Merge branch 'polycam' into main

Author: jgibson2

backends/vulkan/cmake/ShaderLibrary.cmake

@@ -60,6 +60,14 @@ function(gen_vulkan_shader_lib_cpp shaders_path)
     )
   endif()
 
+  # Allow overriding GLSL PRECISION for fp16 shader variants. Empty / unset
+  # keeps upstream default (`highp`). Accepted values: highp, mediump, lowp.
+  if(EXECUTORCH_VULKAN_FP16_PRECISION)
+    list(APPEND GEN_SPV_ARGS "--fp16-precision"
+         "${EXECUTORCH_VULKAN_FP16_PRECISION}"
+    )
+  endif()
+
   # Ninja cannot expand wildcards (*) in DEPENDS lists.
   file(GLOB VULKAN_SHADERS "${shaders_path}/*.glsl" "${shaders_path}/*.glslh"
        "${shaders_path}/*.yaml" "${shaders_path}/*.h"

backends/vulkan/runtime/gen_vulkan_spv.py

@@ -663,6 +663,7 @@ def __init__(
         glslc_path: Optional[str],
         glslc_flags: str = "",
         replace_u16vecn: bool = False,
+        fp16_precision: str = "highp",
     ) -> None:
         if isinstance(src_dir_paths, str):
             self.src_dir_paths = [src_dir_paths]
@@ -678,6 +679,7 @@ def __init__(
         if "-Os" in self.glslc_flags_no_opt:
             self.glslc_flags_no_opt.remove("-Os")
         self.replace_u16vecn = replace_u16vecn
+        self.fp16_precision = fp16_precision
 
         self.src_files: Dict[str, str] = {}
         self.template_yaml_files: List[str] = []
@@ -857,6 +859,17 @@ def create_shader_params(
         for key, value in variant_params.items():
             shader_params[key] = value
 
+        # Optionally override PRECISION for half-precision variants. GLSL
+        # `mediump` is a hint the driver may use fp16 ALUs for arithmetic.
+        # On Mali GPUs it's typically honored; on Adreno it's typically
+        # ignored (harmless). Default is `highp` to match upstream behavior.
+        if (
+            self.fp16_precision != "highp"
+            and shader_params.get("DTYPE") == "half"
+            and shader_params.get("PRECISION") == "highp"
+        ):
+            shader_params["PRECISION"] = self.fp16_precision
+
         return shader_params
 
     def constructOutputMap(self) -> None:
@@ -1488,6 +1501,16 @@ def main(argv: List[str]) -> int:
         default=-1,
         help="Number of threads for shader compilation. -1 (default) uses all available CPU cores, 1 uses sequential compilation.",
     )
+    parser.add_argument(
+        "--fp16-precision",
+        choices=["highp", "mediump", "lowp"],
+        default="highp",
+        help=(
+            "GLSL PRECISION qualifier for DTYPE=half shader variants. "
+            "`mediump` lets drivers (notably Mali) use fp16 ALUs for arithmetic. "
+            "Default `highp` matches upstream behavior. Ignored on fp32 variants."
+        ),
+    )
     options = parser.parse_args()
 
     env = DEFAULT_ENV
@@ -1520,6 +1543,7 @@ def main(argv: List[str]) -> int:
         options.glslc_path,
         glslc_flags=glslc_flags_str,
         replace_u16vecn=options.replace_u16vecn,
+        fp16_precision=options.fp16_precision,
     )
     output_spv_files = shader_generator.generateSPV(
         options.output_path,

extension/android/executorch_android/src/main/java/org/pytorch/executorch/Module.java

@@ -265,6 +265,28 @@ public boolean etdump() {
   @DoNotStrip
   private native boolean etdumpNative();
 
+  /**
+   * Dump the ExecuTorch ETDump file to {@code outputPath}.
+   *
+   * @param outputPath absolute path to write the etdump file to.
+   * @return true if the etdump was successfully written, false otherwise.
+   */
+  @Experimental
+  public boolean etdump(String outputPath) {
+    mLock.lock();
+    try {
+      if (!mHybridData.isValid()) {
+        throw new IllegalStateException("Module has been destroyed");
+      }
+      return etdumpToNative(outputPath);
+    } finally {
+      mLock.unlock();
+    }
+  }
+
+  @DoNotStrip
+  private native boolean etdumpToNative(String outputPath);
+
   /**
    * Explicitly destroys the native Module object. Calling this method is not required, as the
    * native object will be destroyed when this object is garbage-collected. However, the timing of

extension/android/jni/jni_layer.cpp

@@ -472,36 +472,51 @@ class ExecuTorchJni : public facebook::jni::HybridClass<ExecuTorchJni> {
   }
 
   jboolean etdump() {
+    return etdump_to_path("/data/local/tmp/result.etdump");
+  }
+
+  jboolean etdumpTo(facebook::jni::alias_ref<jstring> outputPath) {
+    return etdump_to_path(outputPath->toStdString().c_str());
+  }
+
+ private:
+  jboolean etdump_to_path(const char* path) {
 #ifdef EXECUTORCH_ANDROID_PROFILING
     executorch::etdump::ETDumpGen* etdumpgen =
         (executorch::etdump::ETDumpGen*)module_->event_tracer();
     auto etdump_data = etdumpgen->get_etdump_data();
 
     if (etdump_data.buf != nullptr && etdump_data.size > 0) {
-      int etdump_file =
-          open("/data/local/tmp/result.etdump", O_WRONLY | O_CREAT, 0644);
+      int etdump_file = open(path, O_WRONLY | O_CREAT | O_TRUNC, 0644);
       if (etdump_file == -1) {
-        ET_LOG(Error, "Cannot create result.etdump error: %d", errno);
+        ET_LOG(Error, "Cannot create %s error: %d", path, errno);
+        free(etdump_data.buf);
         return false;
       }
       ssize_t bytes_written =
           write(etdump_file, (uint8_t*)etdump_data.buf, etdump_data.size);
       if (bytes_written == -1) {
-        ET_LOG(Error, "Cannot write result.etdump error: %d", errno);
+        ET_LOG(Error, "Cannot write %s error: %d", path, errno);
+        close(etdump_file);
+        free(etdump_data.buf);
         return false;
       } else {
-        ET_LOG(Info, "ETDump written %d bytes to file.", bytes_written);
+        ET_LOG(Info, "ETDump written %zd bytes to %s.", bytes_written, path);
       }
       close(etdump_file);
       free(etdump_data.buf);
       return true;
     } else {
       ET_LOG(Error, "No ETDump data available!");
     }
+#else
+    (void)path;
 #endif
     return false;
   }
 
+ public:
+
   facebook::jni::local_ref<facebook::jni::JArrayClass<jstring>> getMethods() {
     const auto& names_result = module_->method_names();
     if (!names_result.ok()) {
@@ -571,6 +586,7 @@ class ExecuTorchJni : public facebook::jni::HybridClass<ExecuTorchJni> {
         makeNativeMethod(
             "readLogBufferStaticNative", ExecuTorchJni::readLogBufferStatic),
         makeNativeMethod("etdumpNative", ExecuTorchJni::etdump),
+        makeNativeMethod("etdumpToNative", ExecuTorchJni::etdumpTo),
         makeNativeMethod("getMethodsNative", ExecuTorchJni::getMethods),
         makeNativeMethod("getUsedBackends", ExecuTorchJni::getUsedBackends),
     });

scripts/build_android_library.sh

@@ -40,6 +40,7 @@ build_android_native_library() {
     --preset "android-${ANDROID_ABI}" \
     -DANDROID_PLATFORM=android-26 \
     -DEXECUTORCH_ENABLE_EVENT_TRACER="${EXECUTORCH_ANDROID_PROFILING:-OFF}" \
+    -DEXECUTORCH_ANDROID_PROFILING="${EXECUTORCH_ANDROID_PROFILING:-OFF}" \
     -DEXECUTORCH_BUILD_EXTENSION_LLM="${EXECUTORCH_BUILD_EXTENSION_LLM:-ON}" \
     -DEXECUTORCH_BUILD_EXTENSION_LLM_RUNNER="${EXECUTORCH_BUILD_EXTENSION_LLM:-ON}" \
     -DEXECUTORCH_BUILD_EXTENSION_ASR_RUNNER="${EXECUTORCH_BUILD_EXTENSION_LLM:-ON}" \
@@ -50,7 +51,9 @@ build_android_native_library() {
     -DEXECUTORCH_BUILD_QNN="${EXECUTORCH_BUILD_QNN}" \
     -DQNN_SDK_ROOT="${QNN_SDK_ROOT}" \
     -DEXECUTORCH_BUILD_VULKAN="${EXECUTORCH_BUILD_VULKAN}" \
+    -DEXECUTORCH_VULKAN_FP16_PRECISION="${EXECUTORCH_VULKAN_FP16_PRECISION:-highp}" \
     -DXNNPACK_ENABLE_ARM_SME2="${XNNPACK_ENABLE_ARM_SME2}" \
+    -DFLATCC_ALLOW_WERROR=OFF \
     -DSUPPORT_REGEX_LOOKAHEAD=ON \
     -DCMAKE_BUILD_TYPE="${EXECUTORCH_CMAKE_BUILD_TYPE}" \
     -B"${CMAKE_OUT}"

third-party/CMakeLists.txt

@@ -95,11 +95,12 @@ ExternalProject_Add(
   PREFIX ${CMAKE_CURRENT_BINARY_DIR}/flatcc_ep
   SOURCE_DIR ${PROJECT_SOURCE_DIR}/third-party/flatcc
   BINARY_DIR ${CMAKE_CURRENT_BINARY_DIR}/flatcc_ep/src/build
-  CMAKE_ARGS
+  CMAKE_ARGS 
     -DFLATCC_RTONLY=OFF
     -DFLATCC_TEST=OFF
     -DFLATCC_REFLECTION=OFF
     -DFLATCC_DEBUG_CLANG_SANITIZE=OFF
+    -DFLATCC_ALLOW_WERROR=OFF
     -DFLATCC_INSTALL=ON
     -DCMAKE_POLICY_VERSION_MINIMUM=3.5
     -DCMAKE_INSTALL_PREFIX:PATH=<INSTALL_DIR>

tools/cmake/common/preset.cmake

@@ -82,12 +82,12 @@ macro(define_overridable_option NAME DESCRIPTION VALUE_TYPE DEFAULT_VALUE)
   if(DEFINED ${NAME} AND NOT DEFINED CACHE{${NAME}})
     set(${NAME}
         ${${NAME}}
-        CACHE ${VALUE_TYPE} ${DESCRIPTION} FORCE
+        CACHE ${VALUE_TYPE} "${DESCRIPTION}" FORCE
     )
   else()
     set(${NAME}
         ${DEFAULT_VALUE}
-        CACHE ${VALUE_TYPE} ${DESCRIPTION}
+        CACHE ${VALUE_TYPE} "${DESCRIPTION}"
     )
   endif()
 

tools/cmake/preset/default.cmake

@@ -168,6 +168,12 @@ define_overridable_option(
 define_overridable_option(
   EXECUTORCH_BUILD_VULKAN "Build the Vulkan backend" BOOL OFF
 )
+define_overridable_option(
+  EXECUTORCH_VULKAN_FP16_PRECISION
+  "GLSL PRECISION for Vulkan half-precision shader variants. Accepted values: highp, mediump, lowp. `mediump` lets Mali drivers use fp16 ALUs; ignored on Adreno. Default `highp` matches upstream."
+  STRING
+  highp
+)
 define_overridable_option(
   EXECUTORCH_BUILD_PORTABLE_OPS "Build portable_ops library" BOOL ON
 )