From 14c01b5af544ed14f7f0cd7e259ba9425e8bd99e Mon Sep 17 00:00:00 2001
From: Jacek Galowicz <jacek@galowicz.de>
Date: Mon, 30 Jun 2025 14:21:35 +0200
Subject: [PATCH] test-driver: Parallelize OCR

---
 .../src/test_driver/machine/ocr.py            | 71 +++++++++++--------
 1 file changed, 41 insertions(+), 30 deletions(-)

diff --git a/nixos/lib/test-driver/src/test_driver/machine/ocr.py b/nixos/lib/test-driver/src/test_driver/machine/ocr.py
index 7b1cc62dd7db..c8b19c6e1066 100644
--- a/nixos/lib/test-driver/src/test_driver/machine/ocr.py
+++ b/nixos/lib/test-driver/src/test_driver/machine/ocr.py
@@ -1,3 +1,6 @@
+import itertools
+import multiprocessing
+import os
 import shutil
 import subprocess
 
@@ -9,7 +12,10 @@ def perform_ocr_on_screenshot(screenshot_path: str) -> str:
     Perform OCR on a screenshot that contains text.
     Returns a string with all words that could be found.
     """
-    return perform_ocr_variants_on_screenshot(screenshot_path, False)[0]
+    variants = perform_ocr_variants_on_screenshot(screenshot_path, False)[0]
+    if len(variants) != 1:
+        raise MachineError(f"Received wrong number of OCR results: {len(variants)}")
+    return variants[0]
 
 
 def perform_ocr_variants_on_screenshot(
@@ -29,38 +35,43 @@ def perform_ocr_variants_on_screenshot(
     #  1|lstm_only               Neural nets LSTM engine only.
     #  2|tesseract_lstm_combined Legacy + LSTM engines.
     #  3|default                 Default, based on what is available.
-    model_ids: list[int] = [0, 1, 2] if variants else [3]
+    model_ids: list[int] = [0, 1] if variants else [2]
 
-    image_paths = [
-        screenshot_path,
-        _preprocess_screenshot(screenshot_path, negate=False),
-        _preprocess_screenshot(screenshot_path, negate=True),
-    ]
+    # Tesseract runs parallel on up to 4 cores.
+    # Docs suggest to run it with OMP_THREAD_LIMIT=1 for hundreds of parallel
+    # runs. Our average test run is somewhere inbetween.
+    # https://github.com/tesseract-ocr/tesseract/issues/3109
+    processes = max(1, int(os.process_cpu_count() / 4))
+    with multiprocessing.Pool(processes=processes) as pool:
+        image_paths: list[str] = [screenshot_path]
+        if variants:
+            image_paths.extend(
+                pool.starmap(
+                    _preprocess_screenshot,
+                    [(screenshot_path, False), (screenshot_path, True)],
+                )
+            )
+        return pool.starmap(_run_tesseract, itertools.product(image_paths, model_ids))
 
-    def run_tesseract(image: str, model_id: int) -> str:
-        ret = subprocess.run(
-            [
-                "tesseract",
-                image,
-                "-",
-                "--oem",
-                str(model_id),
-                "-c",
-                "debug_file=/dev/null",
-                "--psm",
-                "11",
-            ],
-            capture_output=True,
-        )
-        if ret.returncode != 0:
-            raise MachineError(f"OCR failed with exit code {ret.returncode}")
-        return ret.stdout.decode("utf-8")
 
-    return [
-        run_tesseract(image, model_id)
-        for image in image_paths
-        for model_id in model_ids
-    ]
+def _run_tesseract(image: str, model_id: int) -> str:
+    ret = subprocess.run(
+        [
+            "tesseract",
+            image,
+            "-",
+            "--oem",
+            str(model_id),
+            "-c",
+            "debug_file=/dev/null",
+            "--psm",
+            "11",
+        ],
+        capture_output=True,
+    )
+    if ret.returncode != 0:
+        raise MachineError(f"OCR failed with exit code {ret.returncode}")
+    return ret.stdout.decode("utf-8")
 
 
 def _preprocess_screenshot(screenshot_path: str, negate: bool = False) -> str: