critical path optimization - fpga fmax @4c = ~212 mhz

This commit is contained in:
Blaise Tine
2020-12-26 03:28:32 -08:00
parent d5c6b9b4d9
commit b459192dec
11 changed files with 216 additions and 169 deletions

View File

@@ -327,7 +327,7 @@
// Size of cache in bytes // Size of cache in bytes
`ifndef SMEM_SIZE `ifndef SMEM_SIZE
`define SMEM_SIZE 4096 `define SMEM_SIZE 8192
`endif `endif
// Number of banks // Number of banks

View File

@@ -307,21 +307,20 @@ module VX_decode #(
/////////////////////////////////////////////////////////////////////////// ///////////////////////////////////////////////////////////////////////////
assign decode_if.valid = ifetch_rsp_if.valid assign decode_if.valid = ifetch_rsp_if.valid;
&& (decode_if.ex_type != `EX_NOP); // skip noop
assign decode_if.wid = ifetch_rsp_if.wid; assign decode_if.wid = ifetch_rsp_if.wid;
assign decode_if.tmask = ifetch_rsp_if.tmask; assign decode_if.tmask = ifetch_rsp_if.tmask;
assign decode_if.PC = ifetch_rsp_if.PC; assign decode_if.PC = ifetch_rsp_if.PC;
assign decode_if.ex_type = is_lsu ? `EX_LSU : assign decode_if.ex_type = is_lsu ? `EX_LSU :
is_csr ? `EX_CSR : is_csr ? `EX_CSR :
is_mul ? `EX_MUL : is_mul ? `EX_MUL :
is_fpu ? `EX_FPU : is_fpu ? `EX_FPU :
is_gpu ? `EX_GPU : is_gpu ? `EX_GPU :
is_br ? `EX_ALU : is_br ? `EX_ALU :
(is_rtype || is_itype || is_lui || is_auipc) ? `EX_ALU : (is_rtype || is_itype || is_lui || is_auipc) ? `EX_ALU :
`EX_NOP; `EX_NOP;
assign decode_if.op_type = is_lsu ? `OP_BITS'(lsu_op) : assign decode_if.op_type = is_lsu ? `OP_BITS'(lsu_op) :
is_csr ? `OP_BITS'(csr_op) : is_csr ? `OP_BITS'(csr_op) :
@@ -367,17 +366,17 @@ module VX_decode #(
/////////////////////////////////////////////////////////////////////////// ///////////////////////////////////////////////////////////////////////////
wire decode_fire = decode_if.valid && decode_if.ready; wire decode_fire_unqual = ifetch_rsp_if.valid && decode_if.ready;
assign join_if.valid = decode_fire && is_gpu && (gpu_op == `GPU_JOIN); assign join_if.valid = decode_fire_unqual && is_gpu && (gpu_op == `GPU_JOIN);
assign join_if.wid = ifetch_rsp_if.wid; assign join_if.wid = ifetch_rsp_if.wid;
assign wstall_if.valid = decode_fire && (is_btype assign wstall_if.valid = decode_fire_unqual && (is_btype
|| is_jal || is_jal
|| is_jalr || is_jalr
|| (is_gpu && (gpu_op == `GPU_TMC || (is_gpu && (gpu_op == `GPU_TMC
|| gpu_op == `GPU_SPLIT || gpu_op == `GPU_SPLIT
|| gpu_op == `GPU_BAR))); || gpu_op == `GPU_BAR)));
assign wstall_if.wid = ifetch_rsp_if.wid; assign wstall_if.wid = ifetch_rsp_if.wid;
/////////////////////////////////////////////////////////////////////////// ///////////////////////////////////////////////////////////////////////////

View File

@@ -11,19 +11,18 @@ module VX_ibuffer #(
VX_decode_if ibuf_enq_if, VX_decode_if ibuf_enq_if,
// outputs // outputs
output wire [`NW_BITS-1:0] deq_wid_next,
VX_decode_if ibuf_deq_if VX_decode_if ibuf_deq_if
); );
localparam DATAW = `NUM_THREADS + 32 + `EX_BITS + `OP_BITS + `FRM_BITS + 1 + (`NR_BITS * 4) + 32 + 1 + 1 + `NUM_REGS; localparam DATAW = `NUM_THREADS + 32 + `EX_BITS + `OP_BITS + `FRM_BITS + 1 + (`NR_BITS * 4) + 32 + 1 + 1 + `NUM_REGS;
localparam SIZE = `IBUF_SIZE; localparam SIZE = `IBUF_SIZE;
localparam SIZEW = $clog2(SIZE+1);
localparam ADDRW = $clog2(SIZE); localparam ADDRW = $clog2(SIZE);
localparam NWARPSW = $clog2(`NUM_WARPS+1); localparam NWARPSW = $clog2(`NUM_WARPS+1);
reg [`NUM_WARPS-1:0][SIZEW-1:0] size_r; reg [`NUM_WARPS-1:0][ADDRW-1:0] used_r;
reg [`NUM_WARPS-1:0] full_r, empty_r, sizeMany_r;
wire [`NUM_WARPS-1:0] q_full; wire [`NUM_WARPS-1:0] q_full, q_empty;
wire [`NUM_WARPS-1:0][SIZEW-1:0] q_size; wire [`NUM_WARPS-1:0] q_sizeMany;
wire [DATAW-1:0] q_data_in; wire [DATAW-1:0] q_data_in;
wire [`NUM_WARPS-1:0][DATAW-1:0] q_data_prev; wire [`NUM_WARPS-1:0][DATAW-1:0] q_data_prev;
reg [`NUM_WARPS-1:0][DATAW-1:0] q_data_out; reg [`NUM_WARPS-1:0][DATAW-1:0] q_data_out;
@@ -36,15 +35,16 @@ module VX_ibuffer #(
wire writing = enq_fire && (i == ibuf_enq_if.wid); wire writing = enq_fire && (i == ibuf_enq_if.wid);
wire reading = deq_fire && (i == ibuf_deq_if.wid); wire reading = deq_fire && (i == ibuf_deq_if.wid);
wire is_slot0 = (0 == size_r[i]) || ((1 == size_r[i]) && reading); wire is_slot0 = (0 == used_r[i]) || ((1 == used_r[i]) && reading);
wire push = writing && !is_slot0; wire push = writing && !is_slot0;
wire pop = reading && (size_r[i] != 1); wire pop = reading && sizeMany_r[i];
VX_generic_queue #( VX_generic_queue #(
.DATAW (DATAW), .DATAW (DATAW),
.SIZE (SIZE), .SIZE (SIZE),
.FASTRAM (1) .BUFFERED (1),
.FASTRAM (1)
) queue ( ) queue (
.clk (clk), .clk (clk),
.reset (reset), .reset (reset),
@@ -59,26 +59,43 @@ module VX_ibuffer #(
always @(posedge clk) begin always @(posedge clk) begin
if (reset) begin if (reset) begin
size_r[i] <= 0; used_r[i] <= 0;
full_r[i] <= 0;
empty_r[i] <= 1;
sizeMany_r[i] <= 0;
end else begin end else begin
if (writing && !reading) begin if (writing && !reading) begin
size_r[i] <= size_r[i] + SIZEW'(1); empty_r[i] <= 0;
if (used_r[i] == ADDRW'(SIZE-1)) begin
full_r[i] <= 1;
end
if (used_r[i] == 1) begin
sizeMany_r[i] <= 1;
end
end end
if (reading && !writing) begin if (reading && !writing) begin
size_r[i] <= size_r[i] - SIZEW'(1); full_r[i] <= 0;
if (used_r[i] == ADDRW'(1)) begin
empty_r[i] <= 1;
end
if (used_r[i] == ADDRW'(2)) begin
sizeMany_r[i] <= 0;
end
end end
used_r[i] <= used_r[i] + ADDRW'($signed(2'(writing) - 2'(reading)));
end end
if (writing && is_slot0) begin if (writing && is_slot0) begin
q_data_out[i] <= q_data_in; q_data_out[i] <= q_data_in;
end end
if (reading && (size_r[i] != 1)) begin if (pop) begin
q_data_out[i] <= q_data_prev[i]; q_data_out[i] <= q_data_prev[i];
end end
end end
assign q_full[i] = (size_r[i] == SIZE); assign q_full[i] = full_r[i];
assign q_size[i] = size_r[i]; assign q_empty[i] = empty_r[i];
assign q_sizeMany[i] = sizeMany_r[i];
end end
/////////////////////////////////////////////////////////////////////////// ///////////////////////////////////////////////////////////////////////////
@@ -93,7 +110,7 @@ module VX_ibuffer #(
always @(*) begin always @(*) begin
valid_table_n = valid_table; valid_table_n = valid_table;
if (deq_fire) begin if (deq_fire) begin
valid_table_n[deq_wid] = (q_size[deq_wid] != SIZEW'(1)); valid_table_n[deq_wid] = q_sizeMany[deq_wid];
end end
if (enq_fire) begin if (enq_fire) begin
valid_table_n[ibuf_enq_if.wid] = 1; valid_table_n[ibuf_enq_if.wid] = 1;
@@ -103,24 +120,24 @@ module VX_ibuffer #(
// schedule the next instruction to issue // schedule the next instruction to issue
// do round-robin when multiple warps are active // do round-robin when multiple warps are active
always @(*) begin always @(*) begin
deq_valid_n = 0; deq_valid_n = 0;
deq_wid_n = 'x; deq_wid_n = 'x;
deq_instr_n = 'x; deq_instr_n = 'x;
schedule_table_n = schedule_table; schedule_table_n = schedule_table;
if (0 == num_warps) begin if (0 == num_warps) begin
deq_valid_n = enq_fire; deq_valid_n = enq_fire;
deq_wid_n = ibuf_enq_if.wid; deq_wid_n = ibuf_enq_if.wid;
deq_instr_n = q_data_in; deq_instr_n = q_data_in;
end else if ((1 == num_warps) || freeze) begin end else if ((1 == num_warps) || freeze) begin
deq_valid_n = (!deq_fire || (q_size[deq_wid] != SIZEW'(1))) || enq_fire; deq_valid_n = (!deq_fire || q_sizeMany[deq_wid]) || enq_fire;
deq_wid_n = (!deq_fire || (q_size[deq_wid] != SIZEW'(1))) ? deq_wid : ibuf_enq_if.wid; deq_wid_n = (!deq_fire || q_sizeMany[deq_wid]) ? deq_wid : ibuf_enq_if.wid;
deq_instr_n = deq_fire ? ((q_size[deq_wid] != SIZEW'(1)) ? q_data_prev[deq_wid] : q_data_in) : q_data_out[deq_wid]; deq_instr_n = deq_fire ? (q_sizeMany[deq_wid] ? q_data_prev[deq_wid] : q_data_in) : q_data_out[deq_wid];
end else begin end else begin
deq_valid_n = (| schedule_table_n);
for (integer i = 0; i < `NUM_WARPS; i++) begin for (integer i = 0; i < `NUM_WARPS; i++) begin
if (schedule_table_n[i]) begin if (schedule_table_n[i]) begin
deq_valid_n = 1;
deq_wid_n = `NW_BITS'(i); deq_wid_n = `NW_BITS'(i);
deq_instr_n = q_data_out[i]; deq_instr_n = q_data_out[i];
schedule_table_n[i] = 0; schedule_table_n[i] = 0;
@@ -130,8 +147,8 @@ module VX_ibuffer #(
end end
end end
wire warp_added = enq_fire && (0 == q_size[ibuf_enq_if.wid]); wire warp_added = enq_fire && q_empty[ibuf_enq_if.wid];
wire warp_removed = deq_fire && ~(enq_fire && ibuf_enq_if.wid == deq_wid) && ~(q_size[deq_wid] != SIZEW'(1)); wire warp_removed = deq_fire && ~(enq_fire && ibuf_enq_if.wid == deq_wid) && ~q_sizeMany[deq_wid];
always @(posedge clk) begin always @(posedge clk) begin
if (reset) begin if (reset) begin
@@ -162,23 +179,21 @@ module VX_ibuffer #(
`ifdef VERILATOR `ifdef VERILATOR
/*if (enq_fire || deq_fire || deq_valid) begin /*if (enq_fire || deq_fire || deq_valid) begin
$display("*** %t: cur=%b(%0d), nxt=%b(%0d), enq=%b(%0d), deq=%b(%0d), nw=%0d(%0d,%0d,%0d,%0d), sched=%b, sched_n=%b", $display("*** %t: cur=%b(%0d), nxt=%b(%0d), enq=%b(%0d), deq=%b(%0d), nw=%0d(%0d,%0d,%0d,%0d), sched=%b, sched_n=%b",
$time, deq_valid, deq_wid, deq_valid_n, deq_wid_n, enq_fire, ibuf_enq_if.wid, deq_fire, ibuf_deq_if.wid, num_warps, size_r[0], size_r[1], size_r[2], size_r[3], schedule_table, schedule_table_n); $time, deq_valid, deq_wid, deq_valid_n, deq_wid_n, enq_fire, ibuf_enq_if.wid, deq_fire, ibuf_deq_if.wid, num_warps, used_r[0], used_r[1], used_r[2], used_r[3], schedule_table, schedule_table_n);
end*/ end*/
begin // verify 'num_warps' begin // verify 'num_warps'
integer nw = 0; integer nw = 0;
for (integer i = 0; i < `NUM_WARPS; i++) begin for (integer i = 0; i < `NUM_WARPS; i++) begin
nw += 32'(q_size[i] != 0); nw += 32'(!q_empty[i]);
end end
assert(nw == 32'(num_warps)) else $error("%t: error: invalid num_warps: nw=%0d, ref=%0d", $time, num_warps, nw); assert(nw == 32'(num_warps)) else $error("%t: error: invalid num_warps: nw=%0d, ref=%0d", $time, num_warps, nw);
assert(~deq_valid || (q_size[deq_wid] != 0)) else $error("%t: error: invalid schedule: wid=%0d", $time, deq_wid); assert(~deq_valid || !q_empty[deq_wid]) else $error("%t: error: invalid schedule: wid=%0d", $time, deq_wid);
assert(~deq_fire || (q_size[deq_wid] != 0)) else $error("%t: error: invalid dequeu: wid=%0d", $time, deq_wid); assert(~deq_fire || !q_empty[deq_wid]) else $error("%t: error: invalid dequeu: wid=%0d", $time, deq_wid);
end end
`endif `endif
end end
end end
assign deq_wid_next = deq_wid_n;
assign ibuf_enq_if.ready = ~q_full[ibuf_enq_if.wid]; assign ibuf_enq_if.ready = ~q_full[ibuf_enq_if.wid];
assign q_data_in = {ibuf_enq_if.tmask, assign q_data_in = {ibuf_enq_if.tmask,
ibuf_enq_if.PC, ibuf_enq_if.PC,

View File

@@ -40,7 +40,8 @@ module VX_ipdom_stack #(
VX_dp_ram #( VX_dp_ram #(
.DATAW(WIDTH * 2), .DATAW(WIDTH * 2),
.SIZE(DEPTH), .SIZE(DEPTH),
.RWCHECK(0) .RWCHECK(1),
.FASTRAM(1)
) store ( ) store (
.clk(clk), .clk(clk),
.waddr(wr_ptr), .waddr(wr_ptr),

View File

@@ -28,7 +28,6 @@ module VX_issue #(
VX_gpr_rsp_if gpr_rsp_if(); VX_gpr_rsp_if gpr_rsp_if();
wire scoreboard_delay; wire scoreboard_delay;
wire [`NW_BITS-1:0] deq_wid_next;
VX_ibuffer #( VX_ibuffer #(
.CORE_ID(CORE_ID) .CORE_ID(CORE_ID)
@@ -37,7 +36,6 @@ module VX_issue #(
.reset (reset), .reset (reset),
.freeze (1'b0), .freeze (1'b0),
.ibuf_enq_if (decode_if), .ibuf_enq_if (decode_if),
.deq_wid_next (deq_wid_next),
.ibuf_deq_if (ibuf_deq_if) .ibuf_deq_if (ibuf_deq_if)
); );
@@ -48,8 +46,6 @@ module VX_issue #(
.reset (reset), .reset (reset),
.ibuf_deq_if (ibuf_deq_if), .ibuf_deq_if (ibuf_deq_if),
.writeback_if (writeback_if), .writeback_if (writeback_if),
.deq_wid_next (deq_wid_next),
.exe_delay (~execute_if.ready),
.delay (scoreboard_delay) .delay (scoreboard_delay)
); );
@@ -95,6 +91,9 @@ module VX_issue #(
.gpu_req_if (gpu_req_if) .gpu_req_if (gpu_req_if)
); );
// issue the instruction
assign ibuf_deq_if.ready = !scoreboard_delay && execute_if.ready;
`SCOPE_ASSIGN (issue_fire, ibuf_deq_if.valid && ibuf_deq_if.ready); `SCOPE_ASSIGN (issue_fire, ibuf_deq_if.valid && ibuf_deq_if.ready);
`SCOPE_ASSIGN (issue_wid, ibuf_deq_if.wid); `SCOPE_ASSIGN (issue_wid, ibuf_deq_if.wid);
`SCOPE_ASSIGN (issue_tmask, ibuf_deq_if.tmask); `SCOPE_ASSIGN (issue_tmask, ibuf_deq_if.tmask);
@@ -123,8 +122,8 @@ module VX_issue #(
`SCOPE_ASSIGN (writeback_data, writeback_if.data); `SCOPE_ASSIGN (writeback_data, writeback_if.data);
`ifdef PERF_ENABLE `ifdef PERF_ENABLE
reg [63:0] perf_ibf_stalls ; reg [63:0] perf_ibf_stalls;
reg [63:0] perf_scb_stalls ; reg [63:0] perf_scb_stalls;
reg [63:0] perf_alu_stalls; reg [63:0] perf_alu_stalls;
reg [63:0] perf_lsu_stalls; reg [63:0] perf_lsu_stalls;
reg [63:0] perf_csr_stalls; reg [63:0] perf_csr_stalls;

View File

@@ -3,21 +3,19 @@
module VX_scoreboard #( module VX_scoreboard #(
parameter CORE_ID = 0 parameter CORE_ID = 0
) ( ) (
input wire clk, input wire clk,
input wire reset, input wire reset,
VX_decode_if ibuf_deq_if, VX_decode_if ibuf_deq_if,
VX_writeback_if writeback_if, VX_writeback_if writeback_if,
input wire [`NW_BITS-1:0] deq_wid_next, output wire delay
input wire exe_delay,
output wire delay
); );
reg [`NUM_THREADS-1:0] inuse_registers [(`NUM_WARPS * `NUM_REGS)-1:0]; reg [`NUM_THREADS-1:0] inuse_registers [(`NUM_WARPS * `NUM_REGS)-1:0];
reg [`NUM_WARPS-1:0][`NUM_REGS-1:0] inuse_reg_mask, inuse_reg_mask_n; reg [`NUM_WARPS-1:0][`NUM_REGS-1:0] inuse_reg_mask;
reg [`NUM_REGS-1:0] deq_used_regs; wire [`NUM_REGS-1:0] inuse_regs;
wire [`NUM_THREADS-1:0] inuse_registers_n;
wire [`NUM_REGS-1:0] inuse_regs = deq_used_regs & ibuf_deq_if.used_regs; assign inuse_regs = inuse_reg_mask[ibuf_deq_if.wid] & ibuf_deq_if.used_regs;
assign delay = (| inuse_regs); assign delay = (| inuse_regs);
@@ -25,17 +23,7 @@ module VX_scoreboard #(
wire release_reg = writeback_if.valid && writeback_if.ready; wire release_reg = writeback_if.valid && writeback_if.ready;
wire [`NUM_THREADS-1:0] inuse_registers_n = inuse_registers[{writeback_if.wid, writeback_if.rd}] & ~writeback_if.tmask; assign inuse_registers_n = inuse_registers[{writeback_if.wid, writeback_if.rd}] & ~writeback_if.tmask;
always @(*) begin
inuse_reg_mask_n = inuse_reg_mask;
if (reserve_reg) begin
inuse_reg_mask_n[ibuf_deq_if.wid][ibuf_deq_if.rd] = 1;
end
if (release_reg) begin
inuse_reg_mask_n[writeback_if.wid][writeback_if.rd] = (| inuse_registers_n);
end
end
always @(posedge clk) begin always @(posedge clk) begin
if (reset) begin if (reset) begin
@@ -48,28 +36,24 @@ module VX_scoreboard #(
end else begin end else begin
if (reserve_reg) begin if (reserve_reg) begin
inuse_registers[{ibuf_deq_if.wid, ibuf_deq_if.rd}] <= ibuf_deq_if.tmask; inuse_registers[{ibuf_deq_if.wid, ibuf_deq_if.rd}] <= ibuf_deq_if.tmask;
inuse_reg_mask[ibuf_deq_if.wid][ibuf_deq_if.rd] <= 1;
end end
if (release_reg) begin if (release_reg) begin
assert(inuse_reg_mask[writeback_if.wid][writeback_if.rd] != 0) assert(inuse_reg_mask[writeback_if.wid][writeback_if.rd] != 0)
else $error("*** %t: core%0d: invalid writeback register: wid=%0d, PC=%0h, rd=%0d", else $error("*** %t: core%0d: invalid writeback register: wid=%0d, PC=%0h, rd=%0d",
$time, CORE_ID, writeback_if.wid, writeback_if.PC, writeback_if.rd); $time, CORE_ID, writeback_if.wid, writeback_if.PC, writeback_if.rd);
inuse_registers[{writeback_if.wid, writeback_if.rd}] <= inuse_registers_n; inuse_registers[{writeback_if.wid, writeback_if.rd}] <= inuse_registers_n;
inuse_reg_mask[writeback_if.wid][writeback_if.rd] <= (| inuse_registers_n);
end end
inuse_reg_mask <= inuse_reg_mask_n;
end end
deq_used_regs <= inuse_reg_mask_n[deq_wid_next];
end end
// issue the instruction
assign ibuf_deq_if.ready = ~(delay || exe_delay);
`ifdef DBG_PRINT_PIPELINE `ifdef DBG_PRINT_PIPELINE
always @(posedge clk) begin always @(posedge clk) begin
if (ibuf_deq_if.valid && ~ibuf_deq_if.ready) begin if (ibuf_deq_if.valid && ~ibuf_deq_if.ready) begin
$display("%t: core%0d-stall: wid=%0d, PC=%0h, rd=%0d, wb=%0d, inuse=%b%b%b%b, exe=%b", $display("%t: core%0d-stall: wid=%0d, PC=%0h, rd=%0d, wb=%0d, inuse=%b%b%b%b",
$time, CORE_ID, ibuf_deq_if.wid, ibuf_deq_if.PC, ibuf_deq_if.rd, ibuf_deq_if.wb, $time, CORE_ID, ibuf_deq_if.wid, ibuf_deq_if.PC, ibuf_deq_if.rd, ibuf_deq_if.wb,
inuse_regs[ibuf_deq_if.rd], inuse_regs[ibuf_deq_if.rs1], inuse_regs[ibuf_deq_if.rs2], inuse_regs[ibuf_deq_if.rs3], exe_delay); inuse_regs[ibuf_deq_if.rd], inuse_regs[ibuf_deq_if.rs1], inuse_regs[ibuf_deq_if.rs2], inuse_regs[ibuf_deq_if.rs3]);
end end
end end
`endif `endif
@@ -80,9 +64,9 @@ module VX_scoreboard #(
stall_ctr <= 0; stall_ctr <= 0;
end else if (ibuf_deq_if.valid && ~ibuf_deq_if.ready) begin end else if (ibuf_deq_if.valid && ~ibuf_deq_if.ready) begin
stall_ctr <= stall_ctr + 1; stall_ctr <= stall_ctr + 1;
assert(stall_ctr < 100000) else $error("*** %t: core%0d-stalled: wid=%0d, PC=%0h, rd=%0d, wb=%0d, inuse=%b%b%b%b, exe=%b", assert(stall_ctr < 100000) else $error("*** %t: core%0d-stalled: wid=%0d, PC=%0h, rd=%0d, wb=%0d, inuse=%b%b%b%b",
$time, CORE_ID, ibuf_deq_if.wid, ibuf_deq_if.PC, ibuf_deq_if.rd, ibuf_deq_if.wb, $time, CORE_ID, ibuf_deq_if.wid, ibuf_deq_if.PC, ibuf_deq_if.rd, ibuf_deq_if.wb,
inuse_regs[ibuf_deq_if.rd], inuse_regs[ibuf_deq_if.rs1], inuse_regs[ibuf_deq_if.rs2], inuse_regs[ibuf_deq_if.rs3], exe_delay); inuse_regs[ibuf_deq_if.rd], inuse_regs[ibuf_deq_if.rs1], inuse_regs[ibuf_deq_if.rs2], inuse_regs[ibuf_deq_if.rs3]);
end else if (ibuf_deq_if.valid && ibuf_deq_if.ready) begin end else if (ibuf_deq_if.valid && ibuf_deq_if.ready) begin
stall_ctr <= 0; stall_ctr <= 0;
end end

View File

@@ -17,7 +17,7 @@ module VX_writeback #(
VX_writeback_if writeback_if VX_writeback_if writeback_if
); );
wire alu_valid = alu_commit_if.valid && alu_commit_if.wb; wire alu_valid = alu_commit_if.valid && alu_commit_if.wb;
wire ld_valid = ld_commit_if.valid /*&& ld_commit_if.wb*/; wire ld_valid = ld_commit_if.valid && ld_commit_if.wb;
wire csr_valid = csr_commit_if.valid && csr_commit_if.wb; wire csr_valid = csr_commit_if.valid && csr_commit_if.wb;
wire mul_valid = mul_commit_if.valid && mul_commit_if.wb; wire mul_valid = mul_commit_if.valid && mul_commit_if.wb;
wire fpu_valid = fpu_commit_if.valid && fpu_commit_if.wb; wire fpu_valid = fpu_commit_if.valid && fpu_commit_if.wb;

View File

@@ -91,6 +91,11 @@ module VX_cache #(
wire [NUM_BANKS-1:0] per_bank_core_req_valid; wire [NUM_BANKS-1:0] per_bank_core_req_valid;
wire [NUM_BANKS-1:0][`REQS_BITS-1:0] per_bank_core_req_tid; wire [NUM_BANKS-1:0][`REQS_BITS-1:0] per_bank_core_req_tid;
wire [NUM_BANKS-1:0] per_bank_core_req_rw;
wire [NUM_BANKS-1:0][WORD_SIZE-1:0] per_bank_core_req_byteen;
wire [NUM_BANKS-1:0][`WORD_ADDR_WIDTH-1:0] per_bank_core_req_addr;
wire [NUM_BANKS-1:0][CORE_TAG_WIDTH-1:0] per_bank_core_req_tag;
wire [NUM_BANKS-1:0][`WORD_WIDTH-1:0] per_bank_core_req_data;
wire [NUM_BANKS-1:0] per_bank_core_req_ready; wire [NUM_BANKS-1:0] per_bank_core_req_ready;
wire [NUM_BANKS-1:0] per_bank_core_rsp_valid; wire [NUM_BANKS-1:0] per_bank_core_rsp_valid;
@@ -122,7 +127,8 @@ module VX_cache #(
.BANK_LINE_SIZE (BANK_LINE_SIZE), .BANK_LINE_SIZE (BANK_LINE_SIZE),
.NUM_BANKS (NUM_BANKS), .NUM_BANKS (NUM_BANKS),
.WORD_SIZE (WORD_SIZE), .WORD_SIZE (WORD_SIZE),
.NUM_REQS (NUM_REQS) .NUM_REQS (NUM_REQS),
.CORE_TAG_WIDTH (CORE_TAG_WIDTH)
) cache_core_req_bank_sel ( ) cache_core_req_bank_sel (
.clk (clk), .clk (clk),
.reset (reset), .reset (reset),
@@ -132,11 +138,20 @@ module VX_cache #(
`UNUSED_PIN (bank_stalls), `UNUSED_PIN (bank_stalls),
`endif `endif
.core_req_valid (core_req_valid), .core_req_valid (core_req_valid),
.core_req_rw (core_req_rw),
.core_req_byteen(core_req_byteen),
.core_req_addr (core_req_addr), .core_req_addr (core_req_addr),
.core_req_data (core_req_data),
.core_req_tag (core_req_tag),
.core_req_ready (core_req_ready), .core_req_ready (core_req_ready),
.per_bank_valid (per_bank_core_req_valid), .per_bank_core_req_valid (per_bank_core_req_valid),
.per_bank_tid (per_bank_core_req_tid), .per_bank_core_req_tid (per_bank_core_req_tid),
.per_bank_ready (per_bank_core_req_ready) .per_bank_core_req_rw (per_bank_core_req_rw),
.per_bank_core_req_byteen(per_bank_core_req_byteen),
.per_bank_core_req_addr (per_bank_core_req_addr),
.per_bank_core_req_tag (per_bank_core_req_tag),
.per_bank_core_req_data (per_bank_core_req_data),
.per_bank_core_req_ready (per_bank_core_req_ready)
); );
assign dram_req_tag = dram_req_addr; assign dram_req_tag = dram_req_addr;
@@ -179,11 +194,11 @@ module VX_cache #(
// Core Req // Core Req
assign curr_bank_core_req_valid = per_bank_core_req_valid[i]; assign curr_bank_core_req_valid = per_bank_core_req_valid[i];
assign curr_bank_core_req_tid = per_bank_core_req_tid[i]; assign curr_bank_core_req_tid = per_bank_core_req_tid[i];
assign curr_bank_core_req_addr = core_req_addr[per_bank_core_req_tid[i]]; assign curr_bank_core_req_addr = per_bank_core_req_addr[i];
assign curr_bank_core_req_rw = core_req_rw[per_bank_core_req_tid[i]]; assign curr_bank_core_req_rw = per_bank_core_req_rw[i];
assign curr_bank_core_req_byteen = core_req_byteen[per_bank_core_req_tid[i]]; assign curr_bank_core_req_byteen = per_bank_core_req_byteen[i];
assign curr_bank_core_req_data = core_req_data[per_bank_core_req_tid[i]]; assign curr_bank_core_req_data = per_bank_core_req_data[i];
assign curr_bank_core_req_tag = core_req_tag[per_bank_core_req_tid[i]]; assign curr_bank_core_req_tag = per_bank_core_req_tag[i];
assign per_bank_core_req_ready[i] = curr_bank_core_req_ready; assign per_bank_core_req_ready[i] = curr_bank_core_req_ready;
// Core WB // Core WB

View File

@@ -8,48 +8,80 @@ module VX_cache_core_req_bank_sel #(
// Number of banks // Number of banks
parameter NUM_BANKS = 1, parameter NUM_BANKS = 1,
// Number of Word requests per cycle // Number of Word requests per cycle
parameter NUM_REQS = 1 parameter NUM_REQS = 1,
// core request tag size
parameter CORE_TAG_WIDTH = 1
) ( ) (
input wire clk, input wire clk,
input wire reset, input wire reset,
input wire [NUM_REQS-1:0] core_req_valid,
input wire [NUM_REQS-1:0][`WORD_ADDR_WIDTH-1:0] core_req_addr, output wire [63:0] bank_stalls,
output wire [NUM_REQS-1:0] core_req_ready,
output wire [NUM_BANKS-1:0] per_bank_valid, input wire [NUM_REQS-1:0] core_req_valid,
output wire [NUM_BANKS-1:0][`REQS_BITS-1:0] per_bank_tid, input wire [NUM_REQS-1:0] core_req_rw,
input wire [NUM_BANKS-1:0] per_bank_ready, input wire [NUM_REQS-1:0][WORD_SIZE-1:0] core_req_byteen,
output wire [63:0] bank_stalls input wire [NUM_REQS-1:0][`WORD_ADDR_WIDTH-1:0] core_req_addr,
input wire [NUM_REQS-1:0][`WORD_WIDTH-1:0] core_req_data,
input wire [NUM_REQS-1:0][CORE_TAG_WIDTH-1:0] core_req_tag,
output wire [NUM_REQS-1:0] core_req_ready,
output wire [NUM_BANKS-1:0] per_bank_core_req_valid,
output wire [NUM_BANKS-1:0][`REQS_BITS-1:0] per_bank_core_req_tid,
output wire [NUM_BANKS-1:0] per_bank_core_req_rw,
output wire [NUM_BANKS-1:0][WORD_SIZE-1:0] per_bank_core_req_byteen,
output wire [NUM_BANKS-1:0][`WORD_ADDR_WIDTH-1:0] per_bank_core_req_addr,
output wire [NUM_BANKS-1:0][CORE_TAG_WIDTH-1:0] per_bank_core_req_tag,
output wire [NUM_BANKS-1:0][`WORD_WIDTH-1:0] per_bank_core_req_data,
input wire [NUM_BANKS-1:0] per_bank_core_req_ready
); );
if (NUM_BANKS > 1) begin if (NUM_BANKS > 1) begin
reg [NUM_BANKS-1:0] per_bank_valid_r;
reg [NUM_BANKS-1:0][`REQS_BITS-1:0] per_bank_tid_r; reg [NUM_BANKS-1:0] per_bank_core_req_valid_r;
reg [NUM_REQS-1:0] core_req_ready_r; reg [NUM_BANKS-1:0][`REQS_BITS-1:0] per_bank_core_req_tid_r;
reg [NUM_BANKS-1:0] core_req_sel_r; reg [NUM_BANKS-1:0] per_bank_core_req_rw_r;
wire [NUM_REQS-1:0][`BANK_BITS-1:0] core_req_bid; reg [NUM_BANKS-1:0][WORD_SIZE-1:0] per_bank_core_req_byteen_r;
reg [NUM_BANKS-1:0][`WORD_ADDR_WIDTH-1:0] per_bank_core_req_addr_r;
reg [NUM_BANKS-1:0][CORE_TAG_WIDTH-1:0] per_bank_core_req_tag_r;
reg [NUM_BANKS-1:0][`WORD_WIDTH-1:0] per_bank_core_req_data_r;
reg [NUM_REQS-1:0] core_req_ready_r;
reg [NUM_BANKS-1:0] core_req_sel_r;
wire [NUM_REQS-1:0][`BANK_BITS-1:0] core_req_bid;
for (genvar i = 0; i < NUM_REQS; ++i) begin for (genvar i = 0; i < NUM_REQS; ++i) begin
assign core_req_bid[i] = core_req_addr[i][`BANK_SELECT_ADDR_RNG]; assign core_req_bid[i] = core_req_addr[i][`BANK_SELECT_ADDR_RNG];
end end
always @(*) begin always @(*) begin
per_bank_valid_r = 0; per_bank_core_req_valid_r = 0;
per_bank_tid_r = 'x; per_bank_core_req_tid_r = 'x;
per_bank_core_req_rw_r = 'x;
per_bank_core_req_byteen_r= 'x;
per_bank_core_req_addr_r = 'x;
per_bank_core_req_tag_r = 'x;
per_bank_core_req_data_r = 'x;
for (integer i = NUM_REQS-1; i >= 0; --i) begin for (integer i = NUM_REQS-1; i >= 0; --i) begin
if (core_req_valid[i]) begin if (core_req_valid[i]) begin
per_bank_valid_r[core_req_bid[i]] = 1; per_bank_core_req_valid_r[core_req_bid[i]] = 1;
per_bank_tid_r[core_req_bid[i]] = `REQS_BITS'(i); per_bank_core_req_tid_r[core_req_bid[i]] = `REQS_BITS'(i);
per_bank_core_req_rw_r[core_req_bid[i]] = core_req_rw[i];
per_bank_core_req_byteen_r[core_req_bid[i]]= core_req_byteen[i];
per_bank_core_req_addr_r[core_req_bid[i]] = core_req_addr[i];
per_bank_core_req_tag_r[core_req_bid[i]] = core_req_tag[i];
per_bank_core_req_data_r[core_req_bid[i]] = core_req_data[i];
end end
end end
end end
always @(*) begin always @(*) begin
core_req_ready_r = 0; core_req_ready_r = 0;
core_req_sel_r = 0; core_req_sel_r = 0;
for (integer j = 0; j < NUM_BANKS; ++j) begin for (integer j = 0; j < NUM_BANKS; ++j) begin
for (integer i = 0; i < NUM_REQS; ++i) begin for (integer i = 0; i < NUM_REQS; ++i) begin
if (core_req_valid[i] && (core_req_bid[i] == `BANK_BITS'(j))) begin if (core_req_valid[i] && (core_req_bid[i] == `BANK_BITS'(j))) begin
core_req_ready_r[i] = per_bank_ready[j]; core_req_ready_r[i] = per_bank_core_req_ready[j];
core_req_sel_r[i] = 1; core_req_sel_r[i] = 1;
break; break;
end end
end end
@@ -65,20 +97,30 @@ module VX_cache_core_req_bank_sel #(
end end
end end
assign per_bank_valid = per_bank_valid_r; assign bank_stalls = bank_stalls_r;
assign per_bank_tid = per_bank_tid_r; assign per_bank_core_req_valid = per_bank_core_req_valid_r;
assign core_req_ready = core_req_ready_r; assign per_bank_core_req_tid = per_bank_core_req_tid_r;
assign bank_stalls = bank_stalls_r; assign per_bank_core_req_rw = per_bank_core_req_rw_r;
assign per_bank_core_req_byteen = per_bank_core_req_byteen_r;
assign per_bank_core_req_addr = per_bank_core_req_addr_r;
assign per_bank_core_req_tag = per_bank_core_req_tag_r;
assign per_bank_core_req_data = per_bank_core_req_data_r;
assign core_req_ready = core_req_ready_r;
end else begin end else begin
`UNUSED_VAR (clk) `UNUSED_VAR (clk)
`UNUSED_VAR (reset) `UNUSED_VAR (reset)
`UNUSED_VAR (core_req_valid) assign bank_stalls = 0;
`UNUSED_VAR (core_req_addr) assign per_bank_core_req_valid = core_req_valid;
assign per_bank_valid = core_req_valid; assign per_bank_core_req_tid[0] = 0;
assign per_bank_tid = 0; assign per_bank_core_req_rw[0] = core_req_rw;
assign core_req_ready[0] = per_bank_ready; assign per_bank_core_req_byteen[0] = core_req_byteen;
assign bank_stalls = 0; assign per_bank_core_req_addr[0] = core_req_addr;
assign per_bank_core_req_tag[0] = core_req_tag;
assign per_bank_core_req_data[0] = core_req_data;
assign core_req_ready[0] = per_bank_core_req_ready;
end end
endmodule endmodule

View File

@@ -37,7 +37,6 @@ module VX_cache_core_rsp_merge #(
if (CORE_TAG_ID_BITS != 0) begin if (CORE_TAG_ID_BITS != 0) begin
reg [CORE_TAG_WIDTH-1:0] core_rsp_tag_unqual; reg [CORE_TAG_WIDTH-1:0] core_rsp_tag_unqual;
reg [CORE_TAG_ID_BITS-1:0] sel_tag_id;
reg core_rsp_valid_unaual_any; reg core_rsp_valid_unaual_any;
wire core_rsp_ready_unqual; wire core_rsp_ready_unqual;
@@ -48,19 +47,16 @@ module VX_cache_core_rsp_merge #(
core_rsp_data_unqual = 'x; core_rsp_data_unqual = 'x;
core_rsp_bank_select = 0; core_rsp_bank_select = 0;
sel_tag_id = 'x;
for (integer i = 0; i < NUM_BANKS; i++) begin for (integer i = 0; i < NUM_BANKS; i++) begin
if (per_bank_core_rsp_valid[i]) begin if (per_bank_core_rsp_valid[i]) begin
core_rsp_tag_unqual = per_bank_core_rsp_tag[i]; core_rsp_tag_unqual = per_bank_core_rsp_tag[i];
sel_tag_id = per_bank_core_rsp_tag[i][CORE_TAG_ID_BITS-1:0];
break; break;
end end
end end
for (integer i = 0; i < NUM_BANKS; i++) begin for (integer i = 0; i < NUM_BANKS; i++) begin
if (per_bank_core_rsp_valid[i] if (per_bank_core_rsp_valid[i]
&& (per_bank_core_rsp_tag[i][CORE_TAG_ID_BITS-1:0] == sel_tag_id)) begin && (per_bank_core_rsp_tag[i][CORE_TAG_ID_BITS-1:0] == core_rsp_tag_unqual[CORE_TAG_ID_BITS-1:0])) begin
core_rsp_valid_unaual_any = 1; core_rsp_valid_unaual_any = 1;
core_rsp_valid_unqual[per_bank_core_rsp_tid[i]] = 1; core_rsp_valid_unqual[per_bank_core_rsp_tid[i]] = 1;
core_rsp_data_unqual[per_bank_core_rsp_tid[i]] = per_bank_core_rsp_data[i]; core_rsp_data_unqual[per_bank_core_rsp_tid[i]] = per_bank_core_rsp_data[i];

View File

@@ -60,25 +60,21 @@ module VX_generic_queue #(
full_r <= 0; full_r <= 0;
used_r <= 0; used_r <= 0;
end else begin end else begin
if (push) begin assert(!push || !full);
assert(!full); assert(!pop || !empty);
if (!pop) begin if (push && !pop) begin
empty_r <= 0; empty_r <= 0;
if (used_r == ADDRW'(SIZE-1)) begin if (used_r == ADDRW'(SIZE-1)) begin
full_r <= 1; full_r <= 1;
end
end end
end end
if (pop) begin if (pop && !push) begin
assert(!empty); full_r <= 0;
if (!push) begin if (used_r == ADDRW'(1)) begin
full_r <= 0; empty_r <= 1;
if (used_r == ADDRW'(1)) begin end;
empty_r <= 1;
end;
end
end end
used_r <= used_r + (ADDRW'(push) - ADDRW'(pop)); used_r <= used_r + ADDRW'($signed(2'(push) - 2'(pop)));
end end
end end