# coding=utf-8
# SPDX-FileCopyrightText: Copyright (c) 2022 The torch-harmonics Authors. All rights reserved.
# SPDX-License-Identifier: BSD-3-Clause
#
# Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions are met:
#
# 1. Redistributions of source code must retain the above copyright notice, this
# list of conditions and the following disclaimer.
#
# 2. Redistributions in binary form must reproduce the above copyright notice,
# this list of conditions and the following disclaimer in the documentation
# and/or other materials provided with the distribution.
#
# 3. Neither the name of the copyright holder nor the names of its
# contributors may be used to endorse or promote products derived from
# this software without specific prior written permission.
#
# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
# DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
# FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
# DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
# SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
# CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
# OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#
from itertools import accumulate
from typing import Optional, Tuple, Union
import torch
from disco_helpers import optimized_kernels_is_available, pack_psi_dense, preprocess_psi
from torch_harmonics.disco._disco_utils import _get_psi
from torch_harmonics.disco.convolution import (
DiscreteContinuousConv,
_kpacked_device_supported_for_tensor,
_precompute_convolution_tensor_s2,
)
from torch_harmonics.disco.kernels_torch.disco_torch import _disco_s2_transpose_contraction_torch
from torch_harmonics.disco.optimized.disco_optimized import _build_kernel_split_csr, _disco_s2_transpose_contraction_optimized, _maybe_kpack_psi, _split_csr_python_offsets
# a2a forward orchestration: standard (fused=False) and reordered (fused=True).
from .kernels import (
_distributed_disco_fwd_a2a,
_distributed_disco_fwd_a2a_reordered,
)
# distributed stuff — relative imports to avoid the circular dependency
# through torch_harmonics.distributed.__init__.
from .primitives import (
compute_split_shapes,
distributed_transpose_azimuth,
gather_from_copy_to_polar_region,
)
from .utils import (
azimuth_group_rank,
azimuth_group_size,
polar_group_rank,
polar_group_size,
)
def _split_distributed_convolution_tensor_s2(
idx: torch.Tensor,
vals: torch.Tensor,
in_shape: Tuple[int],
out_shape: Tuple[int],
):
"""
Splits a pre-computed convolution tensor along the latitude dimension for distributed processing.
This function takes a convolution tensor that was generated by the serial routine and filters
it to only include entries corresponding to the local latitude slice assigned to this process.
The filtering is done based on the polar group rank and the computed split shapes.
Parameters
----------
idx : torch.Tensor
Indices of the pre-computed convolution tensor
vals : torch.Tensor
Values of the pre-computed convolution tensor
in_shape : Tuple[int]
Shape of the input tensor (nlat_in, nlon_in)
out_shape : Tuple[int]
Shape of the output tensor (nlat_out, nlon_out)
Returns
-------
idx : torch.Tensor
Filtered indices corresponding to the local latitude slice
vals : torch.Tensor
Filtered values corresponding to the local latitude slice
"""
nlat_in, nlon_in = in_shape
nlat_out, nlon_out = out_shape
comm_size_polar = polar_group_size()
comm_rank_polar = polar_group_rank()
split_shapes = compute_split_shapes(nlat_in, num_chunks=comm_size_polar)
offsets = [0] + list(accumulate(split_shapes))
start_idx = offsets[comm_rank_polar]
end_idx = offsets[comm_rank_polar + 1]
# once normalization is done we can throw away the entries which correspond to input latitudes we do not care about
lats = idx[2] // nlon_in
lons = idx[2] % nlon_in
ilats = torch.argwhere((lats < end_idx) & (lats >= start_idx)).squeeze()
vals = vals[ilats]
# for the indices we need to recompute them to refer to local indices of the input tenor
idx = torch.stack([idx[0, ilats], idx[1, ilats], (lats[ilats] - start_idx) * nlon_in + lons[ilats]], dim=0)
# make results contiguous
idx = idx.contiguous()
vals = vals.contiguous()
return idx, vals
[docs]
class DistributedDiscreteContinuousConvS2(DiscreteContinuousConv):
"""
Distributed version of Discrete-continuous convolutions (DISCO) on the 2-Sphere as described in :cite:`Ocampo2023`.
We assume the data can be split in polar and azimuthal directions.
.. seealso::
:class:`torch_harmonics.DiscreteContinuousConvS2`
Serial counterpart with full mathematical description and parameter
documentation.
The algorithm is all-to-all (azimuth <-> channel swap so the sparse psi
contraction runs against the full nlon_in row, polar reduce_scatter
completes the H sum, then back to channel-distributed). The ``fused=``
flag mirrors the serial conv:
``fused=False`` (default) — standard a2a: einsum after the
transpose-back; the K-expanded intermediate is saved for backward.
``fused=True`` — reordered a2a: the weight einsum runs before the
collectives on the local azimuth channel shard, via the fused
contraction+einsum op that recomputes the K-expanded in backward
instead of saving it. K× lower activation memory and K× less
collective volume, at the cost of one extra contraction in
backward. CUDA + optimized kernels only.
Parameters
----------
in_channels : int
Number of input channels
out_channels : int
Number of output channels
in_shape : Tuple[int]
Shape of the input tensor
out_shape : Tuple[int]
Shape of the output tensor
kernel_shape : Union[int, Tuple[int], Tuple[int, int]]
Shape of the kernel
basis_type : Optional[str]
Type of basis to use
basis_norm_mode : Optional[str]
Normalization mode for the filter basis
groups : Optional[int]
Number of groups
grid_in : Optional[str]
Grid type for the input tensor
grid_out : Optional[str]
Grid type for the output tensor
bias : Optional[bool]
Whether to use bias
theta_cutoff : Optional[float]
Theta cutoff for the filter basis
optimized_kernel : Optional[bool]
Use the optimized CUDA contraction kernel. Required when ``fused=True``.
fused : bool
Mirrors the serial conv. ``False`` (default): standard all-to-all
(the K-expanded intermediate is saved for backward). ``True``:
reordered all-to-all — the weight einsum runs before the collectives
on the local azimuth channel shard and the K-expanded is recomputed
in backward instead of saved, for K× lower activation memory and K×
less collective volume (CUDA + optimized kernels only).
Returns
-------
torch.Tensor
Output tensor
References
----------
:cite:`Ocampo2023`
"""
def __init__(
self,
in_channels: int,
out_channels: int,
in_shape: Tuple[int],
out_shape: Tuple[int],
kernel_shape: Union[int, Tuple[int], Tuple[int, int]],
basis_type: Optional[str] = "piecewise linear",
basis_norm_mode: Optional[str] = "nodal",
groups: Optional[int] = 1,
grid_in: Optional[str] = "equiangular",
grid_out: Optional[str] = "equiangular",
bias: Optional[bool] = True,
theta_cutoff: Optional[float] = None,
optimized_kernel: Optional[bool] = True,
fused: bool = False,
):
super().__init__(in_channels, out_channels, kernel_shape, basis_type, groups, bias, optimized_kernel)
# fused=True uses the reordered a2a (fused contraction+einsum op with
# K-expanded recompute in backward); it is CUDA + optimized-kernel only.
self.fused = bool(fused)
if self.fused and not (torch.cuda.is_available() and optimized_kernels_is_available() and optimized_kernel):
raise NotImplementedError(
"DistributedDiscreteContinuousConvS2(fused=True) requires CUDA and the "
"optimized DISCO kernels (it uses the fused contraction+einsum op with "
"K-expanded recompute in backward). Use fused=False otherwise."
)
self.nlat_in, self.nlon_in = in_shape
self.nlat_out, self.nlon_out = out_shape
# get the comms grid:
self.comm_size_polar = polar_group_size()
self.comm_rank_polar = polar_group_rank()
self.comm_size_azimuth = azimuth_group_size()
self.comm_rank_azimuth = azimuth_group_rank()
# we need those shapes:
self.lat_in_shapes = compute_split_shapes(self.nlat_in, self.comm_size_polar)
self.lon_in_shapes = compute_split_shapes(self.nlon_in, self.comm_size_azimuth)
self.lat_out_shapes = compute_split_shapes(self.nlat_out, self.comm_size_polar)
self.lon_out_shapes = compute_split_shapes(self.nlon_out, self.comm_size_azimuth)
# compute theta cutoff based on the bandlimit of the input field
if theta_cutoff is None:
self.theta_cutoff = torch.pi / float(self.nlat_out - 1)
else:
self.theta_cutoff = theta_cutoff
if self.theta_cutoff <= 0.0:
raise ValueError("Error, theta_cutoff has to be positive.")
# Note that the psi matrix is of shape nlat_out x nlat_in * nlon_in.
# Since the contraction in nlon direction is a convolution, we keep
# it local to all nodes and split along nlat. We further split the
# input dim because this reduces the number of atomic reduction
# calls inside the actual kernel.
# set local shapes according to distributed mode
self.nlat_in_local = self.lat_in_shapes[self.comm_rank_polar]
self.nlat_out_local = self.nlat_out
self.nlon_in_local = self.lon_in_shapes[self.comm_rank_azimuth]
self.nlon_out_local = self.lon_out_shapes[self.comm_rank_azimuth]
self.kpacked_device_supported = False
# compute global convolution tensor
idx, vals, _ = _precompute_convolution_tensor_s2(
in_shape,
out_shape,
self.filter_basis,
grid_in=grid_in,
grid_out=grid_out,
theta_cutoff=self.theta_cutoff,
transpose_normalization=False,
basis_norm_mode=basis_norm_mode,
merge_quadrature=True,
)
idx, vals = _split_distributed_convolution_tensor_s2(idx, vals, in_shape, out_shape)
self._build_local_psi(idx, vals)
def _build_local_psi(self, idx: torch.Tensor, vals: torch.Tensor):
"""Register psi buffers for the a2a path. The a2a swap makes W
local before the kernel reads col_idx, so no wi pre-shift is
applied here."""
ker_idx = idx[0, ...].contiguous()
row_idx = idx[1, ...].contiguous()
col_idx = idx[2, ...].contiguous()
vals = vals.contiguous()
self.psi_kpacked_K_pad = None
if self.optimized_kernel:
roff_idx = preprocess_psi(
self.kernel_size,
self.nlat_out_local,
ker_idx,
row_idx,
col_idx,
vals,
).contiguous()
self.register_buffer("psi_roff_idx", roff_idx, persistent=False)
split_roff_idx, split_nnz_off, split_ker_idx, split_row_idx, split_col_idx, split_vals = _build_kernel_split_csr(
roff_idx, ker_idx, row_idx, col_idx, vals, self.kernel_size, self.nlat_out_local
)
self.psi_split_row_offsets, self.psi_split_nnz_offsets = _split_csr_python_offsets(split_nnz_off)
self.register_buffer("psi_split_roff_idx", split_roff_idx, persistent=False)
self.register_buffer("psi_split_nnz_off", split_nnz_off, persistent=False)
self.register_buffer("psi_split_ker_idx", split_ker_idx, persistent=False)
self.register_buffer("psi_split_row_idx", split_row_idx, persistent=False)
self.register_buffer("psi_split_col_idx", split_col_idx, persistent=False)
self.register_buffer("psi_split_vals", split_vals, persistent=False)
# optional K-packed dense layout for the WGMMA path (Hopper bf16/fp16).
# A2A makes W local before the kernel, so wi_shift=0 like the serial path.
psi_packed_idx, psi_packed_vals, psi_packed_count = pack_psi_dense(self.kernel_size, self.nlat_out_local, self.nlon_in, 0, ker_idx, row_idx, col_idx, vals, roff_idx)
kpack = _maybe_kpack_psi(psi_packed_idx.contiguous(), psi_packed_vals.contiguous(), psi_packed_count.contiguous())
if kpack is not None:
kpacked_idx, kpacked_vals, kpacked_count, K_pad = kpack
self.register_buffer("psi_kpacked_idx", kpacked_idx, persistent=False)
self.register_buffer("psi_kpacked_vals", kpacked_vals, persistent=False)
self.register_buffer("psi_kpacked_count", kpacked_count, persistent=False)
self.psi_kpacked_K_pad = K_pad
# optional K-packed dense layout for the WGMMA path (Hopper bf16/fp16).
# A2A makes W local before the kernel, so wi_shift=0 like the serial path.
psi_packed_idx, psi_packed_vals, psi_packed_count = pack_psi_dense(self.kernel_size, self.nlat_out_local, self.nlon_in, 0, ker_idx, row_idx, col_idx, vals, roff_idx)
kpack = _maybe_kpack_psi(psi_packed_idx.contiguous(), psi_packed_vals.contiguous(), psi_packed_count.contiguous())
if kpack is not None:
kpacked_idx, kpacked_vals, kpacked_count, K_pad = kpack
self.register_buffer("psi_kpacked_idx", kpacked_idx, persistent=False)
self.register_buffer("psi_kpacked_vals", kpacked_vals, persistent=False)
self.register_buffer("psi_kpacked_count", kpacked_count, persistent=False)
self.psi_kpacked_K_pad = K_pad
self.register_buffer("psi_ker_idx", ker_idx, persistent=False)
self.register_buffer("psi_row_idx", row_idx, persistent=False)
self.register_buffer("psi_col_idx", col_idx, persistent=False)
self.register_buffer("psi_vals", vals, persistent=False)
if not self.optimized_kernel:
self.psi = _get_psi(
self.kernel_size,
self.psi_idx,
self.psi_vals,
self.nlat_in,
self.nlon_in,
self.nlat_out,
self.nlon_out,
self.nlat_in_local,
self.nlat_out_local,
)
def extra_repr(self):
return (
f"in_shape={(self.nlat_in, self.nlon_in)}, "
f"out_shape={(self.nlat_out, self.nlon_out)}, "
f"in_chans={self.groupsize * self.groups}, "
f"out_chans={self.weight.shape[0]}, "
f"filter_basis={self.filter_basis}, "
f"kernel_shape={self.kernel_shape}, "
f"theta_cutoff={self.theta_cutoff}, "
f"groups={self.groups}, fused={self.fused}"
)
@property
def psi_idx(self):
return torch.stack([self.psi_ker_idx, self.psi_row_idx, self.psi_col_idx], dim=0).contiguous()
def _refresh_kpacked_device_supported(self):
if not hasattr(self, "psi_vals"):
self.kpacked_device_supported = False
return
self.kpacked_device_supported = _kpacked_device_supported_for_tensor(self.psi_vals)
def _apply(self, fn):
result = super()._apply(fn)
self._refresh_kpacked_device_supported()
return result
def forward(self, x: torch.Tensor) -> torch.Tensor:
if self.fused:
# reordered a2a: einsum-first on the local channel shard via the
# fused conv op (K-expanded recomputed in backward, not saved).
out = _distributed_disco_fwd_a2a_reordered(
x,
self.weight,
psi_roff_idx=self.psi_roff_idx,
psi_ker_idx=self.psi_ker_idx,
psi_row_idx=self.psi_row_idx,
psi_col_idx=self.psi_col_idx,
psi_vals=self.psi_vals,
psi_split_roff_idx=self.psi_split_roff_idx,
psi_split_nnz_off=self.psi_split_nnz_off,
psi_split_ker_idx=self.psi_split_ker_idx,
psi_split_row_idx=self.psi_split_row_idx,
psi_split_col_idx=self.psi_split_col_idx,
psi_split_vals=self.psi_split_vals,
psi_split_row_offsets=self.psi_split_row_offsets,
psi_split_nnz_offsets=self.psi_split_nnz_offsets,
psi_kpacked_idx=getattr(self, "psi_kpacked_idx", None),
psi_kpacked_vals=getattr(self, "psi_kpacked_vals", None),
psi_kpacked_count=getattr(self, "psi_kpacked_count", None),
psi_kpacked_K_pad=self.psi_kpacked_K_pad,
kpacked_device_supported=self.kpacked_device_supported,
kernel_size=self.kernel_size,
nlat_out_local=self.nlat_out_local,
nlon_out=self.nlon_out,
groups=self.groups,
groupsize=self.groupsize,
comm_size_azimuth=self.comm_size_azimuth,
comm_rank_azimuth=self.comm_rank_azimuth,
lon_in_shapes=self.lon_in_shapes,
)
else:
# standard a2a: contraction then einsum after the transpose-back;
# the K-expanded intermediate is saved for backward.
out = _distributed_disco_fwd_a2a(
x,
self.weight,
psi_roff_idx=getattr(self, "psi_roff_idx", None),
psi_ker_idx=self.psi_ker_idx,
psi_row_idx=self.psi_row_idx,
psi_col_idx=self.psi_col_idx,
psi_vals=self.psi_vals,
psi_kpacked_idx=getattr(self, "psi_kpacked_idx", None),
psi_kpacked_vals=getattr(self, "psi_kpacked_vals", None),
psi_kpacked_count=getattr(self, "psi_kpacked_count", None),
psi_kpacked_K_pad=self.psi_kpacked_K_pad,
kpacked_device_supported=self.kpacked_device_supported,
psi_torch=getattr(self, "psi", None),
optimized_kernel=self.optimized_kernel,
kernel_size=self.kernel_size,
nlat_out_local=self.nlat_out_local,
nlon_out=self.nlon_out,
groups=self.groups,
groupsize=self.groupsize,
comm_size_azimuth=self.comm_size_azimuth,
lon_in_shapes=self.lon_in_shapes,
)
if self.bias is not None:
out = out + self.bias.reshape(1, self.bias.shape[0], 1, 1)
return out
[docs]
class DistributedDiscreteContinuousConvTransposeS2(DiscreteContinuousConv):
"""
Distributed version of discrete-continuous transpose convolutions (DISCO) on the 2-Sphere as described in :cite:`Ocampo2023`.
.. seealso::
:class:`torch_harmonics.DiscreteContinuousConvTransposeS2`
Serial counterpart with full mathematical description and parameter
documentation.
Parameters
----------
in_channels : int
Number of input channels
out_channels : int
Number of output channels
in_shape : Tuple[int]
Shape of the input tensor
out_shape : Tuple[int]
Shape of the output tensor
kernel_shape : Union[int, Tuple[int], Tuple[int, int]]
Shape of the kernel
basis_type : Optional[str]
Type of basis to use
basis_norm_mode : Optional[str]
Normalization mode for the filter basis
groups : Optional[int]
Number of groups
grid_in : Optional[str]
Grid type for the input tensor
grid_out : Optional[str]
Grid type for the output tensor
bias : Optional[bool]
Whether to use bias
theta_cutoff : Optional[float]
Theta cutoff for the filter basis
Returns
-------
torch.Tensor
Output tensor
References
----------
:cite:`Ocampo2023`
"""
def __init__(
self,
in_channels: int,
out_channels: int,
in_shape: Tuple[int],
out_shape: Tuple[int],
kernel_shape: Union[int, Tuple[int], Tuple[int, int]],
basis_type: Optional[str] = "piecewise linear",
basis_norm_mode: Optional[str] = "nodal",
groups: Optional[int] = 1,
grid_in: Optional[str] = "equiangular",
grid_out: Optional[str] = "equiangular",
bias: Optional[bool] = True,
theta_cutoff: Optional[float] = None,
optimized_kernel: Optional[bool] = True,
):
super().__init__(in_channels, out_channels, kernel_shape, basis_type, groups, bias, optimized_kernel)
self.nlat_in, self.nlon_in = in_shape
self.nlat_out, self.nlon_out = out_shape
# get the comms grid:
self.comm_size_polar = polar_group_size()
self.comm_rank_polar = polar_group_rank()
self.comm_size_azimuth = azimuth_group_size()
self.comm_rank_azimuth = azimuth_group_rank()
# we need those shapes:
self.lat_in_shapes = compute_split_shapes(self.nlat_in, self.comm_size_polar)
self.lon_in_shapes = compute_split_shapes(self.nlon_in, self.comm_size_azimuth)
self.lat_out_shapes = compute_split_shapes(self.nlat_out, self.comm_size_polar)
self.lon_out_shapes = compute_split_shapes(self.nlon_out, self.comm_size_azimuth)
# bandlimit
if theta_cutoff is None:
self.theta_cutoff = torch.pi / float(self.nlat_in - 1)
else:
self.theta_cutoff = theta_cutoff
if self.theta_cutoff <= 0.0:
raise ValueError("Error, theta_cutoff has to be positive.")
# Note that the psi matrix is of shape nlat_out x nlat_in * nlon_in. Since the contraction in nlon direction is a convolution,
# we will keep local to all nodes and split the computation up along nlat. We further split the input dim because this reduces the number
# of atomic reduction calls inside the actual kernel
# set local shapes according to distributed mode:
self.nlat_in_local = self.nlat_in
self.nlat_out_local = self.lat_out_shapes[self.comm_rank_polar]
# compute global convolution tensor
# switch in_shape and out_shape since we want transpose conv
# distributed mode here is swapped because of the transpose
idx, vals, _ = _precompute_convolution_tensor_s2(
out_shape,
in_shape,
self.filter_basis,
grid_in=grid_out,
grid_out=grid_in,
theta_cutoff=self.theta_cutoff,
transpose_normalization=True,
basis_norm_mode=basis_norm_mode,
merge_quadrature=True,
)
# split the convolution tensor along latitude, again, we need to swap the meaning
# of in_shape and out_shape
idx, vals = _split_distributed_convolution_tensor_s2(idx, vals, out_shape, in_shape)
# sort the values
ker_idx = idx[0, ...].contiguous()
row_idx = idx[1, ...].contiguous()
col_idx = idx[2, ...].contiguous()
vals = vals.contiguous()
if self.optimized_kernel:
# preprocessed data-structure for GPU kernel
roff_idx = preprocess_psi(self.kernel_size, self.nlat_in_local, ker_idx, row_idx, col_idx, vals).contiguous()
self.register_buffer("psi_roff_idx", roff_idx, persistent=False)
# save all datastructures
self.register_buffer("psi_ker_idx", ker_idx, persistent=False)
self.register_buffer("psi_row_idx", row_idx, persistent=False)
self.register_buffer("psi_col_idx", col_idx, persistent=False)
self.register_buffer("psi_vals", vals, persistent=False)
# store psi as COO
if not self.optimized_kernel:
self.psi_st = _get_psi(
self.kernel_size,
self.psi_idx,
self.psi_vals,
self.nlat_in,
self.nlon_in,
self.nlat_out,
self.nlon_out,
self.nlat_in_local,
self.nlat_out_local,
semi_transposed=True,
)
def extra_repr(self):
return f"in_shape={(self.nlat_in, self.nlon_in)}, out_shape={(self.nlat_out, self.nlon_out)}, in_chans={self.groupsize * self.groups}, out_chans={self.weight.shape[0]}, filter_basis={self.filter_basis}, kernel_shape={self.kernel_shape}, theta_cutoff={self.theta_cutoff}, groups={self.groups}"
@property
def psi_idx(self):
return torch.stack([self.psi_ker_idx, self.psi_row_idx, self.psi_col_idx], dim=0).contiguous()
def forward(self, x: torch.Tensor) -> torch.Tensor:
# extract shape
B, C, H, W = x.shape
x = x.reshape(B, self.groups, self.groupsize, H, W)
# do weight multiplication
x = torch.einsum("bgcxy,gock->bgokxy", x, self.weight.reshape(self.groups, self.out_per_group, self.weight.shape[1], self.weight.shape[2])).contiguous()
x = x.reshape(B, self.weight.shape[0], x.shape[-3], H, W)
num_chans = x.shape[1]
# transpose such that lon is local, channels are split
if self.comm_size_azimuth > 1:
x = distributed_transpose_azimuth(x, (1, -1), self.lon_in_shapes)
# Fused gather + copy on the polar group. Forward is the same
# all_gather along nlat that gather_from_polar_region performed;
# backward replaces all_reduce + slice with a single reduce_scatter
# (half the backward comm volume on the K-expanded tensor).
x = gather_from_copy_to_polar_region(x, -2, self.lat_in_shapes)
if self.optimized_kernel:
out = _disco_s2_transpose_contraction_optimized(
x, self.psi_roff_idx, self.psi_ker_idx, self.psi_row_idx, self.psi_col_idx, self.psi_vals, self.kernel_size, self.nlat_out_local, self.nlon_out
)
else:
out = _disco_s2_transpose_contraction_torch(x, self.psi_st.to(x.device), self.nlon_out)
# now we can transpose back the result, so that lon is split and channels are local
if self.comm_size_azimuth > 1:
chan_shapes = compute_split_shapes(num_chans, self.comm_size_azimuth)
out = distributed_transpose_azimuth(out, (-1, 1), chan_shapes)
if self.bias is not None:
out = out + self.bias.reshape(1, self.bias.shape[0], 1, 1)
return out
