Adapt FP32 operation with TF32

Accelerated Computing GPU-Accelerated Libraries
1

I have 3 matrix A, B & C. A & B are in Half data type and B is Single. The operation is like this:
A = A + BxC.

Now I am using “cublasGemmEx” for this operation, but I want to do this operation in Half and TF32. Can I do this operation without converting from FP32 to TF32? Do we have any replacment for “cublasGemmEx” which accept FP32 and do GEMM in TF32?

I am not sure if “cublasGemmEx” is doing the same now or not!

Regards,
N.S.

2

If I understand your question, you want to do compute in FP16 or TF32? According to the documentation A, B, and C must be be FP32.

3

I want to to do a GEMM with TF32, my data is in FP32.

4

?

5

it was my mistake sorry.

Originaly A,B & C are FP32. but for doing faster GEMM I am converting A&B to half and storing final in FP32 with cublasGemmEx.

But I thinking about doing this GEMM a bit little faster. Maybe TF 32 for all matrixes or some other combination.