[resend PATCH v3 2/4] wined3d: Support Nvidia alpha to coverage state.

[Paul Gofman]

On 2/13/20 19:10, Henri Verbeet wrote:
> On Thu, 13 Feb 2020 at 19:25, Paul Gofman <gofmanp at gmail.com> wrote:
>> On 2/13/20 18:37, Henri Verbeet wrote:
>>> On Thu, 13 Feb 2020 at 18:28, Paul Gofman <gofmanp at gmail.com> wrote:
>>> I was thinking of simply taking WINED3D_RS_ADAPTIVETESS_Y into account
>>> when setting the blend state in wined3d_device_apply_stateblock().
>>> I.e., some variant of the following:
>>>
>>>     if (changed->blend_state || state_changed(changed->renderState,
>>> WINED3D_RS_ADAPTIVETESS_Y))
>>>     {
>>>         blend_state = state->blend_state;
>>>         if (state->rs[WINED3D_RS_ADAPTIVETESS_Y] == WINED3DFMT_ATOC)
>>>             blend_state = device->blend_state_atoc_enabled;
>>>         ...
>>>         wined3d_device_set_blend_state(..., blend_state, ...);
>>>     }
>> This will reset ATOC state whenever application sets _ADAPTIVETESS_Y,
>> while the application could have set ATOC state AMD way previously (not
>> in the current stateblock, so changed->blend_state is 0 and
>> state->blend_state is NULL).
> I don't think that can happen for WINED3D_SBT_PRIMARY stateblocks.
> (I.e., the ones used with wined3d_device_apply_stateblock().) That
> does mean some loss of generality for
> wined3d_device_apply_stateblock(), but I don't think that can be
> maintained once it starts setting state objects anyway. I.e., I think
> it's fine to assume the stateblock contains complete and consistent
> state in wined3d_device_apply_stateblock().


ATOC state can be set / reset with _SBT_RECORDED stateblocks as well.
The logic for this case seems to work now: if the recorded stateblock
had an AMD trigger then state block has changed.blend_state and updates
it, otherwise it does not touch it. It is encoded not exactly
straightforward but looks the same in principle like if we had something
like D3DRS_ATOC separate state. I. e., I think we could introduce
WINED3D_RS_ATOC and translate _POINTSIZE('A2M1') to _RS_ATOC(TRUE) and
_POINTSIZE('A2M0') to _RS_ATOC(FALSE) (I am not suggesting to do so at
least because it won't simplify the things in wined3d anyway due to
d3d11 blend state object). But maybe it worth to change
wined3d_stateblock_set_render_state() and not to record _RS_POINTSIZE in
case it is (re)setting ATOC? This would simplify blend state application
in wined3d_device_apply_stateblock() and wined3d_device_apply_stateblock().

While d3d9 adaptive tesselation itself which could interfere with AMD
ATOC state is barely ever used or supported, my concern was more related
to games which like to reset a bunch of random (or all) states now and
then just in case. But it is not something I have an application to
depend on though.