tilelang.language.math_intrinsics¶

Common math intrinsics exposed on the TileLang language surface.

Functions¶

`ieee_add`(x, y[, rounding_mode])	IEEE-compliant addition with specified rounding mode
`ieee_sub`(x, y[, rounding_mode])	IEEE-compliant subtraction with specified rounding mode
`ieee_mul`(x, y[, rounding_mode])	IEEE-compliant multiplication with specified rounding mode
`ieee_fmaf`(x, y, z[, rounding_mode])	IEEE-compliant fused multiply-add with specified rounding mode
`ieee_frcp`(x[, rounding_mode])	IEEE-compliant reciprocal with specified rounding mode
`ieee_fsqrt`(x[, rounding_mode])	IEEE-compliant square root with specified rounding mode
`ieee_frsqrt`(x)	IEEE-compliant reciprocal square root (round to nearest only)
`ieee_fdiv`(x, y[, rounding_mode])	IEEE-compliant division with specified rounding mode
`fadd2`(x, y)	Packed FP32x2 add.
`fmul2`(x, y)	Packed FP32x2 multiply.
`fma2`(x, y, z)	Packed FP32x2 fused multiply-add (x * y + z).

Module Contents¶

tilelang.language.math_intrinsics.ieee_add(x, y, rounding_mode='rn')¶

IEEE-compliant addition with specified rounding mode

参数:

x (PrimExpr) -- First operand.
y (PrimExpr) -- Second operand.
rounding_mode (str, optional) -- Rounding mode: 'rn' (round to nearest), 'rz' (round toward zero), 'ru' (round toward positive infinity), 'rd' (round toward negative infinity). Default is 'rn'.

返回:

result -- The result.

返回类型:

PrimExpr

tilelang.language.math_intrinsics.ieee_sub(x, y, rounding_mode='rn')¶

IEEE-compliant subtraction with specified rounding mode

参数:

x (PrimExpr) -- First operand.
y (PrimExpr) -- Second operand.
rounding_mode (str, optional) -- Rounding mode: 'rn', 'rz', 'ru', 'rd'. Default is 'rn'.

返回:

result -- The result.

返回类型:

PrimExpr

tilelang.language.math_intrinsics.ieee_mul(x, y, rounding_mode='rn')¶

IEEE-compliant multiplication with specified rounding mode

参数:

x (PrimExpr) -- First operand.
y (PrimExpr) -- Second operand.
rounding_mode (str, optional) -- Rounding mode: 'rn', 'rz', 'ru', 'rd'. Default is 'rn'.

返回:

result -- The result.

返回类型:

PrimExpr

tilelang.language.math_intrinsics.ieee_fmaf(x, y, z, rounding_mode='rn')¶

IEEE-compliant fused multiply-add with specified rounding mode

参数:

x (PrimExpr) -- First operand.
y (PrimExpr) -- Second operand.
z (PrimExpr) -- Third operand (addend).
rounding_mode (str, optional) -- Rounding mode: 'rn', 'rz', 'ru', 'rd'. Default is 'rn'.

返回:

result -- The result of x * y + z.

返回类型:

PrimExpr

tilelang.language.math_intrinsics.ieee_frcp(x, rounding_mode='rn')¶

IEEE-compliant reciprocal with specified rounding mode

参数:

x (PrimExpr) -- Input operand.
rounding_mode (str, optional) -- Rounding mode: 'rn', 'rz', 'ru', 'rd'. Default is 'rn'.

返回:

result -- The result of 1/x.

返回类型:

PrimExpr

tilelang.language.math_intrinsics.ieee_fsqrt(x, rounding_mode='rn')¶

IEEE-compliant square root with specified rounding mode

参数:

x (PrimExpr) -- Input operand.
rounding_mode (str, optional) -- Rounding mode: 'rn', 'rz', 'ru', 'rd'. Default is 'rn'.

返回:

result -- The result of sqrt(x).

返回类型:

PrimExpr

tilelang.language.math_intrinsics.ieee_frsqrt(x)¶

IEEE-compliant reciprocal square root (round to nearest only)

参数:: x (PrimExpr) -- Input operand.
返回:: result -- The result of 1/sqrt(x).
返回类型:: PrimExpr

tilelang.language.math_intrinsics.ieee_fdiv(x, y, rounding_mode='rn')¶

IEEE-compliant division with specified rounding mode

参数:

x (PrimExpr) -- Dividend.
y (PrimExpr) -- Divisor.
rounding_mode (str, optional) -- Rounding mode: 'rn', 'rz', 'ru', 'rd'. Default is 'rn'.

返回:

result -- The result of x/y.

返回类型:

PrimExpr

tilelang.language.math_intrinsics.fadd2(x, y)¶

Packed FP32x2 add.

Lowers to PTX add.rn.f32x2 on supported NVIDIA architectures/toolchains, and falls back to per-lane scalar operations otherwise.

参数:

x (PrimExpr) -- First operand. Must be dtype float32x2.
y (PrimExpr) -- Second operand. Must be dtype float32x2.

返回:

result -- A float32x2 result.

返回类型:

PrimExpr

tilelang.language.math_intrinsics.fmul2(x, y)¶

Packed FP32x2 multiply.

Lowers to PTX mul.rn.f32x2 on supported NVIDIA architectures/toolchains, and falls back to per-lane scalar operations otherwise.

参数:

x (tvm.tir.PrimExpr)
y (tvm.tir.PrimExpr)

返回类型:

tvm.tir.PrimExpr

tilelang.language.math_intrinsics.fma2(x, y, z)¶

Packed FP32x2 fused multiply-add (x * y + z).

Lowers to PTX fma.rn.f32x2 on supported NVIDIA architectures/toolchains, and falls back to per-lane scalar operations otherwise.

参数:

x (tvm.tir.PrimExpr)
y (tvm.tir.PrimExpr)
z (tvm.tir.PrimExpr)

返回类型:

tvm.tir.PrimExpr