spspmm half and version up

CMakeLists.txt

 cmake_minimum_required(VERSION 3.0)
 project(torchsparse)
 set(CMAKE_CXX_STANDARD 14)
-set(TORCHSPARSE_VERSION 0.6.10)
+set(TORCHSPARSE_VERSION 0.6.11)
 
 option(WITH_CUDA "Enable CUDA support" OFF)
 

conda/pytorch-sparse/meta.yaml

 package:
   name: pytorch-sparse
-  version: 0.6.10
+  version: 0.6.11
 
 source:
   path: ../..

setup.py

@@ -102,7 +102,7 @@ tests_require = ['pytest', 'pytest-runner', 'pytest-cov']
 
 setup(
     name='torch_sparse',
-    version='0.6.10',
+    version='0.6.11',
     author='Matthias Fey',
     author_email='matthias.fey@tu-dortmund.de',
     url='https://github.com/rusty1s/pytorch_sparse',

test/test_matmul.py

@@ -47,9 +47,6 @@ def test_spmm(dtype, device, reduce):
 
 @pytest.mark.parametrize('dtype,device', product(grad_dtypes, devices))
 def test_spspmm(dtype, device):
-    if dtype == torch.half:
-        return  # TODO
-
     src = torch.tensor([[1, 0, 0], [0, 1, 0], [0, 0, 1]], dtype=dtype,
                        device=device)
 

test/test_spspmm.py

@@ -9,9 +9,6 @@ from .utils import grad_dtypes, devices, tensor
 
 @pytest.mark.parametrize('dtype,device', product(grad_dtypes, devices))
 def test_spspmm(dtype, device):
-    if dtype == torch.half:
-        return  # TODO
-
     indexA = torch.tensor([[0, 0, 1, 2, 2], [1, 2, 0, 0, 1]], device=device)
     valueA = tensor([1, 2, 3, 4, 5], dtype, device)
     indexB = torch.tensor([[0, 2], [1, 0]], device=device)
@@ -24,9 +21,6 @@ def test_spspmm(dtype, device):
 
 @pytest.mark.parametrize('dtype,device', product(grad_dtypes, devices))
 def test_sparse_tensor_spspmm(dtype, device):
-    if dtype == torch.half:
-        return  # TODO
-
     x = SparseTensor(
         row=torch.tensor(
             [0, 1, 1, 1, 2, 3, 4, 5, 5, 6, 6, 7, 7, 7, 8, 8, 9, 9],
@@ -44,8 +38,8 @@ def test_sparse_tensor_spspmm(dtype, device):
     expected = torch.eye(10, dtype=dtype, device=device)
 
     out = x @ x.to_dense().t()
-    assert torch.allclose(out, expected, atol=1e-7)
+    assert torch.allclose(out, expected, atol=1e-2)
 
     out = x @ x.t()
     out = out.to_dense()
-    assert torch.allclose(out, expected, atol=1e-7)
+    assert torch.allclose(out, expected, atol=1e-2)

torch_sparse/__init__.py

@@ -3,7 +3,7 @@ import os.path as osp
 
 import torch
 
-__version__ = '0.6.10'
+__version__ = '0.6.11'
 
 suffix = 'cuda' if torch.cuda.is_available() else 'cpu'
 

torch_sparse/matmul.py

@@ -78,9 +78,16 @@ def spspmm_sum(src: SparseTensor, other: SparseTensor) -> SparseTensor:
     assert src.sparse_size(1) == other.sparse_size(0)
     rowptrA, colA, valueA = src.csr()
     rowptrB, colB, valueB = other.csr()
+    value = valueA
+    if valueA is not None and valueA.dtype == torch.half:
+        valueA = valueA.to(torch.float)
+    if valueB is not None and valueB.dtype == torch.half:
+        valueB = valueB.to(torch.float)
     M, K = src.sparse_size(0), other.sparse_size(1)
     rowptrC, colC, valueC = torch.ops.torch_sparse.spspmm_sum(
         rowptrA, colA, valueA, rowptrB, colB, valueB, K)
+    if valueC is not None and value is not None:
+        valueC = valueC.to(value.dtype)
     return SparseTensor(row=None, rowptr=rowptrC, col=colC, value=valueC,
                         sparse_sizes=(M, K), is_sorted=True)