This document is relevant for: Inf2, Trn1, Trn1n

nki.language.store#

nki.language.store(dst, value, mask=None, **kwargs)[source]#

Store into a tensor on device memory (HBM) from on-chip memory (SBUF).

See Memory hierarchy for detailed information.

Parameters:

dst – HBM tensor to store the data into.
value – An SBUF tile that contains the values to store.
mask – (optional) a compile-time constant predicate that controls whether/how this instruction is executed (see NKI API Masking for details)

Returns:

import neuronxcc.nki.language as nl

@nki_jit
def example_kernel(in_tensor, out_tensor):
  ...
  # store into out_tensor[P, F] that is on HBM
  # from data_tile[P, F] that is on SBUF
  nl.store(out_tensor, data_tile)

Note

Partition dimension size can’t exceed the hardware limitation of nki.language.tile_size.pmax, see Tile size considerations.

Partition dimension has to be the first dimension in the index tuple of a tile. Therefore, data may need to be split into multiple batches to load/store, for example:

import neuronxcc.nki.language as nl

@nki_jit
def example_load_store_b(in_tensor, out_tensor):
  for i_b in nl.affine_range(4): 
    data_tile = nl.zeros((128, 512), dtype=in_tensor.dtype) 
    ...
    # store into out_tensor[4, 128, 512] one batch at a time
    # from data_tile[128, 512] 
    i_p, i_f = nl.mgrid[0:128, 0:512]
    nl.store(out_tensor[i_b, i_p, i_f], value=data_tile[i_p, i_f]) 

Also supports indirect DMA access with dynamic index values:

import neuronxcc.nki.language as nl
...

  ##################################################################################
  # Indirect DMA write example 1:
  #  - data_tensor has shape [128 x 512].
  #  - idx_tensor on HBM has shape [64] (with values [0, 2, 4, 6, ...]).
  #  - idx_tensor values read from HBM and stored in SBUF idx_tile.
  #  - data_tile of shape [64 x 512] values written into
  #    HBM data_tensor indexed by values in idx_tile.
  ##################################################################################
  i_p = nl.arange(64)[:, None]
  i_f = nl.arange(512)[None, :]
  idx_tile = nl.load(idx_tensor[i_p]) # indices have to be in SB

  nl.store(data_tensor[idx_tile[i_p, 0], i_f], value=data_tile[0:64, 0:512])

import neuronxcc.nki.isa as nisa
import neuronxcc.nki.language as nl
...

  #############################################################################################
  # Indirect DMA write example 2:
  #  - data_tensor has shape [128 x 512].
  #  - idx_tile on SBUF has shape [64 x 1] (with values [[0], [2], [4], ...] generated by iota)
  #  - data_tile of shape [64 x 512] values written into
  #    HBM data_tensor indexed by values in idx_tile.
  #############################################################################################
  idx_expr = 2*nl.arange(64)[:, None]
  idx_tile = nisa.iota(idx_expr, dtype=np.int32)
  
  nl.store(data_tensor[idx_tile, i_f], value=data_tile[0:64, 0:512]) 

This document is relevant for: Inf2, Trn1, Trn1n

nki.language.store

Contents

nki.language.store#