【SCU】【Paddle Tensor 第二期 API 支持 0-size TensorNo.46】paddle.linalg.solve 支持 0-size Tensor (#70575)

* support_0size

* fix codestyle

* Update solve_kernel_impl.h

* update

* fix codestyle

* Update test_solve_op.py

* Update test_solve_op.py

* Update test_solve_op.py

Author: PolaKuma

paddle/phi/kernels/impl/solve_kernel_impl.h

@@ -195,6 +195,40 @@ void SolveKernel(const Context& dev_ctx,
                  const DenseTensor& x,
                  const DenseTensor& y,
                  DenseTensor* out) {
+  if (x.numel() == 0 || y.numel() == 0) {
+    auto x_dims = x.dims();
+    auto y_dims = y.dims();
+    std::vector<int> out_dims;
+    if (y_dims.size() == 1) {
+      out_dims =
+          std::vector<int>(x_dims.Get(), x_dims.Get() + x_dims.size() - 2);
+      out_dims.push_back(y_dims[y_dims.size() - 1]);
+    } else {
+      // broadcast
+      std::vector<int> x_shape(x_dims.Get(), x_dims.Get() + x_dims.size() - 2);
+      std::vector<int> y_shape(y_dims.Get(), y_dims.Get() + y_dims.size() - 2);
+      auto x_it = x_shape.rbegin();
+      auto y_it = y_shape.rbegin();
+      while (x_it != x_shape.rend() || y_it != y_shape.rend()) {
+        int x_dim = (x_it != x_shape.rend()) ? *x_it : 1;
+        int y_dim = (y_it != y_shape.rend()) ? *y_it : 1;
+        if (x_dim == 0 || y_dim == 0) {
+          out_dims.push_back(0);
+        } else {
+          out_dims.push_back(std::max(x_dim, y_dim));
+        }
+        if (x_it != x_shape.rend()) ++x_it;
+        if (y_it != y_shape.rend()) ++y_it;
+      }
+      std::reverse(out_dims.begin(), out_dims.end());
+      out_dims.insert(out_dims.end(),
+                      y_dims.Get() + y_dims.size() - 2,
+                      y_dims.Get() + y_dims.size());
+    }
+    out->Resize(phi::make_ddim(out_dims));
+    dev_ctx.template Alloc<T>(out);
+    return;
+  }
   linalg_solve<Context, T>(dev_ctx, x, y, out);
 }
 

test/legacy_test/test_solve_op.py

@@ -923,5 +923,87 @@ def test_dygraph(self):
                     print("The mat is singular")
 
 
+class TestSolveOpAPIZeroDimCase(unittest.TestCase):
+    def setUp(self):
+        np.random.seed(2021)
+        self.place = []
+        self.dtype = "float32"
+        if (
+            os.environ.get('FLAGS_CI_both_cpu_and_gpu', 'False').lower()
+            in ['1', 'true', 'on']
+            or not core.is_compiled_with_cuda()
+        ):
+            self.place.append(paddle.CPUPlace())
+        if core.is_compiled_with_cuda():
+            self.place.append(paddle.CUDAPlace(0))
+
+    def check_static_result(self, place, x_shape, y_shape, np_y_shape):
+        paddle.enable_static()
+        with base.program_guard(base.Program(), base.Program()):
+            paddle_input_x = paddle.static.data(
+                name="input_x", shape=x_shape, dtype=self.dtype
+            )
+            paddle_input_y = paddle.static.data(
+                name="input_y", shape=y_shape, dtype=self.dtype
+            )
+            paddle_result = paddle.linalg.solve(
+                paddle_input_x, paddle_input_y, left=False
+            )
+
+            np_input_x = np.random.random(x_shape).astype(self.dtype)
+            np_input_y = np.random.random(np_y_shape).astype(self.dtype)
+
+            np_result = np.linalg.solve(np_input_x, np_input_y)
+
+            exe = base.Executor(place)
+            fetches = exe.run(
+                base.default_main_program(),
+                feed={"input_x": np_input_x, "input_y": np_input_y},
+                fetch_list=[paddle_result],
+            )
+            np.testing.assert_allclose(fetches[0], np_result, rtol=0.0001)
+
+    def test_static(self):
+        for place in self.place:
+            self.check_static_result(
+                place=place,
+                x_shape=[10, 0, 0],
+                y_shape=[6, 0, 0],
+                np_y_shape=[10, 0, 0],
+            )
+            with self.assertRaises(ValueError) as context:
+                self.check_static_result(
+                    place=place,
+                    x_shape=[10, 0, 0],
+                    y_shape=[10],
+                    np_y_shape=[10],
+                )
+
+    def test_dygraph(self):
+        def run(place, x_shape, y_shape):
+            with base.dygraph.guard(place):
+                input_x_np = np.random.random(x_shape).astype(self.dtype)
+                input_y_np = np.random.random(y_shape).astype(self.dtype)
+
+                tensor_input_x = paddle.to_tensor(input_x_np)
+                tensor_input_y = paddle.to_tensor(input_y_np)
+
+                numpy_output = np.linalg.solve(input_x_np, input_y_np)
+                paddle_output = paddle.linalg.solve(
+                    tensor_input_x, tensor_input_y, left=False
+                )
+                np.testing.assert_allclose(
+                    numpy_output, paddle_output.numpy(), rtol=0.0001
+                )
+                self.assertEqual(
+                    numpy_output.shape, paddle_output.numpy().shape
+                )
+
+        for place in self.place:
+            run(place, x_shape=[10, 0, 0], y_shape=[10, 0, 0])
+            with self.assertRaises(ValueError) as context:
+                run(place, x_shape=[10, 0, 0], y_shape=[10])
+
+
 if __name__ == "__main__":
     unittest.main()