Merge pull request #107 from issp-center-dev/fix_link_error

Fix HPC build: missing explicit instantiation, and icpx fast-math breaking Inf/NaN handling

Author: yomichi

CMakeLists.txt

@@ -9,6 +9,15 @@ set(CMAKE_CXX_EXTENSIONS OFF)
 
 set(CMAKE_POSITION_INDEPENDENT_CODE ON)
 
+# icpx defaults to -fp-model=fast, whose finite-math-only assumption breaks
+# Inf/NaN handling (e.g. divergent correlation lengths and std::isinf/isnan).
+# Force the precise model; add_compile_options here also covers mptensor.
+# The path match is a fallback for CMake < 3.20, which does not know the
+# IntelLLVM compiler id.
+if(CMAKE_CXX_COMPILER_ID STREQUAL "IntelLLVM" OR CMAKE_CXX_COMPILER MATCHES "icpx")
+    add_compile_options(-fp-model=precise)
+endif()
+
 option(GIT_SUBMODULE "Check submodules during build" ON)
 option(MPTENSOR_ROOT "External directory where mptensor is installed into" OFF)
 option(CPPTOML_ROOT "External directory where cpptoml is download into" OFF)

README.md

@@ -45,6 +45,8 @@ The following tools are required for building TeNeS.
 
 - C++17 compiler
   - For Intel compilers, use the newer `icpx`; the classic `icpc` cannot compile mptensor.
+    `icpx` enables `-fp-model=fast` by default, which breaks Inf/NaN handling; the TeNeS build
+    automatically adds `-fp-model=precise`.
 - CMake (>=3.6.0)
 - BLAS/LAPACK
 

docs/sphinx/en/install.rst

@@ -17,6 +17,8 @@ The following tools are required for building TeNeS.
 1. C++17 compiler
 
    - For Intel compilers, use the newer ``icpx``; the classic ``icpc`` cannot compile mptensor.
+     Note that ``icpx`` enables ``-fp-model=fast`` by default, which breaks Inf/NaN handling;
+     the TeNeS build automatically adds ``-fp-model=precise`` to avoid this.
 
 2. CMake (>=3.6.0)
 3. BLAS and LAPACK

docs/sphinx/ja/install.rst

@@ -17,6 +17,8 @@ TeNeSをコンパイルするには以下のライブラリ・環境が必要で
 1. C++17 compiler
 
    - Intel コンパイラの場合、古い ``icpc`` では mptensor をビルドできないため、新しい ``icpx`` を利用してください。
+     なお、 ``icpx`` はデフォルトで ``-fp-model=fast`` が有効になっており Inf/NaN の扱いが壊れるため、
+     TeNeS のビルドでは自動的に ``-fp-model=precise`` を付加します。
 
 2. CMake (>=3.6.0)
 3. BLAS, LAPACK

src/iTPS/core/ctm_single.cpp

@@ -1057,6 +1057,11 @@ int Calc_CTM_Environment_density(std::vector<tensor> &C1, std::vector<tensor> &C
 
 // template instantiate
 
+template std::vector<real_tensor> Make_single_tensor_density(
+    const std::vector<real_tensor> &Tn);
+template std::vector<complex_tensor> Make_single_tensor_density(
+    const std::vector<complex_tensor> &Tn);
+
 template void Calc_projector_left_block_single(
     const real_tensor &C1, const real_tensor &C4, const real_tensor &eT1,
     const real_tensor &eT6, const real_tensor &eT7, const real_tensor &eT8,