some bugs fix in using AmdFft library

modules/ocl/src/fft.cpp

@@ -45,29 +45,28 @@
 #include <iomanip>
 #include "precomp.hpp"
 
-#ifdef HAVE_CLAMDFFT
-
 using namespace cv;
 using namespace cv::ocl;
 using namespace std;
 
 #if !defined HAVE_OPENCL
-void cv::ocl::dft(const oclMat &src, oclMat &dst, int flags)
+void cv::ocl::dft(const oclMat &, oclMat &, Size , int )
 {
     throw_nogpu();
 }
 #elif !defined HAVE_CLAMDFFT
-void cv::ocl::dft(const oclMat &src, oclMat &dst, int flags)
+void cv::ocl::dft(const oclMat&, oclMat&, Size, int)
 {
     CV_Error(CV_StsNotImplemented, "OpenCL DFT is not implemented");
 }
 #else
-#include <clAmdFft.h>
-
+#include "clAmdFft.h"
 namespace cv
 {
     namespace ocl
     {
+        void fft_setup();
+        void fft_teardown();
         enum FftType
         {
             C2R = 1, // complex to complex
@@ -76,73 +75,94 @@ namespace cv
         };
         struct FftPlan
         {
-            friend void fft_setup();
-            friend void fft_teardown();
-            ~FftPlan();
         protected:
+            clAmdFftPlanHandle plHandle;
+            FftPlan& operator=(const FftPlan&);
+        public:
             FftPlan(Size _dft_size, int _src_step, int _dst_step, int _flags, FftType _type);
+            ~FftPlan();
+            inline clAmdFftPlanHandle getPlanHandle() { return plHandle; }
+
             const Size dft_size;
             const int src_step, dst_step;
             const int flags;
             const FftType type;
-            clAmdFftPlanHandle plHandle;
-            static vector<FftPlan *> planStore;
-            static bool started;
-            static clAmdFftSetupData *setupData;
+        };
+        class PlanCache
+        {
+        protected:
+            PlanCache();
+            ~PlanCache();
+            friend class auto_ptr<PlanCache>;
+            static auto_ptr<PlanCache> planCache;
+
+            bool started;
+            vector<FftPlan *> planStore;
+            clAmdFftSetupData *setupData;
         public:
+            friend void fft_setup();
+            friend void fft_teardown();
+
+            static PlanCache* getPlanCache()
+            {
+                if( NULL == planCache.get())
+                    planCache.reset(new PlanCache());
+                return planCache.get();
+            }
             // return a baked plan->
             // if there is one matched plan, return it
             // if not, bake a new one, put it into the planStore and return it.
-            static clAmdFftPlanHandle getPlan(Size _dft_size, int _src_step, int _dst_step, int _flags, FftType _type);
+            static FftPlan* getPlan(Size _dft_size, int _src_step, int _dst_step, int _flags, FftType _type);
+
+            // remove a single plan from the store
+            // return true if the plan is successfully removed
+            // else
+            static bool removePlan(clAmdFftPlanHandle );
         };
     }
 }
-bool cv::ocl::FftPlan::started = false;
-vector<cv::ocl::FftPlan *> cv::ocl::FftPlan::planStore = vector<cv::ocl::FftPlan *>();
-clAmdFftSetupData *cv::ocl::FftPlan::setupData = 0;
+auto_ptr<PlanCache> PlanCache::planCache;
 
 void cv::ocl::fft_setup()
 {
-    if(FftPlan::started)
+    PlanCache& pCache = *PlanCache::getPlanCache();
+    if(pCache.started)
     {
         return;
     }
-    FftPlan::setupData = new clAmdFftSetupData;
-    openCLSafeCall(clAmdFftInitSetupData( FftPlan::setupData ));
-    FftPlan::started = true;
+    pCache.setupData = new clAmdFftSetupData;
+    openCLSafeCall(clAmdFftInitSetupData( pCache.setupData ));
+    pCache.started = true;
 }
 void cv::ocl::fft_teardown()
 {
-    if(!FftPlan::started)
+    PlanCache& pCache = *PlanCache::getPlanCache();
+    if(!pCache.started)
     {
         return;
     }
-    delete FftPlan::setupData;
-    for(int i = 0; i < FftPlan::planStore.size(); i ++)
+    delete pCache.setupData;
+    for(size_t i = 0; i < pCache.planStore.size(); i ++)
     {
-        delete FftPlan::planStore[i];
+        delete pCache.planStore[i];
     }
-    FftPlan::planStore.clear();
+    pCache.planStore.clear();
     openCLSafeCall( clAmdFftTeardown( ) );
-    FftPlan::started = false;
+    pCache.started = false;
 }
 
 // bake a new plan
 cv::ocl::FftPlan::FftPlan(Size _dft_size, int _src_step, int _dst_step, int _flags, FftType _type)
-    : dft_size(_dft_size), src_step(_src_step), dst_step(_dst_step), flags(_flags), type(_type), plHandle(0)
+    : plHandle(0), dft_size(_dft_size), src_step(_src_step), dst_step(_dst_step), flags(_flags), type(_type)
 {
-    if(!FftPlan::started)
-    {
-        // implicitly do fft setup
     fft_setup();
-    }
 
     bool is_1d_input	= (_dft_size.height == 1);
     int is_row_dft		= flags & DFT_ROWS;
     int is_scaled_dft   = flags & DFT_SCALE;
     int is_inverse		= flags & DFT_INVERSE;
 
-    clAmdFftResultLocation	place;
+    //clAmdFftResultLocation	place;
     clAmdFftLayout			inLayout;
     clAmdFftLayout			outLayout;
     clAmdFftDim				dim = is_1d_input || is_row_dft ? CLFFT_1D : CLFFT_2D;
@@ -150,7 +170,7 @@ cv::ocl::FftPlan::FftPlan(Size _dft_size, int _src_step, int _dst_step, int _fla
     size_t batchSize		 = is_row_dft ? dft_size.height : 1;
     size_t clLengthsIn[ 3 ]  = {1, 1, 1};
     size_t clStridesIn[ 3 ]  = {1, 1, 1};
-    size_t clLengthsOut[ 3 ] = {1, 1, 1};
+    //size_t clLengthsOut[ 3 ] = {1, 1, 1};
     size_t clStridesOut[ 3 ] = {1, 1, 1};
     clLengthsIn[0]			 = dft_size.width;
     clLengthsIn[1]			 = is_row_dft ? 1 : dft_size.height;
@@ -166,14 +186,12 @@ cv::ocl::FftPlan::FftPlan(Size _dft_size, int _src_step, int _dst_step, int _fla
         clStridesOut[1] = clStridesIn[1];
         break;
     case R2C:
-        CV_Assert(!is_row_dft); // this is not supported yet
         inLayout        = CLFFT_REAL;
         outLayout       = CLFFT_HERMITIAN_INTERLEAVED;
         clStridesIn[1]  = src_step / sizeof(float);
         clStridesOut[1] = dst_step / sizeof(std::complex<float>);
         break;
     case C2R:
-        CV_Assert(!is_row_dft); // this is not supported yet
         inLayout        = CLFFT_HERMITIAN_INTERLEAVED;
         outLayout       = CLFFT_REAL;
         clStridesIn[1]  = src_step / sizeof(std::complex<float>);
@@ -197,27 +215,39 @@ cv::ocl::FftPlan::FftPlan(Size _dft_size, int _src_step, int _dst_step, int _fla
 
     openCLSafeCall( clAmdFftSetPlanInStride  ( plHandle, dim, clStridesIn ) );
     openCLSafeCall( clAmdFftSetPlanOutStride ( plHandle, dim, clStridesOut ) );
-    openCLSafeCall( clAmdFftSetPlanDistance  ( plHandle, clStridesIn[ dim ], clStridesIn[ dim ]) );
+    openCLSafeCall( clAmdFftSetPlanDistance  ( plHandle, clStridesIn[ dim ], clStridesOut[ dim ]) );
+
+    float scale_ = is_scaled_dft ? 1.f / _dft_size.area() : 1.f;
+    openCLSafeCall( clAmdFftSetPlanScale  ( plHandle, is_inverse ? CLFFT_BACKWARD : CLFFT_FORWARD, scale_ ) );
+
+    //ready to bake
     openCLSafeCall( clAmdFftBakePlan( plHandle, 1, &(Context::getContext()->impl->clCmdQueue), NULL, NULL ) );
 }
 cv::ocl::FftPlan::~FftPlan()
 {
-    for(int i = 0; i < planStore.size(); i ++)
-    {
-        if(planStore[i]->plHandle == plHandle)
-        {
-            planStore.erase(planStore.begin() + i);
-        }
-    }
     openCLSafeCall( clAmdFftDestroyPlan( &plHandle ) );
 }
 
-clAmdFftPlanHandle cv::ocl::FftPlan::getPlan(Size _dft_size, int _src_step, int _dst_step, int _flags, FftType _type)
+cv::ocl::PlanCache::PlanCache()
+    : started(false),
+      planStore(vector<cv::ocl::FftPlan *>()),
+      setupData(NULL)
+{
+}
+
+cv::ocl::PlanCache::~PlanCache()
+{
+    fft_teardown();
+}
+
+FftPlan* cv::ocl::PlanCache::getPlan(Size _dft_size, int _src_step, int _dst_step, int _flags, FftType _type)
 {
+    PlanCache& pCache = *PlanCache::getPlanCache();
+    vector<FftPlan *>& pStore = pCache.planStore;
     // go through search
-    for(int i = 0; i < planStore.size(); i ++)
+    for(size_t i = 0; i < pStore.size(); i ++)
     {
-        FftPlan *plan = planStore[i];
+        FftPlan *plan = pStore[i];
         if(
             plan->dft_size.width == _dft_size.width &&
             plan->dft_size.height == _dft_size.height &&
@@ -227,13 +257,29 @@ clAmdFftPlanHandle cv::ocl::FftPlan::getPlan(Size _dft_size, int _src_step, int
             plan->type == _type
             )
         {
-            return plan->plHandle;
+            return plan;
         }
     }
     // no baked plan is found
     FftPlan *newPlan = new FftPlan(_dft_size, _src_step, _dst_step, _flags, _type);
-    planStore.push_back(newPlan);
-    return newPlan->plHandle;
+    pStore.push_back(newPlan);
+    return newPlan;
+}
+
+bool cv::ocl::PlanCache::removePlan(clAmdFftPlanHandle plHandle)
+{
+    PlanCache& pCache = *PlanCache::getPlanCache();
+    vector<FftPlan *>& pStore = pCache.planStore;
+    for(size_t i = 0; i < pStore.size(); i ++)
+    {
+        if(pStore[i]->getPlanHandle() == plHandle)
+        {
+            pStore.erase(pStore.begin() + i);
+            delete pStore[i];
+            return true;
+        }
+    }
+    return false;
 }
 
 void cv::ocl::dft(const oclMat &src, oclMat &dst, Size dft_size, int flags)
@@ -245,19 +291,20 @@ void cv::ocl::dft(const oclMat &src, oclMat &dst, Size dft_size, int flags)
     // check if the given dft size is of optimal dft size
     CV_Assert(dft_size.area() == getOptimalDFTSize(dft_size.area()));
 
+    // the two flags are not compatible
+    CV_Assert( !((flags & DFT_SCALE) && (flags & DFT_ROWS)) );
+
     // similar assertions with cuda module
     CV_Assert(src.type() == CV_32F || src.type() == CV_32FC2);
 
-    // we don't support DFT_SCALE flag
-    CV_Assert(!(DFT_SCALE & flags));
-
-    bool is_1d_input	= (src.rows == 1);
-    int is_row_dft		= flags & DFT_ROWS;
-    int is_scaled_dft		= flags & DFT_SCALE;
+    //bool is_1d_input	= (src.rows == 1);
+    //int is_row_dft		= flags & DFT_ROWS;
+    //int is_scaled_dft		= flags & DFT_SCALE;
     int is_inverse			= flags & DFT_INVERSE;
     bool is_complex_input	= src.channels() == 2;
     bool is_complex_output	= !(flags & DFT_REAL_OUTPUT);
 
+
     // We don't support real-to-real transform
     CV_Assert(is_complex_input || is_complex_output);
     FftType type = (FftType)(is_complex_input << 0 | is_complex_output << 1);
@@ -268,12 +315,10 @@ void cv::ocl::dft(const oclMat &src, oclMat &dst, Size dft_size, int flags)
         dst.create(src.rows, src.cols, CV_32FC2);
         break;
     case R2C:
-        CV_Assert(!is_row_dft); // this is not supported yet
         dst.create(src.rows, src.cols / 2 + 1, CV_32FC2);
         break;
     case C2R:
         CV_Assert(dft_size.width / 2 + 1 == src.cols && dft_size.height == src.rows);
-        CV_Assert(!is_row_dft); // this is not supported yet
         dst.create(src.rows, dft_size.width, CV_32FC1);
         break;
     default:
@@ -282,13 +327,14 @@ void cv::ocl::dft(const oclMat &src, oclMat &dst, Size dft_size, int flags)
         throw exception();
         break;
     }
-    clAmdFftPlanHandle plHandle = FftPlan::getPlan(dft_size, src.step, dst.step, flags, type);
+    clAmdFftPlanHandle plHandle = PlanCache::getPlan(dft_size, src.step, dst.step, flags, type)->getPlanHandle();
 
     //get the buffersize
     size_t buffersize = 0;
     openCLSafeCall( clAmdFftGetTmpBufSize(plHandle, &buffersize ) );
 
     //allocate the intermediate buffer
+    // TODO, bind this with the current FftPlan
     cl_mem clMedBuffer = NULL;
     if (buffersize)
     {
@@ -307,7 +353,7 @@ void cv::ocl::dft(const oclMat &src, oclMat &dst, Size dft_size, int flags)
     {
         openCLFree(clMedBuffer);
     }
+    //fft_teardown();
 }
 
 #endif
-#endif //HAVE_CLAMDFFT

modules/ocl/test/test_fft.cpp

@@ -48,50 +48,59 @@ using namespace std;
 #ifdef HAVE_CLAMDFFT
 ////////////////////////////////////////////////////////////////////////////
 // Dft
-PARAM_TEST_CASE(Dft, cv::Size, bool)
+PARAM_TEST_CASE(Dft, cv::Size, int)
 {
     cv::Size dft_size;
-    bool	 dft_rows;
-    //std::vector<cv::ocl::Info> oclinfo;
+    int	 dft_flags;
     virtual void SetUp()
     {
-        //int devnums = getDevice(oclinfo);
-        //   CV_Assert(devnums > 0);
         dft_size  = GET_PARAM(0);
-        dft_rows = GET_PARAM(1);
+        dft_flags = GET_PARAM(1);
     }
 };
 
 TEST_P(Dft, C2C)
 {
-    cv::Mat a = randomMat(dft_size, CV_32FC2, 0.0, 10.0);
+    cv::Mat a = randomMat(dft_size, CV_32FC2, 0.0, 100.0);
     cv::Mat b_gold;
-    int flags = 0;
-    flags |= dft_rows ? cv::DFT_ROWS : 0;
 
     cv::ocl::oclMat d_b;
 
-    cv::dft(a, b_gold, flags);
-    cv::ocl::dft(cv::ocl::oclMat(a), d_b, a.size(), flags);
+    cv::dft(a, b_gold, dft_flags);
+    cv::ocl::dft(cv::ocl::oclMat(a), d_b, a.size(), dft_flags);
     EXPECT_MAT_NEAR(b_gold, cv::Mat(d_b), a.size().area() * 1e-4, "");
 }
 
+TEST_P(Dft, R2C)
+{
+    cv::Mat a = randomMat(dft_size, CV_32FC1, 0.0, 100.0);
+    cv::Mat b_gold, b_gold_roi;
+
+    cv::ocl::oclMat d_b, d_c;
+    cv::ocl::dft(cv::ocl::oclMat(a), d_b, a.size(), dft_flags);
+    cv::dft(a, b_gold, cv::DFT_COMPLEX_OUTPUT | dft_flags);
+
+    b_gold_roi = b_gold(cv::Rect(0, 0, d_b.cols, d_b.rows));
+    EXPECT_MAT_NEAR(b_gold_roi, cv::Mat(d_b), a.size().area() * 1e-4, "");
+
+    cv::Mat c_gold;
+    cv::dft(b_gold, c_gold, cv::DFT_INVERSE | cv::DFT_REAL_OUTPUT | cv::DFT_SCALE);
+    EXPECT_MAT_NEAR(b_gold_roi, cv::Mat(d_b), a.size().area() * 1e-4, "");
+}
 
 TEST_P(Dft, R2CthenC2R)
 {
     cv::Mat a = randomMat(dft_size, CV_32FC1, 0.0, 10.0);
 
-    int flags = 0;
-    //flags |= dft_rows ? cv::DFT_ROWS : 0; // not supported yet
-
     cv::ocl::oclMat d_b, d_c;
-    cv::ocl::dft(cv::ocl::oclMat(a), d_b, a.size(), flags);
-    cv::ocl::dft(d_b, d_c, a.size(), flags + cv::DFT_INVERSE + cv::DFT_REAL_OUTPUT);
+    cv::ocl::dft(cv::ocl::oclMat(a), d_b, a.size(), 0);
+    cv::ocl::dft(d_b, d_c, a.size(), cv::DFT_SCALE | cv::DFT_INVERSE | cv::DFT_REAL_OUTPUT);
     EXPECT_MAT_NEAR(a, d_c, a.size().area() * 1e-4, "");
 }
 
-INSTANTIATE_TEST_CASE_P(ocl_DFT, Dft, testing::Combine(
-                            testing::Values(cv::Size(5, 4), cv::Size(20, 20)),
-                            testing::Values(false, true)));
+
+INSTANTIATE_TEST_CASE_P(OCL_ImgProc, Dft, testing::Combine(
+                            testing::Values(cv::Size(2, 3), cv::Size(5, 4), cv::Size(25, 20), cv::Size(512, 1), cv::Size(1024, 768)),
+                            testing::Values(0, (int)cv::DFT_ROWS, (int)cv::DFT_SCALE) ));
 
 #endif // HAVE_CLAMDFFT