60 Commits
Author SHA1 Message Date
Henrik RydgårdandClaude Opus 5.5 a9eacff1e1 SoftGPU: Keep the compile flushes on builds without the JIT
Skipping them is only a speed-up, but it changes how states get batched and
optimized, which changes the rendered output (NBA 2K13 and Virtua Tennis
frame dumps). Keep the old behaviour until that's understood.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>
2026-09-29 09:35:17 -06:00
Henrik RydgårdandClaude Opus 5.5 a9c782bfcf SoftGPU: Recompute the raster state after a JIT cache clear
A code space clear frees the functions the current state points to, but
the state was kept as long as the GE registers didn't change. Track the
clear generations and recompute, also when a compile during the state
computation clears the caches.

Also skip the binner flush for compiles on builds without the software
JIT, where Compile() does nothing.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>
2026-09-29 09:35:17 -06:00
Henrik Rydgård 3e198c53b2 More include cleanup 2024-12-18 13:57:26 +01:00
Henrik Rydgård c2bf09744a SoftGPU: Fix refactoring mistake where we could return an uninitialized value. Oops. 2023-09-15 10:01:28 +02:00
Henrik Rydgård 10f93875c6 Fix the semantics of DenseHashMap to be consistent even when inserting nulls 2023-09-11 12:07:18 +02:00
Unknown W. Brackets cd3fc26190 samplerjit: Prevent thread local stale cache read.
If the generation count happens to match, would still get a stale pointer
and crash.  Let's just make the generation count static so it always
increases.
2023-02-22 21:15:03 -08:00
Unknown W. Brackets 587a322207 softgpu: Use NEON SIMD for alpha blending, etc. 2023-01-07 19:06:34 -08:00
Unknown W. Brackets d9522a7ac5 softgpu: Avoid clear hazard for last cached funcs. 2022-12-06 21:23:56 -08:00
Unknown W. Brackets eda3ce556e softgpu: Avoid atomic structs.
Apparently we don't link libatomic and rather than fighting that, I'll
just use thread local values.
2022-12-06 20:35:07 -08:00
Unknown W. Brackets 400f6abf9a softgpu: Optimize lookup of last jit func.
This is common (for example, maybe a pixel state is updated but sampler is
not), and reduces time spent in ComputeRasterizerState() quite a bit in
Darkstalkers, where jits are available (i.e. Intel currently.)
2022-12-06 19:16:19 -08:00
Unknown W. Brackets 87fb9eef37 softgpu: Remove std::function usage.
Wanted to avoid coupling these, but don't like the std::function
construct/destructs showing in profiles...
2022-12-06 19:15:57 -08:00
Unknown W. Brackets 38eb0a7a82 softgpu: Check for queued compile.
Rarely, we could have queued compiling the same one, which would crash on
a double insert.
2022-12-03 12:15:58 -08:00
Unknown W. Brackets 778a0487cb softjit: Switch to DenseHashMap. 2022-12-02 20:59:13 -08:00
Unknown W. Brackets 4d06400548 softgpu: Fix compile hazard while running.
This prevents any clearing of cache while other threads may be using
previously cached funcs, and avoids wx exclusive hazards.
2022-11-20 12:04:02 -08:00
Unknown W. Brackets ce51942508 softgpu: Correct WX-exclusive platform hazards.
Should mainly affect BSD at this point.
2022-11-20 10:55:35 -08:00
Unknown W. Brackets c47d7eab38 softgpu: Simply 5551 blending fast path.
Since it only supports multiply and add, let's just stick with that.
2022-09-24 18:55:45 -07:00
Unknown W. Brackets 6e6535c263 softjit: Skip reading dst pixel where blended out.
Sometimes used by blends used purely to multiply the source color by
something, usually prep for bloom.
2022-09-24 02:00:03 -07:00
Unknown W. Brackets e72309745e softjit: Implement accurate fog color blending. 2022-09-11 08:50:07 -07:00
Unknown W. Brackets b90fc7137f softgpu: Correct accuracy of fog calculation.
This matches values from a PSP exactly, with the help of immediate mode
vertex values (since this directly allows specifying the fog factor
without any floating point math.)
2022-09-11 08:24:40 -07:00
Unknown W. Brackets 15d5fa48f7 softgpu: Check depth test early on simple stencil.
If we don't need to write stencil on sfail/zfail, we can do the depthtest
early, which allows us to more often skip texture sampling.

This gives a good improvement in Chains of Olympus.
2022-09-10 21:24:19 -07:00
Unknown W. Brackets 2479d52202 Global: Reduce includes of common headers.
In many places, string, map, or Common.h were included but not needed.
2022-01-30 16:35:33 -08:00
Unknown W. Brackets 6c723c0517 softjit: Fix src blend factor handling.
This was causing us to skip a shift, oops.
2022-01-24 00:05:00 -08:00
Unknown W. Brackets c2985bca31 softjit: Centralize some common funcs from sampler.
No need to duplicate this code.
2022-01-19 00:03:59 -08:00
Unknown W. Brackets ac2b96cec0 softjit: Switch to constant pool.
This is simpler without RIP access checks, and tends to be fast.
2022-01-17 19:50:37 -08:00
Unknown W. Brackets fc292b127b softgpu: Correct dither matrix lookup.
Oops, need to wrap x/y, of course...
2022-01-15 23:51:21 -08:00
Unknown W. Brackets f4f7ea2736 softgpu: Cache colortest params in draw pix state. 2022-01-15 13:03:11 -08:00
Unknown W. Brackets aa9d751248 softgpu: Cache alpha/stencil test masks in state. 2022-01-15 13:03:11 -08:00
Unknown W. Brackets acad2640dd softgpu: Cache logicOp in draw pixel state. 2022-01-15 13:03:10 -08:00
Unknown W. Brackets c0d548846f softgpu: Use cached write mask in draw pixel. 2022-01-15 13:03:10 -08:00
Unknown W. Brackets f1ce2e7715 softgpu: Cache minz/maxz in draw pixel state. 2022-01-15 13:03:10 -08:00
Unknown W. Brackets 0b3f096c01 softgpu: Cache strides in draw pixel state. 2022-01-15 13:03:10 -08:00
Unknown W. Brackets e9f3720e20 softgpu: Cache fog color draw pixel state. 2022-01-15 13:03:10 -08:00
Unknown W. Brackets 9ec7d65c49 softgpu: Use func IDs instead of gstate more. 2022-01-10 22:12:35 -08:00
Unknown W. Brackets 961cfcd75c softjit: Add describes here too.
Helpful to aggregate when there are multiple rasterizers.
2022-01-03 06:45:10 -08:00
Unknown W. Brackets e93c709f5c sofjit: Correctly poison memory.
Noticed this wasn't breakpoints when reviewing some assembly output.
2022-01-02 08:47:04 -08:00
Unknown W. Brackets 355bad666c softjit: Optimize common case bloom blending.
Bloom often uses fixed ONE + ONE, which is a lot less work for us.  And
bloom often runs over and over again on pixels, so saving work is good.
2022-01-02 08:47:04 -08:00
Unknown W. Brackets 2d8fdd8cf4 Math3D: Allow construction from NEON vectors.
This makes it match SSE and easier to keep things generic.  Will impact
alignment of non-packed Vec2/Vec3.
2021-11-28 08:24:53 -08:00
Unknown W. Brackets 96a7554053 sofjit: Move common types to reg cache header.
This makes it easier to use vectors elsewhere.
2021-11-28 08:03:15 -08:00
Unknown W. Brackets 3d5bced296 softjit: Rename reg cache so it can be reused.
Intentionally just the name changes in this commit.
2021-11-28 08:03:15 -08:00
Unknown W. Brackets 4703b6cb56 softjit: Cleanup, add other arch types to regcache. 2021-11-28 08:03:15 -08:00
Unknown W. Brackets c1882fa1c0 softjit: Disallow use of register after unlock. 2021-11-28 08:03:14 -08:00
Unknown W. Brackets 2f039abd13 softjit: Simplify regcache usage as purpose only.
Dealing with types was annoying, and this helps validate the right
register is released.
2021-11-28 08:03:14 -08:00
Unknown W. Brackets e1ed49a3e4 softjit: Ensure all regs are released. 2021-11-28 08:03:14 -08:00
Unknown W. Brackets d53e13b862 softjit: Manage args in the register cache. 2021-11-28 08:03:13 -08:00
Unknown W. Brackets 1cb48a7bd2 softjit: Reduce jit pool size a bit. 2021-11-26 10:30:00 -08:00
Unknown W. Brackets c62457bb33 softjit: Optimize common blend inverse alpha case. 2021-11-26 09:30:48 -08:00
Unknown W. Brackets a07017dbb0 softjit: Prefer easier to refill regs. 2021-11-26 09:30:47 -08:00
Unknown W. Brackets 7f167c3660 softjit: Implement min/max/absdiff blending.
Alpha not yet implemented.
2021-11-26 09:30:47 -08:00
Unknown W. Brackets 2b4b4ae064 softjit: Add config setting to enable/disable.
Also use it for samplerjit.
2021-11-26 08:21:14 -08:00
Unknown W. Brackets 0e63b357b3 softjit: Add dithering. 2021-11-26 08:21:13 -08:00