[Bug 45901] Avoid GPU synchronization due to GPU-CPU transfer (Overwatch)

https://bugs.winehq.org/show_bug.cgi?id=45901

--- Comment #6 from Andrew Wesie  ---
(In reply to Józef Kucia from comment #5)
> I don't think that you are using the classic Radeon driver. You probably use
> one of the Gallium drivers, radeonsi most likely.

You are correct. Those were the wrong parts of Mesa to highlight.

In ./mesa/state_tracker/st_cb_texture.c:st_GetTexSubImage, which is the code
path for Gallium drivers, it will blit into a tmp texture, map the PBO, map the
tmp texture, and memcpy:

st->pipe->blit(st->pipe, &blit);
    pixels = _mesa_map_pbo_dest(ctx, &ctx->Pack, pixels);
    map = pipe_transfer_map_3d(pipe, dst, 0, PIPE_TRANSFER_READ, ...);
    ...
    memcpy(dest, slice_map, bytesPerRow);
    ...
    pipe_transfer_unmap(pipe, tex_xfer);
    _mesa_unmap_pbo_dest(ctx, &ctx->Pack);

In mesa/state_tracker/st_cb_readpixels.c:st_ReadPixels it has a fast path for
PBOs:

if (st->pbo.download_enabled && _mesa_is_bufferobj(pack->BufferObj)) {       
      if (try_pbo_readpixels(st, strb,                                          
                             st_fb_orientation(ctx->ReadBuffer) == Y_0_TOP,     
                             x, y, width, height,                               
                             src_format, dst_format,                            
                             pack, pixels))                                     
         return;                                                                
   }

The rationale for using glReadPixels is the lack of hardware support for
GetTexSubImage in Mesa's Gallium drivers.

While try_pbo_readpixels looks very complex, to me at least, it avoids the very
costly CPU-GPU synchronization.