gentoo/dev-libs/rocm-comgr/files/rocm-comgr-6.4.1-extend-isa-compatibility-check.patch

Load kernels when compatible by ISA, e. g. if AMDGPU_TARGETS is set
to gfx1030 and some application was started on gfx1036, it loads gfx1030 kernel.

Based on Debian patch by Cordell Bloor <cgmb@slerp.xyz>
https://salsa.debian.org/rocm-team/rocm-hipamd/-/blob/master/debian/patches/0026-extend-hip-isa-compatibility-check.patch
--- a/src/comgr-metadata.cpp
+++ b/src/comgr-metadata.cpp
@@ -927,23 +927,86 @@ constexpr size_t OffloadBundleMagicLen =
     strLiteralLength(ClangOffloadBundlerMagic);
 } // namespace
 
-bool isCompatibleIsaName(StringRef IsaName, StringRef CodeObjectIsaName) {
+struct GfxPattern {
+  std::string root;
+  std::string suffixes;
+};
+
+static bool matches(const GfxPattern& p, StringRef s) {
+  if (p.root.size() + 1 != s.size()) {
+    return false;
+  }
+  if (0 != std::memcmp(p.root.data(), s.data(), p.root.size())) {
+    return false;
+  }
+  return p.suffixes.find(s[p.root.size()]) != std::string::npos;
+}
+
+static bool isGfx900EquivalentProcessor(StringRef processor) {
+  return matches(GfxPattern{"gfx90", "029c"}, processor);
+}
+
+static bool isGfx900SupersetProcessor(StringRef processor) {
+  return matches(GfxPattern{"gfx90", "0269c"}, processor);
+}
+
+static bool isGfx1030EquivalentProcessor(StringRef processor) {
+  return matches(GfxPattern{"gfx103", "0123456"}, processor);
+}
+
+static bool isGfx1010EquivalentProcessor(StringRef processor) {
+  return matches(GfxPattern{"gfx101", "0"}, processor);
+}
+
+static bool isGfx1010SupersetProcessor(StringRef processor) {
+  return matches(GfxPattern{"gfx101", "0123"}, processor);
+}
+
+enum CompatibilityScore {
+  CS_EXACT_MATCH           = 1 << 4,
+  CS_PROCESSOR_MATCH       = 1 << 3,
+  CS_PROCESSOR_COMPATIBLE  = 1 << 2,
+  CS_XNACK_SPECIALIZED     = 1 << 1,
+  CS_SRAM_ECC_SPECIALIZED  = 1 << 0,
+  CS_INCOMPATIBLE          = 0,
+};
+
+static int getProcessorCompatibilityScore(StringRef CodeObjectProcessor,
+                                          StringRef AgentProcessor) {
+  if (CodeObjectProcessor == AgentProcessor) {
+    return CS_PROCESSOR_MATCH;
+  }
+
+  bool compatible = false;
+  if (isGfx900SupersetProcessor(AgentProcessor)) {
+    compatible = isGfx900EquivalentProcessor(CodeObjectProcessor);
+  } else if (isGfx1010SupersetProcessor(AgentProcessor)) {
+    compatible = isGfx1010EquivalentProcessor(CodeObjectProcessor);
+  } else if (isGfx1030EquivalentProcessor(AgentProcessor)) {
+    compatible = isGfx1030EquivalentProcessor(CodeObjectProcessor);
+  }
+
+  return compatible ? CS_PROCESSOR_COMPATIBLE : CS_INCOMPATIBLE;
+}
+
+static int getCompatiblityScore(StringRef IsaName, StringRef CodeObjectIsaName) {
   if (IsaName == CodeObjectIsaName) {
-    return true;
+    return CS_EXACT_MATCH;
   }
 
   TargetIdentifier CodeObjectIdent;
   if (parseTargetIdentifier(CodeObjectIsaName, CodeObjectIdent)) {
-    return false;
+    return CS_INCOMPATIBLE;
   }
 
   TargetIdentifier IsaIdent;
   if (parseTargetIdentifier(IsaName, IsaIdent)) {
-    return false;
+    return CS_INCOMPATIBLE;
   }
 
-  if (CodeObjectIdent.Processor != IsaIdent.Processor) {
-    return false;
+  int ProcessorScore = getProcessorCompatibilityScore(CodeObjectIdent.Processor, IsaIdent.Processor);
+  if (ProcessorScore == CS_INCOMPATIBLE) {
+    return CS_INCOMPATIBLE;
   }
 
   char CodeObjectXnack = ' ', CodeObjectSramecc = ' ';
@@ -967,18 +1030,23 @@ bool isCompatibleIsaName(StringRef IsaName, StringRef CodeObjectIsaName) {
     }
   }
 
+  int XnackBonus = 0;
   if (CodeObjectXnack != ' ') {
     if (CodeObjectXnack != IsaXnack) {
-      return false;
+      return CS_INCOMPATIBLE;
     }
+    XnackBonus = CS_XNACK_SPECIALIZED;
   }
 
+  int SrameccBonus = 0;
   if (CodeObjectSramecc != ' ') {
     if (CodeObjectSramecc != IsaSramecc) {
-      return false;
+      return CS_INCOMPATIBLE;
     }
+    SrameccBonus = CS_SRAM_ECC_SPECIALIZED;
   }
-  return true;
+
+  return ProcessorScore + XnackBonus + SrameccBonus;
 }
 
 amd_comgr_status_t
@@ -996,14 +1064,21 @@ lookUpCodeObjectInSharedObject(DataObject *DataP,
     return Status;
   }
 
+  int MaxScore = 0;
+  unsigned MaxScoreItem;
   for (unsigned J = 0; J < QueryListSize; J++) {
-    if (isCompatibleIsaName(QueryList[J].isa, IsaName)) {
-      QueryList[J].offset = 0;
-      QueryList[J].size = DataP->Size;
-      break;
+    int Score = getCompatiblityScore(QueryList[J].isa, IsaName);
+    if (Score > MaxScore) {
+      MaxScore = Score;
+      MaxScoreItem = J;
     }
   }
 
+  if (MaxScore) {
+    QueryList[MaxScoreItem].offset = 0;
+    QueryList[MaxScoreItem].size = DataP->Size;
+  }
+
   return AMD_COMGR_STATUS_SUCCESS;
 }
 
@@ -1015,7 +1090,6 @@ amd_comgr_status_t lookUpCodeObject(DataObject *DataP,
     return lookUpCodeObjectInSharedObject(DataP, QueryList, QueryListSize);
   }
 
-  int Seen = 0;
   BinaryStreamReader Reader(StringRef(DataP->Data, DataP->Size),
                             llvm::endianness::little);
 
@@ -1041,6 +1115,8 @@ amd_comgr_status_t lookUpCodeObject(DataObject *DataP,
     QueryList[I].size = 0;
   }
 
+  std::vector<int> QueryListScores(QueryListSize);
+
   // For each code object, extract BundleEntryID information, and check that
   // against each ISA in the QueryList
   for (uint64_t I = 0; I < NumOfCodeObjects; I++) {
@@ -1073,28 +1149,22 @@ amd_comgr_status_t lookUpCodeObject(DataObject *DataP,
     }
 
     for (unsigned J = 0; J < QueryListSize; J++) {
-      // If this QueryList item has already been found to be compatible with
+      // If this QueryList item has exact match with
       // another BundleEntryID, no need to check against the current
       // BundleEntryID
-      if (QueryList[J].size != 0) {
+      if (QueryListScores[J] == CS_EXACT_MATCH) {
         continue;
       }
 
       // If the QueryList Isa is compatible with the BundleEntryID, set the
       // QueryList offset/size to this BundleEntryID
-      if (isCompatibleIsaName(QueryList[J].isa, OffloadAndTargetId.second)) {
+      int Score = getCompatiblityScore(QueryList[J].isa, OffloadAndTargetId.second);
+      if (Score > QueryListScores[J]) {
+        QueryListScores[J] = Score;
         QueryList[J].offset = BundleEntryCodeObjectOffset;
         QueryList[J].size = BundleEntryCodeObjectSize;
-        Seen++;
-        break;
       }
     }
-
-    // Stop iterating over BundleEntryIDs once we have populated the entire
-    // QueryList
-    if (Seen == (int)QueryListSize) {
-      break;
-    }
   }
 
   return AMD_COMGR_STATUS_SUCCESS;