Added cast op oneDNN kernel for bf16/fp32 datatypes casting(FWD/BWD) (#33056)

jakpiase · web-flow · commit a2a45d8d55d0 · 2021-05-26T11:45:14.000+08:00
* added op cast functionality for fp32/bf16

* added newline

* added entries in static mode white list and unity build

* fixed failing tests

* changes after review

* added formatting

* upgraded tests file as reviewer suggested

* changes after review

* minor change
diff --git a/paddle/fluid/operators/cast_op.cc b/paddle/fluid/operators/cast_op.cc
@@ -27,6 +27,9 @@ class CastOpProtoMaker : public framework::OpProtoAndCheckerMaker {
     AddOutput("Out", "The output tensor of cast op");
     AddAttr<int>("out_dtype", "output data type");
     AddAttr<int>("in_dtype", "input data type");
+    AddAttr<bool>("use_mkldnn",
+                  "(bool, default false) Only used in mkldnn kernel")
+        .SetDefault(false);
     AddComment(R"DOC(
 Cast Operator.
 
@@ -50,6 +53,7 @@ class CastOpGradMaker : public framework::SingleGradOpMaker<T> {
     grad->SetOutput("Out", this->InputGrad("X"));
     grad->SetAttr("out_dtype", this->GetAttr("in_dtype"));
     grad->SetAttr("in_dtype", this->GetAttr("out_dtype"));
+    grad->SetAttr("use_mkldnn", this->GetAttr("use_mkldnn"));
   }
 };
 
@@ -77,6 +81,28 @@ class CastOp : public framework::OperatorWithKernel {
     if (platform::is_cuda_pinned_place(tensor_place)) {
       return framework::OpKernelType(tensor->type(), ctx.device_context());
     }
+
+#ifdef PADDLE_WITH_MKLDNN
+    int in_dtype = ctx.Attr<int>("in_dtype");
+    int out_dtype = ctx.Attr<int>("out_dtype");
+
+    auto MKLDNNSupportsCast = [&]() -> bool {
+      int dtype_fp32 = static_cast<int>(framework::proto::VarType::FP32);
+      int dtype_bf16 = static_cast<int>(framework::proto::VarType::BF16);
+
+      if ((in_dtype != dtype_fp32 && in_dtype != dtype_bf16) ||
+          (out_dtype != dtype_fp32 && out_dtype != dtype_bf16))
+        return false;
+
+      return true;
+    };
+
+    if (this->CanMKLDNNBeUsed(ctx, tensor->type()) && MKLDNNSupportsCast()) {
+      return framework::OpKernelType(tensor->type(), ctx.GetPlace(),
+                                     framework::DataLayout::kMKLDNN,
+                                     framework::LibraryType::kMKLDNN);
+    }
+#endif
     return framework::OpKernelType(tensor->type(), tensor_place);
   }
 };
diff --git a/paddle/fluid/operators/mkldnn/cast_mkldnn_op.cc b/paddle/fluid/operators/mkldnn/cast_mkldnn_op.cc
@@ -0,0 +1,73 @@
+/* Copyright (c) 2021 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/platform/mkldnn_reuse.h"
+
+namespace paddle {
+namespace operators {
+
+using paddle::framework::Tensor;
+
+template <typename T>
+class CastMKLDNNKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    this->RunKernel(ctx);
+  }
+
+  void RunKernel(const framework::ExecutionContext& ctx) const {
+    const auto& dev_ctx =
+        ctx.template device_context<platform::MKLDNNDeviceContext>();
+
+    auto* x = ctx.Input<Tensor>("X");
+    auto* out = ctx.Output<Tensor>("Out");
+
+    int in_dtype = ctx.Attr<int>("in_dtype");
+    int out_dtype = ctx.Attr<int>("out_dtype");
+
+    auto x_paddle_type = framework::proto::VarType::Type(in_dtype);
+    auto out_paddle_type = framework::proto::VarType::Type(out_dtype);
+
+    mkldnn::memory::data_type x_type =
+        framework::ToMKLDNNDataType(x_paddle_type);
+    mkldnn::memory::data_type out_type =
+        framework::ToMKLDNNDataType(out_paddle_type);
+
+    auto x_tz = framework::vectorize(x->dims());
+
+    std::string key =
+        platform::CreateKey(dev_ctx, x_tz, x->format(), x->format(), x_type);
+    platform::ReorderMKLDNNHandler reorder_handler(
+        x_tz, x_paddle_type, x_type, out_paddle_type, out_type, dev_ctx,
+        dev_ctx.GetEngine(), key);
+
+    auto reorder_src_memory_p = reorder_handler.AcquireSrcMemory(
+        x->format(), platform::to_void_cast(x->data<T>()));
+    auto reorder_dst_memory_p =
+        reorder_handler.AcquireDstMemory(out, x->format(), dev_ctx.GetPlace());
+    auto reorder_p = reorder_handler.AcquireReorder(reorder_dst_memory_p,
+                                                    reorder_src_memory_p);
+
+    auto& astream = platform::MKLDNNDeviceContext::tls().get_stream();
+    reorder_p->execute(astream, *reorder_src_memory_p, *reorder_dst_memory_p);
+    astream.wait();
+
+    out->set_layout(framework::DataLayout::kMKLDNN);
+    out->set_format(platform::GetMKLDNNFormat(*reorder_dst_memory_p));
+  }
+};
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_KERNEL(cast, MKLDNN, paddle::platform::CPUPlace,
+                   ops::CastMKLDNNKernel<float>,
+                   ops::CastMKLDNNKernel<paddle::platform::bfloat16>);
diff --git a/paddle/fluid/operators/unity_build_rule.cmake b/paddle/fluid/operators/unity_build_rule.cmake
@@ -30,6 +30,7 @@ register_unity_group(cc
     bmm_op.cc
     bpr_loss_op.cc
     cast_op.cc
+    mkldnn/cast_mkldnn_op.cc
     cholesky_op.cc
     chunk_eval_op.cc
     clip_by_norm_op.cc
diff --git a/paddle/fluid/platform/mkldnn_reuse.h b/paddle/fluid/platform/mkldnn_reuse.h
@@ -926,7 +926,23 @@ class ReorderMKLDNNHandler : public MKLDNNHandler {
       : platform::MKLDNNHandler(dev_ctx, engine, base_key),
         dims_(dims),
         vtype_(vtype),
-        dtype_(dtype) {}
+        vtype_dst_(vtype),
+        dtype_(dtype),
+        dtype_dst_(dtype) {}
+
+  ReorderMKLDNNHandler(std::vector<int64_t>& dims,  // NOLINT
+                       framework::proto::VarType::Type vtype,
+                       mkldnn::memory::data_type dtype,
+                       framework::proto::VarType::Type vtype_dst,
+                       mkldnn::memory::data_type dtype_dst,
+                       const platform::MKLDNNDeviceContext& dev_ctx,
+                       mkldnn::engine engine, const std::string& base_key)
+      : platform::MKLDNNHandler(dev_ctx, engine, base_key),
+        dims_(dims),
+        vtype_(vtype),
+        vtype_dst_(vtype_dst),
+        dtype_(dtype),
+        dtype_dst_(dtype_dst) {}
 
   std::shared_ptr<mkldnn::memory> AcquireSrcMemory(
       const MKLDNNMemoryFormat& fmt, void* ptr) {
@@ -940,15 +956,16 @@ class ReorderMKLDNNHandler : public MKLDNNHandler {
     auto mem_p =
         std::static_pointer_cast<mkldnn::memory>(dev_ctx_.GetBlob(local_key));
     if (mem_p == nullptr) {
-      auto dst_md = platform::MKLDNNMemDesc(dims_, dtype_, fmt);
-      auto dst_data = output->mutable_data(place, vtype_, dst_md.get_size());
+      auto dst_md = platform::MKLDNNMemDesc(dims_, dtype_dst_, fmt);
+      auto dst_data =
+          output->mutable_data(place, vtype_dst_, dst_md.get_size());
 
       mem_p = std::make_shared<mkldnn::memory>(dst_md, engine_, dst_data);
       dev_ctx_.SetBlob(local_key, mem_p);
     } else {
       // Even if memory object exists , we may be using it for diffrent tensor
       auto dst_data =
-          output->mutable_data(place, vtype_, mem_p->get_desc().get_size());
+          output->mutable_data(place, vtype_dst_, mem_p->get_desc().get_size());
       mem_p->set_data_handle(dst_data);
     }
     return mem_p;
@@ -970,8 +987,8 @@ class ReorderMKLDNNHandler : public MKLDNNHandler {
 
  private:
   std::vector<int64_t> dims_;
-  framework::proto::VarType::Type vtype_;
-  mkldnn::memory::data_type dtype_;
+  framework::proto::VarType::Type vtype_, vtype_dst_;
+  mkldnn::memory::data_type dtype_, dtype_dst_;
 };
 
 template <typename T>
diff --git a/python/paddle/fluid/tests/unittests/mkldnn/test_cast_mkldnn_op.py b/python/paddle/fluid/tests/unittests/mkldnn/test_cast_mkldnn_op.py
@@ -0,0 +1,78 @@
+#   Copyright (c) 2021 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 paddle
+import paddle.fluid.core as core
+import paddle.fluid as fluid
+from paddle.fluid import compiler, Program, program_guard
+from paddle.fluid.tests.unittests.op_test import OpTest, convert_float_to_uint16
+
+
+@unittest.skipIf(not core.supports_bfloat16(),
+                 "place does not support BF16 evaluation")
+class TestCastBF16ToFP32MKLDNNOp(OpTest):
+    def init_data(self):
+        self.out = np.random.random(size=[10, 10]).astype("float32")
+        self.x = convert_float_to_uint16(self.out)
+
+    def setUp(self):
+        self.init_data()
+        self.inputs = {'X': self.x}
+        self.outputs = {'Out': self.out}
+        prepare_dtype = lambda x: int(core.VarDesc.VarType.BF16 if x.dtype != np.float32 else core.VarDesc.VarType.FP32)
+        self.attrs = {
+            'in_dtype': prepare_dtype(self.x),
+            'out_dtype': prepare_dtype(self.out),
+            'use_mkldnn': True
+        }
+        self.op_type = 'cast'
+
+    def test_check_output(self):
+        self.check_output(check_dygraph=False)
+
+    def test_check_grad(self):
+        self.check_grad_with_place(
+            core.CPUPlace(), ["X"],
+            "Out",
+            check_dygraph=False,
+            user_defined_grads=[self.inputs['X']],
+            user_defined_grad_outputs=[self.outputs['Out']])
+
+
+class TestCastFP32ToBF16MKLDNNOp(TestCastBF16ToFP32MKLDNNOp):
+    def init_data(self):
+        self.x = np.random.random(size=[2, 6]).astype("float32")
+        self.out = convert_float_to_uint16(self.x)
+
+
+class TestCastBF16ToBF16MKLDNNOp(TestCastBF16ToFP32MKLDNNOp):
+    def init_data(self):
+        self.x = np.random.random(size=[6, 13]).astype("uint16")
+        self.out = self.x
+
+
+class TestCastFP32ToFP32MKLDNNOp(TestCastBF16ToFP32MKLDNNOp):
+    def init_data(self):
+        self.x = np.random.random(size=[7, 15]).astype("float32")
+        self.out = self.x
+
+
+if __name__ == '__main__':
+    paddle.enable_static()
+    unittest.main()
diff --git a/python/paddle/fluid/tests/unittests/op_test.py b/python/paddle/fluid/tests/unittests/op_test.py
@@ -1191,8 +1191,12 @@ def find_actual(target_name, fetch_list):
                         np.float32, np.float64
                 ]:
                     actual_t = convert_uint16_to_float(actual_t)
-                    atol = 0.03
+                    atol = max(atol, 0.03)
 
+                if expect_t.dtype == np.uint16 and actual_t.dtype == np.uint16:
+                    expect_t = convert_uint16_to_float(expect_t)
+                    actual_t = convert_uint16_to_float(actual_t)
+                    atol = max(atol, 0.03)
                 # NOTE(zhiqiu): np.allclose([], [1.]) returns True
                 # see details: https://stackoverflow.com/questions/38331703/why-does-numpys-broadcasting-sometimes-allow-comparing-arrays-of-different-leng
                 if expect_t.size == 0:
@@ -1501,13 +1505,21 @@ def check_grad_with_place(self,
 
         # comparison of bf16 results will happen as fp32
         # loop over list of grads and convert bf16 to fp32
-        fp32_grads = []
+        fp32_analytic_grads = []
         for grad in analytic_grads:
             if grad.dtype == np.uint16:
                 grad = convert_uint16_to_float(grad)
                 max_relative_error = 0.03
-            fp32_grads.append(grad)
-        analytic_grads = fp32_grads
+            fp32_analytic_grads.append(grad)
+        analytic_grads = fp32_analytic_grads
+
+        fp32_numeric_grads = []
+        for grad in numeric_grads:
+            if grad.dtype == np.uint16:
+                grad = convert_uint16_to_float(grad)
+                max_relative_error = 0.03
+            fp32_numeric_grads.append(grad)
+        numeric_grads = fp32_numeric_grads
 
         self._assert_is_close(numeric_grads, analytic_grads, inputs_to_check,
                               max_relative_error,
diff --git a/tools/static_mode_white_list.py b/tools/static_mode_white_list.py
@@ -589,6 +589,7 @@
     'test_matmul_op_with_head',
     'test_var_conv_2d',
     'test_batch_norm_mkldnn_op',
+    'test_cast_mkldnn_op',
     'test_concat_int8_mkldnn_op',
     'test_concat_bf16_mkldnn_op',
     'test_concat_mkldnn_op',
