Electronic – Modelsim – Weird verification problem with DDR and Xilinx UNISIM

modelsimsystem-verilogverificationvhdlxilinx

I am doing verification of VHDL component using OVM and ran into serious problems. I have found that problem is in one specific component and created environment specifically for it. It's a RGMII to internal interface converter (from DDR to SDR) using Xilinx UNISIM components to simulate special FPGA chip parts behavior.

BUFG_inst : BUFG
port map (
    O => INTERNAL_CLK,
    I => RGMII_CLK
);

IDDRDATA0_inst: IDDR
generic map (
    DDR_CLK_EDGE => "SAME_EDGE_PIPELINED",
    INIT_Q1 => '0',
    INIT_Q2 => '0',
    SRTYPE  => "SYNC"
)
port map (
    Q1 => data(0),
    Q2 => data(4),
    C  => INTERNAL_CLK,
    CE => CE,
    D  => RGMII_DATA(0),
    R  => RST,
    S  => '0'
);
-- several more instantiations

I generate transactions and send them to DUT using driver:

forever
begin
my_transaction tx;

@(posedge dut_vi.clock); 
seq_item_port.get(tx);

dut_vi.ctl   = tx.ctl;
dut_vi.reset = tx.reset;
dut_vi.data  = tx.data;

@(negedge dut_vi.clock); 
seq_item_port.get(tx);

dut_vi.ctl   = tx.ctl;
dut_vi.reset = tx.reset;
dut_vi.data  = tx.data;
end

In Modelsim wave, this input (watching DUT input ports) is just fine.. But internally, it takes previous value of data – for example previous value of CTL was 0, on rising edge it's 1, but DUT samples with rising edge that previous 0 as it didn't actualize inside modelsim.

I tried using "little hack":

TMP_CLK <= RGMII_CLK after 1ps;

BUFG_inst : BUFG
port map (
    O => INTERNAL_CLK,
    I => TMP_CLK
);

It solved the problem (partially). DUT was now sampling correct values.. But it created problem with output and monitor. Now everything is correct in wave, but monitor sees wrong values (but only sometimes – usually with beginning of packet – valid signal changed from 0 to 1). So it's similar problem as with input. My monitor looks like this:

forever
begin
dut_out_transaction tx;

@(posedge dut_out_vi.clock);          
tx = dut_out_transaction::type_id::create("tx");

tx.dv    = dut_out_vi.dv;
tx.err   = dut_out_vi.err;
tx.sof   = dut_out_vi.sof;
tx.eof   = dut_out_vi.eof;
tx.data  = dut_out_vi.data;

$display("DUT_OUT EOF: %b", dut_out_vi.eof);

if (dut_out_vi.eof == 1)  // this never happens
    ovm_report_info("out_monitor", "EOF");

aport.write(tx);
end

Since it's similar bug, I tried similar hack – I added #1ps right after @(posedge dut_out_vi.clock); and it truly helped.. But new problem arisen – now monitor never sees EOF signal in 1 even though it is in 1 in wave several times.

I use Modelsim 10.5 with resolution of 1ps. Any ideas what could be causing it or how to fix it?

Best Answer

I finally solved it - problem was in Xilinx unisim library. For some reason, there is 100ps delay (which shouldn't be in behavioral simulation) in IDDR.. So solution is to wait in monitor (I have actually two - for input and output of DUT):

forever
begin
  my_transaction tx;

  @(posedge dut_vi.clock);
  #1ps
  tx = my_transaction::type_id::create("tx");

  tx.ctl   = dut_vi.ctl;
  tx.reset = dut_vi.reset;
  tx.data  = dut_vi.data;

  aport.write(tx);

  @(negedge dut_vi.clock);
  #1ps
  tx = my_transaction::type_id::create("tx");

  tx.ctl   = dut_vi.ctl;
  tx.reset = dut_vi.reset;
  tx.data  = dut_vi.data;

  aport.write(tx);
end

and second one:

forever
begin
  dut_out_transaction tx;

  @(posedge dut_out_vi.clock);
  #101ps   // simulation hack
  tx = dut_out_transaction::type_id::create("tx");

  tx.dv    = dut_out_vi.dv;
  tx.err   = dut_out_vi.err;
  tx.sof   = dut_out_vi.sof;
  tx.eof   = dut_out_vi.eof;
  tx.data  = dut_out_vi.data;
end

In the input monitor, I have to wait 1ps (didn't figure out why) and in output, I have to wait 101ps due to delay inside IDDR (I checked the code). This way, scoreboard gets correct data.

Related Topic