format

.github/workflows/tests.yml

@@ -29,12 +29,11 @@ jobs:
       run: |
         sudo apt-get install -y libopenmpi-dev
         python -m pip install --upgrade pip
-        pip install jax==0.4.35
-        pip install numpy setuptools cython wheel
-        pip install git+https://github.com/MP-Gadget/pfft-python
-        pip install git+https://github.com/MP-Gadget/pmesh
-        pip install git+https://github.com/ASKabalan/fastpm-python --no-build-isolation
-        pip install -r requirements-test.txt
+        pip install jax
+        pip install setuptools cython wheel mpi4py
+        pip install -r requirements-test.txt  --no-build-isolation
+        pip install pytest
+        pip install diffrax
         pip install .
 
     - name: Run Single Device Tests
@@ -43,4 +42,4 @@ jobs:
         pytest -v -m "not distributed"
     - name: Run Distributed tests
       run: |
-        pytest -v -m distributed
+        pytest -v tests/test_distributed_pm.py

jaxpm/distributed.py

@@ -166,7 +166,7 @@ def uniform_particles(mesh_shape, sharding=None):
                          axis=-1)
 
 
-def normal_field(mesh_shape, seed, sharding=None , dtype='float32'):
+def normal_field(mesh_shape, seed, sharding=None, dtype='float32'):
     """Generate a Gaussian random field with the given power spectrum."""
     gpu_mesh = sharding.mesh if sharding is not None else None
     if gpu_mesh is not None and not (gpu_mesh.empty):

jaxpm/growth.py

@@ -1,3 +1,5 @@
+import os
+
 import jax.numpy as np
 from jax.numpy import interp
 from jax_cosmo.background import *
@@ -243,56 +245,61 @@ def _growth_factor_ODE(cosmo, a, log10_amin=-3, steps=256, eps=1e-4):
         Growth factor computed at requested scale factor
     """
     # Check if growth has already been computed
-    #if not "background.growth_factor" in cosmo._workspace.keys():
-    # Compute tabulated array
-    atab = np.logspace(log10_amin, 0.0, steps)
+    CACHING_ACTIVATED = os.environ.get("JC_CACHE", "1") == "1"
+    if CACHING_ACTIVATED and "background.growth_factor" in cosmo._workspace.keys(
+    ):
+        cache = cosmo._workspace["background.growth_factor"]
+    else:
+        # Compute tabulated array
+        atab = np.logspace(log10_amin, 0.0, steps)
 
-    def D_derivs(y, x):
-        q = (2.0 - 0.5 *
-                (Omega_m_a(cosmo, x) +
-                (1.0 + 3.0 * w(cosmo, x)) * Omega_de_a(cosmo, x))) / x
-        r = 1.5 * Omega_m_a(cosmo, x) / x / x
+        def D_derivs(y, x):
+            q = (2.0 - 0.5 *
+                 (Omega_m_a(cosmo, x) +
+                  (1.0 + 3.0 * w(cosmo, x)) * Omega_de_a(cosmo, x))) / x
+            r = 1.5 * Omega_m_a(cosmo, x) / x / x
 
-        g1, g2 = y[0]
-        f1, f2 = y[1]
-        dy1da = [f1, -q * f1 + r * g1]
-        dy2da = [f2, -q * f2 + r * g2 - r * g1**2]
-        return np.array([[dy1da[0], dy2da[0]], [dy1da[1], dy2da[1]]])
+            g1, g2 = y[0]
+            f1, f2 = y[1]
+            dy1da = [f1, -q * f1 + r * g1]
+            dy2da = [f2, -q * f2 + r * g2 - r * g1**2]
+            return np.array([[dy1da[0], dy2da[0]], [dy1da[1], dy2da[1]]])
 
-    y0 = np.array([[atab[0], -3.0 / 7 * atab[0]**2],
-                    [1.0, -6.0 / 7 * atab[0]]])
-    y = odeint(D_derivs, y0, atab)
+        y0 = np.array([[atab[0], -3.0 / 7 * atab[0]**2],
+                       [1.0, -6.0 / 7 * atab[0]]])
+        y = odeint(D_derivs, y0, atab)
 
-    # compute second order derivatives growth
-    dyda2 = D_derivs(np.transpose(y, (1, 2, 0)), atab)
-    dyda2 = np.transpose(dyda2, (2, 0, 1))
+        # compute second order derivatives growth
+        dyda2 = D_derivs(np.transpose(y, (1, 2, 0)), atab)
+        dyda2 = np.transpose(dyda2, (2, 0, 1))
 
-    # Normalize results
-    y1 = y[:, 0, 0]
-    gtab = y1 / y1[-1]
-    y2 = y[:, 0, 1]
-    g2tab = y2 / y2[-1]
-    # To transform from dD/da to dlnD/dlna: dlnD/dlna = a / D dD/da
-    ftab = y[:, 1, 0] / y1[-1] * atab / gtab
-    f2tab = y[:, 1, 1] / y2[-1] * atab / g2tab
-    # Similarly for second order derivatives
-    # Note: these factors are not accessible as parent functions yet
-    # since it is unclear what to refer to them with.
-    htab = dyda2[:, 1, 0] / y1[-1] * atab / gtab
-    h2tab = dyda2[:, 1, 1] / y2[-1] * atab / g2tab
+        # Normalize results
+        y1 = y[:, 0, 0]
+        gtab = y1 / y1[-1]
+        y2 = y[:, 0, 1]
+        g2tab = y2 / y2[-1]
+        # To transform from dD/da to dlnD/dlna: dlnD/dlna = a / D dD/da
+        ftab = y[:, 1, 0] / y1[-1] * atab / gtab
+        f2tab = y[:, 1, 1] / y2[-1] * atab / g2tab
+        # Similarly for second order derivatives
+        # Note: these factors are not accessible as parent functions yet
+        # since it is unclear what to refer to them with.
+        htab = dyda2[:, 1, 0] / y1[-1] * atab / gtab
+        h2tab = dyda2[:, 1, 1] / y2[-1] * atab / g2tab
 
+        cache = {
+            "a": atab,
+            "g": gtab,
+            "f": ftab,
+            "h": htab,
+            "g2": g2tab,
+            "f2": f2tab,
+            "h2": h2tab,
+        }
+        if CACHING_ACTIVATED:
+            cosmo._workspace["background.growth_factor"] = cache
 
-    cache = {
-        "a": atab,
-        "g": gtab,
-        "f": ftab,
-        "h": htab,
-        "g2": g2tab,
-        "f2": f2tab,
-        "h2": h2tab,
-    }
-
-    return np.clip(interp(a, cache["a"], cache["g"]), 0.0, 1.0) , cache
+    return np.clip(interp(a, cache["a"], cache["g"]), 0.0, 1.0), cache
 
 
 def _growth_rate_ODE(cosmo, a):
@@ -317,6 +324,7 @@ def _growth_rate_ODE(cosmo, a):
     cache = _growth_factor_ODE(cosmo, np.atleast_1d(1.0))[1]
     return interp(a, cache["a"], cache["f"])
 
+
 def _growth_factor_second_ODE(cosmo, a):
     """Compute second order growth factor D2(a) at a given scale factor,
     normalised such that D(a=1) = 1.
@@ -384,7 +392,11 @@ def _growth_factor_gamma(cosmo, a, log10_amin=-3, steps=128):
 
     """
     # Check if growth has already been computed, if not, compute it
-    if not "background.growth_factor" in cosmo._workspace.keys():
+    CACHING_ACTIVATED = os.environ.get("JC_CACHE", "1") == "1"
+    if CACHING_ACTIVATED and "background.growth_factor" in cosmo._workspace.keys(
+    ):
+        cache = cosmo._workspace["background.growth_factor"]
+    else:
         # Compute tabulated array
         atab = np.logspace(log10_amin, 0.0, steps)
 
@@ -395,9 +407,8 @@ def _growth_factor_gamma(cosmo, a, log10_amin=-3, steps=128):
         gtab = np.exp(odeint(integrand, np.log(atab[0]), np.log(atab)))
         gtab = gtab / gtab[-1]  # Normalize to a=1.
         cache = {"a": atab, "g": gtab}
-        cosmo._workspace["background.growth_factor"] = cache
-    else:
-        cache = cosmo._workspace["background.growth_factor"]
+        if CACHING_ACTIVATED:
+            cosmo._workspace["background.growth_factor"] = cache
     return np.clip(interp(a, cache["a"], cache["g"]), 0.0, 1.0)
 
 
@@ -522,6 +533,7 @@ def gp(cosmo, a):
     D1f = f1 * g1 / a
     return D1f
 
+
 def dGfa(cosmo, a):
     r""" Derivative of Gf against a
 

jaxpm/painting.py

@@ -240,7 +240,12 @@ def _cic_read_dx_impl(grid_mesh, disp, halo_size):
 def cic_read_dx(grid_mesh, disp, halo_size=0, sharding=None):
 
     halo_size, halo_extents = get_halo_size(halo_size, sharding=sharding)
-    halo_size = jax.tree.map(lambda x: x//2, halo_size)
+    # Halo size is halved for the read operation
+    # We only need to read the density field
+    # while in the painting operation we need to exchange and reduce the halo
+    # We chose to do that since it is much easier to write a custom jvp rule for exchange
+    # while it is a bit harder if there is a reduction involved
+    halo_size = jax.tree.map(lambda x: x // 2, halo_size)
     grid_mesh = slice_pad(grid_mesh, halo_size, sharding=sharding)
     grid_mesh = halo_exchange(grid_mesh,
                               halo_extents=halo_extents,

jaxpm/painting_utils.py

@@ -30,17 +30,14 @@ def enmesh(base_indices, displacements, cell_size, base_shape, offset,
     """Multilinear enmeshing."""
     base_indices = jnp.asarray(base_indices)
     displacements = jnp.asarray(displacements)
-    with jax.experimental.enable_x64():
-        cell_size = jnp.float64(
-            cell_size) if new_cell_size is not None else jnp.array(
-                cell_size, dtype=displacements.dtype)
-        if base_shape is not None:
-            base_shape = jnp.array(base_shape, dtype=base_indices.dtype)
-        offset = jnp.float64(offset)
-        if new_cell_size is not None:
-            new_cell_size = jnp.float64(new_cell_size)
-        if new_shape is not None:
-            new_shape = jnp.array(new_shape, dtype=base_indices.dtype)
+    cell_size = jnp.array(cell_size, dtype=displacements.dtype)
+    if base_shape is not None:
+        base_shape = jnp.array(base_shape, dtype=base_indices.dtype)
+    offset = offset.astype(base_indices.dtype)
+    if new_cell_size is not None:
+        new_cell_size = jnp.array(new_cell_size, dtype=displacements.dtype)
+    if new_shape is not None:
+        new_shape = jnp.array(new_shape, dtype=base_indices.dtype)
 
     spatial_dim = base_indices.shape[1]
     neighbor_offsets = (

requirements-test.txt

@@ -1,5 +1,3 @@
-pytest>=8.0.0
-diffrax
 pfft-python @ git+https://github.com/MP-Gadget/pfft-python
 pmesh @ git+https://github.com/MP-Gadget/pmesh
 fastpm @ git+https://github.com/ASKabalan/fastpm-python

tests/test_distributed_pm.py

@@ -22,7 +22,7 @@ from jaxpm.distributed import fft3d, ifft3d
 from jaxpm.painting import cic_paint, cic_paint_dx  # noqa : E402
 from jaxpm.pm import lpt, make_diffrax_ode, pm_forces  # noqa : E402
 
-_TOLERANCE = 1e-1  # 🙃🙃
+_TOLERANCE = 1e-6  # 🎉🎉🎉
 
 pdims = [(1, 8), (8, 1), (4, 2), (2, 4)]