[PATCH vkd3d 00/11] vkd3d-shader: Fix patch constant handling in hull shaders.

[Philip Rebohle]

Hi,

unfortunately that is currently the case, yes, since some descriptors 
are left completely undefined. On RADV you can work around the system 
hang by setting RADV_DEBUG=zerovram, however you will see rendering 
artifacts.

- Philip

Am 18.10.19 um 15:59 schrieb Conor McCarthy:
> Hi Philip,
> 
> Is the descriptor update issue expected to crash the game? Using these 
> patches and tessellation I get a hard crash requiring reboot at the 
> start of benchmark scene 2. Mesa 19.2.1 and AMD RX 580.
> 
> On Thu, Oct 17, 2019 at 5:57 AM Philip Rebohle 
> <philip.rebohle at tu-dortmund.de <mailto:philip.rebohle at tu-dortmund.de>> 
> wrote:
> 
> 
>     This series mainly changes the way patch constant outputs are generated
>     in hull shaders to always use a private array variable. This is done to
>     fix several issues in the current imlementation, and to avoid excessive
>     complexity in dealing with patch constants which stem from DXBC's
>     design:
> 
>     1) Tessellation factors and regular patch constants share the same
>         register space in DXBC and can both be dynamically or statically
>         indexed. The current implementation only works correctly in case
>         tessellation factors do not share registers with regular outputs,
>         and does not support all indexing patterns, leading to invalid
>         code being generated in some cases.
> 
>     2) Join phases can read patch constants written by a fork phase,
>         also with dynamic indexing. However, the index ranges used by
>         the join phase are not known at the time the fork phase is
>         compiled.
> 
>     With these changes, tessellation shaders work correctly in Shadow of
>     the Tomb Raider, which previously contained invalid SPIR-V and would
>     crash the game if enabled.
> 
>     Note that enabling the tessellation option in the game triggers a
>     separate issue related to descriptor updates, which will have to
>     be addressed separately.
> 
>