[PATCH] perf vendor events power10: Update JSON/events
kajoljain
kjain at linux.ibm.com
Tue Jul 30 23:53:53 AEST 2024
On 7/23/24 12:35, Disha Goel wrote:
> On 23/07/24 10:51 am, Kajol Jain wrote:
>
>> Update JSON/events for power10 platform with additional events.
>> Also move PM_VECTOR_LD_CMPL event from others.json to
>> frontend.json file.
>>
>> Signed-off-by: Kajol Jain <kjain at linux.ibm.com>
>
> I have tested the patch on power10 machine. Looks good to me.
>
Hi Disha,
Thanks for testing this patch.
Thanks,
Kajol Jain
> Tested-by: Disha Goel <disgoel at linux.ibm.com>
>
>> ---
>> .../arch/powerpc/power10/frontend.json | 5 +
>> .../arch/powerpc/power10/others.json | 100 +++++++++++++++++-
>> 2 files changed, 100 insertions(+), 5 deletions(-)
>>
>> diff --git a/tools/perf/pmu-events/arch/powerpc/power10/frontend.json
>> b/tools/perf/pmu-events/arch/powerpc/power10/frontend.json
>> index 5977f5e64212..53660c279286 100644
>> --- a/tools/perf/pmu-events/arch/powerpc/power10/frontend.json
>> +++ b/tools/perf/pmu-events/arch/powerpc/power10/frontend.json
>> @@ -74,6 +74,11 @@
>> "EventName": "PM_ISSUE_KILL",
>> "BriefDescription": "Cycles in which an instruction or group of
>> instructions were cancelled after being issued. This event increments
>> once per occurrence, regardless of how many instructions are included
>> in the issue group."
>> },
>> + {
>> + "EventCode": "0x44054",
>> + "EventName": "PM_VECTOR_LD_CMPL",
>> + "BriefDescription": "Vector load instruction completed."
>> + },
>> {
>> "EventCode": "0x44056",
>> "EventName": "PM_VECTOR_ST_CMPL",
>> diff --git a/tools/perf/pmu-events/arch/powerpc/power10/others.json
>> b/tools/perf/pmu-events/arch/powerpc/power10/others.json
>> index fcf8a8ebe7bd..53ca610152fa 100644
>> --- a/tools/perf/pmu-events/arch/powerpc/power10/others.json
>> +++ b/tools/perf/pmu-events/arch/powerpc/power10/others.json
>> @@ -94,11 +94,6 @@
>> "EventName": "PM_L1_ICACHE_RELOADED_ALL",
>> "BriefDescription": "Counts all instruction cache reloads
>> includes demand, prefetch, prefetch turned into demand and demand
>> turned into prefetch."
>> },
>> - {
>> - "EventCode": "0x44054",
>> - "EventName": "PM_VECTOR_LD_CMPL",
>> - "BriefDescription": "Vector load instruction completed."
>> - },
>> {
>> "EventCode": "0x4D05E",
>> "EventName": "PM_BR_CMPL",
>> @@ -108,5 +103,100 @@
>> "EventCode": "0x400F0",
>> "EventName": "PM_LD_DEMAND_MISS_L1_FIN",
>> "BriefDescription": "Load missed L1, counted at finish time."
>> + },
>> + {
>> + "EventCode": "0x00000038BC",
>> + "EventName": "PM_ISYNC_CMPL",
>> + "BriefDescription": "Isync completion count per thread."
>> + },
>> + {
>> + "EventCode": "0x000000C088",
>> + "EventName": "PM_LD0_32B_FIN",
>> + "BriefDescription": "256-bit load finished in the LD0 load
>> execution unit."
>> + },
>> + {
>> + "EventCode": "0x000000C888",
>> + "EventName": "PM_LD1_32B_FIN",
>> + "BriefDescription": "256-bit load finished in the LD1 load
>> execution unit."
>> + },
>> + {
>> + "EventCode": "0x000000C090",
>> + "EventName": "PM_LD0_UNALIGNED_FIN",
>> + "BriefDescription": "Load instructions in LD0 port that are
>> either unaligned, or treated as unaligned and require an additional
>> recycle through the pipeline using the load gather buffer. This
>> typically adds about 10 cycles to the latency of the instruction. This
>> includes loads that cross the 128 byte boundary, octword loads that
>> are not aligned, and a special forward progress case of a load that
>> does not hit in the L1 and crosses the 32 byte boundary and is
>> launched NTC. Counted at finish time."
>> + },
>> + {
>> + "EventCode": "0x000000C890",
>> + "EventName": "PM_LD1_UNALIGNED_FIN",
>> + "BriefDescription": "Load instructions in LD1 port that are
>> either unaligned, or treated as unaligned and require an additional
>> recycle through the pipeline using the load gather buffer. This
>> typically adds about 10 cycles to the latency of the instruction. This
>> includes loads that cross the 128 byte boundary, octword loads that
>> are not aligned, and a special forward progress case of a load that
>> does not hit in the L1 and crosses the 32 byte boundary and is
>> launched NTC. Counted at finish time."
>> + },
>> + {
>> + "EventCode": "0x000000C0A4",
>> + "EventName": "PM_ST0_UNALIGNED_FIN",
>> + "BriefDescription": "Store instructions in ST0 port that are
>> either unaligned, or treated as unaligned and require an additional
>> recycle through the pipeline. This typically adds about 10 cycles to
>> the latency of the instruction. This only includes stores that cross
>> the 128 byte boundary. Counted at finish time."
>> + },
>> + {
>> + "EventCode": "0x000000C8A4",
>> + "EventName": "PM_ST1_UNALIGNED_FIN",
>> + "BriefDescription": "Store instructions in ST1 port that are
>> either unaligned, or treated as unaligned and require an additional
>> recycle through the pipeline. This typically adds about 10 cycles to
>> the latency of the instruction. This only includes stores that cross
>> the 128 byte boundary. Counted at finish time."
>> + },
>> + {
>> + "EventCode": "0x000000C8B8",
>> + "EventName": "PM_STCX_SUCCESS_CMPL",
>> + "BriefDescription": "STCX instructions that completed
>> successfully. Specifically, counts only when a pass status is returned
>> from the nest."
>> + },
>> + {
>> + "EventCode": "0x000000D0B4",
>> + "EventName": "PM_DC_PREF_STRIDED_CONF",
>> + "BriefDescription": "A demand load referenced a line in an active
>> strided prefetch stream. The stream could have been allocated through
>> the hardware prefetch mechanism or through software."
>> + },
>> + {
>> + "EventCode": "0x000000F880",
>> + "EventName": "PM_SNOOP_TLBIE_CYC",
>> + "BriefDescription": "Cycles in which TLBIE snoops are executed in
>> the LSU."
>> + },
>> + {
>> + "EventCode": "0x000000F084",
>> + "EventName": "PM_SNOOP_TLBIE_CACHE_WALK_CYC",
>> + "BriefDescription": "TLBIE snoop cycles in which the data cache
>> is being walked."
>> + },
>> + {
>> + "EventCode": "0x000000F884",
>> + "EventName": "PM_SNOOP_TLBIE_WAIT_ST_CYC",
>> + "BriefDescription": "TLBIE snoop cycles in which older stores are
>> still draining."
>> + },
>> + {
>> + "EventCode": "0x000000F088",
>> + "EventName": "PM_SNOOP_TLBIE_WAIT_LD_CYC",
>> + "BriefDescription": "TLBIE snoop cycles in which older loads are
>> still draining."
>> + },
>> + {
>> + "EventCode": "0x000000F08C",
>> + "EventName": "PM_SNOOP_TLBIE_WAIT_MMU_CYC",
>> + "BriefDescription": "TLBIE snoop cycles in which the Load-Store
>> unit is waiting for the MMU to finish invalidation."
>> + },
>> + {
>> + "EventCode": "0x0000004884",
>> + "EventName": "PM_NO_FETCH_IBUF_FULL_CYC",
>> + "BriefDescription": "Cycles in which no instructions are fetched
>> because there is no room in the instruction buffers."
>> + },
>> + {
>> + "EventCode": "0x00000048B4",
>> + "EventName": "PM_BR_TKN_UNCOND_FIN",
>> + "BriefDescription": "An unconditional branch finished. All
>> unconditional branches are taken."
>> + },
>> + {
>> + "EventCode": "0x0B0000016080",
>> + "EventName": "PM_L2_TLBIE_SLBIE_START",
>> + "BriefDescription": "NCU Master received a TLBIE/SLBIEG/SLBIAG
>> operation from the core. Event count should be multiplied by 2 since
>> the data is coming from a 2:1 clock domain and the data is time sliced
>> across all 4 threads."
>> + },
>> + {
>> + "EventCode": "0x0B0000016880",
>> + "EventName": "PM_L2_TLBIE_SLBIE_DELAY",
>> + "BriefDescription": "Cycles when a TLBIE/SLBIEG/SLBIAG command
>> was held in a hottemp condition by the NCU Master. Multiply this count
>> by 1000 to obtain the total number of cycles. This can be divided by
>> PM_L2_TLBIE_SLBIE_SENT to obtain the average time a
>> TLBIE/SLBIEG/SLBIAG command was held. Event count should be multiplied
>> by 2 since the data is coming from a 2:1 clock domain and the data is
>> time sliced across all 4 threads."
>> + },
>> + {
>> + "EventCode": "0x0B0000026880",
>> + "EventName": "PM_L2_SNP_TLBIE_SLBIE_DELAY",
>> + "BriefDescription": "Cycles when a TLBIE/SLBIEG/SLBIAG that
>> targets this thread's LPAR was in flight while in a hottemp condition.
>> Multiply this count by 1000 to obtain the total number of cycles. This
>> can be divided by PM_L2_SNP_TLBIE_SLBIE_START to obtain the overall
>> efficiency. Note: ’inflight’ means SnpTLB has been sent to core(ie
>> doesn’t include when SnpTLB is in NCU waiting to be launched serially
>> behind different SnpTLB). The NCU Snooper gets in a ’hottemp’ delay
>> window when it detects it is above its TLBIE/SLBIE threshold for
>> process SnpTLBIE/SLBIE with this core. Event count should be
>> multiplied by 2 since the data is coming from a 2:1 clock domain and
>> the data is time sliced across all 4 threads."
>> }
>> ]
More information about the Linuxppc-dev
mailing list