Merge pull request #80 from AamirRaihan/dev

Merging Tensor Cores code

1 files changed, 70 insertions, 10 deletions

diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h
index 4a87126..3542120 100644
--- a/src/gpgpu-sim/shader.h
+++ b/src/gpgpu-sim/shader.h
@@ -328,11 +328,12 @@ public:
                    register_set* sp_out,
 				   register_set* dp_out,
                    register_set* sfu_out,
+                   register_set* tensor_core_out,
                    register_set* mem_out,
                    int id) 
         : m_supervised_warps(), m_stats(stats), m_shader(shader),
         m_scoreboard(scoreboard), m_simt_stack(simt), /*m_pipeline_reg(pipe_regs),*/ m_warp(warp),
-        m_sp_out(sp_out),m_dp_out(dp_out),m_sfu_out(sfu_out),m_mem_out(mem_out), m_id(id){}
+        m_sp_out(sp_out),m_dp_out(dp_out),m_sfu_out(sfu_out),m_tensor_core_out(tensor_core_out),m_mem_out(mem_out), m_id(id){}
     virtual ~scheduler_unit(){}
     virtual void add_supervised_warp_id(int i) {
         m_supervised_warps.push_back(&warp(i));
@@ -406,6 +407,7 @@ protected:
     register_set* m_sp_out;
     register_set* m_dp_out;
     register_set* m_sfu_out;
+    register_set* m_tensor_core_out;
     register_set* m_mem_out;
 
     int m_id;
@@ -419,9 +421,10 @@ public:
                     register_set* sp_out,
 					register_set* dp_out,
                     register_set* sfu_out,
+                    register_set* tensor_core_out,
                     register_set* mem_out,
                     int id )
-	: scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, dp_out, sfu_out, mem_out, id ){}
+	: scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, dp_out, sfu_out,tensor_core_out, mem_out, id ){}
 	virtual ~lrr_scheduler () {}
 	virtual void order_warps ();
     virtual void done_adding_supervised_warps() {
@@ -437,9 +440,10 @@ public:
                     register_set* sp_out,
 					register_set* dp_out,
                     register_set* sfu_out,
+                    register_set* tensor_core_out,
                     register_set* mem_out,
                     int id )
-	: scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, dp_out, sfu_out, mem_out, id ){}
+	: scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, dp_out, sfu_out,tensor_core_out, mem_out, id ){}
 	virtual ~gto_scheduler () {}
 	virtual void order_warps ();
     virtual void done_adding_supervised_warps() {
@@ -456,9 +460,10 @@ public:
                     register_set* sp_out,
 					register_set* dp_out,
                     register_set* sfu_out,
+                          register_set* tensor_core_out,
                     register_set* mem_out,
                     int id )
-	: scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, dp_out, sfu_out, mem_out, id ){}
+	: scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, dp_out, sfu_out,tensor_core_out, mem_out, id ){}
 	virtual ~oldest_scheduler () {}
 	virtual void order_warps ();
         virtual void done_adding_supervised_warps() {
@@ -475,10 +480,11 @@ public:
                           register_set* sp_out,
 						  register_set* dp_out,
                           register_set* sfu_out,
+                          register_set* tensor_core_out,
                           register_set* mem_out,
                           int id,
                           char* config_str )
-	: scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, dp_out, sfu_out, mem_out, id ),
+	: scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, dp_out, sfu_out,tensor_core_out, mem_out, id ),
 	  m_pending_warps() 
     {
         unsigned inner_level_readin;
@@ -526,6 +532,7 @@ public:
                     register_set* sp_out,
 					register_set* dp_out,
                     register_set* sfu_out,
+                    register_set* tensor_core_out,
                     register_set* mem_out,
                     int id,
                     char* config_string );
@@ -1112,6 +1119,23 @@ public:
     virtual void issue(  register_set& source_reg );
 };
 
+class tensor_core : public pipelined_simd_unit
+{
+public:
+    tensor_core( register_set* result_port, const shader_core_config *config, shader_core_ctx *core );
+    virtual bool can_issue( const warp_inst_t &inst ) const
+    {
+        switch(inst.op) {
+        case TENSOR_CORE_OP: break;
+        default: return false;
+        }
+        return pipelined_simd_unit::can_issue(inst);
+    }
+    virtual void active_lanes_in_pipeline();
+    virtual void issue(  register_set& source_reg );
+};
+
+
 class sp_unit : public pipelined_simd_unit
 {
 public:
@@ -1121,7 +1145,9 @@ public:
         switch(inst.op) {
         case SFU_OP: return false; 
         case LOAD_OP: return false;
+        case TENSOR_CORE_LOAD_OP: return false;
         case STORE_OP: return false;
+        case TENSOR_CORE_STORE_OP: return false;
         case MEMORY_BARRIER_OP: return false;
         case DP_OP: return false;
         default: break;
@@ -1165,7 +1191,9 @@ public:
     {
         switch(inst.op) {
         case LOAD_OP: break;
+        case TENSOR_CORE_LOAD_OP: break;
         case STORE_OP: break;
+        case TENSOR_CORE_STORE_OP: break;
         case MEMORY_BARRIER_OP: break;
         default: return false;
         }
@@ -1264,8 +1292,10 @@ enum pipeline_stage_name_t {
     OC_EX_SFU,
     OC_EX_MEM,
     EX_WB,
+    ID_OC_TENSOR_CORE,  
+    OC_EX_TENSOR_CORE,
     N_PIPELINE_STAGES 
-};
+    };
 
 const char* const pipeline_stage_name_decode[] = {
     "ID_OC_SP",
@@ -1277,6 +1307,8 @@ const char* const pipeline_stage_name_decode[] = {
     "OC_EX_SFU",
     "OC_EX_MEM",
     "EX_WB",
+    "ID_OC_TENSOR_CORE",  
+    "OC_EX_TENSOR_CORE",
     "N_PIPELINE_STAGES" 
 };
 
@@ -1299,16 +1331,24 @@ struct shader_core_config : public core_config
 	char* toks = new char[100];
 	char* tokd = toks;
 	strcpy(toks,pipeline_widths_string);
-
+                                  
 	toks = strtok(toks,",");
-	for (unsigned i = 0; i < N_PIPELINE_STAGES; i++) { 
+
+	/*	Removing the tensorcore pipeline while reading the config files if the tensor core is not available.
+	 	If we won't remove it, old regression will be broken.
+		So to support the legacy config files it's best to handle in this way.
+         */  
+	int num_config_to_read=N_PIPELINE_STAGES-2*(!gpgpu_tensor_core_avail);
+
+        for (unsigned i = 0; i <num_config_to_read; i++) { 
 	    assert(toks);
 	    ntok = sscanf(toks,"%d", &pipe_widths[i]);
 	    assert(ntok == 1); 
 	    toks = strtok(NULL,",");
 	}
-	delete[] tokd;
 
+	delete[] tokd;
+	
         if (n_thread_per_shader > MAX_THREAD_PER_SM) {
            printf("GPGPU-Sim uArch: Error ** increase MAX_THREAD_PER_SM in abstract_hardware_model.h from %u to %u\n", 
                   MAX_THREAD_PER_SM, n_thread_per_shader);
@@ -1318,7 +1358,14 @@ struct shader_core_config : public core_config
         assert( !(n_thread_per_shader % warp_size) );
         max_sfu_latency = 512;
         max_sp_latency = 32;
-        m_L1I_config.init(m_L1I_config.m_config_string,FuncCachePreferNone);
+        
+	max_tensor_core_latency = 64;
+        gpgpu_num_tensor_core_units=4;//It will be (#TENSORCORE INSIDE SM)/2 (One warp is allocated to 2 Tensor Core)
+        gpgpu_operand_collector_num_units_tensor_core=24;
+        gpgpu_operand_collector_num_in_ports_tensor_core=8;
+        gpgpu_operand_collector_num_out_ports_tensor_core=8;
+        
+	m_L1I_config.init(m_L1I_config.m_config_string,FuncCachePreferNone);
         m_L1T_config.init(m_L1T_config.m_config_string,FuncCachePreferNone);
         m_L1C_config.init(m_L1C_config.m_config_string,FuncCachePreferNone);
         m_L1D_config.init(m_L1D_config.m_config_string,FuncCachePreferNone);
@@ -1364,24 +1411,29 @@ struct shader_core_config : public core_config
     int gpgpu_operand_collector_num_units_sp;
     int gpgpu_operand_collector_num_units_dp;
     int gpgpu_operand_collector_num_units_sfu;
+    int gpgpu_operand_collector_num_units_tensor_core;
     int gpgpu_operand_collector_num_units_mem;
     int gpgpu_operand_collector_num_units_gen;
 
     unsigned int gpgpu_operand_collector_num_in_ports_sp;
     unsigned int gpgpu_operand_collector_num_in_ports_dp;
     unsigned int gpgpu_operand_collector_num_in_ports_sfu;
+    unsigned int gpgpu_operand_collector_num_in_ports_tensor_core;
     unsigned int gpgpu_operand_collector_num_in_ports_mem;
     unsigned int gpgpu_operand_collector_num_in_ports_gen;
 
     unsigned int gpgpu_operand_collector_num_out_ports_sp;
     unsigned int gpgpu_operand_collector_num_out_ports_dp;
     unsigned int gpgpu_operand_collector_num_out_ports_sfu;
+    unsigned int gpgpu_operand_collector_num_out_ports_tensor_core;
     unsigned int gpgpu_operand_collector_num_out_ports_mem;
     unsigned int gpgpu_operand_collector_num_out_ports_gen;
 
     int gpgpu_num_sp_units;
+    int gpgpu_tensor_core_avail;
     int gpgpu_num_dp_units;
     int gpgpu_num_sfu_units;
+    int gpgpu_num_tensor_core_units;
     int gpgpu_num_mem_units;
 
     //Shader core resources
@@ -1395,6 +1447,7 @@ struct shader_core_config : public core_config
     
     unsigned max_sp_latency;
     unsigned max_sfu_latency;
+    unsigned max_tensor_core_latency;
     
     unsigned n_simt_cores_per_cluster;
     unsigned n_simt_clusters;
@@ -1436,12 +1489,14 @@ struct shader_core_stats_pod {
     unsigned *m_num_fpdiv_acesses;
     unsigned *m_num_sp_acesses;
     unsigned *m_num_sfu_acesses;
+    unsigned *m_num_tensor_core_acesses;
     unsigned *m_num_trans_acesses;
     unsigned *m_num_mem_acesses;
     unsigned *m_num_sp_committed;
     unsigned *m_num_tlb_hits;
     unsigned *m_num_tlb_accesses;
     unsigned *m_num_sfu_committed;
+    unsigned *m_num_tensor_core_committed;
     unsigned *m_num_mem_committed;
     unsigned *m_read_regfile_acesses;
     unsigned *m_write_regfile_acesses;
@@ -1450,12 +1505,14 @@ struct shader_core_stats_pod {
     unsigned *m_num_imul32_acesses;
     unsigned *m_active_sp_lanes;
     unsigned *m_active_sfu_lanes;
+    unsigned *m_active_tensor_core_lanes;
     unsigned *m_active_fu_lanes;
     unsigned *m_active_fu_mem_lanes;
     unsigned *m_n_diverge;    // number of divergence occurring in this shader
     unsigned gpgpu_n_load_insn;
     unsigned gpgpu_n_store_insn;
     unsigned gpgpu_n_shmem_insn;
+    unsigned gpgpu_n_sstarr_insn;
     unsigned gpgpu_n_tex_insn;
     unsigned gpgpu_n_const_insn;
     unsigned gpgpu_n_param_insn;
@@ -1522,6 +1579,7 @@ public:
         m_num_fpdiv_acesses= (unsigned*) calloc(config->num_shader(),sizeof(unsigned));
         m_num_sp_acesses= (unsigned*) calloc(config->num_shader(),sizeof(unsigned));
         m_num_sfu_acesses= (unsigned*) calloc(config->num_shader(),sizeof(unsigned));
+        m_num_tensor_core_acesses= (unsigned*) calloc(config->num_shader(),sizeof(unsigned));
         m_num_trans_acesses= (unsigned*) calloc(config->num_shader(),sizeof(unsigned));
         m_num_mem_acesses= (unsigned*) calloc(config->num_shader(),sizeof(unsigned));
         m_num_sp_committed= (unsigned*) calloc(config->num_shader(),sizeof(unsigned));
@@ -1529,9 +1587,11 @@ public:
         m_num_tlb_accesses=(unsigned*) calloc(config->num_shader(),sizeof(unsigned));
         m_active_sp_lanes= (unsigned*) calloc(config->num_shader(),sizeof(unsigned));
         m_active_sfu_lanes= (unsigned*) calloc(config->num_shader(),sizeof(unsigned));
+        m_active_tensor_core_lanes= (unsigned*) calloc(config->num_shader(),sizeof(unsigned));
         m_active_fu_lanes= (unsigned*) calloc(config->num_shader(),sizeof(unsigned));
         m_active_fu_mem_lanes= (unsigned*) calloc(config->num_shader(),sizeof(unsigned));
         m_num_sfu_committed= (unsigned*) calloc(config->num_shader(),sizeof(unsigned));
+        m_num_tensor_core_committed= (unsigned*) calloc(config->num_shader(),sizeof(unsigned));
         m_num_mem_committed= (unsigned*) calloc(config->num_shader(),sizeof(unsigned));
         m_read_regfile_acesses= (unsigned*) calloc(config->num_shader(),sizeof(unsigned));
         m_write_regfile_acesses= (unsigned*) calloc(config->num_shader(),sizeof(unsigned));