AMDGPU enable global variable read write

Author: Hugh Delaney

source/adapters/hip/enqueue.cpp

@@ -1545,15 +1545,71 @@ UR_APIEXPORT ur_result_t UR_APICALL urEnqueueUSMMemcpy2D(
 }
 
 UR_APIEXPORT ur_result_t UR_APICALL urEnqueueDeviceGlobalVariableWrite(
-    ur_queue_handle_t, ur_program_handle_t, const char *, bool, size_t, size_t,
-    const void *, uint32_t, const ur_event_handle_t *, ur_event_handle_t *) {
-  return UR_RESULT_ERROR_UNSUPPORTED_FEATURE;
+    ur_queue_handle_t hQueue, ur_program_handle_t hProgram, const char *name,
+    bool blockingWrite, size_t count, size_t offset, const void *pSrc,
+    uint32_t numEventsInWaitList, const ur_event_handle_t *phEventWaitList,
+    ur_event_handle_t *phEvent) {
+  // Since HIP requires a the global variable to be referenced by name, we use
+  // metadata to find the correct name to access it by.
+  auto DeviceGlobalNameIt = hProgram->GlobalIDMD.find(name);
+  if (DeviceGlobalNameIt == hProgram->GlobalIDMD.end())
+    return UR_RESULT_ERROR_INVALID_VALUE;
+  std::string DeviceGlobalName = DeviceGlobalNameIt->second;
+
+  ur_result_t Result = UR_RESULT_SUCCESS;
+  try {
+    hipDeviceptr_t DeviceGlobal = 0;
+    size_t DeviceGlobalSize = 0;
+    UR_CHECK_ERROR(hipModuleGetGlobal(&DeviceGlobal, &DeviceGlobalSize,
+                                      hProgram->get(),
+                                      DeviceGlobalName.c_str()));
+
+    if (offset + count > DeviceGlobalSize)
+      return UR_RESULT_ERROR_INVALID_VALUE;
+
+    return urEnqueueUSMMemcpy(
+        hQueue, blockingWrite,
+        reinterpret_cast<void *>(reinterpret_cast<uint8_t *>(DeviceGlobal) +
+                                 offset),
+        pSrc, count, numEventsInWaitList, phEventWaitList, phEvent);
+  } catch (ur_result_t Err) {
+    Result = Err;
+  }
+  return Result;
 }
 
 UR_APIEXPORT ur_result_t UR_APICALL urEnqueueDeviceGlobalVariableRead(
-    ur_queue_handle_t, ur_program_handle_t, const char *, bool, size_t, size_t,
-    void *, uint32_t, const ur_event_handle_t *, ur_event_handle_t *) {
-  return UR_RESULT_ERROR_UNSUPPORTED_FEATURE;
+    ur_queue_handle_t hQueue, ur_program_handle_t hProgram, const char *name,
+    bool blockingRead, size_t count, size_t offset, void *pDst,
+    uint32_t numEventsInWaitList, const ur_event_handle_t *phEventWaitList,
+    ur_event_handle_t *phEvent) {
+  // Since HIP requires a the global variable to be referenced by name, we use
+  // metadata to find the correct name to access it by.
+  auto DeviceGlobalNameIt = hProgram->GlobalIDMD.find(name);
+  if (DeviceGlobalNameIt == hProgram->GlobalIDMD.end())
+    return UR_RESULT_ERROR_INVALID_VALUE;
+  std::string DeviceGlobalName = DeviceGlobalNameIt->second;
+
+  ur_result_t Result = UR_RESULT_SUCCESS;
+  try {
+    hipDeviceptr_t DeviceGlobal = 0;
+    size_t DeviceGlobalSize = 0;
+    UR_CHECK_ERROR(hipModuleGetGlobal(&DeviceGlobal, &DeviceGlobalSize,
+                                      hProgram->get(),
+                                      DeviceGlobalName.c_str()));
+
+    if (offset + count > DeviceGlobalSize)
+      return UR_RESULT_ERROR_INVALID_VALUE;
+
+    return urEnqueueUSMMemcpy(
+        hQueue, blockingRead, pDst,
+        reinterpret_cast<const void *>(
+            reinterpret_cast<uint8_t *>(DeviceGlobal) + offset),
+        count, numEventsInWaitList, phEventWaitList, phEvent);
+  } catch (ur_result_t Err) {
+    Result = Err;
+  }
+  return Result;
 }
 
 UR_APIEXPORT ur_result_t UR_APICALL urEnqueueReadHostPipe(

source/adapters/hip/program.cpp

@@ -78,17 +78,35 @@ void getCoMgrBuildLog(const amd_comgr_data_set_t BuildDataSet, char *BuildLog,
 } // namespace
 #endif
 
+std::pair<std::string, std::string>
+splitMetadataName(const std::string &metadataName) {
+  size_t splitPos = metadataName.rfind('@');
+  if (splitPos == std::string::npos)
+    return std::make_pair(metadataName, std::string{});
+  return std::make_pair(metadataName.substr(0, splitPos),
+                        metadataName.substr(splitPos, metadataName.length()));
+}
+
 ur_result_t
 ur_program_handle_t_::setMetadata(const ur_program_metadata_t *Metadata,
                                   size_t Length) {
   for (size_t i = 0; i < Length; ++i) {
     const ur_program_metadata_t MetadataElement = Metadata[i];
     std::string MetadataElementName{MetadataElement.pName};
 
+    auto [Prefix, Tag] = splitMetadataName(MetadataElementName);
+
     if (MetadataElementName ==
         __SYCL_UR_PROGRAM_METADATA_TAG_NEED_FINALIZATION) {
       assert(MetadataElement.type == UR_PROGRAM_METADATA_TYPE_UINT32);
       IsRelocatable = MetadataElement.value.data32;
+    } else if (Tag == __SYCL_UR_PROGRAM_METADATA_GLOBAL_ID_MAPPING) {
+      const char *MetadataValPtr =
+          reinterpret_cast<const char *>(MetadataElement.value.pData) +
+          sizeof(std::uint64_t);
+      const char *MetadataValPtrEnd =
+          MetadataValPtr + MetadataElement.size - sizeof(std::uint64_t);
+      GlobalIDMD[Prefix] = std::string{MetadataValPtr, MetadataValPtrEnd};
     }
   }
   return UR_RESULT_SUCCESS;

source/adapters/hip/program.hpp

@@ -29,6 +29,8 @@ struct ur_program_handle_t_ {
   // Metadata
   bool IsRelocatable = false;
 
+  std::unordered_map<std::string, std::string> GlobalIDMD;
+
   constexpr static size_t MAX_LOG_SIZE = 8192u;
 
   char ErrorLog[MAX_LOG_SIZE], InfoLog[MAX_LOG_SIZE];