opencv: support cross-compilation and add sample function (#44)

* opencv: support cross-compilation and add sample function

* nits: clean-up

* func: pass --cmdline

* opencv: add example doing k-NN inference

* gha: re-generate wasm if we add new function

* gha: few nits

* cpp: bump for latest s3

* gha: make cache depend on functions too

.github/workflows/tests.yml

@@ -28,12 +28,12 @@ jobs:
 
   # Work-out whether we need to re-build the examples. We need to re-build the
   # examples if either _any_ fo the examples has changed, or the WASM cache
-  # has expired
+  # has expired, or the ./funcs folder have changed
   needs-build:
     if: github.event.pull_request.draft == false
     runs-on: ubuntu-latest
     outputs:
-      needs-wasm: ${{ steps.wasm-cache.outputs.cache-hit != 'true' }}
+      needs-wasm: ${{ (steps.wasm-cache.outputs.cache-hit != 'true') || (steps.filter.outputs.funcs-changed == 'true') }}
     steps:
       - name: "Checkout code"
         uses: actions/checkout@v4
@@ -52,6 +52,14 @@ jobs:
           path: ./wasm
           key: wasm-${{ steps.submodule-commit.outputs.digest }}
           lookup-only: true
+      # Check if any of the submodules have been modified
+      - uses: dorny/paths-filter@v3
+        id: filter
+        with:
+          filters: |
+            funcs-changed:
+              - 'func/**'
+
 
   build-examples:
     needs: needs-build
@@ -73,6 +81,8 @@ jobs:
           rustup default stable
           rustup target add wasm32-wasip1
           ./bin/inv_wrapper.sh rabe jwt
+      - name: "Build OpenCV"
+        run: ./bin/inv_wrapper.sh opencv
       - name: "Build FFmpeg"
         run: ./bin/inv_wrapper.sh ffmpeg
       - name: "Build ImageMagick"
@@ -93,14 +103,22 @@ jobs:
         run: ./bin/inv_wrapper.sh tensorflow
       - name: "Build PolyBench/C"
         run: ./bin/inv_wrapper.sh polybench polybench --native
+      - name: "Manually re-compile libz (the sysroot has a version with atomics)"
+        run: |
+          git submodule update --init
+          ./bin/inv_wrapper.sh zlib
+        working-directory: ./cpp
       - name: "Build functions used in the tests"
         run: ./bin/inv_wrapper.sh func.tests
       - name: "Get CPP/Python commits"
         id: submodule-commit
         run: |
           apt install -y zstd
           git config --global --add safe.directory "$GITHUB_WORKSPACE"
-          echo "digest=$(git submodule status | awk '{ print $1; }' | md5sum | cut -d' ' -f1)" >> $GITHUB_OUTPUT
+          submodule_digest=$(git submodule status | awk '{ print $1; }' | md5sum | awk '{ print $1}')
+          func_digest=$(find 'func' -type f -exec md5sum {} + | sort | md5sum | awk '{ print $1 }')
+          echo "${submodule_digest}-${func_digest}"
+          echo "digest=$(echo -n '${submodule_digest}-${func_digest}' | md5sum | awk '{ print $1 }')" >> $GITHUB_OUTPUT
           # Also move to a different path to restore from
           mv /usr/local/faasm/wasm ./wasm
       # If we are here we _always_ want to overwrite the cache
@@ -168,7 +186,11 @@ jobs:
         run: |
           sudo apt install -y zstd
           git config --global --add safe.directory "$GITHUB_WORKSPACE"
-          echo "digest=$(git submodule status | awk '{ print $1; }' | md5sum | cut -d' ' -f1)" >> $GITHUB_OUTPUT
+          submodule_digest=$(git submodule status | awk '{ print $1; }' | md5sum | awk '{ print $1}')
+          func_digest=$(find 'func' -type f -exec md5sum {} + | sort | md5sum | awk '{ print $1 }')
+          echo "${submodule_digest}-${func_digest}"
+          echo "digest=$(echo -n '${submodule_digest}-${func_digest}' | md5sum | awk '{ print $1 }')" >> $GITHUB_OUTPUT
+          # echo "digest=$(git submodule status | awk '{ print $1; }' | md5sum | cut -d' ' -f1)" >> $GITHUB_OUTPUT
       - name: "Get WASM cache"
         uses: actions/cache/restore@v4
         id: cpp-wasm-cache
@@ -182,6 +204,14 @@ jobs:
       - name: "Print logs if wasm-upload fails"
         if: failure() && (steps.wasm-upload.outcome == 'failure')
         run: faasmctl logs -s upload
+      - name: "Run OpenCV check"
+        if: "contains(env.FAASM_WASM_VM, 'wamr')"
+        timeout-minutes: 2
+        run: faasmctl invoke opencv pca --cmdline "faasm://opencv/composers"
+      - name: "Run OpenCV k-NN Inference"
+        if: "contains(env.FAASM_WASM_VM, 'wamr')"
+        timeout-minutes: 2
+        run: faasmctl invoke opencv check --cmdline "faasm://opencv/bus_photo.bmp faasm://opencv/out.bmp"
       - name: "Run MPI kernels"
         timeout-minutes: 2
         run: |

.gitmodules

@@ -54,3 +54,6 @@
 [submodule "examples/tless-jwt"]
 	path = examples/tless-jwt
 	url = https://github.com/faasm/tless-jwt.git
+[submodule "examples/opencv"]
+	path = examples/opencv
+	url = https://github.com/opencv/opencv.git

func/CMakeLists.txt

@@ -35,5 +35,6 @@ add_subdirectory(ffmpeg)
 add_subdirectory(jwt)
 add_subdirectory(lammps)
 add_subdirectory(mpi)
+add_subdirectory(opencv)
 add_subdirectory(rabe)
 add_subdirectory(tf)

func/opencv/CMakeLists.txt

@@ -0,0 +1,7 @@
+set(FAASM_USER opencv)
+
+faasm_example_func(check check.cpp)
+target_link_libraries(opencv_check opencv_core opencv_imgcodecs opencv_imgproc)
+
+faasm_example_func(pca pca.cpp)
+target_link_libraries(opencv_pca opencv_core opencv_imgcodecs opencv_imgproc opencv_ml z)

func/opencv/check.cpp

@@ -0,0 +1,46 @@
+// Even when threading is completely disabled, OpenCV assumes the C++ library
+// has been built with threading support, and typedefs (no-ops) these two
+// symbols. To prevent an undefined symbol error, we define them here.
+namespace std {
+    class recursive_mutex {
+    public:
+        void lock() {}
+        bool try_lock() { return true; }
+        void unlock() {}
+    };
+
+    template <typename T>
+    class lock_guard {
+    public:
+        explicit lock_guard(T&) {}
+    };
+}
+
+#include <opencv2/opencv.hpp>
+#include <opencv2/imgcodecs.hpp>
+#include <opencv2/imgproc.hpp>
+#include <iostream>
+
+int main(int argc, char** argv) {
+    if (argc != 3) {
+        std::cerr << "Usage: <image_name_in.bmp> <image_name_out.bmp>" << std::endl;
+        return 1;
+    }
+
+    // Load an image
+    cv::Mat img = cv::imread(argv[1], cv::IMREAD_COLOR);
+    if (img.empty()) {
+        std::cerr << "Error: Could not load image!" << std::endl;
+        return -1;
+    }
+
+    // Convert the image to grayscale
+    cv::Mat gray;
+    cv::cvtColor(img, gray, cv::COLOR_BGR2GRAY);
+
+    // Save the grayscale image
+    cv::imwrite(argv[2], gray);
+    std::cout << "Image has been converted to grayscale and saved as " << argv[2] << std::endl;
+
+    return 0;
+}

func/opencv/pca.cpp

@@ -0,0 +1,88 @@
+// Even when threading is completely disabled, OpenCV assumes the C++ library
+// has been built with threading support, and typedefs (no-ops) these two
+// symbols. To prevent an undefined symbol error, we define them here.
+namespace std {
+    class recursive_mutex {
+    public:
+        void lock() {}
+        bool try_lock() { return true; }
+        void unlock() {}
+    };
+
+    template <typename T>
+    class lock_guard {
+    public:
+        explicit lock_guard(T&) {}
+    };
+}
+
+#include <opencv2/opencv.hpp>
+#include <opencv2/imgcodecs.hpp>
+#include <opencv2/imgproc.hpp>
+#include <opencv2/ml.hpp>
+#include <iostream>
+#include <filesystem>
+
+void loadImages(const std::string& folder, std::vector<cv::Mat>& data, std::vector<int>& labels) {
+    int label = 0;
+    for (const auto& entry : std::filesystem::directory_iterator(folder)) {
+        if (entry.is_regular_file()) {
+
+            std::cout << " - loading image: " << entry.path().string() << " (" << label << ")" << std::endl;
+            cv::Mat img = cv::imread(entry.path().string(), cv::IMREAD_GRAYSCALE);
+            if (!img.empty()) {
+                cv::resize(img, img, cv::Size(64, 64));
+                data.push_back(img.reshape(1, 1));
+                labels.push_back(label);
+            }
+        }
+
+        label++;
+    }
+}
+
+int main(int argc, char** argv) {
+    if (argc != 2) {
+        std::cerr << "Usage: <image_path>" << std::endl;
+        return 1;
+    }
+
+    // Load images
+    std::cout << "Beginning to load images..." << std::endl;
+    std::vector<cv::Mat> images;
+    std::vector<int> labels;
+    loadImages(argv[1], images, labels);
+    std::cout << "Images loaded!" << std::endl;
+
+    // Convert data to a single matrix
+    std::cout << "Converting data..." << std::endl;
+    cv::Mat data;
+    cv::vconcat(images, data);
+    data.convertTo(data, CV_32F);
+    std::cout << "Data converted!" << std::endl;
+
+    // Perform PCA with 10 principal components
+    std::cout << "Performing PCA analysis..." << std::endl;
+    cv::PCA pca(data, cv::Mat(), cv::PCA::DATA_AS_ROW, 10);
+    cv::Mat pcaResult;
+    pca.project(data, pcaResult);
+    std::cout << "PCA on images succeded!" << std::endl;
+
+    // Prepare labels for training
+    cv::Mat labelsMat(labels);
+    labelsMat.convertTo(labelsMat, CV_32S);
+
+    // Train k-NN classifier with PCA result
+    cv::Ptr<cv::ml::KNearest> knn = cv::ml::KNearest::create();
+    knn->setDefaultK(3);
+    knn->train(pcaResult, cv::ml::ROW_SAMPLE, labelsMat);
+    std::cout << "Training k-NN classifier succeeded!" << std::endl;
+
+    // Perform a prediction on the first sample as an example
+    cv::Mat sample = pcaResult.row(0);
+    cv::Mat response;
+    knn->findNearest(sample, knn->getDefaultK(), response);
+    std::cout << "Predicted label: " << response.at<float>(0, 0) << std::endl;
+
+    return 0;
+}

tasks/__init__.py

@@ -13,6 +13,7 @@
 from . import lammps
 from . import libpng
 from . import lulesh
+from . import opencv
 from . import polybench
 from . import rabe
 from . import tensorflow
@@ -32,6 +33,7 @@
     lammps,
     libpng,
     lulesh,
+    opencv,
     polybench,
     rabe,
     tensorflow,

tasks/env.py

@@ -43,6 +43,22 @@
         join(EXAMPLES_DATA_HOST_DIR, "grace_hopper.bmp"),
         join(EXAMPLES_DATA_BASE_DIR, "tflite", "grace_hopper.bmp"),
     ],
+    [
+        join(EXAMPLES_DATA_HOST_DIR, "bus_photo.bmp"),
+        join(EXAMPLES_DATA_BASE_DIR, "opencv", "bus_photo.bmp"),
+    ],
+    [
+        join(EXAMPLES_DATA_HOST_DIR, "tchaikovsky.bmp"),
+        join(EXAMPLES_DATA_BASE_DIR, "opencv", "composers", "tchaikovsky.bmp"),
+    ],
+    [
+        join(EXAMPLES_DATA_HOST_DIR, "wagner.bmp"),
+        join(EXAMPLES_DATA_BASE_DIR, "opencv", "composers", "wagner.bmp"),
+    ],
+    [
+        join(EXAMPLES_DATA_HOST_DIR, "beethoven.bmp"),
+        join(EXAMPLES_DATA_BASE_DIR, "opencv", "composers", "beethoven.bmp"),
+    ],
 ]
 
 

tasks/func.py

@@ -42,6 +42,8 @@ def tests(ctx, clean=False):
     """
     funcs = [
         ["ffmpeg", "check"],
+        ["opencv", "check"],
+        ["opencv", "pca"],
         ["rabe", "test"],
         ["tf", "check"],
         # TODO: this two functions are not used in the tests, as they are used