asc2.inline_vf

asc2.inline_vf(code: str, shape: Tuple[int, ...], dtype: DataType, inputs: Iterable[LocalTensor] | None = None) → LocalTensor

Embed Ascend C VF (vector function) code within a kernel.

This is an escape hatch for advanced users who need to express vector-fusion operations (e.g., Ascend C MicroAPI calls) that are not covered by the built-in API. The provided code string is injected verbatim as the body of a __VEC_SCOPE__ block in the generated Ascend C source.

Tensors are referenced by positional placeholders: $0 is always the output tensor, and $1, $2, … refer to the input tensors in the order they appear in inputs. Zero or more input tensors are allowed. Each placeholder will be replaced with a LocalTensor allocated for a corresponding tensor.

All input tensors must reside in UB memory. The output tensor is always allocated in UB.

Parameters:

  • code – The raw Ascend C code string to embed (treated as a __VEC_SCOPE__ body). Use $0 for the output tensor and $1, $2, … for input tensors.

  • shape – The shape of the output tensor.

  • dtype – The data type of the output tensor.

  • inputs – An optional iterable of zero or more input tensors referenced as $1, $2, … in the code.

Returns:

A new UB tensor ($0) containing the result produced by the inline vector function.

Return type:

LocalTensor

Raises:

  • TypeError – If code is not a str, dtype is not a DataType, or any input is not a LocalTensor.

  • RuntimeError – If any input tensor is not located in UB memory or shape is invalid.

Examples

Embed an inline vector multiply-add (x * y + z) using Ascend C MicroAPI:

out = asc2.inline_vf(
    '''
    auto* out_ptr = reinterpret_cast<__ubuf__ float*>($0.GetPhyAddr());
    auto* x_ptr = reinterpret_cast<__ubuf__ float*>($1.GetPhyAddr());
    auto* y_ptr = reinterpret_cast<__ubuf__ float*>($2.GetPhyAddr());
    auto* z_ptr = reinterpret_cast<__ubuf__ float*>($3.GetPhyAddr());
    AscendC::MicroAPI::RegTensor<float> result_reg;
    . . .
    AscendC::MicroAPI::MaskReg mask_reg = AscendC::MicroAPI::UpdateMask<float>(mask);
    AscendC::MicroAPI::DataCopy(x_reg, x_ptr);
    AscendC::MicroAPI::DataCopy(y_reg, y_ptr);
    AscendC::MicroAPI::Mul(xy_reg, x_reg, y_reg, mask_reg);
    . . .
    ''',
    x.shape, x.dtype, [x, y, z])

In the example above, $0 placeholder refers to a LocalTensor corresponding to out tensor; $1, $2, and $3 refers to x, y, and z respectively.