[Offload] Add OL_DEVICE_INFO_MAX_WORK_SIZE[_PER_DIMENSION] (#155823)

This is the total number of work items that the device supports (the
equivalent work group properties are for only a single work group).

Author: RossBrunton

offload/liboffload/API/Device.td

@@ -33,6 +33,8 @@ def ol_device_info_t : Enum {
     TaggedEtor<"DRIVER_VERSION", "char[]", "Driver version">,
     TaggedEtor<"MAX_WORK_GROUP_SIZE", "uint32_t", "Maximum total work group size in work items">,
     TaggedEtor<"MAX_WORK_GROUP_SIZE_PER_DIMENSION", "ol_dimensions_t", "Maximum work group size in each dimension">,
+    TaggedEtor<"MAX_WORK_SIZE", "uint32_t", "Maximum total work items">,
+    TaggedEtor<"MAX_WORK_SIZE_PER_DIMENSION", "ol_dimensions_t", "Maximum work items in each dimension">,
     TaggedEtor<"VENDOR_ID", "uint32_t", "A unique vendor device identifier assigned by PCI-SIG">,
     TaggedEtor<"NUM_COMPUTE_UNITS", "uint32_t", "The number of parallel compute units available to the device">,
     TaggedEtor<"MAX_CLOCK_FREQUENCY", "uint32_t", "The maximum configured clock frequency of this device in MHz">,

offload/liboffload/src/OffloadImpl.cpp

@@ -456,6 +456,7 @@ Error olGetDeviceInfoImplDetail(ol_device_handle_t Device,
   }
 
   case OL_DEVICE_INFO_MAX_WORK_GROUP_SIZE:
+  case OL_DEVICE_INFO_MAX_WORK_SIZE:
   case OL_DEVICE_INFO_VENDOR_ID:
   case OL_DEVICE_INFO_NUM_COMPUTE_UNITS:
   case OL_DEVICE_INFO_ADDRESS_BITS:
@@ -472,6 +473,7 @@ Error olGetDeviceInfoImplDetail(ol_device_handle_t Device,
     return Info.write(static_cast<uint32_t>(Value));
   }
 
+  case OL_DEVICE_INFO_MAX_WORK_SIZE_PER_DIMENSION:
   case OL_DEVICE_INFO_MAX_WORK_GROUP_SIZE_PER_DIMENSION: {
     // {x, y, z} triples
     ol_dimensions_t Out{0, 0, 0};
@@ -510,6 +512,8 @@ Error olGetDeviceInfoImplDetailHost(ol_device_handle_t Device,
   assert(Device == OffloadContext::get().HostDevice());
   InfoWriter Info(PropSize, PropValue, PropSizeRet);
 
+  constexpr auto uint32_max = std::numeric_limits<uint32_t>::max();
+
   switch (PropName) {
   case OL_DEVICE_INFO_PLATFORM:
     return Info.write<void *>(Device->Platform);
@@ -527,6 +531,11 @@ Error olGetDeviceInfoImplDetailHost(ol_device_handle_t Device,
     return Info.write<uint32_t>(1);
   case OL_DEVICE_INFO_MAX_WORK_GROUP_SIZE_PER_DIMENSION:
     return Info.write<ol_dimensions_t>(ol_dimensions_t{1, 1, 1});
+  case OL_DEVICE_INFO_MAX_WORK_SIZE:
+    return Info.write<uint32_t>(uint32_max);
+  case OL_DEVICE_INFO_MAX_WORK_SIZE_PER_DIMENSION:
+    return Info.write<ol_dimensions_t>(
+        ol_dimensions_t{uint32_max, uint32_max, uint32_max});
   case OL_DEVICE_INFO_VENDOR_ID:
     return Info.write<uint32_t>(0);
   case OL_DEVICE_INFO_NUM_COMPUTE_UNITS:

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

@@ -2924,11 +2924,12 @@ struct AMDGPUDeviceTy : public GenericDeviceTy, AMDGenericDeviceTy {
 
     Status = getDeviceAttrRaw(HSA_AGENT_INFO_GRID_MAX_SIZE, TmpUInt);
     if (Status == HSA_STATUS_SUCCESS)
-      Info.add("Grid Max Size", TmpUInt);
+      Info.add("Grid Max Size", TmpUInt, "", DeviceInfo::MAX_WORK_SIZE);
 
     Status = getDeviceAttrRaw(HSA_AGENT_INFO_GRID_MAX_DIM, GridMaxDim);
     if (Status == HSA_STATUS_SUCCESS) {
-      auto &MaxDim = *Info.add("Grid Max Size per Dimension");
+      auto &MaxDim = *Info.add("Grid Max Size per Dimension", std::monostate{},
+                               "", DeviceInfo::MAX_WORK_SIZE_PER_DIMENSION);
       MaxDim.add("x", GridMaxDim.x);
       MaxDim.add("y", GridMaxDim.y);
       MaxDim.add("z", GridMaxDim.z);

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

@@ -1120,7 +1120,13 @@ struct CUDADeviceTy : public GenericDeviceTy {
     if (Res == CUDA_SUCCESS)
       MaxBlock.add("z", TmpInt);
 
-    auto &MaxGrid = *Info.add("Maximum Grid Dimensions", "");
+    // TODO: I assume CUDA devices have no limit on the amount of threads,
+    // verify this
+    Info.add("Maximum Grid Size", std::numeric_limits<uint32_t>::max(), "",
+             DeviceInfo::MAX_WORK_SIZE);
+
+    auto &MaxGrid = *Info.add("Maximum Grid Dimensions", std::monostate{}, "",
+                              DeviceInfo::MAX_WORK_SIZE_PER_DIMENSION);
     Res = getDeviceAttrRaw(CU_DEVICE_ATTRIBUTE_MAX_GRID_DIM_X, TmpInt);
     if (Res == CUDA_SUCCESS)
       MaxGrid.add("x", TmpInt);

offload/unittests/OffloadAPI/device/olGetDeviceInfo.cpp

@@ -145,6 +145,19 @@ TEST_P(olGetDeviceInfoTest, SuccessMaxWorkGroupSizePerDimension) {
   ASSERT_GT(Value.z, 0u);
 }
 
+OL_DEVICE_INFO_TEST_VALUE_GT(MaxWorkSize, uint32_t,
+                             OL_DEVICE_INFO_MAX_WORK_SIZE, 0);
+
+TEST_P(olGetDeviceInfoTest, SuccessMaxWorkSizePerDimension) {
+  ol_dimensions_t Value{0, 0, 0};
+  ASSERT_SUCCESS(olGetDeviceInfo(Device,
+                                 OL_DEVICE_INFO_MAX_WORK_SIZE_PER_DIMENSION,
+                                 sizeof(Value), &Value));
+  ASSERT_GT(Value.x, 0u);
+  ASSERT_GT(Value.y, 0u);
+  ASSERT_GT(Value.z, 0u);
+}
+
 OL_DEVICE_INFO_TEST_DEVICE_VALUE_GT(VendorId, uint32_t,
                                     OL_DEVICE_INFO_VENDOR_ID, 0);
 OL_DEVICE_INFO_TEST_HOST_SUCCESS(VendorId, uint32_t, OL_DEVICE_INFO_VENDOR_ID);

offload/unittests/OffloadAPI/device/olGetDeviceInfoSize.cpp

@@ -36,6 +36,8 @@ OL_DEVICE_INFO_SIZE_TEST_NONZERO(Vendor, OL_DEVICE_INFO_VENDOR);
 OL_DEVICE_INFO_SIZE_TEST_NONZERO(DriverVersion, OL_DEVICE_INFO_DRIVER_VERSION);
 OL_DEVICE_INFO_SIZE_TEST_EQ(MaxWorkGroupSize, uint32_t,
                             OL_DEVICE_INFO_MAX_WORK_GROUP_SIZE);
+OL_DEVICE_INFO_SIZE_TEST_EQ(MaxWorkSize, uint32_t,
+                            OL_DEVICE_INFO_MAX_WORK_SIZE);
 OL_DEVICE_INFO_SIZE_TEST_EQ(VendorId, uint32_t, OL_DEVICE_INFO_VENDOR_ID);
 OL_DEVICE_INFO_SIZE_TEST_EQ(NumComputeUnits, uint32_t,
                             OL_DEVICE_INFO_NUM_COMPUTE_UNITS);
@@ -77,6 +79,14 @@ TEST_P(olGetDeviceInfoSizeTest, SuccessMaxWorkGroupSizePerDimension) {
   ASSERT_EQ(Size, sizeof(uint32_t) * 3);
 }
 
+TEST_P(olGetDeviceInfoSizeTest, SuccessMaxWorkSizePerDimension) {
+  size_t Size = 0;
+  ASSERT_SUCCESS(olGetDeviceInfoSize(
+      Device, OL_DEVICE_INFO_MAX_WORK_SIZE_PER_DIMENSION, &Size));
+  ASSERT_EQ(Size, sizeof(ol_dimensions_t));
+  ASSERT_EQ(Size, sizeof(uint32_t) * 3);
+}
+
 TEST_P(olGetDeviceInfoSizeTest, InvalidNullHandle) {
   size_t Size = 0;
   ASSERT_ERROR(OL_ERRC_INVALID_NULL_HANDLE,