【0D output】add_0D_output_support (#52857)

* add 0d support for dist, trace, paddle.linalg.cond test=allcase

* add_0d_output_support_for_det

* test=allcase

* support_0d_output_for_linalg.norm

* support linalg.norm 0d output, test=allcase

* fix 0D test

* fix zero dim test, test=allcase

* fix 0D test

* fix tets,test=allcase

* fix error,test=allcase

* fix errors ,test=allcase

* add static backward , test=allcase

* add static backwward test, test=allcase

* fix pr-ci-build error;test=document_fix (#53060)

* [Cherry-Pick] Unique support float16&bfloat16 (#53023)

unique支持float16和bfloat16数据类型，并完善相关单测。

* slogdet_support_0D_output

* add new case

* fix tests, test=allcase

* fix p_norm related test, test=allcase

* fix some err, test=allcase

* test=allcase

* move out trace

* open some case, test=allcase

* fix norm all case, test=allcase

* fix some test error, test=allcase

* fix typro,test=allcase

* fix test err, test=allcase

* test=allcase

* test

* fix test error, test=allcase

* fix test error, test=allcase

* fallback norm, test=allcase

---------

Co-authored-by: tianshuo78520a <707759223@qq.com>
Co-authored-by: Zhang Zheng <32410583+ZzSean@users.noreply.github.com>

paddle/phi/infermeta/binary.cc

@@ -990,7 +990,7 @@ void DistInferMeta(const MetaTensor& x,
                         "The Input(Y) has not been initialized properly. The "
                         "shape of Input(Y) = [%s].",
                         y_dims));
-  out->set_dims({1});
+  out->set_dims(phi::make_ddim({}));
   out->set_dtype(x.dtype());
 }
 

paddle/phi/infermeta/unary.cc

@@ -2767,7 +2767,6 @@ void PNormInferMeta(const MetaTensor& x,
     if (reduce_dims.size() == 0) {
       reduce_dims.emplace_back(1);
     }
-
     x_dim[axis] = 1;
   }
 

paddle/phi/kernels/impl/determinant_grad_kernel_impl.h

@@ -91,10 +91,10 @@ void DeterminantGradKernel(const Context& dev_ctx,
             " input tensor's, but here differ %d",
             input_dims_size - out_grad.dims().size()));
   } else if (input_dims_size == 2) {
-    // input dims size 2 and grad dims size 1 is possible
+    // input dims size 2 and grad dims size 0 is possible
     PADDLE_ENFORCE_EQ(
         out_grad.dims().size(),
-        1,
+        0,
         phi::errors::InvalidArgument(
             "The grad tensor of det dims size should be 2 less than"
             " input tensor's, but here differ %d",

paddle/phi/kernels/impl/determinant_kernel_impl.h

@@ -126,7 +126,7 @@ void DeterminantKernel(const Context& dev_ctx,
     out->Resize(output_dims);
   } else {
     // when input is a two-dimension matrix, The det value is a number.
-    out->Resize({1});
+    out->Resize(phi::make_ddim({}));
   }
   VLOG(10) << "output dim:" << out->dims();
 }

python/paddle/fluid/tests/unittests/test_zero_dim_tensor.py

@@ -2616,6 +2616,107 @@ def test_multi_dot(self):
         self.assertEqual(b.grad.shape, [4, 5])
         self.assertEqual(c.grad.shape, [5])
 
+    def test_cov(self):
+        xt = paddle.randn((3, 4))
+        xt.stop_gradient = False
+        xt_1 = paddle.randn((12,))
+        xt_1.stop_gradient = False
+
+        xt_out = paddle.linalg.cov(xt)
+        xt_out.retain_grads()
+        xt_out.backward()
+        self.assertEqual(xt_out.shape, [3, 3])
+        self.assertEqual(xt.grad.shape, [3, 4])
+
+        xt_1_out = paddle.linalg.cov(xt_1)
+        xt_1.retain_grads()
+        xt_1_out.backward()
+        self.assertEqual(xt_1_out.shape, [])
+        self.assertEqual(xt_1.grad.shape, [12])
+
+    def test_det(self):
+        xt = paddle.randn([3, 3, 3])
+        xt.stop_gradient = False
+        xt_1 = paddle.randn([3, 3])
+        xt_1.stop_gradient = False
+
+        xt_out = paddle.linalg.det(xt)
+        xt.retain_grads()
+        xt_out.backward()
+        self.assertEqual(xt_out.shape, [3])
+        self.assertEqual(xt.grad.shape, [3, 3, 3])
+
+        xt_1_out = paddle.linalg.det(xt_1)
+        xt_1.retain_grads()
+        xt_1_out.backward()
+        self.assertEqual(xt_1_out.shape, [])
+        self.assertEqual(xt_1.grad.shape, [3, 3])
+
+    def test_dist(self):
+        x = paddle.to_tensor([[3, 3], [3, 3]], dtype="float32")
+        y = paddle.to_tensor([[3, 3], [3, 1]], dtype="float32")
+        x.stop_gradient = False
+        y.stop_gradient = False
+        out = paddle.dist(x, y, 0)
+        out.backward()
+
+        self.assertEqual(out.shape, [])
+        np.testing.assert_allclose(out, np.array(1))
+        self.assertEqual(x.grad.shape, [2, 2])
+        self.assertEqual(y.grad.shape, [2, 2])
+
+    def test_linalg_cond(self):
+        def assert_shape(out):
+            self.assertEqual(out.shape, [])
+
+        # x1 = paddle.to_tensor([[1.0, 0, -1], [0, 1, 0], [1, 0, 1]])
+        # x1.stop_gradient = False
+        # p = 2 : use paddle.sum
+        # out = paddle.linalg.cond(x1)
+        # assert_shape(out)
+
+        # p = fro : use paddle.sum
+        # x2 = paddle.to_tensor([[1.0, 0, -1], [0, 1, 0], [1, 0, 1]])
+        # x2.stop_gradient = False
+        # out_fro = paddle.linalg.cond(x2, p='fro')
+        # assert_shape(out_fro)
+
+        # p = nuc : use paddle.sum
+        # x3 = paddle.to_tensor([[1.0, 0, -1], [0, 1, 0], [1, 0, 1]])
+        # x3.stop_gradient = False
+        # out_nuc = paddle.linalg.cond(x3, p='nuc')
+        # assert_shape(out_nuc)
+
+        # p in (-1, 1) : use paddle.sum
+        # x4 = paddle.to_tensor([[1.0, 0, -1], [0, 1, 0], [1, 0, 1]])
+        # x4.stop_gradient = False
+        # out_1 = paddle.linalg.cond(x4, p=1)
+        # assert_shape(out_1)
+        # x5 = paddle.to_tensor([[1.0, 0, -1], [0, 1, 0], [1, 0, 1]])
+        # x5.stop_gradient = False
+        # out_minus_1 = paddle.linalg.cond(x5, p=-1)
+        # assert_shape(out_minus_1)
+
+        # p in (-2, 2)  depends on paddle.sum
+        # x6 = paddle.to_tensor([[1.0, 0, -1], [0, 1, 0], [1, 0, 1]])
+        # x6.stop_gradient = False
+        # out_2 = paddle.linalg.cond(x6, p=2)
+        # assert_shape(out_2)
+
+        # p in (-inf, inf):use paddle.sum
+        # x8 = paddle.to_tensor([[1.0, 0, -1], [0, 1, 0], [1, 0, 1]])
+        # x8.stop_gradient = False
+        # out_inf = paddle.linalg.cond(x8, p=float("inf"))
+        # assert_shape(out_inf)
+
+        # depends on paddle.sum
+        # a = paddle.randn([2, 4, 4])
+        # a.stop_gradient = False
+        # a_cond_fro = paddle.linalg.cond(a, p='fro')
+        # a_cond_fro.backward()
+        # self.assertEqual(len(a_cond_fro.shape), 1)
+        # self.assertEqual(a.grad.shape, [2, 4, 4])
+
     def test_trace(self):
         x = paddle.to_tensor([[3, 2], [1, 9]], dtype="float32")
         x.stop_gradient = False
@@ -4723,6 +4824,148 @@ def test_multi_dot(self):
         self.assertEqual(res[2].shape, (4, 5))
         self.assertEqual(res[3].shape, (5,))
 
+    @prog_scope()
+    def test_cov(self):
+        xt_1 = paddle.randn((12,))
+        xt_1.stop_gradient = False
+
+        out = paddle.linalg.cov(xt_1)
+        paddle.static.append_backward(out)
+
+        prog = paddle.static.default_main_program()
+
+        res = self.exe.run(prog, fetch_list=[out, xt_1.grad_name])
+        self.assertEqual(res[0].shape, ())
+        self.assertEqual(res[1].shape, (12,))
+
+    @prog_scope()
+    def test_det(self):
+        xt_1 = paddle.randn((3, 3))
+        xt_1.stop_gradient = False
+
+        out = paddle.linalg.det(xt_1)
+        paddle.static.append_backward(out.sum())
+
+        prog = paddle.static.default_main_program()
+        res = self.exe.run(prog, fetch_list=[out, xt_1.grad_name])
+        self.assertEqual(res[0].shape, ())
+        self.assertEqual(res[1].shape, (3, 3))
+
+    @prog_scope()
+    def test_dist(self):
+        x = paddle.to_tensor([[3, 3], [3, 3]], dtype="float32")
+        y = paddle.to_tensor([[3, 3], [3, 1]], dtype="float32")
+        x.stop_gradient = False
+        y.stop_gradient = False
+        out = paddle.dist(x, y)
+        paddle.static.append_backward(out)
+
+        prog = paddle.static.default_main_program()
+        res = self.exe.run(prog, fetch_list=[out, x.grad_name, y.grad_name])
+
+        self.assertEqual(res[0].shape, ())
+        self.assertEqual(res[1].shape, (2, 2))
+        self.assertEqual(res[1].shape, (2, 2))
+        np.testing.assert_array_equal(res[0], np.array(2).astype(np.float32))
+
+    @prog_scope()
+    def test_linalg_cond(self):
+        pass
+        # use paddle.sum
+        # x = paddle.to_tensor([[1.0, 0, -1], [0, 1, 0], [1, 0, 1]])
+        # x.stop_gradient = False
+        # out = paddle.linalg.cond(x)
+        # paddle.static.append_backward(out)
+
+        # prog = paddle.static.default_main_program()
+        # res = self.exe.run(prog, fetch_list=[out, x.grad_name])
+
+        # self.assertTrue(res[0].shape, ())
+        # self.assertTrue(res[1].shape, (3, 3))
+
+        # p = fro : use paddle.sum
+        # x2 = paddle.to_tensor([[1.0, 0, -1], [0, 1, 0], [1, 0, 1]])
+        # x2.stop_gradient = False
+        # out_fro = paddle.linalg.cond(x2, p='fro')
+        # paddle.static.append_backward(out_fro)
+
+        # prog = paddle.static.default_main_program()
+        # res = self.exe.run(prog, fetch_list=[out_fro, x.grad_name])
+
+        # self.assertTrue(res[0].shape, ())
+        # self.assertTrue(res[1].shape, (3, 3))
+
+        # p = nuc : use paddle.sum
+        # x3 = paddle.to_tensor([[1.0, 0, -1], [0, 1, 0], [1, 0, 1]])
+        # x3.stop_gradient = False
+        # out_nuc = paddle.linalg.cond(x3, p='nuc')
+        # paddle.static.append_backward(out_nuc)
+
+        # prog = paddle.static.default_main_program()
+        # res = self.exe.run(prog, fetch_list=[out_nuc, x.grad_name])
+
+        # self.assertTrue(res[0].shape, ())
+        # self.assertTrue(res[1].shape, (3, 3))
+
+        # p in (-1, 1) : use paddle.sum
+        # x4 = paddle.to_tensor([[1.0, 0, -1], [0, 1, 0], [1, 0, 1]])
+        # x4.stop_gradient = False
+        # out_1 = paddle.linalg.cond(x4, p=1)
+        # paddle.static.append_backward(out_1)
+
+        # prog = paddle.static.default_main_program()
+        # res = self.exe.run(prog, fetch_list=[out_1, x.grad_name])
+
+        # self.assertTrue(res[0].shape, ())
+        # self.assertTrue(res[1].shape, (3, 3))
+
+        # x5 = paddle.to_tensor([[1.0, 0, -1], [0, 1, 0], [1, 0, 1]])
+        # x5.stop_gradient = False
+        # out_minus_1 = paddle.linalg.cond(x5, p=-1)
+        # paddle.static.append_backward(out_minus_1)
+
+        # prog = paddle.static.default_main_program()
+        # res = self.exe.run(prog, fetch_list=[out_minus_1, x.grad_name])
+
+        # self.assertTrue(res[0].shape, ())
+        # self.assertTrue(res[1].shape, (3, 3))
+
+        # p in (-2, 2)  depends on paddle.sum
+        # x6 = paddle.to_tensor([[1.0, 0, -1], [0, 1, 0], [1, 0, 1]])
+        # x6.stop_gradient = False
+        # out_2 = paddle.linalg.cond(x6, p=2)
+        # paddle.static.append_backward(out_2)
+
+        # prog = paddle.static.default_main_program()
+        # res = self.exe.run(prog, fetch_list=[out_2, x.grad_name])
+
+        # self.assertTrue(res[0].shape, ())
+        # self.assertTrue(res[1].shape, (3, 3))
+
+        # p in (-inf, inf):use paddle.sum
+        # x8 = paddle.to_tensor([[1.0, 0, -1], [0, 1, 0], [1, 0, 1]])
+        # x8.stop_gradient = False
+        # out_inf = paddle.linalg.cond(x8, p=float("inf"))
+        # paddle.static.append_backward(out_inf)
+
+        # prog = paddle.static.default_main_program()
+        # res = self.exe.run(prog, fetch_list=[out_inf, x.grad_name])
+
+        # self.assertTrue(res[0].shape, ())
+        # self.assertTrue(res[1].shape, (3, 3))
+
+        # depends on paddle.sum
+        # a = paddle.randn([2, 4, 4])
+        # a.stop_gradient = False
+        # a_cond_fro = paddle.linalg.cond(a, p='fro')
+        # paddle.static.append_backward(a_cond_fro)
+
+        # prog = paddle.static.default_main_program()
+        # res = self.exe.run(prog, fetch_list=[a_cond_fro, a.grad_name])
+
+        # self.assertEqual(res[0].shape, (2,))
+        # self.assertEqual(res[1].shape, (2, 4, 4))
+
     @prog_scope()
     def test_trace(self):
         x = paddle.to_tensor([[3, 2], [1, 9]], dtype="float32")

test/autograd/test_orig2prim.py

@@ -463,7 +463,12 @@ def init_data(self):
         }
 
         self.orig2prim_args = (X,)
-        self.all_ops = ['p_norm', 'reshape_p', 'abs_p', 'reduce_sum_p']
+        self.all_ops = [
+            'p_norm',
+            'reshape_p',
+            'abs_p',
+            'reduce_sum_p',
+        ]
         self.out_map = {0: self.output['Out']}