SWDEV-373173 - Correct max VGPRs and VGPR Granularity in Occupancy calculation for different asics

Change-Id: I6f595428ed31733cdc04e65e7c3e383f454ab72b


[ROCm/clr commit: 2b50636809]

projects/clr/rocclr/device/device.hpp

@@ -613,6 +613,11 @@ struct Info : public amd::EmbeddedObject {
   size_t virtualMemAllocGranularity_; //!< virtual memory allocation size/addr granularity
 
   uint32_t driverNodeId_;
+  //! Number of Physical SGPRs per SIMD
+  uint32_t sgprsPerSimd_;
+  //! Number of VGPRs per SIMD
+  uint32_t vgprsPerSimd_;
+  uint32_t vgprAllocGranularity_;
 };
 
 //! Device settings

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

@@ -632,7 +632,11 @@ void NullDevice::fillDeviceInfo(const Pal::DeviceProperties& palProp,
     }
   }
   info_.virtualMemoryManagement_ = true;
-  info_.virtualMemAllocGranularity_ = static_cast<size_t>(palProp.gpuMemoryProperties.virtualMemAllocGranularity);
+  info_.virtualMemAllocGranularity_ =
+      static_cast<size_t>(palProp.gpuMemoryProperties.virtualMemAllocGranularity);
+  info_.vgprAllocGranularity_ = palProp.gfxipProperties.shaderCore.vgprAllocGranularity;
+  info_.vgprsPerSimd_ = palProp.gfxipProperties.shaderCore.vgprsPerSimd;
+  info_.sgprsPerSimd_ = palProp.gfxipProperties.shaderCore.sgprsPerSimd;
 }
 
 Device::XferBuffers::~XferBuffers() {

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

@@ -1648,7 +1648,63 @@ bool Device::populateOCLDeviceConstants() {
 
   info_.globalCUMask_ = {};
   info_.virtualMemoryManagement_ = false;
+  switch (isa().versionMajor()) {
+    case (11):
+      if (isa().versionMinor() == 0) {
+        switch (isa().versionStepping()) {
+          case (0):
+          case (1):
+            info_.vgprAllocGranularity_ = 24;
+            info_.vgprsPerSimd_ = 1536;
+            break;
+          case (2):
+          case (3):
+          default:
+            info_.vgprAllocGranularity_ = 16;
+            info_.vgprsPerSimd_ = 1024;
+            break;
+        }
+      }
+      break;
+    case (10):
+      switch (isa().versionMinor()) {
+        case (0):
+        case (1):
+          info_.vgprAllocGranularity_ = 8;
+          info_.vgprsPerSimd_ = 1024;
+          break;
+        case (3):
+        default:
+          info_.vgprAllocGranularity_ = 16;
+          info_.vgprsPerSimd_ = 1024;
+          break;
+      }
+      break;
+    case (9):
+      if ((isa().versionMinor() == 0 && isa().versionStepping() == 10) ||
+          (isa().versionMinor() == 4 && isa().versionStepping() == 0)) {
+        info_.vgprAllocGranularity_ = 8;
+        info_.vgprsPerSimd_ = 512;
+      } else {
+        info_.vgprAllocGranularity_ = 4;
+        info_.vgprsPerSimd_ = 256;
+      }
+      break;
+    default:
+      // For gfx<=8
+      info_.vgprAllocGranularity_ = 4;
+      info_.vgprsPerSimd_ = 256;
+      break;
+  }
 
+  if (isa().versionMajor() < 8) {
+    info_.sgprsPerSimd_ = 512;
+  } else if (isa().versionMajor() < 10) {
+    info_.sgprsPerSimd_ = 800;
+  } else {
+    info_.sgprsPerSimd_ =
+        std::numeric_limits<uint32_t>::max();  // gfx10+ does not share SGPRs between waves
+  }
   return true;
 }