cufft: Use per-thread CUDA stream for cuBLAS calls

diff --git a/src/fft_cufft.cpp b/src/fft_cufft.cpp
index 963c8b7084ff42d3ed24971cb627d2181897bdf9..c7a9e2f1f3197de294b6adde40f60b610462eacb 100644 (file)
--- a/src/fft_cufft.cpp
+++ b/src/fft_cufft.cpp
@@ -4,6 +4,7 @@ cuFFT::cuFFT()
 {
     CudaSafeCall(cudaSetDeviceFlags(cudaDeviceMapHost));
     cudaErrorCheck(cublasCreate(&cublas));
+    cudaErrorCheck(cublasSetStream(cublas, cudaStreamPerThread));
 }
 
 cufftHandle cuFFT::create_plan_fwd(uint howmany) const