How to handle difference in sinf/cosf for M1 #107829

New Issue

Chris Blackbourn · 2023-05-11T03:19:43+02:00

Chris Blackbourn commented

2023-05-11 03:19:43 +02:00

As discovered in #104513 , it appears that the single-precision floating point trig functions are producing slightly different results on apple silicon CPUs (M1, M2) compared with x86_64 CPUs, for certain inputs.
e.g.

  sinf(0.8960554599761962890625) = 0.780868828296661376953125 <- output on x86_64
  sinf(0.8960554599761962890625) = 0.78086888790130615234375 <- output on Apple M1 and M2

This is a design task to help understand this problem, track solutions and mitigations.

See for example, 0eba9e41bf

As discovered in https://projects.blender.org/blender/blender/issues/104513 , it appears that the single-precision floating point trig functions are producing slightly different results on apple silicon CPUs (M1, M2) compared with x86_64 CPUs, for certain inputs. e.g. ``` sinf(0.8960554599761962890625) = 0.780868828296661376953125 <- output on x86_64 sinf(0.8960554599761962890625) = 0.78086888790130615234375 <- output on Apple M1 and M2 ``` This is a design task to help understand this problem, track solutions and mitigations. See for example, 0eba9e41bff97659bdb5e51ba814d3e5dd150b43

Chris Blackbourn added the

labels 2023-05-11 03:19:43 +02:00

Chris Blackbourn added this to the Core project 2023-05-11 03:20:10 +02:00

Chris Blackbourn added

and removed

labels 2023-05-11 03:20:32 +02:00

Chris Blackbourn commented

2023-05-11 03:29:43 +02:00

One thing we should probably do is see if the problem exists on double precision as well.

For example, take all 4 billion single precision floating point numbers, convert them to double, call cos(x) and sin(x) on them, write to a file.
Compare across CPU architectures.
(We can use python for this, the problem is the CPU, not in the c++ math library)

... assuming they are identical, we can then undefine sinf / cosf, and replace with

float blender_sinf(x){
  return float( sin( double(x) ) );
}

... obviously there's a performance hit here, but it may well be small, as long as we store floats as single precision in memory, only promoting the type to call sin/cos

One thing we should probably do is see if the problem exists on `double precision` as well. For example, take all 4 billion single precision floating point numbers, convert them to double, call cos(x) and sin(x) on them, write to a file. Compare across CPU architectures. (We can use python for this, the problem is the CPU, not in the c++ math library) ... assuming they are identical, we can then undefine sinf / cosf, and replace with ``` float blender_sinf(x){ return float( sin( double(x) ) ); } ``` ... obviously there's a performance hit here, but it may well be small, **as long as we store floats as single precision in memory**, only promoting the type to call sin/cos

Chris Blackbourn commented

2023-05-11 03:31:30 +02:00

Another thing we should do is ... google search... and see if other projects are seeing this problem

Chris Blackbourn commented

2023-05-11 03:33:49 +02:00

Also, check the other trig function. atan2f(), asinf(), coshf() etc are probably vulnerable.

sqrtf, fmodf, log10f etc are probably safe. (These functions are not transcendental, so the output is completely determined by the rounding mode.)

Also, check the other trig function. `atan2f(), asinf(), coshf()` etc are probably vulnerable. `sqrtf, fmodf, log10f` etc are probably safe. (These functions are not transcendental, so the output is completely determined by the rounding mode.)

Iliya Katushenock removed the

Status

Needs Triage

label 2023-05-11 07:27:24 +02:00

Bastien Montagne commented

2023-05-24 15:23:59 +02:00

I do not really see why this should be an issue... A float number can only be trusted up to 6 (or 7 in best case) digits, and both results are equals up to their 7th digits.

If an algorithm is sensitive to that level (giving such huge different results caused by such small differences of values), then either this should be documented and accepted as a known limitation of that algorithm, or the algorithm should be revised and made more stable.

Fairly certain that using doubles would not help here, you'd end up facing the same kind of issues, just more rarely probably...

I do not really see why this should be an issue... A float number can only be trusted up to 6 (or 7 in best case) digits, and both results are equals up to their 7th digits. If an algorithm is sensitive to that level (giving such huge different results caused by such small differences of values), then either this should be documented and accepted as a known limitation of that algorithm, or the algorithm should be revised and made more stable. Fairly certain that using doubles would not help here, you'd end up facing the same kind of issues, just more rarely probably...

👍 1

Chris Blackbourn commented

2023-05-25 00:11:28 +02:00

If an algorithm is sensitive to that level (giving such huge different results caused by such small differences of values), then either this should be documented and accepted as a known limitation of that algorithm, or the algorithm should be revised and made more stable.

Yeah, it's an odd situation, where the algorithm is choosing the "longest" side of a square.

In theory, both sides should be equal, no problem.

In floating point math we're only comparing the round-off error, and the two different CPUs disagree about what the round-off error is for these particular inputs.

So on some CPUs, the rectangle is "tall" and stays where it is. On others, the rectangle is "wide" and gets rotated 90degrees to stand up.

Again, normally, this wouldn't be a problem as no-one would ever notice, you run the algorithm, compute the rectangle, save the file to disk and move on.

Except that for geometry nodes, the computation is run a second time when the file is loaded. That is, the same .blend file has dramatically different appearance depending on the CPU it's currently being loaded, and is independent of the CPU of the file when it was saved.

As just one example, if you author a .blend file on a Mac, and try and use an intel render farm, some of your textures might be rotated 90 degrees!

Fairly certain that using doubles would not help here, you'd end up facing the same kind of issues, just more rarely probably...

Well.. surprisingly, we can actually write a computer program to figure this out and know for sure:

float x = (float)sin( (double) y)

That y has 52 bits in the significand, and the x has 22 bits. So we're 2^30 less likely to hit the bug. Given there's only 2^32 single precision floats and not all of them have the problem, it may well be the case that reducing the problem incidence by a factor of 2^30 does indeed eliminate all of the problems rather than just some of them.

We could (i.e "should") write a program to just go and check all 4 billion of them and see what happens...

> If an algorithm is sensitive to that level (giving such huge different results caused by such small differences of values), then either this should be documented and accepted as a known limitation of that algorithm, or the algorithm should be revised and made more stable. Yeah, it's an odd situation, where the algorithm is choosing the "longest" side of a square. In theory, both sides should be equal, no problem. In floating point math we're *only* comparing the round-off error, and the two different CPUs disagree about what the round-off error is for these particular inputs. So on some CPUs, the rectangle is "tall" and stays where it is. On others, the rectangle is "wide" and gets rotated 90degrees to stand up. Again, normally, this wouldn't be a problem as no-one would ever notice, you run the algorithm, compute the rectangle, save the file to disk and move on. Except that for geometry nodes, the computation is run a second time when the file is loaded. That is, the same .blend file has dramatically different appearance depending on the CPU it's currently being *loaded*, and is independent of the CPU of the file when it was *saved*. As just one example, if you author a .blend file on a Mac, and try and use an intel render farm, some of your textures might be rotated 90 degrees! > Fairly certain that using doubles would not help here, you'd end up facing the same kind of issues, just more rarely probably... Well.. surprisingly, we can actually write a computer program to figure this out and know for sure: ``` float x = (float)sin( (double) y) ``` That` y` has 52 bits in the significand, and the `x` has 22 bits. So we're `2^30` less likely to hit the bug. Given there's only `2^32` single precision floats and not all of them have the problem, it may well be the case that reducing the problem incidence by a factor of `2^30` does indeed eliminate *all* of the problems rather than just some of them. We could (i.e "should") write a program to just go and check all 4 billion of them and see what happens...

Bastien Montagne commented

2023-05-25 09:55:14 +02:00

If an algorithm is sensitive to that level (giving such huge different results caused by such small differences of values), then either this should be documented and accepted as a known limitation of that algorithm, or the algorithm should be revised and made more stable.

Yeah, it's an odd situation, where the algorithm is choosing the "longest" side of a square.

In theory, both sides should be equal, no problem.

Can't you simply use a relative epsilon comparison instead of exact equality here?

> > If an algorithm is sensitive to that level (giving such huge different results caused by such small differences of values), then either this should be documented and accepted as a known limitation of that algorithm, or the algorithm should be revised and made more stable. > > Yeah, it's an odd situation, where the algorithm is choosing the "longest" side of a square. > > In theory, both sides should be equal, no problem. Can't you simply use a relative epsilon comparison instead of exact equality here?

Chris Blackbourn commented

2023-05-25 10:59:49 +02:00

In theory, both sides should be equal, no problem.

Can't you simply use a relative epsilon comparison instead of exact equality here?

I think the underlying problem is closer to argmax, so it's something like :

argmax( [0.796857, 0.796855, 0.796856, 0.796854] )  = 0 // x86
argmax( [0.796856, 0.796855, 0.796857, 0.796854] )  = 2 // M1

For a more visual example:

To make a "Tall" rectangle, should the diamond turn +45 degrees or -45 degrees?

Just biasing the outputs one way or the other doesn't really solve the underlying problem, which is that the decision is based purely on the round-off error, and the round-off error is different on different CPUs.

> > In theory, both sides should be equal, no problem. > > Can't you simply use a relative epsilon comparison instead of exact equality here? I think the underlying problem is closer to argmax, so it's something like : ``` argmax( [0.796857, 0.796855, 0.796856, 0.796854] ) = 0 // x86 argmax( [0.796856, 0.796855, 0.796857, 0.796854] ) = 2 // M1 ``` For a more visual example: ![image](/attachments/75267629-3653-4989-bef4-314f62a37e16) To make a "Tall" rectangle, should the diamond turn +45 degrees or -45 degrees? Just biasing the outputs one way or the other doesn't really solve the underlying problem, which is that the decision is based purely on the round-off error, and the round-off error is different on different CPUs.

image.png

34 KiB

Bastien Montagne commented

2023-05-25 11:36:06 +02:00

I don't understand. I'd rather implement own version of argmax for this, which does relative epsilon comparison, and if several items are 'equal enough', just always e.g. return the first one.

That way you get stable result regardless of 'implementation noise'. In that case, it does not matter to get the 'absolute exact' answer to your question, you just need a way to always get the same answer on all platforms.

I don't understand. I'd rather implement own version of argmax for this, which does relative epsilon comparison, and if several items are 'equal enough', just always e.g. return the first one. That way you get stable result regardless of 'implementation noise'. In that case, it does not matter to get the 'absolute exact' answer to your question, you just need a way to always get the same answer on all platforms.

👍 1

Bastien Montagne commented

2023-05-25 11:42:08 +02:00

(To be specific, use regular argmax, and compare found max value too all previous ones, and return first item within epsilon range of found max value)

Chris Blackbourn referenced this issue from a commit

2023-06-08 04:25:08 +02:00

Fix #104513: UV packing produces different results on x86 vs apple silicon

Sign in to join this conversation.

No Label

Download

What's New

Blender Studio

Manual

Developers Blog

Documentation

Benchmark

Blender Conference

Development Fund

One-time Donations

How to handle difference in sinf/cosf for M1 #107829