Sparse Vector Types

Overview

Sparse vector data-types are stored in the vector registers within the AIE-ML. They can be used as the second argument to mac operations.

Sparse Data Format

The supported sparse data layout requires a minimum of 50% sparsity.
Specifically, two zero values within each group of four consecutive values.
This 50% is a lower bound on the sparsity, meaning that further compression is possible if more zeroes are present.

Loading sparse data from memory will interpret the first 64 bits as a mask:

• If a mask bit is not set, then 8 bits will be initialized to zero at its corresponding position.
  
• If a mask bit is set, then the 8 bits will be loaded from the incompressible data that follows the mask.
  

Hence, masks describe the layout of 512 bits after decompression.
Each mask must be aligned to a 32-bit boundary. Failure to meet this requirement will result in incorrect parsing.

When loading sparse data, a partial decompression is carried out to reconstruct the data such that:

• A 64-bit mask is paired with 256-bit partially decompressed data.
  
• 4 bits of the mask are used to represent two elements of the partially decompressed data.

Sparse partial decompression

Mask bits				Partially decompressed data
0	0	0	0	0 , 0
0	0	0	1	0 , A
0	0	1	0	0 , B
0	0	1	1	B , A
0	1	0	0	C , 0
0	1	0	1	C , A
0	1	1	0	C , B
0	1	1	1	C , B
1	0	0	0	D , 0
1	0	0	1	D , A
1	0	1	0	D , B
1	1	0	0	D , C

Topics
	Sparse floating-point vector types
	Sparse integer vector types