Fix correlation on A100 CUTLASS GEMM Kernels (#127)

* 1. Add a queue to promote fairness CTA scheduling
2. Add a Perfect xbar interconnect option

* remove merge queue

* Should not change this

* rename pending cta to allocated_ctas

* run formatter

1 files changed, 15 insertions, 1 deletions

diff --git a/src/gpgpu-sim/local_interconnect.cc b/src/gpgpu-sim/local_interconnect.cc
index ce4415c..93689c7 100644
--- a/src/gpgpu-sim/local_interconnect.cc
+++ b/src/gpgpu-sim/local_interconnect.cc
@@ -116,10 +116,24 @@ void xbar_router::Advance() {
     RR_Advance();
   else if (arbit_type == iSLIP)
     iSLIP_Advance();
+  else if (arbit_type == PERFECT)
+    Perfect_Advance();
   else
     assert(0);
 }
 
+void xbar_router::Perfect_Advance() {
+  for (unsigned node_id = 0; node_id < total_nodes; node_id++) {
+    if (!in_buffers[node_id].empty()) {
+      Packet _packet = in_buffers[node_id].front();
+      if (Has_Buffer_Out(_packet.output_deviceID, 1)) {
+        out_buffers[_packet.output_deviceID].push(_packet);
+        in_buffers[node_id].pop();
+      }
+    }
+  }
+};
+
 void xbar_router::RR_Advance() {
   bool active = false;
   vector<bool> issued(total_nodes, false);
@@ -433,4 +447,4 @@ unsigned LocalInterconnect::GetFlitSize() const { return LOCAL_INCT_FLIT_SIZE; }
 
 void LocalInterconnect::DisplayState(FILE* fp) const {
   fprintf(fp, "GPGPU-Sim uArch: ICNT:Display State: Under implementation\n");
-}
+}
\ No newline at end of file