Dear Champs,
I am asking this for our customer.
The user is evaluating CLA and reports the computation precision between C28 and CLA seem different?
For example,
They tested like this
A = A + B - C
C = A / 1024
in multiple repetitive computation times (say more than several tens of times), they found the result C differed, no matter A, B, C are in fixed point (int32) or floating point (float).
Note that the user already uses the workaround for CLA signed integer comparison (--cla_signed_compare_workaround) in the compiler in this case.
The user said the C28 result was correct (meaning same result as that with PC Excel results), but CLA result was not correct after several tens of times of computation (meaning the result could differ up to integer when float format was used).
Questions:
1. In our understanding, the fixed-point format (int32) and floating-point (float) are in the same format on both C28 and CLA. That is, with floating-point, both are in IEEE 754 single-precision. Is it right?
2. If they are in the same format of both integer and floating-point format, the result should be the same irrespective of computation and times. Is our understanding correct?
3. Is there any limitation when using CLA for such kind of repetitive computation with accumulation?
