[DeepLearning] Add openmp support

[xlinsist]

Add openmp support for batchmatmul op and bert, tinyllama model.

---

To cross-compile libomp.so on RV:

STEP ONE: modify the corresponding part in llvm/openmp/runtime/src/CMakeLists.txt like this:

libomp_get_libflags(LIBOMP_CONFIGURED_LIBFLAGS)
# Build libomp library. Add LLVMSupport dependency if building in-tree with libomptarget profiling enabled.
if(OPENMP_STANDALONE_BUILD OR (NOT OPENMP_ENABLE_LIBOMP_PROFILING))
add_library(omp ${LIBOMP_LIBRARY_KIND} ${LIBOMP_SOURCE_FILES})
target_compile_options(omp PRIVATE ${CMAKE_C_FLAGS}) # new added
# Linking command will include libraries in LIBOMP_CONFIGURED_LIBFLAGS
target_link_libraries(omp ${LIBOMP_CONFIGURED_LIBFLAGS} ${LIBOMP_DL_LIBS})

STEP TWO: use the following configuration to cross compile openmp with clang.

export BUILD_LOCAL_LLVM_DIR=/root/intern/buddy-mlir/llvm/build &&
export BUILD_RISCV_GNU_TOOLCHAIN_DIR=/root/intern/buddy-mlir/thirdparty/riscv-gnu-toolchain/install &&
export RISCV_GNU_TOOLCHAIN_SYSROOT_DIR=${BUILD_RISCV_GNU_TOOLCHAIN_DIR}/sysroot/

cmake -G Ninja ../llvm \
    -DLLVM_ENABLE_PROJECTS="clang;openmp" \
    -DLLVM_TARGETS_TO_BUILD="RISCV" \
    -DCMAKE_SYSTEM_NAME=Linux \
    -DCMAKE_VERBOSE_MAKEFILE=ON \
    -DCMAKE_C_COMPILER=${BUILD_LOCAL_LLVM_DIR}/bin/clang \
    -DCMAKE_CXX_COMPILER=${BUILD_LOCAL_LLVM_DIR}/bin/clang++ \
    -DCMAKE_C_FLAGS="--target=riscv64-unknown-linux-gnu --sysroot=${RISCV_GNU_TOOLCHAIN_SYSROOT_DIR} --gcc-toolchain=${BUILD_RISCV_GNU_TOOLCHAIN_DIR}" \
    -DCMAKE_CXX_FLAGS="--target=riscv64-unknown-linux-gnu --sysroot=${RISCV_GNU_TOOLCHAIN_SYSROOT_DIR} --gcc-toolchain=${BUILD_RISCV_GNU_TOOLCHAIN_DIR}" \
    -DLLVM_TABLEGEN=${BUILD_LOCAL_LLVM_DIR}/bin/llvm-tblgen \
    -DCLANG_TABLEGEN=${BUILD_LOCAL_LLVM_DIR}/bin/clang-tblgen \
    -DLLVM_DEFAULT_TARGET_TRIPLE=riscv64-unknown-linux-gnu \
    -DLLVM_TARGET_ARCH=RISCV64 \
    -DCMAKE_BUILD_TYPE=Release \
    -DOPENMP_ENABLE_LIBOMPTARGET=Off \
	-DLIBOMP_ARCH=riscv64 \
	-DLIBOMP_ENABLE_SHARED=On

ninja  clang lli omp

TinyLlama performance output:

// X86 platform
----------------------------------------------------------------------------
Benchmark                                  Time             CPU   Iterations
----------------------------------------------------------------------------
DL_MODEL_TINYLLAMA/scalar             237142 ms       237115 ms            1
DL_MODEL_TINYLLAMA/matmul_opt          15745 ms        15742 ms            1
DL_MODEL_TINYLLAMA/matmul_opt_omp      13696 ms        13594 ms            1

// RV platform
----------------------------------------------------------------------------
Benchmark                                  Time             CPU   Iterations
----------------------------------------------------------------------------
DL_MODEL_TINYLLAMA/scalar             >1h
DL_MODEL_TINYLLAMA/matmul_opt         342555 ms       342539 ms            1
DL_MODEL_TINYLLAMA/matmul_opt_omp     245805 ms       245794 ms            1

[xlinsist]

Move the original variable setups above.

[xlinsist]

Add openmp support for Bert.

[xlinsist]

Relate commits have been merged, closing it.