add hw-realistic co-issue scheduling policies

1 files changed, 30 insertions, 3 deletions

diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h
index 92d6d33..4734d6a 100644
--- a/src/gpgpu-sim/shader.h
+++ b/src/gpgpu-sim/shader.h
@@ -604,6 +604,32 @@ class scheduler_unit {  // this can be copied freely, so can be used in std
                                    unsigned &available_sets,
                                    unsigned &next_free_set);
 
+  // K-bounded variant of try_utilization_max_coissue: scans only the
+  // first K warps in m_next_cycle_prioritized_warps (K = num_simd_sets).
+  // Hardware-realistic alternative to the all-warps idealization above.
+  void try_utilization_max_coissue_window(warp_inst_t *co_issue_composite,
+                                          exec_unit_type_t co_issue_fu_type,
+                                          unsigned co_issue_primary_warp_id,
+                                          unsigned &available_sets,
+                                          unsigned &next_free_set);
+
+  // Centered-window K-bounded variant: scans K warps centered on the
+  // primary issuer's index in m_supervised_warps (K/2 before, K/2 after).
+  // Models a HW priority encoder local to the primary's natural warp_id
+  // position rather than tied to scheduler priority.
+  void try_utilization_max_coissue_window_centered(
+      warp_inst_t *co_issue_composite, exec_unit_type_t co_issue_fu_type,
+      unsigned co_issue_primary_warp_id, unsigned &available_sets,
+      unsigned &next_free_set);
+
+  // All-before-primary K-bounded variant: scans (K-1) warps strictly
+  // before primary's index in m_supervised_warps. Asymmetric counterpart
+  // to the centered variant.
+  void try_utilization_max_coissue_window_before(
+      warp_inst_t *co_issue_composite, exec_unit_type_t co_issue_fu_type,
+      unsigned co_issue_primary_warp_id, unsigned &available_sets,
+      unsigned &next_free_set);
+
  protected:
   shd_warp_t &warp(int i);
 
@@ -1874,10 +1900,10 @@ class shader_core_config : public core_config {
               gpgpu_compaction_mode);
       abort();
     }
-    if (gpgpu_co_issue_priority > 4) {
+    if (gpgpu_co_issue_priority > 7) {
       fprintf(stderr,
               "GPGPU-Sim: invalid -gpgpu_co_issue_priority %u; must be "
-              "in [0,4]\n",
+              "in [0,7]\n",
               gpgpu_co_issue_priority);
       abort();
     }
@@ -2048,7 +2074,8 @@ class shader_core_config : public core_config {
   bool gpgpu_enable_compaction;
   unsigned gpgpu_compaction_mode;       // 0=none, 1=xor-static, 2=full
   unsigned gpgpu_opndcoll_read_latency; // extra cycles to collector
-  unsigned gpgpu_co_issue_priority;     // 0=greedy..4=same-PC
+  unsigned gpgpu_co_issue_priority;     // 0=greedy..5=K-bounded
+  unsigned gpgpu_coissue_window_K;      // mode-5 window size; 0 = num_simd_sets
   bool gpgpu_simd_partitioning_debug;   // verbose SIMD_SETS printf gate
   bool gpgpu_mem_coissue;               // enable MEM/LDST co-issue path
   unsigned gpgpu_scoreboard_mode;       // 0=legacy, 1=mask-aware, 2=slot-pinned