Use cache config multipilier when possible

4 files changed, 19 insertions, 20 deletions

diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h
index bd10a93..dbe138a 100644
--- a/src/abstract_hardware_model.h
+++ b/src/abstract_hardware_model.h
@@ -374,7 +374,6 @@ class core_config {
   unsigned mem_warp_parts;
   mutable unsigned gpgpu_shmem_size;
   char *gpgpu_shmem_option;
-  unsigned gpgpu_unified_l1d_size;
   unsigned gpgpu_shmem_sizeDefault;
   unsigned gpgpu_shmem_sizePrefL1;
   unsigned gpgpu_shmem_sizePrefShared;
diff --git a/src/gpgpu-sim/gpu-cache.h b/src/gpgpu-sim/gpu-cache.h
index 0162b6c..87a6b13 100644
--- a/src/gpgpu-sim/gpu-cache.h
+++ b/src/gpgpu-sim/gpu-cache.h
@@ -512,6 +512,14 @@ class cache_config {
       exit_parse_error();
     }
 
+    // set * assoc * cacheline size. Then convert Byte to KB
+    unsigned original_size = m_nset * m_assoc * m_line_sz / 1024;
+    if (m_unified_cache_size > 0) {
+      max_cache_multiplier = m_unified_cache_size / original_size;
+    } else {
+      max_cache_multiplier = MAX_DEFAULT_CACHE_SIZE_MULTIBLIER;
+    }
+    
     switch (ct) {
       case 'N':
         m_cache_type = NORMAL;
@@ -588,7 +596,7 @@ class cache_config {
       // https://ieeexplore.ieee.org/document/8344474/
       m_is_streaming = true;
       m_alloc_policy = ON_FILL;
-      m_mshr_entries = m_nset * m_assoc * MAX_DEFAULT_CACHE_SIZE_MULTIBLIER;
+      m_mshr_entries = m_nset * m_assoc * max_cache_multiplier;
       if (m_cache_type == SECTOR) m_mshr_entries *= SECTOR_CHUNCK_SIZE;
       m_mshr_max_merge = MAX_WARP_PER_SM;
     }
@@ -616,7 +624,6 @@ class cache_config {
     m_atom_sz = (m_cache_type == SECTOR) ? SECTOR_SIZE : m_line_sz;
     m_sector_sz_log2 = LOGB2(SECTOR_SIZE);
     original_m_assoc = m_assoc;
-    original_sz = m_nset * original_m_assoc * m_line_sz;
 
 
     // For more details about difference between FETCH_ON_WRITE and WRITE
@@ -706,19 +713,13 @@ class cache_config {
   }
   unsigned get_max_num_lines() const {
     assert(m_valid);
-    return MAX_DEFAULT_CACHE_SIZE_MULTIBLIER * m_nset * original_m_assoc;
+    // gpgpu_unified_cache_size is in KB while original_sz is in B
+    return max_cache_multiplier * m_nset * original_m_assoc;
   }
   unsigned get_max_assoc() const {
     assert(m_valid);
-    return MAX_DEFAULT_CACHE_SIZE_MULTIBLIER * original_m_assoc;
-  }
-  unsigned get_original_assoc() const {
-    assert(m_valid);
-    return original_m_assoc;
-  }
-  unsigned get_original_sz() const {
-    assert(m_valid);
-    return original_sz;
+    // gpgpu_unified_cache_size is in KB while original_sz is in B
+    return max_cache_multiplier * original_m_assoc;
   }
   void print(FILE *fp) const {
     fprintf(fp, "Size = %d B (%d Set x %d-way x %d byte line)\n",
@@ -766,6 +767,7 @@ class cache_config {
   char *m_config_stringPrefShared;
   FuncCache cache_status;
   unsigned m_wr_percent;
+  unsigned m_unified_cache_size;
   write_allocate_policy_t get_write_allocate_policy() {
     return m_write_alloc_policy;
   }
@@ -787,8 +789,8 @@ class cache_config {
   unsigned m_atom_sz;
   unsigned m_sector_sz_log2;
   unsigned original_m_assoc;
-  unsigned original_sz;
   bool m_is_streaming;
+  unsigned max_cache_multiplier;
 
   enum replacement_policy_t m_replacement_policy;  // 'L' = LRU, 'F' = FIFO
   enum write_policy_t
diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc
index a2aa929..df30047 100644
--- a/src/gpgpu-sim/gpu-sim.cc
+++ b/src/gpgpu-sim/gpu-sim.cc
@@ -331,7 +331,7 @@ void shader_core_config::reg_options(class OptionParser *opp) {
       opp, "-gpgpu_shmem_option", OPT_CSTR, &gpgpu_shmem_option,
       "Option list of shared memory sizes", "0");
   option_parser_register(
-      opp, "-gpgpu_unified_l1d_size", OPT_UINT32, &gpgpu_unified_l1d_size,
+      opp, "-gpgpu_unified_l1d_size", OPT_UINT32, &m_L1D_config.m_unified_cache_size,
       "Size of unified data cache(L1D + shared memory) in KB", "0");
   option_parser_register(opp, "-gpgpu_adaptive_cache_config", OPT_BOOL,
                          &adaptive_cache_config, "adaptive_cache_config", "0");
diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc
index 141c700..3efef2b 100644
--- a/src/gpgpu-sim/shader.cc
+++ b/src/gpgpu-sim/shader.cc
@@ -3315,7 +3315,8 @@ unsigned int shader_core_config::max_cta(const kernel_info_t &k) const {
     }
 
     unsigned total_shmem = kernel_info->smem * result;
-    unsigned total_unified = gpgpu_unified_l1d_size * 1024;
+    // Unified cache config is in KB. Converting to B
+    unsigned total_unified = m_L1D_config.m_unified_cache_size * 1024;
     std::sort(shmem_list.begin(), shmem_list.end());
 
     assert(total_shmem >= 0 && total_shmem <= shmem_list.back());
@@ -3325,10 +3326,7 @@ unsigned int shader_core_config::max_cta(const kernel_info_t &k) const {
       case ADAPTIVE_CACHE: {
         // For more info about adaptive cache, see
         bool l1d_configured = false;
-        unsigned l1_defined = m_L1D_config.get_original_sz() / 1024;
-        assert(gpgpu_unified_l1d_size % l1_defined == 0);
-        unsigned max_assoc = m_L1D_config.get_original_assoc() * 
-          gpgpu_unified_l1d_size / l1_defined;
+        unsigned max_assoc = m_L1D_config.get_max_assoc();
 
         if (total_shmem == 0) {
           m_L1D_config.set_assoc(max_assoc);