P4 to Git Change 1455464 by asalmanp@asalmanp-ocl-stg on 2017/09/05 18:07:43

SWDEV-79278 - [OCL][PAL] refactoring PAL Null device create function to account for creating all the gfx9+ subtarget devices such as gfx901/gfx902/etc

	ReviewboardURL = http://ocltc.amd.com/reviews/r/13378/

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/pal/paldefs.hpp#21 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/paldevice.cpp#53 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/paldevice.hpp#18 edit


[ROCm/clr commit: 9772217dcb]

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

@@ -147,33 +147,21 @@ static const AMDDeviceInfo DeviceInfo[] = {
     /* Baffin */ {"Baffin", "baffin", 4, 16, 1, 256, 64 * Ki, 32, 804},
 };
 
-// The GfxIpDeviceInfo table must match with GfxIpLevel enum
-// (located in //depot/stg/pal/inc/core/palDevice.h).
-static const AMDDeviceInfo GfxIpDeviceInfo[] = {
-    /* Unknown  */  {"unknown", "unknown",  4, 16, 1, 256, 64 * Ki, 32, 000},
-    /* GFX600 */    {"gfx600",  "gfx600",   4, 16, 1, 256, 64 * Ki, 32, 600},
-    /* GFX700 */    {"gfx700",  "gfx700",   4, 16, 1, 256, 64 * Ki, 32, 700},
-    /* GFX800 */    {"gfx800",  "gfx800",   4, 16, 1, 256, 64 * Ki, 32, 800},
-    /* GFX810 */    {"gfx810",  "gfx810",   4, 16, 1, 256, 64 * Ki, 32, 810},
-    /* GFX900 */    {"gfx900",  "gfx900",   4, 16, 1, 256, 64 * Ki, 32, 900},
-    /* GFX1000 */   {"gfx1000", "gfx1000",  4, 16, 1, 256, 64 * Ki, 32, 1000 },
-};
-
 // Ordering as per AsicRevision# in //depot/stg/pal/inc/core/palDevice.h and
 // http://confluence.amd.com/pages/viewpage.action?spaceKey=ASLC&title=AMDGPU+Target+Names
 static const AMDDeviceInfo Gfx9PlusSubDeviceInfo[] = {
     /* Vega10       */{"gfx900", "gfx900", 4, 16, 1, 256, 64 * Ki, 32, 900},
     /* Vega10 XNACK */{"gfx901", "gfx901", 4, 16, 1, 256, 64 * Ki, 32, 901},
-    /* Vega12       */{"gfx904", "gfx904", 4, 16, 1, 256, 64 * Ki, 32, 904},
+    /* Vega12       */{ "", "", 4, 16, 1, 256, 64 * Ki, 32, 904},
-    /* Vega12 XNACK */{"gfx905", "gfx905", 4, 16, 1, 256, 64 * Ki, 32, 905},
+    /* Vega12 XNACK */{ "", "", 4, 16, 1, 256, 64 * Ki, 32, 905},
-    /* Vega20       */{"gfx906", "gfx906", 4, 16, 1, 256, 64 * Ki, 32, 906},
+    /* Vega20       */{ "", "", 4, 16, 1, 256, 64 * Ki, 32, 906},
-    /* Vega20 XNACK */{"gfx907", "gfx907", 4, 16, 1, 256, 64 * Ki, 32, 907},
+    /* Vega20 XNACK */{ "", "", 4, 16, 1, 256, 64 * Ki, 32, 907},
     /* Raven        */{"gfx902", "gfx902", 4, 16, 1, 256, 64 * Ki, 32, 902},
     /* Raven XNACK  */{"gfx903", "gfx903", 4, 16, 1, 256, 64 * Ki, 32, 903},
-    /* Raven2       */{0},
+    /* Raven2       */{""      , ""      , 0, 0 , 0, 0  , 0      , 0 , 0},
-    /* Raven2 XNACK */{0},
+    /* Raven2 XNACK */{""      , ""      , 0, 0 , 0, 0  , 0      , 0 , 0},
-    /* Navi10       */{"gfx1000", "gfx1000", 4, 16, 1, 256, 64 * Ki, 32, 1000},
+    /* Navi10       */{ "", "", 4, 16, 1, 256, 64 * Ki, 32, 1000},
-    /* Navi10 XNACK */{"gfx1001", "gfx1001", 4, 16, 1, 256, 64 * Ki, 32, 1001},
+    /* Navi10 XNACK */{ "", "", 4, 16, 1, 256, 64 * Ki, 32, 1001},
 };
 
 enum gfx_handle {

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

@@ -62,6 +62,7 @@ NullDevice::NullDevice()
 
 bool NullDevice::init() {
   std::vector<Device*> devices;
+  std::string driverVersion;
 
   devices = getDevices(CL_DEVICE_TYPE_GPU, false);
 
@@ -74,11 +75,14 @@ bool NullDevice::init() {
       continue;
     }
 
-    // Loop through all active devices and see if we match one
+    // Loop through all active PAL devices and see if we match one
     for (uint i = 0; i < devices.size(); ++i) {
-      if (static_cast<NullDevice*>(devices[i])->asicRevision() == revision) {
+      driverVersion = static_cast<amd::Device*>(devices[i])->info().driverVersion_;
-        foundActive = true;
+      if (driverVersion.find("PAL") != std::string::npos) {
-        break;
+        if (static_cast<NullDevice*>(devices[i])->asicRevision() == revision) {
+            foundActive = true;
+            break;
+        }
       }
     }
 
@@ -98,42 +102,89 @@ bool NullDevice::init() {
   }
 
   // Loop through all supported devices and create each of them
-  for (uint id = static_cast<uint>(Pal::GfxIpLevel::GfxIp7);
+  for (uint id = 0;
-       id <= static_cast<uint>(Pal::GfxIpLevel::GfxIp10); ++id) {
+        id < sizeof(Gfx9PlusSubDeviceInfo)/sizeof(AMDDeviceInfo); ++id) {
-    bool foundActive = false;
+      bool foundActive = false;
-    Pal::GfxIpLevel ipLevel = static_cast<Pal::GfxIpLevel>(id);
+      uint gfxipVersion = pal::Gfx9PlusSubDeviceInfo[id].gfxipVersion_;
 
-    if (pal::GfxIpDeviceInfo[id].targetName_[0] == '\0') {
+      if (pal::Gfx9PlusSubDeviceInfo[id].targetName_[0] == '\0') {
-      continue;
+          continue;
-    }
-
-    // Loop through all active devices and see if we match one
-    for (uint i = 0; i < devices.size(); ++i) {
-      if (static_cast<NullDevice*>(devices[i])->ipLevel() == ipLevel) {
-        foundActive = true;
-        break;
       }
-    }
 
-    // Don't report an offline device if it's active
+      // Loop through all active PAL devices and see if we match one
-    if (foundActive) {
+      for (uint i = 0; i < devices.size(); ++i) {
-      continue;
+        driverVersion = static_cast<amd::Device*>(devices[i])->info().driverVersion_;
-    }
+        if (driverVersion.find("PAL") != std::string::npos) {
-
+          if (static_cast<NullDevice*>(devices[i])->hwInfo()->gfxipVersion_ ==
-    NullDevice* dev = new NullDevice();
+                                                              gfxipVersion) {
-    if (nullptr != dev) {
+              foundActive = true;
-      if (!dev->create(Pal::AsicRevision::Unknown, ipLevel)) {
+              break;
-        delete dev;
+          }
-      } else {
+        }
-        dev->registerDevice();
+      }
+
+      // Don't report an offline device if it's active
+      if (foundActive) {
+          continue;
+      }
+
+      Pal::GfxIpLevel ipLevel = Pal::GfxIpLevel::_None;
+      uint ipLevelMajor = round(pal::Gfx9PlusSubDeviceInfo[id].gfxipVersion_ / 100);
+      switch (ipLevelMajor) {
+      case 9:
+          ipLevel = Pal::GfxIpLevel::GfxIp9;
+          break;
+      case 10:
+          ipLevel = Pal::GfxIpLevel::GfxIp10;
+          break;
+      }
+
+      Pal::AsicRevision revision = Pal::AsicRevision::Unknown;
+      uint xNACKSupported = 0;
+      switch (pal::Gfx9PlusSubDeviceInfo[id].gfxipVersion_) {
+      case 901:
+          xNACKSupported = 1;
+      case 900:
+          revision = Pal::AsicRevision::Vega10;
+          break;
+      case 903:
+          xNACKSupported = 1;
+      case 902:
+          revision = Pal::AsicRevision::Raven;
+          break;
+      case 905:
+          xNACKSupported = 1;
+      case 904:
+          revision = Pal::AsicRevision::Vega12;
+          break;
+      case 907:
+          xNACKSupported = 1;
+      case 906:
+          revision = Pal::AsicRevision::Vega20;
+          break;
+      case 1001:
+          xNACKSupported = 1;
+      case 1000:
+          revision = Pal::AsicRevision::Navi10;
+          break;
+      }
+
+      NullDevice* dev = new NullDevice();
+      if (nullptr != dev) {
+          if (!dev->create(revision, ipLevel, xNACKSupported)) {
+              delete dev;
+          }
+          else {
+              dev->registerDevice();
+          }
       }
-    }
   }
 
   return true;
 }
 
-bool NullDevice::create(Pal::AsicRevision asicRevision, Pal::GfxIpLevel ipLevel) {
+bool NullDevice::create(Pal::AsicRevision asicRevision, Pal::GfxIpLevel ipLevel,
+                        uint xNACKSupported) {
   online_ = false;
   Pal::DeviceProperties properties = {};
 
@@ -142,12 +193,17 @@ bool NullDevice::create(Pal::AsicRevision asicRevision, Pal::GfxIpLevel ipLevel)
   ipLevel_ = ipLevel;
   properties.revision = asicRevision;
   properties.gfxLevel = ipLevel;
+  uint subtarget = 0;
 
   // Update HW info for the device
   if ((GPU_ENABLE_PAL == 1) && (ipLevel == Pal::GfxIpLevel::_None)) {
     hwInfo_ = &DeviceInfo[static_cast<uint>(asicRevision)];
   } else if (ipLevel >= Pal::GfxIpLevel::GfxIp9) {
-    hwInfo_ = &GfxIpDeviceInfo[static_cast<uint>(ipLevel)];
+      subtarget = (static_cast<uint>(asicRevision_) %
+                   static_cast<uint>(Pal::AsicRevision::Vega10))
+                   << 1 | xNACKSupported;
+      hwInfo_ = &Gfx9PlusSubDeviceInfo[subtarget];
+
   } else {
     return false;
   }

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

@@ -44,8 +44,9 @@ class NullDevice : public amd::Device {
 
   //! Creates an offline device with the specified target
   bool create(Pal::AsicRevision asicRevision,  //!< GPU ASIC revision
-              Pal::GfxIpLevel ipLevel          //!< GPU ip level
+              Pal::GfxIpLevel ipLevel,         //!< GPU ip level
-              );
+              uint xNACKSupported = 0          //!< GPU xNACKSupported
+             );
 
   virtual cl_int createSubDevices(device::CreateSubDevicesInfo& create_info, cl_uint num_entries,
                                   cl_device_id* devices, cl_uint* num_devices) {