src/cpu/kernels/fuse_batch_normalization/generic/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/fuse_batch_normalization/generic/impl.h"

namespace arm_compute
{
namespace cpu
{
template <typename T>
void fused_batch_normalization_conv(const ITensor *conv_weights, const ITensor *conv_bias, ITensor *fused_weights, ITensor *fused_bias,
                                    const ITensor *bn_mean, const ITensor *bn_var, const ITensor *bn_beta, const ITensor *bn_gamma, float epsilon, const Window &window)
{
    using ScalarType   = T;
    const int size     = 16 / conv_weights->info()->element_size();
    using ExactTagType = typename wrapper::traits::neon_bitvector_tag_t<T, wrapper::traits::BitWidth::W128>;

    const bool run_in_place_weights = (fused_weights == nullptr) || (fused_weights == conv_weights);
    const bool run_in_place_bias    = (fused_bias == nullptr) || (conv_bias != nullptr && fused_bias == conv_bias);

    // Set build options
    Window win = window;
    win.set(Window::DimX, Window::Dimension(0, 1, 1));

    const int  window_step_x  = size;
    const auto window_start_x = static_cast<int>(window.x().start());
    const auto window_end_x   = static_cast<int>(window.x().end());

    Iterator conv_w_in(conv_weights, win);
    Iterator conv_w_out(run_in_place_weights ? conv_weights : fused_weights, win);

    const auto conv_bias_in  = (conv_bias != nullptr ? reinterpret_cast<ScalarType *>(conv_bias->ptr_to_element(Coordinates(0, 0))) : nullptr);
    auto       conv_bias_out = (run_in_place_bias ? conv_bias_in : reinterpret_cast<ScalarType *>(fused_bias->ptr_to_element(Coordinates(0, 0))));

    const auto input_mean  = reinterpret_cast<const ScalarType *>(bn_mean->ptr_to_element(Coordinates(0, 0)));
    const auto input_var   = reinterpret_cast<const ScalarType *>(bn_var->ptr_to_element(Coordinates(0, 0)));
    const auto input_gamma = (bn_gamma != nullptr) ? reinterpret_cast<const ScalarType *>(bn_gamma->ptr_to_element(Coordinates(0, 0))) : nullptr;
    const auto input_beta  = (bn_beta != nullptr) ? reinterpret_cast<const ScalarType *>(bn_beta->ptr_to_element(Coordinates(0, 0))) : nullptr;

    auto       mean_vec    = wrapper::vdup_n(ScalarType(0), ExactTagType{});
    auto       var_vec     = wrapper::vdup_n(ScalarType(0), ExactTagType{});
    auto       gamma_vec   = wrapper::vdup_n(ScalarType(1), ExactTagType{});
    auto       beta_vec    = wrapper::vdup_n(ScalarType(0), ExactTagType{});
    auto       rvar_vec    = wrapper::vdup_n(ScalarType(0), ExactTagType{});
    const auto epsilon_vec = wrapper::vdup_n(ScalarType(epsilon), ExactTagType{});

    auto mean                = ScalarType(0.0);
    auto var                 = ScalarType(0.0);
    auto gamma               = ScalarType(1.0);
    auto beta                = ScalarType(0.0);
    auto conv_bias_in_scalar = ScalarType(0.0);
    execute_window_loop(win, [&](const Coordinates & id)
    {
        var = input_var[id[3]];
        if(input_gamma != nullptr)
        {
            gamma = input_gamma[id[3]];
        }

        if((id[0] == 0) && (id[1] == 0) && (id[2] == 0))
        {
            if(input_beta != nullptr)
            {
                beta     = input_beta[id[3]];
                beta_vec = wrapper::vdup_n(beta, ExactTagType{});
            }

            // Construct vectors
            mean     = input_mean[id[3]];
            mean_vec = wrapper::vdup_n(mean, ExactTagType{});

            if(conv_bias_in != nullptr)
            {
                conv_bias_in_scalar = conv_bias_in[id[3]];
            }
            auto conv_bias_tmp_scalar = (conv_bias_in_scalar - mean) / std::sqrt(var + ScalarType(epsilon));
            conv_bias_out[id[3]]      = (conv_bias_tmp_scalar * gamma) + beta;
        }

        int  x              = window_start_x;
        auto conv_w_in_ptr  = reinterpret_cast<const ScalarType *>(conv_w_in.ptr());
        auto conv_w_out_ptr = reinterpret_cast<ScalarType *>(conv_w_out.ptr());
        var_vec             = wrapper::vdup_n(var, ExactTagType{});
        gamma_vec           = wrapper::vdup_n(gamma, ExactTagType{});
        rvar_vec            = wrapper::vinvsqrt(wrapper::vadd(var_vec, epsilon_vec));

        for(; x <= (window_end_x - window_step_x); x += window_step_x)
        {
            auto wn = wrapper::vloadq(conv_w_in_ptr + x);
            wn      = wrapper::vmul(wn, rvar_vec);
            wn      = wrapper::vmul(wn, gamma_vec);

            // Store results
            wrapper::vstore(conv_w_out_ptr + x, wn);
        }

        // Compute left-over elements
        for(; x < window_end_x; ++x)
        {
            *(conv_w_out_ptr + x) = *(conv_w_in_ptr + x) / std::sqrt(var + ScalarType(epsilon)) * gamma;
        }
    },
    conv_w_in, conv_w_out);
}

template void fused_batch_normalization_conv<float32_t>(const ITensor *conv_weights, const ITensor *conv_bias, ITensor *fused_weights, ITensor *fused_bias,
                                                        const ITensor *bn_mean, const ITensor *bn_var, const ITensor *bn_beta, const ITensor *bn_gamma, float epsilon, const Window &window);

#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS)
template void fused_batch_normalization_conv<float16_t>(const ITensor *conv_weights, const ITensor *conv_bias, ITensor *fused_weights, ITensor *fused_bias,
                                                        const ITensor *bn_mean, const ITensor *bn_var, const ITensor *bn_beta, const ITensor *bn_gamma, float epsilon, const Window &window);
#endif /* defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) */

} // namespace cpu
} // namespace arm_compute