mfdeakin / adaptive_predicates Goto Github PK

The adaptive_eval_impl class has some complicated template meta-programming which makes it slow to compile and can prevent compilation when used with larger expressions (ie, the determinant of a 3x3 in-sphere matrix) with certain compilers (clang).
I've done some work on replacing the current system with a simpler one and am recording results here.

GCC Results:

Original tests binary size (with debug info):
3.7 MiB

Run-time integer index tests binary size:
3.3 MiB

Compile-time integer index tests binary size:
3.4 MiB

Compile times:
Original:

tests             test_benchmarks
real     user     real     user
12.286   9.774    10.996   8.581
12.020   9.609    11.043   8.612
12.137   9.626    11.030   8.603
12.150   9.648    11.036   8.655

Run-time indexing:

tests             test_benchmarks
real     user     real     user
12.161   9.619    10.209   7.896
11.609   9.226    10.213   7.914
11.595   9.168    10.147   7.814
11.743   9.245    10.163   7.888

Compile-time indexing:

tests             test_benchmarks
real     user     real     user
11.696   9.285    10.622   8.358
11.723   9.323    10.715   8.335
11.770   9.344    10.590   8.275
11.801   9.363    10.627   8.277

Original benchmarks:

Run-time indexing benchmarks:

Compile-time indexing benchmarks:

Clang Results:

Original tests binary size (with debug info):
2.2 MiB

Run-time integer index tests binary size:
2.1 MiB

Compile-time integer index tests binary size:
2.1 MiB

Compile times:
Original:

tests
real     user
12.620   12.385
12.842   12.647
12.802   12.599
12.819   12.608

Run-time indexing:

tests
real     user
12.206   11.937
12.370   12.116
12.429   12.150
12.150   11.935

Compile-time indexing:

tests
real     user
12.468   12.270
12.329   12.130
12.535   12.340
12.314   12.110

Original benchmarks:

Run-time indexing benchmarks:

Compile-time indexing benchmarks:

I'm guessing performance would be much better if shared memory was used, as is register pressure is likely very high

• Verify register pressure is the main bottleneck
• Implement or use an allocator that can choose to make use of the stack or if available, fast shared memory