Asura/tools/gpu-parity
Repository files (latest commit first)
Filename Latest commit message Latest commit date
2026-11-11 12:08:24 +00:00
..
.gitignore gpu-parity: FXC/D3D11 harness for cultmath snoise and snoise_grad 2026-11-11 11:32:51 +00:00
GpuParity.csproj gpu-parity: FXC/D3D11 harness for cultmath snoise and snoise_grad 2026-11-11 11:32:51 +00:00
parity.hlsl gpu-parity: branch codes as small floats, error percentiles 2026-11-11 11:34:17 +00:00
Program.cs gpu-parity: pin check requires a clean packages/cultmath and non-empty SHAs; README matches the run 2026-11-11 12:08:24 +00:00
README.md gpu-parity: pin check requires a clean packages/cultmath and non-empty SHAs; README matches the run 2026-11-11 12:08:24 +00:00

GPU parity: CultMath simplex noise on D3D11

Checks that cultmath_snoise(float2), cultmath_snoise(float3) and cultmath_snoise_grad (float4: gradient in .xyz, value in .w) from CultLib's packages/cultmath/shaders/CultMath.hlsl, compiled by FXC and run as a D3D11 compute shader, match C# math.snoise / math.snoise_grad (packages/cultmath) within a stated tolerance. The shader text under test is CultLib's own file: parity.hlsl #includes it through fxc /I. Nothing is copied.

Windows only (D3D11 hardware device, fxc.exe). Needs a CultLib checkout at tag cultmath-unity-v0.5.0; the run fails its pin check if CultLib HEAD is not that tag's commit.

git clone https://eureka.gamecult.org/forge/GameCult/CultLib.git
git -C CultLib checkout cultmath-unity-v0.5.0
$env:CULTLIB_ROOT = "$PWD\CultLib"          # default: a sibling directory named CultLib
# optional: $env:FXC = "...\fxc.exe"        # default: Windows Kits 10.0.26100.0 x64
dotnet run -c Release --project tools/gpu-parity/GpuParity.csproj

It compiles the kernel twice (default flags, where FMA contraction is allowed, and /Gis IEEE strict), dispatches once per flag (about 164k threads), and prints max abs / max relative error per function, split into tie-constructed points and a seeded random sweep. Exit code 0 means within Tolerance (in Program.cs). Branch mismatches are reported, not failed.

Points: dyadic coordinates on the simplex tie planes (x0.x == x0.y, y == z, x == z, offsets k in -2..2) with 0, 1 and 2 ulp neighbours, plus 3 x 40000 seeded uniform points at scales 4, 64, 1024. parity.hlsl also reads back the ordering comparisons (a re-derivation of the step/> tests, so it witnesses the comparison, not the compiled noise body); a mismatch against C# is reported. Relative error is |gpu - ref| / max(|ref|, 0.01).

The GPU is shared: the run is one short dispatch per flag.

Measured results

Run 2026-11-11 on Starfire, NVIDIA GeForce GTX 1070, D3D11 feature level 11_0, fxc 10.0.26100.0 (cs_5_0), CultLib 7a8933b0 (cultmath-unity-v0.5.0). 163928 points: 43928 tie-constructed, 120000 seeded random (3 x 40000 at scales 4, 64, 1024). Default and /Gis produced the same maxima to three digits; no NaN or Inf.

Maximum absolute error, both flag sets, by max |coordinate| of the input:

tier snoise(float3) snoise(float2) grad .w grad .xyz
<= 8 3.1e-6 3.1e-6 3.1e-6 2.4e-5
8 to 128 5.3e-5 4.2e-5 5.3e-5 3.1e-4
> 128 (to 1024) 1.2e-3 4.9e-4 1.2e-3 5.0e-3

Max relative error (floor 0.01): 4.8e-2 for values and 2.1e-1 for gradients, both at |p| > 128 near zero crossings. Median absolute error is 0 to 2e-7 in every tier.

Stated tolerance (Tolerance in Program.cs): absolute error <= 4.2e-6, 7.2e-5, 1.65e-3 for values and <= 3.3e-5, 4.2e-4, 6.8e-3 for gradient components, per tier above (1.35x the measured maximum). Basis: the error scales with ulp of the input magnitude (x0 = value - i + ... cancels, and GPU and CPU round the first steps differently), not with the noise: 16x more input magnitude gave 17x more error between the first two tiers. Single GPU and driver; a different driver may need the bounds re-measured.

Tie handling: under both flag sets the GPU took a different simplex ordering branch than C# at 190 of 35814 tie points (float3) and 39 (float2) with |p| <= 8, 2 and 4 in the middle tier, 0 above among tie points (the random sweep above 128 had 1 float3 mismatch) (the same counts under /Gis, so it is not FMA contraction; the probe is a separate expression tree, so it is evidence of a GPU/CPU rounding difference in x0, not of the noise body's own path). Every one of those points is inside the value and gradient tolerance: the noise is continuous across the tie, so a flipped branch costs at most the ulp-scale error above. Branch mismatches are reported, not failed.