Fast access to/from GPMC Enabled FPGA Memory

hmmm...maybe to do this with the NEON SIMD unit :o

Eric,

When you run the above code what's the observation? 

There is no need of 32-bit data bus to use ldm/stm instructions. What is the current throughput that you are getting? What is the expected throughput?

Thinking that I needed to specifically implement a 16-bit copy to/from the GPMC CS addressable memory at 0x2b000000, I implemented fast memcpy using the NEON

NEONCopyPLD:
      pld [r1, #0xC0]
      vldm r1!,{d0-d7}
      vstm r0!,{d0-d7}
      subs r2,r2,#0x40
      bge NEONCopyPLD

as is discussed here : http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.faqs/ka13544.html

I tried benchmarking a normal 16-bit memcpy, ldm/stm, and the NEON memcpy, and as expected the NEON won.  For grins, I tried copying to the GPMC device memory using all three.  To my surprise I didn't have to make any specific accommodation for accessing the GPMC device!  The controller took care of it for me!  As expected, access from the NEON was the winner.

How is it that I'm able to transfer the contents of a 32-bit register to two 16-bit locations in my GPMC device?  Is the controller taking care of that for me?  I am impressed

As requested, here's the #s for the throughput I'm getting:

"regular memcpy" is a naive copy short * to short *

Copy 2048 bytes from a DDR word align buffer to a memory mapped GPMC device:

Regular memcpy takes           1033 Usec
Compare result is 0
memcpy with LDM/STM takes 207 Usec
Compare result is 0
neon takes                                  185 Usec

Copy 2048 bytes from a DDR word aligned buffer to another DDR memory mapped word aligned buffer:

Regular memcpy takes       936 Usec
memcpy with ldm/stm  takes 89 Usec
neon takes                               52 Usec