Optimisation Tracks Sample Clauses

Optimisation Tracks. Attempt to utilise full potential of x86 instruction set Early in the hotspot analysis, the compiler was only capable of vectorizing small chunks of the code. About 99% of the executable and with it the computationally extensive regions were composed out of serial instructions. As the computationally most expensive region, the focus is set on the blockwiseBulkCollideAndStream() function in blockLattice3D.hh. It is responsible of advancing the fluid according to the Lattice Boltzmann Method. The rather complex nature of the 6 nested for loops made it necessary to traverse the call stack down to where the actual computations are performed. In the following, the subroutines addNaiveForce() and addGuoForce(), which are implemented in externalForceTemplateeD.h are investigated. Within these functions, the force with respect to each neighbouring cell is computed. The computations differ slightly, depending for instance on the distance of the centre points of the respective cells. To enable vectorisation or fused multiply-add (FMA) operations, we need to unify these computations and even more importantly, the data on the random-access memory (RAM) must be aligned. In principle, we archived this by introducing an additional buffer which stored precomputed results which had to computed individually and once this was done, vector operations were performed on the buffer. static T A[27] = {…} attribute ((aligned(64))); static T B[27] = {…} attribute ((aligned(64))); for(unsigned int i = 0; i < 9; ++i){ #ifdef _USE_BLAS C[i] = blas_ddot(3, &(A[i*3]), 1, &(B[i*3]), 1); #else C[i] = A[I*3] * B[i*3] + A[i*3 + 1] * B[i*3 +1] + A[i*3+2] * B[i*3+2] ; #endif } //some more computations to get to the final result for array f[] Throughout multiple test, the structure was altered to obtain more performance. For instance, instated of statically allocated buffers, a memory bool was used. Also, as shown in the sample code, XXXX level one routines were put in place. However, this technique did not lead to the desired decrease in execution time. Unfortunately, the way the 6 nested loops operate, results in rather short loops. Thus, the vector length is very limited. The benefits of using SIMD instructions were at best about even with the effort to prepare and align the data at each iteration. Additionally, the BLAS routines couldn’t show their full potential, due to the short vector length. Eventually, the idea to vectorise computations within a loop iteration had to be abandoned. How...