powerpc64, fix #2054 (#2058), workaround gcc altivec bug

.github/workflows/linux-ppc64-cpu-gcc.yml

@@ -0,0 +1,45 @@
+name: linux-ppc64-cpu-gcc
+on: [push, pull_request]
+jobs:
+  linux-gcc-ppc64le:
+    runs-on: ubuntu-20.04
+    steps:
+    - uses: actions/checkout@v2
+
+    - name: cache-qemu
+      id: cache-qemu
+      uses: actions/cache@v1
+      with:
+        path: qemu-install
+        key: qemu-ppc64le-install-1
+    - name: checkout-qemu
+      if: steps.cache-qemu.outputs.cache-hit != 'true'
+      uses: actions/checkout@v2
+      with:
+        repository: qemu/qemu
+        path: qemu
+        ref: 8746309137ba470d1b2e8f5ce86ac228625db940
+    - name: qemu
+      if: steps.cache-qemu.outputs.cache-hit != 'true'
+      run: |
+        cd qemu
+        ./configure --prefix=install --target-list=ppc64le-linux-user --disable-system
+        make -j2
+        make install
+        cp -r ppc64le-linux-user/install $GITHUB_WORKSPACE/qemu-install
+
+    - name: powerpc64le-gnu-toolchain
+      run: |
+        sudo apt-get update
+        sudo apt-get install g++-powerpc64le-linux-gnu
+
+    - name: configure
+      run: mkdir build && cd build && cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/powerpc64le-linux-gnu.toolchain.cmake -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF ..
+    - name: build
+      run: cmake --build build -j 2
+
+    - name: test
+      run: |
+        export PATH=$GITHUB_WORKSPACE/qemu-install/bin:$PATH
+        cd build
+        TESTS_EXECUTABLE_LOADER=qemu-ppc64le TESTS_EXECUTABLE_LOADER_ARGUMENTS="-L;/usr/powerpc64le-linux-gnu" ctest --output-on-failure -j 2

CMakeLists.txt

@@ -79,6 +79,8 @@ elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "^(mips)")
     set(NCNN_TARGET_ARCH mips)
 elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "^(riscv)")
     set(NCNN_TARGET_ARCH riscv)
+elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "^(powerpc)")
+    set(NCNN_TARGET_ARCH powerpc)
 else()
     set(NCNN_TARGET_ARCH x86)
     option(NCNN_AVX2 "optimize x86 platform with avx2" ON)

src/layer/interp.cpp

@@ -45,6 +45,11 @@ int Interp::load_param(const ParamDict& pd)
     return 0;
 }
 
+#if defined(__GNUC__) && defined(__powerpc__) && defined(__ALTIVEC__)
+// NOTE gcc altivec optimized version produce wrong result
+// so I have to disable vectorize here  --- nihui
+__attribute__((optimize("no-tree-vectorize")))
+#endif
 static void linear_coeffs(int w, int outw, int* xofs, float* alpha)
 {
     double scale = (double)w / outw;

toolchains/powerpc64le-linux-gnu.toolchain.cmake

@@ -0,0 +1,16 @@
+set(CMAKE_SYSTEM_NAME Linux)
+set(CMAKE_SYSTEM_PROCESSOR powerpc64le)
+
+set(CMAKE_C_COMPILER "powerpc64le-linux-gnu-gcc")
+set(CMAKE_CXX_COMPILER "powerpc64le-linux-gnu-g++")
+
+set(CMAKE_FIND_ROOT_PATH_MODE_PROGRAM NEVER)
+set(CMAKE_FIND_ROOT_PATH_MODE_LIBRARY ONLY)
+set(CMAKE_FIND_ROOT_PATH_MODE_INCLUDE ONLY)
+
+set(CMAKE_C_FLAGS "-mcpu=powerpc64le")
+set(CMAKE_CXX_FLAGS "-mcpu=powerpc64le")
+
+# cache flags
+set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS}" CACHE STRING "c flags")
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS}" CACHE STRING "c++ flags")