diff --git a/bin/clone_aomp.sh b/bin/clone_aomp.sh index fefddbbef9..2d00cff059 100755 --- a/bin/clone_aomp.sh +++ b/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: $manifest_file" diff --git a/bin/run_composable-kernels.sh b/bin/run_composable-kernels.sh index 0184330f2c..574d17077e 100755 --- a/bin/run_composable-kernels.sh +++ b/bin/run_composable-kernels.sh @@ -282,7 +282,14 @@ if [ ! -d "${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 diff --git a/manifests/aomp_23.0.xml b/manifests/aomp_23.0.xml index e0119381b2..656ab15466 100644 --- a/manifests/aomp_23.0.xml +++ b/manifests/aomp_23.0.xml @@ -1,30 +1,20 @@ - + - - - - - - + + + + - - - - - - - - - - - - + + + + diff --git a/manifests/aompi_23.0.xml b/manifests/aompi_23.0.xml deleted file mode 100644 index 656ab15466..0000000000 --- a/manifests/aompi_23.0.xml +++ /dev/null @@ -1,20 +0,0 @@ - - - - - - - - - - - - - - - - - - - - diff --git a/test/smoke-dev/HipDevCov/Makefile b/test/smoke-dev/HipDevCov/Makefile index 749ada2af2..1fc8db22bb 100644 --- a/test/smoke-dev/HipDevCov/Makefile +++ b/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 diff --git a/test/smoke-dev/HipDevCov/check.txt b/test/smoke-dev/HipDevCov/check.txt deleted file mode 100644 index 5eccb5f9a5..0000000000 --- a/test/smoke-dev/HipDevCov/check.txt +++ /dev/null @@ -1,6 +0,0 @@ -Both kernels must appear in the merged host+device profile. host_kernel is the -host-shadow path; mod_kernel (from the hipModuleLoad'd code object, no host -shadow) can only be present if the HSA-introspection drain collected it. - -CHECK-DAG: host_kernel -CHECK-DAG: mod_kernel diff --git a/test/smoke-dev/HipDevCov/run_and_check.sh b/test/smoke-dev/HipDevCov/run_and_check.sh index 9e72f4f453..3fc0a88b5e 100755 --- a/test/smoke-dev/HipDevCov/run_and_check.sh +++ b/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.*.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,"$ROCM/lib" fi -# 3. Run. Device .profraw files are written to CWD (arch-prefixed for the -# host-shadow drain, arch.hsa-prefixed for the HSA drain); the host one -# goes to LLVM_PROFILE_FILE. +# 3. Run with mod.co in CWD. rm -f ./*.profraw LLVM_PROFILE_FILE="$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 +# (.hsa..profraw). Its presence is a direct fingerprint that +# the HSA pass ran (the host-shadow drain names files [:feat].). +hsa=(*.hsa[0-9]*.profraw) +[ ${#hsa[@]} -gt 0 ] || fail "no HSA-drain profraw (*.hsa.*.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)" diff --git a/test/smoke-dev/OmpDevCov/Makefile b/test/smoke-dev/OmpDevCov/Makefile index de216662ca..2a39e2f3e3 100644 --- a/test/smoke-dev/OmpDevCov/Makefile +++ b/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 diff --git a/test/smoke-dev/OmpDevCov/check.txt b/test/smoke-dev/OmpDevCov/check.txt deleted file mode 100644 index 831fad5558..0000000000 --- a/test/smoke-dev/OmpDevCov/check.txt +++ /dev/null @@ -1,6 +0,0 @@ -The merged profile must contain device-side counters that could only have been -collected via the HSA-introspection drain (OpenMP offload has no host-shadow -drain). The device offload entry and the device function classify both appear. - -CHECK-DAG: __omp_offloading_ -CHECK-DAG: classify diff --git a/test/smoke-dev/OmpDevCov/run_and_check.sh b/test/smoke-dev/OmpDevCov/run_and_check.sh index 424b6a0c24..175f70fb40 100755 --- a/test/smoke-dev/OmpDevCov/run_and_check.sh +++ b/test/smoke-dev/OmpDevCov/run_and_check.sh @@ -4,73 +4,82 @@ # Args (passed from the Makefile): # $1 = AOMP (LLVM dir containing bin/clang, llvm-profdata) # $2 = GPU arch (e.g. gfx90a, may include :features) -# $3 = FILECHECK (path to FileCheck) # -# omp_pgo (the -fprofile-generate binary) is built by the harness. Here we: -# 1. run it -> host default.profraw + device amdgcn-amd-amdhsa.*.profraw -# (the device file is produced ONLY by the HSA drain), -# 2. assert a device profraw was produced and merge, -# 3. FileCheck the merged profile for the device offload function, -# 4. rebuild with -fprofile-use and assert it consumes cleanly (no profile -# mismatch / out-of-date diagnostics) and runs. +# OpenMP offload has no host-shadow drain, so device counters can only reach the +# profile through the HSA-introspection drain. omp_pgo (the -fprofile-generate +# binary) is built by the harness. Here we verify, against the actual merged +# profile, that: +# * the HSA drain produced a device profraw (amdgcn-amd-amdhsa.*.profraw), +# * the device function classify has NONZERO counts on BOTH branches of its +# data-dependent if/else -- real device PGO data, not just a present symbol, +# * -fprofile-use consumes the merged profile cleanly and runs. +# The human-readable profile dump is left in devcov-profile.txt for inspection. set -u -AOMP="$1"; ARCH="$2"; FILECHECK="$3" +AOMP="$1"; ARCH="$2" CLANG="$AOMP/bin/clang" PROFDATA="$AOMP/bin/llvm-profdata" here="$(cd "$(dirname "$0")" && pwd)" cd "$here" || exit 1 -# Capability gate: skip cleanly if the toolchain has no amdgcn device profile -# runtime (i.e. no device PGO / HSA drain yet), so CI is not red before the -# feature lands. +fail() { echo "FAIL OmpDevCov: $*"; exit 1; } +run_env() { LD_LIBRARY_PATH="$AOMP/lib:${LD_LIBRARY_PATH:-}" "$@"; } + +# Capability gate: skip cleanly only when there is no amdgcn device profile +# runtime (no device PGO possible). Leave a marker so the skip is visible. resdir="$("$CLANG" -print-resource-dir 2>/dev/null)" +rm -f DEVCOV-SKIPPED.txt if [ ! -f "$resdir/lib/amdgcn-amd-amdhsa/libclang_rt.profile.a" ]; then - echo "SKIP OmpDevCov: no amdgcn device profile runtime in toolchain" + echo "no amdgcn device profile runtime in $resdir" > DEVCOV-SKIPPED.txt + echo "SKIP OmpDevCov: $(cat DEVCOV-SKIPPED.txt)" exit 0 fi set -e -rm -f ./*.profraw merged.profdata omp_pgo_use 2>/dev/null || true +rm -f ./*.profraw merged.profdata devcov-profile.txt omp_pgo_use 2>/dev/null || true -run_env() { LD_LIBRARY_PATH="$AOMP/lib:${LD_LIBRARY_PATH:-}" "$@"; } - -# 1. Run the generate binary (built by the harness as ./omp_pgo). Build it here -# too if missing, so the script also works standalone. +# 1. Run the generate binary (built by the harness as ./omp_pgo). if [ ! -x ./omp_pgo ]; then "$CLANG" -fopenmp --offload-arch="$ARCH" -fprofile-generate \ omp_pgo.c -o omp_pgo -Wl,-rpath,"$AOMP/lib" fi run_env ./omp_pgo -# 2. The HSA drain must have produced at least one device profraw. +# 2. The HSA drain must have produced a device profraw. shopt -s nullglob dev=(amdgcn-amd-amdhsa*.profraw) -if [ ${#dev[@]} -eq 0 ]; then - echo "FAIL OmpDevCov: no device (amdgcn-amd-amdhsa) profraw -- HSA drain did not fire" - ls -1 ./*.profraw 2>/dev/null || true - exit 1 -fi +[ ${#dev[@]} -gt 0 ] || fail "no device (amdgcn-amd-amdhsa) profraw -- HSA drain did not fire" echo "OmpDevCov: device profraw(s): ${dev[*]}" +# 3. Merge and verify real device counts: classify must have both branches +# exercised on-device (nonzero, nonzero). "$PROFDATA" merge -o merged.profdata ./*.profraw +"$PROFDATA" show --all-functions --counts merged.profdata > devcov-profile.txt -# 3. The merged profile must carry the device offload function counters. -"$PROFDATA" show --all-functions merged.profdata | "$FILECHECK" check.txt +# Pull classify's "Block counts: [a, b, ...]" line and require every counter > 0. +bc="$(awk '/[;:]classify:$/ {f=1; next} f && /Block counts:/ {print; exit}' devcov-profile.txt)" +[ -n "$bc" ] || fail "device function classify absent from merged profile (HSA drain did not capture it)" +nums="$(printf '%s\n' "$bc" | grep -oE '[0-9]+')" +[ -n "$nums" ] || fail "could not parse classify block counts from: $bc" +for n in $nums; do + [ "$n" -gt 0 ] 2>/dev/null || fail "classify has a zero block count ($bc) -- device branch not profiled" +done +echo "OmpDevCov: classify $bc (all branches nonzero)" # 4. -fprofile-use round trip: must consume the profile without a mismatch / # out-of-date diagnostic, then run. use_log="$(mktemp)" "$CLANG" -fopenmp --offload-arch="$ARCH" -fprofile-use=merged.profdata \ omp_pgo.c -o omp_pgo_use -Wl,-rpath,"$AOMP/lib" 2> "$use_log" || { - echo "FAIL OmpDevCov: -fprofile-use build failed"; cat "$use_log"; rm -f "$use_log"; exit 1; } + echo "--- profile-use build log ---"; cat "$use_log"; rm -f "$use_log" + fail "-fprofile-use build failed"; } if grep -Eiq "out of date|profile data may be out of date|no profile data available|mismatch" "$use_log"; then - echo "FAIL OmpDevCov: -fprofile-use reported stale/mismatched profile" - cat "$use_log"; rm -f "$use_log"; exit 1 + echo "--- profile-use build log ---"; cat "$use_log"; rm -f "$use_log" + fail "-fprofile-use reported a stale/mismatched profile" fi rm -f "$use_log" run_env ./omp_pgo_use -echo "OmpDevCov PASSED" +echo "OmpDevCov PASSED (classify branches profiled on device, -fprofile-use round trip clean)" diff --git a/test/smoke/helloworld%n/Makefile b/test/smoke/helloworld%n/Makefile index a1a8198587..51d5ebc578 100644 --- a/test/smoke/helloworld%n/Makefile +++ b/test/smoke/helloworld%n/Makefile @@ -5,6 +5,11 @@ TESTSRC_MAIN = helloworld.c TESTSRC_AUX = TESTSRC_ALL = $(TESTSRC_MAIN) $(TESTSRC_AUX) +# The %n printf format does not work with libc printf so force emissary printf +ifeq ($(shell echo 'int main() { return 0; }' | $(CC) -fuse-emissary-print -xc - 2>/dev/null && echo true), true) + EXTRA_CFLAGS=-fuse-emissary-print +endif + CLANG ?= clang OMP_BIN = $(AOMP)/bin/$(CLANG) CC = $(OMP_BIN) $(VERBOSE)