diff --git a/src/layer/convolutiondepthwise.cpp b/src/layer/convolutiondepthwise.cpp
index 206785cb7..a31e03026 100644
--- a/src/layer/convolutiondepthwise.cpp
+++ b/src/layer/convolutiondepthwise.cpp
@@ -760,6 +760,9 @@ int ConvolutionDepthWise::forward(const VkMat& bottom_blob, VkMat& top_blob, VkC
         return 0;
     }
 
+    // record
+    cmd.record_prepare_compute_barrier(top_blob);
+
     const int channels_g = channels / group;
     const int num_output_g = num_output / packing / group;
 
diff --git a/src/layer/deconvolutiondepthwise.cpp b/src/layer/deconvolutiondepthwise.cpp
index 367893625..19773c1a4 100644
--- a/src/layer/deconvolutiondepthwise.cpp
+++ b/src/layer/deconvolutiondepthwise.cpp
@@ -13,6 +13,7 @@
 // specific language governing permissions and limitations under the License.
 
 #include "deconvolutiondepthwise.h"
+#include "layer_type.h"
 
 namespace ncnn {
 
@@ -22,6 +23,22 @@ DeconvolutionDepthWise::DeconvolutionDepthWise()
 {
     one_blob_only = true;
     support_inplace = false;
+    support_vulkan = true;
+
+#if NCNN_VULKAN
+    pipeline_deconvolutiondepthwise = 0;
+    pipeline_deconvolutiondepthwise_pack4 = 0;
+#endif // NCNN_VULKAN
+}
+
+DeconvolutionDepthWise::~DeconvolutionDepthWise()
+{
+#if NCNN_VULKAN
+    for (int i=0; i<(int)deconvolution_group_ops.size(); i++)
+        delete deconvolution_group_ops[i];
+
+    deconvolution_group_ops.clear();
+#endif // NCNN_VULKAN
 }
 
 int DeconvolutionDepthWise::load_param(const ParamDict& pd)
@@ -55,6 +72,65 @@ int DeconvolutionDepthWise::load_model(const ModelBin& mb)
             return -100;
     }
 
+#if NCNN_VULKAN
+    const int maxk = kernel_w * kernel_h;
+    int channels = (weight_data_size / group) / maxk / (num_output / group) * group;
+
+    // group deconvolution
+    if (!(channels == group && group == num_output))
+    {
+        // create Deconvolution op for each group
+
+        for (int i=0; i<(int)deconvolution_group_ops.size(); i++)
+            delete deconvolution_group_ops[i];
+
+        deconvolution_group_ops.clear();
+
+        const int channels_g = channels / group;
+        const int num_output_g = num_output / group;
+
+        deconvolution_group_ops.resize(group);
+
+        for (int g=0; g<group; g++)
+        {
+            Mat weight_data_g = weight_data.range(maxk * channels_g * num_output_g * g, maxk * channels_g * num_output_g);
+            Mat bias_data_g;
+            if (bias_term)
+                bias_data_g = bias_data.range(num_output_g * g, num_output_g);
+
+            ncnn::Layer* op = ncnn::create_layer(ncnn::LayerType::Deconvolution);
+            op->vkdev = vkdev;
+
+            // set param
+            ncnn::ParamDict pd;
+            pd.set(0, num_output_g);// num_output
+            pd.set(1, kernel_w);
+            pd.set(11, kernel_h);
+            pd.set(2, dilation_w);
+            pd.set(12, dilation_h);
+            pd.set(3, stride_w);
+            pd.set(13, stride_h);
+            pd.set(4, 0);// pad_w
+            pd.set(14, 0);// pad_h
+            pd.set(5, bias_term);
+            pd.set(6, maxk * channels_g * num_output_g);// weight_data_size
+
+            pd.use_vulkan_compute = 1;
+
+            op->load_param(pd);
+
+            // set weights
+            ncnn::Mat weights[2];
+            weights[0] = weight_data_g;
+            weights[1] = bias_data_g;
+
+            op->load_model(ModelBinFromMatArray(weights));
+
+            deconvolution_group_ops[g] = op;
+        }
+    }
+#endif // NCNN_VULKAN
+
     return 0;
 }
 
@@ -212,4 +288,204 @@ int DeconvolutionDepthWise::forward(const Mat& bottom_blob, Mat& top_blob, const
     return 0;
 }
 
+#if NCNN_VULKAN
+int DeconvolutionDepthWise::upload_model(VkTransfer& cmd)
+{
+    const int maxk = kernel_w * kernel_h;
+    int channels = (weight_data_size / group) / maxk / (num_output / group) * group;
+
+    Mat weight_data_transposed(weight_data.w);
+    {
+        float* pt = weight_data_transposed;
+        const float* p = weight_data;
+
+        for (int i=0; i<(channels/group)*(num_output/group)*group; i++)
+        {
+            for (int k=0; k<maxk; k++)
+            {
+                pt[maxk-1 - k] = p[k];
+            }
+
+            p += maxk;
+            pt += maxk;
+        }
+    }
+
+    // depth-wise
+    if (channels == group && group == num_output)
+    {
+        cmd.record_upload(weight_data_transposed, weight_data_gpu);
+
+        if (bias_term)
+        {
+            cmd.record_upload(bias_data, bias_data_gpu);
+        }
+
+        // pack4
+        if (channels % 4 == 0 && num_output % 4 == 0)
+        {
+            const int maxk = kernel_w * kernel_h;
+
+            Mat weight_data_pack4;
+            Mat weight_data_r2 = weight_data_transposed.reshape(maxk, group);
+            convert_packing(weight_data_r2, weight_data_pack4, 4);
+
+            weight_data_pack4 = weight_data_pack4.reshape(maxk * (group/4));
+            cmd.record_upload(weight_data_pack4, weight_data_gpu_pack4);
+
+            if (bias_term)
+            {
+                Mat bias_data_pack4;
+                convert_packing(bias_data, bias_data_pack4, 4);
+                cmd.record_upload(bias_data_pack4, bias_data_gpu_pack4);
+            }
+        }
+
+        return 0;
+    }
+
+    for (int g=0; g<group; g++)
+    {
+        deconvolution_group_ops[g]->upload_model(cmd);
+    }
+
+    return 0;
+}
+
+int DeconvolutionDepthWise::create_pipeline()
+{
+    const int maxk = kernel_w * kernel_h;
+    int channels = (weight_data_size / group) / maxk / (num_output / group) * group;
+
+    // depth-wise
+    if (channels == group && group == num_output)
+    {
+        pipeline_deconvolutiondepthwise = new Pipeline(vkdev);
+        pipeline_deconvolutiondepthwise->set_optimal_local_size_xyz(32, 32, num_output);
+
+        std::vector<vk_specialization_type> specializations(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 = group;
+
+        pipeline_deconvolutiondepthwise->create("deconvolutiondepthwise", specializations, 4, 10);
+
+        // pack4
+        if (num_output % 4 == 0)
+        {
+            pipeline_deconvolutiondepthwise_pack4 = new Pipeline(vkdev);
+            pipeline_deconvolutiondepthwise_pack4->set_optimal_local_size_xyz(32, 32, std::max(1, num_output / 4));
+            pipeline_deconvolutiondepthwise_pack4->create("deconvolutiondepthwise_pack4", specializations, 4, 10);
+        }
+
+        return 0;
+    }
+
+    for (int g=0; g<group; g++)
+    {
+        deconvolution_group_ops[g]->create_pipeline();
+    }
+
+    return 0;
+}
+
+int DeconvolutionDepthWise::destroy_pipeline()
+{
+    for (int g=0; g<(int)deconvolution_group_ops.size(); g++)
+    {
+        deconvolution_group_ops[g]->destroy_pipeline();
+    }
+
+    delete pipeline_deconvolutiondepthwise;
+    pipeline_deconvolutiondepthwise = 0;
+
+    delete pipeline_deconvolutiondepthwise_pack4;
+    pipeline_deconvolutiondepthwise_pack4 = 0;
+
+    return 0;
+}
+
+int DeconvolutionDepthWise::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 packing = bottom_blob.packing;
+
+    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
+    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
+
+    int outw = (w - 1) * stride_w + kernel_extent_w;
+    int outh = (h - 1) * stride_h + kernel_extent_h;
+
+    // TODO assert num_output % packing == 0
+
+    top_blob.create(outw, outh, num_output / packing, elemsize, packing, opt.blob_vkallocator, opt.staging_vkallocator);
+    if (top_blob.empty())
+        return -100;
+
+//     fprintf(stderr, "DeconvolutionDepthWise::forward %p %p\n", bottom_blob.buffer(), top_blob.buffer());
+
+    // depth-wise
+    if (channels == group / packing && group / packing == num_output / packing)
+    {
+        std::vector<VkMat> bindings(4);
+        bindings[0] = bottom_blob;
+        bindings[1] = top_blob;
+        bindings[2] = packing == 4 ? weight_data_gpu_pack4 : weight_data_gpu;
+        bindings[3] = bias_term ? (packing == 4 ? bias_data_gpu_pack4 : bias_data_gpu) : weight_data_gpu;// TODO use dummy buffer
+
+        std::vector<vk_constant_type> constants(10);
+        constants[0].i = bottom_blob.dims;
+        constants[1].i = bottom_blob.w;
+        constants[2].i = bottom_blob.h;
+        constants[3].i = bottom_blob.c;
+        constants[4].i = bottom_blob.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;
+
+        const Pipeline* pipeline = packing == 4 ? pipeline_deconvolutiondepthwise_pack4 : pipeline_deconvolutiondepthwise;
+
+        // record
+        cmd.record_prepare_compute_barrier(bottom_blob);
+        cmd.record_prepare_compute_barrier(top_blob);
+        cmd.record_pipeline(pipeline, bindings, constants, top_blob);
+
+        return 0;
+    }
+
+    // record
+    cmd.record_prepare_compute_barrier(top_blob);
+
+    const int channels_g = channels / group;
+    const int num_output_g = num_output / packing / group;
+
+    for (int g=0; g<group; g++)
+    {
+        VkMat bottom_blob_bordered_g = bottom_blob.channel_range(channels_g * g, channels_g);
+        VkMat top_blob_g = top_blob.channel_range(num_output_g * g, num_output_g);
+
+        const ncnn::Layer* op = deconvolution_group_ops[g];
+
+        ncnn::Option opt_g = opt;
+        opt_g.blob_vkallocator = top_blob.allocator;
+        opt_g.staging_vkallocator = top_blob.staging_allocator;
+
+        // forward
+        op->forward(bottom_blob_bordered_g, top_blob_g, cmd, opt_g);
+    }
+
+    return 0;
+}
+#endif // NCNN_VULKAN
+
 } // namespace ncnn
diff --git a/src/layer/deconvolutiondepthwise.h b/src/layer/deconvolutiondepthwise.h
index 674cb2c47..8cab20802 100644
--- a/src/layer/deconvolutiondepthwise.h
+++ b/src/layer/deconvolutiondepthwise.h
@@ -23,6 +23,7 @@ class DeconvolutionDepthWise : public Layer
 {
 public:
     DeconvolutionDepthWise();
+    ~DeconvolutionDepthWise();
 
     virtual int load_param(const ParamDict& pd);
 
@@ -30,6 +31,15 @@ public:
 
     virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
 
+#if NCNN_VULKAN
+    virtual int upload_model(VkTransfer& cmd);
+
+    virtual int create_pipeline();
+    virtual int destroy_pipeline();
+
+    virtual int forward(const VkMat& bottom_blob, VkMat& top_blob, VkCompute& cmd, const Option& opt) const;
+#endif // NCNN_VULKAN
+
 public:
     // param
     int num_output;
@@ -49,6 +59,20 @@ public:
     // model
     Mat weight_data;
     Mat bias_data;
+
+#if NCNN_VULKAN
+    VkMat weight_data_gpu;
+    VkMat bias_data_gpu;
+
+    std::vector<ncnn::Layer*> deconvolution_group_ops;
+
+    Pipeline* pipeline_deconvolutiondepthwise;
+
+    VkMat weight_data_gpu_pack4;
+    VkMat bias_data_gpu_pack4;
+    Pipeline* pipeline_deconvolutiondepthwise_pack4;
+#endif // NCNN_VULKAN
+
 };
 
 } // namespace ncnn
diff --git a/src/layer/shader/deconvolutiondepthwise.comp b/src/layer/shader/deconvolutiondepthwise.comp
new file mode 100644
index 000000000..1e81c4c56
--- /dev/null
+++ b/src/layer/shader/deconvolutiondepthwise.comp
@@ -0,0 +1,105 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
+//
+// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
+// in compliance with the License. You may obtain a copy of the License at
+//
+// https://opensource.org/licenses/BSD-3-Clause
+//
+// Unless required by applicable law or agreed to in writing, software distributed
+// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
+// CONDITIONS OF ANY KIND, either express or implied. See the License for the
+// specific language governing permissions and limitations under the License.
+
+#version 450
+
+layout (constant_id = 0) const int kernel_w = 1;
+layout (constant_id = 1) const int kernel_h = 1;
+layout (constant_id = 2) const int dilation_w = 1;
+layout (constant_id = 3) const int dilation_h = 1;
+layout (constant_id = 4) const int stride_w = 1;
+layout (constant_id = 5) const int stride_h = 1;
+layout (constant_id = 6) const int bias_term = 0;
+layout (constant_id = 7) const int group = 1;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer bottom_blob { float bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { float top_blob_data[]; };
+layout (binding = 2) readonly buffer weight_blob { float weight_data[]; };
+layout (binding = 3) readonly buffer bias_blob { float bias_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.outw || gy >= p.outh || gz >= p.outc)
+        return;
+
+    float sum;
+
+    if (bias_term == 1)
+    {
+        sum = bias_data[gz];
+    }
+    else
+    {
+        sum = 0.f;
+    }
+
+    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
+    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
+
+    // depth-wise deconvolution
+    int v_offset_0 = gz * p.cstep;
+    int w_offset_0 = gz * kernel_w * kernel_h;
+
+    for (int y = 0; y < kernel_h; y++)
+    {
+        int sys = (gy + y * dilation_h - (kernel_extent_h - 1));
+        if (sys % stride_h != 0)
+            continue;
+
+        int sy = sys / stride_h;
+        if (sy < 0 || sy >= p.h)
+            continue;
+
+        for (int x = 0; x < kernel_w; x++)
+        {
+            int sxs = (gx + x * dilation_w - (kernel_extent_w - 1));
+            if (sxs % stride_w != 0)
+                continue;
+
+            int sx = sxs / stride_w;
+            if (sx < 0 || sx >= p.w)
+                continue;
+
+            int v_offset = v_offset_0 + sy * p.w + sx;
+            int w_offset = w_offset_0 + y * kernel_w + x;
+
+            sum += weight_data[w_offset] * bottom_blob_data[v_offset];
+        }
+    }
+
+    top_blob_data[gz * p.outcstep + gy * p.outw + gx] = sum;
+}
diff --git a/src/layer/shader/deconvolutiondepthwise_pack4.comp b/src/layer/shader/deconvolutiondepthwise_pack4.comp
new file mode 100644
index 000000000..8dff3851a
--- /dev/null
+++ b/src/layer/shader/deconvolutiondepthwise_pack4.comp
@@ -0,0 +1,109 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
+//
+// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
+// in compliance with the License. You may obtain a copy of the License at
+//
+// https://opensource.org/licenses/BSD-3-Clause
+//
+// Unless required by applicable law or agreed to in writing, software distributed
+// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
+// CONDITIONS OF ANY KIND, either express or implied. See the License for the
+// specific language governing permissions and limitations under the License.
+
+#version 450
+
+layout (constant_id = 0) const int kernel_w = 1;
+layout (constant_id = 1) const int kernel_h = 1;
+layout (constant_id = 2) const int dilation_w = 1;
+layout (constant_id = 3) const int dilation_h = 1;
+layout (constant_id = 4) const int stride_w = 1;
+layout (constant_id = 5) const int stride_h = 1;
+layout (constant_id = 6) const int bias_term = 0;
+layout (constant_id = 7) const int group = 1;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer bottom_blob { vec4 bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { vec4 top_blob_data[]; };
+layout (binding = 2) readonly buffer weight_blob { vec4 weight_data[]; };
+layout (binding = 3) readonly buffer bias_blob { vec4 bias_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.outw || gy >= p.outh || gz >= p.outc)
+        return;
+
+    vec4 sum;
+
+    if (bias_term == 1)
+    {
+        sum = bias_data[gz];
+    }
+    else
+    {
+        sum = vec4(0.f);
+    }
+
+    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
+    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
+
+    // depth-wise deconvolution
+    int v_offset_0 = gz * p.cstep;
+    int w_offset_0 = gz * kernel_w * kernel_h;
+
+    for (int y = 0; y < kernel_h; y++)
+    {
+        int sys = (gy + y * dilation_h - (kernel_extent_h - 1));
+        if (sys % stride_h != 0)
+            continue;
+
+        int sy = sys / stride_h;
+        if (sy < 0 || sy >= p.h)
+            continue;
+
+        for (int x = 0; x < kernel_w; x++)
+        {
+            int sxs = (gx + x * dilation_w - (kernel_extent_w - 1));
+            if (sxs % stride_w != 0)
+                continue;
+
+            int sx = sxs / stride_w;
+            if (sx < 0 || sx >= p.w)
+                continue;
+
+            int v_offset = v_offset_0 + sy * p.w + sx;
+            int w_offset = w_offset_0 + y * kernel_w + x;
+
+            vec4 v = bottom_blob_data[v_offset];
+
+            vec4 k = weight_data[w_offset];
+
+            sum += v * k;
+        }
+    }
+
+    top_blob_data[gz * p.outcstep + gy * p.outw + gx] = sum;
+}