Merge branch 'feature/KASM-6788-threading_2' into 'master'

KASM-6788 Refactor EncodeManager to use std::execution with parallel... Closes KASM-6788 See merge request kasm-technologies/internal/KasmVNC!168
2025-06-27 05:01:41 +02:00 · 2025-04-09 13:57:45 +00:00 · 2025-04-09 13:57:45 +00:00 · dc75e98344
commit dc75e98344
parent e647af5e28 90c38a0da8
10 changed files with 129 additions and 86 deletions
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@ -21,7 +21,7 @@ include(CheckCSourceRuns)

 include(CMakeMacroLibtoolFile)

-project(kasmvnc)
+project(kasmvnc LANGUAGES C CXX)
 set(VERSION 0.9)

 # The RC version must always be four comma-separated numbers
@ -74,13 +74,10 @@ set(CMAKE_CXX_FLAGS_MINSIZEREL "${CMAKE_CXX_FLAGS_MINSIZEREL} -UNDEBUG")

 # Make sure we get a sane C version
 set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -std=gnu99")
+set(CMAKE_CXX_STANDARD 20)

 # Enable OpenMP
 set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -fopenmp")
-set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fopenmp")
-
-# Enable C++ 11
-set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=gnu++11")

 # Tell the compiler to be stringent
 set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -Wall -Wformat=2")
--- a/builder/dockerfile.debian_bullseye.build
+++ b/builder/dockerfile.debian_bullseye.build
@ -12,9 +12,22 @@ RUN apt-get update && \

 RUN DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends tzdata
 RUN apt-get update && apt-get -y build-dep xorg-server libxfont-dev
-RUN apt-get update && apt-get -y install cmake git libgnutls28-dev vim wget tightvncserver curl
+RUN apt-get update && apt-get -y install git libgnutls28-dev vim wget tightvncserver curl
 RUN apt-get update && apt-get -y install libpng-dev libtiff-dev libgif-dev libavcodec-dev libssl-dev libxrandr-dev libxcursor-dev

+RUN CMAKE_URL="https://cmake.org/files/v3.22/cmake-3.22.0" && \
+    ARCH=$(arch) && \
+    if [ "$ARCH" = "x86_64" ]; then \
+        CMAKE_URL="${CMAKE_URL}-linux-x86_64.sh"; \
+    elif [ "$ARCH" = "aarch64" ]; then \
+        CMAKE_URL="${CMAKE_URL}-linux-aarch64.sh"; \
+    else \
+        echo "Unsupported architecture: $ARCH" && exit 1; \
+    fi && \
+    curl -fsSL $CMAKE_URL -o cmake.sh && \
+    (echo y; echo n) | bash cmake.sh --prefix=/usr/local --skip-license && \
+    rm cmake.sh
+
 ENV SCRIPTS_DIR=/tmp/scripts
 COPY builder/scripts $SCRIPTS_DIR
 RUN $SCRIPTS_DIR/build-webp
--- a/builder/dockerfile.opensuse_15.build
+++ b/builder/dockerfile.opensuse_15.build
@ -13,8 +13,8 @@ RUN zypper install -ny \
  ffmpeg-4-libavcodec-devel \
  fonttosfnt \
  font-util \
-  gcc \
-  gcc-c++ \
+  gcc14 \
+  gcc14-c++ \
  giflib-devel \
  git \
  gzip \
@ -45,17 +45,30 @@ RUN zypper install -ny \
  xorg-x11-util-devel \
  zlib-devel

-ENV SCRIPTS_DIR=/tmp/scripts
-COPY builder/scripts $SCRIPTS_DIR
-RUN $SCRIPTS_DIR/build-webp
-RUN $SCRIPTS_DIR/build-libjpeg-turbo
+RUN update-alternatives --install /usr/bin/c++ c++ /usr/bin/g++-14 100

 RUN useradd -u 1000 docker && \
  groupadd -g 1000 docker && \
  usermod -a -G docker docker

+RUN ARCH=$(arch) && \
+  CMAKE_URL="https://cmake.org/files/v3.22/cmake-3.22.0" && \
+  if [ "$ARCH" = "x86_64" ]; then \
+      CMAKE_URL="${CMAKE_URL}-linux-x86_64.sh"; \
+  elif [ "$ARCH" = "aarch64" ]; then \
+      CMAKE_URL="${CMAKE_URL}-linux-aarch64.sh"; \
+  else \
+      echo "Unsupported architecture: $ARCH" && exit 1; \
+  fi && \
+  curl -fsSL $CMAKE_URL -o cmake.sh && \
+  (echo y; echo n) | bash cmake.sh --prefix=/usr/local --skip-license && \
+  rm cmake.sh
+
+ENV SCRIPTS_DIR=/tmp/scripts
+COPY builder/scripts $SCRIPTS_DIR
+RUN $SCRIPTS_DIR/build-webp && $SCRIPTS_DIR/build-libjpeg-turbo
+
 COPY --chown=docker:docker . /src/

-
 USER docker
-ENTRYPOINT ["/src/builder/build.sh"]
+ENTRYPOINT ["bash", "-l", "-c", "/src/builder/build.sh"]
--- a/builder/dockerfile.oracle_8.build
+++ b/builder/dockerfile.oracle_8.build
@ -15,6 +15,7 @@ RUN \
    dnf-plugins-core \
    gcc \
    gcc-c++ \
+    gcc-toolset-14 \
    git \
    gnutls-devel \
    libjpeg-turbo-devel \
@ -38,6 +39,7 @@ RUN dnf install -y --nogpgcheck https://mirrors.rpmfusion.org/free/el/rpmfusion-

 # Install from new repos
 RUN dnf install -y \
+  tbb-devel \
  ffmpeg-devel \
  giflib-devel \
  lbzip2 \
@ -48,16 +50,16 @@ RUN dnf install -y \
  xorg-x11-xtrans-devel \
  libXrandr-devel \
  libXtst-devel \
-  libXcursor-devel
+  libXcursor-devel \
+  libSM-devel

 ENV SCRIPTS_DIR=/tmp/scripts
 COPY builder/scripts $SCRIPTS_DIR
-RUN $SCRIPTS_DIR/build-webp
-RUN $SCRIPTS_DIR/build-libjpeg-turbo
-
-RUN useradd -m docker && echo "docker:docker" | chpasswd
+RUN echo "source /opt/rh/gcc-toolset-14/enable" > /etc/profile.d/gcc-toolset.sh && \
+  $SCRIPTS_DIR/build-webp && $SCRIPTS_DIR/build-libjpeg-turbo && \
+  useradd -m docker && echo "docker:docker" | chpasswd

 COPY --chown=docker:docker . /src/

 USER docker
-ENTRYPOINT ["/src/builder/build.sh"]
+ENTRYPOINT ["bash", "-l", "-c", "/src/builder/build.sh"]
--- a/builder/dockerfile.oracle_9.build
+++ b/builder/dockerfile.oracle_9.build
@ -15,6 +15,7 @@ RUN \
    dnf-plugins-core \
    gcc \
    gcc-c++ \
+    gcc-toolset-14 \
    git \
    gnutls-devel \
    libjpeg-turbo-devel \
@ -47,17 +48,16 @@ RUN dnf install -y \
  xorg-x11-xtrans-devel \
  libXrandr-devel \
  libXtst-devel \
-  libXcursor-devel
-
+  libXcursor-devel \
+  libSM-devel

 ENV SCRIPTS_DIR=/tmp/scripts
 COPY builder/scripts $SCRIPTS_DIR
-RUN $SCRIPTS_DIR/build-webp
-RUN $SCRIPTS_DIR/build-libjpeg-turbo
-
-RUN useradd -m docker && echo "docker:docker" | chpasswd
+RUN echo "source /opt/rh/gcc-toolset-14/enable" > /etc/profile.d/gcc-toolset.sh && \
+  $SCRIPTS_DIR/build-webp && $SCRIPTS_DIR/build-libjpeg-turbo && \
+  useradd -m docker && echo "docker:docker" | chpasswd

 COPY --chown=docker:docker . /src/

 USER docker
-ENTRYPOINT ["/src/builder/build.sh"]
+ENTRYPOINT ["bash", "-l", "-c", "/src/builder/build.sh"]
--- a/builder/dockerfile.ubuntu_focal.build
+++ b/builder/dockerfile.ubuntu_focal.build
@ -13,7 +13,7 @@ RUN apt-get update && \
 RUN apt-get update && apt-get install -y --no-install-recommends tzdata
 RUN apt-get update && apt-get -y build-dep xorg-server libxfont-dev
 RUN apt-get update && apt-get -y install cmake git vim wget curl
-RUN apt-get update && apt-get -y install libpng-dev libtiff-dev libgif-dev libavcodec-dev libssl-dev libxrandr-dev libxcursor-dev
+RUN apt-get update && apt-get -y install libtbb-dev libpng-dev libtiff-dev libgif-dev libavcodec-dev libssl-dev libxrandr-dev libxcursor-dev

 ENV SCRIPTS_DIR=/tmp/scripts
 COPY builder/scripts $SCRIPTS_DIR
@ -22,6 +22,19 @@ RUN $SCRIPTS_DIR/build-libjpeg-turbo

 RUN useradd -m docker && echo "docker:docker" | chpasswd && adduser docker sudo

+RUN ARCH=$(arch) && \
+  CMAKE_URL="https://cmake.org/files/v3.22/cmake-3.22.0" && \
+  if [ "$ARCH" = "x86_64" ]; then \
+      CMAKE_URL="${CMAKE_URL}-linux-x86_64.sh"; \
+  elif [ "$ARCH" = "aarch64" ]; then \
+      CMAKE_URL="${CMAKE_URL}-linux-aarch64.sh"; \
+  else \
+      echo "Unsupported architecture: $ARCH" && exit 1; \
+  fi && \
+  curl -fsSL $CMAKE_URL -o cmake.sh && \
+  (echo y; echo n) | bash cmake.sh --prefix=/usr/local --skip-license && \
+  rm cmake.sh
+
 COPY --chown=docker:docker . /src/

 USER docker
--- a/builder/dockerfile.ubuntu_focal.deb.build
+++ b/builder/dockerfile.ubuntu_focal.deb.build
@ -3,7 +3,7 @@ FROM ubuntu:focal
 ENV DEBIAN_FRONTEND noninteractive

 RUN apt-get update && \
-      apt-get -y install vim build-essential devscripts equivs
+      apt-get -y install vim build-essential devscripts equivs libtbb-dev

 # Install build-deps for the package.
 COPY ./debian/control /tmp
--- a/common/rfb/CMakeLists.txt
+++ b/common/rfb/CMakeLists.txt
@ -82,6 +82,12 @@ endif(WIN32)

 set(RFB_LIBRARIES ${JPEG_LIBRARIES} ${PNG_LIBRARIES} os rdr Xregion)

+cmake_host_system_information(RESULT DISTRO QUERY DISTRIB_INFO)
+if ((CMAKE_CXX_COMPILER_ID MATCHES "GNU" AND CMAKE_CXX_COMPILER_VERSION VERSION_LESS 10) OR
+(DISTRO_PLATFORM_ID MATCHES "platform:el8"))
+  set(RFB_LIBRARIES ${RFB_LIBRARIES} tbb)
+endif ()
+
 if(HAVE_PAM)
  set(RFB_SOURCES ${RFB_SOURCES} UnixPasswordValidator.cxx
    UnixPasswordValidator.h pam.c pam.h)
--- a/common/rfb/EncodeManager.cxx
+++ b/common/rfb/EncodeManager.cxx
@ -19,8 +19,7 @@
 * USA.
 */

-#include <omp.h>
-#include <stdlib.h>
+#include <cstdlib>

 #include <rfb/cpuid.h>
 #include <rfb/EncCache.h>
@ -44,6 +43,7 @@
 #include <rfb/TightJPEGEncoder.h>
 #include <rfb/TightWEBPEncoder.h>
 #include <rfb/TightQOIEncoder.h>
+#include <execution>

 using namespace rfb;

@ -93,9 +93,9 @@ struct RectInfo {
 };

 struct QualityInfo {
-  struct timeval lastUpdate;
+  struct timeval lastUpdate{};
  Rect rect;
-  unsigned score;
+  unsigned score{};
 };

 };
@ -880,14 +880,12 @@ void EncodeManager::findSolidRect(const Rect& rect, Region *changed,
  }
 }

-void EncodeManager::checkWebpFallback(const struct timeval *start) {
+void EncodeManager::checkWebpFallback(const timeval *start) {
    // Have we taken too long for the frame? If so, drop from WEBP to JPEG
-    if (start && activeEncoders[encoderFullColour] == encoderTightWEBP && !webpTookTooLong) {
-        unsigned us;
-        us = msSince(start) * 1024;
+    if (start && activeEncoders[encoderFullColour] == encoderTightWEBP && !webpTookTooLong.load(std::memory_order_relaxed)) {
+        const auto us = msSince(start) * 1024;
        if (us > webpFallbackUs)
-            #pragma omp atomic
-            webpTookTooLong |= true;
+            webpTookTooLong.store(true, std::memory_order_relaxed);
    }
 }

@ -1122,18 +1120,13 @@ void EncodeManager::writeRects(const Region& changed, const PixelBuffer* pb,
                               const bool mainScreen)
 {
  std::vector<Rect> rects, subrects, scaledrects;
-  std::vector<Rect>::const_iterator rect;
  std::vector<uint8_t> encoderTypes;
  std::vector<uint8_t> isWebp, fromCache;
  std::vector<Palette> palettes;
  std::vector<std::vector<uint8_t> > compresseds;
  std::vector<uint32_t> ms;
-  uint32_t i;

-  if (rfb::Server::rectThreads > 0)
-    omp_set_num_threads(rfb::Server::rectThreads);
-
-  webpTookTooLong = false;
+  webpTookTooLong.store(false, std::memory_order_relaxed);
  changed.get_rects(&rects);

  // Update stats
@ -1148,18 +1141,18 @@ void EncodeManager::writeRects(const Region& changed, const PixelBuffer* pb,

  subrects.reserve(rects.size() * 1.5f);

-  for (rect = rects.begin(); rect != rects.end(); ++rect) {
-    int w, h, sw, sh;
+  for (const auto& rect : rects) {
+    int sw, sh;
    Rect sr;

-    w = rect->width();
-    h = rect->height();
+    const auto w = rect.width();
+    const auto h = rect.height();

    // No split necessary?
    if ((((w*h) < SubRectMaxArea) && (w < SubRectMaxWidth)) ||
        (videoDetected && !encoders[encoderTightWEBP]->isSupported())) {
-      subrects.push_back(*rect);
-      trackRectQuality(*rect);
+      subrects.push_back(rect);
+      trackRectQuality(rect);
      continue;
    }

@ -1170,15 +1163,15 @@ void EncodeManager::writeRects(const Region& changed, const PixelBuffer* pb,

    sh = SubRectMaxArea / sw;

-    for (sr.tl.y = rect->tl.y; sr.tl.y < rect->br.y; sr.tl.y += sh) {
+    for (sr.tl.y = rect.tl.y; sr.tl.y < rect.br.y; sr.tl.y += sh) {
      sr.br.y = sr.tl.y + sh;
-      if (sr.br.y > rect->br.y)
-        sr.br.y = rect->br.y;
+      if (sr.br.y > rect.br.y)
+        sr.br.y = rect.br.y;

-      for (sr.tl.x = rect->tl.x; sr.tl.x < rect->br.x; sr.tl.x += sw) {
+      for (sr.tl.x = rect.tl.x; sr.tl.x < rect.br.x; sr.tl.x += sw) {
        sr.br.x = sr.tl.x + sw;
-        if (sr.br.x > rect->br.x)
-          sr.br.x = rect->br.x;
+        if (sr.br.x > rect.br.x)
+          sr.br.x = rect.br.x;

        subrects.push_back(sr);
        trackRectQuality(sr);
@ -1186,13 +1179,18 @@ void EncodeManager::writeRects(const Region& changed, const PixelBuffer* pb,
    }
  }

-  encoderTypes.resize(subrects.size());
-  isWebp.resize(subrects.size());
-  fromCache.resize(subrects.size());
-  palettes.resize(subrects.size());
-  compresseds.resize(subrects.size());
-  scaledrects.resize(subrects.size());
-  ms.resize(subrects.size());
+  const size_t subrects_size = subrects.size();
+
+  std::vector<size_t> indices(subrects_size);
+  std::iota(std::begin(indices), std::end(indices), 0);
+
+  encoderTypes.resize(subrects_size);
+  isWebp.resize(subrects_size);
+  fromCache.resize(subrects_size);
+  palettes.resize(subrects_size);
+  compresseds.resize(subrects_size);
+  scaledrects.resize(subrects_size);
+  ms.resize(subrects_size);

  // In case the current resolution is above the max video res, and video was detected,
  // scale to that res, keeping aspect ratio
@ -1224,7 +1222,7 @@ void EncodeManager::writeRects(const Region& changed, const PixelBuffer* pb,
      break;
    }

-    for (i = 0; i < subrects.size(); ++i) {
+    for (uint32_t i = 0; i < subrects_size; ++i) {
      const Rect old = scaledrects[i] = subrects[i];
      scaledrects[i].br.x *= diff;
      scaledrects[i].br.y *= diff;
@ -1249,15 +1247,15 @@ void EncodeManager::writeRects(const Region& changed, const PixelBuffer* pb,
  }
  scalingTime = msSince(&scalestart);

-  #pragma omp parallel for schedule(dynamic, 1)
-  for (i = 0; i < subrects.size(); ++i) {
-    encoderTypes[i] = getEncoderType(subrects[i], pb, &palettes[i], compresseds[i],
-                                     &isWebp[i], &fromCache[i],
-                                     scaledpb, scaledrects[i], ms[i]);
-    checkWebpFallback(start);
-  }
+    std::for_each(std::execution::par_unseq, std::begin(indices), std::end(indices), [&](size_t i)
+    {
+        encoderTypes[i] = getEncoderType(subrects[i], pb, &palettes[i], compresseds[i],
+                        &isWebp[i], &fromCache[i],
+                        scaledpb, scaledrects[i], ms[i]);
+        checkWebpFallback(start);
+    });

-  for (i = 0; i < subrects.size(); ++i) {
+  for (uint32_t i = 0; i < subrects_size; ++i) {
    if (encoderTypes[i] == encoderFullColour) {
      if (isWebp[i])
        webpstats.ms += ms[i];
@ -1269,7 +1267,7 @@ void EncodeManager::writeRects(const Region& changed, const PixelBuffer* pb,
  if (start) {
    encodingTime = msSince(start);

-    if (vlog.getLevel() >= vlog.LEVEL_DEBUG) {
+    if (vlog.getLevel() >= rfb::LogWriter::LEVEL_DEBUG) {
      framesSinceEncPrint++;
      if (maxEncodingTime < encodingTime)
        maxEncodingTime = encodingTime;
@ -1284,11 +1282,11 @@ void EncodeManager::writeRects(const Region& changed, const PixelBuffer* pb,
    }
  }

-  if (webpTookTooLong)
+  if (webpTookTooLong.load(std::memory_order_relaxed))
    activeEncoders[encoderFullColour] = encoderTightJPEG;

-  for (i = 0; i < subrects.size(); ++i) {
-    if (encCache->enabled && compresseds[i].size() && !fromCache[i] &&
+  for (uint32_t i = 0; i < subrects_size; ++i) {
+    if (encCache->enabled && !compresseds[i].empty() && !fromCache[i] &&
        !encoders[encoderTightQOI]->isSupported()) {
      void *tmp = malloc(compresseds[i].size());
      memcpy(tmp, &compresseds[i][0], compresseds[i].size());
--- a/common/rfb/EncodeManager.h
+++ b/common/rfb/EncodeManager.h
@ -29,9 +29,9 @@
 #include <rfb/Region.h>
 #include <rfb/Timer.h>
 #include <rfb/UpdateTracker.h>
-#include <rfb/util.h>

 #include <stdint.h>
+#include <atomic>
 #include <sys/time.h>

 namespace rfb {
@ -50,7 +50,7 @@ namespace rfb {
  class EncodeManager: public Timer::Callback {
  public:
    EncodeManager(SConnection* conn, EncCache *encCache);
-    ~EncodeManager();
+    ~EncodeManager() override;

    void logStats();

@ -72,10 +72,10 @@ namespace rfb {
        encodingTime = 0;
    };

-    unsigned getEncodingTime() const {
+    [[nodiscard]] unsigned getEncodingTime() const {
        return encodingTime;
    };
-    unsigned getScalingTime() const {
+    [[nodiscard]] unsigned getScalingTime() const {
        return scalingTime;
    };

@ -124,7 +124,8 @@ namespace rfb {
                           uint8_t *fromCache,
                           const PixelBuffer *scaledpb, const Rect& scaledrect,
                           uint32_t &ms) const;
-    virtual bool handleTimeout(Timer* t);
+
+    bool handleTimeout(Timer* t) override;

    bool checkSolidTile(const Rect& r, const rdr::U8* colourValue,
                        const PixelBuffer *pb);
@ -199,7 +200,7 @@ namespace rfb {
    size_t curMaxUpdateSize;
    unsigned webpFallbackUs;
    unsigned webpBenchResult;
-    bool webpTookTooLong;
+    std::atomic<bool> webpTookTooLong{false};
    unsigned encodingTime;
    unsigned maxEncodingTime, framesSinceEncPrint;
    unsigned scalingTime;
@ -208,14 +209,14 @@ namespace rfb {

    class OffsetPixelBuffer : public FullFramePixelBuffer {
    public:
-      OffsetPixelBuffer() {}
-      virtual ~OffsetPixelBuffer() {}
+      OffsetPixelBuffer() = default;
+      ~OffsetPixelBuffer() override = default;

      void update(const PixelFormat& pf, int width, int height,
                  const rdr::U8* data_, int stride);

    private:
-      virtual rdr::U8* getBufferRW(const Rect& r, int* stride);
+      rdr::U8* getBufferRW(const Rect& r, int* stride) override;
    };
  };