tilelang.ascend.language.random¶

Ascend dialect RNG ops.

Same tl.rng_* emission as the common implementation, but rng_init’s default sequence id derives from the thread context — and on Ascend the thread domain lives in the enclosing T.SimtVF scope, so the derivation must go through this dialect’s thread accessors rather than the launch-frame ones the CUDA module consults.

Functions¶

rng_init(seed[, seq, off, generator])

Initialize the random number generator state.

Module Contents¶

tilelang.ascend.language.random.rng_init(seed, seq=None, off=0, generator='curandStatePhilox4_32_10_t')¶

Initialize the random number generator state.

See tilelang.cuda.language.random.rng_init(); the only difference is that the default seq (one stream per lane) is derived from the Ascend thread scope: the SimtVF lane id plus the AI-core index times the lane count.

Return type:

tvm.tirx.PrimExpr