[Offload] Allow "tagging" device info entries with offload keys

[RossBrunton]

When generating the device info tree, nodes can be marked with an
offload Device Info value. The nodes can also look up children based
on this value.

[llvmbot]

@llvm/pr-subscribers-offload

@llvm/pr-subscribers-backend-amdgpu

Author: Ross Brunton (RossBrunton)

Changes

When generating the device info tree, nodes can be marked with an
offload Device Info value. The nodes can also look up children based
on this value.

---

Full diff: https://github.com/llvm/llvm-project/pull/147317.diff

3 Files Affected:

• (modified) offload/plugins-nextgen/amdgpu/src/rtl.cpp (+7-4)
• (modified) offload/plugins-nextgen/common/include/PluginInterface.h (+24-3)
• (modified) offload/plugins-nextgen/cuda/src/rtl.cpp (+5-3)

diff --git a/offload/plugins-nextgen/amdgpu/src/rtl.cpp b/offload/plugins-nextgen/amdgpu/src/rtl.cpp
index 832c31c43b5d2..52ea3283b24ef 100644
--- a/offload/plugins-nextgen/amdgpu/src/rtl.cpp
+++ b/offload/plugins-nextgen/amdgpu/src/rtl.cpp
@@ -2562,7 +2562,8 @@ struct AMDGPUDeviceTy : public GenericDeviceTy, AMDGenericDeviceTy {
     Status2 = hsa_system_get_info(HSA_SYSTEM_INFO_VERSION_MINOR, &Minor);
     if (Status == HSA_STATUS_SUCCESS && Status2 == HSA_STATUS_SUCCESS)
       Info.add("HSA Runtime Version",
-               std::to_string(Major) + "." + std::to_string(Minor));
+               std::to_string(Major) + "." + std::to_string(Minor), "",
+               DeviceInfo::DRIVER_VERSION);
 
     Info.add("HSA OpenMP Device Number", DeviceId);
 
@@ -2572,11 +2573,11 @@ struct AMDGPUDeviceTy : public GenericDeviceTy, AMDGenericDeviceTy {
 
     Status = getDeviceAttrRaw(HSA_AGENT_INFO_NAME, TmpChar);
     if (Status == HSA_STATUS_SUCCESS)
-      Info.add("Device Name", TmpChar);
+      Info.add("Device Name", TmpChar, "", DeviceInfo::NAME);
 
     Status = getDeviceAttrRaw(HSA_AGENT_INFO_VENDOR_NAME, TmpChar);
     if (Status == HSA_STATUS_SUCCESS)
-      Info.add("Vendor Name", TmpChar);
+      Info.add("Vendor Name", TmpChar, "", DeviceInfo::VENDOR);
 
     hsa_device_type_t DevType;
     Status = getDeviceAttrRaw(HSA_AGENT_INFO_DEVICE, DevType);
@@ -2652,7 +2653,9 @@ struct AMDGPUDeviceTy : public GenericDeviceTy, AMDGenericDeviceTy {
 
     Status = getDeviceAttrRaw(HSA_AGENT_INFO_WORKGROUP_MAX_DIM, WorkgrpMaxDim);
     if (Status == HSA_STATUS_SUCCESS) {
-      auto &MaxSize = *Info.add("Workgroup Max Size per Dimension");
+      auto &MaxSize =
+          *Info.add("Workgroup Max Size per Dimension", std::monostate{}, "",
+                    DeviceInfo::MAX_WORK_GROUP_SIZE);
       MaxSize.add("x", WorkgrpMaxDim[0]);
       MaxSize.add("y", WorkgrpMaxDim[1]);
       MaxSize.add("z", WorkgrpMaxDim[2]);
diff --git a/offload/plugins-nextgen/common/include/PluginInterface.h b/offload/plugins-nextgen/common/include/PluginInterface.h
index b5addc13d6644..9dc01ca0277fe 100644
--- a/offload/plugins-nextgen/common/include/PluginInterface.h
+++ b/offload/plugins-nextgen/common/include/PluginInterface.h
@@ -113,6 +113,12 @@ struct AsyncInfoWrapperTy {
   __tgt_async_info *AsyncInfoPtr;
 };
 
+enum class DeviceInfo {
+#define OFFLOAD_DEVINFO(Name, _, Value) Name = Value,
+#include "OffloadInfo.inc"
+#undef OFFLOAD_DEVINFO
+};
+
 /// Tree node for device information
 ///
 /// This information is either printed or used by liboffload to extract certain
@@ -133,6 +139,8 @@ struct InfoTreeNode {
   // * The same key can appear multiple times
   std::unique_ptr<llvm::SmallVector<InfoTreeNode, 8>> Children;
 
+  std::map<DeviceInfo, size_t> DeviceInfoMap;
+
   InfoTreeNode() : InfoTreeNode("", std::monostate{}, "") {}
   InfoTreeNode(std::string Key, VariantType Value, std::string Units)
       : Key(Key), Value(Value), Units(Units) {}
@@ -140,10 +148,12 @@ struct InfoTreeNode {
   /// Add a new info entry as a child of this node. The entry requires at least
   /// a key string in \p Key. The value in \p Value is optional and can be any
   /// type that is representable as a string. The units in \p Units is optional
-  /// and must be a string.
+  /// and must be a string. Providing a device info key allows liboffload to
+  /// use that value for an appropriate olGetDeviceInfo query
   template <typename T = std::monostate>
   InfoTreeNode *add(std::string Key, T Value = T(),
-                    const std::string &Units = std::string()) {
+                    const std::string &Units = std::string(),
+                    std::optional<DeviceInfo> DeviceInfoKey = std::nullopt) {
     assert(!Key.empty() && "Invalid info key");
 
     if (!Children)
@@ -157,7 +167,12 @@ struct InfoTreeNode {
     else
       ValueVariant = std::string{Value};
 
-    return &Children->emplace_back(Key, ValueVariant, Units);
+    auto Ptr = &Children->emplace_back(Key, ValueVariant, Units);
+
+    if (DeviceInfoKey)
+      DeviceInfoMap[*DeviceInfoKey] = Children->size() - 1;
+
+    return Ptr;
   }
 
   std::optional<InfoTreeNode *> get(StringRef Key) {
@@ -171,6 +186,12 @@ struct InfoTreeNode {
     return It;
   }
 
+  std::optional<InfoTreeNode *> get(DeviceInfo Info) {
+    if (DeviceInfoMap.count(Info))
+      return &(*Children)[DeviceInfoMap[Info]];
+    return std::nullopt;
+  }
+
   /// Print all info entries in the tree
   void print() const {
     // Fake an additional indent so that values are offset from the keys
diff --git a/offload/plugins-nextgen/cuda/src/rtl.cpp b/offload/plugins-nextgen/cuda/src/rtl.cpp
index f1164074f9ea9..2365aa200f584 100644
--- a/offload/plugins-nextgen/cuda/src/rtl.cpp
+++ b/offload/plugins-nextgen/cuda/src/rtl.cpp
@@ -932,13 +932,14 @@ struct CUDADeviceTy : public GenericDeviceTy {
     if (Res == CUDA_SUCCESS)
       // For consistency with other drivers, store the version as a string
       // rather than an integer
-      Info.add("CUDA Driver Version", std::to_string(TmpInt));
+      Info.add("CUDA Driver Version", std::to_string(TmpInt), "",
+               DeviceInfo::DRIVER_VERSION);
 
     Info.add("CUDA OpenMP Device Number", DeviceId);
 
     Res = cuDeviceGetName(TmpChar, 1000, Device);
     if (Res == CUDA_SUCCESS)
-      Info.add("Device Name", TmpChar);
+      Info.add("Device Name", TmpChar, "", DeviceInfo::NAME);
 
     Res = cuDeviceTotalMem(&TmpSt, Device);
     if (Res == CUDA_SUCCESS)
@@ -973,7 +974,8 @@ struct CUDADeviceTy : public GenericDeviceTy {
     if (Res == CUDA_SUCCESS)
       Info.add("Maximum Threads per Block", TmpInt);
 
-    auto &MaxBlock = *Info.add("Maximum Block Dimensions", "");
+    auto &MaxBlock = *Info.add("Maximum Block Dimensions", std::monostate{}, "",
+                               DeviceInfo::MAX_WORK_GROUP_SIZE);
     Res = getDeviceAttrRaw(CU_DEVICE_ATTRIBUTE_MAX_BLOCK_DIM_X, TmpInt);
     if (Res == CUDA_SUCCESS)
       MaxBlock.add("x", TmpInt);

[RossBrunton]

@jhuber6 @arsenm I think I've responded to your feedback, can I get another look?

[RossBrunton]

@jhuber6 @arsenm Can I get this looked at?

[RossBrunton]

@arsenm Going to merge this as is (using find). If you want any more changes, just ping me and I'll do them in a follow up task.