Hardware Locality (hwloc)  PR-737-20250925.0822.gite8f69c77f
opencl.h
1 /*
2  * SPDX-License-Identifier: BSD-3-Clause
3  * Copyright © 2012-2024 Inria. All rights reserved.
4  * Copyright © 2013, 2018 Université Bordeaux. All right reserved.
5  * See COPYING in top-level directory.
6  */
7 
15 #ifndef HWLOC_OPENCL_H
16 #define HWLOC_OPENCL_H
17 
18 #include "hwloc.h"
19 #include "hwloc/autogen/config.h"
20 #include "hwloc/helper.h"
21 #ifdef HWLOC_LINUX_SYS
22 #include "hwloc/linux.h"
23 #endif
24 
25 #ifdef __APPLE__
26 #include <OpenCL/cl.h>
27 #else
28 #include <CL/cl.h>
29 #endif
30 
31 #include <stdio.h>
32 
33 
34 #ifdef __cplusplus
35 extern "C" {
36 #endif
37 
38 
39 /* OpenCL extensions aren't always shipped with default headers, and
40  * they don't always reflect what the installed implementations support.
41  * Try everything and let the implementation return errors when non supported.
42  */
43 /* Copyright (c) 2008-2018 The Khronos Group Inc. */
44 
45 /* needs "cl_khr_pci_bus_info" device extension, but not strictly required for clGetDeviceInfo() */
46 typedef struct {
47  cl_uint pci_domain;
48  cl_uint pci_bus;
49  cl_uint pci_device;
50  cl_uint pci_function;
52 #define HWLOC_CL_DEVICE_PCI_BUS_INFO_KHR 0x410F
53 
54 /* needs "cl_amd_device_attribute_query" device extension, but not strictly required for clGetDeviceInfo() */
55 #define HWLOC_CL_DEVICE_TOPOLOGY_AMD 0x4037
56 typedef union {
57  struct { cl_uint type; cl_uint data[5]; } raw;
58  struct { cl_uint type; cl_char unused[17]; cl_char bus; cl_char device; cl_char function; } pcie;
60 #define HWLOC_CL_DEVICE_TOPOLOGY_TYPE_PCIE_AMD 1
61 
62 /* needs "cl_nv_device_attribute_query" device extension, but not strictly required for clGetDeviceInfo() */
63 #define HWLOC_CL_DEVICE_PCI_BUS_ID_NV 0x4008
64 #define HWLOC_CL_DEVICE_PCI_SLOT_ID_NV 0x4009
65 #define HWLOC_CL_DEVICE_PCI_DOMAIN_ID_NV 0x400A
66 
67 
86 static __hwloc_inline int
88  unsigned *domain, unsigned *bus, unsigned *dev, unsigned *func)
89 {
92  cl_uint nvbus, nvslot, nvdomain;
93  cl_int clret;
94 
95  clret = clGetDeviceInfo(device, HWLOC_CL_DEVICE_PCI_BUS_INFO_KHR, sizeof(khrbusinfo), &khrbusinfo, NULL);
96  if (CL_SUCCESS == clret) {
97  *domain = (unsigned) khrbusinfo.pci_domain;
98  *bus = (unsigned) khrbusinfo.pci_bus;
99  *dev = (unsigned) khrbusinfo.pci_device;
100  *func = (unsigned) khrbusinfo.pci_function;
101  return 0;
102  }
103 
104  clret = clGetDeviceInfo(device, HWLOC_CL_DEVICE_TOPOLOGY_AMD, sizeof(amdtopo), &amdtopo, NULL);
105  if (CL_SUCCESS == clret
106  && HWLOC_CL_DEVICE_TOPOLOGY_TYPE_PCIE_AMD == amdtopo.raw.type) {
107  *domain = 0; /* can't do anything better */
108  /* cl_device_topology_amd stores bus ID in cl_char, dont convert those signed char directly to unsigned int */
109  *bus = (unsigned) (unsigned char) amdtopo.pcie.bus;
110  *dev = (unsigned) (unsigned char) amdtopo.pcie.device;
111  *func = (unsigned) (unsigned char) amdtopo.pcie.function;
112  return 0;
113  }
114 
115  clret = clGetDeviceInfo(device, HWLOC_CL_DEVICE_PCI_BUS_ID_NV, sizeof(nvbus), &nvbus, NULL);
116  if (CL_SUCCESS == clret) {
117  clret = clGetDeviceInfo(device, HWLOC_CL_DEVICE_PCI_SLOT_ID_NV, sizeof(nvslot), &nvslot, NULL);
118  if (CL_SUCCESS == clret) {
119  clret = clGetDeviceInfo(device, HWLOC_CL_DEVICE_PCI_DOMAIN_ID_NV, sizeof(nvdomain), &nvdomain, NULL);
120  if (CL_SUCCESS == clret) { /* available since CUDA 10.2 */
121  *domain = nvdomain;
122  } else {
123  *domain = 0;
124  }
125  *bus = nvbus & 0xff;
126  /* non-documented but used in many other projects */
127  *dev = nvslot >> 3;
128  *func = nvslot & 0x7;
129  return 0;
130  }
131  }
132 
133  return -1;
134 }
135 
156 static __hwloc_inline int
157 hwloc_opencl_get_device_cpuset(hwloc_topology_t topology __hwloc_attribute_unused,
158  cl_device_id device __hwloc_attribute_unused,
159  hwloc_cpuset_t set)
160 {
161 #if (defined HWLOC_LINUX_SYS)
162  /* If we're on Linux, try AMD/NVIDIA extensions + the sysfs mechanism to get the local cpus */
163 #define HWLOC_OPENCL_DEVICE_SYSFS_PATH_MAX 128
164  char path[HWLOC_OPENCL_DEVICE_SYSFS_PATH_MAX];
165  unsigned pcidomain, pcibus, pcidev, pcifunc;
166 
167  if (!hwloc_topology_is_thissystem(topology)) {
168  errno = EINVAL;
169  return -1;
170  }
171 
172  if (hwloc_opencl_get_device_pci_busid(device, &pcidomain, &pcibus, &pcidev, &pcifunc) < 0) {
174  return 0;
175  }
176 
177  sprintf(path, "/sys/bus/pci/devices/%04x:%02x:%02x.%01x/local_cpus", pcidomain, pcibus, pcidev, pcifunc);
178  if (hwloc_linux_read_path_as_cpumask(path, set) < 0
179  || hwloc_bitmap_iszero(set))
181 #else
182  /* Non-Linux systems simply get a full cpuset */
184 #endif
185  return 0;
186 }
187 
203 static __hwloc_inline hwloc_obj_t
205  unsigned platform_index, unsigned device_index)
206 {
207  unsigned x = (unsigned) -1, y = (unsigned) -1;
208  hwloc_obj_t osdev = NULL;
209  while ((osdev = hwloc_get_next_osdev(topology, osdev)) != NULL) {
210  if ((osdev->attr->osdev.types & HWLOC_OBJ_OSDEV_COPROC) /* assume future OpenCL devices will be at least COPROC */
211  && osdev->name
212  && sscanf(osdev->name, "opencl%ud%u", &x, &y) == 2
213  && platform_index == x && device_index == y)
214  return osdev;
215  }
216  return NULL;
217 }
218 
240 static __hwloc_inline hwloc_obj_t
241 hwloc_opencl_get_device_osdev(hwloc_topology_t topology __hwloc_attribute_unused,
242  cl_device_id device __hwloc_attribute_unused)
243 {
244  hwloc_obj_t osdev;
245  unsigned pcidomain, pcibus, pcidevice, pcifunc;
246 
247  if (hwloc_opencl_get_device_pci_busid(device, &pcidomain, &pcibus, &pcidevice, &pcifunc) < 0) {
248  errno = EINVAL;
249  return NULL;
250  }
251 
252  osdev = NULL;
253  while ((osdev = hwloc_get_next_osdev(topology, osdev)) != NULL) {
254  hwloc_obj_t pcidev = osdev->parent;
255  if (strncmp(osdev->name, "opencl", 6))
256  continue;
257  if (pcidev
258  && pcidev->type == HWLOC_OBJ_PCI_DEVICE
259  && pcidev->attr->pcidev.domain == pcidomain
260  && pcidev->attr->pcidev.bus == pcibus
261  && pcidev->attr->pcidev.dev == pcidevice
262  && pcidev->attr->pcidev.func == pcifunc)
263  return osdev;
264  /* if PCI are filtered out, we need a info attr to match on */
265  }
266 
267  return NULL;
268 }
269 
273 #ifdef __cplusplus
274 } /* extern "C" */
275 #endif
276 
277 
278 #endif /* HWLOC_OPENCL_H */
hwloc_cl_device_topology_amd::raw
struct hwloc_cl_device_topology_amd::@3 raw
hwloc_cl_device_topology_amd::type
cl_uint type
Definition: opencl.h:57
HWLOC_OBJ_OSDEV_COPROC
@ HWLOC_OBJ_OSDEV_COPROC
Operating system co-processor device. For instance "opencl0d0" for a OpenCL device,...
Definition: hwloc.h:396
hwloc_bitmap_iszero
int hwloc_bitmap_iszero(hwloc_const_bitmap_t bitmap)
Test whether bitmap bitmap is empty.
hwloc_cl_device_pci_bus_info_khr::pci_bus
cl_uint pci_bus
Definition: opencl.h:48
hwloc_opencl_get_device_osdev
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:241
hwloc_obj
Structure of a topology object.
Definition: hwloc.h:488
hwloc_opencl_get_device_pci_busid
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:87
hwloc_topology_t
struct hwloc_topology * hwloc_topology_t
Topology context.
Definition: hwloc.h:778
hwloc_opencl_get_device_cpuset
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:157
hwloc_linux_read_path_as_cpumask
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.
hwloc_cl_device_pci_bus_info_khr
Definition: opencl.h:46
hwloc_cl_device_topology_amd::function
cl_char function
Definition: opencl.h:58
hwloc_obj_attr_u::hwloc_pcidev_attr_s::dev
unsigned char dev
Device number (zz in the PCI BDF notation xxxx:yy:zz.t).
Definition: hwloc.h:727
hwloc_obj_attr_u::hwloc_pcidev_attr_s::bus
unsigned char bus
Bus number (yy in the PCI BDF notation xxxx:yy:zz.t).
Definition: hwloc.h:726
hwloc_obj::parent
struct hwloc_obj * parent
Parent, NULL if root (Machine object)
Definition: hwloc.h:538
hwloc_obj_attr_u::osdev
struct hwloc_obj_attr_u::hwloc_osdev_attr_s osdev
hwloc_topology_get_complete_cpuset
hwloc_const_cpuset_t hwloc_topology_get_complete_cpuset(hwloc_topology_t topology)
Get complete CPU set.
hwloc_obj::attr
union hwloc_obj_attr_u * attr
Object type-specific Attributes, may be NULL if no attribute value was found.
Definition: hwloc.h:507
hwloc_cl_device_topology_amd::pcie
struct hwloc_cl_device_topology_amd::@4 pcie
hwloc_cl_device_topology_amd::device
cl_char device
Definition: opencl.h:58
hwloc_cl_device_pci_bus_info_khr::pci_function
cl_uint pci_function
Definition: opencl.h:50
hwloc_obj_attr_u::hwloc_osdev_attr_s::types
hwloc_obj_osdev_types_t types
OR'ed set of at least one hwloc_obj_osdev_type_e.
Definition: hwloc.h:761
hwloc_obj::name
char * name
Object-specific name if any. Mostly used for identifying OS devices and Misc objects where a name str...
Definition: hwloc.h:500
hwloc_cl_device_topology_amd
Definition: opencl.h:56
hwloc_cpuset_t
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:162
hwloc_opencl_get_device_osdev_by_index
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:204
hwloc_obj_attr_u::hwloc_pcidev_attr_s::domain
unsigned int domain
Domain number (xxxx in the PCI BDF notation xxxx:yy:zz.t).
Definition: hwloc.h:725
HWLOC_OBJ_PCI_DEVICE
@ HWLOC_OBJ_PCI_DEVICE
PCI device (filtered out by default).
Definition: hwloc.h:320
hwloc_cl_device_topology_amd::bus
cl_char bus
Definition: opencl.h:58
hwloc_bitmap_copy
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.
hwloc_cl_device_pci_bus_info_khr::pci_domain
cl_uint pci_domain
Definition: opencl.h:47
hwloc_obj_attr_u::hwloc_pcidev_attr_s::func
unsigned char func
Function number (t in the PCI BDF notation xxxx:yy:zz.t).
Definition: hwloc.h:728
hwloc_obj_attr_u::pcidev
struct hwloc_obj_attr_u::hwloc_pcidev_attr_s pcidev
hwloc_obj::type
hwloc_obj_type_t type
Type of object.
Definition: hwloc.h:490
hwloc_get_next_osdev
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:1291
hwloc_cl_device_pci_bus_info_khr::pci_device
cl_uint pci_device
Definition: opencl.h:49
hwloc_topology_is_thissystem
int hwloc_topology_is_thissystem(hwloc_topology_t restrict topology)
Does the topology context come from this system?