Grids & blocks

Grids and Blocks

Grids and Blocks are basic CUDA concepts, that permit to organise threads

They are created as kernel launch:

• grids are 1, 2, (or 3) dimensional array of blocks

• blocks are 1, 2, (or 3) dimensionale array of threads

A thread block is a batch of threads that can cooperate with each other by:

• synchronising their execution using barriers

• efficiently sharing data through a low latency shared memory

Two threads from two different blocks cannot cooperate

Thread block organisation

Threads are grouped into blocks

Blocks can have 1, 2, or 3 dimensions

Each block dimension can have a maximum number of threads (e.g. 512, 1024)

Thread block organisation (1D/2D)

Thread block

Block shape doesn't matter, but may be usefull for mapping threads to data

Thread block vs array access

Or mapping nested loops in the original code: each 2D-organised thread taking care of an (i,j) pair:

Thread blocks - mapping loops

Block organisation

Threads are grouped into blocks

Blocks are organised in a grid

A grid can be 1D, 2D (or 3D)

Grid vs blocks