[Offload] Define additional device info properties (llvm#152533)

Add the following properties in Offload device info:
* VENDOR_ID
* NUM_COMPUTE_UNITS
* [SINGLE|DOUBLE|HALF]_FP_CONFIG
* NATIVE_VECTOR_WIDTH_[CHAR|SHORT|INT|LONG|FLOAT|DOUBLE|HALF]
* MAX_CLOCK_FREQUENCY
* MEMORY_CLOCK_RATE
* ADDRESS_BITS
* MAX_MEM_ALLOC_SIZE
* GLOBAL_MEM_SIZE

Add a bitfield option to enumerators, allowing the values to be
bit-shifted instead of incremented. Generate the per-type enums using
`foreach` to reduce code duplication.

Use macros in unit test definitions to reduce code duplication.

Author: rafbiels

offload/liboffload/API/APIDefs.td

@@ -176,6 +176,10 @@ class Enum : APIObject {
   // all Etor values must be TaggedEtor records
   bit is_typed = 0;
 
+  // This refers to whether the enumerator is used to name bits of a bit field,
+  // where consecutive values are bit-shifted rather than incremented.
+  bit is_bit_field = 0;
+
   list<Etor> etors = [];
 }
 

offload/liboffload/API/Device.td

@@ -26,17 +26,49 @@ def DeviceInfo : Enum {
   let name = "ol_device_info_t";
   let desc = "Supported device info.";
   let is_typed = 1;
-  let etors =[
+  list<TaggedEtor> basic_etors =[
     TaggedEtor<"TYPE", "ol_device_type_t", "type of the device">,
     TaggedEtor<"PLATFORM", "ol_platform_handle_t", "the platform associated with the device">,
     TaggedEtor<"NAME", "char[]", "Device name">,
     TaggedEtor<"VENDOR", "char[]", "Device vendor">,
     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<"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">,
+    TaggedEtor<"MEMORY_CLOCK_RATE", "uint32_t", "Memory clock frequency in MHz">,
+    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">,
+  ];
+  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>);
+  let etors = !listconcat(basic_etors, fp_configs, native_vec_widths);
+}
+
+def : Enum {
+  let name = "ol_device_fp_capability_flag_t";
+  let desc = "Device floating-point capability flags";
+  let is_bit_field = 1;
+  let etors =[
+    Etor<"CORRECTLY_ROUNDED_DIVIDE_SQRT", "Support correctly rounded divide and sqrt">,
+    Etor<"ROUND_TO_NEAREST", "Support round to nearest">,
+    Etor<"ROUND_TO_ZERO", "Support round to zero">,
+    Etor<"ROUND_TO_INF", "Support round to infinity">,
+    Etor<"INF_NAN", "Support INF to NAN">,
+    Etor<"DENORM", "Support denorm">,
+    Etor<"FMA", "Support fused multiply-add">,
+    Etor<"SOFT_FLOAT", "Basic floating point operations implemented in software">,
   ];
 }
 
+def : Typedef {
+  let name = "ol_device_fp_capability_flags_t";
+  let desc = "Device floating-point capability flags";
+  let value = "uint32_t";
+}
+
 def : FptrTypedef {
   let name = "ol_device_iterate_cb_t";
   let desc = "User-provided function to be used with `olIterateDevices`";

offload/liboffload/src/OffloadImpl.cpp

@@ -302,10 +302,50 @@ Error olGetDeviceInfoImplDetail(ol_device_handle_t Device,
   };
 
   // These are not implemented by the plugin interface
-  if (PropName == OL_DEVICE_INFO_PLATFORM)
+  switch (PropName) {
+  case OL_DEVICE_INFO_PLATFORM:
     return Info.write<void *>(Device->Platform);
-  if (PropName == OL_DEVICE_INFO_TYPE)
+
+  case OL_DEVICE_INFO_TYPE:
     return Info.write<ol_device_type_t>(OL_DEVICE_TYPE_GPU);
+
+  case OL_DEVICE_INFO_SINGLE_FP_CONFIG:
+  case OL_DEVICE_INFO_DOUBLE_FP_CONFIG: {
+    ol_device_fp_capability_flags_t flags{0};
+    flags |= OL_DEVICE_FP_CAPABILITY_FLAG_CORRECTLY_ROUNDED_DIVIDE_SQRT |
+             OL_DEVICE_FP_CAPABILITY_FLAG_ROUND_TO_NEAREST |
+             OL_DEVICE_FP_CAPABILITY_FLAG_ROUND_TO_ZERO |
+             OL_DEVICE_FP_CAPABILITY_FLAG_ROUND_TO_INF |
+             OL_DEVICE_FP_CAPABILITY_FLAG_INF_NAN |
+             OL_DEVICE_FP_CAPABILITY_FLAG_DENORM |
+             OL_DEVICE_FP_CAPABILITY_FLAG_FMA;
+    return Info.write(flags);
+  }
+
+  case OL_DEVICE_INFO_HALF_FP_CONFIG:
+    return Info.write<ol_device_fp_capability_flags_t>(0);
+
+  case OL_DEVICE_INFO_NATIVE_VECTOR_WIDTH_CHAR:
+  case OL_DEVICE_INFO_NATIVE_VECTOR_WIDTH_SHORT:
+  case OL_DEVICE_INFO_NATIVE_VECTOR_WIDTH_INT:
+  case OL_DEVICE_INFO_NATIVE_VECTOR_WIDTH_LONG:
+  case OL_DEVICE_INFO_NATIVE_VECTOR_WIDTH_FLOAT:
+  case OL_DEVICE_INFO_NATIVE_VECTOR_WIDTH_DOUBLE:
+    return Info.write<uint32_t>(1);
+
+  case OL_DEVICE_INFO_NATIVE_VECTOR_WIDTH_HALF:
+    return Info.write<uint32_t>(0);
+
+  // None of the existing plugins specify a limit on a single allocation,
+  // so return the global memory size instead
+  case OL_DEVICE_INFO_MAX_MEM_ALLOC_SIZE:
+    PropName = OL_DEVICE_INFO_GLOBAL_MEM_SIZE;
+    break;
+
+  default:
+    break;
+  }
+
   if (PropName >= OL_DEVICE_INFO_LAST)
     return createOffloadError(ErrorCode::INVALID_ENUMERATION,
                               "getDeviceInfo enum '%i' is invalid", PropName);
@@ -316,6 +356,7 @@ Error olGetDeviceInfoImplDetail(ol_device_handle_t Device,
                      "plugin did not provide a response for this information");
   auto Entry = *EntryOpt;
 
+  // Retrieve properties from the plugin interface
   switch (PropName) {
   case OL_DEVICE_INFO_NAME:
   case OL_DEVICE_INFO_VENDOR:
@@ -327,7 +368,20 @@ Error olGetDeviceInfoImplDetail(ol_device_handle_t Device,
     return Info.writeString(std::get<std::string>(Entry->Value).c_str());
   }
 
-  case OL_DEVICE_INFO_MAX_WORK_GROUP_SIZE: {
+  case OL_DEVICE_INFO_GLOBAL_MEM_SIZE: {
+    // Uint64 values
+    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_GROUP_SIZE:
+  case OL_DEVICE_INFO_VENDOR_ID:
+  case OL_DEVICE_INFO_NUM_COMPUTE_UNITS:
+  case OL_DEVICE_INFO_ADDRESS_BITS:
+  case OL_DEVICE_INFO_MAX_CLOCK_FREQUENCY:
+  case OL_DEVICE_INFO_MEMORY_CLOCK_RATE: {
     // Uint32 values
     if (!std::holds_alternative<uint64_t>(Entry->Value))
       return makeError(ErrorCode::BACKEND_FAILURE,
@@ -389,9 +443,40 @@ Error olGetDeviceInfoImplDetailHost(ol_device_handle_t Device,
   case OL_DEVICE_INFO_DRIVER_VERSION:
     return Info.writeString(LLVM_VERSION_STRING);
   case OL_DEVICE_INFO_MAX_WORK_GROUP_SIZE:
-    return Info.write<uint64_t>(1);
+    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_VENDOR_ID:
+    return Info.write<uint32_t>(0);
+  case OL_DEVICE_INFO_NUM_COMPUTE_UNITS:
+    return Info.write<uint32_t>(1);
+  case OL_DEVICE_INFO_SINGLE_FP_CONFIG:
+  case OL_DEVICE_INFO_DOUBLE_FP_CONFIG:
+    return Info.write<ol_device_fp_capability_flags_t>(
+        OL_DEVICE_FP_CAPABILITY_FLAG_CORRECTLY_ROUNDED_DIVIDE_SQRT |
+        OL_DEVICE_FP_CAPABILITY_FLAG_ROUND_TO_NEAREST |
+        OL_DEVICE_FP_CAPABILITY_FLAG_ROUND_TO_ZERO |
+        OL_DEVICE_FP_CAPABILITY_FLAG_ROUND_TO_INF |
+        OL_DEVICE_FP_CAPABILITY_FLAG_INF_NAN |
+        OL_DEVICE_FP_CAPABILITY_FLAG_DENORM | OL_DEVICE_FP_CAPABILITY_FLAG_FMA);
+  case OL_DEVICE_INFO_HALF_FP_CONFIG:
+    return Info.write<ol_device_fp_capability_flags_t>(0);
+  case OL_DEVICE_INFO_NATIVE_VECTOR_WIDTH_CHAR:
+  case OL_DEVICE_INFO_NATIVE_VECTOR_WIDTH_SHORT:
+  case OL_DEVICE_INFO_NATIVE_VECTOR_WIDTH_INT:
+  case OL_DEVICE_INFO_NATIVE_VECTOR_WIDTH_LONG:
+  case OL_DEVICE_INFO_NATIVE_VECTOR_WIDTH_FLOAT:
+  case OL_DEVICE_INFO_NATIVE_VECTOR_WIDTH_DOUBLE:
+    return Info.write<uint32_t>(1);
+  case OL_DEVICE_INFO_NATIVE_VECTOR_WIDTH_HALF:
+    return Info.write<uint32_t>(0);
+  case OL_DEVICE_INFO_MAX_CLOCK_FREQUENCY:
+  case OL_DEVICE_INFO_MEMORY_CLOCK_RATE:
+  case OL_DEVICE_INFO_ADDRESS_BITS:
+    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:
+    return Info.write<uint64_t>(0);
   default:
     return createOffloadError(ErrorCode::INVALID_ENUMERATION,
                               "getDeviceInfo enum '%i' is invalid", PropName);

offload/plugins-nextgen/amdgpu/dynamic_hsa/hsa.h

@@ -70,10 +70,16 @@ typedef enum {
   HSA_ISA_INFO_NAME = 1
 } hsa_isa_info_t;
 
+typedef enum {
+  HSA_MACHINE_MODEL_SMALL = 0,
+  HSA_MACHINE_MODEL_LARGE = 1
+} hsa_machine_model_t;
+
 typedef enum {
   HSA_AGENT_INFO_NAME = 0,
   HSA_AGENT_INFO_VENDOR_NAME = 1,
   HSA_AGENT_INFO_FEATURE = 2,
+  HSA_AGENT_INFO_MACHINE_MODEL = 3,
   HSA_AGENT_INFO_PROFILE = 4,
   HSA_AGENT_INFO_WAVEFRONT_SIZE = 6,
   HSA_AGENT_INFO_WORKGROUP_MAX_DIM = 7,

offload/plugins-nextgen/amdgpu/dynamic_hsa/hsa_ext_amd.h

@@ -67,6 +67,7 @@ typedef enum hsa_amd_agent_info_s {
   HSA_AMD_AGENT_INFO_CACHELINE_SIZE = 0xA001,
   HSA_AMD_AGENT_INFO_COMPUTE_UNIT_COUNT = 0xA002,
   HSA_AMD_AGENT_INFO_MAX_CLOCK_FREQUENCY = 0xA003,
+  HSA_AMD_AGENT_INFO_MEMORY_MAX_FREQUENCY = 0xA008,
   HSA_AMD_AGENT_INFO_PRODUCT_NAME = 0xA009,
   HSA_AMD_AGENT_INFO_MAX_WAVES_PER_CU = 0xA00A,
   HSA_AMD_AGENT_INFO_NUM_SIMDS_PER_CU = 0xA00B,

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

@@ -2697,6 +2697,15 @@ struct AMDGPUDeviceTy : public GenericDeviceTy, AMDGenericDeviceTy {
     if (Status == HSA_STATUS_SUCCESS)
       Info.add("Vendor Name", TmpChar, "", DeviceInfo::VENDOR);
 
+    Info.add("Vendor ID", uint64_t{4130}, "", DeviceInfo::VENDOR_ID);
+
+    hsa_machine_model_t MachineModel;
+    Status = getDeviceAttrRaw(HSA_AGENT_INFO_MACHINE_MODEL, MachineModel);
+    if (Status == HSA_STATUS_SUCCESS)
+      Info.add("Memory Address Size",
+               uint64_t{MachineModel == HSA_MACHINE_MODEL_SMALL ? 32u : 64u},
+               "bits", DeviceInfo::ADDRESS_BITS);
+
     hsa_device_type_t DevType;
     Status = getDeviceAttrRaw(HSA_AGENT_INFO_DEVICE, DevType);
     if (Status == HSA_STATUS_SUCCESS) {
@@ -2747,11 +2756,17 @@ struct AMDGPUDeviceTy : public GenericDeviceTy, AMDGenericDeviceTy {
 
     Status = getDeviceAttrRaw(HSA_AMD_AGENT_INFO_MAX_CLOCK_FREQUENCY, TmpUInt);
     if (Status == HSA_STATUS_SUCCESS)
-      Info.add("Max Clock Freq", TmpUInt, "MHz");
+      Info.add("Max Clock Freq", TmpUInt, "MHz",
+               DeviceInfo::MAX_CLOCK_FREQUENCY);
+
+    Status = getDeviceAttrRaw(HSA_AMD_AGENT_INFO_MEMORY_MAX_FREQUENCY, TmpUInt);
+    if (Status == HSA_STATUS_SUCCESS)
+      Info.add("Max Memory Clock Freq", TmpUInt, "MHz",
+               DeviceInfo::MEMORY_CLOCK_RATE);
 
     Status = getDeviceAttrRaw(HSA_AMD_AGENT_INFO_COMPUTE_UNIT_COUNT, TmpUInt);
     if (Status == HSA_STATUS_SUCCESS)
-      Info.add("Compute Units", TmpUInt);
+      Info.add("Compute Units", TmpUInt, "", DeviceInfo::NUM_COMPUTE_UNITS);
 
     Status = getDeviceAttrRaw(HSA_AMD_AGENT_INFO_NUM_SIMDS_PER_CU, TmpUInt);
     if (Status == HSA_STATUS_SUCCESS)
@@ -2833,7 +2848,11 @@ struct AMDGPUDeviceTy : public GenericDeviceTy, AMDGenericDeviceTy {
 
       Status = Pool->getAttrRaw(HSA_AMD_MEMORY_POOL_INFO_SIZE, TmpSt);
       if (Status == HSA_STATUS_SUCCESS)
-        PoolNode.add("Size", TmpSt, "bytes");
+        PoolNode.add(
+            "Size", TmpSt, "bytes",
+            (Pool->isGlobal() && Pool->isCoarseGrained())
+                ? std::optional<DeviceInfo>{DeviceInfo::GLOBAL_MEM_SIZE}
+                : std::nullopt);
 
       Status = Pool->getAttrRaw(HSA_AMD_MEMORY_POOL_INFO_RUNTIME_ALLOC_ALLOWED,
                                 TmpBool);

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

@@ -979,13 +979,20 @@ struct CUDADeviceTy : public GenericDeviceTy {
 
     Info.add("Vendor Name", "NVIDIA", "", DeviceInfo::VENDOR);
 
+    Info.add("Vendor ID", uint64_t{4318}, "", DeviceInfo::VENDOR_ID);
+
+    Info.add("Memory Address Size", std::numeric_limits<CUdeviceptr>::digits,
+             "bits", DeviceInfo::ADDRESS_BITS);
+
     Res = cuDeviceTotalMem(&TmpSt, Device);
     if (Res == CUDA_SUCCESS)
-      Info.add("Global Memory Size", TmpSt, "bytes");
+      Info.add("Global Memory Size", TmpSt, "bytes",
+               DeviceInfo::GLOBAL_MEM_SIZE);
 
     Res = getDeviceAttrRaw(CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT, TmpInt);
     if (Res == CUDA_SUCCESS)
-      Info.add("Number of Multiprocessors", TmpInt);
+      Info.add("Number of Multiprocessors", TmpInt, "",
+               DeviceInfo::NUM_COMPUTE_UNITS);
 
     Res = getDeviceAttrRaw(CU_DEVICE_ATTRIBUTE_GPU_OVERLAP, TmpInt);
     if (Res == CUDA_SUCCESS)
@@ -1046,7 +1053,8 @@ struct CUDADeviceTy : public GenericDeviceTy {
 
     Res = getDeviceAttrRaw(CU_DEVICE_ATTRIBUTE_CLOCK_RATE, TmpInt);
     if (Res == CUDA_SUCCESS)
-      Info.add("Clock Rate", TmpInt, "kHz");
+      Info.add("Clock Rate", TmpInt / 1000, "MHz",
+               DeviceInfo::MAX_CLOCK_FREQUENCY);
 
     Res = getDeviceAttrRaw(CU_DEVICE_ATTRIBUTE_KERNEL_EXEC_TIMEOUT, TmpInt);
     if (Res == CUDA_SUCCESS)
@@ -1083,7 +1091,8 @@ struct CUDADeviceTy : public GenericDeviceTy {
 
     Res = getDeviceAttrRaw(CU_DEVICE_ATTRIBUTE_MEMORY_CLOCK_RATE, TmpInt);
     if (Res == CUDA_SUCCESS)
-      Info.add("Memory Clock Rate", TmpInt, "kHz");
+      Info.add("Memory Clock Rate", TmpInt / 1000, "MHz",
+               DeviceInfo::MEMORY_CLOCK_RATE);
 
     Res = getDeviceAttrRaw(CU_DEVICE_ATTRIBUTE_GLOBAL_MEMORY_BUS_WIDTH, TmpInt);
     if (Res == CUDA_SUCCESS)

offload/tools/offload-tblgen/APIGen.cpp

@@ -131,7 +131,8 @@ static void ProcessEnum(const EnumRec &Enum, raw_ostream &OS) {
   OS << formatv("/// @brief {0}\n", Enum.getDesc());
   OS << formatv("typedef enum {0} {{\n", Enum.getName());
 
-  uint32_t EtorVal = 0;
+  // Bitfields start from 1, other enums from 0
+  uint32_t EtorVal = Enum.isBitField();
   for (const auto &EnumVal : Enum.getValues()) {
     if (Enum.isTyped()) {
       OS << MakeComment(
@@ -141,7 +142,12 @@ static void ProcessEnum(const EnumRec &Enum, raw_ostream &OS) {
       OS << MakeComment(EnumVal.getDesc());
     }
     OS << formatv(TAB_1 "{0}_{1} = {2},\n", Enum.getEnumValNamePrefix(),
-                  EnumVal.getName(), EtorVal++);
+                  EnumVal.getName(), EtorVal);
+    if (Enum.isBitField()) {
+      EtorVal <<= 1u;
+    } else {
+      ++EtorVal;
+    }
   }
 
   // Add last_element/force uint32 val

offload/tools/offload-tblgen/MiscGen.cpp

@@ -107,10 +107,16 @@ void EmitOffloadInfo(const RecordKeeper &Records, raw_ostream &OS) {
 
 )";
 
-  auto ErrorCodeEnum = EnumRec{Records.getDef("DeviceInfo")};
-  uint32_t EtorVal = 0;
-  for (const auto &EnumVal : ErrorCodeEnum.getValues()) {
+  auto Enum = EnumRec{Records.getDef("DeviceInfo")};
+  // Bitfields start from 1, other enums from 0
+  uint32_t EtorVal = Enum.isBitField();
+  for (const auto &EnumVal : Enum.getValues()) {
     OS << formatv(TAB_1 "OFFLOAD_DEVINFO({0}, \"{1}\", {2})\n",
-                  EnumVal.getName(), EnumVal.getDesc(), EtorVal++);
+                  EnumVal.getName(), EnumVal.getDesc(), EtorVal);
+    if (Enum.isBitField()) {
+      EtorVal <<= 1u;
+    } else {
+      ++EtorVal;
+    }
   }
 }

offload/tools/offload-tblgen/RecordTypes.hpp

@@ -92,6 +92,8 @@ class EnumRec {
 
   bool isTyped() const { return rec->getValueAsBit("is_typed"); }
 
+  bool isBitField() const { return rec->getValueAsBit("is_bit_field"); }
+
 private:
   const Record *rec;
   std::vector<EnumValueRec> vals;