Fix T78324: Different Sky Texture results between CPU and GPU
The problem here was numerical precision: The code calculates the angle between sun and view direction, and the usual acos(dot(a, b)) approach for that has poor numerical performance for almost parallel angles. As a result, the generally tiny difference between floating point computation between CPU and GPU was enough to make the sun vanish at different radii, causing different results. The new version fixes the difference by making the computation much more robust on both platforms.
This commit is contained in:
parent
afcb41a0aa
commit
7fcb6bc59c
|
@ -122,6 +122,11 @@ vector geographical_to_direction(float lat, float lon)
|
|||
return vector(cos(lat) * cos(lon), cos(lat) * sin(lon), sin(lat));
|
||||
}
|
||||
|
||||
float precise_angle(vector a, vector b)
|
||||
{
|
||||
return 2.0 * atan2(length(a - b), length(a + b));
|
||||
}
|
||||
|
||||
color sky_radiance_nishita(vector dir, float nishita_data[9], string filename)
|
||||
{
|
||||
/* definitions */
|
||||
|
@ -138,7 +143,7 @@ color sky_radiance_nishita(vector dir, float nishita_data[9], string filename)
|
|||
if (dir[2] >= 0.0) {
|
||||
/* definitions */
|
||||
vector sun_dir = geographical_to_direction(sun_elevation, sun_rotation + M_PI_2);
|
||||
float sun_dir_angle = acos(dot(dir, sun_dir));
|
||||
float sun_dir_angle = precise_angle(dir, sun_dir);
|
||||
float half_angular = angular_diameter / 2.0;
|
||||
float dir_elevation = M_PI_2 - direction[0];
|
||||
|
||||
|
|
|
@ -145,7 +145,7 @@ ccl_device float3 sky_radiance_nishita(KernelGlobals *kg,
|
|||
if (dir.z >= 0.0f) {
|
||||
/* definitions */
|
||||
float3 sun_dir = geographical_to_direction(sun_elevation, sun_rotation + M_PI_2_F);
|
||||
float sun_dir_angle = acos(dot(dir, sun_dir));
|
||||
float sun_dir_angle = precise_angle(dir, sun_dir);
|
||||
float half_angular = angular_diameter / 2.0f;
|
||||
float dir_elevation = M_PI_2_F - direction.x;
|
||||
|
||||
|
|
|
@ -787,6 +787,16 @@ ccl_device_inline float compare_floats(float a, float b, float abs_diff, int ulp
|
|||
return (abs(__float_as_int(a) - __float_as_int(b)) < ulp_diff);
|
||||
}
|
||||
|
||||
/* Calculate the angle between the two vectors a and b.
|
||||
* The usual approach acos(dot(a, b)) has severe precision issues for small angles,
|
||||
* which are avoided by this method.
|
||||
* Based on "Mangled Angles" from https://people.eecs.berkeley.edu/~wkahan/Mindless.pdf
|
||||
*/
|
||||
ccl_device_inline float precise_angle(float3 a, float3 b)
|
||||
{
|
||||
return 2.0f * atan2f(len(a - b), len(a + b));
|
||||
}
|
||||
|
||||
CCL_NAMESPACE_END
|
||||
|
||||
#endif /* __UTIL_MATH_H__ */
|
||||
|
|
Loading…
Reference in New Issue