Hi,
device: A100;
For cudnn function “CuDNN (v8400) function cudnnConvolutionForward()” and convoluton params:
x, w, y dataType =CUDNN_DATA_FLOAT ;
mode: CUDNN_CROSS_CORRELATION;
mathType: CUDNN_FMA_MATH;
layout: NCHW;
I found that cudnn will select fp32 vector core or tf32 tensor for different shape(n c h w, r, s, stride and dilation);
What’s rules that cudnn convoluton select fp32 vector core and tf 32 tensor core based on shapes?