Jay's novel SSA-based register allocator relies on a fixed partition of Intel
GRFs mapping to logical GPRs.
Previously, Jay used a simple partitioning scheme, which was good enough for
simple compute and fragment shaders, but has both limitations preventing new
feature bring-up and performance issues.
Here we rewrite the Jay partitioning code at the heart of the Jay RA in
order to lift these restrictions and allow fully flexible partitions. This
should be easier to reason about, fix a bunch of issues around simd32 payloads,
enable better performance, etc.
The # of stride 16 GRFs reserved is halved in simd32 mode here to match how
multisampling stuff works, which explains the large simd32-only instruction
count reduction.
While churning all this code, I took the opportunity to break off
jay_partition.c... I think that is better organized and the diff was garbage
otherwise.
SIMD16:
Totals from 2189 (82.70% of 2647) affected shaders:
Instrs: 2702159 -> 2670951 (-1.15%); split: -1.41%, +0.26%
CodeSize: 40296128 -> 39850304 (-1.11%); split: -1.40%, +0.30%
SIMD32:
Totals from 2373 (89.65% of 2647) affected shaders:
Instrs: 4559418 -> 4072897 (-10.67%); split: -10.77%, +0.10%
CodeSize: 68185488 -> 60635616 (-11.07%); split: -11.17%, +0.09%
Number of spill instructions: 44069 -> 44055 (-0.03%)
Number of fill instructions: 43292 -> 43278 (-0.03%)
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/41808>
Implement a simple pre-RA bottom-up list scheduler with the goal of decreasing
register pressure. On Xe2, this significantly reduces spilling.
SSA form allows us to estimate register demand cheaply and accurately, which
theoretically [1] gives this algorithm the two Hippocratic properties:
1. Shaders with low register pressure are unaffected.
2. Register pressure can only be decreased, never increased.
In other words: first, do no harm.
The heuristic itself is very simple: greedily choose instructions that decrease
liveness using a backwards list scheduler. This is far from optimal! But thanks
to the above properties, even a heuristic that picked random instructions would
be a win overall - by construction, we can only ever win.
In other words: this scheduler is your older brother powering off the game
console any time he's about to lose a game, maintaining a 100% win rate.
[1] In reality, neither property is strictly satisfied due to the messy details
of mapping our clean logical model onto Intel's many weird physical register
files. Nevertheless, the algorithm is well-motivated and the empirical results
on Xe2 are excellent.
SIMD16:
Totals:
Instrs: 2754194 -> 2753957 (-0.01%); split: -0.23%, +0.22%
CodeSize: 41094768 -> 41092768 (-0.00%); split: -0.23%, +0.23%
Number of spill instructions: 1724 -> 1129 (-34.51%)
Number of fill instructions: 1912 -> 1119 (-41.47%)
Totals from 168 (6.35% of 2647) affected shaders:
Instrs: 850994 -> 850757 (-0.03%); split: -0.75%, +0.73%
CodeSize: 12825680 -> 12823680 (-0.02%); split: -0.74%, +0.73%
Number of spill instructions: 1724 -> 1129 (-34.51%)
Number of fill instructions: 1912 -> 1119 (-41.47%)
SIMD32:
Totals:
Instrs: 4688858 -> 4557800 (-2.80%); split: -3.53%, +0.74%
CodeSize: 70177200 -> 68214816 (-2.80%); split: -3.53%, +0.74%
Number of spill instructions: 50316 -> 45795 (-8.99%); split: -9.56%, +0.57%
Number of fill instructions: 51526 -> 45075 (-12.52%); split: -13.23%, +0.71%
Totals from 819 (30.94% of 2647) affected shaders:
Instrs: 3810182 -> 3679124 (-3.44%); split: -4.35%, +0.91%
CodeSize: 57044000 -> 55081616 (-3.44%); split: -4.35%, +0.91%
Number of spill instructions: 49264 -> 44743 (-9.18%); split: -9.76%, +0.58%
Number of fill instructions: 50182 -> 43731 (-12.86%); split: -13.58%, +0.73%
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/41688>
on top of scheduler changes, compile-time of shaders/blender/1017.shader_test:
Difference at 95.0% confidence
-0.00173202 +/- 0.00116931
-0.791537% +/- 0.532384%
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/41688>
Greedy post-RA substitution pass, similar to IGC's AccSubstitution pass.
Stats together with the previous commits.
SIMD16:
Totals from 2209 (83.45% of 2647) affected shaders:
Instrs: 2701029 -> 2696350 (-0.17%)
CodeSize: 39166720 -> 40372272 (+3.08%); split: -0.36%, +3.44%
SIMD32:
Totals from 2211 (83.53% of 2647) affected shaders:
Instrs: 4691165 -> 4641188 (-1.07%)
CodeSize: 69365792 -> 69341616 (-0.03%); split: -0.50%, +0.47%
The instruction count reduction is from RA shuffle code getting coalesced via
accumulators. The code size changes are from:
* Fewer moves from the instr count reduction (helped)
* Smaller MADs encoded as MACs (helped)
* Fewer SYNC.nop due to fewer scoreboarding annotations (helped)
* Less compaction due to explicit accumulator operands (hurt)
I expect significant cycle count changes from this but we don't have a cycle
model wired up yet, so reading the assembly will have to do.
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/41398>
Now instructions still read/write UFLAG, which preserves the information about
lane 0 we need for proper predication etc.
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/41215>
This is totally broken now that we have a physical CFG for UGPRs. And of course,
UGPRs generally were totally broken without the physical CFG. So I conclude
this code basically never worked. Which is good because it was also basically
always dead too. Just delete it and replace with a clear error message, instead
of pretending it works and either randomly splatting validation or just straight
up miscompiling silently or whatever.
We might need an alternative UGPR->GPR spill path some day but that day is not
today.
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/41215>
Consider:
u0 = foo()
if (divergent) {
u0 = bar()
r0 = baz(u0)
} else {
r0 = quux(u0)
}
Logically, this is fine, there is no interference between bar() and u0. But
physically, both sides of the if execute so the bar() write to u0 overwrites the
variable the else reads. So this is a miscompile.
The solution is to model the extra edges in the physical control flow graph,
which lives next to the existing logical control flow graph. Liveness for UGPRs
now follows the physical CFG, while liveness for GPRs continues to follow the
logical CFG. That models the interference properly, while still allowing phis to
work as before (since phis writing UGPRs follow uniform bits of control flow
that are necessarily critical edge free for the same reason the logical CFG is).
Because our RA copies shuffled registers back at block ends (following
Colombet), there's no issue with live range splits here (unlike aco which
inserts phis for this case and then needs to worry about critical edges around
those phis).
There might still be an extremely-challenging-to-hit bug here with UGPR spilling
which I need to think more about. It might be fine as-is? Not convinced though.
But this is big enough and strictly less broken than what we have right now and
the full solution will build on this, so here we are.
Fixes artefating in SuperTuxKart and Celestia knows what else.
Totals:
Instrs: 2770938 -> 2771269 (+0.01%); split: -0.00%, +0.02%
CodeSize: 40133712 -> 40138480 (+0.01%); split: -0.01%, +0.02%
Totals from 158 (5.97% of 2647) affected shaders:
Instrs: 514523 -> 514854 (+0.06%); split: -0.02%, +0.09%
CodeSize: 7603040 -> 7607808 (+0.06%); split: -0.03%, +0.09%
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/41215>
Jay is a new SSA-based compiler for Intel GPUs. This is an early
work-in-progress. It isn't ready to ship, but we'd like to move development in
tree rather than rebasing the world every week. Please don't bother testing yet
- we know the status and we're working on it!
Jay's design is similar to other modern NIR backends, particularly ACO, NAK and
AGX. It is fully SSA, deconstructing phis after RA. We use a Colombet register
allocator similar to NAK, allowing us to handle Intel's complex register
regioning restrictions in a straightforward way. Spilling logical registers is
straightforward with Braun-Hack.
Thanks to the SSA-based design, the entire backend is essentially linear time,
regardless of register pressure, addressing brw's excessive compile time when
especially spilling with brw.
In this current early draft, we support a limited subset of all three APIs on
Xe2. A lot works and a lot doesn't. The core compiler is there (spilling,
scoreboarding, SIMD32, etc should more or less work), but there are details to
fill in for both performance and correctness. We essentially pass conformance on
OpenGL ES 3.0 and OpenCL 3.0, and we're busy iterating on Vulkan.
Likewise, additional hardware support will come down the line. There's nothing
fundamentally Xe2-specific here. I just have a Lunarlake laptop on my desk, Ken
has a Battlemage card, and we had to pick _something_ as the first target.
Co-authored-by: Kenneth Graunke <kenneth@whitecape.org>
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/40835>
