Open MPI logo

Portable Hardware Locality (hwloc) Documentation: v2.11.2

  |   Home   |   Support   |   FAQ   |  
opencl.h
1/*
2 * Copyright © 2012-2023 Inria. All rights reserved.
3 * Copyright © 2013, 2018 Université Bordeaux. All right reserved.
4 * See COPYING in top-level directory.
5 */
6
14#ifndef HWLOC_OPENCL_H
15#define HWLOC_OPENCL_H
16
17#include "hwloc.h"
18#include "hwloc/autogen/config.h"
19#include "hwloc/helper.h"
20#ifdef HWLOC_LINUX_SYS
21#include "hwloc/linux.h"
22#endif
23
24#ifdef __APPLE__
25#include <OpenCL/cl.h>
26#else
27#include <CL/cl.h>
28#endif
29
30#include <stdio.h>
31
32
33#ifdef __cplusplus
34extern "C" {
35#endif
36
37
38/* OpenCL extensions aren't always shipped with default headers, and
39 * they don't always reflect what the installed implementations support.
40 * Try everything and let the implementation return errors when non supported.
41 */
42/* Copyright (c) 2008-2018 The Khronos Group Inc. */
43
44/* needs "cl_khr_pci_bus_info" device extension, but not strictly required for clGetDeviceInfo() */
45typedef struct {
46 cl_uint pci_domain;
47 cl_uint pci_bus;
48 cl_uint pci_device;
49 cl_uint pci_function;
51#define HWLOC_CL_DEVICE_PCI_BUS_INFO_KHR 0x410F
52
53/* needs "cl_amd_device_attribute_query" device extension, but not strictly required for clGetDeviceInfo() */
54#define HWLOC_CL_DEVICE_TOPOLOGY_AMD 0x4037
55typedef union {
56 struct { cl_uint type; cl_uint data[5]; } raw;
57 struct { cl_uint type; cl_char unused[17]; cl_char bus; cl_char device; cl_char function; } pcie;
59#define HWLOC_CL_DEVICE_TOPOLOGY_TYPE_PCIE_AMD 1
60
61/* needs "cl_nv_device_attribute_query" device extension, but not strictly required for clGetDeviceInfo() */
62#define HWLOC_CL_DEVICE_PCI_BUS_ID_NV 0x4008
63#define HWLOC_CL_DEVICE_PCI_SLOT_ID_NV 0x4009
64#define HWLOC_CL_DEVICE_PCI_DOMAIN_ID_NV 0x400A
65
66
85static __hwloc_inline int
87 unsigned *domain, unsigned *bus, unsigned *dev, unsigned *func)
88{
91 cl_uint nvbus, nvslot, nvdomain;
92 cl_int clret;
93
94 clret = clGetDeviceInfo(device, HWLOC_CL_DEVICE_PCI_BUS_INFO_KHR, sizeof(khrbusinfo), &khrbusinfo, NULL);
95 if (CL_SUCCESS == clret) {
96 *domain = (unsigned) khrbusinfo.pci_domain;
97 *bus = (unsigned) khrbusinfo.pci_bus;
98 *dev = (unsigned) khrbusinfo.pci_device;
99 *func = (unsigned) khrbusinfo.pci_function;
100 return 0;
101 }
102
103 clret = clGetDeviceInfo(device, HWLOC_CL_DEVICE_TOPOLOGY_AMD, sizeof(amdtopo), &amdtopo, NULL);
104 if (CL_SUCCESS == clret
105 && HWLOC_CL_DEVICE_TOPOLOGY_TYPE_PCIE_AMD == amdtopo.raw.type) {
106 *domain = 0; /* can't do anything better */
107 /* cl_device_topology_amd stores bus ID in cl_char, dont convert those signed char directly to unsigned int */
108 *bus = (unsigned) (unsigned char) amdtopo.pcie.bus;
109 *dev = (unsigned) (unsigned char) amdtopo.pcie.device;
110 *func = (unsigned) (unsigned char) amdtopo.pcie.function;
111 return 0;
112 }
113
114 clret = clGetDeviceInfo(device, HWLOC_CL_DEVICE_PCI_BUS_ID_NV, sizeof(nvbus), &nvbus, NULL);
115 if (CL_SUCCESS == clret) {
116 clret = clGetDeviceInfo(device, HWLOC_CL_DEVICE_PCI_SLOT_ID_NV, sizeof(nvslot), &nvslot, NULL);
117 if (CL_SUCCESS == clret) {
118 clret = clGetDeviceInfo(device, HWLOC_CL_DEVICE_PCI_DOMAIN_ID_NV, sizeof(nvdomain), &nvdomain, NULL);
119 if (CL_SUCCESS == clret) { /* available since CUDA 10.2 */
120 *domain = nvdomain;
121 } else {
122 *domain = 0;
123 }
124 *bus = nvbus & 0xff;
125 /* non-documented but used in many other projects */
126 *dev = nvslot >> 3;
127 *func = nvslot & 0x7;
128 return 0;
129 }
130 }
131
132 return -1;
133}
134
155static __hwloc_inline int
156hwloc_opencl_get_device_cpuset(hwloc_topology_t topology __hwloc_attribute_unused,
157 cl_device_id device __hwloc_attribute_unused,
158 hwloc_cpuset_t set)
159{
160#if (defined HWLOC_LINUX_SYS)
161 /* If we're on Linux, try AMD/NVIDIA extensions + the sysfs mechanism to get the local cpus */
162#define HWLOC_OPENCL_DEVICE_SYSFS_PATH_MAX 128
163 char path[HWLOC_OPENCL_DEVICE_SYSFS_PATH_MAX];
164 unsigned pcidomain, pcibus, pcidev, pcifunc;
165
166 if (!hwloc_topology_is_thissystem(topology)) {
167 errno = EINVAL;
168 return -1;
169 }
170
171 if (hwloc_opencl_get_device_pci_busid(device, &pcidomain, &pcibus, &pcidev, &pcifunc) < 0) {
173 return 0;
174 }
175
176 sprintf(path, "/sys/bus/pci/devices/%04x:%02x:%02x.%01x/local_cpus", pcidomain, pcibus, pcidev, pcifunc);
177 if (hwloc_linux_read_path_as_cpumask(path, set) < 0
178 || hwloc_bitmap_iszero(set))
180#else
181 /* Non-Linux systems simply get a full cpuset */
183#endif
184 return 0;
185}
186
202static __hwloc_inline hwloc_obj_t
204 unsigned platform_index, unsigned device_index)
205{
206 unsigned x = (unsigned) -1, y = (unsigned) -1;
207 hwloc_obj_t osdev = NULL;
208 while ((osdev = hwloc_get_next_osdev(topology, osdev)) != NULL) {
210 && osdev->name
211 && sscanf(osdev->name, "opencl%ud%u", &x, &y) == 2
212 && platform_index == x && device_index == y)
213 return osdev;
214 }
215 return NULL;
216}
217
239static __hwloc_inline hwloc_obj_t
240hwloc_opencl_get_device_osdev(hwloc_topology_t topology __hwloc_attribute_unused,
241 cl_device_id device __hwloc_attribute_unused)
242{
243 hwloc_obj_t osdev;
244 unsigned pcidomain, pcibus, pcidevice, pcifunc;
245
246 if (hwloc_opencl_get_device_pci_busid(device, &pcidomain, &pcibus, &pcidevice, &pcifunc) < 0) {
247 errno = EINVAL;
248 return NULL;
249 }
250
251 osdev = NULL;
252 while ((osdev = hwloc_get_next_osdev(topology, osdev)) != NULL) {
253 hwloc_obj_t pcidev = osdev->parent;
254 if (strncmp(osdev->name, "opencl", 6))
255 continue;
256 if (pcidev
257 && pcidev->type == HWLOC_OBJ_PCI_DEVICE
258 && pcidev->attr->pcidev.domain == pcidomain
259 && pcidev->attr->pcidev.bus == pcibus
260 && pcidev->attr->pcidev.dev == pcidevice
261 && pcidev->attr->pcidev.func == pcifunc)
262 return osdev;
263 /* if PCI are filtered out, we need a info attr to match on */
264 }
265
266 return NULL;
267}
268
272#ifdef __cplusplus
273} /* extern "C" */
274#endif
275
276
277#endif /* HWLOC_OPENCL_H */
hwloc_bitmap_t hwloc_cpuset_t
A CPU set is a bitmap whose bits are set according to CPU physical OS indexes.
Definition hwloc.h:161
@ HWLOC_OBJ_OSDEV_COPROC
Operating system co-processor device. For instance "opencl0d0" for a OpenCL device,...
Definition hwloc.h:385
@ HWLOC_OBJ_PCI_DEVICE
PCI device (filtered out by default).
Definition hwloc.h:295
struct hwloc_topology * hwloc_topology_t
Topology context.
Definition hwloc.h:742
int hwloc_topology_is_thissystem(hwloc_topology_t restrict topology)
Does the topology context come from this system?
hwloc_const_cpuset_t hwloc_topology_get_complete_cpuset(hwloc_topology_t topology)
Get complete CPU set.
static hwloc_obj_t hwloc_get_next_osdev(hwloc_topology_t topology, hwloc_obj_t prev)
Get the next OS device in the system.
Definition helper.h:1290
int hwloc_bitmap_iszero(hwloc_const_bitmap_t bitmap)
Test whether bitmap bitmap is empty.
int hwloc_bitmap_copy(hwloc_bitmap_t dst, hwloc_const_bitmap_t src)
Copy the contents of bitmap src into the already allocated bitmap dst.
int hwloc_linux_read_path_as_cpumask(const char *path, hwloc_bitmap_t set)
Convert a linux kernel cpumask file path into a hwloc bitmap set.
static int hwloc_opencl_get_device_pci_busid(cl_device_id device, unsigned *domain, unsigned *bus, unsigned *dev, unsigned *func)
Return the domain, bus and device IDs of the OpenCL device device.
Definition opencl.h:86
static int hwloc_opencl_get_device_cpuset(hwloc_topology_t topology, cl_device_id device, hwloc_cpuset_t set)
Get the CPU set of processors that are physically close to OpenCL device device.
Definition opencl.h:156
static hwloc_obj_t hwloc_opencl_get_device_osdev(hwloc_topology_t topology, cl_device_id device)
Get the hwloc OS device object corresponding to OpenCL device deviceX.
Definition opencl.h:240
static hwloc_obj_t hwloc_opencl_get_device_osdev_by_index(hwloc_topology_t topology, unsigned platform_index, unsigned device_index)
Get the hwloc OS device object corresponding to the OpenCL device for the given indexes.
Definition opencl.h:203
Structure of a topology object.
Definition hwloc.h:431
char * name
Object-specific name if any. Mostly used for identifying OS devices and Misc objects where a name str...
Definition hwloc.h:443
hwloc_obj_type_t type
Type of object.
Definition hwloc.h:433
union hwloc_obj_attr_u * attr
Object type-specific Attributes, may be NULL if no attribute value was found.
Definition hwloc.h:450
struct hwloc_obj * parent
Parent, NULL if root (Machine object)
Definition hwloc.h:481
struct hwloc_obj_attr_u::hwloc_pcidev_attr_s pcidev
struct hwloc_obj_attr_u::hwloc_osdev_attr_s osdev
unsigned char dev
Device number (zz in the PCI BDF notation xxxx:yy:zz.t).
Definition hwloc.h:677
unsigned char func
Function number (t in the PCI BDF notation xxxx:yy:zz.t).
Definition hwloc.h:678
unsigned short domain
Domain number (xxxx in the PCI BDF notation xxxx:yy:zz.t). Only 16bits PCI domains are supported by d...
Definition hwloc.h:670
unsigned char bus
Bus number (yy in the PCI BDF notation xxxx:yy:zz.t).
Definition hwloc.h:676
hwloc_obj_osdev_type_t type
Definition hwloc.h:716
Definition opencl.h:45
cl_uint pci_bus
Definition opencl.h:47
cl_uint pci_device
Definition opencl.h:48
cl_uint pci_domain
Definition opencl.h:46
cl_uint pci_function
Definition opencl.h:49
Definition opencl.h:55
cl_char device
Definition opencl.h:57
cl_char bus
Definition opencl.h:57
cl_char function
Definition opencl.h:57
cl_uint type
Definition opencl.h:56
struct hwloc_cl_device_topology_amd::@0 raw
struct hwloc_cl_device_topology_amd::@1 pcie