co-issue global mem requests

1 files changed, 21 insertions, 2 deletions

diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h
index 698c896..1e3c94e 100644
--- a/src/abstract_hardware_model.h
+++ b/src/abstract_hardware_model.h
@@ -859,6 +859,14 @@ class mem_access_t {
   mem_access_byte_mask_t get_byte_mask() const { return m_byte_mask; }
   mem_access_sector_mask_t get_sector_mask() const { return m_sector_mask; }
 
+  // Source-warp attribution for MEM co-issue composites. Unset by default;
+  // when set, overrides the composite's primary warp_id for mem_fetch
+  // allocation + pending_writes tracking + writeback scoreboard release.
+  unsigned get_source_wid() const { return m_source_wid; }
+  unsigned get_source_split_id() const { return m_source_split_id; }
+  void set_source_wid(unsigned w) { m_source_wid = w; }
+  void set_source_split_id(unsigned s) { m_source_split_id = s; }
+
   void print(FILE *fp) const {
     fprintf(fp, "addr=0x%llx, %s, size=%u, ", m_addr,
             m_write ? "store" : "load ", m_req_size);
@@ -909,6 +917,8 @@ class mem_access_t {
   active_mask_t m_warp_mask;
   mem_access_byte_mask_t m_byte_mask;
   mem_access_sector_mask_t m_sector_mask;
+  unsigned m_source_wid;       // MEM co-issue: originating warp (-1 = unset)
+  unsigned m_source_split_id;  // MEM co-issue: originating split (-1 = unset)
 };
 
 class mem_fetch;
@@ -1113,6 +1123,7 @@ const unsigned MAX_ACCESSES_PER_INSN_PER_THREAD = 8;
 struct simd_set_info {
   unsigned set_id;                      // Set index (0 to num_sets-1)
   unsigned warp_id;                     // Owning warp
+  unsigned split_id;                    // Owning split (-1 = inter-warp / primary)
   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)
@@ -1121,8 +1132,8 @@ struct simd_set_info {
   const inst_t *source_inst;            // Pointer to static decoded instruction (for MIMD)
 
   simd_set_info()
-      : set_id(0), warp_id(0), num_active_threads(0), valid(false),
-        source_inst(NULL) {
+      : set_id(0), warp_id(0), split_id((unsigned)-1),
+        num_active_threads(0), valid(false), source_inst(NULL) {
     set_active_mask.reset();
     active_mask_in_warp.reset();
     for (unsigned i = 0; i < MAX_WARP_SIZE; i++) thread_ids[i] = 0;
@@ -1198,6 +1209,7 @@ class warp_inst_t : public inst_t {
   bool has_simd_sets() const { return !m_simd_sets.empty(); }
   void merge_simd_sets(const std::vector<simd_set_info> &other_sets);
   void set_source_inst_on_sets(const inst_t *src);
+  void set_split_id_on_sets(unsigned split_id);
   static bool simd_sets_overlap(const std::vector<simd_set_info> &a,
                                 const std::vector<simd_set_info> &b);
 
@@ -1323,6 +1335,13 @@ class warp_inst_t : public inst_t {
   unsigned accessq_count() const { return m_accessq.size(); }
   const mem_access_t &accessq_back() { return m_accessq.back(); }
   void accessq_pop_back() { m_accessq.pop_back(); }
+  // Const iteration over the access queue for per-source grouping (MEM co-issue).
+  std::list<mem_access_t>::const_iterator accessq_begin() const {
+    return m_accessq.begin();
+  }
+  std::list<mem_access_t>::const_iterator accessq_end() const {
+    return m_accessq.end();
+  }
 
   bool dispatch_delay() {
     if (cycles > 0) cycles--;