> -----Original Message-----
> From: Wang, Rander
> Sent: Wednesday, March 29, 2017 1:12 PM
> To: Song, Ruiling <[email protected]>; [email protected]
> Subject: RE: [Beignet] [PATCH] utests: add utest to double version of log, 
> log2,
> log10
> 
> As I know ivyBrigde support double, it there anything wrong?
IVB did support many double operations. But division operation on double is not 
supported until GEN8.
And what's more, the double add/mul on IVB and HSW has very low precision which 
is far from required if I remember correctly.
So, we can only support double from GEN8.

Thanks!
Ruiling
> 
> -----Original Message-----
> From: Song, Ruiling
> Sent: Wednesday, March 29, 2017 11:06 AM
> To: Wang, Rander <[email protected]>; [email protected]
> Cc: Wang, Rander <[email protected]>
> Subject: RE: [Beignet] [PATCH] utests: add utest to double version of log, 
> log2,
> log10
> 
> > +static void builtin_double_logx(void) {
> > +  // Setup kernel and buffers
> > +  int k, i, index_cur;
> > +  unsigned long gpu_data[max_function * count_input] = {0};
> > +  float diff;
> > +  char log[256] = {0};
> > +
> > +  OCL_CREATE_KERNEL("builtin_double_logx");
> > +
> 
> I just thought of a problem. Seems that you need to check whether cl_khr_fp64
> extension is supported or not.
> For older generation hardware. It is not possible to support double. Then this
> utest will certainly fail.
> So, I think you need to add the this extension check for all the double test 
> cases.
> 
> Thanks!
> Ruiling
_______________________________________________
Beignet mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/beignet

Reply via email to