GPU bench marking on iMX6

[anusreesarkar]

Hi,

I wrote a very simple kernel to check the performance of the GPU on iMX6 dual core in which I was performing a multiple operation million number of times using variables defined in the private register. I ran the kernel with different number of global work group size using different number of work items. The best performance I got was when I was launching 64 work groups with 64 work items in each of them. The maximum GFLOPS that I got was slightly higher than 3 at 500 MHz whereas Vivante GPU is suppose to work at 24 GFLOPS. I tried this using float16. Is there any way other than increasing the clock frequency for increasing the performance of GPU? Hasty reply will be appreciated.

Thanks,

Anusree Sarkar

[Bio_TICFSL]

Hi Anusree,

Vivante gpu max Gflops with high precision is 24, latest tested BSP 3.14.28 with i.mx6d clock at 594M, numbers should 21.79 GFLOPS for float16. Officials GPU benchmarks are under NDA with freescale if you like to check it you need to contact your local Freescale representative.

Hope this helps

[anusreesarkar]

Hi,

I am running the GPU at 500 MHz. How shall I increase it to 594 MHz that

you have mentioned? Can you refer to the post mentioned in the following

link. I am performing the calculations on a similar line and getting a

similar computation power. Can the reason of limited memory of GPU and

constricted bandwidth be a reason for its low computation power?

https://community.freescale.com/message/320692#320692

Thanks,

Anusree

On Tue, Dec 1, 2015 at 10:18 PM, Bio_TICFSL <admin@community.freescale.com>