add reduce_prod_xpu. fix reduce_mean_xpu bug. (#38481)

* add reduce_prod_xpu. fix reduce_mean_xpu bug.

* iadd reduce_prod_xpu. fix reduce_mean_xpu bug. test=kunlun

paddle/fluid/operators/reduce_ops/reduce_mean_op_xpu.cc

@@ -14,11 +14,12 @@
 
 #ifdef PADDLE_WITH_XPU
 
-#include "paddle/fluid/operators/reduce_ops/reduce_mean_op.h"
 #include <memory>
 #include <string>
 #include <vector>
 
+#include "paddle/fluid/operators/reduce_ops/reduce_mean_op.h"
+
 namespace paddle {
 namespace operators {
 template <typename DeviceContext, typename T>
@@ -41,15 +42,24 @@ class ReduceMeanXPUKernel : public framework::OpKernel<T> {
       xdims.push_back(input->dims()[i]);
     }
     auto rdims = context.Attr<std::vector<int>>("dim");
+    const auto& input_dim_size = input->dims().size();
+    std::vector<int> reduce_dims;
     if (reduce_all) {
-      rdims.clear();
       for (size_t i = 0; i < xdims.size(); i++) {
-        rdims.push_back(static_cast<int>(i));
+        reduce_dims.push_back(static_cast<int>(i));
+      }
+    } else {
+      for (size_t i = 0; i < rdims.size(); ++i) {
+        if (rdims[i] < 0) {
+          reduce_dims.push_back(rdims[i] + input_dim_size);
+        } else {
+          reduce_dims.push_back(rdims[i]);
+        }
       }
     }
     int r = xpu::reduce_mean(
         dev_ctx.x_context(), reinterpret_cast<const XPUType*>(input->data<T>()),
-        reinterpret_cast<XPUType*>(output->data<T>()), xdims, rdims);
+        reinterpret_cast<XPUType*>(output->data<T>()), xdims, reduce_dims);
 
     PADDLE_ENFORCE_EQ(r, XPU_SUCCESS,
                       platform::errors::External(

paddle/fluid/operators/reduce_ops/reduce_prod_op_xpu.cc

+// Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// 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.
+
+#ifdef PADDLE_WITH_XPU
+
+#include <memory>
+#include <vector>
+
+#include "paddle/fluid/operators/reduce_ops/reduce_prod_op.h"
+
+namespace paddle {
+namespace operators {
+template <typename DeviceContext, typename T>
+class ReduceProdXPUKernel : public framework::OpKernel<T> {
+  using XPUType = typename XPUTypeTrait<T>::Type;
+
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    PADDLE_ENFORCE_EQ(
+        platform::is_xpu_place(context.GetPlace()), true,
+        platform::errors::Unavailable("This kernel only runs on XPU."));
+    bool reduce_all = context.Attr<bool>("reduce_all");
+    auto* input = context.Input<Tensor>("X");
+    auto* output = context.Output<Tensor>("Out");
+    output->mutable_data<T>(context.GetPlace());
+    auto& dev_ctx = context.template device_context<DeviceContext>();
+
+    std::vector<int> xdims;
+    for (int i = 0; i < input->dims().size(); i++) {
+      xdims.push_back(input->dims()[i]);
+    }
+    auto rdims = context.Attr<std::vector<int>>("dim");
+    const auto& input_dim_size = input->dims().size();
+
+    std::vector<int> reduce_dims;
+    if (reduce_all) {
+      for (size_t i = 0; i < xdims.size(); i++) {
+        reduce_dims.push_back(static_cast<int>(i));
+      }
+    } else {
+      for (size_t i = 0; i < rdims.size(); ++i) {
+        if (rdims[i] < 0) {
+          reduce_dims.push_back(rdims[i] + input_dim_size);
+        } else {
+          reduce_dims.push_back(rdims[i]);
+        }
+      }
+    }
+    int r = xpu::reduce_prod(
+        dev_ctx.x_context(), reinterpret_cast<const XPUType*>(input->data<T>()),
+        reinterpret_cast<XPUType*>(output->data<T>()), xdims, reduce_dims);
+
+    PADDLE_ENFORCE_EQ(r, XPU_SUCCESS,
+                      platform::errors::External(
+                          "XPU reduce_prod kernel return wrong value[%d %s]", r,
+                          XPUAPIErrorMsg[r]));
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+REGISTER_OP_XPU_KERNEL(
+    reduce_prod,
+    ops::ReduceProdXPUKernel<paddle::platform::XPUDeviceContext, float>);
+
+#endif

paddle/fluid/platform/device/xpu/xpu1_op_list.h

@@ -219,6 +219,7 @@ XPUOpMap& get_kl1_ops() {
       {"reduce_mean", XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace())})},
       {"reduce_mean_grad",
        XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace())})},
+      {"reduce_prod", XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace())})},
       {"reduce_sum_grad",
        XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace())})},
       {"reduce_sum", XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace())})},

paddle/fluid/platform/device/xpu/xpu2_op_list.h

@@ -242,6 +242,7 @@ XPUOpMap& get_kl2_ops() {
       {"reduce_mean_grad",
        XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace())})},
       {"reduce_mean", XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace())})},
+      {"reduce_prod", XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace())})},
       {"reduce_sum_grad",
        XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace())})},
       {"reduce_sum", XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace())})},

python/paddle/fluid/tests/unittests/xpu/test_reduce_mean_op_xpu.py

@@ -29,7 +29,7 @@ from paddle.fluid.framework import convert_np_dtype_to_dtype_
 class TestMeanOp(OpTest):
     def setUp(self):
         self.op_type = "reduce_mean"
-        self.inputs = {'X': np.random.random((5, 6, 10)).astype("float64")}
+        self.inputs = {'X': np.random.random((5, 6, 10)).astype("float32")}
         self.attrs = {'use_xpu': True}
         self.outputs = {'Out': self.inputs['X'].mean(axis=0)}
 
@@ -46,7 +46,7 @@ class TestMeanOp5D(OpTest):
     def setUp(self):
         self.op_type = "reduce_mean"
         self.inputs = {
-            'X': np.random.random((1, 2, 5, 6, 10)).astype("float64")
+            'X': np.random.random((1, 2, 5, 6, 10)).astype("float32")
         }
         self.attrs = {'use_xpu': True}
         self.outputs = {'Out': self.inputs['X'].mean(axis=0)}
@@ -64,7 +64,7 @@ class TestMeanOp6D(OpTest):
     def setUp(self):
         self.op_type = "reduce_mean"
         self.inputs = {
-            'X': np.random.random((1, 1, 2, 5, 6, 10)).astype("float64")
+            'X': np.random.random((1, 1, 2, 5, 6, 10)).astype("float32")
         }
         self.attrs = {'use_xpu': True}
         self.outputs = {'Out': self.inputs['X'].mean(axis=0)}
@@ -82,7 +82,7 @@ class TestMeanOp8D(OpTest):
     def setUp(self):
         self.op_type = "reduce_mean"
         self.inputs = {
-            'X': np.random.random((1, 3, 1, 2, 1, 4, 3, 10)).astype("float64")
+            'X': np.random.random((1, 3, 1, 2, 1, 4, 3, 10)).astype("float32")
         }
         self.attrs = {'dim': (0, 3), 'use_xpu': True}
         self.outputs = {'Out': self.inputs['X'].mean(axis=(0, 3))}
@@ -99,7 +99,7 @@ class TestMeanOp8D(OpTest):
 class Test1DReduce(OpTest):
     def setUp(self):
         self.op_type = "reduce_mean"
-        self.inputs = {'X': np.random.random(120).astype("float64")}
+        self.inputs = {'X': np.random.random(120).astype("float32")}
         self.attrs = {'use_xpu': True}
         self.outputs = {'Out': self.inputs['X'].mean(axis=0)}
 
@@ -116,7 +116,7 @@ class Test2DReduce0(Test1DReduce):
     def setUp(self):
         self.op_type = "reduce_mean"
         self.attrs = {'dim': [0], 'use_xpu': True}
-        self.inputs = {'X': np.random.random((20, 10)).astype("float64")}
+        self.inputs = {'X': np.random.random((20, 10)).astype("float32")}
         self.outputs = {'Out': self.inputs['X'].mean(axis=0)}
 
 
@@ -124,7 +124,7 @@ class Test2DReduce1(Test1DReduce):
     def setUp(self):
         self.op_type = "reduce_mean"
         self.attrs = {'dim': [1], 'use_xpu': True}
-        self.inputs = {'X': np.random.random((20, 10)).astype("float64")}
+        self.inputs = {'X': np.random.random((20, 10)).astype("float32")}
         self.outputs = {
             'Out': self.inputs['X'].mean(axis=tuple(self.attrs['dim']))
         }
@@ -134,7 +134,7 @@ class Test3DReduce0(Test1DReduce):
     def setUp(self):
         self.op_type = "reduce_mean"
         self.attrs = {'dim': [1], 'use_xpu': True}
-        self.inputs = {'X': np.random.random((5, 6, 7)).astype("float64")}
+        self.inputs = {'X': np.random.random((5, 6, 7)).astype("float32")}
         self.outputs = {
             'Out': self.inputs['X'].mean(axis=tuple(self.attrs['dim']))
         }
@@ -144,7 +144,7 @@ class Test3DReduce1(Test1DReduce):
     def setUp(self):
         self.op_type = "reduce_mean"
         self.attrs = {'dim': [2], 'use_xpu': True}
-        self.inputs = {'X': np.random.random((5, 6, 7)).astype("float64")}
+        self.inputs = {'X': np.random.random((5, 6, 7)).astype("float32")}
         self.outputs = {
             'Out': self.inputs['X'].mean(axis=tuple(self.attrs['dim']))
         }
@@ -154,7 +154,7 @@ class Test3DReduce2(Test1DReduce):
     def setUp(self):
         self.op_type = "reduce_mean"
         self.attrs = {'dim': [-2], 'use_xpu': True}
-        self.inputs = {'X': np.random.random((5, 6, 7)).astype("float64")}
+        self.inputs = {'X': np.random.random((5, 6, 7)).astype("float32")}
         self.outputs = {
             'Out': self.inputs['X'].mean(axis=tuple(self.attrs['dim']))
         }
@@ -164,7 +164,7 @@ class Test3DReduce3(Test1DReduce):
     def setUp(self):
         self.op_type = "reduce_mean"
         self.attrs = {'dim': [1, 2], 'use_xpu': True}
-        self.inputs = {'X': np.random.random((5, 6, 7)).astype("float64")}
+        self.inputs = {'X': np.random.random((5, 6, 7)).astype("float32")}
         self.outputs = {
             'Out': self.inputs['X'].mean(axis=tuple(self.attrs['dim']))
         }
@@ -173,7 +173,7 @@ class Test3DReduce3(Test1DReduce):
 class TestKeepDimReduce(Test1DReduce):
     def setUp(self):
         self.op_type = "reduce_mean"
-        self.inputs = {'X': np.random.random((5, 6, 10)).astype("float64")}
+        self.inputs = {'X': np.random.random((5, 6, 10)).astype("float32")}
         self.attrs = {'dim': [1], 'keep_dim': True, 'use_xpu': True}
         self.outputs = {
             'Out': self.inputs['X'].mean(
@@ -185,7 +185,7 @@ class TestKeepDim8DReduce(Test1DReduce):
     def setUp(self):
         self.op_type = "reduce_mean"
         self.inputs = {
-            'X': np.random.random((2, 5, 3, 2, 2, 3, 4, 2)).astype("float64")
+            'X': np.random.random((2, 5, 3, 2, 2, 3, 4, 2)).astype("float32")
         }
         self.attrs = {'dim': (3, 4, 5), 'keep_dim': True, 'use_xpu': True}
         self.outputs = {
@@ -197,7 +197,7 @@ class TestKeepDim8DReduce(Test1DReduce):
 class TestReduceAll(Test1DReduce):
     def setUp(self):
         self.op_type = "reduce_mean"
-        self.inputs = {'X': np.random.random((5, 6, 2, 10)).astype("float64")}
+        self.inputs = {'X': np.random.random((5, 6, 2, 10)).astype("float32")}
         self.attrs = {'reduce_all': True, 'use_xpu': True}
         self.outputs = {'Out': self.inputs['X'].mean()}
 

python/paddle/fluid/tests/unittests/xpu/test_reduce_prod_op_xpu.py

+#   Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# 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.
+
+from __future__ import print_function
+
+import unittest
+import numpy as np
+import sys
+sys.path.append("..")
+from op_test_xpu import OpTest, XPUOpTest
+from op_test import skip_check_grad_ci
+import paddle
+import paddle.fluid.core as core
+import paddle.fluid as fluid
+from paddle.fluid import compiler, Program, program_guard
+from paddle.fluid.framework import convert_np_dtype_to_dtype_
+
+
+class TestXPUReduceProdOp(XPUOpTest):
+    def setUp(self):
+        self.init_op_type()
+        self.initTestCase()
+        self.use_xpu = True
+        self.use_mkldnn = False
+        self.attrs = {
+            'dim': self.axis,
+            'keep_dim': self.keep_dim,
+            'reduce_all': self.reduce_all
+        }
+        self.inputs = {'X': np.random.random(self.shape).astype("float32")}
+        if self.attrs['reduce_all']:
+            self.outputs = {'Out': self.inputs['X'].prod()}
+        else:
+            self.outputs = {
+                'Out': self.inputs['X'].prod(
+                    axis=self.axis, keepdims=self.attrs['keep_dim'])
+            }
+
+    def test_check_output(self):
+        if paddle.is_compiled_with_xpu():
+            paddle.enable_static()
+            place = paddle.XPUPlace(0)
+            self.check_output_with_place(place)
+
+    def test_check_grad(self):
+        if paddle.is_compiled_with_xpu():
+            paddle.enable_static()
+            place = paddle.XPUPlace(0)
+            self.check_grad_with_place(place, ['X'], 'Out')
+
+    def init_op_type(self):
+        self.op_type = "reduce_prod"
+        self.use_mkldnn = False
+        self.keep_dim = False
+        self.reduce_all = False
+
+    def initTestCase(self):
+        self.shape = (5, 6, 10)
+        self.axis = (0, )
+
+
+class TestProdOp5D(TestXPUReduceProdOp):
+    def initTestCase(self):
+        self.shape = (1, 2, 5, 6, 10)
+        self.axis = (0, )
+
+
+class TestProdOp6D(TestXPUReduceProdOp):
+    def initTestCase(self):
+        self.shape = (1, 1, 2, 5, 6, 10)
+        self.axis = (0, )
+
+
+class TestProdOp8D(TestXPUReduceProdOp):
+    def initTestCase(self):
+        self.shape = (1, 3, 1, 2, 1, 4, 3, 10)
+        self.axis = (0, 3)
+
+
+class Test1DReduce(TestXPUReduceProdOp):
+    def initTestCase(self):
+        self.shape = 120
+        self.axis = (0, )
+
+
+class Test2DReduce0(TestXPUReduceProdOp):
+    def initTestCase(self):
+        self.shape = (20, 10)
+        self.axis = (0, )
+
+
+class Test2DReduce1(TestXPUReduceProdOp):
+    def initTestCase(self):
+        self.shape = (20, 10)
+        self.axis = (1, )
+
+
+class Test3DReduce0(TestXPUReduceProdOp):
+    def initTestCase(self):
+        self.shape = (5, 6, 7)
+        self.axis = (1, )
+
+
+class Test3DReduce1(TestXPUReduceProdOp):
+    def initTestCase(self):
+        self.shape = (5, 6, 7)
+        self.axis = (2, )
+
+
+class Test3DReduce2(TestXPUReduceProdOp):
+    def initTestCase(self):
+        self.shape = (5, 6, 7)
+        self.axis = (-2, )
+
+
+class Test3DReduce3(TestXPUReduceProdOp):
+    def initTestCase(self):
+        self.shape = (5, 6, 7)
+        self.axis = (1, 2)
+
+
+class TestKeepDimReduce(TestXPUReduceProdOp):
+    def initTestCase(self):
+        self.shape = (5, 6, 10)
+        self.axis = (1, )
+        self.keep_dim = True
+
+
+class TestKeepDim8DReduce(TestXPUReduceProdOp):
+    def initTestCase(self):
+        self.shape = (2, 5, 3, 2, 2, 3, 4, 2)
+        self.axis = (3, 4, 5)
+        self.keep_dim = True
+
+
+class TestReduceAll(TestXPUReduceProdOp):
+    def initTestCase(self):
+        self.shape = (5, 6, 2, 10)
+        self.axis = (0, )
+        self.reduce_all = True
+
+
+if __name__ == '__main__':
+    unittest.main()