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

pipeline refactoring

Browse Source
This commit is contained in:
Blaise Tine
2020-07-19 05:03:47 -04:00
parent 9cf8bf6149
commit 25f66e6490
71 changed files with 2242 additions and 2379 deletions
Download Patch File Download Diff File Expand all files Collapse all files

528
hw/rtl/VX_decode.v
View File

@@ -1,321 +1,279 @@
`include "VX_define.vh"
module VX_decode(
// Fetch Inputs
VX_inst_meta_if fd_inst_meta_de,
module VX_decode #(
parameter CORE_ID = 0
) (
input wire clk,
input wire reset,
// Outputs
VX_backend_req_if frE_to_bckE_req_if,
VX_wstall_if wstall_if,
VX_join_if join_if
// inputs
VX_ifetch_rsp_if ifetch_rsp_if,
// outputs
VX_decode_if decode_if,
VX_wstall_if wstall_if,
VX_join_if join_if
);
wire in_valid = (| fd_inst_meta_de.valid);
wire[31:0] in_instruction = fd_inst_meta_de.instruction;
wire[31:0] in_curr_PC = fd_inst_meta_de.curr_PC;
wire[`NW_BITS-1:0] in_warp_num = fd_inst_meta_de.warp_num;
wire in_valid = (| ifetch_rsp_if.valid);
wire [31:0] instr = ifetch_rsp_if.instr;
assign frE_to_bckE_req_if.curr_PC = in_curr_PC;
reg [`ALU_BITS-1:0] alu_op;
reg [`BR_BITS-1:0] br_op;
reg [`MUL_BITS-1:0] mul_op;
wire [`LSU_BITS-1:0] lsu_op;
reg [`CSR_BITS-1:0] csr_op;
reg [`GPU_BITS-1:0] gpu_op;
wire[6:0] curr_opcode;
reg [19:0] upper_imm;
reg [31:0] jalx_offset;
reg [31:0] src2_imm;
wire is_itype;
wire is_rtype;
wire is_stype;
wire is_btype;
wire is_linst;
wire is_jal;
wire is_jalr;
wire is_lui;
wire is_auipc;
wire is_csr;
wire is_csr_immed;
wire is_etype;
wire [6:0] opcode = instr[6:0];
wire [2:0] func3 = instr[14:12];
wire [6:0] func7 = instr[31:25];
wire [11:0] u_12 = instr[31:20];
wire is_gpgpu;
wire is_wspawn;
wire is_tmc;
wire is_split;
wire is_join;
wire is_barrier;
wire [`NR_BITS-1:0] rd = instr[11:7];
wire [`NR_BITS-1:0] rs1 = instr[19:15];
wire [`NR_BITS-1:0] rs2 = instr[24:20];
wire[2:0] func3;
wire[6:0] func7;
wire[11:0] u_12;
wire[7:0] jal_b_19_to_12;
wire jal_b_11;
wire[9:0] jal_b_10_to_1;
wire jal_b_20;
wire jal_b_0;
wire[20:0] jal_unsigned_offset;
wire[31:0] jal_1_offset;
wire[11:0] jalr_immed;
wire[31:0] jal_2_offset;
wire jal_sys_cond1;
wire jal_sys_cond2;
wire jal_sys_jal;
wire[31:0] jal_sys_off;
wire csr_cond1;
wire csr_cond2;
wire[11:0] alu_tempp;
wire alu_shift_i;
wire[11:0] alu_shift_i_immed;
wire[1:0] csr_type;
reg[4:0] csr_alu;
reg[4:0] alu_op;
reg[4:0] mul_alu;
reg[19:0] temp_upper_immed;
reg temp_jal;
reg[31:0] temp_jal_offset;
reg[31:0] temp_itype_immed;
reg[2:0] temp_branch_type;
reg temp_branch_stall;
assign frE_to_bckE_req_if.valid = fd_inst_meta_de.valid;
assign frE_to_bckE_req_if.warp_num = in_warp_num;
assign curr_opcode = in_instruction[6:0];
assign frE_to_bckE_req_if.rd = in_instruction[11:7];
assign frE_to_bckE_req_if.rs1 = in_instruction[19:15];
assign frE_to_bckE_req_if.rs2 = in_instruction[24:20];
assign func3 = in_instruction[14:12];
assign func7 = in_instruction[31:25];
assign u_12 = in_instruction[31:20];
assign frE_to_bckE_req_if.next_PC = in_curr_PC + 32'h4;
// Write Back sigal
assign is_rtype = (curr_opcode == `INST_R);
assign is_linst = (curr_opcode == `INST_L);
assign is_itype = (curr_opcode == `INST_ALU) || is_linst;
assign is_stype = (curr_opcode == `INST_S);
assign is_btype = (curr_opcode == `INST_B);
assign is_jal = (curr_opcode == `INST_JAL);
assign is_jalr = (curr_opcode == `INST_JALR);
assign is_lui = (curr_opcode == `INST_LUI);
assign is_auipc = (curr_opcode == `INST_AUIPC);
assign is_csr = (curr_opcode == `INST_SYS) && (func3 != 0);
assign is_csr_immed = is_csr && (func3[2] == 1);
assign is_gpgpu = (curr_opcode == `INST_GPGPU);
assign is_tmc = is_gpgpu && (func3 == 0); // Goes to BE
assign is_wspawn = is_gpgpu && (func3 == 1); // Goes to BE
assign is_barrier = is_gpgpu && (func3 == 4); // Goes to BE
assign is_split = is_gpgpu && (func3 == 2); // Goes to BE
assign is_join = is_gpgpu && (func3 == 3); // Doesn't go to BE
assign join_if.is_join = is_join && in_valid;
assign join_if.warp_num = in_warp_num;
assign frE_to_bckE_req_if.is_wspawn = is_wspawn;
assign frE_to_bckE_req_if.is_tmc = is_tmc;
assign frE_to_bckE_req_if.is_split = is_split;
assign frE_to_bckE_req_if.is_barrier = is_barrier;
assign frE_to_bckE_req_if.csr_immed = is_csr_immed;
assign frE_to_bckE_req_if.is_csr = is_csr;
assign frE_to_bckE_req_if.wb = (is_jal || is_jalr || is_etype) ? `WB_JAL :
is_linst ? `WB_MEM :
(is_itype || is_rtype || is_lui || is_auipc || is_csr) ? `WB_ALU :
`WB_NO;
assign frE_to_bckE_req_if.rs2_src = (is_itype || is_stype) ? `RS2_IMMED : `RS2_REG;
// MEM signals
assign frE_to_bckE_req_if.mem_read = (is_linst) ? func3 : `BYTE_EN_NO;
assign frE_to_bckE_req_if.mem_write = (is_stype) ? func3 : `BYTE_EN_NO;
// UPPER IMMEDIATE
// opcode types
wire is_rtype = (opcode == `INST_R);
wire is_ltype = (opcode == `INST_L);
wire is_itype = (opcode == `INST_I);
wire is_stype = (opcode == `INST_S);
wire is_btype = (opcode == `INST_B);
wire is_jal = (opcode == `INST_JAL);
wire is_jalr = (opcode == `INST_JALR);
wire is_lui = (opcode == `INST_LUI);
wire is_auipc = (opcode == `INST_AUIPC);
wire is_jals = (opcode == `INST_SYS) && (func3 == 0);
wire is_csr = (opcode == `INST_SYS) && (func3 != 0);
wire is_gpu = (opcode == `INST_GPU);
wire is_br = (is_btype || is_jal || is_jalr || is_jals);
wire is_mul = is_rtype && (func7 == 7'h1);
// upper immediate
always @(*) begin
case (curr_opcode)
`INST_LUI: temp_upper_immed = {func7, frE_to_bckE_req_if.rs2, frE_to_bckE_req_if.rs1, func3};
`INST_AUIPC: temp_upper_immed = {func7, frE_to_bckE_req_if.rs2, frE_to_bckE_req_if.rs1, func3};
default: temp_upper_immed = 20'h0;
endcase // curr_opcode
end
assign frE_to_bckE_req_if.upper_immed = temp_upper_immed;
assign jal_b_19_to_12 = in_instruction[19:12];
assign jal_b_11 = in_instruction[20];
assign jal_b_10_to_1 = in_instruction[30:21];
assign jal_b_20 = in_instruction[31];
assign jal_b_0 = 1'b0;
assign jal_unsigned_offset = {jal_b_20, jal_b_19_to_12, jal_b_11, jal_b_10_to_1, jal_b_0};
assign jal_1_offset = {{11{jal_b_20}}, jal_unsigned_offset};
assign jalr_immed = {func7, frE_to_bckE_req_if.rs2};
assign jal_2_offset = {{20{jalr_immed[11]}}, jalr_immed};
assign jal_sys_cond1 = (func3 == 3'h0);
assign jal_sys_cond2 = (u_12 < 12'h2);
assign jal_sys_jal = (jal_sys_cond1 && jal_sys_cond2) ? 1'b1 : 1'b0;
assign jal_sys_off = (jal_sys_cond1 && jal_sys_cond2) ? 32'hb0000000 : 32'hdeadbeef;
// JAL
always @(*) begin
case (curr_opcode)
`INST_JAL:
begin
temp_jal = in_valid;
temp_jal_offset = jal_1_offset;
end
`INST_JALR:
begin
temp_jal = in_valid;
temp_jal_offset = jal_2_offset;
end
`INST_SYS:
begin
// $display("SYS EBREAK %h", (jal_sys_jal && in_valid));
temp_jal = jal_sys_jal && in_valid;
temp_jal_offset = jal_sys_off;
end
default:
begin
temp_jal = 1'b0;
temp_jal_offset = 32'hdeadbeef;
end
case (opcode)
`INST_LUI: upper_imm = {func7, rs2, rs1, func3};
`INST_AUIPC: upper_imm = {func7, rs2, rs1, func3};
default: upper_imm = 20'h0;
endcase
end
end
assign frE_to_bckE_req_if.is_jal = is_jal;
assign frE_to_bckE_req_if.jal = temp_jal;
assign frE_to_bckE_req_if.jal_offset = temp_jal_offset;
// ecall/ebreak
assign is_etype = (curr_opcode == `INST_SYS) && jal_sys_jal;
assign frE_to_bckE_req_if.is_etype = is_etype;
// CSR
assign csr_cond1 = func3 != 3'h0;
assign csr_cond2 = u_12 >= 12'h2;
assign frE_to_bckE_req_if.csr_addr = (csr_cond1 && csr_cond2) ? u_12 : 12'h55;
// ITYPE IMEED
assign alu_shift_i = (func3 == 3'h1) || (func3 == 3'h5);
assign alu_shift_i_immed = {{7{1'b0}}, frE_to_bckE_req_if.rs2};
assign alu_tempp = alu_shift_i ? alu_shift_i_immed : u_12;
always @(*) begin
case (curr_opcode)
`INST_ALU: temp_itype_immed = {{20{alu_tempp[11]}}, alu_tempp};
`INST_S: temp_itype_immed = {{20{func7[6]}}, func7, frE_to_bckE_req_if.rd};
`INST_L: temp_itype_immed = {{20{u_12[11]}}, u_12};
`INST_B: temp_itype_immed = {{20{in_instruction[31]}}, in_instruction[31], in_instruction[7], in_instruction[30:25], in_instruction[11:8]};
default: temp_itype_immed = 32'hdeadbeef;
// JAL
wire [20:0] jal_imm = {instr[31], instr[19:12], instr[20], instr[30:21], 1'b0};
wire [31:0] jal_offset = {{11{jal_imm[20]}}, jal_imm};
wire [11:0] jalr_imm = {func7, rs2};
wire [31:0] jalr_offset = {{20{jalr_imm[11]}}, jalr_imm};
always @(*) begin
case (opcode)
`INST_JAL: jalx_offset = jal_offset;
`INST_JALR: jalx_offset = jalr_offset;
default: jalx_offset = 32'd4;
endcase
end
assign frE_to_bckE_req_if.itype_immed = temp_itype_immed;
end
// I-type immediate
wire alu_shift_i = (func3 == 3'h1) || (func3 == 3'h5);
wire [11:0] alu_shift_imm = {{7{1'b0}}, rs2};
wire [11:0] alu_imm = alu_shift_i ? alu_shift_imm : u_12;
always @(*) begin
case (curr_opcode)
case (opcode)
`INST_I: src2_imm = {{20{alu_imm[11]}}, alu_imm};
`INST_S: src2_imm = {{20{func7[6]}}, func7, rd};
`INST_L: src2_imm = {{20{u_12[11]}}, u_12};
`INST_B: src2_imm = {{20{instr[31]}}, instr[7], instr[30:25], instr[11:8], 1'b0};
default: src2_imm = 32'hdeadbeef;
endcase
end
// BRANCH
always @(*) begin
br_op = `BR_OTHER;
case (opcode)
`INST_B: begin
// $display("BRANCH IN DECODE");
temp_branch_stall = in_valid;
case (func3)
3'h0: temp_branch_type = `BR_EQ;
3'h1: temp_branch_type = `BR_NE;
3'h4: temp_branch_type = `BR_LT;
3'h5: temp_branch_type = `BR_GT;
3'h6: temp_branch_type = `BR_LTU;
3'h7: temp_branch_type = `BR_GTU;
default: temp_branch_type = `BR_NO;
3'h0: br_op = `BR_EQ;
3'h1: br_op = `BR_NE;
3'h4: br_op = `BR_LT;
3'h5: br_op = `BR_GE;
3'h6: br_op = `BR_LTU;
3'h7: br_op = `BR_GEU;
default:;
endcase
end
`INST_JAL: begin
temp_branch_type = `BR_NO;
temp_branch_stall = in_valid;
end
`INST_JALR: begin
temp_branch_type = `BR_NO;
temp_branch_stall = in_valid;
end
default: begin
temp_branch_type = `BR_NO;
temp_branch_stall = 1'b0;
`INST_JAL: br_op = `BR_JAL;
`INST_JALR: br_op = `BR_JALR;
`INST_SYS: begin
if (is_jals && u_12 == 12'h000) br_op = `BR_ECALL;
if (is_jals && u_12 == 12'h001) br_op = `BR_EBREAK;
if (is_jals && u_12 == 12'h302) br_op = `BR_MRET;
if (is_jals && u_12 == 12'h102) br_op = `BR_SRET;
if (is_jals && u_12 == 12'h7B2) br_op = `BR_DRET;
end
default:;
endcase
end
assign frE_to_bckE_req_if.branch_type = temp_branch_type;
assign wstall_if.wstall = (temp_branch_stall || is_tmc || is_split || is_barrier) && in_valid;
assign wstall_if.warp_num = in_warp_num;
// ALU
always @(*) begin
// ALU OP
alu_op = `ALU_OTHER;
if (is_lui) begin
alu_op = `ALU_LUI;
end else if (is_auipc) begin
alu_op = `ALU_AUIPC;
end else if (is_itype || is_rtype) begin
case (func3)
3'h0: alu_op = (is_rtype && func7 == 7'h20) ? `ALU_SUB : `ALU_ADD;
3'h1: alu_op = `ALU_SLL;
3'h2: alu_op = `ALU_SLT;
3'h3: alu_op = `ALU_SLTU;
3'h4: alu_op = `ALU_XOR;
3'h5: alu_op = (func7 == 7'h0) ? `ALU_SRL : `ALU_SRA;
3'h6: alu_op = `ALU_OR;
3'h7: alu_op = `ALU_AND;
default:;
endcase
end
end
// MUL
always @(*) begin
mul_op = `MUL_MUL;
case (func3)
3'h0: alu_op = (curr_opcode == `INST_ALU) ? `ALU_ADD : (func7 == 7'h0 ? `ALU_ADD : `ALU_SUB);
3'h1: alu_op = `ALU_SLLA;
3'h2: alu_op = `ALU_SLT;
3'h3: alu_op = `ALU_SLTU;
3'h4: alu_op = `ALU_XOR;
3'h5: alu_op = (func7 == 7'h0) ? `ALU_SRL : `ALU_SRA;
3'h6: alu_op = `ALU_OR;
3'h7: alu_op = `ALU_AND;
default: alu_op = `ALU_NO;
3'h0: mul_op = `MUL_MUL;
3'h1: mul_op = `MUL_MULH;
3'h2: mul_op = `MUL_MULHSU;
3'h3: mul_op = `MUL_MULHU;
3'h4: mul_op = `MUL_DIV;
3'h5: mul_op = `MUL_DIVU;
3'h6: mul_op = `MUL_REM;
3'h7: mul_op = `MUL_REMU;
default:;
endcase
end
// LSU
wire is_lsu = (is_ltype || is_stype);
assign lsu_op = {is_stype, func3};
// CSR
wire is_csr_imm = is_csr && (func3[2] == 1);
always @(*) begin
// ALU OP
csr_op = `CSR_OTHER;
case (func3[1:0])
2'h1: csr_op = `CSR_RW;
2'h2: csr_op = `CSR_RS;
2'h3: csr_op = `CSR_RC;
default:;
endcase
end
// GPU
always @(*) begin
gpu_op = `GPU_OTHER;
case (func3)
3'h0: mul_alu = `ALU_MUL;
3'h1: mul_alu = `ALU_MULH;
3'h2: mul_alu = `ALU_MULHSU;
3'h3: mul_alu = `ALU_MULHU;
3'h4: mul_alu = `ALU_DIV;
3'h5: mul_alu = `ALU_DIVU;
3'h6: mul_alu = `ALU_REM;
3'h7: mul_alu = `ALU_REMU;
default: mul_alu = `ALU_NO;
3'h0: gpu_op = `GPU_TMC;
3'h1: gpu_op = `GPU_WSPAWN;
3'h2: gpu_op = `GPU_SPLIT;
3'h3: gpu_op = `GPU_JOIN;
3'h4: gpu_op = `GPU_BAR;
default:;
endcase
end
assign csr_type = func3[1:0];
VX_decode_if decode_tmp_if();
always @(*) begin
case (csr_type)
2'h1: csr_alu = `ALU_CSR_RW;
2'h2: csr_alu = `ALU_CSR_RS;
2'h3: csr_alu = `ALU_CSR_RC;
default: csr_alu = `ALU_NO;
endcase
assign decode_tmp_if.valid = ifetch_rsp_if.valid;
assign decode_tmp_if.warp_num = ifetch_rsp_if.warp_num;
assign decode_tmp_if.curr_PC = ifetch_rsp_if.curr_PC;
assign decode_tmp_if.next_PC = ifetch_rsp_if.curr_PC + 32'h4;
assign decode_tmp_if.ex_type = is_br ? `EX_BR :
is_lsu ? `EX_LSU :
is_csr ? `EX_CSR :
is_mul ? `EX_MUL :
is_gpu ? `EX_GPU :
(is_rtype || is_itype || is_lui || is_auipc) ? `EX_ALU :
`EX_NOP;
assign decode_tmp_if.instr_op = is_br ? `OP_BITS'(br_op) :
is_lsu ? `OP_BITS'(lsu_op) :
is_csr ? `OP_BITS'(csr_op) :
is_mul ? `OP_BITS'(mul_op) :
is_gpu ? `OP_BITS'(gpu_op) :
(is_rtype || is_itype || is_lui || is_auipc) ? `OP_BITS'(alu_op) :
0;
assign decode_tmp_if.rd = rd;
assign decode_tmp_if.rs1 = is_lui ? `NR_BITS'(0) : rs1;
assign decode_tmp_if.rs2 = rs2;
assign decode_tmp_if.imm = (is_lui || is_auipc) ? {upper_imm, 12'(0)} :
(is_jal || is_jalr || is_jals) ? jalx_offset :
is_csr ? 32'(u_12) :
src2_imm;
assign decode_tmp_if.rs1_is_PC = is_auipc;
assign decode_tmp_if.rs2_is_imm = is_itype || is_lui || is_auipc || is_csr_imm;
assign decode_tmp_if.use_rs1 = (decode_tmp_if.rs1 != 0)
&& (is_jalr || is_btype || is_ltype || is_stype || is_itype || is_rtype || ~is_csr_imm || is_gpu);
assign decode_tmp_if.use_rs2 = (decode_tmp_if.rs2 != 0)
&& (is_btype || is_stype || is_rtype || (is_gpu && (gpu_op == `GPU_BAR || gpu_op == `GPU_WSPAWN)));
assign decode_tmp_if.wb = (rd == 0) ? `WB_NO : // disable writeback to r0
(is_itype || is_rtype || is_lui || is_auipc || is_csr) ? `WB_ALU :
(is_jal || is_jalr || is_jals) ? `WB_JAL :
is_ltype ? `WB_MEM :
`WB_NO;
assign join_if.is_join = is_gpu && (gpu_op == `GPU_JOIN) && in_valid;
assign join_if.warp_num = ifetch_rsp_if.warp_num;
assign wstall_if.wstall = (is_br || is_gpu) && in_valid;
assign wstall_if.warp_num = ifetch_rsp_if.warp_num;
wire stall = ~decode_if.ready && (| decode_if.valid);
VX_generic_register #(
.N(`NUM_THREADS + `NW_BITS + 32 + 32 + `NR_BITS + `NR_BITS + `NR_BITS + 32 + 1 + 1 + 1 + 1 + `EX_BITS + `OP_BITS + `WB_BITS)
) decode_reg (
.clk (clk),
.reset (reset),
.stall (stall),
.flush (0),
.in ({decode_tmp_if.valid, decode_tmp_if.warp_num, decode_tmp_if.curr_PC, decode_tmp_if.next_PC, decode_tmp_if.rd, decode_tmp_if.rs1, decode_tmp_if.rs2, decode_tmp_if.imm, decode_tmp_if.rs1_is_PC, decode_tmp_if.rs2_is_imm, decode_tmp_if.use_rs1, decode_tmp_if.use_rs2, decode_tmp_if.ex_type, decode_tmp_if.instr_op, decode_tmp_if.wb}),
.out ({decode_if.valid, decode_if.warp_num, decode_if.curr_PC, decode_if.next_PC, decode_if.rd, decode_if.rs1, decode_if.rs2, decode_if.imm, decode_if.rs1_is_PC, decode_if.rs2_is_imm, decode_if.use_rs1, decode_if.use_rs2, decode_if.ex_type, decode_if.instr_op, decode_if.wb})
);
assign ifetch_rsp_if.ready = ~stall;
`ifdef DBG_PRINT_PIPELINE
always @(posedge clk) begin
if ((| decode_tmp_if.valid) && ~stall) begin
$write("%t: Core%0d-Decode: warp=%0d, PC=%0h, ex=", $time, CORE_ID, decode_tmp_if.warp_num, decode_tmp_if.curr_PC);
print_ex_type(decode_tmp_if.ex_type);
$write(", op=");
print_instr_op(decode_tmp_if.ex_type, decode_tmp_if.instr_op);
$write(", wb=");
print_wb(decode_tmp_if.wb);
$write(", rd=%0d, rs1=%0d, rs2=%0d, imm=%0h, use_pc=%b, use_imm=%b, use_rs1=%b, use_rs2=%b\n", decode_tmp_if.rd, decode_tmp_if.rs1, decode_tmp_if.rs2, decode_tmp_if.imm, decode_tmp_if.rs1_is_PC, decode_tmp_if.rs2_is_imm, decode_tmp_if.use_rs1, decode_tmp_if.use_rs2);
// trap unsupported instructions
assert(~(~stall && (decode_tmp_if.ex_type == `EX_ALU) && `ALU_OP(decode_tmp_if.instr_op) == `ALU_OTHER));
assert(~(~stall && (decode_tmp_if.ex_type == `EX_BR) && `BR_OP(decode_tmp_if.instr_op) == `BR_OTHER));
assert(~(~stall && (decode_tmp_if.ex_type == `EX_CSR) && `CSR_OP(decode_tmp_if.instr_op) == `CSR_OTHER));
assert(~(~stall && (decode_tmp_if.ex_type == `EX_GPU) && `GPU_OP(decode_tmp_if.instr_op) == `GPU_OTHER));
end
end
`endif
wire[4:0] temp_final_alu;
assign temp_final_alu = is_btype ? ((frE_to_bckE_req_if.branch_type < `BR_LTU) ? `ALU_SUB : `ALU_SUBU) :
is_lui ? `ALU_LUI :
is_auipc ? `ALU_AUIPC :
is_csr ? csr_alu :
(is_stype || is_linst) ? `ALU_ADD :
alu_op;
assign frE_to_bckE_req_if.alu_op = ((func7[0] == 1'b1) && is_rtype) ? mul_alu : temp_final_alu;
endmodule
endmodule