The light color products went through Vec4<int>'s multiply, a 32-bit
lane multiply that SSE2 emulates, though their factors are at most 10
bits. MulSmall uses CrossSIMD's Mul16 where Vec4<int> has a register.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
GEDot3 and GERowSum called ldexpf for each result, and the UV plane
setup ldexp and ilogb for every triangle and tile, about 2% of the CPU
time between them. Adding to the exponent field gives the same values
while they stay normal (ScaleByPow2Exact, which calls ldexpf otherwise),
and the planes' values and shifts always are. 2-4% fewer instructions
in every game measured.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
ApplyGESkinning took a vertex at a time with its components in three of
four lanes. Now four vertices go through each bone together, a lane
each, with the bone matrix entries, the weights and the positions taken
apart once for the GE's multiply (GEMulOperand). A bone whose weight is
zero in some lanes adds zeros there, leaving the sums as skipping it
would. The per-component order of the adds is the GE's as before.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A draw's vertices now go through the position stage of ReadVertex (the
clip coordinates, the viewport, the screen rounding and the GE's fog)
four at a time, a vertex per lane, before the rest. GERowSumLanes is
GERowSum4 with a lane per vertex instead of per matrix row, sharing the
positions' parts between the five rows, and GERecip4 is GERecip for four
w's. A group with a lane those don't handle (non-finite coordinates, a
w that's zero, denormal or too large) goes through ReadVertex as before.
Wipeout takes 3.8 s for five emulated seconds (from 4.2 s), God of War
4.7 s (from 5.0 s).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
GERowSum4 split every matrix entry into exponent, fraction and sign for
each vertex. GERowSumRows does that once when the transform state is
computed, for the clip matrix, the world matrix's normal rows, the fog
row and the texture matrix.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
GEUVSpan stored the four 64-bit plane values, and the reciprocal table's
entries, to arrays one by one and loaded them as vectors, a store
forwarding stall each time, which took most of its time on x86.
CrossSIMD gets LoadS64Low (the low halves of four 64-bit values) and
LoadPairs (four pairs of 32-bit values split into two vectors) to do it
without memory. Wipeout runs about 2% faster.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A sprite's u gradient comes from the setup reciprocal of its whole area, so it's
exact only for a power of two area, whatever the height: the Δy only cancels on
paper. Checked against exp87's sizes and exp256, where 64x128 CLUT8 sprites with
u = 1 + 2x read the next pixel's low byte on a PSP and 64x272 ones their own
high byte.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The GE doesn't divide by a q <= 0: it samples the last texel in u and v,
whatever s and t are, in clamp and repeat modes alike (gpu/texmtx/negq). The
hardware backends divided, which mirrors a projected texture behind its
projector: Fate/Extra's character shadows ran up the school's walls (#18667).
The software renderer sampled (0, 0) instead.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The fallback (which RISC-V uses) did its integer arithmetic in plain int,
where signed overflow and left shifts of negative values are undefined,
and converted out of range floats with a cast, which is undefined too:
RISC-V gives INT_MAX for NaN where SSE gives INT_MIN and NEON 0. With GCC
on RISC-V that broke GEUVSpanCore in the GEMath unit test. Now the
arithmetic wraps through uint32_t, like the SIMD instructions, and an out
of range or NaN conversion gives INT_MIN, like SSE.
Also skip GERowSum's zero terms before computing their shift, as their
exponent can be INT_MIN. Found with -fsanitize=undefined on the fallback.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
When a triangle's q, s and t plane sums fit 32 bits wherever its spans
are walked, and the planes' exponents keep the GE's UV arithmetic in the
normal float range (GEUVSpanSafe), a span's plane values are one vector
add each and GEUVSpanCore runs without its range checks. Tested bit for
bit against the scalar code under those conditions.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
GEUVSpan does the GE's per-pixel UV arithmetic for four pixels at once:
the plane values to float24s by exponent arithmetic, GERecip with its
table entries gathered, and the truncated 24-bit products from the 16-bit
significands' product. Lanes it doesn't cover (values past 2^24,
denormals, results out of range) send the span to the scalar code.
Tested bit for bit against it.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The rasterizer takes a GERecip per pixel, so it's inline now, with its
table exported. The q for the mip level is the same for the two pixels
of a quad row unless they're in different 4-pixel spans, so it's only
looked up again then.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Every GEAdd4 and GEMulFloat24x4 checked its lanes for the cases it can't do and branched to a
fallback, a vector-to-scalar move and a branch per operation, about 14% of skinning. The unchecked
variants only collect those lanes; a vertex where any appears is skinned again with the checked ones.
The products use GEMulFloat24x4 (the same as a one-term GERowSum4 here), and the position's and
normal's add chains are interleaved. Skinning is about 20% faster.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
GEDot3 forms one term per lane and sums across the lanes, about 1.5-2x
as fast as the scalar GEDot. Adds Vec4S32::GetLane.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Skinning did each component's weighted bone entries and its chain of
GEAdds separately in scalar code. The components now go in lanes, with
GEMulFloat24x4 for the weighted entries, one-term row sums for the
products and GEAdd4 for the sums, in the same order. Checked against the
scalar version on every vertex of 143 skinned frame dumps.
GEMulFloat24x4 matches ProductToFloat24, so inf and NaN bone entries
behave as before. GEAdd4 handles two zero operands without its fallback.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
They're truncated to float24 with one AND, and their exponents, fractions
and signs worked out for all terms at once, each term then broadcasting its
lane (new CrossSIMD Vec4S32::SplatLane). The term count becomes a template
parameter so the lanes are constants. About 75-83 M transforms/s on an M1,
from 68; results unchanged.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The four clip space rows of a vertex, the texture matrix's three, the world
normal's three, the vector to a light, and each row of the combined matrix
now take one GERowSum4 call. Same results (pspautotests, frametests and
the probe regression set unchanged).
GERowSum4 scales its result by adding to the exponent instead of
multiplying, and CrossSIMD gets a Vec4S32::Max for the largest lsb. About
68 M transforms/s against 26 for four scalar GERowSums on an M1 (a pure
NEON version with per-lane shifts did about 79); replaying the frametest
dumps gets only about 1% faster, up to 11% on Taiko no Tatsujin.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Four rows sharing their left operands, like a vertex transform's: with
float24 operands each exact product is a 32-bit integer, (1.a' * 1.b') >> 15
= 32768 + a' + b' + (a' b' >> 15), and the terms are aligned by scaling with
a power of two and truncating. Lanes whose result exponent is out of a
single float scale's reach take the scalar GERowSum. Tested bit for bit
against GERowSum on random operands (arm64 and x86-64). Not used yet.
Also adds an arithmetic Vec4S32::Shr and the float/int bit casts to
CrossSIMD.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The minor-axis offset an antialiased line's alpha comes from is computed
with the triangle setup's reciprocal of the line's major length, walking
from v0, and floored; diagonal lines are y-major. exp112 went from 101
differing pixels to 23; the misses left sit at reciprocal segment starts.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The exponent times the log (the float's bits) keeps units of 16 in the bits,
truncated toward zero, not floored at 1. Fixes the last specular pixels at
small exponents (gpu/probe exp102 e = 0.3, exp3, exp4 e = 0.05); all lighting
probes are now bit exact.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Found by the probe regression run on the first lighting probes:
- The normal stays as the world matrix leaves it, and N.L and N.H are
scaled by its GE reciprocal length, like the spot direction. Exact with
scaled, sheared and rotated world matrices (gpu/probe exp3).
- A negative factor (a spot pointing away with cutoff -1) scales the light
by 0 instead of subtracting (exp3, exp4).
- Shade mapping UVs use the same GE arithmetic as lighting (exp2-4).
- The half vector uses the GE's add.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
GEProduct multiplies the significands as integers instead of going through
frexpf/ldexp/trunc in doubles, and GERowSum aligns and adds them as integers.
About 4.5x faster per matrix row on arm64, same results.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The float24 truncations, reciprocal and rsqrt tables, GE adder and row sums,
the setup reciprocal for planes, the mip log2, texture coordinate precision,
the lighting pow and color scaling, and the antialiased line alpha, from
TransformUnit, Rasterizer and Lighting, so they can be unit tested. No
behavior change.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
