Fix #110255: Cover up CPU/GPU differences with small suns in light tree #110307

Alaska · 2023-07-20T16:28:40+02:00

Alaska commented

2023-07-20 16:28:40 +02:00

This pull request covers up a subtle difference between the CPU and GPU
when rendering with a light tree. Specifically a case where the user
has a sun light with a small angle.

The difference was caused by the dot() function being different between
CPU and GPU backends, with the GPU showing more meaningful
floating-point precision losses when working with small suns.

This pull request covers up a subtle difference between the CPU and GPU when rendering with a light tree. Specifically a case where the user has a sun light with a small angle. The difference was caused by the dot() function being different between CPU and GPU backends, with the GPU showing more meaningful floating-point precision losses when working with small suns.

❤️ 1

Alaska added the

Module

Render & Cycles

label 2023-07-20 16:28:40 +02:00

Alaska requested review from Brecht Van Lommel 2023-07-20 16:29:06 +02:00

Alaska requested review from Weizhen Huang 2023-07-20 16:29:17 +02:00

Alaska commented

2023-07-20 16:31:44 +02:00

This pull request is for the Main branch of Blender, but should be back ported to the blender-v3.6-release branch if accepted.

For reference, here's an example of the issue:
In light_tree_importance there is a section of code that is run if cos_theta >= cos_theta_u. For distant lights this statement is always true. But on the GPU backend under specific circumstances, there were floating point precision losses which resulted in it not being true, resulting in the distant light not being sampled.

This happens because when a sun light has a small angle (0 in this example), it has a high cos_theta_u (1 in this example). cos_theta_u = cos(sun_angle)
cos_theta is calculated as dot(bcone.axis, -point_to_centroid), and in the case of distant lights, bcone.axis = -point_to_centroid and both vectors are normalized, so cos_theta should be 1. But for some orientations of the sun lamp, cos_theta was not equal to 1 on the GPU. Instead cos_theta would be something like 0.99999994. This resulted in the if statement not being true, leading to the distant light having an importance of 0 and not being sampled.

It should also be noted that I went with a "simple" fix to the issue, but it could have subtle effects on lights in other situations. Other fixes could also be implemented if you wish.

Some examples of alternative fixes I can think of are:

if (bcone.axis == -point_to_centroid) then (cos_theta = 1) otherwise use cos_theta = dot(bcone.axis, -point_to_centroid). From a quick look into the code, there's the potential for a floating point precision issue resulting in bcone.axis != -point_to_centroid due to how they are calculated. But it should be fixed with #110374.
if distance light, then (cos_theta = 1)

Note: The floating point precision issues with the dot function aren't just a weird thing with the GPU. It's actually because the GPU and CPU run different code. Specifically the CPU takes advantage of SSE optimizations. If you swap the CPU over to using the same dot function as the GPU, then it will have the exact same issue.

This pull request is for the Main branch of Blender, but should be back ported to the blender-v3.6-release branch if accepted. For reference, here's an example of the issue: In `light_tree_importance` there is a section of code that is run if `cos_theta >= cos_theta_u`. For distant lights this statement is always true. But on the GPU backend under specific circumstances, there were floating point precision losses which resulted in it not being true, resulting in the distant light not being sampled. This happens because when a sun light has a small angle (0 in this example), it has a high `cos_theta_u` (1 in this example). `cos_theta_u = cos(sun_angle)` `cos_theta` is calculated as `dot(bcone.axis, -point_to_centroid)`, and in the case of distant lights, `bcone.axis = -point_to_centroid` and both vectors are normalized, so `cos_theta` should be `1`. But for some orientations of the sun lamp, `cos_theta` was not equal to `1` on the GPU. Instead `cos_theta` would be something like `0.99999994`. This resulted in the if statement not being true, leading to the distant light having an importance of 0 and not being sampled. --- It should also be noted that I went with a "simple" fix to the issue, but it could have subtle effects on lights in other situations. Other fixes could also be implemented if you wish. Some examples of alternative fixes I can think of are: - `if (bcone.axis == -point_to_centroid)` then `(cos_theta = 1)` otherwise use `cos_theta = dot(bcone.axis, -point_to_centroid)`. From a quick look into the code, there's the potential for a floating point precision issue resulting in `bcone.axis != -point_to_centroid` due to how they are calculated. But it should be fixed with #110374. - `if distance light`, then `(cos_theta = 1)` --- Note: The floating point precision issues with the dot function aren't just a weird thing with the GPU. It's actually because the GPU and CPU run different code. Specifically the CPU takes advantage of SSE optimizations. If you swap the CPU over to using the same dot function as the GPU, then it will have the exact same issue.

Alaska changed title from ~~Fix #110255: Cover up CPU/GPU differences in light tree.~~ to Fix #110255: Cover up CPU/GPU differences with small suns in light tree

2023-07-21 03:29:47 +02:00

Alaska force-pushed fix-small-sun-light-tree-sampling from 5010b1d7d5 to 8896301a7a

2023-07-22 14:24:23 +02:00

Compare

Alaska commented

2023-07-26 09:46:28 +02:00

As pointed out earlier, the rendering error occurs on the GPU but not the CPU and this was due differences between the dot() function on the CPU and GPU. Specifically because the SSE optimisations used on the CPU produce a slightly different result due to floating point precision differences. If you remove the SSE code from the dot() function, relying on the "simple math" that the GPU uses, then the CPU will have the exact same issues as the GPU.

I tried to align the way the "simple math" version of dot() worked with the SSE operation. Specifically changing the order of operations and separating operations into different variables. And making those changes would fix the issue on the CPU with the SSE code removed, but wouldn't fix it on the GPU (I was testing the Metal backend), either due to some other issue or compiler optimisations. For reference, here are some examples of the different things I tried:

// Orignal
return a.x * a.x + a.y * b.y + a.z * b.z;
// Test 1 (Brackets to try and influence order of operations)
return a.x * a.x + (a.y * b.y + a.z * b.z);
// Test 2 (Change the order)
return  a.y * b.y + a.z * b.z + a.x * a.x;
// Test 3 (Separating variables)
float x = a.x * a.x;
float y = a.y * b.y;
float z = a.z * b.z;
return x + y + z;
// Test 4 (Adjust order of operations with separated variables)
float x = a.x * a.x;
float y = a.y * b.y;
float z = a.z * b.z;
float yz = y + z;
return yz + x;

As mentioned above, some tests did work to resolve the issue on the CPU with SSE code removed. But it didn't fix the issue on the GPU. Maybe there was a specific combination I missed?

As pointed out earlier, the rendering error occurs on the GPU but not the CPU and this was due differences between the `dot()` function on the CPU and GPU. Specifically because the SSE optimisations used on the CPU produce a slightly different result due to floating point precision differences. If you remove the SSE code from the `dot()` function, relying on the "simple math" that the GPU uses, then the CPU will have the exact same issues as the GPU. I tried to align the way the "simple math" version of `dot()` worked with the SSE operation. Specifically changing the order of operations and separating operations into different variables. And making those changes would fix the issue on the CPU with the SSE code removed, but wouldn't fix it on the GPU (I was testing the Metal backend), either due to some other issue or compiler optimisations. For reference, here are some examples of the different things I tried: ``` // Orignal return a.x * a.x + a.y * b.y + a.z * b.z; // Test 1 (Brackets to try and influence order of operations) return a.x * a.x + (a.y * b.y + a.z * b.z); // Test 2 (Change the order) return a.y * b.y + a.z * b.z + a.x * a.x; // Test 3 (Separating variables) float x = a.x * a.x; float y = a.y * b.y; float z = a.z * b.z; return x + y + z; // Test 4 (Adjust order of operations with separated variables) float x = a.x * a.x; float y = a.y * b.y; float z = a.z * b.z; float yz = y + z; return yz + x; ``` As mentioned above, some tests did work to resolve the issue on the CPU with SSE code removed. But it didn't fix the issue on the GPU. Maybe there was a specific combination I missed?

Weizhen Huang requested changes 2023-07-31 14:24:26 +02:00

Weizhen Huang left a comment

Such value seems quite arbitrary, I would lean towards assigning 1 to cos_theta when isequal(bcone.axis, -point_to_centroid) and comment that dot product doesn't necessarily returns 1 in such cases due to precision issue.
I guess part of the precision is already lost when normalizing the vector, which we can't fully recover in dot().

Such value seems quite arbitrary, I would lean towards assigning 1 to `cos_theta` when `isequal(bcone.axis, -point_to_centroid)` and comment that dot product doesn't necessarily returns 1 in such cases due to precision issue. I guess part of the precision is already lost when normalizing the vector, which we can't fully recover in `dot()`.

intern/cycles/kernel/light/tree.h Outdated

						
				@ -187,3 +187,2 @@

				   * precision losses in the dot() function used to calculate cos_theta. */

				  float cos_min_outgoing_angle;

				  if ((cos_theta >= cos_theta_u) || (cos_theta_minus_theta_u >= cos_theta_o)) {

				    /* theta - theta_o - theta_u <= 0 */