From 067c7316bacb8c99e04df4ea9f9128266ccde5be Mon Sep 17 00:00:00 2001
From: Anusha GodavarthySurya <anusha.godavarthysurya@amd.com>
Date: Fri, 23 Dec 2022 10:37:48 +0000
Subject: [PATCH] SWDEV-373173 - Correct max VGPRs and VGPR Granularity in
 Occupancy calculation for different asics

Change-Id: Ib213ed8767a793a832776fba7c0811c2f023a8ad
---
 hipamd/src/hip_platform.cpp | 68 ++++++++++++++++++++++++++++++-------
 1 file changed, 55 insertions(+), 13 deletions(-)

diff --git a/hipamd/src/hip_platform.cpp b/hipamd/src/hip_platform.cpp
index afcbaab007..93bdf4f837 100644
--- a/hipamd/src/hip_platform.cpp
+++ b/hipamd/src/hip_platform.cpp
@@ -333,20 +333,62 @@ hipError_t ihipOccupancyMaxActiveBlocksPerMultiprocessor(
     MaxWavesPerSimd = 16;
   }
   size_t VgprWaves = MaxWavesPerSimd;
-  size_t maxVGPRs;
-  uint32_t VgprGranularity;
-  if (device.isa().versionMajor() <= 9) {
-    if (device.isa().versionMajor() == 9 && device.isa().versionMinor() == 0 &&
-        device.isa().versionStepping() == 10) {
-      maxVGPRs = 512;
-      VgprGranularity = 8;
-    } else {
-      maxVGPRs = 256;
+  uint32_t VgprGranularity = 0;
+  size_t maxVGPRs = 0;
+  size_t wavefrontSize = wrkGrpInfo->wavefrontSize_;
+  switch (device.isa().versionMajor()) {
+    case (11):
+      if (device.isa().versionMinor() == 0) {
+        switch (device.isa().versionStepping()) {
+          case (0):
+          case (1):
+            VgprGranularity = (wavefrontSize == 32) ? 24 : 12;
+            maxVGPRs = (wavefrontSize == 32) ? 1536 : 768;
+            break;
+          case (2):
+          case (3):
+            VgprGranularity = (wavefrontSize == 32) ? 16 : 8;
+            maxVGPRs = (wavefrontSize == 32) ? 1024 : 512;
+            break;
+          default:
+            VgprGranularity = 8;
+            maxVGPRs = 1024;
+            break;
+        }
+      }
+      break;
+    case (10):
+      switch (device.isa().versionMinor()) {
+        case (0):
+        case (1):
+          VgprGranularity = (wavefrontSize == 32) ? 8 : 4;
+          maxVGPRs = (wavefrontSize == 32) ? 1024 : 512;
+          break;
+        case (3):
+          VgprGranularity = (wavefrontSize == 32) ? 16 : 8;
+          maxVGPRs = (wavefrontSize == 32) ? 1024 : 512;
+          break;
+        default:
+          VgprGranularity = 8;
+          maxVGPRs = 1024;
+          break;
+      }
+      break;
+    case (9):
+      if ((device.isa().versionMinor() == 0 && device.isa().versionStepping() == 10) ||
+          (device.isa().versionMinor() == 4 && device.isa().versionStepping() == 0)) {
+        VgprGranularity = 8;
+        maxVGPRs = 512;
+      } else {
+        VgprGranularity = 4;
+        maxVGPRs = 256;
+      }
+      break;
+    default:
+      // For gfx<=8
       VgprGranularity = 4;
-    }
-  } else {
-    maxVGPRs = 1024;
-    VgprGranularity = 8;
+      maxVGPRs = 256;
+      break;
   }
   if (wrkGrpInfo->usedSGPRs_ > 0) {
     VgprWaves = maxVGPRs / amd::alignUp(wrkGrpInfo->usedVGPRs_, VgprGranularity);