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

adding tensor regression test.

Browse Source
This commit is contained in:
Blaise Tine
2023-11-14 05:37:46 -08:00
parent ecf546bc4a
commit 4e7a536918
22 changed files with 474 additions and 133 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
tests/regression/demo/common.h
View File

@@ -3,6 +3,10 @@
#define KERNEL_ARG_DEV_MEM_ADDR 0x7ffff000
#ifndef TYPE
#define TYPE float
#endif
typedef struct {
uint32_t num_tasks;
uint32_t task_size;
@@ -11,4 +15,4 @@ typedef struct {
uint64_t dst_addr;
} kernel_arg_t;
#endif
#endif

8
tests/regression/demo/kernel.cpp
View File

@@ -4,11 +4,11 @@
#include "common.h"
void kernel_body(int task_id, kernel_arg_t* __UNIFORM__ arg) {
uint32_t count = arg->task_size;
int32_t* src0_ptr = (int32_t*)arg->src0_addr;
int32_t* src1_ptr = (int32_t*)arg->src1_addr;
int32_t* dst_ptr = (int32_t*)arg->dst_addr;
auto src0_ptr = reinterpret_cast<TYPE*>(arg->src0_addr);
auto src1_ptr = reinterpret_cast<TYPE*>(arg->src1_addr);
auto dst_ptr = reinterpret_cast<TYPE*>(arg->dst_addr);
uint32_t count = arg->task_size;
uint32_t offset = task_id * count;
for (uint32_t i = 0; i < count; ++i) {

94
tests/regression/demo/main.cpp
View File

@@ -5,6 +5,8 @@
#include <vortex.h>
#include "common.h"
#define FLOAT_ULP 6
#define RT_CHECK(_expr) \
do { \
int _ret = _expr; \
@@ -17,10 +19,52 @@
///////////////////////////////////////////////////////////////////////////////
union Float_t {
float f;
int i;
struct {
uint32_t man : 23;
uint32_t exp : 8;
uint32_t sign : 1;
} parts;
};
template <typename Type>
class Comparator {};
template <>
class Comparator<int> {
public:
static const char* type_str() {
return "integer";
}
static bool compare(int a, int b) {
return a == b;
}
};
template <>
class Comparator<float> {
public:
static const char* type_str() {
return "float";
}
static bool compare(float a, float b) {
Float_t fa{a}, fb{b};
auto d = std::abs(fa.i - fb.i);
if (d > FLOAT_ULP) {
std::cout << "*** almost_equal_ulp: a=" << a << ", b=" << b << ", ulp=" << d << ", ia=" << std::hex << fa.i << ", ib=" << fb.i << std::endl;
return false;
}
return true;
}
};
const char* kernel_file = "kernel.bin";
uint32_t count = 0;
uint32_t count = 16;
vx_device_h device = nullptr;
std::vector<TYPE> source_data;
std::vector<uint8_t> staging_buf;
kernel_arg_t kernel_arg = {};
@@ -79,11 +123,11 @@ int run_test(const kernel_arg_t& kernel_arg,
std::cout << "verify result" << std::endl;
{
int errors = 0;
auto buf_ptr = (int32_t*)staging_buf.data();
auto buf_ptr = (TYPE*)staging_buf.data();
for (uint32_t i = 0; i < num_points; ++i) {
int ref = i + i;
int cur = buf_ptr[i];
if (cur != ref) {
auto ref = source_data[2 * i + 0] + source_data[2 * i + 1];
auto cur = buf_ptr[i];
if (!Comparator<TYPE>::compare(cur, ref)) {
std::cout << "error at result #" << std::dec << i
<< std::hex << ": actual 0x" << cur << ", expected 0x" << ref << std::endl;
++errors;
@@ -103,9 +147,7 @@ int main(int argc, char *argv[]) {
// parse command arguments
parse_args(argc, argv);
if (count == 0) {
count = 1;
}
std::srand(50);
// open device connection
std::cout << "open device connection" << std::endl;
@@ -118,8 +160,9 @@ int main(int argc, char *argv[]) {
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);
uint32_t buf_size = num_points * sizeof(TYPE);
std::cout << "data type: " << Comparator<TYPE>::type_str() << std::endl;
std::cout << "number of points: " << num_points << std::endl;
std::cout << "buffer size: " << buf_size << " bytes" << std::endl;
@@ -147,18 +190,22 @@ int main(int argc, char *argv[]) {
// upload kernel argument
std::cout << "upload kernel argument" << std::endl;
{
auto buf_ptr = (int*)staging_buf.data();
memcpy(buf_ptr, &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)));
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)));
// generate source data
source_data.resize(2 * num_points);
for (uint32_t i = 0; i < source_data.size(); ++i) {
auto r = static_cast<float>(std::rand()) / RAND_MAX;
source_data[i] = static_cast<TYPE>(r * 2 * num_points);
}
// upload source buffer0
{
std::cout << "upload source buffer0" << std::endl;
auto buf_ptr = (int32_t*)staging_buf.data();
auto buf_ptr = (TYPE*)staging_buf.data();
for (uint32_t i = 0; i < num_points; ++i) {
buf_ptr[i] = i-1;
buf_ptr[i] = source_data[2 * i + 0];
}
RT_CHECK(vx_copy_to_dev(device, kernel_arg.src0_addr, staging_buf.data(), buf_size));
}
@@ -166,23 +213,18 @@ int main(int argc, char *argv[]) {
// upload source buffer1
{
std::cout << "upload source buffer1" << std::endl;
auto buf_ptr = (int32_t*)staging_buf.data();
auto buf_ptr = (TYPE*)staging_buf.data();
for (uint32_t i = 0; i < num_points; ++i) {
buf_ptr[i] = i+1;
buf_ptr[i] = source_data[2 * i + 1];
}
RT_CHECK(vx_copy_to_dev(device, kernel_arg.src1_addr, staging_buf.data(), buf_size));
}
// clear destination buffer
{
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;
memset(staging_buf.data(), 0, num_points * sizeof(TYPE));
RT_CHECK(vx_copy_to_dev(device, kernel_arg.dst_addr, staging_buf.data(), buf_size));
// run tests
std::cout << "run tests" << std::endl;
RT_CHECK(run_test(kernel_arg, buf_size, num_points));