Merge pull request #117 from PSuchita/dev

Texture + Stream Bug Fix 

7 files changed, 104 insertions, 17 deletions

diff --git a/src/abstract_hardware_model.cc b/src/abstract_hardware_model.cc
index 307df40..cebdb25 100644
--- a/src/abstract_hardware_model.cc
+++ b/src/abstract_hardware_model.cc
@@ -88,8 +88,6 @@ void checkpoint::load_global_mem(class memory_space *temp_mem, char * f1name)
       fclose ( fp2 );
 }
 
-
-
 void checkpoint::store_global_mem(class memory_space * mem, char *fname, char * format)
 {
 
@@ -192,6 +190,11 @@ gpgpu_t::gpgpu_t( const gpgpu_functional_sim_config &config )
    checkpoint_CTA_t = m_function_model_config.get_checkpoint_CTA_t();
    checkpoint_insn_Y = m_function_model_config.get_checkpoint_insn_Y();
 
+   // initialize texture mappings to empty
+   m_NameToTextureInfo.clear();
+   m_NameToCudaArray.clear();
+   m_TextureRefToName.clear();
+   m_NameToAttribute.clear();
 
    if(m_function_model_config.get_ptx_inst_debug_to_file() != 0) 
       ptx_inst_debug_file = fopen(m_function_model_config.get_ptx_inst_debug_file(), "w");
@@ -688,7 +691,10 @@ unsigned g_kernel_launch_latency;
 
 unsigned kernel_info_t::m_next_uid = 1;
 
-kernel_info_t::kernel_info_t( dim3 gridDim, dim3 blockDim, class function_info *entry )
+/*A snapshot of the texture mappings needs to be stored in the kernel's info as 
+kernels should use the texture bindings seen at the time of launch and textures
+ can be bound/unbound asynchronously with respect to streams. */
+kernel_info_t::kernel_info_t( dim3 gridDim, dim3 blockDim, class function_info *entry, std::map<std::string, const struct cudaArray*> nameToCudaArray, std::map<std::string, const struct textureInfo*> nameToTextureInfo)   
 {
     m_kernel_entry=entry;
     m_grid_dim=gridDim;
@@ -708,6 +714,8 @@ kernel_info_t::kernel_info_t( dim3 gridDim, dim3 blockDim, class function_info *
     m_launch_latency = g_kernel_launch_latency;
 
     volta_cache_config_set=false;
+    m_NameToCudaArray = nameToCudaArray;
+    m_NameToTextureInfo = nameToTextureInfo;
 }
 
 kernel_info_t::~kernel_info_t()
diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h
index 28a22e0..e8716ab 100644
--- a/src/abstract_hardware_model.h
+++ b/src/abstract_hardware_model.h
@@ -212,7 +212,7 @@ public:
 //      m_num_cores_running=0;
 //      m_param_mem=NULL;
 //   }
-   kernel_info_t( dim3 gridDim, dim3 blockDim, class function_info *entry );
+   kernel_info_t( dim3 gridDim, dim3 blockDim, class function_info *entry, std::map<std::string, const struct cudaArray*> nameToCudaArray, std::map<std::string, const struct textureInfo*> nameToTextureInfo);
    ~kernel_info_t();
 
    void inc_running() { m_num_cores_running++; }
@@ -275,6 +275,23 @@ public:
    std::list<class ptx_thread_info *> &active_threads() { return m_active_threads; }
    class memory_space *get_param_memory() { return m_param_mem; }
 
+   
+   //The following functions access texture bindings present at the kernel's launch
+   
+   const struct cudaArray* get_texarray( const std::string &texname ) const
+   {
+      std::map<std::string,const struct cudaArray*>::const_iterator t=m_NameToCudaArray.find(texname);
+      assert(t != m_NameToCudaArray.end());
+      return t->second;
+   }
+
+   const struct textureInfo* get_texinfo( const std::string &texname ) const
+   {
+      std::map<std::string, const struct textureInfo*>::const_iterator t=m_NameToTextureInfo.find(texname);
+      assert(t != m_NameToTextureInfo.end());
+      return t->second;
+   }
+
 private:
    kernel_info_t( const kernel_info_t & ); // disable copy constructor
    void operator=( const kernel_info_t & ); // disable copy operator
@@ -283,6 +300,10 @@ private:
 
    unsigned m_uid;
    static unsigned m_next_uid;
+   
+   //These maps contain the snapshot of the texture mappings at kernel launch
+   std::map<std::string, const struct cudaArray*> m_NameToCudaArray;
+   std::map<std::string, const struct textureInfo*> m_NameToTextureInfo;
 
    dim3 m_grid_dim;
    dim3 m_block_dim;
@@ -580,8 +601,8 @@ public:
 
     const struct textureInfo* get_texinfo( const std::string &texname ) const
     {
-        std::map<std::string, const struct textureInfo*>::const_iterator t=m_NameToTexureInfo.find(texname);
-        assert(t != m_NameToTexureInfo.end());
+        std::map<std::string, const struct textureInfo*>::const_iterator t=m_NameToTextureInfo.find(texname);
+        assert(t != m_NameToTextureInfo.end());
         return t->second;
     }
 
@@ -594,6 +615,10 @@ public:
 
     const gpgpu_functional_sim_config &get_config() const { return m_function_model_config; }
     FILE* get_ptx_inst_debug_file() { return ptx_inst_debug_file; }
+    
+    //  These maps return the current texture mappings for the GPU at any given time.
+    std::map<std::string, const struct cudaArray*> getNameArrayMapping() {return m_NameToCudaArray;}
+    std::map<std::string, const struct textureInfo*> getNameInfoMapping() {return m_NameToTextureInfo;}
 
 protected:
     const gpgpu_functional_sim_config &m_function_model_config;
@@ -604,11 +629,11 @@ protected:
     class memory_space *m_surf_mem;
 
     unsigned long long m_dev_malloc;
-    
+    //  These maps contain the current texture mappings for the GPU at any given time. 
     std::map<std::string, std::set<const struct textureReference*> > m_NameToTextureRef;
     std::map<const struct textureReference*, std::string> m_TextureRefToName;
     std::map<std::string, const struct cudaArray*> m_NameToCudaArray;
-    std::map<std::string, const struct textureInfo*> m_NameToTexureInfo;
+    std::map<std::string, const struct textureInfo*> m_NameToTextureInfo;
     std::map<std::string, const struct textureReferenceAttr*> m_NameToAttribute;
 };
 
diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc
index ec68b5b..f7bb9cc 100644
--- a/src/cuda-sim/cuda-sim.cc
+++ b/src/cuda-sim/cuda-sim.cc
@@ -218,7 +218,7 @@ void gpgpu_t::gpgpu_ptx_sim_bindTextureToArray(const struct textureReference* te
    texInfo->Ty_numbits = intLOGB2(Ty);
    texInfo->texel_size = texel_size;
    texInfo->texel_size_numbits = intLOGB2(texel_size);
-   m_NameToTexureInfo[texname] = texInfo;
+   m_NameToTextureInfo[texname] = texInfo;
 }
 
 void gpgpu_t::gpgpu_ptx_sim_unbindTexture(const struct textureReference* texref)
@@ -226,7 +226,7 @@ void gpgpu_t::gpgpu_ptx_sim_unbindTexture(const struct textureReference* texref)
    //assumes bind-use-unbind-bind-use-unbind pattern
    std::string texname = gpgpu_ptx_sim_findNamefromTexture(texref);
    m_NameToCudaArray.erase(texname);
-   m_NameToTexureInfo.erase(texname);
+   m_NameToTextureInfo.erase(texname);
 }
 
 unsigned g_assemble_code_next_pc=0; 
@@ -1507,7 +1507,15 @@ static unsigned get_tex_datasize( const ptx_instruction *pI, ptx_thread_info *th
    std::string texname = src1.name();
 
    gpgpu_t *gpu = thread->get_gpu();
-   const struct textureInfo* texInfo = gpu->get_texinfo(texname);
+   /*
+     For programs with many streams, textures can be bound and unbound
+     asynchronously.  This means we need to use the kernel's "snapshot" of
+     the state of the texture mappings when it was launched (so that we
+     don't try to access the incorrect texture mapping if it's been updated,
+     or that we don't access a mapping that has been unbound).
+    */
+   kernel_info_t& k = thread->get_kernel();
+   const struct textureInfo* texInfo = k.get_texinfo(texname);
 
    unsigned data_size = texInfo->texel_size;
    return data_size; 
@@ -1918,7 +1926,7 @@ kernel_info_t *gpgpu_opencl_ptx_sim_init_grid(class function_info *entry,
                                              struct dim3 blockDim,
                                              gpgpu_t *gpu )
 {
-   kernel_info_t *result = new kernel_info_t(gridDim,blockDim,entry);
+   kernel_info_t *result = new kernel_info_t(gridDim,blockDim,entry,gpu->getNameArrayMapping(),gpu->getNameInfoMapping());
    unsigned argcount=args.size();
    unsigned argn=1;
    for( gpgpu_ptx_sim_arg_list_t::iterator a = args.begin(); a != args.end(); a++ ) {
diff --git a/src/cuda-sim/cuda_device_runtime.cc b/src/cuda-sim/cuda_device_runtime.cc
index b399133..917e7a8 100644
--- a/src/cuda-sim/cuda_device_runtime.cc
+++ b/src/cuda-sim/cuda_device_runtime.cc
@@ -198,7 +198,8 @@ void gpgpusim_cuda_launchDeviceV2(const ptx_instruction * pI, ptx_thread_info *
             memory_space *device_kernel_param_mem;
 
             //create child kernel_info_t and index it with parameter_buffer address
-            device_grid = new kernel_info_t(config.grid_dim, config.block_dim, device_kernel_entry);
+	    gpgpu_t* gpu=thread->get_gpu();
+            device_grid = new kernel_info_t(config.grid_dim, config.block_dim, device_kernel_entry, gpu->getNameArrayMapping(), gpu->getNameInfoMapping());
             device_grid->launch_cycle = gpu_sim_cycle + gpu_tot_sim_cycle;
             kernel_info_t & parent_grid = thread->get_kernel();
             DEV_RUNTIME_REPORT("child kernel launched by " << parent_grid.name() << ", cta (" <<
diff --git a/src/cuda-sim/instructions.cc b/src/cuda-sim/instructions.cc
index c85654c..11001d7 100644
--- a/src/cuda-sim/instructions.cc
+++ b/src/cuda-sim/instructions.cc
@@ -5258,11 +5258,18 @@ void tex_impl( const ptx_instruction *pI, ptx_thread_info *thread )
    if (!ptx_tex_regs) ptx_tex_regs = new ptx_reg_t[4];
    unsigned nelem = src2.get_vect_nelem();
    thread->get_vector_operand_values(src2, ptx_tex_regs, nelem); //ptx_reg should be 4 entry vector type...coordinates into texture
-
+   /*
+     For programs with many streams, textures can be bound and unbound
+     asynchronously.  This means we need to use the kernel's "snapshot" of
+     the state of the texture mappings when it was launched (so that we
+     don't try to access the incorrect texture mapping if it's been updated,
+     or that we don't access a mapping that has been unbound).
+   */
    gpgpu_t *gpu = thread->get_gpu();
+   kernel_info_t &k = thread->get_kernel();
    const struct textureReference* texref = gpu->get_texref(texname);
-   const struct cudaArray* cuArray = gpu->get_texarray(texname); 
-   const struct textureInfo* texInfo = gpu->get_texinfo(texname);
+   const struct cudaArray* cuArray = k.get_texarray(texname); 
+   const struct textureInfo* texInfo = k.get_texinfo(texname);
    const struct textureReferenceAttr* texAttr = gpu->get_texattr(texname);
 
    //assume always 2D f32 input
diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc
index ec570bf..3687448 100644
--- a/src/gpgpu-sim/gpu-sim.cc
+++ b/src/gpgpu-sim/gpu-sim.cc
@@ -499,6 +499,12 @@ void gpgpu_sim_config::reg_options(option_parser_t opp)
    option_parser_register(opp, "-liveness_message_freq", OPT_INT64, &liveness_message_freq, 
                "Minimum number of seconds between simulation liveness messages (0 = always print)",
                "1");
+   option_parser_register(opp, "-gpgpu_compute_capability_major", OPT_UINT32, &gpgpu_compute_capability_major,
+                 "Major compute capability version number",
+                 "7");
+   option_parser_register(opp, "-gpgpu_compute_capability_minor", OPT_UINT32, &gpgpu_compute_capability_minor,
+                 "Minor compute capability version number",
+                 "0");
    option_parser_register(opp, "-gpgpu_flush_l1_cache", OPT_BOOL, &gpgpu_flush_l1_cache,
                 "Flush L1 cache at the end of each kernel call",
                 "0");
@@ -532,6 +538,14 @@ void gpgpu_sim_config::reg_options(option_parser_t opp)
    option_parser_register(opp, "-visualizer_zlevel", OPT_INT32,
                           &g_visualizer_zlevel, "Compression level of the visualizer output log (0=no comp, 9=highest)",
                           "6");
+   option_parser_register(opp, "-gpgpu_stack_size_limit", OPT_INT32, &stack_size_limit,
+                          "GPU thread stack size", "1024" );
+   option_parser_register(opp, "-gpgpu_heap_size_limit", OPT_INT32, &heap_size_limit,
+                          "GPU malloc heap size ", "8388608" );
+   option_parser_register(opp, "-gpgpu_runtime_sync_depth_limit", OPT_INT32, &runtime_sync_depth_limit,
+                          "GPU device runtime synchronize depth", "2" );
+   option_parser_register(opp, "-gpgpu_runtime_pending_launch_count_limit", OPT_INT32, &runtime_pending_launch_count_limit,
+                          "GPU device runtime pending launch count", "2048" );
     option_parser_register(opp, "-trace_enabled", OPT_BOOL, 
                           &Trace::enabled, "Turn on traces",
                           "0");
@@ -794,6 +808,16 @@ void gpgpu_sim::set_prop( cudaDeviceProp *prop )
    m_cuda_properties = prop;
 }
 
+int gpgpu_sim::compute_capability_major() const
+{
+   return m_config.gpgpu_compute_capability_major;
+}
+
+int gpgpu_sim::compute_capability_minor() const
+{
+   return m_config.gpgpu_compute_capability_minor;
+}
+
 const struct cudaDeviceProp *gpgpu_sim::get_prop() const
 {
    return m_cuda_properties;
diff --git a/src/gpgpu-sim/gpu-sim.h b/src/gpgpu-sim/gpu-sim.h
index 6ce5524..d4cc18f 100644
--- a/src/gpgpu-sim/gpu-sim.h
+++ b/src/gpgpu-sim/gpu-sim.h
@@ -337,6 +337,11 @@ public:
     unsigned get_max_concurrent_kernel() const { return max_concurrent_kernel; }
     unsigned checkpoint_option;
 
+    size_t stack_limit() const {return stack_size_limit; }
+    size_t heap_limit() const {return heap_size_limit; }
+    size_t sync_depth_limit() const {return runtime_sync_depth_limit; }
+    size_t pending_launch_count_limit() const {return runtime_pending_launch_count_limit;}
+
 private:
     void init_clock_domains(void ); 
 
@@ -377,8 +382,15 @@ private:
     int gpu_stat_sample_freq;
     int gpu_runtime_stat_flag;
 
+    // Device Limits
+    size_t stack_size_limit;
+    size_t heap_size_limit;
+    size_t runtime_sync_depth_limit;
+    size_t runtime_pending_launch_count_limit;	
 
-
+ //gpu compute capability options
+    unsigned int gpgpu_compute_capability_major;
+    unsigned int gpgpu_compute_capability_minor;
     unsigned long long liveness_message_freq; 
 
     friend class gpgpu_sim;
@@ -438,6 +450,8 @@ public:
 
    int shared_mem_size() const;
    int shared_mem_per_block() const;
+   int compute_capability_major() const;
+   int compute_capability_minor() const;
    int num_registers_per_core() const;
    int num_registers_per_block() const;
    int wrp_size() const;