sending cache block byte mask

7 files changed, 91 insertions, 5 deletions

diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h
index 982e416..e09acdb 100644
--- a/src/abstract_hardware_model.h
+++ b/src/abstract_hardware_model.h
@@ -869,6 +869,12 @@ class mem_fetch_allocator {
   virtual mem_fetch *alloc(const class warp_inst_t &inst,
                            const mem_access_t &access,
                            unsigned long long cycle) const = 0;
+  virtual mem_fetch *alloc(new_addr_type addr, mem_access_type type,
+                            const active_mask_t &active_mask,
+                            const mem_access_byte_mask_t &byte_mask,
+                            const mem_access_sector_mask_t &sector_mask,
+                            unsigned size, bool wr,
+                            unsigned long long cycle) const = 0;                    
 };
 
 // the maximum number of destination, source, or address uarch operands in a
diff --git a/src/gpgpu-sim/gpu-cache.cc b/src/gpgpu-sim/gpu-cache.cc
index 763705f..ded8004 100644
--- a/src/gpgpu-sim/gpu-cache.cc
+++ b/src/gpgpu-sim/gpu-cache.cc
@@ -358,8 +358,13 @@ enum cache_request_status tag_array::access(new_addr_type addr, unsigned time,
       if (m_config.m_alloc_policy == ON_MISS) {
         if (m_lines[idx]->is_modified_line()) {
           wb = true;
+          ((sector_cache_block *)m_lines[idx])->set_byte_mask(mf);
           evicted.set_info(m_lines[idx]->m_block_addr,
-                           m_lines[idx]->get_modified_size());
+                           m_lines[idx]->get_modified_size(),
+                           ((sector_cache_block *)m_lines[idx])
+                                              ->get_byte_mask(),
+                            ((sector_cache_block *)m_lines[idx])
+                                              ->get_sector_mask());
         }
         m_lines[idx]->allocate(m_config.tag(addr), m_config.block_addr(addr),
                                time, mf->get_access_sector_mask());
@@ -1464,6 +1469,8 @@ enum cache_request_status data_cache::wr_miss_wa_lazy_fetch_on_read(
       m_tag_array->access(block_addr, time, cache_index, wb, evicted, mf);
   assert(m_status != HIT);
   cache_block_t *block = m_tag_array->get_block(cache_index);
+  block->set_status(MODIFIED, mf->get_access_sector_mask());
+  ((sector_cache_block *)block)->set_byte_mask(mf);
   if (m_status == HIT_RESERVED) {
     block->set_ignore_on_fill(true, mf->get_access_sector_mask());
     block->set_modified_on_fill(true, mf->get_access_sector_mask());
@@ -1484,8 +1491,10 @@ enum cache_request_status data_cache::wr_miss_wa_lazy_fetch_on_read(
     // (already modified lower level)
     if (wb && (m_config.m_write_policy != WRITE_THROUGH)) {
       mem_fetch *wb = m_memfetch_creator->alloc(
-          evicted.m_block_addr, m_wrbk_type, evicted.m_modified_size, true,
-          m_gpu->gpu_tot_sim_cycle + m_gpu->gpu_sim_cycle);
+        evicted.m_block_addr,m_wrbk_type,
+        mf->get_access_warp_mask(), evicted.m_byte_mask,
+        evicted.m_sector_mask, evicted.m_modified_size,
+        true, m_gpu->gpu_tot_sim_cycle + m_gpu->gpu_sim_cycle);
       // the evicted block may have wrong chip id when advanced L2 hashing  is
       // used, so set the right chip address from the original mf
       wb->set_chip(mf->get_tlx_addr().chip);
@@ -1560,8 +1569,10 @@ enum cache_request_status data_cache::rd_miss_base(
     // (already modified lower level)
     if (wb && (m_config.m_write_policy != WRITE_THROUGH)) {
       mem_fetch *wb = m_memfetch_creator->alloc(
-          evicted.m_block_addr, m_wrbk_type, evicted.m_modified_size, true,
-          m_gpu->gpu_tot_sim_cycle + m_gpu->gpu_sim_cycle);
+        evicted.m_block_addr,m_wrbk_type,
+        mf->get_access_warp_mask(), evicted.m_byte_mask,
+        evicted.m_sector_mask, evicted.m_modified_size,
+        true, m_gpu->gpu_tot_sim_cycle + m_gpu->gpu_sim_cycle);
       // the evicted block may have wrong chip id when advanced L2 hashing  is
       // used, so set the right chip address from the original mf
       wb->set_chip(mf->get_tlx_addr().chip);
diff --git a/src/gpgpu-sim/gpu-cache.h b/src/gpgpu-sim/gpu-cache.h
index 381ce94..042c1d6 100644
--- a/src/gpgpu-sim/gpu-cache.h
+++ b/src/gpgpu-sim/gpu-cache.h
@@ -72,14 +72,26 @@ enum cache_event_type {
 struct evicted_block_info {
   new_addr_type m_block_addr;
   unsigned m_modified_size;
+  mem_access_byte_mask_t m_byte_mask;
+  mem_access_sector_mask_t m_sector_mask;
   evicted_block_info() {
     m_block_addr = 0;
     m_modified_size = 0;
+    m_byte_mask.reset();
+    m_sector_mask.reset();
   }
   void set_info(new_addr_type block_addr, unsigned modified_size) {
     m_block_addr = block_addr;
     m_modified_size = modified_size;
   }
+  void set_info(new_addr_type block_addr, unsigned modified_size, 
+                mem_access_byte_mask_t byte_mask,
+                mem_access_sector_mask_t sector_mask) {
+    m_block_addr = block_addr;
+    m_modified_size = modified_size;
+    m_byte_mask = byte_mask;
+    m_sector_mask = sector_mask;
+  }
 };
 
 struct cache_event {
@@ -251,6 +263,7 @@ struct sector_cache_block : public cache_block_t {
     m_line_alloc_time = 0;
     m_line_last_access_time = 0;
     m_line_fill_time = 0;
+    m_byte_mask.reset();
   }
 
   virtual void allocate(new_addr_type tag, new_addr_type block_addr,
@@ -362,6 +375,20 @@ struct sector_cache_block : public cache_block_t {
     m_status[sidx] = status;
   }
 
+  virtual void set_byte_mask(mem_fetch *mf) {
+    m_byte_mask = m_byte_mask | mf->get_access_byte_mask();;
+  }
+  virtual mem_access_byte_mask_t get_byte_mask() {
+    return m_byte_mask;
+  }
+  virtual mem_access_sector_mask_t get_sector_mask() {
+    mem_access_sector_mask_t sector_mask;
+    for (unsigned i = 0; i < SECTOR_CHUNCK_SIZE; i++) {
+      if (m_status[i] == MODIFIED) 
+        sector_mask.set(i);
+    }
+    return sector_mask;
+  }
   virtual unsigned long long get_last_access_time() {
     return m_line_last_access_time;
   }
@@ -429,6 +456,7 @@ struct sector_cache_block : public cache_block_t {
   bool m_set_modified_on_fill[SECTOR_CHUNCK_SIZE];
   bool m_set_readable_on_fill[SECTOR_CHUNCK_SIZE];
   bool m_readable[SECTOR_CHUNCK_SIZE];
+  mem_access_byte_mask_t m_byte_mask;
 
   unsigned get_sector_index(mem_access_sector_mask_t sector_mask) {
     assert(sector_mask.count() == 1);
diff --git a/src/gpgpu-sim/l2cache.cc b/src/gpgpu-sim/l2cache.cc
index ab6e5c2..cd04af5 100644
--- a/src/gpgpu-sim/l2cache.cc
+++ b/src/gpgpu-sim/l2cache.cc
@@ -57,6 +57,20 @@ mem_fetch *partition_mf_allocator::alloc(new_addr_type addr,
   return mf;
 }
 
+mem_fetch *partition_mf_allocator::alloc(new_addr_type addr, 
+                            mem_access_type type,
+                            const active_mask_t &active_mask,
+                            const mem_access_byte_mask_t &byte_mask,
+                            const mem_access_sector_mask_t &sector_mask,
+                            unsigned size, bool wr,
+                            unsigned long long cycle) const {
+  mem_access_t access(type, addr, size, wr, active_mask, byte_mask, 
+                        sector_mask, m_memory_config->gpgpu_ctx);
+  mem_fetch *mf =
+    new mem_fetch(access, NULL, wr ? WRITE_PACKET_SIZE : READ_PACKET_SIZE, -1,
+                  -1, -1, m_memory_config, cycle);
+    return mf;
+}
 memory_partition_unit::memory_partition_unit(unsigned partition_id,
                                              const memory_config *config,
                                              class memory_stats_t *stats,
diff --git a/src/gpgpu-sim/l2cache.h b/src/gpgpu-sim/l2cache.h
index 3152db3..1f5d7c4 100644
--- a/src/gpgpu-sim/l2cache.h
+++ b/src/gpgpu-sim/l2cache.h
@@ -51,6 +51,12 @@ class partition_mf_allocator : public mem_fetch_allocator {
   virtual mem_fetch *alloc(new_addr_type addr, mem_access_type type,
                            unsigned size, bool wr,
                            unsigned long long cycle) const;
+  virtual mem_fetch *alloc(new_addr_type addr, mem_access_type type,
+                            const active_mask_t &active_mask,
+                            const mem_access_byte_mask_t &byte_mask,
+                            const mem_access_sector_mask_t &sector_mask,
+                            unsigned size, bool wr,
+                            unsigned long long cycle) const;
 
  private:
   const memory_config *m_memory_config;
diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc
index 4769ca8..4b4c98d 100644
--- a/src/gpgpu-sim/shader.cc
+++ b/src/gpgpu-sim/shader.cc
@@ -61,6 +61,21 @@ mem_fetch *shader_core_mem_fetch_allocator::alloc(
                     m_core_id, m_cluster_id, m_memory_config, cycle);
   return mf;
 }
+
+mem_fetch *shader_core_mem_fetch_allocator::alloc(
+  new_addr_type addr, mem_access_type type,
+  const active_mask_t &active_mask,
+  const mem_access_byte_mask_t &byte_mask,
+  const mem_access_sector_mask_t &sector_mask,
+  unsigned size, bool wr,
+  unsigned long long cycle) const {
+    mem_access_t access(type, addr, size, wr, active_mask, byte_mask, 
+                          sector_mask, m_memory_config->gpgpu_ctx);
+    mem_fetch *mf =
+      new mem_fetch(access, NULL, wr ? WRITE_PACKET_SIZE : READ_PACKET_SIZE, -1,
+                    m_core_id, m_cluster_id, m_memory_config, cycle);
+      return mf;
+  }
 /////////////////////////////////////////////////////////////////////////////
 
 std::list<unsigned> shader_core_ctx::get_regs_written(const inst_t &fvt) const {
diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h
index 8c02fd7..a7a2c02 100644
--- a/src/gpgpu-sim/shader.h
+++ b/src/gpgpu-sim/shader.h
@@ -1872,6 +1872,12 @@ class shader_core_mem_fetch_allocator : public mem_fetch_allocator {
   }
   mem_fetch *alloc(new_addr_type addr, mem_access_type type, unsigned size,
                    bool wr, unsigned long long cycle) const;
+  mem_fetch *alloc(new_addr_type addr, mem_access_type type,
+                            const active_mask_t &active_mask,
+                            const mem_access_byte_mask_t &byte_mask,
+                            const mem_access_sector_mask_t &sector_mask,
+                            unsigned size, bool wr,
+                            unsigned long long cycle) const;
   mem_fetch *alloc(const warp_inst_t &inst, const mem_access_t &access,
                    unsigned long long cycle) const {
     warp_inst_t inst_copy = inst;