UVM中通过静态类实现对全局资源实现管理

Static class

Systemverilog中可以使用static修饰变量，方法，得到静态变量和静态函数。static也可以直接修饰class，获得静态类。但是static class class_name .... endclass这种直接定义一个静态类的方式是不允许的，只可以在一个class的内部，通过static class_name的方式声明一个静态类。这一点SV与Jave相同。（Java中的 static class[2])

静态类的特征：

1.相比automatic class, 静态类被new创建后，其对应的内存空间一直存在，伴随仿真结束。配合单例模式使用，则是全局唯一类的实例，可以被当作全局变量使用。2.相比静态变量，静态类可以封装更多内容，充当资源容器的角色。

所以UVM中对全局的资源管理，都是通过静态类实现的。

uvm_pool

uvm_pool#(type KEY=int, T=uvm_void)是一个参数化的类，相当于SV中的联合数组。参数KEY是联合数组的索引，参数T是联合数组存储的变量类型。uvm_pool封装了原有联合数组的操作方法，供访问内部的联合数组pool。m_global_pool是uvm_pool #(KEY, T)类型的静态类，通过静态方法get_global_pool可以获得唯一的实例。uvm_queue #(T)和uvm_pool#(type KEY=int, T=uvm_void)类似，实现一个class-based dynamic queue.

# uvm_pool.svh

class uvm_pool #(type KEY=int, T=uvm_void) extends uvm_object;

  const static string type_name = "uvm_pool";

  typedef uvm_pool #(KEY,T) this_type;

  static protected this_type m_global_pool;
  protected T pool[KEY];

  // Function: new
  // Creates a new pool with the given ~name~.
  function new (string name="");
    super.new(name);
  endfunction
  // Function: get_global_pool
  // Returns the singleton global pool for the item type, T. 
  // This allows items to be shared amongst components throughout the
  // verification environment.
  static function this_type get_global_pool ();
    if (m_global_pool==null)
      m_global_pool = new("pool");
    return m_global_pool;
  endfunction

  // Function: get_global
  // Returns the specified item instance from the global item pool. 
  static function T get_global (KEY key);
    this_type gpool;
    gpool = get_global_pool(); 
    return gpool.get(key);
  endfunction
  // Function: get
  // Returns the item with the given ~key~.
  // If no item exists by that key, a new item is created with that key
  // and returned.
  virtual function T get (KEY key);
    if (!pool.exists(key)) begin
      T default_value;
      pool[key] = default_value;
    end
    return pool[key];
  endfunction

  virtual function void add (KEY key, T item);
    pool[key] = item;
  endfunction

  virtual function int num ();
    return pool.num();
  endfunction

......
...... 
......
endclass

uvm_event_pool

uvm_object_string_pool的KEY是string类型。uvm_event_pool由uvm_object_string_pool #(uvm_event)声明，KEY是string类型，T是uvm_event类型。uvm_event是sv中event的class warpper，内建了很多方法。

# uvm_pool.svh

class uvm_object_string_pool #(type T=uvm_object) extends uvm_pool #(string,T);
......
typedef class uvm_barrier;
typedef class uvm_event;

typedef uvm_object_string_pool #(uvm_barrier) uvm_barrier_pool;
typedef uvm_object_string_pool #(uvm_event) uvm_event_pool;

# uvm_event.svh

//------------------------------------------------------------------------------
// CLASS: uvm_event
// The uvm_event class is a wrapper class around the SystemVerilog event
// construct.  It provides some additional services such as setting callbacks
// and maintaining the number of waiters.
//------------------------------------------------------------------------------
class uvm_event extends uvm_object;

  const static string type_name = "uvm_event";

  local event      m_event;
  local int        num_waiters;
  local bit        on;
  local time       trigger_time=0;
  local uvm_object trigger_data;
  local uvm_event_callback  callbacks[$];

  function new (string name="");
    super.new(name);
  endfunction  
......
......
......
  virtual task wait_on (bit delta=0);
  virtual task wait_off (bit delta=0);
  virtual task wait_trigger ();
  virtual task wait_ptrigger ();
  virtual task wait_trigger_data (output uvm_object data);
  virtual task wait_ptrigger_data (output uvm_object data);
  virtual function void trigger (uvm_object data=null);
  virtual function uvm_object get_trigger_data ();
  virtual function time get_trigger_time ();
  virtual function void add_callback (uvm_event_callback cb, bit append=1);
  virtual function int get_num_waiters ();
......
endclass : uvm_event

使用

uvm_event_pool作为全局唯一的uvm_evnet的资源池，可以在不同component中调用，实现事件的同步功能。比如sequence和scoreboard需要通过event来握手：

# event_sequence.sv
class event_sequence extends uvm_sequence;
    ......
    uvm_event_pool events_pool;
    uvm_event sync_e;
    ......
    function new (string name = "event_sequence");
        events_pool = uvm_event_pool::get_global_pool();
        sync_e = events_pool.get("sync_e");
        // 也可以直接调用 sync_e = uvm_event_pool::get_global("sync_e")；
    endfunction
    ......
    virtual task body();
    ......
        sync_e.trigger();
    ......
    endtask
endclass

# event_scb.sv
class event_scb extends uvm_scoreboard;
    ......
    uvm_event_pool events_pool;
    uvm_event sync_e;
    ......
    function new (string name = "event_sequence");
        events_pool = uvm_event_pool::get_global_pool();
        sync_e = events_pool.get("sync_e");
    endfunction
    ......
    virtual task wait_sync_e();
    ......
        sync_e.wait_trigger();
    ......
    endtask
endclass

无论是uvm_component还是uvm_object，是否在同一个package中，都可以通过uvm_evnet_pool::get_global_pool()获得全局唯一的uvm_event资源池。上述的uvm_event并不需要new(),因为调用get()函数时会自动创建：

# uvm_object_string_pool #(uvm_event)

  // Function: get
  // Returns the object item at the given string ~key~.
  // If no item exists by the given ~key~, a new item is created for that key
  // and returned.
  virtual function T get (string key);
    if (!pool.exists(key))
      pool[key] = new (key);
    return pool[key];
  endfunction

uvm_event不仅可以用于event同步，也可以传递一些简单的数据。提供了两个方法：wait_trigger_data (output uvm_object data)trigger (uvm_object data)，可以传递uvm_object类型的数据。也可以add_callback加入回调函数。uvm_event[3]uvm中还提供了uvm_barrier用于多个组件之间的同步，uvm_barrier_pool存放所有的uvm_barrier。

uvm_config_db

结构

uvm_resource#(type T): 各类资源（*scalar objectsclass handlesqueueslistsvirtual interfaces等*）的一个wrapper，内部成员变量val是一个type T类型的句柄，通过调用write函数，将val指向资源的实例（ *!!! 对于class handlesvirtual interfaces，val看作句柄指向实例；对于queueslists等，val是被直接赋值 !!!*）。set函数将uvm_resource加入全局资源池uvm_resources。

class uvm_resource #(type T=int) extends uvm_resource_base;
  typedef uvm_resource#(T) this_type;
  // singleton handle that represents the type of this resource
  static this_type my_type = get_type();
  // Can't be rand since things like rand strings are not legal.
  protected T val;
  ......
  // Function: get_type
  // Static function that returns the static type handle.  The return
  // type is this_type, which is the type of the parameterized class.
  static function this_type get_type();
    if(my_type == null)
      my_type = new();
    return my_type;
  endfunction
  ......
  // Function: set
  // Simply put this resource into the global resource pool
  function void set();
    uvm_resource_pool rp = uvm_resource_pool::get();
    rp.set(this);
  endfunction
  ......
  // Function: write
  // Modify the object stored in this resource container.  If the
  // resource is read-only then issue an error message and return
  // without modifying the object in the container.  If the resource is
  // not read-only and an ~accessor~ object has been supplied then also
  // update the accessor record.  Lastly, replace the object value in
  // the container with the value supplied as the argument, ~t~, and
  // release any processes blocked on
  // .  If the value to be written is
  // the same as the value already present in the resource then the
  // write is not done.  That also means that the accessor record is not
  // updated and the modified bit is not set.
  function void write(T t, uvm_object accessor = null);
    if(is_read_only()) begin
      uvm_report_error("resource", $sformatf("resource %s is read only -- cannot modify", get_name()));
      return;
    end
    // Set the modified bit and record the transaction only if the value
    // has actually changed.
    if(val == t)
      return;
    record_write_access(accessor);
    // set the value and set the dirty bit
    val = t;
    modified = 1;
  endfunction
  ......

uvm_resource_pool: 存放所有的uvm_resource, 全局唯一的实例uvm_resources。uvm_resource_pool::get()得到uvm_resources。（注意：这里的new函数使用了local修饰，而uvm_event_pool中的new没有被local修饰。所以uvm_resource_pool和uvm_root拥有真正的全局唯一实例，而uvm_event_pool可以在外部手动new创建多个实例。）

rtab[string]ttab[uvm_resource_base]分别以nametype为索引存放uvm_resource。

//----------------------------------------------------------------------
// Class: uvm_resource_pool
//
// The global (singleton) resource database.
//
// Each resource is stored both by primary name and by type handle.  The
// resource pool contains two associative arrays, one with name as the
// key and one with the type handle as the key.  Each associative array
// contains a queue of resources.  Each resource has a regular
// expression that represents the set of scopes over with it is visible.
//
//|  +------+------------+                          +------------+------+
//|  | name | rsrc queue |                          | rsrc queue | type |
//|  +------+------------+                          +------------+------+
//|  |      |            |                          |            |      |
//|  +------+------------+                  +-+-+   +------------+------+
//|  |      |            |                  | | |<--+---*        |  T   |
//|  +------+------------+   +-+-+          +-+-+   +------------+------+
//|  |  A   |        *---+-->| | |           |      |            |      |
//|  +------+------------+   +-+-+           |      +------------+------+
//|  |      |            |      |            |      |            |      |
//|  +------+------------+      +-------+  +-+      +------------+------+
//|  |      |            |              |  |        |            |      |
//|  +------+------------+              |  |        +------------+------+
//|  |      |            |              V  V        |            |      |
//|  +------+------------+            +------+      +------------+------+
//|  |      |            |            | rsrc |      |            |      |
//|  +------+------------+            +------+      +------------+------+
//
// The above diagrams illustrates how a resource whose name is A and
// type is T is stored in the pool.  The pool contains an entry in the
// type map for type T and an entry in the name map for name A.  The
// queues in each of the arrays each contain an entry for the resource A
// whose type is T.  The name map can contain in its queue other
// resources whose name is A which may or may not have the same type as
// our resource A.  Similarly, the type map can contain in its queue
// other resources whose type is T and whose name may or may not be A.
//
// Resources are added to the pool by calling ; they are retrieved
// from the pool by calling  or .  When an object 
// creates a new resource and calls  the resource is made available to be
// retrieved by other objects outside of itsef; an object gets a
// resource when it wants to access a resource not currently available
// in its scope.

//----------------------------------------------------------------------
// static global resource pool handle
//----------------------------------------------------------------------
const uvm_resource_pool uvm_resources = uvm_resource_pool::get();

class uvm_resource_pool;

  static bit m_has_wildcard_names;
  static local uvm_resource_pool rp = get();

  uvm_resource_types::rsrc_q_t rtab [string];
  uvm_resource_types::rsrc_q_t ttab [uvm_resource_base];
 ......
  local function new();
  endfunction
  // Function: get
  //
  // Returns the singleton handle to the resource pool
  static function uvm_resource_pool get();
    if(rp == null)
      rp = new();
    return rp;
  endfunction
  ......
  // Function: get_by_name
  //
  // Lookup a resource by ~name~, ~scope~, and ~type_handle~.  Whether
  // the get succeeds or fails, save a record of the get attempt.  The
  // ~rpterr~ flag indicates whether to report errors or not.
  // Essentially, it serves as a verbose flag.  If set then the spell
  // checker will be invoked and warnings about multiple resources will
  // be produced.

  function uvm_resource_base get_by_name(string scope = "",
                                         string name,
                                         uvm_resource_base type_handle,
                                         bit rpterr = 1);

  // Function: get_by_type
  //
  // Lookup a resource by ~type_handle~ and ~scope~.  Insert a record into
  // the get history list whether or not the get succeeded.

  function uvm_resource_base get_by_type(string scope = "",
                                         uvm_resource_base type_handle);
  ......
......
endclass

uvm_resource_db#(type T):访问资源池的接口，内部都是static function。

uvm_config_db#(type T):继承于uvm_resource_db#(type T)，进行了一些功能扩展。内部都是static function。

接下来重点看一下uvm_config_db传入的4个参数：

cntxt:uvm_component类型，context上下文的含义，由cntxt可以确定资源的优先级，对同一个资源set，从UVM树顶层往下，优先级依次降低。cntxt = null时，为uvm_root::get()，优先级最高。

inst_name:string类型，通过{cntxt.get_full_name(), ".", inst_name}组成的字符串设置资源的scope。内部调用DPI，支持通配符。只有get中的scope可以和set中的scope匹配上，才可以正常访问资源。

field_name: 资源名，scope内的资源。也支持通配符，一般setget设置相同即可。

value:type T类型，资源的类型。get中的是inout，调用get函数先给value赋值，结束后再将val传出，但是get中并没有处理传入value的代码，不知道为什么这样用。

class uvm_config_db#(type T=int) extends uvm_resource_db#(T);
......
  static uvm_pool#(string,uvm_resource#(T)) m_rsc[uvm_component];
......

   static function bit get(uvm_component cntxt,
                          string inst_name,
                          string field_name,
                          inout T value);
//TBD: add file/line
    int unsigned p;
    uvm_resource#(T) r, rt;
    uvm_resource_pool rp = uvm_resource_pool::get();
    uvm_resource_types::rsrc_q_t rq;

    if(cntxt == null) 
      cntxt = uvm_root::get();
    if(inst_name == "") 
      inst_name = cntxt.get_full_name();
    else if(cntxt.get_full_name() != "") 
      inst_name = {cntxt.get_full_name(), ".", inst_name};

   ......

    value = r.read(cntxt);

    return 1;
  endfunction

   static function void set(uvm_component cntxt,
                           string inst_name,
                           string field_name,
                           T value);

    uvm_root top;
    uvm_phase curr_phase;
    uvm_resource#(T) r;
    bit exists;
    string lookup;
    uvm_pool#(string,uvm_resource#(T)) pool;

    //take care of random stability during allocation
    process p = process::self();
    string rstate = p.get_randstate();
    top = uvm_root::get();
    curr_phase = top.m_current_phase;

    if(cntxt == null) 
      cntxt = top;
    if(inst_name == "") 
      inst_name = cntxt.get_full_name();
    else if(cntxt.get_full_name() != "") 
      inst_name = {cntxt.get_full_name(), ".", inst_name};

    ......
    r.write(value, cntxt);
    ......
  endfunction

有一种特殊的情况，就是cntxt = null， 而uvm_root::get().get_full_name()为空字符串。其他情况都是以uvm_test_top为开头设置scope。

Verdi也提供了GUI界面，可以看到setget的具体内容，方便debug：

使用

uvm_config_db机制本质就是在一个地方创建资源，通过set以uvm_resource的形式存储到uvm_resource_pool上。然后在另一个地方，通过get获得之前创建的资源。

uvm_config_db机制还为资源加上了scope限制资源的访问，保证数据安全；precedence设置优先级；override资源重写；record记录资源访问历史用于debug等功能。

这里的资源可以是scalar objectsclass handlesqueueslistsvirtual interfaces等。

队列，数组类型

对于队列，数组的传递，int val_q[$]直接通过uvm_config_db#(int) :: set(this, "env.i_agt.drv","val_q",val_q);的方式会编译报错，因为uvm_config_db#(type T=int)中type是int类型，而不是队列类型。需要typedef int t_q[$] ;定义队列类型。t_q val_q；uvm_config_db#(t_q) :: set(this, "env.i_agt.drv","val_q",val_q)；

还有另一种方式，通过class封装队列（uvm已提供了uvm_queue#(type T）。这种方式的好处是，队列在一端改变，另一段也可以看到队列的修改，因为变量对资源的访问是通过句柄的形式，指向同一处资源。而直接传递队列，一端修改，另一端看不到，因为两端的队列是各自class scope中的两份不同的数据，只有再执行一次setget操作才会获得队列的新内容。

sequence中的资源访问

UVM中sequence不属uvm_component，存在固定的生命周期，对资源的访问，分为直接和间接两种类型。以reg_model为例，reg_model在env中create, sequence通过reg_model访问寄存器。

总结如下5种方式：

1.sequence没有出现在树形结构中，难以确定路径参数，可以通过setnull, ""设定全局scope，这样sequence在哪里都可以访问资源。注意：sequence没有phase，get可以放在pre_bodybodypre_randomizepre_start中，只要在调用reg_model之前get被调用即可。

# env
top_reg_model m_regmodel；
m_regmodel = top_reg_model::create("m_regmodel");
uvm_config_db #(top_reg_model)::set(null， ""，"reg_model"，m_regmodel);

# seq
top_reg_model m_regmodel；
uvm_config_db #(top_reg_model)::get(null， ""，"reg_model"，m_regmodel);

// 为什么说 null, ""的组合是全局scope呢？
// TODO

1.sequence的路径可以通过get_full_name()获得。

# env
top_reg_model m_regmodel；
m_regmodel = top_reg_model::create("m_regmodel");
uvm_config_db #(top_reg_model)::set(this， " agt.sqr.* "，"reg_model"，m_regmodel);

# seq
top_reg_model m_regmodel；
uvm_config_db #(top_reg_model)::get(null， get_full_name()，"reg_model"，m_regmodel);

1.sequence虽然不在树形结构上，但是其内部成员变量m_sequencer在树形结构上，可以通过m_sequencer间接访问资源。

# env
top_reg_model m_regmodel；
m_regmodel = top_reg_model::create("m_regmodel");
uvm_config_db #(top_reg_model)::set(this， " agt.sqr"，"reg_model"，m_regmodel);

# seq
top_reg_model m_regmodel；
uvm_config_db #(top_reg_model)::get(m_sequencer， ""，"reg_model"，m_regmodel);

1.直接通过p_sequener访问资源。sequencer中get获得资源，sequence中调用p_sequencer.m_regmodel访问资源。2.直接赋值，在test中例化sequence时，seq.m_regmodel = m_regmodel。也可以封装到sequence中的function。

#seq
virtual function set_regmodel(top_reg_model m_regmodel);
    this.m_regmodel = m_regmodel;
endfunction

当然也可以封装到test中的function:

#test
virtual function set_regmodel(sequence_baes seq);
    seq.m_regmodel = m_regmodel;
endfunction

上述方法，本质都是一样的，将m_model句柄指向在env中创建的top_reg_model实例。

自定义资源池

我们也可以模仿uvm_event_pooluvm_resource_pool的方式，自定义一个资源池。

不使用uvm_config_db，自定义一个interface pool，示例如下（UVM实战 /ch10/section10.6/10.6.2）：

# interface_pool

class if_object extends uvm_object;

   `uvm_object_utils(if_object)
   function new(string name = "if_object");
      super.new(name);
   endfunction

   static if_object me;

   static function if_object get();
      if(me == null) begin
         me = new("me");
      end
      return me;
   endfunction

   virtual my_if input_vif0;
   virtual my_if output_vif0;
   virtual my_if input_vif1;
   virtual my_if output_vif1;
endclass

# top_tb.sv 
module top_tb;
......
my_if input_if0(clk, rst_n);
my_if input_if1(clk, rst_n);
my_if output_if0(clk, rst_n);
my_if output_if1(clk, rst_n);

......
initial begin
   if_object if_obj; 
   if_obj = if_object::get();
   if_obj.input_vif0 = input_if0;
   if_obj.input_vif1 = input_if1;
   if_obj.output_vif0 = output_if0;
   if_obj.output_vif1 = output_if1;
end
endmodule

# base_test.sv
class base_test extends uvm_test;

   my_env         env0;
   my_env         env1;
   my_vsqr        v_sqr;   

   function new(string name = "base_test", uvm_component parent = null);
      super.new(name,parent);
   endfunction

   extern virtual function void build_phase(uvm_phase phase);
   extern virtual function void connect_phase(uvm_phase phase);
   extern virtual function void report_phase(uvm_phase phase);
   `uvm_component_utils(base_test)
endclass

......

function void base_test::connect_phase(uvm_phase phase);
   if_object if_obj; 
   if_obj = if_object::get();
   v_sqr.p_sqr0 = env0.i_agt.sqr;
   v_sqr.p_sqr1 = env1.i_agt.sqr;
   env0.i_agt.drv.vif = if_obj.input_vif0;
   env0.i_agt.mon.vif = if_obj.input_vif0;
   env0.o_agt.mon.vif = if_obj.output_vif0;
   env1.i_agt.drv.vif = if_obj.input_vif1;
   env1.i_agt.mon.vif = if_obj.input_vif1;
   env1.o_agt.mon.vif = if_obj.output_vif1;
endfunction

平台中所有使用uvm_config_db的地方都可以通过这种方式替代，当然并不建议这样使用，uvm_config_db提供了更丰富的功能。这种方式的一个优点是编写错误可以在编译阶段发现，而uvm_config_db中字符串错误不容易发现。

wait_modified

uvm_config_db不仅可以共享资源，也可以像uvm_event那样，用于事件的同步，可以通过wait_modified实现。

drv0_seq在get之前，case0_vseq中必须先set，否则wait_modified会一直阻塞。因为wait_modified中调用@waiter.trigger，trigger是event类型；在set的最后，->w.trigger会触发该event。平台中所有的等待事件都放在了m_waiters[string]中，其中的索引sting对应getset函数中第三个参数field_name。

class drv0_seq extends uvm_sequence #(my_transaction);
   ......
      virtual task body();
      bit send_en = 1;
      fork
         while(1) begin
            uvm_config_db#(bit)::wait_modified(null, get_full_name(), "send_en");
            void'(uvm_config_db#(bit)::get(null, get_full_name, "send_en", send_en)); 
            `uvm_info("drv0_seq", $sformatf("send_en value modified, the new value is %0d", send_en), UVM_LOW)
         end
      join_none
      repeat (10) begin
         `uvm_do(m_trans)
      end
   endtask
endclass

class case0_vseq extends uvm_sequence;
   ......
   virtual task body();
      my_transaction tr;
      drv0_seq seq0;
      drv1_seq seq1;
      if(starting_phase != null) 
         starting_phase.raise_objection(this);
      fork
         `uvm_do_on(seq0, p_sequencer.p_sqr0);
         `uvm_do_on(seq1, p_sequencer.p_sqr1);
         begin
            #10000;
            uvm_config_db#(bit)::set(uvm_root::get(), "uvm_test_top.v_sqr.*", "send_en", 0);
            #10000;
            uvm_config_db#(bit)::set(uvm_root::get(), "uvm_test_top.v_sqr.*", "send_en", 1);
         end
      join 
      #100;
      if(starting_phase != null) 
         starting_phase.drop_objection(this);
   endtask
endclass

class m_uvm_waiter;
  string inst_name;
  string field_name;
  event trigger;
  function new (string inst_name, string field_name);
    this.inst_name = inst_name;
    this.field_name = field_name;
  endfunction
endclass

class uvm_config_db#(type T=int) extends uvm_resource_db#(T);
......
  // Internal waiter list for wait_modified
  static local uvm_queue#(m_uvm_waiter) m_waiters[string];
......
  static task wait_modified(uvm_component cntxt, string inst_name,
      string field_name);
    process p = process::self();
    string rstate = p.get_randstate();
    m_uvm_waiter waiter;

    if(cntxt == null)
      cntxt = uvm_root::get();
    if(cntxt != uvm_root::get()) begin
      if(inst_name != "")
        inst_name = {cntxt.get_full_name(),".",inst_name};
      else
        inst_name = cntxt.get_full_name();
    end

    waiter = new(inst_name, field_name);

    if(!m_waiters.exists(field_name))
      m_waiters[field_name] = new;
    m_waiters[field_name].push_back(waiter);

    p.set_randstate(rstate);

    // wait on the waiter to trigger
    @waiter.trigger;

    // Remove the waiter from the waiter list 
    for(int i=0; iw.trigger;  
      end
    end

   ......
  endfunction

UVM_REGEX_NO_DPI

大量 ”滥用“uvm_config_db配置访问资源，会降低仿真速度。db:set/get()会自动扩展正则匹配，uvm_re_match()调用DPI的方式，增加仿真内核切换时间，每一次get()需要遍历set()的所有资源。而且不是简单的线性降低仿真速度。可以优化原有代码，减少set/get调用；使用UVM_REGEX_NO_DPI禁用正则匹配，仅仅使用*?+这种简单的通配符。

如果在SoC级别的平台，发现build_phase占用太多时间，可以参考SNUG 2018 Auditing the UVM Resource Database[4]进行优化，对源码进行相应修改，更好的Auditing平台的资源配置。

审核编辑：刘清