Merge pull request opencv#10456 from dkurt:dnn_allocate_mem_for_optimized_concat

alalek · alalek · commit 8e7af7f08938 · 2017-12-28T16:04:51.000Z
diff --git a/modules/dnn/CMakeLists.txt b/modules/dnn/CMakeLists.txt
@@ -97,3 +97,8 @@ if(BUILD_PERF_TESTS)
     endif()
   endif()
 endif()
+
+ocv_option(${the_module}_REUSE_MEMORY "Enable reusing strategy of memory management" ON)
+if (${the_module}_REUSE_MEMORY)
+  add_definitions(-DREUSE_DNN_MEMORY=1)
+endif()
diff --git a/modules/dnn/src/dnn.cpp b/modules/dnn/src/dnn.cpp
@@ -367,43 +367,42 @@ struct BlobManager
         }
     }
 
-    void reuseOrCreate(const MatShape& shape, const LayerPin& lp, Mat& dst, bool force)
+    void reuseOrCreate(const MatShape& shape, const LayerPin& lp, Mat& dst)
     {
+#ifdef REUSE_DNN_MEMORY
         Mat bestBlob;
         LayerPin bestBlobPin;
 
-        if( !force )
-        {
-            std::map<LayerPin, Mat>::iterator hostIt;
-            std::map<LayerPin, int>::iterator refIt;
+        std::map<LayerPin, Mat>::iterator hostIt;
+        std::map<LayerPin, int>::iterator refIt;
 
-            const int targetTotal = total(shape);
-            int bestBlobTotal = INT_MAX;
+        const int targetTotal = total(shape);
+        int bestBlobTotal = INT_MAX;
 
-            for (hostIt = memHosts.begin(); hostIt != memHosts.end(); ++hostIt)
+        for (hostIt = memHosts.begin(); hostIt != memHosts.end(); ++hostIt)
+        {
+            refIt = refCounter.find(hostIt->first);
+            // Use only blobs that had references before because if not,
+            // it might be used as output.
+            if (refIt != refCounter.end() && refIt->second == 0)
             {
-                refIt = refCounter.find(hostIt->first);
-                // Use only blobs that had references before because if not,
-                // it might be used as output.
-                if (refIt != refCounter.end() && refIt->second == 0)
+                Mat& unusedBlob = hostIt->second;
+                if (unusedBlob.total() >= targetTotal &&
+                    unusedBlob.total() < bestBlobTotal)
                 {
-                    Mat& unusedBlob = hostIt->second;
-                    if (unusedBlob.total() >= targetTotal &&
-                        unusedBlob.total() < bestBlobTotal)
-                    {
-                        bestBlobPin = hostIt->first;
-                        bestBlob = unusedBlob;
-                        bestBlobTotal = unusedBlob.total();
-                    }
+                    bestBlobPin = hostIt->first;
+                    bestBlob = unusedBlob;
+                    bestBlobTotal = unusedBlob.total();
                 }
             }
         }
         if (!bestBlob.empty())
         {
             reuse(bestBlobPin, lp);
-            dst = Mat(shape, CV_32F, bestBlob.data);
+            dst = bestBlob.reshape(1, 1).colRange(0, targetTotal).reshape(1, shape);
         }
         else
+#endif  // REUSE_DNN_MEMORY
         {
             // if dst already has been allocated with total(shape) elements,
             // it won't be recrreated and pointer of dst.data remains the same.
@@ -412,34 +411,32 @@ struct BlobManager
         }
     }
 
-    void reuseOrCreate(const MatShape& shape, const LayerPin& lp, UMat &umat_dst, bool force)
+    void reuseOrCreate(const MatShape& shape, const LayerPin& lp, UMat &umat_dst)
     {
+#ifdef REUSE_DNN_MEMORY
         UMat bestBlob;
         LayerPin bestBlobPin;
 
-        if( !force )
-        {
-            std::map<LayerPin, UMat>::iterator hostIt;
-            std::map<LayerPin, int>::iterator refIt;
+        std::map<LayerPin, UMat>::iterator hostIt;
+        std::map<LayerPin, int>::iterator refIt;
 
-            const int targetTotal = total(shape);
-            int bestBlobTotal = INT_MAX;
+        const int targetTotal = total(shape);
+        int bestBlobTotal = INT_MAX;
 
-            for (hostIt = umat_memHosts.begin(); hostIt != umat_memHosts.end(); ++hostIt)
+        for (hostIt = umat_memHosts.begin(); hostIt != umat_memHosts.end(); ++hostIt)
+        {
+            refIt = refCounter.find(hostIt->first);
+            // Use only blobs that had references before because if not,
+            // it might be used as output.
+            if (refIt != refCounter.end() && refIt->second == 0)
             {
-                refIt = refCounter.find(hostIt->first);
-                // Use only blobs that had references before because if not,
-                // it might be used as output.
-                if (refIt != refCounter.end() && refIt->second == 0)
+                UMat& unusedBlob = hostIt->second;
+                if (unusedBlob.total() >= targetTotal &&
+                    unusedBlob.total() < bestBlobTotal)
                 {
-                    UMat& unusedBlob = hostIt->second;
-                    if (unusedBlob.total() >= targetTotal &&
-                        unusedBlob.total() < bestBlobTotal)
-                    {
-                        bestBlobPin = hostIt->first;
-                        bestBlob = unusedBlob;
-                        bestBlobTotal = unusedBlob.total();
-                    }
+                    bestBlobPin = hostIt->first;
+                    bestBlob = unusedBlob;
+                    bestBlobTotal = unusedBlob.total();
                 }
             }
         }
@@ -449,6 +446,7 @@ struct BlobManager
             umat_dst.create(shape, CV_32F);
         }
         else
+#endif  // REUSE_DNN_MEMORY
         {
             // if dst already has been allocated with total(shape) elements,
             // it won't be recrreated and pointer of dst.data remains the same.
@@ -458,8 +456,7 @@ struct BlobManager
     }
 
     void allocateBlobsForLayer(LayerData &ld, const LayerShapes& layerShapes,
-                               std::vector<LayerPin>& pinsForInternalBlobs,
-                               bool maximizeReuse)
+                               std::vector<LayerPin>& pinsForInternalBlobs)
     {
         CV_TRACE_FUNCTION();
         bool use_umat = (preferableBackend == DNN_BACKEND_DEFAULT &&
@@ -530,7 +527,6 @@ struct BlobManager
         }
 
         std::map<int, std::vector<int> >::reverse_iterator it;
-        bool force = !maximizeReuse && ld.inputBlobsId.size() > 1;
         for(it = idxSizes.rbegin(); it != idxSizes.rend(); it++)
         {
             for(int j = 0; j < it->second.size(); j++)
@@ -539,7 +535,7 @@ struct BlobManager
                 if (total(shapes[index]))
                 {
                     LayerPin blobPin(ld.id, index);
-                    if (index < outShapes.size() && inPlace && !force)
+                    if (index < outShapes.size() && inPlace)
                     {
                         if (use_umat)
                         {
@@ -558,9 +554,9 @@ struct BlobManager
                     else
                     {
                         if (use_umat)
-                            reuseOrCreate(shapes[index], blobPin, *umat_blobs[index], force);
+                            reuseOrCreate(shapes[index], blobPin, *umat_blobs[index]);
                         else
-                            reuseOrCreate(shapes[index], blobPin, *blobs[index], force);
+                            reuseOrCreate(shapes[index], blobPin, *blobs[index]);
                     }
                 }
             }
@@ -1111,8 +1107,7 @@ struct Net::Impl
         CV_Assert(layerShapesIt != layersShapes.end());
 
         std::vector<LayerPin> pinsForInternalBlobs;
-        bool maximizeReuse = preferableBackend == DNN_BACKEND_HALIDE;
-        blobManager.allocateBlobsForLayer(ld, layerShapesIt->second, pinsForInternalBlobs, maximizeReuse);
+        blobManager.allocateBlobsForLayer(ld, layerShapesIt->second, pinsForInternalBlobs);
         ld.outputBlobsWrappers.resize(ld.outputBlobs.size());
         for (int i = 0; i < ld.outputBlobs.size(); ++i)
         {
@@ -1415,6 +1410,9 @@ struct Net::Impl
 
                     if( i >= ninputs )
                     {
+                        // Allocate new memory to prevent collisions during memory
+                        // reusing (see https://github.com/opencv/opencv/pull/10456).
+                        output = output.clone();
                         Range chrange[] = { Range::all(), Range::all(), Range::all(), Range::all() };
                         int ofs = 0;
                         for( i = 0; i < ninputs; i++ )

Original file line number	Diff line number	Diff line change
`@@ -367,43 +367,42 @@ struct BlobManager`
`367`	`367`	`}`
`368`	`368`	`}`
`369`	`369`
`370`		`- void reuseOrCreate(const MatShape& shape, const LayerPin& lp, Mat& dst, bool force)`
	`370`	`+ void reuseOrCreate(const MatShape& shape, const LayerPin& lp, Mat& dst)`
`371`	`371`	`{`
	`372`	`+#ifdef REUSE_DNN_MEMORY`
`372`	`373`	`Mat bestBlob;`
`373`	`374`	`LayerPin bestBlobPin;`
`374`	`375`
`375`		`- if( !force )`
`376`		`- {`
`377`		`- std::map<LayerPin, Mat>::iterator hostIt;`
`378`		`- std::map<LayerPin, int>::iterator refIt;`
	`376`	`+ std::map<LayerPin, Mat>::iterator hostIt;`
	`377`	`+ std::map<LayerPin, int>::iterator refIt;`
`379`	`378`
`380`		`- const int targetTotal = total(shape);`
`381`		`- int bestBlobTotal = INT_MAX;`
	`379`	`+ const int targetTotal = total(shape);`
	`380`	`+ int bestBlobTotal = INT_MAX;`
`382`	`381`
`383`		`- for (hostIt = memHosts.begin(); hostIt != memHosts.end(); ++hostIt)`
	`382`	`+ for (hostIt = memHosts.begin(); hostIt != memHosts.end(); ++hostIt)`
	`383`	`+ {`
	`384`	`+ refIt = refCounter.find(hostIt->first);`
	`385`	`+ // Use only blobs that had references before because if not,`
	`386`	`+ // it might be used as output.`
	`387`	`+ if (refIt != refCounter.end() && refIt->second == 0)`
`384`	`388`	`{`
`385`		`- refIt = refCounter.find(hostIt->first);`
`386`		`- // Use only blobs that had references before because if not,`
`387`		`- // it might be used as output.`
`388`		`- if (refIt != refCounter.end() && refIt->second == 0)`
	`389`	`+ Mat& unusedBlob = hostIt->second;`
	`390`	`+ if (unusedBlob.total() >= targetTotal &&`
	`391`	`+ unusedBlob.total() < bestBlobTotal)`
`389`	`392`	`{`
`390`		`- Mat& unusedBlob = hostIt->second;`
`391`		`- if (unusedBlob.total() >= targetTotal &&`
`392`		`- unusedBlob.total() < bestBlobTotal)`
`393`		`- {`
`394`		`- bestBlobPin = hostIt->first;`
`395`		`- bestBlob = unusedBlob;`
`396`		`- bestBlobTotal = unusedBlob.total();`
`397`		`- }`
	`393`	`+ bestBlobPin = hostIt->first;`
	`394`	`+ bestBlob = unusedBlob;`
	`395`	`+ bestBlobTotal = unusedBlob.total();`
`398`	`396`	`}`
`399`	`397`	`}`
`400`	`398`	`}`
`401`	`399`	`if (!bestBlob.empty())`
`402`	`400`	`{`
`403`	`401`	`reuse(bestBlobPin, lp);`
`404`		`- dst = Mat(shape, CV_32F, bestBlob.data);`
	`402`	`+ dst = bestBlob.reshape(1, 1).colRange(0, targetTotal).reshape(1, shape);`
`405`	`403`	`}`
`406`	`404`	`else`
	`405`	`+#endif // REUSE_DNN_MEMORY`
`407`	`406`	`{`
`408`	`407`	`// if dst already has been allocated with total(shape) elements,`
`409`	`408`	`// it won't be recrreated and pointer of dst.data remains the same.`
`@@ -412,34 +411,32 @@ struct BlobManager`
`412`	`411`	`}`
`413`	`412`	`}`
`414`	`413`
`415`		`- void reuseOrCreate(const MatShape& shape, const LayerPin& lp, UMat &umat_dst, bool force)`
	`414`	`+ void reuseOrCreate(const MatShape& shape, const LayerPin& lp, UMat &umat_dst)`
`416`	`415`	`{`
	`416`	`+#ifdef REUSE_DNN_MEMORY`
`417`	`417`	`UMat bestBlob;`
`418`	`418`	`LayerPin bestBlobPin;`
`419`	`419`
`420`		`- if( !force )`
`421`		`- {`
`422`		`- std::map<LayerPin, UMat>::iterator hostIt;`
`423`		`- std::map<LayerPin, int>::iterator refIt;`
	`420`	`+ std::map<LayerPin, UMat>::iterator hostIt;`
	`421`	`+ std::map<LayerPin, int>::iterator refIt;`
`424`	`422`
`425`		`- const int targetTotal = total(shape);`
`426`		`- int bestBlobTotal = INT_MAX;`
	`423`	`+ const int targetTotal = total(shape);`
	`424`	`+ int bestBlobTotal = INT_MAX;`
`427`	`425`
`428`		`- for (hostIt = umat_memHosts.begin(); hostIt != umat_memHosts.end(); ++hostIt)`
	`426`	`+ for (hostIt = umat_memHosts.begin(); hostIt != umat_memHosts.end(); ++hostIt)`
	`427`	`+ {`
	`428`	`+ refIt = refCounter.find(hostIt->first);`
	`429`	`+ // Use only blobs that had references before because if not,`
	`430`	`+ // it might be used as output.`
	`431`	`+ if (refIt != refCounter.end() && refIt->second == 0)`
`429`	`432`	`{`
`430`		`- refIt = refCounter.find(hostIt->first);`
`431`		`- // Use only blobs that had references before because if not,`
`432`		`- // it might be used as output.`
`433`		`- if (refIt != refCounter.end() && refIt->second == 0)`
	`433`	`+ UMat& unusedBlob = hostIt->second;`
	`434`	`+ if (unusedBlob.total() >= targetTotal &&`
	`435`	`+ unusedBlob.total() < bestBlobTotal)`
`434`	`436`	`{`
`435`		`- UMat& unusedBlob = hostIt->second;`
`436`		`- if (unusedBlob.total() >= targetTotal &&`
`437`		`- unusedBlob.total() < bestBlobTotal)`
`438`		`- {`
`439`		`- bestBlobPin = hostIt->first;`
`440`		`- bestBlob = unusedBlob;`
`441`		`- bestBlobTotal = unusedBlob.total();`
`442`		`- }`
	`437`	`+ bestBlobPin = hostIt->first;`
	`438`	`+ bestBlob = unusedBlob;`
	`439`	`+ bestBlobTotal = unusedBlob.total();`
`443`	`440`	`}`
`444`	`441`	`}`
`445`	`442`	`}`
`@@ -449,6 +446,7 @@ struct BlobManager`
`449`	`446`	`umat_dst.create(shape, CV_32F);`
`450`	`447`	`}`
`451`	`448`	`else`
	`449`	`+#endif // REUSE_DNN_MEMORY`
`452`	`450`	`{`
`453`	`451`	`// if dst already has been allocated with total(shape) elements,`
`454`	`452`	`// it won't be recrreated and pointer of dst.data remains the same.`
`@@ -458,8 +456,7 @@ struct BlobManager`
`458`	`456`	`}`
`459`	`457`
`460`	`458`	`void allocateBlobsForLayer(LayerData &ld, const LayerShapes& layerShapes,`
`461`		`- std::vector<LayerPin>& pinsForInternalBlobs,`
`462`		`- bool maximizeReuse)`
	`459`	`+ std::vector<LayerPin>& pinsForInternalBlobs)`
`463`	`460`	`{`
`464`	`461`	`CV_TRACE_FUNCTION();`
`465`	`462`	`bool use_umat = (preferableBackend == DNN_BACKEND_DEFAULT &&`
`@@ -530,7 +527,6 @@ struct BlobManager`
`530`	`527`	`}`
`531`	`528`
`532`	`529`	`std::map<int, std::vector<int> >::reverse_iterator it;`
`533`		`- bool force = !maximizeReuse && ld.inputBlobsId.size() > 1;`
`534`	`530`	`for(it = idxSizes.rbegin(); it != idxSizes.rend(); it++)`
`535`	`531`	`{`
`536`	`532`	`for(int j = 0; j < it->second.size(); j++)`
`@@ -539,7 +535,7 @@ struct BlobManager`
`539`	`535`	`if (total(shapes[index]))`
`540`	`536`	`{`
`541`	`537`	`LayerPin blobPin(ld.id, index);`
`542`		`- if (index < outShapes.size() && inPlace && !force)`
	`538`	`+ if (index < outShapes.size() && inPlace)`
`543`	`539`	`{`
`544`	`540`	`if (use_umat)`
`545`	`541`	`{`
`@@ -558,9 +554,9 @@ struct BlobManager`
`558`	`554`	`else`
`559`	`555`	`{`
`560`	`556`	`if (use_umat)`
`561`		`- reuseOrCreate(shapes[index], blobPin, *umat_blobs[index], force);`
	`557`	`+ reuseOrCreate(shapes[index], blobPin, *umat_blobs[index]);`
`562`	`558`	`else`
`563`		`- reuseOrCreate(shapes[index], blobPin, *blobs[index], force);`
	`559`	`+ reuseOrCreate(shapes[index], blobPin, *blobs[index]);`
`564`	`560`	`}`
`565`	`561`	`}`
`566`	`562`	`}`
`@@ -1111,8 +1107,7 @@ struct Net::Impl`
`1111`	`1107`	`CV_Assert(layerShapesIt != layersShapes.end());`
`1112`	`1108`
`1113`	`1109`	`std::vector<LayerPin> pinsForInternalBlobs;`
`1114`		`- bool maximizeReuse = preferableBackend == DNN_BACKEND_HALIDE;`
`1115`		`- blobManager.allocateBlobsForLayer(ld, layerShapesIt->second, pinsForInternalBlobs, maximizeReuse);`
	`1110`	`+ blobManager.allocateBlobsForLayer(ld, layerShapesIt->second, pinsForInternalBlobs);`
`1116`	`1111`	`ld.outputBlobsWrappers.resize(ld.outputBlobs.size());`
`1117`	`1112`	`for (int i = 0; i < ld.outputBlobs.size(); ++i)`
`1118`	`1113`	`{`
`@@ -1415,6 +1410,9 @@ struct Net::Impl`
`1415`	`1410`
`1416`	`1411`	`if( i >= ninputs )`
`1417`	`1412`	`{`
	`1413`	`+ // Allocate new memory to prevent collisions during memory`
	`1414`	`+ // reusing (see https://github.com/opencv/opencv/pull/10456).`
	`1415`	`+ output = output.clone();`
`1418`	`1416`	`Range chrange[] = { Range::all(), Range::all(), Range::all(), Range::all() };`
`1419`	`1417`	`int ofs = 0;`
`1420`	`1418`	`for( i = 0; i < ninputs; i++ )`