How Does CubeCL Compare to Triton for Thread Block Control?

[lucianyao]

Hi everyone,

I'm new to neural networks and the Burn framework.

I've noticed that CubeCL allows GPU kernel programming, which seems similar to Triton.

However, I'm wondering how CubeCL's thread block control (e.g., CUBE_COUNT and CUBE_DIM) compares to Triton's fine-grained control (e.g., grid and BLOCK_SIZE) for performance optimization. Does CubeCL provide enough flexibility for manual tuning, or does it rely mostly on autotuning? Are there plans to add more explicit control features in the future?

I'd really appreciate your insights. Thanks!

-Lucian

[laggui]

CCing @nathanielsimard

Maybe we could add discussions to the cubecl at some point 🤔

[nathanielsimard]

Ho you can modify the grid and block size. The grid is the cube count and the block size is the cube dim.
See the equivalance with cuda and webgpu:

CubeCL	CUDA	WebGPU
CUBE_COUNT	N/A	N/A
CUBE_COUNT_X	gridDim.x	num_workgroups.x
CUBE_COUNT_Y	gridDim.y	num_workgroups.y
CUBE_COUNT_Z	gridDim.z	num_workgroups.z
CUBE_POS	N/A	N/A
CUBE_POS_X	blockIdx.x	workgroup.x
CUBE_POS_Y	blockIdx.y	workgroup.y
CUBE_POS_Z	blockIdx.z	workgroup.z
CUBE_DIM	N/A	N/A
CUBE_DIM_X	blockDim.x	workgroup_size.x
CUBE_DIM_Y	blockDim.y	workgroup_size.y
CUBE_DIM_Z	blockDim.z	workgroup_size.z
UNIT_POS	N/A	local_invocation_index
UNIT_POS_X	threadIdx.x	local_invocation_id.x
UNIT_POS_Y	threadIdx.y	local_invocation_id.y
UNIT_POS_Z	threadIdx.z	local_invocation_id.z
PLANE_DIM	warpSize	subgroup_size
ABSOLUTE_POS	N/A	N/A
ABSOLUTE_POS_X	N/A	global_id.x
ABSOLUTE_POS_Y	N/A	global_id.y
ABSOLUTE_POS_Z	N/A	global_id.z