P4 to Git Change 1338972 by scharupa@scharupa_temp on 2016/11/09 16:33:58

SWDEV-95488 - Limiting max cap Context on ROCm, it is limited to 100

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/device.cpp#204 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/device.hpp#279 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocdevice.cpp#25 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocdevice.hpp#7 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocsettings.cpp#6 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocvirtual.cpp#16 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocvirtual.hpp#4 edit

rocclr/runtime/device/device.cpp

@@ -628,6 +628,8 @@ Settings::Settings()
     waitCommand_         = AMD_OCL_WAIT_COMMAND;
     supportDepthsRGB_    = false;
     enableHwDebug_       = false;
+    commandQueues_        = 200;  //!< Field value set to maximum number
+                                  //!< concurrent Virtual GPUs for default
 }
 
 bool

rocclr/runtime/device/device.hpp

@@ -594,6 +594,8 @@ public:
         uint    value_;
     };
 
+    uint commandQueues_; //!< Field value for maximum number
+                         //!< concurrent Virtual GPUs for each backend
     //! Default constructor
     Settings();
 

rocclr/runtime/device/rocm/rocdevice.cpp

@@ -186,6 +186,7 @@ Device::Device(hsa_agent_t bkendDevice)
     , alloc_granularity_(0)
     , context_(nullptr)
     , xferQueue_(nullptr)
+    , numOfVgpus_(0)
 {
     group_segment_.handle = 0;
     system_segment_.handle = 0;

rocclr/runtime/device/rocm/rocdevice.hpp

@@ -221,6 +221,8 @@ public:
 
     static bool loadHsaModules();
 
+    amd::Atomic<uint>   numOfVgpus_; //!< Virtual gpu unique index
+
     bool create();
 
     //! Construct a new physical HSA device

rocclr/runtime/device/rocm/rocsettings.cpp

@@ -52,6 +52,8 @@ Settings::Settings()
     partialDispatch = getenv("OPENCL_DISABLE_PARTIAL_DISPATCH");
     enablePartialDispatch_ = (partialDispatch) ? false : true;
     partialDispatch_ = (partialDispatch) ? false : true;
+    commandQueues_ = 100 //!< Field value set to maximum number
+	                 //!< concurrent Virtual GPUs for ROCm backend
 }
 
 bool
@@ -97,6 +99,11 @@ Settings::create(bool doublePrecision)
 void
 Settings::override()
 {
+    if (!flagIsDefault(GPU_MAX_COMMAND_QUEUES)) {
+        commandQueues_ = GPU_MAX_COMMAND_QUEUES;
+    }
+}
+
 }
 
 } // namespace roc

rocclr/runtime/device/rocm/rocvirtual.cpp

@@ -450,6 +450,8 @@ bool VirtualGPU::releaseGpuMemoryFence() {
 VirtualGPU::VirtualGPU(Device &device)
     : device::VirtualDevice(device)
     , roc_device_(device)
+    , index_(device_.numOfVgpus_++) // Virtual gpu unique index incrementing
+    , gpuDevice_(device)
 {
     gpu_device_ = device.getBackendDevice();
     // Initialize the last signal and dispatch flags
@@ -477,6 +479,7 @@ VirtualGPU::~VirtualGPU()
     }
 
     tools_lib_ = NULL;
+    --gpuDevice_.numOfVgpus_; // Virtual gpu unique index decrementing
 }
 
 bool
@@ -494,6 +497,11 @@ VirtualGPU::create(bool profilingEna)
         tools_lib_ = amd::Os::loadLibrary(tools_lib_name);
     }
 
+    // Checking Virtual gpu unique index for ROCm backend
+    if (index() > device().settings().commandQueues_) {
+        return false;
+    }
+
     uint32_t queue_max_packets = 0;
     if (HSA_STATUS_SUCCESS !=
         hsa_agent_get_info(

rocclr/runtime/device/rocm/rocvirtual.hpp

@@ -193,7 +193,10 @@ public:
     bool processMemObjects(
         const amd::Kernel&  kernel, //!< AMD kernel object for execution
         const_address       params  //!< Pointer to the param's store
-        );
+	);
+    //Retun the virtual gpu unique index
+    uint index() const { return index_;  }
+
 
 // } roc OpenCL integration
 private:
@@ -244,7 +247,8 @@ private:
     uint  kernarg_pool_cur_offset_;
 
     std::vector<ProfilingSignal> signal_pool_;  //!< Pool of signals for profiling
-
+    Device&        gpuDevice_;   //!< Roc device obj
+    const uint     index_; //!< Virtual gpu unique index
     friend class Timestamp;
 };
 }