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

opae build fixes

Browse Source
This commit is contained in:
Blaise Tine
2020-07-21 05:44:13 -07:00
parent dc7efbcfb4
commit ff7f65bd1f
17 changed files with 192 additions and 166 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
driver/rtlsim/Makefile
View File

@@ -16,12 +16,12 @@ DBG_PRINT_FLAGS += -DDBG_PRINT_OPAE
DBG_FLAGS += $(DBG_PRINT_FLAGS) DBG_FLAGS += $(DBG_PRINT_FLAGS)
DBG_FLAGS += -DDBG_CORE_REQ_INFO DBG_FLAGS += -DDBG_CORE_REQ_INFO
#CONFIGS += -DNUM_CLUSTERS=2 -DNUM_CORES=4 -DL2_ENABLE=1 CONFIGS += -DNUM_CLUSTERS=2 -DNUM_CORES=4 -DL2_ENABLE=1
#CONFIGS += -DNUM_CLUSTERS=1 -DNUM_CORES=4 -DL2_ENABLE=1 #CONFIGS += -DNUM_CLUSTERS=1 -DNUM_CORES=4 -DL2_ENABLE=1
#CONFIGS += -DNUM_CLUSTERS=1 -DNUM_CORES=2 -DL2_ENABLE=0 #CONFIGS += -DNUM_CLUSTERS=1 -DNUM_CORES=2 -DL2_ENABLE=0
CONFIGS += -DNUM_CLUSTERS=1 -DNUM_CORES=1 #CONFIGS += -DNUM_CLUSTERS=1 -DNUM_CORES=1
DEBUG=1 #DEBUG=1
#AFU=1 #AFU=1
CFLAGS += -fPIC CFLAGS += -fPIC

15
hw/opae/Makefile
View File

@@ -4,13 +4,16 @@ FPGA_BUILD_DIR=build_fpga
all: ase-1c all: ase-1c
ase-1c: setup-ase-1c sources.txt:
./gen_sources.sh
ase-1c: setup-ase-1c sources.txt
make -C $(ASE_BUILD_DIR)_1c make -C $(ASE_BUILD_DIR)_1c
ase-2c: setup-ase-2c ase-2c: setup-ase-2c sources.txt
make -C $(ASE_BUILD_DIR)_2c make -C $(ASE_BUILD_DIR)_2c
ase-4c: setup-ase-4c ase-4c: setup-ase-4c sources.txt
make -C $(ASE_BUILD_DIR)_4c make -C $(ASE_BUILD_DIR)_4c
setup-ase-1c: $(ASE_BUILD_DIR)_1c/Makefile setup-ase-1c: $(ASE_BUILD_DIR)_1c/Makefile
@@ -28,13 +31,13 @@ $(ASE_BUILD_DIR)_2c/Makefile:
$(ASE_BUILD_DIR)_4c/Makefile: $(ASE_BUILD_DIR)_4c/Makefile:
afu_sim_setup -s sources_4c.txt $(ASE_BUILD_DIR)_4c afu_sim_setup -s sources_4c.txt $(ASE_BUILD_DIR)_4c
fpga-1c: setup-fpga-1c fpga-1c: setup-fpga-1c sources.txt
cd $(FPGA_BUILD_DIR)_1c && qsub-synth cd $(FPGA_BUILD_DIR)_1c && qsub-synth
fpga-2c: setup-fpga-2c fpga-2c: setup-fpga-2c sources.txt
cd $(FPGA_BUILD_DIR)_2c && qsub-synth cd $(FPGA_BUILD_DIR)_2c && qsub-synth
fpga-4c: setup-fpga-4c fpga-4c: setup-fpga-4c sources.txt
cd $(FPGA_BUILD_DIR)_4c && qsub-synth cd $(FPGA_BUILD_DIR)_4c && qsub-synth
setup-fpga-1c: $(FPGA_BUILD_DIR)_1c/build/dcp.qpf setup-fpga-1c: $(FPGA_BUILD_DIR)_1c/build/dcp.qpf

4
hw/opae/README
View File

@@ -60,8 +60,8 @@ qsub-sim
make ase make ase
# tests # tests
./run_ase.sh build_ase_1c ../../driver/tests/basic/basic ./run_ase.sh build_ase_1c ../../driver/tests/basic/basic -n 256
./run_ase.sh build_ase_1c ../../driver/tests/demo/demo ./run_ase.sh build_ase_1c ../../driver/tests/demo/demo -n 16
./run_ase.sh build_ase_1c ../../benchmarks/opencl/vecadd/vecadd ./run_ase.sh build_ase_1c ../../benchmarks/opencl/vecadd/vecadd
# modify "vsim_run.tcl" to dump VCD trace # modify "vsim_run.tcl" to dump VCD trace

21
hw/opae/gen_sources.sh Executable file
View File

@@ -0,0 +1,21 @@
#!/bin/bash
dir_list='../rtl/libs ../rtl/cache ../rtl/interfaces ../rtl'
inc_list=""
for dir in $dir_list; do
inc_list="$inc_list -I$dir"
done
echo "inc_list=$inc_list"
{
# read design sources
for dir in $dir_list; do
echo "+incdir+$dir"
for file in $(find $dir -maxdepth 1 -name '*.v' -o -name '*.sv' -type f)
do
echo $file
done
done
} > sources.txt

137
hw/opae/sources.txt
View File

@@ -1,46 +1,34 @@
vortex_afu.json
QI:vortex_afu.qsf
#+define+SCOPE
#+define+DBG_PRINT_CORE_ICACHE
#+define+DBG_PRINT_CORE_DCACHE
#+define+DBG_PRINT_CACHE_BANK
#+define+DBG_PRINT_CACHE_SNP
#+define+DBG_PRINT_CACHE_MSRQ
#+define+DBG_PRINT_DRAM
#+define+DBG_PRINT_PIPELINE
#+define+DBG_PRINT_OPAE
#+define+DBG_PRINT_SCOPE
+incdir+.
+incdir+../rtl
+incdir+../rtl/interfaces
+incdir+../rtl/pipe_regs
+incdir+../rtl/cache
+incdir+../rtl/libs +incdir+../rtl/libs
../rtl/libs/VX_countones.v
../rtl/VX_user_config.vh ../rtl/libs/VX_divide.v
../rtl/VX_config.vh ../rtl/libs/VX_fair_arbiter.v
../rtl/VX_define.vh ../rtl/libs/VX_fixed_arbiter.v
../rtl/libs/VX_generic_queue.v
../rtl/cache/VX_cache_config.vh ../rtl/libs/VX_generic_register.v
../rtl/cache/VX_cache.v ../rtl/libs/VX_generic_stack.v
../rtl/cache/VX_cache_core_rsp_merge.v ../rtl/libs/VX_index_queue.v
../rtl/cache/VX_cache_core_req_bank_sel.v ../rtl/libs/VX_matrix_arbiter.v
../rtl/cache/VX_cache_dram_req_arb.v ../rtl/libs/VX_mult.v
../rtl/cache/VX_cache_dram_fill_arb.v ../rtl/libs/VX_priority_encoder.v
../rtl/cache/VX_cache_miss_resrv.v ../rtl/libs/VX_rr_arbiter.v
../rtl/libs/VX_onehot_encooder.v
+incdir+../rtl/cache
../rtl/cache/VX_bank.v ../rtl/cache/VX_bank.v
../rtl/cache/VX_bank_core_req_arb.v ../rtl/cache/VX_bank_core_req_arb.v
../rtl/cache/VX_cache.v
../rtl/cache/VX_cache_core_req_bank_sel.v
../rtl/cache/VX_cache_core_rsp_merge.v
../rtl/cache/VX_cache_dram_fill_arb.v
../rtl/cache/VX_cache_dram_req_arb.v
../rtl/cache/VX_cache_miss_resrv.v
../rtl/cache/VX_prefetcher.v
../rtl/cache/VX_snp_forwarder.v
../rtl/cache/VX_snp_rsp_arb.v ../rtl/cache/VX_snp_rsp_arb.v
../rtl/cache/VX_tag_data_access.v ../rtl/cache/VX_tag_data_access.v
../rtl/cache/VX_tag_data_structure.v ../rtl/cache/VX_tag_data_structure.v
../rtl/cache/VX_snp_forwarder.v +incdir+../rtl/interfaces
../rtl/cache/VX_prefetcher.v ../rtl/interfaces/VX_alu_req_if.v
../rtl/interfaces/VX_branch_ctl_if.v
../rtl/interfaces/VX_branch_rsp_if.v
../rtl/interfaces/VX_cache_core_req_if.v ../rtl/interfaces/VX_cache_core_req_if.v
../rtl/interfaces/VX_cache_core_rsp_if.v ../rtl/interfaces/VX_cache_core_rsp_if.v
../rtl/interfaces/VX_cache_dram_req_if.v ../rtl/interfaces/VX_cache_dram_req_if.v
@@ -48,65 +36,46 @@ QI:vortex_afu.qsf
../rtl/interfaces/VX_cache_snp_req_if.v ../rtl/interfaces/VX_cache_snp_req_if.v
../rtl/interfaces/VX_cache_snp_rsp_if.v ../rtl/interfaces/VX_cache_snp_rsp_if.v
../rtl/interfaces/VX_csr_req_if.v ../rtl/interfaces/VX_csr_req_if.v
../rtl/interfaces/VX_commit_if.v
../rtl/interfaces/VX_csr_io_req_if.v ../rtl/interfaces/VX_csr_io_req_if.v
../rtl/interfaces/VX_csr_io_rsp_if.v ../rtl/interfaces/VX_decode_if.v
../rtl/interfaces/VX_exec_unit_req_if.v ../rtl/interfaces/VX_gpr_data_if.v
../rtl/interfaces/VX_backend_req_if.v ../rtl/interfaces/VX_gpu_req_if.v
../rtl/interfaces/VX_gpr_read_if.v
../rtl/interfaces/VX_gpu_inst_req_if.v
../rtl/interfaces/VX_inst_meta_if.v
../rtl/interfaces/VX_jal_rsp_if.v
../rtl/interfaces/VX_join_if.v ../rtl/interfaces/VX_join_if.v
../rtl/interfaces/VX_lsu_req_if.v ../rtl/interfaces/VX_lsu_req_if.v
../rtl/interfaces/VX_warp_ctl_if.v ../rtl/interfaces/VX_warp_ctl_if.v
../rtl/interfaces/VX_wb_if.v ../rtl/interfaces/VX_wb_if.v
../rtl/interfaces/VX_wstall_if.v ../rtl/interfaces/VX_wstall_if.v
../rtl/interfaces/VX_csr_io_rsp_if.v
../rtl/libs/VX_generic_register.v ../rtl/interfaces/VX_ifetch_req_if.v
../rtl/libs/VX_mult.v ../rtl/interfaces/VX_ifetch_rsp_if.v
../rtl/libs/VX_divide.v ../rtl/interfaces/VX_mul_req_if.v
../rtl/libs/VX_generic_stack.v ../rtl/interfaces/VX_perf_cntrs_if.v
../rtl/libs/VX_priority_encoder.v +incdir+../rtl
../rtl/libs/VX_generic_queue.v ../rtl/VX_alu_unit.v
../rtl/libs/VX_indexable_queue.v ../rtl/VX_commit.v
../rtl/libs/VX_fair_arbiter.v
../rtl/libs/VX_fixed_arbiter.v
../rtl/libs/VX_rr_arbiter.v
../rtl/libs/VX_countones.v
../rtl/libs/VX_scope.v
../rtl/Vortex.v
../rtl/VX_cluster.v ../rtl/VX_cluster.v
../rtl/VX_core.v ../rtl/VX_core.v
../rtl/VX_mem_unit.v
../rtl/VX_pipeline.v
../rtl/VX_front_end.v
../rtl/VX_back_end.v
../rtl/VX_fetch.v
../rtl/VX_scheduler.v
../rtl/VX_exec_unit.v
../rtl/VX_warp.v
../rtl/VX_icache_stage.v
../rtl/VX_gpr_wrapper.v
../rtl/VX_gpu_inst.v
../rtl/VX_writeback.v
../rtl/VX_csr_pipe.v
../rtl/VX_csr_data.v ../rtl/VX_csr_data.v
../rtl/VX_csr_arb.v ../rtl/VX_csr_arb.v
../rtl/VX_dcache_arb.v
../rtl/VX_decode.v
../rtl/VX_csr_io_arb.v ../rtl/VX_csr_io_arb.v
../rtl/VX_warp_sched.v ../rtl/VX_fetch.v
../rtl/VX_csr_unit.v
../rtl/VX_gpr_ram.v ../rtl/VX_gpr_ram.v
../rtl/VX_gpr_stage.v ../rtl/VX_gpr_stage.v
../rtl/VX_alu_unit.v ../rtl/VX_execute.v
../rtl/VX_gpu_unit.v
../rtl/VX_icache_stage.v
../rtl/VX_issue.v
../rtl/VX_lsu_unit.v ../rtl/VX_lsu_unit.v
../rtl/VX_decode.v
../rtl/VX_inst_multiplex.v
../rtl/VX_dcache_arb.v
../rtl/VX_mem_arb.v ../rtl/VX_mem_arb.v
../rtl/VX_f_d_reg.v ../rtl/VX_mem_unit.v
../rtl/VX_i_d_reg.v ../rtl/VX_pipeline.v
../rtl/VX_d_e_reg.v ../rtl/VX_scheduler.v
../rtl/VX_issue_mux.v
ccip_interface_reg.sv ../rtl/VX_warp_sched.v
ccip_std_afu.sv ../rtl/VX_writeback.v
vortex_afu.sv ../rtl/Vortex.v
../rtl/VX_mul_unit.v

18
hw/opae/sources_1c.txt
View File

@@ -1,3 +1,21 @@
+define+NUM_CORES=1 +define+NUM_CORES=1
#+define+SCOPE
#+define+DBG_PRINT_CORE_ICACHE
#+define+DBG_PRINT_CORE_DCACHE
#+define+DBG_PRINT_CACHE_BANK
#+define+DBG_PRINT_CACHE_SNP
#+define+DBG_PRINT_CACHE_MSRQ
#+define+DBG_PRINT_DRAM
#+define+DBG_PRINT_PIPELINE
#+define+DBG_PRINT_OPAE
#+define+DBG_PRINT_SCOPE
vortex_afu.json
QI:vortex_afu.qsf
ccip_interface_reg.sv
ccip_std_afu.sv
vortex_afu.sv
C:sources.txt C:sources.txt

6
hw/opae/sources_2c.txt
View File

@@ -1,4 +1,10 @@
+define+NUM_CORES=2 +define+NUM_CORES=2
+define+L2_ENABLE=0 +define+L2_ENABLE=0
vortex_afu.json
QI:vortex_afu.qsf
ccip_interface_reg.sv
ccip_std_afu.sv
vortex_afu.sv
C:sources.txt C:sources.txt

6
hw/opae/sources_4c.txt
View File

@@ -1,4 +1,10 @@
+define+NUM_CORES=4 +define+NUM_CORES=4
+define+L2_ENABLE=0 +define+L2_ENABLE=0
vortex_afu.json
QI:vortex_afu.qsf
ccip_interface_reg.sv
ccip_std_afu.sv
vortex_afu.sv
C:sources.txt C:sources.txt

4
hw/rtl/VX_alu_unit.v
View File

@@ -13,7 +13,7 @@ module VX_alu_unit #(
VX_branch_ctl_if branch_ctl_if, VX_branch_ctl_if branch_ctl_if,
VX_commit_if alu_commit_if VX_commit_if alu_commit_if
); );
wire [`NUM_THREADS-1:0][31:0] alu_result; reg [`NUM_THREADS-1:0][31:0] alu_result;
wire [`NUM_THREADS-1:0][32:0] sub_result; wire [`NUM_THREADS-1:0][32:0] sub_result;
wire [`NUM_THREADS-1:0][32:0] shift_result; wire [`NUM_THREADS-1:0][32:0] shift_result;
@@ -99,7 +99,7 @@ module VX_alu_unit #(
); );
VX_generic_register #( VX_generic_register #(
.N(`NUM_THREADS + `NW_BITS + 32 + `NR_BITS + `WB_BITS + (`NUM_THREADS * 32)), .N(`NUM_THREADS + `NW_BITS + 32 + `NR_BITS + `WB_BITS + (`NUM_THREADS * 32))
) alu_reg ( ) alu_reg (
.clk (clk), .clk (clk),
.reset (reset), .reset (reset),

4
hw/rtl/VX_define.vh
View File

@@ -76,7 +76,7 @@
`define CSR_WIDTH 12 `define CSR_WIDTH 12
`define DIV_LATENCY 2 `define DIV_LATENCY 21
`define MUL_LATENCY 2 `define MUL_LATENCY 2
@@ -390,6 +390,8 @@
/////////////////////////////////////////////////////////////////////////////// ///////////////////////////////////////////////////////////////////////////////
task print_ex_type; task print_ex_type;
input [`EX_BITS-1:0] ex; input [`EX_BITS-1:0] ex;
begin begin

4
hw/rtl/VX_lsu_unit.v
View File

@@ -30,6 +30,7 @@ module VX_lsu_unit #(
wire [`NW_BITS-1:0] use_warp_num; wire [`NW_BITS-1:0] use_warp_num;
wire [`WB_BITS-1:0] use_wb; wire [`WB_BITS-1:0] use_wb;
wire [31:0] use_pc; wire [31:0] use_pc;
wire mrq_full;
genvar i; genvar i;
@@ -83,8 +84,7 @@ module VX_lsu_unit #(
wire [`LOG2UP(`DCREQ_SIZE)-1:0] mrq_write_addr, dbg_mrq_write_addr; wire [`LOG2UP(`DCREQ_SIZE)-1:0] mrq_write_addr, dbg_mrq_write_addr;
wire [`NUM_THREADS-1:0][1:0] mem_rsp_offset; wire [`NUM_THREADS-1:0][1:0] mem_rsp_offset;
wire [`BYTEEN_BITS-1:0] core_rsp_mem_read; wire [`BYTEEN_BITS-1:0] core_rsp_mem_read;
wire mrq_full;
wire mrq_push = (| dcache_req_if.valid) && dcache_req_if.ready wire mrq_push = (| dcache_req_if.valid) && dcache_req_if.ready
&& (0 == use_req_rw); // only push read requests && (0 == use_req_rw); // only push read requests

10
hw/rtl/VX_mul_unit.v
View File

@@ -12,7 +12,7 @@ module VX_mul_unit #(
// Outputs // Outputs
VX_commit_if mul_commit_if VX_commit_if mul_commit_if
); );
wire [`NUM_THREADS-1:0][31:0] alu_result; reg [`NUM_THREADS-1:0][31:0] alu_result;
wire [`NUM_THREADS-1:0][63:0] mul_result; wire [`NUM_THREADS-1:0][63:0] mul_result;
wire [`NUM_THREADS-1:0][31:0] div_result; wire [`NUM_THREADS-1:0][31:0] div_result;
wire [`NUM_THREADS-1:0][31:0] rem_result; wire [`NUM_THREADS-1:0][31:0] rem_result;
@@ -77,6 +77,8 @@ module VX_mul_unit #(
end end
end end
wire stall;
reg result_avail; reg result_avail;
reg [4:0] pending_ctr; reg [4:0] pending_ctr;
wire [4:0] instr_delay = `IS_DIV_OP(alu_op) ? `DIV_LATENCY : `MUL_LATENCY; wire [4:0] instr_delay = `IS_DIV_OP(alu_op) ? `DIV_LATENCY : `MUL_LATENCY;
@@ -104,13 +106,13 @@ module VX_mul_unit #(
wire pipeline_stall = ~result_avail && (| mul_req_if.valid); wire pipeline_stall = ~result_avail && (| mul_req_if.valid);
wire stall = (~mul_commit_if.ready && (| mul_commit_if.valid)) assign stall = (~mul_commit_if.ready && (| mul_commit_if.valid))
|| pipeline_stall; || pipeline_stall;
wire flush = mul_commit_if.ready && pipeline_stall; wire flush = mul_commit_if.ready && pipeline_stall;
VX_generic_register #( VX_generic_register #(
.N(`NUM_THREADS + `NW_BITS + 32 + `NR_BITS + `WB_BITS + (`NUM_THREADS * 32)), .N(`NUM_THREADS + `NW_BITS + 32 + `NR_BITS + `WB_BITS + (`NUM_THREADS * 32))
) mul_reg ( ) mul_reg (
.clk (clk), .clk (clk),
.reset (reset), .reset (reset),

1
hw/rtl/VX_warp_sched.v
View File

@@ -18,7 +18,6 @@ module VX_warp_sched #(
); );
wire update_use_wspawn; wire update_use_wspawn;
wire update_visible_active; wire update_visible_active;
wire scheduled_warp;
wire [(1+32+`NUM_THREADS-1):0] ipdom[`NUM_WARPS-1:0]; wire [(1+32+`NUM_THREADS-1):0] ipdom[`NUM_WARPS-1:0];

116
hw/rtl/Vortex.v
View File

@@ -139,54 +139,54 @@ module Vortex (
end else begin end else begin
wire per_cluster_dram_req_valid [`NUM_CLUSTERS-1:0]; wire [`NUM_CLUSTERS-1:0] per_cluster_dram_req_valid;
wire per_cluster_dram_req_rw [`NUM_CLUSTERS-1:0]; wire [`NUM_CLUSTERS-1:0] per_cluster_dram_req_rw;
wire [`L2DRAM_BYTEEN_WIDTH-1:0] per_cluster_dram_req_byteen [`NUM_CLUSTERS-1:0]; wire [`NUM_CLUSTERS-1:0][`L2DRAM_BYTEEN_WIDTH-1:0] per_cluster_dram_req_byteen;
wire [`L2DRAM_ADDR_WIDTH-1:0] per_cluster_dram_req_addr [`NUM_CLUSTERS-1:0]; wire [`NUM_CLUSTERS-1:0][`L2DRAM_ADDR_WIDTH-1:0] per_cluster_dram_req_addr;
wire [`L2DRAM_LINE_WIDTH-1:0] per_cluster_dram_req_data [`NUM_CLUSTERS-1:0]; wire [`NUM_CLUSTERS-1:0][`L2DRAM_LINE_WIDTH-1:0] per_cluster_dram_req_data;
wire [`L2DRAM_TAG_WIDTH-1:0] per_cluster_dram_req_tag [`NUM_CLUSTERS-1:0]; wire [`NUM_CLUSTERS-1:0][`L2DRAM_TAG_WIDTH-1:0] per_cluster_dram_req_tag;
wire l3_core_req_ready; wire l3_core_req_ready;
wire per_cluster_dram_rsp_valid [`NUM_CLUSTERS-1:0]; wire [`NUM_CLUSTERS-1:0] per_cluster_dram_rsp_valid;
wire [`L2DRAM_LINE_WIDTH-1:0] per_cluster_dram_rsp_data [`NUM_CLUSTERS-1:0]; wire [`NUM_CLUSTERS-1:0][`L2DRAM_LINE_WIDTH-1:0] per_cluster_dram_rsp_data;
wire [`L2DRAM_TAG_WIDTH-1:0] per_cluster_dram_rsp_tag [`NUM_CLUSTERS-1:0]; wire [`NUM_CLUSTERS-1:0][`L2DRAM_TAG_WIDTH-1:0] per_cluster_dram_rsp_tag;
wire per_cluster_dram_rsp_ready [`NUM_CLUSTERS-1:0]; wire [`NUM_CLUSTERS-1:0] per_cluster_dram_rsp_ready;
wire per_cluster_snp_req_valid [`NUM_CLUSTERS-1:0]; wire [`NUM_CLUSTERS-1:0] per_cluster_snp_req_valid;
wire [`L2DRAM_ADDR_WIDTH-1:0] per_cluster_snp_req_addr [`NUM_CLUSTERS-1:0]; wire [`NUM_CLUSTERS-1:0][`L2DRAM_ADDR_WIDTH-1:0] per_cluster_snp_req_addr;
wire per_cluster_snp_req_invalidate [`NUM_CLUSTERS-1:0]; wire [`NUM_CLUSTERS-1:0] per_cluster_snp_req_invalidate;
wire [`L2SNP_TAG_WIDTH-1:0] per_cluster_snp_req_tag [`NUM_CLUSTERS-1:0]; wire [`NUM_CLUSTERS-1:0][`L2SNP_TAG_WIDTH-1:0] per_cluster_snp_req_tag;
wire per_cluster_snp_req_ready [`NUM_CLUSTERS-1:0]; wire [`NUM_CLUSTERS-1:0] per_cluster_snp_req_ready;
wire per_cluster_snp_rsp_valid [`NUM_CLUSTERS-1:0]; wire [`NUM_CLUSTERS-1:0] per_cluster_snp_rsp_valid;
wire [`L2SNP_TAG_WIDTH-1:0] per_cluster_snp_rsp_tag [`NUM_CLUSTERS-1:0]; wire [`NUM_CLUSTERS-1:0][`L2SNP_TAG_WIDTH-1:0] per_cluster_snp_rsp_tag;
wire per_cluster_snp_rsp_ready [`NUM_CLUSTERS-1:0]; wire [`NUM_CLUSTERS-1:0] per_cluster_snp_rsp_ready;
wire per_cluster_io_req_valid [`NUM_CLUSTERS-1:0]; wire [`NUM_CLUSTERS-1:0] per_cluster_io_req_valid;
wire per_cluster_io_req_rw [`NUM_CLUSTERS-1:0]; wire [`NUM_CLUSTERS-1:0] per_cluster_io_req_rw;
wire [3:0] per_cluster_io_req_byteen [`NUM_CLUSTERS-1:0]; wire [`NUM_CLUSTERS-1:0][3:0] per_cluster_io_req_byteen;
wire [29:0] per_cluster_io_req_addr [`NUM_CLUSTERS-1:0]; wire [`NUM_CLUSTERS-1:0][29:0] per_cluster_io_req_addr;
wire [31:0] per_cluster_io_req_data [`NUM_CLUSTERS-1:0]; wire [`NUM_CLUSTERS-1:0][31:0] per_cluster_io_req_data;
wire [`L2CORE_TAG_WIDTH-1:0] per_cluster_io_req_tag [`NUM_CLUSTERS-1:0]; wire [`NUM_CLUSTERS-1:0][`L2CORE_TAG_WIDTH-1:0] per_cluster_io_req_tag;
wire per_cluster_io_req_ready [`NUM_CLUSTERS-1:0]; wire [`NUM_CLUSTERS-1:0] per_cluster_io_req_ready;
wire per_cluster_io_rsp_valid [`NUM_CLUSTERS-1:0]; wire [`NUM_CLUSTERS-1:0] per_cluster_io_rsp_valid;
wire [`L2CORE_TAG_WIDTH-1:0] per_cluster_io_rsp_tag [`NUM_CLUSTERS-1:0]; wire [`NUM_CLUSTERS-1:0][`L2CORE_TAG_WIDTH-1:0] per_cluster_io_rsp_tag;
wire [31:0] per_cluster_io_rsp_data [`NUM_CLUSTERS-1:0]; wire [`NUM_CLUSTERS-1:0][31:0] per_cluster_io_rsp_data;
wire per_cluster_io_rsp_ready [`NUM_CLUSTERS-1:0]; wire [`NUM_CLUSTERS-1:0] per_cluster_io_rsp_ready;
wire per_cluster_csr_io_req_valid [`NUM_CLUSTERS-1:0]; wire [`NUM_CLUSTERS-1:0] per_cluster_csr_io_req_valid;
wire [11:0] per_cluster_csr_io_req_addr [`NUM_CLUSTERS-1:0]; wire [`NUM_CLUSTERS-1:0][11:0] per_cluster_csr_io_req_addr;
wire per_cluster_csr_io_req_rw [`NUM_CLUSTERS-1:0]; wire [`NUM_CLUSTERS-1:0] per_cluster_csr_io_req_rw;
wire [31:0] per_cluster_csr_io_req_data [`NUM_CLUSTERS-1:0]; wire [`NUM_CLUSTERS-1:0][31:0] per_cluster_csr_io_req_data;
wire per_cluster_csr_io_req_ready [`NUM_CLUSTERS-1:0]; wire [`NUM_CLUSTERS-1:0] per_cluster_csr_io_req_ready;
wire per_cluster_csr_io_rsp_valid [`NUM_CLUSTERS-1:0]; wire [`NUM_CLUSTERS-1:0] per_cluster_csr_io_rsp_valid;
wire [31:0] per_cluster_csr_io_rsp_data [`NUM_CLUSTERS-1:0]; wire [`NUM_CLUSTERS-1:0][31:0] per_cluster_csr_io_rsp_data;
wire per_cluster_csr_io_rsp_ready [`NUM_CLUSTERS-1:0]; wire [`NUM_CLUSTERS-1:0] per_cluster_csr_io_rsp_ready;
wire per_cluster_busy [`NUM_CLUSTERS-1:0]; wire [`NUM_CLUSTERS-1:0] per_cluster_busy;
wire per_cluster_ebreak [`NUM_CLUSTERS-1:0]; wire [`NUM_CLUSTERS-1:0] per_cluster_ebreak;
wire [`CLOG2(`NUM_CLUSTERS)-1:0] csr_io_request_id = `CLOG2(`NUM_CLUSTERS)'(csr_io_req_coreid >> `CLOG2(`NUM_CLUSTERS)); wire [`CLOG2(`NUM_CLUSTERS)-1:0] csr_io_request_id = `CLOG2(`NUM_CLUSTERS)'(csr_io_req_coreid >> `CLOG2(`NUM_CLUSTERS));
wire [`NC_BITS-1:0] per_cluster_csr_io_req_coreid = `NC_BITS'(csr_io_req_coreid); wire [`NC_BITS-1:0] per_cluster_csr_io_req_coreid = `NC_BITS'(csr_io_req_coreid);
@@ -336,27 +336,27 @@ module Vortex (
// L3 Cache /////////////////////////////////////////////////////////// // L3 Cache ///////////////////////////////////////////////////////////
wire l3_core_req_valid [`L3NUM_REQUESTS-1:0]; wire [`L3NUM_REQUESTS-1:0] l3_core_req_valid;
wire l3_core_req_rw [`L3NUM_REQUESTS-1:0]; wire [`L3NUM_REQUESTS-1:0] l3_core_req_rw;
wire [`L2DRAM_BYTEEN_WIDTH-1:0] l3_core_req_byteen [`L3NUM_REQUESTS-1:0]; wire [`L3NUM_REQUESTS-1:0][`L2DRAM_BYTEEN_WIDTH-1:0] l3_core_req_byteen;
wire [`L2DRAM_ADDR_WIDTH-1:0] l3_core_req_addr [`L3NUM_REQUESTS-1:0]; wire [`L3NUM_REQUESTS-1:0][`L2DRAM_ADDR_WIDTH-1:0] l3_core_req_addr;
wire [`L2DRAM_LINE_WIDTH-1:0] l3_core_req_data [`L3NUM_REQUESTS-1:0]; wire [`L3NUM_REQUESTS-1:0][`L2DRAM_LINE_WIDTH-1:0] l3_core_req_data;
wire [`L2DRAM_TAG_WIDTH-1:0] l3_core_req_tag [`L3NUM_REQUESTS-1:0]; wire [`L3NUM_REQUESTS-1:0][`L2DRAM_TAG_WIDTH-1:0] l3_core_req_tag;
wire l3_core_rsp_valid [`L3NUM_REQUESTS-1:0]; wire [`L3NUM_REQUESTS-1:0] l3_core_rsp_valid;
wire [`L2DRAM_LINE_WIDTH-1:0] l3_core_rsp_data [`L3NUM_REQUESTS-1:0]; wire [`L3NUM_REQUESTS-1:0][`L2DRAM_LINE_WIDTH-1:0] l3_core_rsp_data;
wire [`L2DRAM_TAG_WIDTH-1:0] l3_core_rsp_tag [`L3NUM_REQUESTS-1:0]; wire [`L3NUM_REQUESTS-1:0][`L2DRAM_TAG_WIDTH-1:0] l3_core_rsp_tag;
wire l3_core_rsp_ready; wire l3_core_rsp_ready;
wire l3_snp_fwdout_valid [`NUM_CLUSTERS-1:0]; wire [`NUM_CLUSTERS-1:0] l3_snp_fwdout_valid;
wire [`L2DRAM_ADDR_WIDTH-1:0] l3_snp_fwdout_addr [`NUM_CLUSTERS-1:0]; wire [`NUM_CLUSTERS-1:0][`L2DRAM_ADDR_WIDTH-1:0] l3_snp_fwdout_addr;
wire l3_snp_fwdout_invalidate [`NUM_CLUSTERS-1:0]; wire [`NUM_CLUSTERS-1:0] l3_snp_fwdout_invalidate;
wire [`L2SNP_TAG_WIDTH-1:0] l3_snp_fwdout_tag [`NUM_CLUSTERS-1:0]; wire [`NUM_CLUSTERS-1:0][`L2SNP_TAG_WIDTH-1:0] l3_snp_fwdout_tag;
wire l3_snp_fwdout_ready [`NUM_CLUSTERS-1:0]; wire [`NUM_CLUSTERS-1:0] l3_snp_fwdout_ready;
wire l3_snp_fwdin_valid [`NUM_CLUSTERS-1:0]; wire [`NUM_CLUSTERS-1:0] l3_snp_fwdin_valid;
wire [`L2SNP_TAG_WIDTH-1:0] l3_snp_fwdin_tag [`NUM_CLUSTERS-1:0]; wire [`NUM_CLUSTERS-1:0][`L2SNP_TAG_WIDTH-1:0] l3_snp_fwdin_tag;
wire l3_snp_fwdin_ready [`NUM_CLUSTERS-1:0]; wire [`NUM_CLUSTERS-1:0] l3_snp_fwdin_ready;
for (i = 0; i < `L3NUM_REQUESTS; i++) begin for (i = 0; i < `L3NUM_REQUESTS; i++) begin
// Core Request // Core Request

2
hw/rtl/tex_unit/VX_tex_mgr.v
View File

@@ -2,7 +2,7 @@
module VX_tex_mgr ( module VX_tex_mgr (
input wire clk, input wire clk,
input wire reset, input wire reset
); );
//-- //--

2
hw/rtl/tex_unit/VX_tex_unit.v
View File

@@ -11,7 +11,7 @@ module VX_tex_unit #(
parameter MAXAMW = 2, parameter MAXAMW = 2,
parameter TAGW = 16, parameter TAGW = 16,
parameter NUMCRQS = 32, parameter NUMCRQS = 32
) ( ) (
input wire clk, input wire clk,
input wire reset, input wire reset,

2
hw/syn/yosys/synth.sh
View File

@@ -12,7 +12,7 @@ echo "inc_list=$inc_list"
{ {
# read design sources # read design sources
for dir in $dir_list; do for dir in $dir_list; do
for file in $(find $dir -name '*.v' -o -name '*.sv' -type f) for file in $(find $dir -maxdepth 1 -name '*.v' -o -name '*.sv' -type f)
do do
echo "read_verilog -sv $inc_list $file" echo "read_verilog -sv $inc_list $file"
done done