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

multicore fix

Browse Source
This commit is contained in:
Blaise Tine
2020-05-10 08:30:04 -04:00
parent 359601cfd3
commit cc84e0691c
40 changed files with 27930 additions and 28148 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
driver/rtlsim/Makefile
View File

@@ -7,7 +7,7 @@ CFLAGS += -I../../include -I../../../hw/simulate -I../../../runtime
#MULTICORE += -DNUM_CLUSTERS=1 -DNUM_CORES=2
MULTICORE += -DNUM_CLUSTERS=1 -DNUM_CORES=2 -DL2_ENABLE=0
DEBUG = 1
#DEBUG = 1
CFLAGS += -fPIC
@@ -21,6 +21,10 @@ RTL_INCLUDE = -I../../hw/rtl -I../../hw/rtl/libs -I../../hw/rtl/interfaces -I../
VL_FLAGS += --language 1800-2009 --assert -Wall -Wpedantic $(MULTICORE)
# Use 64 bytes DRAM blocks
CFLAGS += -DGLOBAL_BLOCK_SIZE=64
VL_FLAGS += -DGLOBAL_BLOCK_SIZE=64
# Enable Verilator multithreaded simulation
#THREADS ?= $(shell python3 -c 'import multiprocessing as mp; print(max(1, mp.cpu_count() // 2))')
#VL_FLAGS += --threads $(THREADS)

21
driver/rtlsim/vortex.cpp
View File

@@ -60,9 +60,7 @@ private:
class vx_device {
public:
vx_device()
: is_done_(false)
, simulator_(&ram_) {
simulator_.reset();
: is_done_(false) {
thread_ = new std::thread(__thread_proc__, this);
mem_allocation_ = vx_dev_caps(VX_CAPS_ALLOC_BASE_ADDR);
}
@@ -95,7 +93,7 @@ public:
/*printf("VXDRV: upload %d bytes to 0x%x\n", size, dest_addr);
for (int i = 0; i < size; i += 4) {
printf("mem-write: 0x%x <- 0x%x\n", dest_addr + i, *(uint32_t*)((uint8_t*)src + src_offset + i));
printf("mem-write: 0x%x <- 0x%x\n", uint32_t(dest_addr + i), *(uint32_t*)((uint8_t*)src + src_offset + i));
}*/
ram_.write(dest_addr, asize, (uint8_t*)src + src_offset);
@@ -111,7 +109,7 @@ public:
/*printf("VXDRV: download %d bytes from 0x%x\n", size, src_addr);
for (int i = 0; i < size; i += 4) {
printf("mem-read: 0x%x -> 0x%x\n", src_addr + i, *(uint32_t*)((uint8_t*)dest + dest_offset + i));
printf("mem-read: 0x%x -> 0x%x\n", uint32_t(src_addr + i), *(uint32_t*)((uint8_t*)dest + dest_offset + i));
}*/
return 0;
@@ -120,7 +118,9 @@ public:
int flush_caches(size_t dev_maddr, size_t size) {
mutex_.lock();
simulator_.attach_ram(&ram_);
simulator_.flush_caches(dev_maddr, size);
simulator_.attach_ram(nullptr);
mutex_.unlock();
return 0;
@@ -130,7 +130,8 @@ public:
mutex_.lock();
simulator_.reset();
mutex_.unlock();
simulator_.attach_ram(&ram_);
mutex_.unlock();
return 0;
}
@@ -142,8 +143,14 @@ public:
bool is_busy = simulator_.is_busy();
mutex_.unlock();
if (!is_busy || 0 == timeout_sec--)
if (!is_busy || 0 == timeout_sec--) {
if (!is_busy) {
mutex_.lock();
simulator_.attach_ram(nullptr);
mutex_.unlock();
}
break;
}
std::this_thread::sleep_for(std::chrono::seconds(1));
}

31
driver/tests/basic/basic.cpp
View File

@@ -1,6 +1,7 @@
#include <iostream>
#include <unistd.h>
#include <vortex.h>
#include "common.h"
int test = -1;
@@ -61,7 +62,7 @@ int run_memcopy_test(vx_buffer_h sbuf,
int errors = 0;
// write sbuf data
for (int i = 0; i < 8 * num_blocks; ++i) {
for (int i = 0; i < (64 * num_blocks) / 8; ++i) {
((uint64_t*)vx_host_ptr(sbuf))[i] = shuffle(i, value);
}
@@ -75,12 +76,12 @@ int run_memcopy_test(vx_buffer_h sbuf,
// verify result
std::cout << "verify result" << std::endl;
for (int i = 0; i < 8 * num_blocks; ++i) {
for (int i = 0; i < (64 * num_blocks) / 8; ++i) {
auto curr = ((uint64_t*)vx_host_ptr(dbuf))[i];
auto ref = shuffle(i, value);
if (curr != ref) {
std::cout << "error @ " << std::hex << (address + 64 * i)
<< ": actual " << curr << ", expected " << ref << std::endl;
std::cout << "error at 0x" << std::hex << (address + 8 * i)
<< ": actual 0x" << curr << ", expected 0x" << ref << std::endl;
++errors;
}
}
@@ -101,13 +102,13 @@ int run_kernel_test(vx_device_h device,
int errors = 0;
uint64_t seed = 0x0badf00d40ff40ff;
int num_blocks = 4;
unsigned src_dev_addr = 0x10000000;
unsigned dest_dev_addr = 0x20000000;
int src_dev_addr = DEV_MEM_SRC_ADDR;
int dest_dev_addr = DEV_MEM_DST_ADDR;
int num_blocks = NUM_BLOCKS;
// write sbuf data
for (int i = 0; i < 8 * num_blocks; ++i) {
for (int i = 0; i < (64 * num_blocks) / 8; ++i) {
((uint64_t*)vx_host_ptr(sbuf))[i] = shuffle(i, seed);
}
@@ -137,12 +138,12 @@ int run_kernel_test(vx_device_h device,
// verify result
std::cout << "verify result" << std::endl;
for (int i = 0; i < 8 * num_blocks; ++i) {
for (int i = 0; i < (64 * num_blocks) / 8; ++i) {
auto curr = ((uint64_t*)vx_host_ptr(dbuf))[i];
auto ref = shuffle(i, seed);
if (curr != ref) {
std::cout << "error @ " << std::hex << (dest_dev_addr + 64 * i)
<< ": actual " << curr << ", expected " << ref << std::endl;
std::cout << "error at 0x" << std::hex << (dest_dev_addr + 8 * i)
<< ": actual 0x" << curr << ", expected 0x" << ref << std::endl;
++errors;
}
}
@@ -174,10 +175,10 @@ int main(int argc, char *argv[]) {
RT_CHECK(vx_alloc_shared_mem(device, 4096, &dbuf));
// run tests
/*9if (0 == test || -1 == test) {
/*if (0 == test || -1 == test) {
std::cout << "run memcopy test" << std::endl;
RT_CHECK(run_memcopy_test(sbuf, dbuf, 0x10000000, 0x0badf00d00ff00ff, 1));
RT_CHECK(run_memcopy_test(sbuf, dbuf, 0x20000000, 0x0badf00d40ff40ff, 8));
RT_CHECK(run_memcopy_test(sbuf, dbuf, DEV_MEM_SRC_ADDR, 0x0badf00d00ff00ff, 1));
RT_CHECK(run_memcopy_test(sbuf, dbuf, DEV_MEM_DST_ADDR, 0x0badf00d40ff40ff, 8));
}*/
if (1 == test || -1 == test) {

8
driver/tests/basic/common.h Normal file
View File

@@ -0,0 +1,8 @@
#ifndef _COMMON_H_
#define _COMMON_H_
#define DEV_MEM_SRC_ADDR 0x10000000
#define DEV_MEM_DST_ADDR 0x20000000
#define NUM_BLOCKS 16
#endif

BIN
driver/tests/basic/kernel.bin
View File

Binary file not shown.

19
driver/tests/basic/kernel.c
View File

@@ -1,9 +1,20 @@
#include <stdint.h>
#include "config.h"
#include "intrinsics/vx_intrinsics.h"
#include "common.h"
void main() {
int64_t* x = (int64_t*)0x10000000;
int64_t* y = (int64_t*)0x20000000;
for (int i = 0; i < 8 * 4; ++i) {
y[i] = x[i];
int64_t* x = (int64_t*)DEV_MEM_SRC_ADDR;
int64_t* y = (int64_t*)DEV_MEM_DST_ADDR;
int num_words = (NUM_BLOCKS * 64) / 8;
int core_id = vx_core_id();
int num_cores = vx_num_cores();
int num_words_per_core = num_words / num_cores;
int offset = core_id * num_words_per_core;
for (int i = 0; i < num_words_per_core; ++i) {
y[offset + i] = x[offset + i];
}
}

2
driver/tests/demo/Makefile
View File

@@ -49,7 +49,7 @@ run-ase: $(PROJECT)
ASE_LOG=0 LD_LIBRARY_PATH=../../opae/ase:$(LD_LIBRARY_PATH) ./$(PROJECT) -f kernel.bin -n 16
run-rtlsim: $(PROJECT)
LD_LIBRARY_PATH=../../rtlsim:$(LD_LIBRARY_PATH) ./$(PROJECT) -f kernel.bin -n 16
LD_LIBRARY_PATH=../../rtlsim:$(LD_LIBRARY_PATH) ./$(PROJECT) -f kernel.bin -n 1
run-simx: $(PROJECT)
LD_LIBRARY_PATH=../../simx:$(LD_LIBRARY_PATH) ./$(PROJECT) -f kernel.bin -n 16

BIN
driver/tests/demo/kernel.bin
View File

Binary file not shown.

BIN
driver/tests/demo/kernel.elf
View File

Binary file not shown.