fixed some compile warnings in dnn & protobuf; improved convolution layer…

fixed some compile warnings in dnn & protobuf; improved convolution layer performance when blas is not available by parallelizing gemmCPU() function in dnn

modules/dnn/3rdparty/protobuf/src/google/protobuf/descriptor.cc

@@ -223,8 +223,7 @@ struct PointerStringPairHash {
   }
 
   // Used only by MSVC and platforms where hash_map is not available.
-  static const size_t bucket_size = 4;
-  static const size_t min_buckets = 8;
+  enum { bucket_size = 4, min_buckets = 8 };
   inline bool operator()(const PointerStringPair& a,
                          const PointerStringPair& b) const {
     if (a.first < b.first) return true;

modules/dnn/3rdparty/protobuf/src/google/protobuf/stubs/strutil.cc

@@ -729,13 +729,6 @@ char *FastHex32ToBuffer(uint32 value, char* buffer) {
   return InternalFastHexToBuffer(value, buffer, 8);
 }
 
-static inline char* PlaceNum(char* p, int num, char prev_sep) {
-   *p-- = '0' + num % 10;
-   *p-- = '0' + num / 10;
-   *p-- = prev_sep;
-   return p;
-}
-
 // ----------------------------------------------------------------------
 // FastInt32ToBufferLeft()
 // FastUInt32ToBufferLeft()

modules/dnn/3rdparty/protobuf/src/google/protobuf/wire_format.cc

@@ -53,15 +53,6 @@ namespace google {
 namespace protobuf {
 namespace internal {
 
-namespace {
-
-// This function turns out to be convenient when using some macros later.
-inline int GetEnumNumber(const EnumValueDescriptor* descriptor) {
-  return descriptor->number();
-}
-
-}  // anonymous namespace
-
 // ===================================================================
 
 bool UnknownFieldSetFieldSkipper::SkipField(

modules/dnn/samples/fcn_semsegm.cpp

@@ -140,7 +140,10 @@ int main(int argc, char **argv)
     //! [Set input blob]
 
     //! [Make forward pass]
+    double t = (double)cv::getTickCount();
     net.forward();                          //compute output
+    t = (double)cv::getTickCount() - t;
+    printf("processing time: %.1fms\n", t*1000./getTickFrequency());
     //! [Make forward pass]
 
     //! [Gather output]

modules/dnn/src/caffe/layer_loaders.cpp

@@ -44,7 +44,7 @@ Ptr<Layer> createLayerFromCaffe<DeconvolutionLayer>(LayerParams &params)
 template<>
 Ptr<Layer> createLayerFromCaffe<PoolingLayer>(LayerParams &params)
 {
-    int type;
+    int type = PoolingLayer::MAX;
     Size kernel, stride, pad;
     bool globalPooling;
 
@@ -60,10 +60,6 @@ Ptr<Layer> createLayerFromCaffe<PoolingLayer>(LayerParams &params)
         else
             CV_Error(Error::StsBadArg, "Unknown pooling type \"" + pool + "\"");
     }
-    else
-    {
-        type = PoolingLayer::MAX;
-    }
 
     getPoolingKernelParams(params, kernel.height, kernel.width, globalPooling, pad.height, pad.width, stride.height, stride.width);
     //getCaffeConvParams(params, kernel, pad, stride);
@@ -107,7 +103,7 @@ Ptr<Layer> createLayerFromCaffe<InnerProductLayer>(LayerParams &params)
 template<> //LRNLayer specialization
 Ptr<Layer> createLayerFromCaffe<LRNLayer>(LayerParams& params)
 {
-    int type;
+    int type = -1;
     String nrmType = params.get<String>("norm_region", "ACROSS_CHANNELS");
     if (nrmType == "ACROSS_CHANNELS")
         type = LRNLayer::CHANNEL_NRM;

modules/dnn/src/layers/convolution_layer.cpp

@@ -55,7 +55,7 @@ namespace dnn
 
 ConvolutionLayerImpl::ConvolutionLayerImpl()
 {
-    tryUseOpenCL = true;
+    tryUseOpenCL = false; //true;
     numOutput = -1;
     group = -1;
 

modules/dnn/src/layers/op_blas.cpp

@@ -28,8 +28,85 @@ inline void SwapRowCols(const Mat &A, int &rows, int &cols, bool isTrans)
     cols = (isTrans) ? A.rows : A.cols;
 }
 
+
+class GEMMInvoker : public ParallelLoopBody
+{
+public:
+    GEMMInvoker(const Mat* _a, const Mat* _b, double _alpha, Mat* _c, double _beta)
+    {
+        a = _a;
+        b = _b;
+        c = _c;
+        alpha = _alpha;
+        beta = _beta;
+    }
+
+    void operator()(const Range& range) const
+    {
+        int mmax = a->rows;
+        int nmax = range.end - range.start;
+        int kmax = a->cols;
+        int m, n, k;
+        AutoBuffer<float> buf(nmax);
+        float* ptr = buf;
+        if( mmax %2 != 0 )
+            memset(ptr, 0, nmax*sizeof(ptr[0]));
+
+        for( m = 0; m < mmax; m += 2 )
+        {
+            float* dst0 = c->ptr<float>(m) + range.start;
+            float* dst1 = m+1 < mmax ? c->ptr<float>(m+1) + range.start : ptr;
+            const float* aptr0 = a->ptr<float>(m);
+            const float* aptr1 = m+1 < mmax ? a->ptr<float>(m+1) : aptr0;
+
+            if( beta != 1 )
+            {
+                if( beta == 0 )
+                    for( n = 0; n < nmax; n++ )
+                    {
+                        dst0[n] = 0.f;
+                        dst1[n] = 0.f;
+                    }
+                else
+                    for( n = 0; n < nmax; n++ )
+                    {
+                        dst0[n] *= (float)beta;
+                        dst1[n] *= (float)beta;
+                    }
+            }
+
+            for( k = 0; k < kmax; k++ )
+            {
+                float alpha0 = (float)(alpha*aptr0[k]);
+                float alpha1 = (float)(alpha*aptr1[k]);
+                const float* bptr = b->ptr<float>(k) + range.start;
+
+                for( n = 0; n < nmax; n++ )
+                {
+                    float d0 = dst0[n] + alpha0*bptr[n];
+                    float d1 = dst1[n] + alpha1*bptr[n];
+                    dst0[n] = d0;
+                    dst1[n] = d1;
+                }
+            }
+        }
+    }
+
+    const Mat *a, *b;
+    Mat* c;
+    double alpha, beta;
+};
+
 void gemmCPU(const Mat &A, const Mat &B, double alpha, Mat &C, double beta, int flags /*= 0*/)
 {
+    if( C.type() == CV_32F && flags == 0 )
+    {
+        GEMMInvoker invoker(&A, &B, alpha, &C, beta);
+        double granularity = 10000000./((double)A.rows*A.cols);
+        parallel_for_(Range(0, B.cols), invoker, granularity);
+    }
+    else
+    {
     #if HAVE_CBLAS
     bool transA = static_cast<bool>(flags & GEMM_1_T);
     bool transB = static_cast<bool>(flags & GEMM_2_T);
@@ -70,6 +147,7 @@ void gemmCPU(const Mat &A, const Mat &B, double alpha, Mat &C, double beta, int
     #else
     cv::gemm(A, B, alpha, C, beta, C, flags);
     #endif
+    }
 }
 
 int getBlasThreads()