Strange behavior of cvt.sat.f64.f64 in the OptiX shader

Gaming and Visualization Technologies Visualization OptiX
1

It looks like cvt.sat.f64.f64 in addition to clamping double to the range [0, 1] as supposed to do ALSO MISTAKENLY PERFORMS THE BINARIZATION OF THE ARGUMENT (JUST LOOK AT THE PIXEL COLORS BELOW). For example:
For -1 we get 0;
For -0.5 we get 0;
For 0 we get 0;
For 0.25 we get 0; (instead of 0.25)
For 0.75 we get 1; (instead of 0.75)
For 1 we get 1;
For 1.5 we get 1;

I have RTX 4070, OptiX 8.0.0, CUDA 12.4.1 and Windows 10.

It looks like the problem is connected to the JIT compilation by the OptiX function optixModuleCreate. I asked question on the general CUDA forum. The people wrote simple piece of code with cvt.sat.f64.f64 and it worked as supposed to do.

2

The code comes from my OptiX shader. In my ray generation shader I have:

R = MIN_R(MAX_R(R, 0.0), 1.0)
G = MIN_R(MAX_R(G, 0.0), 1.0)
B = MIN_R(MAX_R(B, 0.0), 1.0)

where MIN_R and MAX_R are defined as:

#ifndef RENDERER_OPTIX_USE_DOUBLE_PRECISION
#define MIN_R(x, y) fminf(x, y)
#define MAX_R(x, y) fmaxf(x, y)
#else
#define MIN_R(x, y) fmin(x, y)
#define MAX_R(x, y) fmax(x, y)
#endif

which compiles into three cvt.sat.f64.f64 that don’t work as I described.

But if I change the piece of code into:

R = (R < 0) ? 0 : R;
R = (R > 1) ? 1 : R;
...

it compiles into:

setp.lt.f64 ...
selp.f64 ...
setp.gt.f64 ...
selp.f64 ...

and that works perfectly fine. I don’t have clue what’s going on, since the cvt.sat.f64.f64 theoretically speaking should be equivalent to the code with setp.lt.f64, setp.gt.f64 and self.f64 and it shouldn’t perform the “binarization”.

screenshot
screenshot802×833 31.6 KB

shaders.cu.ptx.txt (30.9 KB)
shaders.cu.txt (10.4 KB)

3

I tried to increase the number of registers in the module compile options as well as the optimization level. Nothing works. Is it possible that optixModuleCreate wrongly translates cvt.sat.f64.f64 into the SASS code? Now the most interesting part… . cvt.sat.f64.f64 works perfectly fine in “pure” CUDA. It looks like the PTX → SASS translation by the optixModuleCreate is messing it up. Check out the discussion on the topic on the general forum:

https://forums.developer.nvidia.com/t/strange-behavior-of-cvt-sat-f64-f64/

4

Could be a bug, I’ll file a report to get it investigated. In the mean time, have you tried switching between PTX and OptiX-IR?

–
David.

5

BTW which driver are you using? Also have you tried the latest CUDA toolkit, and latest driver?

–
David.

6

Thank You from Your answer. As for OptiX IR, the first time I learned about it was from Your post. On the other hand, as far as my specs are concerned, I have RTX 4070, OptiX 8.0.0, CUDA 12.4.1, VS 2019 Enterprise, Windows 10 Professional and 560.96 NVIDIA GPU Driver. Since I am now performing some piece of scientific computation, so far I haven’t considered upgrading my CUDA drivers because I am affraid evething I am working on will mess up. But upgrading GPU driver may be a very good idea since the OptiX is embedded in the drivers.

7

I was able to repro the issue just now using PTX. It does seem like a bug. This means you don’t need to try a new toolkit or new driver.

With OptiX-IR, I do not see the problem occurring. See if you can switch to using OptiX-IR.

We recommend moving to OptiX-IR in general anyway, as in the future it will be much more able to support debugging and profiling tools.

–
David.

8

Thank You one more time for Your help. It’s good to know it’s not my code’s fault… . I will surely consider switching to OptiX IR but since I am new to OptiX in general, it will take me some time to learn this stuff. Actually I am compiling my shaders in somewhat primitive manner by setting the *.ptx file as the output of my compilation and then loading it manually from the compiled *.ptx file using fread function to finally pass the NULL terminated string containing the code as the argument of the optix_ModuleCreate function.

9

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.

10

Just a small update to close out this thread, the bug was identified and fixed a few weeks ago. The fix is scheduled to roll out with the r580 drivers… which is at least a couple of months out. Thanks for reporting it!

–
David.