The same lerps as the GE-exact Bezier path, four at a time. At each v step,
four control columns per vector share the v blending factors; then each row
is evaluated at four u steps, each lane with its own six blending factors.
A vector's u steps can straddle a segment boundary, so the row's points are
gathered per lane, or splatted when all four are in one segment.
The unit test compares it bit for bit with the scalar code, including
tessellations that split a vector across segments. The GE-exact spline
benchmark runs about twice as fast on an M1.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Four of the GE's lerps at once: both operands go to 16-bit fixed point at
the larger exponent by scaling with a power of two and truncating, then the
integer lerp (split into byte halves so both multiplies fit 16 bits), then
back. The patch's four control columns run in the four lanes at each v
step, then four u steps of a row at once, with the row's first level of
lerps converted once per row. Lanes whose exponents are tiny take the scalar
lerp, and the normals' cross product stays scalar for now.
The scalar code stays as the reference (g_splineGEScalar), and the unit test
compares the two bit for bit on random control points. The GE-exact Bezier
benchmark goes from 3.4 to about 8.8 Mverts/s on an M1; the float SIMD path
does about 100.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The tessellator's output was queued as ordinary vertices and run through
a vertex decoder at flush, although the decoded layout is the same as
SimpleVertex. That decode did two things: the UV scale/offset, and
clearing gstate_c.vertexFullAlpha. The tessellator now does both, and
SubmitCurve writes its output where DecodeVerts would have put it and
queues it as predecoded. DecodeVerts only copies such a draw, when it
decodes elsewhere, like the Vulkan and GLES push buffers.
The software renderer decodes by itself and keeps the old path, as do
through mode curves and anything with draws already queued.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Each row of vertices is evaluated by folding the patch's four control point
rows into four columns with the v weights, then weighting those along u,
four vertices per register with one attribute component each. Cross
products, normalization and the pole test become plain per-lane math, and
the result is transposed into SimpleVertex order on the way out.
The weights are stored one row per term so four steps load as one vector,
and the control points are padded to full vectors in a single array.
Colors now round to nearest on every platform. The SSE path already did,
NEON and the scalar path truncated.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
x86 normalizes with a bare _mm_rsqrt_ps, so normal lengths are only good
to about 3.7e-4 there and the 1e-4 check failed on every x86 CI job.
Compare direction after normalizing instead.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
With animated control points the pole is only nearly degenerate, so the
vanishing derivative is rounding noise rather than exactly zero, and the
absolute threshold missed it. The resulting random normals still showed
as dark patches on Pac-Man Arrangement's ghosts (#12354).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Compares the software tessellator's positions, normals, UVs and colors
against a longhand double-precision reference (Bernstein polynomials,
Cox-de Boor with clamped knots, finite-difference normals), across
Bezier and spline surfaces, edge types, poles and patch facing, so it
can be optimized with something independent to be wrong against. Also
reports vertices per second.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
