提交 a29aabed 编写于 作者: R rusty1s

rename

上级 247154f9
......@@ -12,34 +12,36 @@
torch::Tensor partition_cpu(torch::Tensor rowptr, torch::Tensor col,
torch::optional<torch::Tensor> optional_value,
torch::optional<torch::Tensor> vweights,
torch::optional<torch::Tensor> optional_node_weight,
int64_t num_parts, bool recursive) {
#ifdef WITH_METIS
CHECK_CPU(rowptr);
CHECK_CPU(col);
if (optional_value.has_value()) {
CHECK_CPU(optional_value.value());
CHECK_INPUT(optional_value.value().dim() == 1);
CHECK_INPUT(optional_value.value().numel() == col.numel());
}
if (vweights.has_value()) {
CHECK_CPU(vweights.value());
CHECK_INPUT(vweights.value().dim() == 1);
CHECK_INPUT(vweights.value().numel() == rowptr.numel() - 1);
if (optional_node_weight.has_value()) {
CHECK_CPU(optional_node_weight.value());
CHECK_INPUT(optional_node_weight.value().dim() == 1);
CHECK_INPUT(optional_node_weight.value().numel() == rowptr.numel() - 1);
}
int64_t nvtxs = rowptr.numel() - 1;
int64_t ncon = 1;
auto *xadj = rowptr.data_ptr<int64_t>();
auto *adjncy = col.data_ptr<int64_t>();
int64_t *adjwgt = NULL;
if (optional_value.has_value())
adjwgt = optional_value.value().data_ptr<int64_t>();
int64_t *vwgt = NULL;
if (vweights.has_value())
vwgt = vweights.value().data_ptr<int64_t>();
if (optional_node_weight.has_value())
vwgt = optional_node_weight.value().data_ptr<int64_t>();
int64_t objval = -1;
auto part = torch::empty(nvtxs, rowptr.options());
......@@ -62,11 +64,11 @@ torch::Tensor partition_cpu(torch::Tensor rowptr, torch::Tensor col,
// needs mt-metis installed via:
// ./configure --shared --edges64bit --vertices64bit --weights64bit
// --partitions64bit
torch::Tensor mt_partition_cpu(torch::Tensor rowptr, torch::Tensor col,
torch::optional<torch::Tensor> optional_value,
torch::optional<torch::Tensor> vweights,
int64_t num_parts, bool recursive,
int64_t num_workers) {
torch::Tensor
mt_partition_cpu(torch::Tensor rowptr, torch::Tensor col,
torch::optional<torch::Tensor> optional_value,
torch::optional<torch::Tensor> optional_node_weight,
int64_t num_parts, bool recursive, int64_t num_workers) {
#ifdef WITH_MTMETIS
CHECK_CPU(rowptr);
CHECK_CPU(col);
......@@ -76,10 +78,10 @@ torch::Tensor mt_partition_cpu(torch::Tensor rowptr, torch::Tensor col,
CHECK_INPUT(optional_value.value().numel() == col.numel());
}
if (vweights.has_value()) {
CHECK_CPU(vweights.value());
CHECK_INPUT(vweights.value().dim() == 1);
CHECK_INPUT(vweights.value().numel() == rowptr.numel() - 1);
if (optional_node_weight.has_value()) {
CHECK_CPU(optional_node_weight.value());
CHECK_INPUT(optional_node_weight.value().dim() == 1);
CHECK_INPUT(optional_node_weight.value().numel() == rowptr.numel() - 1);
}
mtmetis_vtx_type nvtxs = rowptr.numel() - 1;
......@@ -87,11 +89,14 @@ torch::Tensor mt_partition_cpu(torch::Tensor rowptr, torch::Tensor col,
mtmetis_adj_type *xadj = (mtmetis_adj_type *)rowptr.data_ptr<int64_t>();
mtmetis_vtx_type *adjncy = (mtmetis_vtx_type *)col.data_ptr<int64_t>();
mtmetis_wgt_type *adjwgt = NULL;
if (optional_value.has_value())
adjwgt = optional_value.value().data_ptr<int64_t>();
mtmetis_wgt_type *vwgt = NULL;
if (vweights.has_value())
vwgt = vweights.value().data_ptr<int64_t>();
if (optional_node_weight.has_value())
vwgt = optional_node_weight.value().data_ptr<int64_t>();
mtmetis_pid_type nparts = num_parts;
mtmetis_wgt_type objval = -1;
auto part = torch::empty(nvtxs, rowptr.options());
......
......@@ -4,11 +4,11 @@
torch::Tensor partition_cpu(torch::Tensor rowptr, torch::Tensor col,
torch::optional<torch::Tensor> optional_value,
torch::optional<torch::Tensor> vweights,
torch::optional<torch::Tensor> optional_node_weight,
int64_t num_parts, bool recursive);
torch::Tensor mt_partition_cpu(torch::Tensor rowptr, torch::Tensor col,
torch::optional<torch::Tensor> optional_value,
torch::optional<torch::Tensor> vweights,
int64_t num_parts, bool recursive,
int64_t num_workers);
torch::Tensor
mt_partition_cpu(torch::Tensor rowptr, torch::Tensor col,
torch::optional<torch::Tensor> optional_value,
torch::optional<torch::Tensor> optional_node_weight,
int64_t num_parts, bool recursive, int64_t num_workers);
......@@ -13,7 +13,7 @@ PyMODINIT_FUNC PyInit__metis_cpu(void) { return NULL; }
torch::Tensor partition(torch::Tensor rowptr, torch::Tensor col,
torch::optional<torch::Tensor> optional_value,
torch::optional<torch::Tensor> vweights,
torch::optional<torch::Tensor> optional_node_weight,
int64_t num_parts, bool recursive) {
if (rowptr.device().is_cuda()) {
#ifdef WITH_CUDA
......@@ -22,13 +22,14 @@ torch::Tensor partition(torch::Tensor rowptr, torch::Tensor col,
AT_ERROR("Not compiled with CUDA support");
#endif
} else {
return partition_cpu(rowptr, col, optional_value, vweights, num_parts, recursive);
return partition_cpu(rowptr, col, optional_value, optional_node_weight,
num_parts, recursive);
}
}
torch::Tensor mt_partition(torch::Tensor rowptr, torch::Tensor col,
torch::optional<torch::Tensor> optional_value,
torch::optional<torch::Tensor> vweights,
torch::optional<torch::Tensor> optional_node_weight,
int64_t num_parts, bool recursive,
int64_t num_workers) {
if (rowptr.device().is_cuda()) {
......@@ -38,8 +39,8 @@ torch::Tensor mt_partition(torch::Tensor rowptr, torch::Tensor col,
AT_ERROR("Not compiled with CUDA support");
#endif
} else {
return mt_partition_cpu(rowptr, col, optional_value, vweights, num_parts, recursive,
num_workers);
return mt_partition_cpu(rowptr, col, optional_value, optional_node_weight,
num_parts, recursive, num_workers);
}
}
......
import pytest
from itertools import product
import torch
from torch_sparse.tensor import SparseTensor
......@@ -12,32 +14,24 @@ except RuntimeError:
@pytest.mark.skipif(not with_metis, reason='Not compiled with METIS support')
@pytest.mark.parametrize('device', devices)
def test_metis(device):
value1 = torch.randn(6 * 6, device=device).view(6, 6)
value2 = torch.arange(6 * 6, dtype=torch.long, device=device).view(6, 6)
value3 = torch.ones(6 * 6, device=device).view(6, 6)
vwgts = torch.rand(6, device=device)
for value in [value1, value2, value3]:
for vwgt in [None, vwgts]:
mat = SparseTensor.from_dense(value)
_, partptr, perm = mat.partition(num_parts=2, recursive=False,
vweights=vwgt,
weighted=True)
assert partptr.numel() == 3
assert perm.numel() == 6
_, partptr, perm = mat.partition(num_parts=2, recursive=False,
vweights=vwgt,
weighted=False)
assert partptr.numel() == 3
assert perm.numel() == 6
_, partptr, perm = mat.partition(num_parts=1, recursive=False,
vweights=vwgt,
weighted=True)
assert partptr.numel() == 2
assert perm.numel() == 6
@pytest.mark.parametrize('device,weighted', product(devices, [False, True]))
def test_metis(device, weighted):
mat1 = torch.randn(6 * 6, device=device).view(6, 6)
mat2 = torch.arange(6 * 6, dtype=torch.long, device=device).view(6, 6)
mat3 = torch.ones(6 * 6, device=device).view(6, 6)
vec1 = None
vec2 = torch.rand(6, device=device)
for mat, vec in product([mat1, mat2, mat3], [vec1, vec2]):
mat = SparseTensor.from_dense(mat)
_, partptr, perm = mat.partition(num_parts=1, recursive=False,
weighted=weighted, node_weight=vec)
assert partptr.numel() == 2
assert perm.numel() == 6
_, partptr, perm = mat.partition(num_parts=2, recursive=False,
weighted=weighted, node_weight=vec)
assert partptr.numel() == 3
assert perm.numel() == 6
......@@ -20,9 +20,8 @@ def weight2metis(weight: torch.Tensor) -> Optional[torch.Tensor]:
def partition(
src: SparseTensor, num_parts: int, recursive: bool = False,
vweights: torch.tensor = None,
weighted: bool = False
src: SparseTensor, num_parts: int, recursive: bool = False,
weighted: bool = False, node_weight: torch.tensor = None
) -> Tuple[SparseTensor, torch.Tensor, torch.Tensor]:
assert num_parts >= 1
......@@ -42,14 +41,14 @@ def partition(
else:
value = None
if vweights is not None:
assert vweights.numel() == rowptr.numel() - 1
vweights = vweights.view(-1).detach().cpu()
if vweights.is_floating_point():
vweights = weight2metis(vweights)
if node_weight is not None:
assert node_weight.numel() == rowptr.numel() - 1
node_weight = node_weight.view(-1).detach().cpu()
if node_weight.is_floating_point():
node_weight = weight2metis(node_weight)
cluster = torch.ops.torch_sparse.partition(rowptr, col, value, vweights, num_parts,
recursive)
cluster = torch.ops.torch_sparse.partition(rowptr, col, value, node_weight,
num_parts, recursive)
cluster = cluster.to(src.device())
cluster, perm = cluster.sort()
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册