asc2.zeros_acc

asc2.zeros_acc(shape: Iterable[int], dtype: DataType, *, bias: LocalTensor | None = None) → LocalTensor

Create a zero-initialized accumulator tensor in L0C memory for matrix multiplication.

This tensor is specifically designed for use with matmul_acc() operations and is always located in TensorLocation.L0C.

The supported data type is: float32.

Parameters:

shape – The shape of the accumulator tensor
dtype – The data type of the accumulator
bias – Optional initialization tensor (1D tensor in BT). If provided, the accumulator will be initialized with this value instead of zeros. This is typically used for bias initialization in matrix multiplication. Supported dtypes: float16, bfloat16, or float32. Tensors with float16 or bfloat16 are automatically promoted to float32.

Returns:

A new accumulator tensor in L0C memory (zero-initialized or initialized with the provided value)

Return type:

LocalTensor

Raises:

TypeError – If shape contains non-integer values
RuntimeError – If shape contains non-positive values or bias has wrong shape/dtype

Examples

Create a zero-initialized accumulator:

acc = asc2.zeros_acc([64, 256], dtype=asc2.float32)
for k in range(k_tiles):
    a_k = asc2.copy_in(a_gm, [0, k * 32], [64, 32], asc2.TensorLocation.L0A)
    b_k = asc2.copy_in(b_gm, [k * 32, 0], [32, 256], asc2.TensorLocation.L0B)
    asc2.matmul_acc(acc, a_k, b_k)
asc2.copy_out(acc, c_gm, [0, 0])

Create a bias-initialized accumulator:

bias = asc2.copy_in(bias_gm, [0], [256], asc2.TensorLocation.BT)
acc = asc2.zeros_acc([64, 256], dtype=asc2.float32, bias=bias)
for k in range(k_tiles):
    a_k = asc2.copy_in(a_gm, [0, k * 32], [64, 32], asc2.TensorLocation.L0A)
    b_k = asc2.copy_in(b_gm, [k * 32, 0], [32, 256], asc2.TensorLocation.L0B)
    asc2.matmul_acc(acc, a_k, b_k)
asc2.copy_out(acc, c_gm, [0, 0])