From baf174e85ddb5399355281cd34d0f459d92124a7 Mon Sep 17 00:00:00 2001 From: Georgios Pinitas Date: Fri, 8 Sep 2017 19:47:30 +0100 Subject: COMPMID-485: Memory Manager Change-Id: Ib421b7622838f050038cd81e7426bb1413a7d6e6 Reviewed-on: http://mpd-gerrit.cambridge.arm.com/87376 Tested-by: Kaizen Reviewed-by: Anthony Barbier --- src/runtime/NEON/functions/NEConvolutionLayer.cpp | 31 ++++++++++++++-------- .../NEON/functions/NEFullyConnectedLayer.cpp | 21 +++++++++++---- src/runtime/NEON/functions/NESoftmaxLayer.cpp | 13 +++++++-- 3 files changed, 47 insertions(+), 18 deletions(-) (limited to 'src/runtime/NEON') diff --git a/src/runtime/NEON/functions/NEConvolutionLayer.cpp b/src/runtime/NEON/functions/NEConvolutionLayer.cpp index 1c87f60a29..0466a4a501 100644 --- a/src/runtime/NEON/functions/NEConvolutionLayer.cpp +++ b/src/runtime/NEON/functions/NEConvolutionLayer.cpp @@ -34,8 +34,8 @@ using namespace arm_compute; -NEConvolutionLayerReshapeWeights::NEConvolutionLayerReshapeWeights() - : _weights_reshape_kernel(), _weights_transposed_kernel(), _weights_reshaped(), _transpose1xW(false) +NEConvolutionLayerReshapeWeights::NEConvolutionLayerReshapeWeights(std::shared_ptr memory_manager) + : _memory_group(std::move(memory_manager)), _weights_reshape_kernel(), _weights_transposed_kernel(), _weights_reshaped(), _transpose1xW(false) { } @@ -68,6 +68,7 @@ void NEConvolutionLayerReshapeWeights::configure(const ITensor *weights, const I TensorInfo info_wr(shape_wr, 1, weights->info()->data_type(), weights->info()->fixed_point_position()); _weights_reshaped.allocator()->init(info_wr); + _memory_group.manage(&_weights_reshaped); _weights_reshape_kernel.configure(weights, biases, &_weights_reshaped); _weights_transposed_kernel.configure(&_weights_reshaped, output); _weights_reshaped.allocator()->allocate(); @@ -80,16 +81,20 @@ void NEConvolutionLayerReshapeWeights::configure(const ITensor *weights, const I void NEConvolutionLayerReshapeWeights::run() { + _memory_group.acquire(); + NEScheduler::get().schedule(&_weights_reshape_kernel, 3); if(_transpose1xW) { NEScheduler::get().schedule(&_weights_transposed_kernel, Window::DimY); } + + _memory_group.release(); } -NEConvolutionLayer::NEConvolutionLayer() - : _input_im2col_kernel(), _input_interleave_kernel(), _reshape_weights(), _mm_kernel(), _output_col2im_kernel(), _input_im2col_reshaped(), _input_interleaved_reshaped(), _weights_reshaped(), - _gemm_output(), _has_bias(false), _is_fully_connected_convolution(false), _are_weights_reshaped(false) +NEConvolutionLayer::NEConvolutionLayer(std::shared_ptr memory_manager) + : _memory_group(std::move(memory_manager)), _input_im2col_kernel(), _input_interleave_kernel(), _reshape_weights(), _mm_kernel(), _output_col2im_kernel(), _input_im2col_reshaped(), + _input_interleaved_reshaped(), _weights_reshaped(), _gemm_output(), _has_bias(false), _is_fully_connected_convolution(false), _are_weights_reshaped(false) { } @@ -175,6 +180,7 @@ void NEConvolutionLayer::configure(const ITensor *input, const ITensor *weights, 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)); + _memory_group.manage(&_input_im2col_reshaped); // Create tensor (interleave) to prepare input tensor for GEMM if(!_is_fully_connected_convolution) @@ -183,6 +189,7 @@ void NEConvolutionLayer::configure(const ITensor *input, const ITensor *weights, 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)); + _memory_group.manage(&_input_interleaved_reshaped); } // Create GEMM output tensor @@ -190,6 +197,7 @@ void NEConvolutionLayer::configure(const ITensor *input, const ITensor *weights, 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)); + _memory_group.manage(&_gemm_output); // Configure kernels _input_im2col_kernel.configure(input, &_input_im2col_reshaped, Size2D(kernel_width, kernel_height), conv_info, _has_bias); @@ -201,8 +209,11 @@ void NEConvolutionLayer::configure(const ITensor *input, const ITensor *weights, { _input_interleave_kernel.configure(&_input_im2col_reshaped, &_input_interleaved_reshaped); _mm_kernel.configure(&_input_interleaved_reshaped, weights, &_gemm_output, 1.0f); + _input_interleaved_reshaped.allocator()->allocate(); } + _input_im2col_reshaped.allocator()->allocate(); _output_col2im_kernel.configure(&_gemm_output, output, std::make_pair(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"); @@ -211,12 +222,6 @@ void NEConvolutionLayer::configure(const ITensor *input, const ITensor *weights, { _weights_reshaped.allocator()->allocate(); } - _input_im2col_reshaped.allocator()->allocate(); - if(!_is_fully_connected_convolution) - { - _input_interleaved_reshaped.allocator()->allocate(); - } - _gemm_output.allocator()->allocate(); } void NEConvolutionLayer::run() @@ -228,6 +233,8 @@ void NEConvolutionLayer::run() _reshape_weights.run(); } + _memory_group.acquire(); + // Run input reshaping NEScheduler::get().schedule(&_input_im2col_kernel, Window::DimY); if(!_is_fully_connected_convolution) @@ -241,4 +248,6 @@ void NEConvolutionLayer::run() // Reshape output matrix NEScheduler::get().schedule(&_output_col2im_kernel, Window::DimY); + + _memory_group.release(); } diff --git a/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp b/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp index 39983bf643..2e8d10598d 100644 --- a/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp +++ b/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp @@ -32,8 +32,8 @@ namespace arm_compute { -NEFullyConnectedLayerReshapeWeights::NEFullyConnectedLayerReshapeWeights() - : _transpose_kernel(), _transpose1xW_kernel(), _transpose_output(), _transpose_weights(false), _is_batched_fc_layer(false) +NEFullyConnectedLayerReshapeWeights::NEFullyConnectedLayerReshapeWeights(std::shared_ptr memory_manager) + : _memory_group(std::move(memory_manager)), _transpose_kernel(), _transpose1xW_kernel(), _transpose_output(), _transpose_weights(false), _is_batched_fc_layer(false) { } @@ -58,6 +58,7 @@ void NEFullyConnectedLayerReshapeWeights::configure(const ITensor *input, ITenso // Initialize the output tensor for transpose TensorShape shape_transposed(input->info()->dimension(1), input->info()->dimension(0)); _transpose_output.allocator()->init(TensorInfo(shape_transposed, 1, data_type, fixed_point_position)); + _memory_group.manage(&_transpose_output); _transpose_kernel.configure(input, &_transpose_output); // Configure transpose 1xW kernel @@ -87,6 +88,8 @@ void NEFullyConnectedLayerReshapeWeights::configure(const ITensor *input, ITenso void NEFullyConnectedLayerReshapeWeights::run() { + _memory_group.acquire(); + if(_transpose_weights) { NEScheduler::get().schedule(&_transpose_kernel, Window::DimY); @@ -96,11 +99,13 @@ void NEFullyConnectedLayerReshapeWeights::run() { NEScheduler::get().schedule(&_transpose1xW_kernel, Window::DimY); } + + _memory_group.release(); } -NEFullyConnectedLayer::NEFullyConnectedLayer() - : _im2col_kernel(), _reshape_weights_kernel(), _interleave4x4_kernel(), _mm_kernel(), _accumulate_biases_kernel(), _im2col_output(), _interleave4x4_output(), _reshape_weights_output(), - _are_weights_reshaped(false), _is_batched_fc_layer(false), _linearize_input(false), _accumulate_biases(false) +NEFullyConnectedLayer::NEFullyConnectedLayer(std::shared_ptr memory_manager) + : _memory_group(std::move(memory_manager)), _im2col_kernel(), _reshape_weights_kernel(), _interleave4x4_kernel(), _mm_kernel(), _accumulate_biases_kernel(), _im2col_output(), _interleave4x4_output(), + _reshape_weights_output(), _are_weights_reshaped(false), _is_batched_fc_layer(false), _linearize_input(false), _accumulate_biases(false) { } @@ -191,6 +196,7 @@ void NEFullyConnectedLayer::configure(const ITensor *input, const ITensor *weigh _im2col_output.allocator()->init(TensorInfo(shape_im2col, 1, data_type, fixed_point_position)); // Configure im2col kernel + _memory_group.manage(&_im2col_output); _im2col_kernel.configure(input, &_im2col_output, Size2D(1, 1), PadStrideInfo(1, 1, 0, 0), false); multiply_input = &_im2col_output; @@ -204,6 +210,7 @@ void NEFullyConnectedLayer::configure(const ITensor *input, const ITensor *weigh _interleave4x4_output.allocator()->init(TensorInfo(shape_interleaved, 1, data_type, fixed_point_position)); // Configure interleave4x4 kernel + _memory_group.manage(&_interleave4x4_output); _interleave4x4_kernel.configure(multiply_input, &_interleave4x4_output); multiply_input = &_interleave4x4_output; @@ -248,6 +255,8 @@ void NEFullyConnectedLayer::run() _reshape_weights_kernel.run(); } + _memory_group.acquire(); + // Linearize input if it comes from a convolutional layer if(_linearize_input) { @@ -268,5 +277,7 @@ void NEFullyConnectedLayer::run() { NEScheduler::get().schedule(&_accumulate_biases_kernel, Window::DimY); } + + _memory_group.release(); } } // namespace arm_compute diff --git a/src/runtime/NEON/functions/NESoftmaxLayer.cpp b/src/runtime/NEON/functions/NESoftmaxLayer.cpp index 13dfa4a51e..cc5d4e91c3 100644 --- a/src/runtime/NEON/functions/NESoftmaxLayer.cpp +++ b/src/runtime/NEON/functions/NESoftmaxLayer.cpp @@ -31,8 +31,8 @@ using namespace arm_compute; -NESoftmaxLayer::NESoftmaxLayer() - : _max_kernel(), _shift_exp_sum_kernel(), _norm_kernel(), _fill_border_kernel(), _max(), _sum(), _tmp() +NESoftmaxLayer::NESoftmaxLayer(std::shared_ptr memory_manager) + : _memory_group(std::move(memory_manager)), _max_kernel(), _shift_exp_sum_kernel(), _norm_kernel(), _fill_border_kernel(), _max(), _sum(), _tmp() { } @@ -50,6 +50,11 @@ void NESoftmaxLayer::configure(ITensor *input, ITensor *output) _max.allocator()->init(tensor_info_max_sum); _sum.allocator()->init(tensor_info_max_sum); + // Manage intermediate buffers + _memory_group.manage(&_tmp); + _memory_group.manage(&_max); + _memory_group.manage(&_sum); + // Configure Kernels _max_kernel.configure(input, &_max); _shift_exp_sum_kernel.configure(input, &_max, &_tmp, &_sum); @@ -64,8 +69,12 @@ void NESoftmaxLayer::configure(ITensor *input, ITensor *output) void NESoftmaxLayer::run() { + _memory_group.acquire(); + NEScheduler::get().schedule(&_fill_border_kernel, Window::DimY); NEScheduler::get().schedule(&_max_kernel, Window::DimY); NEScheduler::get().schedule(&_shift_exp_sum_kernel, Window::DimY); NEScheduler::get().schedule(&_norm_kernel, Window::DimY); + + _memory_group.release(); } -- cgit v1.2.1