- Renumber VERT_RESULT_* values so that they match the
numbers of the corresponding FRAG_ATTRIB_ values.
- Add ctx->VertexProgram._Current and FragmentProgram._Current
values which point to either the current client-supplied
program if enabled, or otherwise to the current mesa-internal
program. Thus this program is always the correct one for
the current state, providing that the mesa flags to turn on
automatic generation are enabled.
- Add callbacks to ctx->Driver.BindProgram() in texenvprogram.c
and t_vp_build.c so that the driver knows when the generated
program changes. This is cleaner than trying to code all the
possible _NEW_* flags into the driver, and more precise as well.
- Add a UsesKill flag to identify fragment programs with that
instruction, as these can require special treatment.
- Move the FRAG_OUTPUT values to mtypes.h, near to similar defn's.
-Use 16-bit elts in vtxfmt_a path if possible
-Optimize VSF param uploading
-return in r300DepthMask looks suspicious, use r300Enable instead
-Dont use r300ResetHwState in invalidate state(disabled, missing hooks, possible instabilities)
temporaries, parameters, etc).
glGetProgramivARB() now returns all the right things.
Updated i915 and r300 code to initialize program native limits and
current program's native instruction/temporary/etc counts.
Specifically, don't attach a dummy program.
This change also fixes an occasional segfault.
Some code clean-ups. Use GLboolean instead of GLuint to return pass/fail.
it, and guessing that the two remaining entries in the 3-bit values were the
new funcs. Tested with modified stencilwrap test. Also, remove the commented
fallback stuff -- more modification to stencilwrap suggests that those issues
were just from span readback, not stencil implementation.
stencilwrap reported many issues with various modes. Some of these were
complicated by the fact that spans are broken (Bug #1615), but some appear to be
real bugs. However, while spans remain broken, I found that visual results were
better by avoiding fallbacks rather than avoiding just a broken stencil
implementation. Note that this required changing the depth spans at 24+8bpp
into read-modify-write cycles. It would be nicer as a single write with
a mask, but the kernel span blits turn off masking.
Reviewed by: ajax
back on. Tested using seccolor modified to use the blue channel instead of
green, since green stays in the same place across RGB/BGR mistakes. Also hook
in UpdateSpecular on COLOR_EXT change, which might have resulted in missing
statechanges before.
not respecting the coords (or perhaps interpreting them differently?) in my
testing. However, in the process it led to a fix of a secondary color handling
issue where it would be taken from the wrong offset, I believe, based off of
reading the r200 driver. Also add a minor tweak to save time in the
fog-but-not-specular case.
