diff options
-rw-r--r-- | arm_compute/core/Types.h | 20 | ||||
-rw-r--r-- | arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h | 34 | ||||
-rw-r--r-- | arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h | 1 | ||||
-rw-r--r-- | src/runtime/CL/functions/CLFullyConnectedLayer.cpp | 12 | ||||
-rw-r--r-- | src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp | 25 |
5 files changed, 58 insertions, 34 deletions
diff --git a/arm_compute/core/Types.h b/arm_compute/core/Types.h index 46e6dba1a0..639170f0fd 100644 --- a/arm_compute/core/Types.h +++ b/arm_compute/core/Types.h @@ -946,18 +946,19 @@ class WeightsInfo public: /** Default constructor */ WeightsInfo() - : _are_reshaped(false), _kernel_width(0), _kernel_height(0), _num_kernels(0) + : _are_reshaped(false), _kernel_width(0), _kernel_height(0), _num_kernels(0), _retain_internal_weights(false) { } /** Constructor * - * @param[in] are_reshaped True if the weights have been reshaped - * @param[in] kernel_width Kernel width. - * @param[in] kernel_height Kernel height. - * @param[in] num_kernels Number of convolution kernels. + * @param[in] are_reshaped True if the weights have been reshaped + * @param[in] kernel_width Kernel width. + * @param[in] kernel_height Kernel height. + * @param[in] num_kernels Number of convolution kernels. + * @param[in] retain_internal_weights (Optional) True if internal reshaped weights must be retained. Used for reconfiguration purposes. Default is false. */ - WeightsInfo(bool are_reshaped, unsigned int kernel_width, unsigned int kernel_height, unsigned int num_kernels) - : _are_reshaped(are_reshaped), _kernel_width(kernel_width), _kernel_height(kernel_height), _num_kernels(num_kernels) + WeightsInfo(bool are_reshaped, unsigned int kernel_width, unsigned int kernel_height, unsigned int num_kernels, bool retain_internal_weights = false) + : _are_reshaped(are_reshaped), _kernel_width(kernel_width), _kernel_height(kernel_height), _num_kernels(num_kernels), _retain_internal_weights(retain_internal_weights) { } /** Flag which specifies if the weights tensor has been reshaped. @@ -984,12 +985,17 @@ public: { return std::make_pair(_kernel_width, _kernel_height); } + bool retain_internal_weights() const + { + return _retain_internal_weights; + } private: const bool _are_reshaped; const unsigned int _kernel_width; const unsigned int _kernel_height; const unsigned int _num_kernels; + const bool _retain_internal_weights; }; /** GEMM reshape information class. This class stores the necessary information about matrix A and matrix B reshape. diff --git a/arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h b/arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h index 7fb5af9229..127d8acf10 100644 --- a/arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h +++ b/arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h @@ -86,26 +86,32 @@ public: CLFullyConnectedLayer &operator=(CLFullyConnectedLayer &&) = default; /** Set the input and output tensors. * - * @param[in] input Source tensor. Data type supported: QS8/QASYMM8/QS16/F16/F32. - * @param[in] weights Weights tensor. The weights must be 2 dimensional. Data type supported: Same as @p input - * @param[in] biases Bias tensor. It can be nullptr. Data type supported:Same as @p input. - * @param[out] output Destination tensor. Data type supported: Same as @p input. - * @param[in] transpose_weights (Optional) Transpose weights if true. Defaults to true. - * @param[in] are_weights_reshaped (Optional) Reshape the weights tensor if false. Defaults to false. + * @param[in] input Source tensor. Data type supported: QS8/QASYMM8/QS16/F16/F32. + * @param[in] weights Weights tensor. The weights must be 2 dimensional. Data type supported: Same as @p input + * @param[in] biases Bias tensor. It can be nullptr. Data type supported:Same as @p input. + * @param[out] output Destination tensor. Data type supported: Same as @p input. + * @param[in] transpose_weights (Optional) Transpose weights if true. Defaults to true. + * @param[in] are_weights_reshaped (Optional) Reshape the weights tensor if false. Defaults to false. + * @param[in] retain_internal_weights (Optional) Retain internal reshaped weights. Defaults to false. + * Used for reconfiguration purposes. */ - void configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, bool transpose_weights = true, bool are_weights_reshaped = false); + void configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, bool transpose_weights = true, bool are_weights_reshaped = false, + bool retain_internal_weights = false); /** Static function to check if given info will lead to a valid configuration of @ref CLFullyConnectedLayer * - * @param[in] input Source tensor. Data type supported: QS8/QASYMM8/QS16/F16/F32. - * @param[in] weights Weights tensor. The weights must be 2 dimensional. Data type supported: Same as @p input - * @param[in] biases Bias tensor. It can be nullptr. Data type supported:Same as @p input. - * @param[in] output Destination tensor. Data type supported: Same as @p input. - * @param[in] transpose_weights (Optional) Transpose weights if true. Defaults to true. - * @param[in] are_weights_reshaped (Optional) Reshape the weights tensor if false. Defaults to false. + * @param[in] input Source tensor. Data type supported: QS8/QASYMM8/QS16/F16/F32. + * @param[in] weights Weights tensor. The weights must be 2 dimensional. Data type supported: Same as @p input + * @param[in] biases Bias tensor. It can be nullptr. Data type supported:Same as @p input. + * @param[in] output Destination tensor. Data type supported: Same as @p input. + * @param[in] transpose_weights (Optional) Transpose weights if true. Defaults to true. + * @param[in] are_weights_reshaped (Optional) Reshape the weights tensor if false. Defaults to false. + * @param[in] retain_internal_weights (Optional) Retain internal reshaped weights. Defaults to false. + * Used for reconfiguration purposes. * * @return a status */ - static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, bool transpose_weights = true, bool are_weights_reshaped = false); + static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, bool transpose_weights = true, bool are_weights_reshaped = false, + bool retain_internal_weights = false); //Inherited methods override void run() override; diff --git a/arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h b/arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h index 3dde52989b..aaa432616d 100644 --- a/arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h +++ b/arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h @@ -195,6 +195,7 @@ private: bool _is_quantized; bool _is_activationlayer_enabled; bool _is_prepared; + bool _retain_internal_weights; }; } #endif /* __ARM_COMPUTE_CLGEMMCONVOLUTIONLAYER_H__ */ diff --git a/src/runtime/CL/functions/CLFullyConnectedLayer.cpp b/src/runtime/CL/functions/CLFullyConnectedLayer.cpp index 151fa1b5fa..44bf28374f 100644 --- a/src/runtime/CL/functions/CLFullyConnectedLayer.cpp +++ b/src/runtime/CL/functions/CLFullyConnectedLayer.cpp @@ -133,7 +133,8 @@ void CLFullyConnectedLayer::configure_fc_fc(const ICLTensor *input, const ICLTen configure_mm(input, weights, output); } -void CLFullyConnectedLayer::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, bool transpose_weights, bool are_weights_reshaped) +void CLFullyConnectedLayer::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, bool transpose_weights, bool are_weights_reshaped, + bool retain_internal_weights) { ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output); @@ -143,7 +144,8 @@ void CLFullyConnectedLayer::configure(const ICLTensor *input, const ICLTensor *w biases != nullptr ? biases->info() : nullptr, output->info(), transpose_weights, - are_weights_reshaped)); + are_weights_reshaped, + retain_internal_weights)); _are_weights_reshaped = transpose_weights ? are_weights_reshaped : true; _is_fc_after_conv = true; @@ -220,10 +222,14 @@ void CLFullyConnectedLayer::configure(const ICLTensor *input, const ICLTensor *w _gemmlowp_output_stage.configure(&_gemmlowp_output, biases, output, output_multiplier, output_shift, output->info()->quantization_info().offset); _gemmlowp_output.allocator()->allocate(); } + + _are_weights_reshaped = _are_weights_reshaped || retain_internal_weights; } -Status CLFullyConnectedLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, bool transpose_weights, bool are_weights_reshaped) +Status CLFullyConnectedLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, bool transpose_weights, bool are_weights_reshaped, + bool retain_internal_weights) { + ARM_COMPUTE_UNUSED(retain_internal_weights); ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output); ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QASYMM8, DataType::QS16, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, output); diff --git a/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp b/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp index 610eec4d67..4f87043373 100644 --- a/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp +++ b/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp @@ -91,7 +91,8 @@ void CLConvolutionLayerReshapeWeights::run() CLGEMMConvolutionLayer::CLGEMMConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager) : _memory_group(memory_manager), _reshape_weights(), _im2col_kernel(), _mm_gemm(memory_manager), _mm_gemmlowp(memory_manager), _gemmlowp_output_stage(), _col2im_kernel(), _activationlayer_function(), - _original_weights(nullptr), _im2col_output(), _weights_reshaped(), _gemm_output(), _tmp_output(), _is_quantized(false), _is_activationlayer_enabled(false), _is_prepared(false) + _original_weights(nullptr), _im2col_output(), _weights_reshaped(), _gemm_output(), _tmp_output(), _is_quantized(false), _is_activationlayer_enabled(false), _is_prepared(false), + _retain_internal_weights(false) { } @@ -165,9 +166,10 @@ void CLGEMMConvolutionLayer::configure(const ICLTensor *input, const ICLTensor * dilation, act_info)); - _is_prepared = false; - _original_weights = weights; - _is_quantized = is_data_type_quantized_asymmetric(input->info()->data_type()); + _is_prepared = false; + _original_weights = weights; + _is_quantized = is_data_type_quantized_asymmetric(input->info()->data_type()); + _retain_internal_weights = weights_info.retain_internal_weights(); const DataType dt = input->info()->data_type(); @@ -404,17 +406,20 @@ void CLGEMMConvolutionLayer::prepare() { if(!_is_prepared) { - // Run weights reshaping and mark as unused - ARM_COMPUTE_ERROR_ON(!_original_weights->is_used()); - _weights_reshaped.allocator()->allocate(); - _reshape_weights.run(); - _original_weights->mark_as_unused(); + if(!_retain_internal_weights) + { + // Run weights reshaping and mark as unused + ARM_COMPUTE_ERROR_ON(!_original_weights->is_used()); + _weights_reshaped.allocator()->allocate(); + _reshape_weights.run(); + _original_weights->mark_as_unused(); + } // Run GEMM prepare if(!_is_quantized) { _mm_gemm.prepare(); - if(!_weights_reshaped.is_used()) + if(!_weights_reshaped.is_used() && !_retain_internal_weights) { _weights_reshaped.allocator()->free(); } |