benchmark 腳本

Author: 5000user5000

benchmarks/run_benchmark.py

@@ -0,0 +1,202 @@
+import timeit
+import glob
+import os
+import argparse
+import math
+import numpy as np
+
+# 嘗試 import，如果失敗則提示
+try:
+    import fast_jpeg_decoder as fjd
+except ImportError:
+    print("❌ Error: Could not import 'fast_jpeg_decoder'. Make sure the C++ module is compiled and in path.")
+    exit(1)
+
+try:
+    from python_implementations import numpy_decoder
+except ImportError:
+    print("❌ Error: Could not import 'numpy_decoder' from 'python_implementations'.")
+    exit(1)
+
+# --- Configuration ---
+TEST_DATA_DIR = 'tests/test_data'
+OUTPUT_DIR = 'output'
+NUMBER_OF_RUNS = 10 
+
+def calculate_psnr(img1, img2):
+    """
+    計算兩張圖片的 PSNR (峰值訊噪比)
+    img1: 測試圖片
+    img2: 參考圖片 (Ground Truth, 通常是 PIL)
+    """
+    if img1.shape != img2.shape:
+        return 0.0
+    
+    # 計算 MSE (均方誤差)
+    mse = np.mean((img1.astype(float) - img2.astype(float)) ** 2)
+    if mse == 0:
+        return float('inf') # 完全相同
+    
+    max_pixel = 255.0
+    psnr = 20 * math.log10(max_pixel / math.sqrt(mse))
+    return psnr
+
+def run_cpp_decoder(image_bytes):
+    """Wrapper for the C++ decoder."""
+    return fjd.load_bytes(image_bytes)
+
+def run_numpy_decoder(image_bytes):
+    """Wrapper for the NumPy decoder."""
+    return numpy_decoder.decode(image_bytes)
+
+
+def benchmark_single_image(image_path):
+    """Benchmark a single image (Includes Verify, Output, and PSNR comparison)."""
+    filename = os.path.basename(image_path)
+    print(f"\n{'='*60}")
+    print(f"Processing: {filename}")
+    print(f"{'='*60}")
+
+    try:
+        with open(image_path, 'rb') as f:
+            image_bytes = f.read()
+    except FileNotFoundError:
+        print(f"❌ Error: Test image not found at '{image_path}'")
+        return
+
+    file_size = len(image_bytes) / 1024
+    print(f"File size: {file_size:.1f} KB")
+
+    # 1. 確保輸出目錄存在
+    os.makedirs(OUTPUT_DIR, exist_ok=True)
+
+    # 2. 準備 "原始圖像" (Ground Truth / Reference)
+    # 我們使用 PIL 解碼的結果作為 "標準答案"
+    img_gt = None
+    try:
+        from PIL import Image
+        import io
+        img_gt = np.array(Image.open(io.BytesIO(image_bytes)).convert('RGB'))
+    except ImportError:
+        print("⚠️  PIL not installed, cannot calculate PSNR or save images.")
+        return
+
+    # print(f"\n{'─'*60}")
+    # print("Decoding & Saving:")
+    # print(f"{'─'*60}")
+
+    # --- 收集所有解碼器的結果 ---
+    decoders_result = {}
+    
+    # 1. C++ Decoder
+    avg_cpp_time = float('inf')
+    try:
+        img_cpp = run_cpp_decoder(image_bytes)
+        if img_cpp.size > 0:
+            decoders_result['C++  '] = img_cpp
+            # 存檔
+            out_name = os.path.join(OUTPUT_DIR, f"cpp_{filename}.png")
+            Image.fromarray(img_cpp).save(out_name)
+            # 測速
+            cpp_time = timeit.timeit(lambda: run_cpp_decoder(image_bytes), number=NUMBER_OF_RUNS)
+            avg_cpp_time = (cpp_time / NUMBER_OF_RUNS) * 1000
+        else:
+            print("❌ C++ Decoder returned empty image")
+    except Exception as e:
+        print(f"❌ C++ Decoder Error: {e}")
+
+    # 2. NumPy Decoder
+    avg_numpy_time = float('inf')
+    try:
+        img_numpy = run_numpy_decoder(image_bytes)
+        # Handle list return type if necessary
+        if not isinstance(img_numpy, np.ndarray): 
+            img_numpy = np.array(img_numpy, dtype=np.uint8)
+        
+        if img_numpy.size > 0:
+            decoders_result['NumPy'] = img_numpy
+            # 存檔
+            out_name = os.path.join(OUTPUT_DIR, f"numpy_{filename}.png")
+            Image.fromarray(img_numpy).save(out_name)
+            # 測速
+            numpy_time = timeit.timeit(lambda: run_numpy_decoder(image_bytes), number=NUMBER_OF_RUNS)
+            avg_numpy_time = (numpy_time / NUMBER_OF_RUNS) * 1000
+        else:
+            print("❌ NumPy Decoder returned empty image")
+    except Exception as e:
+        print(f"❌ NumPy Decoder Error: {e}")
+
+    # 3. PIL Decoder (本身也加入比較列表，確認基準)
+    decoders_result['PIL  '] = img_gt
+    # PIL 也可以存一份 png 當作對照組
+    Image.fromarray(img_gt).save(os.path.join(OUTPUT_DIR, f"pil_{filename}.png"))
+
+
+    # --- 統一計算 PSNR (全部 vs 原始圖像) ---
+    print(f"\n{'─'*60}")
+    print("Quality Metrics (vs Original/PIL):")
+    print(f"{'─'*60}")
+
+    for name, img_test in decoders_result.items():
+        try:
+            # 形狀檢查與修正 (針對 flatten array)
+            if img_test.shape != img_gt.shape:
+                if img_test.size == img_gt.size:
+                     img_test = img_test.reshape(img_gt.shape)
+                else:
+                    print(f"{name}: Shape mismatch {img_test.shape} vs {img_gt.shape}")
+                    continue
+
+            # 計算 PSNR
+            psnr = calculate_psnr(img_test, img_gt)
+            
+            # 計算平均像素差異
+            mean_diff = np.abs(img_test.astype(float) - img_gt.astype(float)).mean()
+
+            # 格式化輸出
+            status = "✅" if psnr > 30 or psnr == float('inf') else "⚠️ "
+            psnr_str = "Infinity" if psnr == float('inf') else f"{psnr:6.2f} dB"
+            
+            print(f"{name} Decoder:")
+            print(f"PSNR: {status} {psnr_str}")
+            #print(f"   Mean Diff: {mean_diff:.2f}")
+            print("-" * 20)
+            
+        except Exception as e:
+            print(f"{name} vs Reference: Error ({e})")
+
+    # --- 效能總結 ---
+    print(f"\n{'─'*60}")
+    print("Speed Benchmark Results:")
+    print(f"{'─'*60}")
+    
+    if avg_cpp_time != float('inf'):
+        print(f"🚀 C++ Decoder:   {avg_cpp_time:.2f} ms")
+    if avg_numpy_time != float('inf'):
+        print(f"🐍 NumPy Decoder: {avg_numpy_time:.2f} ms")
+
+    if avg_cpp_time != float('inf') and avg_numpy_time != float('inf'):
+        speedup = avg_numpy_time / avg_cpp_time
+        print(f"\n⚡ Speedup: C++ is {speedup:.2f}x faster than NumPy")
+
+def main():
+    # Find images
+    image_patterns = [
+        os.path.join(TEST_DATA_DIR, '*.jpg'),
+        os.path.join(TEST_DATA_DIR, '*.jpeg'),
+    ]
+    test_images = []
+    for pattern in image_patterns:
+        test_images.extend(glob.glob(pattern))
+    
+    test_images.sort()
+    
+    if not test_images:
+        print(f"❌ No images found in {TEST_DATA_DIR}")
+        return
+
+    for image_path in test_images:
+        benchmark_single_image(image_path)
+
+if __name__ == "__main__":
+    main()