Merge branch 'AcademySoftwareFoundation:main' into CUDA

1div0 · web-flow · commit e28508bb61e4 · 2026-01-10T19:45:12.000+01:00
diff --git a/.github/PULL_REQUEST_TEMPLATE.md b/.github/PULL_REQUEST_TEMPLATE.md
@@ -1,3 +1,8 @@
+
+
+
+------ :scissors: -------------------------------------------------------------------
+
 YOU MAY DELETE ALL OF THIS IF YOU ALREADY HAVE A DESCRIPTIVE COMMIT MESSAGE!
 
 This is just a template and set of reminders about what constitutes a good PR.
diff --git a/.github/workflows/analysis.yml b/.github/workflows/analysis.yml
@@ -26,6 +26,7 @@ on:
       - '!**/scorecard.yml'
       - '!**/wheel.yml'
       - '!**.properties'
+      - '!pyproject.toml'
       - '!docs/**'
   # Run analysis on PRs only if the branch name indicates that the purpose of
   # the PR is related to the Sonar analysis. We don't run on every PR because
diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
@@ -16,6 +16,7 @@ on:
       - '!**/scorecard.yml'
       - '!**/wheel.yml'
       - '!**.properties'
+      - '!pyproject.toml'
       - '!docs/**'
   pull_request:
     paths:
@@ -466,7 +467,7 @@ jobs:
                             PTEX_VERSION=v2.5.0
                             PUGIXML_VERSION=v1.15
                             WEBP_VERSION=v1.6.0
-                            FREETYPE_VERSION=VER-2-14-0
+                            FREETYPE_VERSION=VER-2-14-1
                             USE_OPENVDB=0
             # Ensure we are testing all the deps we think we are. We would
             # like this test to have minimal missing dependencies.
@@ -549,7 +550,7 @@ jobs:
                             PTEX_VERSION=v2.4.3
                             PUGIXML_VERSION=v1.15
                             WEBP_VERSION=v1.6.0
-                            FREETYPE_VERSION=VER-2-14-0
+                            FREETYPE_VERSION=VER-2-14-1
                             USE_OPENVDB=0
           - desc: Linux ARM latest releases clang18 C++20 py3.12 exr3.4 ocio2.4
             nametag: linux-arm-latest-releases-clang
@@ -570,7 +571,7 @@ jobs:
                             PTEX_VERSION=v2.4.3
                             PUGIXML_VERSION=v1.15
                             WEBP_VERSION=v1.6.0
-                            FREETYPE_VERSION=VER-2-14-0
+                            FREETYPE_VERSION=VER-2-14-1
                             USE_OPENVDB=0
 
 
@@ -626,7 +627,7 @@ jobs:
             python_ver: "3.13"
             simd: sse4.2,avx2
             ctest_test_timeout: 1200
-            setenvs: export MACOSX_DEPLOYMENT_TARGET=12.0
+            setenvs: export MACOSX_DEPLOYMENT_TARGET=12.0 INSTALL_QT=0
             benchmark: 1
           - desc: MacOS-14-ARM aclang15/C++20/py3.13
             runner: macos-14
@@ -683,7 +684,7 @@ jobs:
       # built. But we would like to add more dependencies and reduce this list
       # of exceptions in the future.
       required_deps: ${{ matrix.required_deps || 'all' }}
-      optional_deps: ${{ matrix.optional_deps || 'CUDAToolkit;DCMTK;FFmpeg;GIF;JXL;Libheif;LibRaw;Nuke;OpenCV;OpenGL;OpenJPEG;openjph;OpenCV;OpenVDB;Ptex;pystring;Qt5;Qt6;TBB;R3DSDK;${{matrix.optional_deps_append}}' }}
+      optional_deps: ${{ matrix.optional_deps || 'BZip2;CUDAToolkit;DCMTK;FFmpeg;GIF;JXL;Libheif;LibRaw;Nuke;OpenCV;OpenGL;OpenJPEG;openjph;OpenCV;OpenVDB;Ptex;pystring;Qt5;Qt6;TBB;R3DSDK;${{matrix.optional_deps_append}}' }}
     strategy:
       fail-fast: false
       matrix:
diff --git a/.github/workflows/docs.yml b/.github/workflows/docs.yml
@@ -20,6 +20,7 @@ on:
       - '**/run.py'
       - 'src/build-scripts/**'
       - './*.md'
+      - 'pyproject.toml'
   pull_request:
     paths-ignore:
       - '**/ci.yml'
@@ -33,6 +34,7 @@ on:
       - '**/run.py'
       - 'src/build-scripts/**'
       - './*.md'
+      - 'pyproject.toml'
   schedule:
     # Full nightly build
     - cron: "0 8 * * *"
diff --git a/src/build-scripts/build_Freetype.bash b/src/build-scripts/build_Freetype.bash
@@ -11,7 +11,7 @@ set -ex
 
 # Repo and branch/tag/commit of Freetype to download if we don't have it yet
 FREETYPE_REPO=${FREETYPE_REPO:=https://github.com/freetype/freetype.git}
-FREETYPE_VERSION=${FREETYPE_VERSION:=VER-2-13-3}
+FREETYPE_VERSION=${FREETYPE_VERSION:=VER-2-14-1}
 
 # Where to put Freetype repo source (default to the ext area)
 LOCAL_DEPS_DIR=${LOCAL_DEPS_DIR:=${PWD}/ext}
diff --git a/src/cmake/build_Freetype.cmake b/src/cmake/build_Freetype.cmake
@@ -6,10 +6,10 @@
 # Freetype by hand!
 ######################################################################
 
-set_cache (Freetype_BUILD_VERSION 2.13.2 "Freetype version for local builds")
+set_cache (Freetype_BUILD_VERSION 2.14.1 "Freetype version for local builds")
 set (Freetype_GIT_REPOSITORY "https://github.com/freetype/freetype")
-set (Freetype_GIT_TAG "VER-2-13-2")
-set_cache (Freetype_BUILD_SHARED_LIBS  OFF
+set (Freetype_GIT_TAG "VER-2-14-1")
+set_cache (Freetype_BUILD_SHARED_LIBS ${LOCAL_BUILD_SHARED_LIBS_DEFAULT}
            DOC "Should a local Freetype build, if necessary, build shared libraries" ADVANCED)
 # We would prefer to build a static Freetype, but haven't figured out how to make
 # it all work with the static dependencies, it just makes things complicated
diff --git a/src/cmake/externalpackages.cmake b/src/cmake/externalpackages.cmake
@@ -200,11 +200,9 @@ checked_find_package (R3DSDK NO_RECORD_NOTFOUND)  # RED camera
 set (NUKE_VERSION "7.0" CACHE STRING "Nuke version to target")
 checked_find_package (Nuke NO_RECORD_NOTFOUND)
 
-if (FFmpeg_FOUND OR FREETYPE_FOUND)
+if ((FFmpeg_FOUND OR FREETYPE_FOUND OR TARGET Freetype::Freetype)
+    AND NOT TARGET BZip2::BZip2)
     checked_find_package (BZip2)   # Used by ffmpeg and freetype
-    if (NOT BZIP2_FOUND)
-        set (BZIP2_LIBRARIES "")  # TODO: why does it break without this?
-    endif ()
 endif()
 
 
diff --git a/src/ffmpeg.imageio/CMakeLists.txt b/src/ffmpeg.imageio/CMakeLists.txt
@@ -28,7 +28,7 @@ if (FFmpeg_FOUND)
     add_oiio_plugin (ffmpeginput.cpp
                      INCLUDE_DIRS ${FFMPEG_INCLUDES}
                      LINK_LIBRARIES ${FFMPEG_LIBRARIES}
-                                    ${BZIP2_LIBRARIES}
+                                    $<TARGET_NAME_IF_EXISTS:BZip2::BZip2>
                      DEFINITIONS "USE_FFMPEG"
                                  "-DOIIO_FFMPEG_VERSION=\"${FFMPEG_VERSION}\"")
 else()
diff --git a/src/include/OpenImageIO/typedesc.h b/src/include/OpenImageIO/typedesc.h
@@ -409,7 +409,7 @@ template<> struct BaseTypeFromC<uint64_t> { static constexpr TypeDesc::BASETYPE
 template<> struct BaseTypeFromC<const uint64_t> { static constexpr TypeDesc::BASETYPE value = TypeDesc::UINT64; };
 template<> struct BaseTypeFromC<int64_t> { static constexpr TypeDesc::BASETYPE value = TypeDesc::INT64; };
 template<> struct BaseTypeFromC<const int64_t> { static constexpr TypeDesc::BASETYPE value = TypeDesc::INT64; };
-#if defined(__GNUC__) && (ULONG_MAX == 0xffffffffffffffff) && !(defined(__APPLE__) && defined(__MACH__)) || defined(__NetBSD__)
+#if defined(__GNUC__) && (ULONG_MAX == 0xffffffffffffffff) && !(defined(__APPLE__) && defined(__MACH__)) && !defined(__OpenBSD__) || defined(__NetBSD__)
 // Some platforms consider int64_t and long long to be different types, even
 // though they are actually the same size.
 static_assert(!std::is_same_v<unsigned long long, uint64_t>);
diff --git a/src/libOpenImageIO/CMakeLists.txt b/src/libOpenImageIO/CMakeLists.txt
@@ -163,7 +163,7 @@ target_link_libraries (OpenImageIO
             $<TARGET_NAME_IF_EXISTS:pugixml::pugixml>
             $<TARGET_NAME_IF_EXISTS:TBB::tbb>
             $<TARGET_NAME_IF_EXISTS:Freetype::Freetype>
-            ${BZIP2_LIBRARIES}
+            $<TARGET_NAME_IF_EXISTS:BZip2::BZip2>
             ZLIB::ZLIB
             ${CMAKE_DL_LIBS}
         )
diff --git a/src/libOpenImageIO/imagebufalgo_test.cpp b/src/libOpenImageIO/imagebufalgo_test.cpp
@@ -641,6 +641,41 @@ test_zover()
 
 
 
+// Test ImageBuf::resample
+void
+test_resample()
+{
+    std::cout << "test resample\n";
+
+    // Timing
+    Benchmarker bench;
+    bench.units(Benchmarker::Unit::ms);
+
+    ImageSpec spec_hd_rgba_f(1920, 1080, 4, TypeFloat);
+    ImageSpec spec_hd_rgba_u8(1920, 1080, 4, TypeUInt8);
+    ImageBuf buf_hd_rgba_f(spec_hd_rgba_f);
+    ImageBuf buf_hd_rgba_u8(spec_hd_rgba_u8);
+    float red_rgba[] = { 1.0, 0.0, 0.0, 1.0 };
+    ImageBufAlgo::fill(buf_hd_rgba_f, red_rgba);
+    ImageBufAlgo::fill(buf_hd_rgba_u8, red_rgba);
+    ImageBuf smallf(ImageSpec(1024, 512, 4, TypeFloat));
+    ImageBuf smallu8(ImageSpec(1024, 512, 4, TypeUInt8));
+    bench("  IBA::resample HD->1024x512 rgba f->f    interp   ",
+          [&]() { ImageBufAlgo::resample(smallf, buf_hd_rgba_f, true); });
+    bench("  IBA::resample HD->1024x512 rgba f->u8   interp   ",
+          [&]() { ImageBufAlgo::resample(smallu8, buf_hd_rgba_f, true); });
+    bench("  IBA::resample HD->1024x512 rgba u8->u8  interp   ",
+          [&]() { ImageBufAlgo::resample(smallu8, buf_hd_rgba_u8, true); });
+    bench("  IBA::resample HD->1024x512 rgba f->f   no interp ",
+          [&]() { ImageBufAlgo::resample(smallf, buf_hd_rgba_f, false); });
+    bench("  IBA::resample HD->1024x512 rgba f->u8  no interp ",
+          [&]() { ImageBufAlgo::resample(smallu8, buf_hd_rgba_f, false); });
+    bench("  IBA::resample HD->1024x512 rgba u8->u8 no interp ",
+          [&]() { ImageBufAlgo::resample(smallu8, buf_hd_rgba_u8, false); });
+}
+
+
+
 // Tests ImageBufAlgo::compare
 void
 test_compare()
@@ -1581,6 +1616,7 @@ main(int argc, char** argv)
     test_over(TypeFloat);
     test_over(TypeHalf);
     test_zover();
+    test_resample();
     test_compare();
     test_isConstantColor();
     test_isConstantChannel();
diff --git a/src/libOpenImageIO/imagebufalgo_xform.cpp b/src/libOpenImageIO/imagebufalgo_xform.cpp
@@ -1070,6 +1070,35 @@ ImageBufAlgo::fit(const ImageBuf& src, KWArgs options, ROI roi, int nthreads)
 
 
 
+// This operates just like the internals of ImageBuf::interppixel(), but
+// reuses the provided iterator to avoid the overhead of constructing a new
+// one each time. This speeds it up by 20x! The iterator `it` must already be
+// associated with `img`, but it need not be positioned correctly.
+template<class T>
+static bool
+interppixel(const ImageBuf& img, ImageBuf::ConstIterator<T>& it, float x,
+            float y, span<float> pixel, ImageBuf::WrapMode wrap)
+{
+    int n             = std::min(int(pixel.size()), img.spec().nchannels);
+    float* localpixel = OIIO_ALLOCA(float, n * 4);
+    float* p[4]       = { localpixel, localpixel + n, localpixel + 2 * n,
+                          localpixel + 3 * n };
+    x -= 0.5f;
+    y -= 0.5f;
+    int xtexel, ytexel;
+    float xfrac, yfrac;
+    xfrac = floorfrac(x, &xtexel);
+    yfrac = floorfrac(y, &ytexel);
+    it.rerange(xtexel, xtexel + 2, ytexel, ytexel + 2, 0, 1, wrap);
+    for (int i = 0; i < 4; ++i, ++it)
+        for (int c = 0; c < n; ++c)
+            p[i][c] = it[c];  //NOSONAR
+    bilerp(p[0], p[1], p[2], p[3], xfrac, yfrac, n, pixel.data());
+    return true;
+}
+
+
+
 template<typename DSTTYPE, typename SRCTYPE>
 static bool
 resample_(ImageBuf& dst, const ImageBuf& src, bool interpolate, ROI roi,
@@ -1080,7 +1109,6 @@ resample_(ImageBuf& dst, const ImageBuf& src, bool interpolate, ROI roi,
         const ImageSpec& srcspec(src.spec());
         const ImageSpec& dstspec(dst.spec());
         int nchannels = src.nchannels();
-        bool deep     = src.deep();
 
         // Local copies of the source image window, converted to float
         float srcfx = srcspec.full_x;
@@ -1109,25 +1137,10 @@ resample_(ImageBuf& dst, const ImageBuf& src, bool interpolate, ROI roi,
                 float s      = (x - dstfx + 0.5f) * dstpixelwidth;
                 float src_xf = srcfx + s * srcfw;
                 int src_x    = ifloor(src_xf);
-                if (deep) {
-                    srcpel.pos(src_x, src_y, 0);
-                    int nsamps = srcpel.deep_samples();
-                    OIIO_DASSERT(nsamps == out.deep_samples());
-                    if (!nsamps || nsamps != out.deep_samples())
-                        continue;
-                    for (int c = 0; c < nchannels; ++c) {
-                        if (dstspec.channelformat(c) == TypeDesc::UINT32)
-                            for (int samp = 0; samp < nsamps; ++samp)
-                                out.set_deep_value(
-                                    c, samp, srcpel.deep_value_uint(c, samp));
-                        else
-                            for (int samp = 0; samp < nsamps; ++samp)
-                                out.set_deep_value(c, samp,
-                                                   srcpel.deep_value(c, samp));
-                    }
-                } else if (interpolate) {
+                if (interpolate) {
                     // Non-deep image, bilinearly interpolate
-                    src.interppixel(src_xf, src_yf, pel, ImageBuf::WrapClamp);
+                    interppixel(src, srcpel, src_xf, src_yf, pel,
+                                ImageBuf::WrapClamp);
                     for (int c = roi.chbegin; c < roi.chend; ++c)
                         out[c] = pel[c];
                 } else {
@@ -1144,6 +1157,88 @@ resample_(ImageBuf& dst, const ImageBuf& src, bool interpolate, ROI roi,
 
 
 
+static bool
+resample_deep(ImageBuf& dst, const ImageBuf& src, bool interpolate, ROI roi,
+              int nthreads)
+{
+    // If it's deep, figure out the sample allocations first, because
+    // it's not thread-safe to do that simultaneously with copying the
+    // values.
+    const ImageSpec& srcspec(src.spec());
+    const ImageSpec& dstspec(dst.spec());
+    float srcfx          = srcspec.full_x;
+    float srcfy          = srcspec.full_y;
+    float srcfw          = srcspec.full_width;
+    float srcfh          = srcspec.full_height;
+    float dstpixelwidth  = 1.0f / dstspec.full_width;
+    float dstpixelheight = 1.0f / dstspec.full_height;
+    ImageBuf::ConstIterator<float> srcpel(src, roi);
+    ImageBuf::Iterator<float> dstpel(dst, roi);
+    for (; !dstpel.done(); ++dstpel, ++srcpel) {
+        float s   = (dstpel.x() - dstspec.full_x + 0.5f) * dstpixelwidth;
+        float t   = (dstpel.y() - dstspec.full_y + 0.5f) * dstpixelheight;
+        int src_y = ifloor(srcfy + t * srcfh);
+        int src_x = ifloor(srcfx + s * srcfw);
+        srcpel.pos(src_x, src_y, 0);
+        dstpel.set_deep_samples(srcpel.deep_samples());
+    }
+
+    OIIO_ASSERT(src.deep() == dst.deep());
+    ImageBufAlgo::parallel_image(roi, nthreads, [&](ROI roi) {
+        const ImageSpec& srcspec(src.spec());
+        const ImageSpec& dstspec(dst.spec());
+        int nchannels = src.nchannels();
+
+        // Local copies of the source image window, converted to float
+        float srcfx = srcspec.full_x;
+        float srcfy = srcspec.full_y;
+        float srcfw = srcspec.full_width;
+        float srcfh = srcspec.full_height;
+
+        float dstfx          = dstspec.full_x;
+        float dstfy          = dstspec.full_y;
+        float dstfw          = dstspec.full_width;
+        float dstfh          = dstspec.full_height;
+        float dstpixelwidth  = 1.0f / dstfw;
+        float dstpixelheight = 1.0f / dstfh;
+
+        ImageBuf::Iterator<float> out(dst, roi);
+        ImageBuf::ConstIterator<float> srcpel(src);
+        for (int y = roi.ybegin; y < roi.yend; ++y) {
+            // s,t are NDC space
+            float t = (y - dstfy + 0.5f) * dstpixelheight;
+            // src_xf, src_xf are image space float coordinates
+            float src_yf = srcfy + t * srcfh;
+            // src_x, src_y are image space integer coordinates of the floor
+            int src_y = ifloor(src_yf);
+            for (int x = roi.xbegin; x < roi.xend; ++x, ++out) {
+                float s      = (x - dstfx + 0.5f) * dstpixelwidth;
+                float src_xf = srcfx + s * srcfw;
+                int src_x    = ifloor(src_xf);
+                srcpel.pos(src_x, src_y, 0);
+                int nsamps = srcpel.deep_samples();
+                OIIO_DASSERT(nsamps == out.deep_samples());
+                if (!nsamps || nsamps != out.deep_samples())
+                    continue;
+                for (int c = 0; c < nchannels; ++c) {
+                    if (dstspec.channelformat(c) == TypeDesc::UINT32)
+                        for (int samp = 0; samp < nsamps; ++samp)
+                            out.set_deep_value(c, samp,
+                                               srcpel.deep_value_uint(c, samp));
+                    else
+                        for (int samp = 0; samp < nsamps; ++samp)
+                            out.set_deep_value(c, samp,
+                                               srcpel.deep_value(c, samp));
+                }
+            }
+        }
+    });
+
+    return true;
+}
+
+
+
 bool
 ImageBufAlgo::resample(ImageBuf& dst, const ImageBuf& src, bool interpolate,
                        ROI roi, int nthreads)
@@ -1155,27 +1250,7 @@ ImageBufAlgo::resample(ImageBuf& dst, const ImageBuf& src, bool interpolate,
         return false;
 
     if (dst.deep()) {
-        // If it's deep, figure out the sample allocations first, because
-        // it's not thread-safe to do that simultaneously with copying the
-        // values.
-        const ImageSpec& srcspec(src.spec());
-        const ImageSpec& dstspec(dst.spec());
-        float srcfx          = srcspec.full_x;
-        float srcfy          = srcspec.full_y;
-        float srcfw          = srcspec.full_width;
-        float srcfh          = srcspec.full_height;
-        float dstpixelwidth  = 1.0f / dstspec.full_width;
-        float dstpixelheight = 1.0f / dstspec.full_height;
-        ImageBuf::ConstIterator<float> srcpel(src, roi);
-        ImageBuf::Iterator<float> dstpel(dst, roi);
-        for (; !dstpel.done(); ++dstpel, ++srcpel) {
-            float s   = (dstpel.x() - dstspec.full_x + 0.5f) * dstpixelwidth;
-            float t   = (dstpel.y() - dstspec.full_y + 0.5f) * dstpixelheight;
-            int src_y = ifloor(srcfy + t * srcfh);
-            int src_x = ifloor(srcfx + s * srcfw);
-            srcpel.pos(src_x, src_y, 0);
-            dstpel.set_deep_samples(srcpel.deep_samples());
-        }
+        return resample_deep(dst, src, interpolate, roi, nthreads);
     }
 
     bool ok;

Original file line number	Diff line number	Diff line change
`@@ -163,7 +163,7 @@ target_link_libraries (OpenImageIO`
`163`	`163`	`$<TARGET_NAME_IF_EXISTS:pugixml::pugixml>`
`164`	`164`	`$<TARGET_NAME_IF_EXISTS:TBB::tbb>`
`165`	`165`	`$<TARGET_NAME_IF_EXISTS:Freetype::Freetype>`
`166`		`- ${BZIP2_LIBRARIES}`
	`166`	`+ $<TARGET_NAME_IF_EXISTS:BZip2::BZip2>`
`167`	`167`	`ZLIB::ZLIB`
`168`	`168`	`${CMAKE_DL_LIBS}`
`169`	`169`	`)`