Speed up quantiles with sorting

.github/workflows/main.yml

@@ -135,7 +135,7 @@ jobs:
           - tox-env: py310-coverage-lmoments  # No markers -- includes slow tests
             python-version: "3.10"
             os: ubuntu-latest
-          - tox-env: py311-coverage-sbck
+          - tox-env: py311-coverage-sbck-extras
             python-version: "3.11"
             markers: -m 'not slow'
             os: ubuntu-latest

AUTHORS.rst

@@ -29,7 +29,7 @@ Contributors
 * David Caron `@davidcaron <https://github.com/davidcaron>`_
 * Carsten Ehbrecht <ehbrecht@dkrz.de> `@cehbrecht <https://github.com/cehbrecht>`_
 * Jeremy Fyke `@jeremyfyke <https://github.com/jeremyfyke>`_
-* Sarah Gammon <gammon.sarah@ouranos.ca> `@SarahG-579462 <https://github.com/SarahG-579462>`_
+* Sarah Gammon `@SarahG-579462 <https://github.com/SarahG-579462>`_
 * Tom Keel <thomas.keel.18@ucl.ac.uk> `@Thomasjkeel <https://github.com/Thomasjkeel>`_
 * Marie-Pier Labonté <labonte.marie-pier@ouranos.ca> `@marielabonte <https://github.com/marielabonte>`_
 * Ludwig Lierhammer <ludwig.lierhammer@hereon.de> `@ludwiglierhammer <https://github.com/ludwiglierhammer>`_

CHANGELOG.rst

@@ -4,7 +4,11 @@ Changelog
 
 v0.52.0 (unreleased)
 --------------------
-Contributors to this version: David Huard (:user:`huard`), Trevor James Smith (:user:`Zeitsperre`), Hui-Min Wang (:user:`Hem-W`).
+Contributors to this version: David Huard (:user:`huard`), Trevor James Smith (:user:`Zeitsperre`), Hui-Min Wang (:user:`Hem-W`), Éric Dupuis (:user:`coxipi`), Sarah Gammon (:user:`SarahG-579462`).
+
+New features and enhancements
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+* ``xclim.sdba.nbutils.quantile`` and its child functions are now faster. If the module `fastnanquantile` is installed, it is used as the backend for the computation of quantiles and yields even faster results. (:issue:`1255`, :pull:`1513`).
 
 Bug fixes
 ^^^^^^^^^

docs/conf.py

@@ -267,6 +267,7 @@ class XCStyle(AlphaStyle):
     "_build",
     "Thumbs.db",
     ".DS_Store",
+    "notebooks/benchmarks",
     "notebooks/xclim_training",
     "paper/paper.md",
     "**.ipynb_checkpoints",

docs/notebooks/benchmarks/sdba_quantile.ipynb

@@ -0,0 +1,142 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from __future__ import annotations\n",
+    "\n",
+    "import time\n",
+    "\n",
+    "import numpy as np\n",
+    "\n",
+    "import xclim\n",
+    "from xclim import sdba\n",
+    "from xclim.testing import open_dataset\n",
+    "\n",
+    "ds = open_dataset(\"sdba/CanESM2_1950-2100.nc\")\n",
+    "tx = ds.sel(time=slice(\"1950\", \"1980\")).tasmax\n",
+    "kws = {\"dim\": \"time\", \"q\": np.linspace(0, 1, 50)}"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Tests with %%timeit (full 30 years)\n",
+    "\n",
+    "Here `fastnanquantile` is the best algorithm out of \n",
+    "* `xr.DataArray.quantile`\n",
+    "* `nbutils.quantile`, using: \n",
+    "  * `xclim.core.utils.nan_quantile`\n",
+    "  * `fastnanquantile`\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "%%timeit\n",
+    "tx.quantile(**kws).compute()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "%%timeit\n",
+    "sdba.nbutils.USE_FASTNANQUANTILE = False\n",
+    "sdba.nbutils.quantile(tx, **kws).compute()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "! pip install fastnanquantile"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "%%timeit\n",
+    "sdba.nbutils.USE_FASTNANQUANTILE = True\n",
+    "sdba.nbutils.quantile(tx, **kws).compute()"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Test computation time as a function of number of points\n",
+    "\n",
+    "For a smaller number of time steps <=2000, `_sortquantile` is the best algorithm in general"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import time\n",
+    "\n",
+    "import matplotlib.pyplot as plt\n",
+    "import xarray as xr\n",
+    "\n",
+    "num_tests = 500\n",
+    "timed = {}\n",
+    "# fastnanquantile has nothing to do with sortquantile\n",
+    "# I just added a third step using this variable\n",
+    "\n",
+    "for use_fnq in [True, False]:\n",
+    "    sdba.nbutils.USE_FASTNANQUANTILE = use_fnq\n",
+    "    # heat-up the jit\n",
+    "    sdba.nbutils.quantile(\n",
+    "        xr.DataArray(np.array([0, 1.5])), dim=\"dim_0\", q=np.array([0.5])\n",
+    "    )\n",
+    "    for size in np.arange(250, 2000 + 250, 250):\n",
+    "        da = tx.isel(time=slice(0, size))\n",
+    "        t0 = time.time()\n",
+    "        for ii in range(num_tests):\n",
+    "            sdba.nbutils.quantile(da, **kws).compute()\n",
+    "        timed[use_fnq].append([size, time.time() - t0])\n",
+    "\n",
+    "for k, lab in zip([True, False], [\"xclim.core.utils.nan_quantile\", \"fastnanquantile\"]):\n",
+    "    arr = np.array(timed[k])\n",
+    "    plt.plot(arr[:, 0], arr[:, 1] / num_tests, label=lab)\n",
+    "plt.legend()\n",
+    "plt.title(\"Quantile computation, average time vs array size, for 50 quantiles\")\n",
+    "plt.xlabel(\"Number of time steps in the distribution\")\n",
+    "plt.ylabel(\"Computation time (s)\")"
+   ]
+  }
+ ],
+ "metadata": {
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.12.3"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 2
+}

pyproject.toml

@@ -110,7 +110,8 @@ docs = [
   "sphinxcontrib-bibtex",
   "sphinxcontrib-svg2pdfconverter[Cairosvg]"
 ]
-all = ["xclim[dev]", "xclim[docs]"]
+extras = ["fastnanquantile"]
+all = ["xclim[dev]", "xclim[docs]", "xclim[extras]"]
 
 [project.scripts]
 xclim = "xclim.cli:cli"

tests/test_sdba/test_nbutils.py

@@ -0,0 +1,36 @@
+from __future__ import annotations
+
+import numpy as np
+import pytest
+import xarray as xr
+
+from xclim.sdba import nbutils as nbu
+
+
+class TestQuantiles:
+    @pytest.mark.parametrize("uses_dask", [True, False])
+    def test_quantile(self, open_dataset, uses_dask):
+        da = (
+            open_dataset("sdba/CanESM2_1950-2100.nc").sel(time=slice("1950", "1955")).pr
+        ).load()
+        if uses_dask:
+            da = da.chunk({"location": 1})
+        else:
+            da = da.load()
+        q = np.linspace(0.1, 0.99, 50)
+        out_nbu = nbu.quantile(da, q, dim="time").transpose("location", ...)
+        out_xr = da.quantile(q=q, dim="time").transpose("location", ...)
+        np.testing.assert_array_almost_equal(out_nbu.values, out_xr.values)
+
+    def test_edge_cases(self, open_dataset):
+        q = np.linspace(0.1, 0.99, 50)
+
+        # only 1 non-null value
+        da = xr.DataArray([1] + [np.nan] * 100, dims="dim_0")
+        out_nbu = nbu.quantile(da, q, dim="dim_0")
+        np.testing.assert_array_equal(out_nbu.values, np.full_like(q, 1))
+
+        # only NANs
+        da = xr.DataArray([np.nan] * 100, dims="dim_0")
+        out_nbu = nbu.quantile(da, q, dim="dim_0")
+        np.testing.assert_array_equal(out_nbu.values, np.full_like(q, np.nan))

tox.ini

@@ -113,7 +113,9 @@ passenv =
     LD_LIBRARY_PATH
     SKIP_NOTEBOOKS
     XCLIM_*
-extras = dev
+extras =
+    dev
+    extras: extras
 deps =
     upstream: -r CI/requirements_upstream.txt
     sbck: pybind11