Fix RNG type in mtgp32 uniform cuRAND example

[jamespud]

Fixes #146

Summary

Three cuRAND Host API examples create a generator type that does not match the generator they are documented to demonstrate. Each one runs without error because the substituted generator is valid for the requested distribution, but the mismatch is invisible at runtime and silently misleads anyone who copies the example.

File	Documented generator	Was using
cuRAND/Host/mtgp32/curand_mtgp32_uniform_example.cpp	MTGP32	CURAND_RNG_PSEUDO_MRG32K3A
cuRAND/Host/mtgp32/curand_mtgp32_poisson_example.cpp	MTGP32	CURAND_RNG_PSEUDO_MT19937
cuRAND/Host/scrambled_sobol32/curand_scrambled_sobol32_poisson_example.cpp	scrambled Sobol32	CURAND_RNG_QUASI_SCRAMBLED_SOBOL64

Why this is a bug

The documented intent comes from three places for every file: the file name, the folder README (e.g. "generate MTGP32 pseudorandom generated numbers"), and the cuRAND index (e.g. "uniform mtgp32 pseudorandom generation using Host API"). Because the wrongly used generators are still valid and produce the same distribution shape, the samples build and run cleanly — this is the same class of issue as #146, where the maintainer asked "MRG32K3A is a valid generator type, what is the issue?" and the reporter had to clarify that the sample is named mtgp32.

The mtgp32_poisson change was requested in the review of this PR.

Change

For each file, all three generator-construction sites are updated consistently: the device path (curandCreateGenerator), the host path (curandCreateGeneratorHost), and the rng variable in main().

• curand_mtgp32_uniform_example.cpp: MRG32K3A → CURAND_RNG_PSEUDO_MTGP32
• curand_mtgp32_poisson_example.cpp: MT19937 → CURAND_RNG_PSEUDO_MTGP32
• curand_scrambled_sobol32_poisson_example.cpp: SOBOL64 → CURAND_RNG_QUASI_SCRAMBLED_SOBOL32

This matches the sibling examples (curand_mtgp32_normal_example.cpp, curand_mtgp32_lognormal_example.cpp, and the other scrambled_sobol32 samples), which already use the correct generator type in the same three places.

Testing

• Built with CUDA 13.2 on Linux x86_64.
• Ran on NVIDIA GeForce RTX 4060 Laptop GPU (sm_89).
• All four mtgp32 examples and the scrambled_sobol32_poisson example build and run with matching host and device output.
• Before the change, the SOBOL64 + Poisson combination was run as a baseline to confirm the quasi-random + Poisson path is supported.

[greptile-apps]

Greptile Summary

The PR aligns three cuRAND samples’ generator types with the algorithms advertised by their directories and filenames.

• Changes the MTGP32 uniform sample from MRG32K3A to MTGP32.
• Changes the MTGP32 Poisson sample from MT19937 to MTGP32.
• Changes the scrambled Sobol32 Poisson sample from scrambled Sobol64 to scrambled Sobol32.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains.

Important Files Changed

Filename	Overview
cuRAND/Host/mtgp32/curand_mtgp32_poisson_example.cpp	Updates device, host, and main generator-type constants to MTGP32 without introducing a supported-path failure.
cuRAND/Host/mtgp32/curand_mtgp32_uniform_example.cpp	Replaces MRG32K3A with MTGP32 consistently across both execution paths and the main configuration.
cuRAND/Host/scrambled_sobol32/curand_scrambled_sobol32_poisson_example.cpp	Corrects all three generator-type constants to scrambled Sobol32 while preserving valid dimensions, offset, and output handling.

_{Reviews (2): Last reviewed commit: "Fix RNG types in Poisson cuRAND examples" | Re-trigger Greptile}

[dorispnvidia]

Can you also add the changes to curand_mtgp32_poisson_example.cpp please? It looks like the wrong generator is also used in this file. Thanks.

[jamespud]

Thanks @dorispnvidia — added the MTGP32 fix to curand_mtgp32_poisson_example.cpp, and while scanning the family I also fixed the same class of bug in curand_scrambled_sobol32_poisson_example.cpp (SOBOL64 → SOBOL32).

[dorispnvidia]

LGTM