I want to profile my AGX Xavier CUDA Core integer usage with nvprof.
The compute capability of my GPU is 7.2.
However, is there some flag to profile CUDA Core integer usage (not Tensor Core)?
According to:
I found that:
1.
In previous compute capability, we can use --single_precision_fu_utilization flag to profile both FP32 and integer usage in CUDA Core.
But in compute capability 7.X, we can only profile FP32 usage in CUDA Core.
How to profile CUDA Core integer usage?
tensor_precision_fu_utilization: Tensor Core fp16
tensor_int_fu_utilization Tensor: Core int8
single_precision_fu_utilization: Cuda Core fp32
half_precision_fu_utilization: Cuda Core fp16
Cuda Core int8: missing?