PAPI power10 event list presets #252
Conversation
|
@dbarry9 : Requesting you to please review this. |
|
We need to verify the events you added using our benchmark suite. I will look into it this week. |
|
@adanalis - Gentle reminder |
src/papi_events.csv
Outdated
| PRESET,PAPI_L1_DCR,NOT_DERIVED,PM_LD_HIT_L1 | ||
| PRESET,PAPI_L1_DCA,DERIVED_ADD,PM_LD_REF_L1,PM_ST_CMPL | ||
| PRESET,PAPI_L2_DCM,DERIVED_ADD,PM_DATA_FROM_L2MISS,PM_L2_ST_MISS | ||
| PRESET,PAPI_L2_LDM,NOT_DERIVED,PM_L2_LD_MISS |
There was a problem hiding this comment.
PRESET,PAPI_L2_LDM,DERIVED_POSTFIX,N0|2|*|,PM_L2_LD_MISS
There was a problem hiding this comment.
Accepted the change:
PM_L2_LD_MISS event count should be multiplied by 2
src/papi_events.csv
Outdated
| PRESET,PAPI_L1_DCA,DERIVED_ADD,PM_LD_REF_L1,PM_ST_CMPL | ||
| PRESET,PAPI_L2_DCM,DERIVED_ADD,PM_DATA_FROM_L2MISS,PM_L2_ST_MISS | ||
| PRESET,PAPI_L2_LDM,NOT_DERIVED,PM_L2_LD_MISS | ||
| PRESET,PAPI_L2_STM,NOT_DERIVED,PM_L2_ST_MISS |
There was a problem hiding this comment.
PRESET,PAPI_L2_STM,DERIVED_POSTFIX,N0|2|*|,PM_L2_ST_MISS
There was a problem hiding this comment.
Accepted the change:
PM_L2_ST_MISS event count should be multiplied by 2
src/papi_events.csv
Outdated
| PRESET,PAPI_L2_LDM,NOT_DERIVED,PM_L2_LD_MISS | ||
| PRESET,PAPI_L2_STM,NOT_DERIVED,PM_L2_ST_MISS | ||
| PRESET,PAPI_L2_DCR,NOT_DERIVED,PM_DATA_FROM_L2 | ||
| PRESET,PAPI_L2_DCW,NOT_DERIVED,PM_L2_ST_HIT |
There was a problem hiding this comment.
PRESET,PAPI_L2_DCW,DERIVED_POSTFIX,N0|2|*|,PM_L2_ST_ATTEMPT
There was a problem hiding this comment.
We haven't modified the Power9 behavior; it was also mapped to PM_L2_ST_HIT in Power9. By default, all stores are written to L2, so for PM_L2_ST_ATTEMPT, our counter will account for both updates and reloads to L2. PM_L2_ST_HIT will count only updates to L2. So, what do you expect from PAPI_L2_DCW?
There was a problem hiding this comment.
As @jeevitha30 mentioned above, we would like to go with PM_L2_ST_HIT because we expect that reloads should not be accounted.
However, we have multiplied the event result by 2 here also.
PRESET,PAPI_L2_DCW,DERIVED_POSTFIX,N0|2|*|,PM_L2_ST_HIT
There was a problem hiding this comment.
Yes, the suggested event is good if multiplied by two:
PRESET,PAPI_L2_DCW,DERIVED_POSTFIX,N0|2|*|,PM_L2_ST_HIT
| PRESET,PAPI_L3_DCR,NOT_DERIVED,PM_DATA_FROM_L3 | ||
| PRESET,PAPI_L3_DCM,NOT_DERIVED,PM_DATA_FROM_L3MISS | ||
| PRESET,PAPI_L3_LDM,NOT_DERIVED,PM_DATA_FROM_L3MISS | ||
| PRESET,PAPI_L1_ICH,NOT_DERIVED,PM_INST_FROM_L1 |
There was a problem hiding this comment.
The PM_INST_FROM_L1 native event seems to overcount L1I hits.
There was a problem hiding this comment.
PM_INST_FROM_L1 will count all L1I hits that's the assumption.
| PRESET,PAPI_TOT_IIS,NOT_DERIVED,PM_INST_DISP | ||
| PRESET,PAPI_TOT_INS,NOT_DERIVED,PM_INST_CMPL | ||
| PRESET,PAPI_INT_INS,NOT_DERIVED,PM_FXU_ISSUE | ||
| PRESET,PAPI_FP_OPS,NOT_DERIVED,PM_FLOP_CMPL |
There was a problem hiding this comment.
PM_FLOP_CMPL counts instructions, not operations (FL_OPS).
There was a problem hiding this comment.
Yes, PM_FLOP_CMPL counts only instructions, not operations. Currently, we don't have an event to map operations directly. This behavior is consistent with Power9.
| PRESET,PAPI_INT_INS,NOT_DERIVED,PM_FXU_ISSUE | ||
| PRESET,PAPI_FP_OPS,NOT_DERIVED,PM_FLOP_CMPL | ||
| PRESET,PAPI_FP_INS,NOT_DERIVED,PM_FLOP_CMPL | ||
| PRESET,PAPI_DP_OPS,NOT_DERIVED,PM_DPP_FLOP_CMPL |
There was a problem hiding this comment.
PM_DPP_FLOP_CMPL counts instructions, not operations (DP_OPS)
There was a problem hiding this comment.
Yes, PM_DPP_FLOP_CMPL counts only instructions, not operations. Currently, we don't have an event to map operations directly. This behavior is consistent with Power9.
| PRESET,PAPI_FP_OPS,NOT_DERIVED,PM_FLOP_CMPL | ||
| PRESET,PAPI_FP_INS,NOT_DERIVED,PM_FLOP_CMPL | ||
| PRESET,PAPI_DP_OPS,NOT_DERIVED,PM_DPP_FLOP_CMPL | ||
| PRESET,PAPI_SP_OPS,NOT_DERIVED,PM_SP_FLOP_CMPL |
There was a problem hiding this comment.
PM_SP_FLOP_CMPL counts instructions, not operations (SP_OPS)
There was a problem hiding this comment.
Yes, PM_SP_FLOP_CMPL counts only instructions, not operations. Currently, we don't have an event to map operations directly. This behavior is consistent with Power9.
| PRESET,PAPI_DP_OPS,NOT_DERIVED,PM_DPP_FLOP_CMPL | ||
| PRESET,PAPI_SP_OPS,NOT_DERIVED,PM_SP_FLOP_CMPL | ||
| PRESET,PAPI_TOT_CYC,NOT_DERIVED,PM_RUN_CYC | ||
| PRESET,PAPI_HW_INT,NOT_DERIVED,PM_EXT_INT |
There was a problem hiding this comment.
We are currently unable to verify this event.
There was a problem hiding this comment.
@sachinmonga64, after running the Counter Analysis Toolkit, we detected some discrepancies, which we report in individual comments below. Can you please apply these changes to your PR?
There was a problem hiding this comment.
@sachinmonga64, can you please modify the entries to reflect the changes you accepted and then rebase the PR?
There was a problem hiding this comment.
Thanks @adanalis
Do I need to create a separate pull request after the updated patch ?
Let me know if anything else is expected.
There was a problem hiding this comment.
@sachinmonga64, if you commit and push the changed files to your fork, it should automatically update the PR. Also, don't forget to rebase so that there are no conflicts. If all this is too much of a hasle, I can create an up-to-date PR; I just don't want to take credit away from you.
There was a problem hiding this comment.
Thanks @adanalis
I've pushed the changes once again. Let me know if anything is missing.
Added presets for IBM Power 10 Signed-off-by: Jeevitha Palanisamy <jeevitha@linux.ibm.com>
sachinmonga64
force-pushed
the
power10
branch
from
d8dadbc
to
f08df81
Compare
Added presets for IBM Power 10
Added presets for IBM Power 10 in .csv file
Author Checklist
Why this PR exists. Reference all relevant information, including background, issues, test failures, etc
Commits are self contained and only do one thing
Commits have a header of the form:
module: short descriptionCommits have a body (whenever relevant) containing a detailed description of the addressed problem and its solution
The PR needs to pass all the tests