// Tencent is pleased to support the open source community by making ncnn available.
//
// Copyright (C) 2019 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.

#version 450

layout (constant_id = 0) const int kernel_w = 1;
layout (constant_id = 1) const int kernel_h = 1;
layout (constant_id = 2) const int dilation_w = 1;
layout (constant_id = 3) const int dilation_h = 1;
layout (constant_id = 4) const int stride_w = 1;
layout (constant_id = 5) const int stride_h = 1;
layout (constant_id = 6) const int bias_term = 0;
layout (constant_id = 7) const int group = 1;

layout (local_size_x_id = 233) in;
layout (local_size_y_id = 234) in;
layout (local_size_z_id = 235) in;

layout (binding = 0) readonly buffer bottom_blob { vec4 bottom_blob_data[]; };
layout (binding = 1) writeonly buffer top_blob { vec4 top_blob_data[]; };
layout (binding = 2) readonly buffer weight_blob { mat4 weight_data[]; };
layout (binding = 3) readonly buffer bias_blob { vec4 bias_data[]; };

layout (push_constant) uniform parameter
{
    int dims;
    int w;
    int h;
    int c;
    int cstep;

    int outdims;
    int outw;
    int outh;
    int outc;
    int outcstep;
} p;

void main()
{
    int gx = int(gl_GlobalInvocationID.x);
    int gy = int(gl_GlobalInvocationID.y);
    int gz = int(gl_GlobalInvocationID.z);

    if (gx >= p.outw || gy >= p.outh || gz >= p.outc)
        return;

    vec4 sum;

    if (bias_term == 1)
    {
        sum = bias_data[gz];
    }
    else
    {
        sum = vec4(0.f);
    }

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

    // group convolution
    const int channels_g = p.c / group;
    const int num_output_g = p.outc / group;

    // group id
    const int gg = gz / num_output_g;

    int w_offset_0 = gz * channels_g * kernel_w * kernel_h;
    int v_offset_0 = gg * channels_g * p.cstep;

    for (int y = 0; y < kernel_h; y++)
    {
        int sys = (gy + y * dilation_h - (kernel_extent_h - 1));
        if (sys % stride_h != 0)
            continue;

        int sy = sys / stride_h;
        if (sy < 0 || sy >= p.h)
            continue;

        for (int x = 0; x < kernel_w; x++)
        {
            int sxs = (gx + x * dilation_w - (kernel_extent_w - 1));
            if (sxs % stride_w != 0)
                continue;

            int sx = sxs / stride_w;
            if (sx < 0 || sx >= p.w)
                continue;

            int v_offset = v_offset_0 + sy * p.w + sx;
            int w_offset = w_offset_0 + y * kernel_w + x;

            for (int z = 0; z < channels_g; z++)
            {
                vec4 v = bottom_blob_data[v_offset];

                mat4 k = weight_data[w_offset];

                sum += v * k;

                v_offset += p.cstep;
                w_offset += kernel_w * kernel_h;
            }
        }
    }

    top_blob_data[gz * p.outcstep + gy * p.outw + gx] = sum;
}