Moved muldiv result selection to M stage for performance

This commit is contained in:
David Harris 2021-10-02 09:38:02 -04:00
parent fbe6e41169
commit 9d63aa683f
2 changed files with 26 additions and 30 deletions

View File

@ -59,6 +59,7 @@ module intdivrestoring (
assign DAbsB = ~Din;
// *** parameterize steps per cycle
intdivrestoringstep step1(Win, XQin, DAbsB, W1, XQ1);
intdivrestoringstep step2(W1, XQ1, DAbsB, W2, XQshift);

View File

@ -31,7 +31,7 @@ module muldiv (
input logic [31:0] InstrD,
// Execute Stage interface
input logic [`XLEN-1:0] SrcAE, SrcBE,
input logic [2:0] Funct3E,
input logic [2:0] Funct3E, Funct3M,
input logic MulDivE, W64E,
// Writeback stage
output logic [`XLEN-1:0] MulDivResultW,
@ -45,8 +45,8 @@ module muldiv (
generate
if (`M_SUPPORTED) begin
logic [`XLEN-1:0] MulDivResultE, MulDivResultM;
logic [`XLEN-1:0] PrelimResultE;
logic [`XLEN-1:0] QuotE, RemE;
logic [`XLEN-1:0] PrelimResultM;
logic [`XLEN-1:0] QuotM, RemM;
logic [`XLEN*2-1:0] ProdE, ProdM;
logic enable_q;
@ -57,7 +57,9 @@ module muldiv (
// logic gclk;
logic startDivideE, busy;
logic signedDivide;
logic SignedDivideE;
logic W64M;
// Multiplier
mul mul(.*);
@ -65,51 +67,44 @@ module muldiv (
// Divide
/*// *** replace this clock gater
always @(negedge clk) begin
enable_q <= ~StallM;
end
assign gclk = enable_q & clk; */
// Handle sign extension for W-type instructions
if (`XLEN == 64) begin // RV64 has W-type instructions
assign X = W64E ? {{32{SrcAE[31]&signedDivide}}, SrcAE[31:0]} : SrcAE;
assign D = W64E ? {{32{SrcBE[31]&signedDivide}}, SrcBE[31:0]} : SrcBE;
assign X = W64E ? {{32{SrcAE[31]&SignedDivideE}}, SrcAE[31:0]} : SrcAE;
assign D = W64E ? {{32{SrcBE[31]&SignedDivideE}}, SrcBE[31:0]} : SrcBE;
end else begin // RV32 has no W-type instructions
assign X = SrcAE;
assign D = SrcBE;
end
assign signedDivide = ~Funct3E[0]; // simplified from (Funct3E[2]&~Funct3E[1]&~Funct3E[0]) | (Funct3E[2]&Funct3E[1]&~Funct3E[0]);
//intdiv #(`XLEN) div (QuotE, RemE, DivDoneE, DivBusyE, div0error, N, D, gclk, reset, startDivideE, signedDivide);
// intdivrestoring div(.clk, .reset, .StallM, .signedDivide, .start(startDivideE), .X(X), .D(D), .busy(busy), .done(DivDoneE), .Q(QuotE), .REM(RemE));
intdivrestoring div(.clk, .reset, .StallM, .signedDivide, .start(startDivideE), .X(X), .D(D), .busy(busy), .done(DivDoneE), .Q(QuotE), .REM(RemE));
assign SignedDivideE = ~Funct3E[0]; // simplified from (Funct3E[2]&~Funct3E[1]&~Funct3E[0]) | (Funct3E[2]&Funct3E[1]&~Funct3E[0]);
//intdiv #(`XLEN) div (QuotE, RemE, DivDoneE, DivBusyE, div0error, N, D, gclk, reset, startDivideE, SignedDivideE);
intdivrestoring div(.clk, .reset, .StallM, .signedDivide(SignedDivideE), .start(startDivideE), .X(X), .D(D), .busy(busy), .done(DivDoneE), .Q(QuotM), .REM(RemM));
// Start a divide when a new division instruction is received and the divider isn't already busy or finishing
assign startDivideE = MulDivE & Funct3E[2] & ~busy & ~DivDoneE;
assign startDivideE = MulDivE & Funct3E[2] & ~busy & ~DivDoneE; // *** mabye DivDone should be M stage
assign DivBusyE = startDivideE | busy;
// Select result
always_comb
case (Funct3E)
3'b000: PrelimResultE = ProdE[`XLEN-1:0];
3'b001: PrelimResultE = ProdE[`XLEN*2-1:`XLEN];
3'b010: PrelimResultE = ProdE[`XLEN*2-1:`XLEN];
3'b011: PrelimResultE = ProdE[`XLEN*2-1:`XLEN];
3'b100: PrelimResultE = QuotE;
3'b101: PrelimResultE = QuotE;
3'b110: PrelimResultE = RemE;
3'b111: PrelimResultE = RemE;
endcase // case (Funct3E)
case (Funct3M)
3'b000: PrelimResultM = ProdM[`XLEN-1:0];
3'b001: PrelimResultM = ProdM[`XLEN*2-1:`XLEN];
3'b010: PrelimResultM = ProdM[`XLEN*2-1:`XLEN];
3'b011: PrelimResultM = ProdM[`XLEN*2-1:`XLEN];
3'b100: PrelimResultM = QuotM;
3'b101: PrelimResultM = QuotM;
3'b110: PrelimResultM = RemM;
3'b111: PrelimResultM = RemM;
endcase
// Handle sign extension for W-type instructions
flopenrc #(1) W64MReg(clk, reset, FlushM, ~StallM, W64E, W64M);
if (`XLEN == 64) begin // RV64 has W-type instructions
assign MulDivResultE = W64E ? {{32{PrelimResultE[31]}}, PrelimResultE[31:0]} : PrelimResultE;
assign MulDivResultM = W64M ? {{32{PrelimResultM[31]}}, PrelimResultM[31:0]} : PrelimResultM;
end else begin // RV32 has no W-type instructions
assign MulDivResultE = PrelimResultE;
assign MulDivResultM = PrelimResultM;
end
flopenrc #(`XLEN) MulDivResultMReg(clk, reset, FlushM, ~StallM, MulDivResultE, MulDivResultM); // could let part of multiplication spill into Memory stage
flopenrc #(`XLEN) MulDivResultWReg(clk, reset, FlushW, ~StallW, MulDivResultM, MulDivResultW);
end else begin // no M instructions supported