Vector search is a powerful technique to identify similar elements. This is a simple gpu parallel processing implementation to find the closest vector based on the minimum Euclidean Distance.
The architecure of a GPU includes a large amount of processing cores, the picture below from Nvidia's documentation illustrates this.
Green: Processing cores
Purple: L1 cache
Yellow: Control units
The vast amount of processing cores make GPUs useful for Single Instruction, Multiple Data operations.
For quickly iterating through a large matrix and finding Euclidean Distance we can build a block that contains number of threads equivalent to the number of columns and then use many blocks when we launch our kernel.
The picture above illustrates a simple examples of iterating through a matrix with GPU parallel processing vs a regular loop.
CPU Results:
CPU
ROWS: 33554432
COLUMNS: 50
NEAREST VECTOR INDEX: 5021
TIME TAKEN: 59.019957
GPU Results:
GPU
ROWS: 33554432
COLUMNS: 50
NEAREST VECTOR INDEX: 5021
TIME TAKEN: 1.681548
We see about a 35 times faster result from our GPU program.