add simple simd lane set partitioning

1 files changed, 33 insertions, 1 deletions

diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h
index e96ddc4..948e226 100644
--- a/src/abstract_hardware_model.h
+++ b/src/abstract_hardware_model.h
@@ -1109,6 +1109,24 @@ enum divergence_support_t {
 
 const unsigned MAX_ACCESSES_PER_INSN_PER_THREAD = 8;
 
+// SIMD lane partitioning: per-set execution data
+struct simd_set_info {
+  unsigned set_id;                      // Set index (0 to num_sets-1)
+  unsigned warp_id;                     // Owning warp
+  simt_mask_t set_active_mask;          // Active lanes within this set (lower set_width bits)
+  active_mask_t active_mask_in_warp;    // Active mask in original 32-bit warp thread domain
+  unsigned thread_ids[MAX_WARP_SIZE];   // Original thread IDs for each lane (max set_width used)
+  unsigned num_active_threads;          // popcount of set_active_mask
+  bool valid;                           // Whether this set has any active threads
+
+  simd_set_info()
+      : set_id(0), warp_id(0), num_active_threads(0), valid(false) {
+    set_active_mask.reset();
+    active_mask_in_warp.reset();
+    for (unsigned i = 0; i < MAX_WARP_SIZE; i++) thread_ids[i] = 0;
+  }
+};
+
 class warp_inst_t : public inst_t {
  public:
   // constructors
@@ -1152,12 +1170,23 @@ class warp_inst_t : public inst_t {
   void broadcast_barrier_reduction(const active_mask_t &access_mask);
   void do_atomic(bool forceDo = false);
   void do_atomic(const active_mask_t &access_mask, bool forceDo = false);
-  void clear() { m_empty = true; }
+  void clear() {
+    m_empty = true;
+    m_simd_sets.clear();
+  }
 
   void issue(const active_mask_t &mask, unsigned warp_id,
              unsigned long long cycle, int dynamic_warp_id, int sch_id,
              unsigned long long streamID);
 
+  // SIMD lane partitioning
+  void compute_simd_sets(unsigned num_sets, unsigned set_width);
+  const std::vector<simd_set_info> &get_simd_sets() const {
+    return m_simd_sets;
+  }
+  unsigned num_active_simd_sets() const;
+  bool has_simd_sets() const { return !m_simd_sets.empty(); }
+
   const active_mask_t &get_active_mask() const { return m_warp_active_mask; }
   void completed(unsigned long long cycle)
       const;  // stat collection: called when the instruction is completed
@@ -1333,6 +1362,9 @@ class warp_inst_t : public inst_t {
 
   unsigned m_scheduler_id;  // the scheduler that issues this inst
 
+  // SIMD lane partitioning data
+  std::vector<simd_set_info> m_simd_sets;
+
   // Jin: cdp support
  public:
   int m_is_cdp;