wu-arch/kernels
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

minor update

Browse Source
This commit is contained in:
Blaise Tine
2020-06-03 03:05:45 -07:00
parent 9b186dcc6e
commit 04fc34b848
12 changed files with 165 additions and 176 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
driver/opae/vortex.cpp
View File

@@ -246,8 +246,12 @@ extern int vx_ready_wait(vx_device_h hdevice, long long timeout) {
for (;;) { for (;;) {
CHECK_RES(fpgaReadMMIO64(device->fpga, 0, MMIO_CSR_STATUS, &data)); CHECK_RES(fpgaReadMMIO64(device->fpga, 0, MMIO_CSR_STATUS, &data));
if (0 == data || 0 == timeout) if (0 == data || 0 == timeout) {
if (data != 0) {
fprintf(stdout, "ready-wait timed out: status=%ld\n", data);
}
break; break;
}
nanosleep(&sleep_time, nullptr); nanosleep(&sleep_time, nullptr);
timeout -= sleep_time_ms; timeout -= sleep_time_ms;
}; };

2
hw/opae/sources.txt
View File

@@ -7,7 +7,7 @@ vortex_afu.json
+define+NUM_CORES=2 +define+NUM_CORES=2
+define+NUM_WARPS=4 +define+NUM_WARPS=4
+define+NUM_THREADS=4 +define+NUM_THREADS=4
+define+L2_ENABLE=0 +define+L2_ENABLE=1
+define+DNUM_BANKS=4 +define+DNUM_BANKS=4
+define+INUM_BANKS=1 +define+INUM_BANKS=1

32
hw/rtl/VX_config.vh
View File

@@ -12,7 +12,7 @@
`endif `endif
`ifndef NUM_WARPS `ifndef NUM_WARPS
`define NUM_WARPS 8 `define NUM_WARPS 4
`endif `endif
`ifndef NUM_THREADS `ifndef NUM_THREADS
@@ -87,7 +87,7 @@
// Number of banks {1, 2, 4, 8,...} // Number of banks {1, 2, 4, 8,...}
`ifndef DNUM_BANKS `ifndef DNUM_BANKS
`define DNUM_BANKS 8 `define DNUM_BANKS 4
`endif `endif
// Size of a word in bytes // Size of a word in bytes
@@ -107,12 +107,12 @@
// Miss Reserv Queue Knob // Miss Reserv Queue Knob
`ifndef DMRVQ_SIZE `ifndef DMRVQ_SIZE
`define DMRVQ_SIZE (`NUM_WARPS*`NUM_THREADS) `define DMRVQ_SIZE `MAX(`NUM_WARPS*`NUM_THREADS, 8)
`endif `endif
// Dram Fill Rsp Queue Size // Dram Fill Rsp Queue Size
`ifndef DDFPQ_SIZE `ifndef DDFPQ_SIZE
`define DDFPQ_SIZE 32 `define DDFPQ_SIZE 16
`endif `endif
// Snoop Req Queue Size // Snoop Req Queue Size
@@ -137,7 +137,7 @@
// Prefetcher // Prefetcher
`ifndef DPRFQ_SIZE `ifndef DPRFQ_SIZE
`define DPRFQ_SIZE 32 `define DPRFQ_SIZE 16
`endif `endif
`ifndef DPRFQ_STRIDE `ifndef DPRFQ_STRIDE
@@ -178,12 +178,12 @@
// Miss Reserv Queue Knob // Miss Reserv Queue Knob
`ifndef IMRVQ_SIZE `ifndef IMRVQ_SIZE
`define IMRVQ_SIZE `ICREQ_SIZE `define IMRVQ_SIZE `MAX(`ICREQ_SIZE, 8)
`endif `endif
// Dram Fill Rsp Queue Size // Dram Fill Rsp Queue Size
`ifndef IDFPQ_SIZE `ifndef IDFPQ_SIZE
`define IDFPQ_SIZE 32 `define IDFPQ_SIZE 16
`endif `endif
// Core Writeback Queue Size // Core Writeback Queue Size
@@ -203,7 +203,7 @@
// Prefetcher // Prefetcher
`ifndef IPRFQ_SIZE `ifndef IPRFQ_SIZE
`define IPRFQ_SIZE 32 `define IPRFQ_SIZE 16
`endif `endif
`ifndef IPRFQ_STRIDE `ifndef IPRFQ_STRIDE
@@ -276,17 +276,17 @@
// Core Request Queue Size // Core Request Queue Size
`ifndef L2CREQ_SIZE `ifndef L2CREQ_SIZE
`define L2CREQ_SIZE 32 `define L2CREQ_SIZE 16
`endif `endif
// Miss Reserv Queue Knob // Miss Reserv Queue Knob
`ifndef L2MRVQ_SIZE `ifndef L2MRVQ_SIZE
`define L2MRVQ_SIZE 32 `define L2MRVQ_SIZE `MAX(`L2CREQ_SIZE, 8)
`endif `endif
// Dram Fill Rsp Queue Size // Dram Fill Rsp Queue Size
`ifndef L2DFPQ_SIZE `ifndef L2DFPQ_SIZE
`define L2DFPQ_SIZE 32 `define L2DFPQ_SIZE 16
`endif `endif
// Snoop Req Queue Size // Snoop Req Queue Size
@@ -311,7 +311,7 @@
// Prefetcher // Prefetcher
`ifndef L2PRFQ_SIZE `ifndef L2PRFQ_SIZE
`define L2PRFQ_SIZE 32 `define L2PRFQ_SIZE 16
`endif `endif
`ifndef L2PRFQ_STRIDE `ifndef L2PRFQ_STRIDE
@@ -347,17 +347,17 @@
// Core Request Queue Size // Core Request Queue Size
`ifndef L3CREQ_SIZE `ifndef L3CREQ_SIZE
`define L3CREQ_SIZE 32 `define L3CREQ_SIZE 16
`endif `endif
// Miss Reserv Queue Knob // Miss Reserv Queue Knob
`ifndef L3MRVQ_SIZE `ifndef L3MRVQ_SIZE
`define L3MRVQ_SIZE `L3CREQ_SIZE `define L3MRVQ_SIZE `MAX(`L3CREQ_SIZE, 8)
`endif `endif
// Dram Fill Rsp Queue Size // Dram Fill Rsp Queue Size
`ifndef L3DFPQ_SIZE `ifndef L3DFPQ_SIZE
`define L3DFPQ_SIZE 32 `define L3DFPQ_SIZE 16
`endif `endif
// Snoop Req Queue Size // Snoop Req Queue Size
@@ -382,7 +382,7 @@
// Prefetcher // Prefetcher
`ifndef L3PRFQ_SIZE `ifndef L3PRFQ_SIZE
`define L3PRFQ_SIZE 32 `define L3PRFQ_SIZE 16
`endif `endif
`ifndef L3PRFQ_STRIDE `ifndef L3PRFQ_STRIDE

14
hw/rtl/VX_dmem_ctrl.v
View File

@@ -60,13 +60,13 @@ module VX_dmem_ctrl # (
.NUM_REQUESTS (`SNUM_REQUESTS), .NUM_REQUESTS (`SNUM_REQUESTS),
.STAGE_1_CYCLES (`SSTAGE_1_CYCLES), .STAGE_1_CYCLES (`SSTAGE_1_CYCLES),
.CREQ_SIZE (`SCREQ_SIZE), .CREQ_SIZE (`SCREQ_SIZE),
.MRVQ_SIZE (1), .MRVQ_SIZE (8),
.DFPQ_SIZE (0), .DFPQ_SIZE (1),
.SNRQ_SIZE (0), .SNRQ_SIZE (1),
.CWBQ_SIZE (`SCWBQ_SIZE), .CWBQ_SIZE (`SCWBQ_SIZE),
.DWBQ_SIZE (0), .DWBQ_SIZE (1),
.DFQQ_SIZE (0), .DFQQ_SIZE (1),
.PRFQ_SIZE (0), .PRFQ_SIZE (1),
.PRFQ_STRIDE (0), .PRFQ_STRIDE (0),
.SNOOP_FORWARDING (0), .SNOOP_FORWARDING (0),
.DRAM_ENABLE (0), .DRAM_ENABLE (0),
@@ -223,7 +223,7 @@ module VX_dmem_ctrl # (
.CREQ_SIZE (`ICREQ_SIZE), .CREQ_SIZE (`ICREQ_SIZE),
.MRVQ_SIZE (`IMRVQ_SIZE), .MRVQ_SIZE (`IMRVQ_SIZE),
.DFPQ_SIZE (`IDFPQ_SIZE), .DFPQ_SIZE (`IDFPQ_SIZE),
.SNRQ_SIZE (0), .SNRQ_SIZE (1),
.CWBQ_SIZE (`ICWBQ_SIZE), .CWBQ_SIZE (`ICWBQ_SIZE),
.DWBQ_SIZE (`IDWBQ_SIZE), .DWBQ_SIZE (`IDWBQ_SIZE),
.DFQQ_SIZE (`IDFQQ_SIZE), .DFQQ_SIZE (`IDFQQ_SIZE),

6
hw/rtl/cache/VX_cache.v vendored
View File

@@ -125,13 +125,13 @@ module VX_cache #(
`DEBUG_BLOCK( `DEBUG_BLOCK(
wire[31:0] debug_core_req_use_pc; wire[31:0] debug_core_req_use_pc;
wire[1:0] debug_core_req_wb; wire[1:0] debug_core_req_wb;
wire[2:0] debug_core_req_rmask;
wire[4:0] debug_core_req_rd; wire[4:0] debug_core_req_rd;
wire[`NW_BITS-1:0] debug_core_req_warp_num; wire[`NW_BITS-1:0] debug_core_req_warp_num;
wire[`LOG2UP(CREQ_SIZE)-1:0] debug_core_req_idx;
if (WORD_SIZE != `GLOBAL_BLOCK_SIZE) begin if (WORD_SIZE != `GLOBAL_BLOCK_SIZE) begin
assign {debug_core_req_use_pc, debug_core_req_wb, debug_core_req_rmask, debug_core_req_rd, debug_core_req_warp_num} = core_req_tag[0]; assign {debug_core_req_use_pc, debug_core_req_wb, debug_core_req_rd, debug_core_req_warp_num, debug_core_req_idx} = core_req_tag[0];
end end
) )
wire [NUM_BANKS-1:0][NUM_REQUESTS-1:0] per_bank_valid; wire [NUM_BANKS-1:0][NUM_REQUESTS-1:0] per_bank_valid;

4
hw/rtl/cache/VX_cache_miss_resrv.v vendored
View File

@@ -64,8 +64,10 @@ module VX_cache_miss_resrv #(
reg [`LOG2UP(MRVQ_SIZE+1)-1:0] size; reg [`LOG2UP(MRVQ_SIZE+1)-1:0] size;
`STATIC_ASSERT(MRVQ_SIZE > 5, "invalid size");
assign miss_resrv_full = (size == $bits(size)'(MRVQ_SIZE)); assign miss_resrv_full = (size == $bits(size)'(MRVQ_SIZE));
assign miss_resrv_stop = (size > $bits(size)'(MRVQ_SIZE-5)); assign miss_resrv_stop = (size > $bits(size)'(MRVQ_SIZE-1));
wire enqueue_possible = !miss_resrv_full; wire enqueue_possible = !miss_resrv_full;
wire [`LOG2UP(MRVQ_SIZE)-1:0] enqueue_index = tail_ptr; wire [`LOG2UP(MRVQ_SIZE)-1:0] enqueue_index = tail_ptr;

260
hw/rtl/libs/VX_generic_queue.v
View File

@@ -15,166 +15,150 @@ module VX_generic_queue #(
output wire full, output wire full,
output wire [`LOG2UP(SIZE+1)-1:0] size output wire [`LOG2UP(SIZE+1)-1:0] size
); );
`STATIC_ASSERT(0 == SIZE || `ISPOW2(SIZE), "must be 0 or power of 2!"); `STATIC_ASSERT(`ISPOW2(SIZE), "must be 0 or power of 2!");
if (SIZE == 0) begin reg [`LOG2UP(SIZE+1)-1:0] size_r;
wire reading;
wire writing;
assign empty = 1; assign reading = pop && !empty;
assign data_out = 0; assign writing = push && !full;
assign full = 0;
assign size = 0;
`UNUSED_VAR (clk) if (SIZE == 1) begin // (SIZE == 1)
`UNUSED_VAR (reset)
`UNUSED_VAR (push) reg [DATAW-1:0] head_r;
`UNUSED_VAR (pop)
`UNUSED_VAR (data_in) always @(posedge clk) begin
if (reset) begin
head_r <= 0;
size_r <= 0;
end else begin
if (writing && !reading) begin
size_r <= 1;
end else if (reading && !writing) begin
size_r <= 0;
end
if (writing) begin
head_r <= data_in;
end
end
end
assign data_out = head_r;
assign empty = (size_r == 0);
assign full = (size_r != 0);
assign size = size_r;
end else begin // (SIZE > 1)
end else begin // (SIZE > 0)
`ifdef QUEUE_FORCE_MLAB `ifdef QUEUE_FORCE_MLAB
(* syn_ramstyle = "mlab" *) reg [DATAW-1:0] data [SIZE-1:0]; (* syn_ramstyle = "mlab" *) reg [DATAW-1:0] data [SIZE-1:0];
`else `else
reg [DATAW-1:0] data [SIZE-1:0]; reg [DATAW-1:0] data [SIZE-1:0];
`endif `endif
reg [`LOG2UP(SIZE+1)-1:0] size_r; if (0 == BUFFERED_OUTPUT) begin
wire reading;
wire writing;
assign reading = pop && !empty; reg [`LOG2UP(SIZE):0] wr_ptr_r;
assign writing = push && !full; reg [`LOG2UP(SIZE):0] rd_ptr_r;
if (SIZE == 1) begin // (SIZE == 1) wire [`LOG2UP(SIZE)-1:0] wr_ptr_a = wr_ptr_r[`LOG2UP(SIZE)-1:0];
wire [`LOG2UP(SIZE)-1:0] rd_ptr_a = rd_ptr_r[`LOG2UP(SIZE)-1:0];
reg [DATAW-1:0] head_r;
always @(posedge clk) begin always @(posedge clk) begin
if (reset) begin if (reset) begin
head_r <= 0; rd_ptr_r <= 0;
size_r <= 0; wr_ptr_r <= 0;
size_r <= 0;
end else begin end else begin
if (writing && !reading) begin if (writing) begin
size_r <= 1; data[wr_ptr_a] <= data_in;
end else if (reading && !writing) begin wr_ptr_r <= wr_ptr_r + 1;
size_r <= 0;
if (!reading) begin
size_r <= size_r + 1;
end
end end
if (writing) begin if (reading) begin
head_r <= data_in; rd_ptr_r <= rd_ptr_r + 1;
if (!writing) begin
size_r <= size_r - 1;
end
end end
end
end
assign data_out = data[rd_ptr_a];
assign empty = (wr_ptr_r == rd_ptr_r);
assign full = (wr_ptr_a == rd_ptr_a) && (wr_ptr_r[`LOG2UP(SIZE)] != rd_ptr_r[`LOG2UP(SIZE)]);
assign size = size_r;
end else begin
reg [DATAW-1:0] head_r;
reg [DATAW-1:0] curr_r;
reg [`LOG2UP(SIZE)-1:0] wr_ptr_r;
reg [`LOG2UP(SIZE)-1:0] rd_ptr_r;
reg [`LOG2UP(SIZE)-1:0] rd_ptr_next_r;
reg empty_r;
reg full_r;
reg bypass_r;
always @(posedge clk) begin
if (reset) begin
wr_ptr_r <= 0;
rd_ptr_r <= 0;
rd_ptr_next_r <= 1;
empty_r <= 1;
full_r <= 0;
size_r <= 0;
end else begin
if (writing) begin
data[wr_ptr_r] <= data_in;
wr_ptr_r <= wr_ptr_r + 1;
if (!reading) begin
empty_r <= 0;
if (size_r == SIZE-1) begin
full_r <= 1;
end
size_r <= size_r + 1;
end
end
if (reading) begin
rd_ptr_r <= rd_ptr_next_r;
if (SIZE > 2) begin
rd_ptr_next_r <= rd_ptr_r + 2;
end else begin // (SIZE == 2);
rd_ptr_next_r <= ~rd_ptr_next_r;
end
if (!writing) begin
if (size_r == 1) begin
assert(rd_ptr_next_r == wr_ptr_r);
empty_r <= 1;
end;
full_r <= 0;
size_r <= size_r - 1;
end
end
bypass_r <= writing
&& (empty_r || ((1 == size_r) && reading)); // empty or about to go empty
curr_r <= data_in;
head_r <= data[reading ? rd_ptr_next_r : rd_ptr_r];
end end
end end
assign data_out = head_r; assign data_out = bypass_r ? curr_r : head_r;
assign empty = (size_r == 0); assign empty = empty_r;
assign full = (size_r != 0); assign full = full_r;
assign size = size_r; assign size = size_r;
end else begin // (SIZE > 1)
if (0 == BUFFERED_OUTPUT) begin
reg [`LOG2UP(SIZE):0] wr_ptr_r;
reg [`LOG2UP(SIZE):0] rd_ptr_r;
wire [`LOG2UP(SIZE)-1:0] wr_ptr_a = wr_ptr_r[`LOG2UP(SIZE)-1:0];
wire [`LOG2UP(SIZE)-1:0] rd_ptr_a = rd_ptr_r[`LOG2UP(SIZE)-1:0];
always @(posedge clk) begin
if (reset) begin
rd_ptr_r <= 0;
wr_ptr_r <= 0;
size_r <= 0;
end else begin
if (writing) begin
data[wr_ptr_a] <= data_in;
wr_ptr_r <= wr_ptr_r + 1;
if (!reading) begin
size_r <= size_r + 1;
end
end
if (reading) begin
rd_ptr_r <= rd_ptr_r + 1;
if (!writing) begin
size_r <= size_r - 1;
end
end
end
end
assign data_out = data[rd_ptr_a];
assign empty = (wr_ptr_r == rd_ptr_r);
assign full = (wr_ptr_a == rd_ptr_a) && (wr_ptr_r[`LOG2UP(SIZE)] != rd_ptr_r[`LOG2UP(SIZE)]);
assign size = size_r;
end else begin
reg [DATAW-1:0] head_r;
reg [DATAW-1:0] curr_r;
reg [`LOG2UP(SIZE)-1:0] wr_ptr_r;
reg [`LOG2UP(SIZE)-1:0] rd_ptr_r;
reg [`LOG2UP(SIZE)-1:0] rd_ptr_next_r;
reg empty_r;
reg full_r;
reg bypass_r;
always @(posedge clk) begin
if (reset) begin
wr_ptr_r <= 0;
rd_ptr_r <= 0;
rd_ptr_next_r <= 1;
empty_r <= 1;
full_r <= 0;
size_r <= 0;
end else begin
if (writing) begin
data[wr_ptr_r] <= data_in;
wr_ptr_r <= wr_ptr_r + 1;
if (!reading) begin
empty_r <= 0;
if (size_r == SIZE-1) begin
full_r <= 1;
end
size_r <= size_r + 1;
end
end
if (reading) begin
rd_ptr_r <= rd_ptr_next_r;
if (SIZE > 2) begin
rd_ptr_next_r <= rd_ptr_r + 2;
end else begin // (SIZE == 2);
rd_ptr_next_r <= ~rd_ptr_next_r;
end
if (!writing) begin
if (size_r == 1) begin
assert(rd_ptr_next_r == wr_ptr_r);
empty_r <= 1;
end;
full_r <= 0;
size_r <= size_r - 1;
end
end
bypass_r <= writing
&& (empty_r || ((1 == size_r) && reading)); // empty or about to go empty
curr_r <= data_in;
head_r <= data[reading ? rd_ptr_next_r : rd_ptr_r];
end
end
assign data_out = bypass_r ? curr_r : head_r;
assign empty = empty_r;
assign full = full_r;
assign size = size_r;
end
end end
end end

8
hw/syn/quartus/top/Makefile
View File

@@ -1,6 +1,6 @@
PROJECT = Vortex_Socket PROJECT = vortex_afu
TOP_LEVEL_ENTITY = Vortex_Socket TOP_LEVEL_ENTITY = vortex_afu
SRC_FILE = Vortex_Socket.v SRC_FILE = vortex_afu.sv
PROJECT_FILES = $(PROJECT).qpf $(PROJECT).qsf PROJECT_FILES = $(PROJECT).qpf $(PROJECT).qsf
# Part, Family # Part, Family
@@ -49,7 +49,7 @@ smart.log: $(PROJECT_FILES)
# Project initialization # Project initialization
$(PROJECT_FILES): $(PROJECT_FILES):
quartus_sh -t project.tcl -project $(PROJECT) -family $(FAMILY) -device $(DEVICE) -top $(TOP_LEVEL_ENTITY) -src $(SRC_FILE) -sdc project.sdc -inc "../../../rtl;../../../rtl/libs;../../../rtl/interfaces;../../../rtl/pipe_regs;../../../rtl/cache" quartus_sh -t project.tcl -project $(PROJECT) -family $(FAMILY) -device $(DEVICE) -top $(TOP_LEVEL_ENTITY) -src $(SRC_FILE) -sdc project.sdc -inc "../../../rtl;../../../rtl/libs;../../../rtl/interfaces;../../../rtl/pipe_regs;../../../rtl/cache;../../../opae"
syn.chg: syn.chg:
$(STAMP) syn.chg $(STAMP) syn.chg

0
hw/syn/quartus/top/vortex.sdc → hw/syn/quartus/top/project.sdc
View File

6
hw/syn/quartus/vortex/Makefile
View File

@@ -1,6 +1,6 @@
PROJECT = Vortex PROJECT = Vortex_Socket
TOP_LEVEL_ENTITY = Vortex TOP_LEVEL_ENTITY = Vortex_Socket
SRC_FILE = Vortex.v SRC_FILE = Vortex_Socket.v
PROJECT_FILES = $(PROJECT).qpf $(PROJECT).qsf PROJECT_FILES = $(PROJECT).qpf $(PROJECT).qsf
# Part, Family # Part, Family

0
hw/syn/quartus/vortex/vortex.sdc → hw/syn/quartus/vortex/project.sdc
View File

3
hw/syn/quartus/vortex/timing.tcl
View File

@@ -1,4 +1,4 @@
project_open Vortex project_open Vortex_Socket
set_global_assignment -name NUM_PARALLEL_PROCESSORS ALL set_global_assignment -name NUM_PARALLEL_PROCESSORS ALL
@@ -6,7 +6,6 @@ create_timing_netlist
read_sdc read_sdc
update_timing_netlist update_timing_netlist
foreach_in_collection op [get_available_operating_conditions] { foreach_in_collection op [get_available_operating_conditions] {
set_operating_conditions $op set_operating_conditions $op