Home Original page

Android ARM64: xtensor auto Deduction Fails in Release Build - Expression Template Optimization Bug

On Android arm64-v8a,

template <typename T>
xt::xtensor<T, 2> logSoftmax(const xt::xtensor<T, 2> &matrix)
{
    xt::xtensor<T, 2> maxVals = xt::amax(matrix, {1}, xt::keep_dims);
    auto shifted = matrix - maxVals;
    auto expVals = xt::exp(shifted);
    auto sumExp = xt::sum(expVals, {1}, xt::keep_dims);
    return shifted - xt::log(sumExp);
}

logSoftmax() cannot produce right results, sometimes crashes.

template <typename T>
xt::xtensor<T, 2> logSoftmax2(const xt::xtensor<T, 2> &matrix)
{
    xt::xtensor<T, 2> maxVals = xt::amax(matrix, {1}, xt::keep_dims);
    xt::xtensor<T, 2> shifted = matrix - maxVals;
    xt::xtensor<T, 2> expVals = xt::exp(shifted);
    xt::xtensor<T, 2> sumExp = xt::sum(expVals, {1}, xt::keep_dims);
    return shifted - xt::log(sumExp);
}

but logSoftmax2() works properly.

CMakeLists.txt

...
add_library(dp dp.cpp)
target_link_libraries(dp PRIVATE 
    xtensor xtensor::optimize xtensor::use_xsimd xtensor-blas BLAS::BLAS)
if(ANDROID)
    target_link_options(dp PUBLIC -Wl,-z,max-page-size=16384)
else()
    target_compile_options(dp PUBLIC -march=native)
endif()
...

built with -DCMAKE_BUILD_TYPE=Release.

Note: when built with -DCMAKE_BUILD_TYPE=RelWithDebInfo, logSoftmax() works properly too.