| Age | Commit message (Collapse) | Author |
|---|
|
This provides the optimizer with hints about code hotness, which we're
quite certain about for debug printouts (or, rather, while we
developers often hit the checks for debug printouts, we don't care
about performance while doing so).
|
|
|
|
This fixes some insanity that would otherwise be required for GLSL
1.30 bit ops or gen6 integer uniform operations in general, at the
cost of upload-time pain. Given that we only have that pain because
mesa's mangling our integer uniforms to be floats, this something that
should be fixed outside of the shader codegen.
|
|
The assumption is that all stages are the same program or that
varyings are passed between stages using built-in varyings.
|
|
It's a little more painful than before because we don't have the handy
mask register any more, and have to make do with cooking up a value
out of the flag register.
|
|
This doesn't appear to help any testcases I'm looking at, but it looks
like it's required.
|
|
Fixes glsl-fs-uniform-array-5, but not 6 which fails in ir_to_mesa.
|
|
This makes it a lot easier to track down where we failed when some
code emit triggers an assert. Plus, less memory allocation for
codegen.
|
|
Fixes glsl-fs-convolution-2, which was blowing up due to the array
access insanity getting at the uniform values within the loop. Each
temporary was considered live across the whole loop.
|
|
One, it was allocating increments of 1kb, but per thread scratch space
is a power of two. Two, the new FS wasn't getting total_scratch set
at all, so everyone thought they had 1kb and writes beyond 1kb would
go stomping on a neighbor thread.
With this plus the previous register spilling for the new FS,
glsl-fs-convolution-1 passes.
|
|
It can be tested with if (0) replaced with if (1) to force spilling for all
virtual GRFs. Some simple tests work, but large texturing tests fail.
|
|
It's amazing this code worked. Basically, we would get lucky in
register allocation and the tests using frontfacing would happen to
allocate gl_FrontFacing storage and the instructions generating
gl_FrontFacing but pointing at another register to the same hardware
register. Noticed during register spilling debug, when suddenly they
didn't get allocatd the same storage.
|
|
|
|
"Everyone else" does it this way, so follow suit. It's fewer
instructions, anyway.
|
|
|
|
XOR makes much more sense. Note that the previous code would have
failed for not(not(x)), but that gets optimized out.
|
|
|
|
Otherwise, it would try to handle arrays as structures, use
uninitialized memory, and crash.
|
|
We often use reg_null as the destination when setting up the flag
regs. However, on gen6 there aren't general implicit conversions to
destination types from src types, so the comparison to produce the
flag regs would be done on the integer result interpreted as a float.
Hilarity ensued.
Fixes 20 piglit cases.
|
|
|
|
Previously _LinkedShaders was a compact array of the linked shaders
for each shader stage. Now it is arranged such that each slot,
indexed by the MESA_SHADER_* defines, refers to a specific shader
stage. As a result, some slots will be NULL. This makes things a
little more complex in the linker, but it simplifies things in other
places.
As a side effect _NumLinkedShaders is removed.
NOTE: This may be a candidate for the 7.9 branch. If there are other
patches that get backported to 7.9 that use _LinkedShader, this patch
should be cherry picked also.
|
|
I broke it in 06fd639c519214b6ebcbf29127b6d9ed429f8641 by only testing
2 generations of hardware :(
|
|
|
|
It is now to the point where we have no regressing piglit tests. It
also fixes Yo Frankie! and Humus DynamicBranching, probably due to the
piglit bias tests that work that didn't on the Mesa IR backend.
As a downside, performance takes about a 5-10% performance hit at the
moment (e.g. nexuiz 19.8fps -> 18.8fps), which I plan to resolve by
reintroducing 16-wide fragment shaders where possible. It is a win,
though, for fragment shaders using flow control.
|
|
The existing code used RNDD, which rounds down, rather than toward zero.
|
|
Fixes piglit test glsl-fs-ceil.
|
|
This cuts usually 2 out of 3 instructions for flag reg generation (if
statements, conditional assignment) by producing the conditional mod
in the expression representing the boolean value.
Fixes glsl-fs-vec4-indexing-temp-dst-in-nested-loop-combined (register
allocation no longer fails for the conditional generation
proliferation)
|
|
This will be a place to peephole comparisions directly to the flag
regs, and for now avoids using MOV with conditional mod on gen6, which
is now illegal.
|
|
Improves nexuiz performance 0.91% (+/- 0.54%, n=8)
|
|
|
|
So far, I've only seen this be a valgrind warning and not a real failure.
|
|
|
|
|
|
Fixes glsl-fs-i2b.
|
|
It's now much more correct for gen6 than the old backend, with just 2
regressions I've found (one of which is common with pre-gen6 and will
be fixed by an array splitting IR pass).
This does leave the old Mesa IR backend getting used still when we
don't have GLSL IR, but the plan is to get GLSL IR input to the driver
for the ARB programs and fixed function by the next release.
|
|
Pre-gen6, you could mix int and float just fine. Now, you get goofy
results.
Fixes:
glsl-arb-fragment-coord-conventions
glsl-fs-fragcoord
glsl-fs-if-greater
glsl-fs-if-greater-equal
glsl-fs-if-less
glsl-fs-if-less-equal
|
|
This is a hw requirement in math args. This also is inefficient, as
we're calculating the same result 8 times, but then we've been doing
that on pre-gen6 as well. If we're doing math on uniforms, though,
we'd probably be better served by having some sort of mechanism for
precalculating those results into another uniform value to use.
Fixes 7 piglit math tests.
|
|
|
|
This was leftover from the pre-gen6 cleanups. One tests regresses
where compute-to-MRF now occurs.
|
|
This didn't produce a statistically significant performance difference
in my demo (n=4) or nexuiz (n=3), but it still seems like a good idea
and is recommended by the HW team.
|
|
|
|
This is progress towards enabling a compute-to-MRF pass.
|
|
It's time to start splitting some of this up.
|
|
While I don't know of any performance changes from this (once extra
reg available out of 128), it makes the generated asm a lot cleaner
looking.
|
|
Having the single opcode write then read the reg meant that single
instruction opcodes had to consider their source regs to interfere
with their dest regs.
|
|
Improves performance of my GLSL demo 14.3% (+/- 4%, n=4) by
eliminating the moves used in ir_assignment and ir_swizzle handling.
Still 16.5% to go to catch up to the Mesa IR backend, presumably
because instructions are almost perfectly mis-scheduled now.
|
|
We'd overwrite the same element twice.
|
|
Fixes glsl-fs-texturecube-2-*
|
|
|
|
We sensibly only provide it if the FS asks for it. We could actually
skip WPOS unless the FS needed WPOS.zw, but that's something for
later.
Fixes: glsl-texture2d and probably many others.
|
