I wonder what the meaning of varing kernel is. about habitat HOT 6 CLOSED

Thanks for your interest in Habitat!

What we mean by "kernel-varying operation" is that the high-level operation is not necessarily implemented using the exact same kernel(s) on all GPUs. One example is convolutions. For example, if you take a PyTorch model that contains convolutions and run it on GPUs that are from different generations, (e.g., a P100 (Pascal) and a V100 (Volta)) you will see different kernels being used.

Note that in our paper, "operation" is not the same as "kernel". By operation we mean the high-level functions that a user might call in a deep learning framework like PyTorch (e.g., conv2d or relu). These operations are implemented using one or more GPU kernels.

from habitat.

My understanding is that the main reason is for performance. Operations like convolutions can take up a big portion of the model's execution time during training. So, in the case of cuDNN, NVIDIA has different implementations of the convolution kernels for different GPU generations. Presumably they do this so that they can optimize/tune the kernels for specific GPU architectures.

from habitat.

Closing this for now - feel free to reopen to follow up!

from habitat.

Thank you very much for your reply.
I have an additional question.

Why are different kernels used when running PyTorch models with convolution on different generations of GPUs (eg P100 (Pascal) and V100 (Volta))?

I guess I'm not figuring out exactly what the kernel is.

from habitat.

Thanks for your kind reply.
It really helped a lot!!!!
There is one additional confusing thing.

Should warp and wave be considered different?

from habitat.

Yes, they are different. You can think of warps as the "physical unit of execution" on the GPU whereas thread blocks are the "logical" unit of execution. Thread blocks are split up into warps, which are scheduled for execution on the GPU. In the Habitat paper, waves are essentially the maximum number of thread blocks (for a specific kernel) that can run concurrently on a specific GPU.

There’s more information about warps (and their relationship to thread blocks) here: https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#simt-architecture

from habitat.