From 7c23ad01c028f73aef0b439fc5d5d14e92e5f4e2 Mon Sep 17 00:00:00 2001 From: Giorgio Arena Date: Thu, 30 Nov 2017 15:08:38 +0000 Subject: COMPMID-617 Add validation to NEConvolutionLayer Change-Id: I796a13e6ea672e274aaa8234ee0689828fec7292 Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/111348 Tested-by: Jenkins Reviewed-by: Ioan-Cristian Szabo Reviewed-by: Anthony Barbier --- .../core/NEON/kernels/NEGEMMMatrixMultiplyKernel.h | 9 + .../core/NEON/kernels/NEWeightsReshapeKernel.h | 11 + .../runtime/NEON/functions/NEConvolutionLayer.h | 28 ++ .../NEON/kernels/NEGEMMMatrixMultiplyKernel.cpp | 106 ++++--- src/core/NEON/kernels/NEWeightsReshapeKernel.cpp | 99 ++++--- src/runtime/NEON/functions/NEConvolutionLayer.cpp | 305 +++++++++++++++++---- 6 files changed, 433 insertions(+), 125 deletions(-) diff --git a/arm_compute/core/NEON/kernels/NEGEMMMatrixMultiplyKernel.h b/arm_compute/core/NEON/kernels/NEGEMMMatrixMultiplyKernel.h index aff2671441..9923c31018 100644 --- a/arm_compute/core/NEON/kernels/NEGEMMMatrixMultiplyKernel.h +++ b/arm_compute/core/NEON/kernels/NEGEMMMatrixMultiplyKernel.h @@ -61,6 +61,15 @@ public: * @param[in] alpha Weight of the matrix product */ void configure(const ITensor *input0, const ITensor *input1, ITensor *output, float alpha); + /** Static function to check if given info will lead to a valid configuration of @ref NEGEMMMatrixMultiplyKernel + * + * @param[in] input0 Input tensor containing the Matrix A. Data types supported: QS8/QS16/F16/F32 + * @param[in] input1 Input tensor containing the Matrix B. Data type supported: same as @p input0 + * @param[in] output Output tensor to store the result of matrix multiplication. Data type supported: same as @p input0 + * + * @return a status + */ + static Status validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output); // Inherited methods overridden: void run(const Window &window, const ThreadInfo &info) override; diff --git a/arm_compute/core/NEON/kernels/NEWeightsReshapeKernel.h b/arm_compute/core/NEON/kernels/NEWeightsReshapeKernel.h index 5b37d6cc62..84ec736d03 100644 --- a/arm_compute/core/NEON/kernels/NEWeightsReshapeKernel.h +++ b/arm_compute/core/NEON/kernels/NEWeightsReshapeKernel.h @@ -77,6 +77,17 @@ public: * @param[out] output The output tensor. Data types supported: Same as @p input */ void configure(const ITensor *input, const ITensor *bias, ITensor *output); + /** Static function to check if given info will lead to a valid configuration of @ref NEWeightsReshapeKernel + * + * @param[in] input The input tensor to convert. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM] if shared, + * and 5D tensor with dimensions [kernel_x, kernel_y, IFM, OFM, num_patches] if unshared. Data types supported: QS8/QS16/F16/F32 + * @param[in] biases The shared biases tensor to append. Bias is 1D tensor with dimensions [OFM] if shared and 2D tensor with + * dimensions [OFM, num_patches] if unshared. Data types supported: Same as @p input + * @param[in] output The output tensor. Should be a 2D Tensor. Data types supported: Same as @p input + * + * @return a status + */ + static Status validate(const ITensorInfo *input, const ITensorInfo *biases, const ITensorInfo *output); // Inherited methods overridden: void run(const Window &window, const ThreadInfo &info) override; diff --git a/arm_compute/runtime/NEON/functions/NEConvolutionLayer.h b/arm_compute/runtime/NEON/functions/NEConvolutionLayer.h index 893dfa0f9d..a79f21f3ac 100644 --- a/arm_compute/runtime/NEON/functions/NEConvolutionLayer.h +++ b/arm_compute/runtime/NEON/functions/NEConvolutionLayer.h @@ -62,6 +62,17 @@ public: * Data types supported: Same as @p weights. */ void configure(const ITensor *weights, const ITensor *biases, ITensor *output, bool transpose1xW); + /** Static function to check if given info will lead to a valid configuration of @ref NEConvolutionLayerReshapeWeights + * + * @param[in] weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported: QS8/QS16/F16/F32. + * @param[in] biases Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p weights. + * @param[in] output Destination tensor. Data types supported: Same as @p weights. + * @param[in] transpose1xW True if the weights are to undergo a 1xW transposition after reshaping (in case of GEMM operation), false otherwise. + * Data types supported: Same as @p weights. + * + * @return an error status + */ + static Status validate(const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, bool transpose1xW); // Inherited methods overridden: void run() override; @@ -101,6 +112,23 @@ public: * tensor has also been transposed with NEGEMMTranspose1xWKernel. Data type supported: Same as @p input. */ void configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info = WeightsInfo()); + /** Static function to check if given info will lead to a valid configuration of @ref NEConvolutionLayer + * + * @param[in] input Source tensor. 3 lower dimensions represent a single input [width, height, IFM], + * while every optional dimension from 4 and above represent a batch of inputs. + * Data types supported: QS8/QS16/F16/F32. + * @param[in] weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported:Same as @p input. + * @param[in] biases Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. Data type supported:Same as @p input. + * @param[in] output Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs. + * Data types supported: Same as @p input. + * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo. + * @param[in] weights_info Specifies if the weights tensor has been reshaped with NEWeightsReshapeKernel. If this is not part of the fully connected layer the weights + * tensor has also been transposed with NEGEMMTranspose1xWKernel. Data type supported: Same as @p input. + * + * @return a status + */ + static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, + const WeightsInfo &weights_info = WeightsInfo()); // Inherited methods overridden: void run() override; diff --git a/src/core/NEON/kernels/NEGEMMMatrixMultiplyKernel.cpp b/src/core/NEON/kernels/NEGEMMMatrixMultiplyKernel.cpp index a583c1dfd4..aa5e2dd0dd 100644 --- a/src/core/NEON/kernels/NEGEMMMatrixMultiplyKernel.cpp +++ b/src/core/NEON/kernels/NEGEMMMatrixMultiplyKernel.cpp @@ -1408,36 +1408,36 @@ void matrix_matrix_multiply_qs16(const ITensor *input0, const ITensor *input1, I }, ina, inb, out); } -} // namespace - -NEGEMMMatrixMultiplyKernel::NEGEMMMatrixMultiplyKernel() - : _input0(nullptr), _input1(nullptr), _output(nullptr), _alpha(1.0f) -{ -} -void NEGEMMMatrixMultiplyKernel::configure(const ITensor *input0, const ITensor *input1, ITensor *output, float alpha) +Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output) { - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::F16, DataType::F32, DataType::QS8, DataType::QS16); - ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1, output); - ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input0, input1, output); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::F16, DataType::F32, DataType::QS8, DataType::QS16); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1, output); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input0, input1, output); + ARM_COMPUTE_UNUSED(input0); + ARM_COMPUTE_UNUSED(input1); + ARM_COMPUTE_UNUSED(output); - if(output->info()->dimension(1) == 1) + if(output->dimension(1) == 1) { - ARM_COMPUTE_ERROR_ON(input0->info()->dimension(0) != input1->info()->dimension(1)); + ARM_COMPUTE_RETURN_ERROR_ON(input0->dimension(0) != input1->dimension(1)); } - _input0 = input0; - _input1 = input1; - _output = output; - _alpha = alpha; + return Status{}; +} + +std::pair validate_and_configure_window(ITensorInfo *input0, ITensorInfo *input1, ITensorInfo *output) +{ + Window win = Window(); + bool window_changed = false; unsigned int num_elems_processed_per_iteration_x = 0; const unsigned int num_elems_processed_per_iteration_y = 4; // Check if the output tensor is a vector. If so,the kernel runs the vector-matrix multiplication - if((output->info()->dimension(1) == 1)) + if((output->dimension(1) == 1)) { - switch(input0->info()->data_type()) + switch(input0->data_type()) { case DataType::F32: { @@ -1469,24 +1469,22 @@ void NEGEMMMatrixMultiplyKernel::configure(const ITensor *input0, const ITensor } // Configure kernel window - Window win = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration_x)); + win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration_x)); - AccessWindowHorizontal output_access(output->info(), 0, num_elems_processed_per_iteration_x); + AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration_x); - update_window_and_padding(win, - AccessWindowStatic(input0->info(), 0, 0, input0->info()->tensor_shape().x(), 1), - AccessWindowHorizontal(input1->info(), 0, num_elems_processed_per_iteration_x), - output_access); + window_changed = update_window_and_padding(win, + AccessWindowStatic(input0, 0, 0, input0->tensor_shape().x(), 1), + AccessWindowHorizontal(input1, 0, num_elems_processed_per_iteration_x), + output_access); Coordinates coord; - coord.set_num_dimensions(output->info()->num_dimensions()); - output_access.set_valid_region(win, ValidRegion(coord, output->info()->tensor_shape())); - - INEKernel::configure(win); + coord.set_num_dimensions(output->num_dimensions()); + output_access.set_valid_region(win, ValidRegion(coord, output->tensor_shape())); } else { - switch(input0->info()->data_type()) + switch(input0->data_type()) { case DataType::F32: { @@ -1518,19 +1516,51 @@ void NEGEMMMatrixMultiplyKernel::configure(const ITensor *input0, const ITensor } // Configure kernel window - Window win = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y)); + win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y)); - AccessWindowRectangle output_access(output->info(), 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y); + AccessWindowRectangle output_access(output, 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y); - update_window_and_padding(win, - AccessWindowRectangle(input0->info(), 0, 0, 4, 1, 1.f, 0.25f), - AccessWindowStatic(input1->info(), 0, 0, input1->info()->tensor_shape().x(), ceil_to_multiple(input1->info()->tensor_shape().y(), 4)), - output_access); + window_changed = update_window_and_padding(win, + AccessWindowRectangle(input0, 0, 0, 4, 1, 1.f, 0.25f), + AccessWindowStatic(input1, 0, 0, input1->tensor_shape().x(), ceil_to_multiple(input1->tensor_shape().y(), 4)), + output_access); - output_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), output->info()->tensor_shape())); - - INEKernel::configure(win); + output_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), output->tensor_shape())); } + + Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; + return std::make_pair(err, win); +} +} // namespace + +NEGEMMMatrixMultiplyKernel::NEGEMMMatrixMultiplyKernel() + : _input0(nullptr), _input1(nullptr), _output(nullptr), _alpha(1.0f) +{ +} + +void NEGEMMMatrixMultiplyKernel::configure(const ITensor *input0, const ITensor *input1, ITensor *output, float alpha) +{ + // Perform validate step + ARM_COMPUTE_ERROR_ON_NULLPTR(input0, input1, output); + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input0->info(), input1->info(), output->info())); + + _input0 = input0; + _input1 = input1; + _output = output; + _alpha = alpha; + + // Configure kernel window + auto win_config = validate_and_configure_window(input0->info(), input1->info(), output->info()); + ARM_COMPUTE_ERROR_THROW_ON(win_config.first); + INEKernel::configure(win_config.second); +} + +Status NEGEMMMatrixMultiplyKernel::validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output) +{ + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input0, input1, output)); + ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input0->clone().get(), input1->clone().get(), output->clone().get()).first); + + return Status{}; } void NEGEMMMatrixMultiplyKernel::run(const Window &window, const ThreadInfo &info) diff --git a/src/core/NEON/kernels/NEWeightsReshapeKernel.cpp b/src/core/NEON/kernels/NEWeightsReshapeKernel.cpp index d52e88c37a..794c179277 100644 --- a/src/core/NEON/kernels/NEWeightsReshapeKernel.cpp +++ b/src/core/NEON/kernels/NEWeightsReshapeKernel.cpp @@ -86,6 +86,57 @@ void weights_reshape(const ITensor *input, const ITensor *bias, ITensor *output, }, in); } + +TensorShape get_output_shape(const ITensorInfo *input, bool has_bias) +{ + TensorShape output_shape{ input->tensor_shape() }; + + output_shape.collapse(3); + const size_t tmp_dim = output_shape[0]; + output_shape.set(0, output_shape[1]); + output_shape.set(1, tmp_dim + (has_bias ? 1 : 0)); + + return output_shape; +} + +Status validate_arguments(const ITensorInfo *input, const ITensorInfo *biases, const ITensorInfo *output) +{ + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output); + + if(biases != nullptr) + { + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, biases); + ARM_COMPUTE_RETURN_ERROR_ON((input->num_dimensions() == 4) && (biases->num_dimensions() != 1)); + ARM_COMPUTE_RETURN_ERROR_ON((input->num_dimensions() == 5) && (biases->num_dimensions() != 2)); + ARM_COMPUTE_RETURN_ERROR_ON((input->num_dimensions() == 4) && (biases->dimension(0) != input->tensor_shape()[3])); + ARM_COMPUTE_RETURN_ERROR_ON((input->num_dimensions() == 5) && (biases->dimension(0) != input->tensor_shape()[3] || biases->dimension(1) != input->tensor_shape()[4])); + } + + // Checks performed when output is configured + if(output->total_size() != 0) + { + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), get_output_shape(input, biases != nullptr)); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); + } + + return Status{}; +} + +std::pair validate_and_configure_window(ITensorInfo *input, ITensorInfo *output) +{ + Window window = calculate_max_window(*input, Steps()); + window.set(Window::DimX, Window::Dimension(0, input->dimension(0), input->dimension(0))); + window.set(Window::DimY, Window::Dimension(0, input->dimension(1), input->dimension(1))); + window.set(Window::DimZ, Window::Dimension(0, input->dimension(2), input->dimension(2))); + + // The NEConvolutionLayerWeightsReshapeKernel doesn't need padding so update_window_and_padding() can be skipped + output->set_valid_region(ValidRegion(Coordinates(), output->tensor_shape())); + + return std::make_pair(Status{}, window); +} } // namespace NEWeightsReshapeKernel::NEWeightsReshapeKernel() @@ -95,35 +146,15 @@ NEWeightsReshapeKernel::NEWeightsReshapeKernel() void NEWeightsReshapeKernel::configure(const ITensor *input, const ITensor *bias, ITensor *output) { - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32); - ARM_COMPUTE_ERROR_ON_NULLPTR(output); - - const int fixed_point_position = input->info()->fixed_point_position(); - const DataType dt = input->info()->data_type(); - const TensorShape &input_shape = input->info()->tensor_shape(); - TensorShape output_shape{ input_shape }; - output_shape.collapse(3); - - const size_t tmp_dim = output_shape[0]; - output_shape.set(0, output_shape[1]); - output_shape.set(1, tmp_dim + (bias != nullptr ? 1 : 0)); + ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); // Output tensor auto inizialitation if not yet initialized - auto_init_if_empty(*output->info(), output_shape, 1, dt, fixed_point_position); + auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(get_output_shape(input->info(), (bias != nullptr)))); - ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(output->info()->tensor_shape(), output_shape); - ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); - - if(bias != nullptr) - { - ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, bias); - ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, bias); - ARM_COMPUTE_ERROR_ON((input->info()->num_dimensions() == 4) && (bias->info()->num_dimensions() != 1)); - ARM_COMPUTE_ERROR_ON((input->info()->num_dimensions() == 5) && (bias->info()->num_dimensions() != 2)); - ARM_COMPUTE_ERROR_ON((input->info()->num_dimensions() == 4) && (bias->info()->dimension(0) != input->info()->tensor_shape()[3])); - ARM_COMPUTE_ERROR_ON((input->info()->num_dimensions() == 5) && (bias->info()->dimension(0) != input->info()->tensor_shape()[3] || bias->info()->dimension(1) != input->info()->tensor_shape()[4])); - } + // Perform validation step + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), + (bias != nullptr) ? bias->info() : nullptr, + output->info())); _input = input; _bias = bias; @@ -154,15 +185,17 @@ void NEWeightsReshapeKernel::configure(const ITensor *input, const ITensor *bias } // Configure kernel - Window window = calculate_max_window(*input->info(), Steps()); - window.set(Window::DimX, Window::Dimension(0, _input->info()->dimension(0), _input->info()->dimension(0))); - window.set(Window::DimY, Window::Dimension(0, _input->info()->dimension(1), _input->info()->dimension(1))); - window.set(Window::DimZ, Window::Dimension(0, _input->info()->dimension(2), _input->info()->dimension(2))); + auto win_config = validate_and_configure_window(input->info(), output->info()); + ARM_COMPUTE_ERROR_THROW_ON(win_config.first); + INEKernel::configure(win_config.second); +} - // The NEConvolutionLayerWeightsReshapeKernel doesn't need padding so update_window_and_padding() can be skipped - output->info()->set_valid_region(ValidRegion(Coordinates(), output->info()->tensor_shape())); +Status NEWeightsReshapeKernel::validate(const ITensorInfo *input, const ITensorInfo *biases, const ITensorInfo *output) +{ + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, biases, output)); + ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get()).first); - INEKernel::configure(window); + return Status{}; } void NEWeightsReshapeKernel::run(const Window &window, const ThreadInfo &info) diff --git a/src/runtime/NEON/functions/NEConvolutionLayer.cpp b/src/runtime/NEON/functions/NEConvolutionLayer.cpp index 5ca8eb8179..8f7d940fca 100644 --- a/src/runtime/NEON/functions/NEConvolutionLayer.cpp +++ b/src/runtime/NEON/functions/NEConvolutionLayer.cpp @@ -44,6 +44,16 @@ namespace arm_compute namespace arm_compute { +namespace +{ +TensorShape get_reshaped_weights_shape(const ITensorInfo *weights, bool has_bias) +{ + const unsigned int mat_weights_cols = weights->dimension(3); + const unsigned int mat_weights_rows = weights->dimension(0) * weights->dimension(1) * weights->dimension(2) + (has_bias ? 1 : 0); + return TensorShape(mat_weights_cols, mat_weights_rows); +} +} // namespace + NEConvolutionLayerReshapeWeights::NEConvolutionLayerReshapeWeights(std::shared_ptr memory_manager) : _memory_group(std::move(memory_manager)), _weights_reshape_kernel(), _weights_transposed_kernel(), _weights_reshaped(), _transpose1xW(false) { @@ -51,18 +61,12 @@ NEConvolutionLayerReshapeWeights::NEConvolutionLayerReshapeWeights(std::shared_p void NEConvolutionLayerReshapeWeights::configure(const ITensor *weights, const ITensor *biases, ITensor *output, bool transpose1xW) { - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32); - ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(weights, output); - ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(weights, output); - ARM_COMPUTE_ERROR_ON(weights->info()->num_dimensions() > 4); - - if(biases != nullptr) - { - ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases); - ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(weights, biases); - ARM_COMPUTE_ERROR_ON(biases->info()->dimension(0) != weights->info()->dimension(3)); - ARM_COMPUTE_ERROR_ON(biases->info()->num_dimensions() > 1); - } + // Perform validation step + ARM_COMPUTE_ERROR_ON_NULLPTR(weights, output); + ARM_COMPUTE_ERROR_THROW_ON(NEConvolutionLayerReshapeWeights::validate(weights->info(), + (biases != nullptr) ? biases->info() : nullptr, + output->info(), + transpose1xW)); // Check if bias are present, if yes they will be embedded to the weights matrix const bool _has_bias = (biases != nullptr); @@ -72,10 +76,7 @@ void NEConvolutionLayerReshapeWeights::configure(const ITensor *weights, const I if(transpose1xW) { // Create tensor to store the reshaped weights - const unsigned int mat_weights_cols = weights->info()->dimension(3); - const unsigned int mat_weights_rows = weights->info()->dimension(0) * weights->info()->dimension(1) * weights->info()->dimension(2) + (_has_bias ? 1 : 0); - TensorShape shape_wr(mat_weights_cols, mat_weights_rows); - TensorInfo info_wr(shape_wr, 1, weights->info()->data_type(), weights->info()->fixed_point_position()); + TensorInfo info_wr = weights->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(get_reshaped_weights_shape(weights->info(), _has_bias)); _weights_reshaped.allocator()->init(info_wr); _memory_group.manage(&_weights_reshaped); @@ -91,6 +92,46 @@ void NEConvolutionLayerReshapeWeights::configure(const ITensor *weights, const I } } +Status NEConvolutionLayerReshapeWeights::validate(const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, bool transpose1xW) +{ + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, output); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(weights, output); + ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 4); + + if(biases != nullptr) + { + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(weights, biases); + ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(3)); + ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1); + } + + // Check if bias are present, if yes they will be embedded to the weights matrix + const bool has_bias = (biases != nullptr); + + // Checks performed when biases are present + if(has_bias) + { + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases); + ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(3)); + ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1); + } + + if(transpose1xW) + { + TensorInfo weights_reshaped = weights->clone()->set_tensor_shape(get_reshaped_weights_shape(weights, has_bias)); + ARM_COMPUTE_RETURN_ON_ERROR(NEWeightsReshapeKernel::validate(weights, biases, &weights_reshaped)); + ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMTranspose1xWKernel::validate(&weights_reshaped, output)); + } + else + { + ARM_COMPUTE_RETURN_ON_ERROR(NEWeightsReshapeKernel::validate(weights, biases, output)); + } + + return Status{}; +} + void NEConvolutionLayerReshapeWeights::run() { _memory_group.acquire(); @@ -105,50 +146,89 @@ void NEConvolutionLayerReshapeWeights::run() _memory_group.release(); } -NEConvolutionLayer::NEConvolutionLayer(std::shared_ptr memory_manager) - : _memory_group(std::move(memory_manager)), _input_im2col_kernel(), _input_interleave_kernel(), _reshape_weights(), _mm_kernel(), _mm_optimised_kernel(nullptr), _output_col2im_kernel(), - _input_im2col_reshaped(), _input_interleaved_reshaped(), _weights_reshaped(), _gemm_output(), _workspace(), _has_bias(false), _is_fully_connected_convolution(false), _are_weights_reshaped(false) +namespace { +TensorShape get_reshaped_weights_shape_conv(const ITensorInfo *weights, bool has_bias, bool is_fully_connected_convolution) +{ + unsigned int mat_weights_cols = weights->dimension(3); + unsigned int mat_weights_rows = weights->dimension(0) * weights->dimension(1) * weights->dimension(2) + (has_bias ? 1 : 0); + + if(is_fully_connected_convolution) + { + // Create tensor to store the reshaped weights + return TensorShape(mat_weights_cols, mat_weights_rows); + } + else + { + // Create tensor to store transposed weights + const float transpose_width = 16.0f / weights->element_size(); + return TensorShape(mat_weights_rows * static_cast(transpose_width), static_cast(std::ceil(mat_weights_cols / transpose_width))); + } } -void NEConvolutionLayer::configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info) +Status validate_and_initialize_values(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const PadStrideInfo &conv_info, const WeightsInfo &weights_info, DataType &dt, + bool &has_bias, + bool &are_weights_reshaped, unsigned int &kernel_width, unsigned int &kernel_height, bool &is_fully_connected_convolution, unsigned int &mat_weights_cols, unsigned int &mat_weights_rows, + unsigned int &conv_w, unsigned int &conv_h) { - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32); - ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights); - ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, weights); - ARM_COMPUTE_ERROR_ON(!weights_info.are_reshaped() && weights->info()->dimension(2) != input->info()->dimension(2)); - ARM_COMPUTE_ERROR_ON(weights->info()->num_dimensions() > 4); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, weights); + ARM_COMPUTE_RETURN_ERROR_ON(!weights_info.are_reshaped() && weights->dimension(2) != input->dimension(2)); + ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 4); if(biases != nullptr) { - ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases); - ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, biases); - ARM_COMPUTE_ERROR_ON(!weights_info.are_reshaped() && biases->info()->dimension(0) != weights->info()->dimension(3)); - ARM_COMPUTE_ERROR_ON(biases->info()->num_dimensions() > 1); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, biases); + ARM_COMPUTE_RETURN_ERROR_ON(!weights_info.are_reshaped() && biases->dimension(0) != weights->dimension(3)); + ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1); } - const DataType dt = input->info()->data_type(); - const int fixed_point_position = input->info()->fixed_point_position(); + dt = input->data_type(); + has_bias = (biases != nullptr); + are_weights_reshaped = weights_info.are_reshaped(); + kernel_width = (are_weights_reshaped) ? weights_info.kernel_size().first : weights->dimension(0); + kernel_height = (are_weights_reshaped) ? weights_info.kernel_size().second : weights->dimension(1); + mat_weights_cols = weights->dimension(3); + mat_weights_rows = weights->dimension(0) * weights->dimension(1) * weights->dimension(2) + (has_bias ? 1 : 0); - _has_bias = (biases != nullptr); - _are_weights_reshaped = weights_info.are_reshaped(); + std::tie(conv_w, conv_h) = scaled_dimensions(input->dimension(0), input->dimension(1), kernel_width, kernel_height, + conv_info); - // Get parameters from conv_info - unsigned int stride_x = 0; - unsigned int stride_y = 0; - std::tie(stride_x, stride_y) = conv_info.stride(); + is_fully_connected_convolution = ((conv_w == 1) && (conv_h == 1)); - // Get convolved dimensions - unsigned int conv_w = 0; - unsigned int conv_h = 0; + return Status{}; +} +} // namespace - const unsigned int kernel_width = (_are_weights_reshaped) ? weights_info.kernel_size().first : weights->info()->dimension(0); - const unsigned int kernel_height = (_are_weights_reshaped) ? weights_info.kernel_size().second : weights->info()->dimension(1); - std::tie(conv_w, conv_h) = scaled_dimensions(input->info()->dimension(0), input->info()->dimension(1), kernel_width, kernel_height, - conv_info); +NEConvolutionLayer::NEConvolutionLayer(std::shared_ptr memory_manager) + : _memory_group(std::move(memory_manager)), _input_im2col_kernel(), _input_interleave_kernel(), _reshape_weights(), _mm_kernel(), _mm_optimised_kernel(nullptr), _output_col2im_kernel(), + _input_im2col_reshaped(), _input_interleaved_reshaped(), _weights_reshaped(), _gemm_output(), _workspace(), _has_bias(false), _is_fully_connected_convolution(false), _are_weights_reshaped(false) +{ +} + +void NEConvolutionLayer::configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info) +{ + // Perform validate step + ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output); - // Check if its a "fully connected" convolution, i.e. the output size is 1x1xnum_kernels - _is_fully_connected_convolution = ((conv_w == 1) && (conv_h == 1)); + DataType dt{}; + unsigned int kernel_width = 0; + unsigned int kernel_height = 0; + unsigned int mat_weights_cols = 0; + unsigned int mat_weights_rows = 0; + unsigned int conv_w = 0; + unsigned int conv_h = 0; + + Status status = validate_and_initialize_values(input->info(), weights->info(), (biases == nullptr) ? nullptr : biases->info(), conv_info, weights_info, dt, _has_bias, _are_weights_reshaped, + kernel_width, kernel_height, + _is_fully_connected_convolution, + mat_weights_cols, mat_weights_rows, conv_w, conv_h); + + ARM_COMPUTE_ERROR_THROW_ON(status); + + const unsigned int fixed_point_position = input->info()->fixed_point_position(); #if defined(__arm__) if(NEScheduler::get().cpu_info().CPU == CPUTarget::ARMV7 && dt == DataType::F32) @@ -162,9 +242,6 @@ void NEConvolutionLayer::configure(const ITensor *input, const ITensor *weights, } #endif /* defined(__arm__) || defined(__aarch64__) */ - unsigned int mat_weights_cols = weights->info()->dimension(3); - unsigned int mat_weights_rows = weights->info()->dimension(0) * weights->info()->dimension(1) * weights->info()->dimension(2) + (_has_bias ? 1 : 0); - // Reshape weights if needed if(_mm_optimised_kernel != nullptr) { @@ -230,7 +307,7 @@ void NEConvolutionLayer::configure(const ITensor *input, const ITensor *weights, shape_im2col.set(0, mat_input_cols); shape_im2col.set(1, mat_input_rows); shape_im2col.set(2, 1); - _input_im2col_reshaped.allocator()->init(TensorInfo(shape_im2col, 1, dt, fixed_point_position)); + _input_im2col_reshaped.allocator()->init(input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_im2col)); _memory_group.manage(&_input_im2col_reshaped); // Create tensor (interleave) to prepare input tensor for GEMM @@ -239,7 +316,7 @@ void NEConvolutionLayer::configure(const ITensor *input, const ITensor *weights, TensorShape shape_interleaved(shape_im2col); shape_interleaved.set(0, shape_interleaved.x() * 4); shape_interleaved.set(1, std::ceil(shape_interleaved.y() / 4.f)); - _input_interleaved_reshaped.allocator()->init(TensorInfo(shape_interleaved, 1, dt, fixed_point_position)); + _input_interleaved_reshaped.allocator()->init(input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_interleaved)); _memory_group.manage(&_input_interleaved_reshaped); } @@ -247,7 +324,7 @@ void NEConvolutionLayer::configure(const ITensor *input, const ITensor *weights, TensorShape shape_gemm(_input_im2col_reshaped.info()->tensor_shape()); shape_gemm.set(0, mat_weights_cols); shape_gemm.set(1, mat_input_rows); - _gemm_output.allocator()->init(TensorInfo(shape_gemm, 1, dt, fixed_point_position)); + _gemm_output.allocator()->init(_input_im2col_reshaped.info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_gemm)); _memory_group.manage(&_gemm_output); // Configure kernels @@ -296,8 +373,6 @@ void NEConvolutionLayer::configure(const ITensor *input, const ITensor *weights, _output_col2im_kernel.configure(&_gemm_output, output, Size2D(conv_w, conv_h)); _gemm_output.allocator()->allocate(); - ARM_COMPUTE_ERROR_ON_MSG((output->info()->dimension(0) != conv_w) || (output->info()->dimension(1) != conv_h), "Output shape does not match the expected one"); - // Allocate intermediate tensor if(!_are_weights_reshaped) { @@ -305,6 +380,128 @@ void NEConvolutionLayer::configure(const ITensor *input, const ITensor *weights, } } +Status NEConvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, + const WeightsInfo &weights_info) +{ + DataType dt{}; + bool has_bias{}; + bool are_weights_reshaped{}; + bool is_fully_connected_convolution{}; + unsigned int kernel_width = 0; + unsigned int kernel_height = 0; + unsigned int mat_weights_cols = 0; + unsigned int mat_weights_rows = 0; + unsigned int conv_w = 0; + unsigned int conv_h = 0; + + Status status = validate_and_initialize_values(input, weights, biases, conv_info, weights_info, dt, has_bias, are_weights_reshaped, kernel_width, kernel_height, + is_fully_connected_convolution, mat_weights_cols, mat_weights_rows, + conv_w, conv_h); + + ARM_COMPUTE_RETURN_ON_ERROR(status); + + std::unique_ptr reshaped_weights = weights->clone(); + bool optimised_kernel = false; + +#if defined(__arm__) + if(NEScheduler::get().cpu_info().CPU == CPUTarget::ARMV7 && dt == DataType::F32) + { + optimised_kernel = true; + } +#elif defined(__aarch64__) + if(NEScheduler::get().cpu_info().CPU >= CPUTarget::ARMV8 && dt == DataType::F32) + { + optimised_kernel = true; + } +#endif /* defined(__arm__) || defined(__aarch64__) */ + + // Reshape weights if needed + if(optimised_kernel) + { + if(are_weights_reshaped) + { + mat_weights_cols = weights_info.num_kernels(); + mat_weights_rows = weights->dimension(1); + } + else + { + TensorShape reshaped_weights_shape{ mat_weights_cols, mat_weights_rows }; + + // Create tensor to store the reshaped weights + reshaped_weights->set_tensor_shape(get_reshaped_weights_shape_conv(weights, has_bias, is_fully_connected_convolution)); + ARM_COMPUTE_RETURN_ON_ERROR(NEConvolutionLayerReshapeWeights::validate(weights, biases, reshaped_weights.get(), !is_fully_connected_convolution /* 1xW transpose */)); + weights = reshaped_weights.get(); + } + } + else + { + if(are_weights_reshaped) + { + const unsigned int transpose_width = 16 / input->element_size(); + mat_weights_cols = weights_info.num_kernels(); + mat_weights_rows = weights->dimension(0) / transpose_width + (has_bias ? 1 : 0); + } + else + { + TensorShape reshaped_weights_shape; + + if(is_fully_connected_convolution) + { + reshaped_weights_shape = TensorShape{ mat_weights_cols, mat_weights_rows }; + } + else + { + // Create tensor to store transposed weights + const float transpose_width = 16.0f / input->element_size(); + reshaped_weights_shape = TensorShape{ mat_weights_rows *static_cast(transpose_width), + static_cast(std::ceil(mat_weights_cols / transpose_width)) }; + } + + // Create tensor to store the reshaped weights + reshaped_weights->set_tensor_shape(get_reshaped_weights_shape_conv(weights, has_bias, is_fully_connected_convolution)); + ARM_COMPUTE_RETURN_ON_ERROR(NEConvolutionLayerReshapeWeights::validate(weights, biases, reshaped_weights.get(), !is_fully_connected_convolution /* 1xW transpose */)); + weights = reshaped_weights.get(); + } + } + + // Validate im2col + const unsigned int mat_input_cols = mat_weights_rows; + const unsigned int mat_input_rows = conv_w * conv_h; + TensorShape shape_im2col = input->tensor_shape(); + shape_im2col.set(0, mat_input_cols); + shape_im2col.set(1, mat_input_rows); + shape_im2col.set(2, 1); + TensorInfo im2_col_info = input->clone()->set_tensor_shape(shape_im2col); + ARM_COMPUTE_RETURN_ON_ERROR(NEIm2ColKernel::validate(input, &im2_col_info, Size2D(weights->dimension(0), weights->dimension(1)), conv_info, has_bias)); + + // Create GEMM output tensor + TensorShape shape_gemm(im2_col_info.tensor_shape()); + shape_gemm.set(0, mat_weights_cols); + shape_gemm.set(1, mat_input_rows); + TensorInfo gemm_output_info = input->clone()->set_tensor_shape(shape_gemm); + + // Validate GEMM interleave and multiply + if(!is_fully_connected_convolution) + { + TensorShape shape_interleaved = shape_im2col; + shape_interleaved.set(0, shape_interleaved.x() * 4); + shape_interleaved.set(1, std::ceil(shape_interleaved.y() / 4.f)); + TensorInfo input_interleaved_info = input->clone()->set_tensor_shape(shape_interleaved); + ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMInterleave4x4Kernel::validate(&im2_col_info, &input_interleaved_info)); + ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMMatrixMultiplyKernel::validate(&input_interleaved_info, weights, &gemm_output_info)); + } + else + { + ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMMatrixMultiplyKernel::validate(&im2_col_info, weights, &gemm_output_info)); + } + + ARM_COMPUTE_RETURN_ON_ERROR(NECol2ImKernel::validate(&gemm_output_info, output, Size2D(conv_w, conv_h))); + + ARM_COMPUTE_RETURN_ERROR_ON_MSG((output->dimension(0) != conv_w) || (output->dimension(1) != conv_h), "Output shape does not match the expected one"); + + return Status{}; +} + void NEConvolutionLayer::run() { // Run weights reshaping (Runs once for every configure) -- cgit v1.2.1