minor update
This commit is contained in:
@@ -47,8 +47,6 @@ module VX_operands import VX_gpu_pkg::*; #(
|
|||||||
reg [`NUM_THREADS-1:0] cache_tmask_n [ISSUE_RATIO-1:0];
|
reg [`NUM_THREADS-1:0] cache_tmask_n [ISSUE_RATIO-1:0];
|
||||||
reg [ISSUE_RATIO-1:0] cache_eop, cache_eop_n;
|
reg [ISSUE_RATIO-1:0] cache_eop, cache_eop_n;
|
||||||
|
|
||||||
reg valid_out_r;
|
|
||||||
reg [DATAW-1:0] data_out_r;
|
|
||||||
reg [`NUM_THREADS-1:0][`XLEN-1:0] rs1_data, rs1_data_n;
|
reg [`NUM_THREADS-1:0][`XLEN-1:0] rs1_data, rs1_data_n;
|
||||||
reg [`NUM_THREADS-1:0][`XLEN-1:0] rs2_data, rs2_data_n;
|
reg [`NUM_THREADS-1:0][`XLEN-1:0] rs2_data, rs2_data_n;
|
||||||
reg [`NUM_THREADS-1:0][`XLEN-1:0] rs3_data, rs3_data_n;
|
reg [`NUM_THREADS-1:0][`XLEN-1:0] rs3_data, rs3_data_n;
|
||||||
@@ -60,7 +58,7 @@ module VX_operands import VX_gpu_pkg::*; #(
|
|||||||
reg rs3_ready, rs3_ready_n;
|
reg rs3_ready, rs3_ready_n;
|
||||||
reg data_ready, data_ready_n;
|
reg data_ready, data_ready_n;
|
||||||
|
|
||||||
wire ready_out = operands_if[i].ready;
|
wire stg_valid_in, stg_ready_in;
|
||||||
|
|
||||||
wire is_rs1_zero = (scoreboard_if[i].data.rs1 == 0);
|
wire is_rs1_zero = (scoreboard_if[i].data.rs1 == 0);
|
||||||
wire is_rs2_zero = (scoreboard_if[i].data.rs2 == 0);
|
wire is_rs2_zero = (scoreboard_if[i].data.rs2 == 0);
|
||||||
@@ -85,7 +83,7 @@ module VX_operands import VX_gpu_pkg::*; #(
|
|||||||
|
|
||||||
case (state)
|
case (state)
|
||||||
STATE_IDLE: begin
|
STATE_IDLE: begin
|
||||||
if (valid_out_r && ready_out) begin
|
if (operands_if[i].valid && operands_if[i].ready) begin
|
||||||
data_ready_n = 0;
|
data_ready_n = 0;
|
||||||
end
|
end
|
||||||
if (scoreboard_if[i].valid && data_ready_n == 0) begin
|
if (scoreboard_if[i].valid && data_ready_n == 0) begin
|
||||||
@@ -177,33 +175,11 @@ module VX_operands import VX_gpu_pkg::*; #(
|
|||||||
state <= STATE_IDLE;
|
state <= STATE_IDLE;
|
||||||
cache_eop <= {ISSUE_RATIO{1'b1}};
|
cache_eop <= {ISSUE_RATIO{1'b1}};
|
||||||
data_ready <= 0;
|
data_ready <= 0;
|
||||||
valid_out_r <= 0;
|
|
||||||
end else begin
|
end else begin
|
||||||
state <= state_n;
|
state <= state_n;
|
||||||
cache_eop <= cache_eop_n;
|
cache_eop <= cache_eop_n;
|
||||||
data_ready <= data_ready_n;
|
data_ready <= data_ready_n;
|
||||||
if (~valid_out_r) begin
|
|
||||||
valid_out_r <= scoreboard_if[i].valid && data_ready;
|
|
||||||
end else if (ready_out) begin
|
|
||||||
valid_out_r <= 0;
|
|
||||||
end
|
end
|
||||||
end
|
|
||||||
|
|
||||||
if (~valid_out_r) begin
|
|
||||||
data_out_r <= {scoreboard_if[i].data.uuid,
|
|
||||||
scoreboard_if[i].data.wis,
|
|
||||||
scoreboard_if[i].data.tmask,
|
|
||||||
scoreboard_if[i].data.PC,
|
|
||||||
scoreboard_if[i].data.wb,
|
|
||||||
scoreboard_if[i].data.ex_type,
|
|
||||||
scoreboard_if[i].data.op_type,
|
|
||||||
scoreboard_if[i].data.op_mod,
|
|
||||||
scoreboard_if[i].data.use_PC,
|
|
||||||
scoreboard_if[i].data.use_imm,
|
|
||||||
scoreboard_if[i].data.imm,
|
|
||||||
scoreboard_if[i].data.rd};
|
|
||||||
end
|
|
||||||
|
|
||||||
gpr_rd_rid <= gpr_rd_rid_n;
|
gpr_rd_rid <= gpr_rd_rid_n;
|
||||||
gpr_rd_wis <= gpr_rd_wis_n;
|
gpr_rd_wis <= gpr_rd_wis_n;
|
||||||
rs2_ready <= rs2_ready_n;
|
rs2_ready <= rs2_ready_n;
|
||||||
@@ -218,8 +194,33 @@ module VX_operands import VX_gpu_pkg::*; #(
|
|||||||
cache_tmask <= cache_tmask_n;
|
cache_tmask <= cache_tmask_n;
|
||||||
end
|
end
|
||||||
|
|
||||||
assign operands_if[i].valid = valid_out_r;
|
assign stg_valid_in = scoreboard_if[i].valid && data_ready;
|
||||||
assign {operands_if[i].data.uuid,
|
assign scoreboard_if[i].ready = stg_ready_in && data_ready;
|
||||||
|
|
||||||
|
VX_toggle_buffer #(
|
||||||
|
.DATAW (DATAW)
|
||||||
|
) staging_buffer (
|
||||||
|
.clk (clk),
|
||||||
|
.reset (reset),
|
||||||
|
.valid_in (stg_valid_in),
|
||||||
|
.data_in ({
|
||||||
|
scoreboard_if[i].data.uuid,
|
||||||
|
scoreboard_if[i].data.wis,
|
||||||
|
scoreboard_if[i].data.tmask,
|
||||||
|
scoreboard_if[i].data.PC,
|
||||||
|
scoreboard_if[i].data.wb,
|
||||||
|
scoreboard_if[i].data.ex_type,
|
||||||
|
scoreboard_if[i].data.op_type,
|
||||||
|
scoreboard_if[i].data.op_mod,
|
||||||
|
scoreboard_if[i].data.use_PC,
|
||||||
|
scoreboard_if[i].data.use_imm,
|
||||||
|
scoreboard_if[i].data.imm,
|
||||||
|
scoreboard_if[i].data.rd
|
||||||
|
}),
|
||||||
|
.ready_in (stg_ready_in),
|
||||||
|
.valid_out (operands_if[i].valid),
|
||||||
|
.data_out ({
|
||||||
|
operands_if[i].data.uuid,
|
||||||
operands_if[i].data.wis,
|
operands_if[i].data.wis,
|
||||||
operands_if[i].data.tmask,
|
operands_if[i].data.tmask,
|
||||||
operands_if[i].data.PC,
|
operands_if[i].data.PC,
|
||||||
@@ -230,13 +231,15 @@ module VX_operands import VX_gpu_pkg::*; #(
|
|||||||
operands_if[i].data.use_PC,
|
operands_if[i].data.use_PC,
|
||||||
operands_if[i].data.use_imm,
|
operands_if[i].data.use_imm,
|
||||||
operands_if[i].data.imm,
|
operands_if[i].data.imm,
|
||||||
operands_if[i].data.rd} = data_out_r;
|
operands_if[i].data.rd
|
||||||
|
}),
|
||||||
|
.ready_out (operands_if[i].ready)
|
||||||
|
);
|
||||||
|
|
||||||
assign operands_if[i].data.rs1_data = rs1_data;
|
assign operands_if[i].data.rs1_data = rs1_data;
|
||||||
assign operands_if[i].data.rs2_data = rs2_data;
|
assign operands_if[i].data.rs2_data = rs2_data;
|
||||||
assign operands_if[i].data.rs3_data = rs3_data;
|
assign operands_if[i].data.rs3_data = rs3_data;
|
||||||
|
|
||||||
assign scoreboard_if[i].ready = ~valid_out_r && data_ready;
|
|
||||||
|
|
||||||
// GPR banks
|
// GPR banks
|
||||||
|
|
||||||
reg [RAM_ADDRW-1:0] gpr_rd_addr;
|
reg [RAM_ADDRW-1:0] gpr_rd_addr;
|
||||||
|
|||||||
@@ -152,50 +152,46 @@ module VX_scoreboard import VX_gpu_pkg::*; #(
|
|||||||
assign perf_issue_stalls_per_cycle[i] = ibuffer_if[i].valid && ~ibuffer_if[i].ready;
|
assign perf_issue_stalls_per_cycle[i] = ibuffer_if[i].valid && ~ibuffer_if[i].ready;
|
||||||
`endif
|
`endif
|
||||||
|
|
||||||
reg [DATAW-1:0] data_out_r;
|
wire [3:0] operands_busy = {inuse_rd, inuse_rs1, inuse_rs2, inuse_rs3};
|
||||||
reg valid_out_r;
|
wire operands_ready = ~(| operands_busy);
|
||||||
wire ready_out;
|
|
||||||
|
|
||||||
wire [3:0] ready_masks = ~{inuse_rd, inuse_rs1, inuse_rs2, inuse_rs3};
|
wire stg_valid_in, stg_ready_in;
|
||||||
wire deps_ready = (& ready_masks);
|
assign stg_valid_in = ibuffer_if[i].valid && operands_ready;
|
||||||
|
assign ibuffer_if[i].ready = stg_ready_in && operands_ready;
|
||||||
|
|
||||||
wire valid_in = ibuffer_if[i].valid && deps_ready;
|
VX_stream_buffer #(
|
||||||
wire ready_in = ~valid_out_r && deps_ready;
|
.DATAW (DATAW)
|
||||||
wire [DATAW-1:0] data_in = ibuffer_if[i].data;
|
) staging_buffer (
|
||||||
|
.clk (clk),
|
||||||
assign ready_out = scoreboard_if[i].ready;
|
.reset (reset),
|
||||||
|
.valid_in (stg_valid_in),
|
||||||
|
.data_in (ibuffer_if[i].data),
|
||||||
|
.ready_in (stg_ready_in),
|
||||||
|
.valid_out (scoreboard_if[i].valid),
|
||||||
|
.data_out (scoreboard_if[i].data),
|
||||||
|
.ready_out (scoreboard_if[i].ready)
|
||||||
|
);
|
||||||
|
|
||||||
always @(posedge clk) begin
|
always @(posedge clk) begin
|
||||||
if (reset) begin
|
if (reset) begin
|
||||||
valid_out_r <= 0;
|
|
||||||
inuse_regs <= '0;
|
inuse_regs <= '0;
|
||||||
end else begin
|
end else begin
|
||||||
if (writeback_fire) begin
|
if (writeback_fire) begin
|
||||||
inuse_regs[writeback_if[i].data.wis][writeback_if[i].data.rd] <= 0;
|
inuse_regs[writeback_if[i].data.wis][writeback_if[i].data.rd] <= 0;
|
||||||
end
|
end
|
||||||
if (~valid_out_r) begin
|
if (stg_valid_in && stg_ready_in && ibuffer_if[i].data.wb) begin
|
||||||
valid_out_r <= valid_in;
|
inuse_regs[ibuffer_if[i].data.wis][ibuffer_if[i].data.rd] <= 1;
|
||||||
end else if (ready_out) begin
|
end
|
||||||
if (scoreboard_if[i].data.wb) begin
|
end
|
||||||
inuse_regs[scoreboard_if[i].data.wis][scoreboard_if[i].data.rd] <= 1;
|
|
||||||
`ifdef PERF_ENABLE
|
`ifdef PERF_ENABLE
|
||||||
inuse_units[scoreboard_if[i].data.wis][scoreboard_if[i].data.rd] <= scoreboard_if[i].data.ex_type;
|
if (stg_valid_in && stg_ready_in && ibuffer_if[i].data.wb) begin
|
||||||
if (scoreboard_if[i].data.ex_type == `EX_SFU) begin
|
inuse_units[ibuffer_if[i].data.wis][ibuffer_if[i].data.rd] <= ibuffer_if[i].data.ex_type;
|
||||||
inuse_sfu[scoreboard_if[i].data.wis][scoreboard_if[i].data.rd] <= sfu_type;
|
if (ibuffer_if[i].data.ex_type == `EX_SFU) begin
|
||||||
|
inuse_sfu[ibuffer_if[i].data.wis][ibuffer_if[i].data.rd] <= sfu_type;
|
||||||
|
end
|
||||||
end
|
end
|
||||||
`endif
|
`endif
|
||||||
end
|
end
|
||||||
valid_out_r <= 0;
|
|
||||||
end
|
|
||||||
end
|
|
||||||
if (~valid_out_r) begin
|
|
||||||
data_out_r <= data_in;
|
|
||||||
end
|
|
||||||
end
|
|
||||||
|
|
||||||
assign ibuffer_if[i].ready = ready_in;
|
|
||||||
assign scoreboard_if[i].valid = valid_out_r;
|
|
||||||
assign scoreboard_if[i].data = data_out_r;
|
|
||||||
|
|
||||||
`ifdef SIMULATION
|
`ifdef SIMULATION
|
||||||
reg [31:0] timeout_ctr;
|
reg [31:0] timeout_ctr;
|
||||||
@@ -208,7 +204,7 @@ module VX_scoreboard import VX_gpu_pkg::*; #(
|
|||||||
`ifdef DBG_TRACE_CORE_PIPELINE
|
`ifdef DBG_TRACE_CORE_PIPELINE
|
||||||
`TRACE(3, ("%d: *** core%0d-scoreboard-stall: wid=%0d, PC=0x%0h, tmask=%b, cycles=%0d, inuse=%b (#%0d)\n",
|
`TRACE(3, ("%d: *** core%0d-scoreboard-stall: wid=%0d, PC=0x%0h, tmask=%b, cycles=%0d, inuse=%b (#%0d)\n",
|
||||||
$time, CORE_ID, wis_to_wid(ibuffer_if[i].data.wis, i), ibuffer_if[i].data.PC, ibuffer_if[i].data.tmask, timeout_ctr,
|
$time, CORE_ID, wis_to_wid(ibuffer_if[i].data.wis, i), ibuffer_if[i].data.PC, ibuffer_if[i].data.tmask, timeout_ctr,
|
||||||
~ready_masks, ibuffer_if[i].data.uuid));
|
operands_busy, ibuffer_if[i].data.uuid));
|
||||||
`endif
|
`endif
|
||||||
timeout_ctr <= timeout_ctr + 1;
|
timeout_ctr <= timeout_ctr + 1;
|
||||||
end else if (ibuffer_if[i].valid && ibuffer_if[i].ready) begin
|
end else if (ibuffer_if[i].valid && ibuffer_if[i].ready) begin
|
||||||
@@ -220,7 +216,7 @@ module VX_scoreboard import VX_gpu_pkg::*; #(
|
|||||||
`RUNTIME_ASSERT((timeout_ctr < `STALL_TIMEOUT),
|
`RUNTIME_ASSERT((timeout_ctr < `STALL_TIMEOUT),
|
||||||
("%t: *** core%0d-scoreboard-timeout: wid=%0d, PC=0x%0h, tmask=%b, cycles=%0d, inuse=%b (#%0d)",
|
("%t: *** core%0d-scoreboard-timeout: wid=%0d, PC=0x%0h, tmask=%b, cycles=%0d, inuse=%b (#%0d)",
|
||||||
$time, CORE_ID, wis_to_wid(ibuffer_if[i].data.wis, i), ibuffer_if[i].data.PC, ibuffer_if[i].data.tmask, timeout_ctr,
|
$time, CORE_ID, wis_to_wid(ibuffer_if[i].data.wis, i), ibuffer_if[i].data.PC, ibuffer_if[i].data.tmask, timeout_ctr,
|
||||||
~ready_masks, ibuffer_if[i].data.uuid));
|
operands_busy, ibuffer_if[i].data.uuid));
|
||||||
|
|
||||||
`RUNTIME_ASSERT(~writeback_fire || inuse_regs[writeback_if[i].data.wis][writeback_if[i].data.rd] != 0,
|
`RUNTIME_ASSERT(~writeback_fire || inuse_regs[writeback_if[i].data.wis][writeback_if[i].data.rd] != 0,
|
||||||
("%t: *** core%0d: invalid writeback register: wid=%0d, PC=0x%0h, tmask=%b, rd=%0d (#%0d)",
|
("%t: *** core%0d: invalid writeback register: wid=%0d, PC=0x%0h, tmask=%b, rd=%0d (#%0d)",
|
||||||
|
|||||||
Reference in New Issue
Block a user