ncnn/src/layer/vulkan/shader/convolutiondepthwise_pack4.comp

// 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

#if NCNN_fp16_storage
#extension GL_EXT_shader_16bit_storage: require
#endif
#if NCNN_fp16_arithmetic
#extension GL_AMD_gpu_shader_half_float: require
#endif

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 (constant_id = 8) const int activation_type = 0;
layout (constant_id = 9) const float activation_param_0 = 0;
layout (constant_id = 10) const float activation_param_1 = 0;

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 { sfpvec4 bottom_blob_data[]; };
layout (binding = 1) writeonly buffer top_blob { sfpvec4 top_blob_data[]; };
layout (binding = 2) readonly buffer weight_blob { sfpvec4 weight_data[]; };
layout (binding = 3) readonly buffer bias_blob { sfpvec4 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;

    afpvec4 sum;

    if (bias_term == 1)
    {
        sum = afpvec4(bias_data[gz]);
    }
    else
    {
        sum = afpvec4(0.f);
    }

    // depth-wise convolution
    int w_offset = gz * kernel_w * kernel_h;
    int v_offset = gz * p.cstep + gy * stride_h * p.w + gx * stride_w;

    for (int y = 0; y < kernel_h; y++)
    {
        for (int x = 0; x < kernel_w; x++)
        {
            afpvec4 v = afpvec4(bottom_blob_data[v_offset + x * dilation_w]);

            afpvec4 k = afpvec4(weight_data[w_offset + x]);

            sum += v * k;
        }

        v_offset += dilation_h * p.w;
        w_offset += kernel_w;
    }

    if (activation_type == 1)
    {
        sum = max(sum, afp(0.f));
    }
    if (activation_type == 2)
    {
        const afp slope = afp(activation_param_0);
        sum = mix(sum, sum * afp(slope), lessThan(sum, afpvec4(0.f)));
    }
    if (activation_type == 3)
    {
        const afp const_min = afp(activation_param_0);
        const afp const_max = afp(activation_param_1);
        sum = clamp(sum, const_min, const_max);
    }
    if (activation_type == 4)
    {
        sum = afp(1.f) / (afp(1.f) + exp(-sum));
    }

    top_blob_data[gz * p.outcstep + gy * p.outw + gx] = sfpvec4(sum);
}