_compute -#endif /* ARM_COMPUTE_CL_COMPILE_CONTEXT_H */ diff --git a/src/core/gpu/cl/IClKernel.h b/src/core/gpu/cl/IClKernel.h deleted file mode 100644 index 52ea3c9183..0000000000 --- a/src/core/gpu/cl/IClKernel.h +++ /dev/null @@ -1,37 +0,0 @@ -/* - * Copyright (c) 2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#ifndef ARM_COMPUTE_ICL_KERNEL_H -#define ARM_COMPUTE_ICL_KERNEL_H - -#include "arm_compute/core/ITensorInfo.h" -#include "src/core/CL/ICLKernel.h" - -namespace arm_compute -{ -namespace opencl -{ -using IClKernel = arm_compute::ICLKernel; -} // namespace opencl -} // namespace arm_compute -#endif /* ARM_COMPUTE_ICL_KERNEL_H */ diff --git a/src/core/gpu/cl/kernels/ClActivationKernel.cpp b/src/core/gpu/cl/kernels/ClActivationKernel.cpp deleted file mode 100644 index f48425f7cd..0000000000 --- a/src/core/gpu/cl/kernels/ClActivationKernel.cpp +++ /dev/null @@ -1,256 +0,0 @@ -/* - * Copyright (c) 2016-2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#include "src/core/gpu/cl/kernels/ClActivationKernel.h" - -#include "arm_compute/core/CL/CLCoreRuntimeContext.h" -#include "arm_compute/core/CL/CLHelpers.h" -#include "arm_compute/core/CL/ICLTensor.h" -#include "arm_compute/core/TensorInfo.h" -#include "arm_compute/core/Utils.h" -#include "src/core/CL/CLValidate.h" -#include "src/core/helpers/AutoConfiguration.h" -#include "src/core/helpers/WindowHelpers.h" -#include "support/Cast.h" - -#include "support/StringSupport.h" - -#include <set> - -namespace arm_compute -{ -namespace opencl -{ -namespace kernels -{ -namespace -{ -Status validate_arguments(const ITensorInfo *src, const ITensorInfo *dst, const ActivationLayerInfo &act_info) -{ - ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(src); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::QSYMM16, DataType::F16, DataType::F32); - - static std::set<ActivationLayerInfo::ActivationFunction> quantized_supported_activations = - { - ActivationLayerInfo::ActivationFunction::RELU, - ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, - ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, - ActivationLayerInfo::ActivationFunction::LOGISTIC, - ActivationLayerInfo::ActivationFunction::TANH, - ActivationLayerInfo::ActivationFunction::HARD_SWISH, - ActivationLayerInfo::ActivationFunction::LEAKY_RELU, - }; - const DataType data_type = src->data_type(); - const QuantizationInfo &oq_info = (dst != nullptr) ? dst->quantization_info() : src->quantization_info(); - const ActivationLayerInfo::ActivationFunction f_act = act_info.activation(); - - ARM_COMPUTE_RETURN_ERROR_ON_MSG(is_data_type_quantized(data_type) && (quantized_supported_activations.count(f_act) == 0), - "For Quantized data type only hard swish, leaky relu, tanh, logistic, relu and lower/upper bounded relu are supported"); - - ARM_COMPUTE_RETURN_ERROR_ON(data_type == DataType::QASYMM8 && (f_act == ActivationLayerInfo::ActivationFunction::TANH) && (oq_info != QuantizationInfo(1.f / 128.f, 128))); - ARM_COMPUTE_RETURN_ERROR_ON(data_type == DataType::QASYMM8 && (f_act == ActivationLayerInfo::ActivationFunction::LOGISTIC) && (oq_info != QuantizationInfo(1.f / 256.f, 0))); - - ARM_COMPUTE_RETURN_ERROR_ON(is_data_type_quantized_symmetric(data_type) && (f_act == ActivationLayerInfo::ActivationFunction::TANH) && (oq_info != QuantizationInfo(1.f / 32768.f, 0))); - ARM_COMPUTE_RETURN_ERROR_ON(is_data_type_quantized_symmetric(data_type) && (f_act == ActivationLayerInfo::ActivationFunction::LOGISTIC) && (oq_info != QuantizationInfo(1.f / 32768.f, 0))); - - ARM_COMPUTE_RETURN_ERROR_ON(data_type == DataType::QASYMM8_SIGNED && (f_act == ActivationLayerInfo::ActivationFunction::TANH) && (oq_info != QuantizationInfo(1.f / 128.f, 0))); - ARM_COMPUTE_RETURN_ERROR_ON(data_type == DataType::QASYMM8_SIGNED && (f_act == ActivationLayerInfo::ActivationFunction::LOGISTIC) && (oq_info != QuantizationInfo(1.f / 256.f, -128))); - - // Checks performed when destination is configured - if((dst != nullptr) && (dst->total_size() != 0)) - { - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(src, dst); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, dst); - } - - return Status{}; -} -} // namespace - -ClActivationKernel::ClActivationKernel() - : _run_in_place(false) -{ -} - -void ClActivationKernel::configure(const ClCompileContext &compile_context, ITensorInfo *src, ITensorInfo *dst, ActivationLayerInfo act_info) -{ - ARM_COMPUTE_ERROR_ON_NULLPTR(src); - - auto padding_info = get_padding_info({ src, dst }); - - _run_in_place = (dst == nullptr) || (dst == src); - - if(dst != nullptr) - { - // Destination auto inizialitation if not yet initialized - auto_init_if_empty(*dst, *src->clone()); - } - - ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, (dst != nullptr) ? dst : nullptr, act_info)); - - const unsigned int num_elems_processed_per_iteration = adjust_vec_size(16 / src->element_size(), src->dimension(0)); - - const DataType dt = src->data_type(); - float a_const = act_info.a(); - float b_const = act_info.b(); - - const ActivationLayerInfo::ActivationFunction f_act = act_info.activation(); - const bool is_quantized = is_data_type_quantized(dt); - const bool perform_activation_in_float = - (f_act == ActivationLayerInfo::ActivationFunction::LOGISTIC) - || (f_act == ActivationLayerInfo::ActivationFunction::TANH) - || (f_act == ActivationLayerInfo::ActivationFunction::HARD_SWISH) - || (f_act == ActivationLayerInfo::ActivationFunction::LEAKY_RELU); - - // Set build options - CLBuildOptions build_opts; - build_opts.add_option_if(perform_activation_in_float, "-DFLOAT_DOMAIN"); - build_opts.add_option_if(_run_in_place, "-DIN_PLACE"); - build_opts.add_option("-DACT=" + lower_string(string_from_activation_func(f_act))); - build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(dt)); - 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(src->dimension(0) % num_elems_processed_per_iteration)); - - std::string kernel_name = std::string("activation_layer"); - - // Set quantization info build options - if(is_quantized) - { - const UniformQuantizationInfo iq_info = src->quantization_info().uniform(); - - if(!perform_activation_in_float) - { - int a_const_int = 0; - int b_const_int = 0; - - // Create quantized version of constants a, b if needed - switch(dt) - { - case DataType::QASYMM8: - { - a_const_int = quantize_qasymm8(a_const, iq_info); - b_const_int = quantize_qasymm8(b_const, iq_info); - } - break; - case DataType::QASYMM8_SIGNED: - { - a_const_int = quantize_qasymm8_signed(a_const, iq_info); - b_const_int = quantize_qasymm8_signed(b_const, iq_info); - } - break; - case DataType::QSYMM16: - { - a_const_int = quantize_qsymm16(a_const, iq_info); - b_const_int = quantize_qsymm16(b_const, iq_info); - } - break; - default: - break; - } - build_opts.add_option(("-DA_VAL=" + support::cpp11::to_string(a_const_int))); - build_opts.add_option(("-DB_VAL=" + support::cpp11::to_string(b_const_int))); - } - else - { - build_opts.add_option(("-DA_VAL=" + float_to_string_with_full_precision(a_const))); - build_opts.add_option(("-DB_VAL=" + float_to_string_with_full_precision(b_const))); - } - - // Quantized value of 0 corresponds to the offset o1 - build_opts.add_option(("-DCONST_0=" + (is_data_type_quantized_asymmetric(dt) ? support::cpp11::to_string(iq_info.offset) : "0"))); - build_opts.add_option(("-DS1_VAL=" + float_to_string_with_full_precision(iq_info.scale))); - build_opts.add_option_if(is_data_type_quantized_asymmetric(dt), "-DO1_VAL=" + support::cpp11::to_string(iq_info.offset)); - - // Set correct kernel name - kernel_name += perform_activation_in_float ? std::string("_quant_f32") : std::string("_quant"); - - // Set scale and offset of the source and destination if they have different quantization info - if(dst != nullptr) - { - const UniformQuantizationInfo oq_info = dst->quantization_info().uniform(); - - if(iq_info != oq_info) - { - build_opts.add_option(("-DS2_VAL=" + float_to_string_with_full_precision(oq_info.scale))); - build_opts.add_option_if(is_data_type_quantized_asymmetric(dt), "-DO2_VAL=" + support::cpp11::to_string(oq_info.offset)); - } - } - } - else - { - // Set A, B constants in build options for float types - build_opts.add_option(("-DA_VAL=" + float_to_string_with_full_precision(a_const))); - build_opts.add_option(("-DB_VAL=" + float_to_string_with_full_precision(b_const))); - } - - // Create kernel - _kernel = create_kernel(compile_context, kernel_name, build_opts.options()); - - // Configure kernel window - Window win = calculate_max_window(*src, Steps(num_elems_processed_per_iteration)); - ICLKernel::configure_internal(win); - - // Set config_id for enabling LWS tuning - _config_id = "activation_layer_"; - _config_id += lower_string(string_from_data_type(dt)); - _config_id += "_"; - _config_id += support::cpp11::to_string(src->dimension(0)); - _config_id += "_"; - _config_id += support::cpp11::to_string(src->dimension(1)); - - ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info)); -} - -Status ClActivationKernel::validate(const ITensorInfo *src, const ITensorInfo *dst, const ActivationLayerInfo &act_info) -{ - ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, dst, act_info)); - return Status{}; -} - -void ClActivationKernel::run_op(ITensorPack &tensors, const Window &window, ::cl::CommandQueue &queue) -{ - ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); - ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); - - const auto src = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC)); - auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST)); - ARM_COMPUTE_ERROR_ON(_run_in_place && src != dst); - - Window collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ); - Window slice = collapsed.first_slice_window_3D(); - - do - { - unsigned int idx = 0; - add_3D_tensor_argument(idx, src, slice); - if(!_run_in_place) - { - add_3D_tensor_argument(idx, dst, slice); - } - enqueue(queue, *this, slice, lws_hint()); - } - while(collapsed.slide_window_slice_3D(slice)); -} -} // namespace kernels -} // namespace opencl -} // namespace arm_compute diff --git a/src/core/gpu/cl/kernels/ClActivationKernel.h b/src/core/gpu/cl/kernels/ClActivationKernel.h deleted file mode 100644 index 68c309e9e7..0000000000 --- a/src/core/gpu/cl/kernels/ClActivationKernel.h +++ /dev/null @@ -1,74 +0,0 @@ -/* - * Copyright (c) 2016-2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#ifndef ARM_COMPUTE_CL_ACTIVATION_KERNEL_H -#define ARM_COMPUTE_CL_ACTIVATION_KERNEL_H - -#include "src/core/common/Macros.h" -#include "src/core/gpu/cl/ClCompileContext.h" -#include "src/core/gpu/cl/IClKernel.h" - -namespace arm_compute -{ -namespace opencl -{ -namespace kernels -{ -/** Interface for the activation kernel. */ -class ClActivationKernel : public IClKernel -{ -public: - ClActivationKernel(); - ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(ClActivationKernel); - /** Configure kernel for a given list of arguments - * - * @note If the output tensor is a nullptr, the activation function will be performed in-place - * - * @param[in] compile_context The compile context to be used. - * @param[in, out] src Source tensor info. In case of @p dst tensor = nullptr, this tensor will store the result - * of the activation function. Data types supported: QASYMM8/QASYMM8_SIGNED/QSYMM16/F16/F32. - * @param[out] dst Destination tensor info. Data type supported: same as @p src - * @param[in] act_info Activation layer information. - */ - void configure(const ClCompileContext &compile_context, ITensorInfo *src, ITensorInfo *dst, ActivationLayerInfo act_info); - /** Static function to check if given info will lead to a valid configuration of @ref ClActivationKernel - * - * @param[in] src Source tensor info. In case of @p dst tensor info = nullptr, this tensor will store the result - * of the activation function. Data types supported: QASYMM8/QASYMM8_SIGNED/QSYMM16/F16/F32. - * @param[in] dst Destination tensor info. Data type supported: same as @p src - * @param[in] act_info Activation layer information. - * - * @return a status - */ - static Status validate(const ITensorInfo *src, const ITensorInfo *dst, const ActivationLayerInfo &act_info); - - // Inherited methods overridden: - void run_op(ITensorPack &tensors, const Window &window, ::cl::CommandQueue &queue) override; - -private: - bool _run_in_place; -}; -} // namespace kernels -} // namespace opencl -} // namespace arm_compute -#endif /* ARM_COMPUTE_CL_ACTIVATION_KERNEL_H */ diff --git a/src/core/gpu/cl/kernels/ClBatchConcatenateKernel.cpp b/src/core/gpu/cl/kernels/ClBatchConcatenateKernel.cpp deleted file mode 100644 index c16ff1f028..0000000000 --- a/src/core/gpu/cl/kernels/ClBatchConcatenateKernel.cpp +++ /dev/null @@ -1,156 +0,0 @@ -/* - * Copyright (c) 2019-2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#include "src/core/gpu/cl/kernels/ClBatchConcatenateKernel.h" - -#include "arm_compute/core/CL/CLHelpers.h" -#include "arm_compute/core/CL/CLKernelLibrary.h" -#include "arm_compute/core/CL/ICLTensor.h" -#include "arm_compute/core/Helpers.h" -#include "arm_compute/core/Utils.h" -#include "src/core/CL/CLValidate.h" -#include "src/core/helpers/WindowHelpers.h" -#include "support/Cast.h" - -#include "support/StringSupport.h" - -namespace arm_compute -{ -namespace opencl -{ -namespace kernels -{ -namespace -{ -Status validate_arguments(const ITensorInfo *src, unsigned int batch_offset, const ITensorInfo *dst) -{ - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst); - ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(src); - ARM_COMPUTE_RETURN_ERROR_ON(src->data_type() == DataType::UNKNOWN); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, dst); - - ARM_COMPUTE_RETURN_ERROR_ON(src->dimension(Window::DimX) != dst->dimension(Window::DimX)); - ARM_COMPUTE_RETURN_ERROR_ON(src->dimension(Window::DimY) != dst->dimension(Window::DimY)); - ARM_COMPUTE_RETURN_ERROR_ON(src->dimension(Window::DimZ) != dst->dimension(Window::DimZ)); - ARM_COMPUTE_RETURN_ERROR_ON(src->dimension(3) + batch_offset > dst->dimension(3)); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(4, src, dst); - - return Status{}; -} -} // namespace - -ClBatchConcatenateKernel::ClBatchConcatenateKernel() - : _batch_offset(0) -{ -} - -void ClBatchConcatenateKernel::configure(const CLCompileContext &compile_context, ITensorInfo *src, unsigned int batch_offset, ITensorInfo *dst) -{ - ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst); - ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, batch_offset, dst)); - - auto padding_info = get_padding_info({ src, dst }); - - _batch_offset = batch_offset; - - const unsigned int num_elems_processed_per_iteration = adjust_vec_size(16 / src->element_size(), src->dimension(0)); - - // Add build options - CLBuildOptions build_opts; - build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(src->data_type())); - 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(src->dimension(0) % num_elems_processed_per_iteration)); - if(is_data_type_quantized_asymmetric(src->data_type()) && src->quantization_info() != dst->quantization_info()) - { - const UniformQuantizationInfo iq_info = src->quantization_info().uniform(); - const UniformQuantizationInfo oq_info = dst->quantization_info().uniform(); - - build_opts.add_option("-DOFFSET_IN1=" + float_to_string_with_full_precision(iq_info.offset)); - build_opts.add_option("-DOFFSET_OUT=" + float_to_string_with_full_precision(oq_info.offset)); - build_opts.add_option("-DSCALE_IN1=" + float_to_string_with_full_precision(iq_info.scale)); - build_opts.add_option("-DSCALE_OUT=" + float_to_string_with_full_precision(oq_info.scale)); - } - - // Create kernel - _kernel = create_kernel(compile_context, "concatenate", build_opts.options()); - - // Configure kernel window - auto win = calculate_max_window(*dst, Steps(num_elems_processed_per_iteration)); - win.set(3, Window::Dimension(0, src->tensor_shape()[3], 1)); - ICLKernel::configure_internal(win); - - // Set dst valid region - dst->set_valid_region(ValidRegion(Coordinates(), dst->tensor_shape())); - - // Set config_id for enabling LWS tuning - _config_id = "concatenate_"; - _config_id += support::cpp11::to_string(3); - _config_id += "_"; - _config_id += support::cpp11::to_string(batch_offset); - _config_id += "_"; - _config_id += support::cpp11::to_string(src->dimension(0)); - _config_id += "_"; - _config_id += support::cpp11::to_string(src->dimension(1)); - _config_id += "_"; - _config_id += support::cpp11::to_string(src->dimension(2)); - _config_id += "_"; - _config_id += support::cpp11::to_string(src->dimension(3)); - - ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info)); -} - -Status ClBatchConcatenateKernel::validate(const arm_compute::ITensorInfo *src, - unsigned int batch_offset, - const arm_compute::ITensorInfo *dst) -{ - ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, batch_offset, dst)); - return Status{}; -} - -void ClBatchConcatenateKernel::run_op(ITensorPack &tensors, const Window &window, ::cl::CommandQueue &queue) -{ - ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); - ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); - - const auto src = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC)); - auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST)); - - Window slice = window.first_slice_window_3D(); - - const int offset_to_first_elements_in_bytes = _batch_offset * dst->info()->strides_in_bytes()[3]; - - unsigned int idx = 2 * num_arguments_per_3D_tensor(); // Skip the src and dst parameters - _kernel.setArg<cl_int>(idx, offset_to_first_elements_in_bytes); - - do - { - unsigned int idx = 0; - add_3D_tensor_argument(idx, src, slice); - add_3D_tensor_argument(idx, dst, slice); - enqueue(queue, *this, slice, lws_hint()); - } - while(window.slide_window_slice_3D(slice)); -} -} // namespace opencl -} // namespace kernels -} // namespace arm_compute diff --git a/src/core/gpu/cl/kernels/ClBatchConcatenateKernel.h b/src/core/gpu/cl/kernels/ClBatchConcatenateKernel.h deleted file mode 100644 index d9fa905e8e..0000000000 --- a/src/core/gpu/cl/kernels/ClBatchConcatenateKernel.h +++ /dev/null @@ -1,77 +0,0 @@ -/* - * Copyright (c) 2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#ifndef ARM_COMPUTE_CL_BATCH_CONCATENATE_KERNEL_H -#define ARM_COMPUTE_CL_BATCH_CONCATENATE_KERNEL_H - -#include "src/core/common/Macros.h" -#include "src/core/gpu/cl/ClCompileContext.h" -#include "src/core/gpu/cl/IClKernel.h" - -namespace arm_compute -{ -namespace opencl -{ -namespace kernels -{ -/** Interface for the batch concatenate kernel. - * The src tensor will be concatenated into the destination tensor. - */ -class ClBatchConcatenateKernel : public IClKernel -{ -public: - /** Default constructor */ - ClBatchConcatenateKernel(); - ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(ClBatchConcatenateKernel); - /** Initialise the kernel's source and destination - * - * @param[in] compile_context The compile context to be used. - * @param[in] src Source tensor info. Data types supported: All. - * @param[in] batch_offset The offset on axis # 3. - * @param[in,out] dst Destination tensor info. Data types supported: Same as @p src. - * - * @note: The dst tensor's low two dimensions can't be smaller than the src one's. - * @note: The gaps between the two lowest dimensions of src and dst need to be divisible by 2. - * - */ - void configure(const CLCompileContext &compile_context, ITensorInfo *src, unsigned int batch_offset, ITensorInfo *dst); - /** Static function to check if given info will lead to a valid configuration of @ref ClBatchConcatenateKernel - * - * @param[in] src Input tensor info. Data types supported: All. - * @param[in] batch_offset The offset on axis # 3. - * @param[in] dst Destination tensor info. Data types supported: Same as @p src. - * - * @return a status - */ - static Status validate(const ITensorInfo *src, unsigned int batch_offset, const ITensorInfo *dst); - - // Inherited methods overridden: - void run_op(ITensorPack &tensors, const Window &window, ::cl::CommandQueue &queue) override; - -private: - unsigned int _batch_offset; -}; -} // namespace kernels -} // namespace opencl -} // namespace arm_compute -#endif /* ARM_COMPUTE_CL_BATCH_CONCATENATE_KERNEL_H */ diff --git a/src/core/gpu/cl/kernels/ClCopyKernel.cpp b/src/core/gpu/cl/kernels/ClCopyKernel.cpp deleted file mode 100644 index d6c87f8fad..0000000000 --- a/src/core/gpu/cl/kernels/ClCopyKernel.cpp +++ /dev/null @@ -1,170 +0,0 @@ -/* - * Copyright (c) 2018-2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#include "src/core/gpu/cl/kernels/ClCopyKernel.h" - -#include "arm_compute/core/CL/CLHelpers.h" -#include "arm_compute/core/CL/CLKernelLibrary.h" -#include "arm_compute/core/CL/ICLTensor.h" -#include "arm_compute/core/Helpers.h" -#include "arm_compute/core/TensorInfo.h" -#include "arm_compute/core/Utils.h" -#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/Cast.h" -#include "support/StringSupport.h" - -namespace arm_compute -{ -namespace opencl -{ -namespace kernels -{ -namespace -{ -Status validate_arguments(const ITensorInfo *src, const ITensorInfo *dst, Window *dst_window = nullptr) -{ - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst); - - // Validate dst if initialized - if(dst->total_size() != 0) - { - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, dst); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(src, dst); - if(dst_window == nullptr) - { - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(src->tensor_shape(), dst->tensor_shape()); - } - else - { - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(src->tensor_shape(), dst_window->shape()); - } - } - - return Status{}; -} - -} // namespace - -void ClCopyKernel::configure(const CLCompileContext &compile_context, const ITensorInfo *src, ITensorInfo *dst, Window *dst_window) -{ - ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst); - ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, dst, dst_window)); - - auto padding_info = get_padding_info({ src, dst }); - - // Create kernel - CLBuildOptions build_opts; - build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(src->data_type())); - - // Output auto inizialitation if not yet initialized - auto_init_if_empty(*dst, *src); - - // Configure window - const unsigned int vec_size_x = adjust_vec_size(16 / src->element_size(), src->dimension(0)); - - const Window win_config = calculate_max_window(*src, Steps(vec_size_x)); - - if(dst_window != nullptr) - { - _has_dst_window = true; - _dst_window = Window(*dst_window); - const int width_x = dst_window->num_iterations(0); - const int vec_size_x_leftover = width_x % vec_size_x; - const bool multi_access_x = width_x >= static_cast<int32_t>(vec_size_x); - - if(multi_access_x) - { - _dst_window.set(Window::DimX, Window::Dimension(dst_window->x().start(), ceil_to_multiple(dst_window->x().end(), vec_size_x), vec_size_x)); - } - - build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(vec_size_x_leftover)); - } - else - { - const int width_x = src->tensor_shape().x(); - const int vec_size_x_leftover = width_x % vec_size_x; - - build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(vec_size_x_leftover)); - } - - build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(vec_size_x)); - - // Build kernel - _kernel = create_kernel(compile_context, "copy_tensor", build_opts.options()); - - // Validate and set the window - ICLKernel::configure_internal(win_config); - - ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info)); -} - -Status ClCopyKernel::validate(const arm_compute::ITensorInfo *src, const arm_compute::ITensorInfo *dst, Window *dst_window) -{ - ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, dst, dst_window)); - - return Status{}; -} - -void ClCopyKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) -{ - ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); - ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); - - const auto src = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC)); - auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST)); - - Window slice; - - if(_has_dst_window) - { - slice = window.first_slice_window_3D(); - Window out_slice = _dst_window.first_slice_window_3D(); - do - { - unsigned int idx = 0; - add_3D_tensor_argument(idx, src, slice); - add_3D_tensor_argument(idx, dst, out_slice); - enqueue(queue, *this, slice, lws_hint()); - } - while(window.slide_window_slice_3D(slice) && _dst_window.slide_window_slice_3D(out_slice)); - } - else - { - Window collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ); - slice = collapsed.first_slice_window_3D(); - do - { - unsigned int idx = 0; - add_3D_tensor_argument(idx, src, slice); - add_3D_tensor_argument(idx, dst, slice); - enqueue(queue, *this, slice, lws_hint()); - } - while(collapsed.slide_window_slice_3D(slice)); - } -} -} // namespace kernels -} // namespace opencl -} // namespace arm_compute diff --git a/src/core/gpu/cl/kernels/ClCopyKernel.h b/src/core/gpu/cl/kernels/ClCopyKernel.h deleted file mode 100644 index 9d5457d72d..0000000000 --- a/src/core/gpu/cl/kernels/ClCopyKernel.h +++ /dev/null @@ -1,71 +0,0 @@ -/* - * Copyright (c) 2018-2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#ifndef ARM_COMPUTE_CL_COPY_KERNEL_H -#define ARM_COMPUTE_CL_COPY_KERNEL_H - -#include "src/core/common/Macros.h" -#include "src/core/gpu/cl/ClCompileContext.h" -#include "src/core/gpu/cl/IClKernel.h" - -namespace arm_compute -{ -namespace opencl -{ -namespace kernels -{ -/** OpenCL kernel to perform a copy between two tensors */ -class ClCopyKernel : public ICLKernel -{ -public: - ClCopyKernel() = default; - ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(ClCopyKernel); - /** Initialize the kernel's src, dst. - * - * @param[in] compile_context The compile context to be used. - * @param[in] src Source tensor info. Data types supported: All. - * @param[out] dst Destination tensor info. Data types supported: same as @p src. - * @param[in] dst_window (Optional) Window to be used in case only copying into part of a tensor. Default is nullptr. - */ - void configure(const CLCompileContext &compile_context, const ITensorInfo *src, ITensorInfo *dst, Window *dst_window = nullptr); - /** Static function to check if given info will lead to a valid configuration of @ref ClCopyKernel - * - * @param[in] src Source tensor info. Data types supported: All. - * @param[in] dst Destination tensor info. Data types supported: same as @p src. - * @param[in] dst_window (Optional) Window to be used in case only copying into part of a tensor. Default is nullptr. - * - * @return a status - */ - static Status validate(const ITensorInfo *src, const ITensorInfo *dst, Window *dst_window = nullptr); - - // Inherited methods overridden: - void run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) override; - -private: - Window _dst_window{}; - bool _has_dst_window{}; -}; -} // namespace kernels -} // namespace opencl -} // namespace arm_compute -#endif /*ARM_COMPUTE_CL_COPY_KERNEL_H */ diff --git a/src/core/gpu/cl/kernels/ClCropKernel.cpp b/src/core/gpu/cl/kernels/ClCropKernel.cpp deleted file mode 100644 index 20e94b2876..0000000000 --- a/src/core/gpu/cl/kernels/ClCropKernel.cpp +++ /dev/null @@ -1,137 +0,0 @@ -/* - * Copyright (c) 2019-2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#include "src/core/gpu/cl/kernels/ClCropKernel.h" - -#include "arm_compute/core/CL/CLHelpers.h" -#include "arm_compute/core/CL/CLKernelLibrary.h" -#include "arm_compute/core/CL/ICLTensor.h" -#include "arm_compute/core/IAccessWindow.h" -#include "arm_compute/core/TensorInfo.h" -#include "arm_compute/core/Utils.h" -#include "src/core/CPP/Validate.h" -#include "src/core/helpers/WindowHelpers.h" -#include "support/Cast.h" -#include "support/StringSupport.h" - -#include <map> - -namespace arm_compute -{ -namespace opencl -{ -namespace kernels -{ -void ClCropKernel::configure(const ITensorInfo *src, ITensorInfo *dst, Coordinates2D start, Coordinates2D end, uint32_t batch_index, float extrapolation_value, Window *dst_window) -{ - configure(CLKernelLibrary::get().get_compile_context(), src, dst, start, end, batch_index, extrapolation_value, dst_window); -} - -void ClCropKernel::configure(const CLCompileContext &compile_context, const ITensorInfo *src, ITensorInfo *dst, Coordinates2D start, Coordinates2D end, uint32_t batch_index, - float extrapolation_value, Window *dst_window) -{ - ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst); - ARM_COMPUTE_ERROR_THROW_ON(validate(src, dst, start, end, batch_index, extrapolation_value, dst_window)); - - _start = start; - _batch_index = batch_index; - _extrapolation_value = extrapolation_value; - - const int vec_size_x = 4; - // Create and update the window (if needed) - Window win = calculate_max_window(*dst); - - if(dst_window != nullptr) - { - ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(win, *dst_window); - win = *dst_window; - } - - const int dst_width_x = win.num_iterations(0); - const bool multi_access_x = dst_width_x >= vec_size_x; - const bool remainder_x = dst_width_x % vec_size_x > 0; - - if(multi_access_x) - { - win.set(Window::DimX, - Window::Dimension(win.x().start(), ceil_to_multiple(win.x().end(), vec_size_x), vec_size_x)); - } - ICLKernel::configure_internal(win); - - // Create kernel - CLBuildOptions build_opts; - build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(src->data_type())); - build_opts.add_option_if(multi_access_x, "-DVEC_SIZE=" + support::cpp11::to_string(vec_size_x)); - build_opts.add_option_if(multi_access_x && remainder_x, "-DLAST_ACCESSED_X=" + support::cpp11::to_string(std::max<int>(dst_width_x - vec_size_x, 0))); - build_opts.add_option_if(start.x > end.x, "-DWIDTH_FLIPPED="); - build_opts.add_option_if(start.y > end.y, "-DHEIGHT_FLIPPED="); - _kernel = create_kernel(compile_context, "crop_tensor", build_opts.options()); -} - -Status ClCropKernel::validate(const ITensorInfo *src, const ITensorInfo *dst, Coordinates2D start, Coordinates2D end, uint32_t batch_index, float extrapolation_value, Window *dst_window) -{ - ARM_COMPUTE_UNUSED(extrapolation_value, dst_window); - ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(src); - ARM_COMPUTE_RETURN_ERROR_ON(src->data_type() == DataType::UNKNOWN); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_LAYOUT_NOT_IN(src, DataLayout::NHWC); - ARM_COMPUTE_RETURN_ERROR_ON(src->tensor_shape().num_dimensions() > 4); - ARM_COMPUTE_RETURN_ERROR_ON(start.x < 0 || start.y < 0 || end.x < 0 || end.y < 0); - ARM_COMPUTE_RETURN_ERROR_ON(start.x >= static_cast<int32_t>(src->dimension(1)) || start.y >= static_cast<int32_t>(src->dimension(2)) - || end.x >= static_cast<int32_t>(src->dimension(1)) || end.y >= static_cast<int32_t>(src->dimension(2))); - ARM_COMPUTE_RETURN_ERROR_ON(batch_index >= src->dimension(3)); - if(dst_window != nullptr) - { - ARM_COMPUTE_RETURN_ERROR_ON(dst_window->x().step() != 1); - } - if(dst->total_size() > 0) - { - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_NOT_IN(dst, DataType::F32); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(src, dst); - ARM_COMPUTE_RETURN_ERROR_ON(dst->num_dimensions() > 3); - } - return Status{}; -} - -void ClCropKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) -{ - ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); - ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); - - const auto src = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC)); - auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST)); - - Window in_slice = Window(); - in_slice.use_tensor_dimensions(src->info()->tensor_shape()); - in_slice.set(Window::DimX, Window::Dimension(in_slice.x().start(), ceil_to_multiple(in_slice.x().end(), window.x().step()), window.x().step())); - in_slice.set(3, Window::Dimension(_batch_index, _batch_index + 1, 1)); - - unsigned int idx = 0; - add_3D_tensor_argument(idx, src, in_slice); - add_3D_tensor_argument(idx, dst, window); - add_argument(idx, _start.x); - add_argument(idx, _start.y); - enqueue(queue, *this, window, lws_hint()); -} -} // namespace kernels -} // namespace opencl -} // namespace arm_compute diff --git a/src/core/gpu/cl/kernels/ClCropKernel.h b/src/core/gpu/cl/kernels/ClCropKernel.h deleted file mode 100644 index 92f94d471b..0000000000 --- a/src/core/gpu/cl/kernels/ClCropKernel.h +++ /dev/null @@ -1,98 +0,0 @@ -/* - * Copyright (c) 2019-2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#ifndef ARM_COMPUTE_CLCROPKERNEL_H -#define ARM_COMPUTE_CLCROPKERNEL_H - -#include "src/core/common/Macros.h" -#include "src/core/gpu/cl/ClCompileContext.h" -#include "src/core/gpu/cl/IClKernel.h" - -namespace arm_compute -{ -namespace opencl -{ -namespace kernels -{ -/** OpenCL kernel to perform a copy between two tensors */ -class ClCropKernel : public ICLKernel -{ -public: - ClCropKernel() = default; - ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(ClCropKernel); - /** Configure kernel - * - * @note Supported tensor rank: up to 4 - * - * @param[in] src Source tensor info. Data type supported: All. Data layouts supported: NHWC. - * @param[out] dst Destination tensor info. Data type supported: F32 - * @param[in] start Coordinates of where to start cropping the image. - * @param[in] end Coordinates of where to end cropping the image. - * @param[in] batch_index Fourth dimension index of the 3D image to crop in @p src. - * @param[in] extrapolation_value Value to be used for values outside of the image. Default is 0. - * @param[in] dst_window Output window to be used in case cropped image is being copied into a tensor. Default is nullptr. - */ - void configure(const ITensorInfo *src, ITensorInfo *dst, Coordinates2D start, Coordinates2D end, uint32_t batch_index, float extrapolation_value = 0, Window *dst_window = nullptr); - /** Configure kernel - * - * @note Supported tensor rank: up to 4 - * - * @param[in] compile_context The compile context to be used. - * @param[in] src Source tensor info. Data type supported: All. Data layouts supported: NHWC. - * @param[out] dst Destination tensor info. Data type supported: F32 - * @param[in] start Coordinates of where to start cropping the image. - * @param[in] end Coordinates of where to end cropping the image. - * @param[in] batch_index Fourth dimension index of the 3D image to crop in @p src. - * @param[in] extrapolation_value Value to be used for values outside of the image. Default is 0. - * @param[in] dst_window Output window to be used in case cropped image is being copied into a tensor. Default is nullptr. - */ - void configure(const CLCompileContext &compile_context, const ITensorInfo *src, ITensorInfo *dst, Coordinates2D start, Coordinates2D end, uint32_t batch_index, float extrapolation_value = 0, - Window *dst_window = nullptr); - - /** Static function to check if given info will lead to a valid configuration of @ref CLStridedSliceKernel - * - * @note Supported tensor rank: up to 4 - * - * @param[in] src Source tensor info. Data type supported: All. Data layouts supported: NHWC. - * @param[in] dst Destination tensor info. Data type supported: F32 - * @param[in] start Coordinates of where to start cropping the image. - * @param[in] end Coordinates of where to end cropping the image. - * @param[in] batch_index Fourth dimension index of the 3D image to crop in @p src. - * @param[in] extrapolation_value Value to be used for values outside of the image. Default is 0. - * @param[in] dst_window Output window to be used in case cropped image is being copied into a tensor. Default is nullptr. - */ - static Status validate(const ITensorInfo *src, const ITensorInfo *dst, Coordinates2D start, Coordinates2D end, uint32_t batch_index, float extrapolation_value = 0, - Window *dst_window = nullptr); - - // Inherited methods overridden: - void run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) override; - -private: - Coordinates2D _start{}; - uint32_t _batch_index{}; - float _extrapolation_value{}; -}; -} // namespace kernels -} // namespace opencl -} // namespace arm_compute -#endif /*ARM_COMPUTE_CLCROPKERNEL_H */ diff --git a/src/core/gpu/cl/kernels/ClDepthConcatenateKernel.cpp b/src/core/gpu/cl/kernels/ClDepthConcatenateKernel.cpp deleted file mode 100644 index e8893d76d2..0000000000 --- a/src/core/gpu/cl/kernels/ClDepthConcatenateKernel.cpp +++ /dev/null @@ -1,141 +0,0 @@ -/* - * Copyright (c) 2017-2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#include "src/core/gpu/cl/kernels/ClDepthConcatenateKernel.h" - -#include "arm_compute/core/CL/CLHelpers.h" -#include "arm_compute/core/CL/CLKernelLibrary.h" -#include "arm_compute/core/CL/ICLTensor.h" -#include "arm_compute/core/Helpers.h" -#include "arm_compute/core/Utils.h" -#include "src/core/CL/CLValidate.h" -#include "src/core/helpers/WindowHelpers.h" -#include "support/Cast.h" - -#include "support/StringSupport.h" - -namespace arm_compute -{ -namespace opencl -{ -namespace kernels -{ -namespace -{ -Status validate_arguments(const ITensorInfo *src, unsigned int depth_offset, const ITensorInfo *dst) -{ - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst); - ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(src); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::F16, DataType::F32); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, dst); - - ARM_COMPUTE_RETURN_ERROR_ON(src->dimension(Window::DimX) != dst->dimension(Window::DimX)); - ARM_COMPUTE_RETURN_ERROR_ON(src->dimension(Window::DimY) != dst->dimension(Window::DimY)); - ARM_COMPUTE_RETURN_ERROR_ON(src->dimension(2) + depth_offset > dst->dimension(2)); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(3, src, dst); - - return Status{}; -} -} // namespace - -ClDepthConcatenateKernel::ClDepthConcatenateKernel() - : _depth_offset(0) -{ -} - -void ClDepthConcatenateKernel::configure(const CLCompileContext &compile_context, ITensorInfo *src, unsigned int depth_offset, ITensorInfo *dst) -{ - ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst); - ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, depth_offset, dst)); - - auto padding_info = get_padding_info({ src, dst }); - - _depth_offset = depth_offset; - - const unsigned int num_elems_processed_per_iteration = adjust_vec_size(16 / src->element_size(), src->dimension(0)); - - // Add build options - CLBuildOptions build_opts; - build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(src->data_type())); - 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(src->dimension(0) % num_elems_processed_per_iteration)); - if(is_data_type_quantized_asymmetric(src->data_type()) && src->quantization_info() != dst->quantization_info()) - { - const UniformQuantizationInfo iq_info = src->quantization_info().uniform(); - const UniformQuantizationInfo oq_info = dst->quantization_info().uniform(); - - build_opts.add_option("-DOFFSET_IN1=" + float_to_string_with_full_precision(iq_info.offset)); - build_opts.add_option("-DOFFSET_OUT=" + float_to_string_with_full_precision(oq_info.offset)); - build_opts.add_option("-DSCALE_IN1=" + float_to_string_with_full_precision(iq_info.scale)); - build_opts.add_option("-DSCALE_OUT=" + float_to_string_with_full_precision(oq_info.scale)); - } - - // Create kernel - _kernel = create_kernel(compile_context, "concatenate", build_opts.options()); - - // Configure kernel window - auto win = calculate_max_window(*dst, Steps(num_elems_processed_per_iteration)); - win.set(Window::DimZ, Window::Dimension(0, src->tensor_shape().z(), 1)); - ICLKernel::configure_internal(win); - - // Set dst valid region - dst->set_valid_region(ValidRegion(Coordinates(), dst->tensor_shape())); - - ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info)); -} - -Status ClDepthConcatenateKernel::validate(const arm_compute::ITensorInfo *src, - unsigned int depth_offset, - const arm_compute::ITensorInfo *dst) -{ - ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, depth_offset, dst)); - return Status{}; -} - -void ClDepthConcatenateKernel::run_op(ITensorPack &tensors, const Window &window, ::cl::CommandQueue &queue) -{ - ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); - ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); - - const auto src = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC)); - auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST)); - - Window slice = window.first_slice_window_3D(); - - const int offset_to_first_elements_in_bytes = _depth_offset * dst->info()->strides_in_bytes()[2]; - - unsigned int idx = 2 * num_arguments_per_3D_tensor(); // Skip the src and dst parameters - _kernel.setArg<cl_int>(idx, offset_to_first_elements_in_bytes); - - do - { - unsigned int idx = 0; - add_3D_tensor_argument(idx, src, slice); - add_3D_tensor_argument(idx, dst, slice); - enqueue(queue, *this, slice, lws_hint()); - } - while(window.slide_window_slice_3D(slice)); -} -} // namespace kernels -} // namespace opencl -} // namespace arm_compute diff --git a/src/core/gpu/cl/kernels/ClDepthConcatenateKernel.h b/src/core/gpu/cl/kernels/ClDepthConcatenateKernel.h deleted file mode 100644 index 5acfb33199..0000000000 --- a/src/core/gpu/cl/kernels/ClDepthConcatenateKernel.h +++ /dev/null @@ -1,77 +0,0 @@ -/* - * Copyright (c) 2017-2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#ifndef ARM_COMPUTE_CL_DEPTH_CONCATENATE_KERNEL_H -#define ARM_COMPUTE_CL_DEPTH_CONCATENATE_KERNEL_H - -#include "src/core/common/Macros.h" -#include "src/core/gpu/cl/ClCompileContext.h" -#include "src/core/gpu/cl/IClKernel.h" - -namespace arm_compute -{ -namespace opencl -{ -namespace kernels -{ -/** Interface for the depth concatenate kernel. - * The src tensor will be concatenated into the dst tensor. - */ -class ClDepthConcatenateKernel : public ICLKernel -{ -public: - /** Default constructor */ - ClDepthConcatenateKernel(); - ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(ClDepthConcatenateKernel); - /** Initialise the kernel's source and destination - * - * @param[in] compile_context The compile context to be used. - * @param[in] src Source tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32. - * @param[in] depth_offset The offset on the Z axis. - * @param[in,out] dst Destination tensor info. Data types supported: Same as @p src. - * - * @note: The dst tensor's low two dimensions can't be smaller than the src one's. - * @note: The gaps between the two lowest dimensions of src and dst need to be divisible by 2. - * - */ - void configure(const CLCompileContext &compile_context, ITensorInfo *src, unsigned int depth_offset, ITensorInfo *dst); - /** Static function to check if given info will lead to a valid configuration of @ref ClDepthConcatenateKernel - * - * @param[in] src Source tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32 - * @param[in] depth_offset The offset on the Z axis. - * @param[in] dst Destination tensor info. Data types supported: Same as @p src. - * - * @return a status - */ - static Status validate(const ITensorInfo *src, unsigned int depth_offset, const ITensorInfo *dst); - - // Inherited methods overridden: - void run_op(ITensorPack &tensors, const Window &window, ::cl::CommandQueue &queue) override; - -private: - unsigned int _depth_offset; -}; -} // namespace kernels -} // namespace opencl -} // namespace arm_compute -#endif /* ARM_COMPUTE_CL_DEPTH_CONCATENATE_KERNEL_H */ diff --git a/src/core/gpu/cl/kernels/ClElementwiseKernel.cpp b/src/core/gpu/cl/kernels/ClElementwiseKernel.cpp deleted file mode 100644 index 8f12eb2215..0000000000 --- a/src/core/gpu/cl/kernels/ClElementwiseKernel.cpp +++ /dev/null @@ -1,505 +0,0 @@ -/* - * Copyright (c) 2018-2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#include "src/core/gpu/cl/kernels/ClElementwiseKernel.h" - -#include "arm_compute/core/CL/CLHelpers.h" -#include "arm_compute/core/CL/ICLTensor.h" -#include "src/common/utils/Validate.h" -#include "src/core/CL/CLValidate.h" -#include "src/core/helpers/AutoConfiguration.h" -#include "src/core/helpers/WindowHelpers.h" -#include "support/Cast.h" -#include "support/StringSupport.h" -#include <map> - -namespace arm_compute -{ -namespace opencl -{ -namespace kernels -{ -namespace -{ -constexpr unsigned int vector_size_byte_opencl = 16; - -std::map<ArithmeticOperation, std::string> supported_arithmetic_ops = -{ - { ArithmeticOperation::ADD, "ADD" }, - { ArithmeticOperation::SUB, "SUB" }, - { ArithmeticOperation::DIV, "DIV" }, - { ArithmeticOperation::SQUARED_DIFF, "SQUARED_DIFF" }, - { ArithmeticOperation::MIN, "MIN" }, - { ArithmeticOperation::MAX, "MAX" }, - { ArithmeticOperation::POWER, "POWER" }, - { ArithmeticOperation::PRELU, "PRELU" }, -}; - -std::map<ArithmeticOperation, std::string> supported_sat_arithmetic_ops = -{ - { ArithmeticOperation::ADD, "ADD" }, - { ArithmeticOperation::SUB, "SUB" }, -}; - -std::string generate_id_for_tuning_common(const std::string &kernel_name, const ITensorInfo &src1, const ITensorInfo &dst) -{ - std::string config_id; - // Set config_id for enabling LWS tuning - config_id = kernel_name; - config_id += "_"; - config_id += lower_string(string_from_data_type(src1.data_type())); - config_id += "_"; - config_id += support::cpp11::to_string(dst.dimension(0)); - config_id += "_"; - config_id += support::cpp11::to_string(dst.dimension(1)); - return config_id; -} - -Status validate_arguments_with_float_only_supported_rules(const ITensorInfo &src1, const ITensorInfo &src2, const ITensorInfo &dst) -{ - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(&src1, &src2, &dst); - ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(&src1); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&src1, 1, DataType::F16, DataType::F32); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&src1, &src2); - - const TensorShape out_shape = TensorShape::broadcast_shape(src1.tensor_shape(), src2.tensor_shape()); - - ARM_COMPUTE_RETURN_ERROR_ON_MSG(out_shape.total_size() == 0, "Inputs are not broadcast compatible"); - - // Validate in case of configured dst - if(dst.total_size() > 0) - { - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&dst, 1, DataType::F16, DataType::F32); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&src1, &dst); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(detail::have_different_dimensions(out_shape, dst.tensor_shape(), 0), - "Wrong shape for dst"); - } - - return Status{}; -} - -Status validate_arguments_with_arithmetic_rules(const ITensorInfo &src1, const ITensorInfo &src2, const ITensorInfo &dst) -{ - ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(&src1); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&src1, 1, DataType::U8, DataType::QASYMM8, DataType::QASYMM8_SIGNED, - DataType::S16, DataType::QSYMM16, DataType::F16, - DataType::S32, DataType::F32); - ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(&src2); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&src2, 1, DataType::U8, DataType::QASYMM8, DataType::QASYMM8_SIGNED, - DataType::S16, DataType::QSYMM16, DataType::F16, - DataType::S32, DataType::F32); - - const bool is_quantized = is_data_type_quantized(src1.data_type()) || is_data_type_quantized(src2.data_type()); - if(is_quantized) - { - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&src1, &src2); - - if(is_data_type_quantized_symmetric(src1.data_type())) - { - const int32_t in1_offset = src1.quantization_info().uniform().offset; - const int32_t in2_offset = src2.quantization_info().uniform().offset; - ARM_COMPUTE_RETURN_ERROR_ON_MSG(in1_offset != 0, "For quantized symmetric, offset must be zero"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(in2_offset != 0, "For quantized symmetric, offset must be zero"); - } - } - - const TensorShape out_shape = TensorShape::broadcast_shape(src1.tensor_shape(), src2.tensor_shape()); - - ARM_COMPUTE_RETURN_ERROR_ON_MSG(out_shape.total_size() == 0, "Inputs are not broadcast compatible"); - - // Validate in case of configured dst - if(dst.total_size() > 0) - { - ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(&dst); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&dst, 1, DataType::U8, DataType::QASYMM8, DataType::QASYMM8_SIGNED, - DataType::S16, DataType::QSYMM16, DataType::F16, - DataType::S32, DataType::F32); - ARM_COMPUTE_RETURN_ERROR_ON_MSG((dst.data_type() == DataType::U8) && ((src1.data_type() != DataType::U8) || (src2.data_type() != DataType::U8)), - "dst can only be U8 if both inputs are U8"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(detail::have_different_dimensions(out_shape, dst.tensor_shape(), 0), - "Wrong shape for dst"); - - if(is_quantized) - { - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&src1, &dst); - - if(is_data_type_quantized_symmetric(dst.data_type())) - { - const int32_t offset = dst.quantization_info().uniform().offset; - ARM_COMPUTE_RETURN_ERROR_ON_MSG(offset != 0, "For quantized symmetric, offset must be zero"); - } - } - } - return Status{}; -} - -CLBuildOptions generate_build_options_with_arithmetic_rules(const ITensorInfo &src1, const ITensorInfo &src2, const ITensorInfo &dst, const std::string &operation_string) -{ - CLBuildOptions build_opts; - - const unsigned int num_elems_processed_per_iteration = adjust_vec_size(vector_size_byte_opencl / dst.element_size(), dst.dimension(0)); - - build_opts.add_option("-DDATA_TYPE_IN1=" + get_cl_type_from_data_type(src1.data_type())); - build_opts.add_option("-DDATA_TYPE_IN2=" + get_cl_type_from_data_type(src2.data_type())); - build_opts.add_option("-DDATA_TYPE_OUT=" + get_cl_type_from_data_type(dst.data_type())); - build_opts.add_option("-DVEC_SIZE_IN1=" + support::cpp11::to_string(src1.dimension(0) == 1 ? 1 : num_elems_processed_per_iteration)); - build_opts.add_option("-DVEC_SIZE_IN2=" + support::cpp11::to_string(src2.dimension(0) == 1 ? 1 : num_elems_processed_per_iteration)); - build_opts.add_option("-DVEC_SIZE_OUT=" + support::cpp11::to_string(num_elems_processed_per_iteration)); - build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(dst.dimension(0) % num_elems_processed_per_iteration)); - build_opts.add_option("-DOP=" + operation_string); - if(is_data_type_quantized(src1.data_type())) - { - const UniformQuantizationInfo iq1info = src1.quantization_info().uniform(); - const UniformQuantizationInfo iq2info = src2.quantization_info().uniform(); - const UniformQuantizationInfo oqinfo = dst.quantization_info().uniform(); - - build_opts.add_option("-DOFFSET_IN1=" + support::cpp11::to_string(iq1info.offset)); - build_opts.add_option("-DOFFSET_IN2=" + support::cpp11::to_string(iq2info.offset)); - build_opts.add_option("-DOFFSET_OUT=" + support::cpp11::to_string(oqinfo.offset)); - build_opts.add_option("-DSCALE_IN1=" + float_to_string_with_full_precision(iq1info.scale)); - build_opts.add_option("-DSCALE_IN2=" + float_to_string_with_full_precision(iq2info.scale)); - build_opts.add_option("-DSCALE_OUT=" + float_to_string_with_full_precision(oqinfo.scale)); - } - return build_opts; -} - -std::pair<Status, Window> configure_window_arithmetic_common(ITensorInfo &dst) -{ - const unsigned int num_elems_processed_per_iteration = adjust_vec_size(vector_size_byte_opencl / dst.element_size(), dst.dimension(0)); - Window win = calculate_max_window(dst, Steps(num_elems_processed_per_iteration)); - return std::make_pair(Status{}, win); -} - -std::pair<Status, Window> validate_and_configure_window_for_arithmetic_operators(ITensorInfo &src1, ITensorInfo &src2, ITensorInfo &dst) -{ - const std::pair<TensorShape, ValidRegion> broadcast_pair = ITensorInfo::broadcast_shape_and_valid_region(src1, src2); - const TensorShape &out_shape = broadcast_pair.first; - - set_shape_if_empty(dst, out_shape); - - if(src1.data_type() == DataType::S16 || src2.data_type() == DataType::S16) - { - set_format_if_unknown(dst, Format::S16); - } - else if(src1.data_type() == DataType::F16 || src2.data_type() == DataType::F16) - { - set_format_if_unknown(dst, Format::F16); - } - else if(src1.data_type() == DataType::F32 || src2.data_type() == DataType::F32) - { - set_format_if_unknown(dst, Format::F32); - } - else if(src1.data_type() == DataType::QASYMM8 || src2.data_type() == DataType::QASYMM8) - { - set_data_type_if_unknown(dst, DataType::QASYMM8); - } - else if(src1.data_type() == DataType::QASYMM8_SIGNED || src2.data_type() == DataType::QASYMM8_SIGNED) - { - set_data_type_if_unknown(dst, DataType::QASYMM8_SIGNED); - } - else if(src1.data_type() == DataType::QSYMM16 || src2.data_type() == DataType::QSYMM16) - { - set_data_type_if_unknown(dst, DataType::QSYMM16); - } - - return configure_window_arithmetic_common(dst); -} - -std::pair<Status, Window> validate_and_configure_window_for_logical_binary_operators(ITensorInfo &src1, ITensorInfo &src2, ITensorInfo &dst) -{ - const std::pair<TensorShape, ValidRegion> broadcast_pair = ITensorInfo::broadcast_shape_and_valid_region(src1, src2); - const TensorShape &out_shape = broadcast_pair.first; - - set_shape_if_empty(dst, out_shape); - set_data_type_if_unknown(dst, DataType::U8); - - // The arithmetic utility functions can be share - return configure_window_arithmetic_common(dst); -} - -std::pair<Status, Window> validate_and_configure_window_for_division(ITensorInfo &src1, ITensorInfo &src2, ITensorInfo &dst) -{ - const std::pair<TensorShape, ValidRegion> broadcast_pair = ITensorInfo::broadcast_shape_and_valid_region(src1, src2); - const TensorShape &out_shape = broadcast_pair.first; - auto_init_if_empty(dst, out_shape, 1, src1.data_type()); - return configure_window_arithmetic_common(dst); -} -} // namespace - -void ClElementwiseKernel::configure_common(ITensorInfo *src1, ITensorInfo *src2, ITensorInfo *dst) -{ - configure_common(CLKernelLibrary::get().get_compile_context(), src1, src2, dst); -} - -void ClElementwiseKernel::configure_common(const ClCompileContext &compile_context, ITensorInfo *src1, ITensorInfo *src2, ITensorInfo *dst) -{ - // Configure kernel window - auto win_config = validate_and_configure_window(*src1, *src2, *dst); - ARM_COMPUTE_ERROR_THROW_ON(win_config.first); - - _src1 = src1; - _src2 = src2; - _dst = dst; - - std::string kernel_name = "elementwise_operation_" + name(); - if(is_data_type_quantized(src1->data_type())) - { - kernel_name += "_quantized"; - } - - // Set kernel build options - CLBuildOptions build_opts = generate_build_options(*src1, *src2, *dst); - if(_act_info.enabled()) - { - build_opts.add_option("-DACTIVATION_TYPE=" + lower_string(string_from_activation_func(_act_info.activation()))); - build_opts.add_option("-DA_VAL=" + float_to_string_with_full_precision(_act_info.a())); - build_opts.add_option("-DB_VAL=" + float_to_string_with_full_precision(_act_info.b())); - } - - // Create kernel - _kernel = create_kernel(compile_context, kernel_name, build_opts.options()); - - ICLKernel::configure_internal(win_config.second); - - _config_id = generate_id_for_tuning(kernel_name, *src1, *dst); -} - -void ClElementwiseKernel::run_op(ITensorPack &tensors, const Window &window, ::cl::CommandQueue &queue) -{ - ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); - ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); - - const auto src_0 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_0)); - const auto src_1 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_1)); - auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(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; - const bool is_vector = in_shape1.num_dimensions() == 1 || in_shape2.num_dimensions() == 1; - if(std::min(in_shape1.total_size(), in_shape2.total_size()) > 1 && !is_vector) - { - can_collapse = (std::min(in_shape1.num_dimensions(), in_shape2.num_dimensions()) > Window::DimZ); - for(size_t d = Window::DimZ; can_collapse && (d < out_shape.num_dimensions()); d++) - { - can_collapse = (in_shape1[d] == in_shape2[d]); - } - } - - bool has_collapsed = false; - Window collapsed = can_collapse ? window.collapse_if_possible(ICLKernel::window(), Window::DimZ, &has_collapsed) : window; - - const TensorShape &in_shape1_collapsed = has_collapsed ? in_shape1.collapsed_from(Window::DimZ) : in_shape1; - const TensorShape &in_shape2_collapsed = has_collapsed ? in_shape2.collapsed_from(Window::DimZ) : in_shape2; - - Window slice = collapsed.first_slice_window_3D(); - Window slice_src1 = slice.broadcast_if_dimension_le_one(in_shape1_collapsed); - Window slice_src2 = slice.broadcast_if_dimension_le_one(in_shape2_collapsed); - do - { - unsigned int idx = 0; - add_3D_tensor_argument(idx, src_0, slice_src1); - add_3D_tensor_argument(idx, src_1, slice_src2); - add_3D_tensor_argument(idx, dst, slice); - - enqueue(queue, *this, slice, lws_hint()); - ARM_COMPUTE_UNUSED(collapsed.slide_window_slice_3D(slice_src1)); - ARM_COMPUTE_UNUSED(collapsed.slide_window_slice_3D(slice_src2)); - } - while(collapsed.slide_window_slice_3D(slice)); -} - -/** Logical binary */ - -void ClLogicalBinaryKernel::configure(const ClCompileContext &compile_context, LogicalOperation op, ITensorInfo *src1, ITensorInfo *src2, ITensorInfo *dst) -{ - ARM_COMPUTE_ERROR_ON_NULLPTR(src1, src2, dst); - ARM_COMPUTE_ERROR_THROW_ON(ClLogicalBinaryKernel::validate(op, src1, src2, dst)); - _op = op; - configure_common(compile_context, src1, src2, dst); -} - -Status ClLogicalBinaryKernel::validate(LogicalOperation op, const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst) -{ - ARM_COMPUTE_UNUSED(op); - ARM_COMPUTE_ASSERT(op != LogicalOperation::Unknown && op != LogicalOperation::Not); - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src1, src2, dst); - - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src1, 1, DataType::U8); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src1, src2); - - ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_with_arithmetic_rules(*src1, *src2, *dst)); - ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window_for_logical_binary_operators(*src1->clone(), *src2->clone(), *dst->clone()).first); - - return Status{}; -} - -std::string ClLogicalBinaryKernel::name() -{ - switch(_op) - { - case LogicalOperation::And: - return "AND"; - case LogicalOperation::Or: - return "OR"; - case LogicalOperation::Not: - /* fall through */ - default: - ARM_COMPUTE_ASSERT(true); - } - return ""; -} - -std::pair<Status, Window> ClLogicalBinaryKernel::validate_and_configure_window(ITensorInfo &src1, ITensorInfo &src2, ITensorInfo &dst) -{ - return validate_and_configure_window_for_logical_binary_operators(src1, src2, dst); -} - -CLBuildOptions ClLogicalBinaryKernel::generate_build_options(const ITensorInfo &src1, const ITensorInfo &src2, const ITensorInfo &dst) -{ - // The arithmetic utility functions can be share - return generate_build_options_with_arithmetic_rules(src1, src2, dst, name()); -} - -std::string ClLogicalBinaryKernel::generate_id_for_tuning(const std::string &kernel_name, const ITensorInfo &src1, const ITensorInfo &dst) -{ - return generate_id_for_tuning_common(kernel_name, src1, dst); -} - -/** Arithmetic operations with saturation*/ -void ClSaturatedArithmeticKernel::configure(const ClCompileContext &compile_context, ArithmeticOperation op, ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output, - const ConvertPolicy &policy, - const ActivationLayerInfo &act_info) -{ - ARM_COMPUTE_ERROR_ON_NULLPTR(input1, input2, output); - ARM_COMPUTE_ERROR_THROW_ON(ClSaturatedArithmeticKernel::validate(op, input1, input2, output, policy, act_info)); - auto padding_info = get_padding_info({ input1, input2, output }); - - _policy = policy; - _op = op; - _act_info = act_info; - configure_common(compile_context, input1, input2, output); - ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info)); -} - -Status ClSaturatedArithmeticKernel::validate(ArithmeticOperation op, const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, const ConvertPolicy &policy, - const ActivationLayerInfo &act_info) -{ - ARM_COMPUTE_UNUSED(op, policy); - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input1, input2, output); - ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_with_arithmetic_rules(*input1, *input2, *output)); - ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window_for_arithmetic_operators(*input1->clone(), *input2->clone(), *output->clone()).first); - ARM_COMPUTE_RETURN_ERROR_ON(act_info.enabled() && !is_data_type_float(output->data_type())); - - return Status{}; -} - -std::pair<Status, Window> ClSaturatedArithmeticKernel::validate_and_configure_window(ITensorInfo &input1, ITensorInfo &input2, ITensorInfo &output) -{ - return validate_and_configure_window_for_arithmetic_operators(input1, input2, output); -} - -CLBuildOptions ClSaturatedArithmeticKernel::generate_build_options(const ITensorInfo &input1, const ITensorInfo &input2, const ITensorInfo &output) -{ - const bool has_float_out = is_data_type_float(output.data_type()); - auto build_options = generate_build_options_with_arithmetic_rules(input1, input2, output, name()); - build_options.add_option((_policy == ConvertPolicy::WRAP || has_float_out) ? "-DWRAP" : "-DSATURATE"); - return build_options; -} - -std::string ClSaturatedArithmeticKernel::generate_id_for_tuning(const std::string &kernel_name, const ITensorInfo &input1, const ITensorInfo &output) -{ - auto config_id = generate_id_for_tuning_common(kernel_name, input1, output); - config_id += (_policy == ConvertPolicy::WRAP) ? "_wrap_" : "_saturate_"; - config_id += lower_string(string_from_data_layout(input1.data_layout())); - return config_id; -} - -std::string ClSaturatedArithmeticKernel::name() -{ - return supported_sat_arithmetic_ops[_op]; -} - -/** Arithmetic operations*/ -void ClArithmeticKernel::configure(const ClCompileContext &compile_context, ArithmeticOperation op, ITensorInfo *src1, ITensorInfo *src2, ITensorInfo *dst, - const ActivationLayerInfo &act_info) -{ - ARM_COMPUTE_ERROR_ON_NULLPTR(src1, src2, dst); - ARM_COMPUTE_ERROR_THROW_ON(ClArithmeticKernel::validate(op, src1, src2, dst, act_info)); - auto padding_info = get_padding_info({ src1, src2, dst }); - - _op = op; - _act_info = act_info; - configure_common(compile_context, src1, src2, dst); - ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info)); -} - -Status ClArithmeticKernel::validate(ArithmeticOperation op, const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst, const ActivationLayerInfo &act_info) -{ - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src1, src2, dst); - if(op == ArithmeticOperation::DIV || op == ArithmeticOperation::POWER) - { - // Division and Power operators don't support integer arithmetic - ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_with_float_only_supported_rules(*src1, *src2, *dst)); - ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window_for_division(*src1->clone(), *src2->clone(), *dst->clone()).first); - } - else - { - ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_with_arithmetic_rules(*src1, *src2, *dst)); - ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window_for_arithmetic_operators(*src1->clone(), *src2->clone(), *dst->clone()).first); - } - ARM_COMPUTE_RETURN_ERROR_ON(act_info.enabled() && !is_data_type_float(dst->data_type())); - - return Status{}; -} -std::pair<Status, Window> ClArithmeticKernel::validate_and_configure_window(ITensorInfo &src1, ITensorInfo &src2, ITensorInfo &dst) -{ - if(_op == ArithmeticOperation::DIV || _op == ArithmeticOperation::POWER) - { - // Division and Power operators don't support integer arithmetic - return validate_and_configure_window_for_division(src1, src2, dst); - } - else - { - return validate_and_configure_window_for_arithmetic_operators(src1, src2, dst); - } -} - -CLBuildOptions ClArithmeticKernel::generate_build_options(const ITensorInfo &src1, const ITensorInfo &src2, const ITensorInfo &dst) -{ - return generate_build_options_with_arithmetic_rules(src1, src2, dst, name()); -} -std::string ClArithmeticKernel::generate_id_for_tuning(const std::string &kernel_name, const ITensorInfo &src1, const ITensorInfo &dst) -{ - return generate_id_for_tuning_common(kernel_name, src1, dst); -} - -std::string ClArithmeticKernel::name() -{ - return supported_arithmetic_ops[_op]; -} -} // namespace kernels -} // namespace opencl -} // namespace arm_compute diff --git a/src/core/gpu/cl/kernels/ClElementwiseKernel.h b/src/core/gpu/cl/kernels/ClElementwiseKernel.h deleted file mode 100644 index 4ed8ae73ab..0000000000 --- a/src/core/gpu/cl/kernels/ClElementwiseKernel.h +++ /dev/null @@ -1,219 +0,0 @@ -/* - * Copyright (c) 2018-2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#ifndef ARM_COMPUTE_CL_ELEMENTWISE_KERNEL_H -#define ARM_COMPUTE_CL_ELEMENTWISE_KERNEL_H - -#include "src/core/KernelTypes.h" -#include "src/core/common/Macros.h" -#include "src/core/gpu/cl/ClCompileContext.h" -#include "src/core/gpu/cl/IClKernel.h" - -namespace arm_compute -{ -namespace opencl -{ -namespace kernels -{ -/** Interface for an element-wise operation kernel - * - * Element-wise operation is computed by: - * @f[ dst(x,y) = OP(src1(x,y), src2(x,y))@f] - * - */ -class ClElementwiseKernel : public IClKernel -{ -public: - /** Default constructor */ - ClElementwiseKernel() = default; - ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(ClElementwiseKernel); - - // Inherited methods overridden: - void run_op(ITensorPack &tensors, const Window &window, ::cl::CommandQueue &queue) override; - -protected: - /** The name of the operation */ - virtual std::string name() = 0; - - /** Configure kernel for a given list of arguments - * - * @param[in] src1 First source tensor info. Data types supported: U8/S8/QASYMM8/QASYMM8_SIGNED/U16/S16/F16/U32/S32/F32. - * @param[in] src2 Second source tensor info. Data types supported: same as @p src1. - * @param[in] dst Destination tensor info. Data types supported: same as @p src1. - * - * @return a pair of Status and Window - */ - virtual std::pair<Status, Window> validate_and_configure_window(ITensorInfo &src1, ITensorInfo &src2, ITensorInfo &dst) = 0; - - /** Generate the build options for the specific kernel - * - * @reutrn a CLBuildOptions struct - */ - virtual CLBuildOptions generate_build_options(const ITensorInfo &src1, const ITensorInfo &src2, const ITensorInfo &dst) = 0; - - /** Generate the identifier for tuning - * - * @reutrn a string - */ - virtual std::string generate_id_for_tuning(const std::string &kernel_name, const ITensorInfo &src1, const ITensorInfo &dst) = 0; - - /** Commmon configure function for element-wise operators with no additional options (e.g., Div, Min, Max, SquaredDiff) - * - */ - void configure_common(ITensorInfo *src1, ITensorInfo *src2, ITensorInfo *dst); - /** Commmon configure function for element-wise operators with no additional options (e.g., Div, Min, Max, SquaredDiff) - * - */ - void configure_common(const ClCompileContext &compile_context, ITensorInfo *src1, ITensorInfo *src2, ITensorInfo *dst); - - ActivationLayerInfo _act_info{}; - -private: - const ITensorInfo *_src1{ nullptr }; /**< Source tensor info 1 */ - const ITensorInfo *_src2{ nullptr }; /**< Source tensor info 2 */ - ITensorInfo *_dst{ nullptr }; /**< Destination tensor info */ -}; - -class ClLogicalBinaryKernel : public ClElementwiseKernel -{ -public: - /** Default constructor */ - ClLogicalBinaryKernel() = default; - ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(ClLogicalBinaryKernel); - /** Function to configure kernel - * - * @param[in] compile_context The compile context to be used. - * @param[in] op Logical binary operation to be executed. - * @param[in] src1 First source tensor info. Data types supported: U8. - * @param[in] src2 Second source tensor info. Data types supported: same as @p src1. - * @param[in] dst Destination tensor info. Data types supported: same as @p src1. - */ - void configure(const ClCompileContext &compile_context, LogicalOperation op, ITensorInfo *src1, ITensorInfo *src2, ITensorInfo *dst); - /** Static function to check if the given configuration is valid for this kernel - * - * @param[in] op Logical binary operation to be executed. - * @param[in] src1 First source tensor info. Data types supported: U8. - * @param[in] src2 Second source tensor info. Data types supported: same as @p src1. - * @param[in] dst Destination tensor info. Data types supported: same as @p src1. - */ - static Status validate(LogicalOperation op, const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst); - -private: - // Inherited methods overridden: - std::string name() override; - std::pair<Status, Window> validate_and_configure_window(ITensorInfo &src1, ITensorInfo &src2, ITensorInfo &dst) override; - CLBuildOptions generate_build_options(const ITensorInfo &src1, const ITensorInfo &src2, const ITensorInfo &dst) override; - std::string generate_id_for_tuning(const std::string &kernel_name, const ITensorInfo &src1, const ITensorInfo &dst) override; - - LogicalOperation _op{ LogicalOperation::Unknown }; -}; - -/** Addition operation */ -class ClSaturatedArithmeticKernel : public ClElementwiseKernel -{ -public: - ClSaturatedArithmeticKernel() = default; - ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(ClSaturatedArithmeticKernel); - /** Static function to check if given info will lead to a valid configuration of @ref ClSaturatedArithmeticKernel - * - * @param[in] compile_context The compile context to be used. - * @param[in] op Arithmetic operation to be executed. - * @param[in] input1 First tensor input info. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/S32/F32. - * @param[in] input2 Second tensor input info. Data types supported: Same as @p input1. - * @param[in] output Output tensor info. Data types supported: Same as @p input1. - * @param[in] policy Policy to use to handle overflow. - * @param[in] act_info (Optional) Activation layer information in case of a fused activation. - */ - void configure(const ClCompileContext &compile_context, ArithmeticOperation op, ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output, const ConvertPolicy &policy, - const ActivationLayerInfo &act_info = ActivationLayerInfo()); - - /** Static function to check if given info will lead to a valid configuration of @ref ClSaturatedArithmeticKernel - * - * @param[in] op Arithmetic operation to be executed. - * @param[in] input1 First tensor input info info. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/S32/F32. - * @param[in] input2 Second tensor input info info. Data types supported: Same as @p input1. - * @param[in] output Output tensor info info. Data types supported: Same as @p input1. - * @param[in] policy Policy to use to handle overflow. - * @param[in] act_info (Optional) Activation layer information in case of a fused activation. - * - * @return a Status - */ - static Status validate(ArithmeticOperation op, const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, const ConvertPolicy &policy, - const ActivationLayerInfo &act_info = ActivationLayerInfo()); - -protected: - // Inherited methods overridden: - std::string name() override; - std::pair<Status, Window> validate_and_configure_window(ITensorInfo &input1, ITensorInfo &input2, ITensorInfo &output) override; - CLBuildOptions generate_build_options(const ITensorInfo &input1, const ITensorInfo &input2, const ITensorInfo &output) override; - std::string generate_id_for_tuning(const std::string &kernel_name, const ITensorInfo &input1, const ITensorInfo &output) override; - -private: - ConvertPolicy _policy{}; - ArithmeticOperation _op{}; -}; - -class ClArithmeticKernel : public ClElementwiseKernel -{ -public: - ClArithmeticKernel() = default; - ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(ClArithmeticKernel); - - /** Static function to check if given info will lead to a valid configuration of @ref ClArithmeticKernel - * - * @param[in] compile_context The compile context to be used. - * @param[in] op Arithmetic operation to be executed. - * @param[in] src1 First source tensor info. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/S32/F32. - * @param[in] src2 Second source tensor info. Data types supported: same as @p src1. - * @param[in] dst Destination tensor info. Data types supported: same as @p src1. - * @param[in] act_info (Optional) Activation layer information in case of a fused activation. - */ - void configure(const ClCompileContext &compile_context, ArithmeticOperation op, ITensorInfo *src1, ITensorInfo *src2, ITensorInfo *dst, - const ActivationLayerInfo &act_info = ActivationLayerInfo()); - - /** Static function to check if given info will lead to a valid configuration of @ref ClArithmeticKernel - * - * @param[in] op Arithmetic operation to be executed. - * @param[in] src1 First source tensor info. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/S32/F32. - * @param[in] src2 Second source tensor info. Data types supported: same as @p src1. - * @param[in] dst Destination tensor info. Data types supported: same as @p src1. - * @param[in] act_info (Optional) Activation layer information in case of a fused activation. - * - * @return a Status - */ - static Status validate(ArithmeticOperation op, const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst, const ActivationLayerInfo &act_info = ActivationLayerInfo()); - -protected: - // Inherited methods overridden: - std::string name() override; - std::pair<Status, Window> validate_and_configure_window(ITensorInfo &src1, ITensorInfo &src2, ITensorInfo &dst) override; - CLBuildOptions generate_build_options(const ITensorInfo &src1, const ITensorInfo &src2, const ITensorInfo &dst) override; - std::string generate_id_for_tuning(const std::string &kernel_name, const ITensorInfo &src1, const ITensorInfo &dst) override; - -private: - ArithmeticOperation _op{}; -}; -} // namespace kernels -} // namespace opencl -} // namespace arm_compute -#endif /* ARM_COMPUTE_CL_ELEMENTWISE_KERNEL_H */ diff --git a/src/core/gpu/cl/kernels/ClElementwiseUnaryKernel.cpp b/src/core/gpu/cl/kernels/ClElementwiseUnaryKernel.cpp deleted file mode 100644 index 8e9698cd50..0000000000 --- a/src/core/gpu/cl/kernels/ClElementwiseUnaryKernel.cpp +++ /dev/null @@ -1,159 +0,0 @@ -/* - * Copyright (c) 2018-2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#include "src/core/gpu/cl/kernels/ClElementwiseUnaryKernel.h" - -#include "arm_compute/core/CL/CLHelpers.h" -#include "arm_compute/core/CL/ICLTensor.h" -#include "src/core/CL/CLValidate.h" -#include "src/core/helpers/WindowHelpers.h" -#include "support/Cast.h" -#include "support/StringSupport.h" - -namespace arm_compute -{ -namespace opencl -{ -namespace kernels -{ -namespace -{ -Status validate_arguments(const ITensorInfo &src, const ITensorInfo &dst, const ElementWiseUnary op) -{ - ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(&src); - if(op == ElementWiseUnary::LOGICAL_NOT) - { - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&src, 1, DataType::U8); - } - else - { - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&src, 1, DataType::F16, DataType::F32); - } - - // Validate in case of configured dst - if(dst.total_size() > 0) - { - ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(&dst); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&src, &dst); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(&src, &dst); - } - - return Status{}; -} -} // namespace - -void ClElementWiseUnaryKernel::configure(const CLCompileContext &compile_context, const ITensorInfo *src, ITensorInfo *dst, const ElementWiseUnary &op) -{ - ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst); - - auto padding_info = get_padding_info({ src, dst }); - - ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(*src, *dst, op)); - - const std::string kernel_name = "elementwise_unary"; - const int vec_size_x = 16 / dst->element_size(); - const int dst_width_x = dst->tensor_shape().x(); - const bool multi_access_x = (dst_width_x / vec_size_x > 0); - - // Set kernel build options - CLBuildOptions build_opts; - build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(src->data_type())); - build_opts.add_option_if(multi_access_x, "-DVEC_SIZE=" + support::cpp11::to_string(vec_size_x)); - build_opts.add_option_if(multi_access_x, "-DLAST_ACCESSED_X=" + support::cpp11::to_string(std::max<int>(dst_width_x - vec_size_x, 0))); - switch(op) - { - case ElementWiseUnary::RSQRT: - build_opts.add_option("-DOPERATION=rsqrt_op"); - break; - case ElementWiseUnary::EXP: - build_opts.add_option("-DOPERATION=exp_op"); - break; - case ElementWiseUnary::NEG: - build_opts.add_option("-DOPERATION=neg_op"); - break; - case ElementWiseUnary::SIN: - build_opts.add_option("-DOPERATION=sin_op"); - break; - case ElementWiseUnary::ABS: - build_opts.add_option("-DOPERATION=fabs_op"); - break; - case ElementWiseUnary::LOG: - build_opts.add_option("-DOPERATION=natural_log_op"); - break; - case ElementWiseUnary::ROUND: - build_opts.add_option("-DOPERATION=round_op"); - break; - case ElementWiseUnary::LOGICAL_NOT: - build_opts.add_option("-DOPERATION=logical_not_op"); - break; - default: - ARM_COMPUTE_ERROR("Not implemented"); - } - - // Create kernel - _kernel = create_kernel(compile_context, kernel_name, build_opts.options()); - - // Configure kernel window - Window win = calculate_max_window(*dst); - if(multi_access_x) - { - win.set(Window::DimX, - Window::Dimension(win.x().start(), ceil_to_multiple(win.x().end(), vec_size_x), vec_size_x)); - } - ICLKernel::configure_internal(win); - - ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info)); -} - -Status ClElementWiseUnaryKernel::validate(const ITensorInfo *src, const ITensorInfo *dst, const ElementWiseUnary &op) -{ - ARM_COMPUTE_UNUSED(op); - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst); - ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(*src, *dst, op)); - - return Status{}; -} - -void ClElementWiseUnaryKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) -{ - ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); - ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); - - Window collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ); - Window slice = collapsed.first_slice_window_3D(); - - const auto src = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC)); - auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST)); - - do - { - unsigned int idx = 0; - add_3D_tensor_argument(idx, src, slice); - add_3D_tensor_argument(idx, dst, slice); - enqueue(queue, *this, slice, lws_hint()); - } - while(collapsed.slide_window_slice_3D(slice)); -} -} // namespace kernels -} // namespace opencl -} // namespace arm_compute diff --git a/src/core/gpu/cl/kernels/ClElementwiseUnaryKernel.h b/src/core/gpu/cl/kernels/ClElementwiseUnaryKernel.h deleted file mode 100644 index 7e5edef3ee..0000000000 --- a/src/core/gpu/cl/kernels/ClElementwiseUnaryKernel.h +++ /dev/null @@ -1,67 +0,0 @@ -/* - * Copyright (c) 2018-2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#ifndef ARM_COMPUTE_CL_ELEMENTWISE_UNARY_KERNEL_H -#define ARM_COMPUTE_CL_ELEMENTWISE_UNARY_KERNEL_H - -#include "src/core/common/Macros.h" -#include "src/core/gpu/cl/ClCompileContext.h" -#include "src/core/gpu/cl/IClKernel.h" - -namespace arm_compute -{ -namespace opencl -{ -namespace kernels -{ -/** Interface for the elementwise unary operator */ -class ClElementWiseUnaryKernel : public IClKernel -{ -public: - ClElementWiseUnaryKernel() = default; - ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(ClElementWiseUnaryKernel); - /** Initialise the kernel's srcs, dst. - * - * @param[in] compile_context The compile context to be used. - * @param[in] src First source tensor info. Data types supported: F16/F32. - * @param[out] dst Destination tensor info. Data types supported: same as @p src. - * @param[in] op Element wise unary operation to perform. - */ - void configure(const CLCompileContext &compile_context, const ITensorInfo *src, ITensorInfo *dst, const ElementWiseUnary &op); - /** Static function to check if given info will lead to a valid configuration of @ref ClElementWiseUnaryKernel - * - * @param[in] src First source tensor info. Data types supported: F16/F32. - * @param[in] dst Destination tensor info. Data types supported: same as @p src. - * @param[in] op Element wise unary operation to perform. - * - * @return a status - */ - static Status validate(const ITensorInfo *src, const ITensorInfo *dst, const ElementWiseUnary &op); - - // Inherited methods overridden: - void run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) override; -}; -} // namespace kernels -} // namespace opencl -} // namespace arm_compute -#endif /* ARM_COMPUTE_CL_ELEMENTWISE_UNARY_KERNEL_H */ diff --git a/src/core/gpu/cl/kernels/ClFillKernel.cpp b/src/core/gpu/cl/kernels/ClFillKernel.cpp deleted file mode 100644 index b194ee549b..0000000000 --- a/src/core/gpu/cl/kernels/ClFillKernel.cpp +++ /dev/null @@ -1,122 +0,0 @@ -/* - * Copyright (c) 2018-2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#include "src/core/gpu/cl/kernels/ClFillKernel.h" - -#include "arm_compute/core/CL/CLHelpers.h" -#include "arm_compute/core/CL/CLKernelLibrary.h" -#include "arm_compute/core/CL/ICLTensor.h" -#include "arm_compute/core/Helpers.h" -#include "arm_compute/core/TensorInfo.h" -#include "arm_compute/core/Utils.h" -#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/Cast.h" -#include "support/StringSupport.h" - -namespace arm_compute -{ -namespace opencl -{ -namespace kernels -{ -void ClFillKernel::configure(ITensorInfo *tensor, - const PixelValue &constant_value, - Window *window) -{ - configure(CLKernelLibrary::get().get_compile_context(), tensor, constant_value, window); -} - -void ClFillKernel::configure(const CLCompileContext &compile_context, ITensorInfo *tensor, - const PixelValue &constant_value, - Window *window) -{ - ARM_COMPUTE_ERROR_ON_NULLPTR(tensor); - ARM_COMPUTE_ERROR_THROW_ON(validate(tensor, constant_value, window)); - - const DataType data_type = tensor->data_type(); - const int vec_size_x = 16 / tensor->element_size(); - - // Create and update the window (if needed) - _full_window = calculate_max_window(*tensor); - Window win = _full_window; - if(window != nullptr) - { - ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(win, *window); - win = *window; - } - - const int output_width_x = win.num_iterations(0); - const bool multi_access_x = output_width_x >= vec_size_x; - const bool remainder_x = output_width_x % vec_size_x > 0; - - if(multi_access_x) - { - win.set(Window::DimX, Window::Dimension(win.x().start(), ceil_to_multiple(win.x().end(), vec_size_x), vec_size_x)); - } - ICLKernel::configure_internal(win); - - // Create kernel - CLBuildOptions build_opts; - build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(data_type)); - build_opts.add_option("-DCONSTANT_VALUE=" + string_from_pixel_value(constant_value, data_type)); - build_opts.add_option_if(multi_access_x, "-DVEC_SIZE=" + support::cpp11::to_string(vec_size_x)); - build_opts.add_option_if(multi_access_x && remainder_x, "-DLAST_ACCESSED_X=" + support::cpp11::to_string(std::max<int>(output_width_x - vec_size_x, 0))); - _kernel = create_kernel(compile_context, "memset", build_opts.options()); -} - -Status ClFillKernel::validate(const ITensorInfo *tensor, const PixelValue &constant_value, Window *window) -{ - ARM_COMPUTE_UNUSED(tensor); - ARM_COMPUTE_UNUSED(constant_value); - if(window != nullptr) - { - ARM_COMPUTE_RETURN_ERROR_ON(window->x().step() != 1); - } - return Status{}; -} - -void ClFillKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) -{ - ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); - ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); - - const auto tensor = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC)); - - // Collapse all the batches on the third - Window collapsed = window.collapse_if_possible(_full_window, Window::DimZ); - Window slice = collapsed.first_slice_window_3D(); - - do - { - unsigned int idx = 0; - add_3D_tensor_argument(idx, tensor, slice); - enqueue(queue, *this, slice, lws_hint()); - } - while(collapsed.slide_window_slice_3D(slice)); -} -} // namespace kernels -} // namespace opencl -} // namespace arm_compute diff --git a/src/core/gpu/cl/kernels/ClFillKernel.h b/src/core/gpu/cl/kernels/ClFillKernel.h deleted file mode 100644 index 136c1243fe..0000000000 --- a/src/core/gpu/cl/kernels/ClFillKernel.h +++ /dev/null @@ -1,77 +0,0 @@ -/* - * Copyright (c) 2018-2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#ifndef ARM_COMPUTE_CL_FILL_KERNEL_H -#define ARM_COMPUTE_CL_FILL_KERNEL_H - -#include "src/core/common/Macros.h" -#include "src/core/gpu/cl/ClCompileContext.h" -#include "src/core/gpu/cl/IClKernel.h" - -namespace arm_compute -{ -namespace opencl -{ -namespace kernels -{ -/** Interface for filling the planes of a tensor */ -class ClFillKernel : public ICLKernel -{ -public: - ClFillKernel() = default; - ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(ClFillKernel); - /** Initialise the kernel's tensor and filling value - * - * @param[in,out] tensor Input tensor info. Supported data types: All. - * @param[in] constant_value The value used to fill the planes of the tensor - * @param[in] window Window to be used in case setting only part of a tensor. Default is nullptr. - */ - void configure(ITensorInfo *tensor, const PixelValue &constant_value, Window *window = nullptr); - /** Initialise the kernel's tensor and filling value - * - * @param[in] compile_context The compile context to be used. - * @param[in,out] tensor Input tensor info. Supported data types: All. - * @param[in] constant_value The value used to fill the planes of the tensor - * @param[in] window Window to be used in case setting only part of a tensor. Default is nullptr. - */ - void configure(const CLCompileContext &compile_context, ITensorInfo *tensor, const PixelValue &constant_value, Window *window = nullptr); - /** Static function to check if given info will lead to a valid configuration of @ref ClFillKernel - * - * @param[in] tensor Source tensor info. Data types supported: All. - * @param[in] constant_value The value used to fill the planes of the tensor - * @param[in] window Window to be used in case setting only part of a tensor. Default is nullptr. - * - * @return a status - */ - static Status validate(const ITensorInfo *tensor, const PixelValue &constant_value, Window *window = nullptr); - - // Inherited methods overridden: - void run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) override; - -private: - Window _full_window{}; -}; -} // namespace kernels -} // namespace opencl -} // namespace arm_compute -#endif /*ARM_COMPUTE_CLMEMSETRKERNEL_H */ diff --git a/src/core/gpu/cl/kernels/ClFloorKernel.cpp b/src/core/gpu/cl/kernels/ClFloorKernel.cpp deleted file mode 100644 index 7296d40eaf..0000000000 --- a/src/core/gpu/cl/kernels/ClFloorKernel.cpp +++ /dev/null @@ -1,119 +0,0 @@ -/* - * Copyright (c) 2017-2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#include "src/core/gpu/cl/kernels/ClFloorKernel.h" - -#include "arm_compute/core/CL/CLHelpers.h" -#include "arm_compute/core/CL/CLKernelLibrary.h" -#include "arm_compute/core/CL/ICLTensor.h" -#include "arm_compute/core/Helpers.h" -#include "arm_compute/core/TensorInfo.h" -#include "arm_compute/core/Utils.h" -#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/Cast.h" -#include "support/StringSupport.h" - -namespace arm_compute -{ -namespace opencl -{ -namespace kernels -{ -namespace -{ -Status validate_arguments(const ITensorInfo *src, const ITensorInfo *dst) -{ - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst); - ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(src); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::F16, DataType::F32); - - // Validate in case of configured output - if(dst->total_size() > 0) - { - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, dst); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(src, dst); - } - - return Status{}; -} -} // namespace - -void ClFloorKernel::configure(const ClCompileContext &compile_context, const ITensorInfo *src, ITensorInfo *dst) -{ - ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst); - - // Auto initialize output - auto_init_if_empty(*dst, src->tensor_shape(), 1, src->data_type()); - - // Validate - ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, dst)); - auto padding_info = get_padding_info({ src, dst }); - - const unsigned int vec_size_x = adjust_vec_size(max_cl_vector_width / src->element_size(), src->dimension(0)); - const int vec_size_x_leftovers = src->dimension(0) % vec_size_x; - CLBuildOptions build_opts; - build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(src->data_type())); - build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(vec_size_x)); - build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(vec_size_x_leftovers)); - - // Create kernel - _kernel = create_kernel(compile_context, "floor_layer", build_opts.options()); - - // Configure kernel window - Window win = calculate_max_window(*src, Steps(vec_size_x)); - IClKernel::configure_internal(win); - ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info)); -} - -Status ClFloorKernel::validate(const ITensorInfo *src, const ITensorInfo *dst) -{ - ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, dst)); - return Status{}; -} - -void ClFloorKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) -{ - ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); - ARM_COMPUTE_ERROR_ON_MISMATCHING_WINDOWS(IClKernel::window(), window); - - const auto src = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC)); - auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST)); - - Window collapsed = window.collapse_if_possible(IClKernel::window(), Window::DimZ); - Window slice = collapsed.first_slice_window_3D(); - - do - { - unsigned int idx = 0; - add_3D_tensor_argument(idx, src, slice); - add_3D_tensor_argument(idx, dst, slice); - enqueue(queue, *this, slice, lws_hint()); - } - while(collapsed.slide_window_slice_3D(slice)); -} -} // namespace kernels -} // namespace opencl -} // namespace arm_compute diff --git a/src/core/gpu/cl/kernels/ClFloorKernel.h b/src/core/gpu/cl/kernels/ClFloorKernel.h deleted file mode 100644 index 646dfb30d8..0000000000 --- a/src/core/gpu/cl/kernels/ClFloorKernel.h +++ /dev/null @@ -1,66 +0,0 @@ -/* - * Copyright (c) 2017-2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#ifndef ARM_COMPUTE_CL_FLOOR_KERNEL_H -#define ARM_COMPUTE_CL_FLOOR_KERNEL_H - -#include "src/core/common/Macros.h" -#include "src/core/gpu/cl/ClCompileContext.h" -#include "src/core/gpu/cl/IClKernel.h" - -namespace arm_compute -{ -namespace opencl -{ -namespace kernels -{ -/** OpenCL kernel to perform a floor operation */ -class ClFloorKernel : public IClKernel -{ -public: - ClFloorKernel() = default; - ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(ClFloorKernel); - /** Configure kernel for a given list of arguments - * - * @param[in] compile_context The compile context to be used. - * @param[in] src Source tensor info. Data type supported: F16/F32. - * @param[out] dst Destination tensor info. Same as @p src - */ - void configure(const ClCompileContext &compile_context, const ITensorInfo *src, ITensorInfo *dst); - - /** Static function to check if given info will lead to a valid configuration of @ref ClFloorKernel - * - * @param[in] src Source tensor info. Data type supported: F16/F32. - * @param[in] dst Destination tensor info. Same as @p src - * - * @return a status - */ - static Status validate(const ITensorInfo *src, const ITensorInfo *dst); - - // Inherited methods overridden: - void run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) override; -}; -} // namespace kernels -} // namespace opencl -} // namespace arm_compute -#endif /* ARM_COMPUTE_CL_FLOOR_KERNEL_H */ diff --git a/src/core/gpu/cl/kernels/ClHeightConcatenateKernel.cpp b/src/core/gpu/cl/kernels/ClHeightConcatenateKernel.cpp deleted file mode 100644 index 83e976e10f..0000000000 --- a/src/core/gpu/cl/kernels/ClHeightConcatenateKernel.cpp +++ /dev/null @@ -1,134 +0,0 @@ -/* - * Copyright (c) 2019-2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#include "src/core/gpu/cl/kernels/ClHeightConcatenateKernel.h" - -#include "arm_compute/core/CL/CLHelpers.h" -#include "arm_compute/core/CL/CLKernelLibrary.h" -#include "arm_compute/core/CL/ICLTensor.h" -#include "arm_compute/core/Helpers.h" -#include "arm_compute/core/Utils.h" -#include "arm_compute/core/utils/misc/ShapeCalculator.h" -#include "src/core/CL/CLValidate.h" -#include "src/core/helpers/WindowHelpers.h" -#include "support/Cast.h" - -#include "support/StringSupport.h" - -namespace arm_compute -{ -namespace opencl -{ -namespace kernels -{ -namespace -{ -Status validate_arguments(const ITensorInfo *src, unsigned int height_offset, const ITensorInfo *dst) -{ - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst); - ARM_COMPUTE_RETURN_ERROR_ON(src->data_type() == DataType::UNKNOWN); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, dst); - ARM_COMPUTE_RETURN_ERROR_ON(src->dimension(Window::DimY) + height_offset > dst->dimension(Window::DimY)); - - ARM_COMPUTE_RETURN_ERROR_ON(src->dimension(0) != dst->dimension(0)); - for(size_t i = 2; i < Coordinates::num_max_dimensions; ++i) - { - ARM_COMPUTE_RETURN_ERROR_ON(src->dimension(i) != dst->dimension(i)); - } - ARM_COMPUTE_RETURN_ERROR_ON(src->num_dimensions() > 4); - - return Status{}; -} -} // namespace - -ClHeightConcatenateKernel::ClHeightConcatenateKernel() - : _height_offset(0) -{ -} - -Status ClHeightConcatenateKernel::validate(const ITensorInfo *src, unsigned int height_offset, const ITensorInfo *dst) -{ - ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, height_offset, dst)); - return Status{}; -} - -void ClHeightConcatenateKernel::configure(const CLCompileContext &compile_context, ITensorInfo *src, unsigned int height_offset, ITensorInfo *dst) -{ - ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst); - ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, height_offset, dst)); - - auto padding_info = get_padding_info({ src, dst }); - - _height_offset = height_offset; - - // Add build options - const unsigned int num_elems_processed_per_iteration = adjust_vec_size(4, src->dimension(0)); - - CLBuildOptions build_opts; - build_opts.add_option("-DDATA_TYPE=" + get_cl_unsigned_type_from_element_size(src->element_size())); - build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration)); - build_opts.add_option("-DHEIGHT_OFFSET=" + support::cpp11::to_string(_height_offset)); - build_opts.add_option("-DDEPTH=" + support::cpp11::to_string(src->dimension(2))); - build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(src->dimension(0) % num_elems_processed_per_iteration)); - - if(is_data_type_quantized_asymmetric(src->data_type()) && src->quantization_info() != dst->quantization_info()) - { - const UniformQuantizationInfo iq_info = src->quantization_info().uniform(); - const UniformQuantizationInfo oq_info = dst->quantization_info().uniform(); - - build_opts.add_option("-DOFFSET_IN1=" + float_to_string_with_full_precision(iq_info.offset)); - build_opts.add_option("-DOFFSET_OUT=" + float_to_string_with_full_precision(oq_info.offset)); - build_opts.add_option("-DSCALE_IN1=" + float_to_string_with_full_precision(iq_info.scale)); - build_opts.add_option("-DSCALE_OUT=" + float_to_string_with_full_precision(oq_info.scale)); - } - - // Create kernel - _kernel = create_kernel(compile_context, "concatenate_height", build_opts.options()); - // Configure kernel window - - // The window needs to be based on src as we copy all the heights of src - Window win = calculate_max_window(*src, Steps(num_elems_processed_per_iteration)); - ICLKernel::configure_internal(win.collapse(win, Window::DimZ)); - - // Set dst valid region - dst->set_valid_region(ValidRegion(Coordinates(), dst->tensor_shape())); - - ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info)); -} - -void ClHeightConcatenateKernel::run_op(ITensorPack &tensors, const Window &window, ::cl::CommandQueue &queue) -{ - ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); - ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); - - const auto src = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC)); - auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST)); - - unsigned int idx = 0; - add_4D_tensor_argument(idx, src, window); - add_4D_tensor_argument(idx, dst, window); - enqueue(queue, *this, window, lws_hint()); -} -} // namespace kernels -} // namespace opencl -} // namespace arm_compute diff --git a/src/core/gpu/cl/kernels/ClHeightConcatenateKernel.h b/src/core/gpu/cl/kernels/ClHeightConcatenateKernel.h deleted file mode 100644 index 9a4380a5b7..0000000000 --- a/src/core/gpu/cl/kernels/ClHeightConcatenateKernel.h +++ /dev/null @@ -1,74 +0,0 @@ -/* - * Copyright (c) 2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#ifndef ARM_COMPUTE_CL_HEIGHT_CONCATENATE_LAYER_KERNEL_H -#define ARM_COMPUTE_CL_HEIGHT_CONCATENATE_LAYER_KERNEL_H - -#include "src/core/common/Macros.h" -#include "src/core/gpu/cl/ClCompileContext.h" -#include "src/core/gpu/cl/IClKernel.h" - -namespace arm_compute -{ -namespace opencl -{ -namespace kernels -{ -/** Interface for the height concatenate kernel. - * The source tensor will be concatenated into the destination tensor. - */ -class ClHeightConcatenateKernel : public IClKernel -{ -public: - /** Default constructor */ - ClHeightConcatenateKernel(); - ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(ClHeightConcatenateKernel); - /** Initialise the kernel's source and destination - * - * @param[in] compile_context The compile context to be used. - * @param[in] src Source tensor info. Data types supported: All. - * @param[in] height_offset The starting offset on the Y axis for the dst tensor. - * @param[out] dst Destination tensor info. Data types supported: same as @p src. - * - */ - void configure(const CLCompileContext &compile_context, ITensorInfo *src, unsigned int height_offset, ITensorInfo *dst); - /** Static function to check if given info will lead to a valid configuration of @ref ClHeightConcatenateKernel - * - * @param[in] src Source tensor info. Data types supported: All. - * @param[in] height_offset The starting offset on the Y axis for the dst tensor. - * @param[in] dst Destination tensor info. Data types supported: same as @p src. - * - * @return a status - */ - static Status validate(const ITensorInfo *src, unsigned int height_offset, const ITensorInfo *dst); - - // Inherited methods overridden: - void run_op(ITensorPack &tensors, const Window &window, ::cl::CommandQueue &queue) override; - -private: - unsigned int _height_offset; -}; -} // namespace kernels -} // namespace opencl -} // namespace arm_compute -#endif /* ARM_COMPUTE_CL_HEIGHT_CONCATENATE_LAYER_KERNEL_H */ diff --git a/src/core/gpu/cl/kernels/ClPermuteKernel.cpp b/src/core/gpu/cl/kernels/ClPermuteKernel.cpp deleted file mode 100644 index 992c2a89d3..0000000000 --- a/src/core/gpu/cl/kernels/ClPermuteKernel.cpp +++ /dev/null @@ -1,157 +0,0 @@ -/* - * Copyright (c) 2018-2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#include "src/core/gpu/cl/kernels/ClPermuteKernel.h" - -#include "arm_compute/core/CL/CLHelpers.h" -#include "arm_compute/core/CL/CLKernelLibrary.h" -#include "arm_compute/core/CL/ICLTensor.h" -#include "arm_compute/core/Helpers.h" -#include "arm_compute/core/TensorInfo.h" -#include "arm_compute/core/Utils.h" -#include "arm_compute/core/Validate.h" -#include "arm_compute/core/utils/misc/ShapeCalculator.h" -#include "src/core/CL/CLValidate.h" -#include "src/core/helpers/AutoConfiguration.h" -#include "src/core/helpers/WindowHelpers.h" -#include "support/Cast.h" -#include "support/StringSupport.h" - -namespace arm_compute -{ -namespace opencl -{ -namespace kernels -{ -namespace -{ -TensorShape get_dst_shape(const ITensorInfo *src, const PermutationVector &perm) -{ - TensorShape dst_shape = src->tensor_shape(); - permute(dst_shape, perm); - return dst_shape; -} - -Status validate_arguments(const ITensorInfo *src, const ITensorInfo *dst, const PermutationVector &perm) -{ - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst); - ARM_COMPUTE_RETURN_ERROR_ON(src->data_type() == DataType::UNKNOWN); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(src->num_dimensions() < 1 || src->num_dimensions() > 4, - "Permutation upto 4-D src tensor is supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(perm.num_dimensions() < 1 || perm.num_dimensions() > 4, - "Permutation vector size should be less than or equal to 4"); - for(const auto &p : perm) - { - ARM_COMPUTE_RETURN_ERROR_ON_MSG(p >= perm.num_dimensions(), "Permutation vector has invalid values"); - } - - // Validate configured dst - if(dst->total_size() != 0) - { - const TensorShape dst_shape = misc::shape_calculator::compute_permutation_output_shape(*src, perm); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(dst->tensor_shape(), dst_shape); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(src, dst); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, dst); - } - return Status{}; -} -} // namespace - -void ClPermuteKernel::configure(const ITensorInfo *src, ITensorInfo *dst, const PermutationVector &perm) -{ - configure(CLKernelLibrary::get().get_compile_context(), src, dst, perm); -} - -void ClPermuteKernel::configure(const CLCompileContext &compile_context, const ITensorInfo *src, ITensorInfo *dst, const PermutationVector &perm) -{ - ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst); - auto padding_info = get_padding_info({ src, dst }); - const TensorShape dst_shape = get_dst_shape(src, perm); - // Output auto inizialitation if not yet initialized - auto_init_if_empty(*dst, src->clone()->set_tensor_shape(dst_shape)); - - ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, dst, perm)); - - _perm = perm; - - // Create kernel - CLBuildOptions build_opts; - build_opts.add_option("-DDATA_TYPE=" + get_cl_unsigned_type_from_element_size(data_size_from_type(src->data_type()))); - build_opts.add_option("-DDEPTH_IN=" + support::cpp11::to_string(src->dimension(2))); - // New positions of width(W), height(H), channel(C) and batch(D) based on permutation vector - build_opts.add_option("-DP1=" + support::cpp11::to_string((_perm.num_dimensions() >= 1) ? perm[0] : 0)); - build_opts.add_option("-DP2=" + support::cpp11::to_string((_perm.num_dimensions() >= 2) ? perm[1] : 1)); - build_opts.add_option("-DP3=" + support::cpp11::to_string((_perm.num_dimensions() >= 3) ? perm[2] : 2)); - build_opts.add_option("-DP4=" + support::cpp11::to_string((_perm.num_dimensions() >= 4) ? perm[3] : 3)); - - _kernel = create_kernel(compile_context, "permute", build_opts.options()); - - // Configure kernel window - Window win = calculate_max_window(*src, Steps()); - - // The CLPermute doesn't need padding so update_window_and_padding() can be skipped - Coordinates coord; - coord.set_num_dimensions(dst->num_dimensions()); - dst->set_valid_region(ValidRegion(coord, dst->tensor_shape())); - - ICLKernel::configure_internal(win); - ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info)); -} - -Status ClPermuteKernel::validate(const ITensorInfo *src, const ITensorInfo *dst, const PermutationVector &perm) -{ - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst); - ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, dst, perm)); - - return Status{}; -} - -void ClPermuteKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) -{ - ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); - ARM_COMPUTE_ERROR_ON_MISMATCHING_WINDOWS(ICLKernel::window(), window); - - const auto src = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC)); - auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST)); - - Window slice_in = window.first_slice_window_4D().collapse(ICLKernel::window(), 2, 4); - - // Setup dst slice - Window slice_out(slice_in); - slice_out.set(Window::DimX, Window::Dimension(0, 0, 0)); - slice_out.set(Window::DimY, Window::Dimension(0, 0, 0)); - slice_out.set(Window::DimZ, Window::Dimension(0, 0, 0)); - slice_out.set(3, Window::Dimension(0, 0, 0)); - - do - { - unsigned int idx = 0; - add_4D_tensor_argument(idx, src, slice_in); - add_4D_tensor_argument(idx, dst, slice_out); - enqueue(queue, *this, slice_in, lws_hint()); - } - while(window.slide_window_slice_4D(slice_in) && window.slide_window_slice_4D(slice_out)); -} -} // namespace kernels -} // namespace opencl -} // namespace arm_compute
\ No newline at end of file diff --git a/src/core/gpu/cl/kernels/ClPermuteKernel.h b/src/core/gpu/cl/kernels/ClPermuteKernel.h deleted file mode 100644 index 4cc72491bd..0000000000 --- a/src/core/gpu/cl/kernels/ClPermuteKernel.h +++ /dev/null @@ -1,86 +0,0 @@ -/* - * Copyright (c) 2018-2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#ifndef ARM_COMPUTE_CL_PERMUTE_KERNEL_H -#define ARM_COMPUTE_CL_PERMUTE_KERNEL_H - -#include "src/core/common/Macros.h" -#include "src/core/gpu/cl/ClCompileContext.h" -#include "src/core/gpu/cl/IClKernel.h" - -namespace arm_compute -{ -namespace opencl -{ -namespace kernels -{ -/** OpenCL kernel to perform tensor permutation. - * - * Permutes given a permutation vector - */ -class ClPermuteKernel : public ICLKernel -{ -public: - ClPermuteKernel() = default; - ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(ClPermuteKernel); - /** Set the src and dst of the kernel. - * - * @note Arbitrary permutation vectors are supported with rank not greater than 4 - * - * @param[in] src The src tensor info. Data types supported: All. - * @param[in] dst The dst tensor info. Data types supported: Same as @p src - * @param[in] perm Permutation vector - */ - void configure(const ITensorInfo *src, ITensorInfo *dst, const PermutationVector &perm); - /** Set the src and dst of the kernel. - * - * @note Arbitrary permutation vectors are supported with rank not greater than 4 - * - * @param[in] compile_context The compile context to be used. - * @param[in] src The src tensor info. Data types supported: All. - * @param[in] dst The dst tensor info. Data types supported: Same as @p src - * @param[in] perm Permutation vector - */ - void configure(const CLCompileContext &compile_context, const ITensorInfo *src, ITensorInfo *dst, const PermutationVector &perm); - /** Static function to check if given info will lead to a valid configuration of @ref ClPermuteKernel - * - * @note Arbitrary permutation vectors are supported with rank not greater than 4 - * - * @param[in] src The src tensor info. Data types supported: All. - * @param[in] dst The dst tensor info. Data types supported: same as @p src. - * @param[in] perm Permutation vector - * - * @return a status - */ - static Status validate(const ITensorInfo *src, const ITensorInfo *dst, const PermutationVector &perm); - - // Inherited methods overridden: - void run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) override; - -private: - PermutationVector _perm{}; -}; -} // namespace kernels -} // namespace opencl -} // namespace arm_compute -#endif /*ARM_COMPUTE_CL_PERMUTE_KERNEL_H */ diff --git a/src/core/gpu/cl/kernels/ClPoolingKernel.cpp b/src/core/gpu/cl/kernels/ClPoolingKernel.cpp deleted file mode 100644 index 567fec2a37..0000000000 --- a/src/core/gpu/cl/kernels/ClPoolingKernel.cpp +++ /dev/null @@ -1,502 +0,0 @@ -/* - * Copyright (c) 2017-2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#include "src/core/gpu/cl/kernels/ClPoolingKernel.h" - -#include "arm_compute/core/CL/CLHelpers.h" -#include "arm_compute/core/CL/CLKernelLibrary.h" -#include "arm_compute/core/CL/ICLTensor.h" -#include "arm_compute/core/TensorInfo.h" -#include "arm_compute/core/Utils.h" -#include "arm_compute/core/Validate.h" -#include "arm_compute/core/utils/misc/ShapeCalculator.h" -#include "src/core/CL/CLValidate.h" -#include "src/core/helpers/AutoConfiguration.h" -#include "src/core/helpers/WindowHelpers.h" -#include "support/Cast.h" -#include "support/StringSupport.h" - -namespace arm_compute -{ -namespace opencl -{ -namespace kernels -{ -using namespace arm_compute::misc::shape_calculator; - -namespace -{ -// Internal window config info -using ClPoolingConfig = std::pair<unsigned int, BorderSize>; //num_elems_processed_per_iteration, border_size - -void auto_init(const ITensorInfo *src, ITensorInfo *dst, ITensorInfo *indices, PoolingLayerInfo pool_info) -{ - TensorShape out_shape = compute_pool_shape(*src, pool_info); - auto_init_if_empty(*dst, src->clone()->set_tensor_shape(out_shape)); - if(indices) - { - auto_init_if_empty(*indices, src->clone()->set_tensor_shape(out_shape).set_data_type(DataType::U32)); - } -} - -Status validate_arguments(const ITensorInfo *src, const ITensorInfo *dst, const PoolingLayerInfo &pool_info, const ITensorInfo *indices) -{ - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst); - ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(src); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::F16, DataType::F32); - ARM_COMPUTE_RETURN_ERROR_ON_MSG((is_data_type_quantized_asymmetric(src->data_type()) && pool_info.pool_type == PoolingType::L2), - "Unsupported combination of parameters!"); - - // Check indices - if(indices) - { - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::F16, DataType::F32); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(pool_info.pool_type != PoolingType::MAX, "Pooling indices only supported for MAX pooling method"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG((pool_info.pool_size != Size2D(2, 2)), "Pooling indices only supported for pool size 2x2"); - - if(indices->total_size() != 0) - { - TensorInfo idx_info(TensorInfo(compute_pool_shape(*src, pool_info), 1, DataType::U32)); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(indices, &idx_info); - } - } - - // Checks performed when dst is configured - if(dst->total_size() != 0) - { - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, dst); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(src, dst); - TensorInfo out_info(TensorInfo(compute_pool_shape(*src, pool_info), 1, dst->data_type())); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(dst, &out_info); - } - - return Status{}; -} - -std::tuple<Status, Window, ClPoolingConfig> validate_and_configure_window(ITensorInfo *src, ITensorInfo *dst, const PoolingLayerInfo &pool_info, ITensorInfo *indices = nullptr) -{ - ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst); - - // Get data layout - const DataLayout data_layout = pool_info.data_layout == DataLayout::UNKNOWN ? src->data_layout() : pool_info.data_layout; - const int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH); - const int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT); - - int pool_stride_x = 0; - int pool_stride_y = 0; - unsigned int pooled_w = 0; - unsigned int pooled_h = 0; - int pool_size_x = pool_info.is_global_pooling ? src->dimension(idx_width) : pool_info.pool_size.width; - int pool_size_y = pool_info.is_global_pooling ? src->dimension(idx_height) : pool_info.pool_size.height; - const PadStrideInfo pad_stride_info = pool_info.pad_stride_info; - std::tie(pool_stride_x, pool_stride_y) = pad_stride_info.stride(); - const int pool_pad_right = pad_stride_info.pad_right(); - const int pool_pad_top = pad_stride_info.pad_top(); - const int pool_pad_left = pad_stride_info.pad_left(); - const int pool_pad_bottom = pad_stride_info.pad_bottom(); - BorderSize border_size = BorderSize(); - - auto_init(src, dst, indices, pool_info); - pooled_w = dst->tensor_shape()[idx_width]; - pooled_h = dst->tensor_shape()[idx_height]; - - const DataType data_type = src->data_type(); - - const int src_width = src->dimension(idx_width); - const int src_height = src->dimension(idx_height); - - unsigned int num_elems_processed_per_iteration = 0; - bool window_changed = false; - Window win{}; - switch(data_layout) - { - case DataLayout::NCHW: - { - // Initialize border size - border_size = BorderSize(pool_pad_top, pool_pad_right, pool_pad_bottom, pool_pad_left); - // Change the number of elements processed per iteration - // for pooling 3x3 with stride less equal than 3 - const bool can_optimize = (pool_size_x == 3) && (pool_size_y == 3) && (pool_stride_x <= 3) && !is_data_type_quantized(data_type); - num_elems_processed_per_iteration = can_optimize ? 4 : 1; - const unsigned int num_elems_read_per_iteration = (num_elems_processed_per_iteration - 1) * pool_stride_x + pool_size_x; - - // Number of iterations in X dimension - const int num_iterations_x = (pooled_w + num_elems_processed_per_iteration - 1) / num_elems_processed_per_iteration; - - // Upper limit for the number of right/bottom border elements that are accessed - const int upper_bound_w = ((num_iterations_x - 1) * num_elems_processed_per_iteration * pool_stride_x - pool_pad_left + num_elems_read_per_iteration) - src_width; - const int upper_bound_h = ((pooled_h - 1) * pool_stride_y - pool_pad_top + pool_size_y) - src_height; - - border_size.right = std::max(upper_bound_w, pool_pad_right); - border_size.bottom = std::max(upper_bound_h, pool_pad_bottom); - - win = calculate_max_window(*dst, Steps(num_elems_processed_per_iteration)); - - AccessWindowRectangle src_access(src, -pool_pad_left, -pool_pad_top, num_elems_read_per_iteration, pool_size_y, - pool_stride_x, pool_stride_y); - AccessWindowHorizontal dst_access(dst, 0, num_elems_processed_per_iteration); - - // Update indices window - if(indices) - { - AccessWindowHorizontal indices_access(indices, 0, num_elems_processed_per_iteration); - window_changed = update_window_and_padding(win, src_access, dst_access, indices_access); - indices_access.set_valid_region(win, ValidRegion(Coordinates(), indices->tensor_shape())); - } - else - { - window_changed = update_window_and_padding(win, src_access, dst_access); - } - - dst_access.set_valid_region(win, ValidRegion(Coordinates(), dst->tensor_shape())); - break; - } - case DataLayout::NHWC: - { - // Initialize border size - border_size = BorderSize(); - num_elems_processed_per_iteration = adjust_vec_size(4, dst->dimension(0)); - win = calculate_max_window(*dst, Steps(num_elems_processed_per_iteration)); - - if(indices != nullptr) - { - indices->set_valid_region(ValidRegion(Coordinates(), indices->tensor_shape())); - } - - dst->set_valid_region(ValidRegion(Coordinates(), dst->tensor_shape())); - break; - } - default: - ARM_COMPUTE_ERROR("Not implemented"); - } - - Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; - return std::make_tuple(err, win, ClPoolingConfig(num_elems_processed_per_iteration, border_size)); -} -} // namespace - -ClPoolingKernel::ClPoolingKernel() - : _pool_info(), _data_layout(DataLayout::UNKNOWN), _border_size(0), _num_elems_processed_per_iteration(1) -{ -} - -BorderSize ClPoolingKernel::border_size() const -{ - return _border_size; -} - -void ClPoolingKernel::configure(const ClCompileContext &compile_context, ITensorInfo *src, ITensorInfo *dst, const PoolingLayerInfo &pool_info, ITensorInfo *indices) -{ - ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst); - - auto padding_info = get_padding_info({ src, dst, indices }); - - // Set instance variables - _pool_info = pool_info; - _data_layout = pool_info.data_layout == DataLayout::UNKNOWN ? src->data_layout() : pool_info.data_layout; - int pool_stride_x = 0; - int pool_stride_y = 0; - const PoolingType pool_type = pool_info.pool_type; - const int idx_width = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::WIDTH); - const int idx_height = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::HEIGHT); - const int idx_channel = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::CHANNEL); - const int idx_batch_size = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::BATCHES); - const int pool_size_x = pool_info.is_global_pooling ? src->dimension(idx_width) : pool_info.pool_size.width; - const int pool_size_y = pool_info.is_global_pooling ? src->dimension(idx_height) : pool_info.pool_size.height; - const PadStrideInfo pad_stride_info = pool_info.pad_stride_info; - const bool exclude_padding = pool_info.exclude_padding; - std::tie(pool_stride_x, pool_stride_y) = pad_stride_info.stride(); - const int pool_pad_top = pad_stride_info.pad_top(); - const int pool_pad_left = pad_stride_info.pad_left(); - - // Set build options - CLBuildOptions build_opts; - const DataType data_type = src->data_type(); - - // Configure kernel window - auto win_config = validate_and_configure_window(src, dst, pool_info, indices); - - ARM_COMPUTE_ERROR_THROW_ON(std::get<0>(win_config)); - ICLKernel::configure_internal(std::get<1>(win_config)); - - ClPoolingConfig pooling_config = std::get<2>(win_config); - _num_elems_processed_per_iteration = pooling_config.first; - _border_size = pooling_config.second; - - build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(_num_elems_processed_per_iteration)); - - // Tensor paddings are used to calculate the indicies for MAX pooling - if(pool_info.pool_size == Size2D(2, 2) && pool_type == PoolingType::MAX && indices && is_data_type_float(data_type)) - { - build_opts.add_option("-DPAD_TENSOR_LEFT=" + support::cpp11::to_string(src->padding().left)); - build_opts.add_option("-DPAD_TENSOR_RIGHT=" + support::cpp11::to_string(src->padding().right)); - build_opts.add_option("-DPAD_TENSOR_TOP=" + support::cpp11::to_string(src->padding().top)); - build_opts.add_option("-DPAD_TENSOR_BOTTOM=" + support::cpp11::to_string(src->padding().bottom)); - build_opts.add_option("-DTENSOR_CHANNEL=" + support::cpp11::to_string(src->dimension(idx_channel))); - build_opts.add_option("-DTENSOR_WIDTH=" + support::cpp11::to_string(src->dimension(idx_width))); - build_opts.add_option("-DTENSOR_HEIGHT=" + support::cpp11::to_string(src->dimension(idx_height))); - } - - if(is_data_type_quantized_asymmetric(data_type) && src->quantization_info() != dst->quantization_info()) - { - const UniformQuantizationInfo iq_info = src->quantization_info().uniform(); - const UniformQuantizationInfo oq_info = dst->quantization_info().uniform(); - - build_opts.add_option("-DOFFSET_IN1=" + float_to_string_with_full_precision(iq_info.offset)); - build_opts.add_option("-DOFFSET_OUT=" + float_to_string_with_full_precision(oq_info.offset)); - build_opts.add_option("-DSCALE_IN1=" + float_to_string_with_full_precision(iq_info.scale)); - build_opts.add_option("-DSCALE_OUT=" + float_to_string_with_full_precision(oq_info.scale)); - } - - // Check dst dimensions - auto_init(src, dst, indices, pool_info); - - ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, dst, pool_info, indices)); - - build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(data_type)); - build_opts.add_option("-DPOOL_" + string_from_pooling_type(pool_type)); - build_opts.add_option("-DSTRIDE_X=" + support::cpp11::to_string(pool_stride_x)); - build_opts.add_option("-DSTRIDE_Y=" + support::cpp11::to_string(pool_stride_y)); - build_opts.add_option("-DPAD_X=" + support::cpp11::to_string(pool_pad_left)); - build_opts.add_option("-DPAD_Y=" + support::cpp11::to_string(pool_pad_top)); - build_opts.add_option("-DPOOL_SIZE_X=" + support::cpp11::to_string(pool_size_x)); - build_opts.add_option("-DPOOL_SIZE_Y=" + support::cpp11::to_string(pool_size_y)); - - // Set the initial value for the pooling operation accordingly with the data type - if(pool_type == PoolingType::MAX) - { - if(is_data_type_quantized(data_type)) - { - PixelValue type_min{}; - std::tie(type_min, std::ignore) = get_min_max(data_type); - build_opts.add_option("-DINITIAL_VALUE=" + support::cpp11::to_string(type_min.get<int32_t>())); - } - else - { - build_opts.add_option("-DINITIAL_VALUE=" + float_to_string_with_full_precision(std::numeric_limits<float>::lowest())); - } - } - else - { - // Pool AVG and Pool L2 initial value - build_opts.add_option("-DINITIAL_VALUE=0"); - } - - build_opts.add_option("-DMAX_WIDTH=" + support::cpp11::to_string(src->dimension(idx_width) + (exclude_padding ? 0 : pool_pad_left))); - build_opts.add_option("-DMAX_HEIGHT=" + support::cpp11::to_string(src->dimension(idx_height) + (exclude_padding ? 0 : pool_pad_top))); - - // Create kernel - switch(_data_layout) - { - case DataLayout::NCHW: - { - const auto use_fp_mixed_precision = (data_type == DataType::F16) && pool_info.fp_mixed_precision; - const auto use_wider_accumulator = use_fp_mixed_precision && (pool_type != PoolingType::MAX); - const auto acc_data_type = get_cl_type_from_data_type(use_wider_accumulator ? DataType::F32 : data_type); - build_opts.add_option("-DACC_DATA_TYPE=" + acc_data_type); - build_opts.add_option_if(use_wider_accumulator, "-DFP_MIXED_PRECISION"); - - if(pool_type != PoolingType::MAX) - { - build_opts.add_option_if(exclude_padding, "-DEXCLUDE_PADDING"); - } - - if((pool_size_x == 3) && (pool_size_y == 3) && !is_data_type_quantized_asymmetric(data_type)) - { - // Check if we have pool3x3 with stride_x less equal than 3. In these cases, run an optimized OpenCL kernel where - // each thread computes 4 dst elements - const bool is_pool3x3_stride_le3 = (pool_size_x == 3) && (pool_size_y == 3) && (pool_stride_x <= 3); - - std::string kernel_name = ((is_pool3x3_stride_le3) ? "pooling_layer_optimized_" : "pooling_layer_") - + support::cpp11::to_string(pool_size_x); - _kernel = create_kernel(compile_context, kernel_name, build_opts.options()); - } - else if(pool_info.pool_size == Size2D(2, 2) && pool_type == PoolingType::MAX && indices && is_data_type_float(data_type)) - { - // For max pooling with pool2x2, store indicies which will be used in max unpooling - if(data_type == DataType::F32) - { - std::string kernel_name = "pooling_layer_2_nchw_indices_fp32"; - _kernel = create_kernel(compile_context, kernel_name, build_opts.options()); - } - else if(data_type == DataType::F16) - { - std::string kernel_name = "pooling_layer_2_nchw_indices_fp16"; - _kernel = create_kernel(compile_context, kernel_name, build_opts.options()); - } - } - else // Run general case - { - std::string kernel_name = is_data_type_quantized_asymmetric(data_type) ? "pooling_layer_MxN_quantized_nchw" : "pooling_layer_MxN_nchw"; - _kernel = create_kernel(compile_context, kernel_name, build_opts.options()); - } - break; - } - case DataLayout::NHWC: - { - // Floating point mixed precision is support on F16 only - const auto use_fp_mixed_precision = (data_type == DataType::F16) && pool_info.fp_mixed_precision && pool_type != PoolingType::MAX; - - // Wider accumulation is required to avoid accuracy loss - // Case 1: Floating point mixed precision (fp16 src data and fp32 accumulation) - // Cast 2: Quantized (int8/uint8 src data and int32 accumulation ) - DataType acc_data_type = data_type; - - if(use_fp_mixed_precision) - { - acc_data_type = DataType::F32; - } - else if(is_data_type_quantized(data_type) && pool_type != PoolingType::MAX) - { - acc_data_type = DataType::S32; - } - - build_opts.add_option("-DACC_DATA_TYPE=" + get_cl_type_from_data_type(acc_data_type)); - build_opts.add_option_if(use_fp_mixed_precision, "-DFP_MIXED_PRECISION"); - build_opts.add_option_if(exclude_padding, "-DEXCLUDE_PADDING"); - build_opts.add_option("-DSRC_WIDTH=" + support::cpp11::to_string(src->dimension(idx_width))); - build_opts.add_option("-DSRC_HEIGHT=" + support::cpp11::to_string(src->dimension(idx_height))); - build_opts.add_option("-DDST_HEIGHT=" + support::cpp11::to_string(dst->dimension(idx_height))); - build_opts.add_option("-DDST_CHANNELS=" + support::cpp11::to_string(dst->dimension(idx_channel))); - build_opts.add_option("-DDST_BATCH_SIZE=" + support::cpp11::to_string(dst->dimension(idx_batch_size))); - build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(src->dimension(0) % _num_elems_processed_per_iteration)); - if(pool_info.pool_size == Size2D(2, 2) && is_data_type_float(data_type)) - { - build_opts.add_option_if(indices != nullptr && pool_type == PoolingType::MAX, "-DEXTRACT_MAX_INDEX"); - - std::string kernel_name = "pooling_layer_2x2_nhwc"; - _kernel = create_kernel(compile_context, kernel_name, build_opts.options()); - } - else - { - std::string kernel_name = is_data_type_quantized_asymmetric(data_type) ? "pooling_layer_MxN_quantized_nhwc" : "pooling_layer_MxN_nhwc"; - _kernel = create_kernel(compile_context, kernel_name, build_opts.options()); - } - break; - } - default: - ARM_COMPUTE_ERROR("Not implemented"); - } - - // Set config_id for enabling LWS tuning - _config_id = "pooling_layer_"; - _config_id += lower_string(string_from_data_type(data_type)); - _config_id += "_"; - _config_id += lower_string(string_from_data_layout(_data_layout)); - _config_id += "_"; - _config_id += support::cpp11::to_string(dst->dimension(idx_width)); - _config_id += "_"; - _config_id += support::cpp11::to_string(dst->dimension(idx_height)); - _config_id += "_"; - _config_id += support::cpp11::to_string(dst->dimension(idx_channel)); - _config_id += "_"; - _config_id += lower_string(string_from_data_layout(src->data_layout())); - - ARM_COMPUTE_ERROR_ON(src->data_layout() == DataLayout::NHWC && has_padding_changed(padding_info)); -} - -Status ClPoolingKernel::validate(const ITensorInfo *src, const ITensorInfo *dst, const PoolingLayerInfo &pool_info, const ITensorInfo *indices) -{ - ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, dst, pool_info, indices)); - ARM_COMPUTE_RETURN_ON_ERROR(std::get<0>(validate_and_configure_window(src->clone().get(), dst->clone().get(), pool_info))); - - return Status{}; -} - -void ClPoolingKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) -{ - ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); - ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); - - unsigned int pool_stride_x = 0; - unsigned int pool_stride_y = 0; - std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride(); - - const auto src = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC)); - auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST_0)); - auto indices = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST_1)); - - // Collapse window - Window window_collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ); - - switch(_data_layout) - { - case DataLayout::NCHW: - { - Window slice = window_collapsed.first_slice_window_3D(); - do - { - // Upsample src by pool size - Window in_slice(slice); - in_slice.set(Window::DimX, Window::Dimension(in_slice.x().start() - _pool_info.pad_stride_info.pad_left(), - (in_slice.x().end() - _pool_info.pad_stride_info.pad_left()) * pool_stride_x, - pool_stride_x * _num_elems_processed_per_iteration)); - in_slice.set(Window::DimY, Window::Dimension(in_slice.y().start() - _pool_info.pad_stride_info.pad_top(), - (in_slice.y().end() - _pool_info.pad_stride_info.pad_top()) * pool_stride_y, - pool_stride_y)); - - // Set srcs - unsigned int idx = 0; - add_3D_tensor_argument(idx, src, in_slice); - add_3D_tensor_argument(idx, dst, slice); - if(indices && is_data_type_float(src->info()->data_type()) && (_pool_info.pool_size == Size2D(2, 2))) - { - add_3D_tensor_argument(idx, indices, slice); - } - enqueue(queue, *this, slice, lws_hint()); - } - while(window_collapsed.slide_window_slice_3D(slice)); - break; - } - case DataLayout::NHWC: - { - const size_t batch_size = dst->info()->tensor_shape().total_size_upper(3); - - Window slice = window_collapsed.first_slice_window_4D(); - Window in_slice = window_collapsed.first_slice_window_4D(); - in_slice.set(Window::DimX, Window::Dimension(0, src->info()->dimension(0), _num_elems_processed_per_iteration)); - in_slice.set(Window::DimY, Window::Dimension(0, src->info()->dimension(1), pool_stride_x)); - in_slice.set(Window::DimZ, Window::Dimension(0, src->info()->dimension(2), pool_stride_y)); - in_slice.set(3, Window::Dimension(0, batch_size, 1)); - do - { - // Set srcs - unsigned int idx = 0; - add_4D_tensor_argument(idx, src, in_slice); - add_4D_tensor_argument(idx, dst, slice); - if(indices && is_data_type_float(src->info()->data_type()) && (_pool_info.pool_type == PoolingType::MAX) && (_pool_info.pool_size == Size2D(2, 2))) - { - add_4D_tensor_argument(idx, indices, slice); - } - enqueue(queue, *this, slice, lws_hint()); - } - while(window.slide_window_slice_4D(slice) && window.slide_window_slice_4D(in_slice)); - break; - } - default: - ARM_COMPUTE_ERROR("Not implemented"); - } -} -} // namespace kernels -} // namespace opencl -} // namespace arm_compute diff --git a/src/core/gpu/cl/kernels/ClPoolingKernel.h b/src/core/gpu/cl/kernels/ClPoolingKernel.h deleted file mode 100644 index c1ce859e2c..0000000000 --- a/src/core/gpu/cl/kernels/ClPoolingKernel.h +++ /dev/null @@ -1,79 +0,0 @@ -/* - * Copyright (c) 2017-2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#ifndef ARM_COMPUTE_CL_POOLING_KERNEL_H -#define ARM_COMPUTE_CL_POOLING_KERNEL_H - -#include "src/core/common/Macros.h" -#include "src/core/gpu/cl/ClCompileContext.h" -#include "src/core/gpu/cl/IClKernel.h" - -namespace arm_compute -{ -namespace opencl -{ -namespace kernels -{ -/** Interface for the pooling layer kernel */ -class ClPoolingKernel : public IClKernel -{ -public: - /** Default constructor */ - ClPoolingKernel(); - ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(ClPoolingKernel); - - /** Configure kernel for a given list of arguments - * - * - * @param[in] compile_context The compile context to be used. - * @param[in] src Source tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32. - * @param[out] dst Destination tensor info. Data types supported: same as @p src. - * @param[in] pool_info Contains pooling operation information described in @ref PoolingLayerInfo. - * @param[out] indices (optional) The indices of the maximal values. Data type supported: U32. - */ - void configure(const ClCompileContext &compile_context, ITensorInfo *src, ITensorInfo *dst, const PoolingLayerInfo &pool_info, ITensorInfo *indices = nullptr); - /** Static function to check if given info will lead to a valid configuration of @ref ClPoolingKernel - * - * @param[in] src Source tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32. - * @param[in] dst Destination tensor info. Data types supported: same as @p src. - * @param[in] pool_info Contains pooling operation information described in @ref PoolingLayerInfo. - * @param[in] indices (optional) The indices of the maximal values. Data type supported: U32. - * - * @return a status - */ - static Status validate(const ITensorInfo *src, const ITensorInfo *dst, const PoolingLayerInfo &pool_info, const ITensorInfo *indices = nullptr); - - // Inherited methods overridden: - void run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) override; - BorderSize border_size() const override; - -public: - PoolingLayerInfo _pool_info; - DataLayout _data_layout; - BorderSize _border_size; - unsigned int _num_elems_processed_per_iteration; -}; -} // namespace kernels -} // namespace opencl -} // namespace arm_compute -#endif /*ARM_COMPUTE_CL_POOLING_KERNEL_H */ diff --git a/src/core/gpu/cl/kernels/ClReshapeKernel.cpp b/src/core/gpu/cl/kernels/ClReshapeKernel.cpp deleted file mode 100644 index 4da3fa0e03..0000000000 --- a/src/core/gpu/cl/kernels/ClReshapeKernel.cpp +++ /dev/null @@ -1,131 +0,0 @@ -/* - * Copyright (c) 2017-2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#include "src/core/gpu/cl/kernels/ClReshapeKernel.h" - -#include "arm_compute/core/CL/CLHelpers.h" -#include "arm_compute/core/CL/CLKernelLibrary.h" -#include "arm_compute/core/CL/ICLTensor.h" -#include "arm_compute/core/CL/OpenCL.h" -#include "arm_compute/core/Helpers.h" -#include "arm_compute/core/IAccessWindow.h" -#include "arm_compute/core/TensorInfo.h" -#include "arm_compute/core/Utils.h" -#include "src/core/AccessWindowStatic.h" -#include "src/core/CL/CLValidate.h" -#include "src/core/helpers/WindowHelpers.h" -#include "support/Cast.h" - -#include <string> - -/** [ClReshapeKernel Kernel] **/ -namespace arm_compute -{ -namespace opencl -{ -namespace kernels -{ -namespace -{ -Status validate_arguments(const ITensorInfo *src, const ITensorInfo *dst) -{ - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst); - ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(src); - ARM_COMPUTE_RETURN_ERROR_ON(src->data_type() == DataType::UNKNOWN); - - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, dst); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(src, dst); - ARM_COMPUTE_RETURN_ERROR_ON(src->tensor_shape().total_size() != dst->tensor_shape().total_size()); - - return Status{}; -} -} // namespace - -void ClReshapeKernel::configure(const CLCompileContext &compile_context, const ITensorInfo *src, ITensorInfo *dst) -{ - ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst); - ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, dst)); - - auto padding_info = get_padding_info({ src, dst }); - - // Create kernel - std::set<std::string> build_opts = { "-DDATA_TYPE=" + get_cl_unsigned_type_from_element_size(src->element_size()) }; - _kernel = create_kernel(compile_context, "reshape_layer", build_opts); - - // Add static arguments - const cl_int2 src_shape = - { - { - static_cast<cl_int>(src->tensor_shape()[0]), - static_cast<cl_int>(src->tensor_shape()[1]) - } - }; - const cl_int2 dst_shape = - { - { - static_cast<cl_int>(dst->tensor_shape()[0]), - static_cast<cl_int>(dst->tensor_shape()[1]) - } - }; - unsigned int idx = 2 * num_arguments_per_3D_tensor(); // Skip the src and dst parameters - _kernel.setArg<cl_int2>(idx++, src_shape); - _kernel.setArg<cl_int2>(idx++, dst_shape); - - // Configure kernel window - Window win = calculate_max_window(*src); - - // Set the dst valid region - dst->set_valid_region(ValidRegion(Coordinates(), dst->tensor_shape())); - ICLKernel::configure_internal(win); - - ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info)); -} - -Status ClReshapeKernel::validate(const ITensorInfo *src, const ITensorInfo *dst) -{ - ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, dst)); - - return Status{}; -} - -void ClReshapeKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) -{ - ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); - ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window); - - Window window_collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ); - Window slice = window_collapsed.first_slice_window_3D(); - - const auto src = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC)); - auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST)); - - // Set srcs - unsigned int idx = 0; - add_3D_tensor_argument(idx, src, window_collapsed); - add_3D_tensor_argument(idx, dst, window_collapsed); - enqueue(queue, *this, slice, lws_hint()); -} -} // namespace kernels -} // namespace opencl -} // namespace arm_compute -/** [ClReshapeKernel Kernel] **/ diff --git a/src/core/gpu/cl/kernels/ClReshapeKernel.h b/src/core/gpu/cl/kernels/ClReshapeKernel.h deleted file mode 100644 index ee835c0fd3..0000000000 --- a/src/core/gpu/cl/kernels/ClReshapeKernel.h +++ /dev/null @@ -1,66 +0,0 @@ -/* - * Copyright (c) 2017-2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#ifndef ARM_COMPUTE_CL_RESHAPE_KERNEL_H -#define ARM_COMPUTE_CL_RESHAPE_KERNEL_H - -#include "src/core/common/Macros.h" -#include "src/core/gpu/cl/ClCompileContext.h" -#include "src/core/gpu/cl/IClKernel.h" - -namespace arm_compute -{ -namespace opencl -{ -namespace kernels -{ -/** Interface for the kernel to perform tensor reshaping */ -class ClReshapeKernel : public ICLKernel -{ -public: - ClReshapeKernel() = default; - ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(ClReshapeKernel); - /** Set the src and dst of the kernel - * - * @param[in] compile_context The compile context to be used. - * @param[in] src Source tensor info. Data type supported: All. - * @param[out] dst Destination tensor info. Data type supported: Same as @p src - */ - void configure(const CLCompileContext &compile_context, const ITensorInfo *src, ITensorInfo *dst); - - /** Static function to check if given info will lead to a valid configuration of @ref ClReshapeKernel - * - * @param[in] src Source tensor info. Data type supported: All - * @param[in] dst Destination tensor info. Data type supported: Same as @p src - * - * @return a status - */ - static Status validate(const ITensorInfo *src, const ITensorInfo *dst); - - // Inherited methods overridden: - void run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) override; -}; -} // namespace opencl -} // namespace kernels -} // namespace arm_compute -#endif /*ARM_COMPUTE_CL_RESHAPE_KERNEL_H */ diff --git a/src/core/gpu/cl/kernels/ClWidthConcatenate2TensorsKernel.cpp b/src/core/gpu/cl/kernels/ClWidthConcatenate2TensorsKernel.cpp deleted file mode 100644 index 6a2ab3b50f..0000000000 --- a/src/core/gpu/cl/kernels/ClWidthConcatenate2TensorsKernel.cpp +++ /dev/null @@ -1,156 +0,0 @@ -/* - * Copyright (c) 2018-2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#include "src/core/gpu/cl/kernels/ClWidthConcatenate2TensorsKernel.h" - -#include "arm_compute/core/CL/CLHelpers.h" -#include "arm_compute/core/CL/CLKernelLibrary.h" -#include "arm_compute/core/CL/ICLTensor.h" -#include "arm_compute/core/Utils.h" -#include "arm_compute/core/utils/misc/ShapeCalculator.h" -#include "src/core/CL/CLValidate.h" -#include "src/core/helpers/WindowHelpers.h" -#include "src/core/utils/helpers/tensor_info.h" -#include "support/Cast.h" - -#include "support/StringSupport.h" - -namespace arm_compute -{ -namespace opencl -{ -namespace kernels -{ -namespace -{ -Status validate_arguments(const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst) -{ - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src1, src2, dst); - ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(src1); - ARM_COMPUTE_RETURN_ERROR_ON(src1->data_type() == DataType::UNKNOWN); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src1, src2, dst); - ARM_COMPUTE_RETURN_ERROR_ON(src1->dimension(0) + src2->dimension(0) > dst->dimension(0)); - - for(size_t i = 1; i < Coordinates::num_max_dimensions; ++i) - { - ARM_COMPUTE_RETURN_ERROR_ON(src1->dimension(i) != dst->dimension(i)); - ARM_COMPUTE_RETURN_ERROR_ON(src2->dimension(i) != dst->dimension(i)); - } - ARM_COMPUTE_RETURN_ERROR_ON(src1->num_dimensions() > 4); - - return Status{}; -} -} // namespace - -Status ClWidthConcatenate2TensorsKernel::validate(const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst) -{ - ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src1, src2, dst)); - return Status{}; -} - -void ClWidthConcatenate2TensorsKernel::configure(const CLCompileContext &compile_context, ITensorInfo *src1, ITensorInfo *src2, ITensorInfo *dst) -{ - ARM_COMPUTE_ERROR_ON_NULLPTR(src1, src2, dst); - ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src1, src2, dst)); - - auto padding_info = get_padding_info({ src1, src2, dst }); - - const unsigned int min_dimension = std::min(src1->dimension(0), src2->dimension(0)); - const unsigned int num_elems_processed_per_iteration = adjust_vec_size(8, min_dimension); - const unsigned int vec_size_leftover = dst->dimension(0) % num_elems_processed_per_iteration; - - // Add build options - CLBuildOptions build_opts; - build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(src1->data_type())); - 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(vec_size_leftover)); - build_opts.add_option("-DDEPTH=" + support::cpp11::to_string(src1->dimension(2))); - build_opts.add_option("-DINPUT1_WIDTH=" + support::cpp11::to_string(src1->dimension(0))); - build_opts.add_option("-DINPUT2_WIDTH=" + support::cpp11::to_string(src2->dimension(0))); - build_opts.add_option("-DELEMENT_SIZE=" + support::cpp11::to_string(src1->element_size())); - build_opts.add_option("-DINPUT1_ROTATE_N=" + support::cpp11::to_string((src1->dimension(0) - vec_size_leftover) % num_elems_processed_per_iteration)); - - // 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(dst, src1, src2); - if(is_data_type_quantized_asymmetric(src1->data_type()) && have_different_qinfo) - { - const UniformQuantizationInfo iq1_info = src1->quantization_info().uniform(); - const UniformQuantizationInfo iq2_info = src2->quantization_info().uniform(); - const UniformQuantizationInfo oq_info = dst->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)); - build_opts.add_option("-DOFFSET_IN2=" + float_to_string_with_full_precision(iq2_info.offset)); - build_opts.add_option("-DSCALE_IN2=" + float_to_string_with_full_precision(iq2_info.scale)); - build_opts.add_option("-DOFFSET_OUT=" + float_to_string_with_full_precision(oq_info.offset)); - build_opts.add_option("-DSCALE_OUT=" + float_to_string_with_full_precision(oq_info.scale)); - } - - // Create kernel - _kernel = create_kernel(compile_context, "concatenate_width_x2", build_opts.options()); - - // Configure kernel window - Window win = calculate_max_window(*dst, Steps(num_elems_processed_per_iteration)); - ICLKernel::configure_internal(win.collapse(win, Window::DimZ)); - - // Set dst valid region - dst->set_valid_region(ValidRegion(Coordinates(), dst->tensor_shape())); - ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info)); - - // Set config_id for enabling LWS tuning - _config_id = "concatenate_width_x2_"; - _config_id += lower_string(string_from_data_type(src1->data_type())); - _config_id += "_"; - _config_id += support::cpp11::to_string(src1->dimension(0)); - _config_id += "_"; - _config_id += support::cpp11::to_string(src1->dimension(1)); - _config_id += "_"; - _config_id += support::cpp11::to_string(src2->dimension(0)); - _config_id += "_"; - _config_id += support::cpp11::to_string(src2->dimension(1)); -} - -void ClWidthConcatenate2TensorsKernel::run_op(ITensorPack &tensors, 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_4D(); - - const auto src0 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_VEC)); - const auto src1 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_VEC + 1)); - auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST)); - - do - { - unsigned int idx = 0; - add_4D_tensor_argument(idx, src0, slice); - add_4D_tensor_argument(idx, src1, slice); - add_4D_tensor_argument(idx, dst, slice); - enqueue(queue, *this, window, lws_hint()); - } - while(window.slide_window_slice_4D(slice)); -} -} // namespace kernels -} // namespace opencl -} // namespace arm_compute diff --git a/src/core/gpu/cl/kernels/ClWidthConcatenate2TensorsKernel.h b/src/core/gpu/cl/kernels/ClWidthConcatenate2TensorsKernel.h deleted file mode 100644 index ddade29113..0000000000 --- a/src/core/gpu/cl/kernels/ClWidthConcatenate2TensorsKernel.h +++ /dev/null @@ -1,70 +0,0 @@ -/* - * Copyright (c) 2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#ifndef ARM_COMPUTE_CL_WIDTHCONCATENATE_2TENSORS_KERNEL_H -#define ARM_COMPUTE_CL_WIDTHCONCATENATE_2TENSORS_KERNEL_H - -#include "src/core/common/Macros.h" -#include "src/core/gpu/cl/ClCompileContext.h" -#include "src/core/gpu/cl/IClKernel.h" - -namespace arm_compute -{ -namespace opencl -{ -namespace kernels -{ -/** Interface for the width concatenate kernel of 2 tensors. - * The src1 and src2 tensors will be concatenated into the dst tensor. - */ -class ClWidthConcatenate2TensorsKernel : public IClKernel -{ -public: - /** Default constructor */ - ClWidthConcatenate2TensorsKernel() = default; - ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(ClWidthConcatenate2TensorsKernel); - /** Initialise the kernel's sources and destination - * - * @param[in] compile_context The compile context to be used. - * @param[in] src1 First source tensor info. Data types supported: All. - * @param[in] src2 Second source tensor info. Data types supported: same as @p src1 - * @param[out] dst Destination tensor info. Data types supported: Same as @p src1. - */ - void configure(const CLCompileContext &compile_context, ITensorInfo *src1, ITensorInfo *src2, ITensorInfo *dst); - /** Static function to check if given info will lead to a valid configuration of @ref ClWidthConcatenate2TensorsKernel - * - * @param[in] src1 First tensor info. Data types supported: All. - * @param[in] src2 Second tensor info. Data types supported: same as @p src1 - * @param[in] dst Destination tensor info. Data types supported: Same as @p src1. - * - * @return a status - */ - static Status validate(const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst); - - // Inherited methods overridden: - void run_op(ITensorPack &tensors, const Window &window, ::cl::CommandQueue &queue) override; -}; -} // namespace kernels -} // namespace opencl -} // namespace arm_compute -#endif /* ARM_COMPUTE_CL_WIDTH_CONCATENATE_2TENSORS_KERNEL_H */ diff --git a/src/core/gpu/cl/kernels/ClWidthConcatenate4TensorsKernel.cpp b/src/core/gpu/cl/kernels/ClWidthConcatenate4TensorsKernel.cpp deleted file mode 100644 index 4b49652a73..0000000000 --- a/src/core/gpu/cl/kernels/ClWidthConcatenate4TensorsKernel.cpp +++ /dev/null @@ -1,186 +0,0 @@ -/* - * Copyright (c) 2018-2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#include "src/core/gpu/cl/kernels/ClWidthConcatenate4TensorsKernel.h" - -#include "arm_compute/core/CL/CLHelpers.h" -#include "arm_compute/core/CL/CLKernelLibrary.h" -#include "arm_compute/core/CL/ICLTensor.h" -#include "arm_compute/core/Helpers.h" -#include "arm_compute/core/Utils.h" -#include "src/core/CL/CLValidate.h" -#include "src/core/helpers/WindowHelpers.h" -#include "src/core/utils/helpers/tensor_info.h" -#include "support/Cast.h" - -#include "support/StringSupport.h" - -namespace arm_compute -{ -namespace opencl -{ -namespace kernels -{ -namespace -{ -Status validate_arguments(const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *src3, const ITensorInfo *src4, const ITensorInfo *dst) -{ - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src1, src2, src3, src4, dst); - ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(src1); - ARM_COMPUTE_RETURN_ERROR_ON(src1->data_type() == DataType::UNKNOWN); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src1, src2, src3, src4, dst); - ARM_COMPUTE_RETURN_ERROR_ON(src1->dimension(0) + src2->dimension(0) + src3->dimension(0) + src4->dimension(0) > dst->dimension(0)); - - for(size_t i = 1; i < Coordinates::num_max_dimensions; ++i) - { - ARM_COMPUTE_RETURN_ERROR_ON(src1->dimension(i) != dst->dimension(i)); - ARM_COMPUTE_RETURN_ERROR_ON(src2->dimension(i) != dst->dimension(i)); - ARM_COMPUTE_RETURN_ERROR_ON(src3->dimension(i) != dst->dimension(i)); - ARM_COMPUTE_RETURN_ERROR_ON(src4->dimension(i) != dst->dimension(i)); - } - ARM_COMPUTE_RETURN_ERROR_ON(src1->num_dimensions() > 4); - - return Status{}; -} -} // namespace - -ClWidthConcatenate4TensorsKernel::ClWidthConcatenate4TensorsKernel() -{ -} - -Status ClWidthConcatenate4TensorsKernel::validate(const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *src3, const ITensorInfo *src4, const ITensorInfo *dst) -{ - ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src1, src2, src3, src4, dst)); - return Status{}; -} - -void ClWidthConcatenate4TensorsKernel::configure(const CLCompileContext &compile_context, - ITensorInfo *src1, ITensorInfo *src2, - ITensorInfo *src3, ITensorInfo *src4, - ITensorInfo *dst) -{ - ARM_COMPUTE_ERROR_ON_NULLPTR(src1, src2, src3, src4, dst); - ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src1, src2, src3, src4, dst)); - - auto padding_info = get_padding_info({ src1, src2, src3, src4, dst }); - const unsigned int min_dimension = std::min(std::min(src1->dimension(0), src2->dimension(0)), std::min(src3->dimension(0), src4->dimension(0))); - const unsigned int num_elems_processed_per_iteration = adjust_vec_size(8, min_dimension); - const unsigned int vec_size_leftover = dst->dimension(0) % num_elems_processed_per_iteration; - - // Add build options - CLBuildOptions build_opts; - build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(src1->data_type())); - 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(vec_size_leftover)); - build_opts.add_option("-DDEPTH=" + support::cpp11::to_string(src1->dimension(2))); - build_opts.add_option("-DINPUT1_WIDTH=" + support::cpp11::to_string(src1->dimension(0))); - build_opts.add_option("-DINPUT2_WIDTH=" + support::cpp11::to_string(src2->dimension(0))); - build_opts.add_option("-DINPUT3_WIDTH=" + support::cpp11::to_string(src3->dimension(0))); - build_opts.add_option("-DINPUT4_WIDTH=" + support::cpp11::to_string(src4->dimension(0))); - build_opts.add_option("-DELEMENT_SIZE=" + support::cpp11::to_string(src1->element_size())); - build_opts.add_option("-DINPUT1_ROTATE_N=" + support::cpp11::to_string((src1->dimension(0) - vec_size_leftover) % num_elems_processed_per_iteration)); - build_opts.add_option("-DINPUT2_ROTATE_N=" + support::cpp11::to_string((src1->dimension(0) + src2->dimension(0) - vec_size_leftover) % num_elems_processed_per_iteration)); - build_opts.add_option("-DINPUT3_ROTATE_N=" + support::cpp11::to_string((src1->dimension(0) + src2->dimension(0) + src3->dimension(0) - vec_size_leftover) % num_elems_processed_per_iteration)); - - // If soources 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(dst, src1, src2, src3, src4); - if(is_data_type_quantized_asymmetric(src1->data_type()) && have_different_qinfo) - { - const UniformQuantizationInfo iq1_info = src1->quantization_info().uniform(); - const UniformQuantizationInfo iq2_info = src2->quantization_info().uniform(); - const UniformQuantizationInfo iq3_info = src3->quantization_info().uniform(); - const UniformQuantizationInfo iq4_info = src4->quantization_info().uniform(); - const UniformQuantizationInfo oq_info = dst->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)); - build_opts.add_option("-DOFFSET_IN2=" + float_to_string_with_full_precision(iq2_info.offset)); - build_opts.add_option("-DSCALE_IN2=" + float_to_string_with_full_precision(iq2_info.scale)); - build_opts.add_option("-DOFFSET_IN3=" + float_to_string_with_full_precision(iq3_info.offset)); - build_opts.add_option("-DSCALE_IN3=" + float_to_string_with_full_precision(iq3_info.scale)); - build_opts.add_option("-DOFFSET_IN4=" + float_to_string_with_full_precision(iq4_info.offset)); - build_opts.add_option("-DSCALE_IN4=" + float_to_string_with_full_precision(iq4_info.scale)); - build_opts.add_option("-DOFFSET_OUT=" + float_to_string_with_full_precision(oq_info.offset)); - build_opts.add_option("-DSCALE_OUT=" + float_to_string_with_full_precision(oq_info.scale)); - } - - // Create kernel - _kernel = create_kernel(compile_context, "concatenate_width_x4", build_opts.options()); - - // Configure kernel window - Window win = calculate_max_window(*dst, Steps(num_elems_processed_per_iteration)); - ICLKernel::configure_internal(win.collapse(win, Window::DimZ)); - - // Set dst valid region - dst->set_valid_region(ValidRegion(Coordinates(), dst->tensor_shape())); - ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info)); - - // Set config_id for enabling LWS tuning - _config_id = "concatenate_width_x4_"; - _config_id += lower_string(string_from_data_type(src1->data_type())); - _config_id += "_"; - _config_id += support::cpp11::to_string(src1->dimension(0)); - _config_id += "_"; - _config_id += support::cpp11::to_string(src1->dimension(1)); - _config_id += "_"; - _config_id += support::cpp11::to_string(src2->dimension(0)); - _config_id += "_"; - _config_id += support::cpp11::to_string(src2->dimension(1)); - _config_id += "_"; - _config_id += support::cpp11::to_string(src3->dimension(0)); - _config_id += "_"; - _config_id += support::cpp11::to_string(src3->dimension(1)); - _config_id += "_"; - _config_id += support::cpp11::to_string(src4->dimension(0)); - _config_id += "_"; - _config_id += support::cpp11::to_string(src4->dimension(1)); -} - -void ClWidthConcatenate4TensorsKernel::run_op(ITensorPack &tensors, 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<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_VEC)); - const auto src1 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_VEC + 1)); - const auto src2 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_VEC + 2)); - const auto src3 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_VEC + 3)); - auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST)); - - Window slice = window.first_slice_window_4D(); - - do - { - unsigned int idx = 0; - 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)); -} -} // namespace kernels -} // namespace opencl -} // namespace arm_compute diff --git a/src/core/gpu/cl/kernels/ClWidthConcatenate4TensorsKernel.h b/src/core/gpu/cl/kernels/ClWidthConcatenate4TensorsKernel.h deleted file mode 100644 index 19bda65902..0000000000 --- a/src/core/gpu/cl/kernels/ClWidthConcatenate4TensorsKernel.h +++ /dev/null @@ -1,75 +0,0 @@ -/* - * Copyright (c) 2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ - -#ifndef ARM_COMPUTE_CL_WIDTH_CONCATENATE_4TENSORS_KERNEL_H -#define ARM_COMPUTE_CL_WIDTH_CONCATENATE_4TENSORS_KERNEL_H - -#include "src/core/common/Macros.h" -#include "src/core/gpu/cl/ClCompileContext.h" -#include "src/core/gpu/cl/IClKernel.h" - -namespace arm_compute -{ -namespace opencl -{ -namespace kernels -{ -/** Interface for the width concatenate kernel of 4 tensors. - * All source tensors will be concatenated into the destination tensor. - */ -class ClWidthConcatenate4TensorsKernel : public IClKernel -{ -public: - /** Default constructor */ - ClWidthConcatenate4TensorsKernel(); - ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(ClWidthConcatenate4TensorsKernel); - /** Initialise the kernel's sources and destination - * - * @param[in] compile_context The compile context to be used. - * @param[in] src1 First source tensor info. Data types supported: All. - * @param[in] src2 Second source tensor info. Data types supported: same as @p src1 - * @param[in] src3 Third source tensor info. Data types supported: same as @p src1 - * @param[in] src4 Fourth source tensor info. Data types supported: same as @p src1 - * @param[out] dst Destination tensor info. Data types supported: same as @p src1. - */ - void configure(const CLCompileContext &compile_context, ITensorInfo *src1, ITensorInfo *src2, ITensorInfo *src3, ITensorInfo *src4, ITensorInfo *dst); - /** Static function to check if given info will lead to a valid configuration of @ref ClWidthConcatenate4TensorsKernel - * - * @param[in] src1 First tensor info. Data types supported: All. - * @param[in] src2 Second tensor info. Data types supported: same as @p src1 - * @param[in] src3 Third tensor info. Data types supported: same as @p src1 - * @param[in] src4 Fourth tensor info. Data types supported: same as @p src1 - * @param[in] dst Destination tensor info. Data types supported: same as @p src1. - * - * @return a status - */ - static Status validate(const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *src3, const ITensorInfo *src4, const ITensorInfo *dst); - - // Inherited methods overridden: - void run_op(ITensorPack &tensors, const Window &window, ::cl::CommandQueue &queue) override; -}; -} // namespace kernels -} // namespace opencl -} // namespace arm_compute -#endif /* ARM_COMPUTE_CL_WIDTH_CONCATENATE_4TENSORS_KERNEL_H */ diff --git a/src/core/gpu/cl/kernels/ClWidthConcatenateKernel.cpp b/src/core/gpu/cl/kernels/ClWidthConcatenateKernel.cpp deleted file mode 100644 index 8cbbc27444..0000000000 --- a/src/core/gpu/cl/kernels/ClWidthConcatenateKernel.cpp +++ /dev/null @@ -1,129 +0,0 @@ -/* - * Copyright (c) 2018-2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#include "src/core/gpu/cl/kernels/ClWidthConcatenateKernel.h" - -#include "arm_compute/core/CL/CLHelpers.h" -#include "arm_compute/core/CL/CLKernelLibrary.h" -#include "arm_compute/core/CL/ICLTensor.h" -#include "arm_compute/core/Helpers.h" -#include "arm_compute/core/Utils.h" -#include "src/core/CL/CLValidate.h" -#include "src/core/helpers/WindowHelpers.h" -#include "support/Cast.h" - -#include "support/StringSupport.h" - -namespace arm_compute -{ -namespace opencl -{ -namespace kernels -{ -namespace -{ -Status validate_arguments(const ITensorInfo *src, unsigned int width_offset, const ITensorInfo *dst) -{ - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst); - ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(src); - ARM_COMPUTE_RETURN_ERROR_ON(src->data_type() == DataType::UNKNOWN); - - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, dst); - ARM_COMPUTE_RETURN_ERROR_ON(src->dimension(0) + width_offset > dst->dimension(0)); - - for(size_t i = 1; i < Coordinates::num_max_dimensions; ++i) - { - ARM_COMPUTE_RETURN_ERROR_ON(src->dimension(i) != dst->dimension(i)); - } - ARM_COMPUTE_RETURN_ERROR_ON(src->num_dimensions() > 4); - - return Status{}; -} -} // namespace - -ClWidthConcatenateKernel::ClWidthConcatenateKernel() -{ -} - -Status ClWidthConcatenateKernel::validate(const ITensorInfo *src, unsigned int width_offset, const ITensorInfo *dst) -{ - ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, width_offset, dst)); - return Status{}; -} - -void ClWidthConcatenateKernel::configure(const CLCompileContext &compile_context, ITensorInfo *src, unsigned int width_offset, ITensorInfo *dst) -{ - ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst); - ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, width_offset, dst)); - - auto padding_info = get_padding_info({ src, dst }); - - const unsigned int num_elems_processed_per_iteration = adjust_vec_size(16, src->dimension(0)); - - // Add build options - CLBuildOptions build_opts; - build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(src->data_type())); - 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(src->dimension(0) % num_elems_processed_per_iteration)); - build_opts.add_option("-DWIDTH_OFFSET=" + support::cpp11::to_string(width_offset)); - build_opts.add_option("-DDEPTH=" + support::cpp11::to_string(src->dimension(2))); - - if(is_data_type_quantized_asymmetric(src->data_type()) && src->quantization_info() != dst->quantization_info()) - { - const UniformQuantizationInfo iqinfo = src->quantization_info().uniform(); - const UniformQuantizationInfo oqinfo = dst->quantization_info().uniform(); - - build_opts.add_option("-DOFFSET_IN1=" + float_to_string_with_full_precision(iqinfo.offset)); - build_opts.add_option("-DOFFSET_OUT=" + float_to_string_with_full_precision(oqinfo.offset)); - build_opts.add_option("-DSCALE_IN1=" + float_to_string_with_full_precision(iqinfo.scale)); - build_opts.add_option("-DSCALE_OUT=" + float_to_string_with_full_precision(oqinfo.scale)); - } - - // Create kernel - _kernel = create_kernel(compile_context, "concatenate_width", build_opts.options()); - // Configure kernel window - Window win = calculate_max_window(*src, Steps(num_elems_processed_per_iteration)); - ICLKernel::configure_internal(win.collapse(win, Window::DimZ)); - - // Set dst valid region - dst->set_valid_region(ValidRegion(Coordinates(), dst->tensor_shape())); - - ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info)); -} - -void ClWidthConcatenateKernel::run_op(ITensorPack &tensors, const Window &window, ::cl::CommandQueue &queue) -{ - ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); - ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); - - const auto src = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC)); - auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST)); - - unsigned int idx = 0; - add_4D_tensor_argument(idx, src, window); - add_4D_tensor_argument(idx, dst, window); - enqueue(queue, *this, window, lws_hint()); -} -} // namespace kernels -} // namespace opencl -} // namespace arm_compute diff --git a/src/core/gpu/cl/kernels/ClWidthConcatenateKernel.h b/src/core/gpu/cl/kernels/ClWidthConcatenateKernel.h deleted file mode 100644 index 6bc8e57a08..0000000000 --- a/src/core/gpu/cl/kernels/ClWidthConcatenateKernel.h +++ /dev/null @@ -1,71 +0,0 @@ -/* - * Copyright (c) 2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#ifndef ARM_COMPUTE_CL_WIDTH_CONCATENATE_LAYER_KERNEL_H -#define ARM_COMPUTE_CL_WIDTH_CONCATENATE_LAYER_KERNEL_H - -#include "src/core/common/Macros.h" -#include "src/core/gpu/cl/ClCompileContext.h" -#include "src/core/gpu/cl/IClKernel.h" - -namespace arm_compute -{ -namespace opencl -{ -namespace kernels -{ -/** Interface for the width concatenate kernel. - * The source tensor will be concatenated into the destination tensor. - */ -class ClWidthConcatenateKernel : public IClKernel -{ -public: - /** Default constructor */ - ClWidthConcatenateKernel(); - ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(ClWidthConcatenateKernel); - /** Initialise the kernel's source and destination - * - * @param[in] compile_context The compile context to be used. - * @param[in] src Source tensor info. Data types supported: All. - * @param[in] width_offset The offset on the X axis. - * @param[in,out] dst Destination tensor info. Data types supported: same as @p src. - * - */ - void configure(const CLCompileContext &compile_context, ITensorInfo *src, unsigned int width_offset, ITensorInfo *dst); - /** Static function to check if given info will lead to a valid configuration of @ref ClWidthConcatenateKernel - * - * @param[in] src Source tensor info. Data types supported: All. - * @param[in] width_offset The offset on the X axis. - * @param[in] dst Destination tensor info. Data types supported: same as @p src. - * - * @return a status - */ - static Status validate(const ITensorInfo *src, unsigned int width_offset, const ITensorInfo *dst); - - // Inherited methods overridden: - void run_op(ITensorPack &tensors, const Window &window, ::cl::CommandQueue &queue) override; -}; -} // namespace kernels -} // namespace opencl -} // namespace arm_compute -#endif /* ARM_COMPUTE_CL_WIDTH_CONCATENATE_LAYER_KERNEL_H */ |