[Beignet] [PATCH 3/5] Remove intel_gpgpu_check_binded_buf_address()

[Zhenyu Wang]

On 2014.10.24 18:10:36 +0800, Zhigang Gong wrote:
> This assertion is just to make sure we will not get a NULL pointer
> for a normal buffer. The OpenCL spec doesn't give a very specific
> statement about a NULL buffer object. But it does allow to pass
> a NULL to a buffer object. Thus some one may implement the following
> kernel:
> 
> __kernel foo( global uint * input, global uint *output1, global uint *output2)
> {
>   ...
>   if (output1)
>     output1[get_global_id(0)] = result0;
>   if (output2)
>     output2[get_global_id(1)] = result1;
> }
> 
> If we pass in a NULL output1 which should be a normal allocated buffer,
> it breaks the above code.

Looks "runtime_null_kernel_arg" is a good one to test this.

> 
> Without PPGTT, this assertion works fine till now. If with PPGTT,
> we may hit this assertion, but we can't just simply remove this assertion.

Actually with aliasing PPGTT (which is default on for a long time),
that case can still possibly fail. You might hit it more easy when
boot into console mode without X.  The issue is that buffer offset is
zero doesn't mean buffer pointer is NULL.

> 
> Do you have good suggestion to always avoid allocate 0 offset for a
> valid buffer object?
> 

I'm not sure if kernel provides such capability or if it should, or cl implement
should take special value for NULL.

> All the other patches in this series LGTM, I will push them latter.
> Let's defer this one to next week.

sure, thanks

-- 
Open Source Technology Center, Intel ltd.

$gpg --keyserver wwwkeys.pgp.net --recv-keys 4D781827
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 181 bytes
Desc: Digital signature
URL: <http://lists.freedesktop.org/archives/beignet/attachments/20141027/2e467b56/attachment.sig>