Re: [boost] [compute] Review

Subject: Re: [boost] [compute] Review
From: John Bytheway (jbytheway+boost_at_[hidden])
Date: 2014-12-30 23:13:57

On 2014-12-30 21:46, Gruenke,Matt wrote:
> -----Original Message----- From: Boost
> [mailto:boost-bounces_at_[hidden]] On Behalf Of John Bytheway 
> Sent: Tuesday, December 30, 2014 21:23 To: boost_at_[hidden] 
> Subject: Re: [boost] [compute] Review
> 
>> On 2014-12-29 02:01, Kyle Lutz wrote:
> 
>>> Currently it will just buffer the entire range into a
>>> std::vector<T> and then copy that to the device. If this is a
>>> bottleneck it could also be improved to copy the data in smaller
>>> batches. Let me know if you encounter issues with this.
>> 
>> Does this also work with copy_async?  I presume not, because there 
>> would be troubles with the lifetime of the temporary vector...
> 
> Async host <-> device copies only exist at the low level
> (command_queue).  They sidestep the issue by using raw pointers and
> size.

I dispute that.  There's a copy_async to which you need not supply a
command_queue.  It seems high-level to me.
<http://kylelutz.github.io/compute/boost/compute/copy_async.html>

Indeed, the following compiles:

  typedef uint64_t T;
  std::list<T> list = {0, 1, 2};
  compute::vector<T> device_vector(list.size());
  compute::copy_async(list.begin(), list.end(), device_vector.begin());

but then fails at runtime because copy_async doesn't support
non-contiguous iterators.  This is bad.  The condition inside copy_async
is compile-time, so this assert could be a static_assert.  I made an
issue.  https://github.com/kylelutz/compute/issues/401

John