[PATCH 1/2] wined3d: Introduce wined3d_output_get_ordinal().

[Henri Verbeet]

On Mon, 23 Nov 2020 at 18:50, Zhiyi Zhang <zzhang at codeweavers.com> wrote:
> On 2020/11/23 22:54, Henri Verbeet wrote:
> > On Mon, 23 Nov 2020 at 18:06, Zhiyi Zhang <zzhang at codeweavers.com> wrote:
> >> On 2020/11/23 21:32, Henri Verbeet wrote:
> >>> On Fri, 13 Nov 2020 at 11:19, Zhiyi Zhang <zzhang at codeweavers.com> wrote:
> >>>> Secondhand Lands calls d3d9_device_GetDeviceCaps() frequently. Introduce a wined3d_output_get_ordinal()
> >>>> to get output ordinal only instead of using the more expensive wined3d_output_get_desc().
> >>>>
> >>> Avoiding calls to EnumDisplaySettingsExW() is fine in principle, but I
> >>> don't think these two patches are quite enough to avoid the underlying
> >>> issue. In particular, wined3d_output_get_raster_status() is a
> >>> potential source of frequent wined3d_output_get_display_mode() calls,
> >>> either through d3d9_swapchain_GetRasterStatus() or
> >>> ddraw7_GetVerticalBlankStatus()/ddraw7_GetScanLine(). See also bug
> >>> 30538 and related bugs.
> >> Yes. I am aware of the root cause being that EnumDisplaySettingsExW() is too expensive.
> >> I plan to optimize it later. Meanwhile, I think these two patches are reasonable. Are you
> >> proposing another way to fix this?
> >>
> > Mostly just making sure that you're aware there are other places
> > affected by this. I also don't particularly love introducing an extra
> > wined3d export for this though; shouldn't the 2/2 patch on its own be
> > enough to fix the immediate regression?
> Yes, for that bug, the 2/2 patch is enough to make the game run at 30fps limit.
> I just think that d3d9_device_GetDeviceCaps() only needs output ordinals. So calling
> wined3d_output_get_desc() will do a monitor enumeration that can be avoided, which
> should improve performance for other applications.
>
If it ends up taking significant CPU time even after patch 2/2,
perhaps it would make sense to simply store the converted D3DCAPS9
structure in struct d3d9; its contents should be stable across calls.