Change Details

The current method to update the normals does the following: - Allocate an array the same size of mverts regardless of the normals that need to be updated - Loop over the faces to accumulate the normals using an atomic_add into the array - Loop again over vertices to normalize the accumulated normal and store in in the vertex - Free the array If the SculptSession already contains a pmap that vertices to polys, all these steps can be skiped and the accumulation can be do in a single loop over the affected vertices. Using the scale mesh filter in a Threadripper 3990X to modify an entire mesh with different vertex counts I get the following results: Note: I left the timer code in case you want to check if this improvement also applies in other hardware. In order to test, just run the mesh filter after entering sculpt mode to use the current update method, then use smooth to allocate the pmap and run the test again.

The current method to update the normals does the following: - Allocate an array the same size of mverts regardless of the normals that need to be updated - Loop over the faces to accumulate the normals using an atomic_add into the array - Loop again over vertices to normalize the accumulated normal and store it in the vertex - Free the array If the SculptSession already contains a pmap that vertices to polys, all these steps can be skipped and the accumulation can be do in a single loop over the affected vertices. Using the scale mesh filter in a Threadripper 3990X to modify an entire mesh with different vertex counts I get the following results: | vertices | current | pmap | Improvement | |----------|----------|----------|------------------| | 311906 | 0.029462 | 0.001463 | 20.13 | | 1247818 | 0.050463 | 0.004267 | 11.82 | | 4990466 | 0.073909 | 0.014369 | 5.14 | Note: I left the timer code in case you want to check if this improvement also applies in other hardware. In order to test, just run the mesh filter after entering sculpt mode to use the current update method, then use smooth to allocate the pmap and run the test again.

The current method to update the normals does the following: - Allocate an array the same size of mverts regardless of the normals that need to be updated that need to be updated- Loop over the faces to accumulate the normals using an atomic_add into the array - Loop over the faces to accumulate the normals using an atomic_add into the array - Loop again over vertices to normalize the accumulated normal and store in it in the vertex - Free the array If the SculptSession already contains a pmap that vertices to polys, all these steps can be skipped and the accumulation can be do in a single loop over the affected vertices. Using the scale mesh filter in a Threadripper 3990X to modify an entire mesh with different vertex counts I get the following results: | vertices | current | pmap | Improvement | |----------|----------|----------|------------------| | 311906 | 0.029462 | 0.001463 | 20.13 | | 1247818 | 0.050463 | 0.004267 | 11.82 | | 4990466 | 0.073909 | 0.014369 | 5.14 | Note: I left the timer code in case you want to check if this improvement also applies in other hardware. In order to test, just run the mesh filter after entering sculpt mode to use the current update method, then use smooth to allocate the pmap and run the test again.