Optimizing Marshalling by Run-Time Program Generation

Optimization by runtime specialization for sparse matrix-vector multiplication

“…Work in this area specifically addressing highperformance for realistic applications includes work on marshalling [7,14] and on code-optimizing transformations [15]. With runtime specialization, the focus moves toward the efficiency of specialization itself [23,29].…”

“…Specifically, out of our 115 (23×5) trials, the best specializers were: Stencil (11 times), GenOSKI (29), Unfolding (27), CSRbyNZ (20), CSR (13), MKL (2), BiCSB (6), CSX (7) 2 . The main contribution of this paper is a systematic comparison of a number of methods for performing sparse matrix-dense vector multiplication, including methods that are specialized to a particular matrix.…”

Efficient automated marshaling of C++ data structures for MPI applications

“…Many of the proposed approaches of optimizing Java serialization had the aim of making the language more amenable for HPC. The proposals for faster serialization eliminated the overheads of reflection by providing custom code for individual objects [27], using native code [21,22], and finally using runtime code generation to specialize serialization functionality [1]. Because C++ does not have built-in reflective capabilities, none of these approaches are applicable to us.…”

“…Therefore, the field offsets obtained from a generated isomorphic class are guaranteed to be the same as the corresponding field offsets in the original class. 1 The obtained offsets are then inserted into the generated marshaling code to access an object's non-public fields. A slightly simplified example of accessing the protected field force x in class Mass is shown in Figure 3.…”

Staging static analyses for program generation

“…Staging requires that dataflow functions be represented "adequately." (2) We give representations for several dataflow problems, and for staged type checking. (3) We provide experimental evidence of speed-ups from staging.…”