Benchmarks of C6Accel vs IMGLIB

chuan jia

Prodigy 30 points

Other Parts Discussed in Thread: DM3730

Hello,

I compared the benchmarks of C6Accel with IMGLIB, I found that the C6Accel much slower than the IMGLIB2.

Data source:

IMGLIB: TMS320C64+ DSP Image/Video Processing Library(V2.0.1) Programmer's Guide, Appendix B Benchmarks

C6Accel: Benchmarks of C6Accel on DM3730 , Benchmarks_sync_dm3730.pdf

Comparision:

640x480 8bit/pixel b/w Test Image

800MHz DSP

Conv

IMG_conv_3x3_i8_c8s: 6*cols/4+40 cycles ,

(6*640/4+40)*480 / 800 = 600us

C6accel_IMG_conv_3x3_i8_c8s(),

3814us

it cost 6x time.

Is there any mistake when I calculate the time cost?

Or using C6Accel is much slower then using IMGLIB directly.

Thanks

over 14 years ago

0 Rahul Prabhu over 14 years ago

TI__Guru** 116170 points

chuan jia,

C6Accel is an ARM side tool to call into optimized DSP software like the IMGLIB. The benchmarks provided with C6Accel include the round trip over head of calling the DSP from the ARM and also the time required to manage the cache on the ARM and DSP (as data is placed in external memory). The benchmarks in the IMGLIB documentation are on cycle accurate simultor which assumes all data is in the internal memory. You cannot use IMGLIB functions directly on the ARM as the code is written using C6000 intrinsics, however if you run the natural C version of the IMGLIB function on the ARM, you will notice that C6Accel API performs 60% better.

Please look at the following wiki section for more details,

http://processors.wiki.ti.com/index.php/C6Accel_FAQ#What_is_the_inter-processor_overhead_involved_in_C6Accel_.3F

Regards,

Rahul

0 chuan jia over 14 years ago in reply to Rahul Prabhu

Prodigy 30 points

I got it.

Thank you very much.

Processors

Processors forum

Benchmarks of C6Accel vs IMGLIB