[MLU] add mlu kernel for clip (#43229)

paddle/fluid/operators/clip_op_mlu.cc

+/* Copyright (c) 2022 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. */
+
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/operators/mlu/mlu_baseop.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename T>
+class ClipMLUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* x = ctx.Input<Tensor>("X");
+    auto* out = ctx.Output<Tensor>("Out");
+
+    auto min = static_cast<T>(ctx.Attr<float>("min"));
+    auto max = static_cast<T>(ctx.Attr<float>("max"));
+
+    if (ctx.HasInput("Min")) {
+      Tensor min_cpu;
+      auto* min_tensor = ctx.Input<Tensor>("Min");
+      auto* min_data = min_tensor->data<T>();
+      if (platform::is_mlu_place(min_tensor->place())) {
+        paddle::framework::TensorCopySync(*min_tensor, platform::CPUPlace(),
+                                          &min_cpu);
+        min_data = min_cpu.data<T>();
+      }
+      min = min_data[0];
+    }
+
+    if (ctx.HasInput("Max")) {
+      Tensor max_cpu;
+      auto* max_tensor = ctx.Input<Tensor>("Max");
+      auto* max_data = max_tensor->data<T>();
+      if (platform::is_mlu_place(max_tensor->place())) {
+        paddle::framework::TensorCopySync(*max_tensor, platform::CPUPlace(),
+                                          &max_cpu);
+        max_data = max_cpu.data<T>();
+      }
+      max = max_data[0];
+    }
+    out->mutable_data<T>(ctx.GetPlace());
+
+    MLUCnnlTensorDesc x_desc(*x);
+    MLUCnnlTensorDesc out_desc(*out);
+    MLUCnnl::Clip(ctx, x_desc.get(), GetBasePtr(x),
+                  static_cast<const void*>(&min),
+                  static_cast<const void*>(&max), GetBasePtr(out));
+  }
+};
+
+template <typename T>
+class ClipGradMLUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* x = ctx.Input<Tensor>("X");
+    auto* dout = ctx.Input<Tensor>(framework::GradVarName("Out"));
+    auto* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
+    dx->mutable_data<T>(ctx.GetPlace());
+
+    auto* min_tensor = ctx.HasInput("Min") ? ctx.Input<Tensor>("Min") : nullptr;
+    auto* max_tensor = ctx.HasInput("Max") ? ctx.Input<Tensor>("Max") : nullptr;
+
+    auto min_val = ctx.Attr<float>("min");
+    if (min_tensor) {
+      Tensor min_data;
+      framework::TensorCopy(
+          *min_tensor, platform::CPUPlace(),
+          ctx.template device_context<platform::DeviceContext>(), &min_data);
+      ctx.template device_context<paddle::platform::MLUDeviceContext>().Wait();
+      min_val = static_cast<float>(min_data.data<T>()[0]);
+    }
+    auto max_val = ctx.Attr<float>("max");
+    if (max_tensor) {
+      Tensor max_data;
+      framework::TensorCopy(
+          *max_tensor, platform::CPUPlace(),
+          ctx.template device_context<platform::DeviceContext>(), &max_data);
+      ctx.template device_context<paddle::platform::MLUDeviceContext>().Wait();
+      max_val = static_cast<float>(max_data.data<T>()[0]);
+    }
+
+    MLUCnnlTensorDesc x_desc(*x);
+    MLUCnnlTensorDesc dx_desc(*dx);
+    MLUCnnlTensorDesc dout_desc(*dout);
+
+    MLUCnnl::HardtanhBackward(ctx, x_desc.get(), GetBasePtr(x), dout_desc.get(),
+                              GetBasePtr(dout), max_val, min_val, dx_desc.get(),
+                              GetBasePtr(dx));
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+namespace plat = paddle::platform;
+
+REGISTER_OP_MLU_KERNEL(clip, ops::ClipMLUKernel<float>,
+                       ops::ClipMLUKernel<plat::float16>);
+
+REGISTER_OP_MLU_KERNEL(clip_grad, ops::ClipGradMLUKernel<float>,
+                       ops::ClipGradMLUKernel<plat::float16>);

paddle/fluid/operators/mlu/mlu_baseop.cc

@@ -1942,6 +1942,25 @@ MLUCnnlTrigonDesc::~MLUCnnlTrigonDesc() {
                                               cast_type, output_desc, output));
 }
 
+/*static*/ void MLUCnnl::Clip(const ExecutionContext& ctx,
+                              const cnnlTensorDescriptor_t x_desc,
+                              const void* x, const void* min, const void* max,
+                              void* y) {
+  cnnlHandle_t handle = GetHandleFromCTX(ctx);
+  PADDLE_ENFORCE_MLU_SUCCESS(cnnlClip(handle, x_desc, x, min, max, y));
+}
+
+/*static*/ void MLUCnnl::HardtanhBackward(
+    const ExecutionContext& ctx, const cnnlTensorDescriptor_t x_desc,
+    const void* x, const cnnlTensorDescriptor_t diff_y_desc, const void* diff_y,
+    const float max_val, const float min_val,
+    const cnnlTensorDescriptor_t diff_x_desc, void* diff_x) {
+  cnnlHandle_t handle = GetHandleFromCTX(ctx);
+  PADDLE_ENFORCE_MLU_SUCCESS(
+      cnnlHardtanhBackward(handle, x_desc, x, diff_y_desc, diff_y, max_val,
+                           min_val, diff_x_desc, diff_x));
+}
+
 /* static */ void MLUCnnl::PoolingBackward(
     const ExecutionContext& ctx, const cnnlPoolingDescriptor_t pooling_desc,
     const void* alpha, const cnnlTensorDescriptor_t y_desc, const void* y,

paddle/fluid/operators/mlu/mlu_baseop.h

@@ -439,6 +439,16 @@ class MLUCnnl {
                    const cnnlTensorDescriptor_t input_desc, const void* input,
                    const cnnlTensorDescriptor_t output_desc, void* output);
 
+  static void Clip(const ExecutionContext& ctx,
+                   const cnnlTensorDescriptor_t input_desc, const void* input,
+                   const void* min, const void* max, void* y);
+
+  static void HardtanhBackward(
+      const ExecutionContext& ctx, const cnnlTensorDescriptor_t x_desc,
+      const void* x, const cnnlTensorDescriptor_t diff_y_desc,
+      const void* diff_y, const float max_val, const float min_val,
+      const cnnlTensorDescriptor_t diff_x_desc, void* diff_x);
+
   static void Div(const ExecutionContext& ctx,
                   cnnlComputationPreference_t prefer,
                   const cnnlTensorDescriptor_t in0_desc, const void* in0,

python/paddle/fluid/tests/unittests/mlu/test_clip_op_mlu.py

+#  Copyright (c) 2022 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 sys
+
+sys.path.append("..")
+import numpy as np
+import paddle
+import paddle.fluid as fluid
+from paddle.fluid import Program, program_guard
+from op_test import OpTest
+from paddle.fluid.framework import _test_eager_guard
+
+paddle.enable_static()
+
+
+class TestClipOp(OpTest):
+
+    def setUp(self):
+        self.place = paddle.device.MLUPlace(0)
+        self.__class__.use_mlu = True
+        self.max_relative_error = 0.006
+        self.python_api = paddle.clip
+
+        self.inputs = {}
+        self.initTestCase()
+
+        self.op_type = "clip"
+        self.attrs = {}
+        self.attrs['min'] = self.min
+        self.attrs['max'] = self.max
+        if 'Min' in self.inputs:
+            min_v = self.inputs['Min']
+        else:
+            min_v = self.attrs['min']
+
+        if 'Max' in self.inputs:
+            max_v = self.inputs['Max']
+        else:
+            max_v = self.attrs['max']
+
+        input = np.random.random(self.shape).astype(self.dtype)
+        input[np.abs(input - min_v) < self.max_relative_error] = 0.5
+        input[np.abs(input - max_v) < self.max_relative_error] = 0.5
+        self.inputs['X'] = input
+        self.outputs = {'Out': np.clip(self.inputs['X'], min_v, max_v)}
+
+    def test_check_output(self):
+        self.check_output_with_place(self.place)
+
+    def test_check_grad_normal(self):
+        self.check_grad_with_place(self.place, ['X'], 'Out')
+
+    def initTestCase(self):
+        self.dtype = np.float32
+        self.shape = (4, 10, 10)
+        self.max = 0.8
+        self.min = 0.3
+        self.inputs['Max'] = np.array([0.8]).astype(self.dtype)
+        self.inputs['Min'] = np.array([0.1]).astype(self.dtype)
+
+
+class TestCase1(TestClipOp):
+
+    def initTestCase(self):
+        self.dtype = np.float32
+        self.shape = (8, 16, 8)
+        self.max = 0.7
+        self.min = 0.0
+
+
+class TestCase2(TestClipOp):
+
+    def initTestCase(self):
+        self.dtype = np.float32
+        self.shape = (8, 16)
+        self.max = 1.0
+        self.min = 0.0
+
+
+class TestCase3(TestClipOp):
+
+    def initTestCase(self):
+        self.dtype = np.float32
+        self.shape = (4, 8, 16)
+        self.max = 0.7
+        self.min = 0.2
+
+
+class TestCase4(TestClipOp):
+
+    def initTestCase(self):
+        self.dtype = np.float32
+        self.shape = (4, 8, 8)
+        self.max = 0.7
+        self.min = 0.2
+        self.inputs['Max'] = np.array([0.8]).astype(self.dtype)
+        self.inputs['Min'] = np.array([0.3]).astype(self.dtype)
+
+
+class TestCase5(TestClipOp):
+
+    def initTestCase(self):
+        self.dtype = np.float32
+        self.shape = (4, 8, 16)
+        self.max = 0.5
+        self.min = 0.5
+
+
+class TestCase6(TestClipOp):
+
+    def initTestCase(self):
+        self.dtype = np.float16
+        self.shape = (4, 8, 8)
+        self.max = 0.7
+        self.min = 0.2
+        self.inputs['Max'] = np.array([0.8]).astype(self.dtype)
+        self.inputs['Min'] = np.array([0.3]).astype(self.dtype)
+
+
+if __name__ == '__main__':
+    unittest.main()