impl option w/ only 2 SBs; stall so SBs are cleared before div or reconv

1 files changed, 27 insertions, 1 deletions

diff --git a/src/gpgpu-sim/scoreboard.h b/src/gpgpu-sim/scoreboard.h
index 660587e..da34af5 100644
--- a/src/gpgpu-sim/scoreboard.h
+++ b/src/gpgpu-sim/scoreboard.h
@@ -32,6 +32,7 @@
 #include <set>
 #include <utility>
 #include <vector>
+#include <array>
 #include "assert.h"
 
 #ifndef SCOREBOARD_H_
@@ -83,8 +84,29 @@ class Scoreboard {
   // Per-set release dispatcher used at writeback. In mode 0, picks
   // primary vs secondary release based on `is_intra_legacy`. In mode 1,
   // releases (reg, mask) from mask_reg_table; `is_intra_legacy` ignored.
+  // In mode 2, releases (reg) from slot_reg_table[wid][slot_id].
   void releaseSetReg(unsigned wid, unsigned reg, const active_mask_t &mask,
-                     bool is_intra_legacy);
+                     bool is_intra_legacy, unsigned slot_id = 0);
+
+  // Mode 2: slot-pinned per-half scoreboards. Each warp has two
+  // independent scoreboards (one per ibuffer half). Hazard tracking
+  // ignores masks since slot ownership guarantees lane-disjoint
+  // operation between the two halves. See
+  // for_claude/docs/scoreboard_redesign_plan.md.
+  static const unsigned NUM_SLOTS = 2;
+  void reserveRegisterSlot(unsigned wid, unsigned slot, unsigned regnum);
+  void releaseRegisterSlot(unsigned wid, unsigned slot, unsigned regnum);
+  bool checkCollisionSlot(unsigned wid, unsigned slot,
+                          const inst_t *inst) const;
+  bool pendingWritesSlot(unsigned wid, unsigned slot) const {
+    return !slot_reg_table[wid][slot].empty();
+  }
+  bool slotClean(unsigned wid, unsigned slot) const {
+    return slot_reg_table[wid][slot].empty();
+  }
+  bool bothSlotsClean(unsigned wid) const {
+    return slot_reg_table[wid][0].empty() && slot_reg_table[wid][1].empty();
+  }
 
  public:
   // Accounting / leak-detection (called from gpu-sim end-of-run dump).
@@ -113,6 +135,10 @@ class Scoreboard {
   // Secondary path register tracking (for I-Buffer half 1, intra-warp co-issue)
   std::vector<std::set<unsigned> > sec_reg_table;
 
+  // Mode 2: per-warp, per-slot register sets.
+  // slot_reg_table[wid][slot] = set of reserved register numbers.
+  std::vector<std::array<std::set<unsigned>, NUM_SLOTS> > slot_reg_table;
+
   // Mode 1: per-warp list of mask-aware reservations.
   struct mask_resv {
     unsigned reg;