Skip to content

Commit c39fecf

Browse files
vickiegptvickiegpt
committed
update
1 parent fb3420d commit c39fecf

File tree

7 files changed

+1657
-5
lines changed

7 files changed

+1657
-5
lines changed

‎lib/qemu‎

‎qemu_integration/guest_gpu_test.c‎

Lines changed: 330 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,330 @@
1+
/*
2+
* CXL Type 2 GPU Test Program
3+
* Tests the hetGPU backend through CXL Type 2 device
4+
*
5+
* Compile: gcc -o guest_gpu_test guest_gpu_test.c -ldl
6+
* Run: ./guest_gpu_test
7+
*/
8+
9+
#include <stdio.h>
10+
#include <stdlib.h>
11+
#include <string.h>
12+
#include <fcntl.h>
13+
#include <unistd.h>
14+
#include <sys/mman.h>
15+
#include <sys/ioctl.h>
16+
#include <stdint.h>
17+
#include <dlfcn.h>
18+
19+
/* CXL Type 2 device identifiers */
20+
#define CXL_TYPE2_VENDOR_ID 0x8086
21+
#define CXL_TYPE2_DEVICE_ID 0x0d92
22+
23+
/* CUDA-like types */
24+
typedef int CUresult;
25+
typedef void* CUdevice;
26+
typedef void* CUcontext;
27+
typedef void* CUmodule;
28+
typedef void* CUfunction;
29+
typedef uint64_t CUdeviceptr;
30+
31+
#define CUDA_SUCCESS 0
32+
33+
/* Function pointer types */
34+
typedef CUresult (*cuInit_t)(unsigned int);
35+
typedef CUresult (*cuDeviceGetCount_t)(int*);
36+
typedef CUresult (*cuDeviceGet_t)(CUdevice*, int);
37+
typedef CUresult (*cuDeviceGetName_t)(char*, int, CUdevice);
38+
typedef CUresult (*cuDeviceTotalMem_t)(size_t*, CUdevice);
39+
typedef CUresult (*cuCtxCreate_t)(CUcontext*, unsigned int, CUdevice);
40+
typedef CUresult (*cuMemAlloc_t)(CUdeviceptr*, size_t);
41+
typedef CUresult (*cuMemFree_t)(CUdeviceptr);
42+
typedef CUresult (*cuMemcpyHtoD_t)(CUdeviceptr, const void*, size_t);
43+
typedef CUresult (*cuMemcpyDtoH_t)(void*, CUdeviceptr, size_t);
44+
45+
/* Test using direct PCI access */
46+
int test_pci_device(void)
47+
{
48+
char path[256];
49+
int fd;
50+
uint16_t vendor, device;
51+
52+
printf("=== Testing PCI Device Access ===\n");
53+
54+
/* Try to find the CXL Type 2 device */
55+
for (int bus = 0; bus < 256; bus++) {
56+
for (int dev = 0; dev < 32; dev++) {
57+
snprintf(path, sizeof(path),
58+
"/sys/bus/pci/devices/0000:%02x:%02x.0/vendor", bus, dev);
59+
fd = open(path, O_RDONLY);
60+
if (fd < 0) continue;
61+
62+
char buf[16];
63+
if (read(fd, buf, sizeof(buf)) > 0) {
64+
vendor = strtol(buf, NULL, 16);
65+
close(fd);
66+
67+
snprintf(path, sizeof(path),
68+
"/sys/bus/pci/devices/0000:%02x:%02x.0/device", bus, dev);
69+
fd = open(path, O_RDONLY);
70+
if (fd >= 0 && read(fd, buf, sizeof(buf)) > 0) {
71+
device = strtol(buf, NULL, 16);
72+
close(fd);
73+
74+
if (vendor == CXL_TYPE2_VENDOR_ID && device == CXL_TYPE2_DEVICE_ID) {
75+
printf("Found CXL Type 2 device at %02x:%02x.0\n", bus, dev);
76+
printf(" Vendor: 0x%04x, Device: 0x%04x\n", vendor, device);
77+
78+
/* Read resource info */
79+
snprintf(path, sizeof(path),
80+
"/sys/bus/pci/devices/0000:%02x:%02x.0/resource", bus, dev);
81+
fd = open(path, O_RDONLY);
82+
if (fd >= 0) {
83+
char resource[1024];
84+
ssize_t n = read(fd, resource, sizeof(resource)-1);
85+
if (n > 0) {
86+
resource[n] = '\0';
87+
printf(" Resources:\n%s", resource);
88+
}
89+
close(fd);
90+
}
91+
return 0;
92+
}
93+
}
94+
}
95+
close(fd);
96+
}
97+
}
98+
99+
printf("CXL Type 2 device not found\n");
100+
return -1;
101+
}
102+
103+
/* Test using CUDA driver API (if available) */
104+
int test_cuda_api(void)
105+
{
106+
void *handle;
107+
cuInit_t cuInit;
108+
cuDeviceGetCount_t cuDeviceGetCount;
109+
cuDeviceGet_t cuDeviceGet;
110+
cuDeviceGetName_t cuDeviceGetName;
111+
cuDeviceTotalMem_t cuDeviceTotalMem;
112+
cuCtxCreate_t cuCtxCreate;
113+
cuMemAlloc_t cuMemAlloc;
114+
cuMemFree_t cuMemFree;
115+
cuMemcpyHtoD_t cuMemcpyHtoD;
116+
cuMemcpyDtoH_t cuMemcpyDtoH;
117+
118+
int count;
119+
CUdevice dev;
120+
CUcontext ctx;
121+
CUdeviceptr devPtr;
122+
char name[256];
123+
size_t totalMem;
124+
CUresult err;
125+
126+
printf("\n=== Testing CUDA API ===\n");
127+
128+
/* Try to load CUDA library */
129+
handle = dlopen("libcuda.so.1", RTLD_NOW);
130+
if (!handle) {
131+
handle = dlopen("libcuda.so", RTLD_NOW);
132+
}
133+
if (!handle) {
134+
handle = dlopen("libnvcuda.so", RTLD_NOW);
135+
}
136+
if (!handle) {
137+
printf("Could not load CUDA library: %s\n", dlerror());
138+
printf("This is expected if CUDA is not installed in guest\n");
139+
return -1;
140+
}
141+
142+
printf("CUDA library loaded successfully\n");
143+
144+
/* Load functions */
145+
cuInit = (cuInit_t)dlsym(handle, "cuInit");
146+
cuDeviceGetCount = (cuDeviceGetCount_t)dlsym(handle, "cuDeviceGetCount");
147+
cuDeviceGet = (cuDeviceGet_t)dlsym(handle, "cuDeviceGet");
148+
cuDeviceGetName = (cuDeviceGetName_t)dlsym(handle, "cuDeviceGetName");
149+
cuDeviceTotalMem = (cuDeviceTotalMem_t)dlsym(handle, "cuDeviceTotalMem_v2");
150+
cuCtxCreate = (cuCtxCreate_t)dlsym(handle, "cuCtxCreate_v2");
151+
cuMemAlloc = (cuMemAlloc_t)dlsym(handle, "cuMemAlloc_v2");
152+
cuMemFree = (cuMemFree_t)dlsym(handle, "cuMemFree_v2");
153+
cuMemcpyHtoD = (cuMemcpyHtoD_t)dlsym(handle, "cuMemcpyHtoD_v2");
154+
cuMemcpyDtoH = (cuMemcpyDtoH_t)dlsym(handle, "cuMemcpyDtoH_v2");
155+
156+
if (!cuInit) {
157+
printf("Could not find cuInit\n");
158+
dlclose(handle);
159+
return -1;
160+
}
161+
162+
/* Initialize CUDA */
163+
err = cuInit(0);
164+
if (err != CUDA_SUCCESS) {
165+
printf("cuInit failed: %d\n", err);
166+
dlclose(handle);
167+
return -1;
168+
}
169+
printf("CUDA initialized\n");
170+
171+
/* Get device count */
172+
if (cuDeviceGetCount) {
173+
err = cuDeviceGetCount(&count);
174+
if (err == CUDA_SUCCESS) {
175+
printf("Device count: %d\n", count);
176+
}
177+
}
178+
179+
/* Get device info */
180+
if (cuDeviceGet && count > 0) {
181+
err = cuDeviceGet(&dev, 0);
182+
if (err == CUDA_SUCCESS) {
183+
printf("Got device 0\n");
184+
185+
if (cuDeviceGetName) {
186+
err = cuDeviceGetName(name, sizeof(name), dev);
187+
if (err == CUDA_SUCCESS) {
188+
printf("Device name: %s\n", name);
189+
}
190+
}
191+
192+
if (cuDeviceTotalMem) {
193+
err = cuDeviceTotalMem(&totalMem, dev);
194+
if (err == CUDA_SUCCESS) {
195+
printf("Total memory: %zu MB\n", totalMem / (1024*1024));
196+
}
197+
}
198+
}
199+
}
200+
201+
/* Test memory allocation */
202+
if (cuCtxCreate && cuMemAlloc && cuMemFree && count > 0) {
203+
printf("\n=== Testing Memory Operations ===\n");
204+
205+
err = cuCtxCreate(&ctx, 0, dev);
206+
if (err == CUDA_SUCCESS) {
207+
printf("Context created\n");
208+
209+
/* Allocate device memory */
210+
size_t size = 1024 * 1024; /* 1 MB */
211+
err = cuMemAlloc(&devPtr, size);
212+
if (err == CUDA_SUCCESS) {
213+
printf("Allocated %zu bytes at device address 0x%lx\n", size, (unsigned long)devPtr);
214+
215+
/* Test memcpy if available */
216+
if (cuMemcpyHtoD && cuMemcpyDtoH) {
217+
char *hostBuf = malloc(size);
218+
char *resultBuf = malloc(size);
219+
220+
if (hostBuf && resultBuf) {
221+
/* Fill with pattern */
222+
memset(hostBuf, 0xAB, size);
223+
224+
/* Copy to device */
225+
err = cuMemcpyHtoD(devPtr, hostBuf, size);
226+
if (err == CUDA_SUCCESS) {
227+
printf("Host to device copy succeeded\n");
228+
229+
/* Copy back */
230+
memset(resultBuf, 0, size);
231+
err = cuMemcpyDtoH(resultBuf, devPtr, size);
232+
if (err == CUDA_SUCCESS) {
233+
printf("Device to host copy succeeded\n");
234+
235+
/* Verify */
236+
if (memcmp(hostBuf, resultBuf, size) == 0) {
237+
printf("Data verification PASSED!\n");
238+
} else {
239+
printf("Data verification FAILED!\n");
240+
}
241+
}
242+
}
243+
244+
free(hostBuf);
245+
free(resultBuf);
246+
}
247+
}
248+
249+
cuMemFree(devPtr);
250+
printf("Memory freed\n");
251+
} else {
252+
printf("Memory allocation failed: %d\n", err);
253+
}
254+
} else {
255+
printf("Context creation failed: %d\n", err);
256+
}
257+
}
258+
259+
dlclose(handle);
260+
return 0;
261+
}
262+
263+
/* Test CXL memory region directly */
264+
int test_cxl_memory(void)
265+
{
266+
char path[256];
267+
int fd;
268+
void *map;
269+
uint64_t start, end, flags;
270+
271+
printf("\n=== Testing CXL Memory Region ===\n");
272+
273+
/* Find the device's BAR0 */
274+
snprintf(path, sizeof(path), "/sys/bus/pci/devices/0000:0d:00.0/resource0");
275+
fd = open(path, O_RDWR | O_SYNC);
276+
if (fd < 0) {
277+
printf("Could not open resource0: %s\n", path);
278+
printf("Try running as root\n");
279+
return -1;
280+
}
281+
282+
/* Get resource size from sysfs */
283+
FILE *fp = fopen("/sys/bus/pci/devices/0000:0d:00.0/resource", "r");
284+
if (fp) {
285+
if (fscanf(fp, "0x%lx 0x%lx 0x%lx", &start, &end, &flags) == 3) {
286+
size_t size = end - start + 1;
287+
printf("BAR0: start=0x%lx end=0x%lx size=%zu\n", start, end, size);
288+
289+
if (size > 0 && size < 1024*1024*1024) {
290+
map = mmap(NULL, size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
291+
if (map != MAP_FAILED) {
292+
printf("Mapped BAR0 at %p\n", map);
293+
294+
/* Read some registers */
295+
uint32_t *regs = (uint32_t *)map;
296+
printf("Register[0]: 0x%08x\n", regs[0]);
297+
printf("Register[1]: 0x%08x\n", regs[1]);
298+
printf("Register[2]: 0x%08x\n", regs[2]);
299+
printf("Register[3]: 0x%08x\n", regs[3]);
300+
301+
munmap(map, size);
302+
} else {
303+
perror("mmap failed");
304+
}
305+
}
306+
}
307+
fclose(fp);
308+
}
309+
310+
close(fd);
311+
return 0;
312+
}
313+
314+
int main(int argc, char *argv[])
315+
{
316+
printf("CXL Type 2 GPU Test Program\n");
317+
printf("============================\n\n");
318+
319+
/* Test PCI device access */
320+
test_pci_device();
321+
322+
/* Test CXL memory region */
323+
test_cxl_memory();
324+
325+
/* Test CUDA API */
326+
test_cuda_api();
327+
328+
printf("\n=== Test Complete ===\n");
329+
return 0;
330+
}

0 commit comments

Comments
 (0)