Major Performance Regression: pix.color_count is 150x slower in version 1.25.3 compared to 1.23.8

[mehdi-jamaseb]

Description:

I have identified a severe performance regression in the pix.color_count function in PyMuPDF when comparing versions 1.23.8 and 1.25.3. The execution time has increased by approximately 100x over these versions.

PyMuPDF Version	Execution Time for 100 Calls	Performance Drop
1.23.8	0.225 seconds	Baseline
1.23.26	24.41 seconds	~108x slower
1.25.3	33.7 seconds	~150x slower

• In PyMuPDF 1.23.8, executing pix.color_count 100 times took 0.225 seconds.
• In PyMuPDF 1.23.26, the same function took 24.41 seconds.
• In PyMuPDF 1.25.3, the function took 33.7 seconds under identical conditions.

This regression severely impacts workflows that rely on pix.color_count, making it nearly unusable for high-performance image analysis tasks.

---

Steps to Reproduce:

1. Install PyMuPDF 1.23.8: 
   pip install pymupdf==1.23.8

2. Run the following profiling script on a sample PDF file:

   import click
   import fitz  # PyMuPDF
   import cProfile
   import pstats
   import io
   def profile_function(func, iterations=100, *args, **kwargs):
   
   """Profiles a given function by running it multiple times and prints the stats."""
   
   pr = cProfile.Profile()
   
   pr.enable()
   for _ in range(iterations):
       func(*args, **kwargs)
   
   pr.disable()
   
   s = io.StringIO()
   ps = pstats.Stats(pr, stream=s).sort_stats("cumulative")
   ps.print_stats(10)  # Print top 10 time-consuming functions
   print(s.getvalue())
   
   @click.command()
   
   @click.argument("pdf_path", type=click.Path(exists=True, file_okay=True, dir_okay=False))
   
   def profile_color_count(pdf_path):
   
   """Loads a PDF, extracts the first page as an image, and counts colors 100 times using the given approach."""
   
   doc = fitz.open(pdf_path)
   
   p = doc[0]
   
   pix = p.get_pixmap(dpi=120)
   
   mat = p.rect.torect(pix.irect)
   
   rect = fitz.Rect(50.1, 50.1, 200.5, 200.3)
   print("Profiling pix.color_count (executed 100 times)...")
   profile_function(pix.color_count, iterations=100, colors=True, clip=rect * mat)
   
   if name == "main":
   
   profile_color_count()
   
   

3. Run the script with a sample PDF:

   python script.py sample.pdf
   

4. Record the execution time.

5. Upgrade to PyMuPDF 1.23.26 and repeat the test:

   pip install pymupdf==1.23.26
   

6. Upgrade to PyMuPDF 1.25.3 and repeat the test:

   pip install pymupdf==1.25.3
   

7. Compare execution times across versions.

Observed Behavior:

• PyMuPDF 1.23.8: pix.color_count runs 100 times in 0.225 seconds (very fast).
• PyMuPDF 1.23.26: Execution time jumps to 24.41 seconds (~100x slower).
• PyMuPDF 1.25.3: Execution time increases further to 33.7 seconds (~150x slower).

This suggests that a major change between 1.23.8 → 1.23.9 introduced a significant slowdown in pix.color_count.

---

Expected Behavior:

The pix.color_count function should maintain efficient performance across versions. A 100x performance drop severely impacts its usability in real-world applications.

---

System Information:

• Operating System: Windows 10
• Python Version: 3.12
• CPU: AMD Ryzen 5 3600 (6-Core, 3.60 GHz)
• RAM: 16 GB
• PyMuPDF Versions Tested:
  • ✅ 1.23.8 → Fast (0.225 sec for 100 calls)
  • ⚠️ 1.23.26 → Slow (24.41 sec for 100 calls, ~100x slower)
  • ❌ 1.25.3 → Very slow (33.7 sec for 100 calls, ~150x slower)

---

Additional Information:

This drastic performance drop makes pix.color_count nearly unusable for batch image processing. If a specific change in 1.23.9 or later caused this regression, it would be helpful to investigate whether the previous approach can be optimized or restored.

---

📌 Thank you for your time and for maintaining PyMuPDF! 🚀

[julian-smith-artifex-com]

Thanks for reporting this.

I've reproduced, and fixed by adding an internal C++ implementation of the main part of Pixmap.color_count().

My very simple measurements of one test case indicate it might be slightly faster than PyMuPDF-1.23.8.

[mehdi-jamaseb]

Thanks for the quick fix! I appreciate the improvement and the performance enhancement. Looking forward to testing it in the next release.

[julian-smith-artifex-com]

Fixed in PyMuPDF-1.25.4.