mirrors/Iris
1
0
Fork 0
mirror of https://github.com/VolmitSoftware/Iris.git synced 2026-10-12 03:24:14 +00:00
Code Issues Packages Projects Releases Wiki Activity

Drop unwired 2D noise kernels

Browse Source 
- Delete NoiseKernels2D, ScalarNoiseKernels2D and VectorNoiseKernels2D after measuring the vector kernel at 0.07x scalar on 2-lane NEON with no production call site
- Delete the 2D noise parity test and the SimdSupport plumbing that selected between the two kernel sets
- State in the simd-bench README that its 2D noise classes are a candidate Iris does not ship
This commit is contained in:
Brian Neumann-Fopiano
2026-09-08 04:33:32 -04:00
parent 79af11c435
commit 45b9ff8e11
7 changed files with 15 additions and 436 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tools/simd-bench/README.md
+10 -5
View File
@@ -6,9 +6,13 @@ on**. The Volmit dev/test machines are Apple Silicon, where the wide-SIMD path
(4+ double lanes) does not exist. Copy the built artifact to a Windows/x86 box
(AVX2 = 4 lanes, AVX-512 = 8 lanes) and run it to get real numbers for that CPU.
The six kernel classes are copied verbatim (logic byte-for-byte) from
`Iris/core/.../util/simd/`. This tool is intentionally a duplicate so it builds
and runs on its own with no Gradle, no VolmLib, and no Iris on the classpath.
The three array kernel classes are copied verbatim (logic byte-for-byte) from
`Iris/core/.../util/simd/`. The three 2D noise kernel classes are a candidate
implementation that Iris does not ship: it measured 0.07x against scalar on
2-lane NEON and was removed, and this harness keeps it so the same question can be
answered on a wider CPU before anyone rebuilds it. This tool is intentionally a
duplicate so it builds and runs on its own with no Gradle, no VolmLib, and no Iris
on the classpath.
## Requirements
@@ -98,7 +102,8 @@ run.bat --mode both (default)
throughput will be (cache behavior, allocation, and surrounding code differ in
the real engine).
- **Effort 1 array kernels** (`roundToInt` / `sum` / `max`) run unconditionally
in real Iris. **Effort 2 noise** (`simplexFractalFBM`) is currently unwired in
Iris and gated to 4+ double lanes; this tool force-measures it regardless.
in real Iris. **Effort 2 noise** (`simplexFractalFBM`) is not part of
Iris; this tool force-measures the candidate kernel regardless of the 4+ double
lane gate.
- Vector-API auto-vectorization and cost depend heavily on the JDK version and
CPU. Run on the actual target hardware; do not extrapolate across machines.