Cross-regressions between wine and mesa: how to handle them?

[Ruslan Kabatsayev]

On Fri, Jan 31, 2014 at 12:38 PM, Ruslan Kabatsayev
<b7.10110111 at gmail.com> wrote:
> On Fri, Jan 31, 2014 at 2:27 AM, Henri Verbeet <hverbeet at gmail.com> wrote:
>> On 30 January 2014 22:22, Stefan Dösinger <stefandoesinger at gmail.com> wrote:
>>> I don't know about the extend this card support pixel shaders. There
>>> may be a way for the driver and/or wine to work without falling back
>>> to software. But overall I disagree with the premise of this patch.
>>> There's a point in shipping a < GL 2.0 driver for hardware that just
>>> doesn't support GL 2.0. But there is no point in requiring the
>>> application to magically know if the driver's GL 2.0 is really 2.0 or
>>> just something rigged with software fallback mines.
>>>
>>> Can you file a Mesa bug about this issue? I think we should ask the
>>> Mesa devs about their opinion.
>>>
>> I don't know if that's specifically the issue here, but GL 2.0 also
>> introduced point sprite coordinate origin switching. We need that when
>> rendering offscreen, and I seem to recall i915 hitting a software
>> fallback for that. Without GL 2.0 we'd just render incorrectly
>> instead. It should be easy enough to rule that out by commenting out
>> the "glPointParameteriNV(GL_POINT_SPRITE_COORD_ORIGIN, origin);" call
>> in psorigin(). It's also generally useful to set INTEL_DEBUG=perf when
>> investigating these kinds of things.
>
> Indeed, you're right. If I return psorigin() before any GL calls are
> made (first one, not NV one is effective, btw), then it works fast
> enough.

Now, as to INTEL_DEBUG. Here's the output:

intel_copy_texsubimage mismatched formats MESA_FORMAT_ARGB8888,
MESA_FORMAT_RGB565
intelCopyTexSubImage - fallback to swrast
intelReadPixels: fallback to swrast
intel_copy_texsubimage mismatched formats MESA_FORMAT_ARGB8888,
MESA_FORMAT_RGB565
intelCopyTexSubImage - fallback to swrast
intelReadPixels: fallback to swrast
intel_copy_texsubimage mismatched formats MESA_FORMAT_ARGB8888,
MESA_FORMAT_ARGB1555
intelCopyTexSubImage - fallback to swrast
intelReadPixels: fallback to swrast
intel_copy_texsubimage mismatched formats MESA_FORMAT_ARGB8888,
MESA_FORMAT_ARGB4444
intelCopyTexSubImage - fallback to swrast
[snip repeating messages]
ENTER FALLBACK 100000: point sprite coord origin
Mapping a busy BO, causing a stall on the GPU.
trace:fps:glxdrv_SwapBuffers @ approx 0.04fps, total 0.04fps
trace:fps:swapchain_gl_present 0x1a7ff8 @ approx 0.04fps
ENTER FALLBACK 100000: point sprite coord origin
Mapping a busy BO, causing a stall on the GPU.
trace:fps:glxdrv_SwapBuffers @ approx 0.22fps, total 0.44fps
trace:fps:swapchain_gl_present 0x1a7ff8 @ approx 0.22fps
Mapping a busy BO, causing a stall on the GPU.

Without glPointParameter() calls I don't have those errors after
"[snip repeating messages]", but still do have CopyTexSubImage ones
(although I didn't notice such a serious slowdown as was before; and
these have also been before bisected wine commit).