362   PetscCallCUDA(cudaSetDeviceFlags(cudaDeviceBlockingSync));  in main()
433   PetscCallCUDA(cudaMalloc((void **)&d_a, sizeof(float) * N));  in runStream()
434   PetscCallCUDA(cudaMalloc((void **)&d_b, sizeof(float) * N));  in runStream()
435   PetscCallCUDA(cudaMalloc((void **)&d_c, sizeof(float) * N));  in runStream()
493     PetscCallCUDA(cudaEventRecord(stop, 0));  in runStream()
559     PetscCallCUDA(cudaMemcpy(h_a, d_a, sizeof(float) * N, cudaMemcpyDeviceToHost));  in runStream()
560     PetscCallCUDA(cudaMemcpy(h_c, d_c, sizeof(float) * N, cudaMemcpyDeviceToHost));  in runStream()
573     PetscCallCUDA(cudaMemcpy(h_a, d_a, sizeof(float) * N, cudaMemcpyDeviceToHost));  in runStream()
574     PetscCallCUDA(cudaMemcpy(h_c, d_c, sizeof(float) * N, cudaMemcpyDeviceToHost));  in runStream()
587     PetscCallCUDA(cudaMemcpy(h_b, d_b, sizeof(float) * N, cudaMemcpyDeviceToHost));  in runStream()
588     PetscCallCUDA(cudaMemcpy(h_c, d_c, sizeof(float) * N, cudaMemcpyDeviceToHost));  in runStream()
601     PetscCallCUDA(cudaMemcpy(h_a, d_a, sizeof(float) * N, cudaMemcpyDeviceToHost));  in runStream()
602     PetscCallCUDA(cudaMemcpy(h_b, d_b, sizeof(float) * N, cudaMemcpyDeviceToHost));  in runStream()
603     PetscCallCUDA(cudaMemcpy(h_c, d_c, sizeof(float) * N, cudaMemcpyDeviceToHost));  in runStream()
616     PetscCallCUDA(cudaMemcpy(h_a, d_a, sizeof(float) * N, cudaMemcpyDeviceToHost));  in runStream()
617     PetscCallCUDA(cudaMemcpy(h_b, d_b, sizeof(float) * N, cudaMemcpyDeviceToHost));  in runStream()
618     PetscCallCUDA(cudaMemcpy(h_c, d_c, sizeof(float) * N, cudaMemcpyDeviceToHost));  in runStream()
633   PetscCallCUDA(cudaFree(d_a));  in runStream()
634   PetscCallCUDA(cudaFree(d_b));  in runStream()
635   PetscCallCUDA(cudaFree(d_c));  in runStream()
648   PetscCallCUDA(cudaMalloc((void **)&d_a, sizeof(double) * N));  in runStreamDouble()
649   PetscCallCUDA(cudaMalloc((void **)&d_b, sizeof(double) * N));  in runStreamDouble()
650   PetscCallCUDA(cudaMalloc((void **)&d_c, sizeof(double) * N));  in runStreamDouble()
773     PetscCallCUDA(cudaMemcpy(h_a, d_a, sizeof(double) * N, cudaMemcpyDeviceToHost));  in runStreamDouble()
774     PetscCallCUDA(cudaMemcpy(h_c, d_c, sizeof(double) * N, cudaMemcpyDeviceToHost));  in runStreamDouble()
787     PetscCallCUDA(cudaMemcpy(h_a, d_a, sizeof(double) * N, cudaMemcpyDeviceToHost));  in runStreamDouble()
788     PetscCallCUDA(cudaMemcpy(h_c, d_c, sizeof(double) * N, cudaMemcpyDeviceToHost));  in runStreamDouble()
800     PetscCallCUDA(cudaMemcpy(h_b, d_b, sizeof(double) * N, cudaMemcpyDeviceToHost));  in runStreamDouble()
801     PetscCallCUDA(cudaMemcpy(h_c, d_c, sizeof(double) * N, cudaMemcpyDeviceToHost));  in runStreamDouble()
814     PetscCallCUDA(cudaMemcpy(h_a, d_a, sizeof(double) * N, cudaMemcpyDeviceToHost));  in runStreamDouble()
815     PetscCallCUDA(cudaMemcpy(h_b, d_b, sizeof(double) * N, cudaMemcpyDeviceToHost));  in runStreamDouble()
816     PetscCallCUDA(cudaMemcpy(h_c, d_c, sizeof(double) * N, cudaMemcpyDeviceToHost));  in runStreamDouble()
829     PetscCallCUDA(cudaMemcpy(h_a, d_a, sizeof(double) * N, cudaMemcpyDeviceToHost));  in runStreamDouble()
830     PetscCallCUDA(cudaMemcpy(h_b, d_b, sizeof(double) * N, cudaMemcpyDeviceToHost));  in runStreamDouble()
831     PetscCallCUDA(cudaMemcpy(h_c, d_c, sizeof(double) * N, cudaMemcpyDeviceToHost));  in runStreamDouble()
846   PetscCallCUDA(cudaFree(d_a));  in runStreamDouble()
847   PetscCallCUDA(cudaFree(d_b));  in runStreamDouble()
848   PetscCallCUDA(cudaFree(d_c));  in runStreamDouble()