[NPU] add NPU ops of compare, test=develop (#34365)

* [NPU] add NPU ops&uts of compare, test=develop

* testing

* try style-format

* [NPU] update compare_op_npu uts

* [NPU] fix code sytle of test_compare_op_npu.py

paddle/fluid/operators/controlflow/compare_op_npu.cc

@@ -59,6 +59,56 @@ class LessThanNPUKernel : public framework::OpKernel<T> {
   }
 };
 
+template <typename DeviceContext, typename T>
+class LessEqualNPUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* x = ctx.Input<framework::LoDTensor>("X");
+    auto* y = ctx.Input<framework::LoDTensor>("Y");
+    auto* z = ctx.Output<framework::LoDTensor>("Out");
+    z->mutable_data<bool>(ctx.GetPlace());
+    const auto& runner = NpuOpRunner("LessEqual", {*x, *y}, {*z});
+    auto stream =
+        ctx.template device_context<paddle::platform::NPUDeviceContext>()
+            .stream();
+    runner.Run(stream);
+  }
+};
+
+template <typename DeviceContext, typename T>
+class GreaterThanNPUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* x = ctx.Input<framework::LoDTensor>("X");
+    auto* y = ctx.Input<framework::LoDTensor>("Y");
+    auto* z = ctx.Output<framework::LoDTensor>("Out");
+
+    z->mutable_data<bool>(ctx.GetPlace());
+    const auto& runner = NpuOpRunner("Greater", {*x, *y}, {*z});
+    auto stream =
+        ctx.template device_context<paddle::platform::NPUDeviceContext>()
+            .stream();
+    runner.Run(stream);
+  }
+};
+
+template <typename DeviceContext, typename T>
+class GreaterEqualNPUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* x = ctx.Input<framework::LoDTensor>("X");
+    auto* y = ctx.Input<framework::LoDTensor>("Y");
+    auto* z = ctx.Output<framework::LoDTensor>("Out");
+
+    z->mutable_data<bool>(ctx.GetPlace());
+    const auto& runner = NpuOpRunner("GreaterEqual", {*x, *y}, {*z});
+    auto stream =
+        ctx.template device_context<paddle::platform::NPUDeviceContext>()
+            .stream();
+    runner.Run(stream);
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle
 
@@ -75,4 +125,22 @@ REGISTER_OP_NPU_KERNEL(
     ops::LessThanNPUKernel<paddle::platform::NPUDeviceContext,
                            paddle::platform::float16>);
 
+REGISTER_OP_NPU_KERNEL(
+    less_equal,
+    ops::LessEqualNPUKernel<paddle::platform::NPUDeviceContext, float>,
+    ops::LessEqualNPUKernel<paddle::platform::NPUDeviceContext,
+                            paddle::platform::float16>);
+
+REGISTER_OP_NPU_KERNEL(
+    greater_than,
+    ops::GreaterThanNPUKernel<paddle::platform::NPUDeviceContext, float>,
+    ops::GreaterThanNPUKernel<paddle::platform::NPUDeviceContext,
+                              paddle::platform::float16>);
+
+REGISTER_OP_NPU_KERNEL(
+    greater_equal,
+    ops::GreaterEqualNPUKernel<paddle::platform::NPUDeviceContext, float>,
+    ops::GreaterEqualNPUKernel<paddle::platform::NPUDeviceContext,
+                               paddle::platform::float16>);
+
 #endif

python/paddle/fluid/tests/unittests/npu/test_compare_op_npu.py

@@ -21,121 +21,136 @@ sys.path.append("..")
 from op_test import OpTest
 import paddle
 import paddle.fluid as fluid
-
-paddle.enable_static()
-SEED = 2021
-
-
-class TestEqual(OpTest):
-    def setUp(self):
-        self.set_npu()
-        self.op_type = "equal"
-        self.place = paddle.NPUPlace(0)
-
-        self.init_dtype()
-        np.random.seed(SEED)
-        x = np.random.uniform(1, 2, [11, 17]).astype(self.dtype)
-        y = np.random.uniform(1, 2, [11, 17]).astype(self.dtype)
-        out = x == y  # all elements are not equal
-
-        self.inputs = {
-            'X': OpTest.np_dtype_to_fluid_dtype(x),
-            'Y': OpTest.np_dtype_to_fluid_dtype(y)
-        }
-        self.outputs = {'Out': out}
-
-    def set_npu(self):
-        self.__class__.use_npu = True
-
-    def init_dtype(self):
-        self.dtype = np.float32
-
-    def test_check_output(self):
-        self.check_output_with_place(self.place)
-
-
-class TestLessthan(OpTest):
-    def setUp(self):
-        self.set_npu()
-        self.op_type = "less_than"
-        self.place = paddle.NPUPlace(0)
-
-        self.init_dtype()
-        np.random.seed(SEED)
-        x = np.random.uniform(1, 2, [11, 17]).astype(self.dtype)
-        y = np.random.uniform(1, 2, [11, 17]).astype(self.dtype)
-        out = x < y
-
-        self.inputs = {
-            'X': OpTest.np_dtype_to_fluid_dtype(x),
-            'Y': OpTest.np_dtype_to_fluid_dtype(y)
-        }
-        self.outputs = {'Out': out}
-
-    def set_npu(self):
-        self.__class__.use_npu = True
-
-    def init_dtype(self):
-        self.dtype = np.float32
-
-    def test_check_output(self):
-        self.check_output_with_place(self.place)
-
-
-class TestEqual2(TestEqual):
-    def setUp(self):
-        self.set_npu()
-        self.op_type = "equal"
-        self.place = paddle.NPUPlace(0)
-
-        self.init_dtype()
-        np.random.seed(SEED)
-        x = np.random.uniform(1, 2, [11, 17]).astype(self.dtype)
-        y = x.copy()
-        y[0][1] = 1
-        out = x == y  # all elements are equal, except position [0][1]
-
-        self.inputs = {
-            'X': OpTest.np_dtype_to_fluid_dtype(x),
-            'Y': OpTest.np_dtype_to_fluid_dtype(y)
-        }
-        self.outputs = {'Out': out}
-
-
-class TestLessthan2(TestLessthan):
-    def setUp(self):
-        self.set_npu()
-        self.op_type = "less_than"
-        self.place = paddle.NPUPlace(0)
-
-        self.init_dtype()
-        np.random.seed(SEED)
-        x = np.random.uniform(1, 2, [11, 17]).astype(self.dtype)
-        y = x.copy()
-        y[0][1] = 1
-        out = x < y  # all elements are equal, except position [0][1]
-
-        self.inputs = {
-            'X': OpTest.np_dtype_to_fluid_dtype(x),
-            'Y': OpTest.np_dtype_to_fluid_dtype(y)
-        }
-        self.outputs = {'Out': out}
-
-
-class TestEqual2FP16(TestEqual2):
-    def init_dtype(self):
-        self.dtype = np.float16
-
-
-class TestEqual2Int(TestEqual2):
-    def init_dtype(self):
-        self.dtype = np.int32
-
-
-class TestLessthan2FP16(TestLessthan2):
-    def init_dtype(self):
-        self.dtype = np.float16
-
+from paddle.fluid import Program, program_guard
+
+
+def create_test_class(op_type, typename, callback):
+    class Cls(OpTest):
+        def setUp(self):
+            self.set_npu()
+            self.place = paddle.NPUPlace(0)
+            x = np.random.random(size=(10, 7)).astype(typename)
+            y = np.random.random(size=(10, 7)).astype(typename)
+            out = callback(x, y)
+            self.inputs = {'X': x, 'Y': y}
+            self.outputs = {'Out': out}
+            self.op_type = op_type
+
+        def set_npu(self):
+            self.__class__.use_npu = True
+
+        def test_output(self):
+            self.check_output_with_place(place=self.place)
+
+        def test_errors(self):
+            paddle.enable_static()
+            with program_guard(Program(), Program()):
+                a = fluid.layers.data(name='a', shape=[2], dtype='float32')
+                b = fluid.layers.data(name='b', shape=[2], dtype='float32')
+                c = fluid.layers.data(name='c', shape=[2], dtype='int16')
+                d = fluid.create_lod_tensor(np.array([[-1]]), [[1]], self.place)
+
+                op = eval("fluid.layers.%s" % self.op_type)
+                self.assertRaises(TypeError, op, x=a, y=b, axis=True)
+                self.assertRaises(TypeError, op, x=a, y=b, force_cpu=1)
+                self.assertRaises(TypeError, op, x=a, y=b, cond=1)
+                self.assertRaises(TypeError, op, x=a, y=c)
+                self.assertRaises(TypeError, op, x=c, y=a)
+                self.assertRaises(TypeError, op, x=a, y=d)
+                self.assertRaises(TypeError, op, x=d, y=a)
+                self.assertRaises(TypeError, op, x=c, y=d)
+
+        def test_dynamic_api(self):
+            paddle.disable_static()
+            paddle.set_device('npu:0')
+            x = np.random.random(size=(10, 7)).astype(typename)
+            y = np.random.random(size=(10, 7)).astype(typename)
+            real_result = callback(x, y)
+            x = paddle.to_tensor(x, dtype=typename)
+            y = paddle.to_tensor(y, dtype=typename)
+            op = eval("paddle.%s" % (self.op_type))
+            out = op(x, y)
+            self.assertEqual((out.numpy() == real_result).all(), True)
+
+        @unittest.skipIf(typename == 'float16', "float16 is not supported now")
+        def test_broadcast_api_1(self):
+            paddle.enable_static()
+            with program_guard(Program(), Program()):
+                x = paddle.static.data(
+                    name='x', shape=[1, 2, 1, 3], dtype=typename)
+                y = paddle.static.data(
+                    name='y', shape=[1, 2, 3], dtype=typename)
+                op = eval("paddle.%s" % (self.op_type))
+                out = op(x, y)
+                exe = paddle.static.Executor(self.place)
+                input_x = np.arange(1, 7).reshape((1, 2, 1, 3)).astype(typename)
+                input_y = np.arange(0, 6).reshape((1, 2, 3)).astype(typename)
+                real_result = callback(input_x, input_y)
+                res, = exe.run(feed={"x": input_x,
+                                     "y": input_y},
+                               fetch_list=[out])
+            self.assertEqual((res == real_result).all(), True)
+
+        @unittest.skipIf(typename == 'float16', "float16 is not supported now")
+        def test_broadcast_api_2(self):
+            paddle.enable_static()
+            with program_guard(Program(), Program()):
+                x = paddle.static.data(
+                    name='x', shape=[1, 2, 3], dtype=typename)
+                y = paddle.static.data(
+                    name='y', shape=[1, 2, 1, 3], dtype=typename)
+                op = eval("paddle.%s" % (self.op_type))
+                out = op(x, y)
+                exe = paddle.static.Executor(self.place)
+                input_x = np.arange(0, 6).reshape((1, 2, 3)).astype(typename)
+                input_y = np.arange(1, 7).reshape((1, 2, 1, 3)).astype(typename)
+                real_result = callback(input_x, input_y)
+                res, = exe.run(feed={"x": input_x,
+                                     "y": input_y},
+                               fetch_list=[out])
+            self.assertEqual((res == real_result).all(), True)
+
+        @unittest.skipIf(typename == 'float16', "float16 is not supported now")
+        def test_broadcast_api_3(self):
+            paddle.enable_static()
+            with program_guard(Program(), Program()):
+                x = paddle.static.data(name='x', shape=[5], dtype=typename)
+                y = paddle.static.data(name='y', shape=[3, 1], dtype=typename)
+                op = eval("paddle.%s" % (self.op_type))
+                out = op(x, y)
+                exe = paddle.static.Executor(self.place)
+                input_x = np.arange(0, 5).reshape((5)).astype(typename)
+                input_y = np.array([5, 3, 2]).reshape((3, 1)).astype(typename)
+                real_result = callback(input_x, input_y)
+                res, = exe.run(feed={"x": input_x,
+                                     "y": input_y},
+                               fetch_list=[out])
+            self.assertEqual((res == real_result).all(), True)
+
+        @unittest.skipIf(typename == 'float16', "float16 is not supported now")
+        def test_attr_name(self):
+            paddle.enable_static()
+            with program_guard(Program(), Program()):
+                x = fluid.layers.data(name='x', shape=[4], dtype=typename)
+                y = fluid.layers.data(name='y', shape=[4], dtype=typename)
+                op = eval("paddle.%s" % (self.op_type))
+                out = op(x=x, y=y, name="name_%s" % (self.op_type))
+            self.assertEqual("name_%s" % (self.op_type) in out.name, True)
+
+    cls_name = "{0}_{1}".format(op_type, typename)
+    Cls.__name__ = cls_name
+    globals()[cls_name] = Cls
+
+
+for _type_name in {'float16', 'float32', 'int32'}:
+    if _type_name == 'int32':
+        create_test_class('equal', _type_name, lambda _a, _b: _a == _b)
+        continue
+    create_test_class('equal', _type_name, lambda _a, _b: _a == _b)
+    create_test_class('less_than', _type_name, lambda _a, _b: _a < _b)
+    create_test_class('less_equal', _type_name, lambda _a, _b: _a <= _b)
+    create_test_class('greater_than', _type_name, lambda _a, _b: _a > _b)
+    create_test_class('greater_equal', _type_name, lambda _a, _b: _a >= _b)
 
 if __name__ == '__main__':
     unittest.main()