merge main into amd-staging

Author: ronlieb

offload/liboffload/API/Device.td

@@ -43,6 +43,7 @@ def ol_device_info_t : Enum {
     TaggedEtor<"ADDRESS_BITS", "uint32_t", "Number of bits used to represent an address in device memory">,
     TaggedEtor<"MAX_MEM_ALLOC_SIZE", "uint64_t", "The maximum size of memory object allocation in bytes">,
     TaggedEtor<"GLOBAL_MEM_SIZE", "uint64_t", "The size of global device memory in bytes">,
+    TaggedEtor<"WORK_GROUP_LOCAL_MEM_SIZE", "uint64_t", "The maximum size of local shared memory per work group in bytes">,
   ];
   list<TaggedEtor> fp_configs = !foreach(type, ["Single", "Double", "Half"], TaggedEtor<type # "_FP_CONFIG", "ol_device_fp_capability_flags_t", type # " precision floating point capability">);
   list<TaggedEtor> native_vec_widths = !foreach(type, ["char","short","int","long","float","double","half"], TaggedEtor<"NATIVE_VECTOR_WIDTH_" # type, "uint32_t", "Native vector width for " # type>);

offload/liboffload/src/OffloadImpl.cpp

@@ -495,6 +495,13 @@ Error olGetDeviceInfoImplDetail(ol_device_handle_t Device,
     return Info.write(static_cast<uint32_t>(Value));
   }
 
+  case OL_DEVICE_INFO_WORK_GROUP_LOCAL_MEM_SIZE: {
+    if (!std::holds_alternative<uint64_t>(Entry->Value))
+      return makeError(ErrorCode::BACKEND_FAILURE,
+                       "plugin returned incorrect type");
+    return Info.write(std::get<uint64_t>(Entry->Value));
+  }
+
   case OL_DEVICE_INFO_MAX_WORK_SIZE_PER_DIMENSION:
   case OL_DEVICE_INFO_MAX_WORK_GROUP_SIZE_PER_DIMENSION: {
     // {x, y, z} triples
@@ -590,6 +597,7 @@ Error olGetDeviceInfoImplDetailHost(ol_device_handle_t Device,
     return Info.write<uint32_t>(std::numeric_limits<uintptr_t>::digits);
   case OL_DEVICE_INFO_MAX_MEM_ALLOC_SIZE:
   case OL_DEVICE_INFO_GLOBAL_MEM_SIZE:
+  case OL_DEVICE_INFO_WORK_GROUP_LOCAL_MEM_SIZE:
     return Info.write<uint64_t>(0);
   default:
     return createOffloadError(ErrorCode::INVALID_ENUMERATION,

offload/plugins-nextgen/amdgpu/src/rtl.cpp

@@ -3452,6 +3452,15 @@ struct AMDGPUDeviceTy : public GenericDeviceTy, AMDGenericDeviceTy {
     // for map and zero-copy control
     // TODO: put them back in constructor
     //    readEnvVars();
+    // Retrieve the size of the group memory.
+    for (const auto *Pool : AllMemoryPools) {
+      if (Pool->isGroup()) {
+        if (auto Err = Pool->getAttr(HSA_AMD_MEMORY_POOL_INFO_SIZE,
+                                     MaxBlockSharedMemSize))
+          return Err;
+        break;
+      }
+    }
 
     return Plugin::success();
   }
@@ -4327,6 +4336,9 @@ struct AMDGPUDeviceTy : public GenericDeviceTy, AMDGenericDeviceTy {
     if (Status == HSA_STATUS_SUCCESS)
       Info.add("Cacheline Size", TmpUInt);
 
+    Info.add("Max Shared Memory per Work Group", MaxBlockSharedMemSize, "bytes",
+             DeviceInfo::WORK_GROUP_LOCAL_MEM_SIZE);
+
     Status = getDeviceAttrRaw(HSA_AMD_AGENT_INFO_MAX_CLOCK_FREQUENCY, TmpUInt);
     if (Status == HSA_STATUS_SUCCESS)
       Info.add("Max Clock Freq", TmpUInt, "MHz",

offload/plugins-nextgen/common/include/PluginInterface.h

@@ -869,6 +869,10 @@ struct GenericDeviceTy : public DeviceAllocatorTy {
   /// Get the unique identifier of the device.
   const char *getDeviceUid() const { return DeviceUid.c_str(); }
 
+  /// Get the total shared memory per block (in bytes) that can be used in any
+  /// kernel.
+  size_t getMaxBlockSharedMemSize() const { return MaxBlockSharedMemSize; }
+
   /// Set the context of the device if needed, before calling device-specific
   /// functions. Plugins may implement this function as a no-op if not needed.
   virtual Error setContext() = 0;
@@ -1579,6 +1583,9 @@ struct KernelRunRecordTy {
   std::unordered_map<std::string, TuningMetadataTy> TuningData;
   /// Internal representation for OMPT device (initialize & finalize)
   std::atomic<bool> OmptInitialized;
+
+  /// The total per-block native shared memory that a kernel may use.
+  size_t MaxBlockSharedMemSize = 0;
 };
 
 /// Class implementing common functionalities of offload plugins. Each plugin

offload/plugins-nextgen/cuda/src/rtl.cpp

@@ -382,6 +382,12 @@ struct CUDADeviceTy : public GenericDeviceTy {
       return Err;
     HardwareParallelism = NumMuliprocessors * (MaxThreadsPerSM / WarpSize);
 
+    uint32_t MaxSharedMem;
+    if (auto Err = getDeviceAttr(
+            CU_DEVICE_ATTRIBUTE_MAX_SHARED_MEMORY_PER_BLOCK, MaxSharedMem))
+      return Err;
+    MaxBlockSharedMemSize = MaxSharedMem;
+
     return Plugin::success();
   }
 
@@ -1092,10 +1098,8 @@ struct CUDADeviceTy : public GenericDeviceTy {
     if (Res == CUDA_SUCCESS)
       Info.add("Total Constant Memory", TmpInt, "bytes");
 
-    Res = getDeviceAttrRaw(CU_DEVICE_ATTRIBUTE_MAX_SHARED_MEMORY_PER_BLOCK,
-                           TmpInt);
-    if (Res == CUDA_SUCCESS)
-      Info.add("Max Shared Memory per Block", TmpInt, "bytes");
+    Info.add("Max Shared Memory per Block", MaxBlockSharedMemSize, "bytes",
+             DeviceInfo::WORK_GROUP_LOCAL_MEM_SIZE);
 
     Res = getDeviceAttrRaw(CU_DEVICE_ATTRIBUTE_MAX_REGISTERS_PER_BLOCK, TmpInt);
     if (Res == CUDA_SUCCESS)

offload/tools/deviceinfo/llvm-offload-device-info.cpp

@@ -205,6 +205,9 @@ ol_result_t printDevice(std::ostream &S, ol_device_handle_t D) {
       S, D, OL_DEVICE_INFO_MAX_MEM_ALLOC_SIZE, "Max Mem Allocation Size", "B"));
   OFFLOAD_ERR(printDeviceValue<uint64_t>(S, D, OL_DEVICE_INFO_GLOBAL_MEM_SIZE,
                                          "Global Mem Size", "B"));
+  OFFLOAD_ERR(
+      printDeviceValue<uint64_t>(S, D, OL_DEVICE_INFO_WORK_GROUP_LOCAL_MEM_SIZE,
+                                 "Work Group Shared Mem Size", "B"));
   OFFLOAD_ERR(
       (printDeviceValue<ol_device_fp_capability_flags_t, PrintKind::FP_FLAGS>(
           S, D, OL_DEVICE_INFO_SINGLE_FP_CONFIG,

offload/unittests/OffloadAPI/device/olGetDeviceInfo.cpp

@@ -217,6 +217,11 @@ OL_DEVICE_INFO_TEST_DEVICE_VALUE_GT(GlobalMemSize, uint64_t,
                                     OL_DEVICE_INFO_GLOBAL_MEM_SIZE, 0);
 OL_DEVICE_INFO_TEST_HOST_SUCCESS(GlobalMemSize, uint64_t,
                                  OL_DEVICE_INFO_GLOBAL_MEM_SIZE);
+OL_DEVICE_INFO_TEST_DEVICE_VALUE_GT(SharedMemSize, uint64_t,
+                                    OL_DEVICE_INFO_WORK_GROUP_LOCAL_MEM_SIZE,
+                                    0);
+OL_DEVICE_INFO_TEST_HOST_SUCCESS(SharedMemSize, uint64_t,
+                                 OL_DEVICE_INFO_WORK_GROUP_LOCAL_MEM_SIZE);
 
 TEST_P(olGetDeviceInfoTest, InvalidNullHandleDevice) {
   ol_device_type_t DeviceType;

offload/unittests/OffloadAPI/device/olGetDeviceInfoSize.cpp

@@ -71,6 +71,8 @@ OL_DEVICE_INFO_SIZE_TEST_EQ(MaxMemAllocSize, uint64_t,
                             OL_DEVICE_INFO_MAX_MEM_ALLOC_SIZE);
 OL_DEVICE_INFO_SIZE_TEST_EQ(GlobalMemSize, uint64_t,
                             OL_DEVICE_INFO_GLOBAL_MEM_SIZE);
+OL_DEVICE_INFO_SIZE_TEST_EQ(SharedMemSize, uint64_t,
+                            OL_DEVICE_INFO_WORK_GROUP_LOCAL_MEM_SIZE);
 
 TEST_P(olGetDeviceInfoSizeTest, SuccessMaxWorkGroupSizePerDimension) {
   size_t Size = 0;