asymmetric padding parameter for convolution and deconvolution family

src/layer/arm/convolution_arm.cpp

@@ -374,18 +374,18 @@ int Convolution_arm::forwardDilation(const Mat& bottom_blob, Mat& top_blob, conv
     const int kernel_extent = dilation * (kernel_size - 1) + 1;
 
     Mat bottom_blob_bordered = bottom_blob;
-    if (pad_w > 0 || pad_h > 0)
+    if (pad_left > 0 || pad_right > 0 || pad_top > 0 || pad_bottom > 0)
     {
         Option opt_b = opt;
         opt_b.blob_allocator = opt.workspace_allocator;
-        copy_make_border(bottom_blob, bottom_blob_bordered, pad_h, pad_h, pad_w, pad_w, BORDER_CONSTANT, 0.f, opt_b);
+        copy_make_border(bottom_blob, bottom_blob_bordered, pad_top, pad_bottom, pad_left, pad_right, BORDER_CONSTANT, 0.f, opt_b);
         if (bottom_blob_bordered.empty())
             return -100;
 
         w = bottom_blob_bordered.w;
         h = bottom_blob_bordered.h;
     }
-    else if (pad_w == -233 && pad_h == -233)
+    else if (pad_left == -233 && pad_right == -233 && pad_top == -233 && pad_bottom == -233)
     {
         int wpad = kernel_extent + (w - 1) / stride * stride - w;
         int hpad = kernel_extent + (h - 1) / stride * stride - h;
@@ -491,18 +491,18 @@ int Convolution_arm::forward(const Mat& bottom_blob, Mat& top_blob, const Option
     const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
 
     Mat bottom_blob_bordered = bottom_blob;
-    if (pad_w > 0 || pad_h > 0)
+    if (pad_left > 0 || pad_right > 0 || pad_top > 0 || pad_bottom > 0)
     {
         Option opt_b = opt;
         opt_b.blob_allocator = opt.workspace_allocator;
-        copy_make_border(bottom_blob, bottom_blob_bordered, pad_h, pad_h, pad_w, pad_w, BORDER_CONSTANT, 0.f, opt_b);
+        copy_make_border(bottom_blob, bottom_blob_bordered, pad_top, pad_bottom, pad_left, pad_right, BORDER_CONSTANT, 0.f, opt_b);
         if (bottom_blob_bordered.empty())
             return -100;
 
         w = bottom_blob_bordered.w;
         h = bottom_blob_bordered.h;
     }
-    else if (pad_w == -233 && pad_h == -233)
+    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;
@@ -973,18 +973,18 @@ int Convolution_arm::forward(const Mat& bottom_blob, Mat& top_blob, const Option
     }
 
     Mat bottom_blob_bordered = bottom_blob_unbordered;
-    if (pad_w > 0 || pad_h > 0)
+    if (pad_left > 0 || pad_right > 0 || pad_top > 0 || pad_bottom > 0)
     {
         Option opt_b = opt;
         opt_b.blob_allocator = opt.workspace_allocator;
-        copy_make_border(bottom_blob_unbordered, bottom_blob_bordered, pad_h, pad_h, pad_w, pad_w, BORDER_CONSTANT, 0.f, opt_b);
+        copy_make_border(bottom_blob_unbordered, bottom_blob_bordered, pad_top, pad_bottom, pad_left, pad_right, BORDER_CONSTANT, 0.f, opt_b);
         if (bottom_blob_bordered.empty())
             return -100;
 
         w = bottom_blob_bordered.w;
         h = bottom_blob_bordered.h;
     }
-    else if (pad_w == -233 && pad_h == -233)
+    else if (pad_left == -233 && pad_right == -233 && pad_top == -233 && pad_bottom == -233)
     {
         int wpad = kernel_size + (w - 1) / stride * stride - w;
         int hpad = kernel_size + (h - 1) / stride * stride - h;

src/layer/arm/convolutiondepthwise_arm.cpp

@@ -435,18 +435,18 @@ int ConvolutionDepthWise_arm::forward(const Mat& bottom_blob, Mat& top_blob, con
     }    
 
     Mat bottom_blob_bordered = bottom_blob_unbordered;
-    if (pad_w > 0 || pad_h > 0)
+    if (pad_left > 0 || pad_right > 0 || pad_top > 0 || pad_bottom > 0)
     {
         Option opt_b = opt;
         opt_b.blob_allocator = opt.workspace_allocator;
-        copy_make_border(bottom_blob_unbordered, bottom_blob_bordered, pad_h, pad_h, pad_w, pad_w, BORDER_CONSTANT, 0.f, opt_b);
+        copy_make_border(bottom_blob_unbordered, bottom_blob_bordered, pad_top, pad_bottom, pad_left, pad_right, BORDER_CONSTANT, 0.f, opt_b);
         if (bottom_blob_bordered.empty())
             return -100;
 
         w = bottom_blob_bordered.w;
         h = bottom_blob_bordered.h;
     }
-    else if (pad_w == -233 && pad_h == -233)
+    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;

src/layer/arm/deconvolution_arm.cpp

@@ -299,7 +299,7 @@ int Deconvolution_arm::forward(const Mat& bottom_blob, Mat& top_blob, const Opti
     size_t out_elemsize = elemsize / elempack * out_elempack;
 
     Mat top_blob_bordered;
-    if (pad_w > 0 || pad_h > 0)
+    if (pad_left > 0 || pad_right > 0 || pad_top > 0 || pad_bottom > 0)
     {
         top_blob_bordered.create(outw, outh, num_output / out_elempack, out_elemsize, out_elempack, opt.workspace_allocator);
         if (top_blob_bordered.empty())
@@ -629,9 +629,9 @@ int Deconvolution_arm::forward(const Mat& bottom_blob, Mat& top_blob, const Opti
         }
     }
 
-    if (pad_w > 0 || pad_h > 0)
+    if (pad_left > 0 || pad_right > 0 || pad_top > 0 || pad_bottom > 0)
     {
-        copy_cut_border(top_blob_bordered, top_blob, pad_h, pad_h, pad_w, pad_w, opt);
+        copy_cut_border(top_blob_bordered, top_blob, pad_top, pad_bottom, pad_left, pad_right, opt);
         if (top_blob.empty())
             return -100;
 
@@ -690,7 +690,7 @@ int Deconvolution_arm::forward(const Mat& bottom_blob, Mat& top_blob, const Opti
     int outh = (h - 1) * stride + kernel_size;
 
     Mat top_blob_bordered;
-    if (pad_w > 0 || pad_h > 0)
+    if (pad_left > 0 || pad_right > 0 || pad_top > 0 || pad_bottom > 0)
     {
         top_blob_bordered.create(outw, outh, num_output, elemsize, opt.workspace_allocator);
         if (top_blob_bordered.empty())
@@ -706,9 +706,9 @@ int Deconvolution_arm::forward(const Mat& bottom_blob, Mat& top_blob, const Opti
 
     deconv(bottom_blob, top_blob_bordered, weight_data, bias_data, opt);
 
-    if (pad_w > 0 || pad_h > 0)
+    if (pad_left > 0 || pad_right > 0 || pad_top > 0 || pad_bottom > 0)
     {
-        copy_cut_border(top_blob_bordered, top_blob, pad_h, pad_h, pad_w, pad_w, opt);
+        copy_cut_border(top_blob_bordered, top_blob, pad_top, pad_bottom, pad_left, pad_right, opt);
         if (top_blob.empty())
             return -100;
 

src/layer/arm/deconvolutiondepthwise_arm.cpp

@@ -397,7 +397,7 @@ int DeconvolutionDepthWise_arm::forward(const Mat& bottom_blob, Mat& top_blob, c
     {
 
     Mat top_blob_bordered;
-    if (pad_w > 0 || pad_h > 0)
+    if (pad_left > 0 || pad_right > 0 || pad_top > 0 || pad_bottom > 0)
     {
         top_blob_bordered.create(outw, outh, num_output / out_elempack, out_elemsize, out_elempack, opt.workspace_allocator);
         if (top_blob_bordered.empty())
@@ -507,9 +507,9 @@ int DeconvolutionDepthWise_arm::forward(const Mat& bottom_blob, Mat& top_blob, c
             }
         }
 
-        if (pad_w > 0 || pad_h > 0)
+        if (pad_left > 0 || pad_right > 0 || pad_top > 0 || pad_bottom > 0)
         {
-            copy_cut_border(top_blob_bordered, top_blob, pad_h, pad_h, pad_w, pad_w, opt);
+            copy_cut_border(top_blob_bordered, top_blob, pad_top, pad_bottom, pad_left, pad_right, opt);
             if (top_blob.empty())
                 return -100;
 
@@ -888,9 +888,9 @@ int DeconvolutionDepthWise_arm::forward(const Mat& bottom_blob, Mat& top_blob, c
         top_blob_bordered = top_blob_bordered_unpacked;
     }
 
-    if (pad_w > 0 || pad_h > 0)
+    if (pad_left > 0 || pad_right > 0 || pad_top > 0 || pad_bottom > 0)
     {
-        copy_cut_border(top_blob_bordered, top_blob, pad_h, pad_h, pad_w, pad_w, opt);
+        copy_cut_border(top_blob_bordered, top_blob, pad_top, pad_bottom, pad_left, pad_right, opt);
         if (top_blob.empty())
             return -100;
 
@@ -908,7 +908,7 @@ int DeconvolutionDepthWise_arm::forward(const Mat& bottom_blob, Mat& top_blob, c
 #endif // __ARM_NEON
 
     Mat top_blob_bordered;
-    if (pad_w > 0 || pad_h > 0)
+    if (pad_left > 0 || pad_right > 0 || pad_top > 0 || pad_bottom > 0)
     {
         top_blob_bordered.create(outw, outh, num_output, elemsize, opt.workspace_allocator);
         if (top_blob_bordered.empty())
@@ -961,9 +961,9 @@ int DeconvolutionDepthWise_arm::forward(const Mat& bottom_blob, Mat& top_blob, c
         }
     }
 
-    if (pad_w > 0 || pad_h > 0)
+    if (pad_left > 0 || pad_right > 0 || pad_top > 0 || pad_bottom > 0)
     {
-        copy_cut_border(top_blob_bordered, top_blob, pad_h, pad_h, pad_w, pad_w, opt);
+        copy_cut_border(top_blob_bordered, top_blob, pad_top, pad_bottom, pad_left, pad_right, opt);
         if (top_blob.empty())
             return -100;
 

src/layer/convolution.cpp

@@ -38,14 +38,16 @@ int Convolution::load_param(const ParamDict& pd)
     dilation_h = pd.get(12, dilation_w);
     stride_w = pd.get(3, 1);
     stride_h = pd.get(13, stride_w);
-    pad_w = pd.get(4, 0);
-    pad_h = pd.get(14, pad_w);
+    pad_left = pd.get(4, 0);
+    pad_right = pd.get(15, pad_left);
+    pad_top = pd.get(14, pad_left);
+    pad_bottom = pd.get(16, pad_top);
     bias_term = pd.get(5, 0);
     weight_data_size = pd.get(6, 0);
     int8_scale_term = pd.get(8, 0);
     activation_type = pd.get(9, 0);
     activation_params = pd.get(10, Mat());
-    impl_type = pd.get(15, 0);
+    impl_type = pd.get(17, 0);
 
     return 0;
 }
@@ -329,18 +331,18 @@ int Convolution::forward(const Mat& bottom_blob, Mat& top_blob, const Option& op
     }
 
     Mat bottom_blob_bordered = bottom_blob_unbordered;
-    if (pad_w > 0 || pad_h > 0)
+    if (pad_left > 0 || pad_right > 0 || pad_top > 0 || pad_bottom > 0)
     {
         Option opt_b = opt;
         opt_b.blob_allocator = opt.workspace_allocator;
-        copy_make_border(bottom_blob_unbordered, bottom_blob_bordered, pad_h, pad_h, pad_w, pad_w, BORDER_CONSTANT, 0.f, opt_b);
+        copy_make_border(bottom_blob_unbordered, bottom_blob_bordered, pad_top, pad_bottom, pad_left, pad_right, BORDER_CONSTANT, 0.f, opt_b);
         if (bottom_blob_bordered.empty())
             return -100;
 
         w = bottom_blob_bordered.w;
         h = bottom_blob_bordered.h;
     }
-    else if (pad_w == -233 && pad_h == -233)
+    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;

src/layer/convolution.h

@@ -44,8 +44,10 @@ public:
     int dilation_h;
     int stride_w;
     int stride_h;
-    int pad_w;
-    int pad_h;
+    int pad_left;
+    int pad_right;
+    int pad_top;
+    int pad_bottom;
     int bias_term;
 
     int weight_data_size;

src/layer/convolutiondepthwise.cpp

@@ -37,8 +37,10 @@ int ConvolutionDepthWise::load_param(const ParamDict& pd)
     dilation_h = pd.get(12, dilation_w);
     stride_w = pd.get(3, 1);
     stride_h = pd.get(13, stride_w);
-    pad_w = pd.get(4, 0);
-    pad_h = pd.get(14, pad_w);
+    pad_left = pd.get(4, 0);
+    pad_right = pd.get(15, pad_left);
+    pad_top = pd.get(14, pad_left);
+    pad_bottom = pd.get(16, pad_top);
     bias_term = pd.get(5, 0);
     weight_data_size = pd.get(6, 0);
     group = pd.get(7, 1);
@@ -322,18 +324,18 @@ int ConvolutionDepthWise::forward(const Mat& bottom_blob, Mat& top_blob, const O
     }
 
     Mat bottom_blob_bordered = bottom_blob_unbordered;
-    if (pad_w > 0 || pad_h > 0)
+    if (pad_left > 0 || pad_right > 0 || pad_top > 0 || pad_bottom > 0)
     {
         Option opt_b = opt;
         opt_b.blob_allocator = opt.workspace_allocator;
-        copy_make_border(bottom_blob_unbordered, bottom_blob_bordered, pad_h, pad_h, pad_w, pad_w, BORDER_CONSTANT, 0.f, opt_b);
+        copy_make_border(bottom_blob_unbordered, bottom_blob_bordered, pad_top, pad_bottom, pad_left, pad_right, BORDER_CONSTANT, 0.f, opt_b);
         if (bottom_blob_bordered.empty())
             return -100;
 
         w = bottom_blob_bordered.w;
         h = bottom_blob_bordered.h;
     }
-    else if (pad_w == -233 && pad_h == -233)
+    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;

src/layer/convolutiondepthwise.h

@@ -44,8 +44,10 @@ public:
     int dilation_h;
     int stride_w;
     int stride_h;
-    int pad_w;
-    int pad_h;
+    int pad_left;
+    int pad_right;
+    int pad_top;
+    int pad_bottom;
     int bias_term;
 
     int weight_data_size;

src/layer/deconvolution.cpp

@@ -35,8 +35,10 @@ int Deconvolution::load_param(const ParamDict& pd)
     dilation_h = pd.get(12, dilation_w);
     stride_w = pd.get(3, 1);
     stride_h = pd.get(13, stride_w);
-    pad_w = pd.get(4, 0);
-    pad_h = pd.get(14, pad_w);
+    pad_left = pd.get(4, 0);
+    pad_right = pd.get(15, pad_left);
+    pad_top = pd.get(14, pad_left);
+    pad_bottom = pd.get(16, pad_top);
     bias_term = pd.get(5, 0);
     weight_data_size = pd.get(6, 0);
     activation_type = pd.get(9, 0);
@@ -80,7 +82,7 @@ int Deconvolution::forward(const Mat& bottom_blob, Mat& top_blob, const Option&
     int outh = (h - 1) * stride_h + kernel_extent_h;
 
     Mat top_blob_bordered;
-    if (pad_w > 0 || pad_h > 0)
+    if (pad_left > 0 || pad_right > 0 || pad_top > 0 || pad_bottom > 0)
     {
         top_blob_bordered.create(outw, outh, num_output, elemsize, opt.workspace_allocator);
         if (top_blob_bordered.empty())
@@ -198,9 +200,9 @@ int Deconvolution::forward(const Mat& bottom_blob, Mat& top_blob, const Option&
         }
     }
 
-    if (pad_w > 0 || pad_h > 0)
+    if (pad_left > 0 || pad_right > 0 || pad_top > 0 || pad_bottom > 0)
     {
-        copy_cut_border(top_blob_bordered, top_blob, pad_h, pad_h, pad_w, pad_w, opt);
+        copy_cut_border(top_blob_bordered, top_blob, pad_top, pad_bottom, pad_left, pad_right, opt);
         if (top_blob.empty())
             return -100;
 

src/layer/deconvolution.h

@@ -39,8 +39,10 @@ public:
     int dilation_h;
     int stride_w;
     int stride_h;
-    int pad_w;
-    int pad_h;
+    int pad_left;
+    int pad_right;
+    int pad_top;
+    int pad_bottom;
     int bias_term;
 
     int weight_data_size;

src/layer/deconvolutiondepthwise.cpp

@@ -35,8 +35,10 @@ int DeconvolutionDepthWise::load_param(const ParamDict& pd)
     dilation_h = pd.get(12, dilation_w);
     stride_w = pd.get(3, 1);
     stride_h = pd.get(13, stride_w);
-    pad_w = pd.get(4, 0);
-    pad_h = pd.get(14, pad_w);
+    pad_left = pd.get(4, 0);
+    pad_right = pd.get(15, pad_left);
+    pad_top = pd.get(14, pad_left);
+    pad_bottom = pd.get(16, pad_top);
     bias_term = pd.get(5, 0);
     weight_data_size = pd.get(6, 0);
     group = pd.get(7, 1);
@@ -85,7 +87,7 @@ int DeconvolutionDepthWise::forward(const Mat& bottom_blob, Mat& top_blob, const
     int outh = (h - 1) * stride_h + kernel_extent_h;
 
     Mat top_blob_bordered;
-    if (pad_w > 0 || pad_h > 0)
+    if (pad_left > 0 || pad_right > 0 || pad_top > 0 || pad_bottom > 0)
     {
         top_blob_bordered.create(outw, outh, num_output, elemsize, opt.workspace_allocator);
         if (top_blob_bordered.empty())
@@ -294,9 +296,9 @@ int DeconvolutionDepthWise::forward(const Mat& bottom_blob, Mat& top_blob, const
         }
     }
 
-    if (pad_w > 0 || pad_h > 0)
+    if (pad_left > 0 || pad_right > 0 || pad_top > 0 || pad_bottom > 0)
     {
-        copy_cut_border(top_blob_bordered, top_blob, pad_h, pad_h, pad_w, pad_w, opt);
+        copy_cut_border(top_blob_bordered, top_blob, pad_top, pad_bottom, pad_left, pad_right, opt);
         if (top_blob.empty())
             return -100;
 

src/layer/deconvolutiondepthwise.h

@@ -39,8 +39,10 @@ public:
     int dilation_h;
     int stride_w;
     int stride_h;
-    int pad_w;
-    int pad_h;
+    int pad_left;
+    int pad_right;
+    int pad_top;
+    int pad_bottom;
     int bias_term;
 
     int weight_data_size;

src/layer/vulkan/convolution_vulkan.cpp

@@ -55,10 +55,10 @@ int Convolution_vulkan::create_pipeline(const Option& opt)
         padding->vkdev = vkdev;
 
         ncnn::ParamDict pd;
-        pd.set(0, pad_h);
-        pd.set(1, pad_h);
-        pd.set(2, pad_w);
-        pd.set(3, pad_w);
+        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, 0.f);
 
@@ -862,7 +862,7 @@ int Convolution_vulkan::forward(const VkMat& bottom_blob, VkMat& top_blob, VkCom
     const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
 
     VkMat bottom_blob_bordered = bottom_blob;
-    if (pad_w > 0 || pad_h > 0)
+    if (pad_left > 0 || pad_right > 0 || pad_top > 0 || pad_bottom > 0)
     {
         ncnn::Option opt_pad = opt;
         opt_pad.blob_vkallocator = opt.workspace_vkallocator;
@@ -872,7 +872,7 @@ int Convolution_vulkan::forward(const VkMat& bottom_blob, VkMat& top_blob, VkCom
         w = bottom_blob_bordered.w;
         h = bottom_blob_bordered.h;
     }
-    else if (pad_w == -233 && pad_h == -233)
+    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;

src/layer/vulkan/convolutiondepthwise_vulkan.cpp

@@ -44,10 +44,10 @@ int ConvolutionDepthWise_vulkan::create_pipeline(const Option& opt)
         padding->vkdev = vkdev;
 
         ncnn::ParamDict pd;
-        pd.set(0, pad_h);
-        pd.set(1, pad_h);
-        pd.set(2, pad_w);
-        pd.set(3, pad_w);
+        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, 0.f);
 
@@ -467,7 +467,7 @@ int ConvolutionDepthWise_vulkan::forward(const VkMat& bottom_blob, VkMat& top_bl
     const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
 
     VkMat bottom_blob_bordered = bottom_blob;
-    if (pad_w > 0 || pad_h > 0)
+    if (pad_left > 0 || pad_right > 0 || pad_top > 0 || pad_bottom > 0)
     {
         ncnn::Option opt_pad = opt;
         opt_pad.blob_vkallocator = opt.workspace_vkallocator;
@@ -477,7 +477,7 @@ int ConvolutionDepthWise_vulkan::forward(const VkMat& bottom_blob, VkMat& top_bl
         w = bottom_blob_bordered.w;
         h = bottom_blob_bordered.h;
     }
-    else if (pad_w == -233 && pad_h == -233)
+    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;

src/layer/vulkan/deconvolution_vulkan.cpp

@@ -38,8 +38,8 @@ int Deconvolution_vulkan::create_pipeline(const Option& opt)
         crop->vkdev = vkdev;
 
         ncnn::ParamDict pd;
-        pd.set(0, pad_w);
-        pd.set(1, pad_h);
+        pd.set(0, pad_left);
+        pd.set(1, pad_top);
         pd.set(2, 0);
 
         crop->load_param(pd);
@@ -351,7 +351,7 @@ int Deconvolution_vulkan::forward(const VkMat& bottom_blob, VkMat& top_blob, VkC
     }
 
     VkMat top_blob_bordered;
-    if (pad_w > 0 || pad_h > 0)
+    if (pad_left > 0 || pad_right > 0 || pad_top > 0 || pad_bottom > 0)
     {
         top_blob_bordered.create(outw, outh, num_output / out_elempack, out_elemsize, out_elempack, opt.workspace_vkallocator, opt.staging_vkallocator);
         if (top_blob_bordered.empty())
@@ -420,12 +420,12 @@ int Deconvolution_vulkan::forward(const VkMat& bottom_blob, VkMat& top_blob, VkC
 
     cmd.record_pipeline(pipeline, bindings, constants, top_blob_bordered);
 
-    if (pad_w > 0 || pad_h > 0)
+    if (pad_left > 0 || pad_right > 0 || pad_top > 0 || pad_bottom > 0)
     {
         VkMat reference_blob;
         reference_blob.dims = 2;
-        reference_blob.w = top_blob_bordered.w - pad_w - pad_w;
-        reference_blob.h = top_blob_bordered.h - pad_h - pad_h;
+        reference_blob.w = top_blob_bordered.w - pad_left - pad_right;
+        reference_blob.h = top_blob_bordered.h - pad_top - pad_bottom;
 
         std::vector<VkMat> crop_bottom_blobs(2);
         crop_bottom_blobs[0] = top_blob_bordered;

src/layer/vulkan/deconvolutiondepthwise_vulkan.cpp

@@ -44,8 +44,8 @@ int DeconvolutionDepthWise_vulkan::create_pipeline(const Option& opt)
         crop->vkdev = vkdev;
 
         ncnn::ParamDict pd;
-        pd.set(0, pad_w);
-        pd.set(1, pad_h);
+        pd.set(0, pad_left);
+        pd.set(1, pad_top);
         pd.set(2, 0);
 
         crop->load_param(pd);
@@ -495,7 +495,7 @@ int DeconvolutionDepthWise_vulkan::forward(const VkMat& bottom_blob, VkMat& top_
     }
 
     VkMat top_blob_bordered;
-    if (pad_w > 0 || pad_h > 0)
+    if (pad_left > 0 || pad_right > 0 || pad_top > 0 || pad_bottom > 0)
     {
         top_blob_bordered.create(outw, outh, num_output / out_elempack, out_elemsize, out_elempack, opt.workspace_vkallocator, opt.staging_vkallocator);
         if (top_blob_bordered.empty())
@@ -534,12 +534,12 @@ int DeconvolutionDepthWise_vulkan::forward(const VkMat& bottom_blob, VkMat& top_
         // record
         cmd.record_pipeline(pipeline, bindings, constants, top_blob_bordered);
 
-        if (pad_w > 0 || pad_h > 0)
+        if (pad_left > 0 || pad_right > 0 || pad_top > 0 || pad_bottom > 0)
         {
             VkMat reference_blob;
             reference_blob.dims = 2;
-            reference_blob.w = top_blob_bordered.w - pad_w - pad_w;
-            reference_blob.h = top_blob_bordered.h - pad_h - pad_h;
+            reference_blob.w = top_blob_bordered.w - pad_left - pad_right;
+            reference_blob.h = top_blob_bordered.h - pad_top - pad_bottom;
 
             std::vector<VkMat> crop_bottom_blobs(2);
             crop_bottom_blobs[0] = top_blob_bordered;
@@ -646,12 +646,12 @@ int DeconvolutionDepthWise_vulkan::forward(const VkMat& bottom_blob, VkMat& top_
         top_blob_bordered = top_blob_unpacked;
     }
 
-    if (pad_w > 0 || pad_h > 0)
+    if (pad_left > 0 || pad_right > 0 || pad_top > 0 || pad_bottom > 0)
     {
         VkMat reference_blob;
         reference_blob.dims = 2;
-        reference_blob.w = top_blob_bordered.w - pad_w - pad_w;
-        reference_blob.h = top_blob_bordered.h - pad_h - pad_h;
+        reference_blob.w = top_blob_bordered.w - pad_left - pad_right;
+        reference_blob.h = top_blob_bordered.h - pad_top - pad_bottom;
 
         std::vector<VkMat> crop_bottom_blobs(2);
         crop_bottom_blobs[0] = top_blob_bordered;

src/layer/x86/convolution_x86.cpp

@@ -147,18 +147,18 @@ int Convolution_x86::forwardDilation(const Mat& bottom_blob, Mat& top_blob, conv
     const int kernel_extent = dilation * (kernel_size - 1) + 1;
 
     Mat bottom_blob_bordered = bottom_blob;
-    if (pad_w > 0 || pad_h > 0)
+    if (pad_left > 0 || pad_right > 0 || pad_top > 0 || pad_bottom > 0)
     {
         Option opt_b = opt;
         opt_b.blob_allocator = opt.workspace_allocator;
-        copy_make_border(bottom_blob, bottom_blob_bordered, pad_h, pad_h, pad_w, pad_w, BORDER_CONSTANT, 0.f, opt_b);
+        copy_make_border(bottom_blob, bottom_blob_bordered, pad_top, pad_bottom, pad_left, pad_right, BORDER_CONSTANT, 0.f, opt_b);
         if (bottom_blob_bordered.empty())
             return -100;
 
         w = bottom_blob_bordered.w;
         h = bottom_blob_bordered.h;
     }
-    else if (pad_w == -233 && pad_h == -233)
+    else if (pad_left == -233 && pad_right == -233 && pad_top == -233 && pad_bottom == -233)
     {
         int wpad = kernel_extent + (w - 1) / stride * stride - w;
         int hpad = kernel_extent + (h - 1) / stride * stride - h;
@@ -476,18 +476,18 @@ int Convolution_x86::forward(const Mat& bottom_blob, Mat& top_blob, const Option
     }
 
     Mat bottom_blob_bordered = bottom_blob_unbordered;
-    if (pad_w > 0 || pad_h > 0)
+    if (pad_left > 0 || pad_right > 0 || pad_top > 0 || pad_bottom > 0)
     {
         Option opt_b = opt;
         opt_b.blob_allocator = opt.workspace_allocator;
-        copy_make_border(bottom_blob_unbordered, bottom_blob_bordered, pad_h, pad_h, pad_w, pad_w, BORDER_CONSTANT, 0.f, opt_b);
+        copy_make_border(bottom_blob_unbordered, bottom_blob_bordered, pad_top, pad_bottom, pad_left, pad_right, BORDER_CONSTANT, 0.f, opt_b);
         if (bottom_blob_bordered.empty())
             return -100;
 
         w = bottom_blob_bordered.w;
         h = bottom_blob_bordered.h;
     }
-    else if (pad_w == -233 && pad_h == -233)
+    else if (pad_left == -233 && pad_right == -233 && pad_top == -233 && pad_bottom == -233)
     {
         int wpad = kernel_size + (w - 1) / stride * stride - w;
         int hpad = kernel_size + (h - 1) / stride * stride - h;

src/layer/x86/convolutiondepthwise_x86.cpp

@@ -232,18 +232,18 @@ int ConvolutionDepthWise_x86::forward(const Mat& bottom_blob, Mat& top_blob, con
     }     
 
     Mat bottom_blob_bordered = bottom_blob_unbordered;
-    if (pad_w > 0 || pad_h > 0)
+    if (pad_left > 0 || pad_right > 0 || pad_top > 0 || pad_bottom > 0)
     {
         Option opt_b = opt;
         opt_b.blob_allocator = opt.workspace_allocator;
-        copy_make_border(bottom_blob_unbordered, bottom_blob_bordered, pad_h, pad_h, pad_w, pad_w, BORDER_CONSTANT, 0.f, opt_b);
+        copy_make_border(bottom_blob_unbordered, bottom_blob_bordered, pad_top, pad_bottom, pad_left, pad_right, BORDER_CONSTANT, 0.f, opt_b);
         if (bottom_blob_bordered.empty())
             return -100;
 
         w = bottom_blob_bordered.w;
         h = bottom_blob_bordered.h;
     }
-    else if (pad_w == -233 && pad_h == -233)
+    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;

tools/ncnnoptimize.cpp

@@ -1808,14 +1808,16 @@ int NetOptimize::save(const char* parampath, const char* binpath)
             { if (op->dilation_h != op->dilation_w) fprintf(pp, " 12=%d", op->dilation_h); }
             fprintf_param_value(" 3=%d", stride_w)
             { if (op->stride_h != op->stride_w) fprintf(pp, " 13=%d", op->stride_h); }
-            fprintf_param_value(" 4=%d", pad_w)
-            { if (op->pad_h != op->pad_w) fprintf(pp, " 14=%d", op->pad_h); }
+            fprintf_param_value(" 4=%d", pad_left)
+            { if (op->pad_top != op->pad_left) fprintf(pp, " 14=%d", op->pad_top); }
+            { if (op->pad_right != op->pad_left) fprintf(pp, " 15=%d", op->pad_right); }
+            { if (op->pad_bottom != op->pad_top) fprintf(pp, " 16=%d", op->pad_bottom); }
             fprintf_param_value(" 5=%d", bias_term)
             fprintf_param_value(" 6=%d", weight_data_size)
             fprintf_param_value(" 8=%d", int8_scale_term)
             fprintf_param_value(" 9=%d", activation_type)
             { if (!op->activation_params.empty()) fprintf_param_float_array(10, op->activation_params, pp); }
-            fprintf_param_value(" 15=%d", impl_type)
+            fprintf_param_value(" 17=%d", impl_type)
 
             fwrite_weight_tag_data(0, op->weight_data, bp);
             fwrite_weight_data(op->bias_data, bp);
@@ -1832,8 +1834,10 @@ int NetOptimize::save(const char* parampath, const char* binpath)
             { if (op->dilation_h != op->dilation_w) fprintf(pp, " 12=%d", op->dilation_h); }
             fprintf_param_value(" 3=%d", stride_w)
             { if (op->stride_h != op->stride_w) fprintf(pp, " 13=%d", op->stride_h); }
-            fprintf_param_value(" 4=%d", pad_w)
-            { if (op->pad_h != op->pad_w) fprintf(pp, " 14=%d", op->pad_h); }
+            fprintf_param_value(" 4=%d", pad_left)
+            { if (op->pad_top != op->pad_left) fprintf(pp, " 14=%d", op->pad_top); }
+            { if (op->pad_right != op->pad_left) fprintf(pp, " 15=%d", op->pad_right); }
+            { if (op->pad_bottom != op->pad_top) fprintf(pp, " 16=%d", op->pad_bottom); }
             fprintf_param_value(" 5=%d", bias_term)
             fprintf_param_value(" 6=%d", weight_data_size)
             fprintf_param_value(" 7=%d", group)
@@ -1868,8 +1872,10 @@ int NetOptimize::save(const char* parampath, const char* binpath)
             { if (op->dilation_h != op->dilation_w) fprintf(pp, " 12=%d", op->dilation_h); }
             fprintf_param_value(" 3=%d", stride_w)
             { if (op->stride_h != op->stride_w) fprintf(pp, " 13=%d", op->stride_h); }
-            fprintf_param_value(" 4=%d", pad_w)
-            { if (op->pad_h != op->pad_w) fprintf(pp, " 14=%d", op->pad_h); }
+            fprintf_param_value(" 4=%d", pad_left)
+            { if (op->pad_top != op->pad_left) fprintf(pp, " 14=%d", op->pad_top); }
+            { if (op->pad_right != op->pad_left) fprintf(pp, " 15=%d", op->pad_right); }
+            { if (op->pad_bottom != op->pad_top) fprintf(pp, " 16=%d", op->pad_bottom); }
             fprintf_param_value(" 5=%d", bias_term)
             fprintf_param_value(" 6=%d", weight_data_size)
             fprintf_param_value(" 9=%d", activation_type)
@@ -1890,8 +1896,10 @@ int NetOptimize::save(const char* parampath, const char* binpath)
             { if (op->dilation_h != op->dilation_w) fprintf(pp, " 12=%d", op->dilation_h); }
             fprintf_param_value(" 3=%d", stride_w)
             { if (op->stride_h != op->stride_w) fprintf(pp, " 13=%d", op->stride_h); }
-            fprintf_param_value(" 4=%d", pad_w)
-            { if (op->pad_h != op->pad_w) fprintf(pp, " 14=%d", op->pad_h); }
+            fprintf_param_value(" 4=%d", pad_left)
+            { if (op->pad_top != op->pad_left) fprintf(pp, " 14=%d", op->pad_top); }
+            { if (op->pad_right != op->pad_left) fprintf(pp, " 15=%d", op->pad_right); }
+            { if (op->pad_bottom != op->pad_top) fprintf(pp, " 16=%d", op->pad_bottom); }
             fprintf_param_value(" 5=%d", bias_term)
             fprintf_param_value(" 6=%d", weight_data_size)
             fprintf_param_value(" 7=%d", group)