diff options
Diffstat (limited to 'arm_compute')
| -rw-r--r-- | arm_compute/core/ITensorPack.h | 7 | ||||
| -rw-r--r-- | arm_compute/core/KernelDescriptors.h | 1 | ||||
| -rw-r--r-- | arm_compute/core/experimental/Types.h | 1 | ||||
| -rw-r--r-- | arm_compute/runtime/CL/functions/CLSoftmaxLayer.h | 40 |
4 files changed, 20 insertions, 29 deletions
diff --git a/arm_compute/core/ITensorPack.h b/arm_compute/core/ITensorPack.h index c06e1d9a73..8aea880bb6 100644 --- a/arm_compute/core/ITensorPack.h +++ b/arm_compute/core/ITensorPack.h @@ -69,6 +69,13 @@ public: * @param[in] tensor Tensor to add */ void add_tensor(int id, const ITensor *tensor); + + /** Add const tensor to the pack + * + * @param[in] id ID/type of the tensor to add + * @param[in] tensor Tensor to add + */ + void add_const_tensor(int id, const ITensor *tensor); /** Get tensor of a given id from the pac * * @param[in] id ID of tensor to extract diff --git a/arm_compute/core/KernelDescriptors.h b/arm_compute/core/KernelDescriptors.h index e381220695..1f3cee2dd1 100644 --- a/arm_compute/core/KernelDescriptors.h +++ b/arm_compute/core/KernelDescriptors.h @@ -114,6 +114,7 @@ struct SoftmaxKernelInfo float beta{ 1.f }; /**< A scaling factor for the exponent with default value 1.0 */ bool is_log{ false }; /**< Flag used to perform Log Softmax operation */ DataType input_data_type{ DataType::UNKNOWN }; /**< Input tensor data type */ + int32_t axis{ 0 }; /**< The dimension in which to apply softmax. */ }; /** Descriptor used by the direct convolution layer output stage kernels */ diff --git a/arm_compute/core/experimental/Types.h b/arm_compute/core/experimental/Types.h index f615678e31..2a4bd89385 100644 --- a/arm_compute/core/experimental/Types.h +++ b/arm_compute/core/experimental/Types.h @@ -52,6 +52,7 @@ enum TensorType : int32_t ACL_INT_1 = 51, ACL_INT_2 = 52, ACL_INT_3 = 53, + ACL_INT_4 = 54, ACL_SRC_VEC = 256, }; diff --git a/arm_compute/runtime/CL/functions/CLSoftmaxLayer.h b/arm_compute/runtime/CL/functions/CLSoftmaxLayer.h index ab10a64de4..ddb35ae56f 100644 --- a/arm_compute/runtime/CL/functions/CLSoftmaxLayer.h +++ b/arm_compute/runtime/CL/functions/CLSoftmaxLayer.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2020 Arm Limited. + * Copyright (c) 2017-2021 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -24,8 +24,6 @@ #ifndef ARM_COMPUTE_CLSOFTMAXLAYER_H #define ARM_COMPUTE_CLSOFTMAXLAYER_H -#include "arm_compute/runtime/CL/CLTensor.h" -#include "arm_compute/runtime/CL/functions/CLPermute.h" #include "arm_compute/runtime/IFunction.h" #include "arm_compute/runtime/IMemoryManager.h" #include "arm_compute/runtime/MemoryGroup.h" @@ -34,11 +32,9 @@ namespace arm_compute { -class CLCompileContext; -class CLLogits1DMaxShiftExpSumKernel; -class CLLogits1DNormKernel; class ICLTensor; class ITensorInfo; +class CLCompileContext; /** Basic function to compute a SoftmaxLayer. * @@ -48,11 +44,11 @@ class ITensorInfo; * Log Softmax is calculated by : * @f[ out = (x - max(x) * beta) - log(\sum{e^{x - max(x) * beta}}) @f] * - * This function runs the following kernels: + * This function runs the following operators/kernels: * -# If axis is not 0: - * -# @ref CLPermute - * -# @ref CLLogits1DNormKernel - * -# @ref CLLogits1DMaxShiftExpSumKernel + * -# @ref opencl::ClPermute + * -# @ref opencl::kernels::ClLogits1DNormKernel + * -# @ref opencl::kernels::ClLogits1DMaxShiftExpSumKernel */ template <bool IS_LOG = false> class CLSoftmaxLayerGeneric : public IFunction @@ -60,14 +56,6 @@ class CLSoftmaxLayerGeneric : public IFunction public: /** Constructor */ CLSoftmaxLayerGeneric(std::shared_ptr<IMemoryManager> memory_manager = nullptr); - /** Prevent instances of this class from being copied */ - CLSoftmaxLayerGeneric(const CLSoftmaxLayerGeneric &) = delete; - /** Prevent instances of this class from being copied */ - CLSoftmaxLayerGeneric &operator=(const CLSoftmaxLayerGeneric &) = delete; - /** Prevent instances of this class to be moved */ - CLSoftmaxLayerGeneric(CLSoftmaxLayerGeneric &&) = delete; - /** Prevent instances of this class to be moved */ - CLSoftmaxLayerGeneric &operator=(CLSoftmaxLayerGeneric &&) = delete; /** Default destructor */ ~CLSoftmaxLayerGeneric(); /** Set the input and output tensors. @@ -105,17 +93,11 @@ public: void run() override; private: - MemoryGroup _memory_group; - CLPermute _permute_input; - CLPermute _permute_output; - std::unique_ptr<CLLogits1DMaxShiftExpSumKernel> _max_shift_exp_sum_kernel; - std::unique_ptr<CLLogits1DNormKernel> _norm_kernel; - CLTensor _max; - CLTensor _sum; - CLTensor _tmp; - CLTensor _input_permuted; - CLTensor _output_permuted; - bool _needs_permute; + struct Impl; + std::unique_ptr<Impl> _impl; + + /** Allocate workspace required by the operator */ + void allocate_workspace(); }; using CLSoftmaxLayer = CLSoftmaxLayerGeneric<false>; |