Why my CUDA code is slower than CPU code?

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi Guys, I wrote a CUDA code for 2D convolution,
the code is every simple as attached.

However I tested my code on Tesla, it got no misses compare with the CPU result, but it’s much slower than the CPU code:

setting device 0 with name Tesla C1060
GPU Runtime: 0.009131s
CPU Runtime: 0.001287s
Number of misses: 0

But if I ran my code on fermi card, it’s two times faster.
Anybody can tell me why?
2DConvolution.cu (4.08 KB)

2

The code is entirely memory-bandwidth limited, and on pre-Fermi GPUs it makes poor use of memory bandwidth because it reloads the data for every value it calculates.

For good results on compute capability 1.x devices either use a texture, or pre-load data for the whole block to shared memory and work from there.

3

thank you so much for your reply!

4

are you saying that Fermi cards have the ability to using memory better?

5

Lifan Xu ,

(sm_11) GPU : 0 ms
(x2) CPU : 16 ms

Are you sure that your cpu compute this for 1 ms ?

6

Yes, Fermi GPUs cache global memory. However, if you are willing to do the extra programming for using a texture backed by a CUDA array, it might even be faster because it takes advantage of the 2D spatial locality.

7

If you really want to get good performance you should look into using shared memory to preload the data in a block and then reuse the preloaded data from shared memory.