Add xpu support to decoder benchmark scripts

benchmarks/decoders/benchmark_decoders.py

@@ -126,7 +126,10 @@ def main() -> None:
             options = decoder_registry[decoder].default_options
 
         kind = decoder_registry[decoder].kind
-        decoders_to_run[display] = kind(**options)
+        try:
+            decoders_to_run[display] = kind(**options)
+        except (ImportError, RuntimeError):
+            print(f"Warning: unable to create decoder {decoder}. Test will be skipped.")
 
     video_paths = args.video_paths.split(",")
     if args.video_dir:
@@ -158,6 +161,11 @@ def main() -> None:
                     if torch.cuda.is_available()
                     else "not available"
                 ),
+                "xpu": (
+                    torch.xpu.get_device_properties(0).name
+                    if torch.xpu.is_available()
+                    else "not available"
+                ),
             },
         }
         plot_data(data, args.plot_path)

benchmarks/decoders/benchmark_decoders_library.py

@@ -603,6 +603,35 @@ def decode_and_resize(self, video_file, pts_list, height, width, device):
         return frames
 
 
+# check whether specified decoder is supported
+def check_decoder_support(decoder_type):
+    if decoder_type == "cuda":
+        return torch.cuda.is_available()
+
+    if decoder_type == "xpu":
+        return torch.xpu.is_available()
+
+    if decoder_type == "torchaudio":
+        try:
+            import torchaudio  # noqa: F401
+        except ImportError:
+            return False
+        return True
+
+    if decoder_type == "torchvision":
+        try:
+            # also check whether video_decoder backend is installed
+            import torchvision
+
+            torchvision.set_video_backend("video_reader")
+        except (ImportError, RuntimeError):
+            return False
+        return True
+
+    print(f"Warning: unknown decoder_type {decoder_type}")
+    return False
+
+
 def create_torchcodec_core_decode_first_frame(video_file):
     video_decoder = create_from_file(video_file)
     _add_video_stream(video_decoder)

benchmarks/decoders/generate_readme_data.py

@@ -4,6 +4,7 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
+import argparse
 import json
 import os
 import platform
@@ -14,6 +15,7 @@
 
 from benchmark_decoders_library import (
     BatchParameters,
+    check_decoder_support,
     DataLoaderInspiredWorkloadParameters,
     generate_videos,
     retrieve_videos,
@@ -29,6 +31,50 @@
 def main() -> None:
     """Benchmarks the performance of a few video decoders on synthetic videos"""
 
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--decoders",
+        help=(
+            "Comma-separated list of decoders to benchmark. Valid options: cpu, cuda, xpu, torchvision, torchaudio"
+        ),
+        type=str,
+        default=("cpu,cuda,torchvision,torchaudio"),
+    ),
+    parser.add_argument(
+        "--resize_device",
+        help=(
+            "Device for resize. Default: cuda if available, else cpu. Valid options: cpu, cuda, xpu"
+        ),
+        type=str,
+        default=("cuda"),
+    ),
+    args = parser.parse_args()
+
+    # warn if device or module is not available
+    decoder_list = set(args.decoders.split(","))
+    if "xpu" in decoder_list and not check_decoder_support("xpu"):
+        print("Warning: xpu is not available. Test will be skipped.")
+        decoder_list.remove("xpu")
+
+    if "cuda" in decoder_list and not check_decoder_support("cuda"):
+        print("Warning: cuda is not available. Test will be skipped.")
+        decoder_list.remove("cuda")
+
+    if "torchaudio" in decoder_list and not check_decoder_support("torchaudio"):
+        print("Warning: torchaudio is not available. Test will be skipped.")
+        decoder_list.remove("torchaudio")
+
+    if "torchvision" in decoder_list and not check_decoder_support("torchvision"):
+        print("Warning: torchvision is not available. Test will be skipped.")
+        decoder_list.remove("torchvision")
+
+    resize_device = args.resize_device
+    if resize_device == "cuda" and not torch.cuda.is_available():
+        resize_device = "cpu"
+    if resize_device == "xpu" and not torch.xpu.is_available():
+        resize_device = "cpu"
+    print(f"resize_device = {resize_device}")
+
     videos_dir_path = "/tmp/torchcodec_benchmarking_videos"
     if not os.path.exists(videos_dir_path):
         shutil.rmtree(videos_dir_path, ignore_errors=True)
@@ -60,15 +106,27 @@ def main() -> None:
         retrieve_videos(urls_and_dest_paths)
 
     decoder_dict = {}
-    decoder_dict["torchcodec"] = TorchCodecPublic()
-    decoder_dict["torchcodec[approx]"] = TorchCodecPublic(seek_mode="approximate")
-    if torch.cuda.is_available():
+    if "cpu" in decoder_list:
+        decoder_dict["torchcodec"] = TorchCodecPublic()
+        decoder_dict["torchcodec[approx]"] = TorchCodecPublic(seek_mode="approximate")
+
+    if "cuda" in decoder_list:
         decoder_dict["torchcodec[cuda]"] = TorchCodecPublic(device="cuda")
         decoder_dict["torchcodec[cuda,approx]"] = TorchCodecPublic(
             device="cuda", seek_mode="approximate"
         )
-    decoder_dict["torchvision[video_reader]"] = TorchVision("video_reader")
-    decoder_dict["torchaudio"] = TorchAudioDecoder()
+
+    if "xpu" in decoder_list:
+        decoder_dict["torchcodec[xpu]"] = TorchCodecPublic(device="xpu")
+        decoder_dict["torchcodec[xpu,approx]"] = TorchCodecPublic(
+            device="xpu", seek_mode="approximate"
+        )
+
+    if "torchvision" in decoder_list:
+        decoder_dict["torchvision[video_reader]"] = TorchVision("video_reader")
+
+    if "torchaudio" in decoder_list:
+        decoder_dict["torchaudio"] = TorchAudioDecoder()
 
     # These are the number of uniform seeks we do in the seek+decode benchmark.
     num_samples = 10
@@ -85,7 +143,7 @@ def main() -> None:
             batch_parameters=BatchParameters(batch_size=50, num_threads=10),
             resize_height=256,
             resize_width=256,
-            resize_device="cuda" if torch.cuda.is_available() else "cpu",
+            resize_device=resize_device,
         ),
     )
     data_for_writing = {
@@ -100,6 +158,11 @@ def main() -> None:
                 if torch.cuda.is_available()
                 else "not available"
             ),
+            "xpu": (
+                torch.xpu.get_device_properties(0).name
+                if torch.xpu.is_available()
+                else "not available"
+            ),
         },
     }
 

benchmarks/decoders/gpu_benchmark.py

@@ -8,7 +8,12 @@
 import torch.utils.benchmark as benchmark
 
 import torchcodec
-import torchvision.transforms.v2.functional as F
+
+try:
+    import torchvision.transforms.v2.functional as F
+except ImportError:
+    print("Error importing torchvision. Exiting.")
+    exit()
 
 RESIZED_WIDTH = 256
 RESIZED_HEIGHT = 256
@@ -27,7 +32,7 @@ def decode_full_video(video_path, decode_device_string, resize_device_string):
     print(f"{decode_device_string=} {resize_device_string=}")
     decoder = torchcodec._core.create_from_file(video_path)
     num_threads = None
-    if "cuda" in decode_device_string:
+    if "cuda" in decode_device_string or "xpu" in decode_device_string:
         num_threads = 1
     width = None
     height = None
@@ -78,6 +83,8 @@ def decode_videos_using_threads(
         actual_decode_device = decode_device_string
         if "cuda" in decode_device_string and use_multiple_gpus:
             actual_decode_device = f"cuda:{i % torch.cuda.device_count()}"
+        if "xpu" in decode_device_string and use_multiple_gpus:
+            actual_decode_device = f"xpu:{i % torch.xpu.device_count()}"
         executor.submit(
             decode_full_video, video_path, actual_decode_device, resize_device_string
         )
@@ -154,10 +161,14 @@ def main():
             if "cuda" in decode_label:
                 # Shorten "cuda:0" to "cuda"
                 decode_label = "cuda"
+            if "xpu" in decode_label:
+                decode_label = "xpu"
             resize_label = resize_device_string
             if "cuda" in resize_device_string:
                 # Shorten "cuda:0" to "cuda"
                 resize_label = "cuda"
+            if "xpu" in resize_device_string:
+                resize_label = "xpu"
             print("decode_device", decode_device_string)
             print("resize_device", resize_device_string)
             if args.num_threads > 1: