Add Quantize OP

test=develop

paddle/fluid/operators/quantize_mkldnn_op.cc

@@ -0,0 +1,100 @@
+/* Copyright (c) 2016 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 "mkldnn.hpp"
+#include "paddle/fluid/framework/tensor.h"
+#include "paddle/fluid/operators/quantize_op.h"
+#include "paddle/fluid/platform/mkldnn_helper.h"
+
+namespace paddle {
+namespace operators {
+
+using mkldnn::memory;
+using mkldnn::primitive;
+using mkldnn::reorder;
+using platform::to_void_cast;
+using Tensor = framework::Tensor;
+using framework::DataLayout;
+using mkldnn::stream;
+using platform::GetMKLDNNFormat;
+
+template <typename T>
+class QuantOpKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* input = ctx.Input<Tensor>("Input");
+    auto scale_data = ctx.Attr<float>("Scale");
+    auto* output = ctx.Output<Tensor>("Output");
+    auto& dev_ctx =
+        ctx.template device_context<platform::MKLDNNDeviceContext>();
+    const auto& engine = dev_ctx.GetEngine();
+
+    std::vector<primitive> pipeline;
+    std::vector<int> src_tz = paddle::framework::vectorize2int(input->dims());
+    std::vector<int> dst_tz = paddle::framework::vectorize2int(output->dims());
+
+    const T* input_data = input->data<T>();
+
+    mkldnn::primitive_attr attri;
+    int mask = 0;
+    attri.set_output_scales(mask, {scale_data});
+
+    auto src_md = platform::MKLDNNMemDesc({src_tz}, memory::data_type::f32,
+                                          input->format());
+    auto src_pd = mkldnn::memory::primitive_desc(src_md, engine);
+    auto src_memory =
+        std::make_shared<mkldnn::memory>(src_pd, to_void_cast<T>(input_data));
+    std::shared_ptr<primitive::at> src_memory_p =
+        std::shared_ptr<primitive::at>(new primitive::at(*src_memory));
+
+    bool is_negative = ctx.Attr<bool>("is_negative_input");
+    mkldnn::memory::primitive_desc dst_pd;
+    std::shared_ptr<mkldnn::memory> dst_memory;
+    if (is_negative) {
+      int8_t* output_data = output->mutable_data<int8_t>(ctx.GetPlace());
+      auto dst_md = platform::MKLDNNMemDesc({dst_tz}, memory::data_type::s8,
+                                            memory::format::nhwc);
+      dst_pd = mkldnn::memory::primitive_desc(dst_md, engine);
+      dst_memory.reset(
+          new mkldnn::memory(dst_pd, to_void_cast<int8_t>(output_data)));
+    } else {
+      uint8_t* output_data = output->mutable_data<uint8_t>(ctx.GetPlace());
+      auto dst_md = platform::MKLDNNMemDesc({dst_tz}, memory::data_type::u8,
+                                            memory::format::nhwc);
+      dst_pd = mkldnn::memory::primitive_desc(dst_md, engine);
+      dst_memory.reset(
+          new mkldnn::memory(dst_pd, to_void_cast<uint8_t>(output_data)));
+    }
+
+    auto reorder_pd = std::shared_ptr<reorder::primitive_desc>(
+        new reorder::primitive_desc(src_pd, dst_pd, attri));
+    auto reorder_p = std::shared_ptr<reorder>(
+        new reorder(*reorder_pd, *src_memory_p, *dst_memory));
+
+    pipeline.push_back(*reorder_p);
+    stream(stream::kind::eager).submit(pipeline).wait();
+
+    output->set_layout(DataLayout::kMKLDNN);
+    output->set_format(GetMKLDNNFormat(*dst_memory));
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+namespace ops = paddle::operators;
+
+// TODO(Xiaoli) Support FP32->S8 quantization.
+
+REGISTER_OP_KERNEL(quantize, MKLDNN, ::paddle::platform::CPUPlace,
+                   ops::QuantOpKernel<float>);

paddle/fluid/operators/quantize_op.cc

@@ -0,0 +1,47 @@
+/* Copyright (c) 2016 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/operators/quantize_op.h"
+#ifdef PADDLE_WITH_MKLDNN
+#include "paddle/fluid/platform/mkldnn_helper.h"
+#endif
+
+namespace paddle {
+namespace operators {
+
+framework::OpKernelType QuantOp::GetExpectedKernelType(
+    const framework::ExecutionContext& ctx) const {
+  framework::LibraryType library_ = framework::LibraryType::kMKLDNN;
+  framework::DataLayout layout_ = framework::DataLayout::kMKLDNN;
+
+  return framework::OpKernelType(ctx.Input<Tensor>("Input")->type(),
+                                 ctx.GetPlace(), layout_, library_);
+}
+
+void QuantOpMaker::Make() {
+  AddInput("Input", "input data");
+  AddOutput("Output", "output data");
+  AddAttr<bool>("is_negative_input",
+                "(bool, default false) Only used in mkldnn INT8 kernel")
+      .SetDefault(false);
+  AddAttr<float>("Scale", "scale data").SetDefault({1.0f});
+  AddComment(R"DOC(This op will quantize data from FP32 to INT8)DOC");
+}
+
+}  // namespace operators
+}  // namespace paddle
+namespace ops = paddle::operators;
+
+REGISTER_OPERATOR(quantize, ops::QuantOp, ops::QuantOpMaker,
+                  paddle::framework::DefaultGradOpDescMaker<true>);

paddle/fluid/operators/quantize_op.h

@@ -0,0 +1,46 @@
+/* Copyright (c) 2016 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. */
+
+#pragma once
+
+#include <string>
+#include <vector>
+#include "paddle/fluid/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+using framework::OpKernelType;
+using framework::Tensor;
+
+class QuantOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    ctx->SetOutputDim("Output", ctx->GetInputDim("Input"));
+    ctx->ShareLoD("Input", /*->*/ "Output");
+  }
+
+ protected:
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override;
+};
+
+class QuantOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() override;
+};
+}  // namespace operators
+}  // namespace paddle

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

@@ -0,0 +1,76 @@
+#   Copyright (c) 2018 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
+from op_test import OpTest
+
+
+class TestQuantizeOp(OpTest):
+    def setUp(self):
+        self.op_type = 'quantize'
+        self.scale = 2.0
+        self.input_size = [1, 1, 5, 5]  #Naive nChw16c
+        self.is_negative = False
+        self.set_scale()
+        self.set_is_negative()
+
+        if self.is_negative:
+            input = (100 * np.random.random_sample(self.input_size) - 50
+                     ).astype('float32')
+            output = np.round(input * self.scale).astype('int8')
+        else:
+            input = (100 *
+                     np.random.random_sample(self.input_size)).astype('float32')
+            output = np.round(input * self.scale).astype('uint8')
+
+        self.inputs = {'Input': OpTest.np_dtype_to_fluid_dtype(input)}
+
+        self.outputs = {'Output': output}
+
+        self.attrs = {
+            'Scale': self.scale,
+            'is_negative_input': self.is_negative
+        }
+
+    def test_check_output(self):
+        self.check_output()
+
+    def set_scale(self):
+        pass
+
+    def set_is_negative(self):
+        pass
+
+
+class TestQuantizeOp1(TestQuantizeOp):
+    def set_scale(self):
+        self.scale = 1.5
+
+    def set_is_negative(self):
+        self.is_nagative = True
+
+
+class TestQuantizeOp2(TestQuantizeOp):
+    def set_scale(self):
+        self.scale = 0.1
+
+    def set_is_negative(self):
+        self.is_nagative = False
+
+
+if __name__ == '__main__':
+    unittest.main()