ruby: split CPU and GPU latency stats

2 files changed, 222 insertions, 78 deletions

diff --git a/src/mem/ruby/profiler/Profiler.cc b/src/mem/ruby/profiler/Profiler.cc
index 7d3f20982..04e8331c4 100644
--- a/src/mem/ruby/profiler/Profiler.cc
+++ b/src/mem/ruby/profiler/Profiler.cc
@@ -42,6 +42,8 @@
    ----------------------------------------------------------------------
 */
 
+#include "mem/ruby/profiler/Profiler.hh"
+
 #include <sys/types.h>
 #include <unistd.h>
 
@@ -54,7 +56,7 @@
 #include "mem/protocol/RubyRequest.hh"
 #include "mem/ruby/network/Network.hh"
 #include "mem/ruby/profiler/AddressProfiler.hh"
-#include "mem/ruby/profiler/Profiler.hh"
+#include "mem/ruby/system/GPUCoalescer.hh"
 #include "mem/ruby/system/Sequencer.hh"
 
 using namespace std;
@@ -106,131 +108,217 @@ Profiler::regStats(const std::string &pName)
             .flags(Stats::nozero | Stats::pdf | Stats::oneline);
     }
 
-    m_outstandReqHist
+    m_outstandReqHistSeqr
+        .init(10)
+        .name(pName + ".outstanding_req_hist_seqr")
+        .desc("")
+        .flags(Stats::nozero | Stats::pdf | Stats::oneline);
+
+    m_outstandReqHistCoalsr
+        .init(10)
+        .name(pName + ".outstanding_req_hist_coalsr")
+        .desc("")
+        .flags(Stats::nozero | Stats::pdf | Stats::oneline);
+
+    m_latencyHistSeqr
         .init(10)
-        .name(pName + ".outstanding_req_hist")
+        .name(pName + ".latency_hist_seqr")
         .desc("")
         .flags(Stats::nozero | Stats::pdf | Stats::oneline);
 
-    m_latencyHist
+    m_latencyHistCoalsr
         .init(10)
-        .name(pName + ".latency_hist")
+        .name(pName + ".latency_hist_coalsr")
         .desc("")
         .flags(Stats::nozero | Stats::pdf | Stats::oneline);
 
-    m_hitLatencyHist
+    m_hitLatencyHistSeqr
         .init(10)
-        .name(pName + ".hit_latency_hist")
+        .name(pName + ".hit_latency_hist_seqr")
         .desc("")
         .flags(Stats::nozero | Stats::pdf | Stats::oneline);
 
-    m_missLatencyHist
+    m_missLatencyHistSeqr
         .init(10)
-        .name(pName + ".miss_latency_hist")
+        .name(pName + ".miss_latency_hist_seqr")
+        .desc("")
+        .flags(Stats::nozero | Stats::pdf | Stats::oneline);
+
+    m_missLatencyHistCoalsr
+        .init(10)
+        .name(pName + ".miss_latency_hist_coalsr")
         .desc("")
         .flags(Stats::nozero | Stats::pdf | Stats::oneline);
 
     for (int i = 0; i < RubyRequestType_NUM; i++) {
-        m_typeLatencyHist.push_back(new Stats::Histogram());
-        m_typeLatencyHist[i]
+        m_typeLatencyHistSeqr.push_back(new Stats::Histogram());
+        m_typeLatencyHistSeqr[i]
+            ->init(10)
+            .name(pName + csprintf(".%s.latency_hist_seqr",
+                                    RubyRequestType(i)))
+            .desc("")
+            .flags(Stats::nozero | Stats::pdf | Stats::oneline);
+
+        m_typeLatencyHistCoalsr.push_back(new Stats::Histogram());
+        m_typeLatencyHistCoalsr[i]
             ->init(10)
-            .name(pName + csprintf(".%s.latency_hist",
+            .name(pName + csprintf(".%s.latency_hist_coalsr",
                                     RubyRequestType(i)))
             .desc("")
             .flags(Stats::nozero | Stats::pdf | Stats::oneline);
 
-        m_hitTypeLatencyHist.push_back(new Stats::Histogram());
-        m_hitTypeLatencyHist[i]
+        m_hitTypeLatencyHistSeqr.push_back(new Stats::Histogram());
+        m_hitTypeLatencyHistSeqr[i]
             ->init(10)
-            .name(pName + csprintf(".%s.hit_latency_hist",
+            .name(pName + csprintf(".%s.hit_latency_hist_seqr",
                                     RubyRequestType(i)))
             .desc("")
             .flags(Stats::nozero | Stats::pdf | Stats::oneline);
 
-        m_missTypeLatencyHist.push_back(new Stats::Histogram());
-        m_missTypeLatencyHist[i]
+        m_missTypeLatencyHistSeqr.push_back(new Stats::Histogram());
+        m_missTypeLatencyHistSeqr[i]
             ->init(10)
-            .name(pName + csprintf(".%s.miss_latency_hist",
+            .name(pName + csprintf(".%s.miss_latency_hist_seqr",
+                                    RubyRequestType(i)))
+            .desc("")
+            .flags(Stats::nozero | Stats::pdf | Stats::oneline);
+
+        m_missTypeLatencyHistCoalsr.push_back(new Stats::Histogram());
+        m_missTypeLatencyHistCoalsr[i]
+            ->init(10)
+            .name(pName + csprintf(".%s.miss_latency_hist_coalsr",
                                     RubyRequestType(i)))
             .desc("")
             .flags(Stats::nozero | Stats::pdf | Stats::oneline);
     }
 
     for (int i = 0; i < MachineType_NUM; i++) {
-        m_hitMachLatencyHist.push_back(new Stats::Histogram());
-        m_hitMachLatencyHist[i]
+        m_hitMachLatencyHistSeqr.push_back(new Stats::Histogram());
+        m_hitMachLatencyHistSeqr[i]
+            ->init(10)
+            .name(pName + csprintf(".%s.hit_mach_latency_hist_seqr",
+                                    MachineType(i)))
+            .desc("")
+            .flags(Stats::nozero | Stats::pdf | Stats::oneline);
+
+        m_missMachLatencyHistSeqr.push_back(new Stats::Histogram());
+        m_missMachLatencyHistSeqr[i]
             ->init(10)
-            .name(pName + csprintf(".%s.hit_mach_latency_hist",
+            .name(pName + csprintf(".%s.miss_mach_latency_hist_seqr",
                                     MachineType(i)))
             .desc("")
             .flags(Stats::nozero | Stats::pdf | Stats::oneline);
 
-        m_missMachLatencyHist.push_back(new Stats::Histogram());
-        m_missMachLatencyHist[i]
+        m_missMachLatencyHistCoalsr.push_back(new Stats::Histogram());
+        m_missMachLatencyHistCoalsr[i]
             ->init(10)
-            .name(pName + csprintf(".%s.miss_mach_latency_hist",
+            .name(pName + csprintf(".%s.miss_mach_latency_hist_coalsr",
                                     MachineType(i)))
             .desc("")
             .flags(Stats::nozero | Stats::pdf | Stats::oneline);
 
-        m_IssueToInitialDelayHist.push_back(new Stats::Histogram());
-        m_IssueToInitialDelayHist[i]
+        m_IssueToInitialDelayHistSeqr.push_back(new Stats::Histogram());
+        m_IssueToInitialDelayHistSeqr[i]
+            ->init(10)
+            .name(pName + csprintf(
+                ".%s.miss_latency_hist_seqr.issue_to_initial_request",
+                MachineType(i)))
+            .desc("")
+            .flags(Stats::nozero | Stats::pdf | Stats::oneline);
+
+        m_IssueToInitialDelayHistCoalsr.push_back(new Stats::Histogram());
+        m_IssueToInitialDelayHistCoalsr[i]
             ->init(10)
             .name(pName + csprintf(
-                ".%s.miss_latency_hist.issue_to_initial_request",
+                ".%s.miss_latency_hist_coalsr.issue_to_initial_request",
                 MachineType(i)))
             .desc("")
             .flags(Stats::nozero | Stats::pdf | Stats::oneline);
 
-        m_InitialToForwardDelayHist.push_back(new Stats::Histogram());
-        m_InitialToForwardDelayHist[i]
+        m_InitialToForwardDelayHistSeqr.push_back(new Stats::Histogram());
+        m_InitialToForwardDelayHistSeqr[i]
+            ->init(10)
+            .name(pName + csprintf(".%s.miss_latency_hist_seqr.initial_to_forward",
+                                   MachineType(i)))
+            .desc("")
+            .flags(Stats::nozero | Stats::pdf | Stats::oneline);
+
+        m_InitialToForwardDelayHistCoalsr.push_back(new Stats::Histogram());
+        m_InitialToForwardDelayHistCoalsr[i]
             ->init(10)
-            .name(pName + csprintf(".%s.miss_latency_hist.initial_to_forward",
+            .name(pName + csprintf(".%s.miss_latency_hist_coalsr.initial_to_forward",
                                    MachineType(i)))
             .desc("")
             .flags(Stats::nozero | Stats::pdf | Stats::oneline);
 
-        m_ForwardToFirstResponseDelayHist.push_back(new Stats::Histogram());
-        m_ForwardToFirstResponseDelayHist[i]
+        m_ForwardToFirstResponseDelayHistSeqr.push_back(new Stats::Histogram());
+        m_ForwardToFirstResponseDelayHistSeqr[i]
             ->init(10)
             .name(pName + csprintf(
-                ".%s.miss_latency_hist.forward_to_first_response",
+                ".%s.miss_latency_hist_seqr.forward_to_first_response",
                 MachineType(i)))
             .desc("")
             .flags(Stats::nozero | Stats::pdf | Stats::oneline);
 
-        m_FirstResponseToCompletionDelayHist.push_back(new Stats::Histogram());
-        m_FirstResponseToCompletionDelayHist[i]
+        m_ForwardToFirstResponseDelayHistCoalsr.push_back(new Stats::Histogram());
+        m_ForwardToFirstResponseDelayHistCoalsr[i]
             ->init(10)
             .name(pName + csprintf(
-                ".%s.miss_latency_hist.first_response_to_completion",
+                ".%s.miss_latency_hist_coalsr.forward_to_first_response",
                 MachineType(i)))
             .desc("")
             .flags(Stats::nozero | Stats::pdf | Stats::oneline);
 
-        m_IncompleteTimes[i]
-            .name(pName + csprintf(".%s.incomplete_times", MachineType(i)))
+        m_FirstResponseToCompletionDelayHistSeqr.push_back(new Stats::Histogram());
+        m_FirstResponseToCompletionDelayHistSeqr[i]
+            ->init(10)
+            .name(pName + csprintf(
+                ".%s.miss_latency_hist_seqr.first_response_to_completion",
+                MachineType(i)))
+            .desc("")
+            .flags(Stats::nozero | Stats::pdf | Stats::oneline);
+
+        m_FirstResponseToCompletionDelayHistCoalsr.push_back(new Stats::Histogram());
+        m_FirstResponseToCompletionDelayHistCoalsr[i]
+            ->init(10)
+            .name(pName + csprintf(
+                ".%s.miss_latency_hist_coalsr.first_response_to_completion",
+                MachineType(i)))
+            .desc("")
+            .flags(Stats::nozero | Stats::pdf | Stats::oneline);
+
+        m_IncompleteTimesSeqr[i]
+            .name(pName + csprintf(".%s.incomplete_times_seqr", MachineType(i)))
             .desc("")
             .flags(Stats::nozero);
     }
 
     for (int i = 0; i < RubyRequestType_NUM; i++) {
-        m_hitTypeMachLatencyHist.push_back(std::vector<Stats::Histogram *>());
-        m_missTypeMachLatencyHist.push_back(std::vector<Stats::Histogram *>());
+        m_hitTypeMachLatencyHistSeqr.push_back(std::vector<Stats::Histogram *>());
+        m_missTypeMachLatencyHistSeqr.push_back(std::vector<Stats::Histogram *>());
+        m_missTypeMachLatencyHistCoalsr.push_back(std::vector<Stats::Histogram *>());
 
         for (int j = 0; j < MachineType_NUM; j++) {
-            m_hitTypeMachLatencyHist[i].push_back(new Stats::Histogram());
-            m_hitTypeMachLatencyHist[i][j]
+            m_hitTypeMachLatencyHistSeqr[i].push_back(new Stats::Histogram());
+            m_hitTypeMachLatencyHistSeqr[i][j]
+                ->init(10)
+                .name(pName + csprintf(".%s.%s.hit_type_mach_latency_hist_seqr",
+                                       RubyRequestType(i), MachineType(j)))
+                .desc("")
+                .flags(Stats::nozero | Stats::pdf | Stats::oneline);
+
+            m_missTypeMachLatencyHistSeqr[i].push_back(new Stats::Histogram());
+            m_missTypeMachLatencyHistSeqr[i][j]
                 ->init(10)
-                .name(pName + csprintf(".%s.%s.hit_type_mach_latency_hist",
+                .name(pName + csprintf(".%s.%s.miss_type_mach_latency_hist_seqr",
                                        RubyRequestType(i), MachineType(j)))
                 .desc("")
                 .flags(Stats::nozero | Stats::pdf | Stats::oneline);
 
-            m_missTypeMachLatencyHist[i].push_back(new Stats::Histogram());
-            m_missTypeMachLatencyHist[i][j]
+            m_missTypeMachLatencyHistCoalsr[i].push_back(new Stats::Histogram());
+            m_missTypeMachLatencyHistCoalsr[i][j]
                 ->init(10)
-                .name(pName + csprintf(".%s.%s.miss_type_mach_latency_hist",
+                .name(pName + csprintf(".%s.%s.miss_type_mach_latency_hist_coalsr",
                                        RubyRequestType(i), MachineType(j)))
                 .desc("")
                 .flags(Stats::nozero | Stats::pdf | Stats::oneline);
@@ -271,7 +359,11 @@ Profiler::collateStats()
             AbstractController *ctr = (*it).second;
             Sequencer *seq = ctr->getCPUSequencer();
             if (seq != NULL) {
-                m_outstandReqHist.add(seq->getOutstandReqHist());
+                m_outstandReqHistSeqr.add(seq->getOutstandReqHist());
+            }
+            GPUCoalescer *coal = ctr->getGPUCoalescer();
+            if (coal != NULL) {
+                m_outstandReqHistCoalsr.add(coal->getOutstandReqHist());
             }
         }
     }
@@ -285,52 +377,93 @@ Profiler::collateStats()
             Sequencer *seq = ctr->getCPUSequencer();
             if (seq != NULL) {
                 // add all the latencies
-                m_latencyHist.add(seq->getLatencyHist());
-                m_hitLatencyHist.add(seq->getHitLatencyHist());
-                m_missLatencyHist.add(seq->getMissLatencyHist());
+                m_latencyHistSeqr.add(seq->getLatencyHist());
+                m_hitLatencyHistSeqr.add(seq->getHitLatencyHist());
+                m_missLatencyHistSeqr.add(seq->getMissLatencyHist());
 
                 // add the per request type latencies
                 for (uint32_t j = 0; j < RubyRequestType_NUM; ++j) {
-                    m_typeLatencyHist[j]
+                    m_typeLatencyHistSeqr[j]
                         ->add(seq->getTypeLatencyHist(j));
-                    m_hitTypeLatencyHist[j]
+                    m_hitTypeLatencyHistSeqr[j]
                         ->add(seq->getHitTypeLatencyHist(j));
-                    m_missTypeLatencyHist[j]
+                    m_missTypeLatencyHistSeqr[j]
                         ->add(seq->getMissTypeLatencyHist(j));
                 }
 
                 // add the per machine type miss latencies
                 for (uint32_t j = 0; j < MachineType_NUM; ++j) {
-                    m_hitMachLatencyHist[j]
+                    m_hitMachLatencyHistSeqr[j]
                         ->add(seq->getHitMachLatencyHist(j));
-                    m_missMachLatencyHist[j]
+                    m_missMachLatencyHistSeqr[j]
                         ->add(seq->getMissMachLatencyHist(j));
 
-                    m_IssueToInitialDelayHist[j]->add(
+                    m_IssueToInitialDelayHistSeqr[j]->add(
                         seq->getIssueToInitialDelayHist(MachineType(j)));
 
-                    m_InitialToForwardDelayHist[j]->add(
+                    m_InitialToForwardDelayHistSeqr[j]->add(
                         seq->getInitialToForwardDelayHist(MachineType(j)));
-                    m_ForwardToFirstResponseDelayHist[j]->add(seq->
+                    m_ForwardToFirstResponseDelayHistSeqr[j]->add(seq->
                         getForwardRequestToFirstResponseHist(MachineType(j)));
 
-                    m_FirstResponseToCompletionDelayHist[j]->add(seq->
+                    m_FirstResponseToCompletionDelayHistSeqr[j]->add(seq->
                         getFirstResponseToCompletionDelayHist(
                             MachineType(j)));
-                    m_IncompleteTimes[j] +=
+                    m_IncompleteTimesSeqr[j] +=
                         seq->getIncompleteTimes(MachineType(j));
                 }
 
                 // add the per (request, machine) type miss latencies
                 for (uint32_t j = 0; j < RubyRequestType_NUM; j++) {
                     for (uint32_t k = 0; k < MachineType_NUM; k++) {
-                        m_hitTypeMachLatencyHist[j][k]->add(
+                        m_hitTypeMachLatencyHistSeqr[j][k]->add(
                                 seq->getHitTypeMachLatencyHist(j,k));
-                        m_missTypeMachLatencyHist[j][k]->add(
+                        m_missTypeMachLatencyHistSeqr[j][k]->add(
                                 seq->getMissTypeMachLatencyHist(j,k));
                     }
                 }
             }
+
+            GPUCoalescer *coal = ctr->getGPUCoalescer();
+            if (coal != NULL) {
+                // add all the latencies
+                m_latencyHistCoalsr.add(coal->getLatencyHist());
+                m_missLatencyHistCoalsr.add(coal->getMissLatencyHist());
+
+                // add the per request type latencies
+                for (uint32_t j = 0; j < RubyRequestType_NUM; ++j) {
+                    m_typeLatencyHistCoalsr[j]
+                        ->add(coal->getTypeLatencyHist(j));
+                    m_missTypeLatencyHistCoalsr[j]
+                        ->add(coal->getMissTypeLatencyHist(j));
+                }
+
+                // add the per machine type miss latencies
+                for (uint32_t j = 0; j < MachineType_NUM; ++j) {
+                    m_missMachLatencyHistCoalsr[j]
+                        ->add(coal->getMissMachLatencyHist(j));
+
+                    m_IssueToInitialDelayHistCoalsr[j]->add(
+                        coal->getIssueToInitialDelayHist(MachineType(j)));
+
+                    m_InitialToForwardDelayHistCoalsr[j]->add(
+                        coal->getInitialToForwardDelayHist(MachineType(j)));
+                    m_ForwardToFirstResponseDelayHistCoalsr[j]->add(coal->
+                        getForwardRequestToFirstResponseHist(MachineType(j)));
+
+                    m_FirstResponseToCompletionDelayHistCoalsr[j]->add(coal->
+                        getFirstResponseToCompletionDelayHist(
+                            MachineType(j)));
+                }
+
+                // add the per (request, machine) type miss latencies
+                for (uint32_t j = 0; j < RubyRequestType_NUM; j++) {
+                    for (uint32_t k = 0; k < MachineType_NUM; k++) {
+                        m_missTypeMachLatencyHistCoalsr[j][k]->add(
+                                coal->getMissTypeMachLatencyHist(j,k));
+                    }
+                }
+            }
         }
     }
 }
diff --git a/src/mem/ruby/profiler/Profiler.hh b/src/mem/ruby/profiler/Profiler.hh
index 5be75fb65..6ad65f962 100644
--- a/src/mem/ruby/profiler/Profiler.hh
+++ b/src/mem/ruby/profiler/Profiler.hh
@@ -94,38 +94,49 @@ class Profiler
     std::vector<Stats::Histogram *> delayVCHistogram;
 
     //! Histogram for number of outstanding requests per cycle.
-    Stats::Histogram m_outstandReqHist;
+    Stats::Histogram m_outstandReqHistSeqr;
+    Stats::Histogram m_outstandReqHistCoalsr;
 
     //! Histogram for holding latency profile of all requests.
-    Stats::Histogram m_latencyHist;
-    std::vector<Stats::Histogram *> m_typeLatencyHist;
+    Stats::Histogram m_latencyHistSeqr;
+    Stats::Histogram m_latencyHistCoalsr;
+    std::vector<Stats::Histogram *> m_typeLatencyHistSeqr;
+    std::vector<Stats::Histogram *> m_typeLatencyHistCoalsr;
 
     //! Histogram for holding latency profile of all requests that
     //! hit in the controller connected to this sequencer.
-    Stats::Histogram m_hitLatencyHist;
-    std::vector<Stats::Histogram *> m_hitTypeLatencyHist;
+    Stats::Histogram m_hitLatencyHistSeqr;
+    std::vector<Stats::Histogram *> m_hitTypeLatencyHistSeqr;
 
     //! Histograms for profiling the latencies for requests that
     //! did not required external messages.
-    std::vector<Stats::Histogram *> m_hitMachLatencyHist;
-    std::vector< std::vector<Stats::Histogram *> > m_hitTypeMachLatencyHist;
+    std::vector<Stats::Histogram *> m_hitMachLatencyHistSeqr;
+    std::vector< std::vector<Stats::Histogram *> > m_hitTypeMachLatencyHistSeqr;
 
     //! Histogram for holding latency profile of all requests that
     //! miss in the controller connected to this sequencer.
-    Stats::Histogram m_missLatencyHist;
-    std::vector<Stats::Histogram *> m_missTypeLatencyHist;
+    Stats::Histogram m_missLatencyHistSeqr;
+    Stats::Histogram m_missLatencyHistCoalsr;
+    std::vector<Stats::Histogram *> m_missTypeLatencyHistSeqr;
+    std::vector<Stats::Histogram *> m_missTypeLatencyHistCoalsr;
 
     //! Histograms for profiling the latencies for requests that
     //! required external messages.
-    std::vector<Stats::Histogram *> m_missMachLatencyHist;
-    std::vector< std::vector<Stats::Histogram *> > m_missTypeMachLatencyHist;
+    std::vector<Stats::Histogram *> m_missMachLatencyHistSeqr;
+    std::vector< std::vector<Stats::Histogram *> > m_missTypeMachLatencyHistSeqr;
+    std::vector<Stats::Histogram *> m_missMachLatencyHistCoalsr;
+    std::vector< std::vector<Stats::Histogram *> > m_missTypeMachLatencyHistCoalsr;
 
     //! Histograms for recording the breakdown of miss latency
-    std::vector<Stats::Histogram *> m_IssueToInitialDelayHist;
-    std::vector<Stats::Histogram *> m_InitialToForwardDelayHist;
-    std::vector<Stats::Histogram *> m_ForwardToFirstResponseDelayHist;
-    std::vector<Stats::Histogram *> m_FirstResponseToCompletionDelayHist;
-    Stats::Scalar m_IncompleteTimes[MachineType_NUM];
+    std::vector<Stats::Histogram *> m_IssueToInitialDelayHistSeqr;
+    std::vector<Stats::Histogram *> m_InitialToForwardDelayHistSeqr;
+    std::vector<Stats::Histogram *> m_ForwardToFirstResponseDelayHistSeqr;
+    std::vector<Stats::Histogram *> m_FirstResponseToCompletionDelayHistSeqr;
+    Stats::Scalar m_IncompleteTimesSeqr[MachineType_NUM];
+    std::vector<Stats::Histogram *> m_IssueToInitialDelayHistCoalsr;
+    std::vector<Stats::Histogram *> m_InitialToForwardDelayHistCoalsr;
+    std::vector<Stats::Histogram *> m_ForwardToFirstResponseDelayHistCoalsr;
+    std::vector<Stats::Histogram *> m_FirstResponseToCompletionDelayHistCoalsr;
 
     //added by SS
     const bool m_hot_lines;