-
Notifications
You must be signed in to change notification settings - Fork 304
/
wrapnvml.cu
281 lines (237 loc) · 8.87 KB
/
wrapnvml.cu
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
/*
* A trivial little dlopen()-based wrapper library for the
* NVIDIA NVML library, to allow runtime discovery of NVML on an
* arbitrary system. This is all very hackish and simple-minded, but
* it serves my immediate needs in the short term until NVIDIA provides
* a static NVML wrapper library themselves, hopefully in
* CUDA 6.5 or maybe sometime shortly after.
*
* This trivial code is made available under the "new" 3-clause BSD license,
* and/or any of the GPL licenses you prefer.
* Feel free to use the code and modify as you see fit.
*
* John E. Stone - [email protected]
*
*/
#include <stdio.h>
#include <stdlib.h>
#include "wrapnvml.h"
#include "cuda_runtime.h"
/*
* Wrappers to emulate dlopen() on other systems like Windows
*/
#if defined(_MSC_VER) || defined(_WIN32) || defined(_WIN64)
#include <windows.h>
static void *wrap_dlopen(const char *filename) {
return (void *)LoadLibrary(filename);
}
static void *wrap_dlsym(void *h, const char *sym) {
return (void *)GetProcAddress((HINSTANCE)h, sym);
}
static int wrap_dlclose(void *h) {
/* FreeLibrary returns nonzero on success */
return (!FreeLibrary((HINSTANCE)h));
}
#else
/* assume we can use dlopen itself... */
#include <dlfcn.h>
static void *wrap_dlopen(const char *filename) {
return dlopen(filename, RTLD_NOW);
}
static void *wrap_dlsym(void *h, const char *sym) {
return dlsym(h, sym);
}
static int wrap_dlclose(void *h) {
return dlclose(h);
}
#endif
#if defined(__cplusplus)
extern "C" {
#endif
wrap_nvml_handle * wrap_nvml_create() {
int i=0;
wrap_nvml_handle *nvmlh = NULL;
/*
* We use hard-coded library installation locations for the time being...
* No idea where or if libnvidia-ml.so is installed on MacOS X, a
* deep scouring of the filesystem on one of the Mac CUDA build boxes
* I used turned up nothing, so for now it's not going to work on OSX.
*/
#if defined(_WIN64)
/* 64-bit Windows */
#define libnvidia_ml "%PROGRAMFILES%/NVIDIA Corporation/NVSMI/nvml.dll"
#elif defined(_WIN32) || defined(_MSC_VER)
/* 32-bit Windows */
#define libnvidia_ml "%PROGRAMFILES%/NVIDIA Corporation/NVSMI/nvml.dll"
#elif defined(__linux) && (defined(__i386__) || defined(__ARM_ARCH_7A__))
/* 32-bit linux assumed */
#define libnvidia_ml "/usr/lib32/libnvidia-ml.so"
#elif defined(__linux)
/* 64-bit linux assumed */
#define libnvidia_ml "/usr/lib/libnvidia-ml.so"
#else
#error "Unrecognized platform: need NVML DLL path for this platform..."
#endif
#if WIN32
char tmp[512];
ExpandEnvironmentStringsA(libnvidia_ml, tmp, sizeof(tmp));
#else
char tmp[512] = libnvidia_ml;
#endif
void *nvml_dll = wrap_dlopen(tmp);
if (nvml_dll == NULL)
return NULL;
nvmlh = (wrap_nvml_handle *) calloc(1, sizeof(wrap_nvml_handle));
nvmlh->nvml_dll = nvml_dll;
nvmlh->nvmlInit = (wrap_nvmlReturn_t (*)(void))
wrap_dlsym(nvmlh->nvml_dll, "nvmlInit");
nvmlh->nvmlDeviceGetCount = (wrap_nvmlReturn_t (*)(int *))
wrap_dlsym(nvmlh->nvml_dll, "nvmlDeviceGetCount_v2");
nvmlh->nvmlDeviceGetHandleByIndex = (wrap_nvmlReturn_t (*)(int, wrap_nvmlDevice_t *))
wrap_dlsym(nvmlh->nvml_dll, "nvmlDeviceGetHandleByIndex_v2");
nvmlh->nvmlDeviceGetPciInfo = (wrap_nvmlReturn_t (*)(wrap_nvmlDevice_t, wrap_nvmlPciInfo_t *))
wrap_dlsym(nvmlh->nvml_dll, "nvmlDeviceGetPciInfo");
nvmlh->nvmlDeviceGetName = (wrap_nvmlReturn_t (*)(wrap_nvmlDevice_t, char *, int))
wrap_dlsym(nvmlh->nvml_dll, "nvmlDeviceGetName");
nvmlh->nvmlDeviceGetTemperature = (wrap_nvmlReturn_t (*)(wrap_nvmlDevice_t, int, unsigned int *))
wrap_dlsym(nvmlh->nvml_dll, "nvmlDeviceGetTemperature");
nvmlh->nvmlDeviceGetFanSpeed = (wrap_nvmlReturn_t (*)(wrap_nvmlDevice_t, unsigned int *))
wrap_dlsym(nvmlh->nvml_dll, "nvmlDeviceGetFanSpeed");
nvmlh->nvmlDeviceGetPowerUsage = (wrap_nvmlReturn_t (*)(wrap_nvmlDevice_t, unsigned int *))
wrap_dlsym(nvmlh->nvml_dll, "nvmlDeviceGetPowerUsage");
nvmlh->nvmlShutdown = (wrap_nvmlReturn_t (*)())
wrap_dlsym(nvmlh->nvml_dll, "nvmlShutdown");
if (nvmlh->nvmlInit == NULL ||
nvmlh->nvmlShutdown == NULL ||
nvmlh->nvmlDeviceGetCount == NULL ||
nvmlh->nvmlDeviceGetHandleByIndex == NULL ||
nvmlh->nvmlDeviceGetPciInfo == NULL ||
nvmlh->nvmlDeviceGetName == NULL ||
nvmlh->nvmlDeviceGetTemperature == NULL ||
nvmlh->nvmlDeviceGetFanSpeed == NULL ||
nvmlh->nvmlDeviceGetPowerUsage == NULL
) {
#if 0
printf("Failed to obtain all required NVML function pointers\n");
#endif
wrap_dlclose(nvmlh->nvml_dll);
free(nvmlh);
return NULL;
}
nvmlh->nvmlInit();
nvmlh->nvmlDeviceGetCount(&nvmlh->nvml_gpucount);
/* Query CUDA device count, in case it doesn't agree with NVML, since */
/* CUDA will only report GPUs with compute capability greater than 1.0 */
if (cudaGetDeviceCount(&nvmlh->cuda_gpucount) != cudaSuccess) {
#if 0
printf("Failed to query CUDA device count!\n");
#endif
wrap_dlclose(nvmlh->nvml_dll);
free(nvmlh);
return NULL;
}
nvmlh->devs = (wrap_nvmlDevice_t *) calloc(nvmlh->nvml_gpucount, sizeof(wrap_nvmlDevice_t));
nvmlh->nvml_pci_domain_id = (unsigned int*) calloc(nvmlh->nvml_gpucount, sizeof(unsigned int));
nvmlh->nvml_pci_bus_id = (unsigned int*) calloc(nvmlh->nvml_gpucount, sizeof(unsigned int));
nvmlh->nvml_pci_device_id = (unsigned int*) calloc(nvmlh->nvml_gpucount, sizeof(unsigned int));
nvmlh->nvml_cuda_device_id = (int*) calloc(nvmlh->nvml_gpucount, sizeof(int));
nvmlh->cuda_nvml_device_id = (int*) calloc(nvmlh->cuda_gpucount, sizeof(int));
/* Obtain GPU device handles we're going to need repeatedly... */
for (i=0; i<nvmlh->nvml_gpucount; i++) {
nvmlh->nvmlDeviceGetHandleByIndex(i, &nvmlh->devs[i]);
}
/* Query PCI info for each NVML device, and build table for mapping of */
/* CUDA device IDs to NVML device IDs and vice versa */
for (i=0; i<nvmlh->nvml_gpucount; i++) {
wrap_nvmlPciInfo_t pciinfo;
nvmlh->nvmlDeviceGetPciInfo(nvmlh->devs[i], &pciinfo);
nvmlh->nvml_pci_domain_id[i] = pciinfo.domain;
nvmlh->nvml_pci_bus_id[i] = pciinfo.bus;
nvmlh->nvml_pci_device_id[i] = pciinfo.device;
}
/* build mapping of NVML device IDs to CUDA IDs */
for (i=0; i<nvmlh->nvml_gpucount; i++) {
nvmlh->nvml_cuda_device_id[i] = -1;
}
for (i=0; i<nvmlh->cuda_gpucount; i++) {
cudaDeviceProp props;
nvmlh->cuda_nvml_device_id[i] = -1;
if (cudaGetDeviceProperties(&props, i) == cudaSuccess) {
int j;
for (j=0; j<nvmlh->nvml_gpucount; j++) {
if ((nvmlh->nvml_pci_domain_id[j] == props.pciDomainID) &&
(nvmlh->nvml_pci_bus_id[j] == props.pciBusID) &&
(nvmlh->nvml_pci_device_id[j] == props.pciDeviceID)) {
#if 0
printf("CUDA GPU[%d] matches NVML GPU[%d]\n", i, j);
#endif
nvmlh->nvml_cuda_device_id[j] = i;
nvmlh->cuda_nvml_device_id[i] = j;
}
}
}
}
return nvmlh;
}
int wrap_nvml_destroy(wrap_nvml_handle *nvmlh) {
nvmlh->nvmlShutdown();
wrap_dlclose(nvmlh->nvml_dll);
free(nvmlh);
return 0;
}
int wrap_nvml_get_gpucount(wrap_nvml_handle *nvmlh, int *gpucount) {
*gpucount = nvmlh->nvml_gpucount;
return 0;
}
int wrap_cuda_get_gpucount(wrap_nvml_handle *nvmlh, int *gpucount) {
*gpucount = nvmlh->cuda_gpucount;
return 0;
}
int wrap_nvml_get_gpu_name(wrap_nvml_handle *nvmlh,
int cudaindex,
char *namebuf,
int bufsize) {
int gpuindex = nvmlh->cuda_nvml_device_id[cudaindex];
if (gpuindex < 0 || gpuindex >= nvmlh->nvml_gpucount)
return -1;
if (nvmlh->nvmlDeviceGetName(nvmlh->devs[gpuindex], namebuf, bufsize) != WRAPNVML_SUCCESS)
return -1;
return 0;
}
int wrap_nvml_get_tempC(wrap_nvml_handle *nvmlh,
int cudaindex, unsigned int *tempC) {
wrap_nvmlReturn_t rc;
int gpuindex = nvmlh->cuda_nvml_device_id[cudaindex];
if (gpuindex < 0 || gpuindex >= nvmlh->nvml_gpucount)
return -1;
rc = nvmlh->nvmlDeviceGetTemperature(nvmlh->devs[gpuindex], 0u /* NVML_TEMPERATURE_GPU */, tempC);
if (rc != WRAPNVML_SUCCESS) {
return -1;
}
return 0;
}
int wrap_nvml_get_fanpcnt(wrap_nvml_handle *nvmlh,
int cudaindex, unsigned int *fanpcnt) {
wrap_nvmlReturn_t rc;
int gpuindex = nvmlh->cuda_nvml_device_id[cudaindex];
if (gpuindex < 0 || gpuindex >= nvmlh->nvml_gpucount)
return -1;
rc = nvmlh->nvmlDeviceGetFanSpeed(nvmlh->devs[gpuindex], fanpcnt);
if (rc != WRAPNVML_SUCCESS) {
return -1;
}
return 0;
}
int wrap_nvml_get_power_usage(wrap_nvml_handle *nvmlh,
int cudaindex,
unsigned int *milliwatts) {
int gpuindex = nvmlh->cuda_nvml_device_id[cudaindex];
if (gpuindex < 0 || gpuindex >= nvmlh->nvml_gpucount)
return -1;
if (nvmlh->nvmlDeviceGetPowerUsage(nvmlh->devs[gpuindex], milliwatts) != WRAPNVML_SUCCESS)
return -1;
return 0;
}
#if defined(__cplusplus)
}
#endif