Naoki Shibata
Naoki Shibata
Looks good, but I think we need some kind of a tester for making sure that this functionality is not broken.
@fpetrogalli How do you think? If you are happy with this patch, I will merge this and then add a tester.
Hello, As for trig functions, reduction is taking time. Is it okay to have sinpi instead of sin? I need to know use cases.
Could you list specific functions to implement?
What I meant is that sinpi/cospi would be a lot faster to compute than sin/cos. I did a small experiment, and I think 20000 ULP sinpi and cospi are possible.
And, I think 20000 ULP sincospif would be the best choice for trig functions. sinpif and cospif would be not so much faster than sincospif.
There is actually not much room for improvement for u35 single precision functions. For sinf function, the polynomial for the kernel function only have 4 coefficients. I can reduce it...
Okay, so I will check the argument is in the appropriate range, and then just multiply the arguments by PI. Then, it would be much faster.
I was testing this and it turned out that it is not so easy to make the trig functions faster. One thing that would help is to stop measuring error...
ULP is basically relative error. What I am suggesting is absolute error. The problem occurs, for example, when evaluating sin with argument close to n*PI/2. In this case, the result...