noc_async_write_shard
-
template<typename DSpec, bool posted = false>
void noc_async_write_shard(const uint32_t shard_id, const TensorAccessor<DSpec> &s, std::uint32_t src_local_l1_addr, uint8_t noc = noc_index)
-
Initiates an asynchronous write of a shard from a local L1 address to a destination noc address. The size of the transaction and the destination address are determined by the TensorAccessor object. This function only works for sharded tensors.
Return value: None
Argument Description
shard_id Row-major index of a shard in the sharded tensor
s TensorAccessor object
src_local_l1_addr Source address in local L1 memory
noc Which NOC to use for the transaction
DSpec (template parameter) DistributionSpec type
posted (template parameter) Whether the write is posted (i.e. no ack required)