enabling 128-bit dram bus

This commit is contained in:
Blaise Tine
2021-04-24 00:31:27 -04:00
parent 2f5ccdcf45
commit 4cb98a25a7
19 changed files with 344 additions and 198 deletions

View File

@@ -21,10 +21,10 @@ DBG_PRINT_FLAGS += -DDBG_PRINT_SCOPE
DBG_FLAGS += $(DBG_PRINT_FLAGS)
DBG_FLAGS += -DDBG_CACHE_REQ_INFO
#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=2 -DL2_ENABLE=0
CONFIGS ?= -DNUM_CLUSTERS=1 -DNUM_CORES=1
#CONFIGS := -DNUM_CLUSTERS=2 -DNUM_CORES=4 -DL2_ENABLE=1 $(CONFIGS)
#CONFIGS := -DNUM_CLUSTERS=1 -DNUM_CORES=4 -DL2_ENABLE=1 $(CONFIGS)
#CONFIGS := -DNUM_CLUSTERS=1 -DNUM_CORES=2 -DL2_ENABLE=0 $(CONFIGS)
CONFIGS := -DNUM_CLUSTERS=1 -DNUM_CORES=1 $(CONFIGS)
CFLAGS += -fPIC
@@ -47,10 +47,11 @@ FPU_INCLUDE = -I$(RTL_DIR)/fp_cores -I$(DPI_DIR) -I$(RTL_DIR)/fp_cores/fpnew/src
RTL_INCLUDE = -I$(RTL_DIR) -I$(RTL_DIR)/libs -I$(RTL_DIR)/interfaces -I$(RTL_DIR)/cache $(FPU_INCLUDE)
RTL_INCLUDE += -I$(RTL_DIR)/afu -I$(RTL_DIR)/afu/ccip
VL_FLAGS += -O2 --language 1800-2009 --assert -Wall -Wpedantic $(CONFIGS)
VL_FLAGS += -Wno-DECLFILENAME
VL_FLAGS += -O2 --language 1800-2009 --assert -Wall -Wpedantic
VL_FLAGS += -Wno-DECLFILENAME -Wno-REDEFMACRO
VL_FLAGS += --x-initial unique --x-assign unique
VL_FLAGS += verilator.vlt
VL_FLAGS += $(CONFIGS)
# Enable Verilator multithreaded simulation
#THREADS ?= $(shell python3 -c 'import multiprocessing as mp; print(max(1, mp.cpu_count() // 2))')
@@ -87,10 +88,13 @@ VL_FLAGS += -DFPU_DPI
PROJECT = libopae-c-vlsim.so
all: $(PROJECT)
vortex_afu.h : $(RTL_DIR)/afu/vortex_afu.vh
../../../hw/scripts/gen_config.py -i $(RTL_DIR)/afu/vortex_afu.vh -o vortex_afu.h
$(PROJECT): $(SRCS)
$(PROJECT): $(SRCS) vortex_afu.h
verilator --exe --cc $(TOP) --top-module $(TOP) $(RTL_INCLUDE) $(VL_FLAGS) $(SRCS) -CFLAGS '$(CFLAGS)' -LDFLAGS '$(LDFLAGS)' -o ../$(PROJECT)
make -j -C obj_dir -f V$(TOP).mk
clean:
rm -rf $(PROJECT) obj_dir ../scope-defs.h $(RTL_DIR)/scope-defs.vh
rm -rf $(PROJECT) obj_dir ../scope-defs.h $(RTL_DIR)/scope-defs.vh vortex_afu.h

View File

@@ -285,15 +285,15 @@ void opae_sim::avs_bus() {
vortex_afu_->avs_readdatavalid = 0;
if (dram_rd_it != dram_reads_.end()) {
vortex_afu_->avs_readdatavalid = 1;
memcpy(vortex_afu_->avs_readdata, dram_rd_it->data.data(), CACHE_BLOCK_SIZE);
memcpy(vortex_afu_->avs_readdata, dram_rd_it->data.data(), DRAM_BLOCK_SIZE);
uint32_t addr = dram_rd_it->addr;
dram_reads_.erase(dram_rd_it);
/*printf("%0ld: [sim] DRAM Rd Rsp: addr=%x, pending={", timestamp, addr * CACHE_BLOCK_SIZE);
/*printf("%0ld: [sim] DRAM Rd Rsp: addr=%x, pending={", timestamp, addr * DRAM_BLOCK_SIZE);
for (auto& req : dram_reads_) {
if (req.cycles_left != 0)
printf(" !%0x", req.addr * CACHE_BLOCK_SIZE);
printf(" !%0x", req.addr * DRAM_BLOCK_SIZE);
else
printf(" %0x", req.addr * CACHE_BLOCK_SIZE);
printf(" %0x", req.addr * DRAM_BLOCK_SIZE);
}
printf("}\n");*/
}
@@ -315,19 +315,24 @@ void opae_sim::avs_bus() {
if (vortex_afu_->avs_write) {
assert(0 == vortex_afu_->mem_bank_select);
uint64_t byteen = vortex_afu_->avs_byteenable;
unsigned base_addr = (vortex_afu_->avs_address * CACHE_BLOCK_SIZE);
unsigned base_addr = vortex_afu_->avs_address * DRAM_BLOCK_SIZE;
uint8_t* data = (uint8_t*)(vortex_afu_->avs_writedata);
for (int i = 0; i < CACHE_BLOCK_SIZE; i++) {
for (int i = 0; i < DRAM_BLOCK_SIZE; i++) {
if ((byteen >> i) & 0x1) {
ram_[base_addr + i] = data[i];
}
}
/*printf("%0ld: [sim] DRAM Wr Req: addr=%x, data=", timestamp, base_addr);
for (int i = 0; i < DRAM_BLOCK_SIZE; i++) {
printf("%0x", data[(DRAM_BLOCK_SIZE-1)-i]);
}
printf("\n");*/
}
if (vortex_afu_->avs_read) {
assert(0 == vortex_afu_->mem_bank_select);
dram_rd_req_t dram_req;
dram_req.addr = vortex_afu_->avs_address;
ram_.read(vortex_afu_->avs_address * CACHE_BLOCK_SIZE, CACHE_BLOCK_SIZE, dram_req.data.data());
ram_.read(vortex_afu_->avs_address * DRAM_BLOCK_SIZE, DRAM_BLOCK_SIZE, dram_req.data.data());
dram_req.cycles_left = DRAM_LATENCY;
for (auto& rsp : dram_reads_) {
if (dram_req.addr == rsp.addr) {
@@ -336,15 +341,15 @@ void opae_sim::avs_bus() {
}
}
dram_reads_.emplace_back(dram_req);
/*printf("%0ld: [sim] DRAM Rd Req: addr=%x, pending={", timestamp, dram_req.addr * CACHE_BLOCK_SIZE);
/*printf("%0ld: [sim] DRAM Rd Req: addr=%x, pending={", timestamp, dram_req.addr * DRAM_BLOCK_SIZE);
for (auto& req : dram_reads_) {
if (req.cycles_left != 0)
printf(" !%0x", req.addr * CACHE_BLOCK_SIZE);
printf(" !%0x", req.addr * DRAM_BLOCK_SIZE);
else
printf(" %0x", req.addr * CACHE_BLOCK_SIZE);
printf(" %0x", req.addr * DRAM_BLOCK_SIZE);
}
printf("}\n");*/
}
}
}
vortex_afu_->avs_waitrequest = dram_stalled;

View File

@@ -9,6 +9,7 @@
#endif
#include <VX_config.h>
#include "vortex_afu.h"
#include "ram.h"
#include <ostream>
@@ -16,6 +17,8 @@
#include <list>
#include <unordered_map>
#define DRAM_BLOCK_SIZE (PLATFORM_PARAM_LOCAL_MEMORY_DATA_WIDTH / 8)
#define CACHE_BLOCK_SIZE 64
class opae_sim {
@@ -40,7 +43,7 @@ private:
typedef struct {
int cycles_left;
std::array<uint8_t, CACHE_BLOCK_SIZE> data;
std::array<uint8_t, DRAM_BLOCK_SIZE> data;
uint32_t addr;
} dram_rd_req_t;

View File

@@ -1,5 +1,9 @@
`include "VX_define.vh"
`include "VX_platform.vh"
`IGNORE_WARNINGS_BEGIN
`include "vortex_afu.vh"
`IGNORE_WARNINGS_END
`include "VX_define.vh"
/* verilator lint_off IMPORTSTAR */
import ccip_if_pkg::*;
import local_mem_cfg_pkg::*;

View File

@@ -20,10 +20,10 @@ DBG_PRINT_FLAGS += -DDBG_PRINT_SCOPE
DBG_FLAGS += $(DBG_PRINT_FLAGS)
DBG_FLAGS += -DDBG_CACHE_REQ_INFO
#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=2 -DL2_ENABLE=0
CONFIGS ?= -DNUM_CLUSTERS=1 -DNUM_CORES=1
#CONFIGS := -DNUM_CLUSTERS=2 -DNUM_CORES=4 -DL2_ENABLE=1 $(CONFIGS)
#CONFIGS := -DNUM_CLUSTERS=1 -DNUM_CORES=4 -DL2_ENABLE=1 $(CONFIGS)
#CONFIGS := -DNUM_CLUSTERS=1 -DNUM_CORES=2 -DL2_ENABLE=0 $(CONFIGS)
CONFIGS := -DNUM_CLUSTERS=1 -DNUM_CORES=1 $(CONFIGS)
CFLAGS += $(CONFIGS)
@@ -44,7 +44,7 @@ FPU_INCLUDE = -I$(RTL_DIR)/fp_cores -I$(DPI_DIR) -I$(RTL_DIR)/fp_cores/fpnew/src
RTL_INCLUDE = -I$(RTL_DIR) -I$(RTL_DIR)/libs -I$(RTL_DIR)/interfaces -I$(RTL_DIR)/cache $(FPU_INCLUDE)
VL_FLAGS += -O2 --language 1800-2009 --assert -Wall -Wpedantic $(CONFIGS)
VL_FLAGS += -Wno-DECLFILENAME
VL_FLAGS += -Wno-DECLFILENAME -Wno-REDEFMACRO
VL_FLAGS += --x-initial unique --x-assign unique
VL_FLAGS += verilator.vlt

View File

@@ -10,10 +10,10 @@ CXXFLAGS += -fPIC -Wno-aligned-new -Wno-maybe-uninitialized
CXXFLAGS += -I../include -I../../hw -I$(SIMX_DIR)
CXXFLAGS += -DDUMP_PERF_STATS
#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=2 -DL2_ENABLE=0
CONFIGS ?= -DNUM_CLUSTERS=1 -DNUM_CORES=1
#CONFIGS := -DNUM_CLUSTERS=2 -DNUM_CORES=4 -DL2_ENABLE=1 $(CONFIGS)
#CONFIGS := -DNUM_CLUSTERS=1 -DNUM_CORES=4 -DL2_ENABLE=1 $(CONFIGS)
#CONFIGS := -DNUM_CLUSTERS=1 -DNUM_CORES=2 -DL2_ENABLE=0 $(CONFIGS)
CONFIGS := -DNUM_CLUSTERS=1 -DNUM_CORES=1 $(CONFIGS)
CXXFLAGS += $(CONFIGS)