add elementwise max grad op for npu (#34862)

* add elementwise max grad op for npu

* add elementwise max grad op for npu

* add elementwise max grad op for npu

* add elementwise max grad op for npu

* add elementwise max grad op for npu

paddle/fluid/operators/elementwise/elementwise_max_op_npu.cc

@@ -12,10 +12,8 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#include <memory>
-#include <string>
-
 #include "paddle/fluid/operators/elementwise/elementwise_max_op.h"
+#include "paddle/fluid/operators/elementwise/elementwise_npu.h"
 #include "paddle/fluid/operators/npu_op_runner.h"
 
 namespace paddle {
@@ -27,21 +25,202 @@ template <typename DeviceContext, typename T>
 class ElementwiseMaxNPUKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
+    auto& dev_ctx =
+        ctx.template device_context<paddle::platform::NPUDeviceContext>();
+
     auto* x = ctx.Input<Tensor>("X");
     auto* y = ctx.Input<Tensor>("Y");
-
     auto* out = ctx.Output<Tensor>("Out");
-
-    auto place = ctx.GetPlace();
-
-    out->mutable_data<T>(place);
+    out->mutable_data<T>(ctx.GetPlace());
+    int axis = ctx.Attr<int>("axis");
+
+    bool direct_compute = false;
+    auto x_dims = x->dims();
+    auto y_dims = y->dims();
+    axis = (axis == -1 ? std::abs(x_dims.size() - y_dims.size()) : axis);
+    if (x_dims.size() >= y_dims.size()) {
+      direct_compute =
+          y_dims == framework::slice_ddim(x_dims, axis, x_dims.size());
+    } else {
+      direct_compute =
+          x_dims == framework::slice_ddim(y_dims, axis, y_dims.size());
+    }
 
     auto stream =
         ctx.template device_context<paddle::platform::NPUDeviceContext>()
             .stream();
 
-    const auto& runner = NpuOpRunner("Maximum", {*x, *y}, {*out}, {});
-    runner.Run(stream);
+    if (direct_compute) {
+      const auto& runner = NpuOpRunner("Maximum", {*x, *y}, {*out}, {});
+      runner.Run(stream);
+    } else {
+      Tensor transformed_x, transformed_y;
+      NpuElementWiseOpBroadcast<T>(dev_ctx, x, y, axis, &transformed_x,
+                                   &transformed_y);
+      const auto& runner =
+          NpuOpRunner("Maximum", {transformed_x, transformed_y}, {*out}, {});
+      runner.Run(stream);
+    }
+  }
+};
+
+template <typename DeviceContext, typename T>
+class ElementwiseMaxGradNPUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto& dev_ctx =
+        ctx.template device_context<paddle::platform::NPUDeviceContext>();
+    auto* x = ctx.Input<Tensor>("X");
+    auto* y = ctx.Input<Tensor>("Y");
+    auto* dout = ctx.Input<Tensor>(framework::GradVarName("Out"));
+    auto* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
+    auto* dy = ctx.Output<Tensor>(framework::GradVarName("Y"));
+    int axis = ctx.Attr<int>("axis");
+
+    // The ascend elementwise_max_grad op only supports broadcast
+    // when axis is -1, and requires all the inputs must have the
+    // same shape when axis is not -1. For convenience, we should
+    // broadcast the original input x and y to transformed_x and
+    // transformed_x firstly, then use tmp tensor to get the op
+    // output, last reduce the tmp tensor shape to match the
+    // paddle output.
+
+    auto x_dims = x->dims();
+    auto y_dims = y->dims();
+    axis = (axis == -1 ? std::abs(x_dims.size() - y_dims.size()) : axis);
+    Tensor transformed_x, transformed_y;
+    NpuElementWiseOpBroadcast<T>(dev_ctx, x, y, axis, &transformed_x,
+                                 &transformed_y);
+
+    auto dout_dims = dout->dims();
+    auto stream = dev_ctx.stream();
+    framework::NPUAttributeMap attr_input = {{"grad_x", true},
+                                             {"grad_y", true}};
+    // Reshape info vector.
+    std::vector<int> reduce_axes;
+
+    if (dx && dy) {
+      dx->mutable_data<T>(ctx.GetPlace());
+      dy->mutable_data<T>(ctx.GetPlace());
+      Tensor tmp_dx;
+      tmp_dx.mutable_data<T>(dout_dims, ctx.GetPlace());
+      Tensor tmp_dy;
+      tmp_dy.mutable_data<T>(dout_dims, ctx.GetPlace());
+
+      const auto& runner =
+          NpuOpRunner("MaximumGrad", {*dout, transformed_x, transformed_y},
+                      {tmp_dx, tmp_dy}, attr_input);
+      runner.Run(stream);
+
+      if (x_dims != dout_dims) {
+        reduce_axes.clear();
+        int src_axis = (x_dims.size() < dout_dims.size() ? axis : 0);
+        for (int ax = 0; ax < dout_dims.size(); ++ax) {
+          if ((ax < src_axis || ax >= src_axis + x_dims.size()) ||
+              (dout_dims[ax] > 1 && x_dims[ax - src_axis] == 1)) {
+            reduce_axes.push_back(ax);
+          }
+        }
+        if (!reduce_axes.empty()) {
+          const auto& runner =
+              NpuOpRunner("ReduceSumD", {tmp_dx}, {*dx},
+                          {{"axes", reduce_axes}, {"keep_dims", false}});
+          runner.Run(stream);
+        }
+      } else {
+        framework::TensorCopy(tmp_dx, ctx.GetPlace(), dev_ctx, dx);
+      }
+
+      if (y_dims != dout_dims) {
+        reduce_axes.clear();
+        int src_axis = (y_dims.size() < dout_dims.size() ? axis : 0);
+        for (int ax = 0; ax < dout_dims.size(); ++ax) {
+          if ((ax < src_axis || ax >= src_axis + y_dims.size()) ||
+              (dout_dims[ax] > 1 && y_dims[ax - src_axis] == 1)) {
+            reduce_axes.push_back(ax);
+          }
+        }
+        if (!reduce_axes.empty()) {
+          const auto& runner =
+              NpuOpRunner("ReduceSumD", {tmp_dy}, {*dy},
+                          {{"axes", reduce_axes}, {"keep_dims", false}});
+          runner.Run(stream);
+        }
+      } else {
+        framework::TensorCopy(tmp_dy, ctx.GetPlace(), dev_ctx, dy);
+      }
+
+    } else if (dx) {
+      Tensor zero_tensor(dout->type());
+      zero_tensor.mutable_data<T>(dout_dims, ctx.GetPlace());
+      FillNpuTensorWithConstant<T>(&zero_tensor, static_cast<T>(0));
+
+      dx->mutable_data<T>(ctx.GetPlace());
+      Tensor tmp_dx;
+      tmp_dx.mutable_data<T>(dout_dims, ctx.GetPlace());
+
+      const auto& runner =
+          NpuOpRunner("MaximumGrad", {*dout, transformed_x, transformed_y},
+                      {tmp_dx, zero_tensor}, attr_input);
+      runner.Run(stream);
+
+      if (x_dims != dout_dims) {
+        reduce_axes.clear();
+
+        int src_axis = (x_dims.size() < dout_dims.size() ? axis : 0);
+        for (int ax = 0; ax < dout_dims.size(); ++ax) {
+          if ((ax < src_axis || ax >= src_axis + x_dims.size()) ||
+              (dout_dims[ax] > 1 && x_dims[ax - src_axis] == 1)) {
+            reduce_axes.push_back(ax);
+          }
+        }
+        if (!reduce_axes.empty()) {
+          const auto& runner =
+              NpuOpRunner("ReduceSumD", {tmp_dx}, {*dx},
+                          {{"axes", reduce_axes}, {"keep_dims", false}});
+          runner.Run(stream);
+        }
+      } else {
+        framework::TensorCopy(tmp_dx, ctx.GetPlace(), dev_ctx, dx);
+      }
+
+    } else if (dy) {
+      Tensor zero_tensor(dout->type());
+      zero_tensor.mutable_data<T>(dout_dims, ctx.GetPlace());
+      FillNpuTensorWithConstant<T>(&zero_tensor, static_cast<T>(0));
+
+      dy->mutable_data<T>(ctx.GetPlace());
+      Tensor tmp_dy;
+      tmp_dy.mutable_data<T>(dout_dims, ctx.GetPlace());
+
+      const auto& runner =
+          NpuOpRunner("MaximumGrad", {*dout, transformed_x, transformed_y},
+                      {zero_tensor, tmp_dy}, attr_input);
+      runner.Run(stream);
+
+      if (y_dims != dout_dims) {
+        reduce_axes.clear();
+
+        int src_axis = (y_dims.size() < dout_dims.size() ? axis : 0);
+        for (int ax = 0; ax < dout_dims.size(); ++ax) {
+          if ((ax < src_axis || ax >= src_axis + y_dims.size()) ||
+              (dout_dims[ax] > 1 && y_dims[ax - src_axis] == 1)) {
+            reduce_axes.push_back(ax);
+          }
+        }
+        if (!reduce_axes.empty()) {
+          const auto& runner =
+              NpuOpRunner("ReduceSumD", {tmp_dy}, {*dy},
+                          {{"axes", reduce_axes}, {"keep_dims", false}});
+          runner.Run(stream);
+        }
+      } else {
+        framework::TensorCopy(tmp_dy, ctx.GetPlace(), dev_ctx, dy);
+      }
+    } else {
+      PADDLE_THROW(platform::errors::Unavailable(
+          "Do not support all outputs to be empty."));
+    }
   }
 };
 
@@ -49,9 +228,19 @@ class ElementwiseMaxNPUKernel : public framework::OpKernel<T> {
 }  // namespace paddle
 
 namespace ops = paddle::operators;
+namespace plat = paddle::platform;
 
 REGISTER_OP_NPU_KERNEL(
     elementwise_max,
-    ops::ElementwiseMaxNPUKernel<paddle::platform::NPUDeviceContext, float>,
-    ops::ElementwiseMaxNPUKernel<paddle::platform::NPUDeviceContext,
-                                 paddle::platform::float16>);
+    ops::ElementwiseMaxNPUKernel<plat::NPUDeviceContext, plat::float16>,
+    ops::ElementwiseMaxNPUKernel<plat::NPUDeviceContext, float>,
+    ops::ElementwiseMaxNPUKernel<plat::NPUDeviceContext, double>,
+    ops::ElementwiseMaxNPUKernel<plat::NPUDeviceContext, int>,
+    ops::ElementwiseMaxNPUKernel<plat::NPUDeviceContext, int64_t>);
+
+REGISTER_OP_NPU_KERNEL(
+    elementwise_max_grad,
+    ops::ElementwiseMaxGradNPUKernel<plat::NPUDeviceContext, plat::float16>,
+    ops::ElementwiseMaxGradNPUKernel<plat::NPUDeviceContext, float>,
+    ops::ElementwiseMaxGradNPUKernel<plat::NPUDeviceContext, double>,
+    ops::ElementwiseMaxGradNPUKernel<plat::NPUDeviceContext, int>);

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

@@ -26,70 +26,248 @@ paddle.enable_static()
 SEED = 2021
 
 
-class TestElementwiseMax(OpTest):
-    def setUp(self):
-        self.set_npu()
-        self.op_type = "elementwise_max"
-        self.place = paddle.NPUPlace(0)
+def ComputeGrad(x, y, out, axis):
+    grad = 1 / out.size
+    shape_x = x.shape
+    shape_y = y.shape
+    shape_out = out.shape
+    reduce_axes_x = []
+    reduce_axes_y = []
+
+    if shape_x != shape_out:
+        if len(shape_x.shape) < len(shape_out.shape):
+            src_axis = axis
+        else:
+            src_axis = 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 = np.maximum(x, y)
+        for ax in range(len(shape_out)):
+            if (ax < src_axis or ax >= src_axis + len(shape_x)) or (
+                    shape_out[ax] > 1 and shape_x[ax - src_axis] == 1):
+                reduce_axes_x.append(ax)
 
-        self.inputs = {
-            'X': OpTest.np_dtype_to_fluid_dtype(x),
-            'Y': OpTest.np_dtype_to_fluid_dtype(y)
-        }
-        self.attrs = {}
-        self.outputs = {'Out': out}
+    if shape_y != shape_out:
+        if len(shape_y) < len(shape_out):
+            src_axis = axis
+        else:
+            src_axis = 0
 
-    def set_npu(self):
-        self.__class__.use_npu = True
+        for ax in range(len(shape_out)):
+            if (ax < src_axis or ax >= src_axis + len(shape_y)) or (
+                    shape_out[ax] > 1 and shape_y[ax - src_axis] == 1):
+                reduce_axes_y.append(ax)
 
-    def init_dtype(self):
-        self.dtype = np.float32
+    if len(reduce_axes_x) > 0:
+        for i in reduce_axes_x:
+            x = np.expand_dims(x, axis=i)
 
-    def test_check_output(self):
-        self.check_output_with_place(self.place)
+    if len(reduce_axes_y) > 0:
+        for i in reduce_axes_y:
+            y = np.expand_dims(y, axis=i)
+
+    mask = np.sign(np.subtract(x, y))
+    dx = np.maximum(mask, 0) * grad
+    dy = np.abs(np.minimum(mask, 0) * grad)
+
+    if len(reduce_axes_x) > 0:
+        for i, element in enumerate(reduce_axes_x):
+            dx = np.add.reduce(dx, element - i)
+
+    if len(reduce_axes_y) > 0:
+        for i, element in enumerate(reduce_axes_y):
+            dy = np.add.reduce(dy, element - i)
 
-    # TODO(ascendrc): Max grad test
-    # def test_check_grad(self):
-    #     if self.dtype == np.float16:
-    #         return
-    #     self.check_grad(['X'], 'Out')
-    #
+    return dx, dy
 
 
-class TestElementwiseMaxFp16(OpTest):
+class TestElementwiseMaxOp(OpTest):
     def setUp(self):
         self.set_npu()
         self.op_type = "elementwise_max"
         self.place = paddle.NPUPlace(0)
 
         self.init_dtype()
-        np.random.seed(SEED)
-        x = np.random.uniform(1, 2, [3, 4]).astype(self.dtype)
-        y = np.random.uniform(1, 2, [3, 4]).astype(self.dtype)
-        out = np.maximum(x, y)
+        self.init_input_output()
+        self.init_axis()
 
         self.inputs = {
-            'X': OpTest.np_dtype_to_fluid_dtype(x),
-            'Y': OpTest.np_dtype_to_fluid_dtype(y)
+            'X': OpTest.np_dtype_to_fluid_dtype(self.x),
+            'Y': OpTest.np_dtype_to_fluid_dtype(self.y)
         }
-        self.attrs = {}
-        self.outputs = {'Out': out}
+        self.attrs = {'axis': self.axis}
+        self.outputs = {'Out': self.out}
 
     def set_npu(self):
         self.__class__.use_npu = True
-        self.__class__.no_need_check_grad = True
 
     def init_dtype(self):
-        self.dtype = np.float16
+        self.dtype = np.float32
+
+    def init_input_output(self):
+        self.x = np.random.uniform(0.1, 1, [13, 17]).astype(self.dtype)
+        sgn = np.random.choice([-1, 1], [13, 17]).astype(self.dtype)
+        self.y = self.x + sgn * np.random.uniform(0.1, 1,
+                                                  [13, 17]).astype(self.dtype)
+        self.out = np.maximum(self.x, self.y)
+
+    def init_axis(self):
+        self.axis = -1
 
     def test_check_output(self):
-        self.check_output_with_place(self.place, atol=1e-5)
+        self.check_output_with_place(self.place)
+
+    def test_check_grad_normal(self):
+        if self.dtype == np.float16:
+            return
+        self.check_grad_with_place(self.place, ['X', 'Y'], 'Out')
+
+    def test_check_grad_ingore_x(self):
+        if self.dtype == np.float16:
+            return
+        self.check_grad_with_place(
+            self.place, ['Y'], 'Out', no_grad_set=set("X"))
+
+    def test_check_grad_ingore_y(self):
+        if self.dtype == np.float16:
+            return
+        self.check_grad_with_place(
+            self.place, ['X'], 'Out', no_grad_set=set("Y"))
+
+
+class TestElementwiseMaxOp_int32(TestElementwiseMaxOp):
+    def init_dtype(self):
+        self.dtype = np.int32
+
+    # CTest does not support check grad for int32.
+    def test_check_grad_normal(self):
+        pass
+
+    def test_check_grad_ingore_x(self):
+        pass
+
+    def test_check_grad_ingore_y(self):
+        pass
+
+
+class TestElementwiseMaxOp_scalar(TestElementwiseMaxOp):
+    def init_input_output(self):
+        self.x = np.random.random_integers(-5, 5, [2, 3, 20]).astype(self.dtype)
+        self.y = np.array([0.5]).astype(self.dtype)
+        self.out = np.maximum(self.x, self.y)
+
+
+class TestElementwiseMaxOp_vector(TestElementwiseMaxOp):
+    def init_input_output(self):
+        self.x = np.random.random((100, )).astype(self.dtype)
+        sgn = np.random.choice([-1, 1], (100, )).astype(self.dtype)
+        self.y = self.x + sgn * np.random.uniform(0.1, 1,
+                                                  (100, )).astype(self.dtype)
+        self.out = np.maximum(self.x, self.y)
+
+
+class TestElementwiseMaxOp_broadcast_0(TestElementwiseMaxOp):
+    def init_input_output(self):
+        self.x = np.random.uniform(0.5, 1, (100, 5, 2)).astype(self.dtype)
+        sgn = np.random.choice([-1, 1], (100, )).astype(self.dtype)
+        self.y = self.x[:, 0, 0] + sgn * \
+            np.random.uniform(1, 2, (100, )).astype(self.dtype)
+        self.out = np.maximum(self.x, self.y.reshape(100, 1, 1))
+
+    def init_axis(self):
+        self.axis = 0
+
+
+class TestElementwiseMaxOp_broadcast_1(TestElementwiseMaxOp):
+    def init_input_output(self):
+        self.x = np.random.uniform(0.5, 1, (2, 100, 3)).astype(self.dtype)
+        sgn = np.random.choice([-1, 1], (100, )).astype(self.dtype)
+        self.y = self.x[0, :, 0] + sgn * \
+            np.random.uniform(1, 2, (100, )).astype(self.dtype)
+        self.out = np.maximum(self.x, self.y.reshape(1, 100, 1))
+
+    def init_axis(self):
+        self.axis = 1
+
+    def test_check_grad_ingore_x(self):
+        _, dy = ComputeGrad(self.x, self.y, self.out, self.axis)
+        self.check_grad_with_place(
+            self.place, ['Y'],
+            'Out',
+            no_grad_set=set("X"),
+            user_defined_grads=[dy])
+
+    def test_check_grad_ingore_y(self):
+        dx, _ = ComputeGrad(self.x, self.y, self.out, self.axis)
+        self.check_grad_with_place(
+            self.place, ['X'],
+            'Out',
+            no_grad_set=set("Y"),
+            user_defined_grads=[dx])
+
+
+class TestElementwiseMaxOp_broadcast_2(TestElementwiseMaxOp):
+    def init_input_output(self):
+        self.x = np.random.uniform(0.5, 1, (2, 3, 100)).astype(self.dtype)
+        sgn = np.random.choice([-1, 1], (100, )).astype(self.dtype)
+        self.y = self.x[0, 0, :] + sgn * \
+            np.random.uniform(1, 2, (100, )).astype(self.dtype)
+        self.out = np.maximum(self.x, self.y.reshape(1, 1, 100))
+
+    def test_check_grad_normal(self):
+        if self.dtype == np.float16:
+            return
+        dx, dy = ComputeGrad(self.x, self.y, self.out, self.axis)
+        self.check_grad_with_place(
+            self.place, ['X', 'Y'], 'Out', user_defined_grads=[dx, dy])
+
+    def test_check_grad_ingore_x(self):
+        if self.dtype == np.float16:
+            return
+        _, dy = ComputeGrad(self.x, self.y, self.out, self.axis)
+        self.check_grad_with_place(
+            self.place, ['Y'],
+            'Out',
+            no_grad_set=set("X"),
+            user_defined_grads=[dy])
+
+    def test_check_grad_ingore_y(self):
+        if self.dtype == np.float16:
+            return
+        dx, _ = ComputeGrad(self.x, self.y, self.out, self.axis)
+        self.check_grad_with_place(
+            self.place, ['X'],
+            'Out',
+            no_grad_set=set("Y"),
+            user_defined_grads=[dx])
+
+
+class TestElementwiseMaxOp_broadcast_3(TestElementwiseMaxOp):
+    def init_input_output(self):
+        self.x = np.random.uniform(0.5, 1, (2, 50, 2, 1)).astype(self.dtype)
+        sgn = np.random.choice([-1, 1], (50, 2)).astype(self.dtype)
+        self.y = self.x[0, :, :, 0] + sgn * \
+            np.random.uniform(1, 2, (50, 2)).astype(self.dtype)
+        self.out = np.maximum(self.x, self.y.reshape(1, 50, 2, 1))
+
+    def init_axis(self):
+        self.axis = 1
+
+
+class TestElementwiseMaxOp_broadcast_4(TestElementwiseMaxOp):
+    def init_input_output(self):
+        self.x = np.random.uniform(0.5, 1, (2, 3, 4, 5)).astype(self.dtype)
+        sgn = np.random.choice([-1, 1], (2, 3, 1, 5)).astype(self.dtype)
+        self.y = self.x + sgn * \
+            np.random.uniform(1, 2, (2, 3, 1, 5)).astype(self.dtype)
+        self.out = np.maximum(self.x, self.y)
+
+
+class TestElementwiseMaxOp_broadcast_5(TestElementwiseMaxOp):
+    def init_input_output(self):
+        self.x = np.random.uniform(0.5, 1, (2, 3, 4, 5)).astype(self.dtype)
+        sgn = np.random.choice([-1, 1], (2, 3, 1, 1)).astype(self.dtype)
+        self.y = self.x + sgn * \
+            np.random.uniform(1, 2, (2, 3, 1, 1)).astype(self.dtype)
+        self.out = np.maximum(self.x, self.y)
 
 
 class TestElementwiseMaxNet(unittest.TestCase):