OpenCL gives white noise in depth buffer

[julienaubert]

The first screen is using OpenGL (same result with CPU). The second screen is from when using OpenCL.

[kohrt]

Hi,
I looked at it, and for me, the OpenCL version creates a bit more white noise than the OpenGL version, but not as much as in your images. Anyway, I took a closer look at the code and I found the possible cause for this problem in the bilateral filter.
I compared the OpenGL code with the OpenCL code and saw that I mixed up the order of the arguments of the select statement.
Could you replace line 165 by dist_acc += select(dist, (float3)(0.0f), c1); and try if it solves the issue for you? I will then prepare a pull request for it.

@christiankerl: While looking through the code I found some parts that could be improved.
For the bilateral filter:

• Shoudn't it only run when the full gaussian kernel fits in the image, similar to the edge aware filter if(uv.x < 1 || uv.y < 1 || uv.x > 510 || uv.y > 422)

For the edge aware filter:

• FilterDepth = !max_edge_test_ok ? 0.0f : v.x; this statement is inside the else branch of if(max_edge_test_ok), so the inline if is redundant
• FilterDepth = cond0 ? 0.0f : v.x; could be replaced by a else branch for if(!cond0), where FilteredDepth is set to zero

[boulabiar]

I have the same noise issue and after replacing the line 165 as you mentioned, the noise disappeared.

[julienaubert]

Thanks! Good catch. Now the output is more equal.

[kohrt]

added a pull-request #189 to fix it. So this issue might be closed.

[floe]

Merged #189 - but as you mentioned some other issues, I'll leave this open for now.

[floe]

Issues mentioned in #183 (comment) are somewhat related to #404 & #405, needs another look before 0.1.

[floe]

I'll close this now, I think it has been addressed by some of the refactoring steps taken by @alberth.