sw_engine: applied OpenMP to improve post-processing performance.

- OpenMp threads will be allocated alongside the TaskSculeduler::threads()
- Performance improved by 2x in a specific animation.
- Disable the thread feature in the Android build test due to system issues with OpenMP compilation.

.github/workflows/build_android.yml

@@ -34,7 +34,7 @@ jobs:
         API: 21
       run: |
         sed -e "s|NDK|$NDK|g" -e "s|HOST_TAG|linux-x86_64|g" -e "s|API|$API|g" ./cross/android_x86_64.txt > /tmp/android_cross.txt
-        meson setup build -Dlog=true -Dengines=all -Dloaders=all -Dsavers=all -Dbindings=capi -Dstatic=true --cross-file /tmp/android_cross.txt
+        meson setup build -Dlog=true -Dengines=all -Dloaders=all -Dsavers=all -Dbindings=capi -Dstatic=true -Dthreads=false --cross-file /tmp/android_cross.txt
         sudo ninja -C build install
 
     - uses: actions/upload-artifact@v4
@@ -64,7 +64,7 @@ jobs:
         API: 21
       run: |
         sed -e "s|NDK|$NDK|g" -e "s|HOST_TAG|linux-x86_64|g" -e "s|API|$API|g" ./cross/android_aarch64.txt > /tmp/android_cross.txt
-        meson setup build -Dlog=true -Dengines=all -Dloaders=all -Dsavers=all -Dbindings=capi -Dstatic=true --cross-file /tmp/android_cross.txt
+        meson setup build -Dlog=true -Dengines=all -Dloaders=all -Dsavers=all -Dbindings=capi -Dstatic=true -Dthreads=false --cross-file /tmp/android_cross.txt
         sudo ninja -C build install
 
     - uses: actions/upload-artifact@v4

src/meson.build

@@ -19,7 +19,7 @@ if cc.get_id() == 'clang-cl'
     endif
     if get_option('b_sanitize') == 'none'
         override_options += ['cpp_eh=none','cpp_rtti=false']
-        compiler_flags += ['/clang:-fno-math-errno', '/clang:-Woverloaded-virtual',
+        compiler_flags += ['/clang:-fno-math-errno', '/clang:-Woverloaded-virtual', '/clang:-Wno-unknown-pragmas',
                            '/clang:-fno-stack-protector', '/clang:-fno-unwind-tables' ,
                            '/clang:-fno-asynchronous-unwind-tables']
     endif
@@ -33,7 +33,7 @@ elif (cc.get_id() != 'msvc')
     if get_option('b_sanitize') == 'none'
         compiler_flags += ['-fno-exceptions', '-fno-rtti', '-fno-stack-protector', '-fno-math-errno',
                            '-fno-unwind-tables' , '-fno-asynchronous-unwind-tables',
-                           '-Woverloaded-virtual']
+                           '-Woverloaded-virtual', '-Wno-unknown-pragmas']
     endif
 endif
 

src/renderer/sw_engine/meson.build

@@ -17,7 +17,19 @@ source_file = [
    'tvgSwStroke.cpp',
 ]
 
+omp_dep = []
+sw_compiler_args = []
+
+if (get_option('threads'))
+    omp_dep = dependency('openmp', required: false)
+    if (omp_dep.found())
+        sw_compiler_args = '-DTHORVG_SW_OPENMP_SUPPORT=1'
+    endif
+endif
+
 engine_dep += [declare_dependency(
+    compile_args        : sw_compiler_args,
     include_directories : include_directories('.'),
-    sources : source_file
-)]
+    dependencies        : omp_dep,
+    sources             : source_file
+)]

src/renderer/sw_engine/tvgSwPostEffect.cpp

@@ -74,6 +74,7 @@ static void _gaussianBlur(uint8_t* src, uint8_t* dst, int32_t stride, int32_t w,
 
     auto iarr = 1.0f / (dimension + dimension + 1);
 
+    #pragma omp parallel for
     for (int x = 0; x < h; x++) {
         auto p = x * stride;
         auto i = p * 4;                 //current index

src/renderer/sw_engine/tvgSwRaster.cpp

@@ -1832,6 +1832,7 @@ void rasterXYFlip(uint32_t* src, uint32_t* dst, int32_t stride, int32_t w, int32
         dst += ((bbox.min.x * stride) + bbox.min.y);
     }
 
+    #pragma omp parallel for
     for (int x = 0; x < w; x += BLOCK) {
         auto bx = std::min(w, x + BLOCK) - x;
         auto in = &src[x];

src/renderer/sw_engine/tvgSwRenderer.cpp

@@ -20,6 +20,9 @@
  * SOFTWARE.
  */
 
+#ifdef THORVG_SW_OPENMP_SUPPORT
+    #include <omp.h>
+#endif
 #include <algorithm>
 #include "tvgMath.h"
 #include "tvgSwCommon.h"
@@ -762,6 +765,10 @@ bool SwRenderer::init(uint32_t threads)
 
 int32_t SwRenderer::init()
 {
+#ifdef THORVG_SW_OPENMP_SUPPORT
+    omp_set_num_threads(TaskScheduler::threads());
+#endif
+
     return initEngineCnt;
 }