LeiWang1999 opened a new issue, #19546:
URL: https://github.com/apache/tvm/issues/19546
Currently, some standard TIR math intrinsics on CUDA lower to CUDA fast-math
device functions by default.
For example, `tir.exp` / `tirx.exp` on `float32` lowers to `__expf(x)`
instead of the precise CUDA math function expf(x). **This happens even when
--use_fast_math is not passed to NVCC.**
## Why this is a problem
`__expf`,` __logf`, `__sinf`, etc. are CUDA fast-math intrinsics. They trade
accuracy for performance and can introduce visible precision loss in
numerically sensitive kernels.
Users generally expect standard math intrinsics such as `T.exp`, `T.log`,
`T.sin`, and `T.cos` to preserve normal CUDA math semantics unless fast math is
explicitly requested.
Fast-math behavior should ideally be opt-in, for example through a target
option, compiler flag, or an explicit fast-math intrinsic.
Standard TIR math intrinsics should lower to precise CUDA math functions by
default:
| TIR op | Expected CUDA |
| --- | --- |
| tirx.exp | expf |
| tirx.exp10 | exp10f |
| tirx.log | logf |
| tirx.log2 | log2f |
| tirx.log10 | log10f |
| tirx.sin | sinf |
| tirx.cos | cosf |
| tirx.tan | tanf |
Fast-math variants such as __expf, __logf, __sinf, and __cosf should only be
emitted when fast math is explicitly enabled.
Suggested fix: use CUDAMath instead of CUDAFastMath for standard CUDA math
intrinsic lowering:
TVM_REGISTER_OP("tirx.exp")
.set_attr<FLowerIntrinsic>("cuda.FLowerIntrinsic",
DispatchPureExtern<CUDAMath>);
If fast-math lowering is desired, it would be better to gate it behind an
explicit fast-math option rather than making it the default behavior for
standard math intrinsics.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
