WCCV: improving the vectorization of IF-statements with warp-coherent conditions

When vectorizing programs for modern processors with SIMD extensions, IF-statements pose a challenge: existing vectorization approaches often introduce redundant computations or they resort to inefficient masked instructions.

In this paper, we introduce a new notion of warp-coherence for conditions that exhibit coherent run-time behavior on different lanes of a vector register. We demonstrate that warp-coherent conditions appear frequently in practice. We present Warp-Coherent Condition Vectorization (WCCV) - an approach to detecting and optimizing IF-statements with warp-coherent conditions - to efficiently vectorize programs with IF-statements while avoiding the overhead of existing methods. WCCV detects warp-coherent conditions via the affine analysis of conditional boolean expressions and branch predication of IF-statements; the runtime code generated by WCCV avoids redundant computations and masked instructions. We employ auto-tuning to find the optimal benefit-overhead ratio for WCCV. We implement WCCV on top of Region Vectorizer (RV) - an LLVM-based vectorizing compiler, and we conduct experiments on the Rodinia benchmark suite, achieving a mean speedup of 1.14× over the original vectorized and optimized code, and speedup between 0.98× and 7.02× over the scalar code on Skylake with AVX512.