BMMA on CC 7.2 for hamming distance

Robotics & Edge Computing Jetson & Embedded Systems Jetson AGX Xavier
1

Hello,

As far as I know the tensor cores in the GV100 chip only support floating point numerics (HMMA). The following article states, that the GV10B of the Jetson AGX Xavier also supports (U)INT8 (IMMA): https://developer.nvidia.com/blog/nvidia-jetson-agx-xavier-32-teraops-ai-robotics/

So the GV10B tensor cores are a bit like Turing tensor cores. That lead me to the following question:

The CUDA Toolkit 10.2 as in the current JetPack release supports BMMA experimentally, can the Jetson AGX Xavier perform these operations natively? Sadly, the CUDA docs are a bit lacking for compute capability 7.2 and there is no clear statement for this. Only 7.0 and 7.5 are mentioned explicitly as far as I can tell. If I missed something, I am happy to stand corrected :)

Background is that I want to accelerate the hamming distance between two matrices like AxN and NxB, where the distance of the nth col/row is popcount(a_n XOR b_n). That is exactly what the BMMA operation does as far as I understand.

2

To answer my own question: The PTX ISA WMMA specification states that BMMA is not supported on the AGX/NX:

Target ISA Notes

Floating point wmma requires sm_70 or higher.
Integer wmma requires sm_72 or higher.
Sub-byte and single-bit wmma requires sm_75 or higher.
Double precision and alternate floating point precision wmma requires sm_80 or higher.

4

Hi,

As you said, BMMA is not supported on Xavier.
Xavier only support HMMA and IMMA.

Thanks.

5

Thank you for the confirmation. You wrote “currently”, does that mean that the functionality could be extended? I was assuming that this is a HW limitation.

6

Hi,

No. Sorry about the confusion.
BMMA requires a Turning GPU, but Xavier is Volta generation.

Thanks.