Skip to content

fix compatibility with dgl and cuda 12.0+ #102

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
chamikasudusinghe wants to merge 1 commit into
base: main
Choose a base branch
from
Open

fix compatibility with dgl and cuda 12.0+ #102

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion examples/rgms/rgcn/bench_rgcn_baseline.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -46,7 +46,7 @@ def test_rgcn_baseline(g: DGLHeteroGraph, feat_size: int):
feat = th.rand(g.num_src_nodes(), feat_size).to(0) / 100
out = th.zeros(g.num_dst_nodes(), feat_size).to(0) / 100
weight = th.rand(g.num_rels, feat_size, feat_size).to(0)
indptr, indices, eid = g.adj_sparse(fmt="csc")
indptr, indices, eid = g.adj_tensors(fmt="csc")
etype = g.edata[dgl.ETYPE][eid.long()]

# dgl-lowmem
Expand Down
2 changes: 1 addition & 1 deletion examples/rgms/rgcn/bench_rgcn_composable.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -109,7 +109,7 @@ def test_rgcn_composable_format(
dst_type_id = g.get_ntype_id(dst_type)
g_sub = g[etype]
m_sub, n_sub = g_sub.num_dst_nodes(), g_sub.num_src_nodes()
indptr, indices, _ = g_sub.adj_sparse(fmt="csc")
indptr, indices, _ = g_sub.adj_tensors(fmt="csc")
csf_indptr_0.append(csf_indptr_0[-1] + m_sub)
csf_indices_0.append(ntype_node_pointer[dst_type_id] + th.arange(m_sub, dtype=th.int32))
csf_indptr_1.append(csf_indptr_1[-1][-1] + indptr[1:])
Expand Down
2 changes: 1 addition & 1 deletion examples/rgms/rgcn/bench_rgcn_non_composable.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -60,7 +60,7 @@ def test_lower_rgcn_hetero(
src_type_id = g.get_ntype_id(src_type)
dst_type_id = g.get_ntype_id(dst_type)
g_sub = g[etype]
indptr, indices, _ = g_sub.adj_sparse(fmt="csc")
indptr, indices, _ = g_sub.adj_tensors(fmt="csc")

unique_nodes = th.nonzero(indptr[:-1] != indptr[1:]).squeeze(1)
indptr_i.append(th.LongTensor([len(unique_nodes) + indptr_i[-1].item()]))
Expand Down
2 changes: 1 addition & 1 deletion examples/rgms/rgcn/bench_rgcn_tensorcore.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -680,7 +680,7 @@ def rgcn_tensorcore(
dst_type_id = g.get_ntype_id(dst_type)
g_sub = g[etype]
m_sub, n_sub = g_sub.num_dst_nodes(), g_sub.num_src_nodes()
indptr, indices, _ = g_sub.adj_sparse(fmt="csc")
indptr, indices, _ = g_sub.adj_tensors(fmt="csc")
csf_indptr_0.append(csf_indptr_0[-1] + m_sub)
csf_indices_0.append(ntype_node_pointer[dst_type_id] + th.arange(m_sub, dtype=th.int32))
csf_indptr_1.append(csf_indptr_1[-1][-1] + indptr[1:])
Expand Down
2 changes: 1 addition & 1 deletion examples/rgms/sparse_conv/rgms.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -682,7 +682,7 @@ def rgms_tensorcore(
etype_id = g.get_etype_id(str(etype))
g_sub = g[str(etype)]
m_sub, n_sub = g_sub.num_dst_nodes(), g_sub.num_src_nodes()
indptr, indices, _ = g_sub.adj_sparse(fmt="csc")
indptr, indices, _ = g_sub.adj_tensors(fmt="csc")
csf_indptr_0.append(csf_indptr_0[-1] + m_sub)
csf_indices_0.append(th.arange(m_sub, dtype=th.int32))
csf_indptr_1.append(csf_indptr_1[-1][-1] + indptr[1:])
Expand Down
2 changes: 1 addition & 1 deletion examples/sddmm/bench_sddmm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -57,7 +57,7 @@ def func(

def bench_sddmm(g: dgl.DGLGraph, feat_size: int):
global sddmm
indptr, indices, _ = g.adj_sparse("csr")
indptr, indices, _ = g.adj_tensors("csr")
m = g.num_src_nodes()
n = g.num_dst_nodes()
nnz = g.number_of_edges()
Expand Down
2 changes: 1 addition & 1 deletion examples/spmm/bench_spmm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -100,7 +100,7 @@ def bench_hyb(
):
num_buckets = len(bucket_sizes)
coersening_factor = min(coersening_factor, feat_size // 32)
indptr, indices, _ = g.adj_sparse("csc")
indptr, indices, _ = g.adj_tensors("csc")
m = g.num_dst_nodes()
n = g.num_src_nodes()
nnz = g.num_edges()
Expand Down
2 changes: 1 addition & 1 deletion examples/spmm/bench_spmm_naive.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -79,7 +79,7 @@ def bench_hyb(
feat_size=128,
cwm=2,
):
indptr, indices, _ = g.adj_sparse("csc")
indptr, indices, _ = g.adj_tensors("csc")
m = g.num_dst_nodes()
n = g.num_src_nodes()
nnz = g.num_edges()
Expand Down
2 changes: 1 addition & 1 deletion examples/spmm/bench_tc_spmm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -578,7 +578,7 @@ def parse_mma_shape(mma_shape_str: str):

def bench_tc_spmm(g: dgl.DGLHeteroGraph, x: th.Tensor, y_golden: th.Tensor, mma_shape_str: str):
mma_m, mma_n, mma_k = parse_mma_shape(mma_shape_str)
indptr, indices, _ = g.adj_sparse("csc")
indptr, indices, _ = g.adj_tensors("csc")
indptr_nd = tvm.nd.array(indptr.numpy().astype("int32"), device=tvm.cpu())
indices_nd = tvm.nd.array(indices.numpy().astype("int32"), device=tvm.cpu())
tile_size = mma_m
Expand Down
6 changes: 6 additions & 0 deletions python/tvm/contrib/nvcc.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -72,6 +72,12 @@ def compile_cuda(code, target_format="ptx", arch=None, options=None, path_target
temp_code = temp.relpath("my_kernel.cu")
temp_target = temp.relpath("my_kernel.%s" % target_format)

# CUDA 12.0+ defines htanh/htan/hatan/herf/hpow in cuda_fp16.hpp;
# strip TVM's duplicate definitions to avoid redefinition errors.
import re as _re
_half_fns = r"CUDA_UNSUPPORTED_HALF_MATH_(?:UNARY|BINARY)\(h(?:tanh|tan|atan|erf|pow)[^\n]*\n"
code = _re.sub(_half_fns, "", code)

with open(temp_code, "w") as out_file:
out_file.write(code)

Expand Down
6 changes: 3 additions & 3 deletions tests/python/sparsetir/test_format_conversion_routine.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -105,7 +105,7 @@ def test_column_part_hyb():
g = dgl.rand_graph(1000, 10000).int()
column_parts = 4
buckets = [1, 2, 4]
indptr, indices, _ = g.adj_sparse("csc")
indptr, indices, _ = g.adj_tensors("csc")
indptr_nd = tvm.nd.array(indptr.numpy(), device=tvm.cpu())
indices_nd = tvm.nd.array(indices.numpy(), device=tvm.cpu())
# built-in c++ funcion
Expand Down Expand Up @@ -146,7 +146,7 @@ def condense_py(indptr, indices, block_size):
def test_condense():
g = dgl.rand_graph(1000, 10000).int()
t = 4
indptr, indices, _ = g.adj_sparse("csc")
indptr, indices, _ = g.adj_tensors("csc")
indptr = indptr.numpy()
indices = indices.numpy()
indptr_nd = tvm.nd.array(indptr, device=tvm.cpu())
Expand Down Expand Up @@ -197,7 +197,7 @@ def test_hetero_csr_to_ell3d():
g_sub = g[etype]
# print(g_sub)
m, n = g_sub.num_dst_nodes(), g_sub.num_src_nodes()
indptr, indices, _ = g_sub.adj_sparse(fmt="csc")
indptr, indices, _ = g_sub.adj_tensors(fmt="csc")
# print(indptr, indices)
csf_indptr_0.append(csf_indptr_0[-1] + m)
csf_indices_0.append(ntype_node_pointer[src_type_id] + th.arange(m, dtype=th.int32))
Expand Down