src/runtime/NEON/functions/NEConvolutionLayer.cpp

/*
 * 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/NEON/functions/NEConvolutionLayer.h"

#include "arm_compute/core/NEON/kernels/arm32/NEGEMMAArch32Kernel.h"
#include "arm_compute/core/NEON/kernels/arm64/NEGEMMAArch64Kernel.h"
#include "arm_compute/core/PixelValue.h"
#include "arm_compute/core/Size2D.h"
#include "arm_compute/core/Utils.h"
#include "arm_compute/core/Validate.h"
#include "arm_compute/runtime/NEON/NEScheduler.h"
#include "support/ToolchainSupport.h"

namespace arm_compute
{
#include "arm_compute/core/NEON/kernels/assembly/gemm_interleaved.hpp"
#include "arm_compute/core/NEON/kernels/assembly/kernels/a32_sgemm_8x6.hpp"
#include "arm_compute/core/NEON/kernels/assembly/kernels/a64_sgemm_12x8.hpp"
} // namespace arm_compute

#include <cmath>
#include <tuple>

namespace arm_compute
{
NEConvolutionLayerReshapeWeights::NEConvolutionLayerReshapeWeights(std::shared_ptr<IMemoryManager> memory_manager)
    : _memory_group(std::move(memory_manager)), _weights_reshape_kernel(), _weights_transposed_kernel(), _weights_reshaped(), _transpose1xW(false)
{
}

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);
    }

    // Check if bias are present, if yes they will be embedded to the weights matrix
    const bool _has_bias = (biases != nullptr);

    _transpose1xW = transpose1xW;

    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());

        _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();
    }
    else
    {
        _weights_reshape_kernel.configure(weights, biases, output);
    }
}

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(std::shared_ptr<IMemoryManager> 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)
{
    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);

    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);
    }

    const DataType dt                   = input->info()->data_type();
    const int      fixed_point_position = input->info()->fixed_point_position();

    _has_bias             = (biases != nullptr);
    _are_weights_reshaped = weights_info.are_reshaped();

    // Get parameters from conv_info
    unsigned int stride_x = 0;
    unsigned int stride_y = 0;
    unsigned int pad_x    = 0;
    unsigned int pad_y    = 0;
    std::tie(stride_x, stride_y) = conv_info.stride();
    std::tie(pad_x, pad_y)       = conv_info.pad();

    // Get convolved dimensions
    unsigned int conv_w = 0;
    unsigned int conv_h = 0;

    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);

    // 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));

#if defined(__arm__)
    if(NEScheduler::get().cpu_info().CPU == CPUTarget::ARMV7 && dt == DataType::F32)
    {
        _mm_optimised_kernel = support::cpp14::make_unique<NEGEMMAArch32Kernel>();
    }
#elif defined(__aarch64__)
    if(NEScheduler::get().cpu_info().CPU >= CPUTarget::ARMV8 && dt == DataType::F32)
    {
        _mm_optimised_kernel = support::cpp14::make_unique<NEGEMMAArch64Kernel>();
    }
#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)
    {
        if(_are_weights_reshaped)
        {
            mat_weights_cols = weights_info.num_kernels();
            mat_weights_rows = weights->info()->dimension(1);
        }
        else
        {
            TensorShape reshaped_weights_shape{ mat_weights_cols, mat_weights_rows };

            // Create tensor to store the reshaped weights
            _weights_reshaped.allocator()->init(TensorInfo(reshaped_weights_shape, 1, dt, fixed_point_position));
            _reshape_weights.configure(weights, biases, &_weights_reshaped, false /* 1xW transpose */);
            weights = &_weights_reshaped;
        }
    }
    else
    {
        if(_are_weights_reshaped)
        {
            mat_weights_cols = weights_info.num_kernels();
            mat_weights_rows = weights->info()->dimension(0) / 4 + (_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->info()->element_size();
                reshaped_weights_shape      = TensorShape{ mat_weights_rows *static_cast<unsigned int>(transpose_width),
                                                           static_cast<unsigned int>(std::ceil(mat_weights_cols / transpose_width)) };
            }

            // Create tensor to store the reshaped weights
            _weights_reshaped.allocator()->init(TensorInfo(reshaped_weights_shape, 1, dt, fixed_point_position));
            _reshape_weights.configure(weights, biases, &_weights_reshaped, !_is_fully_connected_convolution /* 1xW transpose */);
            weights = &_weights_reshaped;
        }
    }

    // Create tensor to store im2col reshaped inputs
    const unsigned int mat_input_cols = mat_weights_rows;
    const unsigned int mat_input_rows = conv_w * conv_h;

    TensorShape shape_im2col(input->info()->tensor_shape());
    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));
    _memory_group.manage(&_input_im2col_reshaped);

    // Create tensor (interleave) to prepare input tensor for GEMM
    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));
        _input_interleaved_reshaped.allocator()->init(TensorInfo(shape_interleaved, 1, dt, fixed_point_position));
        _memory_group.manage(&_input_interleaved_reshaped);
    }

    // Create GEMM output tensor
    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));
    _memory_group.manage(&_gemm_output);

    // Configure kernels
    _input_im2col_kernel.configure(input, &_input_im2col_reshaped, Size2D(kernel_width, kernel_height), conv_info, _has_bias);

#if defined(__arm__) || defined(__aarch64__)
    if(_mm_optimised_kernel != nullptr)
    {
        struct CPUInfo ci = NEScheduler::get().cpu_info();

        const int M = _gemm_output.info()->tensor_shape().y();
        const int N = _gemm_output.info()->tensor_shape().x();
        const int K = _input_im2col_reshaped.info()->tensor_shape().x();

#if defined(__arm__)
        GemmInterleaved<sgemm_8x6, float, float> gemm(&ci, M, N, K, false, false);
#elif defined(__aarch64__)
        GemmInterleaved<sgemm_12x8, float, float> gemm(&ci, M, N, K, false, false);
#endif /* defined(__arm__) || defined(__aarch64__) */

        constexpr size_t alignment = 4096;
        _workspace.allocator()->init(TensorInfo(TensorShape{ (gemm.get_working_size() + alignment - 1) * NEScheduler::get().num_threads() }, 1, DataType::U8));
        _memory_group.manage(&_workspace);

        // Configure matrix multiplication kernel
        if(_is_fully_connected_convolution)
        {
            _mm_optimised_kernel->configure(&_input_im2col_reshaped, weights, &_gemm_output, &_workspace, 1.f, 0.f, false, false);
        }
        else
        {
            _mm_optimised_kernel->configure(&_input_im2col_reshaped, weights, &_gemm_output, &_workspace);
        }

        _workspace.allocator()->allocate();
    }
    else
#endif /* defined(__arm__) || defined(__aarch64__) */
    {
        if(_is_fully_connected_convolution)
        {
            _mm_kernel.configure(&_input_im2col_reshaped, weights, &_gemm_output, 1.0f);
        }
        else
        {
            _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");

    // Allocate intermediate tensor
    if(!_are_weights_reshaped)
    {
        _weights_reshaped.allocator()->allocate();
    }
}

void NEConvolutionLayer::run()
{
    // Run weights reshaping (Runs once for every configure)
    if(!_are_weights_reshaped)
    {
        _are_weights_reshaped = true;
        _reshape_weights.run();
    }

    _memory_group.acquire();

    // Run input reshaping
    NEScheduler::get().schedule(&_input_im2col_kernel, Window::DimY);

    // Runs matrix multiply on reshaped matrices
    if(_mm_optimised_kernel != nullptr)
    {
        NEScheduler::get().schedule(_mm_optimised_kernel.get(), Window::DimY);
    }
    else
    {
        if(!_is_fully_connected_convolution)
        {
            // Run interleave
            NEScheduler::get().schedule(&_input_interleave_kernel, Window::DimY);
        }

        NEScheduler::get().schedule(&_mm_kernel, Window::DimY);
    }

    // Reshape output matrix
    NEScheduler::get().schedule(&_output_col2im_kernel, Window::DimY);

    _memory_group.release();
}
} // namespace arm_compute