Hi,
I’m a little confused about how the parameter computeType affects the calculation.
For example, in cusparseSpMV, the types of A, X and Y are 16f and computeType is CUDA_R_32F. Does it mean that in the computation, the types of A, X and Y will be converted to f32 first, then they will be calculated, and finally the result Y will be converted to f16?
Thank you.
This would be the typical, original tensor core calculation. The calculation is a 16-bit float by a 16-bit float, yielding a 32-bit float result. Corresponding results (within a single tensor core op) are accumulated in 32-bit float. The accumulated result is converted back to 16-bit float upon storage, i.e. on completion of the underlying sass tensor core operation. You can refer to diagrams 7 and 8 here, where the only deviation from diagram 8 in this case is that the FP32 result gets converted to FP16 at the point of storage of that result.
This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.