[Mesa-dev] Merge of glsl2 branch to master

[Keith Whitwell]

On Mon, 2010-08-09 at 23:48 -0700, Eric Anholt wrote:
> On Mon, 09 Aug 2010 19:05:52 -0700, Ian Romanick <idr at freedesktop.org> wrote:
> > -----BEGIN PGP SIGNED MESSAGE-----
> > Hash: SHA1
> > 
> > Tom Stellard wrote:
> > > On Fri, Aug 06, 2010 at 03:24:50PM -0700, Ian Romanick wrote:
> > >> -----BEGIN PGP SIGNED MESSAGE-----
> > >> Hash: SHA1
> > >>
> > >> As of today the glsl2 branch has three piglit regressions WRT master,
> > >> and there is only one crashing piglit tests.  As before, we have tested
> > >> swrast and i965.  For master I used commit 27041d7cb, and for glsl2 I
> > >> used commit c234d0b2.  The results are available at
> > >> http://people.freedesktop.org/~idr/results/.  The regressed tests cases are:
> > >>
> > >>  * draw_buffers-05.vert - As mentioned previously, the spec is ambiguous
> > >> and Nvidia and AMD produce different results for this test.  Until the
> > >> spec ambiguity is resolved, I don't care about this failure.
> > >>
> > >> * glsl1-texcoord varying - This test accesses gl_TexCoord[] using a
> > >> non-constant index without dimensioning the array. The new compiler
> > >> correctly rejects this shader. Nvidia accepts it, but I believe they're
> > >> going outside the spec (which they do a lot). I haven't tested this on
> > >> AMD or Apple yet. Is this shader from an application?  We've changed the
> > >> test in piglit commit c6146f121, but I think it still isn't quite right.
> > >>
> > >>  * fbo-drawbuffers-maxtargets - Fails on i965 because we don' have loop
> > >> unrolling in the compiler or good array handling in the driver.  We'll
> > >> get loop unolling in the compiler first.  Proper array handling will
> > >> come to the driver once we move away from using Mesa IR.
> > >>
> > >>
> > >> I propose that we merge master to glsl2 on *Friday, August 13th* (one
> > >> week from today).  Barring unforeseen issues, I propose that we merge
> > >> glsl2 to master on *Monday, August 16th*.
> > > 
> > > Here is a summary of the piglit regressions on r300g (with an r500 card)
> > 
> > Thanks for the reply!  I was starting to wonder if my message went to
> > /dev/null. :)
> > 
> > > These fail because r300 does not implement the SSG opcode:
> > > glsl1-acos(vec4) function
> > > glsl1-asin(vec4) function
> > > glsl1-atan(vec4) function
> > > glsl-fs-asin
> > > glsl-fs-atan-1
> > > glsl-fs-sign
> > > glsl-vs-sign
> > 
> > "Set sign".  It's in NV_vertex_program2 (and 3) and NV_gpu_program4.
> > 
> > * SSG:  Adds a new "set sign" operation, which produces a vector holding
> >   negative one for negative components, zero for components with a value
> >   of zero, and positive one for positive components.  Equivalent results
> >   could be achieved (less efficiently) in NV_vertex_program with
> >   multiple SLT, SGE, and arithmetic instructions.
> > 
> > We can either add a lowering pass to break ir_unop_sign into more
> > primitive operations, or drivers can implement the opcode.  I don't have
> > a strong opinion here.
> > 
> > > This one fails because r300 does not implement the DP2 opcode:
> > > glsl-fs-dot-vec2
> > 
> > Can r300 do a DP2?  I looked at the i915 documentation, and we can
> > implement it there using a DP2A.  We can't do a lowering pass for this
> > one, but we could have a flag to cause a different sequence to be
> > generated.  In that case, it may actually be better to just do it in the
> > driver.
> > 
> > > These fail because r300 does not implement the CONT opcode:
> > > glsl1-for-loop with continue
> > > glsl1-while-loop with continue
> > 
> > These should be fixed once loop unrolling is implemented.
> > 
> > > This one was fixed in master after master was merged into glsl2:
> > > glsl1-discard statement in for loop
> > > 
> > > Not sure why this fails yet:
> > > glsl-fs-loop-nested
> > 
> > This might also be unrolling related.  Some more data would be good.
> > 
> > > These fail on both r300g and llvmpipe with this error:
> > > tgsi/tgsi_ureg.c:746:ureg_emit_src: Assertion `src.File != TGSI_FILE_OUTPUT' failed.
> > > 
> > > glsl-orangebook-ch06-bump
> > > glsl-deadcode-varying
> > 
> > I guess that hardware can't read back outputs that have been written?
> > Ugh.  I'm not sure what to do about that one.  Maybe a lowering pass to
> > generate writes to a shadow copy that can be read?
> 
> The previous compiler called _mesa_remove_output_reads unconditionally,
> while the new compiler doesn't.  This should be left up to the backend
> in my opinion, so perhaps a little flag like Shader.EmitNoIfs and
> friends for r300g to set for now to call this for the Mesa IR generation
> would be a good solution.

It seems like this represents a change in the semantics of the mesa
driver interface -- previously mesa had the (perhaps implicit) rule that
OUT registers were write-only.  The glsl2 branch appears not to be
respecting that.

I don't really mind this as a follow-on change, but it seems like it
would be cleaner to keep the glsl2 merge as purely a change on one side
of the interface, not one which requires a change to drivers or the
interface itself.

Keith