Trigonometry without sine/cosine

CoBB · Post by **CoBB** » Mon 19 Sep, 2005 9:10 pm

If the LUTs are big, they cannot be cached, especially if they are traversed in a cache-unfriendly way. Cache misses can be very expensive. You often cannot tell in advance what the fastest method is.

coelurus · Post by **coelurus** » Mon 19 Sep, 2005 9:46 pm

The only way is to write a test-program that does something meaningful. Generally, today's PCs will evaluate trig functions very quickly and with better accuracy than what any LUTs can give you. Not quite like that on calcs of course...

threefingeredguy · Post by **threefingeredguy** » Tue 20 Sep, 2005 2:41 am

How do PCs do it?

CoBB · Post by **CoBB** » Tue 20 Sep, 2005 5:07 am

It's a single assembly instruction. Formerly (pre-Pentium era) this functionality was available in separate coprocessors, but it was integrated in the first Pentium.

Post by **benryves** » Tue 20 Sep, 2005 9:54 am

If the CPU can calculate trig itself, then you should have no real need for a LUT (unless the instruction was very ineffecient).
Of course, using maths to favour a LUT was what gave us the ATi DOOM3 shader tweak.

So LUTs are not always the best.

crzyrbl · Post by **crzyrbl** » Thu 22 Sep, 2005 4:56 am

im happy as long as what im learning in AP calc doesnt become obsolete.

Kerey · Post by **Kerey** » Thu 22 Sep, 2005 9:58 pm

benryves wrote:If the CPU can calculate trig itself, then you should have no real need for a LUT (unless the instruction was very ineffecient).
Of course, using maths to favour a LUT was what gave us the ATi DOOM3 shader tweak. So LUTs are not always the best.

Ok, I guess one assembly instruction might be a tad faster than an LUT, heh. I was thinking they were a Taylor expansion or something.

coelurus · Post by **coelurus** » Fri 23 Sep, 2005 8:51 am

They probably are, somewhere deep down below... Which is a little worrying, so to settle my nerves, I made a benchmark calculating a whole heap of sins and taking values from a high resolution LUT, compiled with all optimizations I could find for my amd64 3000+. The results were a little surprising, the LUT was nearly 100x faster

I've never looked into any asm so I didn't use any weird vector-tricks with SSE to get a whole heap of values in a row.

This doesn't imply LUTs are better overall, remember that