diff options
Diffstat (limited to 'src/Core/performance_CPU/README')
| -rw-r--r-- | src/Core/performance_CPU/README | 25 |
1 files changed, 25 insertions, 0 deletions
diff --git a/src/Core/performance_CPU/README b/src/Core/performance_CPU/README new file mode 100644 index 0000000..948c544 --- /dev/null +++ b/src/Core/performance_CPU/README @@ -0,0 +1,25 @@ +TNV_core.c.v4.stdver: Fully equivalent (signle-threads). There is two potential breakage points. + - If OpenMP is enabled, the acces to div_upd will not be serialized and results will breaj + - The results will slightly differ due to different order of summation if loop summing resprimal/resdual organized in a logical way +TNV_core.c.v15: Multi-threads. Works correctly only in the single-threaded mode (if TNV_NEW_STYLE is disabled). In multi-threaded there results slightly differ due to changed order of operation + - TNV_NEW_STYLE slightly disturbs results in both single- and multi-threaded modes + - Resprimal/resdual are summed in groups (not sequentially) if multiple threads. But his actually should improve precision. Use TNV_CHECK_RES to check conformance + - Afterwards, in multi-threaded moded there is a still minor descripancy which first occurs in resprimal (after a few iterations). This is because of changed order of operations while computing + div_upd (only on the first lines of each new sub-block). Normally, we first compute the vertical and, then, add horizontal. On the border rows, instead we first add horizontals... + To check, div/div_upd changed to double. There is no difference then. +TNV_core.c.v17: Computationaly comptabile with v15. + - Padding actually harms performance + - Intel compiler gives about 10% speed-up +TNV_core.c.v18: Blocking helps to boost performance further but only with Intel Compiler. Gcc/Clang is slightly slower here. + - Padding here doesn't harm performance, but is not helpful either + - Difference between icc and gcc is probably due to auto-vectorization. + - Results slightly changed due to different order of operations +TNV_core.c.v19: Eliminate conditionals in the inner loops to help gcc-autovectorisation + - Last version implementing full algrotithm with backtrack in the middle of iterations. + - Again results slightly diverge from v18 due to different order of operations +TNV_core.c.v27: v18 with backtracking only on first iterations (otherwise warning reported) +TNV_core.c.v32: v19 with backtracking only on first iterations (otherwise warning reported) + + +Repo: + - Padding seems to have effect on the newer AVX2 systems. Re-enabled. |