Removed FWD
This commit is contained in:
@@ -6,7 +6,6 @@ module VX_back_end (
|
||||
|
||||
input wire[31:0] csr_decode_csr_data,
|
||||
output wire execute_branch_stall,
|
||||
input wire in_fwd_stall,
|
||||
|
||||
output wire out_mem_delay,
|
||||
output wire out_gpr_stall,
|
||||
@@ -23,13 +22,6 @@ module VX_back_end (
|
||||
VX_dcache_response_inter VX_dcache_rsp,
|
||||
VX_dcache_request_inter VX_dcache_req,
|
||||
|
||||
VX_forward_reqeust_inter VX_fwd_req_de,
|
||||
VX_forward_response_inter VX_fwd_rsp,
|
||||
|
||||
VX_forward_exe_inter VX_fwd_exe,
|
||||
VX_forward_mem_inter VX_fwd_mem,
|
||||
VX_forward_wb_inter VX_fwd_wb,
|
||||
|
||||
|
||||
VX_csr_write_request_inter VX_csr_w_req
|
||||
);
|
||||
@@ -74,13 +66,10 @@ VX_gpr_stage VX_gpr_stage(
|
||||
.clk (clk),
|
||||
.schedule_delay (schedule_delay),
|
||||
.VX_writeback_inter(VX_writeback_inter),
|
||||
.VX_fwd_rsp (VX_fwd_rsp),
|
||||
.in_fwd_stall (in_fwd_stall),
|
||||
.VX_bckE_req (VX_bckE_req),
|
||||
.VX_warp_ctl (VX_warp_ctl),
|
||||
.VX_bckE_req_out (VX_bckE_req_out),
|
||||
.VX_gpr_data (VX_gpr_data),
|
||||
.VX_fwd_req_de (VX_fwd_req_de),
|
||||
.out_gpr_stall (out_gpr_stall)
|
||||
);
|
||||
|
||||
@@ -97,7 +86,6 @@ VX_lsu load_store_unit(
|
||||
// .clk (clk),
|
||||
.VX_lsu_req (VX_lsu_req),
|
||||
.VX_mem_wb (VX_mem_wb),
|
||||
.VX_fwd_mem (VX_fwd_mem),
|
||||
.VX_dcache_rsp(VX_dcache_rsp),
|
||||
.VX_dcache_req(VX_dcache_req),
|
||||
.out_delay (memory_delay)
|
||||
@@ -111,7 +99,6 @@ VX_execute_unit VX_execUnit(
|
||||
.VX_jal_rsp (VX_jal_rsp),
|
||||
.VX_branch_rsp (VX_branch_rsp),
|
||||
|
||||
.VX_fwd_exe (VX_fwd_exe),
|
||||
.in_csr_data (csr_decode_csr_data),
|
||||
.out_csr_address (VX_csr_w_req.csr_address),
|
||||
.out_is_csr (VX_csr_w_req.is_csr),
|
||||
@@ -123,80 +110,7 @@ VX_writeback VX_wb(
|
||||
.VX_mem_wb (VX_mem_wb),
|
||||
.VX_inst_exec_wb (VX_inst_exec_wb),
|
||||
|
||||
.VX_fwd_wb (VX_fwd_wb),
|
||||
.VX_writeback_inter(VX_writeback_inter)
|
||||
);
|
||||
|
||||
// VX_execute vx_execute(
|
||||
// .VX_bckE_req (VX_bckE_req_out),
|
||||
// .VX_gpr_data (VX_gpr_data),
|
||||
// .VX_fwd_exe (VX_fwd_exe),
|
||||
// .in_csr_data (csr_decode_csr_data),
|
||||
|
||||
// .VX_exe_mem_req (VX_exe_mem_req),
|
||||
// .out_csr_address (execute_csr_address),
|
||||
// .out_is_csr (execute_is_csr),
|
||||
// .out_csr_result (execute_csr_result),
|
||||
// .out_jal (execute_jal),
|
||||
// .out_jal_dest (execute_jal_dest),
|
||||
// .out_branch_stall (execute_branch_stall)
|
||||
// );
|
||||
|
||||
|
||||
// assign VX_jal_rsp.jal_warp_num = VX_mem_req.warp_num;
|
||||
|
||||
// VX_e_m_reg vx_e_m_reg(
|
||||
// .clk (clk),
|
||||
// .reset (reset),
|
||||
// .in_csr_address (execute_csr_address),
|
||||
// .in_is_csr (execute_is_csr),
|
||||
// .in_csr_result (execute_csr_result),
|
||||
// .in_jal (execute_jal),
|
||||
// .in_jal_dest (execute_jal_dest),
|
||||
// .in_freeze (total_freeze),
|
||||
// .VX_exe_mem_req (VX_exe_mem_req),
|
||||
|
||||
// .VX_mem_req (VX_mem_req),
|
||||
// .out_csr_address (VX_csr_w_req.csr_address),
|
||||
// .out_is_csr (VX_csr_w_req.is_csr),
|
||||
// .out_csr_result (VX_csr_w_req.csr_result),
|
||||
// .out_jal (VX_jal_rsp.jal),
|
||||
// .out_jal_dest (VX_jal_rsp.jal_dest)
|
||||
// );
|
||||
|
||||
// VX_memory vx_memory(
|
||||
// .VX_mem_req (VX_mem_req),
|
||||
// .VX_mem_wb (VX_mem_wb),
|
||||
// .VX_fwd_mem (VX_fwd_mem),
|
||||
// .out_delay (memory_delay),
|
||||
|
||||
// .VX_branch_rsp (VX_branch_rsp),
|
||||
|
||||
// .VX_dcache_rsp(VX_dcache_rsp),
|
||||
// .VX_dcache_req (VX_dcache_req)
|
||||
// );
|
||||
|
||||
// // VX_m_w_reg vx_m_w_reg(
|
||||
// // .clk (clk),
|
||||
// // .reset (reset),
|
||||
// // .in_freeze (total_freeze),
|
||||
// // .VX_mem_wb (VX_mem_wb),
|
||||
// // .VX_mw_wb (VX_mw_wb)
|
||||
// // );
|
||||
|
||||
// assign VX_mw_wb.alu_result = VX_mem_wb.alu_result;
|
||||
// assign VX_mw_wb.mem_result = VX_mem_wb.mem_result;
|
||||
// assign VX_mw_wb.rd = VX_mem_wb.rd;
|
||||
// assign VX_mw_wb.wb = VX_mem_wb.wb;
|
||||
// assign VX_mw_wb.PC_next = VX_mem_wb.PC_next;
|
||||
// assign VX_mw_wb.valid = VX_mem_wb.valid;
|
||||
// assign VX_mw_wb.warp_num = VX_mem_wb.warp_num;
|
||||
|
||||
|
||||
// VX_writeback vx_writeback(
|
||||
// .VX_mw_wb (VX_mw_wb),
|
||||
// .VX_fwd_wb (VX_fwd_wb),
|
||||
// .VX_writeback_inter(VX_writeback_inter)
|
||||
// );
|
||||
|
||||
endmodule
|
||||
@@ -5,18 +5,6 @@ module VX_decode(
|
||||
// Fetch Inputs
|
||||
VX_inst_meta_inter fd_inst_meta_de,
|
||||
|
||||
// WriteBack inputs
|
||||
// VX_wb_inter VX_writeback_inter,
|
||||
|
||||
|
||||
// Fwd Request
|
||||
// VX_forward_reqeust_inter VX_fwd_req_de,
|
||||
|
||||
// FORWARDING INPUTS
|
||||
// VX_forward_response_inter VX_fwd_rsp,
|
||||
|
||||
// input wire[`NW_M1:0] in_which_wspawn,
|
||||
|
||||
// Outputs
|
||||
VX_frE_to_bckE_req_inter VX_frE_to_bckE_req,
|
||||
output reg out_gpr_stall,
|
||||
@@ -91,50 +79,6 @@ module VX_decode(
|
||||
reg[4:0] mul_alu;
|
||||
|
||||
|
||||
|
||||
// assign VX_fwd_req_de.src1 = VX_frE_to_bckE_req.rs1;
|
||||
// assign VX_fwd_req_de.src2 = VX_frE_to_bckE_req.rs2;
|
||||
// assign VX_fwd_req_de.warp_num = VX_frE_to_bckE_req.warp_num;
|
||||
|
||||
|
||||
// VX_gpr_read_inter VX_gpr_read();
|
||||
// assign VX_gpr_read.rs1 = VX_frE_to_bckE_req.rs1;
|
||||
// assign VX_gpr_read.rs2 = VX_frE_to_bckE_req.rs2;
|
||||
// assign VX_gpr_read.warp_num = VX_frE_to_bckE_req.warp_num;
|
||||
|
||||
// VX_gpr_jal_inter VX_gpr_jal();
|
||||
// assign VX_gpr_jal.is_jal = is_jal;
|
||||
// assign VX_gpr_jal.curr_PC = in_curr_PC;
|
||||
|
||||
|
||||
// VX_gpr_clone_inter VX_gpr_clone();
|
||||
// assign VX_gpr_clone.is_clone = is_clone;
|
||||
// assign VX_gpr_clone.warp_num = VX_frE_to_bckE_req.warp_num;
|
||||
|
||||
|
||||
// VX_gpr_wspawn_inter VX_gpr_wspawn();
|
||||
// assign VX_gpr_wspawn.is_wspawn = is_wspawn;
|
||||
// assign VX_gpr_wspawn.which_wspawn = in_which_wspawn;
|
||||
// // assign VX_gpr_wspawn.warp_num = VX_frE_to_bckE_req.warp_num;
|
||||
|
||||
// VX_gpr_wrapper vx_grp_wrapper(
|
||||
// .clk (clk),
|
||||
// .VX_writeback_inter(VX_writeback_inter),
|
||||
// .VX_fwd_rsp (VX_fwd_rsp),
|
||||
// .VX_gpr_read (VX_gpr_read),
|
||||
// .VX_gpr_jal (VX_gpr_jal),
|
||||
// .VX_gpr_clone (VX_gpr_clone),
|
||||
// .VX_gpr_wspawn (VX_gpr_wspawn),
|
||||
|
||||
// .out_a_reg_data (VX_frE_to_bckE_req.a_reg_data),
|
||||
// .out_b_reg_data (VX_frE_to_bckE_req.b_reg_data),
|
||||
// .out_gpr_stall(out_gpr_stall)
|
||||
// );
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
assign VX_frE_to_bckE_req.valid = fd_inst_meta_de.valid;
|
||||
|
||||
assign VX_frE_to_bckE_req.warp_num = in_warp_num;
|
||||
|
||||
@@ -11,9 +11,6 @@ module VX_execute_unit (
|
||||
VX_jal_response_inter VX_jal_rsp,
|
||||
// Branch Response
|
||||
VX_branch_response_inter VX_branch_rsp,
|
||||
// Forward data
|
||||
VX_forward_exe_inter VX_fwd_exe,
|
||||
|
||||
|
||||
input wire[31:0] in_csr_data,
|
||||
output wire[11:0] out_csr_address,
|
||||
@@ -24,12 +21,6 @@ module VX_execute_unit (
|
||||
|
||||
);
|
||||
|
||||
assign VX_fwd_exe.dest = 0;
|
||||
assign VX_fwd_exe.wb = 0;
|
||||
assign VX_fwd_exe.alu_result = 0;
|
||||
assign VX_fwd_exe.PC_next = 0;
|
||||
assign VX_fwd_exe.warp_num = 0;
|
||||
|
||||
|
||||
|
||||
wire[`NT_M1:0][31:0] in_a_reg_data;
|
||||
|
||||
@@ -5,7 +5,6 @@ module VX_fetch (
|
||||
input wire clk,
|
||||
input wire in_memory_delay,
|
||||
input wire in_branch_stall,
|
||||
input wire in_fwd_stall,
|
||||
input wire in_branch_stall_exe,
|
||||
input wire in_gpr_stall,
|
||||
input wire schedule_delay,
|
||||
@@ -29,7 +28,7 @@ module VX_fetch (
|
||||
wire warp_stall;
|
||||
|
||||
|
||||
assign pipe_stall = in_gpr_stall || in_fwd_stall || in_freeze || schedule_delay;
|
||||
assign pipe_stall = in_gpr_stall || in_freeze || schedule_delay;
|
||||
|
||||
assign warp_stall = in_branch_stall || (in_branch_stall_exe && 0);
|
||||
|
||||
|
||||
@@ -1,179 +0,0 @@
|
||||
|
||||
`include "VX_define.v"
|
||||
|
||||
module VX_forwarding (
|
||||
// INFO FROM DECODE
|
||||
VX_forward_reqeust_inter VX_fwd_req_de,
|
||||
VX_forward_exe_inter VX_fwd_exe,
|
||||
VX_forward_mem_inter VX_fwd_mem,
|
||||
VX_forward_wb_inter VX_fwd_wb,
|
||||
|
||||
VX_forward_response_inter VX_fwd_rsp,
|
||||
output wire out_fwd_stall
|
||||
);
|
||||
|
||||
|
||||
wire[4:0] in_decode_src1 = VX_fwd_req_de.src1;
|
||||
wire[4:0] in_decode_src2 = VX_fwd_req_de.src2;
|
||||
wire[`NW_M1:0] in_decode_warp_num = VX_fwd_req_de.warp_num;
|
||||
|
||||
wire[4:0] in_execute_dest = VX_fwd_exe.dest;
|
||||
wire[1:0] in_execute_wb = VX_fwd_exe.wb;
|
||||
wire[`NT_M1:0][31:0] in_execute_alu_result = VX_fwd_exe.alu_result;
|
||||
wire[31:0] in_execute_PC_next = VX_fwd_exe.PC_next;
|
||||
wire[`NW_M1:0] in_execute_warp_num = VX_fwd_exe.warp_num;
|
||||
|
||||
wire[4:0] in_memory_dest = VX_fwd_mem.dest;
|
||||
wire[1:0] in_memory_wb = VX_fwd_mem.wb;
|
||||
wire[`NT_M1:0][31:0] in_memory_alu_result = VX_fwd_mem.alu_result;
|
||||
wire[`NT_M1:0][31:0] in_memory_mem_data = VX_fwd_mem.mem_data;
|
||||
wire[31:0] in_memory_PC_next = VX_fwd_mem.PC_next;
|
||||
wire[`NW_M1:0] in_memory_warp_num = VX_fwd_mem.warp_num;
|
||||
|
||||
wire[4:0] in_writeback_dest = VX_fwd_wb.dest;
|
||||
wire[1:0] in_writeback_wb = VX_fwd_wb.wb;
|
||||
wire[`NT_M1:0][31:0] in_writeback_alu_result = VX_fwd_wb.alu_result;
|
||||
wire[`NT_M1:0][31:0] in_writeback_mem_data = VX_fwd_wb.mem_data;
|
||||
wire[31:0] in_writeback_PC_next = VX_fwd_wb.PC_next;
|
||||
wire[`NW_M1:0] in_writeback_warp_num = VX_fwd_wb.warp_num;
|
||||
|
||||
|
||||
wire out_src1_fwd;
|
||||
wire out_src2_fwd;
|
||||
wire[`NT_M1:0][31:0] out_src1_fwd_data;
|
||||
wire[`NT_M1:0][31:0] out_src2_fwd_data;
|
||||
|
||||
|
||||
assign VX_fwd_rsp.src1_fwd = out_src1_fwd;
|
||||
assign VX_fwd_rsp.src2_fwd = out_src2_fwd;
|
||||
assign VX_fwd_rsp.src1_fwd_data = out_src1_fwd_data;
|
||||
assign VX_fwd_rsp.src2_fwd_data = out_src2_fwd_data;
|
||||
|
||||
|
||||
|
||||
|
||||
wire exe_mem_read;
|
||||
wire mem_mem_read;
|
||||
wire wb_mem_read ;
|
||||
wire exe_jal;
|
||||
wire mem_jal;
|
||||
wire wb_jal ;
|
||||
wire src1_exe_fwd;
|
||||
wire src1_mem_fwd;
|
||||
wire src1_wb_fwd;
|
||||
wire src2_exe_fwd;
|
||||
wire src2_mem_fwd;
|
||||
wire src2_wb_fwd;
|
||||
|
||||
wire[`NT_M1:0][31:0] use_execute_PC_next;
|
||||
wire[`NT_M1:0][31:0] use_memory_PC_next;
|
||||
wire[`NT_M1:0][31:0] use_writeback_PC_next;
|
||||
|
||||
|
||||
genvar index;
|
||||
generate
|
||||
for (index=0; index < `NT; index=index+1)
|
||||
begin: gen_code_label
|
||||
assign use_execute_PC_next[index] = in_execute_PC_next;
|
||||
assign use_memory_PC_next[index] = in_memory_PC_next;
|
||||
assign use_writeback_PC_next[index] = in_writeback_PC_next;
|
||||
end
|
||||
endgenerate
|
||||
|
||||
|
||||
assign exe_mem_read = (in_execute_wb == `WB_MEM);
|
||||
assign mem_mem_read = (in_memory_wb == `WB_MEM);
|
||||
assign wb_mem_read = (in_writeback_wb == `WB_MEM);
|
||||
|
||||
assign exe_jal = (in_execute_wb == `WB_JAL);
|
||||
assign mem_jal = (in_memory_wb == `WB_JAL);
|
||||
assign wb_jal = (in_writeback_wb == `WB_JAL);
|
||||
|
||||
|
||||
|
||||
// SRC1
|
||||
assign src1_exe_fwd = ((in_decode_src1 == in_execute_dest) &&
|
||||
(in_decode_src1 != `ZERO_REG) &&
|
||||
(in_execute_wb != `NO_WB)) &&
|
||||
(in_decode_warp_num == in_execute_warp_num);
|
||||
|
||||
assign src1_mem_fwd = ((in_decode_src1 == in_memory_dest) &&
|
||||
(in_decode_src1 != `ZERO_REG) &&
|
||||
(in_memory_wb != `NO_WB) &&
|
||||
(!src1_exe_fwd)) &&
|
||||
(in_decode_warp_num == in_memory_warp_num);
|
||||
|
||||
assign src1_wb_fwd = ((in_decode_src1 == in_writeback_dest) &&
|
||||
(in_decode_src1 != `ZERO_REG) &&
|
||||
(in_writeback_wb != `NO_WB) &&
|
||||
(in_writeback_warp_num == in_decode_warp_num) &&
|
||||
(!src1_exe_fwd) &&
|
||||
(!src1_mem_fwd));
|
||||
|
||||
|
||||
// assign out_src1_fwd = src1_exe_fwd || src1_mem_fwd || (src1_wb_fwd && 0);
|
||||
assign out_src1_fwd = 0;
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
// SRC2
|
||||
assign src2_exe_fwd = ((in_decode_src2 == in_execute_dest) &&
|
||||
(in_decode_src2 != `ZERO_REG) &&
|
||||
(in_execute_wb != `NO_WB)) &&
|
||||
(in_decode_warp_num == in_execute_warp_num);
|
||||
|
||||
assign src2_mem_fwd = ((in_decode_src2 == in_memory_dest) &&
|
||||
(in_decode_src2 != `ZERO_REG) &&
|
||||
(in_memory_wb != `NO_WB) &&
|
||||
(!src2_exe_fwd)) &&
|
||||
(in_decode_warp_num == in_memory_warp_num);
|
||||
|
||||
assign src2_wb_fwd = ((in_decode_src2 == in_writeback_dest) &&
|
||||
(in_decode_src2 != `ZERO_REG) &&
|
||||
(in_writeback_wb != `NO_WB) &&
|
||||
(!src2_exe_fwd) &&
|
||||
(!src2_mem_fwd)) &&
|
||||
(in_writeback_warp_num == in_decode_warp_num);
|
||||
|
||||
|
||||
// assign out_src2_fwd = src2_exe_fwd || src2_mem_fwd || (src2_wb_fwd && 0);
|
||||
assign out_src2_fwd = 0;
|
||||
|
||||
|
||||
|
||||
|
||||
// wire exe_mem_read_stall = ((src1_exe_fwd || src2_exe_fwd) && exe_mem_read) ? `STALL : `NO_STALL;
|
||||
// wire mem_mem_read_stall = ((src1_mem_fwd || src2_mem_fwd) && mem_mem_read) ? `STALL : `NO_STALL;
|
||||
wire exe_mem_read_stall = `NO_STALL;
|
||||
wire mem_mem_read_stall = `NO_STALL;
|
||||
|
||||
// assign out_fwd_stall = exe_mem_read_stall || mem_mem_read_stall;
|
||||
assign out_fwd_stall = 0;
|
||||
|
||||
// always @(*) begin
|
||||
// if (out_fwd_stall) $display("FWD STALL");
|
||||
// end
|
||||
|
||||
assign out_src1_fwd_data = src1_exe_fwd ? ((exe_jal) ? use_execute_PC_next : in_execute_alu_result) :
|
||||
(src1_mem_fwd) ? ((mem_jal) ? use_memory_PC_next : (mem_mem_read ? in_memory_mem_data : in_memory_alu_result)) :
|
||||
( src1_wb_fwd ) ? (wb_jal ? use_writeback_PC_next : (wb_mem_read ? in_writeback_mem_data : in_writeback_alu_result)) :
|
||||
in_execute_alu_result; // last one should be deadbeef
|
||||
|
||||
assign out_src2_fwd_data = src2_exe_fwd ? ((exe_jal) ? use_execute_PC_next : in_execute_alu_result) :
|
||||
(src2_mem_fwd) ? ((mem_jal) ? use_memory_PC_next : (mem_mem_read ? in_memory_mem_data : in_memory_alu_result)) :
|
||||
( src2_wb_fwd ) ? (wb_jal ? use_writeback_PC_next : (wb_mem_read ? in_writeback_mem_data : in_writeback_alu_result)) :
|
||||
in_execute_alu_result; // last one should be deadbeef
|
||||
|
||||
|
||||
|
||||
|
||||
endmodule // VX_forwarding
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
@@ -4,7 +4,6 @@ module VX_front_end (
|
||||
input wire clk,
|
||||
input wire reset,
|
||||
|
||||
input wire forwarding_fwd_stall,
|
||||
input wire memory_delay,
|
||||
|
||||
input wire execute_branch_stall,
|
||||
@@ -48,7 +47,6 @@ VX_fetch vx_fetch(
|
||||
.clk (clk),
|
||||
.in_memory_delay (memory_delay),
|
||||
.in_branch_stall (decode_branch_stall),
|
||||
.in_fwd_stall (forwarding_fwd_stall),
|
||||
.schedule_delay (schedule_delay),
|
||||
.in_branch_stall_exe(execute_branch_stall),
|
||||
.in_gpr_stall (decode_gpr_stall),
|
||||
@@ -66,7 +64,6 @@ VX_fetch vx_fetch(
|
||||
VX_f_d_reg vx_f_d_reg(
|
||||
.clk (clk),
|
||||
.reset (reset),
|
||||
.in_fwd_stall (forwarding_fwd_stall),
|
||||
.in_freeze (total_freeze),
|
||||
.in_gpr_stall (decode_gpr_stall),
|
||||
.fe_inst_meta_fd(fe_inst_meta_fd),
|
||||
@@ -82,16 +79,11 @@ VX_decode vx_decode(
|
||||
.out_ebreak (fetch_ebreak)
|
||||
);
|
||||
|
||||
wire special_what = total_freeze || forwarding_fwd_stall;
|
||||
|
||||
wire temp_fwd_stall = 0;
|
||||
|
||||
VX_d_e_reg vx_d_e_reg(
|
||||
.clk (clk),
|
||||
.reset (reset),
|
||||
.in_fwd_stall (temp_fwd_stall),
|
||||
.in_branch_stall(execute_branch_stall),
|
||||
.in_freeze (special_what),
|
||||
.in_freeze (total_freeze),
|
||||
.in_gpr_stall (decode_gpr_stall),
|
||||
.VX_frE_to_bckE_req(VX_frE_to_bckE_req),
|
||||
.VX_bckE_req (VX_bckE_req)
|
||||
|
||||
@@ -1,21 +1,16 @@
|
||||
module VX_gpr_stage (
|
||||
input wire clk,
|
||||
input wire in_fwd_stall,
|
||||
input wire schedule_delay,
|
||||
// inputs
|
||||
// Instruction Information
|
||||
VX_frE_to_bckE_req_inter VX_bckE_req,
|
||||
// WriteBack inputs
|
||||
VX_wb_inter VX_writeback_inter,
|
||||
// FORWARDING INPUTS
|
||||
VX_forward_response_inter VX_fwd_rsp,
|
||||
|
||||
|
||||
|
||||
|
||||
// Outputs
|
||||
// Fwd Request
|
||||
VX_forward_reqeust_inter VX_fwd_req_de,
|
||||
// Warp Control
|
||||
VX_warp_ctl_inter VX_warp_ctl,
|
||||
// Original Request 1 cycle later
|
||||
@@ -32,11 +27,6 @@ module VX_gpr_stage (
|
||||
|
||||
wire jalQual = VX_bckE_req.jalQual;
|
||||
|
||||
|
||||
assign VX_fwd_req_de.src1 = VX_bckE_req.rs1;
|
||||
assign VX_fwd_req_de.src2 = VX_bckE_req.rs2;
|
||||
assign VX_fwd_req_de.warp_num = VX_bckE_req.warp_num;
|
||||
|
||||
VX_gpr_read_inter VX_gpr_read();
|
||||
assign VX_gpr_read.rs1 = VX_bckE_req.rs1;
|
||||
assign VX_gpr_read.rs2 = VX_bckE_req.rs2;
|
||||
@@ -53,7 +43,6 @@ module VX_gpr_stage (
|
||||
VX_gpr_wrapper vx_grp_wrapper(
|
||||
.clk (clk),
|
||||
.VX_writeback_inter(VX_writeback_inter),
|
||||
.VX_fwd_rsp (VX_fwd_rsp),
|
||||
.VX_gpr_read (VX_gpr_read),
|
||||
.VX_gpr_jal (VX_gpr_jal),
|
||||
|
||||
@@ -77,13 +66,11 @@ module VX_gpr_stage (
|
||||
.out ({VX_gpr_data.a_reg_data, VX_gpr_data.b_reg_data})
|
||||
);
|
||||
|
||||
wire stall = in_fwd_stall || schedule_delay;
|
||||
|
||||
VX_d_e_reg gpr_stage_reg(
|
||||
.clk (clk),
|
||||
.reset (zero_temp),
|
||||
.in_fwd_stall (stall),
|
||||
.in_branch_stall (zero_temp),
|
||||
.in_branch_stall (schedule_delay),
|
||||
.in_freeze (zero_temp),
|
||||
.in_gpr_stall (out_gpr_stall),
|
||||
.VX_frE_to_bckE_req(VX_bckE_req),
|
||||
|
||||
@@ -3,9 +3,7 @@
|
||||
module VX_gpr_wrapper (
|
||||
input wire clk,
|
||||
VX_gpr_read_inter VX_gpr_read,
|
||||
VX_wb_inter VX_writeback_inter,
|
||||
VX_forward_response_inter VX_fwd_rsp,
|
||||
|
||||
VX_wb_inter VX_writeback_inter,
|
||||
VX_gpr_jal_inter VX_gpr_jal,
|
||||
|
||||
output wire[`NT_M1:0][31:0] out_a_reg_data,
|
||||
@@ -22,8 +20,8 @@ module VX_gpr_wrapper (
|
||||
for (index = 0; index <= `NT_M1; index = index + 1) assign jal_data[index] = VX_gpr_jal.curr_PC;
|
||||
|
||||
|
||||
assign out_a_reg_data = (VX_gpr_jal.is_jal ? jal_data : (VX_fwd_rsp.src1_fwd ? VX_fwd_rsp.src1_fwd_data : temp_a_reg_data[VX_gpr_read.warp_num]));
|
||||
assign out_b_reg_data = (VX_fwd_rsp.src2_fwd ? VX_fwd_rsp.src2_fwd_data : temp_b_reg_data[VX_gpr_read.warp_num]);
|
||||
assign out_a_reg_data = (VX_gpr_jal.is_jal ? jal_data : (temp_a_reg_data[VX_gpr_read.warp_num]));
|
||||
assign out_b_reg_data = (temp_b_reg_data[VX_gpr_read.warp_num]);
|
||||
|
||||
genvar warp_index;
|
||||
generate
|
||||
|
||||
12
rtl/VX_lsu.v
12
rtl/VX_lsu.v
@@ -9,10 +9,6 @@ module VX_lsu (
|
||||
// Write back to GPR
|
||||
VX_inst_mem_wb_inter VX_mem_wb,
|
||||
|
||||
// FWD info
|
||||
VX_forward_mem_inter VX_fwd_mem,
|
||||
|
||||
|
||||
VX_dcache_response_inter VX_dcache_rsp,
|
||||
VX_dcache_request_inter VX_dcache_req,
|
||||
output wire out_delay
|
||||
@@ -64,14 +60,6 @@ module VX_lsu (
|
||||
// .out ({VX_mem_wb.loaded_data , VX_mem_wb.rd , VX_mem_wb.wb , VX_mem_wb.wb_valid , VX_mem_wb.wb_warp_num })
|
||||
// );
|
||||
|
||||
// Delete
|
||||
assign VX_fwd_mem.dest = 0;
|
||||
assign VX_fwd_mem.wb = 0;
|
||||
assign VX_fwd_mem.alu_result = 0;
|
||||
assign VX_fwd_mem.mem_data = 0;
|
||||
assign VX_fwd_mem.PC_next = 0;
|
||||
assign VX_fwd_mem.warp_num = 0;
|
||||
|
||||
|
||||
endmodule // Memory
|
||||
|
||||
|
||||
@@ -8,18 +8,10 @@ module VX_writeback (
|
||||
// EXEC Unit WB info
|
||||
VX_inst_exec_wb_inter VX_inst_exec_wb,
|
||||
|
||||
VX_forward_wb_inter VX_fwd_wb,
|
||||
// Actual WB to GPR
|
||||
VX_wb_inter VX_writeback_inter
|
||||
);
|
||||
|
||||
assign VX_fwd_wb.dest = 0;
|
||||
assign VX_fwd_wb.wb = 0;
|
||||
assign VX_fwd_wb.alu_result = 0;
|
||||
assign VX_fwd_wb.mem_data = 0;
|
||||
assign VX_fwd_wb.PC_next = 0;
|
||||
assign VX_fwd_wb.warp_num = 0;
|
||||
|
||||
|
||||
|
||||
wire exec_wb = (VX_inst_exec_wb.wb != 0) && (|VX_inst_exec_wb.wb_valid);
|
||||
|
||||
26
rtl/Vortex.v
26
rtl/Vortex.v
@@ -53,16 +53,6 @@ VX_jal_response_inter VX_jal_rsp(); // Jump resolution to Fetch
|
||||
wire execute_branch_stall;
|
||||
wire memory_delay;
|
||||
|
||||
// Forwarding Buses
|
||||
VX_forward_reqeust_inter VX_fwd_req_de(); // Forward request
|
||||
VX_forward_response_inter VX_fwd_rsp(); // Forward Response
|
||||
VX_forward_exe_inter VX_fwd_exe(); // Data available in EXE
|
||||
VX_forward_mem_inter VX_fwd_mem(); // Data available in MEM
|
||||
VX_forward_wb_inter VX_fwd_wb(); // Data available in WB
|
||||
wire forwarding_fwd_stall;
|
||||
|
||||
|
||||
|
||||
// CSR Buses
|
||||
VX_csr_write_request_inter VX_csr_w_req();
|
||||
wire[31:0] csr_decode_csr_data;
|
||||
@@ -80,7 +70,6 @@ VX_front_end vx_front_end(
|
||||
.clk (clk),
|
||||
.reset (reset),
|
||||
.VX_warp_ctl (VX_warp_ctl),
|
||||
.forwarding_fwd_stall(forwarding_fwd_stall),
|
||||
.execute_branch_stall(execute_branch_stall),
|
||||
.VX_bckE_req (VX_bckE_req),
|
||||
.decode_csr_address (decode_csr_address),
|
||||
@@ -107,35 +96,20 @@ VX_back_end vx_back_end(
|
||||
.reset (reset),
|
||||
.schedule_delay (schedule_delay),
|
||||
.fetch_delay (fetch_delay),
|
||||
.in_fwd_stall (forwarding_fwd_stall),
|
||||
.VX_fwd_req_de (VX_fwd_req_de),
|
||||
.VX_fwd_rsp (VX_fwd_rsp),
|
||||
.VX_warp_ctl (VX_warp_ctl),
|
||||
.VX_bckE_req (VX_bckE_req),
|
||||
.VX_fwd_exe (VX_fwd_exe),
|
||||
.csr_decode_csr_data (csr_decode_csr_data),
|
||||
.execute_branch_stall(execute_branch_stall),
|
||||
.VX_jal_rsp (VX_jal_rsp),
|
||||
.VX_branch_rsp (VX_branch_rsp),
|
||||
.VX_dcache_rsp (VX_dcache_rsp),
|
||||
.VX_dcache_req (VX_dcache_req),
|
||||
.VX_fwd_mem (VX_fwd_mem),
|
||||
.VX_fwd_wb (VX_fwd_wb),
|
||||
.VX_csr_w_req (VX_csr_w_req),
|
||||
.VX_writeback_inter (VX_writeback_inter),
|
||||
.out_mem_delay (memory_delay),
|
||||
.out_gpr_stall (out_gpr_stall)
|
||||
);
|
||||
|
||||
VX_forwarding vx_forwarding(
|
||||
.VX_fwd_req_de(VX_fwd_req_de),
|
||||
.VX_fwd_exe (VX_fwd_exe),
|
||||
.VX_fwd_mem (VX_fwd_mem),
|
||||
.VX_fwd_wb (VX_fwd_wb),
|
||||
.VX_fwd_rsp (VX_fwd_rsp),
|
||||
.out_fwd_stall(forwarding_fwd_stall)
|
||||
);
|
||||
|
||||
VX_csr_handler vx_csr_handler(
|
||||
.clk (clk),
|
||||
.in_decode_csr_address(decode_csr_address),
|
||||
|
||||
@@ -1,14 +0,0 @@
|
||||
|
||||
`include "../VX_define.v"
|
||||
|
||||
`ifndef VX_FWD_CSR_RSP
|
||||
|
||||
`define VX_FWD_CSR_RSP
|
||||
|
||||
interface VX_forward_csr_response_inter ();
|
||||
wire csr_fwd;
|
||||
wire[31:0] csr_fwd_data;
|
||||
endinterface
|
||||
|
||||
|
||||
`endif
|
||||
@@ -1,19 +0,0 @@
|
||||
|
||||
`include "../VX_define.v"
|
||||
|
||||
`ifndef VX_FWD_EXE
|
||||
|
||||
`define VX_FWD_EXE
|
||||
|
||||
interface VX_forward_exe_inter ();
|
||||
|
||||
wire[4:0] dest;
|
||||
wire[1:0] wb;
|
||||
wire[`NT_M1:0][31:0] alu_result;
|
||||
wire[31:0] PC_next;
|
||||
wire[`NW_M1:0] warp_num;
|
||||
|
||||
endinterface
|
||||
|
||||
|
||||
`endif
|
||||
@@ -1,20 +0,0 @@
|
||||
|
||||
`include "../VX_define.v"
|
||||
|
||||
`ifndef VX_FWD_MEM
|
||||
|
||||
`define VX_FWD_MEM
|
||||
|
||||
interface VX_forward_mem_inter ();
|
||||
|
||||
wire[4:0] dest;
|
||||
wire[1:0] wb;
|
||||
wire[`NT_M1:0][31:0] alu_result;
|
||||
wire[`NT_M1:0][31:0] mem_data;
|
||||
wire[31:0] PC_next;
|
||||
wire[`NW_M1:0] warp_num;
|
||||
|
||||
endinterface
|
||||
|
||||
|
||||
`endif
|
||||
@@ -1,18 +0,0 @@
|
||||
|
||||
`include "../VX_define.v"
|
||||
|
||||
`ifndef VX_FWD_REQ
|
||||
|
||||
`define VX_FWD_REQ
|
||||
|
||||
interface VX_forward_reqeust_inter ();
|
||||
|
||||
wire[4:0] src1;
|
||||
wire[4:0] src2;
|
||||
wire[`NW_M1:0] warp_num;
|
||||
|
||||
|
||||
endinterface
|
||||
|
||||
|
||||
`endif
|
||||
@@ -1,18 +0,0 @@
|
||||
|
||||
`include "../VX_define.v"
|
||||
|
||||
`ifndef VX_FWD_RSP
|
||||
|
||||
`define VX_FWD_RSP
|
||||
|
||||
interface VX_forward_response_inter ();
|
||||
|
||||
wire src1_fwd;
|
||||
wire src2_fwd;
|
||||
wire[`NT_M1:0][31:0] src1_fwd_data;
|
||||
wire[`NT_M1:0][31:0] src2_fwd_data;
|
||||
|
||||
endinterface
|
||||
|
||||
|
||||
`endif
|
||||
@@ -1,21 +0,0 @@
|
||||
|
||||
`include "../VX_define.v"
|
||||
|
||||
`ifndef VX_FWD_WB
|
||||
|
||||
`define VX_FWD_WB
|
||||
|
||||
interface VX_forward_wb_inter ();
|
||||
|
||||
wire[4:0] dest;
|
||||
wire[1:0] wb;
|
||||
wire[`NT_M1:0][31:0] alu_result;
|
||||
wire[`NT_M1:0][31:0] mem_data;
|
||||
wire[31:0] PC_next;
|
||||
wire[`NW_M1:0] warp_num;
|
||||
|
||||
|
||||
endinterface
|
||||
|
||||
|
||||
`endif
|
||||
@@ -5,7 +5,6 @@
|
||||
module VX_d_e_reg (
|
||||
input wire clk,
|
||||
input wire reset,
|
||||
input wire in_fwd_stall,
|
||||
input wire in_branch_stall,
|
||||
input wire in_freeze,
|
||||
input wire in_gpr_stall,
|
||||
@@ -17,7 +16,7 @@ module VX_d_e_reg (
|
||||
|
||||
|
||||
wire stall = in_freeze;
|
||||
wire flush = (in_fwd_stall == `STALL) || (in_branch_stall == `STALL) || (in_gpr_stall == `STALL);
|
||||
wire flush = (in_branch_stall == `STALL) || (in_gpr_stall == `STALL);
|
||||
|
||||
|
||||
VX_generic_register #(.N(237)) d_e_reg
|
||||
|
||||
@@ -3,7 +3,6 @@
|
||||
module VX_f_d_reg (
|
||||
input wire clk,
|
||||
input wire reset,
|
||||
input wire in_fwd_stall,
|
||||
input wire in_freeze,
|
||||
input wire in_gpr_stall,
|
||||
|
||||
@@ -13,7 +12,7 @@ module VX_f_d_reg (
|
||||
);
|
||||
|
||||
wire flush = 1'b0;
|
||||
wire stall = in_fwd_stall == 1'b1 || in_freeze == 1'b1 || in_gpr_stall;
|
||||
wire stall = in_freeze == 1'b1 || in_gpr_stall;
|
||||
|
||||
|
||||
|
||||
|
||||
@@ -3,7 +3,7 @@ set link_library [concat * sc12mc_cln28hpm_base_ulvt_c35_ssg_typical_max_0p81v_
|
||||
set symbol_library {}
|
||||
set target_library [concat sc12mc_cln28hpm_base_ulvt_c35_ssg_typical_max_0p81v_m40c.db]
|
||||
|
||||
set verilog_files [ list VX_inst_exec_wb_inter.v VX_lsu.v VX_execute_unit.v VX_lsu_addr_gen.v VX_inst_multiplex.v VX_exec_unit_req_inter.v VX_lsu_req_inter.v VX_alu.v VX_back_end.v VX_gpr_stage.v VX_gpr_data_inter.v VX_csr_handler.v VX_decode.v VX_define.v VX_scheduler.v VX_fetch.v VX_forwarding.v VX_front_end.v VX_generic_register.v VX_gpr.v VX_gpr_wrapper.v VX_one_counter.v VX_priority_encoder.v VX_warp.v VX_warp_scheduler.v VX_writeback.v Vortex.v byte_enabled_simple_dual_port_ram.v VX_branch_response_inter.v VX_csr_write_request_inter.v VX_dcache_request_inter.v VX_dcache_response_inter.v VX_forward_csr_response_inter.v VX_forward_exe_inter.v VX_forward_mem_inter.v VX_forward_reqeust_inter.v VX_forward_response_inter.v VX_forward_wb_inter.v VX_frE_to_bckE_req_inter.v VX_gpr_clone_inter.v VX_gpr_jal_inter.v VX_gpr_read_inter.v VX_gpr_wspawn_inter.v VX_icache_request_inter.v VX_icache_response_inter.v VX_inst_mem_wb_inter.v VX_inst_meta_inter.v VX_jal_response_inter.v VX_mem_req_inter.v VX_mw_wb_inter.v VX_warp_ctl_inter.v VX_wb_inter.v VX_d_e_reg.v VX_f_d_reg.v \
|
||||
set verilog_files [ list VX_inst_exec_wb_inter.v VX_lsu.v VX_execute_unit.v VX_lsu_addr_gen.v VX_inst_multiplex.v VX_exec_unit_req_inter.v VX_lsu_req_inter.v VX_alu.v VX_back_end.v VX_gpr_stage.v VX_gpr_data_inter.v VX_csr_handler.v VX_decode.v VX_define.v VX_scheduler.v VX_fetch.v VX_front_end.v VX_generic_register.v VX_gpr.v VX_gpr_wrapper.v VX_one_counter.v VX_priority_encoder.v VX_warp.v VX_warp_scheduler.v VX_writeback.v Vortex.v byte_enabled_simple_dual_port_ram.v VX_branch_response_inter.v VX_csr_write_request_inter.v VX_dcache_request_inter.v VX_dcache_response_inter.v VX_frE_to_bckE_req_inter.v VX_gpr_clone_inter.v VX_gpr_jal_inter.v VX_gpr_read_inter.v VX_gpr_wspawn_inter.v VX_icache_request_inter.v VX_icache_response_inter.v VX_inst_mem_wb_inter.v VX_inst_meta_inter.v VX_jal_response_inter.v VX_mem_req_inter.v VX_mw_wb_inter.v VX_warp_ctl_inter.v VX_wb_inter.v VX_d_e_reg.v VX_f_d_reg.v \
|
||||
]
|
||||
|
||||
analyze -format sverilog $verilog_files
|
||||
|
||||
Reference in New Issue
Block a user