fixed nnz value (#17)

fixes for the rspmm kernels
DeepGraphLearning · Mar 30, 2024 · c414f83 · c414f83
1 parent 871add0
commit c414f83
Show file tree

Hide file tree

Showing 3 changed files with 8 additions and 5 deletions.
diff --git a/ultra/rspmm/rspmm.py b/ultra/rspmm/rspmm.py
@@ -181,7 +181,10 @@ def generalized_rspmm(edge_index, edge_type, edge_weight, relation, input, sum="
 def load_extension(name, sources, extra_cflags=None, extra_cuda_cflags=None, **kwargs):
     if extra_cflags is None:
         extra_cflags = ["-Ofast"]
-        if torch.backends.openmp.is_available():
+        # PyTorch 2.2.1+ on Apple Silicon is now compiled by default with OpenMP
+        # However, installing OpenMP on macs properly and wiring it together to the compiler is tedious
+        # So on macs we turn off OpenMP (as the default behavior in all torch < 2.2.1 versions)
+        if torch.backends.openmp.is_available() and not sys.platform.startswith('darwin'):
             extra_cflags += ["-fopenmp", "-DAT_PARALLEL_OPENMP"]
         else:
             extra_cflags.append("-DAT_PARALLEL_NATIVE")

diff --git a/ultra/rspmm/source/rspmm.cpp b/ultra/rspmm/source/rspmm.cpp
@@ -135,7 +135,7 @@ Tensor rspmm_forward_cpu(const Tensor &edge_index_, const Tensor &edge_type_, co
     const Tensor relation = relation_.contiguous();
     const Tensor input = input_.contiguous();
 
-    int64_t nnz = edge_index.size(0);
+    int64_t nnz = edge_index.size(1);
     int64_t num_row = input.size(0);
     int64_t dim = input.size(1);
     Tensor output = at::empty({num_row, dim}, input.options());
@@ -183,7 +183,7 @@ std::tuple<Tensor, Tensor, Tensor> rspmm_backward_cpu(
     const Tensor output = output_.contiguous();
     const Tensor output_grad = output_grad_.contiguous();
 
-    int64_t nnz = edge_index.size(0);
+    int64_t nnz = edge_index.size(1);
     int64_t num_row = input.size(0);
     int64_t dim = input.size(1);
     Tensor weight_grad = at::zeros_like(edge_weight);

diff --git a/ultra/rspmm/source/rspmm.cu b/ultra/rspmm/source/rspmm.cu
@@ -230,7 +230,7 @@ Tensor rspmm_forward_cuda(const Tensor &edge_index_, const Tensor &edge_type_, c
     const Tensor relation = relation_.contiguous();
     const Tensor input = input_.contiguous();
 
-    int64_t nnz = edge_index.size(0);
+    int64_t nnz = edge_index.size(1);
     int64_t num_row = input.size(0);
     int64_t dim = input.size(1);
     Tensor output = at::empty({num_row, dim}, input.options());
@@ -289,7 +289,7 @@ std::tuple<Tensor, Tensor, Tensor> rspmm_backward_cuda(
     const Tensor output = output_.contiguous();
     const Tensor output_grad = output_grad_.contiguous();
 
-    int64_t nnz = edge_index.size(0);
+    int64_t nnz = edge_index.size(1);
     int64_t num_row = input.size(0);
     int64_t dim = input.size(1);
     Tensor weight_grad = at::zeros_like(edge_weight);