src/layer/vulkan/convolution_vulkan.cpp

// 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.

#include "convolution_vulkan.h"

#include "layer_shader_type.h"
#include "layer_type.h"

namespace ncnn {

Convolution_vulkan::Convolution_vulkan()
{
    support_vulkan = true;
    support_image_storage = true;

    padding = 0;

    pipeline_convolution = 0;
    pipeline_convolution_1x1s1d1 = 0;

    pipeline_convolution_gemm = 0;

    pipeline_convolution_3x3s1d1_winograd23_transform_input = 0;
    pipeline_convolution_3x3s1d1_winograd23_gemm = 0;
    pipeline_convolution_3x3s1d1_winograd23_transform_output = 0;

    pipeline_convolution_3x3s1d1_winograd43_transform_input = 0;
    pipeline_convolution_3x3s1d1_winograd43_gemm = 0;
    pipeline_convolution_3x3s1d1_winograd43_transform_output = 0;

    reshape_1x1xw = 0;
    reshape_w = 0;
}

int Convolution_vulkan::create_pipeline(const Option& _opt)
{
    if (dynamic_weight)
    {
        support_vulkan = false;
        support_image_storage = false;
        return 0;
    }

    Option opt = _opt;
    const Mat& shape = bottom_shapes.empty() ? Mat() : bottom_shapes[0];
    const Mat& out_shape = top_shapes.empty() ? Mat() : top_shapes[0];

    const int maxk = kernel_w * kernel_h;
    int num_input = weight_data_size / maxk / num_output;

    // the shape after padding
    Mat shape_bordered;
    if (shape.dims != 0)
    {
        if (pad_left > 0 || pad_right > 0 || pad_top > 0 || pad_bottom > 0)
        {
            shape_bordered = Mat(shape.w + pad_left + pad_right, shape.h + pad_top + pad_bottom, shape.c, (void*)0);
        }
        else if ((pad_left == -233 && pad_right == -233 && pad_top == -233 && pad_bottom == -233)
                 || (pad_left == -234 && pad_right == -234 && pad_top == -234 && pad_bottom == -234))
        {
            const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
            const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;

            int wpad = kernel_extent_w + (shape.w - 1) / stride_w * stride_w - shape.w;
            int hpad = kernel_extent_h + (shape.h - 1) / stride_h * stride_h - shape.h;
            if (wpad > 0 || hpad > 0)
            {
                shape_bordered = Mat(shape.w + wpad, shape.h + hpad, shape.c, (void*)0);
            }
        }
        else
        {
            shape_bordered = shape;
        }
    }

    int elempack = opt.use_shader_pack8 && num_input % 8 == 0 ? 8 : num_input % 4 == 0 ? 4 : 1;
    int out_elempack = opt.use_shader_pack8 && num_output % 8 == 0 ? 8 : num_output % 4 == 0 ? 4 : 1;

    size_t elemsize;
    size_t out_elemsize;
    if (opt.use_fp16_storage)
    {
        elemsize = elempack * 2u;
        out_elemsize = out_elempack * 2u;
    }
    else if (opt.use_fp16_packed)
    {
        elemsize = elempack == 1 ? 4u : elempack * 2u;
        out_elemsize = out_elempack == 1 ? 4u : out_elempack * 2u;
    }
    else
    {
        elemsize = elempack * 4u;
        out_elemsize = out_elempack * 4u;
    }

    Mat shape_bordered_packed;
    if (shape_bordered.dims == 3) shape_bordered_packed = Mat(shape_bordered.w, shape_bordered.h, num_input / elempack, (void*)0, elemsize, elempack);

    Mat out_shape_packed;
    if (out_shape.dims == 3) out_shape_packed = Mat(out_shape.w, out_shape.h, num_output / out_elempack, (void*)0, out_elemsize, out_elempack);

    // fc
    if (kernel_w == 1 && kernel_h == 1)
    {
        {
            reshape_1x1xw = ncnn::create_layer(ncnn::LayerType::Reshape);
            reshape_1x1xw->vkdev = vkdev;

            reshape_1x1xw->bottom_shapes.resize(1);
            reshape_1x1xw->bottom_shapes[0] = Mat(num_input, (void*)0);
            reshape_1x1xw->top_shapes.resize(1);
            reshape_1x1xw->top_shapes[0] = Mat(1, 1, num_input, (void*)0);

            ncnn::ParamDict pd;
            pd.set(0, 1);         // w
            pd.set(1, 1);         // h
            pd.set(2, num_input); // c

            reshape_1x1xw->load_param(pd);

            reshape_1x1xw->create_pipeline(opt);
        }

        {
            reshape_w = ncnn::create_layer(ncnn::LayerType::Reshape);
            reshape_w->vkdev = vkdev;

            reshape_w->bottom_shapes.resize(1);
            reshape_w->bottom_shapes[0] = Mat(1, 1, num_output, (void*)0);
            reshape_w->top_shapes.resize(1);
            reshape_w->top_shapes[0] = Mat(num_output, (void*)0);

            ncnn::ParamDict pd;
            pd.set(0, num_output); // w

            reshape_w->load_param(pd);

            reshape_w->create_pipeline(opt);
        }
    }

    bool is_conv1x1s1d1 = kernel_w == 1 && kernel_h == 1 && stride_w == 1 && stride_h == 1 && dilation_w == 1 && dilation_h == 1;
    bool is_conv3x3s1d1 = kernel_w == 3 && kernel_h == 3 && stride_w == 1 && stride_h == 1 && dilation_w == 1 && dilation_h == 1;

    {
        padding = ncnn::create_layer(ncnn::LayerType::Padding);
        padding->vkdev = vkdev;

        padding->bottom_shapes.resize(1);
        padding->bottom_shapes[0] = shape;
        padding->top_shapes.resize(1);
        padding->top_shapes[0] = shape_bordered;

        ncnn::ParamDict pd;
        pd.set(0, pad_top);
        pd.set(1, pad_bottom);
        pd.set(2, pad_left);
        pd.set(3, pad_right);
        pd.set(4, 0);
        pd.set(5, pad_value);

        padding->load_param(pd);

        padding->create_pipeline(opt);
    }

    if (opt.use_winograd_convolution && (opt.use_winograd23_convolution || opt.use_winograd43_convolution) && is_conv3x3s1d1 && num_input >= 16 && num_output >= 16)
    {
        bool use_cooperative_matrix_16_8_8 = vkdev->info.support_cooperative_matrix_16_8_8() && opt.use_cooperative_matrix && !opt.use_image_storage && !opt.use_shader_pack8 && opt.use_fp16_storage && num_input % 8 == 0 && num_output % 8 == 0;
        bool use_cooperative_matrix_16_16_16 = vkdev->info.support_cooperative_matrix_16_16_16() && opt.use_cooperative_matrix && !opt.use_image_storage && !opt.use_shader_pack8 && opt.use_fp16_storage && num_input % 16 == 0 && num_output % 16 == 0;

        // winograd43 transform kernel
        if (opt.use_winograd43_convolution)
        {
            Mat weight_data_tm;
            weight_data_tm.create(6 * 6, num_input, num_output);

            const float sq2 = 1.41421356237f;
            const float ktm[6][3] = {
                {1.0f, 0.0f, 0.0f},
                {-2.0f / 3, -sq2 / 3, -1.0f / 3},
                {-2.0f / 3, sq2 / 3, -1.0f / 3},
                {1.0f / 6, sq2 / 6, 1.0f / 3},
                {1.0f / 6, -sq2 / 6, 1.0f / 3},
                {0.0f, 0.0f, 1.0f}
            };

            #pragma omp parallel for num_threads(opt.num_threads)
            for (int p = 0; p < num_output; p++)
            {
                for (int q = 0; q < num_input; q++)
                {
                    const float* kernel0 = (const float*)weight_data + p * num_input * 9 + q * 9;
                    float* kernel_tm0 = weight_data_tm.channel(p).row(q);

                    // transform kernel
                    const float* k0 = kernel0;
                    const float* k1 = kernel0 + 3;
                    const float* k2 = kernel0 + 6;

                    // h
                    float tmp[6][3];
                    for (int i = 0; i < 6; i++)
                    {
                        tmp[i][0] = k0[0] * ktm[i][0] + k0[1] * ktm[i][1] + k0[2] * ktm[i][2];
                        tmp[i][1] = k1[0] * ktm[i][0] + k1[1] * ktm[i][1] + k1[2] * ktm[i][2];
                        tmp[i][2] = k2[0] * ktm[i][0] + k2[1] * ktm[i][1] + k2[2] * ktm[i][2];
                    }

                    // U
                    for (int j = 0; j < 6; j++)
                    {
                        float* tmpp = &tmp[j][0];

                        for (int i = 0; i < 6; i++)
                        {
                            kernel_tm0[j * 6 + i] = tmpp[0] * ktm[i][0] + tmpp[1] * ktm[i][1] + tmpp[2] * ktm[i][2];
                        }
                    }
                }
            }

            if (use_cooperative_matrix_16_8_8)
            {
                // src = 36-inch-outch
                // dst = 8b-8a-inch/8a-outch/8b-36
                weight_winograd43_data_packed.create(num_input / 8, num_output / 8, 36, (size_t)4 * 8 * 8, 8 * 8);

                for (int k = 0; k < 36; k++)
                {
                    float* g00 = weight_winograd43_data_packed.channel(k);

                    for (int q = 0; q + (8 - 1) < num_output; q += 8)
                    {
                        for (int p = 0; p + (8 - 1) < num_input; p += 8)
                        {
                            for (int i = 0; i < 8; i++)
                            {
                                for (int j = 0; j < 8; j++)
                                {
                                    const float* k00 = weight_data_tm.channel(q + j).row(p + i);
                                    g00[0] = k00[k];
                                    g00++;
                                }
                            }
                        }
                    }
                }
            }
            else if (use_cooperative_matrix_16_16_16)
            {
                // src = 36-inch-outch
                // dst = 16b-16a-inch/16a-outch/16b-36
                weight_winograd43_data_packed.create(num_input / 16, num_output / 16, 36, (size_t)4 * 16 * 16, 16 * 16);

                for (int k = 0; k < 36; k++)
                {
                    float* g00 = weight_winograd43_data_packed.channel(k);

                    for (int q = 0; q + (16 - 1) < num_output; q += 16)
                    {
                        for (int p = 0; p + (16 - 1) < num_input; p += 16)
                        {
                            for (int i = 0; i < 16; i++)
                            {
                                for (int j = 0; j < 16; j++)
                                {
                                    const float* k00 = weight_data_tm.channel(q + j).row(p + i);
                                    g00[0] = k00[k];
                                    g00++;
                                }
                            }
                        }
                    }
                }
            }
            else
            {
                // src = 36-inch-outch
                // dst = 8a-8b-inch/8a-outch/8b-36
                weight_winograd43_data_packed.create(num_input / elempack, num_output / out_elempack, 36, (size_t)4 * elempack * out_elempack, elempack * out_elempack);

                for (int k = 0; k < 36; k++)
                {
                    float* g00 = weight_winograd43_data_packed.channel(k);

                    for (int q = 0; q + (out_elempack - 1) < num_output; q += out_elempack)
                    {
                        for (int p = 0; p + (elempack - 1) < num_input; p += elempack)
                        {
                            for (int i = 0; i < out_elempack; i++)
                            {
                                const Mat k0 = weight_data_tm.channel(q + i);

                                for (int j = 0; j < elempack; j++)
                                {
                                    const float* k00 = k0.row(p + j);
                                    g00[0] = k00[k];
                                    g00++;
                                }
                            }
                        }
                    }
                }
            }
        }

        // winograd43
        if (opt.use_winograd43_convolution)
        {
            int block_x = 0;
            int block_y = 0;
            Mat shape_winograd_input_transformed;
            Mat shape_winograd_gemm;
            Mat shape_winograd_input_transformed_packed;
            Mat shape_winograd_gemm_packed;

            if (out_shape.dims != 0)
            {
                int block_x = (out_shape.w + 3) / 4;
                int block_y = (out_shape.h + 3) / 4;

                shape_winograd_input_transformed = Mat(block_x * block_y, shape.c, 36, (void*)0);
                shape_winograd_gemm = Mat(block_x * block_y, out_shape.c, 36, (void*)0);
            }

            if (shape_winograd_input_transformed.dims == 3) shape_winograd_input_transformed_packed = Mat(shape_winograd_input_transformed.w, shape_winograd_input_transformed.h / elempack, 36, (void*)0, elemsize, elempack);

            if (shape_winograd_gemm.dims == 3) shape_winograd_gemm_packed = Mat(shape_winograd_gemm.w, shape_winograd_gemm.h / out_elempack, 36, (void*)0, out_elemsize, out_elempack);

            // check blob shape
            if (!vkdev->shape_support_image_storage(shape_winograd_input_transformed_packed) || !vkdev->shape_support_image_storage(shape_winograd_gemm_packed))
            {
                support_image_storage = false;
                opt.use_image_storage = false;
            }

            Mat weight_data_packed_tm(num_input / elempack, num_output / out_elempack, 36, (size_t)4 * elempack * out_elempack, elempack * out_elempack);
            if (!vkdev->shape_support_image_storage(weight_data_packed_tm))
            {
                support_image_storage = false;
                opt.use_image_storage = false;
            }

            if (vkdev->info.vendor_id() == 0x5143 && vkdev->info.api_version() < VK_MAKE_VERSION(1, 0, 66))
            {
                // FIXME workaround qcom adreno image shader produce wrong result on old drivers
                support_image_storage = false;
                opt.use_image_storage = false;
            }

            {
                std::vector<vk_specialization_type> specializations(0 + 7);
                specializations[0 + 0].i = shape_bordered_packed.w;
                specializations[0 + 1].i = shape_bordered_packed.h;
                specializations[0 + 2].i = shape_bordered_packed.c;
                specializations[0 + 3].i = shape_bordered_packed.cstep;
                specializations[0 + 4].i = shape_winograd_input_transformed_packed.cstep;
                specializations[0 + 5].i = block_x;
                specializations[0 + 6].i = block_y;

                int shader_type_index = -1;
                if (elempack == 1) shader_type_index = LayerShaderType::convolution_3x3s1d1_winograd43_transform_input;
                if (elempack == 4) shader_type_index = LayerShaderType::convolution_pack4_3x3s1d1_winograd43_transform_input;
                if (elempack == 8) shader_type_index = LayerShaderType::convolution_pack8_3x3s1d1_winograd43_transform_input;

                pipeline_convolution_3x3s1d1_winograd43_transform_input = new Pipeline(vkdev);
                pipeline_convolution_3x3s1d1_winograd43_transform_input->set_local_size_xyz(4, 4, 1);
                pipeline_convolution_3x3s1d1_winograd43_transform_input->create(shader_type_index, opt, specializations);
            }

            {
                std::vector<vk_specialization_type> specializations(1 + 5);
                specializations[0].i = 36;
                specializations[1 + 0].i = shape_winograd_input_transformed_packed.h;
                specializations[1 + 1].i = shape_winograd_input_transformed_packed.cstep;
                specializations[1 + 2].i = shape_winograd_gemm_packed.w;
                specializations[1 + 3].i = shape_winograd_gemm_packed.h;
                specializations[1 + 4].i = shape_winograd_gemm_packed.cstep;

                int shader_type_index = -1;
                if (elempack == 1 && out_elempack == 1) shader_type_index = LayerShaderType::convolution_3x3s1d1_winograd_gemm;
                if (elempack == 4 && out_elempack == 4) shader_type_index = LayerShaderType::convolution_pack4_3x3s1d1_winograd_gemm;
                if (elempack == 1 && out_elempack == 4) shader_type_index = LayerShaderType::convolution_pack1to4_3x3s1d1_winograd_gemm;
                if (elempack == 4 && out_elempack == 1) shader_type_index = LayerShaderType::convolution_pack4to1_3x3s1d1_winograd_gemm;
                if (elempack == 8 && out_elempack == 8) shader_type_index = LayerShaderType::convolution_pack8_3x3s1d1_winograd_gemm;
                if (elempack == 1 && out_elempack == 8) shader_type_index = LayerShaderType::convolution_pack1to8_3x3s1d1_winograd_gemm;
                if (elempack == 8 && out_elempack == 1) shader_type_index = LayerShaderType::convolution_pack8to1_3x3s1d1_winograd_gemm;
                if (elempack == 4 && out_elempack == 8) shader_type_index = LayerShaderType::convolution_pack4to8_3x3s1d1_winograd_gemm;
                if (elempack == 8 && out_elempack == 4) shader_type_index = LayerShaderType::convolution_pack8to4_3x3s1d1_winograd_gemm;

                if (use_cooperative_matrix_16_8_8)
                {
                    if (vkdev->info.support_VK_KHR_cooperative_matrix())
                        shader_type_index = LayerShaderType::convolution_pack4_3x3s1d1_winograd_gemm_khr_cm_16_8_8;
                    else
                        shader_type_index = LayerShaderType::convolution_pack4_3x3s1d1_winograd_gemm_nv_cm_16_8_8;
                }
                else if (use_cooperative_matrix_16_16_16)
                {
                    if (vkdev->info.support_VK_KHR_cooperative_matrix())
                        shader_type_index = LayerShaderType::convolution_pack4_3x3s1d1_winograd_gemm_khr_cm_16_16_16;
                    else
                        shader_type_index = LayerShaderType::convolution_pack4_3x3s1d1_winograd_gemm_nv_cm_16_16_16;
                }

                pipeline_convolution_3x3s1d1_winograd43_gemm = new Pipeline(vkdev);
                if (use_cooperative_matrix_16_8_8)
                {
                    pipeline_convolution_3x3s1d1_winograd43_gemm->set_local_size_xyz(32, 1, 1);
                }
                else if (use_cooperative_matrix_16_16_16)
                {
                    pipeline_convolution_3x3s1d1_winograd43_gemm->set_local_size_xyz(32, 1, 1);
                }
                else if (opt.use_shader_local_memory)
                {
                    pipeline_convolution_3x3s1d1_winograd43_gemm->set_local_size_xyz(8, 8, 1);
                }
                else
                {
                    pipeline_convolution_3x3s1d1_winograd43_gemm->set_local_size_xyz(4, std::min(4, num_output / out_elempack), 4);
                }
                pipeline_convolution_3x3s1d1_winograd43_gemm->create(shader_type_index, opt, specializations);
            }

            {
                std::vector<vk_specialization_type> specializations(4 + 7);
                specializations[0].i = bias_term;
                specializations[1].i = activation_type;
                specializations[2].f = activation_params.w >= 1 ? activation_params[0] : 0.f;
                specializations[3].f = activation_params.w == 2 ? activation_params[1] : 0.f;
                specializations[4 + 0].i = shape_winograd_gemm_packed.h;
                specializations[4 + 1].i = shape_winograd_gemm_packed.cstep;
                specializations[4 + 2].i = block_x;
                specializations[4 + 3].i = block_y;
                specializations[4 + 4].i = out_shape_packed.w;
                specializations[4 + 5].i = out_shape_packed.h;
                specializations[4 + 6].i = out_shape_packed.cstep;

                int shader_type_index = -1;
                if (out_elempack == 1) shader_type_index = LayerShaderType::convolution_3x3s1d1_winograd43_transform_output;
                if (out_elempack == 4) shader_type_index = LayerShaderType::convolution_pack4_3x3s1d1_winograd43_transform_output;
                if (out_elempack == 8) shader_type_index = LayerShaderType::convolution_pack8_3x3s1d1_winograd43_transform_output;

                pipeline_convolution_3x3s1d1_winograd43_transform_output = new Pipeline(vkdev);
                pipeline_convolution_3x3s1d1_winograd43_transform_output->set_local_size_xyz(4, 4, 1);
                pipeline_convolution_3x3s1d1_winograd43_transform_output->create(shader_type_index, opt, specializations);
            }
        }

        // winograd23 transform kernel
        if (opt.use_winograd23_convolution)
        {
            Mat weight_data_tm;
            weight_data_tm.create(4 * 4, num_input, num_output);

            // G
            const float ktm[4][3] = {
                {1.0f, 0.0f, 0.0f},
                {1.0f / 2, 1.0f / 2, 1.0f / 2},
                {1.0f / 2, -1.0f / 2, 1.0f / 2},
                {0.0f, 0.0f, 1.0f}
            };

            #pragma omp parallel for num_threads(opt.num_threads)
            for (int p = 0; p < num_output; p++)
            {
                for (int q = 0; q < num_input; q++)
                {
                    const float* kernel0 = (const float*)weight_data + p * num_input * 9 + q * 9;
                    float* kernel_tm0 = weight_data_tm.channel(p).row(q);

                    // transform kernel
                    const float* k0 = kernel0;
                    const float* k1 = kernel0 + 3;
                    const float* k2 = kernel0 + 6;

                    // h
                    float tmp[4][3];
                    for (int i = 0; i < 4; i++)
                    {
                        tmp[i][0] = k0[0] * ktm[i][0] + k0[1] * ktm[i][1] + k0[2] * ktm[i][2];
                        tmp[i][1] = k1[0] * ktm[i][0] + k1[1] * ktm[i][1] + k1[2] * ktm[i][2];
                        tmp[i][2] = k2[0] * ktm[i][0] + k2[1] * ktm[i][1] + k2[2] * ktm[i][2];
                    }

                    // U
                    for (int j = 0; j < 4; j++)
                    {
                        float* tmpp = &tmp[j][0];

                        for (int i = 0; i < 4; i++)
                        {
                            kernel_tm0[j * 4 + i] = tmpp[0] * ktm[i][0] + tmpp[1] * ktm[i][1] + tmpp[2] * ktm[i][2];
                        }
                    }
                }
            }

            if (use_cooperative_matrix_16_8_8)
            {
                // src = 16-inch-outch
                // dst = 8b-8a-inch/8a-outch/8b-16
                weight_winograd23_data_packed.create(num_input / 8, num_output / 8, 16, (size_t)4 * 8 * 8, 8 * 8);

                for (int k = 0; k < 16; k++)
                {
                    float* g00 = weight_winograd23_data_packed.channel(k);

                    for (int q = 0; q + (8 - 1) < num_output; q += 8)
                    {
                        for (int p = 0; p + (8 - 1) < num_input; p += 8)
                        {
                            for (int i = 0; i < 8; i++)
                            {
                                for (int j = 0; j < 8; j++)
                                {
                                    const float* k00 = weight_data_tm.channel(q + j).row(p + i);
                                    g00[0] = k00[k];
                                    g00++;
                                }
                            }
                        }
                    }
                }
            }
            else if (use_cooperative_matrix_16_16_16)
            {
                // src = 16-inch-outch
                // dst = 16b-16a-inch/16a-outch/16b-16
                weight_winograd23_data_packed.create(num_input / 16, num_output / 16, 16, (size_t)4 * 16 * 16, 16 * 16);

                for (int k = 0; k < 16; k++)
                {
                    float* g00 = weight_winograd23_data_packed.channel(k);

                    for (int q = 0; q + (16 - 1) < num_output; q += 16)
                    {
                        for (int p = 0; p + (16 - 1) < num_input; p += 16)
                        {
                            for (int i = 0; i < 16; i++)
                            {
                                for (int j = 0; j < 16; j++)
                                {
                                    const float* k00 = weight_data_tm.channel(q + j).row(p + i);
                                    g00[0] = k00[k];
                                    g00++;
                                }
                            }
                        }
                    }
                }
            }
            else
            {
                // src = 16-inch-outch
                // dst = 8a-8b-inch/8a-outch/8b-16
                weight_winograd23_data_packed.create(num_input / elempack, num_output / out_elempack, 16, (size_t)4 * elempack * out_elempack, elempack * out_elempack);

                for (int k = 0; k < 16; k++)
                {
                    float* g00 = weight_winograd23_data_packed.channel(k);

                    for (int q = 0; q + (out_elempack - 1) < num_output; q += out_elempack)
                    {
                        for (int p = 0; p + (elempack - 1) < num_input; p += elempack)
                        {
                            for (int i = 0; i < out_elempack; i++)
                            {
                                const Mat k0 = weight_data_tm.channel(q + i);

                                for (int j = 0; j < elempack; j++)
                                {
                                    const float* k00 = k0.row(p + j);
                                    g00[0] = k00[k];
                                    g00++;
                                }
                            }
                        }
                    }
                }
            }
        }

        // winograd23
        if (opt.use_winograd23_convolution)
        {
            int block_x = 0;
            int block_y = 0;
            Mat shape_winograd_input_transformed;
            Mat shape_winograd_gemm;
            Mat shape_winograd_input_transformed_packed;
            Mat shape_winograd_gemm_packed;

            if (out_shape.dims != 0)
            {
                int block_x = (out_shape.w + 1) / 2;
                int block_y = (out_shape.h + 1) / 2;

                shape_winograd_input_transformed = Mat(block_x * block_y, shape.c, 16, (void*)0);
                shape_winograd_gemm = Mat(block_x * block_y, out_shape.c, 16, (void*)0);
            }

            if (shape_winograd_input_transformed.dims == 3) shape_winograd_input_transformed_packed = Mat(shape_winograd_input_transformed.w, shape_winograd_input_transformed.h / elempack, 16, (void*)0, elemsize, elempack);

            if (shape_winograd_gemm.dims == 3) shape_winograd_gemm_packed = Mat(shape_winograd_gemm.w, shape_winograd_gemm.h / out_elempack, 16, (void*)0, out_elemsize, out_elempack);

            // check blob shape
            if (!vkdev->shape_support_image_storage(shape_winograd_input_transformed_packed) || !vkdev->shape_support_image_storage(shape_winograd_gemm_packed))
            {
                support_image_storage = false;
                opt.use_image_storage = false;
            }

            Mat weight_data_packed_tm(num_input / elempack, num_output / out_elempack, 16, (size_t)4 * elempack * out_elempack, elempack * out_elempack);
            if (!vkdev->shape_support_image_storage(weight_data_packed_tm))
            {
                support_image_storage = false;
                opt.use_image_storage = false;
            }

            if (vkdev->info.vendor_id() == 0x5143 && vkdev->info.api_version() < VK_MAKE_VERSION(1, 0, 66))
            {
                // FIXME workaround qcom adreno image shader produce wrong result on old drivers
                support_image_storage = false;
                opt.use_image_storage = false;
            }

            {
                std::vector<vk_specialization_type> specializations(0 + 7);
                specializations[0 + 0].i = shape_bordered_packed.w;
                specializations[0 + 1].i = shape_bordered_packed.h;
                specializations[0 + 2].i = shape_bordered_packed.c;
                specializations[0 + 3].i = shape_bordered_packed.cstep;
                specializations[0 + 4].i = shape_winograd_input_transformed_packed.cstep;
                specializations[0 + 5].i = block_x;
                specializations[0 + 6].i = block_y;

                int shader_type_index = -1;
                if (elempack == 1) shader_type_index = LayerShaderType::convolution_3x3s1d1_winograd23_transform_input;
                if (elempack == 4) shader_type_index = LayerShaderType::convolution_pack4_3x3s1d1_winograd23_transform_input;
                if (elempack == 8) shader_type_index = LayerShaderType::convolution_pack8_3x3s1d1_winograd23_transform_input;

                pipeline_convolution_3x3s1d1_winograd23_transform_input = new Pipeline(vkdev);
                pipeline_convolution_3x3s1d1_winograd23_transform_input->set_local_size_xyz(8, 8, 1);
                pipeline_convolution_3x3s1d1_winograd23_transform_input->create(shader_type_index, opt, specializations);
            }

            {
                std::vector<vk_specialization_type> specializations(1 + 5);
                specializations[0].i = 16;
                specializations[1 + 0].i = shape_winograd_input_transformed_packed.h;
                specializations[1 + 1].i = shape_winograd_input_transformed_packed.cstep;
                specializations[1 + 2].i = shape_winograd_gemm_packed.w;
                specializations[1 + 3].i = shape_winograd_gemm_packed.h;
                specializations[1 + 4].i = shape_winograd_gemm_packed.cstep;

                int shader_type_index = -1;
                if (elempack == 1 && out_elempack == 1) shader_type_index = LayerShaderType::convolution_3x3s1d1_winograd_gemm;
                if (elempack == 4 && out_elempack == 4) shader_type_index = LayerShaderType::convolution_pack4_3x3s1d1_winograd_gemm;
                if (elempack == 1 && out_elempack == 4) shader_type_index = LayerShaderType::convolution_pack1to4_3x3s1d1_winograd_gemm;
                if (elempack == 4 && out_elempack == 1) shader_type_index = LayerShaderType::convolution_pack4to1_3x3s1d1_winograd_gemm;
                if (elempack == 8 && out_elempack == 8) shader_type_index = LayerShaderType::convolution_pack8_3x3s1d1_winograd_gemm;
                if (elempack == 1 && out_elempack == 8) shader_type_index = LayerShaderType::convolution_pack1to8_3x3s1d1_winograd_gemm;
                if (elempack == 8 && out_elempack == 1) shader_type_index = LayerShaderType::convolution_pack8to1_3x3s1d1_winograd_gemm;
                if (elempack == 4 && out_elempack == 8) shader_type_index = LayerShaderType::convolution_pack4to8_3x3s1d1_winograd_gemm;
                if (elempack == 8 && out_elempack == 4) shader_type_index = LayerShaderType::convolution_pack8to4_3x3s1d1_winograd_gemm;

                if (use_cooperative_matrix_16_8_8)
                {
                    if (vkdev->info.support_VK_KHR_cooperative_matrix())
                        shader_type_index = LayerShaderType::convolution_pack4_3x3s1d1_winograd_gemm_khr_cm_16_8_8;
                    else
                        shader_type_index = LayerShaderType::convolution_pack4_3x3s1d1_winograd_gemm_nv_cm_16_8_8;
                }
                else if (use_cooperative_matrix_16_16_16)
                {
                    if (vkdev->info.support_VK_KHR_cooperative_matrix())
                        shader_type_index = LayerShaderType::convolution_pack4_3x3s1d1_winograd_gemm_khr_cm_16_16_16;
                    else
                        shader_type_index = LayerShaderType::convolution_pack4_3x3s1d1_winograd_gemm_nv_cm_16_16_16;
                }

                pipeline_convolution_3x3s1d1_winograd23_gemm = new Pipeline(vkdev);
                if (use_cooperative_matrix_16_8_8)
                {
                    pipeline_convolution_3x3s1d1_winograd23_gemm->set_local_size_xyz(32, 1, 1);
                }
                else if (use_cooperative_matrix_16_16_16)
                {
                    pipeline_convolution_3x3s1d1_winograd23_gemm->set_local_size_xyz(32, 1, 1);
                }
                else if (opt.use_shader_local_memory)
                {
                    pipeline_convolution_3x3s1d1_winograd23_gemm->set_local_size_xyz(8, 8, 1);
                }
                else
                {
                    pipeline_convolution_3x3s1d1_winograd23_gemm->set_local_size_xyz(4, std::min(4, num_output / out_elempack), 4);
                }
                pipeline_convolution_3x3s1d1_winograd23_gemm->create(shader_type_index, opt, specializations);
            }

            {
                std::vector<vk_specialization_type> specializations(4 + 7);
                specializations[0].i = bias_term;
                specializations[1].i = activation_type;
                specializations[2].f = activation_params.w >= 1 ? activation_params[0] : 0.f;
                specializations[3].f = activation_params.w == 2 ? activation_params[1] : 0.f;
                specializations[4 + 0].i = shape_winograd_gemm_packed.h;
                specializations[4 + 1].i = shape_winograd_gemm_packed.cstep;
                specializations[4 + 2].i = block_x;
                specializations[4 + 3].i = block_y;
                specializations[4 + 4].i = out_shape_packed.w;
                specializations[4 + 5].i = out_shape_packed.h;
                specializations[4 + 6].i = out_shape_packed.cstep;

                int shader_type_index = -1;
                if (out_elempack == 1) shader_type_index = LayerShaderType::convolution_3x3s1d1_winograd23_transform_output;
                if (out_elempack == 4) shader_type_index = LayerShaderType::convolution_pack4_3x3s1d1_winograd23_transform_output;
                if (out_elempack == 8) shader_type_index = LayerShaderType::convolution_pack8_3x3s1d1_winograd23_transform_output;

                pipeline_convolution_3x3s1d1_winograd23_transform_output = new Pipeline(vkdev);
                pipeline_convolution_3x3s1d1_winograd23_transform_output->set_local_size_xyz(8, 8, 1);
                pipeline_convolution_3x3s1d1_winograd23_transform_output->create(shader_type_index, opt, specializations);
            }
        }
    }
    else
    {
        // src = kw-kh-inch-outch
        // dst = pa-pb-kw-kh-inch/pa-outch/pb
        if (opt.use_sgemm_convolution && !is_conv1x1s1d1 && num_input >= 16 && num_output >= 16)
        {
            bool use_cooperative_matrix_16_8_8 = vkdev->info.support_cooperative_matrix_16_8_8() && opt.use_cooperative_matrix && !opt.use_image_storage && !opt.use_shader_pack8 && opt.use_fp16_storage && num_input % 8 == 0 && num_output % 8 == 0;
            bool use_cooperative_matrix_16_16_16 = vkdev->info.support_cooperative_matrix_16_16_16() && opt.use_cooperative_matrix && !opt.use_image_storage && !opt.use_shader_pack8 && opt.use_fp16_storage && num_input % 16 == 0 && num_output % 16 == 0;

            if (use_cooperative_matrix_16_8_8)
            {
                // dst = 8b-8a-maxk-inch/8a-outch/8b
                Mat weight_data_r2 = weight_data.reshape(maxk, num_input, num_output);

                weight_data_packed.create(maxk * num_input / 8, num_output / 8, (size_t)4 * 8 * 8, 8 * 8);

                for (int q = 0; q + 7 < num_output; q += 8)
                {
                    float* g00 = weight_data_packed.row(q / 8);

                    for (int p = 0; p + 7 < num_input; p += 8)
                    {
                        for (int k = 0; k < maxk; k++)
                        {
                            for (int i = 0; i < 8; i++)
                            {
                                for (int j = 0; j < 8; j++)
                                {
                                    const float* k00 = weight_data_r2.channel(q + j).row(p + i);
                                    g00[0] = k00[k];
                                    g00++;
                                }
                            }
                        }
                    }
                }
            }
            else if (use_cooperative_matrix_16_16_16)
            {
                // dst = 16b-16a-maxk-inch/16a-outch/16b
                Mat weight_data_r2 = weight_data.reshape(maxk, num_input, num_output);

                weight_data_packed.create(maxk * num_input / 16, num_output / 16, (size_t)4 * 16 * 16, 16 * 16);

                for (int q = 0; q + 15 < num_output; q += 16)
                {
                    float* g00 = weight_data_packed.row(q / 16);

                    for (int p = 0; p + 15 < num_input; p += 16)
                    {
                        for (int k = 0; k < maxk; k++)
                        {
                            for (int i = 0; i < 16; i++)
                            {
                                for (int j = 0; j < 16; j++)
                                {
                                    const float* k00 = weight_data_r2.channel(q + j).row(p + i);
                                    g00[0] = k00[k];
                                    g00++;
                                }
                            }
                        }
                    }
                }
            }
            else
            {
                Mat weight_data_r2 = weight_data.reshape(maxk, num_input, num_output);

                weight_data_packed.create(maxk * num_input / elempack, num_output / out_elempack, (size_t)4 * elempack * out_elempack, elempack * out_elempack);

                for (int q = 0; q + (out_elempack - 1) < num_output; q += out_elempack)
                {
                    float* g00 = weight_data_packed.row(q / out_elempack);

                    for (int p = 0; p + (elempack - 1) < num_input; p += elempack)
                    {
                        for (int k = 0; k < maxk; k++)
                        {
                            for (int i = 0; i < out_elempack; i++)
                            {
                                const Mat k0 = weight_data_r2.channel(q + i);

                                for (int j = 0; j < elempack; j++)
                                {
                                    const float* k00 = k0.row(p + j);
                                    g00[0] = k00[k];
                                    g00++;
                                }
                            }
                        }
                    }
                }
            }
        }
        else
        {
            bool use_cooperative_matrix_16_8_8 = vkdev->info.support_cooperative_matrix_16_8_8() && opt.use_cooperative_matrix && is_conv1x1s1d1 && !opt.use_image_storage && !opt.use_shader_pack8 && opt.use_fp16_storage && num_input % 8 == 0 && num_output % 8 == 0;
            bool use_cooperative_matrix_16_16_16 = vkdev->info.support_cooperative_matrix_16_16_16() && opt.use_cooperative_matrix && is_conv1x1s1d1 && !opt.use_image_storage && !opt.use_shader_pack8 && opt.use_fp16_storage && num_input % 16 == 0 && num_output % 16 == 0;

            if (use_cooperative_matrix_16_8_8)
            {
                // dst = 8b-8a-inch/8a-outch/8b
                Mat weight_data_r2 = weight_data.reshape(maxk, num_input, num_output);

                weight_data_packed.create(maxk, num_input / 8, num_output / 8, (size_t)4 * 8 * 8, 8 * 8);

                for (int q = 0; q + 7 < num_output; q += 8)
                {
                    float* g00 = weight_data_packed.channel(q / 8);

                    for (int p = 0; p + 7 < num_input; p += 8)
                    {
                        for (int k = 0; k < maxk; k++)
                        {
                            for (int i = 0; i < 8; i++)
                            {
                                for (int j = 0; j < 8; j++)
                                {
                                    const float* k00 = weight_data_r2.channel(q + j).row(p + i);
                                    g00[0] = k00[k];
                                    g00++;
                                }
                            }
                        }
                    }
                }
            }
            else if (use_cooperative_matrix_16_16_16)
            {
                // dst = 16b-16a-inch/16a-outch/16b
                Mat weight_data_r2 = weight_data.reshape(maxk, num_input, num_output);

                weight_data_packed.create(maxk, num_input / 16, num_output / 16, (size_t)4 * 16 * 16, 16 * 16);

                for (int q = 0; q + 15 < num_output; q += 16)
                {
                    float* g00 = weight_data_packed.channel(q / 16);

                    for (int p = 0; p + 15 < num_input; p += 16)
                    {
                        for (int k = 0; k < maxk; k++)
                        {
                            for (int i = 0; i < 16; i++)
                            {
                                for (int j = 0; j < 16; j++)
                                {
                                    const float* k00 = weight_data_r2.channel(q + j).row(p + i);
                                    g00[0] = k00[k];
                                    g00++;
                                }
                            }
                        }
                    }
                }
            }
            else
            {
                Mat weight_data_r2 = weight_data.reshape(maxk, num_input, num_output);

                weight_data_packed.create(maxk, num_input / elempack, num_output / out_elempack, (size_t)4 * elempack * out_elempack, elempack * out_elempack);

                for (int q = 0; q + (out_elempack - 1) < num_output; q += out_elempack)
                {
                    float* g00 = weight_data_packed.channel(q / out_elempack);

                    for (int p = 0; p + (elempack - 1) < num_input; p += elempack)
                    {
                        for (int k = 0; k < maxk; k++)
                        {
                            for (int i = 0; i < out_elempack; i++)
                            {
                                const Mat k0 = weight_data_r2.channel(q + i);

                                for (int j = 0; j < elempack; j++)
                                {
                                    const float* k00 = k0.row(p + j);
                                    g00[0] = k00[k];
                                    g00++;
                                }
                            }
                        }
                    }
                }
            }
        }
    }

    if (bias_term)
    {
        convert_packing(bias_data, bias_data_packed, out_elempack, opt);
    }

    if (opt.use_winograd_convolution && (opt.use_winograd23_convolution || opt.use_winograd43_convolution) && is_conv3x3s1d1 && num_input >= 16 && num_output >= 16)
    {
        // pass
    }
    else if (opt.use_sgemm_convolution && !is_conv1x1s1d1 && num_input >= 16 && num_output >= 16)
    {
        bool use_cooperative_matrix_16_8_8 = vkdev->info.support_cooperative_matrix_16_8_8() && opt.use_cooperative_matrix && !opt.use_image_storage && !opt.use_shader_pack8 && opt.use_fp16_storage && num_input % 8 == 0 && num_output % 8 == 0;
        bool use_cooperative_matrix_16_16_16 = vkdev->info.support_cooperative_matrix_16_16_16() && opt.use_cooperative_matrix && !opt.use_image_storage && !opt.use_shader_pack8 && opt.use_fp16_storage && num_input % 16 == 0 && num_output % 16 == 0;

        // check blob shape
        if (!vkdev->shape_support_image_storage(shape_bordered_packed) || !vkdev->shape_support_image_storage(out_shape_packed))
        {
            support_image_storage = false;
            opt.use_image_storage = false;
        }

        // check weight shape
        Mat weight_data_packed_shape(maxk, num_input / elempack, num_output / out_elempack, (void*)0, (size_t)4 * elempack * out_elempack, elempack * out_elempack);
        if (!vkdev->shape_support_image_storage(weight_data_packed_shape))
        {
            support_image_storage = false;
            opt.use_image_storage = false;
        }

        std::vector<vk_specialization_type> specializations(10 + 8);
        specializations[0].i = kernel_w;
        specializations[1].i = kernel_h;
        specializations[2].i = dilation_w;
        specializations[3].i = dilation_h;
        specializations[4].i = stride_w;
        specializations[5].i = stride_h;
        specializations[6].i = bias_term;
        specializations[7].i = activation_type;
        specializations[8].f = activation_params.w >= 1 ? activation_params[0] : 0.f;
        specializations[9].f = activation_params.w == 2 ? activation_params[1] : 0.f;
        specializations[10 + 0].i = shape_bordered_packed.w;
        specializations[10 + 1].i = shape_bordered_packed.h;
        specializations[10 + 2].i = shape_bordered_packed.c;
        specializations[10 + 3].i = shape_bordered_packed.cstep;
        specializations[10 + 4].i = out_shape_packed.w;
        specializations[10 + 5].i = out_shape_packed.h;
        specializations[10 + 6].i = out_shape_packed.c;
        specializations[10 + 7].i = out_shape_packed.cstep;

        Mat local_size_xyz(16, std::min(4, num_output / out_elempack), 1, (void*)0);
        if (out_shape_packed.dims != 0)
        {
            local_size_xyz.w = std::min(16, out_shape_packed.w * out_shape_packed.h);
            local_size_xyz.h = std::min(4, out_shape_packed.c);
        }

        int shader_type_index = -1;
        if (elempack == 1 && out_elempack == 1) shader_type_index = LayerShaderType::convolution_gemm;
        if (elempack == 4 && out_elempack == 4) shader_type_index = LayerShaderType::convolution_pack4_gemm;
        if (elempack == 1 && out_elempack == 4) shader_type_index = LayerShaderType::convolution_pack1to4_gemm;
        if (elempack == 4 && out_elempack == 1) shader_type_index = LayerShaderType::convolution_pack4to1_gemm;
        if (elempack == 8 && out_elempack == 8) shader_type_index = LayerShaderType::convolution_pack8_gemm;
        if (elempack == 1 && out_elempack == 8) shader_type_index = LayerShaderType::convolution_pack1to8_gemm;
        if (elempack == 8 && out_elempack == 1) shader_type_index = LayerShaderType::convolution_pack8to1_gemm;
        if (elempack == 4 && out_elempack == 8) shader_type_index = LayerShaderType::convolution_pack4to8_gemm;
        if (elempack == 8 && out_elempack == 4) shader_type_index = LayerShaderType::convolution_pack8to4_gemm;

        if (use_cooperative_matrix_16_8_8)
        {
            if (vkdev->info.support_VK_KHR_cooperative_matrix())
                shader_type_index = LayerShaderType::convolution_pack4_gemm_khr_cm_16_8_8;
            else
                shader_type_index = LayerShaderType::convolution_pack4_gemm_nv_cm_16_8_8;
        }
        else if (use_cooperative_matrix_16_16_16)
        {
            if (vkdev->info.support_VK_KHR_cooperative_matrix())
                shader_type_index = LayerShaderType::convolution_pack4_gemm_khr_cm_16_16_16;
            else
                shader_type_index = LayerShaderType::convolution_pack4_gemm_nv_cm_16_16_16;
        }

        pipeline_convolution_gemm = new Pipeline(vkdev);
        if (use_cooperative_matrix_16_8_8)
        {
            pipeline_convolution_gemm->set_local_size_xyz(32, 1, 1); // 16_8_8
        }
        else if (use_cooperative_matrix_16_16_16)
        {
            pipeline_convolution_gemm->set_local_size_xyz(32, 1, 1); // 16_16_16
        }
        else if (opt.use_shader_local_memory)
        {
            pipeline_convolution_gemm->set_local_size_xyz(8, 8, 1);
        }
        else
        {
            pipeline_convolution_gemm->set_optimal_local_size_xyz(local_size_xyz);
        }
        pipeline_convolution_gemm->create(shader_type_index, opt, specializations);
    }
    else if (is_conv1x1s1d1)
    {
        bool use_cooperative_matrix_16_8_8 = vkdev->info.support_cooperative_matrix_16_8_8() && opt.use_cooperative_matrix && !opt.use_image_storage && !opt.use_shader_pack8 && opt.use_fp16_storage && num_input % 8 == 0 && num_output % 8 == 0;
        bool use_cooperative_matrix_16_16_16 = vkdev->info.support_cooperative_matrix_16_16_16() && opt.use_cooperative_matrix && !opt.use_image_storage && !opt.use_shader_pack8 && opt.use_fp16_storage && num_input % 16 == 0 && num_output % 16 == 0;

        std::vector<vk_specialization_type> specializations(4 + 8);
        specializations[0].i = bias_term;
        specializations[1].i = activation_type;
        specializations[2].f = activation_params.w >= 1 ? activation_params[0] : 0.f;
        specializations[3].f = activation_params.w == 2 ? activation_params[1] : 0.f;
        specializations[4 + 0].i = shape_bordered_packed.w;
        specializations[4 + 1].i = shape_bordered_packed.h;
        specializations[4 + 2].i = shape_bordered_packed.c;
        specializations[4 + 3].i = shape_bordered_packed.cstep;
        specializations[4 + 4].i = out_shape_packed.w;
        specializations[4 + 5].i = out_shape_packed.h;
        specializations[4 + 6].i = out_shape_packed.c;
        specializations[4 + 7].i = out_shape_packed.cstep;

        int shader_type_index = -1;
        if (elempack == 1 && out_elempack == 1) shader_type_index = LayerShaderType::convolution_1x1s1d1;
        if (elempack == 4 && out_elempack == 4) shader_type_index = LayerShaderType::convolution_pack4_1x1s1d1;
        if (elempack == 1 && out_elempack == 4) shader_type_index = LayerShaderType::convolution_pack1to4_1x1s1d1;
        if (elempack == 4 && out_elempack == 1) shader_type_index = LayerShaderType::convolution_pack4to1_1x1s1d1;
        if (elempack == 8 && out_elempack == 8) shader_type_index = LayerShaderType::convolution_pack8_1x1s1d1;
        if (elempack == 1 && out_elempack == 8) shader_type_index = LayerShaderType::convolution_pack1to8_1x1s1d1;
        if (elempack == 8 && out_elempack == 1) shader_type_index = LayerShaderType::convolution_pack8to1_1x1s1d1;
        if (elempack == 4 && out_elempack == 8) shader_type_index = LayerShaderType::convolution_pack4to8_1x1s1d1;
        if (elempack == 8 && out_elempack == 4) shader_type_index = LayerShaderType::convolution_pack8to4_1x1s1d1;

        if (use_cooperative_matrix_16_8_8)
        {
            if (vkdev->info.support_VK_KHR_cooperative_matrix())
                shader_type_index = LayerShaderType::convolution_pack4_1x1s1d1_khr_cm_16_8_8;
            else
                shader_type_index = LayerShaderType::convolution_pack4_1x1s1d1_nv_cm_16_8_8;
        }
        else if (use_cooperative_matrix_16_16_16)
        {
            if (vkdev->info.support_VK_KHR_cooperative_matrix())
                shader_type_index = LayerShaderType::convolution_pack4_1x1s1d1_khr_cm_16_16_16;
            else
                shader_type_index = LayerShaderType::convolution_pack4_1x1s1d1_nv_cm_16_16_16;
        }

        pipeline_convolution_1x1s1d1 = new Pipeline(vkdev);
        if (use_cooperative_matrix_16_8_8)
        {
            pipeline_convolution_1x1s1d1->set_local_size_xyz(32, 1, 1); // 16_8_8
        }
        else if (use_cooperative_matrix_16_16_16)
        {
            pipeline_convolution_1x1s1d1->set_local_size_xyz(32, 1, 1); // 16_16_16
        }
        else if (opt.use_shader_local_memory)
        {
            pipeline_convolution_1x1s1d1->set_local_size_xyz(8, 8, 1);
        }
        else
        {
            pipeline_convolution_1x1s1d1->set_local_size_xyz(8, std::min(8, num_output / out_elempack), 1);
        }
        pipeline_convolution_1x1s1d1->create(shader_type_index, opt, specializations);
    }
    else
    {
        std::vector<vk_specialization_type> specializations(10 + 10);
        specializations[0].i = kernel_w;
        specializations[1].i = kernel_h;
        specializations[2].i = dilation_w;
        specializations[3].i = dilation_h;
        specializations[4].i = stride_w;
        specializations[5].i = stride_h;
        specializations[6].i = bias_term;
        specializations[7].i = activation_type;
        specializations[8].f = activation_params.w >= 1 ? activation_params[0] : 0.f;
        specializations[9].f = activation_params.w == 2 ? activation_params[1] : 0.f;
        specializations[10 + 0].i = shape_bordered_packed.dims;
        specializations[10 + 1].i = shape_bordered_packed.w;
        specializations[10 + 2].i = shape_bordered_packed.h;
        specializations[10 + 3].i = shape_bordered_packed.c;
        specializations[10 + 4].i = shape_bordered_packed.cstep;
        specializations[10 + 5].i = out_shape_packed.dims;
        specializations[10 + 6].i = out_shape_packed.w;
        specializations[10 + 7].i = out_shape_packed.h;
        specializations[10 + 8].i = out_shape_packed.c;
        specializations[10 + 9].i = out_shape_packed.cstep;

        Mat local_size_xyz(8, 8, std::min(4, (num_output / out_elempack + 1) / 2), (void*)0);
        if (out_shape_packed.dims != 0)
        {
            local_size_xyz.w = std::min(8, out_shape_packed.w);
            local_size_xyz.h = std::min(8, out_shape_packed.h);
            local_size_xyz.c = std::min(4, (out_shape_packed.c + 1) / 2);
        }

        int shader_type_index = -1;
        if (elempack == 1 && out_elempack == 1) shader_type_index = LayerShaderType::convolution;
        if (elempack == 4 && out_elempack == 4) shader_type_index = LayerShaderType::convolution_pack4;
        if (elempack == 1 && out_elempack == 4) shader_type_index = LayerShaderType::convolution_pack1to4;
        if (elempack == 4 && out_elempack == 1) shader_type_index = LayerShaderType::convolution_pack4to1;
        if (elempack == 8 && out_elempack == 8) shader_type_index = LayerShaderType::convolution_pack8;
        if (elempack == 1 && out_elempack == 8) shader_type_index = LayerShaderType::convolution_pack1to8;
        if (elempack == 8 && out_elempack == 1) shader_type_index = LayerShaderType::convolution_pack8to1;
        if (elempack == 4 && out_elempack == 8) shader_type_index = LayerShaderType::convolution_pack4to8;
        if (elempack == 8 && out_elempack == 4) shader_type_index = LayerShaderType::convolution_pack8to4;

        pipeline_convolution = new Pipeline(vkdev);
        pipeline_convolution->set_optimal_local_size_xyz(local_size_xyz);
        pipeline_convolution->create(shader_type_index, opt, specializations);
    }

    return 0;
}

int Convolution_vulkan::destroy_pipeline(const Option& opt)
{
    if (padding)
    {
        padding->destroy_pipeline(opt);
        delete padding;
        padding = 0;
    }

    delete pipeline_convolution;
    pipeline_convolution = 0;

    delete pipeline_convolution_1x1s1d1;
    pipeline_convolution_1x1s1d1 = 0;

    delete pipeline_convolution_gemm;
    pipeline_convolution_gemm = 0;

    delete pipeline_convolution_3x3s1d1_winograd23_transform_input;
    delete pipeline_convolution_3x3s1d1_winograd23_gemm;
    delete pipeline_convolution_3x3s1d1_winograd23_transform_output;
    pipeline_convolution_3x3s1d1_winograd23_transform_input = 0;
    pipeline_convolution_3x3s1d1_winograd23_gemm = 0;
    pipeline_convolution_3x3s1d1_winograd23_transform_output = 0;

    delete pipeline_convolution_3x3s1d1_winograd43_transform_input;
    delete pipeline_convolution_3x3s1d1_winograd43_gemm;
    delete pipeline_convolution_3x3s1d1_winograd43_transform_output;
    pipeline_convolution_3x3s1d1_winograd43_transform_input = 0;
    pipeline_convolution_3x3s1d1_winograd43_gemm = 0;
    pipeline_convolution_3x3s1d1_winograd43_transform_output = 0;

    // fc
    if (reshape_1x1xw)
    {
        reshape_1x1xw->destroy_pipeline(opt);
        delete reshape_1x1xw;
        reshape_1x1xw = 0;
    }

    if (reshape_w)
    {
        reshape_w->destroy_pipeline(opt);
        delete reshape_w;
        reshape_w = 0;
    }

    return 0;
}

int Convolution_vulkan::upload_model(VkTransfer& cmd, const Option& opt)
{
    if (padding)
    {
        padding->upload_model(cmd, opt);
    }

    const int maxk = kernel_w * kernel_h;
    int num_input = weight_data_size / maxk / num_output;

    bool is_conv3x3s1d1 = kernel_w == 3 && kernel_h == 3 && stride_w == 1 && stride_h == 1 && dilation_w == 1 && dilation_h == 1;

    if (opt.use_winograd_convolution && (opt.use_winograd23_convolution || opt.use_winograd43_convolution) && is_conv3x3s1d1 && num_input >= 16 && num_output >= 16)
    {
        // winograd43
        if (opt.use_winograd43_convolution)
        {
            if (support_image_storage && opt.use_image_storage)
            {
                cmd.record_upload(weight_winograd43_data_packed, weight_data_gpu_tm_winograd43_image, opt);
            }
            else
            {
                cmd.record_upload(weight_winograd43_data_packed, weight_data_gpu_tm_winograd43, opt);
            }

            weight_winograd43_data_packed.release();
        }

        // winograd23
        if (opt.use_winograd23_convolution)
        {
            if (support_image_storage && opt.use_image_storage)
            {
                cmd.record_upload(weight_winograd23_data_packed, weight_data_gpu_tm_winograd23_image, opt);
            }
            else
            {
                cmd.record_upload(weight_winograd23_data_packed, weight_data_gpu_tm_winograd23, opt);
            }

            weight_winograd23_data_packed.release();
        }
    }
    else
    {
        if (support_image_storage && opt.use_image_storage)
        {
            cmd.record_upload(weight_data_packed, weight_data_gpu_image, opt);
        }
        else
        {
            cmd.record_upload(weight_data_packed, weight_data_gpu, opt);
        }

        weight_data_packed.release();
    }

    if (bias_term)
    {
        if (support_image_storage && opt.use_image_storage)
        {
            cmd.record_upload(bias_data_packed, bias_data_gpu_image, opt);
        }
        else
        {
            cmd.record_upload(bias_data_packed, bias_data_gpu, opt);
        }

        bias_data_packed.release();
    }

    return 0;
}

int Convolution_vulkan::forward(const VkMat& bottom_blob, VkMat& top_blob, VkCompute& cmd, 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;
    int elempack = bottom_blob.elempack;

    // 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)
        {
            VkMat bottom_blob_1x1xw;
            {
                Option opt_reshape = opt;
                opt_reshape.blob_vkallocator = opt.workspace_vkallocator;
                reshape_1x1xw->forward(bottom_blob, bottom_blob_1x1xw, cmd, opt_reshape);
            }

            if (bottom_blob_1x1xw.empty())
                return -100;

            VkMat top_blob_1x1xw;
            int ret = forward(bottom_blob_1x1xw, top_blob_1x1xw, cmd, opt);
            if (ret != 0)
                return ret;

            return reshape_w->forward(top_blob_1x1xw, top_blob, cmd, opt);
        }
    }

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

    VkMat bottom_blob_bordered = bottom_blob;
    if (pad_left > 0 || pad_right > 0 || pad_top > 0 || pad_bottom > 0)
    {
        Option opt_pad = opt;
        opt_pad.blob_vkallocator = opt.workspace_vkallocator;

        padding->forward(bottom_blob, bottom_blob_bordered, cmd, opt_pad);
    }
    else if (pad_left == -233 && pad_right == -233 && pad_top == -233 && pad_bottom == -233)
    {
        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_pad = opt;
            opt_pad.blob_vkallocator = opt.workspace_vkallocator;

            VkMat padding_param_blob(6, (size_t)4u, 1, opt.staging_vkallocator);
            int* padding_params = padding_param_blob.mapped();

            padding_params[0] = hpad / 2;
            padding_params[1] = hpad - hpad / 2;
            padding_params[2] = wpad / 2;
            padding_params[3] = wpad - wpad / 2;
            padding_params[4] = 0;
            padding_params[5] = 0;
            std::vector<VkMat> padding_inputs(2);
            padding_inputs[0] = bottom_blob;
            padding_inputs[1] = padding_param_blob;

            std::vector<VkMat> padding_outputs(1);
            padding->forward(padding_inputs, padding_outputs, cmd, opt_pad);
            bottom_blob_bordered = padding_outputs[0];
        }
    }
    else if (pad_left == -234 && pad_right == -234 && pad_top == -234 && pad_bottom == -234)
    {
        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_pad = opt;
            opt_pad.blob_vkallocator = opt.workspace_vkallocator;

            VkMat padding_param_blob(6, (size_t)4u, 1, opt.staging_vkallocator);
            int* padding_params = padding_param_blob.mapped();

            padding_params[0] = hpad - hpad / 2;
            padding_params[1] = hpad / 2;
            padding_params[2] = wpad - wpad / 2;
            padding_params[3] = wpad / 2;
            padding_params[4] = 0;
            padding_params[5] = 0;

            std::vector<VkMat> padding_inputs(2);
            padding_inputs[0] = bottom_blob;
            padding_inputs[1] = padding_param_blob;

            std::vector<VkMat> padding_outputs(1);
            padding->forward(padding_inputs, padding_outputs, cmd, opt_pad);
            bottom_blob_bordered = padding_outputs[0];
        }
    }

    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;
    int out_elempack = opt.use_shader_pack8 && num_output % 8 == 0 ? 8 : num_output % 4 == 0 ? 4 : 1;
    size_t out_elemsize = elemsize / elempack * out_elempack;

    if (opt.use_fp16_packed && !opt.use_fp16_storage)
    {
        if (out_elempack == 8) out_elemsize = 8 * 2u;
        if (out_elempack == 4) out_elemsize = 4 * 2u;
        if (out_elempack == 1) out_elemsize = 4u;
    }

    bool is_conv1x1s1d1 = kernel_w == 1 && kernel_h == 1 && stride_w == 1 && stride_h == 1 && dilation_w == 1 && dilation_h == 1;
    bool is_conv3x3s1d1 = kernel_w == 3 && kernel_h == 3 && stride_w == 1 && stride_h == 1 && dilation_w == 1 && dilation_h == 1;

    if (opt.use_winograd_convolution && (opt.use_winograd23_convolution || opt.use_winograd43_convolution) && is_conv3x3s1d1 && channels * elempack >= 16 && num_output >= 16)
    {
        bool use_cooperative_matrix_16_8_8 = vkdev->info.support_cooperative_matrix_16_8_8() && opt.use_cooperative_matrix && !opt.use_image_storage && !opt.use_shader_pack8 && opt.use_fp16_storage && channels * elempack % 8 == 0 && num_output % 8 == 0;
        bool use_cooperative_matrix_16_16_16 = vkdev->info.support_cooperative_matrix_16_16_16() && opt.use_cooperative_matrix && !opt.use_image_storage && !opt.use_shader_pack8 && opt.use_fp16_storage && channels * elempack % 16 == 0 && num_output % 16 == 0;

        bool pre_winograd43 = opt.use_winograd43_convolution;
        if (opt.use_winograd23_convolution)
        {
            if (vkdev->info.type() == 0 && ((w <= 18 && h <= 18) || ((w >= 23 && w <= 24) && (h >= 23 && h <= 24))))
                pre_winograd43 = false;
            if (vkdev->info.type() != 0 && (w <= 12 && h <= 12))
                pre_winograd43 = false;

            if (use_cooperative_matrix_16_8_8 && (w <= 18 && h <= 18))
                pre_winograd43 = false;
            else if (use_cooperative_matrix_16_16_16 && (w <= 18 && h <= 18))
                pre_winograd43 = false;
        }

        if (pre_winograd43)
        {
            // winograd43
            int block_x = (outw + 3) / 4;
            int block_y = (outh + 3) / 4;

            // transform input
            VkMat bottom_tm_blob;
            {
                bottom_tm_blob.create(block_x * block_y, channels, 36, elemsize, elempack, opt.workspace_vkallocator);
                if (bottom_tm_blob.empty())
                    return -100;

                std::vector<VkMat> bindings(2);
                bindings[0] = bottom_blob_bordered;
                bindings[1] = bottom_tm_blob;

                std::vector<vk_constant_type> constants(7);
                constants[0].i = bottom_blob_bordered.w;
                constants[1].i = bottom_blob_bordered.h;
                constants[2].i = bottom_blob_bordered.c;
                constants[3].i = bottom_blob_bordered.cstep;
                constants[4].i = bottom_tm_blob.cstep;
                constants[5].i = block_x;
                constants[6].i = block_y;

                VkMat dispatcher;
                dispatcher.w = block_x;
                dispatcher.h = block_y;
                dispatcher.c = bottom_tm_blob.h;

                cmd.record_pipeline(pipeline_convolution_3x3s1d1_winograd43_transform_input, bindings, constants, dispatcher);
            }

            // gemm
            VkMat top_tm_blob;
            {
                top_tm_blob.create(block_x * block_y, num_output / out_elempack, 36, out_elemsize, out_elempack, opt.workspace_vkallocator);
                if (top_tm_blob.empty())
                    return -100;

                std::vector<VkMat> bindings(3);
                bindings[0] = bottom_tm_blob;
                bindings[1] = top_tm_blob;
                bindings[2] = weight_data_gpu_tm_winograd43;

                std::vector<vk_constant_type> constants(5);
                constants[0].i = bottom_tm_blob.h;
                constants[1].i = bottom_tm_blob.cstep;
                constants[2].i = top_tm_blob.w;
                constants[3].i = top_tm_blob.h;
                constants[4].i = top_tm_blob.cstep;

                VkMat dispatcher;
                dispatcher.w = (top_tm_blob.w + 3) / 4;
                dispatcher.h = top_tm_blob.h;
                dispatcher.c = 36;

                if (use_cooperative_matrix_16_8_8)
                {
                    dispatcher.w = ((top_tm_blob.w + 15) / 16 + 1) / 2 * 32;
                    dispatcher.h = ((top_tm_blob.h + 1) / 2 + 3) / 4;
                    dispatcher.c = 36;
                }
                else if (use_cooperative_matrix_16_16_16)
                {
                    dispatcher.w = ((top_tm_blob.w + 15) / 16 + 1) / 2 * 32;
                    dispatcher.h = ((top_tm_blob.h + 3) / 4 + 1) / 2;
                    dispatcher.c = 36;
                }

                cmd.record_pipeline(pipeline_convolution_3x3s1d1_winograd43_gemm, bindings, constants, dispatcher);
            }

            // transform output
            {
                top_blob.create(outw, outh, num_output / out_elempack, out_elemsize, out_elempack, opt.blob_vkallocator);
                if (top_blob.empty())
                    return -100;

                std::vector<VkMat> bindings(3);
                bindings[0] = top_tm_blob;
                bindings[1] = top_blob;
                bindings[2] = bias_data_gpu;

                std::vector<vk_constant_type> constants(7);
                constants[0].i = top_tm_blob.h;
                constants[1].i = top_tm_blob.cstep;
                constants[2].i = block_x;
                constants[3].i = block_y;
                constants[4].i = top_blob.w;
                constants[5].i = top_blob.h;
                constants[6].i = top_blob.cstep;

                VkMat dispatcher;
                dispatcher.w = block_x;
                dispatcher.h = block_y;
                dispatcher.c = top_blob.c;

                cmd.record_pipeline(pipeline_convolution_3x3s1d1_winograd43_transform_output, bindings, constants, dispatcher);
            }
        }
        else
        {
            // winograd23
            int block_x = (outw + 1) / 2;
            int block_y = (outh + 1) / 2;

            // transform input
            VkMat bottom_tm_blob;
            {
                bottom_tm_blob.create(block_x * block_y, channels, 16, elemsize, elempack, opt.workspace_vkallocator);
                if (bottom_tm_blob.empty())
                    return -100;

                std::vector<VkMat> bindings(2);
                bindings[0] = bottom_blob_bordered;
                bindings[1] = bottom_tm_blob;

                std::vector<vk_constant_type> constants(7);
                constants[0].i = bottom_blob_bordered.w;
                constants[1].i = bottom_blob_bordered.h;
                constants[2].i = bottom_blob_bordered.c;
                constants[3].i = bottom_blob_bordered.cstep;
                constants[4].i = bottom_tm_blob.cstep;
                constants[5].i = block_x;
                constants[6].i = block_y;

                VkMat dispatcher;
                dispatcher.w = block_x;
                dispatcher.h = block_y;
                dispatcher.c = bottom_tm_blob.h;

                cmd.record_pipeline(pipeline_convolution_3x3s1d1_winograd23_transform_input, bindings, constants, dispatcher);
            }

            // gemm
            VkMat top_tm_blob;
            {
                top_tm_blob.create(block_x * block_y, num_output / out_elempack, 16, out_elemsize, out_elempack, opt.workspace_vkallocator);
                if (top_tm_blob.empty())
                    return -100;

                std::vector<VkMat> bindings(3);
                bindings[0] = bottom_tm_blob;
                bindings[1] = top_tm_blob;
                bindings[2] = weight_data_gpu_tm_winograd23;

                std::vector<vk_constant_type> constants(5);
                constants[0].i = bottom_tm_blob.h;
                constants[1].i = bottom_tm_blob.cstep;
                constants[2].i = top_tm_blob.w;
                constants[3].i = top_tm_blob.h;
                constants[4].i = top_tm_blob.cstep;

                VkMat dispatcher;
                dispatcher.w = (top_tm_blob.w + 3) / 4;
                dispatcher.h = top_tm_blob.h;
                dispatcher.c = 16;

                if (use_cooperative_matrix_16_8_8)
                {
                    dispatcher.w = ((top_tm_blob.w + 15) / 16 + 1) / 2 * 32;
                    dispatcher.h = ((top_tm_blob.h + 1) / 2 + 3) / 4;
                    dispatcher.c = 16;
                }
                else if (use_cooperative_matrix_16_16_16)
                {
                    dispatcher.w = ((top_tm_blob.w + 15) / 16 + 1) / 2 * 32;
                    dispatcher.h = ((top_tm_blob.h + 3) / 4 + 1) / 2;
                    dispatcher.c = 16;
                }

                cmd.record_pipeline(pipeline_convolution_3x3s1d1_winograd23_gemm, bindings, constants, dispatcher);
            }

            // transform output
            {
                top_blob.create(outw, outh, num_output / out_elempack, out_elemsize, out_elempack, opt.blob_vkallocator);
                if (top_blob.empty())
                    return -100;

                std::vector<VkMat> bindings(3);
                bindings[0] = top_tm_blob;
                bindings[1] = top_blob;
                bindings[2] = bias_data_gpu;

                std::vector<vk_constant_type> constants(7);
                constants[0].i = top_tm_blob.h;
                constants[1].i = top_tm_blob.cstep;
                constants[2].i = block_x;
                constants[3].i = block_y;
                constants[4].i = top_blob.w;
                constants[5].i = top_blob.h;
                constants[6].i = top_blob.cstep;

                VkMat dispatcher;
                dispatcher.w = block_x;
                dispatcher.h = block_y;
                dispatcher.c = top_blob.c;

                cmd.record_pipeline(pipeline_convolution_3x3s1d1_winograd23_transform_output, bindings, constants, dispatcher);
            }
        }

        return 0;
    }
    if (opt.use_sgemm_convolution && !is_conv1x1s1d1 && channels * elempack >= 16 && num_output >= 16)
    {
        bool use_cooperative_matrix_16_8_8 = vkdev->info.support_cooperative_matrix_16_8_8() && opt.use_cooperative_matrix && !opt.use_image_storage && !opt.use_shader_pack8 && opt.use_fp16_storage && channels * elempack % 8 == 0 && num_output % 8 == 0;
        bool use_cooperative_matrix_16_16_16 = vkdev->info.support_cooperative_matrix_16_16_16() && opt.use_cooperative_matrix && !opt.use_image_storage && !opt.use_shader_pack8 && opt.use_fp16_storage && channels * elempack % 16 == 0 && num_output % 16 == 0;

        // gemm
        top_blob.create(outw, outh, num_output / out_elempack, out_elemsize, out_elempack, opt.blob_vkallocator);
        if (top_blob.empty())
            return -100;

        std::vector<VkMat> bindings(4);
        bindings[0] = bottom_blob_bordered;
        bindings[1] = top_blob;
        bindings[2] = weight_data_gpu;
        bindings[3] = bias_data_gpu;

        std::vector<vk_constant_type> constants(8);
        constants[0].i = bottom_blob_bordered.w;
        constants[1].i = bottom_blob_bordered.h;
        constants[2].i = bottom_blob_bordered.c;
        constants[3].i = bottom_blob_bordered.cstep;
        constants[4].i = top_blob.w;
        constants[5].i = top_blob.h;
        constants[6].i = top_blob.c;
        constants[7].i = top_blob.cstep;

        VkMat dispatcher;
        dispatcher.w = (top_blob.w * top_blob.h + 3) / 4;
        dispatcher.h = top_blob.c;
        dispatcher.c = 1;

        if (use_cooperative_matrix_16_8_8)
        {
            dispatcher.w = ((top_blob.w * top_blob.h + 15) / 16 + 1) / 2 * 32;
            dispatcher.h = ((top_blob.c + 1) / 2 + 3) / 4;
            dispatcher.c = 1;
        }
        else if (use_cooperative_matrix_16_16_16)
        {
            dispatcher.w = ((top_blob.w * top_blob.h + 15) / 16 + 1) / 2 * 32;
            dispatcher.h = ((top_blob.c + 3) / 4 + 1) / 2;
            dispatcher.c = 1;
        }

        cmd.record_pipeline(pipeline_convolution_gemm, bindings, constants, dispatcher);

        return 0;
    }
    if (is_conv1x1s1d1)
    {
        bool use_cooperative_matrix_16_8_8 = vkdev->info.support_cooperative_matrix_16_8_8() && opt.use_cooperative_matrix && !opt.use_image_storage && !opt.use_shader_pack8 && opt.use_fp16_storage && channels * elempack % 8 == 0 && num_output % 8 == 0;
        bool use_cooperative_matrix_16_16_16 = vkdev->info.support_cooperative_matrix_16_16_16() && opt.use_cooperative_matrix && !opt.use_image_storage && !opt.use_shader_pack8 && opt.use_fp16_storage && channels * elempack % 16 == 0 && num_output % 16 == 0;

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

        std::vector<VkMat> bindings(4);
        bindings[0] = bottom_blob_bordered;
        bindings[1] = top_blob;
        bindings[2] = weight_data_gpu;
        bindings[3] = bias_data_gpu;

        std::vector<vk_constant_type> constants(8);
        constants[0].i = bottom_blob_bordered.w;
        constants[1].i = bottom_blob_bordered.h;
        constants[2].i = bottom_blob_bordered.c;
        constants[3].i = bottom_blob_bordered.cstep;
        constants[4].i = top_blob.w;
        constants[5].i = top_blob.h;
        constants[6].i = top_blob.c;
        constants[7].i = top_blob.cstep;

        VkMat dispatcher;
        dispatcher.w = (top_blob.w * top_blob.h + 3) / 4;
        dispatcher.h = top_blob.c;
        dispatcher.c = 1;

        if (use_cooperative_matrix_16_8_8)
        {
            dispatcher.w = ((top_blob.w * top_blob.h + 15) / 16 + 1) / 2 * 32;
            dispatcher.h = ((top_blob.c + 1) / 2 + 3) / 4;
            dispatcher.c = 1;
        }
        else if (use_cooperative_matrix_16_16_16)
        {
            dispatcher.w = ((top_blob.w * top_blob.h + 15) / 16 + 1) / 2 * 32;
            dispatcher.h = ((top_blob.c + 3) / 4 + 1) / 2;
            dispatcher.c = 1;
        }

        cmd.record_pipeline(pipeline_convolution_1x1s1d1, bindings, constants, dispatcher);

        return 0;
    }

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

    std::vector<VkMat> bindings(4);
    bindings[0] = bottom_blob_bordered;
    bindings[1] = top_blob;
    bindings[2] = weight_data_gpu;
    bindings[3] = bias_data_gpu;

    std::vector<vk_constant_type> constants(10);
    constants[0].i = bottom_blob_bordered.dims;
    constants[1].i = bottom_blob_bordered.w;
    constants[2].i = bottom_blob_bordered.h;
    constants[3].i = bottom_blob_bordered.c;
    constants[4].i = bottom_blob_bordered.cstep;
    constants[5].i = top_blob.dims;
    constants[6].i = top_blob.w;
    constants[7].i = top_blob.h;
    constants[8].i = top_blob.c;
    constants[9].i = top_blob.cstep;

    VkMat dispatcher;
    dispatcher.w = (top_blob.w + 1) / 2;
    dispatcher.h = (top_blob.h + 1) / 2;
    dispatcher.c = (top_blob.c + 1) / 2;

    cmd.record_pipeline(pipeline_convolution, bindings, constants, dispatcher);

    return 0;
}

int Convolution_vulkan::forward(const VkImageMat& bottom_blob, VkImageMat& top_blob, VkCompute& cmd, 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;
    int elempack = bottom_blob.elempack;

    // 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)
        {
            VkImageMat bottom_blob_1x1xw;
            {
                Option opt_reshape = opt;
                opt_reshape.blob_vkallocator = opt.workspace_vkallocator;
                reshape_1x1xw->forward(bottom_blob, bottom_blob_1x1xw, cmd, opt_reshape);
            }

            if (bottom_blob_1x1xw.empty())
                return -100;

            VkImageMat top_blob_1x1xw;
            int ret = forward(bottom_blob_1x1xw, top_blob_1x1xw, cmd, opt);
            if (ret != 0)
                return ret;

            return reshape_w->forward(top_blob_1x1xw, top_blob, cmd, opt);
        }
    }

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

    VkImageMat bottom_blob_bordered = bottom_blob;
    if (pad_left > 0 || pad_right > 0 || pad_top > 0 || pad_bottom > 0)
    {
        Option opt_pad = opt;
        opt_pad.blob_vkallocator = opt.workspace_vkallocator;

        padding->forward(bottom_blob, bottom_blob_bordered, cmd, opt_pad);
    }
    else if (pad_left == -233 && pad_right == -233 && pad_top == -233 && pad_bottom == -233)
    {
        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_pad = opt;
            opt_pad.blob_vkallocator = opt.workspace_vkallocator;

            VkImageMat padding_param_blob(6, (size_t)4u, 1, opt.staging_vkallocator);
            int* padding_params = padding_param_blob.mapped();

            padding_params[0] = hpad / 2;
            padding_params[1] = hpad - hpad / 2;
            padding_params[2] = wpad / 2;
            padding_params[3] = wpad - wpad / 2;
            padding_params[4] = 0;
            padding_params[5] = 0;

            std::vector<VkImageMat> padding_inputs(2);
            padding_inputs[0] = bottom_blob;
            padding_inputs[1] = padding_param_blob;

            std::vector<VkImageMat> padding_outputs(1);
            padding->forward(padding_inputs, padding_outputs, cmd, opt_pad);
            bottom_blob_bordered = padding_outputs[0];
        }
    }
    else if (pad_left == -234 && pad_right == -234 && pad_top == -234 && pad_bottom == -234)
    {
        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_pad = opt;
            opt_pad.blob_vkallocator = opt.workspace_vkallocator;

            VkImageMat padding_param_blob(6, (size_t)4u, 1, opt.staging_vkallocator);
            int* padding_params = padding_param_blob.mapped();

            padding_params[0] = hpad - hpad / 2;
            padding_params[1] = hpad / 2;
            padding_params[2] = wpad - wpad / 2;
            padding_params[3] = wpad / 2;
            padding_params[4] = 0;
            padding_params[5] = 0;

            std::vector<VkImageMat> padding_inputs(2);
            padding_inputs[0] = bottom_blob;
            padding_inputs[1] = padding_param_blob;

            std::vector<VkImageMat> padding_outputs(1);
            padding->forward(padding_inputs, padding_outputs, cmd, opt_pad);
            bottom_blob_bordered = padding_outputs[0];
        }
    }

    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;
    int out_elempack = opt.use_shader_pack8 && num_output % 8 == 0 ? 8 : num_output % 4 == 0 ? 4 : 1;
    size_t out_elemsize = elemsize / elempack * out_elempack;

    if (opt.use_fp16_packed && !opt.use_fp16_storage)
    {
        if (out_elempack == 8) out_elemsize = 8 * 2u;
        if (out_elempack == 4) out_elemsize = 4 * 2u;
        if (out_elempack == 1) out_elemsize = 4u;
    }

    bool is_conv1x1s1d1 = kernel_w == 1 && kernel_h == 1 && stride_w == 1 && stride_h == 1 && dilation_w == 1 && dilation_h == 1;
    bool is_conv3x3s1d1 = kernel_w == 3 && kernel_h == 3 && stride_w == 1 && stride_h == 1 && dilation_w == 1 && dilation_h == 1;

    if (opt.use_winograd_convolution && (opt.use_winograd23_convolution || opt.use_winograd43_convolution) && is_conv3x3s1d1 && channels * elempack >= 16 && num_output >= 16)
    {
        bool pre_winograd43 = opt.use_winograd43_convolution;
        if (opt.use_winograd23_convolution)
        {
            if (vkdev->info.type() == 0 && ((w <= 18 && h <= 18) || ((w >= 23 && w <= 24) && (h >= 23 && h <= 24))))
                pre_winograd43 = false;
            if (vkdev->info.type() != 0 && (w <= 12 && h <= 12))
                pre_winograd43 = false;
        }

        if (pre_winograd43)
        {
            // winograd43
            int block_x = (outw + 3) / 4;
            int block_y = (outh + 3) / 4;

            // transform input
            VkImageMat bottom_tm_blob;
            {
                bottom_tm_blob.create(block_x * block_y, channels, 36, elemsize, elempack, opt.workspace_vkallocator);
                if (bottom_tm_blob.empty())
                    return -100;

                std::vector<VkImageMat> bindings(2);
                bindings[0] = bottom_blob_bordered;
                bindings[1] = bottom_tm_blob;

                std::vector<vk_constant_type> constants(7);
                constants[0].i = bottom_blob_bordered.w;
                constants[1].i = bottom_blob_bordered.h;
                constants[2].i = bottom_blob_bordered.c;
                constants[3].i = 0; //bottom_blob_bordered.cstep;
                constants[4].i = 0; //bottom_tm_blob.cstep;
                constants[5].i = block_x;
                constants[6].i = block_y;

                VkImageMat dispatcher;
                dispatcher.w = block_x;
                dispatcher.h = block_y;
                dispatcher.c = bottom_tm_blob.h;

                cmd.record_pipeline(pipeline_convolution_3x3s1d1_winograd43_transform_input, bindings, constants, dispatcher);
            }

            // gemm
            VkImageMat top_tm_blob;
            {
                top_tm_blob.create(block_x * block_y, num_output / out_elempack, 36, out_elemsize, out_elempack, opt.workspace_vkallocator);
                if (top_tm_blob.empty())
                    return -100;

                std::vector<VkImageMat> bindings(3);
                bindings[0] = bottom_tm_blob;
                bindings[1] = top_tm_blob;
                bindings[2] = weight_data_gpu_tm_winograd43_image;

                std::vector<vk_constant_type> constants(5);
                constants[0].i = bottom_tm_blob.h;
                constants[1].i = 0; //bottom_tm_blob.cstep;
                constants[2].i = top_tm_blob.w;
                constants[3].i = top_tm_blob.h;
                constants[4].i = 0; //top_tm_blob.cstep;

                VkImageMat dispatcher;
                dispatcher.w = (top_tm_blob.w + 3) / 4;
                dispatcher.h = top_tm_blob.h;
                dispatcher.c = 36;

                cmd.record_pipeline(pipeline_convolution_3x3s1d1_winograd43_gemm, bindings, constants, dispatcher);
            }

            // transform output
            {
                top_blob.create(outw, outh, num_output / out_elempack, out_elemsize, out_elempack, opt.blob_vkallocator);
                if (top_blob.empty())
                    return -100;

                std::vector<VkImageMat> bindings(3);
                bindings[0] = top_tm_blob;
                bindings[1] = top_blob;
                bindings[2] = bias_data_gpu_image;

                std::vector<vk_constant_type> constants(7);
                constants[0].i = top_tm_blob.h;
                constants[1].i = 0; //top_tm_blob.cstep;
                constants[2].i = block_x;
                constants[3].i = block_y;
                constants[4].i = top_blob.w;
                constants[5].i = top_blob.h;
                constants[6].i = 0; //top_blob.cstep;

                VkImageMat dispatcher;
                dispatcher.w = block_x;
                dispatcher.h = block_y;
                dispatcher.c = top_blob.c;

                cmd.record_pipeline(pipeline_convolution_3x3s1d1_winograd43_transform_output, bindings, constants, dispatcher);
            }
        }
        else
        {
            // winograd23
            int block_x = (outw + 1) / 2;
            int block_y = (outh + 1) / 2;

            // transform input
            VkImageMat bottom_tm_blob;
            {
                bottom_tm_blob.create(block_x * block_y, channels, 16, elemsize, elempack, opt.workspace_vkallocator);
                if (bottom_tm_blob.empty())
                    return -100;

                std::vector<VkImageMat> bindings(2);
                bindings[0] = bottom_blob_bordered;
                bindings[1] = bottom_tm_blob;

                std::vector<vk_constant_type> constants(7);
                constants[0].i = bottom_blob_bordered.w;
                constants[1].i = bottom_blob_bordered.h;
                constants[2].i = bottom_blob_bordered.c;
                constants[3].i = 0; //bottom_blob_bordered.cstep;
                constants[4].i = 0; //bottom_tm_blob.cstep;
                constants[5].i = block_x;
                constants[6].i = block_y;

                VkImageMat dispatcher;
                dispatcher.w = block_x;
                dispatcher.h = block_y;
                dispatcher.c = bottom_tm_blob.h;

                cmd.record_pipeline(pipeline_convolution_3x3s1d1_winograd23_transform_input, bindings, constants, dispatcher);
            }

            // gemm
            VkImageMat top_tm_blob;
            {
                top_tm_blob.create(block_x * block_y, num_output / out_elempack, 16, out_elemsize, out_elempack, opt.workspace_vkallocator);
                if (top_tm_blob.empty())
                    return -100;

                std::vector<VkImageMat> bindings(3);
                bindings[0] = bottom_tm_blob;
                bindings[1] = top_tm_blob;
                bindings[2] = weight_data_gpu_tm_winograd23_image;

                std::vector<vk_constant_type> constants(5);
                constants[0].i = bottom_tm_blob.h;
                constants[1].i = 0; //bottom_tm_blob.cstep;
                constants[2].i = top_tm_blob.w;
                constants[3].i = top_tm_blob.h;
                constants[4].i = 0; //top_tm_blob.cstep;

                VkImageMat dispatcher;
                dispatcher.w = (top_tm_blob.w + 3) / 4;
                dispatcher.h = top_tm_blob.h;
                dispatcher.c = 16;

                cmd.record_pipeline(pipeline_convolution_3x3s1d1_winograd23_gemm, bindings, constants, dispatcher);
            }

            // transform output
            {
                top_blob.create(outw, outh, num_output / out_elempack, out_elemsize, out_elempack, opt.blob_vkallocator);
                if (top_blob.empty())
                    return -100;

                std::vector<VkImageMat> bindings(3);
                bindings[0] = top_tm_blob;
                bindings[1] = top_blob;
                bindings[2] = bias_data_gpu_image;

                std::vector<vk_constant_type> constants(7);
                constants[0].i = top_tm_blob.h;
                constants[1].i = 0; //top_tm_blob.cstep;
                constants[2].i = block_x;
                constants[3].i = block_y;
                constants[4].i = top_blob.w;
                constants[5].i = top_blob.h;
                constants[6].i = 0; //top_blob.cstep;

                VkImageMat dispatcher;
                dispatcher.w = block_x;
                dispatcher.h = block_y;
                dispatcher.c = top_blob.c;

                cmd.record_pipeline(pipeline_convolution_3x3s1d1_winograd23_transform_output, bindings, constants, dispatcher);
            }
        }

        return 0;
    }
    if (opt.use_sgemm_convolution && !is_conv1x1s1d1 && channels * elempack >= 16 && num_output >= 16)
    {
        // gemm
        top_blob.create(outw, outh, num_output / out_elempack, out_elemsize, out_elempack, opt.blob_vkallocator);
        if (top_blob.empty())
            return -100;

        std::vector<VkImageMat> bindings(4);
        bindings[0] = bottom_blob_bordered;
        bindings[1] = top_blob;
        bindings[2] = weight_data_gpu_image;
        bindings[3] = bias_data_gpu_image;

        std::vector<vk_constant_type> constants(8);
        constants[0].i = bottom_blob_bordered.w;
        constants[1].i = bottom_blob_bordered.h;
        constants[2].i = bottom_blob_bordered.c;
        constants[3].i = 0; // bottom_blob_bordered.cstep;
        constants[4].i = top_blob.w;
        constants[5].i = top_blob.h;
        constants[6].i = top_blob.c;
        constants[7].i = 0; // top_blob.cstep;

        VkImageMat dispatcher;
        dispatcher.w = (top_blob.w * top_blob.h + 3) / 4;
        dispatcher.h = top_blob.c;
        dispatcher.c = 1;

        cmd.record_pipeline(pipeline_convolution_gemm, bindings, constants, dispatcher);

        return 0;
    }
    if (is_conv1x1s1d1)
    {
        top_blob.create(outw, outh, num_output / out_elempack, out_elemsize, out_elempack, opt.blob_vkallocator);
        if (top_blob.empty())
            return -100;

        std::vector<VkImageMat> bindings(4);
        bindings[0] = bottom_blob_bordered;
        bindings[1] = top_blob;
        bindings[2] = weight_data_gpu_image;
        bindings[3] = bias_data_gpu_image;

        std::vector<vk_constant_type> constants(8);
        constants[0].i = bottom_blob_bordered.w;
        constants[1].i = bottom_blob_bordered.h;
        constants[2].i = bottom_blob_bordered.c;
        constants[3].i = 0; // bottom_blob_bordered.cstep;
        constants[4].i = top_blob.w;
        constants[5].i = top_blob.h;
        constants[6].i = top_blob.c;
        constants[7].i = 0; // top_blob.cstep;

        VkImageMat dispatcher;
        dispatcher.w = (top_blob.w * top_blob.h + 3) / 4;
        dispatcher.h = top_blob.c;
        dispatcher.c = 1;

        cmd.record_pipeline(pipeline_convolution_1x1s1d1, bindings, constants, dispatcher);

        return 0;
    }

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

    std::vector<VkImageMat> bindings(4);
    bindings[0] = bottom_blob_bordered;
    bindings[1] = top_blob;
    bindings[2] = weight_data_gpu_image;
    bindings[3] = bias_data_gpu_image;

    std::vector<vk_constant_type> constants(10);
    constants[0].i = bottom_blob_bordered.dims;
    constants[1].i = bottom_blob_bordered.w;
    constants[2].i = bottom_blob_bordered.h;
    constants[3].i = bottom_blob_bordered.c;
    constants[4].i = 0; //bottom_blob_bordered.cstep;
    constants[5].i = top_blob.dims;
    constants[6].i = top_blob.w;
    constants[7].i = top_blob.h;
    constants[8].i = top_blob.c;
    constants[9].i = 0; //top_blob.cstep;

    VkImageMat dispatcher;
    dispatcher.w = (top_blob.w + 1) / 2;
    dispatcher.h = (top_blob.h + 1) / 2;
    dispatcher.c = (top_blob.c + 1) / 2;

    cmd.record_pipeline(pipeline_convolution, bindings, constants, dispatcher);

    return 0;
}

} // namespace ncnn