For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /max/get-started.md).

Mojo function

gevm_kernel

def gevm_kernel[c_type: DType, a_type: DType, b_type: DType, *, tile_size: Int, elementwise_lambda_fn: Optional[def[dtype: DType, width: SIMDLength, *, alignment: Int = Int(1)](IndexList[Int(2)], SIMD[dtype, width]) capturing thin -> None] = None, accum_type: DType = get_accum_type[c_type](), pdl_level: PDLLevel = PDLLevel()](c: Pointer[Scalar[c_type], MutUnsafeAnyOrigin, _safe=False], a: Pointer[Scalar[a_type], ImmUnsafeAnyOrigin, _safe=False], b: Pointer[Scalar[b_type], ImmUnsafeAnyOrigin, _safe=False], m: Int, n: Int, k: Int)

GPU kernel for row-vector times matrix multiplication (GEVM).

Each CTA block computes tile_size // WARP_SIZE consecutive output elements. Threads in each warp collaboratively accumulate the dot product of the input row vector with columns of the matrix, then reduce through shared memory.

Parameters:

c_type (DType): Output element type.
a_type (DType): Input vector (row) element type.
b_type (DType): Input matrix element type.
tile_size (Int): Total threads per block; must be a multiple of WARP_SIZE.
elementwise_lambda_fn (Optional[def[dtype: DType, width: SIMDLength, *, alignment: Int = Int(1)](IndexList[Int(2)], SIMD[dtype, width]) capturing thin -> None]): Optional epilogue applied to each output element.
accum_type (DType): Accumulation precision type.
pdl_level (PDLLevel): Programmatic dependent launch level.

Args:

c (Pointer[Scalar[c_type], MutUnsafeAnyOrigin, _safe=False]): Output pointer of length n.
a (Pointer[Scalar[a_type], ImmUnsafeAnyOrigin, _safe=False]): Input row vector pointer of length k.
b (Pointer[Scalar[b_type], ImmUnsafeAnyOrigin, _safe=False]): Input matrix pointer of shape (k, n).
m (Int): Unused; retained for interface consistency.
n (Int): Number of output columns.
k (Int): Shared reduction dimension.