RGB to HSY fixed functions using CPU

[larochj]

No description provided.

[doug-walker]

You need an applyHSYToRGB, with a corresponding implementation above.

[larochj]

Oh true, I only made one part of the convertion, thanks!

[larochj]

Fixed, but I still get a failure for the test style_RGB_TO_HSY_LIN_inv.
Any Idea why?

[doug-walker]

I looked at the code but did not notice a problem. The tolerance you're using is tighter than SynColor. It might help if you paste the results into the thread here.

[larochj]

Here it is:

[ 49/230] [FixedFunction / style_RGB_TO_HSY_LIN_inv          ] - FAILED - 
Maximum error: 45990.25 at pixel: 63694 on component 2 larger than epsilon.
scr = {1.915679932, 1.915691376, 1.91570282, 1.915714264}
cpu = {-1317085.375, 736788.9375, -3420181.75, 1.915714264}
gpu = {-1334796.875, 746696.5625, -3466172, 1.915714264}
absolute tolerance=9.999999747e-05

[doug-walker]

Please let me know which tests pass and paste the results for any that don't.

[doug-walker]

The tolerance in SynColor was 1e-4, is it passing at 1e-6?

[larochj]

Yes they all pass exept style_RGB_TO_HSY_LIN_inv.
And the error margin is bigger than 1e-4:

[ 49/230] [FixedFunction / style_RGB_TO_HSY_LIN_inv          ] - FAILED - 
Maximum error: 45990.25 at pixel: 63694 on component 2 larger than epsilon.
scr = {1.915679932, 1.915691376, 1.91570282, 1.915714264}
cpu = {-1317085.375, 736788.9375, -3420181.75, 1.915714264}
gpu = {-1334796.875, 746696.5625, -3466172, 1.915714264}
absolute tolerance=9.999999747e-05

[doug-walker]

Hi Jonathan. Looks like you pasted the result of the GPU test. Let's start with the CPU tests. Would you please paste the four input and output values?

[larochj]

All CPU test passes:

[ 667/1091] [FixedFunctionOpCPU / RGB_TO_HSY_LIN                         ] - PASSED
[ 668/1091] [FixedFunctionOpCPU / RGB_TO_HSY_LOG                         ] - PASSED
[ 669/1091] [FixedFunctionOpCPU / RGB_TO_HSY_VID                         ] - PASSED
[ 688/1091] [FixedFunctionOps / RGB_TO_HSY_LIN                           ] - PASSED
[ 689/1091] [FixedFunctionOps / RGB_TO_HSY_LOG                           ] - PASSED
[ 690/1091] [FixedFunctionOps / RGB_TO_HSY_VID                           ] - PASSED

For RGB_TO_HSY_LIN here are the value used in the CPU test in file
tests/cpu/ops/fixedfunction/FixedFunctionOpCPU_tests.cpp

    const std::vector<float> hsyFrame {
         4.70554752e-01f, 9.12594033f, 3.26650218e-02f, 0.f,
         0.75f,           0.22196741f, 0.38596f,        1.f,
         0.08333333f,     0.12976444f, 0.034974f,       0.f,
         0.96296296f,     9.7034f,     -0.1862f,        1.f };

    const std::vector<float> rgbFrame {
        -0.075290f,  0.078996f, -0.108397f, 0.f,
         0.3f,       0.4f,       0.5f,      1.f,
         0.05f,      0.03f,      0.04f,     0.f,
         0.3f,      -0.4f,       0.5f,      1.f };

[doug-walker]

OK, that's confusing. Above you wrote that style_RGB_TO_HSY_LIN_inv was not passing, but now it is?

If it is, that means we should be able to rely on the CPU values for comparison against the GPU tests.

The GPU result you pasted in a different thread had a very extreme output value. If the GPU matches the CPU through the normal range of values, that is the main criteria. If they diverge only at extreme values, that would not be too surprising given that it's a separate implementation. We don't need to fix that. You could limit the GPU test input values or perhaps adjust the parameter values so that it passes.

[larochj]

Alright, I changed it to use custom values for this specific test, thanks!

[doug-walker]

I looked at the code but did not notice a problem. The tolerance you're using is tighter than SynColor. It might help if you paste the results into the thread here.