I want to konw if the Atype, the Btype and the Ctype are all FP16, what is the compute type when compute with alpha, bata in cublasGemmEx? Is it also FP16 or it depends on the computeType parameter? If the computeType is FP32, will the C matrix be converted to FP32 to participate in the calculation of beta, and then the result of the multiplication and addition will be converted to FP16 output?
What is the compute type in cublasGemmEx when compute with alpha, beta?
It is still possible for 32-bit calculations to be done with 16 bit inputs and outputs. See here
Hello, I konw that the computeType parameter will affect the Matrix-Matrix multiplication operations. But I am more interested in the alpha, beta compute process. The result of matrix multiplication is converted to fp16, and then the multiplication and addition operation of alpha and beta is performed, and then the result is output? Or keep the result of matrix multiplication as fp32 first, perform alpha and beta multiplication and addition operations, and then convert the result to fp16 output?