merge from planB

Author: zheliu137

.gitignore

@@ -1,6 +1,2 @@
-*.i
-*.ii
-*.gpu
-*.ptx
-*.cubin
-*.fatbin
+*.o
+*.x

Makefile

@@ -0,0 +1,20 @@
+LIBS+=-L/usr/local/cuda/lib64 -lcudart -lcublas -lcusolver -lcurand
+
+#objects:=cusolver_kernel_wrapper.o cusolver_kernel.o main.o
+objects:=cusolver_kernel.o curand_kernel.o main.o
+
+all: main.x
+main.x: $(objects)
+	nvcc -g -O3 -o $@ $^ $(LDFLAGS) $(LIBS)
+
+cusolver_kernel.o: cusolver_kernel.cu
+	nvcc -g -O2 -gencode=arch=compute_70,code=sm_70 -c -o $@ $^
+curand_kernel.o: curand_kernel.cu
+	nvcc -g -O2 -gencode=arch=compute_70,code=sm_70 -c -o $@ $^
+main.o: main.cu
+	nvcc -g -O2 -gencode=arch=compute_70,code=sm_70 -c -o $@ $^
+
+%.o : %.c
+	icc -c -o $*.o $<
+clean:
+	rm -f ./cu*_kernel_f *.mod *.o

README.md

@@ -1,2 +1,7 @@
-# cusolver_stream
-test stream of cusolver
+# Introduction
+
+A demo of cusolver stream on solving eigen problem of a large number of small matrix.
+
+# Conclusion
+
+After analyses with Nsight system, I found that both jacobi and QR method are unable to employ multi-stream. As there are unavoidable pageable memory copies in these two function, which is different from [cublas](https://github.com/zheliu137/cublas_stream)

curand_kernel.cu

@@ -0,0 +1,18 @@
+#include <cstdio>
+#include <cstdlib>
+#include <algorithm>
+
+#include <cuda_runtime.h>
+#include <curand.h>
+//#include "cuda_settings.h"
+
+void createRandoms(int size, double *h_randomArray){
+    curandGenerator_t generator;
+    double *randomArray;
+    cudaMalloc((void**)&randomArray, size*size*sizeof(double));
+    // cudaMallocHost((void**)&h_randomArray, size*size*sizeof(double));
+    curandCreateGenerator(&generator,CURAND_RNG_PSEUDO_XORWOW);
+    curandSetPseudoRandomGeneratorSeed(generator,(int)time(NULL));
+    curandGenerateUniformDouble(generator,randomArray,size*size);
+    cudaMemcpy(h_randomArray, randomArray, sizeof(double) * size * size, cudaMemcpyDeviceToHost);
+}