SWDEV-301947 - Use new enum for CU count

Use HSA_AMD_AGENT_INFO_COOPERATIVE_COMPUTE_UNIT_COUNT to get compute
units. This is needed to work around assymentric CU harvesting bug on
gfx90a. Add a new device property to get the max available CUs on the
device.

Change-Id: I878f38f14f16c1af01fc0a77157aea1e816a63b8


[ROCm/clr commit: 33aca5a4a6]

projects/clr/rocclr/device/device.hpp

@@ -234,9 +234,12 @@ struct Info : public amd::EmbeddedObject {
   //! A unique device vendor identifier.
   uint32_t vendorId_;
 
-  //! The number of parallel compute cores on the compute device.
+  //! The available number of parallel compute cores on the compute device.
   uint32_t maxComputeUnits_;
 
+  //! The max number of parallel compute cores on the compute device.
+  uint32_t maxBoostComputeUnits_;
+
   //! Maximum dimensions that specify the global and local work-item IDs
   //  used by the data-parallel execution model.
   uint32_t maxWorkItemDimensions_;

projects/clr/rocclr/device/gpu/gpudevice.cpp

@@ -364,6 +364,7 @@ void NullDevice::fillDeviceInfo(const CALdeviceattribs& calAttr, const gslMemInf
   info_.type_ = CL_DEVICE_TYPE_GPU;
   info_.vendorId_ = 0x1002;
   info_.maxComputeUnits_ = calAttr.numberOfSIMD;
+  info_.maxBoostComputeUnits_ = calAttr.numberOfSIMD;
   info_.maxWorkItemDimensions_ = 3;
   info_.numberOfShaderEngines = calAttr.numberOfShaderEngines;
 

projects/clr/rocclr/device/pal/paldevice.cpp

@@ -320,7 +320,7 @@ void NullDevice::fillDeviceInfo(const Pal::DeviceProperties& palProp,
   info_.maxComputeUnits_ = settings().enableWgpMode_
       ? palProp.gfxipProperties.shaderCore.numAvailableCus / 2
       : palProp.gfxipProperties.shaderCore.numAvailableCus;
-
+  info_.maxBoostComputeUnits_ = info_.maxComputeUnits_;
   info_.numberOfShaderEngines = palProp.gfxipProperties.shaderCore.numShaderEngines;
 
   // SI parts are scalar.  Also, reads don't need to be 128-bits to get peak rates.

projects/clr/rocclr/device/rocm/rocdevice.cpp

@@ -1076,7 +1076,8 @@ bool Device::populateOCLDeviceConstants() {
   }
 
   if (HSA_STATUS_SUCCESS !=
-      hsa_agent_get_info(_bkendDevice, (hsa_agent_info_t)HSA_AMD_AGENT_INFO_COMPUTE_UNIT_COUNT,
+      hsa_agent_get_info(_bkendDevice,
+                         (hsa_agent_info_t)HSA_AMD_AGENT_INFO_COOPERATIVE_COMPUTE_UNIT_COUNT,
                          &info_.maxComputeUnits_)) {
     return false;
   }
@@ -1086,6 +1087,17 @@ bool Device::populateOCLDeviceConstants() {
       ? info_.maxComputeUnits_ / 2
       : info_.maxComputeUnits_;
 
+  if (HSA_STATUS_SUCCESS !=
+      hsa_agent_get_info(_bkendDevice, (hsa_agent_info_t)HSA_AMD_AGENT_INFO_COMPUTE_UNIT_COUNT,
+                         &info_.maxBoostComputeUnits_)) {
+    return false;
+  }
+  assert(info_.maxBoostComputeUnits_ > 0);
+
+  info_.maxBoostComputeUnits_ = settings().enableWgpMode_
+      ? info_.maxBoostComputeUnits_ / 2
+      : info_.maxBoostComputeUnits_;
+
   if (HSA_STATUS_SUCCESS != hsa_agent_get_info(_bkendDevice,
                                                (hsa_agent_info_t)HSA_AMD_AGENT_INFO_CACHELINE_SIZE,
                                                &info_.globalMemCacheLineSize_)) {
@@ -2106,7 +2118,7 @@ bool Device::IpcAttach(const void* handle, size_t mem_size, size_t mem_offset,
   void* orig_dev_ptr = nullptr;
 
   // Retrieve the devPtr from the handle
-  hsa_status_t hsa_status = 
+  hsa_status_t hsa_status =
       hsa_amd_ipc_memory_attach(reinterpret_cast<const hsa_amd_ipc_memory_t*>(handle),
                                 mem_size, (1 + p2p_agents_.size()), p2p_agents_list_,
                                 &orig_dev_ptr);