diff options
Diffstat (limited to 'runtimes/libs/ARMComputeEx/arm_compute/runtime/misc/functions')
3 files changed, 283 insertions, 0 deletions
diff --git a/runtimes/libs/ARMComputeEx/arm_compute/runtime/misc/functions/GenericGather.h b/runtimes/libs/ARMComputeEx/arm_compute/runtime/misc/functions/GenericGather.h new file mode 100644 index 000000000..3db0c7e5e --- /dev/null +++ b/runtimes/libs/ARMComputeEx/arm_compute/runtime/misc/functions/GenericGather.h @@ -0,0 +1,84 @@ +/* + * Copyright (c) 2019 Samsung Electronics Co., Ltd. All Rights Reserved + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/** + * @file GenericGather.h + * @brief This file contains GenericGather class + * @ingroup COM_AI_RUNTIME + */ + +#ifndef __ARM_COMPUTE_MISC_GENERIC_GATHER_H__ +#define __ARM_COMPUTE_MISC_GENERIC_GATHER_H__ + +#include <arm_compute/runtime/Tensor.h> +#include <arm_compute/runtime/CL/CLTensor.h> + +#include <arm_compute/runtime/CL/functions/CLPermute.h> +#include <arm_compute/runtime/CL/functions/CLGatherEx.h> + +#include "Utils.h" + +namespace arm_compute +{ +namespace misc +{ + +/** + * @brief Class to run Gather with both CPU and GPU + */ +class GenericGather : public arm_compute::IFunction +{ +public: + GenericGather(void) + { + // DO NOTHING + } + +public: + /** + * @brief Configure the layer + * @param[in] input The source tensor + * @param[in] indices The indices tensor + * @param[in] output The destination tensor + * @param[in] axis (Optional) The axis in input to gather indices from + * @return N/A + */ + void configure(arm_compute::ITensor *input, arm_compute::ITensor *indices, + arm_compute::ITensor *output, int axis = 0); + +public: + /** + * @brief Run the operation. Must be called after configure(). + * @return N/A + */ + void run(void) override; + +private: + arm_compute::ITensor *_input{nullptr}; + arm_compute::ITensor *_indices{nullptr}; + arm_compute::ITensor *_output{nullptr}; + int _axis{0}; + arm_compute::CLTensor _cl_permuted; + +private: + arm_compute::CLPermute _cl_permute; + arm_compute::CLGatherEx _cl_gather; +}; + +} // namespace misc +} // namespace arm_compute + +#endif // __ARM_COMPUTE_MISC_GENERIC_GATHER_H__ diff --git a/runtimes/libs/ARMComputeEx/arm_compute/runtime/misc/functions/GenericReshapeLayer.h b/runtimes/libs/ARMComputeEx/arm_compute/runtime/misc/functions/GenericReshapeLayer.h new file mode 100644 index 000000000..ab2fdc71d --- /dev/null +++ b/runtimes/libs/ARMComputeEx/arm_compute/runtime/misc/functions/GenericReshapeLayer.h @@ -0,0 +1,87 @@ +/* + * Copyright (c) 2018 Samsung Electronics Co., Ltd. All Rights Reserved + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/** + * @file GenericReshapeLayer.h + * @brief This file contains GenericReshapeLayer class + * @ingroup COM_AI_RUNTIME + */ + +#ifndef __ARM_COMPUTE_MISC_GENERIC_RESHAPE_LAYER_H__ +#define __ARM_COMPUTE_MISC_GENERIC_RESHAPE_LAYER_H__ + +#include <arm_compute/runtime/Tensor.h> +#include <arm_compute/runtime/CL/CLTensor.h> + +#include <arm_compute/runtime/CL/functions/CLPermute.h> +#include <arm_compute/runtime/CL/functions/CLReshapeLayer.h> +#include <arm_compute/runtime/NEON/functions/NEPermute.h> +#include <arm_compute/runtime/NEON/functions/NEReshapeLayer.h> + +#include "Utils.h" + +namespace arm_compute +{ +namespace misc +{ + +/** + * @brief Class to run Reshape Layer with both CPU and GPU + */ +class GenericReshapeLayer : public arm_compute::IFunction +{ +public: + GenericReshapeLayer(void) + : _input(nullptr), _output(nullptr), _cl_permuted{}, _neon_permuted{}, _cl_permute{}, + _cl_reshape{}, _neon_permute{}, _neon_reshape{} + { + // DO NOTHING + } + +public: + /** + * @brief Configure the layer + * @param[in] input The source tensor + * @param[in] output The destination tensor + * @return N/A + */ + void configure(const arm_compute::ITensor *input, arm_compute::ITensor *output); + +public: + /** + * @brief Run the operation. Must be called after configure(). + * @return N/A + */ + void run(void) override; + +private: + const arm_compute::ITensor *_input; + arm_compute::ITensor *_output; + arm_compute::CLTensor _cl_permuted; + arm_compute::Tensor _neon_permuted; + +private: + arm_compute::CLPermute _cl_permute; + arm_compute::CLReshapeLayer _cl_reshape; + + arm_compute::NEPermute _neon_permute; + arm_compute::NEReshapeLayer _neon_reshape; +}; + +} // namespace misc +} // namespace arm_compute + +#endif // __ARM_COMPUTE_MISC_GENERIC_RESHAPE_LAYER_H__ diff --git a/runtimes/libs/ARMComputeEx/arm_compute/runtime/misc/functions/Utils.h b/runtimes/libs/ARMComputeEx/arm_compute/runtime/misc/functions/Utils.h new file mode 100644 index 000000000..53736f55f --- /dev/null +++ b/runtimes/libs/ARMComputeEx/arm_compute/runtime/misc/functions/Utils.h @@ -0,0 +1,112 @@ +/* + * Copyright (c) 2019 Samsung Electronics Co., Ltd. All Rights Reserved + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/** + * @file utils.h + * @ingroup COM_AI_RUNTIME + * @brief This file contains utils for arm compute library + */ +#ifndef __ARM_COMPUTE_MISC_UTILS_H__ +#define __ARM_COMPUTE_MISC_UTILS_H__ + +#include <string> +#include <cassert> +#include <arm_compute/runtime/CL/CLTensor.h> + +#include <arm_compute/core/Coordinates.h> +#include <arm_compute/core/TensorInfo.h> +#include <arm_compute/core/TensorShape.h> +#include <arm_compute/core/Types.h> + +// TODO : It should be extracted to independent module. + +namespace arm_compute +{ +namespace misc +{ +namespace utils +{ + +/** + * @brief Check if this runtime runs on GPU or NEON + * @return @c true if GPU mode, otherwise @c false + */ +bool isGpuMode(); + +#ifndef CAST_CL +#define CAST_CL(tensor) static_cast<::arm_compute::CLTensor *>(tensor) +#endif + +#ifndef CAST_NE +#define CAST_NE(tensor) static_cast<::arm_compute::Tensor *>(tensor) +#endif + +/** +* @brief Generate arm compute permutation vector from runtime permutation vector +* @param[in] rank Rank number supported upto 4 +* @param[in] runtime_pv Integer array for runtime permutation vector +* @return Permutation vector of arm compute +*/ +arm_compute::PermutationVector getARMComputePermutationVector(uint32_t rank, + const int32_t *runtime_pv); + +/** + * @brief Set value to arm compute tensor with casting + * @param[in] value Value to set + * @param[out] to Target tensor of arm compute + * @param[in] id Position of element + * @return N/A + */ +template <typename FromT> +void copyCast(const FromT value, arm_compute::ITensor *to, const arm_compute::Coordinates &id) +{ + switch (to->info()->data_type()) + { + case arm_compute::DataType::F32: + { + *reinterpret_cast<float *>(to->ptr_to_element(id)) = static_cast<float>(value); + break; + } + case arm_compute::DataType::S32: + { + *reinterpret_cast<int32_t *>(to->ptr_to_element(id)) = static_cast<int32_t>(value); + break; + } + case arm_compute::DataType::U32: + { + *reinterpret_cast<uint32_t *>(to->ptr_to_element(id)) = static_cast<uint32_t>(value); + break; + } + case arm_compute::DataType::QASYMM8: + { + float realValue = static_cast<float>(value); + // NOTE We haven't known the policy of rounding for quantization. + // So this is set to a temporary value. + *(to->ptr_to_element(id)) = + to->info()->quantization_info().quantize(realValue, arm_compute::RoundingPolicy::TO_ZERO); + break; + } + default: + throw std::runtime_error("Not supported, yet"); + break; + } +} + +} // namespace utils +} // namespace misc +} // namespace arm_compute + +#endif // __ARM_COMPUTE_MISC_UTILS_H__ |