What is f16f16_f16f16_f16 in cudnn gemm kernel

AI & Data Science Deep Learning (Training & Inference) TensorRT
1

I run aconv3d, I see the kernel:
sm80_xmma_fprop_implicit_gemm_indexed_f16f16_f16f32_f32_nhwckrsc_nhwc_tilesize256x32x32_stage4_warpsize4x1x1_g1_tensor16x8x16_kernel_cudnn
but when I use tensorrt,the kernel is:
sm80_xmma_fprop_implicit_gemm_indexed_f16f16_f16f16_f16_nhwckrsc_nhwc_tilesize256x32x32_stage4_warpsize4x1x1_g1_tensor16x8x16
The different of two kernel is "f16f16_f16f16_f16 ", I want to know the five “fp16” means, and why it is different.

4

Hi,

We are moving this post to the cuDNN forum to get better help.

Thank you.