Prefer Array class over register arrays.

[csarofeen]

Prefer Array class over register arrays.
For example in code generation use Array<float, 2> not float[2].

[csarofeen]

!test

[github-actions]

PR Reviewer Guide 🔍

(Review updated until commit 1e7014a)

Here are some key observations to aid the review process:

⏱️ Estimated effort to review: 3 🔵🔵🔵⚪⚪

🧪 No relevant tests

⚡ Recommended focus areas for review Logic Change The aligned_array_of_regs_ set has been removed, and the genVariableNameConvertAlignedArray function has been replaced with genVariableName. This change may affect the handling of aligned arrays in the code generation process. void initStringStreamFormat(std::stringstream& ss) { Type Change The philox function now returns an Array<uint32_t, 4> instead of a uint4. This change may affect the usage of the philox function in other parts of the code. __device__ Array<uint32_t, 4> philox(

[csarofeen]

!test --pybench

[csarofeen]

All tests passing.
Benchmarks do not appear to regress. https://nv/ei9/22766159

[naoyam]

Prefer Array class over register arrays. For example in code generation use Array<float, 2> not float[2].

Any specific reason?

[zasdfgbnm]

Prefer Array class over register arrays. For example in code generation use Array<float, 2> not float[2].

Any specific reason?

Likely because currently we are using float[2] in some cases, but Array<float, 2> in other cases, and sometimes there are compatibility issue for these two types? I remembered Liqiang fixed one such issue on vectorization.

[zasdfgbnm]

Possibly related: #1475

[zasdfgbnm]

Seems that aligned_array_of_regs_ can be removed now.

[csarofeen]

Makes sense, will remove.

[csarofeen]

@naoyam I thought it would help to make things consistent. I also wanted to convert RNG to all use the class instead of literal arrays. This would help do something like generate bit array with Philox and reinterpret it as more lower precision integer types.

[naoyam]

LGTM

[csarofeen]

!test

[csarofeen]

!test