tilelang.cuda.language.math¶

CUDA-specific floating-point arithmetic intrinsics.

Functions¶

fmul(x, y)

Multiply with CUDA round-to-nearest semantics.

fma(x, y, z)

Compute x * y + z as one CUDA fused multiply-add operation.

Module Contents¶

tilelang.cuda.language.math.fmul(x, y)¶

Multiply with CUDA round-to-nearest semantics.

Unlike the * operator, this intrinsic preserves an explicit multiply boundary during CUDA lowering. Vectorized loops may lower it to packed multiply instructions when the target supports them.

Parameters:

  • x (tvm.tirx.PrimExpr)

  • y (tvm.tirx.PrimExpr)

Return type:

tvm.tirx.PrimExpr

tilelang.cuda.language.math.fma(x, y, z)¶

Compute x * y + z as one CUDA fused multiply-add operation.

The result uses round-to-nearest semantics. Vectorized loops may lower the operation to packed FMA instructions when the target supports them.

Parameters:

  • x (tvm.tirx.PrimExpr)

  • y (tvm.tirx.PrimExpr)

  • z (tvm.tirx.PrimExpr)

Return type:

tvm.tirx.PrimExpr