ogg dev - Feb 2008 - OggPCM: support for little-endianness only?

Jean-Marc Valin wrote:
> Really? How low does it take to swap bytes vs. the time it takes to read
> from disk?
In CPU cycles? The difference is enormous.

If no byte swapping is required the vast majority of the work is done by
the DMA engine, from disk straight into user space memory and while the
read is underway, the CPU can do other work.

For the byte swapping case, every sample needs to be loaded into the cache
and CPU, byte swapped and then written back to memory.

Erik
-- 
-----------------------------------------------------------------
Erik de Castro Lopo
-----------------------------------------------------------------
I have now spent two days debugging C# on dotnot and all I can say
is OMFG!!!! The MS tools are completely fucking abysmal. Its the
most developer hostile environment I've ever worked on. The fact
that people are actually able to release software that sort of
works using these tools completely boggles my mind.

>> Really? How low does it take to swap bytes vs. the time it takes to
read
>> from disk?
> 
> In CPU cycles? The difference is enormous.
> 
> If no byte swapping is required the vast majority of the work is done by
> the DMA engine, from disk straight into user space memory and while the
> read is underway, the CPU can do other work.
> 
> For the byte swapping case, every sample needs to be loaded into the cache
> and CPU, byte swapped and then written back to memory.
Come on. Given a disk bandwidth of 40 MB/s and a memory bandwidth
(that's the bottleneck here) of 4 GB/s, you can saturate the disk with
just one percent CPU. Not to mention the fact that if you're doing
anything useful with the data, you're already paying for the memory
bandwidth. I'm pretty sure the overhead of the context switches and
libogg both dwarf the amount of CPU it takes to swap bytes (you can swap
4 16-bit samples per cycle with SSE).

I'd be quite interested if you could provide a benchmark that really
shows by swapping to make a difference.

	Jean-Marc

Jean-Marc Valin wrote:
> I'd be quite interested if you could provide a benchmark that really
> shows by swapping to make a difference.
Yes, I agree that the speed difference between the swapping and the
no-swapping case is negligible, but I'm talking about CPU cycles.

With libsndfile people do stream 10, 20 or more channels of audio
off disk at once, with soft realtime requirements. In situations
like this, avoiding the CPU cycles of end swapping is probably a
good idea.

Erik
-- 
-----------------------------------------------------------------
Erik de Castro Lopo
-----------------------------------------------------------------
"You can have quality software, or you can have pointer arithmetic;
but you cannot have both at the same time." -- Bertrand Meyer, 1989