Site Home Archive Home FAQ Home How to search the Archive How to Navigate the Archive
Compare FPGA features and resources
Threads starting:
Authors:A B C D E F G H I J K L M N O P Q R S T U V W X Y Z
On Mar 1, 12:59=A0pm, a s <nospa...@gmail.com> wrote: > Dear all, > > I have come up with 2 solutions in VHDL, how to count number of bits > in input data. > The thing I don't understand is why the 2 solutions produce different > results, at least with Xilinx ISE and its XST. > There is quite a substantial difference in required number of slices/ > LUTs. > > 1. solution with unrolled loop: =A0 =A0 =A0 =A041 slices, =A073 LUTs > 2. solution with loop: =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A054 slices, = 100 LUTs > > The entity of both architectures is the same: > > entity one_count is > =A0 Port ( din : in =A0STD_LOGIC_vector(31 downto 0); > =A0 =A0 =A0 =A0 =A0dout : out =A0STD_LOGIC_vector(5 downto 0) > =A0 =A0 =A0 =A0 ); > end one_count; > > The architecture with an unrolled loop is the following: > > library IEEE; > use IEEE.STD_LOGIC_1164.ALL; > use IEEE.NUMERIC_STD.ALL; > > entity one_count is > =A0 Port ( din : in =A0STD_LOGIC_vector(31 downto 0); > =A0 =A0 =A0 =A0 =A0dout : out =A0STD_LOGIC_vector(5 downto 0) > =A0 =A0 =A0 =A0 ); > end one_count; > > architecture one_count_unrolled_arch of one_count is > > =A0 signal =A0cnt : integer range 0 to 32; > > begin > > =A0 =A0cnt <=3D to_integer(unsigned(din( 0 downto =A00))) + > =A0 =A0 =A0 =A0 =A0 to_integer(unsigned(din( 1 downto =A01))) + > =A0 =A0 =A0 =A0 =A0 to_integer(unsigned(din( 2 downto =A02))) + > =A0 =A0 =A0 =A0 =A0 to_integer(unsigned(din( 3 downto =A03))) + > =A0 =A0 =A0 =A0 =A0 to_integer(unsigned(din( 4 downto =A04))) + > =A0 =A0 =A0 =A0 =A0 to_integer(unsigned(din( 5 downto =A05))) + > =A0 =A0 =A0 =A0 =A0 to_integer(unsigned(din( 6 downto =A06))) + > =A0 =A0 =A0 =A0 =A0 to_integer(unsigned(din( 7 downto =A07))) + > =A0 =A0 =A0 =A0 =A0 to_integer(unsigned(din( 8 downto =A08))) + > =A0 =A0 =A0 =A0 =A0 to_integer(unsigned(din( 9 downto =A09))) + > =A0 =A0 =A0 =A0 =A0 to_integer(unsigned(din(10 downto 10))) + > =A0 =A0 =A0 =A0 =A0 to_integer(unsigned(din(11 downto 11))) + > =A0 =A0 =A0 =A0 =A0 to_integer(unsigned(din(12 downto 12))) + > =A0 =A0 =A0 =A0 =A0 to_integer(unsigned(din(13 downto 13))) + > =A0 =A0 =A0 =A0 =A0 to_integer(unsigned(din(14 downto 14))) + > =A0 =A0 =A0 =A0 =A0 to_integer(unsigned(din(15 downto 15))) + > =A0 =A0 =A0 =A0 =A0 to_integer(unsigned(din(16 downto 16))) + > =A0 =A0 =A0 =A0 =A0 to_integer(unsigned(din(17 downto 17))) + > =A0 =A0 =A0 =A0 =A0 to_integer(unsigned(din(18 downto 18))) + > =A0 =A0 =A0 =A0 =A0 to_integer(unsigned(din(19 downto 19))) + > =A0 =A0 =A0 =A0 =A0 to_integer(unsigned(din(20 downto 20))) + > =A0 =A0 =A0 =A0 =A0 to_integer(unsigned(din(21 downto 21))) + > =A0 =A0 =A0 =A0 =A0 to_integer(unsigned(din(22 downto 22))) + > =A0 =A0 =A0 =A0 =A0 to_integer(unsigned(din(23 downto 23))) + > =A0 =A0 =A0 =A0 =A0 to_integer(unsigned(din(24 downto 24))) + > =A0 =A0 =A0 =A0 =A0 to_integer(unsigned(din(25 downto 25))) + > =A0 =A0 =A0 =A0 =A0 to_integer(unsigned(din(26 downto 26))) + > =A0 =A0 =A0 =A0 =A0 to_integer(unsigned(din(27 downto 27))) + > =A0 =A0 =A0 =A0 =A0 to_integer(unsigned(din(28 downto 28))) + > =A0 =A0 =A0 =A0 =A0 to_integer(unsigned(din(29 downto 29))) + > =A0 =A0 =A0 =A0 =A0 to_integer(unsigned(din(30 downto 30))) + > =A0 =A0 =A0 =A0 =A0 to_integer(unsigned(din(31 downto 31))); > > =A0 =A0dout <=3D std_logic_vector(to_unsigned(cnt,6)); > > end one_count_unrolled_arch ; > > And the architecture with a loop is the following: > > architecture one_count_loop_arch of one_count_loop is > > signal =A0cnt : integer range 0 to 32; > > begin > > =A0 process(din) is > =A0 =A0 variable =A0tmp : integer range 0 to 32; > =A0 =A0 begin > =A0 =A0 =A0 tmp :=3D to_integer(unsigned(din(0 downto 0))); > =A0 =A0 =A0 for i in 1 to 31 loop > =A0 =A0 =A0 =A0 =A0 tmp :=3D tmp + to_integer(unsigned(din(i downto i))); > =A0 =A0 =A0 end loop; > =A0 =A0 =A0 cnt <=3D tmp; > =A0 end process; > > =A0 dout <=3D std_logic_vector(to_unsigned(cnt,6)); > > end one_count_loop_arch ; > > I would be really grateful if somebody could point out what I did > wrong with the 2. solution with loop. > It certainly must be my mistake, but I can not find it... > > Additionally, I know that this "brute-force" one counting might not be > the optimal approach, > but this is just my first attempt to get the job done. If somebody has > a better solution, I would > appreciate it if you could share it. > > Regards, > Peter see what you get with this function instead (a function I have used before): function count_ones(slv : std_logic_vector) return natural is varaible n_ones : natural :=3D 0; begin for i in slv'range loop if slv(i) =3D '1' then n_ones :=3D n_ones + 1; end if; end loop; return n_ones; end function count_ones; .... inside architecture, no process needed: dout <=3D std_logic_vector( to_unsigned(count_ones(din), dout'length) ); The beauty with this is the function will work with a std_logic_vector of any length.Article: 151026
On Feb 27, 11:02 am, Patrick <Patr...@hotmail.com> wrote: > Hi again, > > Thanks to all your input, I implemented your suggestions, however the > problem remains the same. The result in simulation works fine, but the > hardware > outputs something different. Just to briefly recap, I have two ctrl > signals that determine the behaviour of the entity: > > GET (ctrl = "00000000") sets register tx to input of op1 > SH1_L (ctrl = "00000001") res := (op1 << 1) | tx; > tx := tx >> 31; > > library ieee; > use ieee.std_logic_1164.all; > > entity test is > port > ( > op1 : in std_logic_vector(31 downto 0); -- Input operand > ctrl : in std_logic_vector(7 downto 0); -- Control signal > clk : in std_logic; -- clock > res : out std_logic_vector(31 downto 0) -- Result > ); > end; > > architecture rtl of test is > > type res_sel_type is (GET, SH1_L); > > constant Z : std_logic_vector(31 downto 0) := (others => '0'); > > signal res_sel : res_sel_type; > signal load : std_logic := '0'; > signal shl : std_logic := '0'; > > signal tx : std_logic_vector(31 downto 0) := (others => '0'); > signal inp1 : std_logic_vector(31 downto 0) := (others => '0'); > > begin > > dec_op: process (ctrl, op1) > begin > > res_sel <= GET; > load <= '0'; > shl <= '0'; > inp1 <= ( others => '0'); > > case ctrl is > > -- store operand > when "00000000" => > inp1 <= op1; > load <= '1'; > res_sel <= GET; > > -- 1-bit left-shift with carry > when "00000001" => > inp1 <= op1; > shl <= '1'; > res_sel <= SH1_L; > > when others => > -- Leave default values > > end case; > > end process; > > sel_out: process (res_sel, inp1, tx) > begin > > case res_sel is > > when SH1_L => > res <= ( inp1(30 downto 0) & '0' ) or tx; > > when others => > res <= (others => '0'); > > end case; > > end process; > > sync: process(clk) > begin > if clk'event and clk = '1' then > if load = '1' then > tx <= op1; > elsif shl = '1' then > tx <= Z(30 downto 0) & op1(31); > end if; > end if; > end process; > > end rtl; > > TESTPROGRAM > > GET 0 (this sets tx <= 0 ) > SH1_L 0xfedcba90 exp. output: 0xfdb97520 act. output = 0xfdb97521 > SH1_L 0x7654321f exp. output: 0xeca8643f act. output = 0xeca8643f > SH1_L 0x71234567 exp. output: 0xe2468ace act. output = 0xe2468ace > > As you can see, the last bit is wrong for the first SH1_L operation. The > first SH1_L operation produces a carry for the NEXT SH1_L operation since > the MSB is set to one of the input, however, it seems that this carry is > already considered in the current SH1_L operation, which is wrong (tx > should be zero). > I checked the synthesis report and there are no latches, so I am a bit > clueless and almost desperate what is going wrong here. I use Xilinx ISE > 12.1 for > synthesis, could there be a problem because I do not have a reset signal > in my architecture, that the wrong kind of latches are instantiated? > > Many thanks for further helpful comments to solve this issue, > Patrick I'm not sure what your results mean really. It depends on when you are looking at the output. If you are applying the inputs to the UUT (unit under test) and toggle the clock, then look at the outputs, then I think this result is correct. As soon as you get a rising edge on the clock, tx will change and your output will change. The tx register is the only register in your design. The rest of the circuit is combinatorial, sot that those signals all change as soon as the inputs change. So what is the timing between your input stimulus, the output observation and the clock? RickArticle: 151027
A good synthesis tool should be able to optimize either version to the same implementation. But there are semantic differences that Xilinx may be getting hung up on. In the unrolled version, you have a long expression, and there is freedom within vhdl to evaluate it in different orders or groups of operations. In the loop version, since you are continually updating tmp, you are describing an explicitly sequential order in which the calculation takes place. Like I said, a good synthesis tool should be able to handle either one equally well, but you get what you pay for in synthesis tools. If you are looking for a general solution to the problem for any size vector, try a recursive function implentation of a binary tree and see what happens. Just for kicks, you might also put the whole calculation in the loop (0 to 31), with temp set to 0 before the loop. Shouldn't make any difference, but then again, we're already seeing differences where there should be none. On the other hand, if what you have works (fits and meets timing), use the most maintainable, understandable version. It will save you time (=money) in the long run. It is often interesting to find out what is "the optimal" way to code something such that it results in the smallest/fastest circuit. But in the big picture, it most often does not matter, especially when you write some cryptic code to squeeze the last pS/LUT out of it, and you had plenty of slack and space to spare anyway. Nevertheless, knowing how to squeeze every pS/LUT comes in handy every once in a while. AndyArticle: 151028
On Feb 28, 7:56=A0am, Serkan <ok...@su.sabanciuniv.edu> wrote: > I need to route a FAST CLK (that is used for deserializing and input > to only one bank) to another bank's IODELAY2 and IOSERDES2 elements. > Is this possible? > > Please remember that I also need to send signals like, > > -serdesstrobe, > -fast ioclk(sampling fast serial data), > -parallel clk(clk whose frequency is the same as parallel > data(deserialized data)) to these elements. > > Serkan This isn't possible. These clocks can only operate within one bank. You may be able to just use a BUFG if the data rate isn't too high. Ed McGettigan -- Xilinx Inc.Article: 151029
On Feb 28, 11:32=A0am, Tim Wescott <t...@seemywebsite.com> wrote: > On 02/28/2011 06:36 AM, Eugen_pcad_ru wrote: > > > > > > > Hello all! > > I need pll which can: > > 1) 40 MHz -> =A0320 MHz (0 deg), > > =A0 =A0 =A0 =A0 =A0 =A0 =A0 320 MHz (15 deg), > > =A0 =A0 =A0 =A0 =A0 =A0 =A0 320 MHz (30 deg), > > =A0 =A0 =A0 =A0 =A0 =A0 =A0 320 MHz (45 deg), > > =A0 =A0 =A0 =A0 =A0 =A0 =A0 320 MHz (60 deg) > > =A0 =A0 =A0 =A0 =A0 =A0 =A0 320 MHz (75 deg), > > =A0 =A0 =A0 =A0 =A0 =A0 =A0 320 MHz (90 deg), > > =A0 =A0 =A0 =A0 =A0 =A0 =A0 320 MHz (105 deg), > > =A0 =A0 =A0 =A0 =A0 =A0 =A0 320 MHz (120 deg), > > =A0 =A0 =A0 =A0 =A0 =A0 =A0 320 MHz (135 deg), > > =A0 =A0 =A0 =A0 =A0 =A0 =A0 320 MHz (150 deg), > > =A0 =A0 =A0 =A0 =A0 =A0 =A0 320 MHz (165 deg), > > =A0 =A0 =A0 =A0 =A0 =A0 =A0 320 MHz (180 deg). > > They can be together or not. > > And I have two fpgas: Cyclone III (Altera), Spartan-3AN. > > What fpga is better for me? Why? Or its no difference? > > I don't know about the Altera part, but Xilinx is too cool to use > phase-locked loops -- they use delay-locked loops instead (see the data > sheet). =A0This means that they can maybe generate the clock you need, bu= t > they'll do it by delaying the 40MHz clock, and they'll demand that your > clock edges have no more than 150ps of jitter. =A0In other words, you nee= d > to feed it a 40MHz clock that jitters no worse than a good 320MHz clock. > > -- > > Tim Wescott > Wescott Design Serviceshttp://www.wescottdesign.com > > Do you need to implement control loops in software? > "Applied Control Theory for Embedded Systems" was written for you. > See details athttp://www.wescottdesign.com/actfes/actfes.html- Hide quote= d text - > > - Show quoted text - In the Spartan-3A family the DCM CLKIN jitter is specified at +/-300pS at 40 MHz. > but Xilinx is too cool to use phase-locked loops While this is true for older families, Virtex-5, Virtex-6, and Spartan-6 all include PLL clocking elements. Ed McGettigan -- Xilinx Inc.Article: 151030
Dear Andy, Tricky, thank you both for your valuable input. Please find my comments below. On Mar 1, 3:49=A0pm, Andy <jonesa...@comcast.net> wrote: > A good synthesis tool should be able to optimize either version to the > same implementation. But there are semantic differences that Xilinx > may be getting hung up on. Aha, that's a good thing, it means that I did not make some obvious mistake. ;-) > If you are looking for a general solution to the problem for any size > vector, try a recursive function implentation of a binary tree and see > what happens. OK, I didn't quite get that but will consider it again. > Just for kicks, you might also put the whole calculation in the loop > (0 to 31), with temp set to 0 before the loop. Shouldn't make any > difference, but then again, we're already seeing differences where > there should be none. Sorry I didn't tell you before. I have already tried that and in this case XST produces the same result. > On the other hand, if what you have works (fits and meets timing), use > the most maintainable, understandable version. It will save you time > (=3Dmoney) in the long run. It is often interesting to find out what is > "the optimal" way to code something such that it results in the > smallest/fastest circuit. But in the big picture, it most often does > not matter, especially when you write some cryptic code to squeeze the > last pS/LUT out of it, and you had plenty of slack and space to spare > anyway. Nevertheless, knowing how to squeeze every pS/LUT comes in > handy every once in a while. Andy, I completely agree with what you have written above. One should strive for maintainable and understandable version. Although, on my particular case, I have to find a good solution in terms of LUT resources, because I need 8 instances of one counters with 64-bit input data. And the device is getting full... Tricky, your approach does indeed look very neat. I like it. Although it is far less efficient than mine. For the same input/output ports, your version with function requires 171 Slices in 313 LUTs. (The minimum that I get with unrolled version is 41 Slices and 73 LUTs).Article: 151031
On Mar 1, 6:26=A0pm, Ed McGettigan <ed.mcgetti...@xilinx.com> wrote: > On Feb 28, 7:56=A0am, Serkan <ok...@su.sabanciuniv.edu> wrote: > > > I need to route a FAST CLK (that is used for deserializing and input > > to only one bank) to another bank's IODELAY2 and IOSERDES2 elements. > > Is this possible? > > > Please remember that I also need to send signals like, > > > -serdesstrobe, > > -fast ioclk(sampling fast serial data), > > -parallel clk(clk whose frequency is the same as parallel > > data(deserialized data)) to these elements. > > > Serkan > > This isn't possible. =A0These clocks can only operate within one bank. > > You may be able to just use a BUFG if the data rate isn't too high. > > Ed McGettigan > -- > Xilinx Inc. Dear Ed, Are you sure about this? Because I could be able to drive two PLLs with BUFIO2s using two ISERDES(DFB) outputs. Maybe I'm missing something but below config placed and routed. same GCLK --> same IBUFGDS-->first ISERDES2(DFB)=3D=3D>first bufio2=3D=3D= =3D> PLL 1 same GCLK --> same IBUFGDS-->2nd ISERDES2(DFB)=3D=3D>2nd bufio2=3D=3D=3D> P= LL 2 I'm trying to do this because I do not want to be limited to 400Mhz of Spartan 6 BUFGs while deserializing.Article: 151032
On 03/01/2011 08:34 AM, Ed McGettigan wrote: > On Feb 28, 11:32 am, Tim Wescott<t...@seemywebsite.com> wrote: >> On 02/28/2011 06:36 AM, Eugen_pcad_ru wrote: >> >> >> >> >> >>> Hello all! >>> I need pll which can: >>> 1) 40 MHz -> 320 MHz (0 deg), >>> 320 MHz (15 deg), >>> 320 MHz (30 deg), >>> 320 MHz (45 deg), >>> 320 MHz (60 deg) >>> 320 MHz (75 deg), >>> 320 MHz (90 deg), >>> 320 MHz (105 deg), >>> 320 MHz (120 deg), >>> 320 MHz (135 deg), >>> 320 MHz (150 deg), >>> 320 MHz (165 deg), >>> 320 MHz (180 deg). >>> They can be together or not. >>> And I have two fpgas: Cyclone III (Altera), Spartan-3AN. >>> What fpga is better for me? Why? Or its no difference? >> >> I don't know about the Altera part, but Xilinx is too cool to use >> phase-locked loops -- they use delay-locked loops instead (see the data >> sheet). This means that they can maybe generate the clock you need, but >> they'll do it by delaying the 40MHz clock, and they'll demand that your >> clock edges have no more than 150ps of jitter. In other words, you need >> to feed it a 40MHz clock that jitters no worse than a good 320MHz clock. > > In the Spartan-3A family the DCM CLKIN jitter is specified at +/-300pS > at 40 MHz. By the data sheet that I looked at, that is true unless you're asking for synthesized frequencies > 150MHz, in which case it needs to be 150ps. >> but Xilinx is too cool to use phase-locked loops > > While this is true for older families, Virtex-5, Virtex-6, and > Spartan-6 all include PLL clocking elements. > That's good to know -- sometimes a real PLL is a good thing, when you have a clock to clean up. -- Tim Wescott Wescott Design Services http://www.wescottdesign.com Do you need to implement control loops in software? "Applied Control Theory for Embedded Systems" was written for you. See details at http://www.wescottdesign.com/actfes/actfes.htmlArticle: 151033
Here's a slightly different approach to your problem.... It tries to take advantage of the fact that the LUTs are pretty good as lookup tables. It's in Verilog, but you should easily be able to convert it to VHDL. `define V1 module tst ( input [31:0] data, output [5:0] cnt ); `ifdef V1 /* Device utilization summary: --------------------------- Selected Device : 3s50pq208-5 Number of Slices: 35 out of 768 4% Number of 4 input LUTs: 62 out of 1536 4% Number of IOs: 38 Number of bonded IOBs: 0 out of 124 0% */ function [1:0] cnt3; input [2:0] d_in; begin case (d_in) 3'h0: cnt3 = 2'h0; 3'h1: cnt3 = 2'h1; 3'h2: cnt3 = 2'h1; 3'h3: cnt3 = 2'h2; 3'h4: cnt3 = 2'h1; 3'h5: cnt3 = 2'h2; 3'h6: cnt3 = 2'h2; 3'h7: cnt3 = 2'h3; endcase end endfunction assign cnt = cnt3(data[2:0]) + cnt3(data[5:3]) + cnt3(data[8:6]) + cnt3(data[11:9]) + cnt3(data[14:12]) + cnt3(data[17:15]) + cnt3(data[20:18]) + cnt3(data[23:21]) + cnt3(data[26:24]) + cnt3(data[29:27]) + cnt3({1'b0, data[31:30]}) ; `endif `ifdef V2 /* Selected Device : 3s50pq208-5 Number of Slices: 44 out of 768 5% Number of 4 input LUTs: 79 out of 1536 5% Number of IOs: 38 Number of bonded IOBs: 0 out of 124 0% */ function [2:0] cnt4; input [3:0] d_in; begin case (d_in) 4'h0: cnt4 = 3'h0; 4'h1: cnt4 = 3'h1; 4'h2: cnt4 = 3'h1; 4'h3: cnt4 = 3'h2; 4'h4: cnt4 = 3'h1; 4'h5: cnt4 = 3'h2; 4'h6: cnt4 = 3'h2; 4'h7: cnt4 = 3'h3; 4'h8: cnt4 = 3'h1; 4'h9: cnt4 = 3'h2; 4'ha: cnt4 = 3'h2; 4'hb: cnt4 = 3'h3; 4'hc: cnt4 = 3'h2; 4'hd: cnt4 = 3'h3; 4'he: cnt4 = 3'h3; 4'hf: cnt4 = 3'h4; endcase end endfunction assign cnt = cnt4(data[3:0]) + cnt4(data[7:4]) + cnt4(data[11:8]) + cnt4(data[15:12]) + cnt4(data[19:16]) + cnt4(data[23:20]) + cnt4(data[27:24]) + cnt4(data[31:28]) ; `endif endmodule John ProvidenzaArticle: 151034
Dear all, I am using XST 12.1 for sythensis and for the first time I had a look at how many slices are occupied by my design. I have to say, I am bit overwhelmed by the results, with Virtex II you essentially got one figure which was the total of the slices. In Virtex 5, XST seems break this down as follows: Design Summary: Slice Logic Utilization: Number of Slice Registers: 3,127 out of 28,800 9% Number used as Flip Flops: 3,127 Number of Slice LUTs: 7,521 out of 28,800 25% Number used as logic: 7,256 out of 28,800 25% Number using O6 output only: 7,001 Number using O5 output only: 68 Number using O5 and O6: 75 Number used as Memory: 37 out of 7,680 1% Number used as Shift Register: 37 Number using O6 output only: 37 Number used as exclusive route-thru: 3 Number of route-thrus: 72 Number using O6 output only: 60 Number using O5 output only: 9 Slice Logic Distribution: Number of occupied Slices: 2,451 out of 7,200 32% Number of LUT Flip Flop pairs used: 8,625 Number with an unused Flip Flop: 5,001 out of 7,707 63% Number with an unused LUT: 507 out of 7,707 5% Number of fully used LUT-FF pairs: 2,532 out of 7,707 31% Number of unique control sets: 233 So what it now the number of total slices occupied on my board? Which ones do I have to add up? Maybe: Number of occupied slices + Number of Slice Registers + Number Used as Logic? thanks, TurjArticle: 151035
Generally a post of this sort begs the question: what are you doing and why= do you think you need 15 degree phase increments of a 320MHz clock? Do you= need them all at once or can you switch between them? Do you need those ph= ases to be exact? Do you need exactly 15 degree increments or just some swe= ep from 0 to 180. I can't off hand recall the features of the DCMs in the Xilinx part, but so= mething to consider is also whether you have the clocking resources to get = what you want. Typically in a single region in the FPGA, you have access to= something like 8-12 regional or global clocks. So if you want all 13 phase= s going to some logic via the clock network - it might be tough or impossib= le to do. Going via general routing resources would create horrible skew an= d it would probably defeat the purpose of having precise phases. I think in general you need to elaborate on what you are trying to do. ChrisArticle: 151036
> > If you are looking for a general solution to the problem for any size > > vector, try a recursive function implentation of a binary tree and see > > what happens. > > OK, I didn't quite get that but will consider it again. > The synthesis tool may not be able to figure out that it need not carry all the bits of the sum through every caculation in the loop. A binary tree implementation can manage the sum width at every stage. For a recursive binary tree implementation, define a function that divides the vector into two ~equal parts (i.e. n/2 and n/2 + n mod 2), calls itself on each one, and returns the sum of the two results. Stop recursion when the size of the incoming vector is 1 (just return 1 if the bit is set, and 0 if not). This is synthesizeable as long as the recursion is statically bound (which it is, by the size of the original vector). It should work out pretty close to what johnp's approach does, except work for any size input vector. AndyArticle: 151037
"Number of occupied slices" is the total number of slices used. "Number of = slice registers" is the number of FFs used in the whole FPGA, there are 4 F= Fs and 4 LUTs to a slice in a V5 (7200*4 =3D 28800). "Number used as logic" is the number of LUTs used as a true LUT. A LUT can = also be used as distributed RAM (Number used as Memory) (see Xilinx docs) o= r as a shift register (See SRL in the docs). A slice can have anywhere from 0 to all 4 of it's FFs used, same goes for t= he LUTs. Sometimes a slice is partially empty because the placer chooses to= group the LUTs/FFs that way for no good reason, sometimes there are legiti= mate reasons why certain groups of FFs can't be put in the same slice as so= me others (look for "control sets" in the Xilinx documentation). If you are trying to get an idea of how big your design is, the number of L= UTs (Number of Slice LUTs) or the number of FFs (Number of Slice Registers)= is usually a good start - In my experience, MAP/PAR are able to fill up to= about 75-80% of the FFs or LUTs in a chip before it becomes difficult to f= it everything. ChrisArticle: 151038
In article <ikjqfr$uah$1@speranza.aioe.org>, Turj <Turj@hotmail.com> wrote: >Dear all, > >I am using XST 12.1 for sythensis and for the first time I had a look at >how many slices are occupied by my design. I have to say, I am bit >overwhelmed by the results, with Virtex II you essentially got one >figure which was the total of the slices. >In Virtex 5, XST seems break this down as follows: >Design Summary: >Number of Slice Registers: 3,127 out of 28,800 9% >Number of Slice LUTs: 7,521 out of 28,800 25% > >Slice Logic Distribution: >Number of occupied Slices: 2,451 out of 7,200 32% > >So what it now the number of total slices occupied on my board? Which >ones do I have to add up? 'Number of occupied slices' is the number you want; the slices are quite big (each contains four registers and four LUTs, with not all of them necessarily being used). At the moment, you could double the size of your design with no trouble; you might be able to triple it assuming things pack nicely (you'd be using 75% of the LUTs and maybe the software would be able to pack them into less than the 96% of the slices which simply writing the current design down three times would occupy); it would be unwise to attempt to quadruple it. TomArticle: 151039
In comp.arch.fpga Andy <jonesandy@comcast.net> wrote: (snip) > The synthesis tool may not be able to figure out that it need not > carry all the bits of the sum through every caculation in the loop. A > binary tree implementation can manage the sum width at every stage. Is that the same as a Carry Save Adder tree? Maybe not. The CSA has three inputs and two outputs, so it isn't exactly binary. -- glenArticle: 151040
Hello Today I finally got my Spartan 3E 1600 eval board from Digilent. The problem is that that I can't make it work with the USB cable programmer. I think it's something wrong with the drivers. When connecting it to Windows XP SP3 it reports the VID/PID pair as 03FD / 000D. On the other hand, in the dmodusb.inf from the Adept drivers I found 4 pairs: VID_1443 PID_0007 VID_1443 PID_0005 VID_1443 PID_0003 VID_1443 PID_0001 This means that it can't recognize my board and fails to install any driver. Replacing any pair with 03FD/000D will allow the driver to be installed but Adept still can't find my board and the green LED next to the USB connector is not lit. I guess I should replace all pairs with something meaningful, but I don't know exactly the values. On the other hand, Windows 7 recognizes the board without any driver (as a Xilinx USB cable), the green LED is lit but Adept can't see it ("No devices connected"). The funny thing is that in Win7 the VID/PID pair is 03FD/0008. Rather strange I might say... The board seems to be functional because when in normal operating mode it writes text on the LCD and blinks LED's. Does anybody have any advice? (searching for the Ubuntu install CD..., let's see what VID/PID I get there :P ) Thanks, Fili --------------------------------------- Posted through http://www.FPGARelated.comArticle: 151041
Peter wrote: > > If somebody has a better solution, I would appreciate it if you could share it. > When I looked at this some years back, XST worked well enough at creating an adder cascade using the old "mask and add" software trick that I never bothered writing something more optimal: gen_bcnt32: if ( ALU_WIDTH = 32 ) generate begin process(din) -- multiple variables not needed, but make intermediate steps visible in simulation variable temp : unsigned (ALU_MSB downto 0); variable temp1 : unsigned (ALU_MSB downto 0); variable temp2 : unsigned (ALU_MSB downto 0); variable temp3 : unsigned (ALU_MSB downto 0); variable temp4 : unsigned (ALU_MSB downto 0); variable temp5 : unsigned (ALU_MSB downto 0); begin temp := unsigned(din); temp1 := (temp AND X"5555_5555") + ( ( temp srl 1) AND X"5555_5555"); -- 0..2 out x16 temp2 := (temp1 AND X"3333_3333") + ( ( temp1 srl 2) AND X"3333_3333"); -- 0..4 out x8 temp3 := (temp2 AND X"0707_0707") + ( ( temp2 srl 4) AND X"0707_0707"); -- 0..8 out x4 temp4 := (temp3 AND X"001f_001f") + ( ( temp3 srl 8) AND X"001f_001f"); -- 0..16 out x2 temp5 := (temp4 AND X"0000_003f") + ( ( temp4 srl 16) AND X"0000_003f"); -- 0..32 out cnt <= std_logic_vector(temp5(5 downto 0)); end process; end generate gen_bcnt32; BrianArticle: 151042
On Mar 1, 6:37=A0pm, "RCIngham" <robert.ingham@n_o_s_p_a_m.n_o_s_p_a_m.gmail.com> wrote: > >my project is implementing of blowfish algorithm in FPGA and sending > >the data from PC through FPGA and encrypt the data.for this which > >protocols i can use.please tell me some links related to it. > > What is the data rate you need to send from PC to FPGA? > Is it streaming? > Are you sending any data back from FPGA to PC? > > --------------------------------------- =A0 =A0 =A0 =A0 > Posted throughhttp://www.FPGARelated.com I m using Virtex II pro board. so i can use one side high speed data transfer and one side serial data transfer. and i need to send a file .but i dnt knw wht you are asking streaming...and yes i want to send it to file back to PC.Article: 151043
In comp.arch.fpga Brian Davis <brimdavis@aol.com> wrote: (snip) > When I looked at this some years back, XST worked well enough > at creating an adder cascade using the old "mask and add" software > trick that I never bothered writing something more optimal: (snip) > temp1 := (temp AND X"5555_5555") + ( ( temp srl 1) AND > X"5555_5555"); -- 0..2 out x16 > > temp2 := (temp1 AND X"3333_3333") + ( ( temp1 srl 2) AND > X"3333_3333"); -- 0..4 out x8 OK, that would be a binary tree. I believe the CSA adder tree is slightly more efficient, though it might depend on the number of inputs. The binary tree cascade works especially well on word oriented machines, and can be easily written in many high-level languages (with the assumption of knowing the word size). The first stage of a CSA tree starts with N inputs, and generates N/3 two bit outputs that are the sums and carries from N/3 full adders. (If N isn't a multiple of three, then one bit may bypass the stage, and two bits go into a half adder.) The next stage takes the N/3 ones and N/3 twos, and generates N/9 ones, 2N/9 twos, and N/9 fours. You can continue until there is only one bit of each, or sometimes there are other optimizations near the end. Last time I did one, I only needed to know zero, one, two, three, or more than three, which simplifies it slightly. It also pipelines well, but then so does the binary tree. -- glenArticle: 151044
Andy nailed it again when he said you get what you pay for regarding synthesis tool. Initially I was synthesising with ISE12.4 and the results were, hm, inconsistent. After switching the synthesis tool to SynplifyPro v2010.03 the results were as expected and, of course, even better than that. Please see the table below. Tricky's version is denoted "funct", where there are major differences between the 2 tools: ---------- 32-bit input data -------- unrolled: XST 74 LUTs, 41 slices unrolled: SynplifyPro 57 LUTs, 34 slices loop: XST 100 LUTs, 54 slices loop: SynplifyPro 57 LUTs, 34 slices funct: XST 317 LUTs, 161 slices funct: SynplifyPro 58 LUTs, 34 slices ---------- 64-bit input data -------- unrolled: XST 174 LUTs, 96 slices unrolled: SynplifyPro 129 LUTs, 80 slices loop: XST 227 LUTs, 121 slices loop: SynplifyPro 129 LUTs, 80 slices funct: XST 813 LUTs, 436 slices funct: SynplifyPro 130 LUTs, 82 slices PeterArticle: 151045
It's me again :) I finally got it working, but not the way I expected. I've uninstalled all drivers and installed Xilinx ISE Webpack. This one came with drivers that have the correct VID/PID pairs and now I'm able to program the board via Impact. Digilent's Adept still can't see the board. I'm still interested if there's a way to program the board via Adept (I don't want to install webpack on all computers) but at least I'm not desperate anymore, as I have a way to do it. Thanks... err.. for listening my problems :) Fili --------------------------------------- Posted through http://www.FPGARelated.comArticle: 151046
Johnp, Brian, thank you too for your input! Much appreciated. I have ran your code through 2 synthesisers and have updated the table of required resources. -------------- 32-bit input data -------------- unrolled: XST 74 LUTs, 41 slices unrolled: SynplifyPro 57 LUTs, 34 slices loop: XST 100 LUTs, 54 slices loop: SynplifyPro 57 LUTs, 34 slices funct: XST 317 LUTs, 161 slices funct: SynplifyPro 58 LUTs, 34 slices JohnpV1: XST 62 LUTs, 35 slices JohnpV1: SynplifyPro 57 LUTs, 33 slices JohnpV2: XST 78 LUTs, 43 slices JohnpV2: SynplifyPro 54 LUTs, 32 slices Brian: XST 57 LUTs, 39 slices Brian: SynplifyPro 57 LUTs, 41 slices The latest 3 pairs of results are interesting because even XST produces good results, especially in Brian's version where XST is surprisingly even slightly better. But anyway, it's not that XST is so clever, it is a clever coding of the design. Regards, PeterArticle: 151047
>On Mar 1, 6:37=A0pm, "RCIngham" ><robert.ingham@n_o_s_p_a_m.n_o_s_p_a_m.gmail.com> wrote: >> >my project is implementing of blowfish algorithm in FPGA and sending >> >the data from PC through FPGA and encrypt the data.for this which >> >protocols i can use.please tell me some links related to it. >> >> What is the data rate you need to send from PC to FPGA? >> Is it streaming? >> Are you sending any data back from FPGA to PC? >> >> --------------------------------------- =A0 =A0 =A0 =A0 >> Posted throughhttp://www.FPGARelated.com > >I m using Virtex II pro board. so i can use one side high speed data >transfer and one side serial data transfer. and i need to send a >file .but i dnt knw wht you are asking streaming...and yes i want to >send it to file back to PC. > So, not streaming, then. How big a file? How long can you wait for it to be sent and returned? --------------------------------------- Posted through http://www.FPGARelated.comArticle: 151048
Hello! I'm new in the forum and just done an FPGA university course, very very small...we have only turned on/off led with finite state machine and so on...now i'm tryng to develope an IIR filter with XSA50 board form Xess with spartanIIe50 fpga. FIlter kernel is just a 2 pole system with a zero in 0, i would do a bandpass with changable passaband with pushbuttons; i 've idealized that main structure of the program would be a module with a counter for generating clock for the ADC/DAC, a module that pass this samples in the filtern kernel, the filter kernel iir itself and a module that passes filtererd samples to DAC; mainly i have 2 problems: 1) i can do only operation with radix-2 coefficient, so i can use only 1/2, 1/4 an so on. i don't understand how to pass a float value and multiply it 2)do i need a ram to store at least y[n-2] sample? I know myquestions sounds stupid, maybe i have not a good idea of what i have to do, if you could illuminate me on this... best regards francesco --------------------------------------- Posted through http://www.FPGARelated.comArticle: 151049
> >1) i can do only operation with radix-2 coefficient, so i can use only 1/2, >1/4 an so on. i don't understand how to pass a float value and multiply it You probably don't need floating point. Floating-point is difficult on FPGAs. Imagine your ADC data as between -1 and +1. Then, when you multiply, the result is still between -1 and +1. After additions, scale your data so it is again between -1 and +1. There is a difficulty with IIR filters in digital hardware, which is arithmetic saturation. You will need to do simulations to prove that either it doesn't happen or that you detect and mitigate it. > >2)do i need a ram to store at least y[n-2] sample? Unless the number and size of samples is very large, it could be done with registers (flip/flops). You need to sketch out how the data flows through the arithmetic elements. --------------------------------------- Posted through http://www.FPGARelated.com
Site Home Archive Home FAQ Home How to search the Archive How to Navigate the Archive
Compare FPGA features and resources
Threads starting:
Authors:A B C D E F G H I J K L M N O P Q R S T U V W X Y Z