c7fc448869
softgpu: Use some SSE4 in triangle interpolation.
2022-01-08 11:38:07 -08:00
3b1cc0d3b8
softgpu: Limit minX/maxX per line.
...
Only helps when single-threaded, though.
2022-01-08 10:04:52 -08:00
7594187538
softgpu: Skip sample lookup if masked.
...
Was hoping making other things faster would make this unnecessary or
worse, but it hasn't seemed to. This gives a pretty decent improvement in
most places (~4%.)
2022-01-02 13:47:14 -08:00
a0fe4d06bf
softgpu: Stop specializing on miplevels.
...
Now that samplerjit is processing mips, it no longer helps. Just
complexity now.
2022-01-02 13:47:14 -08:00
d3f0af7458
Merge pull request #15273 from unknownbrackets/softjit-bloom
...
Optimize software renderer handling of common bloom operations
2022-01-02 18:11:07 +01:00
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
496545e55c
softgpu: Add code for tracking GPU writes.
...
Unfortunately, it has a pretty noticeable speed impact, even at the basic
"assume everything's written" level. Compiled off by default, but at
least it's there.
Doesn't account for tests (i.e. alpha test skipping write) so still not
perfectly accurate.
2022-01-02 08:28:30 -08:00
cb1f26122d
Merge pull request #15269 from unknownbrackets/softgpu-opt
...
softgpu: Reduce interpolation if not needed
2022-01-02 09:47:19 +01:00
da38c027b5
Merge pull request #15268 from unknownbrackets/samplerjit-nearest
...
Implement nearest in samplerjit, like linear
2022-01-02 09:46:29 +01:00
025ac99f2f
softgpu: Reduce interpolation if not needed.
...
About 3% gain in some areas.
2022-01-01 18:34:04 -08:00
40240be91c
samplerjit: Update nearest args, temp disable jit.
...
This temporarily disables jit for nearest, but refactors to use the new
arg structure. It now matches linear.
2022-01-01 16:58:05 -08:00
06e954fe2a
samplerjit: Create a separate fetch func.
...
This allows nearest to become more similar to linear, where it applies the
texture function.
2022-01-01 16:58:04 -08:00
d41e42d247
softgpu: Correct off-by-one scissor mask.
...
Fixes Brave Story in the software renderer. Was overwriting display list
data in the stride gap.
2022-01-01 16:42:36 -08:00
b35ca3d472
softgpu: Cleanup min/max tri range handling.
...
The previous looked like it had off by one errors. This is simpler.
2022-01-01 16:42:36 -08:00
12405709f0
softgpu: Skip processing scissored triangles.
...
If only one side was scissored (common), we might even put it on a thread,
which ended up as a lot of overhead. Gives 3-4% improvement in some
places.
2022-01-01 16:40:34 -08:00
33e9841a4a
softgpu: Skip zero size triangles.
...
These were drawing before, incorrectly, which caused artifacts.
Noticeable in Blade Dancer.
2021-12-31 00:20:12 -08:00
4bd94a4e5e
samplerjit: Pass funcs as an argument.
...
Seeing computing the ID in some profiles, so want to avoid computing per
thread/invocation.
2021-12-29 07:11:53 -08:00
74eb450e76
samplerjit: Move texture function into jit.
...
Could do this also for nearest, might end up with a third set of functions
there for a direct sample lookup (for debug funcs.)
2021-12-28 17:52:17 -08:00
940e6bb1d7
samplerjit: Lookup both mip tex values.
2021-12-28 16:22:54 -08:00
6b55d328e5
samplerjit: Use regcache for linear filtering.
...
This makes it easier to reuse for mipmap filtering.
2021-12-28 15:37:25 -08:00
a4558a5736
samplerjit: Take texptr/bufw as arrays.
...
Prep for moving mip map sampling into linear.
2021-12-28 12:04:16 -08:00
a84accf713
samplerjit: Move S/T calculation into jit.
...
Gives a pretty decent 5-10% improvement in many places.
2021-12-28 09:58:23 -08:00
9cc0883d53
softgpu: Correct non-SSE T clamp.
2021-12-27 15:31:37 -08:00
39d5b1c221
softgpu: Reduce mipmap fraction to 4 bits.
...
For CONST (and SLOPE with flat w), this produces accurate values.
SLOPE is still wrong in its handling of w, and AUTO seems to calculate
using a different and less accurate ramp. But they both produce values
with 16 steps, in any case.
2021-12-27 11:37:33 -08:00
d6b6ef4cb1
softgpu: Correct nearest filtering too.
...
Turns out to have the same behavior as linear, when it comes to the
subpixel offset.
2021-12-27 11:37:33 -08:00
1dfaea9062
softgpu: Remove no longer possible report.
...
Also, it's known how this behaves, now.
2021-12-27 11:37:33 -08:00
75f105f84b
softgpu: Make linear filtering more accurate.
...
This matches tests for various u/v offsets and x/y subpixel offsets.
Mipmaps are probably still wrong.
2021-12-27 11:37:32 -08:00
b00a66e34c
samplerjit: Pass u/v coords as vector.
2021-12-27 11:37:32 -08:00
3180e6c043
softgpu: Correct alpha on add + invalid texfuncs.
2021-12-05 16:28:37 -08:00
325a1f75aa
softgpu: Match texenv blend texfunc accurately.
2021-12-05 16:09:26 -08:00
0b6e7c421f
softgpu: Make decal tex func more accurate.
...
Tested for all values of A * B + 0 * (255 - B), as well as A * 127 + B *
(255 - 127), and matches accurately. Spot checked other values, but not
exhaustively.
2021-12-05 13:34:19 -08:00
73460f7461
softgpu: Correct accuracy of MODULATE texfunc.
...
This matches hardware tests for every value of A * B.
Interesting that it's a different formula than alpha blend.
2021-12-05 12:06:52 -08:00
891fa8c613
softgpu: Template away uncommon mip usage.
...
Improves general case about 10%.
2021-12-04 15:45:06 -08:00
48e9404419
softgpu: Remove useless switch by UV gen mode.
...
They're all handled earlier now, and the switch is on a value & 3, so the
default wasn't even possible.
2021-12-04 15:45:06 -08:00
823c4adb15
softgpu: Keep arguments in vectors for sampling.
2021-12-04 15:45:06 -08:00
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
91787e63d9
softjit: Switch to the __vectorcall convention.
2021-11-26 08:21:11 -08:00
ce5ae95854
softgpu: Correct alpha blend subtract on negative.
...
Oops, we need to subtract signed, but then clamp to unsigned.
2021-11-25 22:06:48 -08:00
35444b3051
softgpu: Accurately alpha blend.
2021-11-25 19:23:41 -08:00
876c8cd368
softgpu: Fix PixelFuncID size.
...
Oops, can't use unions in bitfields. Also improve typesafety.
2021-11-21 09:40:13 -08:00
09dc38080a
softgpu: Move draw pixel code to separate file.
...
This separates things better anyway. No major perf impact.
2021-11-21 08:23:32 -08:00
824805ec1e
Merge pull request #15154 from unknownbrackets/softjit
...
Use a pixel func ID in software rendering
2021-11-21 10:50:06 +01:00
e2f0713cc2
softgpu: Clamp and round fog by mantissa bits.
...
This matches hardware calculated fog values much better.
2021-11-20 20:54:52 -08:00
9abf2a4725
softgpu: Confirm mask doesn't hit stencil REPLACE.
2021-11-20 18:53:51 -08:00
f7a31c992d
softgpu: Use pixel func ID to draw pixels.
...
This just reduces reliance on gstate directly, and should help keep things
consistent.
2021-11-20 15:27:04 -08:00
b6bdd69572
softgpu: Clear by dividing out subpixel first.
2021-11-15 06:26:11 -08:00
f802c3bc6d
softgpu: Add some comments and cleanup.
2021-11-15 06:09:12 -08:00
babd63c644
softgpu: Tune thread minimums better.
...
Darkstalkers seems more sensitive to these than many other games, this
improves performance more.
2021-11-14 18:44:30 -08:00
66f635cba0
sfotgpu: Use threads to apply clears.
2021-11-14 18:31:46 -08:00
f66e243727
softgpu: Correct scissor for pixel centers.
2021-11-07 11:19:41 -08:00
Unknown W. Brackets