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

gemv_kernel

def gemv_kernel[c_type: DType, a_type: DType, b_type: DType, *, transpose_b: Bool = False, 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 matrix-vector multiplication using scalar warp-level reduction.

Each warp computes one output row by accumulating a dot product over the K dimension with one scalar element per thread, then reducing across the warp.

Parameters:

c_type (DType): Output element type.
a_type (DType): A (matrix) element type.
b_type (DType): B (vector) element type.
transpose_b (Bool): When True, writes the result to a transposed output index.
elementwise_lambda_fn (Optional[def[dtype: DType, width: SIMDLength, *, alignment: Int = Int(1)](IndexList[Int(2)], SIMD[dtype, width]) capturing thin -> None]): Optional epilogue function applied to each output element.
accum_type (DType): Accumulation precision type.
pdl_level (PDLLevel): Programmatic dependent launch level for PDL barriers.

Args:

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