src/layer/convolution.cpp

// Tencent is pleased to support the open source community by making ncnn available.
//
// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
//
// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
// in compliance with the License. You may obtain a copy of the License at
//
// https://opensource.org/licenses/BSD-3-Clause
//
// Unless required by applicable law or agreed to in writing, software distributed
// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
// CONDITIONS OF ANY KIND, either express or implied. See the License for the
// specific language governing permissions and limitations under the License.

#include "convolution.h"

#include "layer_type.h"

namespace ncnn {

Convolution::Convolution()
{
    one_blob_only = true;
    support_inplace = false;
}

int Convolution::load_param(const ParamDict& pd)
{
    num_output = pd.get(0, 0);
    kernel_w = pd.get(1, 0);
    kernel_h = pd.get(11, kernel_w);
    dilation_w = pd.get(2, 1);
    dilation_h = pd.get(12, dilation_w);
    stride_w = pd.get(3, 1);
    stride_h = pd.get(13, stride_w);
    pad_left = pd.get(4, 0);
    pad_right = pd.get(15, pad_left);
    pad_top = pd.get(14, pad_left);
    pad_bottom = pd.get(16, pad_top);
    pad_value = pd.get(18, 0.f);
    bias_term = pd.get(5, 0);
    weight_data_size = pd.get(6, 0);
    int8_scale_term = pd.get(8, 0);
    activation_type = pd.get(9, 0);
    activation_params = pd.get(10, Mat());
    impl_type = pd.get(17, 0);

    if (int8_scale_term)
    {
#if NCNN_INT8
        support_int8_storage = true;
#else
        NCNN_LOGE("please build ncnn with NCNN_INT8 enabled for int8 inference");
        return -1;
#endif
    }

    return 0;
}

int Convolution::load_model(const ModelBin& mb)
{
    weight_data = mb.load(weight_data_size, 0);
    if (weight_data.empty())
        return -100;

    if (bias_term)
    {
        bias_data = mb.load(num_output, 1);
        if (bias_data.empty())
            return -100;
    }

#if NCNN_INT8
    if (int8_scale_term)
    {
        weight_data_int8_scales = mb.load(num_output, 1);
        bottom_blob_int8_scales = mb.load(1, 1);
    }

    if (int8_scale_term > 100)
    {
        top_blob_int8_scales = mb.load(1, 1);
    }
#endif // NCNN_INT8

    return 0;
}

int Convolution::create_pipeline(const Option& opt)
{
#if NCNN_INT8
    // runtime quantize the weight data
    if (opt.use_int8_inference && weight_data.elemsize == (size_t)4u && int8_scale_term)
    {
        const int maxk = kernel_w * kernel_h;
        const int num_input = weight_data_size / num_output / maxk;

        Mat weight_data_r2 = weight_data.reshape(maxk, num_input, num_output);

        Mat weight_data_int8;

        Option opt_q = opt;
        opt_q.blob_allocator = weight_data.allocator;
        opt_q.use_packing_layout = false;
        quantize_to_int8(weight_data_r2, weight_data_int8, weight_data_int8_scales, opt_q);
        if (weight_data_int8.empty())
            return -100;

        weight_data = weight_data_int8.reshape(weight_data_size);
    }
#endif // NCNN_INT8

    return 0;
}

int Convolution::forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const
{
    // convolv with NxN kernel
    // value = value + bias

#if NCNN_INT8
    if (opt.use_int8_inference && weight_data.elemsize == (size_t)1u)
    {
        return forward_int8(bottom_blob, top_blob, opt);
    }
#endif

    // flattened blob, implement as InnerProduct
    if (bottom_blob.dims == 1 && kernel_w == 1 && kernel_h == 1)
    {
        int num_input = weight_data_size / num_output;
        if (bottom_blob.w * bottom_blob.elempack == num_input)
        {
            // call InnerProduct
            ncnn::Layer* op = ncnn::create_layer(ncnn::LayerType::InnerProduct);

            // set param
            ncnn::ParamDict pd;
            pd.set(0, num_output);
            pd.set(1, bias_term);
            pd.set(2, weight_data_size);
            pd.set(8, int8_scale_term);
            pd.set(9, activation_type);
            pd.set(10, activation_params);

            op->load_param(pd);

            // set weights
            ncnn::Mat weights[4];
            weights[0] = weight_data;
            weights[1] = bias_data;

#if NCNN_INT8
            if (int8_scale_term)
            {
                weights[2] = weight_data_int8_scales;
                weights[3] = bottom_blob_int8_scales;
            }
#endif

            op->load_model(ModelBinFromMatArray(weights));

            op->create_pipeline(opt);

            // forward
            op->forward(bottom_blob, top_blob, opt);

            op->destroy_pipeline(opt);

            delete op;

            return 0;
        }
    }

    int w = bottom_blob.w;
    int h = bottom_blob.h;
    int channels = bottom_blob.c;
    size_t elemsize = bottom_blob.elemsize;

    //     NCNN_LOGE("Convolution input %d x %d  pad = %d %d  ksize=%d %d  stride=%d %d", w, h, pad_w, pad_h, kernel_w, kernel_h, stride_w, stride_h);

    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;

    Mat bottom_blob_bordered;
    make_padding(bottom_blob, bottom_blob_bordered, opt);
    if (bottom_blob_bordered.empty())
        return -100;

    w = bottom_blob_bordered.w;
    h = bottom_blob_bordered.h;

    int outw = (w - kernel_extent_w) / stride_w + 1;
    int outh = (h - kernel_extent_h) / stride_h + 1;

    const int maxk = kernel_w * kernel_h;

    // kernel offsets
    std::vector<int> _space_ofs(maxk);
    int* space_ofs = &_space_ofs[0];
    {
        int p1 = 0;
        int p2 = 0;
        int gap = w * dilation_h - kernel_w * dilation_w;
        for (int i = 0; i < kernel_h; i++)
        {
            for (int j = 0; j < kernel_w; j++)
            {
                space_ofs[p1] = p2;
                p1++;
                p2 += dilation_w;
            }
            p2 += gap;
        }
    }

    // float32
    top_blob.create(outw, outh, num_output, elemsize, opt.blob_allocator);
    if (top_blob.empty())
        return -100;

    // num_output
    #pragma omp parallel for num_threads(opt.num_threads)
    for (int p = 0; p < num_output; p++)
    {
        float* outptr = top_blob.channel(p);

        for (int i = 0; i < outh; i++)
        {
            for (int j = 0; j < outw; j++)
            {
                float sum = 0.f;

                if (bias_term)
                    sum = bias_data[p];

                const float* kptr = (const float*)weight_data + maxk * channels * p;

                // channels
                for (int q = 0; q < channels; q++)
                {
                    const Mat m = bottom_blob_bordered.channel(q);
                    const float* sptr = m.row(i * stride_h) + j * stride_w;

                    for (int k = 0; k < maxk; k++) // 29.23
                    {
                        float val = sptr[space_ofs[k]]; // 20.72
                        float wt = kptr[k];
                        sum += val * wt; // 41.45
                    }

                    kptr += maxk;
                }

                if (activation_type == 1)
                {
                    sum = std::max(sum, 0.f);
                }
                else if (activation_type == 2)
                {
                    float slope = activation_params[0];
                    sum = sum > 0.f ? sum : sum * slope;
                }
                else if (activation_type == 3)
                {
                    float min = activation_params[0];
                    float max = activation_params[1];
                    if (sum < min)
                        sum = min;
                    if (sum > max)
                        sum = max;
                }
                else if (activation_type == 4)
                {
                    sum = static_cast<float>(1.f / (1.f + exp(-sum)));
                }
                else if (activation_type == 5)
                {
                    const float MISH_THRESHOLD = 20;
                    float x = sum, y;
                    if (x > MISH_THRESHOLD)
                        y = x;
                    else if (x < -MISH_THRESHOLD)
                        y = expf(x);
                    else
                        y = logf(expf(x) + 1);
                    sum = static_cast<float>(x * tanh(y));
                }

                outptr[j] = sum;
            }

            outptr += outw;
        }
    }

    return 0;
}

void Convolution::make_padding(const Mat& bottom_blob, Mat& bottom_blob_bordered, const Option& opt) const
{
    int w = bottom_blob.w;
    int h = bottom_blob.h;

    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;

    bottom_blob_bordered = bottom_blob;
    if (pad_left > 0 || pad_right > 0 || pad_top > 0 || pad_bottom > 0)
    {
        Option opt_b = opt;
        opt_b.blob_allocator = opt.workspace_allocator;
        copy_make_border(bottom_blob, bottom_blob_bordered, pad_top, pad_bottom, pad_left, pad_right, BORDER_CONSTANT, pad_value, opt_b);
    }
    else if (pad_left == -233 && pad_right == -233 && pad_top == -233 && pad_bottom == -233)
    {
        // tensorflow padding=SAME or onnx padding=SAME_UPPER
        int wpad = kernel_extent_w + (w - 1) / stride_w * stride_w - w;
        int hpad = kernel_extent_h + (h - 1) / stride_h * stride_h - h;
        if (wpad > 0 || hpad > 0)
        {
            Option opt_b = opt;
            opt_b.blob_allocator = opt.workspace_allocator;
            copy_make_border(bottom_blob, bottom_blob_bordered, hpad / 2, hpad - hpad / 2, wpad / 2, wpad - wpad / 2, BORDER_CONSTANT, pad_value, opt_b);
        }
    }
    else if (pad_left == -234 && pad_right == -234 && pad_top == -234 && pad_bottom == -234)
    {
        // onnx padding=SAME_LOWER
        int wpad = kernel_extent_w + (w - 1) / stride_w * stride_w - w;
        int hpad = kernel_extent_h + (h - 1) / stride_h * stride_h - h;
        if (wpad > 0 || hpad > 0)
        {
            Option opt_b = opt;
            opt_b.blob_allocator = opt.workspace_allocator;
            copy_make_border(bottom_blob, bottom_blob_bordered, hpad - hpad / 2, hpad / 2, wpad - wpad / 2, wpad / 2, BORDER_CONSTANT, pad_value, opt_b);
        }
    }
}

#if NCNN_INT8
static inline signed char float2int8(float v)
{
    int int32 = static_cast<int>(round(v));
    if (int32 > 127) return 127;
    if (int32 < -127) return -127;
    return (signed char)int32;
}

int Convolution::forward_int8(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const
{
    int w = bottom_blob.w;
    int h = bottom_blob.h;
    int channels = bottom_blob.c;
    size_t elemsize = bottom_blob.elemsize;

    //     NCNN_LOGE("Convolution input %d x %d  ksize=%d %d  stride=%d %d", w, h, kernel_w, kernel_h, stride_w, stride_h);

    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;

    Mat bottom_blob_unbordered = bottom_blob;
    if (elemsize != 1)
    {
        Option opt_g = opt;
        opt_g.blob_allocator = opt.workspace_allocator;

        quantize_to_int8(bottom_blob, bottom_blob_unbordered, bottom_blob_int8_scales, opt_g);
    }

    Mat bottom_blob_bordered;
    make_padding(bottom_blob_unbordered, bottom_blob_bordered, opt);
    if (bottom_blob_bordered.empty())
        return -100;

    w = bottom_blob_bordered.w;
    h = bottom_blob_bordered.h;

    int outw = (w - kernel_extent_w) / stride_w + 1;
    int outh = (h - kernel_extent_h) / stride_h + 1;

    const int maxk = kernel_w * kernel_h;

    // kernel offsets
    std::vector<int> _space_ofs(maxk);
    int* space_ofs = &_space_ofs[0];
    {
        int p1 = 0;
        int p2 = 0;
        int gap = w * dilation_h - kernel_w * dilation_w;
        for (int i = 0; i < kernel_h; i++)
        {
            for (int j = 0; j < kernel_w; j++)
            {
                space_ofs[p1] = p2;
                p1++;
                p2 += dilation_w;
            }
            p2 += gap;
        }
    }

    // int8
    bool use_int8_requantize = int8_scale_term > 100;
    size_t out_elemsize = use_int8_requantize ? 1u : 4u;

    top_blob.create(outw, outh, num_output, out_elemsize, opt.blob_allocator);
    if (top_blob.empty())
        return -100;

    // num_output
    #pragma omp parallel for num_threads(opt.num_threads)
    for (int p = 0; p < num_output; p++)
    {
        signed char* outptr = top_blob.channel(p);

        for (int i = 0; i < outh; i++)
        {
            for (int j = 0; j < outw; j++)
            {
                int sum = 0;

                const signed char* kptr = (const signed char*)weight_data + maxk * channels * p;

                // channels
                for (int q = 0; q < channels; q++)
                {
                    const Mat m = bottom_blob_bordered.channel(q);
                    const signed char* sptr = m.row<signed char>(i * stride_h) + j * stride_w;

                    for (int k = 0; k < maxk; k++)
                    {
                        int val = sptr[space_ofs[k]];
                        int wt = kptr[k];
                        sum += val * wt;
                    }

                    kptr += maxk;
                }

                float scale_in;
                if (weight_data_int8_scales[p] == 0)
                    scale_in = 0;
                else
                    scale_in = 1.f / (bottom_blob_int8_scales[0] * weight_data_int8_scales[p]);

                float sumfp32 = sum * scale_in;

                if (bias_term)
                    sumfp32 += bias_data[p];

                if (activation_type == 1)
                {
                    sumfp32 = std::max(sumfp32, 0.f);
                }
                else if (activation_type == 2)
                {
                    float slope = activation_params[0];
                    sumfp32 = sumfp32 > 0.f ? sumfp32 : sumfp32 * slope;
                }
                else if (activation_type == 3)
                {
                    float min = activation_params[0];
                    float max = activation_params[1];
                    if (sumfp32 < min)
                        sumfp32 = min;
                    if (sumfp32 > max)
                        sumfp32 = max;
                }
                else if (activation_type == 4)
                {
                    sumfp32 = static_cast<float>(1.f / (1.f + exp(-sumfp32)));
                }
                else if (activation_type == 5)
                {
                    const float MISH_THRESHOLD = 20;
                    float x = sumfp32, y;
                    if (x > MISH_THRESHOLD)
                        y = x;
                    else if (x < -MISH_THRESHOLD)
                        y = expf(x);
                    else
                        y = logf(expf(x) + 1);
                    sumfp32 = static_cast<float>(x * tanh(y));
                }

                if (use_int8_requantize)
                {
                    // requantize
                    float scale_out = top_blob_int8_scales[0];
                    signed char sums8 = float2int8(sumfp32 * scale_out);
                    outptr[0] = sums8;
                    outptr += 1;
                }
                else
                {
                    // dequantize
                    ((float*)outptr)[0] = sumfp32;
                    outptr += 4;
                }
            }
        }
    }

    return 0;
}
#endif // NCNN_INT8

} // namespace ncnn