[TOGSim] Do not rescan the blocked head of the ready queue on every issue

TOGSim/include/Core.h

@@ -137,6 +137,9 @@ class Core {
   // to its accumulated footprint bytes (freed when its last reader issues).
   size_t _sram_used = 0;
   size_t _sram_capacity = 0;
+  // Free SA weight slots. With _sram_used it keys the issue scan's cursor: an
+  // instruction it walked past wakes only if one of the two grows.
+  int _weight_free = 0;
   std::unordered_map<int64_t, size_t> _sram_allocs;
 
   // SA weight-buffer throttle (sec 10.4). _weight_slots_used[s] = weights resident

TOGSim/include/Tile.h

@@ -39,6 +39,25 @@ class Tile : public std::enable_shared_from_this<Tile> {
   std::deque<std::shared_ptr<Instruction>>& get_instructions() { return _instructions; } 
   void enqueue_ready(const std::shared_ptr<Instruction>& inst) { _ready_queue.push_back(inst); }
   std::list<std::shared_ptr<Instruction>>& get_ready_instructions() { return _ready_queue; }
+
+  // Issue-scan cursor (Core::cycle): the scan only walks past instructions blocked on
+  // free spad bytes or a free weight slot, so it can resume where it left off.
+  using ReadyIt = std::list<std::shared_ptr<Instruction>>::iterator;
+  ReadyIt scan_from(size_t sram_used, int weight_free) {
+    if (_scan_blocked && sram_used >= _scan_sram_used && weight_free <= _scan_weight_free)
+      return std::next(_scan_blocked_at);
+    return _ready_queue.begin();   // a resource grew -> the prefix may issue now
+  }
+  void note_blocked(ReadyIt it, size_t sram_used, int weight_free) {
+    _scan_blocked_at = it;
+    _scan_blocked = true;
+    _scan_sram_used = sram_used;     // it could not issue with this much spad free...
+    _scan_weight_free = weight_free; // ...nor with this many weight slots free
+  }
+  // Kept across a rescan (the prefix blocks again once the freed resource is taken);
+  // it only dies with its instruction.
+  bool is_scan_cursor(ReadyIt it) const { return _scan_blocked && it == _scan_blocked_at; }
+  void drop_scan_cursor() { _scan_blocked = false; }
   void print();
   size_t nr_insts() { return _nr_insts; }
   size_t nr_finshed_insts() { return _nr_finished_insts; }
@@ -61,6 +80,12 @@ class Tile : public std::enable_shared_from_this<Tile> {
   size_t _nr_finished_insts = 0;
   std::deque<std::shared_ptr<Instruction>> _instructions;
   std::list<std::shared_ptr<Instruction>> _ready_queue;
+  // Only dereferenced while resources are no better than at note_blocked, and a
+  // still-blocked instruction is never erased.
+  ReadyIt _scan_blocked_at;
+  bool _scan_blocked = false;
+  size_t _scan_sram_used = 0;
+  int _scan_weight_free = 0;
   std::vector<std::shared_ptr<Tile>> _child_tiles;
   void *_custom_data=NULL;
   bool _stonne_tile=false;

TOGSim/src/Core.cc

@@ -24,6 +24,7 @@ Core::Core(uint32_t id, SimulationConfig config)
     exit(EXIT_FAILURE);
   }
   _weight_slots_used.resize(_num_systolic_array_per_core, 0);
+  _weight_free = _num_systolic_array_per_core * (int)_weight_slot_depth;
 }
 
 // Round-robin a systolic array that still has a free weight slot; -1 if all full
@@ -42,7 +43,7 @@ int Core::pick_free_weight_sa() {
 void Core::apply_due(const DueAction& a) {
   switch (a.kind) {
     case DueAction::FreeWeightSlot:
-      if (--a.token->refcount <= 0) { _weight_slots_used[a.token->sa]--; _issue_dirty = true; }  // weight slot freed -> re-arm
+      if (--a.token->refcount <= 0) { _weight_slots_used[a.token->sa]--; _weight_free++; _issue_dirty = true; }  // weight slot freed -> re-arm
       break;
     case DueAction::WakeBar: {
       auto bar = a.bar;            // async load data arrived -> fire its MEMORY_BAR
@@ -286,7 +287,8 @@ void Core::cycle() {
 
   for (int i=0; i<_tiles.size() && !issued; i++) {
     auto& instructions = _tiles[i]->get_ready_instructions();
-    for (auto it=instructions.begin(); it!=instructions.end();) {
+    // Resume after the prefix already known to be blocked (Tile::scan_from).
+    for (auto it=_tiles[i]->scan_from(_sram_used, _weight_free); it!=instructions.end();) {
       auto& inst = *it;
 
       switch (inst->get_opcode()) {
@@ -365,6 +367,7 @@ void Core::cycle() {
                   exit(EXIT_FAILURE);
                 }
                 _weight_slots_used[sa_idx]++;
+                _weight_free--;
                 auto tok = std::make_shared<WeightToken>(WeightToken{sa_idx, n_consumers});
                 for (auto& c : inst->get_deps(DepEvent::ISSUE))
                   if (c->get_compute_type() == MATMUL) {
@@ -408,6 +411,7 @@ void Core::cycle() {
               inst->finish_instruction();
               static_cast<Tile*>(inst->get_owner())->inc_finished_inst();
               _stat_tot_skipped_inst.at(static_cast<size_t>(inst->get_opcode()))++;
+              if (_tiles[i]->is_scan_cursor(it)) _tiles[i]->drop_scan_cursor();
               it = instructions.erase(it);   // erase returns the next iterator; the
               continue;                      // old code fell through to it++ on the
                                              // erased (invalidated) iterator -> UB
@@ -461,9 +465,12 @@ void Core::cycle() {
 
       if (issued) {
         _stat_inst_count.at(static_cast<size_t>(inst->get_opcode()))++;
+        if (_tiles[i]->is_scan_cursor(it)) _tiles[i]->drop_scan_cursor();
         instructions.erase(it);
         break;
       }
+      // Did not issue -> blocked on spad or a weight slot (the only two stalls).
+      _tiles[i]->note_blocked(it, _sram_used, _weight_free);
       it++;
     }
   }