reduce memory usage of weight packing

src/layer/batchnorm.cpp

@@ -147,9 +147,11 @@ int BatchNorm::upload_model(VkTransfer& cmd)
     // pack4
     if (channels % 4 == 0)
     {
+        Mat a_data_pack4;
         convert_packing(a_data, a_data_pack4, 4);
         cmd.record_upload(a_data_pack4, a_data_gpu_pack4);
 
+        Mat b_data_pack4;
         convert_packing(b_data, b_data_pack4, 4);
         cmd.record_upload(b_data_pack4, b_data_gpu_pack4);
     }

src/layer/batchnorm.h

@@ -56,11 +56,8 @@ public:
 #if NCNN_VULKAN
     VkMat a_data_gpu;
     VkMat b_data_gpu;
-
     Pipeline* pipeline_batchnorm;
 
-    Mat a_data_pack4;
-    Mat b_data_pack4;
     VkMat a_data_gpu_pack4;
     VkMat b_data_gpu_pack4;
     Pipeline* pipeline_batchnorm_pack4;

src/layer/convolution.cpp

@@ -446,10 +446,11 @@ int Convolution::upload_model(VkTransfer& cmd)
     {
         // src = kw-kh-inch-outch
         // dst = 4a-4b-kw-kh-inch/4a-outch/4b
+        Mat weight_data_pack4;
         {
             Mat weight_data_r2 = weight_data.reshape(maxk, num_input, num_output);
 
-            weight_data_pack4 = Mat(16*maxk, num_input/4, num_output/4);
+            weight_data_pack4.create(16*maxk, num_input/4, num_output/4);
 
             for (int q=0; q+3<num_output; q+=4)
             {
@@ -521,10 +522,11 @@ int Convolution::upload_model(VkTransfer& cmd)
     {
         // src = kw-kh-inch-outch
         // dst = 4b-kw-kh-inch-outch/4b
+        Mat weight_data_pack1to4;
         {
             Mat weight_data_r2 = weight_data.reshape(maxk, num_input, num_output);
 
-            weight_data_pack1to4 = Mat(4*maxk, num_input, num_output/4);
+            weight_data_pack1to4.create(4*maxk, num_input, num_output/4);
 
             for (int q=0; q+3<num_output; q+=4)
             {
@@ -566,10 +568,11 @@ int Convolution::upload_model(VkTransfer& cmd)
     {
         // src = kw-kh-inch-outch
         // dst = 4a-kw-kh-inch/4a-outch
+        Mat weight_data_pack4to1;
         {
             Mat weight_data_r2 = weight_data.reshape(maxk, num_input, num_output);
 
-            weight_data_pack4to1 = Mat(4*maxk, num_input/4, num_output);
+            weight_data_pack4to1.create(4*maxk, num_input/4, num_output);
 
             for (int q=0; q<num_output; q++)
             {
@@ -606,6 +609,7 @@ int Convolution::upload_model(VkTransfer& cmd)
     {
         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);
         }

src/layer/convolution.h

@@ -71,21 +71,17 @@ public:
     Pipeline* pipeline_convolution;
     Pipeline* pipeline_convolution_1x1s1d1;
 
-    Mat bias_data_pack4;
     VkMat bias_data_gpu_pack4;
 
     // pack4
-    Mat weight_data_pack4;
     VkMat weight_data_gpu_pack4;
     Pipeline* pipeline_convolution_pack4;
 
     // pack1to4
-    Mat weight_data_pack1to4;
     VkMat weight_data_gpu_pack1to4;
     Pipeline* pipeline_convolution_pack1to4;
 
     // pack4to1
-    Mat weight_data_pack4to1;
     VkMat weight_data_gpu_pack4to1;
     Pipeline* pipeline_convolution_pack4to1;
 #endif // NCNN_VULKAN

src/layer/convolutiondepthwise.cpp

@@ -606,6 +606,7 @@ int ConvolutionDepthWise::upload_model(VkTransfer& cmd)
         {
             const int maxk = kernel_w * kernel_h;
 
+            Mat weight_data_pack4;
             Mat weight_data_r2 = weight_data.reshape(maxk, group);
             convert_packing(weight_data_r2, weight_data_pack4, 4);
 
@@ -614,6 +615,7 @@ int ConvolutionDepthWise::upload_model(VkTransfer& cmd)
 
             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);
             }

src/layer/convolutiondepthwise.h

@@ -72,8 +72,6 @@ public:
 
     Pipeline* pipeline_convolutiondepthwise;
 
-    Mat weight_data_pack4;
-    Mat bias_data_pack4;
     VkMat weight_data_gpu_pack4;
     VkMat bias_data_gpu_pack4;
     Pipeline* pipeline_convolutiondepthwise_pack4;

src/layer/innerproduct.cpp

@@ -248,10 +248,11 @@ int InnerProduct::upload_model(VkTransfer& cmd)
     {
         // src = inch-outch
         // dst = 4a-4b-inch/4a-outch/4b
+        Mat weight_data_pack4;
         {
             Mat weight_data_r2 = weight_data.reshape(num_input, num_output);
 
-            weight_data_pack4 = Mat(16, num_input/4, num_output/4);
+            weight_data_pack4.create(16, num_input/4, num_output/4);
 
             for (int q=0; q+3<num_output; q+=4)
             {
@@ -302,10 +303,11 @@ int InnerProduct::upload_model(VkTransfer& cmd)
     {
         // src = inch-outch
         // dst = 4b-inch-outch/4b
+        Mat weight_data_pack1to4;
         {
             Mat weight_data_r2 = weight_data.reshape(num_input, num_output);
 
-            weight_data_pack1to4 = Mat(4, num_input, num_output/4);
+            weight_data_pack1to4.create(4, num_input, num_output/4);
 
             for (int q=0; q+3<num_output; q+=4)
             {
@@ -337,10 +339,11 @@ int InnerProduct::upload_model(VkTransfer& cmd)
     {
         // src = inch-outch
         // dst = 4a-inch/4a-outch
+        Mat weight_data_pack4to1;
         {
             Mat weight_data_r2 = weight_data.reshape(num_input, num_output);
 
-            weight_data_pack4to1 = Mat(4, num_input/4, num_output);
+            weight_data_pack4to1.create(4, num_input/4, num_output);
 
             for (int q=0; q<num_output; q++)
             {
@@ -369,6 +372,7 @@ int InnerProduct::upload_model(VkTransfer& cmd)
     {
         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);
         }

src/layer/innerproduct.h

@@ -61,21 +61,17 @@ public:
 
     Pipeline* pipeline_innerproduct;
 
-    Mat bias_data_pack4;
     VkMat bias_data_gpu_pack4;
 
     // pack4
-    Mat weight_data_pack4;
     VkMat weight_data_gpu_pack4;
     Pipeline* pipeline_innerproduct_pack4;
 
     // pack1to4
-    Mat weight_data_pack1to4;
     VkMat weight_data_gpu_pack1to4;
     Pipeline* pipeline_innerproduct_pack1to4;
 
     // pack4to1
-    Mat weight_data_pack4to1;
     VkMat weight_data_gpu_pack4to1;
     Pipeline* pipeline_innerproduct_pack4to1;
 #endif // NCNN_VULKAN

src/layer/scale.cpp

@@ -188,6 +188,7 @@ int Scale::upload_model(VkTransfer& cmd)
 
         // pack4
         {
+            Mat scale_data_pack4;
             convert_packing(scale_data, scale_data_pack4, 4);
             cmd.record_upload(scale_data_pack4, scale_data_gpu_pack4);
         }
@@ -199,6 +200,7 @@ int Scale::upload_model(VkTransfer& cmd)
 
         // pack4
         {
+            Mat bias_data_pack4;
             convert_packing(bias_data, bias_data_pack4, 4);
             cmd.record_upload(bias_data_pack4, bias_data_gpu_pack4);
         }

src/layer/scale.h

@@ -53,11 +53,8 @@ public:
 #if NCNN_VULKAN
     VkMat scale_data_gpu;
     VkMat bias_data_gpu;
-
     Pipeline* pipeline_scale;
 
-    Mat scale_data_pack4;
-    Mat bias_data_pack4;
     VkMat scale_data_gpu_pack4;
     VkMat bias_data_gpu_pack4;
     Pipeline* pipeline_scale_pack4;