Skip to main content

/

Docs

Nightly

v25.1

Log in

Mojo function

tc_reduce_gevm_8x

tc_reduce_gevm_8x[out_type: DType, in_type: DType, simd_width: Int](val1: SIMD[in_type, simd_width], val2: SIMD[in_type, simd_width]) -> SIMD[out_type, simd_width]

Performs an 8x GEVM reduction using tensor cores.

Note: Currently only supports bfloat16 input to float32 output conversion. Uses tensor core matrix multiply-accumulate (MMA) operations for reduction.

Parameters:

  • ​out_type (DType): The output data type for the reduction result (must be float32).
  • ​in_type (DType): The input data type of the vectors to reduce (must be bfloat16).
  • ​simd_width (Int): The width of the SIMD vectors.

Args:

  • ​val1 (SIMD[in_type, simd_width]): First input SIMD vector to reduce.
  • ​val2 (SIMD[in_type, simd_width]): Second input SIMD vector to reduce.

Returns:

SIMD vector containing the reduced result.