Language Assumptions Sample Clauses

Language Assumptions. Throughout the paper, we use Xxxxxxx to illustrate the functional programming style, but disregard laziness issues and use lists in- stead of performance-oriented special types like vectors or arrays for the sake of clarity. When discussing the imperative programming model, we use Fortran77 uniformly, because: (i) it accurately illustrates the orig- inal C code of the pricing algorithm, and, if anything, (ii) it elimi- nates the maybe-aliasing issue, which is a major hindrance to au- tomatic parallelization. Furthermore, (iii) a vast amount of work in autoparallelization targets Fortran77. As a fourth point, Fortran77 code resembles the GPU API OpenCL which we use, in that it sup- ports neither recurrence nor dynamic allocation (static arrays only). Another aspect to be taken into account when discussing op- timizations is data locality and thread grouping on GPUs. A GPU operates in thread blocks, and threads are grouped to SIMD groups (so-called warps) executed on one SIMD unit comprising multiple cores. To simplify our argument, we consider that each SIMD unit comprises 32 hardware cores. Technically this is not correct, as a warp resides on only 8 cores, which execute four-cycle instructions and need four threads to amortize the cost, but the analogy is valid for the points we are making. A block of size B yields B/32 hard- ware threads per core, which we call “virtual cores”.