fixed gpr_ram bug + io bus arbitration

This commit is contained in:
Blaise Tine
2020-06-13 05:26:29 -07:00
parent d6b0ef2b3c
commit 4fa540575c
15 changed files with 542 additions and 377 deletions

View File

@@ -74,30 +74,35 @@ struct scope_signal_t {
static const scope_signal_t scope_signals[] = {
{ 32, "icache_req_addr" },
{ 2 , "icache_req_tag" },
{ 2, "icache_req_warp_num" },
{ 2, "icache_req_tag" },
{ 32, "icache_rsp_data" },
{ 2 , "icache_rsp_tag" },
{ 2, "icache_rsp_tag" },
{ 32, "dcache_req_addr" },
{ 2 , "dcache_req_tag" },
{ 2, "dcache_req_warp_num" },
{ 2, "dcache_req_tag" },
{ 32, "dcache_rsp_data" },
{ 2 , "dcache_rsp_tag" },
{ 32, "dram_req_addr" },
{ 29, "dram_req_tag" },
{ 29, "dram_rsp_tag" },
{ 2 , "icache_req_warp_num" },
{ 2 , "dcache_req_warp_num" },
{ 29, "dram_rsp_tag" },
{ 32, "snp_req_addr" },
{ 1, "snp_req_invalidate" },
{ 16, "snp_req_tag" },
{ 16, "snp_rsp_tag" },
{ 2, "decode_warp_num" },
{ 32, "decode_curr_PC" },
{ 5 , "execute_rd" },
{ 2 , "execute_warp_num" },
{ 1, "decode_is_jal" },
{ 5, "decode_rs1" },
{ 5, "decode_rs2" },
{ 2, "execute_warp_num" },
{ 5, "execute_rd" },
{ 32, "execute_a" },
{ 32, "execute_b" },
{ 5 , "writeback_rd" },
{ 2 , "writeback_warp_num" },
{ 32, "writeback_data" },
{ 2 , "decode_warp_num" },
{ 1 , "decode_is_jal" },
{ 5 , "decode_rs1" },
{ 5 , "decode_rs2" },
{ 2 , "writeback_wb" },
{ 32, "execute_b" },
{ 2, "writeback_warp_num" },
{ 2, "writeback_wb" },
{ 5, "writeback_rd" },
{ 32, "writeback_data" },
{ 1, "icache_req_valid" },
{ 1, "icache_req_ready" },
@@ -111,6 +116,10 @@ static const scope_signal_t scope_signals[] = {
{ 1, "dram_req_ready" },
{ 1, "dram_rsp_valid" },
{ 1, "dram_rsp_ready" },
{ 1, "snp_req_valid" },
{ 1, "snp_req_ready" },
{ 1, "snp_rsp_valid" },
{ 1, "snp_rsp_ready" },
{ 4, "decode_valid" },
{ 4, "execute_valid" },
{ 4, "writeback_valid" },

View File

@@ -66,6 +66,11 @@ int run_memcopy_test(vx_buffer_h sbuf,
((uint64_t*)vx_host_ptr(sbuf))[i] = shuffle(i, value);
}
// clear dbuf data
for (int i = 0; i < (64 * num_blocks) / 8; ++i) {
((uint64_t*)vx_host_ptr(dbuf))[i] = 0;
}
// write buffer to local memory
std::cout << "write buffer to local memory" << std::endl;
RT_CHECK(vx_copy_to_dev(sbuf, address, 64 * num_blocks, 0));
@@ -112,6 +117,11 @@ int run_kernel_test(vx_device_h device,
((uint64_t*)vx_host_ptr(sbuf))[i] = shuffle(i, seed);
}
// clear dbuf data
for (int i = 0; i < (64 * num_blocks) / 8; ++i) {
((uint64_t*)vx_host_ptr(dbuf))[i] = 0;
}
// write buffer to local memory
std::cout << "write buffer to local memory" << std::endl;
RT_CHECK(vx_copy_to_dev(sbuf, src_dev_addr, 64 * num_blocks, 0));
@@ -175,11 +185,11 @@ int main(int argc, char *argv[]) {
RT_CHECK(vx_alloc_shared_mem(device, 4096, &dbuf));
// run tests
/*if (0 == test || -1 == test) {
if (0 == test || -1 == test) {
std::cout << "run memcopy test" << std::endl;
RT_CHECK(run_memcopy_test(sbuf, dbuf, DEV_MEM_SRC_ADDR, 0x0badf00d00ff00ff, 1));
RT_CHECK(run_memcopy_test(sbuf, dbuf, DEV_MEM_SRC_ADDR, 0x0badf00d40ff40ff, 64));
}*/
}
if (1 == test || -1 == test) {
std::cout << "run kernel test" << std::endl;

Binary file not shown.

View File

@@ -99,7 +99,7 @@ vortex_afu.json
../rtl/VX_inst_multiplex.v
../rtl/VX_lsu_addr_gen.v
../rtl/VX_dcache_io_arb.v
../rtl/VX_dram_arb.v
../rtl/VX_mem_arb.v
../rtl/pipe_regs/VX_f_d_reg.v
../rtl/pipe_regs/VX_i_d_reg.v

View File

@@ -102,6 +102,7 @@ logic vx_dram_rsp_ready;
logic vx_snp_req_valid;
logic [`VX_DRAM_ADDR_WIDTH-1:0] vx_snp_req_addr;
logic vx_snp_req_invalidate = 0;
logic [`VX_SNP_TAG_WIDTH-1:0] vx_snp_req_tag;
logic vx_snp_req_ready;
@@ -798,26 +799,38 @@ end
`SCOPE_SIGNALS_DECL
`SCOPE_ASSIGN(scope_dram_req_valid, vx_dram_req_valid);
`SCOPE_ASSIGN(scope_dram_req_addr, {vx_dram_req_addr, 4'b0});
`SCOPE_ASSIGN(scope_dram_req_tag, vx_dram_req_tag);
`SCOPE_ASSIGN(scope_dram_req_ready, vx_dram_req_ready);
`SCOPE_ASSIGN(scope_dram_rsp_valid, vx_dram_rsp_valid);
`SCOPE_ASSIGN(scope_dram_rsp_tag, vx_dram_rsp_tag);
`SCOPE_ASSIGN(scope_dram_rsp_ready, vx_dram_rsp_ready);
`STATIC_ASSERT($bits({`SCOPE_SIGNALS_DATA_LIST `SCOPE_SIGNALS_UPD_LIST}) == 389, "oops!")
`SCOPE_ASSIGN(scope_snp_req_valid, vx_snp_req_valid);
`SCOPE_ASSIGN(scope_snp_req_addr, {vx_snp_req_addr, 4'b0});
`SCOPE_ASSIGN(scope_snp_req_invalidate, vx_snp_req_invalidate);
`SCOPE_ASSIGN(scope_snp_req_tag, vx_snp_req_tag);
`SCOPE_ASSIGN(scope_snp_req_ready, vx_snp_req_ready);
`SCOPE_ASSIGN(scope_snp_rsp_valid, vx_snp_rsp_valid);
`SCOPE_ASSIGN(scope_snp_rsp_tag, vx_snp_rsp_tag);
`SCOPE_ASSIGN(scope_snp_rsp_ready, vx_snp_rsp_ready);
`STATIC_ASSERT($bits({`SCOPE_SIGNALS_DATA_LIST `SCOPE_SIGNALS_UPD_LIST}) == 490, "oops!")
wire force_changed = (scope_icache_req_valid && scope_icache_req_ready)
|| (scope_icache_rsp_valid && scope_icache_rsp_ready)
|| ((| scope_dcache_req_valid) && scope_dcache_req_ready)
|| ((| scope_dcache_rsp_valid) && scope_dcache_rsp_ready)
|| (scope_dram_req_valid && scope_dram_req_ready)
|| (scope_dram_rsp_valid && scope_dram_rsp_ready);
|| (scope_dram_rsp_valid && scope_dram_rsp_ready)
|| (scope_snp_req_valid && scope_snp_req_ready)
|| (scope_snp_rsp_valid && scope_snp_rsp_ready);
VX_scope #(
.DATAW ($bits({`SCOPE_SIGNALS_DATA_LIST `SCOPE_SIGNALS_UPD_LIST})),
.BUSW (64),
.SIZE (4096),
.UPDW ($bits({`SCOPE_SIGNALS_UPD_LIST}))
.DATAW ($bits({`SCOPE_SIGNALS_DATA_LIST `SCOPE_SIGNALS_UPD_LIST})),
.BUSW (64),
.SIZE (4096),
.UPDW ($bits({`SCOPE_SIGNALS_UPD_LIST}))
) scope (
.clk (clk),
.reset (SoftReset),
@@ -864,7 +877,7 @@ Vortex_Socket #() vx_socket (
// Snoop request
.snp_req_valid (vx_snp_req_valid),
.snp_req_addr (vx_snp_req_addr),
.snp_req_invalidate(0),
.snp_req_invalidate(vx_snp_req_invalidate),
.snp_req_tag (vx_snp_req_tag),
.snp_req_ready (vx_snp_req_ready),

View File

@@ -141,6 +141,6 @@ module VX_back_end #(
`SCOPE_ASSIGN(scope_writeback_wb, writeback_if.wb);
`SCOPE_ASSIGN(scope_writeback_warp_num, writeback_if.warp_num);
`SCOPE_ASSIGN(scope_writeback_rd, writeback_if.rd);
`SCOPE_ASSIGN(scope_writeback_data, writeback_if.data[0]);
`SCOPE_ASSIGN(scope_writeback_data, writeback_if.data[0]);
endmodule

View File

@@ -228,6 +228,9 @@
// Cache ID
`define L2CACHE_ID (`L3_ENABLE ? 1 : 0)
// Core request tag bits
`define L2CORE_TAG_WIDTH (`DCORE_TAG_WIDTH + `CLOG2(`NUM_CORES))
// DRAM request data bits
`define L2DRAM_LINE_WIDTH (`L2_ENABLE ? (`L2BANK_LINE_SIZE * 8) : `DDRAM_LINE_WIDTH)
@@ -251,6 +254,9 @@
// Cache ID
`define L3CACHE_ID 0
// Core request tag bits
`define L3CORE_TAG_WIDTH (`L2CORE_TAG_WIDTH + `CLOG2(`NUM_CLUSTERS))
// DRAM request data bits
`define L3DRAM_LINE_WIDTH (`L3_ENABLE ? (`L3BANK_LINE_SIZE * 8) : `L2DRAM_LINE_WIDTH)
@@ -285,30 +291,36 @@
`ifdef SCOPE
`define SCOPE_SIGNALS_DATA_LIST \
scope_icache_req_addr, \
scope_icache_req_warp_num, \
scope_icache_req_tag, \
scope_icache_rsp_data, \
scope_icache_rsp_tag, \
scope_dcache_req_addr, \
scope_dcache_req_warp_num, \
scope_dcache_req_tag, \
scope_dcache_rsp_data, \
scope_dcache_rsp_tag, \
scope_dram_req_addr, \
scope_dram_req_tag, \
scope_dram_rsp_tag, \
scope_icache_req_warp_num, \
scope_dcache_req_warp_num, \
scope_decode_curr_PC, \
scope_execute_rd, \
scope_execute_warp_num, \
scope_execute_a, \
scope_execute_b, \
scope_writeback_rd, \
scope_writeback_warp_num, \
scope_writeback_data, \
scope_snp_req_addr, \
scope_snp_req_invalidate, \
scope_snp_req_tag, \
scope_snp_rsp_tag, \
scope_decode_warp_num, \
scope_decode_curr_PC, \
scope_decode_is_jal, \
scope_decode_rs1, \
scope_decode_rs2, \
scope_writeback_wb,
scope_execute_warp_num, \
scope_execute_rd, \
scope_execute_a, \
scope_execute_b, \
scope_writeback_warp_num, \
scope_writeback_wb, \
scope_writeback_rd, \
scope_writeback_data,
`define SCOPE_SIGNALS_UPD_LIST \
scope_icache_req_valid, \
@@ -323,6 +335,10 @@
scope_dram_req_ready, \
scope_dram_rsp_valid, \
scope_dram_rsp_ready, \
scope_snp_req_valid, \
scope_snp_req_ready, \
scope_snp_rsp_valid, \
scope_snp_rsp_ready, \
scope_decode_valid, \
scope_execute_valid, \
scope_writeback_valid, \
@@ -351,11 +367,20 @@
wire [`DCORE_TAG_WIDTH-1:0] scope_dcache_rsp_tag; \
wire scope_dcache_rsp_ready; \
wire scope_dram_req_valid; \
wire [31:0] scope_dram_req_addr; \
wire [`VX_DRAM_TAG_WIDTH-1:0] scope_dram_req_tag; \
wire scope_dram_req_ready; \
wire scope_dram_rsp_valid; \
wire [`VX_DRAM_TAG_WIDTH-1:0] scope_dram_rsp_tag; \
wire scope_dram_rsp_ready; \
wire scope_snp_req_valid; \
wire [31:0] scope_snp_req_addr; \
wire scope_snp_req_invalidate; \
wire [`VX_SNP_TAG_WIDTH-1:0] scope_snp_req_tag; \
wire scope_snp_req_ready; \
wire scope_snp_rsp_valid; \
wire [`VX_SNP_TAG_WIDTH-1:0] scope_snp_rsp_tag; \
wire scope_snp_rsp_ready; \
wire scope_schedule_delay; \
wire scope_memory_delay; \
wire scope_exec_delay; \
@@ -406,6 +431,7 @@
`define SCOPE_SIGNALS_DRAM_IO \
/* verilator lint_off UNDRIVEN */ \
output wire scope_dram_req_valid, \
output wire [31:0] scope_dram_req_addr, \
output wire [`VX_DRAM_TAG_WIDTH-1:0] scope_dram_req_tag, \
output wire scope_dram_req_ready, \
output wire scope_dram_rsp_valid, \
@@ -413,6 +439,18 @@
output wire scope_dram_rsp_ready, \
/* verilator lint_on UNDRIVEN */
`define SCOPE_SIGNALS_SNP_IO \
/* verilator lint_off UNDRIVEN */ \
output wire scope_snp_req_valid, \
output wire [31:0] scope_snp_req_addr, \
output wire scope_snp_req_invalidate, \
output wire [`VX_SNP_TAG_WIDTH-1:0] scope_snp_req_tag, \
output wire scope_snp_req_ready, \
output wire scope_snp_rsp_valid, \
output wire [`VX_SNP_TAG_WIDTH-1:0] scope_snp_rsp_tag, \
output wire scope_snp_rsp_ready, \
/* verilator lint_on UNDRIVEN */
`define SCOPE_SIGNALS_CORE_IO \
/* verilator lint_off UNDRIVEN */ \
output wire scope_schedule_delay, \
@@ -438,7 +476,7 @@
output wire [1:0] scope_writeback_warp_num, \
output wire [1:0] scope_writeback_wb, \
output wire [4:0] scope_writeback_rd, \
output wire [31:0] scope_writeback_data, \
output wire [31:0] scope_writeback_data,
/* verilator lint_on UNDRIVEN */
`define SCOPE_SIGNALS_ICACHE_ATTACH \
@@ -465,12 +503,23 @@
`define SCOPE_SIGNALS_DRAM_ATTACH \
.scope_dram_req_valid (scope_dram_req_valid), \
.scope_dram_req_addr (scope_dram_req_addr), \
.scope_dram_req_tag (scope_dram_req_tag), \
.scope_dram_req_ready (scope_dram_req_ready), \
.scope_dram_rsp_valid (scope_dram_rsp_valid), \
.scope_dram_rsp_tag (scope_dram_rsp_tag), \
.scope_dram_rsp_ready (scope_dram_rsp_ready),
`define SCOPE_SIGNALS_SNP_ATTACH \
.scope_snp_req_valid (scope_snp_req_valid), \
.scope_snp_req_addr (scope_snp_req_addr), \
.scope_snp_req_invalidate(scope_snp_req_invalidate), \
.scope_snp_req_tag (scope_snp_req_tag), \
.scope_snp_req_ready (scope_snp_req_ready), \
.scope_snp_rsp_valid (scope_snp_rsp_valid), \
.scope_snp_rsp_tag (scope_snp_rsp_tag), \
.scope_snp_rsp_ready (scope_snp_rsp_ready),
`define SCOPE_SIGNALS_CORE_ATTACH \
.scope_schedule_delay (scope_schedule_delay), \
.scope_memory_delay (scope_memory_delay), \

View File

@@ -1,77 +0,0 @@
`include "VX_define.vh"
module VX_dram_arb #(
parameter NUM_REQUESTS = 1,
parameter DRAM_LINE_SIZE = 1,
parameter CORE_TAG_WIDTH = 1,
parameter DRAM_TAG_WIDTH = 1,
parameter DRAM_LINE_WIDTH = DRAM_LINE_SIZE * 8,
parameter DRAM_ADDR_WIDTH = 32 - `CLOG2(DRAM_LINE_SIZE)
) (
input wire clk,
input wire reset,
// Core request
input wire [NUM_REQUESTS-1:0] in_dram_req_valid,
input wire [NUM_REQUESTS-1:0] in_dram_req_rw,
input wire [NUM_REQUESTS-1:0][DRAM_LINE_SIZE-1:0] in_dram_req_byteen,
input wire [NUM_REQUESTS-1:0][DRAM_ADDR_WIDTH-1:0] in_dram_req_addr,
input wire [NUM_REQUESTS-1:0][DRAM_LINE_WIDTH-1:0] in_dram_req_data,
input wire [NUM_REQUESTS-1:0][CORE_TAG_WIDTH-1:0] in_dram_req_tag,
output wire [NUM_REQUESTS-1:0] in_dram_req_ready,
// Core response
output wire [NUM_REQUESTS-1:0] in_dram_rsp_valid,
output wire [NUM_REQUESTS-1:0][DRAM_LINE_WIDTH-1:0] in_dram_rsp_data,
output wire [NUM_REQUESTS-1:0][CORE_TAG_WIDTH-1:0] in_dram_rsp_tag,
input wire [NUM_REQUESTS-1:0] in_dram_rsp_ready,
// DRAM request
output wire out_dram_req_valid,
output wire out_dram_req_rw,
output wire [DRAM_LINE_SIZE-1:0] out_dram_req_byteen,
output wire [DRAM_ADDR_WIDTH-1:0] out_dram_req_addr,
output wire [DRAM_LINE_WIDTH-1:0] out_dram_req_data,
output wire [DRAM_TAG_WIDTH-1:0] out_dram_req_tag,
input wire out_dram_req_ready,
// DRAM response
input wire out_dram_rsp_valid,
input wire [DRAM_LINE_WIDTH-1:0] out_dram_rsp_data,
input wire [DRAM_TAG_WIDTH-1:0] out_dram_rsp_tag,
output wire out_dram_rsp_ready
);
reg [`REQS_BITS-1:0] bus_req_sel;
always @(posedge clk) begin
if (reset) begin
bus_req_sel <= 0;
end else begin
bus_req_sel <= bus_req_sel + 1;
end
end
assign out_dram_req_valid = in_dram_req_valid [bus_req_sel];
assign out_dram_req_rw = in_dram_req_rw [bus_req_sel];
assign out_dram_req_byteen= in_dram_req_byteen [bus_req_sel];
assign out_dram_req_addr = in_dram_req_addr [bus_req_sel];
assign out_dram_req_data = in_dram_req_data [bus_req_sel];
assign out_dram_req_tag = {in_dram_req_tag [bus_req_sel], (`REQS_BITS)'(bus_req_sel)};
genvar i;
for (i = 0; i < NUM_REQUESTS; i++) begin
assign in_dram_req_ready[i] = out_dram_req_ready && (bus_req_sel == `REQS_BITS'(i));
end
wire [`REQS_BITS-1:0] bus_rsp_sel = out_dram_rsp_tag[`REQS_BITS-1:0];
for (i = 0; i < NUM_REQUESTS; i++) begin
assign in_dram_rsp_valid[i] = out_dram_rsp_valid && (bus_rsp_sel == `REQS_BITS'(i));
assign in_dram_rsp_data[i] = out_dram_rsp_data;
assign in_dram_rsp_tag[i] = out_dram_rsp_tag[`REQS_BITS +: CORE_TAG_WIDTH];
end
assign out_dram_rsp_ready = in_dram_rsp_ready[bus_rsp_sel];
endmodule

View File

@@ -18,8 +18,8 @@ module VX_gpr (
wire write_enable = write_ce && ((writeback_if.wb != 0));
`ifndef ASIC
`ifndef ASIC
VX_gpr_ram gpr_ram (
.we (write_enable),
.clk (clk),

View File

@@ -12,24 +12,19 @@ module VX_gpr_ram (
output reg [`NUM_THREADS-1:0][31:0] q1,
output reg [`NUM_THREADS-1:0][31:0] q2
);
// Thread Byte Bit
logic [`NUM_THREADS-1:0][3:0][7:0] ram[31:0];
reg [`NUM_THREADS-1:0][31:0] ram[31:0];
integer i;
`UNUSED_VAR(reset)
always @(posedge clk) begin
if (reset) begin
//--
end else begin
if (we) begin
integer i;
for (i = 0; i < `NUM_THREADS; i++) begin
if (be[i]) begin
ram[waddr][i][0] <= wdata[i][7:0];
ram[waddr][i][1] <= wdata[i][15:8];
ram[waddr][i][2] <= wdata[i][23:16];
ram[waddr][i][3] <= wdata[i][31:24];
end
if (we) begin
for (i = 0; i < `NUM_THREADS; i++) begin
if (be[i]) begin
ram[waddr][i] <= wdata[i];
end
end
end
end
end

View File

@@ -8,7 +8,7 @@ module VX_gpr_wrapper (
VX_gpr_jal_if gpr_jal_if,
output wire [`NUM_THREADS-1:0][31:0] a_reg_data,
output wire [`NUM_THREADS-1:0][31:0] b_reg_data
output wire [`NUM_THREADS-1:0][31:0] b_reg_data
);
wire [`NUM_WARPS-1:0][`NUM_THREADS-1:0][31:0] tmp_a_reg_data;
wire [`NUM_WARPS-1:0][`NUM_THREADS-1:0][31:0] tmp_b_reg_data;

100
hw/rtl/VX_mem_arb.v Normal file
View File

@@ -0,0 +1,100 @@
`include "VX_define.vh"
module VX_mem_arb #(
parameter NUM_REQUESTS = 1,
parameter WORD_SIZE = 1,
parameter TAG_IN_WIDTH = 1,
parameter TAG_OUT_WIDTH = 1,
parameter WORD_WIDTH = WORD_SIZE * 8,
parameter ADDR_WIDTH = 32 - `CLOG2(WORD_SIZE),
parameter REQS_BITS = `CLOG2(NUM_REQUESTS)
) (
input wire clk,
input wire reset,
// input requests
input wire [NUM_REQUESTS-1:0] in_mem_req_valid,
input wire [NUM_REQUESTS-1:0] in_mem_req_rw,
input wire [NUM_REQUESTS-1:0][WORD_SIZE-1:0] in_mem_req_byteen,
input wire [NUM_REQUESTS-1:0][ADDR_WIDTH-1:0] in_mem_req_addr,
input wire [NUM_REQUESTS-1:0][WORD_WIDTH-1:0] in_mem_req_data,
input wire [NUM_REQUESTS-1:0][TAG_IN_WIDTH-1:0] in_mem_req_tag,
output wire [NUM_REQUESTS-1:0] in_mem_req_ready,
// input response
output wire [NUM_REQUESTS-1:0] in_mem_rsp_valid,
output wire [NUM_REQUESTS-1:0][WORD_WIDTH-1:0] in_mem_rsp_data,
output wire [NUM_REQUESTS-1:0][TAG_IN_WIDTH-1:0] in_mem_rsp_tag,
input wire [NUM_REQUESTS-1:0] in_mem_rsp_ready,
// output request
output wire out_mem_req_valid,
output wire out_mem_req_rw,
output wire [WORD_SIZE-1:0] out_mem_req_byteen,
output wire [ADDR_WIDTH-1:0] out_mem_req_addr,
output wire [WORD_WIDTH-1:0] out_mem_req_data,
output wire [TAG_OUT_WIDTH-1:0] out_mem_req_tag,
input wire out_mem_req_ready,
// output response
input wire out_mem_rsp_valid,
input wire [WORD_WIDTH-1:0] out_mem_rsp_data,
input wire [TAG_OUT_WIDTH-1:0] out_mem_rsp_tag,
output wire out_mem_rsp_ready
);
if (NUM_REQUESTS == 1) begin
`UNUSED_VAR (clk)
`UNUSED_VAR (reset)
assign out_mem_req_valid = in_mem_req_valid;
assign out_mem_req_rw = in_mem_req_rw;
assign out_mem_req_byteen = in_mem_req_byteen;
assign out_mem_req_addr = in_mem_req_addr;
assign out_mem_req_data = in_mem_req_data;
assign out_mem_req_tag = in_mem_req_tag;
assign in_mem_req_ready = out_mem_req_ready;
assign in_mem_rsp_valid = out_mem_rsp_valid;
assign in_mem_rsp_data = out_mem_rsp_data;
assign in_mem_rsp_tag = out_mem_rsp_tag;
assign out_mem_rsp_ready = in_mem_rsp_ready;
end else begin
reg [REQS_BITS-1:0] bus_req_sel;
always @(posedge clk) begin
if (reset) begin
bus_req_sel <= 0;
end else begin
bus_req_sel <= bus_req_sel + 1;
end
end
assign out_mem_req_valid = in_mem_req_valid [bus_req_sel];
assign out_mem_req_rw = in_mem_req_rw [bus_req_sel];
assign out_mem_req_byteen = in_mem_req_byteen [bus_req_sel];
assign out_mem_req_addr = in_mem_req_addr [bus_req_sel];
assign out_mem_req_data = in_mem_req_data [bus_req_sel];
assign out_mem_req_tag = {in_mem_req_tag [bus_req_sel], REQS_BITS'(bus_req_sel)};
genvar i;
for (i = 0; i < NUM_REQUESTS; i++) begin
assign in_mem_req_ready[i] = out_mem_req_ready && (bus_req_sel == REQS_BITS'(i));
end
wire [REQS_BITS-1:0] bus_rsp_sel = out_mem_rsp_tag[REQS_BITS-1:0];
for (i = 0; i < NUM_REQUESTS; i++) begin
assign in_mem_rsp_valid[i] = out_mem_rsp_valid && (bus_rsp_sel == REQS_BITS'(i));
assign in_mem_rsp_data[i] = out_mem_rsp_data;
assign in_mem_rsp_tag[i] = out_mem_rsp_tag[REQS_BITS +: TAG_IN_WIDTH];
end
assign out_mem_rsp_ready = in_mem_rsp_ready[bus_rsp_sel];
end
endmodule

View File

@@ -56,16 +56,16 @@ module Vortex #(
// I/O request
output wire io_req_valid,
output wire io_req_rw,
output wire[3:0] io_req_byteen,
output wire[29:0] io_req_addr,
output wire[31:0] io_req_data,
output wire[`DCORE_TAG_WIDTH-1:0] io_req_tag,
output wire [3:0] io_req_byteen,
output wire [29:0] io_req_addr,
output wire [31:0] io_req_data,
output wire [`DCORE_TAG_WIDTH-1:0] io_req_tag,
input wire io_req_ready,
// I/O response
input wire io_rsp_valid,
input wire[31:0] io_rsp_data,
input wire[`DCORE_TAG_WIDTH-1:0] io_rsp_tag,
input wire [31:0] io_rsp_data,
input wire [`DCORE_TAG_WIDTH-1:0] io_rsp_tag,
output wire io_rsp_ready,
// Status

View File

@@ -15,98 +15,100 @@ module Vortex_Cluster #(
// DRAM request
output wire dram_req_valid,
output wire dram_req_rw,
output wire[`L2DRAM_BYTEEN_WIDTH-1:0] dram_req_byteen,
output wire[`L2DRAM_ADDR_WIDTH-1:0] dram_req_addr,
output wire[`L2DRAM_LINE_WIDTH-1:0] dram_req_data,
output wire[`L2DRAM_TAG_WIDTH-1:0] dram_req_tag,
output wire [`L2DRAM_BYTEEN_WIDTH-1:0] dram_req_byteen,
output wire [`L2DRAM_ADDR_WIDTH-1:0] dram_req_addr,
output wire [`L2DRAM_LINE_WIDTH-1:0] dram_req_data,
output wire [`L2DRAM_TAG_WIDTH-1:0] dram_req_tag,
input wire dram_req_ready,
// DRAM response
input wire dram_rsp_valid,
input wire[`L2DRAM_LINE_WIDTH-1:0] dram_rsp_data,
input wire[`L2DRAM_TAG_WIDTH-1:0] dram_rsp_tag,
input wire [`L2DRAM_LINE_WIDTH-1:0] dram_rsp_data,
input wire [`L2DRAM_TAG_WIDTH-1:0] dram_rsp_tag,
output wire dram_rsp_ready,
// Snoop request
input wire snp_req_valid,
input wire[`L2DRAM_ADDR_WIDTH-1:0] snp_req_addr,
input wire [`L2DRAM_ADDR_WIDTH-1:0] snp_req_addr,
input wire snp_req_invalidate,
input wire[`L2SNP_TAG_WIDTH-1:0] snp_req_tag,
input wire [`L2SNP_TAG_WIDTH-1:0] snp_req_tag,
output wire snp_req_ready,
// Snoop response
output wire snp_rsp_valid,
output wire[`L2SNP_TAG_WIDTH-1:0] snp_rsp_tag,
output wire [`L2SNP_TAG_WIDTH-1:0] snp_rsp_tag,
input wire snp_rsp_ready,
// I/O request
output wire io_req_valid,
output wire io_req_rw,
output wire[3:0] io_req_byteen,
output wire[29:0] io_req_addr,
output wire[31:0] io_req_data,
output wire[`DCORE_TAG_WIDTH-1:0] io_req_tag,
output wire [3:0] io_req_byteen,
output wire [29:0] io_req_addr,
output wire [31:0] io_req_data,
output wire [`DCORE_TAG_WIDTH-1:0] io_req_tag,
input wire io_req_ready,
// I/O response
input wire io_rsp_valid,
input wire[31:0] io_rsp_data,
input wire[`DCORE_TAG_WIDTH-1:0] io_rsp_tag,
input wire [31:0] io_rsp_data,
input wire [`DCORE_TAG_WIDTH-1:0] io_rsp_tag,
output wire io_rsp_ready,
// Status
output wire busy,
output wire ebreak
);
wire[`NUM_CORES-1:0] per_core_D_dram_req_valid;
wire[`NUM_CORES-1:0] per_core_D_dram_req_rw;
wire[`NUM_CORES-1:0][`DDRAM_BYTEEN_WIDTH-1:0] per_core_D_dram_req_byteen;
wire[`NUM_CORES-1:0][`DDRAM_ADDR_WIDTH-1:0] per_core_D_dram_req_addr;
wire[`NUM_CORES-1:0][`DDRAM_LINE_WIDTH-1:0] per_core_D_dram_req_data;
wire[`NUM_CORES-1:0][`DDRAM_TAG_WIDTH-1:0] per_core_D_dram_req_tag;
wire[`NUM_CORES-1:0] per_core_D_dram_req_ready;
wire [`NUM_CORES-1:0] per_core_D_dram_req_valid;
wire [`NUM_CORES-1:0] per_core_D_dram_req_rw;
wire [`NUM_CORES-1:0][`DDRAM_BYTEEN_WIDTH-1:0] per_core_D_dram_req_byteen;
wire [`NUM_CORES-1:0][`DDRAM_ADDR_WIDTH-1:0] per_core_D_dram_req_addr;
wire [`NUM_CORES-1:0][`DDRAM_LINE_WIDTH-1:0] per_core_D_dram_req_data;
wire [`NUM_CORES-1:0][`DDRAM_TAG_WIDTH-1:0] per_core_D_dram_req_tag;
wire [`NUM_CORES-1:0] per_core_D_dram_req_ready;
wire[`NUM_CORES-1:0] per_core_D_dram_rsp_valid;
wire[`NUM_CORES-1:0][`DDRAM_LINE_WIDTH-1:0] per_core_D_dram_rsp_data;
wire[`NUM_CORES-1:0][`DDRAM_TAG_WIDTH-1:0] per_core_D_dram_rsp_tag;
wire[`NUM_CORES-1:0] per_core_D_dram_rsp_ready;
wire [`NUM_CORES-1:0] per_core_D_dram_rsp_valid;
wire [`NUM_CORES-1:0][`DDRAM_LINE_WIDTH-1:0] per_core_D_dram_rsp_data;
wire [`NUM_CORES-1:0][`DDRAM_TAG_WIDTH-1:0] per_core_D_dram_rsp_tag;
wire [`NUM_CORES-1:0] per_core_D_dram_rsp_ready;
wire[`NUM_CORES-1:0] per_core_I_dram_req_valid;
wire[`NUM_CORES-1:0] per_core_I_dram_req_rw;
wire[`NUM_CORES-1:0][`IDRAM_BYTEEN_WIDTH-1:0] per_core_I_dram_req_byteen;
wire[`NUM_CORES-1:0][`IDRAM_ADDR_WIDTH-1:0] per_core_I_dram_req_addr;
wire[`NUM_CORES-1:0][`IDRAM_LINE_WIDTH-1:0] per_core_I_dram_req_data;
wire[`NUM_CORES-1:0][`IDRAM_TAG_WIDTH-1:0] per_core_I_dram_req_tag;
wire[`NUM_CORES-1:0] per_core_I_dram_req_ready;
wire [`NUM_CORES-1:0] per_core_I_dram_req_valid;
wire [`NUM_CORES-1:0] per_core_I_dram_req_rw;
wire [`NUM_CORES-1:0][`IDRAM_BYTEEN_WIDTH-1:0] per_core_I_dram_req_byteen;
wire [`NUM_CORES-1:0][`IDRAM_ADDR_WIDTH-1:0] per_core_I_dram_req_addr;
wire [`NUM_CORES-1:0][`IDRAM_LINE_WIDTH-1:0] per_core_I_dram_req_data;
wire [`NUM_CORES-1:0][`IDRAM_TAG_WIDTH-1:0] per_core_I_dram_req_tag;
wire [`NUM_CORES-1:0] per_core_I_dram_req_ready;
wire[`NUM_CORES-1:0] per_core_I_dram_rsp_valid;
wire[`NUM_CORES-1:0][`IDRAM_LINE_WIDTH-1:0] per_core_I_dram_rsp_data;
wire[`NUM_CORES-1:0][`IDRAM_TAG_WIDTH-1:0] per_core_I_dram_rsp_tag;
wire[`NUM_CORES-1:0] per_core_I_dram_rsp_ready;
wire [`NUM_CORES-1:0] per_core_I_dram_rsp_valid;
wire [`NUM_CORES-1:0][`IDRAM_LINE_WIDTH-1:0] per_core_I_dram_rsp_data;
wire [`NUM_CORES-1:0][`IDRAM_TAG_WIDTH-1:0] per_core_I_dram_rsp_tag;
wire [`NUM_CORES-1:0] per_core_I_dram_rsp_ready;
wire[`NUM_CORES-1:0] per_core_snp_req_valid;
wire[`NUM_CORES-1:0][`DDRAM_ADDR_WIDTH-1:0] per_core_snp_req_addr;
wire[`NUM_CORES-1:0] per_core_snp_req_invalidate;
wire[`NUM_CORES-1:0][`DSNP_TAG_WIDTH-1:0] per_core_snp_req_tag;
wire[`NUM_CORES-1:0] per_core_snp_req_ready;
wire [`NUM_CORES-1:0] per_core_snp_req_valid;
wire [`NUM_CORES-1:0][`DDRAM_ADDR_WIDTH-1:0] per_core_snp_req_addr;
wire [`NUM_CORES-1:0] per_core_snp_req_invalidate;
wire [`NUM_CORES-1:0][`DSNP_TAG_WIDTH-1:0] per_core_snp_req_tag;
wire [`NUM_CORES-1:0] per_core_snp_req_ready;
wire[`NUM_CORES-1:0] per_core_snp_rsp_valid;
wire[`NUM_CORES-1:0][`DSNP_TAG_WIDTH-1:0] per_core_snp_rsp_tag;
wire[`NUM_CORES-1:0] per_core_snp_rsp_ready;
wire [`NUM_CORES-1:0] per_core_snp_rsp_valid;
wire [`NUM_CORES-1:0][`DSNP_TAG_WIDTH-1:0] per_core_snp_rsp_tag;
wire [`NUM_CORES-1:0] per_core_snp_rsp_ready;
`IGNORE_WARNINGS_BEGIN
wire[`NUM_CORES-1:0] per_core_io_req_valid;
wire[`NUM_CORES-1:0] per_core_io_req_rw;
wire[`NUM_CORES-1:0][3:0] per_core_io_req_byteen;
wire[`NUM_CORES-1:0][29:0] per_core_io_req_addr;
wire[`NUM_CORES-1:0][31:0] per_core_io_req_data;
wire[`NUM_CORES-1:0][`DCORE_TAG_WIDTH-1:0] per_core_io_req_tag;
wire [`NUM_CORES-1:0] per_core_io_req_valid;
wire [`NUM_CORES-1:0] per_core_io_req_rw;
wire [`NUM_CORES-1:0][3:0] per_core_io_req_byteen;
wire [`NUM_CORES-1:0][29:0] per_core_io_req_addr;
wire [`NUM_CORES-1:0][31:0] per_core_io_req_data;
wire [`NUM_CORES-1:0][`DCORE_TAG_WIDTH-1:0] per_core_io_req_tag;
wire [`NUM_CORES-1:0] per_core_io_req_ready;
wire[`NUM_CORES-1:0] per_core_io_rsp_ready;
`IGNORE_WARNINGS_END
wire [`NUM_CORES-1:0] per_core_io_rsp_valid;
wire [`NUM_CORES-1:0][`DCORE_TAG_WIDTH-1:0] per_core_io_rsp_tag;
wire [`NUM_CORES-1:0][31:0] per_core_io_rsp_data;
wire [`NUM_CORES-1:0] per_core_io_rsp_ready;
wire[`NUM_CORES-1:0] per_core_busy;
wire[`NUM_CORES-1:0] per_core_ebreak;
wire [`NUM_CORES-1:0] per_core_busy;
wire [`NUM_CORES-1:0] per_core_ebreak;
genvar i;
for (i = 0; i < `NUM_CORES; i++) begin
@@ -120,6 +122,7 @@ module Vortex_Cluster #(
.clk (clk),
.reset (reset),
.D_dram_req_valid (per_core_D_dram_req_valid [i]),
.D_dram_req_rw (per_core_D_dram_req_rw [i]),
.D_dram_req_byteen (per_core_D_dram_req_byteen [i]),
@@ -160,27 +163,57 @@ module Vortex_Cluster #(
.io_req_addr (per_core_io_req_addr [i]),
.io_req_data (per_core_io_req_data [i]),
.io_req_tag (per_core_io_req_tag [i]),
.io_req_ready (io_req_ready),
.io_req_ready (per_core_io_req_ready [i]),
.io_rsp_valid (io_rsp_valid),
.io_rsp_data (io_rsp_data),
.io_rsp_tag (io_rsp_tag),
.io_rsp_valid (per_core_io_rsp_valid [i]),
.io_rsp_data (per_core_io_rsp_data [i]),
.io_rsp_tag (per_core_io_rsp_tag [i]),
.io_rsp_ready (per_core_io_rsp_ready [i]),
.busy (per_core_busy [i]),
.ebreak (per_core_ebreak [i])
);
end
end
assign io_req_valid = per_core_io_req_valid[0];
assign io_req_rw = per_core_io_req_rw[0];
assign io_req_byteen = per_core_io_req_byteen[0];
assign io_req_addr = per_core_io_req_addr[0];
assign io_req_data = per_core_io_req_data[0];
assign io_req_byteen = per_core_io_req_byteen[0];
assign io_req_tag = per_core_io_req_tag[0];
VX_mem_arb #(
.NUM_REQUESTS (`NUM_CLUSTERS),
.WORD_SIZE (4),
.TAG_IN_WIDTH (`DCORE_TAG_WIDTH),
.TAG_OUT_WIDTH (`L2CORE_TAG_WIDTH)
) io_arb (
.clk (clk),
.reset (reset),
assign io_rsp_ready = per_core_io_rsp_ready[0];
// input requests
.in_mem_req_valid (per_core_io_req_valid),
.in_mem_req_rw (per_core_io_req_rw),
.in_mem_req_byteen (per_core_io_req_byteen),
.in_mem_req_addr (per_core_io_req_addr),
.in_mem_req_data (per_core_io_req_data),
.in_mem_req_tag (per_core_io_req_tag),
.in_mem_req_ready (per_core_io_req_ready),
// input responses
.in_mem_rsp_valid (per_core_io_rsp_valid),
.in_mem_rsp_data (per_core_io_rsp_data),
.in_mem_rsp_tag (per_core_io_rsp_tag),
.in_mem_rsp_ready (per_core_io_rsp_ready),
// output request
.out_mem_req_valid (io_req_valid),
.out_mem_req_rw (io_req_rw),
.out_mem_req_byteen (io_req_byteen),
.out_mem_req_addr (io_req_addr),
.out_mem_req_data (io_req_data),
.out_mem_req_tag (io_req_tag),
.out_mem_req_ready (io_req_ready),
// output response
.out_mem_rsp_valid (io_rsp_valid),
.out_mem_rsp_tag (io_rsp_tag),
.out_mem_rsp_data (io_rsp_data),
.out_mem_rsp_ready (io_rsp_ready)
);
assign busy = (| per_core_busy);
assign ebreak = (& per_core_ebreak);
@@ -343,18 +376,18 @@ module Vortex_Cluster #(
end else begin
wire[`L2NUM_REQUESTS-1:0] arb_core_req_valid;
wire[`L2NUM_REQUESTS-1:0] arb_core_req_rw;
wire[`L2NUM_REQUESTS-1:0][`DDRAM_BYTEEN_WIDTH-1:0] arb_core_req_byteen;
wire[`L2NUM_REQUESTS-1:0][`DDRAM_ADDR_WIDTH-1:0] arb_core_req_addr;
wire[`L2NUM_REQUESTS-1:0][`DDRAM_TAG_WIDTH-1:0] arb_core_req_tag;
wire[`L2NUM_REQUESTS-1:0][`DDRAM_LINE_WIDTH-1:0] arb_core_req_data;
wire[`L2NUM_REQUESTS-1:0] arb_core_req_ready;
wire[`L2NUM_REQUESTS-1:0] arb_dram_req_valid;
wire[`L2NUM_REQUESTS-1:0] arb_dram_req_rw;
wire[`L2NUM_REQUESTS-1:0][`DDRAM_BYTEEN_WIDTH-1:0] arb_dram_req_byteen;
wire[`L2NUM_REQUESTS-1:0][`DDRAM_ADDR_WIDTH-1:0] arb_dram_req_addr;
wire[`L2NUM_REQUESTS-1:0][`DDRAM_TAG_WIDTH-1:0] arb_dram_req_tag;
wire[`L2NUM_REQUESTS-1:0][`DDRAM_LINE_WIDTH-1:0] arb_dram_req_data;
wire[`L2NUM_REQUESTS-1:0] arb_dram_req_ready;
wire[`L2NUM_REQUESTS-1:0] arb_core_rsp_valid;
wire[`L2NUM_REQUESTS-1:0][`DDRAM_LINE_WIDTH-1:0] arb_core_rsp_data;
wire[`L2NUM_REQUESTS-1:0][`DDRAM_TAG_WIDTH-1:0] arb_core_rsp_tag;
wire[`L2NUM_REQUESTS-1:0] arb_core_rsp_ready;
wire[`L2NUM_REQUESTS-1:0] arb_dram_rsp_valid;
wire[`L2NUM_REQUESTS-1:0][`DDRAM_LINE_WIDTH-1:0] arb_dram_rsp_data;
wire[`L2NUM_REQUESTS-1:0][`DDRAM_TAG_WIDTH-1:0] arb_dram_rsp_tag;
wire[`L2NUM_REQUESTS-1:0] arb_dram_rsp_ready;
wire[`NUM_CORES-1:0] arb_snp_fwdout_valid;
wire[`NUM_CORES-1:0][`DDRAM_ADDR_WIDTH-1:0] arb_snp_fwdout_addr;
@@ -367,38 +400,38 @@ module Vortex_Cluster #(
wire[`NUM_CORES-1:0] arb_snp_fwdin_ready;
for (i = 0; i < `L2NUM_REQUESTS; i = i + 2) begin
assign arb_core_req_valid [i] = per_core_D_dram_req_valid[(i/2)];
assign arb_core_req_valid [i+1] = per_core_I_dram_req_valid[(i/2)];
assign arb_dram_req_valid [i] = per_core_D_dram_req_valid[(i/2)];
assign arb_dram_req_valid [i+1] = per_core_I_dram_req_valid[(i/2)];
assign arb_core_req_rw [i] = per_core_D_dram_req_rw[(i/2)];
assign arb_core_req_rw [i+1] = per_core_I_dram_req_rw[(i/2)];
assign arb_dram_req_rw [i] = per_core_D_dram_req_rw[(i/2)];
assign arb_dram_req_rw [i+1] = per_core_I_dram_req_rw[(i/2)];
assign arb_core_req_byteen[i] = per_core_D_dram_req_byteen[(i/2)];
assign arb_core_req_byteen[i+1] = per_core_I_dram_req_byteen[(i/2)];
assign arb_dram_req_byteen[i] = per_core_D_dram_req_byteen[(i/2)];
assign arb_dram_req_byteen[i+1] = per_core_I_dram_req_byteen[(i/2)];
assign arb_core_req_addr [i] = per_core_D_dram_req_addr[(i/2)];
assign arb_core_req_addr [i+1] = per_core_I_dram_req_addr[(i/2)];
assign arb_dram_req_addr [i] = per_core_D_dram_req_addr[(i/2)];
assign arb_dram_req_addr [i+1] = per_core_I_dram_req_addr[(i/2)];
assign arb_core_req_data [i] = per_core_D_dram_req_data[(i/2)];
assign arb_core_req_data [i+1] = per_core_I_dram_req_data[(i/2)];
assign arb_dram_req_data [i] = per_core_D_dram_req_data[(i/2)];
assign arb_dram_req_data [i+1] = per_core_I_dram_req_data[(i/2)];
assign arb_core_req_tag [i] = per_core_D_dram_req_tag[(i/2)];
assign arb_core_req_tag [i+1] = per_core_I_dram_req_tag[(i/2)];
assign arb_dram_req_tag [i] = per_core_D_dram_req_tag[(i/2)];
assign arb_dram_req_tag [i+1] = per_core_I_dram_req_tag[(i/2)];
assign per_core_D_dram_req_ready [(i/2)] = arb_core_req_ready[i];
assign per_core_I_dram_req_ready [(i/2)] = arb_core_req_ready[i+1];
assign per_core_D_dram_req_ready [(i/2)] = arb_dram_req_ready[i];
assign per_core_I_dram_req_ready [(i/2)] = arb_dram_req_ready[i+1];
assign per_core_D_dram_rsp_valid [(i/2)] = arb_core_rsp_valid[i];
assign per_core_I_dram_rsp_valid [(i/2)] = arb_core_rsp_valid[i+1];
assign per_core_D_dram_rsp_valid [(i/2)] = arb_dram_rsp_valid[i];
assign per_core_I_dram_rsp_valid [(i/2)] = arb_dram_rsp_valid[i+1];
assign per_core_D_dram_rsp_data [(i/2)] = arb_core_rsp_data[i];
assign per_core_I_dram_rsp_data [(i/2)] = arb_core_rsp_data[i+1];
assign per_core_D_dram_rsp_data [(i/2)] = arb_dram_rsp_data[i];
assign per_core_I_dram_rsp_data [(i/2)] = arb_dram_rsp_data[i+1];
assign per_core_D_dram_rsp_tag [(i/2)] = arb_core_rsp_tag[i];
assign per_core_I_dram_rsp_tag [(i/2)] = arb_core_rsp_tag[i+1];
assign per_core_D_dram_rsp_tag [(i/2)] = arb_dram_rsp_tag[i];
assign per_core_I_dram_rsp_tag [(i/2)] = arb_dram_rsp_tag[i+1];
assign arb_core_rsp_ready [i] = per_core_D_dram_rsp_ready[(i/2)];
assign arb_core_rsp_ready [i+1] = per_core_I_dram_rsp_ready[(i/2)];
assign arb_dram_rsp_ready [i] = per_core_D_dram_rsp_ready[(i/2)];
assign arb_dram_rsp_ready [i+1] = per_core_I_dram_rsp_ready[(i/2)];
assign per_core_snp_req_valid [(i/2)] = arb_snp_fwdout_valid [(i/2)];
assign per_core_snp_req_addr [(i/2)] = arb_snp_fwdout_addr [(i/2)];
@@ -456,44 +489,44 @@ module Vortex_Cluster #(
assign arb_snp_fwdin_ready = snp_rsp_ready;
end
VX_dram_arb #(
.NUM_REQUESTS (`L2NUM_REQUESTS),
.DRAM_LINE_SIZE (`L2BANK_LINE_SIZE),
.CORE_TAG_WIDTH (`DDRAM_TAG_WIDTH),
.DRAM_TAG_WIDTH (`L2DRAM_TAG_WIDTH)
VX_mem_arb #(
.NUM_REQUESTS (`L2NUM_REQUESTS),
.WORD_SIZE (`L2BANK_LINE_SIZE),
.TAG_IN_WIDTH (`DDRAM_TAG_WIDTH),
.TAG_OUT_WIDTH (`L2DRAM_TAG_WIDTH)
) dram_arb (
.clk (clk),
.reset (reset),
.clk (clk),
.reset (reset),
// Core request
.in_dram_req_valid (arb_core_req_valid),
.in_dram_req_rw (arb_core_req_rw),
.in_dram_req_byteen (arb_core_req_byteen),
.in_dram_req_addr (arb_core_req_addr),
.in_dram_req_data (arb_core_req_data),
.in_dram_req_tag (arb_core_req_tag),
.in_dram_req_ready (arb_core_req_ready),
.in_mem_req_valid (arb_dram_req_valid),
.in_mem_req_rw (arb_dram_req_rw),
.in_mem_req_byteen (arb_dram_req_byteen),
.in_mem_req_addr (arb_dram_req_addr),
.in_mem_req_data (arb_dram_req_data),
.in_mem_req_tag (arb_dram_req_tag),
.in_mem_req_ready (arb_dram_req_ready),
// Core response
.in_dram_rsp_valid (arb_core_rsp_valid),
.in_dram_rsp_data (arb_core_rsp_data),
.in_dram_rsp_tag (arb_core_rsp_tag),
.in_dram_rsp_ready (arb_core_rsp_ready),
.in_mem_rsp_valid (arb_dram_rsp_valid),
.in_mem_rsp_data (arb_dram_rsp_data),
.in_mem_rsp_tag (arb_dram_rsp_tag),
.in_mem_rsp_ready (arb_dram_rsp_ready),
// DRAM request
.out_dram_req_valid (dram_req_valid),
.out_dram_req_rw (dram_req_rw),
.out_dram_req_byteen (dram_req_byteen),
.out_dram_req_addr (dram_req_addr),
.out_dram_req_data (dram_req_data),
.out_dram_req_tag (dram_req_tag),
.out_dram_req_ready (dram_req_ready),
.out_mem_req_valid (dram_req_valid),
.out_mem_req_rw (dram_req_rw),
.out_mem_req_byteen (dram_req_byteen),
.out_mem_req_addr (dram_req_addr),
.out_mem_req_data (dram_req_data),
.out_mem_req_tag (dram_req_tag),
.out_mem_req_ready (dram_req_ready),
// DRAM response
.out_dram_rsp_valid (dram_rsp_valid),
.out_dram_rsp_tag (dram_rsp_tag),
.out_dram_rsp_data (dram_rsp_data),
.out_dram_rsp_ready (dram_rsp_ready)
.out_mem_rsp_valid (dram_rsp_valid),
.out_mem_rsp_tag (dram_rsp_tag),
.out_mem_rsp_data (dram_rsp_data),
.out_mem_rsp_ready (dram_rsp_ready)
);
end

View File

@@ -13,43 +13,43 @@ module Vortex_Socket (
// DRAM request
output wire dram_req_valid,
output wire dram_req_rw,
output wire[`VX_DRAM_BYTEEN_WIDTH-1:0] dram_req_byteen,
output wire[`VX_DRAM_ADDR_WIDTH-1:0] dram_req_addr,
output wire[`VX_DRAM_LINE_WIDTH-1:0] dram_req_data,
output wire[`VX_DRAM_TAG_WIDTH-1:0] dram_req_tag,
output wire [`VX_DRAM_BYTEEN_WIDTH-1:0] dram_req_byteen,
output wire [`VX_DRAM_ADDR_WIDTH-1:0] dram_req_addr,
output wire [`VX_DRAM_LINE_WIDTH-1:0] dram_req_data,
output wire [`VX_DRAM_TAG_WIDTH-1:0] dram_req_tag,
input wire dram_req_ready,
// DRAM response
input wire dram_rsp_valid,
input wire[`VX_DRAM_LINE_WIDTH-1:0] dram_rsp_data,
input wire[`VX_DRAM_TAG_WIDTH-1:0] dram_rsp_tag,
input wire [`VX_DRAM_LINE_WIDTH-1:0] dram_rsp_data,
input wire [`VX_DRAM_TAG_WIDTH-1:0] dram_rsp_tag,
output wire dram_rsp_ready,
// Snoop request
input wire snp_req_valid,
input wire[`VX_DRAM_ADDR_WIDTH-1:0] snp_req_addr,
input wire [`VX_DRAM_ADDR_WIDTH-1:0] snp_req_addr,
input wire snp_req_invalidate,
input wire[`VX_SNP_TAG_WIDTH-1:0] snp_req_tag,
input wire [`VX_SNP_TAG_WIDTH-1:0] snp_req_tag,
output wire snp_req_ready,
// Snoop response
output wire snp_rsp_valid,
output wire[`VX_SNP_TAG_WIDTH-1:0] snp_rsp_tag,
output wire [`VX_SNP_TAG_WIDTH-1:0] snp_rsp_tag,
input wire snp_rsp_ready,
// I/O request
output wire io_req_valid,
output wire io_req_rw,
output wire[3:0] io_req_byteen,
output wire[29:0] io_req_addr,
output wire[31:0] io_req_data,
output wire[`VX_CORE_TAG_WIDTH-1:0] io_req_tag,
output wire [3:0] io_req_byteen,
output wire [29:0] io_req_addr,
output wire [31:0] io_req_data,
output wire [`VX_CORE_TAG_WIDTH-1:0] io_req_tag,
input wire io_req_ready,
// I/O response
input wire io_rsp_valid,
input wire[31:0] io_rsp_data,
input wire[`VX_CORE_TAG_WIDTH-1:0] io_rsp_tag,
input wire [31:0] io_rsp_data,
input wire [`VX_CORE_TAG_WIDTH-1:0] io_rsp_tag,
output wire io_rsp_ready,
// Status
@@ -111,42 +111,44 @@ module Vortex_Socket (
end else begin
wire[`NUM_CLUSTERS-1:0] per_cluster_dram_req_valid;
wire[`NUM_CLUSTERS-1:0] per_cluster_dram_req_rw;
wire[`NUM_CLUSTERS-1:0][`L2DRAM_BYTEEN_WIDTH-1:0] per_cluster_dram_req_byteen;
wire[`NUM_CLUSTERS-1:0][`L2DRAM_ADDR_WIDTH-1:0] per_cluster_dram_req_addr;
wire[`NUM_CLUSTERS-1:0][`L2DRAM_LINE_WIDTH-1:0] per_cluster_dram_req_data;
wire[`NUM_CLUSTERS-1:0][`L2DRAM_TAG_WIDTH-1:0] per_cluster_dram_req_tag;
wire [`NUM_CLUSTERS-1:0] per_cluster_dram_req_valid;
wire [`NUM_CLUSTERS-1:0] per_cluster_dram_req_rw;
wire [`NUM_CLUSTERS-1:0][`L2DRAM_BYTEEN_WIDTH-1:0] per_cluster_dram_req_byteen;
wire [`NUM_CLUSTERS-1:0][`L2DRAM_ADDR_WIDTH-1:0] per_cluster_dram_req_addr;
wire [`NUM_CLUSTERS-1:0][`L2DRAM_LINE_WIDTH-1:0] per_cluster_dram_req_data;
wire [`NUM_CLUSTERS-1:0][`L2DRAM_TAG_WIDTH-1:0] per_cluster_dram_req_tag;
wire l3_core_req_ready;
wire[`NUM_CLUSTERS-1:0] per_cluster_dram_rsp_valid;
wire[`NUM_CLUSTERS-1:0][`L3DRAM_LINE_WIDTH-1:0] per_cluster_dram_rsp_data;
wire[`NUM_CLUSTERS-1:0][`L3DRAM_TAG_WIDTH-1:0] per_cluster_dram_rsp_tag;
wire[`NUM_CLUSTERS-1:0] per_cluster_dram_rsp_ready;
wire [`NUM_CLUSTERS-1:0] per_cluster_dram_rsp_valid;
wire [`NUM_CLUSTERS-1:0][`L3DRAM_LINE_WIDTH-1:0] per_cluster_dram_rsp_data;
wire [`NUM_CLUSTERS-1:0][`L3DRAM_TAG_WIDTH-1:0] per_cluster_dram_rsp_tag;
wire [`NUM_CLUSTERS-1:0] per_cluster_dram_rsp_ready;
wire[`NUM_CLUSTERS-1:0] per_cluster_snp_req_valid;
wire[`NUM_CLUSTERS-1:0][`L2DRAM_ADDR_WIDTH-1:0] per_cluster_snp_req_addr;
wire[`NUM_CLUSTERS-1:0] per_cluster_snp_req_invalidate;
wire[`NUM_CLUSTERS-1:0][`L2SNP_TAG_WIDTH-1:0] per_cluster_snp_req_tag;
wire[`NUM_CLUSTERS-1:0] per_cluster_snp_req_ready;
wire [`NUM_CLUSTERS-1:0] per_cluster_snp_req_valid;
wire [`NUM_CLUSTERS-1:0][`L2DRAM_ADDR_WIDTH-1:0] per_cluster_snp_req_addr;
wire [`NUM_CLUSTERS-1:0] per_cluster_snp_req_invalidate;
wire [`NUM_CLUSTERS-1:0][`L2SNP_TAG_WIDTH-1:0] per_cluster_snp_req_tag;
wire [`NUM_CLUSTERS-1:0] per_cluster_snp_req_ready;
wire[`NUM_CLUSTERS-1:0] per_cluster_snp_rsp_valid;
wire[`NUM_CLUSTERS-1:0][`L2SNP_TAG_WIDTH-1:0] per_cluster_snp_rsp_tag;
wire[`NUM_CLUSTERS-1:0] per_cluster_snp_rsp_ready;
wire [`NUM_CLUSTERS-1:0] per_cluster_snp_rsp_valid;
wire [`NUM_CLUSTERS-1:0][`L2SNP_TAG_WIDTH-1:0] per_cluster_snp_rsp_tag;
wire [`NUM_CLUSTERS-1:0] per_cluster_snp_rsp_ready;
`IGNORE_WARNINGS_BEGIN
wire[`NUM_CLUSTERS-1:0] per_cluster_io_req_valid;
wire[`NUM_CLUSTERS-1:0] per_cluster_io_req_rw;
wire[`NUM_CLUSTERS-1:0][3:0] per_cluster_io_req_byteen;
wire[`NUM_CLUSTERS-1:0][29:0] per_cluster_io_req_addr;
wire[`NUM_CLUSTERS-1:0][31:0] per_cluster_io_req_data;
wire[`NUM_CLUSTERS-1:0][`DCORE_TAG_WIDTH-1:0] per_cluster_io_req_tag;
wire [`NUM_CLUSTERS-1:0] per_cluster_io_req_valid;
wire [`NUM_CLUSTERS-1:0] per_cluster_io_req_rw;
wire [`NUM_CLUSTERS-1:0][3:0] per_cluster_io_req_byteen;
wire [`NUM_CLUSTERS-1:0][29:0] per_cluster_io_req_addr;
wire [`NUM_CLUSTERS-1:0][31:0] per_cluster_io_req_data;
wire [`NUM_CLUSTERS-1:0][`L2CORE_TAG_WIDTH-1:0] per_cluster_io_req_tag;
wire [`NUM_CLUSTERS-1:0] per_cluster_io_req_ready;
wire[`NUM_CLUSTERS-1:0] per_cluster_io_rsp_ready;
`IGNORE_WARNINGS_END
wire [`NUM_CLUSTERS-1:0] per_cluster_io_rsp_valid;
wire [`NUM_CLUSTERS-1:0][`L2CORE_TAG_WIDTH-1:0] per_cluster_io_rsp_tag;
wire [`NUM_CLUSTERS-1:0][31:0] per_cluster_io_rsp_data;
wire [`NUM_CLUSTERS-1:0] per_cluster_io_rsp_ready;
wire[`NUM_CLUSTERS-1:0] per_cluster_busy;
wire[`NUM_CLUSTERS-1:0] per_cluster_ebreak;
wire [`NUM_CLUSTERS-1:0] per_cluster_busy;
wire [`NUM_CLUSTERS-1:0] per_cluster_ebreak;
genvar i;
for (i = 0; i < `NUM_CLUSTERS; i++) begin
@@ -190,62 +192,93 @@ module Vortex_Socket (
.io_req_addr (per_cluster_io_req_addr [i]),
.io_req_data (per_cluster_io_req_data [i]),
.io_req_tag (per_cluster_io_req_tag [i]),
.io_req_ready (io_req_ready),
.io_req_ready (per_cluster_io_req_ready [i]),
.io_rsp_valid (io_rsp_valid),
.io_rsp_data (io_rsp_data),
.io_rsp_tag (io_rsp_tag),
.io_rsp_valid (per_cluster_io_rsp_valid [i]),
.io_rsp_data (per_cluster_io_rsp_data [i]),
.io_rsp_tag (per_cluster_io_rsp_tag [i]),
.io_rsp_ready (per_cluster_io_rsp_ready [i]),
.busy (per_cluster_busy [i]),
.ebreak (per_cluster_ebreak [i])
);
end
end
assign io_req_valid = per_cluster_io_req_valid[0];
assign io_req_rw = per_cluster_io_req_rw[0];
assign io_req_byteen = per_cluster_io_req_byteen[0];
assign io_req_addr = per_cluster_io_req_addr[0];
assign io_req_data = per_cluster_io_req_data[0];
assign io_req_tag = per_cluster_io_req_tag[0];
VX_mem_arb #(
.NUM_REQUESTS (`NUM_CLUSTERS),
.WORD_SIZE (4),
.TAG_IN_WIDTH (`L2CORE_TAG_WIDTH),
.TAG_OUT_WIDTH (`L3CORE_TAG_WIDTH)
) io_arb (
.clk (clk),
.reset (reset),
assign io_rsp_ready = per_cluster_io_rsp_ready[0];
// input requests
.in_mem_req_valid (per_cluster_io_req_valid),
.in_mem_req_rw (per_cluster_io_req_rw),
.in_mem_req_byteen (per_cluster_io_req_byteen),
.in_mem_req_addr (per_cluster_io_req_addr),
.in_mem_req_data (per_cluster_io_req_data),
.in_mem_req_tag (per_cluster_io_req_tag),
.in_mem_req_ready (per_cluster_io_req_ready),
// input responses
.in_mem_rsp_valid (per_cluster_io_rsp_valid),
.in_mem_rsp_data (per_cluster_io_rsp_data),
.in_mem_rsp_tag (per_cluster_io_rsp_tag),
.in_mem_rsp_ready (per_cluster_io_rsp_ready),
// output request
.out_mem_req_valid (io_req_valid),
.out_mem_req_rw (io_req_rw),
.out_mem_req_byteen (io_req_byteen),
.out_mem_req_addr (io_req_addr),
.out_mem_req_data (io_req_data),
.out_mem_req_tag (io_req_tag),
.out_mem_req_ready (io_req_ready),
// output response
.out_mem_rsp_valid (io_rsp_valid),
.out_mem_rsp_tag (io_rsp_tag),
.out_mem_rsp_data (io_rsp_data),
.out_mem_rsp_ready (io_rsp_ready)
);
assign busy = (| per_cluster_busy);
assign ebreak = (& per_cluster_ebreak);
// L3 Cache ///////////////////////////////////////////////////////////
wire[`L3NUM_REQUESTS-1:0] l3_core_req_valid;
wire[`L3NUM_REQUESTS-1:0] l3_core_req_rw;
wire[`L3NUM_REQUESTS-1:0][`L2DRAM_BYTEEN_WIDTH-1:0] l3_core_req_byteen;
wire[`L3NUM_REQUESTS-1:0][`L2DRAM_ADDR_WIDTH-1:0] l3_core_req_addr;
wire[`L3NUM_REQUESTS-1:0][`L2DRAM_LINE_WIDTH-1:0] l3_core_req_data;
wire[`L3NUM_REQUESTS-1:0][`L2DRAM_TAG_WIDTH-1:0] l3_core_req_tag;
wire [`L3NUM_REQUESTS-1:0] l3_core_req_valid;
wire [`L3NUM_REQUESTS-1:0] l3_core_req_rw;
wire [`L3NUM_REQUESTS-1:0][`L2DRAM_BYTEEN_WIDTH-1:0] l3_core_req_byteen;
wire [`L3NUM_REQUESTS-1:0][`L2DRAM_ADDR_WIDTH-1:0] l3_core_req_addr;
wire [`L3NUM_REQUESTS-1:0][`L2DRAM_LINE_WIDTH-1:0] l3_core_req_data;
wire [`L3NUM_REQUESTS-1:0][`L2DRAM_TAG_WIDTH-1:0] l3_core_req_tag;
wire[`L3NUM_REQUESTS-1:0] l3_core_rsp_valid;
wire[`L3NUM_REQUESTS-1:0][`L2DRAM_LINE_WIDTH-1:0] l3_core_rsp_data;
wire[`L3NUM_REQUESTS-1:0][`L2DRAM_TAG_WIDTH-1:0] l3_core_rsp_tag;
wire l3_core_rsp_ready;
wire [`L3NUM_REQUESTS-1:0] l3_core_rsp_valid;
wire [`L3NUM_REQUESTS-1:0][`L2DRAM_LINE_WIDTH-1:0] l3_core_rsp_data;
wire [`L3NUM_REQUESTS-1:0][`L2DRAM_TAG_WIDTH-1:0] l3_core_rsp_tag;
wire l3_core_rsp_ready;
wire[`NUM_CLUSTERS-1:0] l3_snp_fwdout_valid;
wire[`NUM_CLUSTERS-1:0][`L2DRAM_ADDR_WIDTH-1:0] l3_snp_fwdout_addr;
wire[`NUM_CLUSTERS-1:0] l3_snp_fwdout_invalidate;
wire[`NUM_CLUSTERS-1:0][`L2SNP_TAG_WIDTH-1:0] l3_snp_fwdout_tag;
wire[`NUM_CLUSTERS-1:0] l3_snp_fwdout_ready;
wire [`NUM_CLUSTERS-1:0] l3_snp_fwdout_valid;
wire [`NUM_CLUSTERS-1:0][`L2DRAM_ADDR_WIDTH-1:0] l3_snp_fwdout_addr;
wire [`NUM_CLUSTERS-1:0] l3_snp_fwdout_invalidate;
wire [`NUM_CLUSTERS-1:0][`L2SNP_TAG_WIDTH-1:0] l3_snp_fwdout_tag;
wire [`NUM_CLUSTERS-1:0] l3_snp_fwdout_ready;
wire[`NUM_CLUSTERS-1:0] l3_snp_fwdin_valid;
wire[`NUM_CLUSTERS-1:0][`L2SNP_TAG_WIDTH-1:0] l3_snp_fwdin_tag;
wire[`NUM_CLUSTERS-1:0] l3_snp_fwdin_ready;
wire [`NUM_CLUSTERS-1:0] l3_snp_fwdin_valid;
wire [`NUM_CLUSTERS-1:0][`L2SNP_TAG_WIDTH-1:0] l3_snp_fwdin_tag;
wire [`NUM_CLUSTERS-1:0] l3_snp_fwdin_ready;
for (i = 0; i < `L3NUM_REQUESTS; i++) begin
// Core Request
assign l3_core_req_valid [i] = per_cluster_dram_req_valid [i];
assign l3_core_req_rw [i] = per_cluster_dram_req_rw [i];
assign l3_core_req_byteen[i] = per_cluster_dram_req_byteen[i];
assign l3_core_req_addr [i] = per_cluster_dram_req_addr [i];
assign l3_core_req_tag [i] = per_cluster_dram_req_tag [i];
assign l3_core_req_data [i] = per_cluster_dram_req_data [i];
assign l3_core_req_valid [i] = per_cluster_dram_req_valid [i];
assign l3_core_req_rw [i] = per_cluster_dram_req_rw [i];
assign l3_core_req_byteen [i] = per_cluster_dram_req_byteen[i];
assign l3_core_req_addr [i] = per_cluster_dram_req_addr [i];
assign l3_core_req_tag [i] = per_cluster_dram_req_tag [i];
assign l3_core_req_data [i] = per_cluster_dram_req_data [i];
// Core Response
assign per_cluster_dram_rsp_valid [i] = l3_core_rsp_valid [i] && l3_core_rsp_ready;
@@ -268,31 +301,31 @@ module Vortex_Socket (
assign l3_core_rsp_ready = (& per_cluster_dram_rsp_ready);
VX_cache #(
.CACHE_ID (0),
.CACHE_SIZE (`L3CACHE_SIZE),
.BANK_LINE_SIZE (`L3BANK_LINE_SIZE),
.NUM_BANKS (`L3NUM_BANKS),
.WORD_SIZE (`L3WORD_SIZE),
.NUM_REQUESTS (`L3NUM_REQUESTS),
.STAGE_1_CYCLES (`L3STAGE_1_CYCLES),
.CREQ_SIZE (`L3CREQ_SIZE),
.MRVQ_SIZE (`L3MRVQ_SIZE),
.DFPQ_SIZE (`L3DFPQ_SIZE),
.SNRQ_SIZE (`L3SNRQ_SIZE),
.CWBQ_SIZE (`L3CWBQ_SIZE),
.DWBQ_SIZE (`L3DWBQ_SIZE),
.DFQQ_SIZE (`L3DFQQ_SIZE),
.PRFQ_SIZE (`L3PRFQ_SIZE),
.PRFQ_STRIDE (`L3PRFQ_STRIDE),
.DRAM_ENABLE (1),
.WRITE_ENABLE (1),
.SNOOP_FORWARDING (1),
.CORE_TAG_WIDTH (`L2DRAM_TAG_WIDTH),
.CORE_TAG_ID_BITS (0),
.DRAM_TAG_WIDTH (`L3DRAM_TAG_WIDTH),
.NUM_SNP_REQUESTS (`NUM_CLUSTERS),
.SNP_REQ_TAG_WIDTH (`L3SNP_TAG_WIDTH),
.SNP_FWD_TAG_WIDTH (`L2SNP_TAG_WIDTH)
.CACHE_ID (0),
.CACHE_SIZE (`L3CACHE_SIZE),
.BANK_LINE_SIZE (`L3BANK_LINE_SIZE),
.NUM_BANKS (`L3NUM_BANKS),
.WORD_SIZE (`L3WORD_SIZE),
.NUM_REQUESTS (`L3NUM_REQUESTS),
.STAGE_1_CYCLES (`L3STAGE_1_CYCLES),
.CREQ_SIZE (`L3CREQ_SIZE),
.MRVQ_SIZE (`L3MRVQ_SIZE),
.DFPQ_SIZE (`L3DFPQ_SIZE),
.SNRQ_SIZE (`L3SNRQ_SIZE),
.CWBQ_SIZE (`L3CWBQ_SIZE),
.DWBQ_SIZE (`L3DWBQ_SIZE),
.DFQQ_SIZE (`L3DFQQ_SIZE),
.PRFQ_SIZE (`L3PRFQ_SIZE),
.PRFQ_STRIDE (`L3PRFQ_STRIDE),
.DRAM_ENABLE (1),
.WRITE_ENABLE (1),
.SNOOP_FORWARDING (1),
.CORE_TAG_WIDTH (`L2DRAM_TAG_WIDTH),
.CORE_TAG_ID_BITS (0),
.DRAM_TAG_WIDTH (`L3DRAM_TAG_WIDTH),
.NUM_SNP_REQUESTS (`NUM_CLUSTERS),
.SNP_REQ_TAG_WIDTH (`L3SNP_TAG_WIDTH),
.SNP_FWD_TAG_WIDTH (`L2SNP_TAG_WIDTH)
) gpu_l3cache (
.clk (clk),
.reset (reset),