Replace Eigen block operations with small GEMM and GEMV loops.
1. Add Matrix-Matrix and Matrix-Vector multiply functions.
2. Replace Eigen usage in SchurEliminator with these custom
matrix operations.
3. Save on some memory allocations in ChunkOuterProduct.
4. Replace LDLT with LLT.
As a result on problem-16-22106-pre.txt, the linear solver time
goes down from 1.2s to 0.64s.
Change-Id: I2daa667960e0a1e8834489965a30be31f37fd87f
diff --git a/internal/ceres/schur_eliminator.h b/internal/ceres/schur_eliminator.h
index 877621b..e0c7fda 100644
--- a/internal/ceres/schur_eliminator.h
+++ b/internal/ceres/schur_eliminator.h
@@ -321,9 +321,25 @@
// see the documentation of the Chunk object above.
vector<Chunk> chunks_;
+ // TODO(sameeragarwal): The following two arrays contain per-thread
+ // storage. They should be refactored into a per thread struct.
+
// Buffer to store the products of the y and z blocks generated
- // during the elimination phase.
+ // during the elimination phase. buffer_ is of size num_threads *
+ // buffer_size_. Each thread accesses the chunk
+ //
+ // [thread_id * buffer_size_ , (thread_id + 1) * buffer_size_]
+ //
scoped_array<double> buffer_;
+
+ // Buffer to store per thread matrix matrix products used by
+ // ChunkOuterProduct. Like buffer_ it is of size num_threads *
+ // buffer_size_. Each thread accesses the chunk
+ //
+ // [thread_id * buffer_size_ , (thread_id + 1) * buffer_size_]
+ //
+ scoped_array<double> chunk_outer_product_buffer_;
+
int buffer_size_;
int num_threads_;
int uneliminated_row_begins_;
