Share the Native vision runner across CPU and Vulkan engines (#23936)

backends/native/README.md

@@ -252,29 +252,44 @@ prefix (or the corresponding `CMAKE_INSTALL_DATADIR` location).
 CPU engine tests, including quantized linear execution, graph caching, and
 mutable buffers.
 
-### Vision classification runner
+### Vision classification runners
 
-Enable `EXECUTORCH_BUILD_NATIVE_VISION_RUNNER=ON` together with
-`EXECUTORCH_BUILD_NATIVE_VULKAN=ON` to build `run_classification` from
-`examples/vision`. The runner additionally requires gflags; an OSS build uses
-the vendored copy, and other builds resolve an installed gflags package.
-Running the runner requires a Vulkan device.
+`examples/vision` shares model loading, FP32 input/output handling, top-5 display,
+and eager-reference comparison between the CPU and Vulkan engines.
+`run_classification_cpu` uses the CPU engine and `run_classification` the Vulkan
+engine; `run_classification_vulkan` is a Buck alias for `run_classification`.
+
+For a CPU-only CMake build:
 
 ```sh
 cmake -S . -B cmake-out -DCMAKE_BUILD_TYPE=Release \
-  -DEXECUTORCH_BUILD_NATIVE=ON -DEXECUTORCH_BUILD_NATIVE_VULKAN=ON \
+  -DEXECUTORCH_BUILD_NATIVE=ON -DEXECUTORCH_BUILD_NATIVE_CPU=ON \
+  -DEXECUTORCH_BUILD_XNNPACK=ON \
+  -DEXECUTORCH_BUILD_KERNELS_OPTIMIZED=ON \
   -DEXECUTORCH_BUILD_NATIVE_VISION_RUNNER=ON
-cmake --build cmake-out --target run_classification
+cmake --build cmake-out --target run_classification_cpu
 cmake --install cmake-out --prefix install
-install/bin/run_classification \
+install/bin/run_classification_cpu \
   --package mv2.ptn --input mv2_input.bin \
   --expected mv2_expected.bin --categories categories.txt
 ```
 
+`EXECUTORCH_BUILD_NATIVE_CPU` provides `native_cpu_engine`, using
+XNNPACK and the Reference provider with optimized/portable operator registrations.
+The CPU build does not need a Vulkan SDK or device. Enable
+`EXECUTORCH_BUILD_NATIVE_VULKAN=ON` to also build `run_classification`, which
+runs on the host's Vulkan device; both engines can be enabled together. The
+runners require gflags.
+
 Export the `.ptn`, input, reference logits, and labels with
-[`export.py`](../../examples/native/vision/export.py). The runner prints the
-top-5 logits and, when `--expected` is given, exits nonzero if the logits
-differ from the reference beyond `--rtol`/`--atol` (both default to `1e-3`).
+[`export.py`](../../examples/native/vision/export.py). The same package and FP32
+artifacts are accepted by either runner, subject to each engine's supported
+operators. Each runner expects a model with one FP32 input and one FP32 output,
+and prints the selected engine and device and the top-5 logits. When
+`--expected` is given, it compares the logits with the reference using
+`--rtol`/`--atol` (both default to `1e-3`). Exit codes are 0 on success, 1 for
+invalid arguments, 2 for a loading or execution error, and 3 for a reference
+mismatch.
 
 ## Native LLM runner
 

backends/native/examples/vision/BUCK

@@ -3,20 +3,47 @@ load("@fbsource//xplat/executorch/build:runtime_wrapper.bzl", "runtime")
 
 oncall("executorch")
 
-# Run an exported .ptn on the Vulkan engine and check the logits against the
-# eager reference. Runs on the host's system Vulkan device.
-fbcode_target(
-    _kind = runtime.cxx_binary,
-    name = "run_classification",
-    srcs = ["run_classification.cpp"],
+runtime.cxx_library(
+    name = "classification_runner",
+    srcs = ["ClassificationRunner.cpp"],
+    exported_headers = ["ClassificationRunner.h"],
+    exported_deps = ["//executorch/backends/native/runtime/engine:engine"],
     deps = [
         "//executorch/backends/native/examples/runner_utils:runner_utils",
         "//executorch/backends/native/runtime:runtime",
         "//executorch/backends/native/runtime/deserialize:package",
-        "//executorch/backends/native/runtime/engine:engine",
         "//executorch/backends/native/runtime/graph:scalar_type",
+    ],
+    external_deps = ["gflags"],
+    visibility = ["PUBLIC"],
+)
+
+runtime.cxx_binary(
+    name = "run_classification_cpu",
+    srcs = ["run_classification_cpu.cpp"],
+    deps = [
+        ":classification_runner",
+        "//executorch/backends/native/runtime/cpu:cpu_engine",
+    ],
+    external_deps = ["gflags"],
+    visibility = ["PUBLIC"],
+)
+
+fbcode_target(
+    _kind = runtime.cxx_binary,
+    name = "run_classification",
+    srcs = ["run_classification.cpp"],
+    deps = [
+        ":classification_runner",
         "//executorch/backends/native/runtime/vulkan:vulkan_engine",
     ],
     external_deps = ["gflags"],
     visibility = ["PUBLIC"],
 )
+
+fbcode_target(
+    _kind = runtime.command_alias,
+    name = "run_classification_vulkan",
+    exe = ":run_classification",
+    visibility = ["PUBLIC"],
+)

backends/native/examples/vision/CMakeLists.txt

@@ -13,6 +13,8 @@ if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR)
   set(EXECUTORCH_ROOT "${executorch_native_PREFIX}/include/executorch")
   add_subdirectory(../runner_utils runner_utils)
   set(_native_vision_runtime executorch_native::runtime)
+  set(_native_vision_vulkan_engine executorch_native::runtime)
+  set(EXECUTORCH_BUILD_NATIVE_VULKAN ON)
 
   if(NOT TARGET gflags AND NOT TARGET gflags::gflags)
     find_package(gflags REQUIRED)
@@ -23,19 +25,43 @@ if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR)
     set(_native_runner_gflags gflags)
   endif()
 else()
-  set(_native_vision_runtime native_engine native_package native_runtime
-                             native_vulkan_engine
-  )
+  set(_native_vision_runtime native_engine)
+  set(_native_vision_vulkan_engine native_vulkan_engine)
 endif()
 
-add_executable(run_classification run_classification.cpp)
+add_library(native_classification_runner STATIC ClassificationRunner.cpp)
 target_link_libraries(
-  run_classification PRIVATE ${_native_runner_gflags} native_runner_utils
-                             ${_native_vision_runtime}
+  native_classification_runner
+  PUBLIC ${_native_vision_runtime}
+  PRIVATE ${_native_runner_gflags} native_runner_utils
 )
 target_compile_options(
-  run_classification PRIVATE $<$<CXX_COMPILER_ID:MSVC>:/EHsc>
-                             $<$<NOT:$<CXX_COMPILER_ID:MSVC>>:-fexceptions>
+  native_classification_runner
+  PRIVATE $<$<CXX_COMPILER_ID:MSVC>:/EHsc>
+          $<$<NOT:$<CXX_COMPILER_ID:MSVC>>:-fexceptions>
 )
 
-install(TARGETS run_classification RUNTIME DESTINATION ${CMAKE_INSTALL_BINDIR})
+function(add_native_classification_runner target source engine)
+  add_executable(${target} ${source})
+  target_link_libraries(
+    ${target} PRIVATE native_classification_runner ${_native_runner_gflags}
+                      ${engine}
+  )
+  target_compile_options(
+    ${target} PRIVATE $<$<CXX_COMPILER_ID:MSVC>:/EHsc>
+                      $<$<NOT:$<CXX_COMPILER_ID:MSVC>>:-fexceptions>
+  )
+  install(TARGETS ${target} RUNTIME DESTINATION ${CMAKE_INSTALL_BINDIR})
+endfunction()
+
+if(EXECUTORCH_BUILD_NATIVE_VULKAN)
+  add_native_classification_runner(
+    run_classification run_classification.cpp ${_native_vision_vulkan_engine}
+  )
+endif()
+
+if(EXECUTORCH_BUILD_NATIVE_CPU)
+  add_native_classification_runner(
+    run_classification_cpu run_classification_cpu.cpp native_cpu_engine
+  )
+endif()

backends/native/examples/vision/ClassificationRunner.cpp

@@ -0,0 +1,178 @@
+// Copyright (c) Meta Platforms, Inc. and affiliates.
+// All rights reserved.
+//
+// This source code is licensed under the BSD-style license found in the
+// LICENSE file in the root directory of this source tree.
+
+#include <executorch/backends/native/examples/vision/ClassificationRunner.h>
+
+#include <cstdint>
+#include <cstdio>
+#include <exception>
+#include <memory>
+#include <string>
+#include <vector>
+
+#include <executorch/backends/native/examples/runner_utils/runner_utils.h>
+#include <executorch/backends/native/runtime/Program.h>
+#include <executorch/backends/native/runtime/deserialize/Package.h>
+#include <executorch/backends/native/runtime/engine/Engine.h>
+#include <executorch/backends/native/runtime/graph/ScalarType.h>
+#include <gflags/gflags.h>
+
+DEFINE_string(package, "", "Path to the *.ptn to run. Required.");
+DEFINE_string(input, "", "Path to the input tensor, raw float32. Required.");
+DEFINE_string(
+    expected,
+    "",
+    "Reference logits to compare against, raw float32.");
+DEFINE_string(categories, "", "Newline-separated class labels, for display.");
+DEFINE_string(method, "", "Method to run. Defaults to the first one.");
+DEFINE_double(rtol, 1e-3, "Relative tolerance for the logits comparison.");
+DEFINE_double(atol, 1e-3, "Absolute tolerance for the logits comparison.");
+namespace {
+
+std::string label_of(const std::vector<std::string>& labels, int index) {
+  if (index >= 0 && index < static_cast<int>(labels.size())) {
+    return labels[index];
+  }
+  return "class_" + std::to_string(index);
+}
+
+void print_top5(
+    const std::vector<float>& logits,
+    const std::vector<std::string>& labels) {
+  const std::vector<int> order = ptn::runner_utils::top_k(logits, 5);
+  std::printf("top-5 (class : logit):\n");
+  for (const int index : order) {
+    std::printf(
+        "  %4d  %-28s %10.4f\n",
+        index,
+        label_of(labels, index).c_str(),
+        logits[index]);
+  }
+}
+
+} // namespace
+
+namespace ptn {
+
+int run_classification(
+    const std::function<std::unique_ptr<EngineHost>()>& make_host,
+    const std::function<ConstantBytes(const EngineContext&)>& constant_bytes) {
+  if (FLAGS_package.empty() || FLAGS_input.empty()) {
+    std::fprintf(stderr, "error: --package and --input are required\n");
+    return 1;
+  }
+
+  try {
+    const auto package =
+        std::make_shared<const ptn::Package>(ptn::Package::load(FLAGS_package));
+    const ptn::ByteSpan program_bytes = package->program_bytes();
+    const auto program = std::make_shared<const ptn::Program>(
+        ptn::Program::load(program_bytes.data(), program_bytes.size()));
+
+    const std::vector<std::string> names = program->method_names();
+    if (names.empty()) {
+      std::fprintf(stderr, "error: package holds no methods\n");
+      return 2;
+    }
+    const std::string method_name =
+        FLAGS_method.empty() ? names[0] : FLAGS_method;
+    std::printf(
+        "package:   %s (program %zu bytes, %zu constants / %zu bytes)\n",
+        FLAGS_package.c_str(),
+        program_bytes.size(),
+        package->keys().size(),
+        package->constant_bytes());
+
+    const auto host = make_host();
+    const auto context = host->create_context(program, package);
+    std::printf(
+        "engine:    %s on %s\n",
+        host->name().c_str(),
+        host->device_name().c_str());
+
+    const std::unique_ptr<ptn::EngineExecutable> executable =
+        context->compile(method_name);
+    std::printf(
+        "method:    %s (%zu inputs, %zu outputs)\n",
+        method_name.c_str(),
+        executable->num_inputs(),
+        executable->num_outputs());
+    const ConstantBytes constants = constant_bytes(*context);
+    std::printf(
+        "constants: %zu unique source bytes, %zu materialized\n",
+        constants.unique,
+        constants.materialized);
+
+    if (executable->num_inputs() != 1 || executable->num_outputs() != 1) {
+      std::fprintf(
+          stderr,
+          "error: expected 1 input and 1 output, got %zu and %zu\n",
+          executable->num_inputs(),
+          executable->num_outputs());
+      return 2;
+    }
+    if (executable->input_dtype(0) != ptn::kFloat ||
+        executable->output_dtype(0) != ptn::kFloat) {
+      std::fprintf(stderr, "error: this runner handles fp32 models only\n");
+      return 2;
+    }
+
+    const std::vector<int64_t> in_sizes = executable->input_sizes(0);
+    const std::vector<int64_t> out_sizes = executable->output_sizes(0);
+    std::printf(
+        "  input:   %s\n  output:  %s\n",
+        ptn::runner_utils::shape_str(in_sizes).c_str(),
+        ptn::runner_utils::shape_str(out_sizes).c_str());
+
+    const std::vector<float> input =
+        ptn::runner_utils::read_floats(FLAGS_input);
+    if (static_cast<int64_t>(input.size()) !=
+        ptn::runner_utils::numel_of(in_sizes)) {
+      std::fprintf(
+          stderr,
+          "error: input has %zu floats, the model wants %lld\n",
+          input.size(),
+          static_cast<long long>(ptn::runner_utils::numel_of(in_sizes)));
+      return 2;
+    }
+
+    executable->set_input(0, input.data(), input.size(), ptn::kFloat);
+    executable->execute();
+
+    std::vector<float> logits(
+        static_cast<size_t>(ptn::runner_utils::numel_of(out_sizes)));
+    executable->get_output(0, logits.data(), logits.size(), ptn::kFloat);
+    if (logits.empty()) {
+      std::fprintf(stderr, "error: model produced no logits\n");
+      return 2;
+    }
+
+    std::vector<std::string> labels;
+    if (!FLAGS_categories.empty()) {
+      labels = ptn::runner_utils::read_lines(FLAGS_categories);
+    }
+    print_top5(logits, labels);
+    const int argmax = ptn::runner_utils::top_k(logits, 1).front();
+    std::printf(
+        "argmax:    %d (%s)\n", argmax, label_of(labels, argmax).c_str());
+
+    if (FLAGS_expected.empty()) {
+      std::printf("result:    ran (no reference given)\n");
+      return 0;
+    }
+    const std::vector<float> reference =
+        ptn::runner_utils::read_floats(FLAGS_expected);
+    const bool pass = ptn::runner_utils::compare_floats(
+        logits, reference, FLAGS_rtol, FLAGS_atol);
+    std::printf("result:    %s\n", pass ? "PASS" : "FAIL");
+    return pass ? 0 : 3;
+  } catch (const std::exception& e) {
+    std::fprintf(stderr, "error: %s\n", e.what());
+    return 2;
+  }
+}
+
+} // namespace ptn

backends/native/examples/vision/ClassificationRunner.h

@@ -0,0 +1,32 @@
+// Copyright (c) Meta Platforms, Inc. and affiliates.
+// All rights reserved.
+//
+// This source code is licensed under the BSD-style license found in the
+// LICENSE file in the root directory of this source tree.
+
+#pragma once
+
+#include <cstddef>
+#include <functional>
+#include <memory>
+
+#include <executorch/backends/native/runtime/engine/Engine.h>
+
+namespace ptn {
+
+struct ConstantBytes {
+  size_t unique = 0;
+  size_t materialized = 0;
+};
+
+template <typename Context>
+ConstantBytes constant_bytes(const EngineContext& context) {
+  const auto& typed = dynamic_cast<const Context&>(context);
+  return {typed.unique_constant_bytes(), typed.materialized_constant_bytes()};
+}
+
+int run_classification(
+    const std::function<std::unique_ptr<EngineHost>()>& make_host,
+    const std::function<ConstantBytes(const EngineContext&)>& constant_bytes);
+
+} // namespace ptn