ROCr changes to enable small BAR P2P over xGMI

Change-Id: I6aaa3fe2565cdf7e15d58a7484d6bd5916ffff64

runtime/hsa-runtime/core/runtime/amd_memory_region.cpp

@@ -408,7 +408,7 @@ hsa_status_t MemoryRegion::GetAgentPoolInfo(
       ((IsSystem() && (agent.device_type() == core::Agent::kAmdCpuDevice)) ||
        (agent.node_id() == owner()->node_id()))
           ? HSA_AMD_MEMORY_POOL_ACCESS_ALLOWED_BY_DEFAULT
-          : (IsSystem() || (IsPublic() && link_info.num_hop > 0))
+          : (IsSystem() || (IsLocalMemory() && link_info.num_hop > 0))
                 ? HSA_AMD_MEMORY_POOL_ACCESS_DISALLOWED_BY_DEFAULT
                 : HSA_AMD_MEMORY_POOL_ACCESS_NEVER_ALLOWED;
 
@@ -502,7 +502,7 @@ hsa_status_t MemoryRegion::AllowAccess(uint32_t num_agents,
 
   // If this is a local memory region, the owning gpu always needs to be in
   // the whitelist.
-  if (IsPublic() &&
+  if (IsLocalMemory() &&
       std::find(whitelist_nodes.begin(), whitelist_nodes.end(), owner()->node_id()) ==
           whitelist_nodes.end()) {
     whitelist_nodes.push_back(owner()->node_id());

runtime/hsa-runtime/core/runtime/amd_topology.cpp

@@ -120,6 +120,12 @@ void RegisterLinkInfo(uint32_t node_id, uint32_t num_link) {
         link_info.link_type = HSA_AMD_LINK_INFO_TYPE_INFINBAND;
         debug_print("IOLINK is missing atomic and coherency defaults.\n");
         break;
+      case HSA_IOLINK_TYPE_XGMI:
+        link_info.link_type = HSA_AMD_LINK_INFO_TYPE_XGMI;
+        link_info.atomic_support_32bit = true;
+        link_info.atomic_support_64bit = true;
+        link_info.coherent_support = true;
+        break;
       default:
         debug_print("Unrecognized IOLINK type.\n");
         break;

runtime/hsa-runtime/core/runtime/runtime.cpp

@@ -252,9 +252,8 @@ void Runtime::DestroyAgents() {
   system_regions_coarse_.clear();
 }
 
-void Runtime::SetLinkCount(size_t num_link) {
-  const size_t last_index = GetIndexLinkInfo(0, num_link);
-  link_matrix_.resize(last_index);
+void Runtime::SetLinkCount(size_t num_nodes) {
+  link_matrix_.resize(num_nodes * num_nodes);
 }
 
 void Runtime::RegisterLinkInfo(uint32_t node_id_from, uint32_t node_id_to,
@@ -277,9 +276,7 @@ const Runtime::LinkInfo Runtime::GetLinkInfo(uint32_t node_id_from,
 }
 
 uint32_t Runtime::GetIndexLinkInfo(uint32_t node_id_from, uint32_t node_id_to) {
-  const uint32_t node_id_max = std::max(node_id_from, node_id_to) - 1;
-  const uint32_t node_id_min = std::min(node_id_from, node_id_to);
-  return ((node_id_max * (node_id_max + 1) / 2) + node_id_min);
+  return ((node_id_from * agents_by_node_.size()) + node_id_to);
 }
 
 hsa_status_t Runtime::IterateAgent(hsa_status_t (*callback)(hsa_agent_t agent,

runtime/hsa-runtime/inc/hsa_ext_amd.h

@@ -962,7 +962,12 @@ typedef enum {
   /**
   * Infiniband bus type.
   */
-  HSA_AMD_LINK_INFO_TYPE_INFINBAND = 3
+  HSA_AMD_LINK_INFO_TYPE_INFINBAND = 3,
+
+  /**
+  * xGMI link type.
+  */
+  HSA_AMD_LINK_INFO_TYPE_XGMI = 4
 
 } hsa_amd_link_info_type_t;