driver basic test and demo test refactoring
This commit is contained in:
@@ -44,16 +44,16 @@ $(PROJECT): $(SRCS)
|
||||
$(CXX) $(CXXFLAGS) $^ $(LDFLAGS) -L../../stub -lvortex -o $@
|
||||
|
||||
run-fpga: $(PROJECT)
|
||||
LD_LIBRARY_PATH=../../opae:$(LD_LIBRARY_PATH) ./$(PROJECT)
|
||||
LD_LIBRARY_PATH=../../opae:$(LD_LIBRARY_PATH) ./$(PROJECT) -n 32
|
||||
|
||||
run-ase: $(PROJECT)
|
||||
ASE_LOG=0 LD_LIBRARY_PATH=../../opae/ase:$(LD_LIBRARY_PATH) ./$(PROJECT)
|
||||
ASE_LOG=0 LD_LIBRARY_PATH=../../opae/ase:$(LD_LIBRARY_PATH) ./$(PROJECT) -n 16
|
||||
|
||||
run-rtlsim: $(PROJECT)
|
||||
LD_LIBRARY_PATH=../../rtlsim:$(LD_LIBRARY_PATH) ./$(PROJECT)
|
||||
LD_LIBRARY_PATH=../../rtlsim:$(LD_LIBRARY_PATH) ./$(PROJECT) -n 16
|
||||
|
||||
run-simx: $(PROJECT)
|
||||
LD_LIBRARY_PATH=../../simx:$(LD_LIBRARY_PATH) ./$(PROJECT)
|
||||
LD_LIBRARY_PATH=../../simx:$(LD_LIBRARY_PATH) ./$(PROJECT) -n 16
|
||||
|
||||
.depend: $(SRCS)
|
||||
$(CXX) $(CXXFLAGS) -MM $^ > .depend;
|
||||
|
||||
@@ -1,10 +1,9 @@
|
||||
#include <iostream>
|
||||
#include <unistd.h>
|
||||
#include <string.h>
|
||||
#include <vortex.h>
|
||||
#include "common.h"
|
||||
|
||||
int test = -1;
|
||||
|
||||
#define RT_CHECK(_expr) \
|
||||
do { \
|
||||
int _ret = _expr; \
|
||||
@@ -15,79 +14,84 @@ int test = -1;
|
||||
exit(-1); \
|
||||
} while (false)
|
||||
|
||||
const char* kernel_file = "kernel.bin";
|
||||
int test = -1;
|
||||
uint32_t count = 0;
|
||||
|
||||
vx_device_h device = nullptr;
|
||||
vx_buffer_h buffer = nullptr;
|
||||
|
||||
static void show_usage() {
|
||||
std::cout << "Vortex Driver Test." << std::endl;
|
||||
std::cout << "Usage: [-t testno][-k: kernel][-n words][-h: help]" << std::endl;
|
||||
}
|
||||
|
||||
static void parse_args(int argc, char **argv) {
|
||||
int c;
|
||||
while ((c = getopt(argc, argv, "t:h?")) != -1) {
|
||||
while ((c = getopt(argc, argv, "n:t:k:h?")) != -1) {
|
||||
switch (c) {
|
||||
case 't': {
|
||||
case 'n':
|
||||
count = atoi(optarg);
|
||||
break;
|
||||
case 't':
|
||||
test = atoi(optarg);
|
||||
} break;
|
||||
break;
|
||||
case 'k':
|
||||
kernel_file = optarg;
|
||||
break;
|
||||
case 'h':
|
||||
case '?': {
|
||||
std::cout << "Test." << std::endl;
|
||||
std::cout << "Usage: [-t testno][-h: help]" << std::endl;
|
||||
show_usage();
|
||||
exit(0);
|
||||
} break;
|
||||
default:
|
||||
show_usage();
|
||||
exit(-1);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
void cleanup() {
|
||||
if (buffer) {
|
||||
vx_buf_release(buffer);
|
||||
}
|
||||
if (device) {
|
||||
vx_dev_close(device);
|
||||
}
|
||||
}
|
||||
|
||||
uint64_t shuffle(int i, uint64_t value) {
|
||||
return (value << i) | (value & ((1 << i)-1));;
|
||||
}
|
||||
|
||||
vx_device_h device = nullptr;
|
||||
vx_buffer_h sbuf = nullptr;
|
||||
vx_buffer_h dbuf = nullptr;
|
||||
|
||||
int total_blocks = NUM_BLOCKS;
|
||||
|
||||
void cleanup() {
|
||||
if (sbuf) {
|
||||
vx_buf_release(sbuf);
|
||||
}
|
||||
if (dbuf) {
|
||||
vx_buf_release(dbuf);
|
||||
}
|
||||
if (device) {
|
||||
vx_dev_close(device);
|
||||
}
|
||||
}
|
||||
|
||||
int run_memcopy_test(vx_buffer_h sbuf,
|
||||
vx_buffer_h dbuf,
|
||||
uint32_t address,
|
||||
uint64_t value,
|
||||
int num_blocks) {
|
||||
int run_memcopy_test(uint32_t dev_addr, uint64_t value, int num_blocks) {
|
||||
int errors = 0;
|
||||
|
||||
// write sbuf data
|
||||
// update source buffer
|
||||
for (int i = 0; i < (64 * num_blocks) / 8; ++i) {
|
||||
((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;
|
||||
((uint64_t*)vx_host_ptr(buffer))[i] = shuffle(i, value);
|
||||
}
|
||||
|
||||
// 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));
|
||||
RT_CHECK(vx_copy_to_dev(buffer, dev_addr, 64 * num_blocks, 0));
|
||||
|
||||
// clear destination buffer
|
||||
for (int i = 0; i < (64 * num_blocks) / 8; ++i) {
|
||||
((uint64_t*)vx_host_ptr(buffer))[i] = 0;
|
||||
}
|
||||
|
||||
// read buffer from local memory
|
||||
std::cout << "read buffer from local memory" << std::endl;
|
||||
RT_CHECK(vx_copy_from_dev(dbuf, address, 64 * num_blocks, 0));
|
||||
RT_CHECK(vx_copy_from_dev(buffer, dev_addr, 64 * num_blocks, 0));
|
||||
|
||||
// verify result
|
||||
std::cout << "verify result" << std::endl;
|
||||
for (int i = 0; i < (64 * num_blocks) / 8; ++i) {
|
||||
auto curr = ((uint64_t*)vx_host_ptr(dbuf))[i];
|
||||
auto curr = ((uint64_t*)vx_host_ptr(buffer))[i];
|
||||
auto ref = shuffle(i, value);
|
||||
if (curr != ref) {
|
||||
std::cout << "error at 0x" << std::hex << (address + 8 * i)
|
||||
std::cout << "error at 0x" << std::hex << (dev_addr + 8 * i)
|
||||
<< ": actual 0x" << curr << ", expected 0x" << ref << std::endl;
|
||||
++errors;
|
||||
}
|
||||
@@ -102,35 +106,19 @@ int run_memcopy_test(vx_buffer_h sbuf,
|
||||
return 0;
|
||||
}
|
||||
|
||||
int run_kernel_test(vx_device_h device,
|
||||
vx_buffer_h sbuf,
|
||||
vx_buffer_h dbuf,
|
||||
const char* program,
|
||||
int num_blocks) {
|
||||
int errors = 0;
|
||||
|
||||
uint64_t seed = 0x0badf00d40ff40ff;
|
||||
int run_kernel_test(const kernel_arg_t& kernel_arg,
|
||||
uint32_t buf_size,
|
||||
uint32_t num_points) {
|
||||
int errors = 0;
|
||||
|
||||
int src_dev_addr = DEV_MEM_SRC_ADDR;
|
||||
int dest_dev_addr = DEV_MEM_DST_ADDR;
|
||||
|
||||
// write sbuf data
|
||||
for (int i = 0; i < (64 * num_blocks) / 8; ++i) {
|
||||
((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;
|
||||
// update source buffer
|
||||
for (uint32_t i = 0; i < num_points; ++i) {
|
||||
((int32_t*)vx_host_ptr(buffer))[i] = i;
|
||||
}
|
||||
|
||||
// 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));
|
||||
|
||||
// upload program
|
||||
std::cout << "upload program" << std::endl;
|
||||
RT_CHECK(vx_upload_kernel_file(device, program));
|
||||
RT_CHECK(vx_copy_to_dev(buffer, kernel_arg.src_ptr, buf_size, 0));
|
||||
|
||||
// start device
|
||||
std::cout << "start device" << std::endl;
|
||||
@@ -142,19 +130,24 @@ int run_kernel_test(vx_device_h device,
|
||||
|
||||
// flush the caches
|
||||
std::cout << "flush the caches" << std::endl;
|
||||
RT_CHECK(vx_flush_caches(device, dest_dev_addr, 64 * num_blocks));
|
||||
RT_CHECK(vx_flush_caches(device, kernel_arg.dst_ptr, buf_size));
|
||||
|
||||
// clear destination buffer
|
||||
for (uint32_t i = 0; i < num_points; ++i) {
|
||||
((int32_t*)vx_host_ptr(buffer))[i] = 0;
|
||||
}
|
||||
|
||||
// read buffer from local memory
|
||||
std::cout << "read buffer from local memory" << std::endl;
|
||||
RT_CHECK(vx_copy_from_dev(dbuf, dest_dev_addr, 64 * num_blocks, 0));
|
||||
RT_CHECK(vx_copy_from_dev(buffer, kernel_arg.dst_ptr, buf_size, 0));
|
||||
|
||||
// verify result
|
||||
std::cout << "verify result" << std::endl;
|
||||
for (int i = 0; i < (64 * num_blocks) / 8; ++i) {
|
||||
auto curr = ((uint64_t*)vx_host_ptr(dbuf))[i];
|
||||
auto ref = shuffle(i, seed);
|
||||
for (uint32_t i = 0; i < num_points; ++i) {
|
||||
int32_t curr = ((int32_t*)vx_host_ptr(buffer))[i];
|
||||
int32_t ref = i;
|
||||
if (curr != ref) {
|
||||
std::cout << "error at 0x" << std::hex << (dest_dev_addr + 8 * i)
|
||||
std::cout << "error at value " << i
|
||||
<< ": actual 0x" << curr << ", expected 0x" << ref << std::endl;
|
||||
++errors;
|
||||
}
|
||||
@@ -170,33 +163,66 @@ int run_kernel_test(vx_device_h device,
|
||||
}
|
||||
|
||||
int main(int argc, char *argv[]) {
|
||||
size_t value;
|
||||
kernel_arg_t kernel_arg;
|
||||
|
||||
// parse command arguments
|
||||
parse_args(argc, argv);
|
||||
|
||||
std::cout << "total blocks: " << total_blocks << std::endl;
|
||||
if (count == 0) {
|
||||
count = 1;
|
||||
}
|
||||
|
||||
uint32_t max_cores = vx_dev_caps(VX_CAPS_MAX_CORES);
|
||||
uint32_t num_points = max_cores * count;
|
||||
uint32_t num_blocks = (num_points * sizeof(uint32_t) + 63) / 64;
|
||||
uint32_t buf_size = num_blocks * 64;
|
||||
|
||||
std::cout << "number of points: " << num_points << std::endl;
|
||||
std::cout << "buffer size: " << buf_size << " bytes" << std::endl;
|
||||
|
||||
// open device connection
|
||||
std::cout << "open device connection" << std::endl;
|
||||
RT_CHECK(vx_dev_open(&device));
|
||||
|
||||
// create source buffer
|
||||
std::cout << "create source buffer" << std::endl;
|
||||
RT_CHECK(vx_alloc_shared_mem(device, total_blocks * 64, &sbuf));
|
||||
|
||||
// create destination buffer
|
||||
std::cout << "create destination buffer" << std::endl;
|
||||
RT_CHECK(vx_alloc_shared_mem(device, total_blocks * 64, &dbuf));
|
||||
// allocate device memory
|
||||
RT_CHECK(vx_alloc_dev_mem(device, buf_size, &value));
|
||||
kernel_arg.src_ptr = value;
|
||||
RT_CHECK(vx_alloc_dev_mem(device, buf_size, &value));
|
||||
kernel_arg.dst_ptr = value;
|
||||
|
||||
kernel_arg.count = count;
|
||||
|
||||
std::cout << "dev_src=" << std::hex << kernel_arg.src_ptr << std::endl;
|
||||
std::cout << "dev_dst=" << std::hex << kernel_arg.dst_ptr << std::endl;
|
||||
|
||||
// allocate shared memory
|
||||
std::cout << "allocate shared memory" << std::endl;
|
||||
uint32_t alloc_size = std::max<uint32_t>(buf_size, sizeof(kernel_arg_t));
|
||||
RT_CHECK(vx_alloc_shared_mem(device, alloc_size, &buffer));
|
||||
|
||||
// run tests
|
||||
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, total_blocks));
|
||||
RT_CHECK(run_memcopy_test(kernel_arg.src_ptr, 0x0badf00d00ff00ff, 1));
|
||||
RT_CHECK(run_memcopy_test(kernel_arg.src_ptr, 0x0badf00d40ff40ff, num_blocks));
|
||||
}
|
||||
|
||||
if (1 == test || -1 == test) {
|
||||
// upload program
|
||||
std::cout << "upload program" << std::endl;
|
||||
RT_CHECK(vx_upload_kernel_file(device, kernel_file));
|
||||
|
||||
// upload kernel argument
|
||||
std::cout << "upload kernel argument" << std::endl;
|
||||
{
|
||||
auto buf_ptr = (void*)vx_host_ptr(buffer);
|
||||
memcpy(buf_ptr, &kernel_arg, sizeof(kernel_arg_t));
|
||||
RT_CHECK(vx_copy_to_dev(buffer, KERNEL_ARG_DEV_MEM_ADDR, sizeof(kernel_arg_t), 0));
|
||||
}
|
||||
|
||||
std::cout << "run kernel test" << std::endl;
|
||||
RT_CHECK(run_kernel_test(device, sbuf, dbuf, "kernel.bin", total_blocks));
|
||||
RT_CHECK(run_kernel_test(kernel_arg, buf_size, num_points));
|
||||
}
|
||||
|
||||
// cleanup
|
||||
|
||||
@@ -1,8 +1,12 @@
|
||||
#ifndef _COMMON_H_
|
||||
#define _COMMON_H_
|
||||
|
||||
#define DEV_MEM_SRC_ADDR 0x10000040
|
||||
#define DEV_MEM_DST_ADDR 0x20000080
|
||||
#define NUM_BLOCKS 16
|
||||
#define KERNEL_ARG_DEV_MEM_ADDR 0x7ffff000
|
||||
|
||||
struct kernel_arg_t {
|
||||
uint32_t count;
|
||||
uint32_t src_ptr;
|
||||
uint32_t dst_ptr;
|
||||
};
|
||||
|
||||
#endif
|
||||
Binary file not shown.
@@ -4,17 +4,14 @@
|
||||
#include "common.h"
|
||||
|
||||
void main() {
|
||||
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;
|
||||
struct kernel_arg_t* arg = (struct kernel_arg_t*)KERNEL_ARG_DEV_MEM_ADDR;
|
||||
uint32_t count = arg->count;
|
||||
int32_t* src_ptr = (int32_t*)arg->src_ptr;
|
||||
int32_t* dst_ptr = (int32_t*)arg->dst_ptr;
|
||||
|
||||
for (int i = 0; i < num_words_per_core; ++i) {
|
||||
y[offset + i] = x[offset + i];
|
||||
uint32_t offset = vx_core_id() * count;
|
||||
|
||||
for (uint32_t i = 0; i < count; ++i) {
|
||||
dst_ptr[offset + i] = src_ptr[offset + i];
|
||||
}
|
||||
}
|
||||
Reference in New Issue
Block a user