Enable unit test for wasi-nn WinML backend.

This test was disabled because GitHub Actions Windows Server image
doesn't have desktop experience included. But it looks like we can have
a standalone WinML binary downloaded from ONNX Runtime project.

Wasi-nn WinML backend and ONNX Runtime backend now share the same test
code as they accept the same input, and they are expected to produce the
same result.

This change also make wasi-nn WinML backend as a default feature.

prtest:full

.github/workflows/main.yml

@@ -480,6 +480,15 @@ jobs:
     - uses: abrown/install-openvino-action@v8
       if: runner.arch == 'X64'
 
+    # Install WinML for testing wasi-nn WinML backend. WinML is only available
+    # on Windows clients and Windows Server with desktop experience enabled.
+    # GitHub Actions Window Server image doesn't have desktop experience
+    # enabled, so we download the standalone library from ONNX Runtime project.
+    - uses: nuget/setup-nuget@v2
+      if: matrix.os == 'windows-latest'
+    - run: nuget install Microsoft.AI.MachineLearning
+      if: matrix.os == 'windows-latest'
+
     # Fix an ICE for now in gcc when compiling zstd with debuginfo (??)
     - run: echo CFLAGS=-g0 >> $GITHUB_ENV
       if: matrix.target == 'x86_64-pc-windows-gnu'

crates/test-programs/Cargo.toml

@@ -20,3 +20,6 @@ futures = { workspace = true, default-features = false, features = ['alloc'] }
 url = { workspace = true }
 sha2 = "0.10.2"
 base64 = "0.21.0"
+# image and ndarray are used by nn_image_classification_onnx for image preprocessing.
+image = { version = "0.24.6", default-features = false, features = ["jpeg"] }
+ndarray = "0.15.3"

crates/test-programs/src/bin/nn_image_classification_onnx.rs

@@ -1,6 +1,8 @@
 use anyhow::Result;
 use std::fs;
 use wasi_nn::*;
+use ndarray::Array;
+use image::{DynamicImage, RgbImage};
 
 pub fn main() -> Result<()> {
     let model = fs::read("fixture/model.onnx").unwrap();
@@ -17,8 +19,16 @@ pub fn main() -> Result<()> {
 
     // Prepare WASI-NN tensor - Tensor data is always a bytes vector
     // Load a tensor that precisely matches the graph input tensor
-    let data = fs::read("fixture/tensor.bgr").unwrap();
+    let data = fs::read("fixture/dog.jpg").unwrap();
     println!("[ONNX] Read input tensor, size in bytes: {}", data.len());
+    let data = preprocess(
+        data.as_slice(),
+        224,
+        224,
+        &[0.485, 0.456, 0.406],
+        &[0.229, 0.224, 0.225],
+    );
+
     context.set_input(0, wasi_nn::TensorType::F32, &[1, 3, 224, 224], &data)?;
 
     // Execute the inferencing
@@ -28,11 +38,26 @@ pub fn main() -> Result<()> {
     // Retrieve the output.
     let mut output_buffer = vec![0f32; 1000];
     context.get_output(0, &mut output_buffer[..])?;
+
+    // Post-processing.
+    let output_shape = [1, 1000, 1, 1];
+    let output_tensor = Array::from_shape_vec(output_shape, output_buffer).unwrap();
+
+    let exp_output = output_tensor.mapv(|x| x.exp());
+    let sum_exp_output = exp_output.sum_axis(ndarray::Axis(1));
+    let softmax_output = exp_output / &sum_exp_output;
+
+    let sorted = sort_results(&softmax_output.into_raw_vec());
+
     println!(
         "[ONNX] Found results, sorted top 5: {:?}",
-        &sort_results(&output_buffer)[..5]
+        &sorted[..5]
     );
 
+    // Index 207 is curly-coated retriever.
+    // https://github.com/onnx/models/blob/bec48b6a70e5e9042c0badbaafefe4454e072d08/validated/vision/classification/synset.txt#L207
+    assert_eq!(sorted[0].0, 207);
+
     Ok(())
 }
 
@@ -50,6 +75,43 @@ fn sort_results(buffer: &[f32]) -> Vec<InferenceResult> {
     results
 }
 
+// Take the image located at 'path', open it, resize it to height x width, and then converts
+// the pixel precision to FP32. The resulting RGB pixel vector is then returned.
+fn preprocess(
+    image: &[u8],
+    height: u32,
+    width: u32,
+    mean: &[f32],
+    std: &[f32],
+) -> Vec<u8> {
+    let dyn_img: DynamicImage = image::load_from_memory(image).unwrap().resize_exact(width, height, image::imageops::Triangle);
+    let rgb_img: RgbImage = dyn_img.to_rgb8();
+
+    // Get an array of the pixel values
+    let raw_u8_arr: &[u8] = &rgb_img.as_raw()[..];
+
+    // Create an array to hold the f32 value of those pixels
+    let bytes_required = raw_u8_arr.len() * 4;
+    let mut u8_f32_arr: Vec<u8> = vec![0; bytes_required];
+
+    // Read the number as a f32 and break it into u8 bytes
+    for i in 0..raw_u8_arr.len() {
+        let u8_f32: f32 = raw_u8_arr[i] as f32;
+        let rgb_iter = i % 3;
+
+        // Normalize the pixel
+        let norm_u8_f32: f32 = (u8_f32 / 255.0 - mean[rgb_iter]) / std[rgb_iter];
+
+        // Convert it to u8 bytes and write it with new shape
+        let u8_bytes = norm_u8_f32.to_ne_bytes();
+        for j in 0..4 {
+            u8_f32_arr[(raw_u8_arr.len() * 4 * rgb_iter / 3) + (i / 3) * 4 + j] = u8_bytes[j];
+        }
+    }
+
+    return u8_f32_arr;
+}
+
 // A wrapper for class ID and match probabilities.
 #[derive(Debug, PartialEq)]
 struct InferenceResult(usize, f32);

crates/wasi-nn/Cargo.toml

@@ -50,7 +50,7 @@ wasi-common = { workspace = true, features = ["sync"] }
 wasmtime = { workspace = true, features = ["cranelift"] }
 
 [features]
-default = ["openvino"]
+default = ["openvino", "winml"]
 # openvino is available on all platforms, it requires openvino installed.
 openvino = ["dep:openvino"]
 # onnx is available on all platforms.