ADD: add cdp latency

5 files changed, 67 insertions, 11 deletions

diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h
index 3a268ad..45334b6 100644
--- a/src/abstract_hardware_model.h
+++ b/src/abstract_hardware_model.h
@@ -867,7 +867,7 @@ public:
         m_mem_accesses_created=false;
         m_cache_hit=false;
         m_is_printf=false;
-        m_is_cdp = false;
+        m_is_cdp = 0;
     }
     virtual ~warp_inst_t(){
     }
@@ -1020,7 +1020,6 @@ protected:
     unsigned cycles; // used for implementing initiation interval delay
     bool m_isatomic;
     bool m_is_printf;
-    bool m_is_cdp;
     unsigned m_warp_id;
     unsigned m_dynamic_warp_id; 
     const core_config *m_config; 
@@ -1041,6 +1040,11 @@ protected:
     std::list<mem_access_t> m_accessq;
 
     static unsigned sm_next_uid;
+
+    //Jin: cdp support
+public:
+    int m_is_cdp;
+    
 };
 
 void move_warp( warp_inst_t *&dst, warp_inst_t *&src );
diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc
index 276cb9d..9ecd92b 100644
--- a/src/cuda-sim/cuda-sim.cc
+++ b/src/cuda-sim/cuda-sim.cc
@@ -64,6 +64,8 @@ unsigned gpgpu_param_num_shaders = 0;
 
 char *opcode_latency_int, *opcode_latency_fp, *opcode_latency_dp;
 char *opcode_initiation_int, *opcode_initiation_fp, *opcode_initiation_dp;
+char *cdp_latency_str;
+unsigned cdp_latency[4];
 
 void ptx_opcocde_latency_options (option_parser_t opp) {
 	option_parser_register(opp, "-ptx_opcode_latency_int", OPT_CSTR, &opcode_latency_int,
@@ -90,6 +92,11 @@ void ptx_opcocde_latency_options (option_parser_t opp) {
 			"Opcode initiation intervals for double precision floating points <ADD,MAX,MUL,MAD,DIV>"
 			"Default 8,8,8,8,130",
 			"8,8,8,8,130");
+	option_parser_register(opp, "-cdp_latency", OPT_CSTR, &cdp_latency_str,
+			"CDP API latency <cudaGetParameterBufferV2, cudaStreamCreateWithFlags, \
+                cudaLaunchDeviceV2_initCTA, cudaLaunchDevicV2_perKernel>"
+			"Default 1,7200,19320,1680",
+			"1,7200,19320,1680");
 }
 
 static address_type get_converge_point(address_type pc);
@@ -609,6 +616,8 @@ void ptx_instruction::set_opcode_and_latency()
 	sscanf(opcode_initiation_dp, "%u,%u,%u,%u,%u",
 			&dp_init[0],&dp_init[1],&dp_init[2],
 			&dp_init[3],&dp_init[4]);
+	sscanf(cdp_latency_str, "%u,%u,%u,%u",
+			&cdp_latency[0],&cdp_latency[1],&cdp_latency[2], &cdp_latency[3]);
 
 	if(!m_operands.empty()){
 		std::vector<operand_info>::iterator it;
@@ -639,19 +648,21 @@ void ptx_instruction::set_opcode_and_latency()
    case MEMBAR_OP: op = MEMORY_BARRIER_OP; break;
    case CALL_OP:
    {
-       if(m_is_printf || m_is_cdp)
+       if(m_is_printf || m_is_cdp) {
            op = ALU_OP;
+       }
        else
            op = CALL_OPS;
        break;
    }
    case CALLP_OP:
    {
-       if(m_is_printf || m_is_cdp)
+       if(m_is_printf || m_is_cdp) {
                op = ALU_OP;
-           else
-               op = CALL_OPS;
-           break;
+       }
+       else
+           op = CALL_OPS;
+       break;
    }
    case RET_OP: case RETP_OP:  op = RET_OPS;break;
    case ADD_OP: case ADDP_OP: case ADDC_OP: case SUB_OP: case SUBC_OP:
diff --git a/src/cuda-sim/ptx_ir.cc b/src/cuda-sim/ptx_ir.cc
index 176eb14..4931213 100644
--- a/src/cuda-sim/ptx_ir.cc
+++ b/src/cuda-sim/ptx_ir.cc
@@ -1241,10 +1241,12 @@ ptx_instruction::ptx_instruction( int opcode,
        if (fname =="vprintf"){
            m_is_printf = true;
        }
-       if (fname == "cudaGetParameterBufferV2"
-        || fname == "cudaLaunchDeviceV2"
-        || fname == "cudaStreamCreateWithFlags")
-           m_is_cdp = true;
+       if(fname == "cudaGetParameterBufferV2")
+           m_is_cdp = 1;
+       if(fname == "cudaStreamCreateWithFlags")
+           m_is_cdp = 2;
+       if(fname == "cudaLaunchDeviceV2")
+           m_is_cdp = 3;
 
    }
 }
diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc
index cd38cb7..e85c4a8 100644
--- a/src/gpgpu-sim/shader.cc
+++ b/src/gpgpu-sim/shader.cc
@@ -841,6 +841,13 @@ void scheduler_unit::cycle()
         unsigned max_issue = m_shader->m_config->gpgpu_max_insn_issue_per_warp;
         while( !warp(warp_id).waiting() && !warp(warp_id).ibuffer_empty() && (checked < max_issue) && (checked <= issued) && (issued < max_issue) ) {
             const warp_inst_t *pI = warp(warp_id).ibuffer_next_inst();
+            //Jin: handle cdp latency;
+            if(pI->m_is_cdp && warp(warp_id).m_cdp_latency > 0) {
+                assert(warp(warp_id).m_cdp_dummy);
+                warp(warp_id).m_cdp_latency--;
+                break;
+            }
+
             bool valid = warp(warp_id).ibuffer_next_valid();
             bool warp_inst_issued = false;
             unsigned pc,rpc;
@@ -875,6 +882,25 @@ void scheduler_unit::cycle()
                             bool sp_pipe_avail = m_sp_out->has_free();
                             bool sfu_pipe_avail = m_sfu_out->has_free();
                             if( sp_pipe_avail && (pI->op != SFU_OP) ) {
+                                
+                                //Jin: special for CDP api
+                                if(pI->m_is_cdp && !warp(warp_id).m_cdp_dummy) {
+                                    assert(warp(warp_id).m_cdp_latency == 0);
+                                    
+                                    extern unsigned cdp_latency[3];
+                                    if(pI->m_is_cdp != 3)
+                                        warp(warp_id).m_cdp_latency = cdp_latency[pI->m_is_cdp - 1];
+                                    else //cudaLaunchDeviceV2
+                                        warp(warp_id).m_cdp_latency = cdp_latency[pI->m_is_cdp - 1]
+                                            + cdp_latency[pI->m_is_cdp] * active_mask.count();
+                                    warp(warp_id).m_cdp_dummy = true;
+                                    break;
+                                }
+                                else if(pI->m_is_cdp && warp(warp_id).m_cdp_dummy) {
+                                    assert(warp(warp_id).m_cdp_latency == 0);
+                                    warp(warp_id).m_cdp_dummy = false;
+                                }
+
                                 // always prefer SP pipe for operations that can use both SP and SFU pipelines
                                 m_shader->issue_warp(*m_sp_out,pI,active_mask,warp_id);
                                 issued++;
diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h
index fcbc8aa..882868e 100644
--- a/src/gpgpu-sim/shader.h
+++ b/src/gpgpu-sim/shader.h
@@ -108,6 +108,10 @@ public:
         m_last_fetch=0;
         m_next=0;
         m_inst_at_barrier=NULL;
+
+        //Jin: cdp support
+        m_cdp_latency = 0;
+        m_cdp_dummy = false;
     }
     void init( address_type start_pc,
                unsigned cta_id,
@@ -124,6 +128,10 @@ public:
         n_completed   -= active.count(); // active threads are not yet completed
         m_active_threads = active;
         m_done_exit=false;
+
+        //Jin: cdp support
+        m_cdp_latency = 0;
+        m_cdp_dummy = false;
     }
 
     bool functional_done() const;
@@ -260,6 +268,11 @@ private:
 
     unsigned m_stores_outstanding; // number of store requests sent but not yet acknowledged
     unsigned m_inst_in_pipeline;
+
+    //Jin: cdp support
+public:
+    unsigned int m_cdp_latency;
+    bool m_cdp_dummy;
 };