Sign Up to our social questions and Answers Engine to ask questions, answer people’s questions, and connect with other people.
Login to our social questions & Answers Engine to ask questions answer people’s questions & connect with other people.
Lost your password? Please enter your email address. You will receive a link and will create a new password via email.
Please briefly explain why you feel this question should be reported.
Please briefly explain why you feel this answer should be reported.
Please briefly explain why you feel this user should be reported.
what is better? I need to process data in several steps and it appears to me that I’ve 2 options:
1) use one big kernel
2) use streams with one kernel for each step
There is some latency before a kernel is executed, but does it really matter in this case? Is latency for a big kernel same as sum of latencies for several smaller kernels?
Are there any advantages one way compared to the other one?
Thanks guys.
Launch latency for a kernel on a Fermi card is on the order of 10us, so nothing to worry about. It makes sense — to render a scene in a game, one has to run many different shaders (which are kernels).
A kernel has to read the data that it will process from global memory and write the results back to global memory. So each separate kernel implies that full read/write cycle. You may be able to speed things up if you are able to chain multiple steps together in a big kernel, still bracketed by a single read/write cycle.
As an example, if you need to perform operations A, B and C, chaining them might give you READ – A – B – C – WRITE while separate kernels would give you READ – A – WRITE – READ – B – WRITE – READ – C – WRITE.
Remember, even if you run even a single kernel, you can still keep your code readable by breaking the separate steps out to separate device functions.