From the documents, I read that the attribution of cycles to instructions or source code lines can sometimes be one off. A line too early, or too late.
In my run, I see very inaccurate mappings. Both between source code and assembly code. And also between cycles spent and source code. See the attached screen shot. My source code is mostly intrinsics, so I have a pretty good understanding to what it should map. Yet, a simple intrinsic sometimes maps to a whole page of assembly? How could this be?
In the example from the screenshot: a single _mm256_load_ps() gets mapped to 29 assembly instructions?
And the cycles do not evenly distribute over my source code lines: almost all lines get no cycles attributed, only a sparse few get all the cycles?
| Attachment | Size |
|---|---|
Download mapping.png | 438.23 KB |