mirror of
https://github.com/vortexgpgpu/vortex.git
synced 2025-04-24 22:07:41 -04:00
enabling explicit kernel address and arguments allocation
This commit is contained in:
parent
b32ea5b750
commit
25d0c76d14
66 changed files with 1223 additions and 1084 deletions
|
@ -1,8 +1,6 @@
|
|||
#ifndef _COMMON_H_
|
||||
#define _COMMON_H_
|
||||
|
||||
#define KERNEL_ARG_DEV_MEM_ADDR 0x7ffff000
|
||||
|
||||
typedef struct {
|
||||
uint32_t num_points;
|
||||
uint64_t src_addr;
|
||||
|
|
|
@ -77,7 +77,7 @@ void kernel_body(int task_id, kernel_arg_t* __UNIFORM__ arg) {
|
|||
}
|
||||
|
||||
int main() {
|
||||
kernel_arg_t* arg = (kernel_arg_t*)KERNEL_ARG_DEV_MEM_ADDR;
|
||||
kernel_arg_t* arg = (kernel_arg_t*)csr_read(VX_CSR_MSCRATCH);
|
||||
vx_spawn_tasks(arg->num_points, (vx_spawn_tasks_cb)kernel_body, arg);
|
||||
return 0;
|
||||
}
|
||||
|
|
|
@ -26,6 +26,8 @@ std::vector<int> ref_data;
|
|||
|
||||
vx_device_h device = nullptr;
|
||||
std::vector<uint8_t> staging_buf;
|
||||
uint64_t kernel_prog_addr;
|
||||
uint64_t kernel_args_addr;
|
||||
kernel_arg_t kernel_arg = {};
|
||||
|
||||
static void show_usage() {
|
||||
|
@ -59,6 +61,8 @@ void cleanup() {
|
|||
if (device) {
|
||||
vx_mem_free(device, kernel_arg.src_addr);
|
||||
vx_mem_free(device, kernel_arg.dst_addr);
|
||||
vx_mem_free(device, kernel_prog_addr);
|
||||
vx_mem_free(device, kernel_args_addr);
|
||||
vx_dev_close(device);
|
||||
}
|
||||
}
|
||||
|
@ -143,45 +147,6 @@ void gen_ref_data(uint32_t num_points) {
|
|||
}
|
||||
}
|
||||
|
||||
int run_test(const kernel_arg_t& kernel_arg,
|
||||
uint32_t buf_size,
|
||||
uint32_t num_points) {
|
||||
// start device
|
||||
std::cout << "start device" << std::endl;
|
||||
RT_CHECK(vx_start(device));
|
||||
|
||||
// wait for completion
|
||||
std::cout << "wait for completion" << std::endl;
|
||||
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(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*)staging_buf.data();
|
||||
for (uint32_t i = 0; i < num_points; ++i) {
|
||||
int ref = ref_data.at(i);
|
||||
int cur = buf_ptr[i];
|
||||
if (cur != ref) {
|
||||
std::cout << "error at result #" << std::dec << i
|
||||
<< std::hex << ": actual 0x" << cur << ", expected 0x" << ref << std::endl;
|
||||
++errors;
|
||||
}
|
||||
}
|
||||
if (errors != 0) {
|
||||
std::cout << "Found " << std::dec << errors << " errors!" << std::endl;
|
||||
std::cout << "FAILED!" << std::endl;
|
||||
return 1;
|
||||
}
|
||||
}
|
||||
|
||||
return 0;
|
||||
}
|
||||
|
||||
int main(int argc, char *argv[]) {
|
||||
// parse command arguments
|
||||
parse_args(argc, argv);
|
||||
|
@ -212,7 +177,7 @@ int main(int argc, char *argv[]) {
|
|||
|
||||
// upload program
|
||||
std::cout << "upload program" << std::endl;
|
||||
RT_CHECK(vx_upload_kernel_file(device, kernel_file));
|
||||
RT_CHECK(vx_upload_kernel_file(device, kernel_file, &kernel_prog_addr));
|
||||
|
||||
// allocate device memory
|
||||
std::cout << "allocate device memory" << std::endl;
|
||||
|
@ -226,15 +191,12 @@ int main(int argc, char *argv[]) {
|
|||
|
||||
// allocate staging buffer
|
||||
std::cout << "allocate staging buffer" << std::endl;
|
||||
uint32_t staging_buf_size = std::max<uint32_t>(src_buf_size,
|
||||
std::max<uint32_t>(dst_buf_size,
|
||||
sizeof(kernel_arg_t)));
|
||||
uint32_t staging_buf_size = std::max<uint32_t>(src_buf_size, dst_buf_size);
|
||||
staging_buf.resize(staging_buf_size);
|
||||
|
||||
// upload kernel argument
|
||||
std::cout << "upload kernel argument" << std::endl;
|
||||
memcpy(staging_buf.data(), &kernel_arg, sizeof(kernel_arg_t));
|
||||
RT_CHECK(vx_copy_to_dev(device, KERNEL_ARG_DEV_MEM_ADDR, staging_buf.data(), sizeof(kernel_arg_t)));
|
||||
RT_CHECK(vx_upload_bytes(device, &kernel_arg, sizeof(kernel_arg_t), &kernel_args_addr));
|
||||
|
||||
// upload source buffer
|
||||
{
|
||||
|
@ -254,9 +216,38 @@ int main(int argc, char *argv[]) {
|
|||
RT_CHECK(vx_copy_to_dev(device, kernel_arg.dst_addr, staging_buf.data(), dst_buf_size));
|
||||
}
|
||||
|
||||
// run tests
|
||||
std::cout << "run tests" << std::endl;
|
||||
RT_CHECK(run_test(kernel_arg, dst_buf_size, num_points));
|
||||
// start device
|
||||
std::cout << "start device" << std::endl;
|
||||
RT_CHECK(vx_start(device, kernel_prog_addr, kernel_args_addr));
|
||||
|
||||
// wait for completion
|
||||
std::cout << "wait for completion" << std::endl;
|
||||
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(device, staging_buf.data(), kernel_arg.dst_addr, dst_buf_size));
|
||||
|
||||
// verify result
|
||||
std::cout << "verify result" << std::endl;
|
||||
{
|
||||
int errors = 0;
|
||||
auto buf_ptr = (int32_t*)staging_buf.data();
|
||||
for (uint32_t i = 0; i < num_points; ++i) {
|
||||
int ref = ref_data.at(i);
|
||||
int cur = buf_ptr[i];
|
||||
if (cur != ref) {
|
||||
std::cout << "error at result #" << std::dec << i
|
||||
<< std::hex << ": actual 0x" << cur << ", expected 0x" << ref << std::endl;
|
||||
++errors;
|
||||
}
|
||||
}
|
||||
if (errors != 0) {
|
||||
std::cout << "Found " << std::dec << errors << " errors!" << std::endl;
|
||||
std::cout << "FAILED!" << std::endl;
|
||||
return 1;
|
||||
}
|
||||
}
|
||||
|
||||
// cleanup
|
||||
std::cout << "cleanup" << std::endl;
|
||||
|
|
Loading…
Add table
Add a link
Reference in a new issue