Add Requantize OP (#15318)

* Enable INT8 ReQuantize OP
test=develop

* Clean code
test=develop

* Add comments
test=develop

* Revert "Clean code"
test=develop

This reverts commit a7a49b8.

* Modify requantize op test
test=develop

* fix requantize UT by moving public function to public test file.
test=develop

* Fix test fail due to file address change.
test=develop

* Change file address for requantize op.
test=develop

Author: xiaolil1

paddle/fluid/operators/mkldnn/requantize_mkldnn_op.cc

@@ -0,0 +1,94 @@
+/* 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/data_layout_transform.h"
+#include "paddle/fluid/framework/tensor.h"
+#include "paddle/fluid/operators/requantize_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 ReQuantOpKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* input = ctx.Input<Tensor>("Input");
+    auto scale_in = ctx.Attr<float>("Scale_in");
+    auto scale_out = ctx.Attr<float>("Scale_out");
+    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());
+    mkldnn::memory::data_type src_dt =
+        paddle::framework::ToMKLDNNDataType(input->type());
+    mkldnn::memory::data_type dst_dt = src_dt;  // TODO(Xiaoli) support
+                                                // requantize from different
+                                                // data type (e.g., s8 to u8)
+    mkldnn::memory::format src_fmt = memory::format::nhwc;
+    mkldnn::memory::format dst_fmt = memory::format::nhwc;
+
+    const T* input_data = input->data<T>();
+    T* output_data = output->mutable_data<T>(ctx.GetPlace());
+    float scale_shift = scale_out / scale_in;
+
+    mkldnn::primitive_attr attri;
+    int mask = 0;
+    attri.set_output_scales(mask, {scale_shift});
+
+    auto src_md = platform::MKLDNNMemDesc({src_tz}, src_dt, src_fmt);
+    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));
+
+    auto dst_md = platform::MKLDNNMemDesc({dst_tz}, dst_dt, dst_fmt);
+    auto dst_pd = mkldnn::memory::primitive_desc(dst_md, engine);
+    auto dst_memory = mkldnn::memory(dst_pd, to_void_cast<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;
+
+REGISTER_OP_KERNEL(requantize, MKLDNN, ::paddle::platform::CPUPlace,
+                   ops::ReQuantOpKernel<int8_t>, ops::ReQuantOpKernel<uint8_t>);

paddle/fluid/operators/requantize_op.cc

@@ -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. */
+
+#include "paddle/fluid/operators/requantize_op.h"
+#ifdef PADDLE_WITH_MKLDNN
+#include "paddle/fluid/platform/mkldnn_helper.h"
+#endif
+
+namespace paddle {
+namespace operators {
+
+framework::OpKernelType ReQuantOp::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 ReQuantOpMaker::Make() {
+  AddInput("Input", "input data");
+  AddOutput("Output", "output data");
+  AddAttr<float>("Scale_in", "scale in data").SetDefault({1.0f});
+  AddAttr<float>("Scale_out", "scale out data").SetDefault({1.0f});
+  AddComment(
+      R"DOC(This op will re-quantize data from INT8 with scale_in to INT8 with scale_out)DOC");
+}
+
+}  // namespace operators
+}  // namespace paddle
+namespace ops = paddle::operators;
+
+REGISTER_OPERATOR(requantize, ops::ReQuantOp, ops::ReQuantOpMaker,
+                  paddle::framework::DefaultGradOpDescMaker<true>);

paddle/fluid/operators/requantize_op.h

@@ -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. */
+
+#pragma once
+
+#include <string>
+#include <vector>
+#include "paddle/fluid/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+using framework::OpKernelType;
+using framework::Tensor;
+
+class ReQuantOp : 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 ReQuantOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() override;
+};
+
+}  // namespace operators
+}  // namespace paddle

python/paddle/fluid/tests/unittests/mkldnn/mkldnn_op_test.py

@@ -70,3 +70,17 @@ def __assert_close(tensor, np_array, msg, atol=1e-4):
                 fetch_list=['x@GRAD', 'out'])
 
         __assert_close(x_grad, out[0], 'x@GRAD')
+
+
+def format_reorder(out, size):
+    in_n = size[0]
+    out_h = size[2]
+    out_w = size[3]
+    out_c = size[1]
+    out_tmp = np.zeros((in_n, out_h, out_w, out_c))
+    for n in range(in_n):
+        for i in range(out_h):
+            for j in range(out_w):
+                for m in range(out_c):
+                    out_tmp[n, i, j, m] = out[n, m, i, j]
+    return out_tmp.reshape(in_n, out_c, out_h, out_w)

python/paddle/fluid/tests/unittests/mkldnn/test_conv2d_int8_mkldnn_op.py

@@ -20,6 +20,7 @@
 import paddle.fluid.core as core
 from paddle.fluid.tests.unittests.op_test import OpTest
 from paddle.fluid.tests.unittests.test_conv2d_op import conv2d_forward_naive, TestConv2dOp
+from mkldnn_op_test import format_reorder
 
 
 def conv2d_forward_refer(input, filter, group, conv_param):
@@ -29,20 +30,6 @@ def conv2d_forward_refer(input, filter, group, conv_param):
     return format_reorder(out, size)
 
 
-def format_reorder(out, size):
-    in_n = size[0]
-    out_h = size[2]
-    out_w = size[3]
-    out_c = size[1]
-    out_tmp = np.zeros((in_n, out_h, out_w, out_c))
-    for n in range(in_n):
-        for i in range(out_h):
-            for j in range(out_w):
-                for m in range(out_c):
-                    out_tmp[n, i, j, m] = out[n, m, i, j]
-    return out_tmp.reshape(in_n, out_c, out_h, out_w)
-
-
 class TestConv2dInt8Op(TestConv2dOp):
     def setUp(self):
         self.op_type = "conv2d"

python/paddle/fluid/tests/unittests/mkldnn/test_requantize_mkldnn_op.py

@@ -0,0 +1,93 @@
+#   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 paddle.fluid.tests.unittests.op_test import OpTest
+from mkldnn_op_test import format_reorder
+
+
+class TestReQuantizeOp(OpTest):
+    def setUp(self):
+        self.op_type = 'requantize'
+        self.scale_in = 2.0
+        self.scale_out = 1.5
+        self.input_size = [1, 1, 5, 5]
+        self.data_type = 'int8'
+        self.set_scale()
+        self.set_data_type()
+
+        scale_shift = self.scale_out / self.scale_in
+
+        if self.data_type == 'int8':
+            input = (np.random.randint(0, 100, self.input_size) - 50
+                     ).astype(self.data_type)
+            output_tmp = np.round(input.astype('float32') *
+                                  scale_shift).astype('int8')
+        else:
+            input = (np.random.randint(0, 100,
+                                       self.input_size)).astype(self.data_type)
+            output_tmp = np.round(input.astype('float32') *
+                                  scale_shift).astype('uint8')
+
+        output = format_reorder(output_tmp, self.input_size)
+
+        self.inputs = {'Input': OpTest.np_dtype_to_fluid_dtype(input)}
+
+        self.outputs = {'Output': output}
+
+        self.attrs = {'Scale_in': self.scale_in, 'Scale_out': self.scale_out}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def set_scale(self):
+        pass
+
+    def set_data_type(OpTest):
+        pass
+
+
+#--------------------test requantize with s8 input--------------------
+
+
+class TestReQuantizeOp1(TestReQuantizeOp):
+    def set_scale(self):
+        self.scale_in = 1.5
+        self.scale_out = 1.5
+
+
+class TestReQuantizeOp2(TestReQuantizeOp):
+    def set_scale(self):
+        self.scale_in = 0.1
+        self.scale_out = 0.2
+
+
+#--------------------test requantize with u8 input--------------------
+
+
+class TestReQuantizeOp3(TestReQuantizeOp1):
+    def set_data_type(self):
+        self.data_type = 'uint8'
+
+
+class TestReQuantizeOp4(TestReQuantizeOp2):
+    def set_data_type(self):
+        self.data_type = 'uint8'
+
+
+if __name__ == '__main__':
+    unittest.main()