ALADIN HLLD v05.2: GPU = CPU Bit-for-Bit HLLD Flux Kernel for Native-Cylindrical Z-Pinch MHD (Deterministic log10, IEEE Subnormals, FP64 and FP32)

ALADIN HLLD v05.2 is the HLLD flux kernel (slots 04–06: HLLD-R, HLLD-φ, HLLD-Z) of the ALADIN native-cylindrical GPU MHD Driver v4.0 for Z-pinch and J × B research. One CUDA/C++ source, compiled for the GPU (NVRTC) and the CPU (g++), FP64 and FP32. Why v05.2. Driver v4.0's kernel-wise GPU-vs-CPU probe (KW-00) found the HLLD-φ fluxes differing by a few ulp on blended faces of a running Z-pinch, while every other kernel was bit-identical. The cause was the math-library log10() in the blend weight, whose last-bit rounding differs between CUDA and glibc. In a Z-pinch only the φ sweep has B_n ≠ 0 and reaches it. Random-state certification (HLLD-04 G5) reached that branch about once in 10^5 problems and could not see it. What changed, and nothing else:• log10 replaced by log10p6, built only from + − × ÷ and √ (two square roots and a fixed 17-term atanh series); accuracy 6.7e-16 (FP64), monotone.• The kernel input W (cell-centred B) is built inside the IEEE module on the GPU. CuPy's own elementwise arithmetic flushes FP32 subnormals (1e-40f × 1 → 0), which changed the kernel's inputs on the GPU (HLLD-05 run 1, G7: 9 537 subnormal values).• Direct NVRTC compilation with -fmad=false -ftz=false -prec-div=true -prec-sqrt=true; PTX checked free of .ftz. FP32 on the GPU is IEEE now. Campaign HLLD-05, run 2 (NVIDIA L4 and x86-64 g++ 13): 43 of 44 checks VERIFIED on both legs.• All 30 HLLD-04 tests pass with their original thresholds.• Off the log10 ramp, v05.2 is bit-identical to the certified v04 (R00).• GPU = CPU bit for bit, FP64 and FP32, on log10-ramp Riemann problems and on 13 frozen states of the Driver's own run (G7: 0 differences), and on the kernel input with 19 498 subnormal values present (G9: 0).• compute-sanitizer: 0 errors, 0 hazards.• In Driver v4.0 the S6 GPU-vs-CPU gap went from 5.4e-20 to 0, and the CPU and GPU Drivers run 30 steps in lockstep without one differing bit. One predeclared check FAILED and stays FAILED. L02's error vs an 80-bit oracle on log10-ramp problems was 6.8e-10 against 1e-10. The threshold was copied from moderate states, while the ramp problems sit at |B_n| up to ~10^6 (β ~ 1e-15). Certified v05.1 has the identical worst error, and v05.2 vs v05.1 differs by ≤ 3.9e-16. An addendum declared after the result shows 2.3e-11 at |B_n| ≤ 10^3. Run 1, with its four BLOCKERs and both diagnostics, is included unchanged. The README benchmark (B_φ = r/2, J_z = 1, f_r = −r/2, ∇·B = 0, P = P0 − r²/4, μ0 = 1) gives identical results through the certified kernels UCT_CURL v1.3, UCT-EMF v02.1, EGLM v02 and HLLD v05.2. Author: Mihai Alexandru Bucurenciu (ALADIN), ORCID 0009-0006-4457-557X.AI assistant Team: SuperGrok, Claude, ChatGPT

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-20
DOI
https://doi.org/10.5281/zenodo.24401082
Primary Topic
Parallel Computing and Optimization Techniques
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

ALADIN HLLD v05.2: GPU = CPU Bit-for-Bit HLLD Flux Kernel for Native-Cylindrical Z-Pinch MHD (Deterministic log10, IEEE Subnormals, FP64 and FP32)

Mihai Alexandru Bucurenciu
7 links
Zenodo (CERN European Organization for Nuclear Research)
Parallel Computing and Optimization Techniques
article

ALADIN HLLD v05.2: GPU = CPU Bit-for-Bit HLLD Flux Kernel for Native-Cylindrical Z-Pinch MHD (Deterministic log10, IEEE Subnormals, FP64 and FP32)

Mihai Alexandru Bucurenciu
article en

Abstract

ALADIN HLLD v05.2 is the HLLD flux kernel (slots 04–06: HLLD-R, HLLD-φ, HLLD-Z) of the ALADIN native-cylindrical GPU MHD Driver v4.0 for Z-pinch and J × B research. One CUDA/C++ source, compiled for the GPU (NVRTC) and the CPU (g++), FP64 and FP32. Why v05.2. Driver v4.0's kernel-wise GPU-vs-CPU probe (KW-00) found the HLLD-φ fluxes differing by a few ulp on blended faces of a running Z-pinch, while every other kernel was bit-identical. The cause was the math-library log10() in the blend weight, whose last-bit rounding differs between CUDA and glibc. In a Z-pinch only the φ sweep has B_n ≠ 0 and reaches it. Random-state certification (HLLD-04 G5) reached that branch about once in 10^5 problems and could not see it. What changed, and nothing else:• log10 replaced by log10p6, built only from + − × ÷ and √ (two square roots and a fixed 17-term atanh series); accuracy 6.7e-16 (FP64), monotone.• The kernel input W (cell-centred B) is built inside the IEEE module on the GPU. CuPy's own elementwise arithmetic flushes FP32 subnormals (1e-40f × 1 → 0), which changed the kernel's inputs on the GPU (HLLD-05 run 1, G7: 9 537 subnormal values).• Direct NVRTC compilation with -fmad=false -ftz=false -prec-div=true -prec-sqrt=true; PTX checked free of .ftz. FP32 on the GPU is IEEE now. Campaign HLLD-05, run 2 (NVIDIA L4 and x86-64 g++ 13): 43 of 44 checks VERIFIED on both legs.• All 30 HLLD-04 tests pass with their original thresholds.• Off the log10 ramp, v05.2 is bit-identical to the certified v04 (R00).• GPU = CPU bit for bit, FP64 and FP32, on log10-ramp Riemann problems and on 13 frozen states of the Driver's own run (G7: 0 differences), and on the kernel input with 19 498 subnormal values present (G9: 0).• compute-sanitizer: 0 errors, 0 hazards.• In Driver v4.0 the S6 GPU-vs-CPU gap went from 5.4e-20 to 0, and the CPU and GPU Drivers run 30 steps in lockstep without one differing bit. One predeclared check FAILED and stays FAILED. L02's error vs an 80-bit oracle on log10-ramp problems was 6.8e-10 against 1e-10. The threshold was copied from moderate states, while the ramp problems sit at |B_n| up to ~10^6 (β ~ 1e-15). Certified v05.1 has the identical worst error, and v05.2 vs v05.1 differs by ≤ 3.9e-16. An addendum declared after the result shows 2.3e-11 at |B_n| ≤ 10^3. Run 1, with its four BLOCKERs and both diagnostics, is included unchanged. The README benchmark (B_φ = r/2, J_z = 1, f_r = −r/2, ∇·B = 0, P = P0 − r²/4, μ0 = 1) gives identical results through the certified kernels UCT_CURL v1.3, UCT-EMF v02.1, EGLM v02 and HLLD v05.2. Author: Mihai Alexandru Bucurenciu (ALADIN), ORCID 0009-0006-4457-557X.AI assistant Team: SuperGrok, Claude, ChatGPT

Zenodo (CERN European Organization for Nuclear Research)
Parallel Computing and Optimization Techniques
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon