Vortex 2.0 changes:
+ Microarchitecture optimizations + 64-bit support + Xilinx FPGA support + LLVM-16 support + Refactoring and quality control fixes
This commit is contained in:
@@ -1,6 +1,7 @@
|
||||
#include <iostream>
|
||||
#include <unistd.h>
|
||||
#include <string.h>
|
||||
#include <vector>
|
||||
#include <vortex.h>
|
||||
#include "common.h"
|
||||
|
||||
@@ -20,8 +21,8 @@ const char* kernel_file = "kernel.bin";
|
||||
uint32_t count = 0;
|
||||
|
||||
vx_device_h device = nullptr;
|
||||
vx_buffer_h staging_buf = nullptr;
|
||||
kernel_arg_t kernel_arg;
|
||||
std::vector<uint8_t> staging_buf;
|
||||
kernel_arg_t kernel_arg = {};
|
||||
|
||||
static void show_usage() {
|
||||
std::cout << "Vortex Test." << std::endl;
|
||||
@@ -51,9 +52,6 @@ static void parse_args(int argc, char **argv) {
|
||||
}
|
||||
|
||||
void cleanup() {
|
||||
if (staging_buf) {
|
||||
vx_buf_free(staging_buf);
|
||||
}
|
||||
if (device) {
|
||||
vx_mem_free(device, kernel_arg.src0_addr);
|
||||
vx_mem_free(device, kernel_arg.src1_addr);
|
||||
@@ -71,17 +69,17 @@ int run_test(const kernel_arg_t& kernel_arg,
|
||||
|
||||
// wait for completion
|
||||
std::cout << "wait for completion" << std::endl;
|
||||
RT_CHECK(vx_ready_wait(device, MAX_TIMEOUT));
|
||||
RT_CHECK(vx_ready_wait(device, VX_MAX_TIMEOUT));
|
||||
|
||||
// download destination buffer
|
||||
std::cout << "download destination buffer" << std::endl;
|
||||
RT_CHECK(vx_copy_from_dev(staging_buf, kernel_arg.dst_addr, buf_size, 0));
|
||||
RT_CHECK(vx_copy_from_dev(device, staging_buf.data(), kernel_arg.dst_addr, buf_size));
|
||||
|
||||
// verify result
|
||||
std::cout << "verify result" << std::endl;
|
||||
{
|
||||
int errors = 0;
|
||||
auto buf_ptr = (int32_t*)vx_host_ptr(staging_buf);
|
||||
auto buf_ptr = (int32_t*)staging_buf.data();
|
||||
for (uint32_t i = 0; i < num_points; ++i) {
|
||||
int ref = i + i;
|
||||
int cur = buf_ptr[i];
|
||||
@@ -101,9 +99,7 @@ int run_test(const kernel_arg_t& kernel_arg,
|
||||
return 0;
|
||||
}
|
||||
|
||||
int main(int argc, char *argv[]) {
|
||||
size_t value;
|
||||
|
||||
int main(int argc, char *argv[]) {
|
||||
// parse command arguments
|
||||
parse_args(argc, argv);
|
||||
|
||||
@@ -115,12 +111,12 @@ int main(int argc, char *argv[]) {
|
||||
std::cout << "open device connection" << std::endl;
|
||||
RT_CHECK(vx_dev_open(&device));
|
||||
|
||||
uint64_t max_cores, max_warps, max_threads;
|
||||
RT_CHECK(vx_dev_caps(device, VX_CAPS_MAX_CORES, &max_cores));
|
||||
RT_CHECK(vx_dev_caps(device, VX_CAPS_MAX_WARPS, &max_warps));
|
||||
RT_CHECK(vx_dev_caps(device, VX_CAPS_MAX_THREADS, &max_threads));
|
||||
uint64_t num_cores, num_warps, num_threads;
|
||||
RT_CHECK(vx_dev_caps(device, VX_CAPS_NUM_CORES, &num_cores));
|
||||
RT_CHECK(vx_dev_caps(device, VX_CAPS_NUM_WARPS, &num_warps));
|
||||
RT_CHECK(vx_dev_caps(device, VX_CAPS_NUM_THREADS, &num_threads));
|
||||
|
||||
uint32_t num_tasks = max_cores * max_warps * max_threads;
|
||||
uint32_t num_tasks = num_cores * num_warps * num_threads;
|
||||
uint32_t num_points = count * num_tasks;
|
||||
uint32_t buf_size = num_points * sizeof(int32_t);
|
||||
|
||||
@@ -132,64 +128,60 @@ int main(int argc, char *argv[]) {
|
||||
RT_CHECK(vx_upload_kernel_file(device, kernel_file));
|
||||
|
||||
// allocate device memory
|
||||
std::cout << "allocate device memory" << std::endl;
|
||||
|
||||
RT_CHECK(vx_mem_alloc(device, buf_size, &value));
|
||||
kernel_arg.src0_addr = value;
|
||||
RT_CHECK(vx_mem_alloc(device, buf_size, &value));
|
||||
kernel_arg.src1_addr = value;
|
||||
RT_CHECK(vx_mem_alloc(device, buf_size, &value));
|
||||
kernel_arg.dst_addr = value;
|
||||
std::cout << "allocate device memory" << std::endl;
|
||||
RT_CHECK(vx_mem_alloc(device, buf_size, VX_MEM_TYPE_GLOBAL, &kernel_arg.src0_addr));
|
||||
RT_CHECK(vx_mem_alloc(device, buf_size, VX_MEM_TYPE_GLOBAL, &kernel_arg.src1_addr));
|
||||
RT_CHECK(vx_mem_alloc(device, buf_size, VX_MEM_TYPE_GLOBAL, &kernel_arg.dst_addr));
|
||||
|
||||
kernel_arg.num_tasks = num_tasks;
|
||||
kernel_arg.task_size = count;
|
||||
|
||||
std::cout << "dev_src0=" << std::hex << kernel_arg.src0_addr << std::endl;
|
||||
std::cout << "dev_src1=" << std::hex << kernel_arg.src1_addr << std::endl;
|
||||
std::cout << "dev_dst=" << std::hex << kernel_arg.dst_addr << std::endl;
|
||||
std::cout << "dev_src0=0x" << std::hex << kernel_arg.src0_addr << std::endl;
|
||||
std::cout << "dev_src1=0x" << std::hex << kernel_arg.src1_addr << std::endl;
|
||||
std::cout << "dev_dst=0x" << std::hex << kernel_arg.dst_addr << std::endl;
|
||||
|
||||
// allocate shared memory
|
||||
std::cout << "allocate shared memory" << std::endl;
|
||||
// allocate staging buffer
|
||||
std::cout << "allocate staging buffer" << std::endl;
|
||||
uint32_t alloc_size = std::max<uint32_t>(buf_size, sizeof(kernel_arg_t));
|
||||
RT_CHECK(vx_buf_alloc(device, alloc_size, &staging_buf));
|
||||
staging_buf.resize(alloc_size);
|
||||
|
||||
// upload kernel argument
|
||||
std::cout << "upload kernel argument" << std::endl;
|
||||
{
|
||||
auto buf_ptr = (int*)vx_host_ptr(staging_buf);
|
||||
auto buf_ptr = (int*)staging_buf.data();
|
||||
memcpy(buf_ptr, &kernel_arg, sizeof(kernel_arg_t));
|
||||
RT_CHECK(vx_copy_to_dev(staging_buf, KERNEL_ARG_DEV_MEM_ADDR, sizeof(kernel_arg_t), 0));
|
||||
RT_CHECK(vx_copy_to_dev(device, KERNEL_ARG_DEV_MEM_ADDR, staging_buf.data(), sizeof(kernel_arg_t)));
|
||||
}
|
||||
|
||||
// upload source buffer0
|
||||
{
|
||||
auto buf_ptr = (int32_t*)vx_host_ptr(staging_buf);
|
||||
std::cout << "upload source buffer0" << std::endl;
|
||||
auto buf_ptr = (int32_t*)staging_buf.data();
|
||||
for (uint32_t i = 0; i < num_points; ++i) {
|
||||
buf_ptr[i] = i-1;
|
||||
}
|
||||
}
|
||||
RT_CHECK(vx_copy_to_dev(device, kernel_arg.src0_addr, staging_buf.data(), buf_size));
|
||||
}
|
||||
std::cout << "upload source buffer0" << std::endl;
|
||||
RT_CHECK(vx_copy_to_dev(staging_buf, kernel_arg.src0_addr, buf_size, 0));
|
||||
|
||||
// upload source buffer1
|
||||
{
|
||||
auto buf_ptr = (int32_t*)vx_host_ptr(staging_buf);
|
||||
std::cout << "upload source buffer1" << std::endl;
|
||||
auto buf_ptr = (int32_t*)staging_buf.data();
|
||||
for (uint32_t i = 0; i < num_points; ++i) {
|
||||
buf_ptr[i] = i+1;
|
||||
}
|
||||
}
|
||||
RT_CHECK(vx_copy_to_dev(device, kernel_arg.src1_addr, staging_buf.data(), buf_size));
|
||||
}
|
||||
std::cout << "upload source buffer1" << std::endl;
|
||||
RT_CHECK(vx_copy_to_dev(staging_buf, kernel_arg.src1_addr, buf_size, 0));
|
||||
|
||||
// clear destination buffer
|
||||
{
|
||||
auto buf_ptr = (int32_t*)vx_host_ptr(staging_buf);
|
||||
std::cout << "clear destination buffer" << std::endl;
|
||||
auto buf_ptr = (int32_t*)staging_buf.data();
|
||||
for (uint32_t i = 0; i < num_points; ++i) {
|
||||
buf_ptr[i] = 0xdeadbeef;
|
||||
}
|
||||
}
|
||||
RT_CHECK(vx_copy_to_dev(device, kernel_arg.dst_addr, staging_buf.data(), buf_size));
|
||||
}
|
||||
std::cout << "clear destination buffer" << std::endl;
|
||||
RT_CHECK(vx_copy_to_dev(staging_buf, kernel_arg.dst_addr, buf_size, 0));
|
||||
|
||||
// run tests
|
||||
std::cout << "run tests" << std::endl;
|
||||
|
||||
Reference in New Issue
Block a user