P4 to Git Change 1737265 by skudchad@skudchad_test2_win_opencl on 2019/01/29 16:13:19

SWDEV-145570 - [HIP] - Fix maxThreadsPerStreamProcessor info parameter for HIP-VDI

	ReviewBoardURL = http://ocltc.amd.com/reviews/r/16626/diff/

Affected files ...

... //depot/stg/opencl/drivers/opencl/api/hip/hip_device.cpp#17 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/device.hpp#331 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpudevice.cpp#601 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/paldevice.cpp#123 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocdevice.cpp#113 edit


[ROCm/clr commit: 13323b876e]

projects/clr/rocclr/runtime/device/device.hpp

@@ -500,6 +500,9 @@ struct Info : public amd::EmbeddedObject {
   uint32_t pcieDeviceId_;
   //! Revision ID
   uint32_t pcieRevisionId_;
+
+  //! Max numbers of threads per CU
+  cl_uint maxThreadsPerCU_;
 };
 
 //! Device settings

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

@@ -585,6 +585,7 @@ void NullDevice::fillDeviceInfo(const CALdeviceattribs& calAttr, const gslMemInf
 
     info_.pcieDeviceId_ = calAttr.pcieDeviceID;
     info_.pcieRevisionId_ = calAttr.pcieRevisionID;
+    info_.maxThreadsPerCU_ = info_.wavefrontWidth_ * hwInfo()->simdPerCU_ * 10;
   }
 }
 

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

@@ -356,7 +356,6 @@ void NullDevice::fillDeviceInfo(const Pal::DeviceProperties& palProp,
     palProp.gfxipProperties.shaderCore.numAvailableCus / 2 :
     palProp.gfxipProperties.shaderCore.numAvailableCus;
 
-
   info_.numberOfShaderEngines = palProp.gfxipProperties.shaderCore.numShaderEngines;
 
   // SI parts are scalar.  Also, reads don't need to be 128-bits to get peak rates.
@@ -636,6 +635,8 @@ void NullDevice::fillDeviceInfo(const Pal::DeviceProperties& palProp,
 
     info_.pcieDeviceId_ = palProp.deviceId;
     info_.pcieRevisionId_ = palProp.revisionId;
+    info_.maxThreadsPerCU_ = info_.wavefrontWidth_ * hwInfo()->simdPerCU_ *
+                             palProp.gfxipProperties.shaderCore.numWavefrontsPerSimd;
   }
 }
 

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

@@ -1226,6 +1226,14 @@ bool Device::populateOCLDeviceConstants() {
         hsa_agent_get_info(_bkendDevice, (hsa_agent_info_t)HSA_AMD_AGENT_INFO_MEMORY_WIDTH, &info_.vramBusBitWidth_)) {
       return false;
     }
+
+    uint32_t max_waves_per_cu;
+    if (HSA_STATUS_SUCCESS !=
+        hsa_agent_get_info(_bkendDevice, (hsa_agent_info_t)HSA_AMD_AGENT_INFO_MAX_WAVES_PER_CU, &max_waves_per_cu)) {
+      return false;
+    }
+
+    info_.maxThreadsPerCU_ = info_.wavefrontWidth_ * max_waves_per_cu;
     uint32_t cache_sizes[4];
     /* FIXIT [skudchad] -  Seems like hardcoded in HSA backend so 0*/
     if (HSA_STATUS_SUCCESS !=