| Age | Commit message (Collapse) | Author |
|---|
|
|
|
The original allocations use regs->regs as the context, so talloc will
happily ignore the context given here. Change it to match to clarify
that it isn't changing.
|
|
This revealed a bug in ra_get_spill_benefit where we only considered
the benefit of the first adjacency we were to remove, explaining some
of the ugly spilling I've seen in shaders. Because of the reduced
spilling, it reduces the runtime of glsl-fs-convolution-1 36.9% +/-
0.9% (n=5).
|
|
|
|
Reduces runtime of glsl-fs-convolution-1 another 13.9% +/- 0.6% (n=5).
|
|
This was recommended in the original paper, but I figued "make it run"
before "make it fast". Now we make it fast. Reduces the runtime of
glsl-fs-convolution-1 by 12.7% +/- 0.6% (n=5).
|
|
Our use of the register allocator in i965 is somewhat unusual.
Whereas most architectures would have a smaller set of registers with
fewer register classes and reuse that across compilation, we have 1,
2, and 4-register classes (usually) and a variable number up to 128
registers per compile depending on how many setup parameters and push
constants are present. As a result, when compiling large numbers of
programs (as with glean texCombine going through ff_fragment_shader),
we spent much of our CPU time in computing the q[] array. By keeping
a separate list of what the conflicts are for a particular reg, we
reduce glean texCombine time 17.0% +/- 2.3% (n=5).
We don't expect this optimization to be useful for 915, which will
have a constant register set, but it would be useful if we were switch
to this register allocator for Mesa IR.
|
|
|
|
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.
|
|
Notably missing is choice of registers to spill.
|
