From: Michal Sojka <michal.sojka@cvut.cz>
Date: Thu, 11 Oct 2018 14:39:54 +0000 (+0200)
Subject: cufft: Use per-thread CUDA stream for cuBLAS calls
X-Git-Url: https://rtime.felk.cvut.cz/gitweb/hercules2020/kcf.git/commitdiff_plain/7c16a96d82c181a319c35a438f22342c1f5bbbe2

cufft: Use per-thread CUDA stream for cuBLAS calls
---

diff --git a/src/fft_cufft.cpp b/src/fft_cufft.cpp
index 963c8b7..c7a9e2f 100644
--- 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