Fixed ASIC GPR warp number delay

This commit is contained in:
felsabbagh3
2019-11-03 15:56:18 -05:00
parent 95d8a251db
commit a39979a844
9 changed files with 473 additions and 437 deletions

View File

@@ -51,7 +51,8 @@ module VX_gpr (
wire cenb = !going_to_write;
// wire cenb = !going_to_write;
wire cenb = 0;
// wire cena_1 = (VX_gpr_read.rs1 == 0);
// wire cena_2 = (VX_gpr_read.rs2 == 0);
@@ -68,8 +69,8 @@ module VX_gpr (
begin
for (curr_bit = 0; curr_bit < 32; curr_bit=curr_bit+1)
begin
assign out_a_reg_data[thread][curr_bit] = (temp_a[thread][curr_bit] === 1'dx) ? 1'b0 : temp_a[thread][curr_bit];
assign out_b_reg_data[thread][curr_bit] = (temp_b[thread][curr_bit] === 1'dx) ? 1'b0 : temp_b[thread][curr_bit];
assign out_a_reg_data[thread][curr_bit] = ((temp_a[thread][curr_bit] === 1'dx) || cena_1 )? 1'b0 : temp_a[thread][curr_bit];
assign out_b_reg_data[thread][curr_bit] = ((temp_b[thread][curr_bit] === 1'dx) || cena_2) ? 1'b0 : temp_b[thread][curr_bit];
end
end

View File

@@ -22,8 +22,27 @@ module VX_gpr_wrapper (
end
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]);
`ifndef ASIC
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]);
`else
wire zer = 0;
wire[`NW_M1:0] old_warp_num;
VX_generic_register #(`NW_M1+1) store_wn(
.clk (clk),
.reset(reset),
.stall(zer),
.flush(zer),
.in (VX_gpr_read.warp_num),
.out (old_warp_num)
);
assign out_a_reg_data = (VX_gpr_jal.is_jal ? jal_data : (temp_a_reg_data[old_warp_num]));
assign out_b_reg_data = (temp_b_reg_data[old_warp_num]);
`endif
genvar warp_index;
generate

View File

@@ -16,40 +16,52 @@ module VX_scheduler (
reg rename_table[31:0];
reg[31:0] rename_table[`NW-1:0];
wire valid_wb = (VX_writeback_inter.wb != 0) && (|VX_writeback_inter.wb_valid) && (VX_writeback_inter.rd != 0);
wire wb_inc = (VX_bckE_req.wb != 0) && (VX_bckE_req.rd != 0);
// wire pass_through = ((VX_bckE_req.rs1 == VX_writeback_inter.rd) || (VX_bckE_req.rs2 == VX_writeback_inter.rd)) && valid_wb;
// wire pass_through = 0;
wire rs1_rename = rename_table[VX_bckE_req.rs1];
wire rs2_rename = rename_table[VX_bckE_req.rs2];
wire rs1_rename = rename_table[VX_bckE_req.warp_num][VX_bckE_req.rs1];
wire rs2_rename = rename_table[VX_bckE_req.warp_num][VX_bckE_req.rs2];
wire is_store = (VX_bckE_req.mem_write != `NO_MEM_WRITE);
wire is_load = (VX_bckE_req.mem_read != `NO_MEM_READ);
wire is_mem = is_store || is_load;
wire rs1_rename_qual = (rs1_rename && (VX_bckE_req.rs1 != 0));
wire rs2_rename_qual = (rs2_rename && (VX_bckE_req.rs2 != 0) && ((VX_bckE_req.rs2_src == `RS2_REG) || is_store)) || (VX_bckE_req.is_barrier) || (VX_bckE_req.is_wspawn);
wire rs1_pass = ((valid_wb && (VX_writeback_inter.rd == VX_bckE_req.rs1)));
wire rs2_pass = ((valid_wb && (VX_writeback_inter.rd == VX_bckE_req.rs2)));
// wire rs1_pass = 0;
// wire rs2_pass = 0;
wire using_rs2 = (VX_bckE_req.rs2_src == `RS2_REG) || is_store || VX_bckE_req.is_barrier || VX_bckE_req.is_wspawn;
wire rs1_rename_qual = ((rs1_rename || (rs1_pass && 0)) && (VX_bckE_req.rs1 != 0));
wire rs2_rename_qual = ((rs2_rename || (rs2_pass && 0)) && (VX_bckE_req.rs2 != 0 && using_rs2));
wire rename_valid = rs1_rename_qual || rs2_rename_qual ;
assign schedule_delay = (rename_valid) && (|VX_bckE_req.valid) || (memory_delay && (is_mem)) || (gpr_stage_delay && is_mem);
assign schedule_delay = ((rename_valid) && (|VX_bckE_req.valid)) || (memory_delay && (is_mem)) || (gpr_stage_delay && is_mem);
integer i;
integer w;
always @(posedge clk or posedge reset) begin
if (reset) begin
for (i = 0; i < 32; i = i + 1) rename_table[i] <= 0;
for (w = 0; w < `NW; w=w+1)
begin
for (i = 0; i < 32; i = i + 1)
begin
rename_table[w][i] <= 0;
end
end
end else begin
if (valid_wb ) rename_table[VX_writeback_inter.rd] <= 0;
if (!schedule_delay && wb_inc) rename_table[VX_bckE_req.rd] <= 1;
if (valid_wb ) rename_table[VX_writeback_inter.wb_warp_num][VX_writeback_inter.rd] <= 0;
if (!schedule_delay && wb_inc) rename_table[VX_bckE_req.warp_num ][VX_bckE_req.rd] <= 1;
end
end

View File

@@ -103,6 +103,9 @@ LOG=
# setup: source cshrc.modelsim
# vlib
lib:
vlib vortex_lib
comp:
vlog $(OPT) -work $(LIB) $(SRC)
# vlog -O0 -dpiheader vortex_dpi.h $(OPT) -work $(LIB) $(SRC)