[Bug 32096] New: drawStridedSlow can be more efficient

http://bugs.winehq.org/show_bug.cgi?id=32096

Bug #: 32096
           Summary: drawStridedSlow can be more efficient
           Product: Wine
           Version: unspecified
          Platform: x86
        OS/Version: Linux
            Status: UNCONFIRMED
          Severity: enhancement
          Priority: P2
         Component: directx-d3d
        AssignedTo: [e-mail]
        ReportedBy: [e-mail]
    Classification: Unclassified

Hi,

Recently I noticed that World of Warcaft is far more slow on d3d9 mode than
using the OpenGL mode on my computer (pentium 4 and geforce 5 fx).
Unfortunately the OpenGL mode is unusable since last WoW update (bug is inside
WoW, not wine)

D3D mode is slower because an important amount of the primitives are rendered
using drawStridedSlow (dlls/wined3d/drawprim.c) which generates a lot of opengl
call (here, 600K calls compared to 12K in OpenGL mode).

It appears that if the computer can handle vertex array (ie:
glDrawElements/glDrawArrays/gl[Vertex|Color|TexCoords|Fog]Pointer and
glClientActiveTexture) the number of openGL call can be dramatically reduced,
leading to an important increase in FPS (8 to 25 in important city in WoW).

On my computer, I hacked inside drawStritedSlow and I have only one issue, the
format of color is packed in BGRA, which is not compatible with the format
awaited by glColorPointer. This can be fixed using the extension
ARB_vertex_array_BGRA, which is unfortunately not available on my computer. I
solved this issue by looping over each vertex and building a custom array in
memory that I submitted to glColorPointer. I got approximately the same FPS
increase using this approach.

My conclusions here are that the fallback to drawStridedSlow is a bit extreme
and in most case unnecessary. In my configuration everything works except the
conversion from the d3d BGRA format for color in RGBA. This part can be easily
done in software to create a suitable color array. Also, I did not dig inside
the shader code of wined3d, but perhaps this conversion can be simply
implemented as a swizzle in the custom vertex shader. (am I right guessing that
wined3d does not uses the default shader of the fixed pipeline ?).

So this bug report is more a way of opening the discussion. I see many options
to improve the situation:

a) Tweak the heuristics which detects and fallbacks to drawStridedSlow. If
necessary implements the needed conversions in software. But this will not
generates OpenGL function call and will increase performance a bit
b) Tweak the vertex shader to do the conversion in the vertex shader, I think
it is our best option
c) Tweak drawStridedSlow to still rely on vertex array when possible (I have an
example of code which is doing this. I'm playing WoW with this since 3 days
without any issue and with a FPS increase of 200%). I don't think that honestly
it is the good idea.
d) do something else. Because I did not read anything else than the
drawStritedSlow function, I may have missed something important.

I'm motivated to help in the DX -> GL code. I'm a skilled C and GL coder, I
don't know DX, but can learn enough. Thank you for reading.