1 files changed, 34 insertions, 8 deletions

diff --git a/src/abstract_hardware_model.cc b/src/abstract_hardware_model.cc
index 7922bc5..e0e1d23 100644
--- a/src/abstract_hardware_model.cc
+++ b/src/abstract_hardware_model.cc
@@ -205,8 +205,8 @@ gpgpu_t::gpgpu_t(const gpgpu_functional_sim_config &config, gpgpu_context *ctx)
   gpu_tot_sim_cycle = 0;
 }
 
-address_type line_size_based_tag_func(new_addr_type address,
-                                      new_addr_type line_size) {
+new_addr_type line_size_based_tag_func(new_addr_type address,
+                                       new_addr_type line_size) {
   // gives the tag for an address based on a given line size
   return address & ~(line_size - 1);
 }
@@ -257,6 +257,7 @@ void warp_inst_t::do_atomic(bool forceDo) {
 
 void warp_inst_t::do_atomic(const active_mask_t &access_mask, bool forceDo) {
   assert(m_isatomic && (!m_empty || forceDo));
+  if (!should_do_atomic) return;
   for (unsigned i = 0; i < m_config->warp_size; i++) {
     if (access_mask.test(i)) {
       dram_callback_t &cb = m_per_scalar_thread[i].callback;
@@ -447,7 +448,7 @@ void warp_inst_t::generate_mem_accesses() {
     for (unsigned thread = 0; thread < m_config->warp_size; thread++) {
       if (!active(thread)) continue;
       new_addr_type addr = m_per_scalar_thread[thread].memreqaddr[0];
-      unsigned block_address = line_size_based_tag_func(addr, cache_block_size);
+      new_addr_type block_address = line_size_based_tag_func(addr, cache_block_size);
       accesses[block_address].set(thread);
       unsigned idx = addr - block_address;
       for (unsigned i = 0; i < data_size; i++) byte_mask.set(idx + i);
@@ -529,20 +530,37 @@ void warp_inst_t::memory_coalescing_arch(bool is_write,
            (m_per_scalar_thread[thread].memreqaddr[access] != 0);
            access++) {
         new_addr_type addr = m_per_scalar_thread[thread].memreqaddr[access];
-        unsigned block_address = line_size_based_tag_func(addr, segment_size);
+        new_addr_type block_address = line_size_based_tag_func(addr, segment_size);
         unsigned chunk =
             (addr & 127) / 32;  // which 32-byte chunk within in a 128-byte
                                 // chunk does this thread access?
         transaction_info &info = subwarp_transactions[block_address];
 
         // can only write to one segment
-        assert(block_address == line_size_based_tag_func(
-                                    addr + data_size_coales - 1, segment_size));
+        // it seems like in trace driven, a thread can write to more than one
+        // segment assert(block_address ==
+        // line_size_based_tag_func(addr+data_size_coales-1,segment_size));
 
         info.chunks.set(chunk);
         info.active.set(thread);
         unsigned idx = (addr & 127);
-        for (unsigned i = 0; i < data_size_coales; i++) info.bytes.set(idx + i);
+        for (unsigned i = 0; i < data_size_coales; i++)
+          if ((idx + i) < MAX_MEMORY_ACCESS_SIZE) info.bytes.set(idx + i);
+
+        // it seems like in trace driven, a thread can write to more than one
+        // segment handle this special case
+        if (block_address != line_size_based_tag_func(
+                                 addr + data_size_coales - 1, segment_size)) {
+          addr = addr + data_size_coales - 1;
+          new_addr_type block_address = line_size_based_tag_func(addr, segment_size);
+          unsigned chunk = (addr & 127) / 32;
+          transaction_info &info = subwarp_transactions[block_address];
+          info.chunks.set(chunk);
+          info.active.set(thread);
+          unsigned idx = (addr & 127);
+          for (unsigned i = 0; i < data_size_coales; i++)
+            if ((idx + i) < MAX_MEMORY_ACCESS_SIZE) info.bytes.set(idx + i);
+        }
       }
     }
 
@@ -607,7 +625,7 @@ void warp_inst_t::memory_coalescing_arch_atomic(bool is_write,
       if (!active(thread)) continue;
 
       new_addr_type addr = m_per_scalar_thread[thread].memreqaddr[0];
-      unsigned block_address = line_size_based_tag_func(addr, segment_size);
+      new_addr_type block_address = line_size_based_tag_func(addr, segment_size);
       unsigned chunk =
           (addr & 127) / 32;  // which 32-byte chunk within in a 128-byte chunk
                               // does this thread access?
@@ -746,6 +764,10 @@ kernel_info_t::kernel_info_t(dim3 gridDim, dim3 blockDim,
   // Jin: launch latency management
   m_launch_latency = entry->gpgpu_ctx->device_runtime->g_kernel_launch_latency;
 
+  m_kernel_TB_latency =
+      entry->gpgpu_ctx->device_runtime->g_kernel_launch_latency +
+      num_blocks() * entry->gpgpu_ctx->device_runtime->g_TB_launch_latency;
+
   cache_config_set = false;
 }
 
@@ -773,6 +795,10 @@ kernel_info_t::kernel_info_t(
   // Jin: launch latency management
   m_launch_latency = entry->gpgpu_ctx->device_runtime->g_kernel_launch_latency;
 
+  m_kernel_TB_latency =
+      entry->gpgpu_ctx->device_runtime->g_kernel_launch_latency +
+      num_blocks() * entry->gpgpu_ctx->device_runtime->g_TB_launch_latency;
+
   cache_config_set = false;
   m_NameToCudaArray = nameToCudaArray;
   m_NameToTextureInfo = nameToTextureInfo;