Merge pull request #237 from intel/develop

Develop

src/common/common_utils.cpp

@@ -43,7 +43,7 @@ namespace yask {
     // for numbers above 9 (at least up to 99).
 
     // Format: "major.minor.patch".
-    const string version = "3.02.00";
+    const string version = "3.02.01";
 
     string yask_get_version_string() {
         return version;

src/kernel/lib/settings.cpp

@@ -325,7 +325,7 @@ namespace yask {
         // Following options are in the 'yask' namespace, i.e., no object.
         parser.add_option(new CommandLineParser::BoolOption
                           ("print_suffixes",
-                           "Format output with suffixes for human readibility, e.g., 6.15K, 12.3GiB, 7.45m."
+                           "Format output with suffixes for human readibility, e.g., 6.15K, 12.3G, 7.45m."
                            " If disabled, prints without suffixes for computer parsing, e.g., 6150, 1.23e+10, 7.45e-3.",
                            yask::is_suffix_print_enabled));
 
@@ -344,11 +344,6 @@ namespace yask {
 #endif
         _add_domain_option(parser, "mp", "Minimum var-padding size (including halo)", _min_pad_sizes);
         _add_domain_option(parser, "ep", "Extra var-padding size (beyond halo)", _extra_pad_sizes);
-        parser.add_option(new CommandLineParser::BoolOption
-                          ("allow_addl_padding",
-                           "Allow automatic extension of padding beyond what is needed for"
-                           " vector alignment for additional performance reasons",
-                           _allow_addl_pad));
 #ifdef USE_MPI
         _add_domain_option(parser, "nr", "Num ranks", _num_ranks);
         _add_domain_option(parser, "ri", "This rank's logical index (0-based)", _rank_indices);
@@ -365,6 +360,11 @@ namespace yask {
                            "Minimum width of MPI exterior section to compute before starting MPI communication.",
                            _min_exterior));
 #endif
+        parser.add_option(new CommandLineParser::BoolOption
+                          ("allow_addl_padding",
+                           "Allow automatic extension of padding beyond what is needed for"
+                           " vector alignment for additional performance reasons.",
+                           _allow_addl_pad));
         parser.add_option(new CommandLineParser::BoolOption
                           ("force_scalar",
                            "Evaluate every var point with scalar stencil operations (for debug).",
@@ -400,7 +400,7 @@ namespace yask {
                            "This setting may increase cache locality when using multiple "
                            "block-threads when scratch vars are used and/or "
                            "when temporal blocking is active. "
-                           "This option is ignored if there are fewer than two block threads.",
+                           "This option is ignored if there is only one thread per block.",
                            bind_block_threads));
 #ifdef USE_NUMA
         stringstream msg;

src/kernel/lib/soln_apis.cpp

@@ -272,7 +272,9 @@ namespace yask {
 #ifdef USE_MPI
         DEBUG_MSG(" num-ranks:             " << opts->_num_ranks.makeDimValStr(" * ") << endl <<
                   " rank-indices:          " << opts->_rank_indices.makeDimValStr() << endl <<
-                  " local-domain-offsets:  " << rank_domain_offsets.makeDimValStr(dims->_domain_dims));
+                  " local-domain-offsets:  " << rank_domain_offsets.makeDimValStr(dims->_domain_dims) << endl <<
+                  " use-shm:               " << opts->use_shm << endl <<
+                  " overlap-comms:         " << opts->overlap_comms);
         if (opts->overlap_comms)
             DEBUG_MSG(" mpi-interior:          " << mpi_interior.make_range_string(domain_dims));
 #endif

src/kernel/yask.sh

@@ -317,7 +317,7 @@ else
 fi
 
 # Commands to capture some important system status and config info for benchmark documentation.
-config_cmds="sleep 1; uptime; lscpu; sed '/^$/q' /proc/cpuinfo; uname -a; $dump /etc/system-release; $dump /proc/cmdline; $dump /proc/meminfo; free -gt; numactl -H; ulimit -a"
+config_cmds="sleep 1; uptime; lscpu; cpuinfo -A; sed '/^$/q' /proc/cpuinfo; $dump /sys/devices/system/cpu/cpufreq/boost; uname -a; $dump /etc/system-release; $dump /proc/cmdline; $dump /proc/meminfo; free -gt; numactl -H; ulimit -a"
 
 # Command sequence to be run in a shell.
 # Captures

utils/bin/yask_log_to_csv.pl

@@ -51,9 +51,16 @@
 for my $arg (@ARGV) {
   for my $fn (glob $arg) {
     my %results;
-    YaskUtils::getResultsFromFile(\%results, $fn);
 
-    YaskUtils::printCsvValues(\%results, $outFH);
-    print $outFH ",\"$fn\"\n";
+    # Parse file.
+    if (YaskUtils::getResultsFromFile(\%results, $fn)) {
+
+      # Print results if ok.
+      YaskUtils::printCsvValues(\%results, $outFH);
+      print $outFH ",\"$fn\"\n";
+    }
+    else {
+      warn "Warning: log file '$fn' does not seem to contain a successful YASK run; ignoring.\n";
+    }
   }
 }

utils/lib/YaskUtils.pm

@@ -32,10 +32,12 @@ use Carp;
 # Special keys.
 my $linux_key = "Linux kernel";
 my $nodes_key = "MPI node(s)";
+my $auto_tuner_key = "auto-tuner used";
 our @special_log_keys =
   (
    $linux_key,
    $nodes_key,
+   $auto_tuner_key,
    );
 
 # Values to get from log file.
@@ -88,13 +90,19 @@ our @log_keys =
    'L2 prefetch distance',
    'num temporal block steps',
    'num wave front steps',
+   'use shm',
+   'overlap comms',
 
    # other values from log file.
+   'vendor ID',
    'model name',
    'CPU(s)',
+   'thread(s) per core',
    'core(s) per socket',
    'socket(s)',
    'NUMA node(s)',
+   'CPU MHz',
+   'CPU max MHz',
    'MemTotal',
    'MemFree',
    'ShMem',
@@ -213,7 +221,6 @@ sub getResultsFromLine($$) {
   $line =~ s/target.ISA/target/g;
 
   # special cases for manual parsing...
-  # TODO: catch output of auto-tuner and update relevant results.
 
   # Output of 'uname -a'
   if ($line =~ /^\s*Linux\s/) {
@@ -234,6 +241,8 @@ sub getResultsFromLine($$) {
   elsif ($line =~ /^auto-tuner(.).*size:/) {
     my $c = $1;
 
+    $results->{$auto_tuner_key} = 'true';
+
     # If colon found above, tuner is global.
     my $onep = ($c eq ':');
 
@@ -248,6 +257,14 @@ sub getResultsFromLine($$) {
     }
   }
 
+  # Shared-mem & overlap backward-compat.
+  elsif ($line =~ /^Allocating.*MPI buffer/) {
+    $results->{'use shm'} = ($line =~ /using MPI shm/) ? 'true' : 'false';
+  }
+  elsif ($line =~ /^\s*mpi-interior:/) {
+    $results->{'overlap comms'} = 'true';
+  }
+
   # look for matches to all other keys.
   else {
     my ($key, $val) = split /:/,$line,2;
@@ -288,15 +305,32 @@ sub getResultsFromFile($$) {
   my $results = shift;          # ref to hash.
   my $fname = shift;            # filename.
 
+  # Init values.
+  $results->{$auto_tuner_key} = 'false';
+  $results->{'use shm'} = 'false';
+  $results->{'overlap comms'} = 'false';
+  my $ok = 0;
+
+  # Open file.
   my $fh = new FileHandle;
   if (!$fh->open("<$fname")) {
     carp "error: cannot open '$fname'";
   } else {
+
+    # Parse each line.
     while (<$fh>) {
+      $ok = 1 if /DONE/;
+      if (/^Setup for validation/) {
+        $ok = 1;
+        last;
+      }
+
       getResultsFromLine($results, $_);
     }
     $fh->close();
   }
+
+  return $ok;
 }
 
 # Print standard CSV header to given file.
@@ -317,8 +351,8 @@ sub printCsvValues($$) {
   for my $m (@all_log_keys) {
     my $r = $results->{$m};
     $r = '' if !defined $r;
-    $r = '"'.$r.'"'  # add quotes if not a number.
-      if $r !~ /^[0-9.e+-]+$/ || $r =~ /[.].*[.]/;
+    $r = '"'.$r.'"'  # add quotes if not simple chars.
+      if $r !~ /^[\w.]+$/ || $r =~ /[.].*[.]/;
     push @cols, $r;
   }
   print $fh join(',', @cols);