[MPS] Add support for aten::masked_select on mps (#119) (pytorch#85818)

Reuse the `index.Tensor_out` implementation since it's already expanding the bool/byte indices to long tensors.
Pull Request resolved: pytorch#85818
Approved by: https://github.com/kulinseth

Author: DenisVieriu97

aten/src/ATen/mps/MPSFallback.mm

@@ -48,7 +48,6 @@ void mps_error_fallback(const c10::OperatorHandle& op, torch::jit::Stack* stack)
   m.impl("linalg_vector_norm", torch::CppFunction::makeFromBoxedFunction<&mps_fallback>());
   m.impl("sgn.out", torch::CppFunction::makeFromBoxedFunction<&mps_fallback>());
   m.impl("nonzero", torch::CppFunction::makeFromBoxedFunction<&mps_fallback>());
-  m.impl("masked_select", torch::CppFunction::makeFromBoxedFunction<&mps_fallback>());
 }
 
 } // namespace at

aten/src/ATen/native/mps/operations/Indexing.mm

@@ -146,7 +146,6 @@ bool dispatchIndexKernel(TensorIteratorBase& iter,
   return true;
 }
 
-
 static void validateInputData(const TensorIteratorBase& iter, IntArrayRef index_size, IntArrayRef index_stride, const std::string& op, bool accumulate) {
   using namespace mps;
 
@@ -186,6 +185,43 @@ void index_put_kernel_mps(TensorIterator& iter, IntArrayRef index_size, IntArray
   }
 }
 
+static Tensor & masked_select_out_mps_impl(Tensor & result, const Tensor & self, const Tensor & mask) {
+  NoNamesGuard guard;
+
+  TORCH_CHECK(mask.scalar_type() == ScalarType::Byte || mask.scalar_type() == ScalarType::Bool,
+              "masked_select: expected BoolTensor or ByteTensor for mask");
+  TORCH_CHECK(self.scalar_type() == result.scalar_type(),
+              "masked_select(): self and result must have the same scalar type");
+
+  auto mask_temp = (mask.dim() == 0)
+    ? c10::MaybeOwned<Tensor>::owned(mask.unsqueeze(0))
+    : c10::MaybeOwned<Tensor>::borrowed(mask);
+  auto self_temp = (self.dim() == 0)
+    ? c10::MaybeOwned<Tensor>::owned(self.unsqueeze(0))
+    : c10::MaybeOwned<Tensor>::borrowed(self);
+
+  // Cannot reassign to mask_temp and self_temp here! if they are
+  // owning and expand_outplace returns a borrow, the returned borrow
+  // would dangle.
+  auto mask_self_expanded = expand_outplace(*mask_temp, *self_temp);
+  at::index_out(
+      result, *std::get<1>(mask_self_expanded),
+      c10::List<c10::optional<at::Tensor>>({*std::move(std::get<0>(mask_self_expanded))}));
+
+  return result;
+}
+
+Tensor masked_select_mps(const Tensor & self, const Tensor & mask) {
+  namedinference::compute_broadcast_outnames(self, mask);
+  Tensor result = at::empty({0}, self.options());
+  return masked_select_out_mps_impl(result, self, mask);
+}
+
+Tensor & masked_select_out_mps(const Tensor & self, const Tensor & mask, Tensor & result) {
+  namedinference::compute_broadcast_outnames(self, mask);
+  return masked_select_out_mps_impl(result, self, mask);
+}
+
 Tensor flip_mps(const Tensor& self, IntArrayRef dims) {
   using namespace mps;
 

aten/src/ATen/native/native_functions.yaml

@@ -8026,13 +8026,15 @@
   dispatch:
     CPU: masked_select_out_cpu
     CUDA: masked_select_out_cuda
+    MPS: masked_select_out_mps
   tags: dynamic_output_shape
 
 - func: masked_select(Tensor self, Tensor mask) -> Tensor
   variants: method, function
   dispatch:
     CPU: masked_select_cpu
     CUDA: masked_select_cuda
+    MPS: masked_select_mps
   tags: dynamic_output_shape
 
 - func: masked_select_backward(Tensor grad, Tensor input, Tensor mask) -> Tensor

test/test_mps.py

@@ -6048,6 +6048,17 @@ class TestAdvancedIndexing(TestCase):
     supported_dtypes = [torch.float32, torch.float16, torch.int64, torch.int32, torch.int16, torch.uint8]
     supported_np_dtypes = [np.float32, np.float16, np.int64, np.int32, np.int16, np.uint8]
 
+    def test_masked_select(self):
+        x = torch.randn(3, 4)
+        x_mps = x.to("mps")
+        mask = x.ge(0.5)
+        mask_mps = x_mps.ge(0.5)
+
+        res = torch.masked_select(x, mask)
+        res_mps = torch.masked_select(x_mps, mask_mps)
+
+        self.assertEqual(res, res_mps)
+
     # examples from https://www.tutorialspoint.com/numpy/numpy_advanced_indexing.htm
     def test_indexing_get(self):
         def helper(dtype):
@@ -6390,10 +6401,10 @@ def test_index_put_accumulate_duplicate_indices(self, device="mps"):
             delta = torch.empty(i, dtype=torch.float32, device=device).uniform_(-1, 1)
 
             # cumsum not supported on 'mps', fallback on 'cpu'
-            indices = delta.to("cpu").cumsum(0).long().to("mps")
+            indices = delta.cpu().cumsum(0).long().to("mps")
 
             # abs for int64 is not supported on mps, fallback on 'cpu' to calculate it
-            input = torch.randn(indices.to("cpu").abs().to("mps").max() + 1, device=device)
+            input = torch.randn(indices.cpu().abs().max().to("mps") + 1, device=device)
             values = torch.randn(indices.size(0), device=device)
             output = input.index_put((indices,), values, accumulate=True)