Give a phone without a Qualcomm SoC the generic WebGPU runtime

The APK's ONNX Runtime is the QNN build, which carries no WebGPU, so a
non-Qualcomm phone had nothing above the CPU provider. The APK now also
carries Microsoft's stock onnxruntime-android 1.29.0 (32 MB) as
libonnxruntime_generic.so, and the app offers it after the QNN build.

The runtime search stops at a perfect fit, so on a Qualcomm device the
QNN build — listed first — is all that is opened, and the generic build
never loads beside it. A Qualcomm SoC is read from ro.soc.manufacturer
or, before Android 12, from Qualcomm's FastRPC library being present:
a Qualcomm device mistaken for another would trade its Hexagon for the
generic rung.

apps/darkroom-android/src/lib.rs

@@ -451,8 +451,13 @@ fn unpack_bundled_models(app: &slint::android::AndroidApp) {
     // on a first launch they were not on disk until this line. The runtime
     // is in the APK's native library directory beside `libdarkroom.so`,
     // which is also where Qualcomm's DSP loader has to be pointed for the
-    // Hexagon skel (docs/dev/inference.md §3, §8).
+    // Hexagon skel (docs/dev/inference.md §3, §8). Two of them: the QNN
-    dr_ui::inference::init(native_library_dir().into_iter().collect());
+    // build, and the generic WebGPU build by its file name, which the
+    // engine opens only when the first does not fit the SoC (§3.2).
+    let runtimes = native_library_dir()
+        .map(|dir| vec![dir.clone(), dir.join("libonnxruntime_generic.so")])
+        .unwrap_or_default();
+    dr_ui::inference::init(runtimes);
 }
 
 /// The directory the system unpacked this APK's native libraries into.

core/dr-inference-engine/src/api.rs

@@ -58,7 +58,10 @@ pub fn ensure_installed() {
 /// Best fit, not first found (§3.2): a device can hold several runtimes —
 /// the package's OpenVINO build, a CUDA build the user fetched, the
 /// distribution's ROCm build — and each carries one vendor's providers.
-/// Between equals, the earlier directory wins, as it always has.
+/// Between equals, the earlier directory wins, as it always has, and a
+/// runtime that fits perfectly ends the search: the APK's QNN build on a
+/// Qualcomm tablet is found first, and the generic build beside it is
+/// never opened there.
 /// `DARKROOM_ORT_DIR`, when it loads, wins outright: it is how a person
 /// says which runtime they mean.
 pub fn install(dirs: &[PathBuf]) -> Runtime {
@@ -110,7 +113,11 @@ fn install_best(dirs: &[PathBuf]) -> Option<Runtime> {
                     found.push(f);
                     break;
                 }
+                let perfect = gpus.score(&f.providers) >= crate::hardware::PERFECT;
                 found.push(f);
+                if perfect {
+                    break;
+                }
             }
             Err(e) => log::info!("inference: no runtime in {}: {e}", dir.display()),
         }

core/dr-inference-engine/src/hardware.rs

@@ -19,6 +19,10 @@ pub struct Gpus {
     pub qualcomm: bool,
 }
 
+/// The score of a runtime whose vendor rung matches the device's GPU.
+/// Nothing beats it, so the search stops there.
+pub const PERFECT: u32 = 3;
+
 impl Gpus {
     /// How well a runtime offering `providers` fits this device. The vendor
     /// rungs score above OpenVINO because a machine with an Intel iGPU and
@@ -29,10 +33,10 @@ impl Gpus {
         providers
             .iter()
             .map(|p| match p.as_str() {
-                "TensorrtExecutionProvider" | "CUDAExecutionProvider" if self.nvidia => 3,
+                "TensorrtExecutionProvider" | "CUDAExecutionProvider" if self.nvidia => PERFECT,
-                "MIGraphXExecutionProvider" if self.amd_rocm => 3,
+                "MIGraphXExecutionProvider" if self.amd_rocm => PERFECT,
-                "QNNExecutionProvider" if self.qualcomm => 3,
+                "QNNExecutionProvider" if self.qualcomm => PERFECT,
-                "CoreMLExecutionProvider" => 3,
+                "CoreMLExecutionProvider" => PERFECT,
                 "OpenVINOExecutionProvider" if self.intel => 2,
                 "WebGpuExecutionProvider" => 1,
                 _ => 0,
@@ -91,9 +95,21 @@ pub fn detect() -> Gpus {
 
 #[cfg(target_os = "android")]
 pub fn detect() -> Gpus {
+    // `ro.soc.manufacturer` exists from Android 12. Qualcomm's FastRPC
+    // library, which the Hexagon path loads anyway, is the second witness:
+    // a Qualcomm device read as anything else would trade its NPU for the
+    // generic rung.
     let soc = crate::probe::system_property("ro.soc.manufacturer");
+    let fastrpc = [
+        "/vendor/lib64/libcdsprpc.so",
+        "/system/vendor/lib64/libcdsprpc.so",
+    ]
+    .iter()
+    .any(|p| std::path::Path::new(p).exists());
     Gpus {
-        qualcomm: soc.eq_ignore_ascii_case("QTI") || soc.eq_ignore_ascii_case("Qualcomm"),
+        qualcomm: soc.eq_ignore_ascii_case("QTI")
+            || soc.eq_ignore_ascii_case("Qualcomm")
+            || fastrpc,
         ..Gpus::default()
     }
 }

docker/android/assemble-apk.sh

@@ -263,10 +263,12 @@ cp "${DEX}" "${OUT}/staging/classes.dex"
 # native library directory. The build links none of it — the app dlopens
 # `libonnxruntime.so` at launch and runs on tract if it is not there — so an
 # APK without these is a slower app, not a broken one, and `RUNTIME_DIR=none`
-# builds exactly that. 174 MB for the default set; the script says which
+# builds exactly that. 206 MB for the default set — 32 MB of it the generic
-# Hexagon generations that buys.
+# WebGPU build for a phone without a Qualcomm SoC; the script says which
+# Hexagon generations the rest buys.
 if [[ "${RUNTIME_DIR}" != "none" ]]; then
-    if [[ ! -f "${RUNTIME_DIR}/lib/libonnxruntime.so" ]]; then
+    if [[ ! -f "${RUNTIME_DIR}/lib/libonnxruntime.so" \
+          || ! -f "${RUNTIME_DIR}/lib/libonnxruntime_generic.so" ]]; then
         "${REPO}/tools/fetch-android-runtime.sh" "${RUNTIME_DIR}"
     fi
     cp "${RUNTIME_DIR}"/lib/*.so "${OUT}/staging/lib/${ABI}/"

docs/dev/traceability.md

@@ -141,7 +141,7 @@ _None._
 | FR-PLAT-AND-3 | [`core/dr-catalog/src/jobs.rs:1`](../../core/dr-catalog/src/jobs.rs#L1), [`core/dr-catalog/src/runner.rs:1`](../../core/dr-catalog/src/runner.rs#L1), [`core/dr-catalog/tests/every_queued_kind_has_a_consumer.rs:1`](../../core/dr-catalog/tests/every_queued_kind_has_a_consumer.rs#L1) |
 | FR-PLAT-AND-4 | [`core/dr-catalog/src/runner.rs:1`](../../core/dr-catalog/src/runner.rs#L1) |
 | FR-PLAT-AND-5 | [`apps/darkroom-android/src/lib.rs:166`](../../apps/darkroom-android/src/lib.rs#L166), [`core/dr-gpu/src/adjust.rs:1472`](../../core/dr-gpu/src/adjust.rs#L1472), [`core/dr-gpu/src/detail.rs:161`](../../core/dr-gpu/src/detail.rs#L161), [`core/dr-gpu/src/detail.rs:533`](../../core/dr-gpu/src/detail.rs#L533), [`core/dr-gpu/tests/detail_stage.rs:415`](../../core/dr-gpu/tests/detail_stage.rs#L415), [`ui/dr-ui/src/develop/render.rs:654`](../../ui/dr-ui/src/develop/render.rs#L654), [`ui/dr-ui/src/identity_ui.rs:125`](../../ui/dr-ui/src/identity_ui.rs#L125), [`ui/dr-ui/src/lib.rs:1309`](../../ui/dr-ui/src/lib.rs#L1309), [`ui/dr-ui/src/lib.rs:1745`](../../ui/dr-ui/src/lib.rs#L1745), [`ui/dr-ui/src/memory.rs:163`](../../ui/dr-ui/src/memory.rs#L163), [`ui/dr-ui/src/memory.rs:1`](../../ui/dr-ui/src/memory.rs#L1) |
-| FR-PLAT-AND-6 | [`apps/darkroom-android/src/lib.rs:472`](../../apps/darkroom-android/src/lib.rs#L472) |
+| FR-PLAT-AND-6 | [`apps/darkroom-android/src/lib.rs:477`](../../apps/darkroom-android/src/lib.rs#L477) |
 | FR-PLAT-LIN-1 | [`core/dr-types/src/settings.rs:1`](../../core/dr-types/src/settings.rs#L1), [`platform/dr-plat/src/dirs.rs:1`](../../platform/dr-plat/src/dirs.rs#L1), [`platform/dr-plat/src/storage.rs:344`](../../platform/dr-plat/src/storage.rs#L344), [`ui/dr-ui/src/lib.rs:1486`](../../ui/dr-ui/src/lib.rs#L1486), [`ui/dr-ui/src/preset_store.rs:1`](../../ui/dr-ui/src/preset_store.rs#L1), [`ui/dr-ui/src/settings_store.rs:1`](../../ui/dr-ui/src/settings_store.rs#L1) |
 | FR-PLAT-LIN-2 | [`platform/dr-plat/src/display.rs:1`](../../platform/dr-plat/src/display.rs#L1), [`platform/dr-plat/src/display/wayland.rs:1`](../../platform/dr-plat/src/display/wayland.rs#L1), [`platform/dr-plat/src/display/x11.rs:1`](../../platform/dr-plat/src/display/x11.rs#L1) |
 | FR-PLAT-WIN-1 | [`platform/dr-plat/src/dirs.rs:1`](../../platform/dr-plat/src/dirs.rs#L1) |

tools/fetch-android-runtime.sh

@@ -3,17 +3,23 @@
 #
 #   ./tools/fetch-android-runtime.sh [DEST]
 #
-# Two Maven artefacts, pinned to each other by ONNX Runtime's own POM:
+# Three Maven artefacts, the first two pinned to each other by ONNX Runtime's
+# own POM:
 #
 #   com.microsoft.onnxruntime:onnxruntime-android-qnn   MIT
 #   com.qualcomm.qti:qnn-runtime                        Qualcomm AI Engine Direct SDK licence
+#   com.microsoft.onnxruntime:onnxruntime-android       MIT
 #
-# The first is ONNX Runtime built with the CPU, QNN, XNNPACK, NNAPI and WebGPU
+# The first is ONNX Runtime built with the CPU and QNN providers; the second
-# providers; the second is Qualcomm's HTP backend — the ARM-side compiler and
+# is Qualcomm's HTP backend — the ARM-side compiler and the per-generation
-# the per-generation Hexagon "skel" the DSP loads. Both ship as AARs whose
+# Hexagon "skel" the DSP loads. The third is Microsoft's stock build, which
-# `jni/arm64-v8a/` is what an APK's `lib/arm64-v8a/` wants, so this script
+# carries WebGPU (Dawn on Vulkan) and the QNN build does not: the generic
-# unpacks exactly that and nothing else, plus the licence texts, which travel
+# rung for a phone without a Qualcomm SoC (§3.2). It lands as
-# with the libraries (§3.1).
+# `libonnxruntime_generic.so` beside the QNN build; the engine opens the QNN
+# build first, and on a Qualcomm device stops there. All three ship as AARs
+# whose `jni/arm64-v8a/` is what an APK's `lib/arm64-v8a/` wants, so this
+# script unpacks exactly that and nothing else, plus the licence texts, which
+# travel with the libraries (§3.1).
 #
 # ## What is and is not taken from the Qualcomm package
 #
@@ -65,6 +71,8 @@ fetch com.microsoft.onnxruntime onnxruntime-android-qnn "${ORT_VERSION}" \
     "${DEST}/aar/onnxruntime-android-qnn-${ORT_VERSION}.aar"
 fetch com.qualcomm.qti qnn-runtime "${QNN_VERSION}" \
     "${DEST}/aar/qnn-runtime-${QNN_VERSION}.aar"
+fetch com.microsoft.onnxruntime onnxruntime-android "${ORT_VERSION}" \
+    "${DEST}/aar/onnxruntime-android-${ORT_VERSION}.aar"
 
 # The ONNX Runtime POM names the QNN version it was built against; a pair
 # that disagrees loads and then fails at the first graph, which is the kind
@@ -82,6 +90,8 @@ rm -rf "${DEST}/lib"
 mkdir -p "${DEST}/lib"
 unzip -q -o -j "${DEST}/aar/onnxruntime-android-qnn-${ORT_VERSION}.aar" \
     'jni/arm64-v8a/libonnxruntime.so' -d "${DEST}/lib"
+unzip -q -o -p "${DEST}/aar/onnxruntime-android-${ORT_VERSION}.aar" \
+    'jni/arm64-v8a/libonnxruntime.so' > "${DEST}/lib/libonnxruntime_generic.so"
 members=(jni/arm64-v8a/libQnnHtp.so jni/arm64-v8a/libQnnHtpPrepare.so jni/arm64-v8a/libQnnSystem.so)
 for arch in ${QNN_HTP_ARCHS}; do
     members+=("jni/arm64-v8a/libQnnHtpV${arch}Skel.so" "jni/arm64-v8a/libQnnHtpV${arch}Stub.so")