URL
https://opencores.org/ocsvn/ft816float/ft816float/trunk
Subversion Repositories ft816float
[/] [ft816float/] [trunk/] [rtl/] [verilog/] [FT816Float.v] - Rev 77
Go to most recent revision | Compare with Previous | Blame | View Log
`timescale 1ns / 1ps // ============================================================================ // __ // \\__/ o\ (C) 2014 Robert Finch, Stratford // \ __ / All rights reserved. // \/_// robfinch<remove>@finitron.ca // || // // FT816Float.v // - Triple precision floating point accelerator // // This source file is free software: you can redistribute it and/or modify // it under the terms of the GNU Lesser General Public License as published // by the Free Software Foundation, either version 3 of the License, or // (at your option) any later version. // // This source file is distributed in the hope that it will be useful, // but WITHOUT ANY WARRANTY; without even the implied warranty of // MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the // GNU General Public License for more details. // // You should have received a copy of the GNU General Public License // along with this program. If not, see <http://www.gnu.org/licenses/>. // // 1600 LUTs 350 FF's // 140 MHz // ============================================================================ // `define SIMULATION 1'b1 module FT816Float(rst, clk, vda, rw, ad, db, rdy); parameter pIOAddress = 24'hFEA200; parameter pRdyStyle = 1'b1; parameter EMSB = 15; parameter FMSB = 79; parameter TRUE = 1'b1; parameter FALSE = 1'b0; parameter FADD = 8'd1; parameter FSUB = 8'd2; parameter FMUL = 8'd3; parameter FDIV = 8'd4; parameter FIX2FLT = 8'd5; parameter FLT2FIX = 8'd6; parameter FABS = 8'd7; parameter ABS = 8'd7; parameter NABS = 8'd8; parameter FNABS = 8'd8; parameter MD1 = 8'd10; parameter ABSSWP = 8'd11; parameter ABSSWP1 = 8'd12; parameter NORM1 = 8'd13; parameter NORM = 8'd14; parameter ADD = 8'd15; parameter FCOMPL = 8'd16; parameter FNEG = 8'd16; parameter SWAP = 8'd17; parameter FIXED_ADD = 8'h81; parameter FIXED_SUB = 8'h82; parameter FIXED_MUL = 8'h83; parameter FIXED_DIV = 8'h84; parameter FIXED_ABS = 8'h87; parameter FIXED_NEG = 8'h90; parameter SWPALG = 8'd18; parameter ADDEND = 8'd19; parameter ALGNSW = 8'd20; parameter RTLOG = 8'd22; parameter FMUL1 = 8'd24; parameter FMUL2 = 8'd25; parameter MUL1 = 8'd26; parameter FMUL3 = 8'd27; parameter MUL2 = 8'd28; parameter MDEND = 8'd29; parameter FDIV1 = 8'd30; parameter MD2 = 8'd31; parameter MD3 = 8'd32; parameter OVCHK = 8'd34; parameter OVFL = 8'd35; parameter DIV1 = 8'd36; parameter IDLE = 8'd62; parameter RESET = 8'd63; input rst; input clk; input vda; input rw; input [23:0] ad; inout tri [7:0] db; output rdy; reg [7:0] cmd; reg [7:0] state; reg [5:0] state_stk [3:0]; //reg [3:0] sp; reg [1:0] sign; reg [EMSB:0] acc; reg [7:0] y; reg [EMSB+FMSB+1:0] FAC1; reg [EMSB+FMSB+1:0] FAC2; reg [FMSB:0] E; wire [EMSB:0] FAC1_exp = FAC1[EMSB+FMSB+1:FMSB+1]; wire [FMSB:0] FAC1_man = FAC1[FMSB:0]; wire [EMSB:0] FAC2_exp = FAC2[EMSB+FMSB+1:FMSB+1]; wire [FMSB:0] FAC2_man = FAC2[FMSB:0]; reg addOrSub; wire [FMSB+1:0] sum = addOrSub ? FAC2_man - FAC1_man : FAC2_man + FAC1_man; wire [FMSB+1:0] dif = FAC2_man - E; wire [FMSB+1:0] neg = {FMSB+1{1'b0}} - FAC1_man; wire [EMSB+1:0] expdif = FAC2_exp - FAC1_exp; // Note the carry flag must be extended manually! reg cf,vf,nf; wire [EMSB+1:0] exp_sum = acc + FAC1_exp + {15'd0,cf}; // FMUL wire [EMSB+1:0] exp_dif = acc - FAC1_exp - {15'd0,~cf}; // FDIV reg [FMSB:0] rem; reg isRTAR; reg busy; reg shiftBy16; reg isFixedPoint; reg [7:0] dbo; wire eq = FAC1==FAC2; wire gt = (FAC1[FMSB]^FAC2[FMSB]) ? FAC2[FMSB] : // If the signs are different, whichever one is positive FAC1_exp==FAC2_exp ? (FAC1_man > FAC2_man) : // if exponents are equal check mantissa FAC1_exp > FAC2_exp; // else compare exponents wire lt = !(gt|eq); wire zf = ~|FAC1; wire cs = vda && (ad[23:8]==pIOAddress[23:8]); reg rdy1,rdy2; always @(posedge clk) if (rst) begin rdy1 <= 1'b1; rdy2 <= 1'b1; end else begin rdy1 <= cs & ~rdy1; rdy2 <= cs & rdy1; end assign rdy = cs ? (rw ? rdy2 : 1'b1) : pRdyStyle; assign db = cs & rw ? dbo : {8{1'bz}}; // This is a clock cycle counter used in simulation to determine the number of // cycles a given operation takes to complete. reg [11:0] cyccnt; always @(posedge clk) if (rst) begin next_state(RESET); `ifdef SIMULATION FAC1 <= 96'd0; // for simulation FAC2 <= 96'd0; `endif end else begin `ifdef SIMULATION cyccnt <= cyccnt + 1; `endif cmd <= 8'h00; if (cs & ~rw) case(ad[7:0]) 8'h00: FAC1[7:0] <= db; 8'h01: FAC1[15:8] <= db; 8'h02: FAC1[23:16] <= db; 8'h03: FAC1[31:24] <= db; 8'h04: FAC1[39:32] <= db; 8'h05: FAC1[47:40] <= db; 8'h06: FAC1[55:48] <= db; 8'h07: FAC1[63:56] <= db; 8'h08: FAC1[71:64] <= db; 8'h09: FAC1[79:72] <= db; 8'h0A: FAC1[87:80] <= db; 8'h0B: FAC1[95:88] <= db; 8'h0E: cmd <= db; 8'h10: FAC2[7:0] <= db; 8'h11: FAC2[15:8] <= db; 8'h12: FAC2[23:16] <= db; 8'h13: FAC2[31:24] <= db; 8'h14: FAC2[39:32] <= db; 8'h15: FAC2[47:40] <= db; 8'h16: FAC2[55:48] <= db; 8'h17: FAC2[63:56] <= db; 8'h18: FAC2[71:64] <= db; 8'h19: FAC2[79:72] <= db; 8'h1A: FAC2[87:80] <= db; 8'h1B: FAC2[95:88] <= db; endcase case(ad[7:0]) 8'h00: dbo <= FAC1[7:0]; 8'h01: dbo <= FAC1[15:8]; 8'h02: dbo <= FAC1[23:16]; 8'h03: dbo <= FAC1[31:24]; 8'h04: dbo <= FAC1[39:32]; 8'h05: dbo <= FAC1[47:40]; 8'h06: dbo <= FAC1[55:48]; 8'h07: dbo <= FAC1[63:56]; 8'h08: dbo <= FAC1[71:64]; 8'h09: dbo <= FAC1[79:72]; 8'h0A: dbo <= FAC1[87:80]; 8'h0B: dbo <= FAC1[95:88]; 8'h0E: dbo <= {busy,2'b00,lt,eq,gt,zf,vf}; 8'h10: dbo <= FAC2[7:0]; 8'h11: dbo <= FAC2[15:8]; 8'h12: dbo <= FAC2[23:16]; 8'h13: dbo <= FAC2[31:24]; 8'h14: dbo <= FAC2[39:32]; 8'h15: dbo <= FAC2[47:40]; 8'h16: dbo <= FAC2[55:48]; 8'h17: dbo <= FAC2[63:56]; 8'h18: dbo <= FAC2[71:64]; 8'h19: dbo <= FAC2[79:72]; 8'h1A: dbo <= FAC2[87:80]; 8'h1B: dbo <= FAC2[95:88]; endcase case(state) RESET: begin // sp <= 4'h0; next_state(IDLE); end //----------------------------------------------------------------------------- //----------------------------------------------------------------------------- IDLE: begin `ifdef SIMULATION if (cyccnt > 0) $display("Cycle Count=%d", cyccnt); cyccnt <= 12'h0; `endif busy <= 1'b0; // sp <= 4'h0; isFixedPoint <= FALSE; case(cmd) FADD: begin push_state(IDLE); next_state(FADD); busy <= 1'b1; addOrSub <= 1'b0; end FSUB: begin push_state(IDLE); next_state(FSUB); busy <= 1'b1; addOrSub <= 1'b0; end FMUL: begin push_state(IDLE); next_state(FMUL); busy <= 1'b1; addOrSub <= 1'b0; end FDIV: begin push_state(IDLE); next_state(FDIV); busy <= 1'b1; end FIX2FLT: begin push_state(IDLE); next_state(FIX2FLT); busy <= 1'b1; end FLT2FIX: begin push_state(IDLE); next_state(FLT2FIX); busy <= 1'b1; end FNEG: begin push_state(IDLE); next_state(FCOMPL); busy <= 1'b1; end FABS: begin push_state(IDLE); next_state(ABS); busy <= 1'b1; end FNABS: begin push_state(IDLE); next_state(NABS); busy <= 1'b1; end SWAP: begin push_state(IDLE); next_state(SWAP); busy <= 1'b1; end // Fixed point operations FIXED_ADD: begin push_state(IDLE); next_state(FADD); busy <= 1'b1; isFixedPoint <= TRUE; addOrSub <= 1'b0; end FIXED_SUB: begin push_state(IDLE); next_state(FSUB); busy <= 1'b1; isFixedPoint <= TRUE; addOrSub <= 1'b0; end FIXED_MUL: begin push_state(IDLE); next_state(FMUL); busy <= 1'b1; isFixedPoint <= TRUE; addOrSub <= 1'b0; end FIXED_DIV: begin push_state(IDLE); next_state(FDIV); busy <= 1'b1; isFixedPoint <= TRUE; end FIXED_NEG: begin push_state(IDLE); next_state(FCOMPL); busy <= 1'b1; isFixedPoint <= TRUE; end FIXED_ABS: begin push_state(IDLE); next_state(ABS); busy <= 1'b1; isFixedPoint <= TRUE; end endcase end //----------------------------------------------------------------------------- //----------------------------------------------------------------------------- MD1: begin $display("MD1"); sign <= {sign[1:0],1'b0}; next_state(ABSSWP); push_state(ABSSWP); end ABSSWP: begin if (~FAC1_man[FMSB]) begin next_state(ABSSWP1); end else begin push_state(ABSSWP1); sign <= sign + 2'd1; next_state(FCOMPL); end end ABSSWP1: begin cf <= 1'b1; next_state(SWAP); end //----------------------------------------------------------------------------- // Take the absolute value of FAC1 //----------------------------------------------------------------------------- ABS: begin if (FAC1_man[FMSB]) next_state(FCOMPL); else pop_state(); end //----------------------------------------------------------------------------- // Take the negative absolute value of FAC1 //----------------------------------------------------------------------------- NABS: begin if (~FAC1_man[FMSB]) next_state(FCOMPL); else pop_state(); end //----------------------------------------------------------------------------- // Normalize // - Decrement exponent and shift left // - Normalization is normally the last step of an operation. // - If possible the FAC is shifted by 16 bits at a time. This helps with // the many small constants that are usually present. //----------------------------------------------------------------------------- NORM: begin if (isFixedPoint) // nothing to do for fixed point pop_state(); else begin $display("Normalize FAC1H %h", FAC1[FMSB:FMSB-15]); if (FAC1[FMSB]!=FAC1[FMSB-1] || ~|FAC1_exp) begin $display("Normal: %h",FAC1); pop_state(); end // If the mantissa is zero, set the the exponent to zero. Otherwise // normalization could spin for thousands of clock cycles decrementing // the exponent to zero. else if (~|FAC1_man) begin FAC1[EMSB+FMSB+1:FMSB+1] <= 16'h0; pop_state(); end else if (FAC1[FMSB:FMSB-15]=={16{FAC1[FMSB]}}) begin $display("shift by 16"); FAC1[EMSB+FMSB+1:FMSB+1] <= FAC1[EMSB+FMSB+1:FMSB+1] - 16'd16; FAC1[FMSB:0] <= {FAC1[FMSB-16:0],16'h0}; end else begin FAC1[EMSB+FMSB+1:FMSB+1] <= FAC1[EMSB+FMSB+1:FMSB+1] - 16'd1; FAC1[FMSB:0] <= {FAC1[FMSB-1:0],1'b0}; end end end //----------------------------------------------------------------------------- // Add mantissa's and compute carry and overflow. // This is used by both ADD and MUL. //----------------------------------------------------------------------------- ADD: begin FAC1[FMSB:0] <= sum[FMSB:0]; cf <= sum[FMSB+1]; vf <= (sum[FMSB] ^ FAC2[FMSB]) & (1'b1 ^ FAC1[FMSB] ^ FAC2[FMSB]); pop_state(); end //----------------------------------------------------------------------------- // Negate //----------------------------------------------------------------------------- // Complement FAC1 FCOMPL: begin $display("FCOMPL"); FAC1[FMSB:0] <= neg[FMSB:0]; cf <= ~neg[FMSB+1]; vf <= neg[FMSB]==FAC1[FMSB]; if (isFixedPoint) pop_state(); else next_state(ADDEND); end //----------------------------------------------------------------------------- // Swap FAC1 and FAC2 //----------------------------------------------------------------------------- SWAP: begin $display("Swapping FAC1 and FAC2"); FAC1 <= FAC2; FAC2 <= FAC1; E <= FAC2[FMSB:0]; acc <= FAC1_exp; pop_state(); end //----------------------------------------------------------------------------- // Subtract // - subtract first complements the FAC then performs an ADD operation. //----------------------------------------------------------------------------- FSUB: begin // if (isFixedPoint) // push_state(FADD); // else // push_state(SWPALG); push_state(FADD); next_state(FCOMPL); end SWPALG: begin push_state(FADD); next_state(ALGNSW); end //----------------------------------------------------------------------------- // Addition //----------------------------------------------------------------------------- FADD: begin cf <= ~expdif[EMSB+1]; // Must set carry flag from compare // If the exponents are too different then one of the values will // become zero, so the result is just the larger value. This check // is to prevent shifting thousands of times. if (expdif[15] ? expdif < 16'hFFB0 : expdif[15:0] > 16'h0050) begin FAC1 <= expdif[15] ? FAC2 : FAC1; pop_state(); end else if (|expdif[15:0] & !isFixedPoint) next_state(SWPALG); else begin if (!isFixedPoint) push_state(ADDEND); next_state(ADD); end end ADDEND: begin if (!vf) next_state(NORM); else begin isRTAR <= FALSE; next_state(RTLOG); end end ALGNSW: begin if (!cf) next_state(SWAP); else begin isRTAR <= TRUE; next_state(RTLOG); end end //----------------------------------------------------------------------------- // Right shift, logical or arithmetic. //----------------------------------------------------------------------------- RTLOG: begin FAC1[EMSB+FMSB+1:FMSB+1] <= FAC1[EMSB+FMSB+1:FMSB+1] + 16'd1; if (FAC1[EMSB+FMSB+1:FMSB+1]==16'hFFFF) next_state(OVFL); else begin FAC1[FMSB:0] <= {isRTAR ? FAC1_man[FMSB] : cf,FAC1[FMSB:1]}; E[FMSB:0] <= {FAC1[0],E[FMSB-1:1]}; cf <= E[0]; pop_state(); end end //----------------------------------------------------------------------------- // Mulyiply //----------------------------------------------------------------------------- FMUL: begin next_state(MD1); push_state(FMUL1); end FMUL1: begin acc <= exp_sum[EMSB:0]; cf <= exp_sum[EMSB+1]; push_state(MUL1); next_state(MD2); end MUL1: begin // inline RTLOG1 code FAC1[FMSB:0] <= {1'b0,FAC1[FMSB:1]}; E[FMSB:0] <= {FAC1[0],E[FMSB-1:1]}; cf <= E[0]; next_state(FMUL3); //push_state(FMUL3); //next_state(RTLOG1); end FMUL3: begin if (cf) begin FAC1[FMSB:0] <= sum[FMSB:0]; cf <= sum[FMSB+1]; vf <= (sum[FMSB] ^ FAC2[FMSB]) & (1'b1 ^ FAC1[FMSB] ^ FAC2[FMSB]); end y <= y - 8'd1; if (y==8'd0) next_state(MDEND); else next_state(MUL1); end MDEND: begin sign <= {1'b0,sign[1]}; if (~sign[0]) next_state(NORM); else next_state(FCOMPL); end //----------------------------------------------------------------------------- // Divide //----------------------------------------------------------------------------- FDIV: begin push_state(FDIV1); next_state(MD1); end FDIV1: begin acc <= exp_dif[EMSB:0]; cf <= ~exp_dif[EMSB+1]; $display("acc=%h %h %h", exp_dif, acc, FAC1_exp); push_state(DIV1); next_state(MD2); end DIV1: begin $display("FAC1=%h, FAC2=%h, E=%h", FAC1, FAC2, E); y <= y - 8'd1; FAC1[FMSB:0] <= {FAC1[FMSB:0],~dif[FMSB+1]}; if (dif[FMSB+1]) begin FAC2[FMSB:0] <= {FAC2[FMSB-1:0],1'b0}; if (FAC2[FMSB]) begin next_state(OVFL); end else if (y!=8'd1) next_state(DIV1); else begin rem <= dif; next_state(MDEND); end end else begin FAC2[FMSB:0] <= {dif[FMSB-1:0],1'b0}; if (dif[FMSB]) begin next_state(OVFL); end else if (y!=8'd1) next_state(DIV1); else begin rem <= dif; next_state(MDEND); end end end //----------------------------------------------------------------------------- //----------------------------------------------------------------------------- MD2: begin FAC1[FMSB:0] <= 80'h0; if (isFixedPoint) begin y <= 8'h4F; pop_state(); end else if (cf) next_state(OVCHK); else if (acc[EMSB]) next_state(MD3); else begin pop_state(); next_state(NORM); end end MD3: begin acc[EMSB] <= ~acc[EMSB]; FAC1[EMSB+FMSB+1:FMSB+1] <= {~acc[EMSB],acc[EMSB-1:0]}; y <= 8'h4F; pop_state(); end OVCHK: begin if (~acc[EMSB]) next_state(MD3); else next_state(OVFL); end OVFL: begin vf <= 1'b1; next_state(IDLE); end //----------------------------------------------------------------------------- // FIX2FLT // - convert 64 bit fixed point number to floating point //----------------------------------------------------------------------------- FIX2FLT: begin FAC1[EMSB+FMSB+1:FMSB+1] <= 16'h804E; // exponent = 78 next_state(NORM); end //----------------------------------------------------------------------------- // FLT2FIX // - convert floating point number to fixed point. //----------------------------------------------------------------------------- FLT2FIX: begin // If the exponent is too small then no amount of shifting will // result in a non-zero number. In this case we just set the // FAC to zero. Otherwise FLT2FIX would spin for thousands of cycles // until the exponent incremented finally to 803Eh. if (FAC1_exp < 16'h7FB0) begin FAC1[79:0] <= 80'd0; FAC1[95:80] <= 16'h804E; pop_state(); end // If the exponent is too large, we can't right shift and the value // would overflow a 64-bit integer, so we just set it to the max. else if (FAC1_exp > 16'h804E) begin vf <= 1'b1; FAC1[95:80] <= 16'h804E; FAC1[79:0] <= FAC1[79] ? 80'h80000000000000000000 : 80'h7FFFFFFFFFFFFFFFFFFF; pop_state(); end else if (FAC1_exp==16'h804E) pop_state(); else begin push_state(FLT2FIX); isRTAR <= TRUE; next_state(RTLOG); end end endcase end /* DIVBY10: begin FAC2[EMSB+FMSB+1:FMSB+1] <= 16'h8003; FAC2[FMSB] <= 1'b0; // +ve FAC2[FMSB-1:75] <= 4'hA; // 10 FAC2[74:0] <= 75'd0; next_state(FDIV); end */ //----------------------------------------------------------------------------- //----------------------------------------------------------------------------- task push_state; input [5:0] st; begin state_stk[0] <= st; state_stk[1] <= state_stk[0]; state_stk[2] <= state_stk[1]; state_stk[3] <= state_stk[2]; // state_stk[sp-4'd1] <= st; // sp <= sp - 4'd1; end endtask task pop_state; begin state <= state_stk[0]; state_stk[0] <= state_stk[1]; state_stk[1] <= state_stk[2]; state_stk[2] <= state_stk[3]; state_stk[3] <= IDLE; // next_state(state_stk[sp]); // sp <= sp + 4'd1; end endtask task next_state; input [7:0] st; begin state <= st; end endtask endmodule
Go to most recent revision | Compare with Previous | Blame | View Log