- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
I found these events on v3. And test xhpl_hybrid_intel64 (parallel_studio_xe_2015.tgz).
"AVX_INSTS_LOADS", "AVX_INSTS_STORES", "AVX_INSTS_CALC", "AVX_INSTS_ALL",
Link Copied
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
You probably have these events from the LIKWID tool suite. Have you read the related wiki page https://github.com/RRZE-HPC/likwid/wiki/FlopsHaswell ?
For pure AVX codes without split cache line loads the AVX_INSTS_* events are accurate but as soon as you mix scalar, SSE and AVX they are probably off. The Wiki page mentions the insertf128 instruction but there might be more.
To refer to your question: The AVX_INSTS_LOADS should count only AVX loads, AVX_INSTS_STORES should count only AVX stores, the AVX_INSTS_CALC should count only AVX instruction (it counts more) and AVX_INSTS_ALL counts all AVX instructions.
- Subscribe to RSS Feed
- Mark Topic as New
- Mark Topic as Read
- Float this Topic for Current User
- Bookmark
- Subscribe
- Printer Friendly Page