Not all performance groups can be tested for accuracy. We don't have a test application for each performance group. Here only the groups are listed that can be verified. Each group is followed by the low-level benchmarks that are performed for comparison.<br>
Not all performance groups can be tested for accuracy. We don't have a test application for each performance group. Here only the groups are listed that can be verified. Each group is followed by the low-level benchmarks that are performed for comparison.
| Group | Tests |
| Group | Tests |
|-------|-------|
|-------|-------|
| L2 | load, triad, copy, stream, store |
| L2 | load, triad, copy, stream, store |
| L3 | load, triad, copy, stream, store |
| L3 | load, triad, copy, stream, store |
| DATA | triad, copy, stream, store |
| BRANCH | load, triad, copy, stream, store |
For each varification group, the tests are performed twice. Once in a plain manner without measuring but calculating and scaling the resulting values and once through an instumented code with LIKWID.<br>
For each varification group, the tests are performed twice. Once in a plain manner without measuring but calculating and scaling the resulting values and once through an instumented code with LIKWID.<br>
## Information about the pictures
For memory bandwidth accuracy tests, the result of likwid-bench needs to be scaled to represent the real values. The reason is traffic caused by write-allocates (the store operation on a cache line that is not present at the L1 cache must be loaded first before the store operation can complete). In these cases, the figures contain three curves, the original output of likwid-bench, the scaled version and the derived metrics of likwid-perfctr measurements.
### Verification of Group L2
### Verification of Group L2
#### Verification of Group L2 with Test load
#### Verification of Group L2 with Test load
...
@@ -25,19 +31,20 @@ For each varification group, the tests are performed twice. Once in a plain mann
...
@@ -25,19 +31,20 @@ For each varification group, the tests are performed twice. Once in a plain mann
Benchmark performed with data sizes: 12kB, 1MB, 4MB, 1GB.
Benchmark performed with data sizes: 12kB, 1MB, 4MB, 1GB.
Each data size is tested 5 times, hence the first 5 entries on the x-axis correspond to the 5 runs for the first data size of 12kB and so on.<br>
Each data size is tested 5 times, hence the first 5 entries on the x-axis correspond to the 5 runs for the first data size of 12kB and so on.