/*************************************************************************
 * Copyright (c) 2016-2020, NVIDIA CORPORATION. All rights reserved.
 * Modifications Copyright (c) 2019-2020 Advanced Micro Devices, Inc. All rights reserved.
 *
 * See LICENSE.txt for license information
 ************************************************************************/

#include "core.h"
#include "graph.h"
#include "topo.h"
#include "comm.h"
#include "nvmlwrap.h"
#include "net.h"
#include "coll_net.h"
#include <sys/stat.h>
#include <fcntl.h>
#if defined(__HIP_PLATFORM_HCC__) || defined(__HCC__) || defined(__HIPCC__)
#include <hsa/hsa.h>
#include <hsa/hsa_ext_amd.h>
#endif
#include "xml.h"
#include "cpuset.h"
#include <numa.h>

#define BUSID_SIZE (sizeof("0000:00:00.0"))
#define BUSID_REDUCED_SIZE (sizeof("0000:00"))

const char* topoNodeTypeStr[] = { "GPU", "PCI", "NVS", "CPU", "NIC", "NET" };
#if defined(__HIP_PLATFORM_HCC__) || defined(__HCC__) || defined(__HIPCC__)
const char* topoLinkTypeStr[] = { "LOC", "XGMI", "PCI", "",    "",    "SYS", "NET" };
const char* topoPathTypeStr[] = { "LOC", "XGMI", "PIX", "PXB", "PHB", "SYS", "NET" };
#else
const char* topoLinkTypeStr[] = { "LOC", "NVL", "PCI", "",    "",    "SYS", "NET" };
const char* topoPathTypeStr[] = { "LOC", "NVL", "PIX", "PXB", "PHB", "SYS", "NET" };
#endif

/******************************************************************/
/******************* Graph Creation Functions *********************/
/******************************************************************/

// Get an int64 from a PCI path. For example, sys/class/pci0000:00/0000:00:02.0/0000:02:00.0/ will return 0x000002000.
ncclResult_t pciPathToInt64(char* path, int offset, int minOffset, int64_t* id) {
  char* str = path+offset;
  // Remove trailing "/"
  if (*str == '/') str--;
  // Find next /
  while (*str != '/') str--;
  str++;
  int64_t numid;
  NCCLCHECK(busIdToInt64(str, &numid));
  // Ignore subdevice because those should use the same PCI link so we want to merge nodes.
  numid -= numid & 0xf;
  *id = numid;
  return ncclSuccess;
}

static ncclResult_t findLocalCpu(struct ncclTopoNode* node, struct ncclTopoNode** cpu) {
  *cpu = NULL;
  if (node->type == CPU) {
    *cpu = node;
    return ncclSuccess;
  }
  for (int l=0; l<node->nlinks; l++) {
    if (node->links[l].type == LINK_PCI) NCCLCHECK(findLocalCpu(node->links[l].remNode, cpu));
    if (*cpu != NULL) return ncclSuccess;
  }
  return ncclSuccess;
}

int interCpuWidth = 0;
int cpuPciWidth = 0;

static ncclResult_t ncclTopoGetInterCpuWidth(struct ncclTopoNode* cpu, float* width) {
  *width = LOC_WIDTH;
  if (cpu->cpu.arch == NCCL_TOPO_CPU_ARCH_POWER) {
    *width = P9_WIDTH;
    return ncclSuccess;
  }
  if (cpu->cpu.arch == NCCL_TOPO_CPU_ARCH_ARM) {
    *width = ARM_WIDTH;
    return ncclSuccess;
  }
  if (cpu->cpu.arch == NCCL_TOPO_CPU_ARCH_X86 && cpu->cpu.vendor == NCCL_TOPO_CPU_VENDOR_INTEL) {
    *width = cpu->cpu.model == NCCL_TOPO_CPU_TYPE_SKL ? SKL_QPI_WIDTH : QPI_WIDTH;
  }
  return ncclSuccess;
}

enum ncclNvLinkDeviceType {
  ncclNvLinkDeviceUnknown,
  ncclNvLinkDeviceGpu,
  ncclNvLinkDeviceSwitch,
  ncclNvLinkDeviceBridge, // IBM/Power NVLink bridge (Device 04ea)
};

ncclResult_t ncclTopoGetNode(struct ncclTopoSystem* system, struct ncclTopoNode** node, int type, uint64_t id) {
  for (int i=0; i<system->nodes[type].count; i++) {
    if (system->nodes[type].nodes[i].id == id) {
      *node = system->nodes[type].nodes+i;
      return ncclSuccess;
    }
  }
  return ncclSuccess;
}

ncclResult_t ncclTopoCreateNode(struct ncclTopoSystem* system, struct ncclTopoNode** node, int type, uint64_t id) {
  if (system->nodes[type].count == NCCL_TOPO_MAX_NODES) {
    WARN("Error : tried to create too many nodes of type %d\n", type);
    return ncclInternalError;
  }
  struct ncclTopoNode* n = system->nodes[type].nodes+system->nodes[type].count;
  system->nodes[type].count++;
  n->type = type;
  n->id = id;
  if (type == GPU) {
    // Create link to itself (used in some corner cases)
    n->nlinks=1;
    n->links[0].type = LINK_LOC;
    n->links[0].remNode = n;
    n->links[0].width = LOC_WIDTH;
    n->gpu.dev = NCCL_TOPO_UNDEF;
    n->gpu.rank = NCCL_TOPO_UNDEF;
    n->gpu.cudaCompCap = NCCL_TOPO_UNDEF;
  } else if (type == CPU) {
    n->cpu.arch = NCCL_TOPO_UNDEF;
    n->cpu.vendor = NCCL_TOPO_UNDEF;
    n->cpu.model = NCCL_TOPO_UNDEF;
  } else if (type == NET) {
    n->net.asic = 0ULL;
    n->net.port = NCCL_TOPO_UNDEF;
    n->net.width = 0.0;
  }
  *node = n;
  return ncclSuccess;
}

ncclResult_t ncclTopoRemoveNode(struct ncclTopoSystem* system, int type, int index) {
  struct ncclTopoNode* delNode = system->nodes[type].nodes+index;
  for (int t=0; t<NCCL_TOPO_NODE_TYPES; t++) {
    free(delNode->paths[t]);
    for (int n=0; n<system->nodes[t].count; n++) {
      struct ncclTopoNode* node = system->nodes[t].nodes+n;
      if (node == delNode) continue;
      for (int l=0; l<node->nlinks; l++) {
        while (l<node->nlinks && node->links[l].remNode == delNode) {
          memmove(node->links+l, node->links+l+1, (node->nlinks-l-1)*sizeof(struct ncclTopoLink));
          node->nlinks--;
        }
        if (l<node->nlinks && node->links[l].remNode->type == type && node->links[l].remNode >= delNode) {
          node->links[l].remNode--;
        }
      }
    }
  }
  memmove(delNode, delNode+1, (system->nodes[type].count-index-1)*sizeof(struct ncclTopoNode));
  system->nodes[type].count--;
  return ncclSuccess;
}

ncclResult_t ncclTopoConnectNodes(struct ncclTopoNode* node, struct ncclTopoNode* remNode, int type, float width) {
  // Aggregate links into higher width for NVLink
  struct ncclTopoLink* link;
  for (link = node->links; link->remNode; link++) {
    if (link->remNode == remNode && link->type == type) break;
  }
  if (link->remNode == NULL) node->nlinks++;
  link->type = type;
  link->remNode = remNode;
  link->width += width;

  // Sort links in BW descending order
  struct ncclTopoLink linkSave;
  memcpy(&linkSave, link, sizeof(struct ncclTopoLink));
  while (link != node->links) {
    if ((link-1)->width >= linkSave.width) break;
    memcpy(link, link-1, sizeof(struct ncclTopoLink));
    link--;
  }
  memcpy(link, &linkSave, sizeof(struct ncclTopoLink));
  return ncclSuccess;
}

ncclResult_t ncclTopoConnectCpus(struct ncclTopoSystem* system) {
  // And connect all CPU nodes together
  for (int n=0; n<system->nodes[CPU].count; n++) {
    for (int p=0; p<system->nodes[CPU].count; p++) {
      if (n == p) continue;
      float width;
      NCCLCHECK(ncclTopoGetInterCpuWidth(system->nodes[CPU].nodes+n, &width));
      NCCLCHECK(ncclTopoConnectNodes(system->nodes[CPU].nodes+n, system->nodes[CPU].nodes+p, LINK_SYS, width));
    }
  }
  return ncclSuccess;
}

static ncclResult_t ncclTopoPrintRec(struct ncclTopoNode* node, struct ncclTopoNode* prevNode, char* line, int offset) {
  if (node->type == GPU) {
    sprintf(line+offset, "%s/%lX (%d)", topoNodeTypeStr[node->type], node->id, node->gpu.rank);
  } else if (node->type == CPU) {
    sprintf(line+offset, "%s/%lX (%d/%d/%d)", topoNodeTypeStr[node->type], node->id, node->cpu.arch, node->cpu.vendor, node->cpu.model);
  } else {
    sprintf(line+offset, "%s/%lX", topoNodeTypeStr[node->type], node->id);
  }
  INFO(NCCL_GRAPH, "%s", line);
  for (int i=0; i<offset; i++) line[i] = ' ';

  for (int l=0; l<node->nlinks; l++) {
    struct ncclTopoLink* link = node->links+l;
    if (link->type == LINK_LOC) continue;
    if (link->type != LINK_PCI || link->remNode != prevNode) {
      sprintf(line+offset, "+ %s[%2.1f] - ", topoLinkTypeStr[link->type], link->width);
      int nextOffset = strlen(line);
      if (link->type == LINK_PCI) {
        NCCLCHECK(ncclTopoPrintRec(link->remNode, node, line, nextOffset));
      } else {
        if (link->remNode->type == NET) {
          sprintf(line+nextOffset, "%s/%lX (%lx/%d/%f)", topoNodeTypeStr[link->remNode->type], link->remNode->id, link->remNode->net.asic, link->remNode->net.port, link->remNode->net.width);
        } else {
          sprintf(line+nextOffset, "%s/%lX", topoNodeTypeStr[link->remNode->type], link->remNode->id);
        }
        INFO(NCCL_GRAPH, "%s", line);
      }
    }
  }
  return ncclSuccess;
}

ncclResult_t ncclTopoPrint(struct ncclTopoSystem* s) {
  INFO(NCCL_GRAPH, "=== System : maxWidth %2.1f ===", s->maxWidth);
  char line[1024];
  for (int n=0; n<s->nodes[CPU].count; n++) NCCLCHECK(ncclTopoPrintRec(s->nodes[CPU].nodes+n, NULL, line, 0));
  INFO(NCCL_GRAPH, "==========================================");
  NCCLCHECK(ncclTopoPrintPaths(s));
  return ncclSuccess;
}

static ncclResult_t ncclTopoSort(struct ncclTopoNode* node, struct ncclTopoNode* upNode) {
  // Shift all links to have upLink as last link
  if (upNode) {
    int l=0;
    while (node->links[l].remNode != upNode) l++;
    struct ncclTopoLink upLink;
    memcpy(&upLink, node->links+l, sizeof(struct ncclTopoLink));
    while (node->links[l+1].remNode) {
      memcpy(node->links+l, node->links+l+1, sizeof(struct ncclTopoLink));
      l++;
    }
    memcpy(node->links+l, &upLink, sizeof(struct ncclTopoLink));
  }

  // Recursively sort the PCI tree
  for (int l=0; l<node->nlinks; l++) {
    struct ncclTopoLink* link = node->links+l;
    if (link->type == LINK_PCI && link->remNode != upNode) NCCLCHECK(ncclTopoSort(link->remNode, node));
  }
  return ncclSuccess;
}

// We want the graph to be organized to ease/accelerate traversal :
// 1. NVLinks (already the case)
// 2. PCI down
// 3. PCI up
// 4. SYS (already the case)
ncclResult_t ncclTopoSortSystem(struct ncclTopoSystem* system) {
  for (int n=0; n<system->nodes[CPU].count; n++) NCCLCHECK(ncclTopoSort(system->nodes[CPU].nodes+n, NULL));
  return ncclSuccess;
}

ncclResult_t ncclTopoAddNet(struct ncclXmlNode* xmlNet, struct ncclTopoSystem* system, struct ncclTopoNode* nic) {
  int dev;
  NCCLCHECK(xmlGetAttrInt(xmlNet, "dev", &dev));

  struct ncclTopoNode* net;
  NCCLCHECK(ncclTopoCreateNode(system, &net, NET, dev));
  const char* str;
  NCCLCHECK(xmlGetAttr(xmlNet, "guid", &str));
  if (str) sscanf(str, "0x%lx", &net->net.asic);
  else net->net.asic = dev;

  ncclDebugNoWarn = NCCL_GRAPH;
  int mbps;
  if (xmlGetAttrInt(xmlNet, "speed", &mbps) != ncclSuccess) mbps = 0;
  if (mbps <= 0) mbps = 10000; // Some NICs define speed = -1
  net->net.width = mbps / 8000.0;
  if (xmlGetAttrInt(xmlNet, "port", &net->net.port) != ncclSuccess) net->net.port = 0;
  if (xmlGetAttrInt(xmlNet, "gdr", &net->net.gdrSupport) != ncclSuccess) net->net.gdrSupport = 0;
  if (xmlGetAttrInt(xmlNet, "maxconn", &net->net.maxChannels) != ncclSuccess) net->net.maxChannels = MAXCHANNELS;
  if (ncclCollNet && xmlGetAttrInt(xmlNet, "coll", &net->net.collSupport) != ncclSuccess) net->net.collSupport = 0;
  ncclDebugNoWarn = 0;

  NCCLCHECK(ncclTopoConnectNodes(nic, net, LINK_NET, net->net.width));
  NCCLCHECK(ncclTopoConnectNodes(net, nic, LINK_NET, net->net.width));
  return ncclSuccess;
}

ncclResult_t ncclTopoAddNic(struct ncclXmlNode* xmlNic, struct ncclTopoSystem* system, struct ncclTopoNode* nic) {
  for (int s=0; s<xmlNic->nSubs; s++) {
    struct ncclXmlNode* xmlNet = xmlNic->subs[s];
    if (strcmp(xmlNet->name, "net") != 0) continue;
    int index;
    NCCLCHECK(xmlGetAttrIndex(xmlNet, "dev", &index));
    if (index == -1) continue;
    NCCLCHECK(ncclTopoAddNet(xmlNet, system, nic));
  }
  return ncclSuccess;
}

ncclResult_t ncclTopoAddGpu(struct ncclXmlNode* xmlGpu, struct ncclTopoSystem* system, struct ncclTopoNode* gpu) {
  NCCLCHECK(xmlGetAttrInt(xmlGpu, "sm", &gpu->gpu.cudaCompCap));
  NCCLCHECK(xmlGetAttrInt(xmlGpu, "gcn", &gpu->gpu.gcn));
  rcclHipDeviceArch_t arch;
  NCCLCHECK(xmlGetAttrInt(xmlGpu, "arch", &arch.value));
  memcpy(&gpu->gpu.arch, &arch.arch, sizeof(hipDeviceArch_t));
  NCCLCHECK(xmlGetAttrInt(xmlGpu, "rank", &gpu->gpu.rank));
  NCCLCHECK(xmlGetAttrInt(xmlGpu, "dev", &gpu->gpu.dev));
  NCCLCHECK(xmlGetAttrInt(xmlGpu, "gdr", &gpu->gpu.gdrSupport));
  // Do not go any further, nvlinks will be added in a second pass
  return ncclSuccess;
}

struct kvDict kvDictPciClass[] = { { "0x060400", PCI }, { "0x068000", NVS }, { "0x068001", CPU }, { "0x030200", GPU }, { "0x030000", GPU }, { "0x038000", GPU }, { "0x020700", NIC }, { "0x020000", NIC }, { NULL, PCI /* Default fallback value */ } };
struct kvDict kvDictPciGen[] = { { "2.5 GT/s", 15 }, { "5 GT/s", 30 }, { "8 GT/s", 60 }, { "16 GT/s", 120 }, { NULL, 60 /* Default fallback */ } }; // x100 Mbps per lane
ncclResult_t ncclTopoAddPci(struct ncclXmlNode* xmlPci, struct ncclTopoSystem* system, struct ncclTopoNode* parent) {
  const char* str;

  int type;
  NCCLCHECK(xmlGetAttrStr(xmlPci, "class", &str));
  NCCLCHECK(kvConvertToInt(str, &type, kvDictPciClass));

  int64_t busId;
  NCCLCHECK(xmlGetAttrStr(xmlPci, "busid", &str));
  NCCLCHECK(busIdToInt64(str, &busId));

  struct ncclTopoNode* node = NULL;
  struct ncclXmlNode* xmlGpu = NULL;
  NCCLCHECK(xmlGetSub(xmlPci, "gpu", &xmlGpu));
  if (xmlGpu != NULL) {
    type = GPU;
    int index;
    NCCLCHECK(xmlGetAttrIndex(xmlGpu, "rank", &index));
    if (index == -1) return ncclSuccess;
    NCCLCHECK(ncclTopoCreateNode(system, &node, type, busId));
    NCCLCHECK(ncclTopoAddGpu(xmlGpu, system, node));
  }
  struct ncclXmlNode* xmlNic = NULL;
  NCCLCHECK(xmlGetSub(xmlPci, "nic", &xmlNic));
  if (xmlNic != NULL) {
    type = NIC;
    // Ignore sub device ID and merge multi-port NICs into one PCI device.
    busId &= 0xfffffffffffffff0;
    struct ncclTopoNode* nicNode = NULL;
    NCCLCHECK(ncclTopoGetNode(system, &nicNode, type, busId));
    if (nicNode == NULL) {
      NCCLCHECK(ncclTopoCreateNode(system, &nicNode, type, busId));
      node = nicNode; // Connect it to parent later on
    }
    NCCLCHECK(ncclTopoAddNic(xmlNic, system, nicNode));
  } else if (type == PCI) {
    NCCLCHECK(ncclTopoCreateNode(system, &node, type, busId));
    for (int s=0; s<xmlPci->nSubs; s++) {
      struct ncclXmlNode* xmlSubPci = xmlPci->subs[s];
      NCCLCHECK(ncclTopoAddPci(xmlSubPci, system, node));
    }
  }

  if (node) {
    int width, speed;
    NCCLCHECK(xmlGetAttrInt(xmlPci, "link_width", &width));
    NCCLCHECK(xmlGetAttrStr(xmlPci, "link_speed", &str));

    // Manage cases where speed was not indicated in /sys
    if (width == 0) width = 16;
    NCCLCHECK(kvConvertToInt(str, &speed, kvDictPciGen)); // Values in 100Mbps, per lane (we want GB/s in the end)

    NCCLCHECK(ncclTopoConnectNodes(node, parent, LINK_PCI, width*speed/80.0));
    NCCLCHECK(ncclTopoConnectNodes(parent, node, LINK_PCI, width*speed/80.0));
  }
  return ncclSuccess;
}

struct kvDict kvDictCpuArch[] = { { "x86_64", NCCL_TOPO_CPU_ARCH_X86 }, { "arm64", NCCL_TOPO_CPU_ARCH_ARM }, { "ppc64", NCCL_TOPO_CPU_ARCH_POWER }, { NULL, 0 } };
struct kvDict kvDictCpuVendor[] = { { "GenuineIntel", NCCL_TOPO_CPU_VENDOR_INTEL }, { "AuthenticAMD", NCCL_TOPO_CPU_VENDOR_AMD }, { NULL, 0 } };

ncclResult_t ncclTopoAddCpu(struct ncclXmlNode* xmlCpu, struct ncclTopoSystem* system) {
  int numaId;
  NCCLCHECK(xmlGetAttrInt(xmlCpu, "numaid", &numaId));
  struct ncclTopoNode* cpu;
  NCCLCHECK(ncclTopoCreateNode(system, &cpu, CPU, numaId));
  const char* str;
  NCCLCHECK(xmlGetAttr(xmlCpu, "affinity", &str));
  if (str != NULL) {
    NCCLCHECK(ncclStrToCpuset(str, &cpu->cpu.affinity));
  }

  NCCLCHECK(xmlGetAttrStr(xmlCpu, "arch", &str));
  NCCLCHECK(kvConvertToInt(str, &cpu->cpu.arch, kvDictCpuArch));
  if (cpu->cpu.arch == NCCL_TOPO_CPU_ARCH_X86) {
    NCCLCHECK(xmlGetAttrStr(xmlCpu, "vendor", &str));
    NCCLCHECK(kvConvertToInt(str, &cpu->cpu.vendor, kvDictCpuVendor));
    if (cpu->cpu.vendor == NCCL_TOPO_CPU_VENDOR_INTEL) {
      int familyId, modelId;
      NCCLCHECK(xmlGetAttrInt(xmlCpu, "familyid", &familyId));
      NCCLCHECK(xmlGetAttrInt(xmlCpu, "modelid", &modelId));
      cpu->cpu.model = (familyId == 6 && modelId >= 0x55) ? NCCL_TOPO_CPU_TYPE_SKL : NCCL_TOPO_CPU_INTEL_BDW;
    }
    if (cpu->cpu.vendor == NCCL_TOPO_CPU_VENDOR_AMD) {
      int familyId, modelId;
      NCCLCHECK(xmlGetAttrInt(xmlCpu, "familyid", &familyId));
      NCCLCHECK(xmlGetAttrInt(xmlCpu, "modelid", &modelId));
      cpu->cpu.model = (familyId == 143 && modelId >= 49) ? NCCL_TOPO_CPU_TYPE_ROME : NCCL_TOPO_CPU_TYPE_ZEN;
    }
  }
  for (int s=0; s<xmlCpu->nSubs; s++) {
    struct ncclXmlNode* node = xmlCpu->subs[s];
    if (strcmp(node->name, "pci") == 0) NCCLCHECK(ncclTopoAddPci(node, system, cpu));
    if (strcmp(node->name, "nic") == 0) {
      struct ncclTopoNode* nic = NULL;
      NCCLCHECK(ncclTopoGetNode(system, &nic, NIC, 0));
      if (nic == NULL) {
        NCCLCHECK(ncclTopoCreateNode(system, &nic, NIC, 0));
        NCCLCHECK(ncclTopoConnectNodes(cpu, nic, LINK_PCI, LOC_WIDTH));
        NCCLCHECK(ncclTopoConnectNodes(nic, cpu, LINK_PCI, LOC_WIDTH));
      }
      NCCLCHECK(ncclTopoAddNic(node, system, nic));
    }
  }
  return ncclSuccess;
}

#if defined(__HIP_PLATFORM_HCC__) || defined(__HCC__) || defined(__HIPCC__)
ncclResult_t ncclTopoAddXGMI(struct ncclXmlNode* node, struct ncclTopoSystem* system, const char* parentBusId) {
  if (strcmp(node->name, "xgmi") == 0) {
    struct ncclTopoNode* gpu = NULL;
    int64_t pBusId;
    NCCLCHECK(busIdToInt64(parentBusId, &pBusId));
    NCCLCHECK(ncclTopoGetNode(system, &gpu, GPU, pBusId));
    if (gpu == NULL) {
      WARN("Add XGMI error : could not find GPU %lx\n", pBusId);
      return ncclInternalError;
    }
    int count;
    NCCLCHECK(xmlGetAttrInt(node, "count", &count));
    const char* targetClass;
    NCCLCHECK(xmlGetAttrStr(node, "tclass", &targetClass));
    int targetType;
    NCCLCHECK(kvConvertToInt(targetClass, &targetType, kvDictPciClass));
    struct ncclTopoNode* remote = NULL;
    if (targetType == GPU) {
      // NVL P2P connection to another GPU
      const char* target;
      NCCLCHECK(xmlGetAttrStr(node, "target", &target));
      int64_t busId;
      NCCLCHECK(busIdToInt64(target, &busId));
      NCCLCHECK(ncclTopoGetNode(system, &remote, GPU, busId));
    } else if (targetType == CPU) {
      // NVL connection to the local CPU
      NCCLCHECK(findLocalCpu(gpu, &remote));
    } else {
      if (system->nodes[NVS].count == 0) {
        NCCLCHECK(ncclTopoCreateNode(system, &remote, NVS, 0));
      } else {
        remote = system->nodes[NVS].nodes;
      }
    }
    if (remote) {
      int nvlSpeed = VEGA_XGMI_WIDTH;
      NCCLCHECK(ncclTopoConnectNodes(gpu, remote, LINK_NVL, count*nvlSpeed));
      if (remote->type != GPU) {
        NCCLCHECK(ncclTopoConnectNodes(remote, gpu, LINK_NVL, count*nvlSpeed));
      }
    }
  } else {
    const char* busId;
    NCCLCHECK(xmlGetAttr(node, "busid", &busId));
    for (int s=0; s<node->nSubs; s++) {
      NCCLCHECK(ncclTopoAddXGMI(node->subs[s], system, busId ? busId : parentBusId));
    }
  }
  return ncclSuccess;
}
#else
ncclResult_t ncclTopoAddNvLinks(struct ncclXmlNode* node, struct ncclTopoSystem* system, const char* parentBusId) {
  if (strcmp(node->name, "nvlink") == 0) {
    struct ncclTopoNode* gpu = NULL;
    int64_t pBusId;
    NCCLCHECK(busIdToInt64(parentBusId, &pBusId));
    NCCLCHECK(ncclTopoGetNode(system, &gpu, GPU, pBusId));
    if (gpu == NULL) {
      WARN("Add NVLink error : could not find GPU %lx\n", pBusId);
      return ncclInternalError;
    }
    int count;
    NCCLCHECK(xmlGetAttrInt(node, "count", &count));
    const char* targetClass;
    NCCLCHECK(xmlGetAttrStr(node, "tclass", &targetClass));
    int targetType;
    NCCLCHECK(kvConvertToInt(targetClass, &targetType, kvDictPciClass));
    struct ncclTopoNode* remote = NULL;
    if (targetType == GPU) {
      // NVL P2P connection to another GPU
      const char* target;
      NCCLCHECK(xmlGetAttrStr(node, "target", &target));
      int64_t busId;
      NCCLCHECK(busIdToInt64(target, &busId));
      NCCLCHECK(ncclTopoGetNode(system, &remote, GPU, busId));
    } else if (targetType == CPU) {
      // NVL connection to the local CPU
      NCCLCHECK(findLocalCpu(gpu, &remote));
    } else {
      if (system->nodes[NVS].count == 0) {
        NCCLCHECK(ncclTopoCreateNode(system, &remote, NVS, 0));
      } else {
        remote = system->nodes[NVS].nodes;
      }
    }
    if (remote) {
      int nvlSpeed = gpu->gpu.cudaCompCap == 60 ? PASCAL_NVLINK_WIDTH : VOLTA_NVLINK_WIDTH;
      NCCLCHECK(ncclTopoConnectNodes(gpu, remote, LINK_NVL, count*nvlSpeed));
      if (remote->type != GPU) {
        NCCLCHECK(ncclTopoConnectNodes(remote, gpu, LINK_NVL, count*nvlSpeed));
      }
    }
  } else {
    const char* busId;
    NCCLCHECK(xmlGetAttr(node, "busid", &busId));
    for (int s=0; s<node->nSubs; s++) {
      NCCLCHECK(ncclTopoAddNvLinks(node->subs[s], system, busId ? busId : parentBusId));
    }
  }
  return ncclSuccess;
}
#endif

ncclResult_t ncclTopoGetSystemFromXml(struct ncclXml* xml, struct ncclTopoSystem** topoSystem) {
  NCCLCHECK(ncclCalloc(topoSystem, 1));
  struct ncclXmlNode* topNode;
  NCCLCHECK(xmlFindTag(xml, "system", &topNode));
  for (int s=0; s<topNode->nSubs; s++) {
    struct ncclXmlNode* node = topNode->subs[s];
    if (strcmp(node->name, "cpu") == 0) NCCLCHECK(ncclTopoAddCpu(node, *topoSystem));
  }
#if defined(__HIP_PLATFORM_HCC__) || defined(__HCC__) || defined(__HIPCC__)
  NCCLCHECK(ncclTopoAddXGMI(topNode, *topoSystem, NULL));
#else
  NCCLCHECK(ncclTopoAddNvLinks(topNode, *topoSystem, NULL));
#endif

  NCCLCHECK(ncclTopoConnectCpus(*topoSystem));
  NCCLCHECK(ncclTopoSortSystem(*topoSystem));

  return ncclSuccess;
}

NCCL_PARAM(TopoDumpFileRank, "TOPO_DUMP_FILE_RANK", 0);

// Only set values if not already set
static ncclResult_t xmlInitAttrInt(struct ncclXmlNode* node, const char* attrName, const int value) {
  int index;
  NCCLCHECK(xmlGetAttrIndex(node, attrName, &index));
  if (index == -1) {
    index = node->nAttrs++;
    strncpy(node->attrs[index].key, attrName, MAX_STR_LEN);
    snprintf(node->attrs[index].value, MAX_STR_LEN, "%d", value);
  }
  return ncclSuccess;
}
static ncclResult_t xmlInitAttrUint64(struct ncclXmlNode* node, const char* attrName, const uint64_t value) {
  int index;
  NCCLCHECK(xmlGetAttrIndex(node, attrName, &index));
  if (index == -1) {
    index = node->nAttrs++;
    strncpy(node->attrs[index].key, attrName, MAX_STR_LEN);
    snprintf(node->attrs[index].value, MAX_STR_LEN, "0x%lx", value);
  }
  return ncclSuccess;
}


ncclResult_t ncclTopoGetSystem(struct ncclComm* comm, struct ncclTopoSystem** system) {
  struct ncclXml* xml;
  NCCLCHECK(ncclCalloc(&xml, 1));
  char* xmlTopoFile = getenv("NCCL_TOPO_FILE");
  if (xmlTopoFile) {
    INFO(NCCL_ENV, "NCCL_TOPO_FILE set by environment to %s", xmlTopoFile);
    NCCLCHECK(ncclTopoGetXmlFromFile(xmlTopoFile, xml));
  }
  if (xml->maxIndex == 0) {
    // Create top tag
    struct ncclXmlNode* top;
    NCCLCHECK(xmlAddNode(xml, NULL, "system", &top));
    NCCLCHECK(xmlSetAttrInt(top, "version", NCCL_TOPO_XML_VERSION));
  }

  // Auto-detect GPUs if needed
  for (int r=0; r<comm->nRanks; r++) {
    if (comm->peerInfo[r].hostHash == comm->peerInfo[comm->rank].hostHash) {
      char busId[NVML_DEVICE_PCI_BUS_ID_BUFFER_SIZE];
      NCCLCHECK(int64ToBusId(comm->peerInfo[r].busId, busId));
      struct ncclXmlNode* node;
      NCCLCHECK(ncclTopoFillGpu(xml, busId, &node));
      if (node == NULL) continue;
      NCCLCHECK(xmlSetAttrInt(node, "rank", r));
      NCCLCHECK(xmlInitAttrInt(node, "gdr", comm->peerInfo[r].gdrSupport));
    }
  }
  // Auto-detect NICs if needed. net/collnet share the same xml/graph nodes,
  // so we start with collnet so that it has precedence.
  int netDevCount = 0;
  if (ncclCollNet) {
    NCCLCHECK(collNetDevices(&netDevCount));
    for (int n=0; n<netDevCount; n++) {
      ncclNetProperties_t props;
      NCCLCHECK(collNetGetProperties(n, &props));
      struct ncclXmlNode* netNode;
      NCCLCHECK(ncclTopoFillNet(xml, props.pciPath, props.name, &netNode));
      NCCLCHECK(xmlSetAttrInt(netNode, "dev", n));
      NCCLCHECK(xmlInitAttrInt(netNode, "speed", props.speed));
      NCCLCHECK(xmlInitAttrInt(netNode, "port", props.port));
      NCCLCHECK(xmlInitAttrUint64(netNode, "guid", props.guid));
      NCCLCHECK(xmlInitAttrInt(netNode, "maxconn", props.maxComms));
      NCCLCHECK(xmlInitAttrInt(netNode, "gdr", props.ptrSupport & NCCL_PTR_CUDA ? 1 : 0));
      NCCLCHECK(xmlInitAttrInt(netNode, "coll", 1));
    }
  }
  if (netDevCount == 0) {
    NCCLCHECK(ncclNetDevices(&netDevCount));
  }
  for (int n=0; n<netDevCount; n++) {
    ncclNetProperties_t props;
    NCCLCHECK(ncclNetGetProperties(n, &props));
    struct ncclXmlNode* netNode;
    NCCLCHECK(ncclTopoFillNet(xml, props.pciPath, props.name, &netNode));
    NCCLCHECK(xmlSetAttrInt(netNode, "dev", n));
    NCCLCHECK(xmlInitAttrInt(netNode, "speed", props.speed));
    NCCLCHECK(xmlInitAttrInt(netNode, "port", props.port));
    NCCLCHECK(xmlInitAttrUint64(netNode, "guid", props.guid));
    NCCLCHECK(xmlInitAttrInt(netNode, "maxconn", props.maxComms));
    NCCLCHECK(xmlInitAttrInt(netNode, "gdr", props.ptrSupport & NCCL_PTR_CUDA ? 1 : 0));
  }

  xmlTopoFile = getenv("NCCL_TOPO_DUMP_FILE");
  if (xmlTopoFile && comm->rank == ncclParamTopoDumpFileRank()) {
    INFO(NCCL_ENV, "NCCL_TOPO_DUMP_FILE set by environment to %s", xmlTopoFile);
    NCCLCHECK(ncclTopoDumpXmlToFile(xmlTopoFile, xml));
  }

  NCCLCHECK(ncclTopoGetSystemFromXml(xml, system));
  free(xml);
  return ncclSuccess;
}

ncclResult_t ncclTopoGetLocalNet(struct ncclTopoSystem* system, int rank, int64_t* id, int rr) {
  int g;
  NCCLCHECK(ncclTopoRankToIndex(system, rank, &g));
  int minType = PATH_SYS;
  float maxWidth = 0;
  int count = 0;
  int* nets;
  NCCLCHECK(ncclCalloc(&nets, system->nodes[NET].count));
  for (int n=0; n<system->nodes[NET].count; n++) {
    struct ncclTopoLinkList* path = system->nodes[NET].nodes[n].paths[GPU]+g;
    if (path->width > maxWidth || (path->width == maxWidth && path->type < minType)) {
      maxWidth = path->width;
      minType = path->type;
      count = 0;
    }
    if (path->width == maxWidth && path->type == minType) nets[count++] = system->nodes[NET].nodes[n].id;
  }
  *id = nets[rr % count];
  free(nets);
  return ncclSuccess;
}

/****************************/
/* External query functions */
/****************************/

ncclResult_t ncclTopoCpuType(struct ncclTopoSystem* system, int* arch, int* vendor, int* model) {
  *arch = system->nodes[CPU].nodes[0].cpu.arch;
  *vendor = system->nodes[CPU].nodes[0].cpu.vendor;
  *model = system->nodes[CPU].nodes[0].cpu.model;
  return ncclSuccess;
}

NCCL_PARAM(IgnoreCpuAffinity, "IGNORE_CPU_AFFINITY", 0);

ncclResult_t ncclTopoSetAffinity(struct ncclTopoSystem* system, int rank) {
  struct ncclTopoNode* cpu = NULL, *gpu = NULL;
  for (int g=0; g<system->nodes[GPU].count; g++) {
    if (system->nodes[GPU].nodes[g].gpu.rank == rank) {
      gpu = system->nodes[GPU].nodes+g;
      // Find closer CPU
      int cpuIndex = -1, minHops = 0;
      for (int c=0; c<system->nodes[CPU].count; c++) {
        int nHops = system->nodes[GPU].nodes[g].paths[CPU][c].count;
        if (cpuIndex == -1 || nHops < minHops) {
          cpuIndex = c;
          minHops = nHops;
        }
      }
      cpu = system->nodes[CPU].nodes+cpuIndex;
    }
  }
  if (cpu == NULL) {
    WARN("Set CPU affinity : unable to find GPU/CPU for rank %d", rank);
    return ncclInternalError;
  }

  // Query the CPU affinity set we were provided
  cpu_set_t mask;
  SYSCHECK(sched_getaffinity(0, sizeof(cpu_set_t), &mask), "sched_getaffinity");

#ifdef ENABLE_TRACE
  {
    char affinityStr[sizeof(cpu_set_t)*2];
    NCCLCHECK(ncclCpusetToStr(&mask, affinityStr));
    TRACE(NCCL_INIT, "Current affinity for GPU %d is %s", gpu->gpu.dev, affinityStr);
  }
#endif

  // Get the affinity of the CPU close to our GPU.
  cpu_set_t cpuMask = cpu->cpu.affinity;

#ifdef ENABLE_TRACE
  {
    char affinityStr[sizeof(cpu_set_t)*2];
    NCCLCHECK(ncclCpusetToStr(&cpuMask, affinityStr));
    TRACE(NCCL_INIT, "CPU GPU affinity for GPU %d is %s", gpu->gpu.dev, affinityStr);
  }
#endif

  cpu_set_t finalMask;
  if (ncclParamIgnoreCpuAffinity())
    // Ignore the CPU affinity set and use the GPU one instead
    finalMask = cpuMask;
  else
    // Use a subset of the GPU affinity set
    CPU_AND(&finalMask, &mask, &cpuMask);

  // If there is a non empty set, use it to set affinity
  if (CPU_COUNT(&finalMask)) {
    char affinityStr[sizeof(cpu_set_t)*2];
    NCCLCHECK(ncclCpusetToStr(&finalMask, affinityStr));
    INFO(NCCL_INIT, "Setting affinity for GPU %d to %s", gpu->gpu.dev, affinityStr);
    SYSCHECK(sched_setaffinity(0, sizeof(cpu_set_t), &finalMask), "sched_setaffinity");
  }
  return ncclSuccess;
}