Age | Commit message (Collapse) | Author |
|
|
|
Share almost all code with r500_fragprog now.
This also fixes Piglit's texrect-many test, which means that the compiz
bicubic plugin should work with hardware acceleration now.
|
|
|
|
|
|
This fixes a regression introduced by dea8719f0...
|
|
|
|
Missed the homogenous divide of R by Q before...
|
|
This new generic transform replaces "special" instructions by more generic
variants. Hopefully, we will be able to share this code between r300 and r500.
|
|
MOV, ADD and MUL do not fit the hardware as well as MAD, but they are less
complex and thus leave more room for future optimizations.
|
|
|
|
|
|
Streamlining source and destination registers, as well as texcoord scaling for
RECT textures is now done in a radeon_program based transformation.
The idea is that this will allow us to optimize away unnecessary indirections
more easily.
|
|
The idea/hope is that radeon_program will serve as an intermediate
representation for r3xx up to r6xx fragment and vertex programs.
Right now, it is nothing more than a simplistic wrapper around Mesa's
prog_instruction, together with the notion of clauses, taken from r6xx docs.
The clauses will eventually be used to represent the nodes that are used in
r300 family fragment programs.
|
|
|
|
|
|
Refactor so that r300_pfs_compile_state "owns"/holds a pointer to
r300_fragment_program instead of the other way round. This is more natural from
an object orientation point of view.
Move the compiled hardware state into r300_fragment_program_code, in
anticipation of on-the-fly program recompilation based on external OpenGL
state.
|
|
|
|
r500 code still used r300_pfs_compile_state, which contains some fields that
really only make sense on r3xx type hardware. In order to allow both fragprog
implementations to go forward without disturbing each other, I've pushed this
structure down into the respective r[3|5]00_fragprog.c
|
|
Setup fg_depth_src for depth writing programs and change early Z (ztop)
semantics.
Piglit's version of glean/fragprog test passes now (unlike Glean, its
dependency on EXT_fog_coord, which we don't support, is optional).
R3xx only at the moment, but should be straightforward to adapt to R5xx
(I don't own an R5xx, and I don't want to break anything.)
|
|
The KIL instruction only works if at least one texture unit is enabled
in hardware.
Texture instructions do not support swizzles, negations etc. natively,
so we now emit an explicit swizzling etc. operation when the texture coordinate
requires it.
This fixes the Piglit fp-kil test.
|
|
|
|
Includes fallback shader and a handful of working opcodes.
|
|
This bug was introduced by commit 978145a075255ae153ee05c2a037400e61558079.
|
|
|
|
Just for consistency; most of the code already uses __FUNCTION__.
|
|
|
|
|
|
|
|
d28f6d91760374e2eb71b541b0f259f81dd73c69.
|
|
Note there might be some calls to the old function names in conditionally
disabled code, but I think I've got them all.
|
|
|
|
|
|
|
|
|
|
The texture_rectangle fix introduced a bug where every texture instruction
caused a new indirection.
|
|
R300 hardware takes texcoords in the range 0..1 even for rectangle
textures. Previously, the necessary texcoord conversion was applied
to the texture coordinate during vertex processing in a render stage.
This is obviously wrong when fragment programs are used, which can
calculate arbitrary coordinates for TEX instructions. Therefore,
we now inject an appropriate MUL instruction before a TEX that
reference a rectangle texture.
|
|
|
|
The index is no longer necessary to share constants between multiple
SIN/COS/SCS instructions inside a single fragment program, and storing
a tiny implementation detail like this in the fragment_program structure
itself was just nasty.
|
|
The constant/parameter allocation was significantly simplified, removing
one unnecessary copy operation of parameters. The dirty state tracking is
unchanged and far from optimal, since all state is always re-fetched.
Constants and parameters are now emitted only once, which significantly
reduces the resource pressure on larger programs.
|
|
Also, fix a typo in a related comment.
|
|
Make sure that instruction slots are fully initialized with NOPs during
find_and_prepare_slot(). This fixes a bug when a fragment program was
translated more than once (e.g. due to a second call to glProgramStringARB).
This partially fixes glean/fragProg1.
|
|
|
|
|
|
This is a necessary change to emit the right instructions when writing
to result.depth.
However, even with this test, Z-write doesn't work properly, and I don't
fully understand why. In addition to this, we'll at least have to disable
early-Z, but even that doesn't seem to be enough.
|
|
Fix a bug in the LIT implementation (clamp exponent to 128, not 0.5)
and change the implementation around. In theory, the new implementation
needs as little as 5 instruction slots. Unfortunately, the dependency
analysis in find_and_replace_slot is not strong enough to look at
individual components of a register yet.
|
|
Do not move an instruction that writes to a temp forward past an instruction
that reads the same temporary.
|
|
|
|
Dumps of fragment programs were incorrect when the program consisted of multiple
nodes.
Also, improved the formatting a bit.
|
|
There were a number of bugs related to the pairing of vector and scalar
operations where swizzles ended up using the wrong source register,
or an instruction was moved forward and ended up overwriting an aliased
register.
The new algorithm for register allocation is quite conservative and may
run out of registers before necessary. On the plus side, It Just Works.
Pairing is done whenever possible, and in more cases than before, so
in practice this change should be a net win.
|
|
|