JaxPM/jaxpm/distributed.py
Wassim KABALAN df8602b318 jaxdecomp proto (#21)
* adding example of distributed solution

* put back old functgion

* update formatting

* add halo exchange and slice pad

* apply formatting

* implement distributed optimized cic_paint

* Use new cic_paint with halo

* Fix seed for distributed normal

* Wrap interpolation function to avoid all gather

* Return normal order frequencies for single GPU

* add example

* format

* add optimised bench script

* times in ms

* add lpt2

* update benchmark and add slurm

* Visualize only final field

* Update scripts/distributed_pm.py

Co-authored-by: Francois Lanusse <EiffL@users.noreply.github.com>

* Adjust pencil type for frequencies

* fix painting issue with slabs

* Shared operation in fourrier space now take inverted sharding axis for
slabs

* add assert to make pyright happy

* adjust test for hpc-plotter

* add PMWD test

* bench

* format

* added github workflow

* fix formatting from main

* Update for jaxDecomp pure JAX

* revert single halo extent change

* update for latest jaxDecomp

* remove fourrier_space in autoshmap

* make normal_field work with single controller

* format

* make distributed pm work in single controller

* merge bench_pm

* update to leapfrog

* add a strict dependency on jaxdecomp

* global mesh no longer needed

* kernels.py no longer uses global mesh

* quick fix in distributed

* pm.py no longer uses global mesh

* painting.py no longer uses global mesh

* update demo script

* quick fix in kernels

* quick fix in distributed

* update demo

* merge hugos LPT2 code

* format

* Small fix

* format

* remove duplicate get_ode_fn

* update visualizer

* update compensate CIC

* By default check_rep is false for shard_map

* remove experimental distributed code

* update PGDCorrection and neural ode to use new fft3d

* jaxDecomp pfft3d promotes to complex automatically

* remove deprecated stuff

* fix painting issue with read_cic

* use jnp interp instead of jc interp

* delete old slurms

* add notebook examples

* apply formatting

* add distributed zeros

* fix code in LPT2

* jit cic_paint

* update notebooks

* apply formating

* get local shape and zeros can be used by users

* add a user facing function to create uniform particle grid

* use jax interp instead of jax_cosmo

* use float64 for enmeshing

* Allow applying weights with relative cic paint

* Weights can be traced

* remove script folder

* update example notebooks

* delete outdated design file

* add readme for tutorials

* update readme

* fix small error

* forgot particles in multi host

* clarifying why cic_paint_dx is slower

* clarifying the halo size dependence on the box size

* ability to choose snapshots number with MultiHost script

* Adding animation notebook

* Put plotting in package

* Add finite difference laplace kernel + powerspec functions from Hugo

Co-authored-by: Hugo Simonfroy <hugo.simonfroy@gmail.com>

* Put plotting utils in package

* By default use absoulute painting with

* update code

* update notebooks

* add tests

* Upgrade setup.py to pyproject

* Format

* format tests

* update test dependencies

* add test workflow

* fix deprecated FftType in jaxpm.kernels

* Add aboucaud comments

* JAX version is 0.4.35 until Diffrax new release

* add numpy explicitly as dependency for tests

* fix install order for tests

* add numpy to be installed

* enforce no build isolation for fastpm

* pip install jaxpm test without build isolation

* bump jaxdecomp version

* revert test workflow

* remove outdated tests

---------

Co-authored-by: EiffL <fr.eiffel@gmail.com>
Co-authored-by: Francois Lanusse <EiffL@users.noreply.github.com>
Co-authored-by: Wassim KABALAN <wassim@apc.in2p3.fr>
Co-authored-by: Hugo Simonfroy <hugo.simonfroy@gmail.com>
Former-commit-id: 8c2e823d4669eac712089bf7f85ffb7912e8232d
2024-12-20 05:44:02 -05:00

198 lines
6.4 KiB
Python

from typing import Any, Callable, Hashable
Specs = Any
AxisName = Hashable
from functools import partial
import jax
import jax.numpy as jnp
import jaxdecomp
from jax import lax
from jax.experimental.shard_map import shard_map
from jax.sharding import AbstractMesh, Mesh
from jax.sharding import PartitionSpec as P
def autoshmap(
f: Callable,
gpu_mesh: Mesh | AbstractMesh | None,
in_specs: Specs,
out_specs: Specs,
check_rep: bool = False,
auto: frozenset[AxisName] = frozenset()) -> Callable:
"""Helper function to wrap the provided function in a shard map if
the code is being executed in a mesh context."""
if gpu_mesh is None or gpu_mesh.empty:
return f
else:
return shard_map(f, gpu_mesh, in_specs, out_specs, check_rep, auto)
def fft3d(x):
return jaxdecomp.pfft3d(x)
def ifft3d(x):
return jaxdecomp.pifft3d(x).real
def get_halo_size(halo_size, sharding):
gpu_mesh = sharding.mesh if sharding is not None else None
if gpu_mesh is None or gpu_mesh.empty:
zero_ext = (0, 0)
zero_tuple = (0, 0)
return (zero_tuple, zero_tuple, zero_tuple), zero_ext
else:
pdims = gpu_mesh.devices.shape
halo_x = (0, 0) if pdims[0] == 1 else (halo_size, halo_size)
halo_y = (0, 0) if pdims[1] == 1 else (halo_size, halo_size)
halo_x_ext = 0 if pdims[0] == 1 else halo_size // 2
halo_y_ext = 0 if pdims[1] == 1 else halo_size // 2
return ((halo_x, halo_y, (0, 0)), (halo_x_ext, halo_y_ext))
def halo_exchange(x, halo_extents, halo_periods=(True, True)):
if (halo_extents[0] > 0 or halo_extents[1] > 0):
return jaxdecomp.halo_exchange(x, halo_extents, halo_periods)
else:
return x
def slice_unpad_impl(x, pad_width):
halo_x, _ = pad_width[0]
halo_y, _ = pad_width[1]
# Apply corrections along x
x = x.at[halo_x:halo_x + halo_x // 2].add(x[:halo_x // 2])
x = x.at[-(halo_x + halo_x // 2):-halo_x].add(x[-halo_x // 2:])
# Apply corrections along y
x = x.at[:, halo_y:halo_y + halo_y // 2].add(x[:, :halo_y // 2])
x = x.at[:, -(halo_y + halo_y // 2):-halo_y].add(x[:, -halo_y // 2:])
unpad_slice = [slice(None)] * 3
if halo_x > 0:
unpad_slice[0] = slice(halo_x, -halo_x)
if halo_y > 0:
unpad_slice[1] = slice(halo_y, -halo_y)
return x[tuple(unpad_slice)]
def slice_pad(x, pad_width, sharding):
gpu_mesh = sharding.mesh if sharding is not None else None
if gpu_mesh is not None and not (gpu_mesh.empty) and (
pad_width[0][0] > 0 or pad_width[1][0] > 0):
assert sharding is not None
spec = sharding.spec
return shard_map((partial(jnp.pad, pad_width=pad_width)),
mesh=gpu_mesh,
in_specs=spec,
out_specs=spec)(x)
else:
return x
def slice_unpad(x, pad_width, sharding):
mesh = sharding.mesh if sharding is not None else None
if mesh is not None and not (mesh.empty) and (pad_width[0][0] > 0
or pad_width[1][0] > 0):
assert sharding is not None
spec = sharding.spec
return shard_map(partial(slice_unpad_impl, pad_width=pad_width),
mesh=mesh,
in_specs=spec,
out_specs=spec)(x)
else:
return x
def get_local_shape(mesh_shape, sharding=None):
""" Helper function to get the local size of a mesh given the global size.
"""
gpu_mesh = sharding.mesh if sharding is not None else None
if gpu_mesh is None or gpu_mesh.empty:
return mesh_shape
else:
pdims = gpu_mesh.devices.shape
return [
mesh_shape[0] // pdims[0], mesh_shape[1] // pdims[1],
*mesh_shape[2:]
]
def _axis_names(spec):
if len(spec) == 1:
x_axis, = spec
y_axis = None
single_axis = True
elif len(spec) == 2:
x_axis, y_axis = spec
if y_axis == None:
single_axis = True
elif x_axis == None:
x_axis = y_axis
single_axis = True
else:
single_axis = False
else:
raise ValueError("Only 1 or 2 axis sharding is supported")
return x_axis, y_axis, single_axis
def uniform_particles(mesh_shape, sharding=None):
gpu_mesh = sharding.mesh if sharding is not None else None
if gpu_mesh is not None and not (gpu_mesh.empty):
local_mesh_shape = get_local_shape(mesh_shape, sharding)
spec = sharding.spec
x_axis, y_axis, single_axis = _axis_names(spec)
def particles():
x_indx = lax.axis_index(x_axis)
y_indx = 0 if single_axis else lax.axis_index(y_axis)
x = jnp.arange(local_mesh_shape[0]) + x_indx * local_mesh_shape[0]
y = jnp.arange(local_mesh_shape[1]) + y_indx * local_mesh_shape[1]
z = jnp.arange(local_mesh_shape[2])
return jnp.stack(jnp.meshgrid(x, y, z, indexing='ij'), axis=-1)
return shard_map(particles, mesh=gpu_mesh, in_specs=(),
out_specs=spec)()
else:
return jnp.stack(jnp.meshgrid(*[jnp.arange(s) for s in mesh_shape],
indexing='ij'),
axis=-1)
def normal_field(mesh_shape, seed, sharding=None):
"""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):
local_mesh_shape = get_local_shape(mesh_shape, sharding)
size = jax.device_count()
# rank = jax.process_index()
# process_index is multi_host only
# to make the code work both in multi host and single controller we can do this trick
keys = jax.random.split(seed, size)
spec = sharding.spec
x_axis, y_axis, single_axis = _axis_names(spec)
def normal(keys, shape, dtype):
idx = lax.axis_index(x_axis)
if not single_axis:
y_index = lax.axis_index(y_axis)
x_size = lax.psum(1, axis_name=x_axis)
idx += y_index * x_size
return jax.random.normal(key=keys[idx], shape=shape, dtype=dtype)
return shard_map(
partial(normal, shape=local_mesh_shape, dtype='float32'),
mesh=gpu_mesh,
in_specs=P(None),
out_specs=spec)(keys) # yapf: disable
else:
return jax.random.normal(shape=mesh_shape, key=seed)