Comment update suggestion

[etasnadi]

In function runCublasTF32 the comment is misleading/incomplete. Based on the cublas docs, the effect of CUBLAS_COMPUTE_32F_FAST_TF32 is that it will use reduced precision TF32 math with tensor cores for faster GEMM. Based on the documentation of wmma ops in the CUDA programming guide, the input will be converted with __float_to_tf32 to a float (of numerically reduced tf32 precision).

As tensor cores in recent architectures support fp64 natively, I am curious what is the performance benefit of their usage over plain fp64 CUDA computation.

SGEMM_CUDA/src/runner.cu

Line 145 in 60cba6f

// This runs cuBLAS with mixed precision (performing the mul with operands