From 7068f9900d136312318ff430aef588b14e0c87ad Mon Sep 17 00:00:00 2001 From: Anthony Barbier Date: Thu, 26 Oct 2017 15:23:08 +0100 Subject: COMPMID-631: Merge branches/gles_compute branch Last commit: commit b25c5f68042b0c81bf611d59a1bb8535e1c42497 Author: Xinghang Zhou Date: Wed Oct 25 18:48:10 2017 +0800 Synced validation's tolerances of GCSoftmax from cl side Change-Id: Ibe72054205c1c8721845d679a31af7ed0a7c5cf6 Reviewed-on: http://mpd-gerrit.cambridge.arm.com/93283 Reviewed-by: Anthony Barbier Tested-by: Kaizen --- .../functions/GCFullyConnectedLayer.cpp | 177 +++++++++++++++++++++ 1 file changed, 177 insertions(+) create mode 100644 src/runtime/GLES_COMPUTE/functions/GCFullyConnectedLayer.cpp (limited to 'src/runtime/GLES_COMPUTE/functions/GCFullyConnectedLayer.cpp') diff --git a/src/runtime/GLES_COMPUTE/functions/GCFullyConnectedLayer.cpp b/src/runtime/GLES_COMPUTE/functions/GCFullyConnectedLayer.cpp new file mode 100644 index 0000000000..63cb40e616 --- /dev/null +++ b/src/runtime/GLES_COMPUTE/functions/GCFullyConnectedLayer.cpp @@ -0,0 +1,177 @@ +/* + * Copyright (c) 2017 ARM Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "arm_compute/runtime/GLES_COMPUTE/functions/GCFullyConnectedLayer.h" + +#include "arm_compute/core/Validate.h" +#include "arm_compute/runtime/GLES_COMPUTE/GCScheduler.h" +#include "support/ToolchainSupport.h" + +#include + +using namespace arm_compute; + +void GCFullyConnectedLayerReshapeWeights::configure(const IGCTensor *input, IGCTensor *output) +{ + auto k = arm_compute::support::cpp14::make_unique(); + k->configure(input, output); + _kernel = std::move(k); +} + +GCFullyConnectedLayer::GCFullyConnectedLayer() + : _im2col_kernel(), _reshape_weights_kernel(), _mm_kernel(), _accumulate_biases_kernel(), _im2col_output(), _reshape_weights_output(), _are_weights_reshaped(true), _is_fc_after_conv(true), + _accumulate_biases(false) +{ +} + +void GCFullyConnectedLayer::configure_conv_fc(const IGCTensor *input, const IGCTensor *weights, IGCTensor *output) +{ + ARM_COMPUTE_ERROR_ON((weights->info()->dimension(1) != (input->info()->dimension(0) * input->info()->dimension(1) * input->info()->dimension(2)))); + + const DataType dt = input->info()->data_type(); + + // If the fully connected layer is called after a convolution layer, the input tensor must be linearized + + // Initialize output tensor for im2col + TensorShape shape_im2col; + shape_im2col.set(0, input->info()->dimension(0) * input->info()->dimension(1) * input->info()->dimension(2)); + shape_im2col.set(1, input->info()->dimension(3)); + shape_im2col.set(2, input->info()->dimension(4)); + shape_im2col.set(3, input->info()->dimension(5)); + _im2col_output.allocator()->init(TensorInfo(shape_im2col, 1, dt)); + + // Configure im2col kernel + _im2col_kernel.configure(input, &_im2col_output, std::make_pair(1, 1), PadStrideInfo(1, 1, 0, 0), false); + + // Configure matrix multiply kernel + _mm_kernel.configure(&_im2col_output, weights, output, 1.0f, false); + + // Allocate the output tensor for im2col once all the configure methods have been called + _im2col_output.allocator()->allocate(); +} + +void GCFullyConnectedLayer::configure_fc_fc(const IGCTensor *input, const IGCTensor *weights, IGCTensor *output) +{ + ARM_COMPUTE_ERROR_ON(input->info()->dimension(0) != weights->info()->dimension(1)); + + // Configure matrix multiply kernel + _mm_kernel.configure(input, weights, output, 1.0f, false); +} + +void GCFullyConnectedLayer::configure(const IGCTensor *input, const IGCTensor *weights, const IGCTensor *biases, IGCTensor *output, bool transpose_weights, bool are_weights_reshaped) +{ + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32, DataType::F16); + ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, output); + ARM_COMPUTE_ERROR_ON(weights->info()->num_dimensions() > 2); + + _are_weights_reshaped = transpose_weights ? are_weights_reshaped : true; + _is_fc_after_conv = true; + _accumulate_biases = false; + + if(biases != nullptr) + { + ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases); + + _accumulate_biases = true; + + // Configure accumulate biases kernel + _accumulate_biases_kernel.configure(output, biases); + } + + // With the Fully Connected layer we can have 4 different cases: + // 1) Convolution layer -> Fully Connected layer without batches + // 2) Fully Connected layer -> Fully Connected layer without batches + // 3) Convolution layer -> Fully Connected layer with batches + // 4) Fully Connected layer -> Fully Connected layer with batches + + const IGCTensor *weights_to_use = weights; + + if(!_are_weights_reshaped) + { + weights_to_use = &_reshape_weights_output; + + // Reshape the weights + _reshape_weights_kernel.configure(weights, &_reshape_weights_output); + } + + // Check if we have a fully connected layer with batches + const bool is_batched_fc_layer = output->info()->dimension(1) > 1; + + if(is_batched_fc_layer) + { + _is_fc_after_conv = (TensorShape::num_max_dimensions >= 4) && (std::equal(input->info()->tensor_shape().cbegin() + 3, + input->info()->tensor_shape().cend(), + output->info()->tensor_shape().cbegin() + 1)); + } + else + { + _is_fc_after_conv = input->info()->num_dimensions() > 1; + } + + if(_is_fc_after_conv) + { + // Fully Connected layer after a Convolution Layer without batches + configure_conv_fc(input, weights_to_use, output); + } + else + { + // Fully Connected layer after a Fully Connected Layer without batches + configure_fc_fc(input, weights_to_use, output); + } + + // Allocate the transpose tensor if the are_weights_reshaped flag is false and once all the configure methods have been called + if(!_are_weights_reshaped) + { + // Allocate the tensor for the weights reshaped + _reshape_weights_output.allocator()->allocate(); + } +} + +void GCFullyConnectedLayer::run() +{ + // Reshape of the weights (happens only once) + if(!_are_weights_reshaped) + { + _are_weights_reshaped = true; + _reshape_weights_kernel.run(); + } + + // Linearize input if it comes from a convolutional layer + if(_is_fc_after_conv) + { + GCScheduler::get().enqueue(_im2col_kernel, false); + } + + GCScheduler::get().sync(); + + // Run matrix multiply + GCScheduler::get().enqueue(_mm_kernel, !_accumulate_biases); + + // Accumulate biases if provided + if(_accumulate_biases) + { + GCScheduler::get().sync(); + + GCScheduler::get().enqueue(_accumulate_biases_kernel); + } +} -- cgit v1.2.1