diff --git a/driver/common/vx_utils.cpp b/driver/common/vx_utils.cpp
index 4a4eb62d..0e590415 100644
--- a/driver/common/vx_utils.cpp
+++ b/driver/common/vx_utils.cpp
@@ -91,22 +91,22 @@ extern int vx_upload_kernel_file(vx_device_h device, const char* filename) {
   return err;
 }
 
-extern int vx_get_perf(vx_device_h device, size_t* cycles, size_t* instrs) {
+extern int vx_get_perf(vx_device_h device, int core_id, size_t* cycles, size_t* instrs) {
   int ret = 0;
 
   unsigned value;
 
   if (cycles) {
-    ret |= vx_csr_get(device, 0, CSR_CYCLE_H, &value);
+    ret |= vx_csr_get(device, core_id, CSR_CYCLE_H, &value);
     *cycles = value;
-    ret |= vx_csr_get(device, 0, CSR_CYCLE, &value);
+    ret |= vx_csr_get(device, core_id, CSR_CYCLE, &value);
     *cycles = (*cycles << 32) | value;
   }
   
   if (instrs) {
-    ret |= vx_csr_get(device, 0, CSR_INSTRET_H, &value);
+    ret |= vx_csr_get(device, core_id, CSR_INSTRET_H, &value);
     *instrs = value;
-    ret |= vx_csr_get(device, 0, CSR_INSTRET, &value);
+    ret |= vx_csr_get(device, core_id, CSR_INSTRET, &value);
     *instrs = (*instrs << 32) | value;
   }
 
diff --git a/driver/include/vortex.h b/driver/include/vortex.h
index ecdd0542..e5aa9eb6 100644
--- a/driver/include/vortex.h
+++ b/driver/include/vortex.h
@@ -58,10 +58,10 @@ int vx_start(vx_device_h hdevice);
 int vx_ready_wait(vx_device_h hdevice, long long timeout);
 
 // set device constant registers
-int vx_csr_set(vx_device_h hdevice, int core, int address, unsigned value);
+int vx_csr_set(vx_device_h hdevice, int core_id, int addr, unsigned value);
 
 // get device constant registers
-int vx_csr_get(vx_device_h hdevice, int core, int address, unsigned* value);
+int vx_csr_get(vx_device_h hdevice, int core_id, int addr, unsigned* value);
 
 ////////////////////////////// UTILITY FUNCIONS ///////////////////////////////
 
@@ -72,7 +72,7 @@ int vx_upload_kernel_bytes(vx_device_h device, const void* content, size_t size)
 int vx_upload_kernel_file(vx_device_h device, const char* filename);
 
 // get performance counters
-int vx_get_perf(vx_device_h device, size_t* cycles, size_t* instrs);
+int vx_get_perf(vx_device_h device, int core_id, size_t* cycles, size_t* instrs);
 
 #ifdef __cplusplus
 }
diff --git a/driver/opae/Makefile b/driver/opae/Makefile
index b13b897d..9946470f 100644
--- a/driver/opae/Makefile
+++ b/driver/opae/Makefile
@@ -17,6 +17,9 @@ CXXFLAGS +=-fstack-protector
 # Position independent code
 CXXFLAGS += -fPIC
 
+# Dump perf stats
+CXXFLAGS += -DDUMP_PERF_STATS
+
 # Enable scope analyzer
 #CXXFLAGS += -DSCOPE
 
diff --git a/driver/opae/vortex.cpp b/driver/opae/vortex.cpp
index c8bf410b..e66b9afe 100755
--- a/driver/opae/vortex.cpp
+++ b/driver/opae/vortex.cpp
@@ -211,14 +211,29 @@ extern int vx_dev_close(vx_device_h hdevice) {
     vx_scope_stop(device->fpga, 0);
 #endif
 
-    {   
-        // Dump perf stats
+#ifdef DUMP_PERF_STATS
+    // Dump perf stats
+    if (device->num_cores > 1) {
+        uint64_t total_instrs = 0, total_cycles = 0;
+        for (unsigned core_id = 0; core_id < device->num_cores; ++core_id) {           
+            uint64_t instrs, cycles;
+            int ret = vx_get_perf(hdevice, core_id, &instrs, &cycles);
+            assert(ret == 0);
+            float IPC = (float)(double(instrs) / double(cycles));
+            fprintf(stdout, "PERF: core%d: instrs=%ld, cycles=%ld, IPC=%f\n", core_id, instrs, cycles, IPC);            
+            total_instrs += instrs;
+            total_cycles = std::max<uint64_t>(total_cycles, cycles);
+        }
+        float IPC = (float)(double(total_instrs) / double(total_cycles));
+        fprintf(stdout, "PERF: instrs=%ld, cycles=%ld, IPC=%f\n", total_instrs, total_cycles, IPC);    
+    } else {
         uint64_t instrs, cycles;
-        int ret = vx_get_perf(hdevice, &instrs, &cycles);
+        int ret = vx_get_perf(hdevice, 0, &instrs, &cycles);
         float IPC = (float)(double(instrs) / double(cycles));
-        fprintf(stdout, "PERF: instrs=%ld, cycles=%ld, IPC=%f\n", instrs, cycles, IPC);
         assert(ret == 0);
+        fprintf(stdout, "PERF: instrs=%ld, cycles=%ld, IPC=%f\n", instrs, cycles, IPC);        
     }
+#endif
 
     fpgaClose(device->fpga);
 
@@ -480,7 +495,7 @@ extern int vx_start(vx_device_h hdevice) {
 }
 
 // set device constant registers
-extern int vx_csr_set(vx_device_h hdevice, int core, int address, unsigned value) {
+extern int vx_csr_set(vx_device_h hdevice, int core_id, int addr, unsigned value) {
     if (nullptr == hdevice)
         return -1;
 
@@ -491,8 +506,8 @@ extern int vx_csr_set(vx_device_h hdevice, int core, int address, unsigned value
         return -1;    
   
     // write CSR value
-    CHECK_RES(fpgaWriteMMIO64(device->fpga, 0, MMIO_CSR_CORE, core));
-    CHECK_RES(fpgaWriteMMIO64(device->fpga, 0, MMIO_CSR_ADDR, address));
+    CHECK_RES(fpgaWriteMMIO64(device->fpga, 0, MMIO_CSR_CORE, core_id));
+    CHECK_RES(fpgaWriteMMIO64(device->fpga, 0, MMIO_CSR_ADDR, addr));
     CHECK_RES(fpgaWriteMMIO64(device->fpga, 0, MMIO_CSR_DATA, value));
     CHECK_RES(fpgaWriteMMIO64(device->fpga, 0, MMIO_CMD_TYPE, CMD_CSR_WRITE));
 
@@ -500,7 +515,7 @@ extern int vx_csr_set(vx_device_h hdevice, int core, int address, unsigned value
 }
 
 // get device constant registers
-extern int vx_csr_get(vx_device_h hdevice, int core, int address, unsigned* value) {
+extern int vx_csr_get(vx_device_h hdevice, int core_id, int addr, unsigned* value) {
     if (nullptr == hdevice || nullptr == value)
         return -1;
 
@@ -512,8 +527,8 @@ extern int vx_csr_get(vx_device_h hdevice, int core, int address, unsigned* valu
 
     
     // write CSR value    
-    CHECK_RES(fpgaWriteMMIO64(device->fpga, 0, MMIO_CSR_CORE, core));
-    CHECK_RES(fpgaWriteMMIO64(device->fpga, 0, MMIO_CSR_ADDR, address));
+    CHECK_RES(fpgaWriteMMIO64(device->fpga, 0, MMIO_CSR_CORE, core_id));
+    CHECK_RES(fpgaWriteMMIO64(device->fpga, 0, MMIO_CSR_ADDR, addr));
     CHECK_RES(fpgaWriteMMIO64(device->fpga, 0, MMIO_CMD_TYPE, CMD_CSR_READ));    
 
     // Ensure ready for new command
diff --git a/driver/rtlsim/Makefile b/driver/rtlsim/Makefile
index ea4f3c22..d3d0df94 100644
--- a/driver/rtlsim/Makefile
+++ b/driver/rtlsim/Makefile
@@ -28,6 +28,8 @@ CFLAGS += -fPIC
 
 CFLAGS += -DUSE_RTLSIM $(CONFIGS)
 
+CFLAGS += -DDUMP_PERF_STATS
+
 LDFLAGS += -shared -pthread
 # LDFLAGS += -dynamiclib -pthread
 
diff --git a/driver/rtlsim/vortex.cpp b/driver/rtlsim/vortex.cpp
index 80167966..03d6c294 100644
--- a/driver/rtlsim/vortex.cpp
+++ b/driver/rtlsim/vortex.cpp
@@ -69,7 +69,28 @@ public:
     } 
 
     ~vx_device() {
-        simulator_.print_stats(std::cout);     
+    #ifdef DUMP_PERF_STATS
+        unsigned num_cores;
+        this->get_csr(0, CSR_NC, &num_cores);
+        if (num_cores > 1) {
+            uint64_t total_instrs = 0, total_cycles = 0;
+            for (unsigned core_id = 0; core_id < num_cores; ++core_id) {           
+                uint64_t instrs, cycles;
+                vx_get_perf(this, core_id, &instrs, &cycles);
+                float IPC = (float)(double(instrs) / double(cycles));
+                fprintf(stdout, "PERF: core%d: instrs=%ld, cycles=%ld, IPC=%f\n", core_id, instrs, cycles, IPC);            
+                total_instrs += instrs;
+                total_cycles = std::max<uint64_t>(total_cycles, cycles);
+            }
+            float IPC = (float)(double(total_instrs) / double(total_cycles));
+            fprintf(stdout, "PERF: instrs=%ld, cycles=%ld, IPC=%f\n", total_instrs, total_cycles, IPC);    
+        } else {
+            uint64_t instrs, cycles;
+            vx_get_perf(this, 0, &instrs, &cycles);
+            float IPC = (float)(double(instrs) / double(cycles));
+            fprintf(stdout, "PERF: instrs=%ld, cycles=%ld, IPC=%f\n", instrs, cycles, IPC);        
+        }
+    #endif    
         if (future_.valid()) {
             future_.wait();
         }
@@ -152,6 +173,28 @@ public:
         return 0;
     }
 
+    int set_csr(int core_id, int addr, unsigned value) {
+        if (future_.valid()) {
+            future_.wait(); // ensure prior run completed
+        }        
+        simulator_.set_csr(core_id, addr, value);        
+        while (simulator_.is_busy()) {
+            simulator_.step();
+        };
+        return 0;
+    }
+
+    int get_csr(int core_id, int addr, unsigned *value) {
+        if (future_.valid()) {
+            future_.wait(); // ensure prior run completed
+        }        
+        simulator_.get_csr(core_id, addr, value);        
+        while (simulator_.is_busy()) {
+            simulator_.step();
+        };
+        return 0;
+    }
+
 private:
 
     size_t mem_allocation_;     
@@ -324,10 +367,20 @@ extern int vx_ready_wait(vx_device_h hdevice, long long timeout) {
     return device->wait(timeout);
 }
 
-extern int vx_csr_set(vx_device_h /*hdevice*/, int /*core*/, int /*address*/, unsigned /*value*/) {
-    return -1;
+extern int vx_csr_set(vx_device_h hdevice, int core_id, int addr, unsigned value) {
+    if (nullptr == hdevice)
+        return -1;
+
+    vx_device *device = ((vx_device*)hdevice);
+
+    return device->set_csr(core_id, addr, value);
 }
 
-extern int vx_csr_get(vx_device_h /*hdevice*/, int /*core*/, int /*address*/, unsigned* /*value*/) {
-    return -1;
+extern int vx_csr_get(vx_device_h hdevice, int core_id, int addr, unsigned* value) {
+    if (nullptr == hdevice)
+        return -1;
+
+    vx_device *device = ((vx_device*)hdevice);
+
+    return device->get_csr(core_id, addr, value);
 }
\ No newline at end of file
diff --git a/driver/simx/vortex.cpp b/driver/simx/vortex.cpp
index 292c410f..b499803d 100644
--- a/driver/simx/vortex.cpp
+++ b/driver/simx/vortex.cpp
@@ -358,10 +358,10 @@ extern int vx_ready_wait(vx_device_h hdevice, long long timeout) {
     return device->wait(timeout);
 }
 
-extern int vx_csr_set(vx_device_h /*hdevice*/, int /*core*/, int /*address*/, unsigned /*value*/) {
+extern int vx_csr_set(vx_device_h /*hdevice*/, int /*core_id*/, int /*addr*/, unsigned /*value*/) {
     return -1;
 }
 
-extern int vx_csr_get(vx_device_h /*hdevice*/, int /*core*/, int /*address*/, unsigned* /*value*/) {
+extern int vx_csr_get(vx_device_h /*hdevice*/, int /*core_id*/, int /*addr*/, unsigned* /*value*/) {
     return -1;
 }
\ No newline at end of file
diff --git a/driver/stub/vortex.cpp b/driver/stub/vortex.cpp
index 007bce0e..f4a101f0 100644
--- a/driver/stub/vortex.cpp
+++ b/driver/stub/vortex.cpp
@@ -48,10 +48,10 @@ extern int vx_ready_wait(vx_device_h /*hdevice*/, long long /*timeout*/) {
     return -1;
 }
 
-extern int vx_csr_set(vx_device_h /*hdevice*/, int /*core*/, int /*address*/, unsigned /*value*/) {
+extern int vx_csr_set(vx_device_h /*hdevice*/, int /*core_id*/, int /*addr*/, unsigned /*value*/) {
     return -1;
 }
 
-extern int vx_csr_get(vx_device_h /*hdevice*/, int /*core*/, int /*address*/, unsigned* /*value*/) {
+extern int vx_csr_get(vx_device_h /*hdevice*/, int /*core_id*/, int /*addr*/, unsigned* /*value*/) {
     return -1;
 }
\ No newline at end of file
diff --git a/hw/simulate/simulator.cpp b/hw/simulate/simulator.cpp
index b4b1200e..99a5599f 100644
--- a/hw/simulate/simulator.cpp
+++ b/hw/simulate/simulator.cpp
@@ -22,6 +22,7 @@ Simulator::Simulator() {
 
   dram_rsp_active_ = false;
   snp_req_active_ = false;
+  csr_req_active_ = false;
 
 #ifdef VCD_OUTPUT
   Verilated::traceEverOn(true);
@@ -163,15 +164,6 @@ void Simulator::eval_io_bus() {
   vortex_->io_rsp_valid = 0;
 }
 
-void Simulator::eval_csr_bus() {
-  vortex_->csr_io_req_valid = 0;
-  vortex_->csr_io_req_coreid = 0;
-  vortex_->csr_io_req_addr = 0;
-  vortex_->csr_io_req_rw = 0;
-  vortex_->csr_io_req_data = 0;  
-  vortex_->csr_io_rsp_ready = 1;
-}
-
 void Simulator::eval_snp_bus() {
   if (snp_req_active_) {       
     if (vortex_->snp_rsp_valid) {      
@@ -204,6 +196,27 @@ void Simulator::eval_snp_bus() {
   }
 }
 
+void Simulator::eval_csr_bus() {
+  if (csr_req_active_) { 
+    if (vortex_->csr_io_req_rw) {
+      if (vortex_->csr_io_req_ready) {
+        vortex_->snp_req_valid = 0;
+        csr_req_active_ = false;
+      }
+    } else { 
+      if (vortex_->csr_io_rsp_valid) {
+        *csr_rsp_value_ = vortex_->csr_io_rsp_data;
+        vortex_->snp_req_valid = 0;
+        vortex_->csr_io_rsp_ready = 0;
+        csr_req_active_ = false;
+      }
+    }
+  } else {
+    vortex_->csr_io_req_valid = 0;
+    vortex_->csr_io_rsp_ready = 0;
+  }
+}
+
 void Simulator::wait(uint32_t cycles) {
   for (int i = 0; i < cycles; ++i) {
     this->step();
@@ -211,7 +224,9 @@ void Simulator::wait(uint32_t cycles) {
 }
 
 bool Simulator::is_busy() const {
-  return vortex_->busy || snp_req_active_;
+  return vortex_->busy 
+      || snp_req_active_ 
+      || csr_req_active_;
 }
 
 void Simulator::flush_caches(uint32_t mem_addr, uint32_t size) {  
@@ -221,22 +236,52 @@ void Simulator::flush_caches(uint32_t mem_addr, uint32_t size) {
   if (0 == size)
     return;
 
-  snp_req_active_ = true;
-  snp_req_size_   = (size + GLOBAL_BLOCK_SIZE - 1) / GLOBAL_BLOCK_SIZE; 
-  
   vortex_->snp_req_addr  = mem_addr / GLOBAL_BLOCK_SIZE;
   vortex_->snp_req_tag   = 0;
   vortex_->snp_req_valid = 1;
   vortex_->snp_rsp_ready = 1;  
 
+  snp_req_size_   = (size + GLOBAL_BLOCK_SIZE - 1) / GLOBAL_BLOCK_SIZE; 
   --snp_req_size_;
   pending_snp_reqs_ = 1;
+
+  snp_req_active_ = true;
     
   #ifdef DBG_PRINT_CACHE_SNP
     std::cout << timestamp << ": [sim] snp req: addr=" << std::hex << vortex_->snp_req_addr << std::dec << " tag=" << vortex_->snp_req_tag << " remain=" << snp_req_size_ << std::endl;
   #endif  
 }
 
+void Simulator::set_csr(int core_id, int addr, unsigned value) {
+#ifndef NDEBUG
+  std::cout << timestamp << ": [sim] set_csr()" << std::endl;
+#endif
+
+  vortex_->csr_io_req_valid  = 1;
+  vortex_->csr_io_req_coreid = core_id;
+  vortex_->csr_io_req_addr   = addr;
+  vortex_->csr_io_req_rw     = 1;
+  vortex_->csr_io_req_data   = value;  
+  vortex_->csr_io_rsp_ready  = 0;
+
+  csr_req_active_ = true;
+}
+
+void Simulator::get_csr(int core_id, int addr, unsigned *value) {
+#ifndef NDEBUG
+  std::cout << timestamp << ": [sim] get_csr()" << std::endl;
+#endif
+
+  vortex_->csr_io_req_valid  = 1;
+  vortex_->csr_io_req_coreid = core_id;
+  vortex_->csr_io_req_addr   = addr;
+  vortex_->csr_io_req_rw     = 0;
+  vortex_->csr_io_rsp_ready  = 1;
+
+  csr_rsp_value_ = value;
+  csr_req_active_ = true;  
+}
+
 void Simulator::run() {
 #ifndef NDEBUG
   std::cout << timestamp << ": [sim] run()" << std::endl;
diff --git a/hw/simulate/simulator.h b/hw/simulate/simulator.h
index e104b66b..a16218bb 100644
--- a/hw/simulate/simulator.h
+++ b/hw/simulate/simulator.h
@@ -31,6 +31,8 @@ public:
   Simulator();
   virtual ~Simulator();
 
+  void attach_ram(RAM* ram);
+
   void load_bin(const char* program_file);
   void load_ihex(const char* program_file);
   
@@ -39,12 +41,14 @@ public:
   void reset();
   void step();
   void wait(uint32_t cycles);
+  
   void flush_caches(uint32_t mem_addr, uint32_t size);  
-
-  void attach_ram(RAM* ram);
+  void set_csr(int core_id, int addr, unsigned value);
+  void get_csr(int core_id, int addr, unsigned *value);
 
   void run();  
   int get_last_wb_value(int reg) const;  
+
   void print_stats(std::ostream& out);
 
 private:  
@@ -60,8 +64,11 @@ private:
   int dram_rsp_active_;
   
   bool snp_req_active_;
+  bool csr_req_active_;
+
   uint32_t snp_req_size_;
   uint32_t pending_snp_reqs_;
+  uint32_t* csr_rsp_value_;
 
   RAM *ram_;
   VVortex *vortex_;