fix(ai_denoise): probe Vulkan loader before NCNN init (refs #30, #31)

NCNN's create_gpu_instance() dispatches through function pointers that
its internal simplevk fills only when libvulkan can be dlopened. On
macOS without MoltenVK / LunarG SDK installed, dlopen fails silently
and the next call dereferences uninitialised pointers, crashing with
SIGBUS in the .data segment (instruction abort, as reported in #31)
after three failed dlopen attempts (the symptom reported in #30).

Add a pre-flight platform-specific dlopen probe matching NCNN's own
loader names: vulkan-1.dll on Windows, libvulkan.1.dylib /
libvulkan.dylib on macOS, libvulkan.so.1 / libvulkan.so on Linux. If
the probe fails, skip create_gpu_instance() entirely and route the
denoiser through the CPU code path. macOS gets a tailored log message
naming MoltenVK / LunarG so users have a clear remediation.

CI: add scripts/ci_smoke_ai_denoise.sh -- generates a 512x512 PNG and
runs the binary with --force --region br:auto --denoise ai through
end-to-end. Wire it into all six build jobs (full + gwt-mini, x
windows/linux/macos) after the existing --version smoke test. This
exercises NcnnDenoiser::initialize() and would have caught the issues
on the macOS runner immediately.

Author: allenk

.github/workflows/build.yml

@@ -52,9 +52,13 @@ jobs:
         cmake --preset windows-x64-Release
         cmake --build --preset windows-x64-Release
 
-    - name: Test binary
+    - name: Test binary (--version)
       run: ./out/build/windows-x64-Release/GeminiWatermarkTool.exe --version
 
+    - name: Test binary (AI denoise pipeline)
+      shell: bash
+      run: bash scripts/ci_smoke_ai_denoise.sh ./out/build/windows-x64-Release/GeminiWatermarkTool.exe
+
     - name: Upload artifact
       uses: actions/upload-artifact@v4
       with:
@@ -103,9 +107,12 @@ jobs:
         cmake --preset linux-x64-Release
         cmake --build --preset linux-x64-Release
 
-    - name: Test binary
+    - name: Test binary (--version)
       run: ./out/build/linux-x64-Release/GeminiWatermarkTool --version
 
+    - name: Test binary (AI denoise pipeline)
+      run: bash scripts/ci_smoke_ai_denoise.sh ./out/build/linux-x64-Release/GeminiWatermarkTool
+
     - name: Upload artifact
       uses: actions/upload-artifact@v4
       with:
@@ -169,9 +176,12 @@ jobs:
         mv dist/GeminiWatermarkTool .
         rmdir dist
 
-    - name: Test binary
+    - name: Test binary (--version)
       run: ./GeminiWatermarkTool --version
 
+    - name: Test binary (AI denoise pipeline)
+      run: bash scripts/ci_smoke_ai_denoise.sh ./GeminiWatermarkTool
+
     - name: Upload artifact
       uses: actions/upload-artifact@v4
       with:
@@ -233,10 +243,14 @@ jobs:
           echo "--- size after UPX ---"
           ls -lh gwt-mini.exe
 
-      - name: Smoke test
+      - name: Smoke test (--version)
         shell: bash
         run: ./gwt-mini.exe --version
 
+      - name: Smoke test (AI denoise pipeline)
+        shell: bash
+        run: bash scripts/ci_smoke_ai_denoise.sh ./gwt-mini.exe
+
       - name: Upload artifact
         uses: actions/upload-artifact@v4
         with:
@@ -284,9 +298,12 @@ jobs:
           echo "--- size after UPX ---"
           ls -lh gwt-mini
 
-      - name: Smoke test
+      - name: Smoke test (--version)
         run: ./gwt-mini --version
 
+      - name: Smoke test (AI denoise pipeline)
+        run: bash scripts/ci_smoke_ai_denoise.sh ./gwt-mini
+
       - name: Upload artifact
         uses: actions/upload-artifact@v4
         with:
@@ -346,9 +363,12 @@ jobs:
           echo "--- final size (no UPX on macOS) ---"
           ls -lh gwt-mini
 
-      - name: Smoke test
+      - name: Smoke test (--version)
         run: ./gwt-mini --version
 
+      - name: Smoke test (AI denoise pipeline)
+        run: bash scripts/ci_smoke_ai_denoise.sh ./gwt-mini
+
       - name: Upload artifact
         uses: actions/upload-artifact@v4
         with:

scripts/ci_smoke_ai_denoise.sh

@@ -0,0 +1,78 @@
+#!/usr/bin/env bash
+# =============================================================================
+# CI smoke test: AI denoise initialisation + end-to-end run
+# =============================================================================
+# Exercises the NcnnDenoiser::initialize() path, including the Vulkan loader
+# probe and CPU fallback. Generates a synthetic 512x512 image, runs the binary
+# with --denoise ai through the full pipeline, and asserts that the process
+# exits successfully and produces an output file.
+#
+# This catches the class of issues seen in #30 / #31 -- where Vulkan loading
+# fails on macOS without MoltenVK and the AI denoise init path crashes with
+# SIGBUS instead of falling back to CPU. A plain `--version` smoke test does
+# not exercise this code path.
+#
+# Usage: ci_smoke_ai_denoise.sh <path-to-binary>
+# =============================================================================
+set -euo pipefail
+
+BIN="${1:?usage: $0 <path-to-binary>}"
+
+if [[ ! -x "$BIN" && ! -f "$BIN" ]]; then
+    echo "ERROR: binary not found: $BIN" >&2
+    exit 2
+fi
+
+WORKDIR="$(mktemp -d)"
+trap 'rm -rf "$WORKDIR"' EXIT
+
+INPUT="$WORKDIR/smoke_in.png"
+OUTPUT="$WORKDIR/smoke_out.png"
+
+PY=python3
+command -v "$PY" >/dev/null 2>&1 || PY=python
+
+echo "--- generating 512x512 synthetic test image (using $PY) ---"
+"$PY" - "$INPUT" <<'PY_EOF'
+import struct, sys, zlib
+
+def png(w, h, color=(128, 128, 128)):
+    def chunk(t, d):
+        return struct.pack(">I", len(d)) + t + d + struct.pack(">I", zlib.crc32(t + d))
+    sig = b"\x89PNG\r\n\x1a\n"
+    ihdr = chunk(b"IHDR", struct.pack(">IIBBBBB", w, h, 8, 2, 0, 0, 0))
+    row = b"\0" + bytes(color) * w
+    raw = row * h
+    idat = chunk(b"IDAT", zlib.compress(raw))
+    iend = chunk(b"IEND", b"")
+    return sig + ihdr + idat + iend
+
+with open(sys.argv[1], "wb") as f:
+    f.write(png(512, 512))
+PY_EOF
+ls -lh "$INPUT"
+
+echo "--- running AI denoise pipeline ---"
+# --force         skip detection (no real watermark on a uniform image)
+# --region br:auto  process the bottom-right region at the Gemini default position
+# --denoise ai    exercise the NCNN/Vulkan/CPU dispatch path
+"$BIN" --force --region br:auto --denoise ai -i "$INPUT" -o "$OUTPUT"
+RC=$?
+
+if [[ $RC -ne 0 ]]; then
+    echo "ERROR: binary exited with code $RC" >&2
+    exit $RC
+fi
+
+if [[ ! -f "$OUTPUT" ]]; then
+    echo "ERROR: expected output file not produced: $OUTPUT" >&2
+    exit 3
+fi
+
+OUT_SIZE=$(wc -c <"$OUTPUT")
+if [[ $OUT_SIZE -lt 1000 ]]; then
+    echo "ERROR: output file suspiciously small ($OUT_SIZE bytes)" >&2
+    exit 4
+fi
+
+echo "--- AI denoise smoke test PASSED (output=$OUT_SIZE bytes) ---"

src/core/ai_denoise.cpp

@@ -28,6 +28,13 @@
 #include <algorithm>
 #include <chrono>
 
+#if defined(_WIN32)
+#  define WIN32_LEAN_AND_MEAN
+#  include <windows.h>
+#else
+#  include <dlfcn.h>
+#endif
+
 // Embedded model data (defined in ai_denoise_model.cpp)
 namespace gwt::ai_model {
     const unsigned char* param_data();
@@ -36,6 +43,35 @@ namespace gwt::ai_model {
 
 namespace gwt {
 
+namespace {
+
+// Probe whether the Vulkan loader is dynamically loadable on this system.
+// NCNN's internal simplevk dlopens the loader by these exact names; if the
+// probe fails the loader is absent and ncnn::create_gpu_instance() will
+// dispatch through uninitialised function pointers and crash (SIGBUS on
+// macOS Apple Silicon, observed in issues #30 / #31).
+bool vulkan_loader_present() {
+#if defined(_WIN32)
+    HMODULE h = LoadLibraryA("vulkan-1.dll");
+    if (h) { FreeLibrary(h); return true; }
+    return false;
+#elif defined(__APPLE__)
+    // macOS does not ship a Vulkan loader. The user must install MoltenVK
+    // via the LunarG Vulkan SDK (or it must be bundled into the app).
+    void* h = dlopen("libvulkan.1.dylib", RTLD_LAZY | RTLD_LOCAL);
+    if (!h) h = dlopen("libvulkan.dylib", RTLD_LAZY | RTLD_LOCAL);
+    if (h) { dlclose(h); return true; }
+    return false;
+#else
+    void* h = dlopen("libvulkan.so.1", RTLD_LAZY | RTLD_LOCAL);
+    if (!h) h = dlopen("libvulkan.so", RTLD_LAZY | RTLD_LOCAL);
+    if (h) { dlclose(h); return true; }
+    return false;
+#endif
+}
+
+}  // namespace
+
 // Blob indices from binary param (string names not available in binary format)
 // These correspond to model_core.id.h: BLOB_in0 = 0, BLOB_out0 = 20
 static constexpr int BLOB_INPUT  = 0;
@@ -292,7 +328,22 @@ bool NcnnDenoiser::initialize() {
         return true;
     }
 
-    ncnn::create_gpu_instance();
+    // Probe the Vulkan loader BEFORE calling into NCNN. On systems without
+    // it (macOS without MoltenVK is the common case), NCNN's
+    // create_gpu_instance() dispatches through uninitialised function
+    // pointers and crashes -- see issues #30 and #31.
+    const bool vulkan_available = vulkan_loader_present();
+    if (vulkan_available) {
+        ncnn::create_gpu_instance();
+    } else {
+#if defined(__APPLE__)
+        spdlog::info("NcnnDenoiser: Vulkan loader unavailable on macOS "
+                     "(install MoltenVK or the LunarG Vulkan SDK for GPU "
+                     "acceleration); falling back to CPU");
+#else
+        spdlog::info("NcnnDenoiser: Vulkan loader unavailable, falling back to CPU");
+#endif
+    }
 
     // FP16 storage for efficiency, FP32 compute for accuracy
     m_impl->net.opt.use_fp16_packed = true;
@@ -301,7 +352,7 @@ bool NcnnDenoiser::initialize() {
     m_impl->net.opt.use_packing_layout = true;
 
     // Try GPU first, fall back to CPU
-    m_impl->gpu_enabled = m_impl->init_gpu();
+    m_impl->gpu_enabled = vulkan_available && m_impl->init_gpu();
     if (!m_impl->gpu_enabled) {
         m_impl->init_cpu();
     }