merge aomp-dev into amd-staging

bin/clone_aomp.sh

@@ -218,26 +218,15 @@ if [[ "$AOMP_VERSION" == "13.1" ]] || [[ $AOMP_MAJOR_VERSION -gt 13 ]] ; then
    # However, we gave up on using the repo command to clone the repos. 
    # That is all done here by parsing the manifest file.
 
-   # According to git documentation this ssh command should return 1 if authentication is successful
-   ssh -T $AOMP_GIT_INTERNAL_IP 2> /dev/null
-   if [ $? == 1 ] && [ "$AOMP_EXTERNAL_MANIFEST" != 1 ]; then
-      # AMD internal repo file
-      if [ "$AOMP_NEW" == "1" ]; then
-         manifest_file=$thisdir/../manifests/aompi_new_${AOMP_VERSION}.xml
-      else
-         manifest_file=$thisdir/../manifests/aompi_${AOMP_VERSION}.xml
-      fi
+   abranch=$(git branch | awk '/\*/ { print $2; }')
+   # Use release manifest if on release branch
+   if [ "$abranch" == "aomp-${AOMP_VERSION_STRING}" ]; then
+      manifest_file=$thisdir/../manifests/aomp_${AOMP_VERSION_STRING}.xml
    else
-      abranch=$(git branch | awk '/\*/ { print $2; }')
-      # Use release manifest if on release branch
-      if [ "$abranch" == "aomp-${AOMP_VERSION_STRING}" ]; then
-         manifest_file=$thisdir/../manifests/aomp_${AOMP_VERSION_STRING}.xml
+      if [ "$AOMP_NEW" == "1" ]; then
+        manifest_file=$thisdir/../manifests/aomp_new_${AOMP_VERSION}.xml
       else
-	 if [ "$AOMP_NEW" == "1" ]; then
-            manifest_file=$thisdir/../manifests/aomp_new_${AOMP_VERSION}.xml
-         else
-            manifest_file=$thisdir/../manifests/aomp_${AOMP_VERSION}.xml
-	 fi
+        manifest_file=$thisdir/../manifests/aomp_${AOMP_VERSION}.xml
       fi
    fi
    echo "USED manifest file: $manifest_file"

bin/run_composable-kernels.sh

@@ -282,7 +282,14 @@ if [ ! -d "${CK_TOP}" ]; then
   mkdir -p "${CK_TOP}" || exit 1
 fi
 
-if [ ! -d "${CK_REPO}" ]; then
+# Validate that CK_REPO is an actual git checkout, not just an existing
+# directory. Alola periodically wipes the *contents* of old node-local dirs
+# while leaving the directory shells behind, which yields a CK_REPO that
+# exists but has no .git (and no CMakeLists.txt), breaking the build. A plain
+# "-d" test would skip the clone and then fail. Re-clone when the tree is
+# missing or not a valid repo; only pull when it is a healthy checkout.
+if ! git -C "${CK_REPO}" rev-parse --is-inside-work-tree >/dev/null 2>&1; then
+  rm -rf "${CK_REPO}"
   git clone --single-branch --depth 1 ${CKRepoURL} "${CK_REPO}"
 elif [ "${ShouldUpdateCKRepo}" == 'yes' ]; then
   pushd "${CK_REPO}" || exit 1

manifests/aomp_23.0.xml

@@ -1,30 +1,20 @@
 <?xml version="1.0" encoding="UTF-8"?>
 <manifest>
-  <!-- Manifest for AOMP 23.0-x which uses ROCM 7.2 release branches of external repositories -->
+  <!-- Manifest for AOMP 23.0-x which uses ROCM 7.13 release branches of external repositories -->
 
-    <remote name="gerritgit" review="git.amd.com:8080" fetch="ssh://gerritgit/" />
-    <default revision="release/rocm-rel-7.2" remote="gerritgit" sync-j="4" sync-c="true" />
     <remote name="roc"  fetch="https://github.com/ROCm" />
     <remote name="simde"  fetch="https://github.com/simd-everywhere" />
 
-<!-- These first 4 repos are NOT rocm 7.2. They are compiler developer branches -->
-    <project remote="roc" path="llvm-project" name="llvm-project"    revision="amd-staging" groups="unlocked" />
-    <project remote="roc" path="SPIRV-LLVM-Translator" name="SPIRV-LLVM-Translator"    revision="amd-staging" groups="unlocked" />
-    <project remote="roc" path="hipify" name="hipify"    revision="amd-staging" groups="unlocked" />
+<!-- These first 4 repos are NOT rocm 7.13. They are compiler developer branches -->
+    <project remote="roc" path="llvm-project" name="llvm-project.git" revision="amd-staging" groups="unlocked" />
+    <project remote="roc" path="SPIRV-LLVM-Translator" name="spirv-llvm-translator.git" revision="amd-staging" groups="unlocked" />
+    <project remote="roc" path="hipify" name="hipify.git" revision="amd-staging" groups="unlocked" />
     <project remote="simde" path="simde" name="simde" revision="master" groups="unlocked" />
 
     <project remote="roc" path="aomp" name="aomp"                 revision="aomp-dev" groups="unlocked" />
 
-    <project remote="roc" path="rocprofiler-sdk" name="rocprofiler-sdk"              revision="release/rocm-rel-7.2" groups="unlocked" />
-    <project remote="roc" path="ROCdbgapi" name="ROCdbgapi"                  revision="release/rocm-rel-7.2" groups="unlocked" />
-    <project remote="roc" path="ROCgdb" name="ROCgdb"                        revision="release/rocm-rel-7.2" groups="unlocked" />
-    <project remote="roc" path="hip"    name="hip"                           revision="release/rocm-rel-7.2" groups="unlocked" />
-    <project remote="roc" path="clr" name="clr"                        revision="release/rocm-rel-7.2" groups="unlocked" />
-    <project remote="roc" path="rocminfo" name="rocminfo"                         revision="release/rocm-rel-7.2" groups="unlocked" />
-    <project remote="roc" path="rocm_smi_lib" name="rocm_smi_lib"                 revision="release/rocm-rel-7.2" groups="unlocked" />
-    <project remote="roc" path="amdsmi" name="amdsmi"                             revision="release/rocm-rel-7.2" groups="unlocked" />
-    <project remote="roc" path="rocm-cmake" name="rocm-cmake"                     revision="release/rocm-rel-7.2" groups="unlocked" />
-    <project remote="roc" path="rocr-runtime" name="ROCR-Runtime"                 revision="release/rocm-rel-7.2" groups="unlocked" />
-    <project remote="roc" path="rocprofiler-register" name="rocprofiler-register" revision="release/rocm-rel-7.2" groups="unlocked" />
-    <project remote="roc" path="hipfort" name="hipfort"                 revision="release/rocm-rel-7.2" groups="unlocked" />
+    <project remote="roc" path="ROCgdb" name="ROCgdb"       upstream="amd-staging"     revision="therock-7.13" tag="true" groups="revlocked" />
+    <project remote="roc" path="rocm-cmake" name="rocm-cmake"  upstream="develop"  revision="therock-7.13" tag="true" groups="revlocked" />
+    <project remote="roc" path="hipfort" name="hipfort"          revision="release/rocm-rel-7.2" groups="unlocked" />
+    <project remote="roc" path="rocm-systems" name="rocm-systems" upstream="develop" revision="therock-7.13" tag="true" groups="revlocked" />
 </manifest>

test/smoke-dev/HipDevCov/Makefile

@@ -21,10 +21,10 @@ LINK_FLAGS   = -L$(ROCM_PATH)/lib -lamdhip64 -Wl,-rpath,$(ROCM_PATH)/lib
 
 CC           = $(AOMP)/bin/clang $(VERBOSE)
 
-RUNCMD       = ./run_and_check.sh "$(AOMP)" "$(GPU_W_FEATURES)" "$(FILECHECK)" "$(ROCM_PATH)"
+RUNCMD       = ./run_and_check.sh "$(AOMP)" "$(GPU_W_FEATURES)" "$(ROCM_PATH)"
 
 include ../Makefile.rules
 
 clean::
 	rm -f mod.co builder builder.*.host-* builder.*.hip-amdgcn-amd-amdhsa--* \
-	      *.profraw merged.profdata
+	      *.profraw merged.profdata devcov-profile.txt DEVCOV-SKIPPED.txt

test/smoke-dev/HipDevCov/run_and_check.sh

@@ -4,40 +4,50 @@
 # Args (passed from the Makefile):
 #   $1 = AOMP        (LLVM dir containing bin/clang, llvm-profdata, llvm-objdump)
 #   $2 = GPU arch    (e.g. gfx90a, may include :features)
-#   $3 = FILECHECK   (path to FileCheck)
-#   $4 = ROCM_PATH   (HIP/ROCm install for --rocm-path and libamdhip64)
+#   $3 = ROCM_PATH   (HIP/ROCm install for --rocm-path and libamdhip64)
 #
-# Returns 0 only if BOTH the host-shadow kernel and the module-only kernel are
-# present in the merged profile (the latter requires the HSA drain).
+# Proves the HSA-introspection drain captured device coverage that the HIP
+# host-shadow drain cannot see. Verifies, against the actual merged profile:
+#   * the HSA drain produced its own profraw (uniquely named *.hsa<N>.*.profraw),
+#   * mod_kernel (loaded via hipModuleLoad, no host shadow) is present with a
+#     NONZERO function count -- only possible via the HSA drain,
+#   * host_kernel (host-shadow path) is also present with a nonzero count.
+# The human-readable profile dump is left in devcov-profile.txt for inspection.
 
 set -u
-AOMP="$1"; ARCH="$2"; FILECHECK="$3"; ROCM="$4"
+AOMP="$1"; ARCH="$2"; ROCM="$3"
 CLANG="$AOMP/bin/clang"
 PROFDATA="$AOMP/bin/llvm-profdata"
 OBJDUMP="$AOMP/bin/llvm-objdump"
-COV="$AOMP/bin/llvm-cov"
 
 here="$(cd "$(dirname "$0")" && pwd)"
 cd "$here" || exit 1
 
-# Capability gate: skip cleanly on toolchains that do not ship the device
-# profile runtime / HSA drain yet (so this does not red-fail CI before the
-# feature lands). A real, drain-capable toolchain has the amdgcn device profile
-# RT and the host drain symbol.
+fail() { echo "FAIL HipDevCov: $*"; exit 1; }
+
+# Function count for a front-end-instrumented function in the --counts dump.
+func_count() {
+  awk -v fn="$1" '
+    $0 ~ ("^[ \t]*" fn ":$") { inblk = 1; next }
+    inblk && /Function count:/ { print $3; exit }
+  ' "$2"
+}
+
+# Capability gate: skip cleanly only when the toolchain truly cannot do device
+# PGO (no amdgcn device profile runtime / no host HSA-drain runtime). Leaves a
+# marker so the skip is visible in archived results rather than a silent pass.
 resdir="$("$CLANG" -print-resource-dir 2>/dev/null)"
 devrt="$resdir/lib/amdgcn-amd-amdhsa/libclang_rt.profile.a"
-if [ ! -f "$devrt" ]; then
-  echo "SKIP HipDevCov: no device profile runtime at $devrt"
-  exit 0
-fi
-if ! ls "$resdir"/lib/*/libclang_rt.profile_rocm*.a >/dev/null 2>&1; then
-  echo "SKIP HipDevCov: no host profile_rocm runtime (HSA drain) in toolchain"
+rm -f DEVCOV-SKIPPED.txt
+if [ ! -f "$devrt" ] || ! ls "$resdir"/lib/*/libclang_rt.profile_rocm*.a >/dev/null 2>&1; then
+  echo "no device profile runtime / HSA-drain runtime in $resdir" > DEVCOV-SKIPPED.txt
+  echo "SKIP HipDevCov: $(cat DEVCOV-SKIPPED.txt)"
   exit 0
 fi
 
 set -e
-rm -f ./*.profraw mod.co merged.profdata builder builder.*.host-* \
-      builder.*.hip-amdgcn-amd-amdhsa--* 2>/dev/null || true
+rm -f ./*.profraw mod.co merged.profdata devcov-profile.txt builder \
+      builder.*.host-* builder.*.hip-amdgcn-amd-amdhsa--* 2>/dev/null || true
 
 # 1. Build mod.hip as a full executable so the device link gets the device
 #    profile RT, then extract its device code object as the loadable mod.co.
@@ -47,35 +57,38 @@ rm -f ./*.profraw mod.co merged.profdata builder builder.*.host-* \
 "$OBJDUMP" --offloading builder >/dev/null 2>&1 || true
 shopt -s nullglob
 extracted=(builder*.hip-amdgcn-amd-amdhsa--*gfx*)
-if [ ${#extracted[@]} -eq 0 ]; then
-  echo "FAIL HipDevCov: could not extract device code object from builder"
-  exit 1
-fi
+[ ${#extracted[@]} -gt 0 ] || fail "could not extract device code object from builder"
 cp "${extracted[0]}" mod.co
 
-# 2. main is built by the smoke harness (TESTNAME=main). Build it here too if it
-#    is missing, so the script also works when run standalone.
+# 2. main is built by the harness (TESTNAME=main); build it here if missing.
 if [ ! -x ./main ]; then
   "$CLANG" -x hip --offload-arch="$ARCH" -fno-gpu-rdc \
     -fprofile-instr-generate -fcoverage-mapping --rocm-path="$ROCM" \
     main.hip -o main -L"$ROCM/lib" -lamdhip64 -Wl,-rpath,"$ROCM/lib"
 fi
 
-# 3. Run. Device .profraw files are written to CWD (arch-prefixed for the
-#    host-shadow drain, arch.hsa<N>-prefixed for the HSA drain); the host one
-#    goes to LLVM_PROFILE_FILE.
+# 3. Run with mod.co in CWD.
 rm -f ./*.profraw
 LLVM_PROFILE_FILE="$here/host.profraw" \
   LD_LIBRARY_PATH="$ROCM/lib:${LD_LIBRARY_PATH:-}" \
   ./main
 
-# 4. Merge host + all device profraws and assert both kernels are present.
+# 4. The HSA-introspection drain writes its own, uniquely-named profraw
+#    (<arch>.hsa<N>.<file>.profraw). Its presence is a direct fingerprint that
+#    the HSA pass ran (the host-shadow drain names files <arch>[:feat].<file>).
+hsa=(*.hsa[0-9]*.profraw)
+[ ${#hsa[@]} -gt 0 ] || fail "no HSA-drain profraw (*.hsa<N>.*.profraw) produced"
+echo "HipDevCov: HSA-drain profraw(s): ${hsa[*]}"
+
+# 5. Merge and verify real counts from the dump.
 "$PROFDATA" merge -sparse -o merged.profdata ./*.profraw
-"$PROFDATA" show --all-functions merged.profdata | "$FILECHECK" check.txt
+"$PROFDATA" show --all-functions --counts merged.profdata > devcov-profile.txt
 
-# 5. Sanity (non-fatal): llvm-cov can consume the merged device+host profile.
-#    main's covmap only describes host_kernel/main, so llvm-cov warns about the
-#    module-only function; that is expected, hence non-fatal.
-"$COV" report ./main -instr-profile=merged.profdata >/dev/null 2>&1 || true
+mc="$(func_count mod_kernel devcov-profile.txt)"
+hc="$(func_count _Z11host_kernelPii devcov-profile.txt)"
+[ -n "$mc" ] || fail "mod_kernel absent from merged profile (HSA drain did not capture it)"
+[ -n "$hc" ] || fail "host_kernel absent from merged profile"
+[ "$mc" -gt 0 ] 2>/dev/null || fail "mod_kernel function count is '$mc' (expected > 0)"
+[ "$hc" -gt 0 ] 2>/dev/null || fail "host_kernel function count is '$hc' (expected > 0)"
 
-echo "HipDevCov PASSED"
+echo "HipDevCov PASSED (mod_kernel=$mc via HSA drain, host_kernel=$hc)"

test/smoke-dev/OmpDevCov/Makefile

@@ -18,9 +18,9 @@ LINK_FLAGS   = -Wl,-rpath,$(AOMP)/lib
 
 CC           = $(AOMP)/bin/clang $(VERBOSE)
 
-RUNCMD       = ./run_and_check.sh "$(AOMP)" "$(GPU_W_FEATURES)" "$(FILECHECK)"
+RUNCMD       = ./run_and_check.sh "$(AOMP)" "$(GPU_W_FEATURES)"
 
 include ../Makefile.rules
 
 clean::
-	rm -f omp_pgo omp_pgo_use *.profraw merged.profdata
+	rm -f omp_pgo omp_pgo_use *.profraw merged.profdata devcov-profile.txt DEVCOV-SKIPPED.txt