add avx256 and 512 metrics. fix non-contiguous device enumeration.

_version.txt

@@ -1 +1 @@
-1.3.9
+1.3.10

events/icx.txt

@@ -30,7 +30,6 @@ cpu-cycles,
 ref-cycles,
 instructions;
 
-# OCR group 1 (ICX PMU supports a maximum of two OCR counters per group)
 cpu/event=0xb7,umask=0x01,offcore_rsp=0x104000477,name='OCR.READS_TO_CORE.LOCAL_DRAM'/,
 cpu/event=0xb7,umask=0x01,offcore_rsp=0x84002380,name='OCR.HWPF_L3.L3_MISS_LOCAL'/,
 cpu/event=0x85,umask=0x0e,period=100003,name='ITLB_MISSES.WALK_COMPLETED'/,
@@ -39,7 +38,6 @@ cpu-cycles,
 ref-cycles,
 instructions;
 
-# OCR group 2 (ICX PMU supports a maximum of two OCR counters per group)
 cpu/event=0xb7,umask=0x01,offcore_rsp=0x730000477,name='OCR.READS_TO_CORE.REMOTE_DRAM'/,
 cpu/event=0xb7,umask=0x01,offcore_rsp=0x90002380,name='OCR.HWPF_L3.REMOTE'/,
 cpu/event=0x08,umask=0x04,period=100003,name='DTLB_LOAD_MISSES.WALK_COMPLETED_2M_4M'/,
@@ -69,6 +67,17 @@ cpu-cycles,
 ref-cycles,
 instructions;
 
+#TMA AVX512 related
+cpu/event=0xc7,umask=0x80,cmask=0x00,period=100003,name='FP_ARITH_INST_RETIRED.512B_PACKED_SINGLE'/,
+cpu/event=0xc7,umask=0x40,cmask=0x00,period=100003,name='FP_ARITH_INST_RETIRED.512B_PACKED_DOUBLE'/,
+cpu/event=0xc7,umask=0x20,cmask=0x00,period=100003,name='FP_ARITH_INST_RETIRED.256B_PACKED_SINGLE'/,
+cpu/event=0xc7,umask=0x10,cmask=0x00,period=100003,name='FP_ARITH_INST_RETIRED.256B_PACKED_DOUBLE'/,
+cpu/event=0xc7,umask=0x08,cmask=0x00,period=100003,name='FP_ARITH_INST_RETIRED.128B_PACKED_SINGLE'/,
+cpu/event=0xc7,umask=0x04,cmask=0x00,period=100003,name='FP_ARITH_INST_RETIRED.128B_PACKED_DOUBLE'/,
+cpu-cycles,
+ref-cycles,
+instructions;
+
 cpu/event=0x79,umask=0x04,cmask=0x01,period=2000003,name='IDQ.MITE_CYCLES_ANY'/,
 cpu/event=0x79,umask=0x04,cmask=0x05,period=2000003,name='IDQ.MITE_CYCLES_OK'/,
 cpu/event=0x79,umask=0x08,cmask=0x01,period=2000003,name='IDQ.DSB_CYCLES_ANY'/,
@@ -129,10 +138,6 @@ cpu/event=0xd3,umask=0x02,cmask=0x00,period=100007,name='MEM_LOAD_L3_MISS_RETIRE
 cpu/event=0xd3,umask=0x01,cmask=0x00,period=100007,name='MEM_LOAD_L3_MISS_RETIRED.LOCAL_DRAM'/,
 cpu/event=0xd3,umask=0x08,cmask=0x00,period=100007,name='MEM_LOAD_L3_MISS_RETIRED.REMOTE_FWD'/,
 cpu/event=0xd3,umask=0x04,cmask=0x00,period=100007,name='MEM_LOAD_L3_MISS_RETIRED.REMOTE_HITM'/,
-cpu/event=0xb1,umask=0x10,cmask=0x00,period=2000003,name='UOPS_EXECUTED.X87'/,
-cpu/event=0xb1,umask=0x01,cmask=0x00,period=2000003,name='UOPS_EXECUTED.THREAD'/,
-cpu/event=0xc7,umask=0x03,period=100003,name='FP_ARITH_INST_RETIRED.SCALAR_SINGLE:u0x03'/,
-cpu/event=0xc7,umask=0xfc,period=100003,name='FP_ARITH_INST_RETIRED.128B_PACKED_DOUBLE:u0xfc'/,
 cpu-cycles,
 ref-cycles,
 instructions;
@@ -196,4 +201,4 @@ imc/event=0x04,umask=0x30,name='UNC_M_CAS_COUNT.WR'/;
 
 #power related
 power/energy-pkg/,
-power/energy-ram/;
+power/energy-ram/;

events/metric_icx.json

@@ -118,21 +118,6 @@
         "name": "metric_% Uops delivered from legacy decode pipeline (MITE)",
         "expression": "100 * ([IDQ.MITE_UOPS] / ([IDQ.DSB_UOPS] + [IDQ.MITE_UOPS] + [IDQ.MS_UOPS] + [LSD.UOPS]) )"
     },
-    {
-        "name": "metric_core % cycles in non AVX license",
-        "expression": "(100 * [CORE_POWER.LVL0_TURBO_LICENSE]) / ([CORE_POWER.LVL0_TURBO_LICENSE] + [CORE_POWER.LVL1_TURBO_LICENSE] + [CORE_POWER.LVL2_TURBO_LICENSE])",
-        "origin": "perfspect"
-    },
-    {
-        "name": "metric_core % cycles in AVX2 license",
-        "expression": "(100 * [CORE_POWER.LVL1_TURBO_LICENSE]) / ([CORE_POWER.LVL0_TURBO_LICENSE] + [CORE_POWER.LVL1_TURBO_LICENSE] + [CORE_POWER.LVL2_TURBO_LICENSE])",
-        "origin": "perfspect"
-    },
-    {
-        "name": "metric_core % cycles in AVX-512 license",
-        "expression": "(100 * [CORE_POWER.LVL2_TURBO_LICENSE]) / ([CORE_POWER.LVL0_TURBO_LICENSE] + [CORE_POWER.LVL1_TURBO_LICENSE] + [CORE_POWER.LVL2_TURBO_LICENSE])",
-        "origin": "perfspect"
-    },
     {
         "name": "metric_core initiated local dram read bandwidth (MB/sec)",
         "expression": "(([OCR.READS_TO_CORE.LOCAL_DRAM] + [OCR.HWPF_L3.L3_MISS_LOCAL]) * 64 / 1000000) / 1"
@@ -366,16 +351,16 @@
         "expression": "100 * ( max( 0 , ( [PERF_METRICS.RETIRING] / ( [PERF_METRICS.FRONTEND_BOUND] + [PERF_METRICS.BAD_SPECULATION] + [PERF_METRICS.RETIRING] + [PERF_METRICS.BACKEND_BOUND] ) ) - ( ( ( ( ( [PERF_METRICS.RETIRING] / ( [PERF_METRICS.FRONTEND_BOUND] + [PERF_METRICS.BAD_SPECULATION] + [PERF_METRICS.RETIRING] + [PERF_METRICS.BACKEND_BOUND] ) ) * ( [TOPDOWN.SLOTS] ) ) / [UOPS_ISSUED.ANY] ) * [IDQ.MS_UOPS] / ( [TOPDOWN.SLOTS] ) ) + ( [PERF_METRICS.RETIRING] / ( [PERF_METRICS.FRONTEND_BOUND] + [PERF_METRICS.BAD_SPECULATION] + [PERF_METRICS.RETIRING] + [PERF_METRICS.BACKEND_BOUND] ) ) * ( [UOPS_DECODED.DEC0] - [UOPS_DECODED.DEC0:c1] ) / [IDQ.MITE_UOPS] ) ) )"
     },
     {
-        "name": "metric_TMA_....FP_Arith(%)",
-        "expression": "100 * ( ( ( [PERF_METRICS.RETIRING] / ( [PERF_METRICS.FRONTEND_BOUND] + [PERF_METRICS.BAD_SPECULATION] + [PERF_METRICS.RETIRING] + [PERF_METRICS.BACKEND_BOUND] ) ) * [UOPS_EXECUTED.X87] / [UOPS_EXECUTED.THREAD] ) + ( ( [FP_ARITH_INST_RETIRED.SCALAR_SINGLE:u0x03] ) / ( ( [PERF_METRICS.RETIRING] / ( [PERF_METRICS.FRONTEND_BOUND] + [PERF_METRICS.BAD_SPECULATION] + [PERF_METRICS.RETIRING] + [PERF_METRICS.BACKEND_BOUND] ) ) * ( [TOPDOWN.SLOTS] ) ) ) + ( min( ( ( [FP_ARITH_INST_RETIRED.128B_PACKED_DOUBLE:u0xfc] ) / ( ( [PERF_METRICS.RETIRING] / ( [PERF_METRICS.FRONTEND_BOUND] + [PERF_METRICS.BAD_SPECULATION] + [PERF_METRICS.RETIRING] + [PERF_METRICS.BACKEND_BOUND] ) ) * ( [TOPDOWN.SLOTS] ) ) ) , ( 1 ) ) ) )"
+        "name": "metric_TMA_........FP_Vector_128b(%)",
+        "expression": "100 * ( min( ( ( [FP_ARITH_INST_RETIRED.128B_PACKED_DOUBLE] + [FP_ARITH_INST_RETIRED.128B_PACKED_SINGLE] ) / ( ( [PERF_METRICS.RETIRING] / ( [PERF_METRICS.FRONTEND_BOUND] + [PERF_METRICS.BAD_SPECULATION] + [PERF_METRICS.RETIRING] + [PERF_METRICS.BACKEND_BOUND] ) ) * ( [TOPDOWN.SLOTS] ) ) ) , ( 1 ) ) )"
     },
     {
-        "name": "metric_TMA_......FP_Scalar(%)",
-        "expression": "100 * ( ( [FP_ARITH_INST_RETIRED.SCALAR_SINGLE:u0x03] ) / ( ( [PERF_METRICS.RETIRING] / ( [PERF_METRICS.FRONTEND_BOUND] + [PERF_METRICS.BAD_SPECULATION] + [PERF_METRICS.RETIRING] + [PERF_METRICS.BACKEND_BOUND] ) ) * ( [TOPDOWN.SLOTS] ) ) )"
+        "name": "metric_TMA_........FP_Vector_256b(%)",
+        "expression": "100 * ( min( ( ( [FP_ARITH_INST_RETIRED.256B_PACKED_DOUBLE] + [FP_ARITH_INST_RETIRED.256B_PACKED_SINGLE] ) / ( ( [PERF_METRICS.RETIRING] / ( [PERF_METRICS.FRONTEND_BOUND] + [PERF_METRICS.BAD_SPECULATION] + [PERF_METRICS.RETIRING] + [PERF_METRICS.BACKEND_BOUND] ) ) * ( [TOPDOWN.SLOTS] ) ) ) , ( 1 ) ) )"
     },
     {
-        "name": "metric_TMA_......FP_Vector(%)",
-        "expression": "100 * ( min( ( ( [FP_ARITH_INST_RETIRED.128B_PACKED_DOUBLE:u0xfc] ) / ( ( [PERF_METRICS.RETIRING] / ( [PERF_METRICS.FRONTEND_BOUND] + [PERF_METRICS.BAD_SPECULATION] + [PERF_METRICS.RETIRING] + [PERF_METRICS.BACKEND_BOUND] ) ) * ( [TOPDOWN.SLOTS] ) ) ) , ( 1 ) ) )"
+        "name": "metric_TMA_........FP_Vector_512b(%)",
+        "expression": "100 * ( min( ( ( [FP_ARITH_INST_RETIRED.512B_PACKED_DOUBLE] + [FP_ARITH_INST_RETIRED.512B_PACKED_SINGLE] ) / ( ( [PERF_METRICS.RETIRING] / ( [PERF_METRICS.FRONTEND_BOUND] + [PERF_METRICS.BAD_SPECULATION] + [PERF_METRICS.RETIRING] + [PERF_METRICS.BACKEND_BOUND] ) ) * ( [TOPDOWN.SLOTS] ) ) ) , ( 1 ) ) )"
     },
     {
         "name": "metric_TMA_..Heavy_Operations(%)",

events/metric_spr_emr.json

@@ -367,8 +367,16 @@
         "expression": "100 * ( max( 0 , ( [PERF_METRICS.RETIRING] / ( [PERF_METRICS.FRONTEND_BOUND] + [PERF_METRICS.BAD_SPECULATION] + [PERF_METRICS.RETIRING] + [PERF_METRICS.BACKEND_BOUND] ) ) - ( [PERF_METRICS.HEAVY_OPERATIONS] / ( [PERF_METRICS.FRONTEND_BOUND] + [PERF_METRICS.BAD_SPECULATION] + [PERF_METRICS.RETIRING] + [PERF_METRICS.BACKEND_BOUND] ) ) ) )"
     },
     {
-        "name": "metric_TMA_....FP_Arith(%)",
-        "expression": "100 * ( ( ( [PERF_METRICS.RETIRING] / ( [PERF_METRICS.FRONTEND_BOUND] + [PERF_METRICS.BAD_SPECULATION] + [PERF_METRICS.RETIRING] + [PERF_METRICS.BACKEND_BOUND] ) ) * [UOPS_EXECUTED.X87] / [UOPS_EXECUTED.THREAD] ) + ( ( [FP_ARITH_INST_RETIRED.SCALAR_SINGLE:u0x03] + [FP_ARITH_INST_RETIRED2.SCALAR] ) / ( ( [PERF_METRICS.RETIRING] / ( [PERF_METRICS.FRONTEND_BOUND] + [PERF_METRICS.BAD_SPECULATION] + [PERF_METRICS.RETIRING] + [PERF_METRICS.BACKEND_BOUND] ) ) * ( [TOPDOWN.SLOTS] ) ) ) + ( min( ( ( [FP_ARITH_INST_RETIRED.128B_PACKED_DOUBLE:u0x3c] + [FP_ARITH_INST_RETIRED2.VECTOR] ) / ( ( [PERF_METRICS.RETIRING] / ( [PERF_METRICS.FRONTEND_BOUND] + [PERF_METRICS.BAD_SPECULATION] + [PERF_METRICS.RETIRING] + [PERF_METRICS.BACKEND_BOUND] ) ) * ( [TOPDOWN.SLOTS] ) ) ) , ( 1 ) ) ) + ( [AMX_OPS_RETIRED.BF16:c1] / ( ( [PERF_METRICS.RETIRING] / ( [PERF_METRICS.FRONTEND_BOUND] + [PERF_METRICS.BAD_SPECULATION] + [PERF_METRICS.RETIRING] + [PERF_METRICS.BACKEND_BOUND] ) ) * ( [TOPDOWN.SLOTS] ) ) ) )"
+        "name": "metric_TMA_........FP_Vector_256b(%)",
+        "expression": "100 * ( min( ( ( [FP_ARITH_INST_RETIRED.256B_PACKED_DOUBLE] + [FP_ARITH_INST_RETIRED.256B_PACKED_SINGLE] + [FP_ARITH_INST_RETIRED2.256B_PACKED_HALF] ) / ( ( [PERF_METRICS.RETIRING] / ( [PERF_METRICS.FRONTEND_BOUND] + [PERF_METRICS.BAD_SPECULATION] + [PERF_METRICS.RETIRING] + [PERF_METRICS.BACKEND_BOUND] ) ) * ( [TOPDOWN.SLOTS] ) ) ) , ( 1 ) ) )"
+    },
+    {
+        "name": "metric_TMA_........FP_Vector_512b(%)",
+        "expression": "100 * ( min( ( ( [FP_ARITH_INST_RETIRED.512B_PACKED_DOUBLE] + [FP_ARITH_INST_RETIRED.512B_PACKED_SINGLE] + [FP_ARITH_INST_RETIRED2.512B_PACKED_HALF] ) / ( ( [PERF_METRICS.RETIRING] / ( [PERF_METRICS.FRONTEND_BOUND] + [PERF_METRICS.BAD_SPECULATION] + [PERF_METRICS.RETIRING] + [PERF_METRICS.BACKEND_BOUND] ) ) * ( [TOPDOWN.SLOTS] ) ) ) , ( 1 ) ) )"
+    },
+    {
+        "name": "metric_TMA_......Int_Vector_256b(%)",
+        "expression": "100 * ( ( [INT_VEC_RETIRED.ADD_256] + [INT_VEC_RETIRED.MUL_256] + [INT_VEC_RETIRED.VNNI_256] ) / ( ( [PERF_METRICS.RETIRING] / ( [PERF_METRICS.FRONTEND_BOUND] + [PERF_METRICS.BAD_SPECULATION] + [PERF_METRICS.RETIRING] + [PERF_METRICS.BACKEND_BOUND] ) ) * ( [TOPDOWN.SLOTS] ) ) )"
     },
     {
         "name": "metric_TMA_..Heavy_Operations(%)",
@@ -388,4 +396,4 @@
         "expression": "(1 - [CPU_CLK_UNHALTED.ONE_THREAD_ACTIVE] / [CPU_CLK_UNHALTED.REF_DISTRIBUTED]) if [SOCKET_COUNT] > 1 else 0",
         "origin": "perfspect"
     }
-]
+]

events/spr_emr.txt

@@ -57,10 +57,20 @@ cpu-cycles,
 ref-cycles,
 instructions;
 
+cpu/event=0xcf,umask=0x10,cmask=0x00,period=100003,name='FP_ARITH_INST_RETIRED2.512B_PACKED_HALF'/,
+cpu/event=0xcf,umask=0x08,cmask=0x00,period=100003,name='FP_ARITH_INST_RETIRED2.256B_PACKED_HALF'/,
+cpu/event=0xc7,umask=0x80,cmask=0x00,period=100003,name='FP_ARITH_INST_RETIRED.512B_PACKED_SINGLE'/,
+cpu/event=0xc7,umask=0x40,cmask=0x00,period=100003,name='FP_ARITH_INST_RETIRED.512B_PACKED_DOUBLE'/,
+cpu/event=0xc7,umask=0x20,cmask=0x00,period=100003,name='FP_ARITH_INST_RETIRED.256B_PACKED_SINGLE'/,
+cpu/event=0xc7,umask=0x10,cmask=0x00,period=100003,name='FP_ARITH_INST_RETIRED.256B_PACKED_DOUBLE'/,
+cpu-cycles,
+ref-cycles,
+instructions;
+
 cpu/event=0x47,umask=0x03,cmask=0x03,period=1000003,name='MEMORY_ACTIVITY.STALLS_L1D_MISS'/,
+cpu/event=0x47,umask=0x05,cmask=0x05,period=1000003,name='MEMORY_ACTIVITY.STALLS_L2_MISS'/,
 cpu/event=0x12,umask=0x20,cmask=0x01,period=100003,name='DTLB_LOAD_MISSES.STLB_HIT:c1'/,
 cpu/event=0x12,umask=0x10,cmask=0x01,period=100003,name='DTLB_LOAD_MISSES.WALK_ACTIVE'/,
-cpu/event=0x47,umask=0x05,cmask=0x05,period=1000003,name='MEMORY_ACTIVITY.STALLS_L2_MISS'/,
 cpu/event=0xa3,umask=0x10,cmask=0x10,period=1000003,name='CYCLE_ACTIVITY.CYCLES_MEM_ANY'/,
 cpu/event=0xb0,umask=0x09,cmask=0x01,period=1000003,name='ARITH.DIV_ACTIVE'/,
 cpu/event=0xad,umask=0x80,period=500009,name='INT_MISC.CLEAR_RESTEER_CYCLES'/,
@@ -69,7 +79,6 @@ cpu-cycles,
 ref-cycles,
 instructions;
 
-cpu/event=0xce,umask=0x02,cmask=0x01,period=100003,name='AMX_OPS_RETIRED.BF16:c1'/,
 cpu/event=0xd3,umask=0x10,cmask=0x00,period=100007,name='MEM_LOAD_L3_MISS_RETIRED.REMOTE_PMM'/,
 cpu/event=0xd1,umask=0x08,cmask=0x00,period=200003,name='MEM_LOAD_RETIRED.L1_MISS'/,
 cpu/event=0xd1,umask=0x80,cmask=0x00,period=1000003,name='MEM_LOAD_RETIRED.LOCAL_PMM'/,
@@ -85,9 +94,7 @@ cpu/event=0xd0,umask=0x21,cmask=0x00,period=1000003,name='MEM_INST_RETIRED.LOCK_
 cpu/event=0xd0,umask=0x82,cmask=0x00,period=1000003,name='MEM_INST_RETIRED.ALL_STORES'/,
 cpu/event=0x24,umask=0xe2,cmask=0x00,period=2000003,name='L2_RQSTS.ALL_RFO'/,
 cpu/event=0x24,umask=0xc2,cmask=0x00,period=2000003,name='L2_RQSTS.RFO_HIT'/,
-cpu/event=0xcf,umask=0x03,cmask=0x00,period=100003,name='FP_ARITH_INST_RETIRED2.SCALAR'/,
 cpu/event=0xcf,umask=0x1c,cmask=0x00,period=100003,name='FP_ARITH_INST_RETIRED2.VECTOR'/,
-cpu/event=0xc7,umask=0x03,period=100003,name='FP_ARITH_INST_RETIRED.SCALAR_SINGLE:u0x03'/,
 cpu/event=0xc7,umask=0x3c,period=100003,name='FP_ARITH_INST_RETIRED.128B_PACKED_DOUBLE:u0x3c'/,
 cpu-cycles,
 ref-cycles,
@@ -135,6 +142,9 @@ cpu-cycles,
 ref-cycles,
 instructions;
 
+cpu/event=0xe7,umask=0x0c,cmask=0x00,period=100003,name='INT_VEC_RETIRED.ADD_256'/,
+cpu/event=0xe7,umask=0x20,cmask=0x00,period=100003,name='INT_VEC_RETIRED.VNNI_256'/,
+cpu/event=0xe7,umask=0x80,cmask=0x00,period=100003,name='INT_VEC_RETIRED.MUL_256'/,
 cpu/event=0x79,umask=0x08,cmask=0x00,period=2000003,name='IDQ.DSB_UOPS'/,
 cpu/event=0x79,umask=0x04,period=100003,name='IDQ.MITE_UOPS'/,
 cpu/event=0x79,umask=0x20,period=100003,name='IDQ.MS_UOPS'/,

src/perf_helpers.py

@@ -112,13 +112,13 @@ def get_imc_cha_upi_count():
     return imc_count, cha_count, upi_count
 
 
-# get imc channel ids, channel ids are not consecutive in some cases (observed on bdw)
-def get_channel_ids():
+# device ids are not consecutive in some cases
+def get_channel_ids(pattern):
     sysdevices = os.listdir("/sys/bus/event_source/devices")
-    imc = "uncore_imc_[0-9]*"
+    devices = pattern + "[0-9]*"
     ids = []
     for entry in sysdevices:
-        if fnmatch.fnmatch(entry, imc):
+        if fnmatch.fnmatch(entry, devices):
             words = entry.split("_")
             ids.append(int(words[-1]))
     ids = sorted(ids)

src/prepare_perf_events.py

@@ -71,17 +71,11 @@ def is_cpu_event(line):
     return False
 
 
-# save the last group names in a list when it is cha or imc
-# test for cha or imc event. append with count value
-# once reaches new group, start looping through all imc/cha counts to finish up
-def enumerate_uncore(group, pattern, n, default_range=True):
+# enumerate uncore events across all devices
+def enumerate_uncore(group, pattern, count):
     uncore_group = ""
-    ids = []
-    if default_range:
-        ids = range(n)
-    else:
-        ids = helper.get_channel_ids()
-    for i in range(n - 1):
+    ids = helper.get_channel_ids(pattern)
+    for i in range(count - 1):
         old = pattern + str(ids[i])
         new = pattern + str(ids[i + 1])
         group = group.replace(old, new)
@@ -216,8 +210,7 @@ def prepare_perf_events(
         # enumerate all uncore units
         if new_group and (unc_count > 1):
             name = helper.get_dev_name(line.split("/")[0].strip())
-            default_range = name != "uncore_imc"
-            group += enumerate_uncore(prev_group, name + "_", unc_count, default_range)
+            group += enumerate_uncore(prev_group, name + "_", unc_count)
 
     group = group[:-1]
     if len(event_names) == 0: