examples

MARTA - Examples

In order to illustrate the capabilities of MARTA, we have developed a set of cases of study which answer different research questions. For each case, we provide the full experimental setup needed to reproduce these experiments.

Empirical throughput of FMA instructions

RQ.- How many independent FMA instructions can be executed in just one cycle, assuming hot cache, regardless of data type, vector size and ISA?

STREAM: influence of access pattern on memory bandwidth

RQ.- Consider a vector operation of the form: $c(f(i)) = a(g(i)) ∗ b(h(i))$ How does the memory bandwidth vary with the access patterns, influenced by the access functions f, g, and h?

Micro-benchmarking gather and scatter instructions

RQ.- Can we build an analytical model of the performance of gather/scatter based on number (and set it maps to?) of cache lines touched (hot cache)?

Loop interchange/permutation cost model

RQ.- How to relate source code metrics (e.g., loop depth, nesting, trip count, type of array access: affine vs. non-affine, reuse direction/distance, etc.) with the application of loop interchange in GCC, when using the flag -floop-interchange? In other words, can we reverse-engineer the cost model of GCC to apply interchange?

Auto-vectorization of reductions

RQ.- Does -ffast-math flag always enable vectorization of reductions of floats in GCC, even for scalars?

---

We also provide other examples of the tool and the integration with other tools.

Sparse matrices vectorization with MACVETH

MACVETH is a novel compiler for auto-vectorizing irregular codes. MARTA supports this compiler.

Multithread

Just a dummy benchmark for testing the multithreading support of the framework.