critical fix - add option to have masks for SB entries

1 files changed, 48 insertions, 2 deletions

diff --git a/src/gpgpu-sim/scoreboard.h b/src/gpgpu-sim/scoreboard.h
index 3dbb1ce..660587e 100644
--- a/src/gpgpu-sim/scoreboard.h
+++ b/src/gpgpu-sim/scoreboard.h
@@ -41,7 +41,12 @@
 
 class Scoreboard {
  public:
-  Scoreboard(unsigned sid, unsigned n_warps, class gpgpu_t *gpu);
+  // mode: 0 = legacy (primary + secondary tables), 1 = mask-aware,
+  // 2 = slot-pinned. See scoreboard_redesign_plan.md.
+  Scoreboard(unsigned sid, unsigned n_warps, class gpgpu_t *gpu,
+             unsigned mode = 0);
+
+  unsigned mode() const { return m_mode; }
 
   void reserveRegisters(const warp_inst_t *inst);
   void releaseRegisters(const warp_inst_t *inst);
@@ -52,12 +57,35 @@ class Scoreboard {
   void printContents() const;
   const bool islongop(unsigned warp_id, unsigned regnum);
 
-  // Secondary scoreboard for intra-warp co-issue (split-level tracking)
+  // Secondary scoreboard for intra-warp co-issue (split-level tracking).
+  // Mode 0 (legacy) only — bypassed under modes 1 and 2.
   void reserveRegisterSecondary(unsigned wid, unsigned regnum);
   void releaseRegisterSecondary(unsigned wid, unsigned regnum);
   bool checkCollisionSecondary(unsigned wid, const inst_t *inst) const;
   void clearSecondary(unsigned wid);
 
+  // Mode 1: mask-aware single scoreboard.
+  // Each reservation entry tracks (reg, active_mask, ref_count). Hazard
+  // checks return true iff any of the inst's read/written regs has an
+  // entry whose mask overlaps the candidate's active_mask. See
+  // for_claude/docs/scoreboard_redesign_plan.md for design.
+  void reserveRegisterMask(unsigned wid, unsigned reg,
+                           const active_mask_t &mask);
+  void releaseRegisterMask(unsigned wid, unsigned reg,
+                           const active_mask_t &mask);
+  bool checkCollisionMask(unsigned wid, const inst_t *inst,
+                          const active_mask_t &mask) const;
+  // Mask-agnostic helper: any in-flight writes for this warp under the
+  // active scoreboard table. In mode 1 this checks mask_reg_table; in
+  // mode 0 it falls through to legacy reg_table emptiness.
+  bool pendingWritesAny(unsigned wid) const;
+
+  // 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.
+  void releaseSetReg(unsigned wid, unsigned reg, const active_mask_t &mask,
+                     bool is_intra_legacy);
+
  public:
   // Accounting / leak-detection (called from gpu-sim end-of-run dump).
   // Aggregates reserve/release counts, reservation-duration histogram,
@@ -74,6 +102,7 @@ class Scoreboard {
   int get_sid() const { return m_sid; }
 
   unsigned m_sid;
+  unsigned m_mode;  // 0=legacy, 1=mask-aware, 2=slot-pinned
 
   // keeps track of pending writes to registers
   // indexed by warp id, reg_id => pending write count
@@ -84,6 +113,23 @@ class Scoreboard {
   // Secondary path register tracking (for I-Buffer half 1, intra-warp co-issue)
   std::vector<std::set<unsigned> > sec_reg_table;
 
+  // Mode 1: per-warp list of mask-aware reservations.
+  struct mask_resv {
+    unsigned reg;
+    active_mask_t mask;
+    unsigned ref_count;
+    unsigned long long resv_cycle;
+  };
+  std::vector<std::vector<mask_resv> > mask_reg_table;
+
+  // Mode 1 accounting (used by dumpAccounting + dump_warp_state for leak
+  // detection on the deadlock path).
+  unsigned long long m_mask_reserve_inserts;       // new entry pushes
+  unsigned long long m_mask_reserve_inc_refcount;  // ref_count++ on existing
+  unsigned long long m_mask_release_match;          // found and decremented
+  unsigned long long m_mask_release_nomatch;        // no matching entry
+  unsigned long long m_mask_release_erase;          // ref_count hit 0 + erased
+
   class gpgpu_t *m_gpu;
 
   // ---- Accounting (per-shader, scalar; cheap to update) ----