René's URL Explorer Experiment

{"@context":"https://schema.org","@type":"DiscussionForumPosting","headline":"[MNT] Switch from `np.column_stack()` to `np.vstack().T` for performance","articleBody":"From the discussion in https://github.com/matplotlib/matplotlib/pull/31001#discussion_r2790413429, it looks like `np.column_stack()` is generally a slow operation compared to `np.vstack().T`. This is because the former has to interleave elements in memory whereas the second does contiguous memory copies and returns a view.\n\nIt's unclear to me how many of these operations are driving time spent in the hot paths of the code, but we should see a performance improvement by switching things over.\n\nMarking this as an easy first issue since it's largely a find-and-replace.\n\n```\n10,000 elements: 10 runs x 10,000 iterations\n\nWith broadcast:\n- `np.column_stack(np.broadcast_arrays(x, y))`: 36.47 us\n- `np.vstack(np.broadcast_arrays(x, y)).T`: 27.67 us\n- `np.empty + assign`: 30.09 us\n\nWithout broadcast:\n- `np.column_stack([x, y])`: 20.63 us\n- `np.vstack((x, y)).T`: 13.18 us\n```\n\n```python\n\"\"\"Benchmark different array combination methods.\"\"\"\n\nimport timeit\nimport numpy as np\n\nN = 10_000\nNUMBER = 10_000\nREPEAT = 10\n\nx = np.linspace(0, 1, N)\ny = np.float64(2.0)\n\n# Pre-matched arrays for non-broadcast cases\nx_full = x\ny_full = np.full(N, 2.0)\n\n\ndef broadcast_column_stack():\n    return np.column_stack(np.broadcast_arrays(x, y))\n\ndef broadcast_vstack_T():\n    return np.vstack(np.broadcast_arrays(x, y)).T\n\ndef broadcast_empty_assign():\n    out = np.empty((N, 2))\n    bx, by = np.broadcast_arrays(x, y)\n    out[:, 0] = bx\n    out[:, 1] = by\n    return out\n\ndef no_broadcast_column_stack():\n    return np.column_stack([x_full, y_full])\n\ndef no_broadcast_vstack_T():\n    return np.vstack((x_full, y_full)).T\n\n\ndef bench(func):\n    times = timeit.repeat(func, number=NUMBER, repeat=REPEAT)\n    return min(times) / NUMBER\n\n\nif __name__ == \"__main__\":\n    print(f\"{N:,} elements: {REPEAT} runs x {NUMBER:,} iterations\")\n\n    broadcast_cases = [\n        (\"np.column_stack(np.broadcast_arrays(x, y))\", broadcast_column_stack),\n        (\"np.vstack(np.broadcast_arrays(x, y)).T\", broadcast_vstack_T),\n        (\"np.empty + assign\", broadcast_empty_assign),\n    ]\n\n    no_broadcast_cases = [\n        (\"np.column_stack([x, y])\", no_broadcast_column_stack),\n        (\"np.vstack((x, y)).T\", no_broadcast_vstack_T),\n    ]\n\n    print(\"\\nWith broadcast:\")\n    for label, func in broadcast_cases:\n        t = bench(func)\n        print(f\"- `{label}`: {t * 1e6:.2f} us\")\n\n    print(\"\\nWithout broadcast:\")\n    for label, func in no_broadcast_cases:\n        t = bench(func)\n        print(f\"- `{label}`: {t * 1e6:.2f} us\")\n```","author":{"url":"https://github.com/scottshambaugh","@type":"Person","name":"scottshambaugh"},"datePublished":"2026-02-10T21:53:30.000Z","interactionStatistic":{"@type":"InteractionCounter","interactionType":"https://schema.org/CommentAction","userInteractionCount":10},"url":"https://github.com/31130/matplotlib/issues/31130"}