diff --git a/src/command.cpp b/src/command.cpp
index ba1d7a074..3aacef64f 100644
--- a/src/command.cpp
+++ b/src/command.cpp
@@ -183,6 +183,9 @@ VkCompute::~VkCompute()
 
 void VkCompute::record_upload(const VkMat& m)
 {
+    if (m.allocator->mappable)
+        return;
+
     record_prepare_transfer_barrier(m);
 
     if (vkdev->info.support_VK_KHR_push_descriptor)
@@ -200,6 +203,9 @@ void VkCompute::record_upload(const VkMat& m)
 
 void VkCompute::record_download(const VkMat& m)
 {
+    if (m.allocator->mappable)
+        return;
+
     record_prepare_transfer_barrier(m);
 
     if (vkdev->info.support_VK_KHR_push_descriptor)
diff --git a/src/net.cpp b/src/net.cpp
index d46237645..5da55951c 100644
--- a/src/net.cpp
+++ b/src/net.cpp
@@ -1308,17 +1308,10 @@ int Net::forward_layer(int layer_index, std::vector<Mat>& blob_mats, std::vector
                     VkMat& bottom_blob = blob_mats_gpu[bottom_blob_index];
                     bottom_blob.create_like(bottom_blob_cpu_packed, opt.blob_vkallocator, opt.staging_vkallocator);
 
-                    if (!bottom_blob.allocator->mappable)
-                    {
-                        bottom_blob.prepare_staging_buffer();
-                    }
-
+                    bottom_blob.prepare_staging_buffer();
                     bottom_blob.upload(bottom_blob_cpu_packed);
 
-                    if (!bottom_blob.allocator->mappable)
-                    {
-                        cmd.record_upload(bottom_blob);
-                    }
+                    cmd.record_upload(bottom_blob);
 
 //                     fprintf(stderr, "upload %d %d %d %d  %lu %d\n", bottom_blob.total() * bottom_blob.elemsize, bottom_blob.w, bottom_blob.h, bottom_blob.c, bottom_blob.elemsize, bottom_blob.packing);
                 }
@@ -1415,17 +1408,10 @@ int Net::forward_layer(int layer_index, std::vector<Mat>& blob_mats, std::vector
                         VkMat& bottom_blob = blob_mats_gpu[bottom_blob_index];
                         bottom_blob.create_like(bottom_blob_cpu_packed, opt.blob_vkallocator, opt.staging_vkallocator);
 
-                        if (!bottom_blob.allocator->mappable)
-                        {
-                            bottom_blob.prepare_staging_buffer();
-                        }
-
+                        bottom_blob.prepare_staging_buffer();
                         bottom_blob.upload(bottom_blob_cpu_packed);
 
-                        if (!bottom_blob.allocator->mappable)
-                        {
-                            cmd.record_upload(bottom_blob);
-                        }
+                        cmd.record_upload(bottom_blob);
 
 //                         fprintf(stderr, "upload %d %d %d %d  %lu %d\n", bottom_blob.total() * bottom_blob.elemsize, bottom_blob.w, bottom_blob.h, bottom_blob.c, bottom_blob.elemsize, bottom_blob.packing);
                     }
@@ -1535,11 +1521,8 @@ int Net::forward_layer(int layer_index, std::vector<Mat>& blob_mats, std::vector
                     // download
                     VkMat& bottom_blob = blob_mats_gpu[bottom_blob_index];
 
-                    if (!bottom_blob.allocator->mappable)
-                    {
-                        bottom_blob.prepare_staging_buffer();
-                        cmd.record_download(bottom_blob);
-                    }
+                    bottom_blob.prepare_staging_buffer();
+                    cmd.record_download(bottom_blob);
 
                     cmd.submit();
 
@@ -1551,10 +1534,7 @@ int Net::forward_layer(int layer_index, std::vector<Mat>& blob_mats, std::vector
                     bottom_blob_cpu_packed.create_like(bottom_blob, opt.blob_allocator);
                     bottom_blob.download(bottom_blob_cpu_packed);
 
-                    if (!bottom_blob.allocator->mappable)
-                    {
-                        bottom_blob.discard_staging_buffer();
-                    }
+                    bottom_blob.discard_staging_buffer();
 
                     Mat& bottom_blob_cpu = blob_mats[bottom_blob_index];
                     convert_packing(bottom_blob_cpu_packed, bottom_blob_cpu, 1);
@@ -1625,11 +1605,8 @@ int Net::forward_layer(int layer_index, std::vector<Mat>& blob_mats, std::vector
                         // download
                         VkMat& bottom_blob = blob_mats_gpu[bottom_blob_index];
 
-                        if (!bottom_blob.allocator->mappable)
-                        {
-                            bottom_blob.prepare_staging_buffer();
-                            cmd.record_download(bottom_blob);
-                        }
+                        bottom_blob.prepare_staging_buffer();
+                        cmd.record_download(bottom_blob);
                     }
                 }
             }
@@ -1655,10 +1632,7 @@ int Net::forward_layer(int layer_index, std::vector<Mat>& blob_mats, std::vector
                     bottom_blob_cpu_packed.create_like(bottom_blob, opt.blob_allocator);
                     bottom_blob.download(bottom_blob_cpu_packed);
 
-                    if (!bottom_blob.allocator->mappable)
-                    {
-                        bottom_blob.discard_staging_buffer();
-                    }
+                    bottom_blob.discard_staging_buffer();
 
                     // TODO use gpu packing
                     Mat& bottom_blob_cpu = blob_mats[bottom_blob_index];
@@ -1847,11 +1821,8 @@ int Extractor::extract(int blob_index, Mat& feat)
             if (blob_mats[blob_index].dims == 0 && feat_gpu.dims != 0)
             {
                 // download
-                if (!feat_gpu.allocator->mappable)
-                {
-                    feat_gpu.prepare_staging_buffer();
-                    cmd.record_download(feat_gpu);
-                }
+                feat_gpu.prepare_staging_buffer();
+                cmd.record_download(feat_gpu);
 
                 cmd.submit();
 
@@ -1861,10 +1832,7 @@ int Extractor::extract(int blob_index, Mat& feat)
                 feat_cpu_packed.create_like(feat_gpu, opt.blob_allocator);
                 feat_gpu.download(feat_cpu_packed);
 
-                if (!feat_gpu.allocator->mappable)
-                {
-                    feat_gpu.discard_staging_buffer();
-                }
+                feat_gpu.discard_staging_buffer();
 
                 // TODO use gpu packing
                 Mat& feat_cpu = blob_mats[blob_index];