Learned parameters from the paper

[andreas-abel]

Are the learned parameters that were used for Table IV in the paper available in this repository (or elsewhere)?

[alexrenda]

Hey Andreas,

I think I may have overwritten the repeated trials, but here are DiffTune's learned parameters for one trial for each microarchitecture: https://www.dropbox.com/s/vskm0umpha822sp/hsw?dl=0, https://www.dropbox.com/s/d41qc0606h56yca/ivb?dl=0, https://www.dropbox.com/s/gvyyhb95j9itddp/skl?dl=0, https://www.dropbox.com/s/4ma1onalwpypl0u/amd?dl=0. Note that they all use the Haswell resource names despite being trained for different microarchitectures, as our implementation of this sentence in the paper:

Following the default value in llvm-mca for Haswell, we assume that there are 10 execution ports available for dispatch for all microarchitectures.

[andreas-abel]

Hi Alex,

thanks! How do I tell llvm-mca to use these parameters instead of the default ones?

[alexrenda]

If you compile the parameterized version of llvm-mca with https://github.com/ithemal/DiffTune/blob/master/llvm-mca-parametric/build.sh, then you should be able to use these parameters with:

llvm-mca -parameters PARAMETERS_FILE_NAME -mtriple=x86_64-unknown-unknown -march=x86-64 -mcpu=hsw (as in https://github.com/ithemal/DiffTune/blob/master/difftune/runner.py#L485. Also note mcpu=hsw is necessary for all architectures for llvm-mca to correctly read parameters from the Haswell resource names)

[andreas-abel]

With mcpu=hsw I get 'hsw' is not a recognized processor for this target (ignoring processor). Should this be -mcpu=haswell?

With echo "nop" | build/bin/llvm-mca -parameters hsw -mtriple=x86_64-unknown-unknown -march=x86-64 -mcpu=haswell I get

Iterations:        100
Instructions:      100
Total Cycles:      27
Total uOps:        100

Dispatch Width:    4
uOps Per Cycle:    3.70
IPC:               3.70
Block RThroughput: 0.3
llvm-mca: DiffTune/llvm-mca-parametric/llvm/lib/MC/MCSchedule.cpp:118: static double llvm::MCSchedModel::getReciprocalThroughput(const llvm::MCSubtargetInfo&, const llvm::MCSchedClassDesc&, unsigned int): Assertion `false' failed.
Stack dump:
0.	Program arguments: build/bin/llvm-mca -parameters hsw -mtriple=x86_64-unknown-unknown -march=x86-64 -mcpu=haswell 
build/bin/llvm-mca(+0xa48f55)[0x561cd72d2f55]
build/bin/llvm-mca(+0xa48fe8)[0x561cd72d2fe8]
build/bin/llvm-mca(+0xa46bcc)[0x561cd72d0bcc]
build/bin/llvm-mca(+0xa48977)[0x561cd72d2977]
/lib/x86_64-linux-gnu/libpthread.so.0(+0x12980)[0x7f2171690980]
/lib/x86_64-linux-gnu/libc.so.6(gsignal+0xc7)[0x7f2170545fb7]
/lib/x86_64-linux-gnu/libc.so.6(abort+0x141)[0x7f2170547921]
/lib/x86_64-linux-gnu/libc.so.6(+0x3048a)[0x7f217053748a]
/lib/x86_64-linux-gnu/libc.so.6(+0x30502)[0x7f2170537502]
build/bin/llvm-mca(+0x897e1a)[0x561cd7121e1a]
build/bin/llvm-mca(+0x3ca550)[0x561cd6c54550]
build/bin/llvm-mca(+0x3c7d48)[0x561cd6c51d48]
build/bin/llvm-mca(+0x3a5700)[0x561cd6c2f700]
/lib/x86_64-linux-gnu/libc.so.6(__libc_start_main+0xe7)[0x7f2170528bf7]
build/bin/llvm-mca(+0x3a31aa)[0x561cd6c2d1aa]
Aborted (core dumped)

So it outputs the throughput but crashes afterwards. Is this the expected behavior, or is something wrong?

[alexrenda]

Sorry, yes that was a typo -- should be --mcpu=haswell.

If you add --all-views=0 --summary-view do you still get the error? If not, then it's probably fine — llvm-mca calculates some extra summary statistics (e.g., pipeline utilization) that aren't actually necessary for getting throughput numbers.

[andreas-abel]

With --all-views=0 --summary-view I don't get the error.

[andreas-abel]

With these parameters, I get values for Table V that are in several cases quite different from those reported in the paper:

Are the values in Table V also the means over different runs of Difftune (like in Table IV), and would this explain the differences?

[alexrenda]

Unfortunately I don’t remember for sure and it seems like it’s not explicit specified in the paper, but it’s very likely that those are from multiple trials (they’d be from the same raw data as Table IV)