Merge pull request #12 from barnes88/sub_core_devel

sub_core_model operand collector unit partitioning

3 files changed, 202 insertions, 79 deletions

diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h
index c012de0..982e416 100644
--- a/src/abstract_hardware_model.h
+++ b/src/abstract_hardware_model.h
@@ -1291,6 +1291,7 @@ class register_set {
     }
     m_name = name;
   }
+  const char *get_name() { return m_name; }
   bool has_free() {
     for (unsigned i = 0; i < regs.size(); i++) {
       if (regs[i]->empty()) {
@@ -1315,7 +1316,35 @@ class register_set {
     }
     return false;
   }
+  bool has_ready(bool sub_core_model, unsigned reg_id) {
+    if (!sub_core_model) return has_ready();
+    assert(reg_id < regs.size());
+    return (not regs[reg_id]->empty());
+  }
 
+  unsigned get_ready_reg_id() {
+    // for sub core model we need to figure which reg_id has the ready warp
+    // this function should only be called if has_ready() was true
+    assert(has_ready());
+    warp_inst_t **ready;
+    ready = NULL;
+    unsigned reg_id;
+    for (unsigned i = 0; i < regs.size(); i++) {
+      if (not regs[i]->empty()) {
+        if (ready and (*ready)->get_uid() < regs[i]->get_uid()) {
+          // ready is oldest
+        } else {
+          ready = &regs[i];
+          reg_id = i;
+        }
+      }
+    }
+    return reg_id;
+  }
+  unsigned get_schd_id(unsigned reg_id) {
+    assert(not regs[reg_id]->empty());
+    return regs[reg_id]->get_schd_id();
+  }
   void move_in(warp_inst_t *&src) {
     warp_inst_t **free = get_free();
     move_warp(*free, src);
@@ -1323,10 +1352,29 @@ class register_set {
   // void copy_in( warp_inst_t* src ){
   //   src->copy_contents_to(*get_free());
   //}
+  void move_in(bool sub_core_model, unsigned reg_id, warp_inst_t *&src) {
+    warp_inst_t **free;
+    if (!sub_core_model) {
+      free = get_free();
+    } else {
+      assert(reg_id < regs.size());
+      free = get_free(sub_core_model, reg_id);
+    }
+    move_warp(*free, src);
+  }
+
   void move_out_to(warp_inst_t *&dest) {
     warp_inst_t **ready = get_ready();
     move_warp(dest, *ready);
   }
+  void move_out_to(bool sub_core_model, unsigned reg_id, warp_inst_t *&dest) {
+    if (!sub_core_model) {
+      return move_out_to(dest);
+    }
+    warp_inst_t **ready = get_ready(sub_core_model, reg_id);
+    assert(ready != NULL);
+    move_warp(dest, *ready);
+  }
 
   warp_inst_t **get_ready() {
     warp_inst_t **ready;
@@ -1342,6 +1390,14 @@ class register_set {
     }
     return ready;
   }
+  warp_inst_t **get_ready(bool sub_core_model, unsigned reg_id) {
+    if (!sub_core_model) return get_ready();
+    warp_inst_t **ready;
+    ready = NULL;
+    assert(reg_id < regs.size());
+    if (not regs[reg_id]->empty()) ready = &regs[reg_id];
+    return ready;
+  }
 
   void print(FILE *fp) const {
     fprintf(fp, "%s : @%p\n", m_name, this);
diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc
index c6e7b8f..14d9044 100644
--- a/src/gpgpu-sim/shader.cc
+++ b/src/gpgpu-sim/shader.cc
@@ -166,18 +166,15 @@ void shader_core_ctx::create_schedulers() {
   // must currently occur after all inputs have been initialized.
   std::string sched_config = m_config->gpgpu_scheduler_string;
   const concrete_scheduler scheduler =
-      sched_config.find("lrr") != std::string::npos
-          ? CONCRETE_SCHEDULER_LRR
-          : sched_config.find("two_level_active") != std::string::npos
-                ? CONCRETE_SCHEDULER_TWO_LEVEL_ACTIVE
-                : sched_config.find("gto") != std::string::npos
-                      ? CONCRETE_SCHEDULER_GTO
-                      : sched_config.find("old") != std::string::npos
-                            ? CONCRETE_SCHEDULER_OLDEST_FIRST
-                            : sched_config.find("warp_limiting") !=
-                                      std::string::npos
-                                  ? CONCRETE_SCHEDULER_WARP_LIMITING
-                                  : NUM_CONCRETE_SCHEDULERS;
+      sched_config.find("lrr") != std::string::npos ? CONCRETE_SCHEDULER_LRR
+      : sched_config.find("two_level_active") != std::string::npos
+          ? CONCRETE_SCHEDULER_TWO_LEVEL_ACTIVE
+      : sched_config.find("gto") != std::string::npos ? CONCRETE_SCHEDULER_GTO
+      : sched_config.find("old") != std::string::npos
+          ? CONCRETE_SCHEDULER_OLDEST_FIRST
+      : sched_config.find("warp_limiting") != std::string::npos
+          ? CONCRETE_SCHEDULER_WARP_LIMITING
+          : NUM_CONCRETE_SCHEDULERS;
   assert(scheduler != NUM_CONCRETE_SCHEDULERS);
 
   for (unsigned i = 0; i < m_config->gpgpu_num_sched_per_core; i++) {
@@ -377,41 +374,41 @@ void shader_core_ctx::create_exec_pipeline() {
 
   // m_fu = new simd_function_unit*[m_num_function_units];
 
-  for (int k = 0; k < m_config->gpgpu_num_sp_units; k++) {
-    m_fu.push_back(new sp_unit(&m_pipeline_reg[EX_WB], m_config, this));
+  for (unsigned k = 0; k < m_config->gpgpu_num_sp_units; k++) {
+    m_fu.push_back(new sp_unit(&m_pipeline_reg[EX_WB], m_config, this, k));
     m_dispatch_port.push_back(ID_OC_SP);
     m_issue_port.push_back(OC_EX_SP);
   }
 
-  for (int k = 0; k < m_config->gpgpu_num_dp_units; k++) {
-    m_fu.push_back(new dp_unit(&m_pipeline_reg[EX_WB], m_config, this));
+  for (unsigned k = 0; k < m_config->gpgpu_num_dp_units; k++) {
+    m_fu.push_back(new dp_unit(&m_pipeline_reg[EX_WB], m_config, this, k));
     m_dispatch_port.push_back(ID_OC_DP);
     m_issue_port.push_back(OC_EX_DP);
   }
-  for (int k = 0; k < m_config->gpgpu_num_int_units; k++) {
-    m_fu.push_back(new int_unit(&m_pipeline_reg[EX_WB], m_config, this));
+  for (unsigned k = 0; k < m_config->gpgpu_num_int_units; k++) {
+    m_fu.push_back(new int_unit(&m_pipeline_reg[EX_WB], m_config, this, k));
     m_dispatch_port.push_back(ID_OC_INT);
     m_issue_port.push_back(OC_EX_INT);
   }
 
-  for (int k = 0; k < m_config->gpgpu_num_sfu_units; k++) {
-    m_fu.push_back(new sfu(&m_pipeline_reg[EX_WB], m_config, this));
+  for (unsigned k = 0; k < m_config->gpgpu_num_sfu_units; k++) {
+    m_fu.push_back(new sfu(&m_pipeline_reg[EX_WB], m_config, this, k));
     m_dispatch_port.push_back(ID_OC_SFU);
     m_issue_port.push_back(OC_EX_SFU);
   }
 
-  for (int k = 0; k < m_config->gpgpu_num_tensor_core_units; k++) {
-    m_fu.push_back(new tensor_core(&m_pipeline_reg[EX_WB], m_config, this));
+  for (unsigned k = 0; k < m_config->gpgpu_num_tensor_core_units; k++) {
+    m_fu.push_back(new tensor_core(&m_pipeline_reg[EX_WB], m_config, this, k));
     m_dispatch_port.push_back(ID_OC_TENSOR_CORE);
     m_issue_port.push_back(OC_EX_TENSOR_CORE);
   }
 
-  for (int j = 0; j < m_config->m_specialized_unit.size(); j++) {
+  for (unsigned j = 0; j < m_config->m_specialized_unit.size(); j++) {
     for (unsigned k = 0; k < m_config->m_specialized_unit[j].num_units; k++) {
       m_fu.push_back(new specialized_unit(
           &m_pipeline_reg[EX_WB], m_config, this, SPEC_UNIT_START_ID + j,
           m_config->m_specialized_unit[j].name,
-          m_config->m_specialized_unit[j].latency));
+          m_config->m_specialized_unit[j].latency, k));
       m_dispatch_port.push_back(m_config->m_specialized_unit[j].ID_OC_SPEC_ID);
       m_issue_port.push_back(m_config->m_specialized_unit[j].OC_EX_SPEC_ID);
     }
@@ -1669,8 +1666,15 @@ void shader_core_ctx::execute() {
     m_fu[n]->active_lanes_in_pipeline();
     unsigned issue_port = m_issue_port[n];
     register_set &issue_inst = m_pipeline_reg[issue_port];
-    warp_inst_t **ready_reg = issue_inst.get_ready();
-    if (issue_inst.has_ready() && m_fu[n]->can_issue(**ready_reg)) {
+    unsigned reg_id;
+    bool partition_issue =
+        m_config->sub_core_model && m_fu[n]->is_issue_partitioned();
+    if (partition_issue) {
+      reg_id = m_fu[n]->get_issue_reg_id();
+    }
+    warp_inst_t **ready_reg = issue_inst.get_ready(partition_issue, reg_id);
+    if (issue_inst.has_ready(partition_issue, reg_id) &&
+        m_fu[n]->can_issue(**ready_reg)) {
       bool schedule_wb_now = !m_fu[n]->stallable();
       int resbus = -1;
       if (schedule_wb_now &&
@@ -2112,22 +2116,32 @@ simd_function_unit::simd_function_unit(const shader_core_config *config) {
   m_dispatch_reg = new warp_inst_t(config);
 }
 
+void simd_function_unit::issue(register_set &source_reg) {
+  bool partition_issue =
+      m_config->sub_core_model && this->is_issue_partitioned();
+  source_reg.move_out_to(partition_issue, this->get_issue_reg_id(),
+                         m_dispatch_reg);
+  occupied.set(m_dispatch_reg->latency);
+}
+
 sfu::sfu(register_set *result_port, const shader_core_config *config,
-         shader_core_ctx *core)
-    : pipelined_simd_unit(result_port, config, config->max_sfu_latency, core) {
+         shader_core_ctx *core, unsigned issue_reg_id)
+    : pipelined_simd_unit(result_port, config, config->max_sfu_latency, core,
+                          issue_reg_id) {
   m_name = "SFU";
 }
 
 tensor_core::tensor_core(register_set *result_port,
                          const shader_core_config *config,
-                         shader_core_ctx *core)
+                         shader_core_ctx *core, unsigned issue_reg_id)
     : pipelined_simd_unit(result_port, config, config->max_tensor_core_latency,
-                          core) {
+                          core, issue_reg_id) {
   m_name = "TENSOR_CORE";
 }
 
 void sfu::issue(register_set &source_reg) {
-  warp_inst_t **ready_reg = source_reg.get_ready();
+  warp_inst_t **ready_reg =
+      source_reg.get_ready(m_config->sub_core_model, m_issue_reg_id);
   // m_core->incexecstat((*ready_reg));
 
   (*ready_reg)->op_pipe = SFU__OP;
@@ -2136,7 +2150,8 @@ void sfu::issue(register_set &source_reg) {
 }
 
 void tensor_core::issue(register_set &source_reg) {
-  warp_inst_t **ready_reg = source_reg.get_ready();
+  warp_inst_t **ready_reg =
+      source_reg.get_ready(m_config->sub_core_model, m_issue_reg_id);
   // m_core->incexecstat((*ready_reg));
 
   (*ready_reg)->op_pipe = TENSOR_CORE__OP;
@@ -2208,34 +2223,39 @@ void tensor_core::active_lanes_in_pipeline() {
 }
 
 sp_unit::sp_unit(register_set *result_port, const shader_core_config *config,
-                 shader_core_ctx *core)
-    : pipelined_simd_unit(result_port, config, config->max_sp_latency, core) {
+                 shader_core_ctx *core, unsigned issue_reg_id)
+    : pipelined_simd_unit(result_port, config, config->max_sp_latency, core,
+                          issue_reg_id) {
   m_name = "SP ";
 }
 
 specialized_unit::specialized_unit(register_set *result_port,
                                    const shader_core_config *config,
                                    shader_core_ctx *core, unsigned supported_op,
-                                   char *unit_name, unsigned latency)
-    : pipelined_simd_unit(result_port, config, latency, core) {
+                                   char *unit_name, unsigned latency,
+                                   unsigned issue_reg_id)
+    : pipelined_simd_unit(result_port, config, latency, core, issue_reg_id) {
   m_name = unit_name;
   m_supported_op = supported_op;
 }
 
 dp_unit::dp_unit(register_set *result_port, const shader_core_config *config,
-                 shader_core_ctx *core)
-    : pipelined_simd_unit(result_port, config, config->max_dp_latency, core) {
+                 shader_core_ctx *core, unsigned issue_reg_id)
+    : pipelined_simd_unit(result_port, config, config->max_dp_latency, core,
+                          issue_reg_id) {
   m_name = "DP ";
 }
 
 int_unit::int_unit(register_set *result_port, const shader_core_config *config,
-                   shader_core_ctx *core)
-    : pipelined_simd_unit(result_port, config, config->max_int_latency, core) {
+                   shader_core_ctx *core, unsigned issue_reg_id)
+    : pipelined_simd_unit(result_port, config, config->max_int_latency, core,
+                          issue_reg_id) {
   m_name = "INT ";
 }
 
 void sp_unit ::issue(register_set &source_reg) {
-  warp_inst_t **ready_reg = source_reg.get_ready();
+  warp_inst_t **ready_reg =
+      source_reg.get_ready(m_config->sub_core_model, m_issue_reg_id);
   // m_core->incexecstat((*ready_reg));
   (*ready_reg)->op_pipe = SP__OP;
   m_core->incsp_stat(m_core->get_config()->warp_size, (*ready_reg)->latency);
@@ -2243,7 +2263,8 @@ void sp_unit ::issue(register_set &source_reg) {
 }
 
 void dp_unit ::issue(register_set &source_reg) {
-  warp_inst_t **ready_reg = source_reg.get_ready();
+  warp_inst_t **ready_reg =
+      source_reg.get_ready(m_config->sub_core_model, m_issue_reg_id);
   // m_core->incexecstat((*ready_reg));
   (*ready_reg)->op_pipe = DP__OP;
   m_core->incsp_stat(m_core->get_config()->warp_size, (*ready_reg)->latency);
@@ -2251,7 +2272,8 @@ void dp_unit ::issue(register_set &source_reg) {
 }
 
 void specialized_unit ::issue(register_set &source_reg) {
-  warp_inst_t **ready_reg = source_reg.get_ready();
+  warp_inst_t **ready_reg = 
+      source_reg.get_ready(m_config->sub_core_model, m_issue_reg_id);
   // m_core->incexecstat((*ready_reg));
   (*ready_reg)->op_pipe = SPECIALIZED__OP;
   m_core->incsp_stat(m_core->get_config()->warp_size, (*ready_reg)->latency);
@@ -2259,7 +2281,8 @@ void specialized_unit ::issue(register_set &source_reg) {
 }
 
 void int_unit ::issue(register_set &source_reg) {
-  warp_inst_t **ready_reg = source_reg.get_ready();
+  warp_inst_t **ready_reg =
+      source_reg.get_ready(m_config->sub_core_model, m_issue_reg_id);
   // m_core->incexecstat((*ready_reg));
   (*ready_reg)->op_pipe = INTP__OP;
   m_core->incsp_stat(m_core->get_config()->warp_size, (*ready_reg)->latency);
@@ -2269,7 +2292,8 @@ void int_unit ::issue(register_set &source_reg) {
 pipelined_simd_unit::pipelined_simd_unit(register_set *result_port,
                                          const shader_core_config *config,
                                          unsigned max_latency,
-                                         shader_core_ctx *core)
+                                         shader_core_ctx *core,
+                                         unsigned issue_reg_id)
     : simd_function_unit(config) {
   m_result_port = result_port;
   m_pipeline_depth = max_latency;
@@ -2277,6 +2301,7 @@ pipelined_simd_unit::pipelined_simd_unit(register_set *result_port,
   for (unsigned i = 0; i < m_pipeline_depth; i++)
     m_pipeline_reg[i] = new warp_inst_t(config);
   m_core = core;
+  m_issue_reg_id = issue_reg_id;
   active_insts_in_pipeline = 0;
 }
 
@@ -2303,7 +2328,10 @@ void pipelined_simd_unit::cycle() {
 
 void pipelined_simd_unit::issue(register_set &source_reg) {
   // move_warp(m_dispatch_reg,source_reg);
-  warp_inst_t **ready_reg = source_reg.get_ready();
+  bool partition_issue =
+      m_config->sub_core_model && this->is_issue_partitioned();
+  warp_inst_t **ready_reg =
+      source_reg.get_ready(partition_issue, m_issue_reg_id);
   m_core->incexecstat((*ready_reg));
   // source_reg.move_out_to(m_dispatch_reg);
   simd_function_unit::issue(source_reg);
@@ -2360,7 +2388,7 @@ ldst_unit::ldst_unit(mem_fetch_interface *icnt,
                      Scoreboard *scoreboard, const shader_core_config *config,
                      const memory_config *mem_config, shader_core_stats *stats,
                      unsigned sid, unsigned tpc)
-    : pipelined_simd_unit(NULL, config, config->smem_latency, core),
+    : pipelined_simd_unit(NULL, config, config->smem_latency, core, 0),
       m_next_wb(config) {
   assert(config->smem_latency > 1);
   init(icnt, mf_allocator, core, operand_collector, scoreboard, config,
@@ -2388,7 +2416,7 @@ ldst_unit::ldst_unit(mem_fetch_interface *icnt,
                      Scoreboard *scoreboard, const shader_core_config *config,
                      const memory_config *mem_config, shader_core_stats *stats,
                      unsigned sid, unsigned tpc, l1_cache *new_l1d_cache)
-    : pipelined_simd_unit(NULL, config, 3, core),
+    : pipelined_simd_unit(NULL, config, 3, core, 0),
       m_L1D(new_l1d_cache),
       m_next_wb(config) {
   init(icnt, mf_allocator, core, operand_collector, scoreboard, config,
@@ -3867,15 +3895,23 @@ void opndcoll_rfu_t::init(unsigned num_banks, shader_core_ctx *shader) {
   assert((m_bank_warp_shift == 5) || (m_warp_size != 32));
 
   sub_core_model = shader->get_config()->sub_core_model;
-  m_num_warp_sceds = shader->get_config()->gpgpu_num_sched_per_core;
-  if (sub_core_model)
+  m_num_warp_scheds = shader->get_config()->gpgpu_num_sched_per_core;
+  unsigned reg_id;
+  if (sub_core_model) {
     assert(num_banks % shader->get_config()->gpgpu_num_sched_per_core == 0);
+    assert(m_num_warp_scheds <= m_cu.size() &&
+           m_cu.size() % m_num_warp_scheds == 0);
+  }
   m_num_banks_per_sched =
       num_banks / shader->get_config()->gpgpu_num_sched_per_core;
 
   for (unsigned j = 0; j < m_cu.size(); j++) {
+    if (sub_core_model) {
+      unsigned cusPerSched = m_cu.size() / m_num_warp_scheds;
+      reg_id = j / cusPerSched;
+    }
     m_cu[j]->init(j, num_banks, m_bank_warp_shift, shader->get_config(), this,
-                  sub_core_model, m_num_banks_per_sched);
+                  sub_core_model, reg_id, m_num_banks_per_sched);
   }
   m_initialized = true;
 }
@@ -3974,7 +4010,22 @@ void opndcoll_rfu_t::allocate_cu(unsigned port_num) {
       for (unsigned j = 0; j < inp.m_cu_sets.size(); j++) {
         std::vector<collector_unit_t> &cu_set = m_cus[inp.m_cu_sets[j]];
         bool allocated = false;
-        for (unsigned k = 0; k < cu_set.size(); k++) {
+        unsigned cuLowerBound = 0;
+        unsigned cuUpperBound = cu_set.size();
+        unsigned schd_id;
+        if (sub_core_model) {
+          // Sub core model only allocates on the subset of CUs assigned to the
+          // scheduler that issued
+          unsigned reg_id = (*inp.m_in[i]).get_ready_reg_id();
+          schd_id = (*inp.m_in[i]).get_schd_id(reg_id);
+          assert(cu_set.size() % m_num_warp_scheds == 0 &&
+                 cu_set.size() >= m_num_warp_scheds);
+          unsigned cusPerSched = cu_set.size() / m_num_warp_scheds;
+          cuLowerBound = schd_id * cusPerSched;
+          cuUpperBound = cuLowerBound + cusPerSched;
+          assert(0 <= cuLowerBound && cuUpperBound <= cu_set.size());
+        }
+        for (unsigned k = cuLowerBound; k < cuUpperBound; k++) {
           if (cu_set[k].is_free()) {
             collector_unit_t *cu = &cu_set[k];
             allocated = cu->allocate(inp.m_in[i], inp.m_out[i]);
@@ -3984,8 +4035,9 @@ void opndcoll_rfu_t::allocate_cu(unsigned port_num) {
         }
         if (allocated) break;  // cu has been allocated, no need to search more.
       }
-      break;  // can only service a single input, if it failed it will fail for
-              // others.
+      // break;  // can only service a single input, if it failed it will fail
+      // for
+      // others.
     }
   }
 }
@@ -4032,7 +4084,8 @@ void opndcoll_rfu_t::allocate_reads() {
 }
 
 bool opndcoll_rfu_t::collector_unit_t::ready() const {
-  return (!m_free) && m_not_ready.none() && (*m_output_register).has_free();
+  return (!m_free) && m_not_ready.none() &&
+         (*m_output_register).has_free(m_sub_core_model, m_reg_id);
 }
 
 void opndcoll_rfu_t::collector_unit_t::dump(
@@ -4050,12 +4103,10 @@ void opndcoll_rfu_t::collector_unit_t::dump(
   }
 }
 
-void opndcoll_rfu_t::collector_unit_t::init(unsigned n, unsigned num_banks,
-                                            unsigned log2_warp_size,
-                                            const core_config *config,
-                                            opndcoll_rfu_t *rfu,
-                                            bool sub_core_model,
-                                            unsigned banks_per_sched) {
+void opndcoll_rfu_t::collector_unit_t::init(
+    unsigned n, unsigned num_banks, unsigned log2_warp_size,
+    const core_config *config, opndcoll_rfu_t *rfu, bool sub_core_model,
+    unsigned reg_id, unsigned banks_per_sched) {
   m_rfu = rfu;
   m_cuid = n;
   m_num_banks = num_banks;
@@ -4063,6 +4114,7 @@ void opndcoll_rfu_t::collector_unit_t::init(unsigned n, unsigned num_banks,
   m_warp = new warp_inst_t(config);
   m_bank_warp_shift = log2_warp_size;
   m_sub_core_model = sub_core_model;
+  m_reg_id = reg_id;
   m_num_banks_per_sched = banks_per_sched;
 }
 
@@ -4097,8 +4149,7 @@ bool opndcoll_rfu_t::collector_unit_t::allocate(register_set *pipeline_reg_set,
 
 void opndcoll_rfu_t::collector_unit_t::dispatch() {
   assert(m_not_ready.none());
-  // move_warp(*m_output_register,m_warp);
-  m_output_register->move_in(m_warp);
+  m_output_register->move_in(m_sub_core_model, m_reg_id, m_warp);
   m_free = true;
   m_output_register = NULL;
   for (unsigned i = 0; i < MAX_REG_OPERANDS * 2; i++) m_src_op[i].reset();
diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h
index 6481790..8c02fd7 100644
--- a/src/gpgpu-sim/shader.h
+++ b/src/gpgpu-sim/shader.h
@@ -238,7 +238,10 @@ class shd_warp_t {
   unsigned get_dynamic_warp_id() const { return m_dynamic_warp_id; }
   unsigned get_warp_id() const { return m_warp_id; }
 
-  class shader_core_ctx * get_shader() { return m_shader; }
+  class shader_core_ctx *get_shader() {
+    return m_shader;
+  }
+
  private:
   static const unsigned IBUFFER_SIZE = 2;
   class shader_core_ctx *m_shader;
@@ -878,11 +881,13 @@ class opndcoll_rfu_t {  // operand collector based register file unit
     }
     unsigned get_sp_op() const { return m_warp->sp_op; }
     unsigned get_id() const { return m_cuid; }  // returns CU hw id
+    unsigned get_reg_id() const { return m_reg_id; }
 
     // modifiers
     void init(unsigned n, unsigned num_banks, unsigned log2_warp_size,
               const core_config *config, opndcoll_rfu_t *rfu,
-              bool m_sub_core_model, unsigned num_banks_per_sched);
+              bool m_sub_core_model, unsigned reg_id,
+              unsigned num_banks_per_sched);
     bool allocate(register_set *pipeline_reg, register_set *output_reg);
 
     void collect_operand(unsigned op) { m_not_ready.reset(op); }
@@ -906,6 +911,7 @@ class opndcoll_rfu_t {  // operand collector based register file unit
 
     unsigned m_num_banks_per_sched;
     bool m_sub_core_model;
+    unsigned m_reg_id;  // if sub_core_model enabled, limit regs this cu can r/w
   };
 
   class dispatch_unit_t {
@@ -947,7 +953,7 @@ class opndcoll_rfu_t {  // operand collector based register file unit
   arbiter_t m_arbiter;
 
   unsigned m_num_banks_per_sched;
-  unsigned m_num_warp_sceds;
+  unsigned m_num_warp_scheds;
   bool sub_core_model;
 
   // unsigned m_num_ports;
@@ -1039,10 +1045,7 @@ class simd_function_unit {
   ~simd_function_unit() { delete m_dispatch_reg; }
 
   // modifiers
-  virtual void issue(register_set &source_reg) {
-    source_reg.move_out_to(m_dispatch_reg);
-    occupied.set(m_dispatch_reg->latency);
-  }
+  virtual void issue(register_set &source_reg);
   virtual void cycle() = 0;
   virtual void active_lanes_in_pipeline() = 0;
 
@@ -1051,6 +1054,8 @@ class simd_function_unit {
   virtual bool can_issue(const warp_inst_t &inst) const {
     return m_dispatch_reg->empty() && !occupied.test(inst.latency);
   }
+  virtual bool is_issue_partitioned() = 0;
+  virtual unsigned get_issue_reg_id() = 0;
   virtual bool stallable() const = 0;
   virtual void print(FILE *fp) const {
     fprintf(fp, "%s dispatch= ", m_name.c_str());
@@ -1070,7 +1075,7 @@ class pipelined_simd_unit : public simd_function_unit {
  public:
   pipelined_simd_unit(register_set *result_port,
                       const shader_core_config *config, unsigned max_latency,
-                      shader_core_ctx *core);
+                      shader_core_ctx *core, unsigned issue_reg_id);
 
   // modifiers
   virtual void cycle();
@@ -1091,6 +1096,8 @@ class pipelined_simd_unit : public simd_function_unit {
   virtual bool can_issue(const warp_inst_t &inst) const {
     return simd_function_unit::can_issue(inst);
   }
+  virtual bool is_issue_partitioned() = 0;
+  unsigned get_issue_reg_id() { return m_issue_reg_id; }
   virtual void print(FILE *fp) const {
     simd_function_unit::print(fp);
     for (int s = m_pipeline_depth - 1; s >= 0; s--) {
@@ -1106,6 +1113,8 @@ class pipelined_simd_unit : public simd_function_unit {
   warp_inst_t **m_pipeline_reg;
   register_set *m_result_port;
   class shader_core_ctx *m_core;
+  unsigned m_issue_reg_id;  // if sub_core_model is enabled we can only issue
+                            // from a subset of operand collectors
 
   unsigned active_insts_in_pipeline;
 };
@@ -1113,7 +1122,7 @@ class pipelined_simd_unit : public simd_function_unit {
 class sfu : public pipelined_simd_unit {
  public:
   sfu(register_set *result_port, const shader_core_config *config,
-      shader_core_ctx *core);
+      shader_core_ctx *core, unsigned issue_reg_id);
   virtual bool can_issue(const warp_inst_t &inst) const {
     switch (inst.op) {
       case SFU_OP:
@@ -1129,12 +1138,13 @@ class sfu : public pipelined_simd_unit {
   }
   virtual void active_lanes_in_pipeline();
   virtual void issue(register_set &source_reg);
+  bool is_issue_partitioned() { return true; }
 };
 
 class dp_unit : public pipelined_simd_unit {
  public:
   dp_unit(register_set *result_port, const shader_core_config *config,
-          shader_core_ctx *core);
+          shader_core_ctx *core, unsigned issue_reg_id);
   virtual bool can_issue(const warp_inst_t &inst) const {
     switch (inst.op) {
       case DP_OP:
@@ -1146,12 +1156,13 @@ class dp_unit : public pipelined_simd_unit {
   }
   virtual void active_lanes_in_pipeline();
   virtual void issue(register_set &source_reg);
+  bool is_issue_partitioned() { return true; }
 };
 
 class tensor_core : public pipelined_simd_unit {
  public:
   tensor_core(register_set *result_port, const shader_core_config *config,
-              shader_core_ctx *core);
+              shader_core_ctx *core, unsigned issue_reg_id);
   virtual bool can_issue(const warp_inst_t &inst) const {
     switch (inst.op) {
       case TENSOR_CORE_OP:
@@ -1163,12 +1174,13 @@ class tensor_core : public pipelined_simd_unit {
   }
   virtual void active_lanes_in_pipeline();
   virtual void issue(register_set &source_reg);
+  bool is_issue_partitioned() { return true; }
 };
 
 class int_unit : public pipelined_simd_unit {
  public:
   int_unit(register_set *result_port, const shader_core_config *config,
-           shader_core_ctx *core);
+           shader_core_ctx *core, unsigned issue_reg_id);
   virtual bool can_issue(const warp_inst_t &inst) const {
     switch (inst.op) {
       case SFU_OP:
@@ -1194,12 +1206,13 @@ class int_unit : public pipelined_simd_unit {
   }
   virtual void active_lanes_in_pipeline();
   virtual void issue(register_set &source_reg);
+  bool is_issue_partitioned() { return true; }
 };
 
 class sp_unit : public pipelined_simd_unit {
  public:
   sp_unit(register_set *result_port, const shader_core_config *config,
-          shader_core_ctx *core);
+          shader_core_ctx *core, unsigned issue_reg_id);
   virtual bool can_issue(const warp_inst_t &inst) const {
     switch (inst.op) {
       case SFU_OP:
@@ -1223,13 +1236,14 @@ class sp_unit : public pipelined_simd_unit {
   }
   virtual void active_lanes_in_pipeline();
   virtual void issue(register_set &source_reg);
+  bool is_issue_partitioned() { return true; }
 };
 
 class specialized_unit : public pipelined_simd_unit {
  public:
   specialized_unit(register_set *result_port, const shader_core_config *config,
                    shader_core_ctx *core, unsigned supported_op,
-                   char *unit_name, unsigned latency);
+                   char *unit_name, unsigned latency, unsigned issue_reg_id);
   virtual bool can_issue(const warp_inst_t &inst) const {
     if (inst.op != m_supported_op) {
       return false;
@@ -1238,6 +1252,7 @@ class specialized_unit : public pipelined_simd_unit {
   }
   virtual void active_lanes_in_pipeline();
   virtual void issue(register_set &source_reg);
+  bool is_issue_partitioned() { return true; }
 
  private:
   unsigned m_supported_op;
@@ -1259,6 +1274,7 @@ class ldst_unit : public pipelined_simd_unit {
 
   // modifiers
   virtual void issue(register_set &inst);
+  bool is_issue_partitioned() { return false; }
   virtual void cycle();
 
   void fill(mem_fetch *mf);
@@ -2133,8 +2149,8 @@ class shader_core_ctx : public core_t {
   friend class TwoLevelScheduler;
   friend class LooseRoundRobbinScheduler;
   virtual void issue_warp(register_set &warp, const warp_inst_t *pI,
-                  const active_mask_t &active_mask, unsigned warp_id,
-                  unsigned sch_id);
+                          const active_mask_t &active_mask, unsigned warp_id,
+                          unsigned sch_id);
 
   void create_front_pipeline();
   void create_schedulers();