wasm build

block-hczhai · Oct 28, 2024 · 104ebab · 104ebab
1 parent 44ed696
commit 104ebab
Show file tree

Hide file tree

Showing 19 changed files with 534 additions and 345 deletions.
diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml
@@ -18,7 +18,7 @@ permissions:
 
 jobs:
   test:
-    name: test ${{ matrix.python }} ${{ matrix.os }} ${{ matrix.parallel }}
+    name: test ${{ matrix.python }} ${{ matrix.os }} ${{ matrix.parallel }} ${{ matrix.plat }}
     runs-on: ${{ matrix.os }}
 
     strategy:
@@ -34,6 +34,11 @@ jobs:
             parallel: mpi
           - os: windows-2022
             parallel: mpi
+        include:
+          - os: ubuntu-20.04
+            python: '3.12'
+            plat: wasm
+            parallel: serial
       fail-fast: false
 
     steps:
@@ -43,13 +48,32 @@ jobs:
           python-version: ${{ matrix.python }}
 
       - name: install requirements (linux / macos)
-        if: matrix.os == 'macos-13' || matrix.os == 'ubuntu-20.04'
+        if: (matrix.os == 'macos-13' || matrix.os == 'ubuntu-20.04') && matrix.plat == 'x86_64'
         env:
           MKLROOT: ~/.local
         run: |
           python -m pip install pip build twine setuptools --upgrade
           python -m pip install mkl==2021.4 mkl-include intel-openmp numpy 'cmake>=3.19' pybind11==2.12.0
 
+      - name: install requirements (linux emscripten)
+        if: matrix.os == 'ubuntu-20.04' && matrix.plat == 'wasm'
+        run: |
+          python -m pip install pip build twine setuptools --upgrade
+          python -m pip install pyodide-build numpy 'cmake>=3.19' pybind11==2.12.0
+
+      - name: setup emscripten (linux emscripten)
+        if: matrix.os == 'ubuntu-20.04' && matrix.plat == 'wasm'
+        run: |
+          git clone -b main https://github.com/emscripten-core/emsdk.git
+          cd emsdk
+          PYODIDE_EMSCRIPTEN_VERSION=$(pyodide config get emscripten_version)
+          echo $PYODIDE_EMSCRIPTEN_VERSION
+          ./emsdk install ${PYODIDE_EMSCRIPTEN_VERSION}
+          ./emsdk activate ${PYODIDE_EMSCRIPTEN_VERSION}
+          source ./emsdk_env.sh
+          cd ..
+          em++ -v
+
       - name: install requirements (linux / macos)
         if: matrix.os == 'macos-14'
         run: |
@@ -68,6 +92,25 @@ jobs:
           python -m pip install pip build twine setuptools --upgrade
           python -m pip install intel-openmp==2024.1.0 numpy 'cmake>=3.19' pybind11==2.12.0
 
+      - name: Download OpenBLAS (linux emscripten)
+        if: matrix.os == 'ubuntu-20.04' && matrix.plat == 'wasm'
+        run: |
+          wget -q https://github.com/pyodide/pyodide/releases/download/0.26.3/xbuildenv-0.26.3.tar.bz2
+          tar xjf xbuildenv-0.26.3.tar.bz2
+
+      - name: build gtest (linux emscripten)
+        if: matrix.os == 'ubuntu-20.04' && matrix.plat == 'wasm'
+        env:
+          GTESTROOT: ~/gtest
+        run: |
+          source emsdk/emsdk_env.sh
+          git clone -b release-1.8.0 https://github.com/google/googletest
+          cd googletest
+          export GTEST_ROOT=$PWD/googletest/install
+          CC=emcc CXX=em++ emcmake cmake CMakeLists.txt -DCMAKE_INSTALL_PREFIX=${GTESTROOT} -Dgtest_disable_pthreads=ON
+          emmake make -j 4
+          emmake make install
+
       - name: Download OpenBLAS (windows)
         if: matrix.os == 'windows-2022'
         env:
@@ -83,14 +126,14 @@ jobs:
           GTEST_ROOT: D:\a\gtest
         run: |
           Get-WmiObject -Class Win32_Processor
-          git clone  -b release-1.8.0 https://github.com/google/googletest.git
+          git clone -b release-1.8.0 https://github.com/google/googletest.git
           cd googletest
           cmake -G "Visual Studio 17 2022" -S . -B . -DCMAKE_CXX_FLAGS="-D_SILENCE_TR1_NAMESPACE_DEPRECATION_WARNING" -Dgtest_force_shared_crt=ON
           cmake --build . -j 2 --config Release
           cmake --install . --prefix $env:GTEST_ROOT
 
       - name: build gtest (linux)
-        if: matrix.os == 'ubuntu-20.04'
+        if: matrix.os == 'ubuntu-20.04' && matrix.plat == 'x86_64'
         env:
           CC: gcc-9
           CXX: g++-9
@@ -103,7 +146,7 @@ jobs:
           sudo cp lib/*.a /usr/lib
 
       - name: fix mkl for amd cpu
-        if: matrix.os == 'ubuntu-20.04'
+        if: matrix.os == 'ubuntu-20.04' && matrix.plat == 'x86_64'
         env:
           CC: gcc-9
           CXX: g++-9
@@ -146,7 +189,9 @@ jobs:
           brew install libomp
 
       - name: build test (serial-pytest, linux | macos)
-        if: (matrix.parallel == 'serial-pytest' || matrix.parallel == 'main-test') && (matrix.os == 'ubuntu-20.04' || matrix.os == 'macos-13')
+        if: |
+          (matrix.parallel == 'serial-pytest' || matrix.parallel == 'main-test') &&
+          (matrix.os == 'ubuntu-20.04' || matrix.os == 'macos-13') && matrix.plat == 'x86_64'
         env:
           MKLROOT: ~/.local
           MACOSX_DEPLOYMENT_TARGET: '11.0'
@@ -188,7 +233,7 @@ jobs:
           cmake --build . -j 2 --config Release -- '/v:d'
 
       - name: build test (any-symm-pytest, linux | macos)
-        if: matrix.parallel == 'any-symm-pytest' && (matrix.os == 'ubuntu-20.04' || matrix.os == 'macos-13')
+        if: matrix.parallel == 'any-symm-pytest' && (matrix.os == 'ubuntu-20.04' || matrix.os == 'macos-13') && matrix.plat == 'x86_64'
         env:
           MKLROOT: ~/.local
           MACOSX_DEPLOYMENT_TARGET: '11.0'
@@ -241,8 +286,23 @@ jobs:
           cmake -G "Visual Studio 17 2022" -S .. -B . -DUSE_MKL=OFF -DBUILD_TEST=ON -DLARGE_BOND=ON -DUSE_COMPLEX=ON -DUSE_SG=ON
           cmake --build . -j 2 --config Release -- '/v:d'
 
+      - name: build test (serial, linux emscripten)
+        if: matrix.parallel == 'serial' && matrix.os == 'ubuntu-20.04' && matrix.plat == 'wasm'
+        env:
+          GTESTROOT: ~/gtest
+        run: |
+          source emsdk/emsdk_env.sh
+          export BLAS_ROOT=$PWD/xbuildenv/pyodide-root/packages/.libs
+          mkdir build_test
+          cd build_test
+          cmake --version
+          CC=emcc CXX=em++ cmake .. -DUSE_MKL=OFF -DARCH_EMSCRIPTEN=ON -DF77UNDERSCORE=ON -DBUILD_TEST=ON \
+              -DGTEST_ROOT=${GTESTROOT} -DLARGE_BOND=ON -DUSE_COMPLEX=OFF -DUSE_DMRG=ON \
+              -DSIMPLE_TEST=ON -DCMAKE_BUILD_TYPE=Release -DOMP_LIB=NONE
+          emmake make -j 2
+
       - name: build test (serial, linux)
-        if: matrix.parallel == 'serial' && matrix.os == 'ubuntu-20.04'
+        if: matrix.parallel == 'serial' && matrix.os == 'ubuntu-20.04' && matrix.plat == 'x86_64'
         env:
           MKLROOT: ~/.local
           CC: gcc-9
@@ -296,8 +356,15 @@ jobs:
           & $dumpbin /DEPENDENTS .\block2_tests.exe
           .\block2_tests.exe
 
+      - name: run test (serial, linux emscripten)
+        if: matrix.parallel == 'serial' && matrix.os == 'ubuntu-20.04' && matrix.plat == 'wasm'
+        run: |
+          source emsdk/emsdk_env.sh
+          cd build_test
+          node block2_tests.js
+
       - name: run test (serial, linux)
-        if: matrix.parallel == 'serial' && matrix.os == 'ubuntu-20.04'
+        if: matrix.parallel == 'serial' && matrix.os == 'ubuntu-20.04' && matrix.plat == 'x86_64'
         run: |
           export CPUTYPE=$(lscpu | grep 'Vendor ID' | awk '{print $3}')
           if [ "$CPUTYPE" = "AuthenticAMD" ]; then

diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -17,6 +17,7 @@ OPTION(USE_SG "General Spin Symmetry" OFF)
 OPTION(USE_SINGLE_PREC "Single Precision" OFF)
 OPTION(USE_BLIS "BLIS" OFF)
 OPTION(ARCH_ARM64 "MacOS arch arm64" OFF)
+OPTION(ARCH_EMSCRIPTEN "Build for WebAssembly" OFF)
 OPTION(APPLE_ACC_SINGLE_PREC "Fix Apple Accelerate single prec" ON)
 OPTION(BUILD_LIB "Build python block2.so" OFF)
 OPTION(BUILD_CLIB "Build C++ block2.so" OFF)
@@ -419,6 +420,11 @@ IF (CMAKE_CXX_COMPILER_ID STREQUAL "MSVC")
         FIND_LIBRARY(OMP_LIB_NAME NAMES iomp5md PATHS $ENV{OMPROOT} $ENV{OMPROOT}/lib $ENV{OMPROOT}/lib/intel64_win)
         SET(MKL_OMP_VALUE 2)
     ENDIF()
+ELSEIF ("${OMP_LIB}" STREQUAL "NONE")
+    SET(OMP_LIB_NAME "")
+    SET(OMP_FLAG "")
+    UNSET(OMP_LIB_NAME)
+    SET(MKL_OMP_VALUE 4)
 ELSEIF (WIN32)
     SET(OMP_LIB_NAME "")
     SET(PTHREAD "")
@@ -520,6 +526,13 @@ IF ((NOT APPLE) AND (NOT WIN32) AND (CMAKE_CXX_COMPILER_ID STREQUAL "IntelLLVM")
     SET(OPT_FLAG ${OPT_FLAG} -Wno-unused-command-line-argument -Wno-tautological-constant-compare)
 ENDIF()
 
+IF (${ARCH_EMSCRIPTEN})
+    SET(CMAKE_EXECUTABLE_SUFFIX ".js")
+    SET(WASM_LINK_FLAGS "-sALLOW_MEMORY_GROWTH -sSTACK_SIZE=64mb -sSTACK_OVERFLOW_CHECK --preload-file ${CMAKE_SOURCE_DIR}/data@data")
+ELSE()
+    SET(WASM_LINK_FLAGS "")
+ENDIF()
+
 IF (${ARCH_ARM64})
     SET(ARCH_FLAG -arch arm64)
     SET(ARCH_LINK_FLAGS "-arch arm64")
@@ -769,7 +782,7 @@ FOREACH (target ${TARGETS})
     SET_TARGET_PROPERTIES(${target} PROPERTIES POSITION_INDEPENDENT_CODE 1)
     TARGET_LINK_LIBRARIES(${target} PRIVATE ${OMP_LIB_NAME} ${PTHREAD})
     TARGET_LINK_LIBRARIES(${target} PRIVATE ${PTHREAD} ${LAPACK_LIBRARIES} ${BLAS_LIBRARIES} ${MKL_LIBS} ${MPI_LIBS} ${TBB_LIBS} ${BLIS_LIBS})
-    SET_TARGET_PROPERTIES(${target} PROPERTIES LINK_FLAGS "${ARCH_LINK_FLAGS} ${MPI_LINK_FLAGS}")
+    SET_TARGET_PROPERTIES(${target} PROPERTIES LINK_FLAGS "${ARCH_LINK_FLAGS} ${WASM_LINK_FLAGS} ${MPI_LINK_FLAGS}")
 
     TARGET_INCLUDE_DIRECTORIES(${target} PUBLIC ${PYTHON_INCLUDE_DIRS} ${PYBIND_INCLUDE_DIRS}
         ${MKL_INCLUDE_DIR} ${MPI_INCLUDE_DIR} ${TBB_INCLUDE_DIR} ${BLIS_INCLUDE_DIR})
@@ -787,13 +800,20 @@ IF (${BUILD_TEST})
     INCLUDE_DIRECTORIES(${GTEST_INCLUDE_DIRS})
 
     IF (${SIMPLE_TEST})
-        FILE(GLOB TSRCS unit_test/test_wick_*.cpp unit_test/test_npdm_*.cpp unit_test/test_dmrg_*.cpp)
+        FILE(GLOB TSRCS unit_test/test_batch_gemm.cpp unit_test/test_matrix.cpp
+            unit_test/test_cg.cpp unit_test/test_dmrg_n2_sto3g.cpp
+            unit_test/test_dmrg_1site_n2_sto3g.cpp unit_test/test_dmrg_sa_1site_n2_sto3g.cpp
+            unit_test/test_dmrg_sa_n2_sto3g.cpp unit_test/test_dmrg_sa_tto_n2_sto3g.cpp
+            unit_test/test_csf_space.cpp unit_test/test_det_n2_sto3g.cpp
+            unit_test/test_wick_ccsd.cpp unit_test/test_wick_ghf.cpp 
+            unit_test/test_wick_ic_nevpt2.cpp unit_test/test_wick_sc_nevpt2.cpp
+            unit_test/test_wick_uga_ccsd.cpp unit_test/test_npdm_*.cpp)
     ELSE()
         FILE(GLOB TSRCS unit_test/test_*.cpp)
     ENDIF()
     MESSAGE(STATUS "TSRCS = ${TSRCS}")
 
-    IF ((NOT APPLE) AND (NOT WIN32))
+    IF ((NOT APPLE) AND (NOT WIN32) AND (NOT ${ARCH_EMSCRIPTEN}))
         SET(OPT_FLAG ${OPT_FLAG} -fcompare-debug-second)
     ENDIF()
 
@@ -804,7 +824,7 @@ IF (${BUILD_TEST})
         ${TMPL_FLAG} ${BOND_FLAG} ${SCI_FLAG} ${CORE_FLAG} ${DMRG_FLAG} ${BIG_SITE_FLAG}
         ${SP_DMRG_FLAG} ${IC_FLAG} ${KSYMM_FLAG} ${SG_FLAG} ${COMPLEX_FLAG} ${SINGLE_PREC_FLAG} ${TBB_FLAG} ${BLIS_FLAG}
         ${SU2SZ_FLAG} ${SANY_FLAG})
-    SET_TARGET_PROPERTIES(${PROJECT_NAME}_tests PROPERTIES LINK_FLAGS "${ARCH_LINK_FLAGS} ${MPI_LINK_FLAGS}")
+    SET_TARGET_PROPERTIES(${PROJECT_NAME}_tests PROPERTIES LINK_FLAGS "${ARCH_LINK_FLAGS} ${WASM_LINK_FLAGS} ${MPI_LINK_FLAGS}")
 
     IF ((NOT APPLE) AND (NOT WIN32))
         TARGET_LINK_LIBRARIES(${PROJECT_NAME}_tests rt)

diff --git a/src/core/allocator.hpp b/src/core/allocator.hpp
@@ -29,9 +29,11 @@
 #include "tbb/scalable_allocator.h"
 #endif
 #include <algorithm>
+#ifndef __EMSCRIPTEN__
 #ifdef __unix__
 #include <execinfo.h>
 #endif
+#endif
 #include <cassert>
 #include <complex>
 #include <cstdint>
@@ -54,13 +56,15 @@ namespace block2 {
  * Not working for non-unix systems.
  */
 inline void print_trace() {
+#ifndef __EMSCRIPTEN__
 #ifdef __unix__
     void *array[32];
     size_t size = backtrace(array, 32);
     char **strings = backtrace_symbols(array, size);
 
     for (size_t i = 0; i < size; i++)
         fprintf(stderr, "%s\n", strings[i]);
+#endif
 #endif
     abort();
 }

diff --git a/src/core/clebsch_gordan.hpp b/src/core/clebsch_gordan.hpp
@@ -107,7 +107,11 @@ struct SU2CG {
             rst = sqrt_fact[t] * sqrt_fact[t - alpha1] * sqrt_fact[t - alpha2] *
                   sqrt_fact[beta1 - t] * sqrt_fact[beta2 - t] *
                   sqrt_fact[beta3 - t];
+#ifndef __EMSCRIPTEN__
             r += factor / (rst * rst);
+#else
+            r += factor / (rst * rst + 1E-33L);
+#endif
         }
         return r;
     }
@@ -139,7 +143,12 @@ struct SU2CG {
                   sqrt_fact[t - alpha3] * sqrt_fact[t - alpha4] *
                   sqrt_fact[beta1 - t] * sqrt_fact[beta2 - t] *
                   sqrt_fact[beta3 - t];
+#ifndef __EMSCRIPTEN__
             r += factor * sqrt_fact[t + 1] * sqrt_fact[t + 1] / (rst * rst);
+#else
+            r += factor * sqrt_fact[t + 1] * sqrt_fact[t + 1] /
+                 (rst * rst + 1E-33L);
+#endif
         }
         return r;
     }
@@ -191,7 +200,11 @@ struct SU2CG {
         for (int s = min_s; s <= max_s; ++s, factor *= rr) {
             rst = sqrt_fact[jps - s] * sqrt_fact[s] * sqrt_fact[jpp - jps + s] *
                   sqrt_fact[jmp - s];
+#ifndef __EMSCRIPTEN__
             r += factor / (rst * rst);
+#else
+            r += factor / (rst * rst + 1E-33L);
+#endif
         }
         return r * sqrt_fact[jpp] * sqrt_fact[jmp] * sqrt_fact[jps] *
                sqrt_fact[jms];