cuda: starting point for using Stream for thread-safe concurrency whe…

…n using GPU Signed-off-by: deadprogram <[email protected]>
hybridgroup · Aug 12, 2021 · 0385a5b · 0385a5b
1 parent e118065
commit 0385a5b
Show file tree

Hide file tree

Showing 24 changed files with 899 additions and 97 deletions.
diff --git a/Dockerfile.gpu b/Dockerfile.gpu
@@ -1,12 +1,12 @@
 # to build this docker image:
 #   docker build -f Dockerfile.gpu .
-FROM gocv/opencv:4.5.3-gpu AS gocv-gpu
+FROM gocv/opencv:4.5.3-gpu-cuda-11 AS gocv-gpu
 
 ENV GOPATH /go
 
 COPY . /go/src/gocv.io/x/gocv/
 
 WORKDIR /go/src/gocv.io/x/gocv
-RUN go build -tags example -o /build/gocv_cuda_version ./cmd/cuda/
+RUN go build -tags cuda -o /build/gocv_cuda_version ./cmd/cuda/
 
 CMD ["/build/gocv_cuda_version"]
diff --git a/cuda/arithm.cpp b/cuda/arithm.cpp
@@ -2,14 +2,27 @@
 #include "arithm.h"
 #include <string.h>
 
-void GpuAbs(GpuMat src, GpuMat dst) {
-    cv::cuda::abs(*src, *dst);
+void GpuAbs(GpuMat src, GpuMat dst, Stream s) {
+    if (s == NULL) {
+        cv::cuda::abs(*src, *dst);
+        return;
+    }
+    cv::cuda::abs(*src, *dst, *s);
 }
 
-void GpuThreshold(GpuMat src, GpuMat dst, double thresh, double maxval, int typ) {
-    cv::cuda::threshold(*src, *dst, thresh, maxval, typ);
+void GpuThreshold(GpuMat src, GpuMat dst, double thresh, double maxval, int typ, Stream s) {
+    if (s == NULL) {
+        cv::cuda::threshold(*src, *dst, thresh, maxval, typ);
+        return;
+    }
+
+    cv::cuda::threshold(*src, *dst, thresh, maxval, typ, *s);
 }
 
-void GpuFlip(GpuMat src, GpuMat dst, int flipCode) {
-    cv::cuda::flip(*src, *dst, flipCode);
+void GpuFlip(GpuMat src, GpuMat dst, int flipCode, Stream s) {
+    if (s == NULL) {
+        cv::cuda::flip(*src, *dst, flipCode);
+        return;
+    }
+    cv::cuda::flip(*src, *dst, flipCode, *s);
 }
diff --git a/cuda/arithm.go b/cuda/arithm.go
@@ -16,7 +16,17 @@ import "gocv.io/x/gocv"
 // https://docs.opencv.org/master/d8/d34/group__cudaarithm__elem.html#ga54a72bd772494ab34d05406fd76df2b6
 //
 func Abs(src GpuMat, dst *GpuMat) {
-	C.GpuAbs(src.p, dst.p)
+	C.GpuAbs(src.p, dst.p, nil)
+}
+
+// AbsWithStream computes an absolute value of each matrix element
+// using a Stream for concurrency.
+//
+// For further details, please see:
+// https://docs.opencv.org/master/d8/d34/group__cudaarithm__elem.html#ga54a72bd772494ab34d05406fd76df2b6
+//
+func AbsWithStream(src GpuMat, dst *GpuMat, stream Stream) {
+	C.GpuAbs(src.p, dst.p, stream.p)
 }
 
 // Threshold applies a fixed-level threshold to each array element.
@@ -25,7 +35,17 @@ func Abs(src GpuMat, dst *GpuMat) {
 // https://docs.opencv.org/master/d8/d34/group__cudaarithm__elem.html#ga40f1c94ae9a9456df3cad48e3cb008e1
 //
 func Threshold(src GpuMat, dst *GpuMat, thresh, maxval float64, typ gocv.ThresholdType) {
-	C.GpuThreshold(src.p, dst.p, C.double(thresh), C.double(maxval), C.int(typ))
+	C.GpuThreshold(src.p, dst.p, C.double(thresh), C.double(maxval), C.int(typ), nil)
+}
+
+// ThresholdWithStream applies a fixed-level threshold to each array element
+// using a Stream for concurrency.
+//
+// For further details, please see:
+// https://docs.opencv.org/master/d8/d34/group__cudaarithm__elem.html#ga40f1c94ae9a9456df3cad48e3cb008e1
+//
+func ThresholdWithStream(src GpuMat, dst *GpuMat, thresh, maxval float64, typ gocv.ThresholdType, s Stream) {
+	C.GpuThreshold(src.p, dst.p, C.double(thresh), C.double(maxval), C.int(typ), s.p)
 }
 
 // Flip flips a 2D matrix around vertical, horizontal, or both axes.
@@ -34,5 +54,15 @@ func Threshold(src GpuMat, dst *GpuMat, thresh, maxval float64, typ gocv.Thresho
 // https://docs.opencv.org/master/de/d09/group__cudaarithm__core.html#ga4d0a3f2b46e8f0f1ec2b5ac178dcd871
 //
 func Flip(src GpuMat, dst *GpuMat, flipCode int) {
-	C.GpuFlip(src.p, dst.p, C.int(flipCode))
+	C.GpuFlip(src.p, dst.p, C.int(flipCode), nil)
+}
+
+// FlipWithStream flips a 2D matrix around vertical, horizontal, or both axes
+// using a Stream for concurrency.
+//
+// For further details, please see:
+// https://docs.opencv.org/master/de/d09/group__cudaarithm__core.html#ga4d0a3f2b46e8f0f1ec2b5ac178dcd871
+//
+func FlipWithStream(src GpuMat, dst *GpuMat, flipCode int, stream Stream) {
+	C.GpuFlip(src.p, dst.p, C.int(flipCode), stream.p)
 }
diff --git a/cuda/arithm.h b/cuda/arithm.h
@@ -11,9 +11,9 @@ extern "C" {
 #endif
 #include "cuda.h"
 
-void GpuAbs(GpuMat src, GpuMat dst);
-void GpuThreshold(GpuMat src, GpuMat dst, double thresh, double maxval, int typ);
-void GpuFlip(GpuMat src, GpuMat dst, int flipCode);
+void GpuAbs(GpuMat src, GpuMat dst, Stream s);
+void GpuThreshold(GpuMat src, GpuMat dst, double thresh, double maxval, int typ, Stream s);
+void GpuFlip(GpuMat src, GpuMat dst, int flipCode, Stream s);
 
 #ifdef __cplusplus
 }

diff --git a/cuda/arithm_test.go b/cuda/arithm_test.go
@@ -29,6 +29,30 @@ func TestAbs(t *testing.T) {
 	}
 }
 
+func TestAbsWithStream(t *testing.T) {
+	src := gocv.IMRead("../images/gocvlogo.jpg", gocv.IMReadColor)
+	if src.Empty() {
+		t.Error("Invalid read of Mat in Abs test")
+	}
+	defer src.Close()
+
+	var cimg, dimg, s = NewGpuMat(), NewGpuMat(), NewStream()
+	defer cimg.Close()
+	defer dimg.Close()
+	defer s.Close()
+
+	cimg.Upload(src)
+
+	dest := gocv.NewMat()
+	defer dest.Close()
+
+	AbsWithStream(cimg, &dimg, s)
+	dimg.Download(&dest)
+	if dest.Empty() || src.Rows() != dest.Rows() || src.Cols() != dest.Cols() {
+		t.Error("Invalid Abs test")
+	}
+}
+
 func TestThreshold(t *testing.T) {
 	src := gocv.IMRead("../images/gocvlogo.jpg", gocv.IMReadColor)
 	if src.Empty() {
@@ -52,6 +76,30 @@ func TestThreshold(t *testing.T) {
 	}
 }
 
+func TestThresholdWithStream(t *testing.T) {
+	src := gocv.IMRead("../images/gocvlogo.jpg", gocv.IMReadColor)
+	if src.Empty() {
+		t.Error("Invalid read of Mat in Threshold test")
+	}
+	defer src.Close()
+
+	var cimg, dimg, s = NewGpuMat(), NewGpuMat(), NewStream()
+	defer cimg.Close()
+	defer dimg.Close()
+	defer s.Close()
+
+	cimg.Upload(src)
+
+	dest := gocv.NewMat()
+	defer dest.Close()
+
+	ThresholdWithStream(cimg, &dimg, 25, 255, gocv.ThresholdBinary, s)
+	dimg.Download(&dest)
+	if dest.Empty() || src.Rows() != dest.Rows() || src.Cols() != dest.Cols() {
+		t.Error("Invalid Threshold test")
+	}
+}
+
 func TestFlip(t *testing.T) {
 	src := gocv.IMRead("../images/gocvlogo.jpg", gocv.IMReadColor)
 	if src.Empty() {
@@ -74,3 +122,27 @@ func TestFlip(t *testing.T) {
 		t.Error("Invalid Flip test")
 	}
 }
+
+func TestFlipWithStream(t *testing.T) {
+	src := gocv.IMRead("../images/gocvlogo.jpg", gocv.IMReadColor)
+	if src.Empty() {
+		t.Error("Invalid read of Mat in Flip test")
+	}
+	defer src.Close()
+
+	var cimg, dimg, s = NewGpuMat(), NewGpuMat(), NewStream()
+	defer cimg.Close()
+	defer dimg.Close()
+	defer s.Close()
+
+	cimg.Upload(src)
+
+	dest := gocv.NewMat()
+	defer dest.Close()
+
+	FlipWithStream(cimg, &dimg, 0, s)
+	dimg.Download(&dest)
+	if dest.Empty() || src.Rows() != dest.Rows() || src.Cols() != dest.Cols() {
+		t.Error("Invalid Flip test")
+	}
+}
diff --git a/cuda/bgsegm.cpp b/cuda/bgsegm.cpp
@@ -8,8 +8,12 @@ void CudaBackgroundSubtractorMOG2_Close(CudaBackgroundSubtractorMOG2 b) {
     delete b;
 }
 
-void CudaBackgroundSubtractorMOG2_Apply(CudaBackgroundSubtractorMOG2 b, GpuMat src, GpuMat dst) {
-    (*b)->apply(*src, *dst);
+void CudaBackgroundSubtractorMOG2_Apply(CudaBackgroundSubtractorMOG2 b, GpuMat src, GpuMat dst, Stream s) {
+    if (s == NULL) {
+        (*b)->apply(*src, *dst);
+        return;
+    }
+    (*b)->apply(*src, *dst, -1.0, *s);
 }
 
 CudaBackgroundSubtractorMOG CudaBackgroundSubtractorMOG_Create() {
@@ -20,6 +24,10 @@ void CudaBackgroundSubtractorMOG_Close(CudaBackgroundSubtractorMOG b) {
     delete b;
 }
 
-void CudaBackgroundSubtractorMOG_Apply(CudaBackgroundSubtractorMOG b, GpuMat src, GpuMat dst) {
-    (*b)->apply(*src, *dst);
+void CudaBackgroundSubtractorMOG_Apply(CudaBackgroundSubtractorMOG b, GpuMat src, GpuMat dst, Stream s) {
+    if (s == NULL) {
+        (*b)->apply(*src, *dst);
+        return;
+    }
+    (*b)->apply(*src, *dst, -1.0, *s);
 }
diff --git a/cuda/bgsegm.go b/cuda/bgsegm.go
@@ -43,7 +43,18 @@ func (b *BackgroundSubtractorMOG2) Close() error {
 // https://docs.opencv.org/master/df/d23/classcv_1_1cuda_1_1BackgroundSubtractorMOG2.html#a92408f07bf1268c1b778cb186b3113b0
 //
 func (b *BackgroundSubtractorMOG2) Apply(src GpuMat, dst *GpuMat) {
-	C.CudaBackgroundSubtractorMOG2_Apply((C.CudaBackgroundSubtractorMOG2)(b.p), src.p, dst.p)
+	C.CudaBackgroundSubtractorMOG2_Apply((C.CudaBackgroundSubtractorMOG2)(b.p), src.p, dst.p, nil)
+	return
+}
+
+// ApplyWithStream computes a foreground mask using the current BackgroundSubtractorMOG2
+// using a Stream for concurrency.
+//
+// For further details, please see:
+// https://docs.opencv.org/master/df/d23/classcv_1_1cuda_1_1BackgroundSubtractorMOG2.html#a92408f07bf1268c1b778cb186b3113b0
+//
+func (b *BackgroundSubtractorMOG2) ApplyWithStream(src GpuMat, dst *GpuMat, s Stream) {
+	C.CudaBackgroundSubtractorMOG2_Apply((C.CudaBackgroundSubtractorMOG2)(b.p), src.p, dst.p, s.p)
 	return
 }
 
@@ -71,6 +82,17 @@ func (b *BackgroundSubtractorMOG) Close() error {
 // https://docs.opencv.org/master/d1/dfe/classcv_1_1cuda_1_1BackgroundSubtractorMOG.html#a8f52d2f7abd1c77c84243efc53972cbf
 //
 func (b *BackgroundSubtractorMOG) Apply(src GpuMat, dst *GpuMat) {
-	C.CudaBackgroundSubtractorMOG_Apply((C.CudaBackgroundSubtractorMOG)(b.p), src.p, dst.p)
+	C.CudaBackgroundSubtractorMOG_Apply((C.CudaBackgroundSubtractorMOG)(b.p), src.p, dst.p, nil)
+	return
+}
+
+// ApplyWithStream computes a foreground mask using the current BackgroundSubtractorMOG
+// using a Stream for concurrency.
+//
+// For further details, please see:
+// https://docs.opencv.org/master/d1/dfe/classcv_1_1cuda_1_1BackgroundSubtractorMOG.html#a8f52d2f7abd1c77c84243efc53972cbf
+//
+func (b *BackgroundSubtractorMOG) ApplyWithStream(src GpuMat, dst *GpuMat, s Stream) {
+	C.CudaBackgroundSubtractorMOG_Apply((C.CudaBackgroundSubtractorMOG)(b.p), src.p, dst.p, s.p)
 	return
 }
diff --git a/cuda/bgsegm.h b/cuda/bgsegm.h
@@ -21,11 +21,11 @@ typedef void* CudaBackgroundSubtractorMOG;
 
 CudaBackgroundSubtractorMOG2 CudaBackgroundSubtractorMOG2_Create();
 void CudaBackgroundSubtractorMOG2_Close(CudaBackgroundSubtractorMOG2 b);
-void CudaBackgroundSubtractorMOG2_Apply(CudaBackgroundSubtractorMOG2 b, GpuMat src, GpuMat dst);
+void CudaBackgroundSubtractorMOG2_Apply(CudaBackgroundSubtractorMOG2 b, GpuMat src, GpuMat dst, Stream s);
 
 CudaBackgroundSubtractorMOG CudaBackgroundSubtractorMOG_Create();
 void CudaBackgroundSubtractorMOG_Close(CudaBackgroundSubtractorMOG b);
-void CudaBackgroundSubtractorMOG_Apply(CudaBackgroundSubtractorMOG b, GpuMat src, GpuMat dst);
+void CudaBackgroundSubtractorMOG_Apply(CudaBackgroundSubtractorMOG b, GpuMat src, GpuMat dst, Stream s);
 
 #ifdef __cplusplus
 }

diff --git a/cuda/bgsegm_test.go b/cuda/bgsegm_test.go
@@ -1,8 +1,9 @@
 package cuda
 
 import (
-	"gocv.io/x/gocv"
 	"testing"
+
+	"gocv.io/x/gocv"
 )
 
 func TestCudaMOG2(t *testing.T) {
@@ -33,6 +34,35 @@ func TestCudaMOG2(t *testing.T) {
 	}
 }
 
+func TestCudaMOG2WithStream(t *testing.T) {
+	img := gocv.IMRead("../images/face.jpg", gocv.IMReadColor)
+	if img.Empty() {
+		t.Error("Invalid Mat in CudaMOG2 test")
+	}
+	defer img.Close()
+
+	var cimg, dimg, s = NewGpuMat(), NewGpuMat(), NewStream()
+	defer cimg.Close()
+	defer dimg.Close()
+	defer s.Close()
+
+	cimg.Upload(img)
+
+	dst := gocv.NewMat()
+	defer dst.Close()
+
+	mog2 := NewBackgroundSubtractorMOG2()
+	defer mog2.Close()
+
+	mog2.ApplyWithStream(cimg, &dimg, s)
+
+	dimg.Download(&dst)
+
+	if dst.Empty() {
+		t.Error("Error in TestCudaMOG2 test")
+	}
+}
+
 func TestCudaMOG(t *testing.T) {
 	img := gocv.IMRead("../images/face.jpg", gocv.IMReadColor)
 	if img.Empty() {
@@ -60,3 +90,32 @@ func TestCudaMOG(t *testing.T) {
 		t.Error("Error in TestCudaMOG test")
 	}
 }
+
+func TestCudaMOGWithStream(t *testing.T) {
+	img := gocv.IMRead("../images/face.jpg", gocv.IMReadColor)
+	if img.Empty() {
+		t.Error("Invalid Mat in CudaMOG test")
+	}
+	defer img.Close()
+
+	var cimg, dimg, s = NewGpuMat(), NewGpuMat(), NewStream()
+	defer cimg.Close()
+	defer dimg.Close()
+	defer s.Close()
+
+	cimg.Upload(img)
+
+	dst := gocv.NewMat()
+	defer dst.Close()
+
+	mog2 := NewBackgroundSubtractorMOG()
+	defer mog2.Close()
+
+	mog2.ApplyWithStream(cimg, &dimg, s)
+
+	dimg.Download(&dst)
+
+	if dst.Empty() {
+		t.Error("Error in TestCudaMOG test")
+	}
+}
diff --git a/cuda/cuda.cpp b/cuda/cuda.cpp
@@ -55,3 +55,11 @@ int GpuMat_Channels(GpuMat m) {
 int GpuMat_Type(GpuMat m) {
     return m->type();
 }
+
+Stream Stream_New() {
+    return new cv::cuda::Stream();
+}
+
+void Stream_Close(Stream s){
+    delete s;
+}