LatticeDeform: Performance
ClosedPublic
Actions

Authored by Jeroen Bakker (jbakker) on Oct 2 2020, 9:17 AM.

Details

Reviewers

Campbell Barton (campbellbarton)

Group Reviewers

Modeling

Commits

rB042143440d76: LatticeDeform: Performance

Summary

This patch improves the single core performance of the lattice deform.

Prefetching deform vert during initialization. This data is constant for each innerloop. This reduces the complexity of the inner loop what makes more CPU resources free for other optimizations.
Prefetching the Lattice instance. It was constant. Although performance wise this isn't noticeable it is always good to free some space in the branch prediction tables.
Remove branching in all loops by not exiting when the effect of the loop isn't there. The checks in the inner loops detected if this loop didn't have any effect on the final result and then continue to the next loop. This made the branch prediction unpredictable and a lot of mis predictions were done. For smaller inner loops it is always better to remove unpredictable if statements by using branchless code patterns.
Use SSE2 instruction when available.

This gives 50% performance increase measured on a
Intel(R) Core(TM) i7-6700 CPU @ 3.40GHz with GCC 9.3.
Also check other compilers.

Before:

performance_no_dvert_10000 (4 ms)
performance_no_dvert_100000 (30 ms)
performance_no_dvert_1000000 (268 ms)
performance_no_dvert_10000000 (2637 ms)

After:

performance_no_dvert_10000 (3 ms)
performance_no_dvert_100000 (21 ms)
performance_no_dvert_1000000 (180 ms)
performance_no_dvert_10000000 (1756 ms)

Generally looks fine, only minor things noted.

source/blender/blenkernel/intern/lattice_deform.c
97	`+ 1` should be `+ sizeof(float)`.
124	Could just use `int` here?, though `int32_t` was to be used when using 4 bytes is significant (if it's converted to char[4] for example).
227	Is `CLAMPIS` necessary here? It looks like the value will never be zero, `min_ii(ww * w_stride, idx_w_max);` seems to work from a quick test. Same for the two uses below.
255	picky The cast can be avoided here `&co_vec[0]`, avoids having to make sure casts are OK - in general.
source/blender/blenkernel/intern/lattice_deform_test.cc
39	picky - prefer `float (*coords)[3]`, makes indexing the values read nicer.

source/blender/blenkernel/intern/lattice_deform.c
227	Correction "It looks like the value will never be below zero"

source/blender/blenkernel/intern/lattice_deform.c
174	Should be `ifdef`, same below.

source/blender/blenkernel/intern/lattice_deform.c
255	Correcting own statement, cast can't be removed as it's `__m128`.

source/blender/blenkernel/intern/lattice_deform.c
227	`wi` can be 0, selecting the first element of a lattice dimension. In this case `ww` is -1. In the previous implementation this case was handled in 204: if (ww > 0) { ... 212: } else { 213: idx_w = 0; 214: } `min_ii(ww * w_stride, idx_w_max);` would still need to be clamped.