[llvm] [PATCH 2/2] AMDGPU: Add image attribute reader intrinsic.

Zoltán Gilián zoltan.gilian at gmail.com
Sat Jul 11 12:40:53 PDT 2015


> Why is this +16 and not +8?

The grid offset consists of 3 dwords for the 3 coordinates. Grid dim +
grid offset = 4 dwords = 16 bytes.

On Sat, Jul 11, 2015 at 12:17 AM, Matt Arsenault <arsenm2 at gmail.com> wrote:
>
> On Jul 10, 2015, at 2:08 PM, Tom Stellard <tom at stellard.net> wrote:
>
> This should use the amdgpu target prefix.
>
>
> It actually shouldn’t. The current situation is kind of broken. We have a
> mix of AMDIL, AMDGPU, r600 and SI. These are supposed to match the target
> arch name as in the triple. We really should have __builtin_r600 _and
> __builtin_amdgcn_ but right now most are using __builtin_amdgpu/llvm.AMDGPU
>
> -Matt




More information about the llvm-commits mailing list