Add QNN context binary test

Author: ciaranbor

models/qnn/mobilenet_v3_small.bin

@@ -49,6 +49,7 @@ endif()
 if(edgerunner_ENABLE_NPU)
     list(APPEND TEST_SOURCES source/tflite_npu_test.cpp
          source/qnn_shared_library_npu_test.cpp
+         source/qnn_context_binary_npu_test.cpp
     )
 endif()
 

test/CMakeLists.txt

@@ -0,0 +1,79 @@
+#include <algorithm>
+#include <cstddef>
+#include <iterator>
+#include <string>
+#include <vector>
+
+#include <catch2/benchmark/catch_benchmark.hpp>
+#include <catch2/catch_message.hpp>
+#include <catch2/catch_test_macros.hpp>
+
+#include "edgerunner/edgerunner.hpp"
+#include "edgerunner/model.hpp"
+#include "edgerunner/tensor.hpp"
+#include "utils.hpp"
+
+TEST_CASE("QNN context binary NPU runtime", "[qnn][context][npu]") {
+    const std::string modelPath = "models/qnn/mobilenet_v3_small.bin";
+
+    auto model = edge::createModel(modelPath);
+    REQUIRE(model != nullptr);
+    REQUIRE(std::string {"mobilenet_v3_small"} == model->name());
+    REQUIRE(model->getDelegate() == edge::DELEGATE::CPU);
+
+    /* ensure CPU and NPU inference have the same inputs */
+    auto cpuInputData = model->getInput(0)->getTensorAs<float>();
+    std::fill(cpuInputData.begin(), cpuInputData.end(), 0);
+
+    auto executionStatus = model->execute();
+    CHECK(executionStatus == edge::STATUS::SUCCESS);
+
+    const auto cpuOutput = model->getOutput(0)->getTensorAs<float>();
+
+    /* applying a new delegate releases memory, so need to copy CPU output to
+     * compare later */
+    std::vector<float> cpuResult;
+    cpuResult.reserve(cpuOutput.size());
+    std::copy(
+        cpuOutput.cbegin(), cpuOutput.cend(), std::back_inserter(cpuResult));
+
+    const auto delegateStatus = model->applyDelegate(edge::DELEGATE::NPU);
+    REQUIRE(delegateStatus == edge::STATUS::SUCCESS);
+    REQUIRE(model->getDelegate() == edge::DELEGATE::NPU);
+
+    const auto numInputs = model->getNumInputs();
+    REQUIRE(numInputs == 1);
+
+    const auto numOutputs = model->getNumOutputs();
+    REQUIRE(numOutputs == 1);
+
+    auto input = model->getInput(0);
+    REQUIRE(input->getName() == "image_tensor");
+    REQUIRE(input->getDimensions() == std::vector<size_t> {1, 224, 224, 3});
+    REQUIRE(input->getType() == edge::TensorType::FLOAT32);
+
+    auto inputData = input->getTensorAs<float>();
+    REQUIRE(inputData.size() == input->getSize());
+
+    /* ensure CPU and NPU inference have the same inputs */
+    std::fill(inputData.begin(), inputData.end(), 0);
+
+    executionStatus = model->execute();
+    REQUIRE(executionStatus == edge::STATUS::SUCCESS);
+
+    BENCHMARK("execution") {
+        return model->execute();
+    };
+
+    auto output = model->getOutput(0);
+    REQUIRE(output->getName() == "class_logits");
+    REQUIRE(output->getDimensions() == std::vector<size_t> {1, 1000});
+    REQUIRE(output->getType() == edge::TensorType::FLOAT32);
+
+    auto outputData = output->getTensorAs<float>();
+    REQUIRE(outputData.size() == output->getSize());
+
+    const auto mse = meanSquaredError(cpuResult, outputData);
+    CAPTURE(mse);
+    REQUIRE(mse < MseThreshold);
+}

test/source/qnn_context_binary_npu_test.cpp

@@ -13,7 +13,7 @@
 #include "edgerunner/tensor.hpp"
 #include "utils.hpp"
 
-TEST_CASE("QNN NPU runtime", "[qnn][npu]") {
+TEST_CASE("QNN shared library NPU runtime", "[qnn][shared][npu]") {
     const std::string modelPath = "models/qnn/mobilenet_v3_small.so";
 
     auto model = edge::createModel(modelPath);
@@ -66,7 +66,7 @@ TEST_CASE("QNN NPU runtime", "[qnn][npu]") {
     };
 
     auto output = model->getOutput(0);
-    REQUIRE(output->getName() == "output_0");
+    REQUIRE(output->getName() == "class_logits");
     REQUIRE(output->getDimensions() == std::vector<size_t> {1, 1000});
     REQUIRE(output->getType() == edge::TensorType::FLOAT32);