Merge pull request #911 from CHIP-SPV/cpu-usm

Properly annotate Intel USM kernels
CHIP-SPV · Aug 17, 2024 · ab44488 · ab44488
2 parents 4813847 + 14c74fd
commit ab44488
Show file tree

Hide file tree

Showing 4 changed files with 63 additions and 15 deletions.
diff --git a/src/backend/OpenCL/CHIPBackendOpenCL.cc b/src/backend/OpenCL/CHIPBackendOpenCL.cc
@@ -281,6 +281,28 @@ annotateIndirectPointers(const CHIPContextOpenCL &Ctx,
                                    AnnotationList.size() * sizeof(void *),
                                    AnnotationList.data());
     CHIPERR_CHECK_LOG_AND_THROW_TABLE(clSetKernelExecInfo);
+
+    if (Ctx.getAllocStrategy() == AllocationStrategy::IntelUSM) {
+      cl_bool param = CL_TRUE;
+      if (Ctx.MemManager_.isHostAllocUsed()) {
+        clStatus = clSetKernelExecInfo(
+            KernelAPIHandle, CL_KERNEL_EXEC_INFO_INDIRECT_HOST_ACCESS_INTEL,
+            sizeof(cl_bool), &param);
+        CHIPERR_CHECK_LOG_AND_THROW_TABLE(clSetKernelExecInfo);
+      }
+      if (Ctx.MemManager_.isDeviceAllocUsed()) {
+        clStatus = clSetKernelExecInfo(
+            KernelAPIHandle, CL_KERNEL_EXEC_INFO_INDIRECT_DEVICE_ACCESS_INTEL,
+            sizeof(cl_bool), &param);
+        CHIPERR_CHECK_LOG_AND_THROW_TABLE(clSetKernelExecInfo);
+      }
+      if (Ctx.MemManager_.isSharedAllocUsed()) {
+        clStatus = clSetKernelExecInfo(
+            KernelAPIHandle, CL_KERNEL_EXEC_INFO_INDIRECT_SHARED_ACCESS_INTEL,
+            sizeof(cl_bool), &param);
+        CHIPERR_CHECK_LOG_AND_THROW_TABLE(clSetKernelExecInfo);
+      }
+    }
   }
 
   return AllocKeepAlives;
@@ -1259,13 +1281,8 @@ CHIPQueueOpenCL::launchImpl(chipstar::ExecItem *ExecItem) {
   LOCK(Backend->DubiousLockOpenCL);
 #endif
 
-  std::unique_ptr<std::vector<std::shared_ptr<void>>> AllocationsToKeepAlive;
-  // Disable annotation for CPU device. Runtime errror.
-  // https://community.intel.com/t5/OpenCL-for-CPU/OpenCL-Using-USM-on-Intel-CPU-Runtime-along-with/m-p/1621136#M7350
-  if (ChipEnvVars.getDevice().getType() != DeviceType::CPU) {
-    AllocationsToKeepAlive = annotateIndirectPointers(
-        *OclContext, Kernel->getModule()->getInfo(), KernelHandle);
-  }
+  auto AllocationsToKeepAlive = annotateIndirectPointers(
+      *OclContext, Kernel->getModule()->getInfo(), KernelHandle);
 
   auto [EventsToWait, EventLocks] = getSyncQueuesLastEvents(LaunchEvent, false);
   std::vector<cl_event> SyncQueuesEventHandles = getOpenCLHandles(EventsToWait);
@@ -1891,7 +1908,8 @@ void CHIPBackendOpenCL::initializeImpl() {
   std::vector<cl::Device> SupportedDevices;
   std::vector<cl::Device> Dev;
   clStatus = SelectedPlatform.getDevices(SelectedDevType, &Dev);
-  CHIPERR_CHECK_LOG_AND_THROW_TABLE(clGetPlatformIDs);
+  CHIPERR_CHECK_LOG_AND_THROW_TABLE(
+      clGetDeviceIDs); // C equivalent of getDevices
 
   for (auto D : Dev) {
 

diff --git a/src/backend/OpenCL/CHIPBackendOpenCL.hh b/src/backend/OpenCL/CHIPBackendOpenCL.hh
@@ -188,6 +188,11 @@ private:
   /// The allocation strategy to use.
   AllocationStrategy AllocStrategy_;
 
+  // Add these three bools
+  bool hostAllocUsed;
+  bool deviceAllocUsed;
+  bool sharedAllocUsed;
+
 public:
   void init(CHIPContextOpenCL *ChipCtxCl);
   MemoryManager &operator=(MemoryManager &&Rhs);
@@ -212,6 +217,11 @@ public:
     return AllocStrategy_;
   }
 
+  // Add getter methods for the new bools
+  bool isHostAllocUsed() const { return hostAllocUsed; }
+  bool isDeviceAllocUsed() const { return deviceAllocUsed; }
+  bool isSharedAllocUsed() const { return sharedAllocUsed; }
+
   std::pair<cl_mem, size_t> translateDevPtrToBuffer(const void *DevPtr) const;
 
 private:

diff --git a/src/backend/OpenCL/MemoryManager.cc b/src/backend/OpenCL/MemoryManager.cc
@@ -106,6 +106,11 @@ void MemoryManager::init(CHIPContextOpenCL *ChipCtxCl_) {
   } else {
     std::memset(&USM_, 0, sizeof(USM_));
   }
+
+  // Initialize the allocation usage bools
+  hostAllocUsed = false;
+  deviceAllocUsed = false;
+  sharedAllocUsed = false;
 }
 
 MemoryManager &MemoryManager::operator=(MemoryManager &&Rhs) {
@@ -241,6 +246,25 @@ void *MemoryManager::allocate(size_t Size, size_t Alignment,
   logTrace("Memory allocated: {} / {}\n", Ptr.get(), Size);
   assert(Allocations_.find(Ptr) == Allocations_.end());
   Allocations_.emplace(Ptr, Size);
+
+  // Set the appropriate bool based on the MemType
+  switch (MemType) {
+  case hipMemoryTypeHost:
+    hostAllocUsed = true;
+    break;
+  case hipMemoryTypeDevice:
+    deviceAllocUsed = true;
+    break;
+  case hipMemoryTypeManaged:
+  case hipMemoryTypeUnified:
+    sharedAllocUsed = true;
+    break;
+  default:
+    // Handle unexpected memory type
+    assert(!"Unexpected memory type!");
+    break;
+  }
+
   return Ptr.get();
 }
 
@@ -320,4 +344,4 @@ MemoryManager::translateDevPtrToBuffer(const void *DevPtr) const {
   }
 
   return {DevPtrToBuffer_.at(BasePtr), Offset};
-}
+}
diff --git a/src/backend/OpenCL/clHipErrorConversion.hh b/src/backend/OpenCL/clHipErrorConversion.hh
@@ -94,7 +94,8 @@ const std::unordered_map<void *, cl_hip_error_map_t> CL_HIP_ERROR_MAPS = {
       {CL_OUT_OF_RESOURCES, hipErrorOutOfMemory},
       {CL_MEM_OBJECT_ALLOCATION_FAILURE, hipErrorOutOfMemory},
       {CL_INVALID_EVENT_WAIT_LIST, hipErrorInvalidResourceHandle},
-      {CL_OUT_OF_HOST_MEMORY, hipErrorOutOfMemory}}},
+      {CL_OUT_OF_HOST_MEMORY, hipErrorOutOfMemory},
+      {CL_INVALID_OPERATION, hipErrorInvalidValue}}},
 
     {(void *)&clEnqueueReadBuffer,
      {{CL_SUCCESS, hipSuccess},
@@ -447,11 +448,6 @@ const std::unordered_map<void *, cl_hip_error_map_t> CL_HIP_ERROR_MAPS = {
       {CL_INVALID_EVENT_WAIT_LIST, hipErrorInvalidResourceHandle},
       {CL_OUT_OF_HOST_MEMORY, hipErrorOutOfMemory}}},
 
-    {(void *)&clGetPlatformIDs,
-     {{CL_SUCCESS, hipSuccess},
-      {CL_INVALID_VALUE, hipErrorInvalidValue},
-      {CL_OUT_OF_HOST_MEMORY, hipErrorOutOfMemory}}},
-
     {(void *)&clCreateProgramWithIL,
      {{CL_SUCCESS, hipSuccess},
       {CL_INVALID_CONTEXT, hipErrorInvalidResourceHandle},