Chiebot-Mirror
/
opencv
8
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Allocate new memory for optimized concat to prevent collisions.

Browse Source 
Add a flag to disable memory reusing in dnn module.
pull/10456/head
Dmitry Kurtaev 7 years ago
parent 3542c98d45
commit a9807d8f54
2 changed files with 52 additions and 49 deletions
Whitespace
Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL
Split View
Diff Options
Show Stats Download Patch File Download Diff File
  1. 5
      modules/dnn/CMakeLists.txt
  2. 96
      modules/dnn/src/dnn.cpp

5
modules/dnn/CMakeLists.txt
Unescape View File

@ -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()

96
modules/dnn/src/dnn.cpp
Unescape View File

@ -367,43 +367,42 @@ public:
}
}
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 @@ public:
}
}
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 @@ public:
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 @@ public:
}
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 @@ public:
}
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 @@ public:
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 @@ public:
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++ )

Write Preview
Loading…
Cancel
Save