Minor cleanups.
1. Further BLAS and heap allocation cleanups in schur_eliminator_impl.h
2. Modularize blas.h using macros.
3. Lint cleanups from William Rucklidge.
4. Small changes to jet.h
5. ResidualBlock now uses blas.h
Performance improvements:
For static and dynamic sized blocks, the peformance is not changed much.
-use_quaternions -ordering user -linear_solver sparse_schur
master change
problem: 16-22106
gcc 3.4 3.3
clang 2.8 2.7
problem: 49-7776
gcc 1.7 1.7
clang 1.4 1.4
problem: 245-198739
gcc 80.1 79.6
clang 80.6 76.2
problem: 257-65132
gcc 12.2 12.0
clang 10.4 10.2
problem: 356-226730
gcc 99.0 96.8
clang 88.9 88.3
problem: 744-543562
gcc 361.5 356.2
clang 352.7 343.5
problem: 1024-110968
gcc 45.9 45.6
clang 42.6 42.1
However, performance when using local parameterizations is
significantly improved due to residual_block.cc using blas.h
-use_quaternions -use_local_parameterization -ordering user -linear_solver sparse_schur
master change
problem: 16-22106
gcc 3.6 3.3
clang 3.5 2.8
problem: 49-7776
gcc 1.8 1.6
clang 1.7 1.4
problem: 245-198739
gcc 79.7 76.1
clang 79.7 73.0
problem: 257-65132
gcc 12.8 11.9
clang 12.3 9.8
problem: 356-226730
gcc 101.9 93.5
clang 105.0 86.8
problem: 744-543562
gcc 367.9 350.5
clang 355.3 323.1
problem: 1024-110968
gcc 43.0 40.3
clang 41.0 37.5
Change-Id: I6dcf7476ddaa77cb116558d112a9cf1e832f5fc9
diff --git a/include/ceres/jet.h b/include/ceres/jet.h
index 1238123..000bd1c 100644
--- a/include/ceres/jet.h
+++ b/include/ceres/jet.h
@@ -348,8 +348,8 @@
// b + v (b + v)(b - v) b^2
//
// which holds because v*v = 0.
- h.a = f.a / g.a;
- const T g_a_inverse = 1.0 / g.a;
+ const T g_a_inverse = T(1.0) / g.a;
+ h.a = f.a * g_a_inverse;
const T f_a_by_g_a = f.a * g_a_inverse;
for (int i = 0; i < N; ++i) {
h.v[i] = (f.v[i] - f_a_by_g_a * g.v[i]) * g_a_inverse;
@@ -450,7 +450,7 @@
Jet<T, N> sqrt(const Jet<T, N>& f) {
Jet<T, N> g;
g.a = sqrt(f.a);
- const T two_a_inverse = 1.0 / (T(2.0) * g.a);
+ const T two_a_inverse = T(1.0) / (T(2.0) * g.a);
g.v = f.v * two_a_inverse;
return g;
}
