search

The-OpenROAD-Project / RePlAce

RePlAce global placement tool
BSD 3-Clause "New" or "Revised" License
206 stars 75 forks source link

negative HPWL #62

Closed MartinGeisse closed 4 years ago

MartinGeisse commented 4 years ago

The following design crashes replace (I just copied it in place of the gcd demo; I know that some other tool is going to choke on the missing clk signal later in the toolchain, but replace seems to have another problem):

module gcd( input[3:0] encoded, output reg[6:0] decoded ); always @(*) begin case (encoded) 4'd0: decoded <= 7'b1110111; 4'd1: decoded <= 7'b0100100; 4'd2: decoded <= 7'b1011101; 4'd3: decoded <= 7'b1101101; 4'd4: decoded <= 7'b0101110; 4'd5: decoded <= 7'b1101011; 4'd6: decoded <= 7'b1111011; 4'd7: decoded <= 7'b0100101; 4'd8: decoded <= 7'b1111111; 4'd9: decoded <= 7'b1101111; 4'd10: decoded <= 7'b0111111; 4'd11: decoded <= 7'b1111010; 4'd12: decoded <= 7'b1010011; 4'd13: decoded <= 7'b1111100; 4'd14: decoded <= 7'b1011011; 4'd15: decoded <= 7'b0011011; default: decoded <= 0; endcase end endmodule

At some point, this outputs:

[INFO] Nesterov: 10 OverFlow: 0.5945 ScaledHpwl: 5583.1606 ... (repeated many times, but the OverFlow and ScaledHpwl values don't change) [INFO] Nesterov: 2100 OverFlow: 0.5945 ScaledHpwl: 5583.1606 [INFO] Timing: WNS = 1e+30 [INFO] Timing: TNS = 0 [INFO] Nesterov: 2110 OverFlow: 0.4528 ScaledHpwl: -40109.2656 ... (repeated many times, but the OverFlow and ScaledHpwl values don't change) [INFO] Nesterov: 2500 OverFlow: 0.4528 ScaledHpwl: -40109.2656 [INFO] Timing: WNS = 1e+30 [INFO] Timing: TNS = 0 NEGATIVE HPWL ERROR! 00 72.0000 587.4428 NEGATIVE HPWL ERROR! 00 79.2643 592.2643 NEGATIVE HPWL ERROR! 00-80.1800 -79.8000

I have seen a commit that mentions the "negative HPWL" error, but the commit is stuck in the alpha2.1 branch. Is that commit relevant to the error I'm getting, and is the commit planned to be merged at some point in the future?

mgwoo commented 4 years ago

That's a divergence problem in the previous RePlAce. Usually, very small designs like gcd are not converged well. The RePlAce engine recommends having at least 1K instances to place cells. I think you could tune the "-bin_grid_count" as higher(64, 128), or tune the -init_density_penalty(or -init_lambda_coeff in previous replace) parameter as higher(0.01, 0.1).

MartinGeisse commented 4 years ago

Thanks for your quick reply. I suspected this to be a problem with small designs, but then couldn't reproduce that reliably. I'll try something larger for experiments instead.

MartinGeisse commented 4 years ago

"-bin_grid_count 64" did the job. Thanks again!