Merge branch 'fpga_synthesis' of https://github.gatech.edu/casl/Vortex into fpga_synthesis
t # the commit.
This commit is contained in:
@@ -18,7 +18,7 @@ DBG_PRINT_FLAGS = -DDBG_PRINT_CORE_ICACHE \
|
||||
#MULTICORE += -DNUM_CLUSTERS=1 -DNUM_CORES=4
|
||||
MULTICORE += -DNUM_CLUSTERS=1 -DNUM_CORES=2
|
||||
|
||||
#DEBUG = 1
|
||||
#DEBUG=1
|
||||
#AFU=1
|
||||
|
||||
CFLAGS += -fPIC
|
||||
|
||||
@@ -138,8 +138,11 @@ public:
|
||||
int flush_caches(size_t dev_maddr, size_t size) {
|
||||
if (future_.valid()) {
|
||||
future_.wait(); // ensure prior run completed
|
||||
}
|
||||
simulator_.flush_caches(dev_maddr, size);
|
||||
}
|
||||
simulator_.flush_caches(dev_maddr, size);
|
||||
while (simulator_.is_busy()) {
|
||||
simulator_.step();
|
||||
};
|
||||
return 0;
|
||||
}
|
||||
|
||||
|
||||
@@ -172,7 +172,7 @@
|
||||
`define DNUM_REQUESTS `NUM_THREADS
|
||||
|
||||
// Snoop request tag bits
|
||||
`define DSNP_TAG_WIDTH `LOG2UP(`L2SNRQ_SIZE)
|
||||
`define DSNP_TAG_WIDTH ((`NUM_CORES > 1) ? `LOG2UP(`L2SNRQ_SIZE) : `L2SNP_TAG_WIDTH)
|
||||
|
||||
////////////////////////// Icache Configurable Knobs //////////////////////////
|
||||
|
||||
|
||||
@@ -392,6 +392,7 @@ module Vortex_Cluster #(
|
||||
assign per_core_snp_rsp_ready [(i/2)] = arb_snp_fwdin_ready [(i/2)];
|
||||
end
|
||||
|
||||
if (`NUM_CORES > 1) begin
|
||||
VX_snp_forwarder #(
|
||||
.CACHE_ID (`L2CACHE_ID),
|
||||
.BANK_LINE_SIZE (`L2BANK_LINE_SIZE),
|
||||
@@ -421,6 +422,16 @@ module Vortex_Cluster #(
|
||||
.snp_fwdin_tag (arb_snp_fwdin_tag),
|
||||
.snp_fwdin_ready (arb_snp_fwdin_ready)
|
||||
);
|
||||
end else begin
|
||||
assign arb_snp_fwdout_valid = snp_req_valid;
|
||||
assign arb_snp_fwdout_addr = snp_req_addr;
|
||||
assign arb_snp_fwdout_tag = snp_req_tag;
|
||||
assign snp_req_ready = arb_snp_fwdout_ready;
|
||||
|
||||
assign snp_rsp_valid = arb_snp_fwdin_valid;
|
||||
assign snp_rsp_tag = arb_snp_fwdin_tag;
|
||||
assign arb_snp_fwdin_ready = snp_rsp_ready;
|
||||
end
|
||||
|
||||
VX_dram_arb #(
|
||||
.NUM_REQUESTS (`L2NUM_REQUESTS),
|
||||
|
||||
30
hw/rtl/cache/VX_bank.v
vendored
30
hw/rtl/cache/VX_bank.v
vendored
@@ -459,7 +459,7 @@ module VX_bank #(
|
||||
end
|
||||
)
|
||||
wire qual_valid_st1e_2 = valid_st1e && !is_fill_st1[STAGE_1_CYCLES-1];
|
||||
wire from_mrvq_st1e_st2 = from_mrvq_st1e && !is_snp_st1e;
|
||||
wire from_mrvq_st1e_st2 = from_mrvq_st1e;
|
||||
|
||||
wire valid_st2;
|
||||
wire [`UP(`WORD_SELECT_WIDTH)-1:0] wsel_st2;
|
||||
@@ -721,49 +721,49 @@ module VX_bank #(
|
||||
if (NUM_BANKS == 1) begin
|
||||
always_ff @(posedge clk) begin
|
||||
if (core_req_valid && core_req_ready) begin
|
||||
$display("%t: bank%01d%01d core req: addr=%0h, tag=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR0(core_req_addr), core_req_tag);
|
||||
$display("%t: bank%0d-%0d core req: addr=%0h, tag=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR0(core_req_addr), core_req_tag);
|
||||
end
|
||||
if (core_rsp_valid && core_rsp_ready) begin
|
||||
$display("%t: bank%01d%01d core rsp: tag=%0h, data=%0h", $time, CACHE_ID, BANK_ID, core_rsp_tag, core_rsp_data);
|
||||
$display("%t: bank%0d-%0d core rsp: tag=%0h, data=%0h", $time, CACHE_ID, BANK_ID, core_rsp_tag, core_rsp_data);
|
||||
end
|
||||
if (dram_fill_req_valid && dram_fill_req_ready) begin
|
||||
$display("%t: bank%01d%01d dram_fill req: addr=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR0(dram_fill_req_addr));
|
||||
$display("%t: bank%0d-%0d dram_fill req: addr=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR0(dram_fill_req_addr));
|
||||
end
|
||||
if (dram_wb_req_valid && dram_wb_req_ready) begin
|
||||
$display("%t: bank%01d%01d dram_wb req: addr=%0h, data=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR0(dram_wb_req_addr), dram_wb_req_data);
|
||||
$display("%t: bank%0d-%0d dram_wb req: addr=%0h, data=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR0(dram_wb_req_addr), dram_wb_req_data);
|
||||
end
|
||||
if (dram_fill_rsp_valid && dram_fill_rsp_ready) begin
|
||||
$display("%t: bank%01d%01d dram_fill rsp: addr=%0h, data=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR0(dram_fill_rsp_addr), dram_fill_rsp_data);
|
||||
$display("%t: bank%0d-%0d dram_fill rsp: addr=%0h, data=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR0(dram_fill_rsp_addr), dram_fill_rsp_data);
|
||||
end
|
||||
if (snp_req_valid && snp_req_ready) begin
|
||||
$display("%t: bank%01d%01d snp req: addr=%0h, tag=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR0(snp_req_addr), snp_req_tag);
|
||||
$display("%t: bank%0d-%0d snp req: addr=%0h, tag=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR0(snp_req_addr), snp_req_tag);
|
||||
end
|
||||
if (snp_rsp_valid && snp_rsp_ready) begin
|
||||
$display("%t: bank%01d%01d snp rsp: tag=%0h", $time, CACHE_ID, BANK_ID, snp_rsp_tag);
|
||||
$display("%t: bank%0d-%0d snp rsp: tag=%0h", $time, CACHE_ID, BANK_ID, snp_rsp_tag);
|
||||
end
|
||||
end
|
||||
end else begin
|
||||
always_ff @(posedge clk) begin
|
||||
if ((|core_req_valid) && core_req_ready) begin
|
||||
$display("%t: bank%01d%01d core req: addr=%0h, tag=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(core_req_addr, BANK_ID), core_req_tag);
|
||||
$display("%t: bank%0d-%0d core req: addr=%0h, tag=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(core_req_addr, BANK_ID), core_req_tag);
|
||||
end
|
||||
if (core_rsp_valid && core_rsp_ready) begin
|
||||
$display("%t: bank%01d%01d core rsp: tag=%0h, data=%0h", $time, CACHE_ID, BANK_ID, core_rsp_tag, core_rsp_data);
|
||||
$display("%t: bank%0d-%0d core rsp: tag=%0h, data=%0h", $time, CACHE_ID, BANK_ID, core_rsp_tag, core_rsp_data);
|
||||
end
|
||||
if (dram_fill_req_valid && dram_fill_req_ready) begin
|
||||
$display("%t: bank%01d%01d dram_fill req: addr=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(dram_fill_req_addr, BANK_ID));
|
||||
$display("%t: bank%0d-%0d dram_fill req: addr=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(dram_fill_req_addr, BANK_ID));
|
||||
end
|
||||
if (dram_wb_req_valid && dram_wb_req_ready) begin
|
||||
$display("%t: bank%01d%01d dram_wb req: addr=%0h, data=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(dram_wb_req_addr, BANK_ID), dram_wb_req_data);
|
||||
$display("%t: bank%0d-%0d dram_wb req: addr=%0h, data=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(dram_wb_req_addr, BANK_ID), dram_wb_req_data);
|
||||
end
|
||||
if (dram_fill_rsp_valid && dram_fill_rsp_ready) begin
|
||||
$display("%t: bank%01d%01d dram_fill rsp: addr=%0h, data=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(dram_fill_rsp_addr, BANK_ID), dram_fill_rsp_data);
|
||||
$display("%t: bank%0d-%0d dram_fill rsp: addr=%0h, data=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(dram_fill_rsp_addr, BANK_ID), dram_fill_rsp_data);
|
||||
end
|
||||
if (snp_req_valid && snp_req_ready) begin
|
||||
$display("%t: bank%01d%01d snp req: addr=%0h, tag=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(snp_req_addr, BANK_ID), snp_req_tag);
|
||||
$display("%t: bank%0d-%0d snp req: addr=%0h, tag=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(snp_req_addr, BANK_ID), snp_req_tag);
|
||||
end
|
||||
if (snp_rsp_valid && snp_rsp_ready) begin
|
||||
$display("%t: bank%01d%01d snp rsp: tag=%0h", $time, CACHE_ID, BANK_ID, snp_rsp_tag);
|
||||
$display("%t: bank%0d-%0d snp rsp: tag=%0h", $time, CACHE_ID, BANK_ID, snp_rsp_tag);
|
||||
end
|
||||
end
|
||||
end
|
||||
|
||||
22
hw/rtl/cache/VX_cache.v
vendored
22
hw/rtl/cache/VX_cache.v
vendored
@@ -208,13 +208,13 @@ module VX_cache #(
|
||||
assign snp_req_addr_qual = snp_req_addr;
|
||||
assign snp_req_tag_qual = snp_req_tag;
|
||||
assign snp_req_ready = snp_req_ready_qual;
|
||||
end
|
||||
end
|
||||
|
||||
assign dram_req_tag = dram_req_addr;
|
||||
|
||||
assign core_req_ready = (& per_bank_core_req_ready);
|
||||
assign dram_rsp_ready = (| per_bank_dram_fill_rsp_ready);
|
||||
assign snp_req_ready_qual = (& per_bank_snp_req_ready);
|
||||
if (NUM_BANKS == 1) begin
|
||||
assign snp_req_ready_qual = per_bank_snp_req_ready;
|
||||
end else begin
|
||||
assign snp_req_ready_qual = per_bank_snp_req_ready[`DRAM_ADDR_BANK(snp_req_addr_qual)];
|
||||
end
|
||||
|
||||
VX_cache_core_req_bank_sel #(
|
||||
.BANK_LINE_SIZE (BANK_LINE_SIZE),
|
||||
@@ -223,11 +223,17 @@ module VX_cache #(
|
||||
.NUM_REQUESTS (NUM_REQUESTS)
|
||||
) cache_core_req_bank_sel (
|
||||
.core_req_valid (core_req_valid),
|
||||
.per_bank_ready (per_bank_core_req_ready),
|
||||
.core_req_addr (core_req_addr),
|
||||
.per_bank_valid (per_bank_valid)
|
||||
.per_bank_valid (per_bank_valid),
|
||||
.core_req_ready (core_req_ready)
|
||||
);
|
||||
|
||||
assign dram_req_tag = dram_req_addr;
|
||||
assign dram_rsp_ready = (| per_bank_dram_fill_rsp_ready);
|
||||
|
||||
genvar i;
|
||||
|
||||
generate
|
||||
for (i = 0; i < NUM_BANKS; i++) begin
|
||||
wire [NUM_REQUESTS-1:0] curr_bank_core_req_valid;
|
||||
@@ -270,7 +276,7 @@ module VX_cache #(
|
||||
wire curr_bank_core_req_ready;
|
||||
|
||||
// Core Req
|
||||
assign curr_bank_core_req_valid = per_bank_valid[i] & {NUM_REQUESTS{core_req_ready}};
|
||||
assign curr_bank_core_req_valid = (per_bank_valid[i] & {NUM_REQUESTS{core_req_ready}});
|
||||
assign curr_bank_core_req_addr = core_req_addr;
|
||||
assign curr_bank_core_req_rw = core_req_rw;
|
||||
assign curr_bank_core_req_byteen = core_req_byteen;
|
||||
|
||||
21
hw/rtl/cache/VX_cache_core_req_bank_sel.v
vendored
21
hw/rtl/cache/VX_cache_core_req_bank_sel.v
vendored
@@ -11,28 +11,35 @@ module VX_cache_core_req_bank_sel #(
|
||||
// Number of Word requests per cycle {1, 2, 4, 8, ...}
|
||||
parameter NUM_REQUESTS = 0
|
||||
) (
|
||||
input wire [NUM_REQUESTS-1:0] core_req_valid,
|
||||
input wire [NUM_REQUESTS-1:0] core_req_valid,
|
||||
`IGNORE_WARNINGS_BEGIN
|
||||
input wire [NUM_REQUESTS-1:0][`WORD_ADDR_WIDTH-1:0] core_req_addr,
|
||||
input wire [NUM_REQUESTS-1:0][`WORD_ADDR_WIDTH-1:0] core_req_addr,
|
||||
`IGNORE_WARNINGS_END
|
||||
output reg [NUM_BANKS-1:0][NUM_REQUESTS-1:0] per_bank_valid
|
||||
input wire [NUM_BANKS-1:0] per_bank_ready,
|
||||
output reg [NUM_BANKS-1:0][NUM_REQUESTS-1:0] per_bank_valid,
|
||||
output wire core_req_ready
|
||||
);
|
||||
integer i;
|
||||
|
||||
if (NUM_BANKS == 1) begin
|
||||
always @(*) begin
|
||||
always @(*) begin
|
||||
per_bank_valid = 0;
|
||||
for (i = 0; i < NUM_REQUESTS; i++) begin
|
||||
per_bank_valid[0][i] = core_req_valid[i];
|
||||
end
|
||||
end
|
||||
end else begin
|
||||
end
|
||||
assign core_req_ready = per_bank_ready;
|
||||
end else begin
|
||||
reg [NUM_BANKS-1:0] per_bank_ready_sel;
|
||||
always @(*) begin
|
||||
per_bank_valid = 0;
|
||||
per_bank_ready_sel = {NUM_BANKS{1'b1}};
|
||||
for (i = 0; i < NUM_REQUESTS; i++) begin
|
||||
per_bank_valid[core_req_addr[i][`BANK_SELECT_ADDR_RNG]][i] = core_req_valid[i];
|
||||
per_bank_ready_sel[core_req_addr[i][`BANK_SELECT_ADDR_RNG]] = 0;
|
||||
end
|
||||
end
|
||||
end
|
||||
assign core_req_ready = & (per_bank_ready | per_bank_ready_sel);
|
||||
end
|
||||
|
||||
endmodule
|
||||
25
hw/rtl/cache/VX_cache_miss_resrv.v
vendored
25
hw/rtl/cache/VX_cache_miss_resrv.v
vendored
@@ -72,10 +72,6 @@ module VX_cache_miss_resrv #(
|
||||
wire enqueue_possible = !miss_resrv_full;
|
||||
wire [`LOG2UP(MRVQ_SIZE)-1:0] enqueue_index = tail_ptr;
|
||||
|
||||
`IGNORE_WARNINGS_BEGIN
|
||||
wire [31:0] make_ready_push_full;
|
||||
`IGNORE_WARNINGS_END
|
||||
|
||||
reg [MRVQ_SIZE-1:0] make_ready;
|
||||
reg [MRVQ_SIZE-1:0] make_ready_push;
|
||||
reg [MRVQ_SIZE-1:0] valid_address_match;
|
||||
@@ -93,24 +89,21 @@ module VX_cache_miss_resrv #(
|
||||
wire dequeue_possible = valid_table[schedule_ptr] && ready_table[schedule_ptr];
|
||||
wire [`LOG2UP(MRVQ_SIZE)-1:0] dequeue_index = schedule_ptr;
|
||||
|
||||
assign miss_resrv_valid_st0 = (MRVQ_SIZE != 2) && dequeue_possible;
|
||||
assign miss_resrv_valid_st0 = dequeue_possible;
|
||||
assign miss_resrv_addr_st0 = addr_table[dequeue_index];
|
||||
assign {miss_resrv_data_st0, miss_resrv_tid_st0, miss_resrv_tag_st0, miss_resrv_rw_st0, miss_resrv_byteen_st0, miss_resrv_wsel_st0, miss_resrv_is_snp_st0} = metadata_table[dequeue_index];
|
||||
|
||||
wire mrvq_push = miss_add && enqueue_possible && !from_mrvq && (MRVQ_SIZE != 2);
|
||||
wire mrvq_push = miss_add && enqueue_possible && !from_mrvq;
|
||||
wire mrvq_pop = miss_resrv_pop && dequeue_possible;
|
||||
|
||||
|
||||
wire recover_state = miss_add && from_mrvq;
|
||||
wire increment_head = !miss_add && from_mrvq;
|
||||
|
||||
|
||||
wire update_ready = (|make_ready);
|
||||
|
||||
wire qual_mrvq_init = mrvq_push && mrvq_init_ready_state;
|
||||
|
||||
assign make_ready_push_full = ({31'b0, qual_mrvq_init} << enqueue_index);
|
||||
assign make_ready_push = make_ready_push_full[MRVQ_SIZE-1:0];
|
||||
assign make_ready_push = (MRVQ_SIZE'(qual_mrvq_init)) << enqueue_index;
|
||||
|
||||
always @(posedge clk) begin
|
||||
if (reset) begin
|
||||
@@ -160,12 +153,12 @@ module VX_cache_miss_resrv #(
|
||||
integer j;
|
||||
if (NUM_BANKS == 1) begin
|
||||
always_ff @(posedge clk) begin
|
||||
if (mrvq_push || mrvq_pop) begin
|
||||
$write("%t: bank%02d:%01d msrq: push=%b pop=%b", $time, CACHE_ID, BANK_ID, mrvq_push, mrvq_pop);
|
||||
if (mrvq_push || mrvq_pop || increment_head || recover_state) begin
|
||||
$write("%t: bank%0d-%0d msrq: push=%b pop=%b incr=%d recv=%d", $time, CACHE_ID, BANK_ID, mrvq_push, mrvq_pop, increment_head, recover_state);
|
||||
for (j = 0; j < MRVQ_SIZE; j++) begin
|
||||
if (valid_table[j]) begin
|
||||
$write(" ");
|
||||
if (schedule_ptr == $bits(schedule_ptr)'(j)) $write("*");
|
||||
if (schedule_ptr == $bits(schedule_ptr)'(j)) $write("*");
|
||||
if (~ready_table[j]) $write("!");
|
||||
$write("addr%0d=%0h", j, {addr_table[j], `BASE_ADDR_BITS'(0)});
|
||||
end
|
||||
@@ -175,12 +168,12 @@ module VX_cache_miss_resrv #(
|
||||
end
|
||||
end else begin
|
||||
always_ff @(posedge clk) begin
|
||||
if (mrvq_push || mrvq_pop) begin
|
||||
$write("%t: bank%02d:%01d msrq: push=%b pop=%b", $time, CACHE_ID, BANK_ID, mrvq_push, mrvq_pop);
|
||||
if (mrvq_push || mrvq_pop || increment_head || recover_state) begin
|
||||
$write("%t: bank%0d-%0d msrq: push=%b pop=%b incr=%d recv=%d", $time, CACHE_ID, BANK_ID, mrvq_push, mrvq_pop, increment_head, recover_state);
|
||||
for (j = 0; j < MRVQ_SIZE; j++) begin
|
||||
if (valid_table[j]) begin
|
||||
$write(" ");
|
||||
if (schedule_ptr == $bits(schedule_ptr)'(j)) $write("*");
|
||||
if (schedule_ptr == $bits(schedule_ptr)'(j)) $write("*");
|
||||
if (~ready_table[j]) $write("!");
|
||||
$write("addr%0d=%0h", j, `LINE_TO_BYTE_ADDR(addr_table[j], BANK_ID));
|
||||
end
|
||||
|
||||
18
hw/rtl/cache/VX_snp_forwarder.v
vendored
18
hw/rtl/cache/VX_snp_forwarder.v
vendored
@@ -34,6 +34,8 @@ module VX_snp_forwarder #(
|
||||
input wire [NUM_REQUESTS-1:0][`LOG2UP(SNRQ_SIZE)-1:0] snp_fwdin_tag,
|
||||
output wire [NUM_REQUESTS-1:0] snp_fwdin_ready
|
||||
);
|
||||
`STATIC_ASSERT(NUM_REQUESTS > 1, "invalid value");
|
||||
|
||||
reg [`REQS_BITS:0] pending_cntrs [SNRQ_SIZE-1:0];
|
||||
reg [`REQS_BITS-1:0] fwdin_sel;
|
||||
|
||||
@@ -43,12 +45,12 @@ module VX_snp_forwarder #(
|
||||
wire fwdin_valid;
|
||||
wire [`LOG2UP(SNRQ_SIZE)-1:0] fwdin_tag;
|
||||
|
||||
wire fwdin_ready = snp_rsp_ready;
|
||||
wire fwdin_taken = fwdin_valid && fwdin_ready;
|
||||
wire fwdin_ready = snp_rsp_ready || (1 != pending_cntrs[sfq_read_addr]);
|
||||
wire fwdin_fire = fwdin_valid && fwdin_ready;
|
||||
|
||||
wire fwdout_ready = (& snp_fwdout_ready);
|
||||
|
||||
assign snp_rsp_valid = fwdin_taken && (1 == pending_cntrs[sfq_read_addr]); // send response
|
||||
assign snp_rsp_valid = fwdin_valid && (1 == pending_cntrs[sfq_read_addr]); // send response
|
||||
|
||||
assign sfq_read_addr = fwdin_tag;
|
||||
|
||||
@@ -77,7 +79,7 @@ module VX_snp_forwarder #(
|
||||
if (sfq_push) begin
|
||||
pending_cntrs[sfq_write_addr] <= NUM_REQUESTS;
|
||||
end
|
||||
if (fwdin_taken) begin
|
||||
if (fwdin_fire) begin
|
||||
pending_cntrs[sfq_read_addr] <= pending_cntrs[sfq_read_addr] - 1;
|
||||
assert(sfq_read_addr == dbg_sfq_write_addr);
|
||||
end
|
||||
@@ -112,16 +114,16 @@ module VX_snp_forwarder #(
|
||||
`ifdef DBG_PRINT_CACHE_SNP
|
||||
always_ff @(posedge clk) begin
|
||||
if (snp_req_valid && snp_req_ready) begin
|
||||
$display("%t: cache%01d snp req: addr=%0h, tag=%0h", $time, CACHE_ID, `DRAM_TO_BYTE_ADDR(snp_req_addr), snp_req_tag);
|
||||
$display("%t: cache%0d snp req: addr=%0h, tag=%0h", $time, CACHE_ID, `DRAM_TO_BYTE_ADDR(snp_req_addr), snp_req_tag);
|
||||
end
|
||||
if (snp_fwdout_valid[0] && snp_fwdout_ready[0]) begin
|
||||
$display("%t: cache%01d snp fwd_out: addr=%0h, tag=%0h", $time, CACHE_ID, `DRAM_TO_BYTE_ADDR(snp_fwdout_addr[0]), snp_fwdout_tag[0]);
|
||||
$display("%t: cache%0d snp fwd_out: addr=%0h, tag=%0h", $time, CACHE_ID, `DRAM_TO_BYTE_ADDR(snp_fwdout_addr[0]), snp_fwdout_tag[0]);
|
||||
end
|
||||
if (fwdin_valid && fwdin_ready) begin
|
||||
$display("%t: cache%01d snp fwd_in[%01d]: tag=%0h", $time, CACHE_ID, fwdin_sel, fwdin_tag);
|
||||
$display("%t: cache%0d snp fwd_in[%01d]: tag=%0h", $time, CACHE_ID, fwdin_sel, fwdin_tag);
|
||||
end
|
||||
if (snp_rsp_valid && snp_rsp_ready) begin
|
||||
$display("%t: cache%01d snp rsp: addr=%0h, tag=%0h", $time, CACHE_ID, snp_rsp_addr, snp_rsp_tag);
|
||||
$display("%t: cache%0d snp rsp: addr=%0h, tag=%0h", $time, CACHE_ID, snp_rsp_addr, snp_rsp_tag);
|
||||
end
|
||||
end
|
||||
`endif
|
||||
|
||||
@@ -11,11 +11,13 @@ double sc_time_stamp() {
|
||||
|
||||
Simulator::Simulator() {
|
||||
// force random values for unitialized signals
|
||||
Verilated::randReset(2);
|
||||
Verilated::randReset(1);
|
||||
|
||||
ram_ = nullptr;
|
||||
vortex_ = new VVortex_Socket();
|
||||
|
||||
snp_req_active_ = false;
|
||||
|
||||
#ifdef VCD_OUTPUT
|
||||
Verilated::traceEverOn(true);
|
||||
trace_ = new VerilatedVcdC;
|
||||
@@ -36,6 +38,214 @@ void Simulator::attach_ram(RAM* ram) {
|
||||
dram_rsp_vec_.clear();
|
||||
}
|
||||
|
||||
void Simulator::reset() {
|
||||
#ifndef NDEBUG
|
||||
std::cout << timestamp << ": [sim] reset()" << std::endl;
|
||||
#endif
|
||||
|
||||
vortex_->reset = 1;
|
||||
this->step();
|
||||
vortex_->reset = 0;
|
||||
|
||||
dram_rsp_vec_.clear();
|
||||
}
|
||||
|
||||
void Simulator::step() {
|
||||
vortex_->clk = 0;
|
||||
this->eval();
|
||||
|
||||
vortex_->clk = 1;
|
||||
this->eval();
|
||||
|
||||
this->eval_dram_bus();
|
||||
this->eval_io_bus();
|
||||
this->eval_snp_bus();
|
||||
}
|
||||
|
||||
void Simulator::eval() {
|
||||
vortex_->eval();
|
||||
#ifdef VCD_OUTPUT
|
||||
trace_->dump(timestamp);
|
||||
#endif
|
||||
++timestamp;
|
||||
}
|
||||
|
||||
void Simulator::eval_dram_bus() {
|
||||
if (ram_ == nullptr) {
|
||||
vortex_->dram_req_ready = 0;
|
||||
return;
|
||||
}
|
||||
|
||||
// handle DRAM response cycle
|
||||
int dequeue_index = -1;
|
||||
for (int i = 0; i < dram_rsp_vec_.size(); i++) {
|
||||
if (dram_rsp_vec_[i].cycles_left > 0) {
|
||||
dram_rsp_vec_[i].cycles_left -= 1;
|
||||
}
|
||||
if ((dequeue_index == -1)
|
||||
&& (dram_rsp_vec_[i].cycles_left == 0)) {
|
||||
dequeue_index = i;
|
||||
}
|
||||
}
|
||||
|
||||
// handle DRAM response message
|
||||
if ((dequeue_index != -1)
|
||||
&& vortex_->dram_rsp_ready) {
|
||||
vortex_->dram_rsp_valid = 1;
|
||||
memcpy((uint8_t*)vortex_->dram_rsp_data, dram_rsp_vec_[dequeue_index].data, GLOBAL_BLOCK_SIZE);
|
||||
vortex_->dram_rsp_tag = dram_rsp_vec_[dequeue_index].tag;
|
||||
free(dram_rsp_vec_[dequeue_index].data);
|
||||
dram_rsp_vec_.erase(dram_rsp_vec_.begin() + dequeue_index);
|
||||
} else {
|
||||
vortex_->dram_rsp_valid = 0;
|
||||
}
|
||||
|
||||
// handle DRAM stalls
|
||||
bool dram_stalled = false;
|
||||
#ifdef ENABLE_DRAM_STALLS
|
||||
if (0 == ((timestamp/2) % DRAM_STALLS_MODULO)) {
|
||||
dram_stalled = true;
|
||||
} else
|
||||
if (dram_rsp_vec_.size() >= DRAM_RQ_SIZE) {
|
||||
dram_stalled = true;
|
||||
}
|
||||
#endif
|
||||
|
||||
// handle DRAM requests
|
||||
if (!dram_stalled) {
|
||||
if (vortex_->dram_req_valid) {
|
||||
if (vortex_->dram_req_rw) {
|
||||
uint64_t byteen = vortex_->dram_req_byteen;
|
||||
unsigned base_addr = (vortex_->dram_req_addr * GLOBAL_BLOCK_SIZE);
|
||||
uint8_t* data = (uint8_t*)(vortex_->dram_req_data);
|
||||
for (int i = 0; i < GLOBAL_BLOCK_SIZE; i++) {
|
||||
if ((byteen >> i) & 0x1) {
|
||||
(*ram_)[base_addr + i] = data[i];
|
||||
}
|
||||
}
|
||||
} else {
|
||||
dram_req_t dram_req;
|
||||
dram_req.cycles_left = DRAM_LATENCY;
|
||||
dram_req.data = (uint8_t*)malloc(GLOBAL_BLOCK_SIZE);
|
||||
dram_req.tag = vortex_->dram_req_tag;
|
||||
ram_->read(vortex_->dram_req_addr * GLOBAL_BLOCK_SIZE, GLOBAL_BLOCK_SIZE, dram_req.data);
|
||||
dram_rsp_vec_.push_back(dram_req);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
vortex_->dram_req_ready = ~dram_stalled;
|
||||
}
|
||||
|
||||
void Simulator::eval_io_bus() {
|
||||
if (vortex_->io_req_valid
|
||||
&& vortex_->io_req_rw
|
||||
&& ((vortex_->io_req_addr << 2) == IO_BUS_ADDR_COUT)) {
|
||||
uint32_t data_write = (uint32_t)vortex_->io_req_data;
|
||||
char c = (char)data_write;
|
||||
std::cout << c;
|
||||
}
|
||||
vortex_->io_req_ready = 1;
|
||||
vortex_->io_rsp_valid = 0;
|
||||
}
|
||||
|
||||
void Simulator::eval_snp_bus() {
|
||||
if (snp_req_active_) {
|
||||
if (vortex_->snp_rsp_valid) {
|
||||
assert(pending_snp_reqs_ > 0);
|
||||
--pending_snp_reqs_;
|
||||
#ifdef DBG_PRINT_CACHE_SNP
|
||||
std::cout << timestamp << ": [sim] snp rsp: tag=" << vortex_->snp_rsp_tag << " pending=" << pending_snp_reqs_ << std::endl;
|
||||
#endif
|
||||
}
|
||||
if (vortex_->snp_req_valid && vortex_->snp_req_ready) {
|
||||
if (snp_req_size_) {
|
||||
vortex_->snp_req_addr += 1;
|
||||
vortex_->snp_req_tag += 1;
|
||||
--snp_req_size_;
|
||||
++pending_snp_reqs_;
|
||||
#ifdef DBG_PRINT_CACHE_SNP
|
||||
std::cout << timestamp << ": [sim] snp req: addr=" << std::hex << vortex_->snp_req_addr << std::dec << " tag=" << vortex_->snp_req_tag << " remain=" << snp_req_size_ << std::endl;
|
||||
#endif
|
||||
} else {
|
||||
vortex_->snp_req_valid = 0;
|
||||
}
|
||||
}
|
||||
if (!vortex_->snp_req_valid
|
||||
&& 0 == pending_snp_reqs_) {
|
||||
snp_req_active_ = false;
|
||||
}
|
||||
} else {
|
||||
vortex_->snp_req_valid = 0;
|
||||
vortex_->snp_rsp_ready = 0;
|
||||
}
|
||||
}
|
||||
|
||||
void Simulator::wait(uint32_t cycles) {
|
||||
for (int i = 0; i < cycles; ++i) {
|
||||
this->step();
|
||||
}
|
||||
}
|
||||
|
||||
bool Simulator::is_busy() {
|
||||
return vortex_->busy || snp_req_active_;
|
||||
}
|
||||
|
||||
void Simulator::flush_caches(uint32_t mem_addr, uint32_t size) {
|
||||
#ifndef NDEBUG
|
||||
std::cout << timestamp << ": [sim] flush_caches()" << std::endl;
|
||||
#endif
|
||||
if (0 == size)
|
||||
return;
|
||||
|
||||
snp_req_active_ = true;
|
||||
snp_req_size_ = (size + GLOBAL_BLOCK_SIZE - 1) / GLOBAL_BLOCK_SIZE;
|
||||
|
||||
vortex_->snp_req_addr = mem_addr / GLOBAL_BLOCK_SIZE;
|
||||
vortex_->snp_req_tag = 0;
|
||||
vortex_->snp_req_valid = 1;
|
||||
vortex_->snp_rsp_ready = 1;
|
||||
|
||||
--snp_req_size_;
|
||||
pending_snp_reqs_ = 1;
|
||||
|
||||
#ifdef DBG_PRINT_CACHE_SNP
|
||||
std::cout << timestamp << ": [sim] snp req: addr=" << std::hex << vortex_->snp_req_addr << std::dec << " tag=" << vortex_->snp_req_tag << " remain=" << snp_req_size_ << std::endl;
|
||||
#endif
|
||||
}
|
||||
|
||||
bool Simulator::run() {
|
||||
#ifndef NDEBUG
|
||||
std::cout << timestamp << ": [sim] run()" << std::endl;
|
||||
#endif
|
||||
|
||||
// reset the device
|
||||
this->reset();
|
||||
|
||||
// execute program
|
||||
while (vortex_->busy
|
||||
&& !vortex_->ebreak) {
|
||||
this->step();
|
||||
}
|
||||
|
||||
// wait 5 cycles to flush the pipeline
|
||||
this->wait(5);
|
||||
|
||||
// check riscv-tests PASSED/FAILED status
|
||||
#if (NUM_CLUSTERS == 1 && NUM_CORES == 1)
|
||||
int status = (int)vortex_->Vortex_Socket->genblk1__DOT__Vortex_Cluster->genblk1__BRA__0__KET____DOT__vortex_core->back_end->writeback->last_data_wb & 0xf;
|
||||
#else
|
||||
#if (NUM_CLUSTERS == 1)
|
||||
int status = (int)vortex_->Vortex_Socket->genblk1__DOT__Vortex_Cluster->genblk1__BRA__0__KET____DOT__vortex_core->back_end->writeback->last_data_wb & 0xf;
|
||||
#else
|
||||
int status = (int)vortex_->Vortex_Socket->genblk2__DOT__genblk1__BRA__0__KET____DOT__Vortex_Cluster->genblk1__BRA__0__KET____DOT__vortex_core->back_end->writeback->last_data_wb & 0xf;
|
||||
#endif
|
||||
#endif
|
||||
|
||||
return (status == 1);
|
||||
}
|
||||
|
||||
|
||||
void Simulator::load_bin(const char* program_file) {
|
||||
if (ram_ == nullptr)
|
||||
return;
|
||||
@@ -123,202 +333,4 @@ void Simulator::load_ihex(const char* program_file) {
|
||||
void Simulator::print_stats(std::ostream& out) {
|
||||
out << std::left;
|
||||
out << std::setw(24) << "# of total cycles:" << std::dec << timestamp/2 << std::endl;
|
||||
}
|
||||
|
||||
void Simulator::dbus_driver() {
|
||||
if (ram_ == nullptr) {
|
||||
vortex_->dram_req_ready = 0;
|
||||
return;
|
||||
}
|
||||
|
||||
// handle DRAM response cycle
|
||||
int dequeue_index = -1;
|
||||
for (int i = 0; i < dram_rsp_vec_.size(); i++) {
|
||||
if (dram_rsp_vec_[i].cycles_left > 0) {
|
||||
dram_rsp_vec_[i].cycles_left -= 1;
|
||||
}
|
||||
if ((dequeue_index == -1)
|
||||
&& (dram_rsp_vec_[i].cycles_left == 0)) {
|
||||
dequeue_index = i;
|
||||
}
|
||||
}
|
||||
|
||||
// handle DRAM response message
|
||||
if ((dequeue_index != -1)
|
||||
&& vortex_->dram_rsp_ready) {
|
||||
vortex_->dram_rsp_valid = 1;
|
||||
memcpy((uint8_t*)vortex_->dram_rsp_data, dram_rsp_vec_[dequeue_index].data, GLOBAL_BLOCK_SIZE);
|
||||
vortex_->dram_rsp_tag = dram_rsp_vec_[dequeue_index].tag;
|
||||
free(dram_rsp_vec_[dequeue_index].data);
|
||||
dram_rsp_vec_.erase(dram_rsp_vec_.begin() + dequeue_index);
|
||||
} else {
|
||||
vortex_->dram_rsp_valid = 0;
|
||||
}
|
||||
|
||||
// handle DRAM stalls
|
||||
bool dram_stalled = false;
|
||||
#ifdef ENABLE_DRAM_STALLS
|
||||
if (0 == ((timestamp/2) % DRAM_STALLS_MODULO)) {
|
||||
dram_stalled = true;
|
||||
} else
|
||||
if (dram_rsp_vec_.size() >= DRAM_RQ_SIZE) {
|
||||
dram_stalled = true;
|
||||
}
|
||||
#endif
|
||||
|
||||
// handle DRAM requests
|
||||
if (!dram_stalled) {
|
||||
if (vortex_->dram_req_valid) {
|
||||
if (vortex_->dram_req_rw) {
|
||||
uint64_t byteen = vortex_->dram_req_byteen;
|
||||
unsigned base_addr = (vortex_->dram_req_addr * GLOBAL_BLOCK_SIZE);
|
||||
uint8_t* data = (uint8_t*)(vortex_->dram_req_data);
|
||||
for (int i = 0; i < GLOBAL_BLOCK_SIZE; i++) {
|
||||
if ((byteen >> i) & 0x1) {
|
||||
(*ram_)[base_addr + i] = data[i];
|
||||
}
|
||||
}
|
||||
} else {
|
||||
dram_req_t dram_req;
|
||||
dram_req.cycles_left = DRAM_LATENCY;
|
||||
dram_req.data = (uint8_t*)malloc(GLOBAL_BLOCK_SIZE);
|
||||
dram_req.tag = vortex_->dram_req_tag;
|
||||
ram_->read(vortex_->dram_req_addr * GLOBAL_BLOCK_SIZE, GLOBAL_BLOCK_SIZE, dram_req.data);
|
||||
dram_rsp_vec_.push_back(dram_req);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
vortex_->dram_req_ready = ~dram_stalled;
|
||||
}
|
||||
|
||||
void Simulator::io_driver() {
|
||||
if (vortex_->io_req_valid
|
||||
&& vortex_->io_req_rw
|
||||
&& ((vortex_->io_req_addr << 2) == IO_BUS_ADDR_COUT)) {
|
||||
uint32_t data_write = (uint32_t)vortex_->io_req_data;
|
||||
char c = (char)data_write;
|
||||
std::cout << c;
|
||||
}
|
||||
vortex_->io_req_ready = 1;
|
||||
vortex_->io_rsp_valid = 0;
|
||||
}
|
||||
|
||||
void Simulator::reset() {
|
||||
#ifndef NDEBUG
|
||||
std::cout << timestamp << ": [sim] reset()" << std::endl;
|
||||
#endif
|
||||
vortex_->reset = 1;
|
||||
this->step();
|
||||
vortex_->reset = 0;
|
||||
|
||||
dram_rsp_vec_.clear();
|
||||
}
|
||||
|
||||
void Simulator::step() {
|
||||
vortex_->clk = 0;
|
||||
this->eval();
|
||||
|
||||
vortex_->clk = 1;
|
||||
this->eval();
|
||||
|
||||
dbus_driver();
|
||||
io_driver();
|
||||
}
|
||||
|
||||
void Simulator::eval() {
|
||||
vortex_->eval();
|
||||
#ifdef VCD_OUTPUT
|
||||
trace_->dump(timestamp);
|
||||
#endif
|
||||
++timestamp;
|
||||
}
|
||||
|
||||
void Simulator::wait(uint32_t cycles) {
|
||||
for (int i = 0; i < cycles; ++i) {
|
||||
this->step();
|
||||
}
|
||||
}
|
||||
|
||||
bool Simulator::is_busy() {
|
||||
return vortex_->busy;
|
||||
}
|
||||
|
||||
void Simulator::flush_caches(uint32_t mem_addr, uint32_t size) {
|
||||
#ifndef NDEBUG
|
||||
std::cout << timestamp << ": [sim] flush_caches()" << std::endl;
|
||||
#endif
|
||||
// align address to LLC block boundaries
|
||||
auto aligned_addr_start = mem_addr / GLOBAL_BLOCK_SIZE;
|
||||
auto aligned_addr_end = (mem_addr + size + GLOBAL_BLOCK_SIZE - 1) / GLOBAL_BLOCK_SIZE;
|
||||
|
||||
// submit snoop requests for the needed blocks
|
||||
vortex_->snp_req_addr = aligned_addr_start;
|
||||
vortex_->snp_req_tag = 0;
|
||||
vortex_->snp_req_valid = 1;
|
||||
vortex_->snp_rsp_ready = 1;
|
||||
|
||||
int pending_snp_reqs = 1;
|
||||
|
||||
#ifdef DBG_PRINT_CACHE_SNP
|
||||
std::cout << timestamp << ": [sim] snp req: addr=" << std::hex << vortex_->snp_req_addr << std::dec << " tag=" << vortex_->snp_req_tag << " remain=" << (aligned_addr_end - vortex_->snp_req_addr - 1) << std::endl;
|
||||
#endif
|
||||
|
||||
for (;;) {
|
||||
this->step();
|
||||
if (vortex_->snp_rsp_valid) {
|
||||
assert(pending_snp_reqs > 0);
|
||||
--pending_snp_reqs;
|
||||
#ifdef DBG_PRINT_CACHE_SNP
|
||||
std::cout << timestamp << ": [sim] snp rsp: tag=" << vortex_->snp_rsp_tag << " pending=" << pending_snp_reqs << std::endl;
|
||||
#endif
|
||||
}
|
||||
if (vortex_->snp_req_valid && vortex_->snp_req_ready) {
|
||||
if (vortex_->snp_req_addr + 1 < aligned_addr_end) {
|
||||
vortex_->snp_req_addr += 1;
|
||||
vortex_->snp_req_tag += 1;
|
||||
++pending_snp_reqs;
|
||||
#ifdef DBG_PRINT_CACHE_SNP
|
||||
std::cout << timestamp << ": [sim] snp req: addr=" << std::hex << vortex_->snp_req_addr << std::dec << " tag=" << vortex_->snp_req_tag << " remain=" << (aligned_addr_end - vortex_->snp_req_addr - 1) << std::endl;
|
||||
#endif
|
||||
} else {
|
||||
vortex_->snp_req_valid = 0;
|
||||
}
|
||||
}
|
||||
if (!vortex_->snp_req_valid
|
||||
&& 0 == pending_snp_reqs) {
|
||||
break;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
bool Simulator::run() {
|
||||
#ifndef NDEBUG
|
||||
std::cout << timestamp << ": [sim] run()" << std::endl;
|
||||
#endif
|
||||
|
||||
// reset the device
|
||||
this->reset();
|
||||
|
||||
// execute program
|
||||
while (vortex_->busy
|
||||
&& !vortex_->ebreak) {
|
||||
this->step();
|
||||
}
|
||||
|
||||
// wait 5 cycles to flush the pipeline
|
||||
this->wait(5);
|
||||
|
||||
// check riscv-tests PASSED/FAILED status
|
||||
#if (NUM_CLUSTERS == 1 && NUM_CORES == 1)
|
||||
int status = (int)vortex_->Vortex_Socket->genblk1__DOT__Vortex_Cluster->genblk1__BRA__0__KET____DOT__vortex_core->back_end->writeback->last_data_wb & 0xf;
|
||||
#else
|
||||
#if (NUM_CLUSTERS == 1)
|
||||
int status = (int)vortex_->Vortex_Socket->genblk1__DOT__Vortex_Cluster->genblk1__BRA__0__KET____DOT__vortex_core->back_end->writeback->last_data_wb & 0xf;
|
||||
#else
|
||||
int status = (int)vortex_->Vortex_Socket->genblk2__DOT__genblk1__BRA__0__KET____DOT__Vortex_Cluster->genblk1__BRA__0__KET____DOT__vortex_core->back_end->writeback->last_data_wb & 0xf;
|
||||
#endif
|
||||
#endif
|
||||
|
||||
return (status == 1);
|
||||
}
|
||||
@@ -35,6 +35,7 @@ public:
|
||||
void load_ihex(const char* program_file);
|
||||
|
||||
bool is_busy();
|
||||
|
||||
void reset();
|
||||
void step();
|
||||
void wait(uint32_t cycles);
|
||||
@@ -48,14 +49,19 @@ public:
|
||||
private:
|
||||
|
||||
void eval();
|
||||
void dbus_driver();
|
||||
void io_driver();
|
||||
|
||||
void eval_dram_bus();
|
||||
void eval_io_bus();
|
||||
void eval_snp_bus();
|
||||
|
||||
std::vector<dram_req_t> dram_rsp_vec_;
|
||||
|
||||
uint32_t snp_req_active_;
|
||||
uint32_t snp_req_size_;
|
||||
uint32_t pending_snp_reqs_;
|
||||
|
||||
RAM *ram_;
|
||||
VVortex_Socket *vortex_;
|
||||
bool enable_;
|
||||
#ifdef VCD_OUTPUT
|
||||
VerilatedVcdC *trace_;
|
||||
#endif
|
||||
|
||||
Reference in New Issue
Block a user