Add support for using the SIMDE library for implementing SIMD instructions on non-x86 machines

.github/workflows/macos.yml

@@ -106,6 +106,7 @@ jobs:
           sudo port install adwaita-icon-theme
           sudo port install meson
           sudo port install dbus
+          sudo port install simde
 
       - name: Install LibRaw
         run: |
@@ -195,6 +196,11 @@ jobs:
           export PATH=/opt/local/libexec/gnubin:/opt/local/bin:/opt/local/sbin:$PATH
           mkdir build
           cd build
+          if [ "${{ matrix.arch }}" = "arm64" ]; then
+              ART_SIMDE=1
+          else
+              ART_SIMDE=0
+          fi
           export PKG_CONFIG_PATH=/opt/art-deps/lib/pkgconfig
           /opt/local/bin/cmake \
             -DCMAKE_C_COMPILER=/opt/local/bin/clang-mp-17 \
@@ -206,6 +212,7 @@ jobs:
             -DCMAKE_LIBRARY_PATH=/opt/art-deps/lib \
             -DMACOS_LEGACY_BUNDLE=0 \
             -DCMAKE_OSX_DEPLOYMENT_TARGET=$OS_VERSION \
+            -DENABLE_SIMDE=${ART_SIMDE} \
             ../repo
           cd ..
 

.github/workflows/windows.yml

@@ -134,15 +134,28 @@ jobs:
           ninja install
           cd ..
 
+      - if: ${{ matrix.arch == 'arm64' }}
+        name: Install SIMDE
+        run: |
+          wget https://github.com/mesonbuild/meson/releases/download/1.10.0/meson-1.10.0.tar.gz
+          tar xzf meson-1.10.0.tar.gz
+          wget https://github.com/simd-everywhere/simde/archive/refs/tags/v0.8.2.tar.gz
+          tar xzf v0.8.2.tar.gz
+          cd simde-0.8.2
+          python ../meson-1.10.0/meson.py setup build -Dbuildtype=release -Dtests=false -Dprefix=c:/msys2/msys64/usr/local
+          python ../meson-1.10.0/meson.py install -C build
+
       - name: Configure build
         run: |
           mkdir build
           cd build
           export PKG_CONFIG_PATH=c:/msys2/msys64/usr/local/lib/pkgconfig
           if [ "${{ matrix.msystem }}" = "UCRT64" ]; then
              ART_CXX_FLAGS=
+             ART_SIMDE=0
           else
              ART_CXX_FLAGS="-Wno-ignored-attributes -Wno-unused-command-line-argument"
+             ART_SIMDE=1
           fi
           cmake \
             -GNinja \
@@ -153,6 +166,7 @@ jobs:
             -DCTL_INCLUDE_DIR=/usr/local/include/CTL \
             -DCMAKE_LIBRARY_PATH=/usr/local/lib \
             -DBUILD_BUNDLE=1 \
+            -DENABLE_SIMDE=${ART_SIMDE} \
             -DCMAKE_CXX_FLAGS="${ART_CXX_FLAGS}" \
             ../repo
           cd ..

CMakeLists.txt

@@ -152,6 +152,7 @@ option(ENABLE_MIMALLOC "Use the mimalloc library if available" ON)
 option(ENABLE_LIBRAW "Use libraw for decoding" ON)
 option(ENABLE_OCIO "Use OpenColorIOv2 for LUT application" ON)
 option(ENABLE_CTL "Enable support for the ACES Color Transformation Language" OFF)
+option(ENABLE_SIMDE "Build with SIMD Everywhere support" OFF)
 
 option(MACOS_LEGACY_BUNDLE "Use legacy method for building a macOS bundle using the tools/osx/macos_bundle.sh script" OFF)
 
@@ -835,8 +836,9 @@ if(WIN32)
     endif()
 endif()
 
+include(CheckSymbolExists)
+
 if(APPLE)
-    include(CheckSymbolExists)
     set(CMAKE_REQUIRED_INCLUDES ${GTK_INCLUDE_DIRS})
     message(STATUS "Checking whether GDK supports specifying custom color spaces")
     check_symbol_exists(GDK_QUARTZ_WINDOW_SUPPORTS_COLORSPACE "gdk/gdkquartz.h" HAVE_GDK_QUARTZ_COLORSPACE)
@@ -848,6 +850,30 @@ if(APPLE)
     endif()
 endif()
 
+if(ENABLE_SIMDE)
+    pkg_check_modules(SIMDE REQUIRED simde)
+    add_compile_definitions(-DART_USE_SIMDE)
+    add_compile_definitions(-DSIMDE_ENABLE_NATIVE_ALIASES)
+    add_compile_definitions(-DART_SIMD)
+
+    set(CMAKE_REQUIRED_INCLUDES ${SIMDE_INCLUDE_DIRS})
+    set(CMAKE_REQUIRED_DEFINITIONS -DSIMDE_ENABLE_NATIVE_ALIASES)
+    check_symbol_exists(_MM_GET_FLUSH_ZERO_MODE "simde/x86/sse.h" HAVE_SIMDE_MM_GET_FLUSH_ZERO_MODE)
+    if(HAVE_SIMDE_MM_GET_FLUSH_ZERO_MODE)
+        add_definitions(-DART_SIMDE_GET_MM_FLUSH_ZERO_MODE)
+    endif()
+else()
+    check_cxx_source_compiles("#if !defined(__SSE2__)
+#error
+#endif
+
+int main() { return 0; }
+" HAVE_SSE2)
+    if(HAVE_SSE2)
+        add_definitions(-DART_SIMD)
+    endif()
+endif()
+
 add_subdirectory(rtengine)
 add_subdirectory(rtgui)
 add_subdirectory(rtdata)

rtengine/CA_correct_RT.cc

@@ -349,15 +349,15 @@ float *RawImageSource::CA_correct_RT(bool autoCA, size_t autoIterations,
                         // rgb values should be floating point numbers between 0
                         // and 1 after white balance multipliers are applied
 
-#ifdef __SSE2__
+#ifdef ART_SIMD
                         vfloat c65535v = F2V(65535.f);
 #endif
 
                         for (int rr = rrmin; rr < rrmax; rr++) {
                             int row = rr + top;
                             int cc = ccmin;
                             int col = cc + left;
-#ifdef __SSE2__
+#ifdef ART_SIMD
                             int c0 = fc(cfa, rr, cc);
                             if (c0 == 1) {
                                 rgb[c0][rr * ts + cc] =
@@ -488,7 +488,7 @@ float *RawImageSource::CA_correct_RT(bool autoCA, size_t autoIterations,
                         // end of border fill
                         // end of initialization
 
-#ifdef __SSE2__
+#ifdef ART_SIMD
                         vfloat onev = F2V(1.f);
                         vfloat epsv = F2V(eps);
 #endif
@@ -497,7 +497,7 @@ float *RawImageSource::CA_correct_RT(bool autoCA, size_t autoIterations,
                             int cc = 3 + (fc(cfa, rr, 3) & 1);
                             int indx = rr * ts + cc;
                             int c = fc(cfa, rr, cc);
-#ifdef __SSE2__
+#ifdef ART_SIMD
                             for (; cc < cc1 - 9; cc += 8, indx += 8) {
                                 // compute directional weights using image
                                 // gradients
@@ -601,7 +601,7 @@ float *RawImageSource::CA_correct_RT(bool autoCA, size_t autoIterations,
                                 int col = max(left + 3, 0) + offset;
                                 int indx = rr * ts + 3 -
                                            (left < 0 ? (left + 3) : 0) + offset;
-#ifdef __SSE2__
+#ifdef ART_SIMD
                                 for (; col < min(cc1 + left - 3, width) - 7;
                                      col += 8, indx += 8) {
                                     STVFU(Gtmp[(row * width + col) >> 1],
@@ -616,14 +616,14 @@ float *RawImageSource::CA_correct_RT(bool autoCA, size_t autoIterations,
                             }
                         }
 
-#ifdef __SSE2__
+#ifdef ART_SIMD
                         vfloat zd25v = F2V(0.25f);
 #endif
                         for (int rr = 4; rr < rr1 - 4; rr++) {
                             int cc = 4 + (fc(cfa, rr, 2) & 1);
                             int indx = rr * ts + cc;
                             int c = fc(cfa, rr, cc);
-#ifdef __SSE2__
+#ifdef ART_SIMD
                             for (; cc < cc1 - 10; cc += 8, indx += 8) {
                                 vfloat rgb1v = LC2VFU(rgb[1][indx]);
                                 vfloat rgbcv = LVFU(rgb[c][indx >> 1]);
@@ -757,7 +757,7 @@ float *RawImageSource::CA_correct_RT(bool autoCA, size_t autoIterations,
                             }
                         }
 
-#ifdef __SSE2__
+#ifdef ART_SIMD
                         vfloat zd3v = F2V(0.3f);
                         vfloat zd1v = F2V(0.1f);
                         vfloat zd5v = F2V(0.5f);
@@ -771,7 +771,7 @@ float *RawImageSource::CA_correct_RT(bool autoCA, size_t autoIterations,
                             int cc = 8 + (fc(cfa, rr, 2) & 1);
                             int indx = rr * ts + cc;
                             int c = fc(cfa, rr, cc);
-#ifdef __SSE2__
+#ifdef ART_SIMD
                             vfloat coeff00v = ZEROV;
                             vfloat coeff01v = ZEROV;
                             vfloat coeff02v = ZEROV;
@@ -1229,7 +1229,7 @@ float *RawImageSource::CA_correct_RT(bool autoCA, size_t autoIterations,
                         // rgb values should be floating point number between 0
                         // and 1 after white balance multipliers are applied
 
-#ifdef __SSE2__
+#ifdef ART_SIMD
                         vfloat c65535v = F2V(65535.f);
                         vmask gmask =
                             _mm_set_epi32(0, 0xffffffff, 0, 0xffffffff);
@@ -1240,7 +1240,7 @@ float *RawImageSource::CA_correct_RT(bool autoCA, size_t autoIterations,
                             int col = cc + left;
                             int indx = row * width + col;
                             int indx1 = rr * ts + cc;
-#ifdef __SSE2__
+#ifdef ART_SIMD
                             int c = fc(cfa, rr, cc);
                             if (c & 1) {
                                 rgb[1][indx1] = rawData[row][col] / 65535.f;
@@ -1431,7 +1431,7 @@ float *RawImageSource::CA_correct_RT(bool autoCA, size_t autoIterations,
                         // end of border fill
 
                         if (!autoCA || fitParamsIn) {
-#ifdef __SSE2__
+#ifdef ART_SIMD
                             const vfloat onev = F2V(1.f);
                             const vfloat epsv = F2V(eps);
 #endif
@@ -1440,7 +1440,7 @@ float *RawImageSource::CA_correct_RT(bool autoCA, size_t autoIterations,
                             for (int rr = 3; rr < rr1 - 3; rr++) {
                                 int cc = 3 + fc(cfa, rr, 1),
                                     c = fc(cfa, rr, cc), indx = rr * ts + cc;
-#ifdef __SSE2__
+#ifdef ART_SIMD
                                 for (; cc < cc1 - 10; cc += 8, indx += 8) {
                                     // compute directional weights using image
                                     // gradients
@@ -1649,7 +1649,7 @@ float *RawImageSource::CA_correct_RT(bool autoCA, size_t autoIterations,
                                 (rr + shiftvceil[c]) * ts + cc + shifthceil[c];
                             int indxcf =
                                 (rr + shiftvceil[c]) * ts + cc + shifthfloor[c];
-#ifdef __SSE2__
+#ifdef ART_SIMD
                             vfloat shifthfracv = F2V(shifthfrac[c]);
                             vfloat shiftvfracv = F2V(shiftvfrac[c]);
                             for (; cc < cc1 - 10; cc += 8, indxfc += 8,
@@ -1708,7 +1708,7 @@ float *RawImageSource::CA_correct_RT(bool autoCA, size_t autoIterations,
                         shiftvfrac[0] /= 2.f;
                         shiftvfrac[2] /= 2.f;
 
-#ifdef __SSE2__
+#ifdef ART_SIMD
                         vfloat zd25v = F2V(0.25f);
                         vfloat onev = F2V(1.f);
                         vfloat zd5v = F2V(0.5f);
@@ -1719,7 +1719,7 @@ float *RawImageSource::CA_correct_RT(bool autoCA, size_t autoIterations,
                             int c = fc(cfa, rr, cc);
                             int GRBdir0 = GRBdir[0][c];
                             int GRBdir1 = GRBdir[1][c];
-#ifdef __SSE2__
+#ifdef ART_SIMD
                             vfloat shifthfracc = F2V(shifthfrac[c]);
                             vfloat shiftvfracc = F2V(shiftvfrac[c]);
                             for (int indx = rr * ts + cc; cc < cc1 - 14;
@@ -1911,7 +1911,7 @@ float *RawImageSource::CA_correct_RT(bool autoCA, size_t autoIterations,
                             int cc = border + (fc(cfa, rr, 2) & 1);
                             int indx = (row * width + cc + left) >> 1;
                             int indx1 = (rr * ts + cc) >> 1;
-#ifdef __SSE2__
+#ifdef ART_SIMD
                             for (; indx <
                                    (row * width + cc1 - border - 7 + left) >> 1;
                                  indx += 4, indx1 += 4) {
@@ -1951,7 +1951,7 @@ float *RawImageSource::CA_correct_RT(bool autoCA, size_t autoIterations,
                 for (int row = cb; row < height - cb; row++) {
                     int col = cb + (fc(cfa, row, 0) & 1);
                     int indx = (row * width + col) >> 1;
-#ifdef __SSE2__
+#ifdef ART_SIMD
                     for (; col < width - 7 - cb; col += 8, indx += 4) {
                         const vfloat val = vmaxf(LVFU(RawDataTmp[indx]), ZEROV);
                         STC2VFU(rawData[row][col], val);
@@ -1975,7 +1975,7 @@ float *RawImageSource::CA_correct_RT(bool autoCA, size_t autoIterations,
 #pragma omp parallel
 #endif
             {
-#ifdef __SSE2__
+#ifdef ART_SIMD
                 const vfloat onev = F2V(1.f);
                 const vfloat twov = F2V(2.f);
                 const vfloat zd5v = F2V(0.5f);
@@ -1989,7 +1989,7 @@ float *RawImageSource::CA_correct_RT(bool autoCA, size_t autoIterations,
                     const array2D<float> *nonGreen =
                         colour == 0 ? redFactor : blueFactor;
                     int j = firstCol;
-#ifdef __SSE2__
+#ifdef ART_SIMD
                     for (; j < W - 7 - 2 * cb; j += 8) {
                         const vfloat newvals = LC2VFU(rawData[i + cb][j + cb]);
                         const vfloat oldvals = LVFU((*oldraw)[i][j / 2]);

rtengine/CMakeLists.txt

@@ -41,6 +41,10 @@ if(CTL_FOUND)
     link_directories(${CTL_LIBRARY_DIRS})
 endif()
 
+if(SIMDE_INCLUDE_DIRS)
+    include_directories("${SIMDE_INCLUDE_DIRS}")
+endif()
+
 set(CAMCONSTSFILE "camconst.json")
 
 set(RTENGINESOURCEFILES