gpu: fix bugs with MemFence, Flat Instrs and Resource utilization

Both Memory Fence is now flagged as Global Memory only to avoid resource
oversubscribing.
Flat instructions now check for Shared Memory resource busy to avoid
oversubscribing resources.
All WaitClass resources now use cycles (not ticks) to register the number
of pipe stages between Scoreboard and Execute to be consistent with
instruction scheduling logic which always used clock cycles.

3 files changed, 9 insertions, 11 deletions

diff --git a/src/gpu-compute/code_enums.hh b/src/gpu-compute/code_enums.hh
index 126cf6c50..6cd9bfe26 100644
--- a/src/gpu-compute/code_enums.hh
+++ b/src/gpu-compute/code_enums.hh
@@ -84,6 +84,7 @@
                     ||(a)==Enums::OT_PRIVATE_ATOMIC \
                     ||(a)==Enums::OT_SPILL_ATOMIC \
                     ||(a)==Enums::OT_READONLY_ATOMIC \
+                    ||(a)==Enums::OT_BOTH_MEMFENCE \
                     ||(a)==Enums::OT_FLAT_ATOMIC)
 
 #define IS_OT_ATOMIC_GM(a) ((a)==Enums::OT_GLOBAL_ATOMIC \
@@ -93,8 +94,7 @@
                     ||(a)==Enums::OT_BOTH_MEMFENCE)
 
 #define IS_OT_ATOMIC_LM(a) ((a)==Enums::OT_SHARED_ATOMIC \
-                    ||(a)==Enums::OT_SHARED_MEMFENCE \
-                    ||(a)==Enums::OT_BOTH_MEMFENCE)
+                    ||(a)==Enums::OT_SHARED_MEMFENCE)
 
 #define IS_OT_ATOMIC_PM(a) ((a)==Enums::OT_PRIVATE_ATOMIC)
 
diff --git a/src/gpu-compute/compute_unit.cc b/src/gpu-compute/compute_unit.cc
index d3622007a..63f3e8fb5 100644
--- a/src/gpu-compute/compute_unit.cc
+++ b/src/gpu-compute/compute_unit.cc
@@ -587,8 +587,8 @@ void
 ComputeUnit::init()
 {
     // Initialize CU Bus models
-    glbMemToVrfBus.init(&shader->tick_cnt, 1);
-    locMemToVrfBus.init(&shader->tick_cnt, 1);
+    glbMemToVrfBus.init(&shader->tick_cnt, shader->ticks(1));
+    locMemToVrfBus.init(&shader->tick_cnt, shader->ticks(1));
     nextGlbMemBus = 0;
     nextLocMemBus = 0;
     fatal_if(numGlbMemUnits > 1,
@@ -596,7 +596,7 @@ ComputeUnit::init()
     vrfToGlobalMemPipeBus.resize(numGlbMemUnits);
     for (int j = 0; j < numGlbMemUnits; ++j) {
         vrfToGlobalMemPipeBus[j] = WaitClass();
-        vrfToGlobalMemPipeBus[j].init(&shader->tick_cnt, 1);
+        vrfToGlobalMemPipeBus[j].init(&shader->tick_cnt, shader->ticks(1));
     }
 
     fatal_if(numLocMemUnits > 1,
@@ -604,7 +604,7 @@ ComputeUnit::init()
     vrfToLocalMemPipeBus.resize(numLocMemUnits);
     for (int j = 0; j < numLocMemUnits; ++j) {
         vrfToLocalMemPipeBus[j] = WaitClass();
-        vrfToLocalMemPipeBus[j].init(&shader->tick_cnt, 1);
+        vrfToLocalMemPipeBus[j].init(&shader->tick_cnt, shader->ticks(1));
     }
     vectorRegsReserved.resize(numSIMDs, 0);
     aluPipe.resize(numSIMDs);
@@ -612,12 +612,12 @@ ComputeUnit::init()
 
     for (int i = 0; i < numSIMDs + numLocMemUnits + numGlbMemUnits; ++i) {
         wfWait[i] = WaitClass();
-        wfWait[i].init(&shader->tick_cnt, 1);
+        wfWait[i].init(&shader->tick_cnt, shader->ticks(1));
     }
 
     for (int i = 0; i < numSIMDs; ++i) {
         aluPipe[i] = WaitClass();
-        aluPipe[i].init(&shader->tick_cnt, 1);
+        aluPipe[i].init(&shader->tick_cnt, shader->ticks(1));
     }
 
     // Setup space for call args
diff --git a/src/gpu-compute/wavefront.cc b/src/gpu-compute/wavefront.cc
index 0aa033db1..ed13b22c7 100644
--- a/src/gpu-compute/wavefront.cc
+++ b/src/gpu-compute/wavefront.cc
@@ -162,7 +162,6 @@ Wavefront::isGmInstruction(GPUDynInstPtr ii)
 
     if (IS_OT_READ_GM(ii->opType()) || IS_OT_WRITE_GM(ii->opType()) ||
         IS_OT_ATOMIC_GM(ii->opType())) {
-
         return true;
     }
 
@@ -349,7 +348,7 @@ Wavefront::ready(itype_e type)
     }
     bool locMemBusRdy = false;
     bool locMemIssueRdy = false;
-    if (type == I_SHARED) {
+    if (type == I_SHARED || type == I_FLAT) {
         for (int j=0; j < computeUnit->numLocMemUnits; ++j) {
             if (computeUnit->vrfToLocalMemPipeBus[j].prerdy())
                 locMemBusRdy = true;
@@ -598,7 +597,6 @@ Wavefront::ready(itype_e type)
 
     DPRINTF(GPUExec, "CU%d: WF[%d][%d]: Ready Inst : %s\n", computeUnit->cu_id,
             simdId, wfSlotId, ii->disassemble());
-
     return 1;
 }