src/cpu/kernels/directconv2d/nhwc/neon/impl.cpp

/*
 * Copyright (c) 2018-2022 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 "src/cpu/kernels/directconv2d/nhwc/neon/impl.h"

#include "src/core/NEON/kernels/detail/NEDirectConvolutionDetail.h"
#include "src/core/NEON/wrapper/wrapper.h"

#include "arm_compute/core/Error.h"
#include "arm_compute/core/Helpers.h"
#include "arm_compute/core/IAccessWindow.h"
#include "arm_compute/core/ITensor.h"
#include "arm_compute/core/Types.h"
#include "arm_compute/core/Utils.h"
#include "src/core/helpers/WindowHelpers.h"

#include <algorithm>

using namespace arm_compute::detail;

namespace arm_compute
{
namespace cpu
{
namespace kernels
{
namespace
{
bool have_zero_x_internal_padding(ITensorInfo *src, const ITensorInfo *weights)
{
    return (src->padding().left == 0 && weights->padding().left == 0 && src->padding().right == 0 && weights->padding().right == 0);
}
}

template <typename T>
void convolve_nhwc(const Window &window, const ITensor *src, const ITensor *weights, ITensor *dst, const PadStrideInfo &conv_info)
{
    // Declare useful types
    using vtype       = wrapper::traits::neon_bitvector<T, wrapper::traits::BitWidth::W128>;
    using vector_type = typename vtype::type;
    using tag_type    = typename vtype::tag_type;

    // Scalar quantities
    const int element_size   = src->info()->element_size();
    const int input_stride_w = src->info()->strides_in_bytes().y() / element_size;
    const int input_stride_h = src->info()->strides_in_bytes().z() / element_size;
    const int input_stride_n = src->info()->strides_in_bytes()[3] / element_size;
    const int input_dim_w    = src->info()->dimension(1);
    const int input_dim_h    = src->info()->dimension(2);

    const int output_stride_c = dst->info()->strides_in_bytes().x();

    const unsigned int kernel_stride_w = weights->info()->strides_in_bytes().y() / element_size;
    const unsigned int kernel_stride_h = weights->info()->strides_in_bytes().z() / element_size;
    const int          kernel_dim_w    = weights->info()->dimension(1);
    const int          kernel_dim_h    = weights->info()->dimension(2);

    const int conv_pad_top  = conv_info.pad_top();
    const int conv_pad_left = conv_info.pad_left();
    const int conv_stride_w = std::get<0>(conv_info.stride());
    const int conv_stride_h = std::get<1>(conv_info.stride());

    // Setup input window for the output iterator
    Window window_out = window;
    window_out.set(Window::DimX, Window::Dimension(0, 1, 1));

    // Setup input window for the weights iterator
    Window window_w = calculate_max_window(*weights->info(), Steps());
    window_w.set(Window::DimX, Window::Dimension(0, 1, 1));
    window_w.set(Window::DimY, Window::Dimension(0, 1, 1));
    window_w.set(Window::DimZ, Window::Dimension(0, 1, 1));

    Iterator out(dst, window_out);
    Iterator wei(weights, window_w);

    constexpr int num_elems_read_per_iteration = 16 / sizeof(T);

    // nhwc optimized
    if(have_zero_x_internal_padding(src->info(), weights->info()))
    {
        // This function assumes that input and weights have not padding in channel

        /*
        * This implementation parallelize the full WC plane of input and weights by
        * treating them as series of elements. So for example, a 3x3 weights and
        * floating point vector operations of 4 elements per time, the first 3
        * channel elements of the first row would be taken and additionally the first
        * element of the second row. The 9 elements in each single WC weight plane
        * would require 2 4-element vector operations and a last single element operation.
        *
        * This works since when we create the input vector to multiply with the weights,
        * the exact required elements are loaded in the same order. Therefore the
        * multiplication works on the correct input/weight elements.
        */
        execute_window_loop(
            window_out, [&](const Coordinates & id)
        {
            /*
            * In here we create theoretical indexes which then we validate for both
            * inputs and weights.
            * As a reminder, this loop take each output point in NHW, C is treated
            * in the weights loop.
            */
            // We are computing the theoretical starting input starting points
            const int in_w_start_t = static_cast<int>(id.y()) * conv_stride_w - conv_pad_left;
            const int in_h_start_t = static_cast<int>(id.z()) * conv_stride_h - conv_pad_top;
            const int in_w_end_t   = in_w_start_t + kernel_dim_w;
            const int in_h_end_t   = in_h_start_t + kernel_dim_h;

            // We are computing the valid initial and ending input points by checking the borders
            const int in_w_start = std::max(in_w_start_t, 0);
            const int in_h_start = std::max(in_h_start_t, 0);
            const int in_w_end   = std::min(in_w_end_t, input_dim_w);
            const int in_h_end   = std::min(in_h_end_t, input_dim_h);

            // We use the input points to select the valid weight points to use
            const int index_wc_start = (in_w_start - in_w_start_t) * kernel_stride_w;
            const int index_h_start  = in_h_start - in_h_start_t;
            const int index_wc_end   = (kernel_dim_w - (in_w_end_t - in_w_end)) * kernel_stride_w;
            const int index_h_end    = kernel_dim_h - (in_h_end_t - in_h_end);

            execute_window_loop(
                window_w, [&](const Coordinates & id_w)
            {
                /*
                * This is the loop in the weights, and it goes along N (the batches)
                * As a reminder, the batches of the weights are translated into the
                * channels of the output
                */
                const T *in_ptr_row = reinterpret_cast<const T *>(src->buffer() + src->info()->offset_first_element_in_bytes())
                                      + id[3] * input_stride_n + in_w_start * input_stride_w + in_h_start * input_stride_h;
                const T *weights_ptr_row = reinterpret_cast<const T *>(wei.ptr()) + index_h_start * kernel_stride_h;
                uint8_t *out_ptr         = out.ptr() + id_w[3] * output_stride_c;

                T out_temp = static_cast<T>(0);
                for(int index_h = index_h_start; index_h < index_h_end; ++index_h, in_ptr_row += input_stride_h, weights_ptr_row += kernel_stride_h)
                {
                    const T    *in_ptr_mover = in_ptr_row;
                    int         index_wc     = index_wc_start;
                    vector_type out_temp_vec = wrapper::vdup_n(static_cast<T>(0), tag_type());
                    for(; index_wc <= index_wc_end - num_elems_read_per_iteration; index_wc += num_elems_read_per_iteration, in_ptr_mover += num_elems_read_per_iteration)
                    {
                        const auto src_vec = wrapper::vloadq(in_ptr_mover);
                        const auto w_vec   = wrapper::vloadq(weights_ptr_row + index_wc);
                        out_temp_vec       = wrapper::vmla(out_temp_vec, w_vec, src_vec);
                    }
                    out_temp += vreduce(out_temp_vec);
                    for(; index_wc < index_wc_end; ++index_wc, ++in_ptr_mover)
                    {
                        const auto src_val = *(in_ptr_mover);
                        const auto w_val   = *(weights_ptr_row + index_wc);
                        out_temp += src_val * w_val;
                    }
                }
                *(reinterpret_cast<T *>(out_ptr)) = out_temp;
            },
            wei);
        },
        out);
    }
    else // nhwc non optimized
    {
        execute_window_loop(
            window_out, [&](const Coordinates & id)
        {
            // We are computing the theoretical starting input starting points
            const int in_w_start_t = static_cast<int>(id.y()) * conv_stride_w - conv_pad_left;
            const int in_h_start_t = static_cast<int>(id.z()) * conv_stride_h - conv_pad_top;
            const int in_w_end_t   = in_w_start_t + kernel_dim_w;
            const int in_h_end_t   = in_h_start_t + kernel_dim_h;

            // We are computing the valid initial and ending input points by checking the borders
            const int in_w_start = std::max(in_w_start_t, 0);
            const int in_h_start = std::max(in_h_start_t, 0);
            const int in_w_end   = std::min(in_w_end_t, input_dim_w);
            const int in_h_end   = std::min(in_h_end_t, input_dim_h);

            // We use the input points to select the valid weight points to use
            const int wei_w_start = in_w_start - in_w_start_t;
            const int wei_h_start = in_h_start - in_h_start_t;
            const int wei_w_end   = kernel_dim_w - (in_w_end_t - in_w_end);
            const int wei_h_end   = kernel_dim_h - (in_h_end_t - in_h_end);

            const int      index_c_end  = weights->info()->dimension(0);
            const T *const in_ptr_start = reinterpret_cast<const T *>(src->buffer() + src->info()->offset_first_element_in_bytes()) + id[3] * input_stride_n;

            execute_window_loop(
                window_w, [&](const Coordinates & id_w)
            {
                const T *const weights_ptr_start = reinterpret_cast<const T *>(wei.ptr());
                uint8_t       *out_ptr           = out.ptr() + id_w[3] * output_stride_c;

                T out_temp = static_cast<T>(0);
                for(int index_wei_h = wei_h_start, index_in_h = in_h_start; index_wei_h < wei_h_end; ++index_wei_h, ++index_in_h)
                {
                    const T *const in_ptr_row      = in_ptr_start + index_in_h * input_stride_h;
                    const T *const weights_ptr_row = weights_ptr_start + index_wei_h * kernel_stride_h;
                    for(int index_wei_w = wei_w_start, index_in_w = in_w_start; index_wei_w < wei_w_end; ++index_wei_w, ++index_in_w)
                    {
                        const T    *in_ptr_mover      = in_ptr_row + index_in_w * input_stride_w;
                        const T    *weights_ptr_mover = weights_ptr_row + index_wei_w * kernel_stride_w;
                        int         index_c           = 0;
                        vector_type out_temp_vec      = wrapper::vdup_n(static_cast<T>(0), tag_type());
                        for(; index_c <= index_c_end - num_elems_read_per_iteration; index_c += num_elems_read_per_iteration, in_ptr_mover += num_elems_read_per_iteration, weights_ptr_mover += num_elems_read_per_iteration)
                        {
                            const auto src_vec = wrapper::vloadq(in_ptr_mover);
                            const auto w_vec   = wrapper::vloadq(weights_ptr_mover);
                            out_temp_vec       = wrapper::vmla(out_temp_vec, w_vec, src_vec);
                        }
                        out_temp += vreduce(out_temp_vec);
                        for(; index_c < index_c_end; ++index_c, ++in_ptr_mover, ++weights_ptr_mover)
                        {
                            const auto src_val = *(in_ptr_mover);
                            const auto w_val   = *(weights_ptr_mover);
                            out_temp += src_val * w_val;
                        }
                    }
                }
                *(reinterpret_cast<T *>(out_ptr)) = out_temp;
            },
            wei);
        },
        out);
    }
}

template void convolve_nhwc<float>(const Window &window, const ITensor *src, const ITensor *weights, ITensor *dst, const PadStrideInfo &conv_info);

} // namespace kernels
} // namespace cpu
} // namespace arm_compute