On Jun 18, 2007, at 3:32 PM, Ben Harris wrote:

> In article <9BCD26A4-4861-46EE-8FF8-A5DEDDCA0405@invisible.ca> you  
> write:
>> On 18-Jun-07, at 8:16 AM, Pawel Jakub Dawidek wrote:
>>> I'd suggest correctness verification. ~15GB/s seems to be waaaay
>>> too fast:)
>>
>> You're right, the results vary greatly from run to run. I've seen
>> numbers considerably lower than this as well as numbers considerably
>> higher. Not exactly sure why, maybe openssl has the timing wrong --
>> it claims that it's running for 3s on n size blocks, but then later
>> claims that it only ran for a fraction of that period of time, ie:
>
> I haven't checked the code, but my guess would be that 'openssl speed'
> runs for 3s of real time, but then checks the amount of CPU time it's
> used.  Since the CPU is idle waiting for the result from the
> accelerator, the CPU time is roughly zero, hence the ludicrous  
> numbers.

That said, my examination of the code revealed a busy-wait in the  
glxsb driver waiting for the AES operation to complete, which includes  
a bus_space_read_4() in the tight (without delay) loop.

Seems like that ought to use a callout instead.

-- thorpej