diff options
Diffstat (limited to 'src/abstract_hardware_model.h')
| -rw-r--r-- | src/abstract_hardware_model.h | 35 |
1 files changed, 28 insertions, 7 deletions
diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h index aaa4b00..8333247 100644 --- a/src/abstract_hardware_model.h +++ b/src/abstract_hardware_model.h @@ -76,6 +76,7 @@ enum uarch_op_t { NO_OP=-1, ALU_OP=1, SFU_OP, + DP_OP, ALU_SFU_OP, LOAD_OP, STORE_OP, @@ -131,6 +132,7 @@ typedef enum special_operations_t special_ops; // Required to identify for the p enum operation_pipeline_t { UNKOWN_OP, SP__OP, + DP__OP, SFU__OP, MEM__OP }; @@ -299,6 +301,8 @@ public: unsigned long long start_cycle; unsigned long long end_cycle; unsigned m_launch_latency; + + mutable bool volta_cache_config_set; }; struct core_config { @@ -338,6 +342,7 @@ struct core_config { unsigned gpgpu_cache_constl1_linesize; unsigned gpgpu_max_insn_issue_per_warp; + bool gmem_skip_L1D; // on = global memory access always skip the L1 cache }; // bounded stack that implements simt reconvergence using pdom mechanism from MICRO'07 paper @@ -383,12 +388,13 @@ protected: std::deque<simt_stack_entry> m_stack; }; -#define GLOBAL_HEAP_START 0x703E20000 +#define GLOBAL_HEAP_START 0xC0000000 // start allocating from this address (lower values used for allocating globals in .ptx file) #define SHARED_MEM_SIZE_MAX (64*1024) #define LOCAL_MEM_SIZE_MAX (8*1024) #define MAX_STREAMING_MULTIPROCESSORS 64 #define MAX_THREAD_PER_SM 2048 +#define MAX_WARP_PER_SM 64 #define TOTAL_LOCAL_MEM_PER_SM (MAX_THREAD_PER_SM*LOCAL_MEM_SIZE_MAX) #define TOTAL_SHARED_MEM (MAX_STREAMING_MULTIPROCESSORS*SHARED_MEM_SIZE_MAX) #define TOTAL_LOCAL_MEM (MAX_STREAMING_MULTIPROCESSORS*MAX_THREAD_PER_SM*LOCAL_MEM_SIZE_MAX) @@ -514,7 +520,14 @@ public: const struct textureReference* get_texref(const std::string &texname) const { std::map<std::string, const struct textureReference*>::const_iterator t=m_NameToTextureRef.find(texname); - assert( t != m_NameToTextureRef.end() ); + if( t == m_NameToTextureRef.end() ) { + // search for :: prefixed names + std::string temp("::" + texname); + t=m_NameToTextureRef.find(temp); + } + + assert(t != m_NameToTextureRef.end()); + return t->second; } const struct cudaArray* get_texarray( const struct textureReference *texref ) const @@ -565,6 +578,7 @@ struct gpgpu_ptx_sim_info int cmem; int gmem; int regs; + unsigned maxthreads; unsigned ptx_version; unsigned sm_target; }; @@ -615,6 +629,9 @@ private: const unsigned MAX_MEMORY_ACCESS_SIZE = 128; typedef std::bitset<MAX_MEMORY_ACCESS_SIZE> mem_access_byte_mask_t; +const unsigned SECTOR_CHUNCK_SIZE = 4; //four sectors +const unsigned SECTOR_SIZE = 32 ; //sector is 32 bytes width +typedef std::bitset<SECTOR_CHUNCK_SIZE> mem_access_sector_mask_t; #define NO_PARTIAL_WRITE (mem_access_byte_mask_t()) #define MEM_ACCESS_TYPE_TUP_DEF \ @@ -650,6 +667,7 @@ enum cache_operator_type { CACHE_ALL, // .ca CACHE_LAST_USE, // .lu CACHE_VOLATILE, // .cv + CACHE_L1, // .nc // loads and stores CACHE_STREAMING, // .cs @@ -679,8 +697,9 @@ public: unsigned size, bool wr, const active_mask_t &active_mask, - const mem_access_byte_mask_t &byte_mask ) - : m_warp_mask(active_mask), m_byte_mask(byte_mask) + const mem_access_byte_mask_t &byte_mask, + const mem_access_sector_mask_t §or_mask) + : m_warp_mask(active_mask), m_byte_mask(byte_mask), m_sector_mask(sector_mask) { init(); m_type = type; @@ -696,6 +715,7 @@ public: bool is_write() const { return m_write; } enum mem_access_type get_type() const { return m_type; } mem_access_byte_mask_t get_byte_mask() const { return m_byte_mask; } + mem_access_sector_mask_t get_sector_mask() const { return m_sector_mask; } void print(FILE *fp) const { @@ -729,6 +749,7 @@ private: mem_access_type m_type; active_mask_t m_warp_mask; mem_access_byte_mask_t m_byte_mask; + mem_access_sector_mask_t m_sector_mask; static unsigned sm_next_access_uid; }; @@ -938,9 +959,9 @@ public: }; void generate_mem_accesses(); - void memory_coalescing_arch_13( bool is_write, mem_access_type access_type ); - void memory_coalescing_arch_13_atomic( bool is_write, mem_access_type access_type ); - void memory_coalescing_arch_13_reduce_and_send( bool is_write, mem_access_type access_type, const transaction_info &info, new_addr_type addr, unsigned segment_size ); + void memory_coalescing_arch( bool is_write, mem_access_type access_type ); + void memory_coalescing_arch_atomic( bool is_write, mem_access_type access_type ); + void memory_coalescing_arch_reduce_and_send( bool is_write, mem_access_type access_type, const transaction_info &info, new_addr_type addr, unsigned segment_size ); void add_callback( unsigned lane_id, void (*function)(const class inst_t*, class ptx_thread_info*), |