warp.tile_matmul#

warp.tile_matmul(
a: Tile[Float, tuple[int, int]],
b: Tile[Float, tuple[int, int]],
out: Tile[Float, tuple[int, int]],
alpha: Float,
beta: Float,
) → None#

  • Kernel

  • Differentiable

Computes the matrix product and accumulates out = alpha * a*b + beta * out.

Supported datatypes are:

  • fp16, fp32, fp64 (real)

  • vec2h, vec2f, vec2d (complex)

All input and output tiles must have the same datatype. Tile data will automatically be migrated to shared memory if necessary and will use TensorCore operations when available.

Note that computing the adjoints of alpha and beta are not yet supported.

param a:

A tile with shape=(M, K)

param b:

A tile with shape=(K, N)

param out:

A tile with shape=(M, N)

param alpha:

Scaling factor (default 1.0)

param beta:

Accumulator factor (default 1.0)

warp.tile_matmul(
a: Tile[Float, tuple[int, int]],
b: Tile[Float, tuple[int, int]],
alpha: Float,
) → Tile[Float, tuple[int, int]]

  • Kernel

  • Differentiable

Computes the matrix product out = alpha * a*b.

Supported datatypes are:

  • fp16, fp32, fp64 (real)

  • vec2h, vec2f, vec2d (complex)

Both input tiles must have the same datatype. Tile data will automatically be migrated to shared memory if necessary and will use TensorCore operations when available.

Note that computing the adjoints of alpha is not yet supported.

param a:

A tile with shape=(M, K)

param b:

A tile with shape=(K, N)

param alpha:

Scaling factor (default 1.0)

returns:

A tile with shape=(M, N)