Merge tag '2022-05-04' into push-2022-05-04

rdementi · rdementi · commit edbdc7371fa2 · 2022-05-04T12:39:08.000+02:00
2022-05-04

Change-Id: I445612fad539ae5bb82388800fdb43c4a10a5b47
diff --git a/.bdsignore.all b/.bdsignore.all
@@ -30,3 +30,8 @@ My Inspector XE Results - pcm
 .*\.vcxproj.user
 .*\.log
 .*\.vcxproj.filters
+.*\.md
+.codedocs
+.cproject
+.project
+makefile
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -52,9 +52,9 @@ if(UNIX)  # APPLE, LINUX, FREE_BSD
     message(STATUS "CMAKE_C_FLAGS: ${CMAKE_C_FLAGS}")
 
     set(PCM_OPTIONAL_FLAGS "-Wall")
-    set(CMAKE_CXX_FLAGS_RELEASE "${PCM_OPTIONAL_FLAGS} -O3")
+    set(CMAKE_CXX_FLAGS_RELEASE "${PCM_OPTIONAL_FLAGS} -O3 -D_FORTIFY_SOURCE=1")
     set(CMAKE_CXX_FLAGS_DEBUG "${PCM_OPTIONAL_FLAGS} -O0 -g")
-    set(CMAKE_CXX_FLAGS_RELWITHDEBINFO "${PCM_OPTIONAL_FLAGS} -O3 -g")
+    set(CMAKE_CXX_FLAGS_RELWITHDEBINFO "${PCM_OPTIONAL_FLAGS} -O3 -g -D_FORTIFY_SOURCE=1")
     set(CMAKE_C_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE}")
     set(CMAKE_C_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG}")
     set(CMAKE_C_FLAGS_RELWITHDEBINFO "${CMAKE_CXX_FLAGS_RELWITHDEBINFO}")
diff --git a/src/cpucounters.cpp b/src/cpucounters.cpp
@@ -2322,14 +2322,37 @@ bool PCM::isCPUModelSupported(const int model_)
 
 bool PCM::checkModel()
 {
-    if (cpu_model == NEHALEM) cpu_model = NEHALEM_EP;
-    if (cpu_model == ATOM_2) cpu_model = ATOM;
-    if (cpu_model == HASWELL_ULT || cpu_model == HASWELL_2) cpu_model = HASWELL;
-    if (cpu_model == BROADWELL_XEON_E3) cpu_model = BROADWELL;
-    if (cpu_model == ICX_D) cpu_model = ICX;
-    if (cpu_model == CML_1) cpu_model = CML;
-    if (cpu_model == ICL_1) cpu_model = ICL;
-    if (cpu_model == TGL_1) cpu_model = TGL;
+    switch (cpu_model)
+    {
+        case NEHALEM:
+            cpu_model = NEHALEM_EP;
+            break;
+        case ATOM_2:
+            cpu_model = ATOM;
+            break;
+        case HASWELL_ULT:
+        case HASWELL_2:
+            cpu_model = HASWELL;
+            break;
+        case BROADWELL_XEON_E3:
+            cpu_model = BROADWELL;
+            break;
+        case ICX_D:
+            cpu_model = ICX;
+            break;
+        case CML_1:
+            cpu_model = CML;
+            break;
+        case ICL_1:
+            cpu_model = ICL;
+            break;
+        case TGL_1:
+            cpu_model = TGL;
+            break;
+        case ADL_1:
+            cpu_model = ADL;
+            break;
+    }
 
     if(!isCPUModelSupported((int)cpu_model))
     {
diff --git a/src/cpucounters.h b/src/cpucounters.h
@@ -1401,6 +1401,7 @@ class PCM_API PCM
         TGL = 140,
         TGL_1 = 141,
         ADL = 151,
+        ADL_1 = 154,
         BDX = 79,
         KNL = 87,
         SKL = 94,
diff --git a/src/opCode-106.txt b/src/opCode-106.txt
@@ -5,19 +5,35 @@ ctr=0,ev_sel=0x83,umask=0x1,en=1,ch_mask=1,fc_mask=0x7,multiplier=4,divider=1,hn
 ctr=1,ev_sel=0x83,umask=0x1,en=1,ch_mask=2,fc_mask=0x7,multiplier=4,divider=1,hname=IB write,vname=Part1 (2nd x4)
 ctr=0,ev_sel=0x83,umask=0x1,en=1,ch_mask=4,fc_mask=0x7,multiplier=4,divider=1,hname=IB write,vname=Part2 (2nd x8/3rd x4)
 ctr=1,ev_sel=0x83,umask=0x1,en=1,ch_mask=8,fc_mask=0x7,multiplier=4,divider=1,hname=IB write,vname=Part3 (4th x4)
+ctr=0,ev_sel=0x83,umask=0x1,en=1,ch_mask=16,fc_mask=0x7,multiplier=4,divider=1,hname=IB write,vname=Part4 (1st x16/x8/x4)
+ctr=1,ev_sel=0x83,umask=0x1,en=1,ch_mask=32,fc_mask=0x7,multiplier=4,divider=1,hname=IB write,vname=Part5 (2nd x4)
+ctr=0,ev_sel=0x83,umask=0x1,en=1,ch_mask=64,fc_mask=0x7,multiplier=4,divider=1,hname=IB write,vname=Part6 (2nd x8/3rd x4)
+ctr=1,ev_sel=0x83,umask=0x1,en=1,ch_mask=128,fc_mask=0x7,multiplier=4,divider=1,hname=IB write,vname=Part7 (4th x4)
 ctr=0,ev_sel=0x83,umask=0x4,en=1,ch_mask=1,fc_mask=0x7,multiplier=4,divider=1,hname=IB read,vname=Part0 (1st x16/x8/x4)
 ctr=1,ev_sel=0x83,umask=0x4,en=1,ch_mask=2,fc_mask=0x7,multiplier=4,divider=1,hname=IB read,vname=Part1 (2nd x4)
 ctr=0,ev_sel=0x83,umask=0x4,en=1,ch_mask=4,fc_mask=0x7,multiplier=4,divider=1,hname=IB read,vname=Part2 (2nd x8/3rd x4)
 ctr=1,ev_sel=0x83,umask=0x4,en=1,ch_mask=8,fc_mask=0x7,multiplier=4,divider=1,hname=IB read,vname=Part3 (4th x4)
+ctr=0,ev_sel=0x83,umask=0x4,en=1,ch_mask=16,fc_mask=0x7,multiplier=4,divider=1,hname=IB read,vname=Part4 (1st x16/x8/x4)
+ctr=1,ev_sel=0x83,umask=0x4,en=1,ch_mask=32,fc_mask=0x7,multiplier=4,divider=1,hname=IB read,vname=Part5 (2nd x4)
+ctr=0,ev_sel=0x83,umask=0x4,en=1,ch_mask=64,fc_mask=0x7,multiplier=4,divider=1,hname=IB read,vname=Part6 (2nd x8/3rd x4)
+ctr=1,ev_sel=0x83,umask=0x4,en=1,ch_mask=128,fc_mask=0x7,multiplier=4,divider=1,hname=IB read,vname=Part7 (4th x4)
 # Outbound (CPU MMIO to the PCIe device) payload events
 ctr=2,ev_sel=0x83,umask=0x80,en=1,ch_mask=1,fc_mask=0x7,multiplier=4,divider=1,hname=OB read,vname=Part0 (1st x16/x8/x4)
 ctr=3,ev_sel=0x83,umask=0x80,en=1,ch_mask=2,fc_mask=0x7,multiplier=4,divider=1,hname=OB read,vname=Part1 (2nd x4)
 ctr=2,ev_sel=0x83,umask=0x80,en=1,ch_mask=4,fc_mask=0x7,multiplier=4,divider=1,hname=OB read,vname=Part2 (2nd x8/3rd x4)
 ctr=3,ev_sel=0x83,umask=0x80,en=1,ch_mask=8,fc_mask=0x7,multiplier=4,divider=1,hname=OB read,vname=Part3 (4th x4)
+ctr=2,ev_sel=0x83,umask=0x80,en=1,ch_mask=16,fc_mask=0x7,multiplier=4,divider=1,hname=OB read,vname=Part4 (1st x16/x8/x4)
+ctr=3,ev_sel=0x83,umask=0x80,en=1,ch_mask=32,fc_mask=0x7,multiplier=4,divider=1,hname=OB read,vname=Part5 (2nd x4)
+ctr=2,ev_sel=0x83,umask=0x80,en=1,ch_mask=64,fc_mask=0x7,multiplier=4,divider=1,hname=OB read,vname=Part6 (2nd x8/3rd x4)
+ctr=3,ev_sel=0x83,umask=0x80,en=1,ch_mask=128,fc_mask=0x7,multiplier=4,divider=1,hname=OB read,vname=Part7 (4th x4)
 ctr=2,ev_sel=0xc0,umask=0x1,en=1,ch_mask=1,fc_mask=0x7,multiplier=4,divider=1,hname=OB write,vname=Part0 (1st x16/x8/x4)
 ctr=3,ev_sel=0xc0,umask=0x1,en=1,ch_mask=2,fc_mask=0x7,multiplier=4,divider=1,hname=OB write,vname=Part1 (2nd x4)
 ctr=2,ev_sel=0xc0,umask=0x1,en=1,ch_mask=4,fc_mask=0x7,multiplier=4,divider=1,hname=OB write,vname=Part2 (2nd x8/3rd x4)
 ctr=3,ev_sel=0xc0,umask=0x1,en=1,ch_mask=8,fc_mask=0x7,multiplier=4,divider=1,hname=OB write,vname=Part3 (4th x4)
+ctr=2,ev_sel=0xc0,umask=0x1,en=1,ch_mask=16,fc_mask=0x7,multiplier=4,divider=1,hname=OB write,vname=Part4 (1st x16/x8/x4)
+ctr=3,ev_sel=0xc0,umask=0x1,en=1,ch_mask=32,fc_mask=0x7,multiplier=4,divider=1,hname=OB write,vname=Part5 (2nd x4)
+ctr=2,ev_sel=0xc0,umask=0x1,en=1,ch_mask=64,fc_mask=0x7,multiplier=4,divider=1,hname=OB write,vname=Part6 (2nd x8/3rd x4)
+ctr=3,ev_sel=0xc0,umask=0x1,en=1,ch_mask=128,fc_mask=0x7,multiplier=4,divider=1,hname=OB write,vname=Part7 (4th x4)
 # IOMMU events
 ctr=0,ev_sel=0x40,umask=0x02,en=1,ch_mask=0x0,fc_mask=0x0,multiplier=1,divider=1,hname=IOTLB Lookup,vname=Total
 ctr=1,ev_sel=0x40,umask=0x20,en=1,ch_mask=0x0,fc_mask=0x0,multiplier=1,divider=1,hname=IOTLB Miss,vname=Total
@@ -26,4 +42,4 @@ ctr=3,ev_sel=0x41,umask=0x10,en=1,ch_mask=0x0,fc_mask=0x0,multiplier=1,divider=1
 ctr=0,ev_sel=0x41,umask=0x08,en=1,ch_mask=0x0,fc_mask=0x0,multiplier=1,divider=1,hname=1G Cache Hit,vname=Total
 ctr=1,ev_sel=0x41,umask=0x04,en=1,ch_mask=0x0,fc_mask=0x0,multiplier=1,divider=1,hname=2M Cache Hit,vname=Total
 ctr=2,ev_sel=0x41,umask=0x02,en=1,ch_mask=0x0,fc_mask=0x0,multiplier=1,divider=1,hname=4K Cache Hit,vname=Total
-ctr=3,ev_sel=0x41,umask=0x40,en=1,ch_mask=0x0,fc_mask=0x0,multiplier=1,divider=1,hname=IOMMU Mem Access,vname=Total
+ctr=3,ev_sel=0x41,umask=0x40,en=1,ch_mask=0x0,fc_mask=0x0,multiplier=1,divider=1,hname=IOMMU Mem Access,vname=Total
diff --git a/src/pcm-iio.cpp b/src/pcm-iio.cpp
@@ -82,6 +82,15 @@ static const std::string icx_iio_stack_names[6] = {
     "IIO Stack 5 - CBDMA/DMI      "
 };
 
+static const std::string icx_d_iio_stack_names[6] = {
+    "IIO Stack 0 - MCP            ",
+    "IIO Stack 1 - PCIe0          ",
+    "IIO Stack 2 - CBDMA/DMI      ",
+    "IIO Stack 3 - PCIe2          ",
+    "IIO Stack 4 - PCIe3          ",
+    "IIO Stack 5 - PCIe1          "
+};
+
 static const std::string snr_iio_stack_names[5] = {
     "IIO Stack 0 - QAT            ",
     "IIO Stack 1 - CBDMA/DMI      ",
@@ -111,6 +120,15 @@ static const std::map<int, int> icx_sad_to_pmu_id_mapping = {
     { 5,                    4 }
 };
 
+static const std::map<int, int> icx_d_sad_to_pmu_id_mapping = {
+    { ICX_CBDMA_DMI_SAD_ID, 2 },
+    { 1,                    5 },
+    { 2,                    1 },
+    { ICX_MCP_SAD_ID,       0 },
+    { 4,                    3 },
+    { 5,                    4 }
+};
+
 #define SNR_ACCELERATOR_PART_ID 4
 
 #define SNR_ROOT_PORT_A_DID 0x334A
@@ -545,8 +563,16 @@ bool IPlatformMapping10Nm::getSadIdRootBusMap(uint32_t socket_id, std::map<uint8
 // Mapping for IceLake Server.
 class WhitleyPlatformMapping: public IPlatformMapping10Nm {
 private:
+    const bool icx_d;
+    const std::map<int, int>& sad_to_pmu_id_mapping;
+    const std::string * iio_stack_names;
 public:
-    WhitleyPlatformMapping() = default;
+    WhitleyPlatformMapping() :
+        icx_d(PCM::getInstance()->getCPUModelFromCPUID() == PCM::ICX_D),
+        sad_to_pmu_id_mapping(icx_d ? icx_d_sad_to_pmu_id_mapping : icx_sad_to_pmu_id_mapping),
+        iio_stack_names(icx_d ? icx_d_iio_stack_names : icx_iio_stack_names)
+    {
+    }
     ~WhitleyPlatformMapping() = default;
     bool pciTreeDiscover(std::vector<struct iio_stacks_on_socket>& iios, uint32_t sockets_count) override;
 };
@@ -563,15 +589,15 @@ bool WhitleyPlatformMapping::pciTreeDiscover(std::vector<struct iio_stacks_on_so
 
         {
             struct iio_stack stack;
-            stack.iio_unit_id = icx_sad_to_pmu_id_mapping.at(ICX_MCP_SAD_ID);
-            stack.stack_name = icx_iio_stack_names[stack.iio_unit_id];
+            stack.iio_unit_id = sad_to_pmu_id_mapping.at(ICX_MCP_SAD_ID);
+            stack.stack_name = iio_stack_names[stack.iio_unit_id];
             iio_on_socket.stacks.push_back(stack);
         }
 
         for (auto sad_id_bus_pair = sad_id_bus_map.cbegin(); sad_id_bus_pair != sad_id_bus_map.cend(); ++sad_id_bus_pair) {
             int sad_id = sad_id_bus_pair->first;
-            if (icx_sad_to_pmu_id_mapping.find(sad_id) ==
-                icx_sad_to_pmu_id_mapping.end()) {
+            if (sad_to_pmu_id_mapping.find(sad_id) ==
+                sad_to_pmu_id_mapping.end()) {
                 cerr << "Unknown SAD ID: " << sad_id << endl;
                 return false;
             }
@@ -584,9 +610,9 @@ bool WhitleyPlatformMapping::pciTreeDiscover(std::vector<struct iio_stacks_on_so
             int root_bus = sad_id_bus_pair->second;
             if (sad_id == ICX_CBDMA_DMI_SAD_ID) {
                 // There is one DMA Controller on each socket
-                stack.iio_unit_id = icx_sad_to_pmu_id_mapping.at(sad_id);
+                stack.iio_unit_id = sad_to_pmu_id_mapping.at(sad_id);
                 stack.busno = root_bus;
-                stack.stack_name = icx_iio_stack_names[stack.iio_unit_id];
+                stack.stack_name = iio_stack_names[stack.iio_unit_id];
 
                 // PCH is on socket 0 only
                 if (socket == 0) {
@@ -629,8 +655,8 @@ bool WhitleyPlatformMapping::pciTreeDiscover(std::vector<struct iio_stacks_on_so
                 continue;
             }
             stack.busno = root_bus;
-            stack.iio_unit_id = icx_sad_to_pmu_id_mapping.at(sad_id);
-            stack.stack_name = icx_iio_stack_names[stack.iio_unit_id];
+            stack.iio_unit_id = sad_to_pmu_id_mapping.at(sad_id);
+            stack.stack_name = iio_stack_names[stack.iio_unit_id];
             for (int slot = 2; slot < 6; slot++) {
                 struct pci pci;
                 pci.bdf.busno = root_bus;
@@ -1009,7 +1035,7 @@ result_content get_IIO_Samples(PCM *m, const std::vector<struct iio_stacks_on_so
     uint64 rawEvents[4] = {0};
     std::unique_ptr<ccr> pccr(get_ccr(m, ctr.ccr));
     rawEvents[ctr.idx] = pccr->get_ccr_value();
-    int stacks_count = (int)iios[0].stacks.size();
+    const int stacks_count = (int)m->getMaxNumOfIIOStacks();
     before = new IIOCounterState[iios.size() * stacks_count];
     after = new IIOCounterState[iios.size() * stacks_count];
 
diff --git a/src/pcm-memory.cpp b/src/pcm-memory.cpp
@@ -94,7 +94,7 @@ void print_help(const string prog_name)
     cerr << "  -pmm | /pmm | -pmem | /pmem        => monitor PMM memory bandwidth and DRAM cache hit rate in Memory Mode (default on systems with PMM support).\n";
     cerr << "  -mm                                => monitor detailed PMM Memory Mode metrics per-socket.\n";
     cerr << "  -mixed                             => monitor PMM mixed mode (AppDirect + Memory Mode).\n";
-    cerr << "  -partial                           => monitor monitor partial writes instead of PMM (default on systems without PMM support).\n";
+    cerr << "  -partial                           => monitor partial writes instead of PMM (default on systems without PMM support).\n";
     cerr << "  -nc   | --nochannel | /nc          => suppress output for individual channels.\n";
     cerr << "  -csv[=file.csv] | /csv[=file.csv]  => output compact CSV format to screen or\n"
          << "                                        to a file, in case filename is provided\n";
diff --git a/src/pcm-numa.cpp b/src/pcm-numa.cpp
@@ -54,6 +54,7 @@ void print_usage(const string progname)
     cerr << "                                        will read counters only after external program finishes\n";
     cerr << " Supported <options> are: \n";
     cerr << "  -h    | --help  | /h               => print this help and exit\n";
+    cerr << "  -pid PID | /pid PID                => collect core metrics only for specified process ID\n";
     cerr << "  -csv[=file.csv] | /csv[=file.csv]  => output compact CSV format to screen or\n"
          << "                                        to a file, in case filename is provided\n";
     cerr << "  -i[=number] | /i[=number]          => allow to determine number of iterations\n";
diff --git a/src/pcm-tsx.cpp b/src/pcm-tsx.cpp
@@ -163,6 +163,7 @@ void print_usage(const string progname)
     cerr << " Supported <options> are: \n";
     cerr << "  -h    | --help  | /h               => print this help and exit\n";
     cerr << "  -F    | -force                     => force running this program despite lack of HW RTM support (optional)\n";
+    cerr << "  -pid PID | /pid PID                => collect core metrics only for specified process ID\n";
     cerr << "  -csv[=file.csv] | /csv[=file.csv]  => output compact CSV format to screen or\n"
          << "                                        to a file, in case filename is provided\n";
     cerr << "  -i[=number] | /i[=number]          => allow to determine number of iterations\n";
diff --git a/tests/test.sh b/tests/test.sh
@@ -221,14 +221,14 @@ done
 
 
 # Check pcm-raw with TSV files
-./pcm-raw -r -e LD_BLOCKS.STORE_FORWARD -e CPU_CLK_UNHALTED.THREAD_ANY -e INST_RETIRED.ANY -e UNC_CHA_CLOCKTICKS -- sleep 1
-
-if [ "$?" -ne "0" ]; then
-    echo "Error in pcm-raw"
-    rm -rf mapfile.csv
-    cp "mapfile.csv_orig" "mapfile.csv"
-    exit 1
-fi
+#./pcm-raw -r -e LD_BLOCKS.STORE_FORWARD -e CPU_CLK_UNHALTED.THREAD_ANY -e INST_RETIRED.ANY -e UNC_CHA_CLOCKTICKS -- sleep 1
+
+#if [ "$?" -ne "0" ]; then
+#    echo "Error in pcm-raw"
+#    rm -rf mapfile.csv
+#    cp "mapfile.csv_orig" "mapfile.csv"
+#    exit 1
+#fi
 rm -rf mapfile.csv
 cp "mapfile.csv_orig" "mapfile.csv"
 
