merge aomp-dev into amd-staging

bin/run_miniQMC.sh

@@ -23,8 +23,12 @@ export AOMP_USE_CCACHE=0
 # Control how many OpenMP threads are used by MiniQMCPack
 : "${MQMC_OMP_NUM_THREADS:=32}"
 
+if [ -z "$AOMP_GPU" ]; then
+  echo "Error: Set AOMP_GPU to the target GPU architecture (e.g., gfx90a)"
+  exit 1
+fi
+
 export PATH=$AOMP/bin:$PATH
-#export PATH=/home/janplehr/rocm/trunk/bin:$PATH
 
 # We export all these paths, so they will be picked-up in the CMake command.
 # For libraries that AOMP provides (hsa-runtime, hip, comgr, device-libs), use AOMP.
@@ -43,17 +47,23 @@ export rocsolver_DIR=${ROCM}/lib/cmake/rocsolver/
 : "${MQMC_BUILD_PREFIX:=$AOMP_REPOS_TEST/miniqmc_build}"
 # Set the default build directory name
 : "${MQMC_BUILD_DIR:=${MQMC_BUILD_PREFIX}/build_aomp_clang}"
-# Path to the miniqmc source directory
-: "${MQMC_SOURCE_DIR:=$AOMP_REPOS_TEST/miniqmc_src}"
 # how many threads should be used for building miniqmc
 : "${MQMC_NUM_BUILD_PROCS:=32}"
 # We pin the version by default, so we have only AOMP as moving target
 : "${MQMC_GIT_TAG:=9d9d7d3}"
 
+# Path to the miniqmc source directory
+: "${MQMC_SOURCE_DIR:=$AOMP_REPOS_TEST/miniqmc_src}"
 
 if [ ! -d "$MQMC_SOURCE_DIR" ]; then
-  git clone https://github.com/QMCPACK/miniqmc.git "$MQMC_SOURCE_DIR"
-  git checkout "${MQMC_GIT_TAG}"
+  git clone https://github.com/ye-luo/miniqmc.git "$MQMC_SOURCE_DIR"
+  pushd "$MQMC_SOURCE_DIR" || exit
+  git checkout OMP_offload
+  popd || exit
+else
+  pushd "$MQMC_SOURCE_DIR" || exit
+  git pull
+  popd || exit
 fi
 
 rm -rf "${MQMC_BUILD_DIR}"
@@ -63,8 +73,8 @@ rm -rf "${MQMC_BUILD_DIR}"
 cmake -B "${MQMC_BUILD_DIR}" -S "${MQMC_SOURCE_DIR}" \
   -DCMAKE_PREFIX_PATH="${AOMP}/lib/cmake;${ROCM}/lib/cmake" \
   -DCMAKE_CXX_COMPILER=clang++ \
-  -DENABLE_OFFLOAD=ON \
-  -DQMC_ENABLE_ROCM=ON \
+  -DQMC_GPU="openmp" \
+  -DQMC_GPU_ARCHS="${AOMP_GPU}" \
   -DCMAKE_CXX_FLAGS='-fopenmp-assume-no-nested-parallelism -DCUDART_VERSION=10000 -DcudaMemoryTypeManaged=hipMemoryTypeManaged ' \
   -DAMDGPU_DISABLE_HOST_DEVMEM=ON \
   -DCMAKE_VERBOSE_MAKEFILE=ON
@@ -81,13 +91,26 @@ echo "Running Tests"
 # Ensure AOMP runtime libraries are found before /opt/rocm libraries to avoid ABI mismatches
 export LD_LIBRARY_PATH="${AOMP}/lib:${AOMP}/lib/llvm/lib:${LD_LIBRARY_PATH}"
 
-echo "OMP_NUM_THREADS=${MQMC_OMP_NUM_THREADS} ${MQMC_BUILD_DIR}/bin/check_spo_batched_reduction -n 10"
-OMP_NUM_THREADS=${MQMC_OMP_NUM_THREADS} "${MQMC_BUILD_DIR}"/bin/check_spo_batched_reduction -n 10
+# We intentionally continue running even if some binaries are missing.
+if [ ! -f "${MQMC_BUILD_DIR}/bin/check_spo_batched_reduction" ]; then
+  echo "Error: check_spo_batched_reduction binary not found in ${MQMC_BUILD_DIR}/bin"
+else
+  echo "OMP_NUM_THREADS=${MQMC_OMP_NUM_THREADS} ${MQMC_BUILD_DIR}/bin/check_spo_batched_reduction -n 10"
+  OMP_NUM_THREADS=${MQMC_OMP_NUM_THREADS} "${MQMC_BUILD_DIR}"/bin/check_spo_batched_reduction -n 10
+fi
 
-echo ""
-echo "OMP_NUM_THREADS=${MQMC_OMP_NUM_THREADS} ${MQMC_BUILD_DIR}/bin/miniqmc"
-OMP_NUM_THREADS=${MQMC_OMP_NUM_THREADS} "${MQMC_BUILD_DIR}"/bin/miniqmc -v
+if [ ! -f "${MQMC_BUILD_DIR}/bin/miniqmc" ]; then
+  echo "Error: miniqmc binary not found in ${MQMC_BUILD_DIR}/bin"
+else
+  echo ""
+  echo "OMP_NUM_THREADS=${MQMC_OMP_NUM_THREADS} ${MQMC_BUILD_DIR}/bin/miniqmc"
+  OMP_NUM_THREADS=${MQMC_OMP_NUM_THREADS} "${MQMC_BUILD_DIR}"/bin/miniqmc -v
+fi
 
-echo ""
-echo "OMP_NUM_THREADS=${MQMC_OMP_NUM_THREADS} ${MQMC_BUILD_DIR}/bin/check_spo -n 10"
-OMP_NUM_THREADS=${MQMC_OMP_NUM_THREADS} "${MQMC_BUILD_DIR}"/bin/check_spo -n 10 -v
+if [ ! -f "${MQMC_BUILD_DIR}/bin/check_spo" ]; then
+  echo "Error: check_spo binary not found in ${MQMC_BUILD_DIR}/bin"
+else
+  echo ""
+  echo "OMP_NUM_THREADS=${MQMC_OMP_NUM_THREADS} ${MQMC_BUILD_DIR}/bin/check_spo -n 10"
+  OMP_NUM_THREADS=${MQMC_OMP_NUM_THREADS} "${MQMC_BUILD_DIR}"/bin/check_spo -n 10 -v
+fi

test/fortran-map-tests/initialized-common-block-mapping.f90

@@ -0,0 +1,110 @@
+! Copyright © Advanced Micro Devices, Inc., or its affiliates.
+!
+! SPDX-License-Identifier:  MIT
+
+! Common blocks can have a special "edge-case" where they are initialized
+! common blocks, which basically just means we give them some values on
+! initialize. This, however, means the LLVM-IR for a common block is actually
+! generated differently, the common block goes from a byte array into a typed
+! tuple, so we need to be sure this still lowers and handles fine.
+
+block data init_blocks
+  implicit none
+
+  integer :: var_a
+  real :: var_b(5)
+  common /cblock_a/ var_a, var_b
+
+  integer :: var_c
+  real :: var_d(5)
+  common /cblock_b/ var_c, var_d
+
+  data var_a / 100 /
+  data var_b / 1.0, 2.0, 3.0, 4.0, 5.0 /
+
+  data var_c / 200 /
+  data var_d / 10.0, 20.0, 30.0, 40.0, 50.0 /
+
+end block data
+
+program prog_a
+  implicit none
+
+  integer :: var_a
+  real :: var_b(5)
+  common /cblock_a/ var_a, var_b
+  !$omp declare target link(/cblock_a/)
+
+  integer :: var_c
+  real :: var_d(5)
+  common /cblock_b/ var_c, var_d
+
+  integer :: i
+
+  print *, "=== Initial values (before target region) ==="
+  print *, ""
+  print *, "cblock_a (with declare target link):"
+  print *, "  var_a =", var_a
+  print *, "  var_b =", var_b
+  print *, ""
+  print *, "cblock_b (without declare target):"
+  print *, "  var_c =", var_c
+  print *, "  var_d =", var_d
+  print *, ""
+
+  !$omp target map(tofrom: /cblock_a/, /cblock_b/)
+    var_a = var_a + 1000
+    do i = 1, 5
+      var_b(i) = var_b(i) * 2.0
+    end do
+
+    var_c = var_c + 2000
+    do i = 1, 5
+      var_d(i) = var_d(i) * 3.0
+    end do
+  !$omp end target
+
+  print *, "=== Final values (after target region) ==="
+  print *, ""
+  print *, "cblock_a (with declare target link):"
+  print *, "  var_a =", var_a
+  print *, "  var_b =", var_b
+  print *, ""
+  print *, "cblock_b (without declare target):"
+  print *, "  var_c =", var_c
+  print *, "  var_d =", var_d
+  print *, ""
+
+  print *, "=== Validating results ==="
+
+  if (var_a /= 1100) then
+    print *, "FAILED: var_a expected 1100, got", var_a
+    print *, "======= FORTRAN Test Failed! ======="
+    stop 1
+  end if
+
+  if (var_b(1) /= 2.0 .or. var_b(2) /= 4.0 .or. &
+      var_b(3) /= 6.0 .or. var_b(4) /= 8.0 .or. &
+      var_b(5) /= 10.0) then
+    print *, "FAILED: var_b expected [2.0, 4.0, 6.0, 8.0, 10.0], got", var_b
+    print *, "======= FORTRAN Test Failed! ======="
+    stop 1
+  end if
+
+  if (var_c /= 2200) then
+    print *, "FAILED: var_c expected 2200, got", var_c
+    print *, "======= FORTRAN Test Failed! ======="
+    stop 1
+  end if
+
+  if (var_d(1) /= 30.0 .or. var_d(2) /= 60.0 .or. &
+      var_d(3) /= 90.0 .or. var_d(4) /= 120.0 .or. &
+      var_d(5) /= 150.0) then
+    print *, "FAILED: var_d expected [30.0, 60.0, 90.0, 120.0, 150.0], got", var_d
+    print *, "======= FORTRAN Test Failed! ======="
+    stop 1
+  end if
+
+  print *, "======= FORTRAN Test Passed! ======="
+
+end program prog_a

test/fortran-map-tests/size-zero-presence-check.f90

@@ -0,0 +1,44 @@
+! Copyright © Advanced Micro Devices, Inc., or its affiliates.
+!
+! SPDX-License-Identifier:  MIT
+
+program prog_a
+    implicit none
+    integer :: var_a
+    integer :: var_b
+    real, allocatable :: var_c(:)
+    integer :: var_d
+
+    var_a = 10
+    var_b = 9
+
+    ! Make it size 0, the main purpose of this test is to make sure
+    ! the presence check on a size 0 array returns true and lets us
+    ! pass to the end of the program.
+    allocate(var_c(var_a:var_b))
+
+    do var_d = var_a, var_b
+        var_c(var_d) = real(var_d)
+    end do
+
+    !$omp target enter data map(to: var_c)
+
+    !$omp target teams distribute parallel do map(present,alloc: var_c)
+    do var_d = var_a, var_b
+        var_c(var_d) = var_c(var_d) * 2.0
+    end do
+    !$omp end target teams distribute parallel do
+
+    !$omp target exit data map(from: var_c)
+
+    do var_d = var_a, var_b
+        if (var_c(var_d) /= real(var_d) * 2.0) then
+            print *, "======= FORTRAN Test Failed! ======="
+            stop 1
+        end if
+    end do
+
+    deallocate(var_c)
+
+    print*, "======= FORTRAN Test Passed! ======="
+end program