AM3703 VFP Clock Cycles

Paul Joseph

Does it really take on the order of 18-21 clock cycles to perform one single precision floating point operation via the VFP (VFPv3 lite)?

I was under the assumption that it was more like 1-2 clock cycles. Apparently there are a number of versions of the VFP (according to the ARM site), some of which can do over 1 FLOP per clock cycle, but not the one in the 3703?

Can you confirm the number of clock cycles for the VFP when performing a single precision floating point operation via the VFP on the AM37073 please.

over 14 years ago

0 Jeff L over 14 years ago

TI__Expert 5960 points

Paul,

The VFP unit on the Cortex-A8 is not fully pipelined, so you can not get a floating point instructions to execute in 1-2 cycles. This is a known issue on the Cortex-A8. ARM does not publish cycle times on the Cortex-A8 or higher application processors because they can vary depending on the operation and current state of the Cortex-A8.

NEON is SIMD and is fully pipelined and does support single precision float. If the customer's code can utilize NEON, then they can expect 1 or 2 cycles per Single precision floating point instructions when the pipeline is full. NEON does not support double precision floating point.

The link below is a high level overview of how to utilize NEON.

http://processors.wiki.ti.com/index.php/Cortex-A8

Also see in this same article a direct comparison using VFP and NEON for single precision floating point.

http://processors.wiki.ti.com/index.php/Cortex-A8#Neon_and_VFP_both_support_floating_point.2C_which_should_I_use.3F

Processors

Processors forum

AM3703 VFP Clock Cycles