wu-arch/kernels
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

decode optimization

Browse Source
This commit is contained in:
Blaise Tine
2021-04-01 19:08:15 -07:00
parent 41413a51ba
commit 638614fd6d
1 changed files with 95 additions and 76 deletions
Download Patch File Download Diff File Expand all files Collapse all files

151
hw/rtl/VX_decode.v
View File

@@ -22,10 +22,12 @@ module VX_decode #(
reg [`EX_BITS-1:0] ex_type; reg [`EX_BITS-1:0] ex_type;
reg [`OP_BITS-1:0] op_type; reg [`OP_BITS-1:0] op_type;
reg [`MOD_BITS-1:0] op_mod; reg [`MOD_BITS-1:0] op_mod;
reg [4:0] rd_r, rs1_r, rs2_r, rs3_r;
reg [31:0] imm; reg [31:0] imm;
reg use_rd, use_rs1, use_rs2, use_rs3, use_PC, use_imm; reg use_rd, use_PC, use_imm;
reg rd_fp, rs1_fp, rs2_fp; reg rd_fp, rs1_fp, rs2_fp;
reg is_join, is_wstall; reg is_join, is_wstall;
reg [`NUM_REGS-1:0] used_regs;
wire [31:0] instr = ifetch_rsp_if.instr; wire [31:0] instr = ifetch_rsp_if.instr;
wire [6:0] opcode = instr[6:0]; wire [6:0] opcode = instr[6:0];
@@ -45,14 +47,11 @@ module VX_decode #(
always @(*) begin always @(*) begin
ex_type = `EX_NOP; ex_type = 0;
op_type = 'x; op_type = 'x;
op_mod = 'x; op_mod = 'x;
imm = 'x; imm = 'x;
use_rd = 0; use_rd = 0;
use_rs1 = 0;
use_rs2 = 0;
use_rs3 = 0;
use_PC = 0; use_PC = 0;
use_imm = 0; use_imm = 0;
rd_fp = 0; rd_fp = 0;
@@ -60,6 +59,11 @@ module VX_decode #(
rs2_fp = 0; rs2_fp = 0;
is_join = 0; is_join = 0;
is_wstall = 0; is_wstall = 0;
used_regs = 0;
rd_r = rd;
rs1_r = rs1;
rs2_r = rs2;
rs3_r = rs3;
case (opcode) case (opcode)
`INST_I: begin `INST_I: begin
@@ -78,8 +82,9 @@ module VX_decode #(
op_mod = 0; op_mod = 0;
imm = {{20{alu_imm[11]}}, alu_imm}; imm = {{20{alu_imm[11]}}, alu_imm};
use_rd = 1; use_rd = 1;
use_rs1 = 1;
use_imm = 1; use_imm = 1;
used_regs[{1'b0, rd}] = 1;
used_regs[{1'b0, rs1}] = 1;
end end
`INST_R: begin `INST_R: begin
ex_type = `EX_ALU; ex_type = `EX_ALU;
@@ -113,18 +118,21 @@ module VX_decode #(
endcase endcase
op_mod = 0; op_mod = 0;
end end
use_rd = 1; use_rd = 1;
use_rs1 = 1; used_regs[{1'b0, rd}] = 1;
use_rs2 = 1; used_regs[{1'b0, rs1}] = 1;
used_regs[{1'b0, rs2}] = 1;
end end
`INST_LUI: begin `INST_LUI: begin
ex_type = `EX_ALU; ex_type = `EX_ALU;
op_type = `OP_BITS'(`ALU_LUI); op_type = `OP_BITS'(`ALU_LUI);
op_mod = 0; op_mod = 0;
rs1_r = 0;
imm = {upper_imm, 12'(0)}; imm = {upper_imm, 12'(0)};
use_rd = 1; use_rd = 1;
use_rs1 = 1;
use_imm = 1; use_imm = 1;
used_regs[{1'b0, rd}] = 1;
used_regs[{1'b0, 5'b0}] = 1;
end end
`INST_AUIPC: begin `INST_AUIPC: begin
ex_type = `EX_ALU; ex_type = `EX_ALU;
@@ -134,6 +142,7 @@ module VX_decode #(
use_rd = 1; use_rd = 1;
use_PC = 1; use_PC = 1;
use_imm = 1; use_imm = 1;
used_regs[{1'b0, rd}] = 1;
end end
`INST_JAL: begin `INST_JAL: begin
ex_type = `EX_ALU; ex_type = `EX_ALU;
@@ -144,6 +153,7 @@ module VX_decode #(
use_PC = 1; use_PC = 1;
use_imm = 1; use_imm = 1;
is_wstall = 1; is_wstall = 1;
used_regs[{1'b0, rd}] = 1;
end end
`INST_JALR: begin `INST_JALR: begin
ex_type = `EX_ALU; ex_type = `EX_ALU;
@@ -151,9 +161,10 @@ module VX_decode #(
op_mod = 1; op_mod = 1;
imm = {{20{jalr_imm[11]}}, jalr_imm}; imm = {{20{jalr_imm[11]}}, jalr_imm};
use_rd = 1; use_rd = 1;
use_rs1 = 1;
use_imm = 1; use_imm = 1;
is_wstall = 1; is_wstall = 1;
used_regs[{1'b0, rd}] = 1;
used_regs[{1'b0, rs1}] = 1;
end end
`INST_B: begin `INST_B: begin
ex_type = `EX_ALU; ex_type = `EX_ALU;
@@ -168,11 +179,11 @@ module VX_decode #(
endcase endcase
op_mod = 1; op_mod = 1;
imm = {{20{instr[31]}}, instr[7], instr[30:25], instr[11:8], 1'b0}; imm = {{20{instr[31]}}, instr[7], instr[30:25], instr[11:8], 1'b0};
use_rs1 = 1;
use_rs2 = 1;
use_PC = 1; use_PC = 1;
use_imm = 1; use_imm = 1;
is_wstall = 1; is_wstall = 1;
used_regs[{1'b0, rs1}] = 1;
used_regs[{1'b0, rs2}] = 1;
end end
`INST_SYS : begin `INST_SYS : begin
if (func3 == 0) begin if (func3 == 0) begin
@@ -190,6 +201,7 @@ module VX_decode #(
use_rd = 1; use_rd = 1;
use_PC = 1; use_PC = 1;
use_imm = 1; use_imm = 1;
used_regs[{1'b0, rd}] = 1;
end else begin end else begin
ex_type = `EX_CSR; ex_type = `EX_CSR;
case (func3[1:0]) case (func3[1:0])
@@ -201,8 +213,10 @@ module VX_decode #(
endcase endcase
imm = 32'(u_12); imm = 32'(u_12);
use_rd = 1; use_rd = 1;
use_rs1 = !func3[2];
use_imm = func3[2]; use_imm = func3[2];
used_regs[{1'b0, rd}] = 1;
if (!func3[2])
used_regs[{1'b0, rs1}] = 1;
end end
end end
`ifdef EXT_F_ENABLE `ifdef EXT_F_ENABLE
@@ -213,9 +227,12 @@ module VX_decode #(
op_type = `OP_BITS'({1'b0, func3}); op_type = `OP_BITS'({1'b0, func3});
imm = {{20{u_12[11]}}, u_12}; imm = {{20{u_12[11]}}, u_12};
use_rd = 1; use_rd = 1;
use_rs1 = 1; used_regs[{1'b0, rs1}] = 1;
`ifdef EXT_F_ENABLE `ifdef EXT_F_ENABLE
used_regs[{(opcode == `INST_FL), rd}] = 1;
rd_fp = (opcode == `INST_FL); rd_fp = (opcode == `INST_FL);
`else
used_regs[{1'b0, rd}] = 1;
`endif `endif
end end
`ifdef EXT_F_ENABLE `ifdef EXT_F_ENABLE
@@ -225,8 +242,8 @@ module VX_decode #(
ex_type = `EX_LSU; ex_type = `EX_LSU;
op_type = `OP_BITS'({1'b1, func3}); op_type = `OP_BITS'({1'b1, func3});
imm = {{20{func7[6]}}, func7, rd}; imm = {{20{func7[6]}}, func7, rd};
use_rs1 = 1; used_regs[{1'b0, rs1}] = 1;
use_rs2 = 1; used_regs[{1'b0, rs2}] = 1;
`ifdef EXT_F_ENABLE `ifdef EXT_F_ENABLE
rs2_fp = (opcode == `INST_FS); rs2_fp = (opcode == `INST_FS);
`endif `endif
@@ -240,12 +257,13 @@ module VX_decode #(
op_type = `OP_BITS'(opcode[3:0]); op_type = `OP_BITS'(opcode[3:0]);
op_mod = func3; op_mod = func3;
use_rd = 1; use_rd = 1;
use_rs1 = 1;
use_rs2 = 1;
use_rs3 = 1;
rd_fp = 1; rd_fp = 1;
rs1_fp = 1; rs1_fp = 1;
rs2_fp = 1; rs2_fp = 1;
used_regs[{1'b1, rd}] = 1;
used_regs[{1'b1, rs1}] = 1;
used_regs[{1'b1, rs2}] = 1;
used_regs[{1'b1, rs3}] = 1;
end end
`INST_FCI: begin `INST_FCI: begin
ex_type = `EX_FPU; ex_type = `EX_FPU;
@@ -258,55 +276,61 @@ module VX_decode #(
7'h0C: // FDIV 7'h0C: // FDIV
begin begin
op_type = `OP_BITS'(func7[3:0]); op_type = `OP_BITS'(func7[3:0]);
use_rd = 1;
use_rs1 = 1;
use_rs2 = 1;
rd_fp = 1; rd_fp = 1;
rs1_fp = 1; rs1_fp = 1;
rs2_fp = 1; rs2_fp = 1;
used_regs[{1'b1, rd}] = 1;
used_regs[{1'b1, rs1}] = 1;
used_regs[{1'b1, rs2}] = 1;
end end
7'h2C: begin 7'h2C: begin
op_type = `OP_BITS'(`FPU_SQRT); op_type = `OP_BITS'(`FPU_SQRT);
use_rs1 = 1;
rd_fp = 1; rd_fp = 1;
rs1_fp = 1; rs1_fp = 1;
used_regs[{1'b1, rd}] = 1;
used_regs[{1'b1, rs1}] = 1;
end end
7'h50: begin 7'h50: begin
op_type = `OP_BITS'(`FPU_CMP); op_type = `OP_BITS'(`FPU_CMP);
use_rs1 = 1;
use_rs2 = 1;
rs1_fp = 1; rs1_fp = 1;
rs2_fp = 1; rs2_fp = 1;
used_regs[{1'b0, rd}] = 1;
used_regs[{1'b1, rs1}] = 1;
used_regs[{1'b1, rs2}] = 1;
end end
7'h60: begin 7'h60: begin
op_type = (instr[20]) ? `OP_BITS'(`FPU_CVTWUS) : `OP_BITS'(`FPU_CVTWS); op_type = (instr[20]) ? `OP_BITS'(`FPU_CVTWUS) : `OP_BITS'(`FPU_CVTWS);
use_rs1 = 1;
rs1_fp = 1; rs1_fp = 1;
used_regs[{1'b0, rd}] = 1;
used_regs[{1'b1, rs1}] = 1;
end end
7'h68: begin 7'h68: begin
op_type = (instr[20]) ? `OP_BITS'(`FPU_CVTSWU) : `OP_BITS'(`FPU_CVTSW); op_type = (instr[20]) ? `OP_BITS'(`FPU_CVTSWU) : `OP_BITS'(`FPU_CVTSW);
use_rs1 = 1;
rd_fp = 1; rd_fp = 1;
used_regs[{1'b1, rd}] = 1;
used_regs[{1'b0, rs1}] = 1;
end end
7'h10: begin 7'h10: begin
// FSGNJ=0, FSGNJN=1, FSGNJX=2 // FSGNJ=0, FSGNJN=1, FSGNJX=2
op_type = `OP_BITS'(`FPU_MISC); op_type = `OP_BITS'(`FPU_MISC);
op_mod = {1'b0, func3[1:0]}; op_mod = {1'b0, func3[1:0]};
use_rs1 = 1;
use_rs2 = 1;
rd_fp = 1; rd_fp = 1;
rs1_fp = 1; rs1_fp = 1;
rs2_fp = 1; rs2_fp = 1;
used_regs[{1'b1, rd}] = 1;
used_regs[{1'b1, rs1}] = 1;
used_regs[{1'b1, rs2}] = 1;
end end
7'h14: begin 7'h14: begin
// FMIN=3, FMAX=4 // FMIN=3, FMAX=4
op_type = `OP_BITS'(`FPU_MISC); op_type = `OP_BITS'(`FPU_MISC);
op_mod = func3[0] ? 4 : 3; op_mod = func3[0] ? 4 : 3;
use_rs1 = 1;
use_rs2 = 1;
rd_fp = 1; rd_fp = 1;
rs1_fp = 1; rs1_fp = 1;
rs2_fp = 1; rs2_fp = 1;
used_regs[{1'b1, rd}] = 1;
used_regs[{1'b1, rs1}] = 1;
used_regs[{1'b1, rs2}] = 1;
end end
7'h70: begin 7'h70: begin
if (func3[0]) begin if (func3[0]) begin
@@ -317,14 +341,16 @@ module VX_decode #(
op_type = `OP_BITS'(`FPU_MISC); op_type = `OP_BITS'(`FPU_MISC);
op_mod = 5; op_mod = 5;
end end
use_rs1 = 1;
rs1_fp = 1; rs1_fp = 1;
used_regs[{1'b0, rd}] = 1;
used_regs[{1'b0, rs1}] = 1;
end end
7'h78: begin 7'h78: begin
// FMV.W.X=6 // FMV.W.X=6
op_type = `OP_BITS'(`FPU_MISC); op_type = `OP_BITS'(`FPU_MISC);
op_mod = 6; op_mod = 6;
rd_fp = 1; rd_fp = 1;
used_regs[{1'b1, rd}] = 1;
end end
default:; default:;
endcase endcase
@@ -335,18 +361,18 @@ module VX_decode #(
case (func3) case (func3)
3'h0: begin 3'h0: begin
op_type = `OP_BITS'(`GPU_TMC); op_type = `OP_BITS'(`GPU_TMC);
use_rs1 = 1;
is_wstall = 1; is_wstall = 1;
used_regs[{1'b0, rs1}] = 1;
end end
3'h1: begin 3'h1: begin
op_type = `OP_BITS'(`GPU_WSPAWN); op_type = `OP_BITS'(`GPU_WSPAWN);
use_rs1 = 1; used_regs[{1'b0, rs1}] = 1;
use_rs2 = 1; used_regs[{1'b0, rs2}] = 1;
end end
3'h2: begin 3'h2: begin
op_type = `OP_BITS'(`GPU_SPLIT); op_type = `OP_BITS'(`GPU_SPLIT);
use_rs1 = 1;
is_wstall = 1; is_wstall = 1;
used_regs[{1'b0, rs1}] = 1;
end end
3'h3: begin 3'h3: begin
op_type = `OP_BITS'(`GPU_JOIN); op_type = `OP_BITS'(`GPU_JOIN);
@@ -354,9 +380,9 @@ module VX_decode #(
end end
3'h4: begin 3'h4: begin
op_type = `OP_BITS'(`GPU_BAR); op_type = `OP_BITS'(`GPU_BAR);
use_rs1 = 1;
use_rs2 = 1;
is_wstall = 1; is_wstall = 1;
used_regs[{1'b0, rs1}] = 1;
used_regs[{1'b0, rs2}] = 1;
end end
default:; default:;
endcase endcase
@@ -366,10 +392,7 @@ module VX_decode #(
end end
// disable write to integer register r0 // disable write to integer register r0
wire use_rd_qual = use_rd && (rd_fp || (rd != 0)); wire wb = use_rd && (rd_fp || (rd_r != 0));
// EX_ALU needs rs1=0 for LUI operation
wire [4:0] rs1_qual = (opcode == `INST_LUI) ? 5'h0 : rs1;
assign decode_if.valid = ifetch_rsp_if.valid; assign decode_if.valid = ifetch_rsp_if.valid;
assign decode_if.wid = ifetch_rsp_if.wid; assign decode_if.wid = ifetch_rsp_if.wid;
@@ -378,31 +401,27 @@ module VX_decode #(
assign decode_if.ex_type = ex_type; assign decode_if.ex_type = ex_type;
assign decode_if.op_type = op_type; assign decode_if.op_type = op_type;
assign decode_if.op_mod = op_mod; assign decode_if.op_mod = op_mod;
assign decode_if.wb = use_rd_qual; assign decode_if.wb = wb;
`ifdef EXT_F_ENABLE `ifdef EXT_F_ENABLE
assign decode_if.rd = {rd_fp, rd}; assign decode_if.rd = {rd_fp, rd_r};
assign decode_if.rs1 = {rs1_fp, rs1_qual}; assign decode_if.rs1 = {rs1_fp, rs1_r};
assign decode_if.rs2 = {rs2_fp, rs2}; assign decode_if.rs2 = {rs2_fp, rs2_r};
assign decode_if.rs3 = {1'b1, rs3}; assign decode_if.rs3 = {1'b1, rs3_r};
`else `else
`UNUSED_VAR (rd_fp) `UNUSED_VAR (rd_fp)
`UNUSED_VAR (rs1_fp) `UNUSED_VAR (rs1_fp)
`UNUSED_VAR (rs2_fp) `UNUSED_VAR (rs2_fp)
assign decode_if.rd = rd; assign decode_if.rd = rd_r;
assign decode_if.rs1 = rs1_qual; assign decode_if.rs1 = rs1_r;
assign decode_if.rs2 = rs2; assign decode_if.rs2 = rs2_r;
assign decode_if.rs3 = rs3; assign decode_if.rs3 = rs3_r;
`endif `endif
assign decode_if.imm = imm; assign decode_if.imm = imm;
assign decode_if.use_PC = use_PC; assign decode_if.use_PC = use_PC;
assign decode_if.use_imm = use_imm; assign decode_if.use_imm = use_imm;
assign decode_if.used_regs = used_regs;
assign decode_if.used_regs = (`NUM_REGS'(use_rd) << decode_if.rd)
| (`NUM_REGS'(use_rs1) << decode_if.rs1)
| (`NUM_REGS'(use_rs2) << decode_if.rs2)
| (`NUM_REGS'(use_rs3) << decode_if.rs3);
/////////////////////////////////////////////////////////////////////////// ///////////////////////////////////////////////////////////////////////////