Hardware Locality (hwloc)
v2.1-20200123.0330.git0a8b367
|
00001 /* 00002 * Copyright © 2012-2018 Inria. All rights reserved. 00003 * Copyright © 2013, 2018 Université Bordeaux. All right reserved. 00004 * See COPYING in top-level directory. 00005 */ 00006 00014 #ifndef HWLOC_OPENCL_H 00015 #define HWLOC_OPENCL_H 00016 00017 #include "hwloc.h" 00018 #include "hwloc/autogen/config.h" 00019 #include "hwloc/helper.h" 00020 #ifdef HWLOC_LINUX_SYS 00021 #include "hwloc/linux.h" 00022 #endif 00023 00024 #ifdef __APPLE__ 00025 #include <OpenCL/cl.h> 00026 #else 00027 #include <CL/cl.h> 00028 #endif 00029 00030 #include <stdio.h> 00031 00032 00033 #ifdef __cplusplus 00034 extern "C" { 00035 #endif 00036 00037 00038 /* OpenCL extensions aren't always shipped with default headers, and 00039 * they don't always reflect what the installed implementations support. 00040 * Try everything and let the implementation return errors when non supported. 00041 */ 00042 /* Copyright (c) 2008-2018 The Khronos Group Inc. */ 00043 00044 /* needs "cl_amd_device_attribute_query" device extension, but not strictly required for clGetDeviceInfo() */ 00045 #define HWLOC_CL_DEVICE_TOPOLOGY_AMD 0x4037 00046 typedef union { 00047 struct { cl_uint type; cl_uint data[5]; } raw; 00048 struct { cl_uint type; cl_char unused[17]; cl_char bus; cl_char device; cl_char function; } pcie; 00049 } hwloc_cl_device_topology_amd; 00050 #define HWLOC_CL_DEVICE_TOPOLOGY_TYPE_PCIE_AMD 1 00051 00052 /* needs "cl_nv_device_attribute_query" device extension, but not strictly required for clGetDeviceInfo() */ 00053 #define HWLOC_CL_DEVICE_PCI_BUS_ID_NV 0x4008 00054 #define HWLOC_CL_DEVICE_PCI_SLOT_ID_NV 0x4009 00055 00056 00072 static __hwloc_inline int 00073 hwloc_opencl_get_device_pci_busid(cl_device_id device, 00074 unsigned *domain, unsigned *bus, unsigned *dev, unsigned *func) 00075 { 00076 hwloc_cl_device_topology_amd amdtopo; 00077 cl_uint nvbus, nvslot; 00078 cl_int clret; 00079 00080 clret = clGetDeviceInfo(device, HWLOC_CL_DEVICE_TOPOLOGY_AMD, sizeof(amdtopo), &amdtopo, NULL); 00081 if (CL_SUCCESS == clret 00082 && HWLOC_CL_DEVICE_TOPOLOGY_TYPE_PCIE_AMD == amdtopo.raw.type) { 00083 *domain = 0; /* can't do anything better */ 00084 *bus = (unsigned) amdtopo.pcie.bus; 00085 *dev = (unsigned) amdtopo.pcie.device; 00086 *func = (unsigned) amdtopo.pcie.function; 00087 return 0; 00088 } 00089 00090 clret = clGetDeviceInfo(device, HWLOC_CL_DEVICE_PCI_BUS_ID_NV, sizeof(nvbus), &nvbus, NULL); 00091 if (CL_SUCCESS == clret) { 00092 clret = clGetDeviceInfo(device, HWLOC_CL_DEVICE_PCI_SLOT_ID_NV, sizeof(nvslot), &nvslot, NULL); 00093 if (CL_SUCCESS == clret) { 00094 /* FIXME: PCI bus only uses 8bit, assume nvidia hardcodes the domain in higher bits */ 00095 *domain = nvbus >> 8; 00096 *bus = nvbus & 0xff; 00097 /* non-documented but used in many other projects */ 00098 *dev = nvslot >> 3; 00099 *func = nvslot & 0x7; 00100 return 0; 00101 } 00102 } 00103 00104 return -1; 00105 } 00106 00124 static __hwloc_inline int 00125 hwloc_opencl_get_device_cpuset(hwloc_topology_t topology __hwloc_attribute_unused, 00126 cl_device_id device __hwloc_attribute_unused, 00127 hwloc_cpuset_t set) 00128 { 00129 #if (defined HWLOC_LINUX_SYS) 00130 /* If we're on Linux, try AMD/NVIDIA extensions + the sysfs mechanism to get the local cpus */ 00131 #define HWLOC_OPENCL_DEVICE_SYSFS_PATH_MAX 128 00132 char path[HWLOC_OPENCL_DEVICE_SYSFS_PATH_MAX]; 00133 unsigned pcidomain, pcibus, pcidev, pcifunc; 00134 00135 if (!hwloc_topology_is_thissystem(topology)) { 00136 errno = EINVAL; 00137 return -1; 00138 } 00139 00140 if (hwloc_opencl_get_device_pci_busid(device, &pcidomain, &pcibus, &pcidev, &pcifunc) < 0) { 00141 hwloc_bitmap_copy(set, hwloc_topology_get_complete_cpuset(topology)); 00142 return 0; 00143 } 00144 00145 sprintf(path, "/sys/bus/pci/devices/%04x:%02x:%02x.%01x/local_cpus", pcidomain, pcibus, pcidev, pcifunc); 00146 if (hwloc_linux_read_path_as_cpumask(path, set) < 0 00147 || hwloc_bitmap_iszero(set)) 00148 hwloc_bitmap_copy(set, hwloc_topology_get_complete_cpuset(topology)); 00149 #else 00150 /* Non-Linux systems simply get a full cpuset */ 00151 hwloc_bitmap_copy(set, hwloc_topology_get_complete_cpuset(topology)); 00152 #endif 00153 return 0; 00154 } 00155 00171 static __hwloc_inline hwloc_obj_t 00172 hwloc_opencl_get_device_osdev_by_index(hwloc_topology_t topology, 00173 unsigned platform_index, unsigned device_index) 00174 { 00175 unsigned x = (unsigned) -1, y = (unsigned) -1; 00176 hwloc_obj_t osdev = NULL; 00177 while ((osdev = hwloc_get_next_osdev(topology, osdev)) != NULL) { 00178 if (HWLOC_OBJ_OSDEV_COPROC == osdev->attr->osdev.type 00179 && osdev->name 00180 && sscanf(osdev->name, "opencl%ud%u", &x, &y) == 2 00181 && platform_index == x && device_index == y) 00182 return osdev; 00183 } 00184 return NULL; 00185 } 00186 00207 static __hwloc_inline hwloc_obj_t 00208 hwloc_opencl_get_device_osdev(hwloc_topology_t topology __hwloc_attribute_unused, 00209 cl_device_id device __hwloc_attribute_unused) 00210 { 00211 hwloc_obj_t osdev; 00212 unsigned pcidomain, pcibus, pcidevice, pcifunc; 00213 00214 if (hwloc_opencl_get_device_pci_busid(device, &pcidomain, &pcibus, &pcidevice, &pcifunc) < 0) { 00215 errno = EINVAL; 00216 return NULL; 00217 } 00218 00219 osdev = NULL; 00220 while ((osdev = hwloc_get_next_osdev(topology, osdev)) != NULL) { 00221 hwloc_obj_t pcidev = osdev->parent; 00222 if (strncmp(osdev->name, "opencl", 6)) 00223 continue; 00224 if (pcidev 00225 && pcidev->type == HWLOC_OBJ_PCI_DEVICE 00226 && pcidev->attr->pcidev.domain == pcidomain 00227 && pcidev->attr->pcidev.bus == pcibus 00228 && pcidev->attr->pcidev.dev == pcidevice 00229 && pcidev->attr->pcidev.func == pcifunc) 00230 return osdev; 00231 /* if PCI are filtered out, we need a info attr to match on */ 00232 } 00233 00234 return NULL; 00235 } 00236 00240 #ifdef __cplusplus 00241 } /* extern "C" */ 00242 #endif 00243 00244 00245 #endif /* HWLOC_OPENCL_H */