Added LSTM BF16 and fixed GRU BF16 (#31234)

paddle/fluid/operators/fused/fusion_lstm_op.cc

@@ -249,6 +249,10 @@ void FusionLSTMOpMaker::Make() {
   AddAttr<bool>("use_mkldnn",
                 "(bool, default false) Only used in mkldnn kernel")
       .SetDefault(false);
+  AddAttr<bool>("force_fp32_output",
+                "(bool, default false) Force INT8 kernel output FP32, only "
+                "used in MKL-DNN INT8")
+      .SetDefault(false);
   AddComment(R"DOC(
 Fusion Long-Short Term Memory (LSTM) Operator.
 This operator fuse the X into LSTM, more details can refer to LSTM op.

paddle/fluid/operators/fused/mkldnn/fusion_gru_mkldnn_op.cc

@@ -89,6 +89,7 @@ class GRUMKLDNNHandler : public RNNMKLDNNHandler<T, dnnl::gru_forward, T_out> {
     }
   }
 
+  template <typename U>
   std::shared_ptr<dnnl::memory> AcquireWeightXMemory(const Tensor* weight_x,
                                                      const bool origin_mode) {
     const std::string wx_key = this->memory_key_ + "@weight_x";
@@ -98,18 +99,18 @@ class GRUMKLDNNHandler : public RNNMKLDNNHandler<T, dnnl::gru_forward, T_out> {
     if (!memory_p) {
       auto user_md =
           MKLDNNMemDesc({1, 1, this->IC, this->G, this->OC},
-                        MKLDNNGetDataType<float>(), MKLDNNMemoryFormat::ldigo);
+                        MKLDNNGetDataType<U>(), MKLDNNMemoryFormat::ldigo);
       auto user_memory = dnnl::memory(user_md, this->engine_);
 
-      auto* weight_x_data =
-          reinterpret_cast<float*>(user_memory.get_data_handle());
-      memcpy(weight_x_data, weight_x->data<float>(),
-             sizeof(float) * this->IC * this->G * this->OC);
+      auto* weight_x_data = reinterpret_cast<U*>(user_memory.get_data_handle());
+      memcpy(weight_x_data, weight_x->data<U>(),
+             sizeof(U) * this->IC * this->G * this->OC);
 
       if (origin_mode == false) {
         for (int64_t i = 0; i < this->IC; ++i) {
           for (int64_t j = 0; j < this->OC; ++j) {
-            weight_x_data[j] *= -1;
+            U minus_one(-1.0f);
+            weight_x_data[j] = minus_one * weight_x_data[j];
           }
           weight_x_data += 3 * this->OC;
         }
@@ -127,6 +128,7 @@ class GRUMKLDNNHandler : public RNNMKLDNNHandler<T, dnnl::gru_forward, T_out> {
     return memory_p;
   }
 
+  template <typename U>
   std::shared_ptr<dnnl::memory> AcquireWeightHMemory(const Tensor* weight_h,
                                                      const bool origin_mode) {
     const std::string wh_key = this->memory_key_ + "@weight_h";
@@ -136,34 +138,33 @@ class GRUMKLDNNHandler : public RNNMKLDNNHandler<T, dnnl::gru_forward, T_out> {
     if (!memory_p) {
       auto user_md =
           MKLDNNMemDesc({1, 1, this->OC, this->G, this->OC},
-                        MKLDNNGetDataType<float>(), MKLDNNMemoryFormat::ldigo);
+                        MKLDNNGetDataType<U>(), MKLDNNMemoryFormat::ldigo);
       auto user_memory = dnnl::memory(user_md, this->engine_);
 
       // Reorder weights_h from PP format [OC, 2OC] + [OC, OC] to
       // oneDNN format [OC, 3OC]
-      auto* weight_h_data =
-          reinterpret_cast<float*>(user_memory.get_data_handle());
-      auto* user_weight_h_data = weight_h->data<float>();
+      auto* weight_h_data = reinterpret_cast<U*>(user_memory.get_data_handle());
+      auto* user_weight_h_data = weight_h->data<U>();
 
       auto src1_iter = user_weight_h_data;
       auto src2_iter = user_weight_h_data + 2 * this->OC * this->OC;
 
       for (int64_t c = 0; c < this->OC; ++c) {
-        memcpy(weight_h_data, src1_iter, 2 * this->OC * sizeof(float));
-        memcpy(weight_h_data + 2 * this->OC, src2_iter,
-               this->OC * sizeof(float));
+        memcpy(weight_h_data, src1_iter, 2 * this->OC * sizeof(U));
+        memcpy(weight_h_data + 2 * this->OC, src2_iter, this->OC * sizeof(U));
 
         src1_iter += 2 * this->OC;
         src2_iter += this->OC;
         weight_h_data += 3 * this->OC;
       }
 
-      weight_h_data = reinterpret_cast<float*>(user_memory.get_data_handle());
+      weight_h_data = reinterpret_cast<U*>(user_memory.get_data_handle());
 
       if (origin_mode == false) {
         for (int64_t i = 0; i < this->OC; ++i) {
           for (int64_t j = 0; j < this->OC; ++j) {
-            weight_h_data[j] *= -1;
+            U minus_one(-1.0f);
+            weight_h_data[j] = minus_one * weight_h_data[j];
           }
           weight_h_data += 3 * this->OC;
         }
@@ -273,11 +274,34 @@ class FusionGRUMKLDNNKernel : public framework::OpKernel<T> {
 
     auto input_memory_p =
         handler.AcquireInputMemoryWithReorder(input, is_reverse);
-    auto h0_memory_p = handler.AcquireH0Memory(h0);
-    auto weight_x_memory_p =
-        handler.AcquireWeightXMemory(weight_x, origin_mode);
-    auto weight_h_memory_p =
-        handler.AcquireWeightHMemory(weight_h, origin_mode);
+
+    std::shared_ptr<dnnl::memory> h0_memory_p, weight_h_memory_p,
+        weight_x_memory_p;
+
+    if (weight_h->type() == paddle::framework::proto::VarType_Type_FP32) {
+      h0_memory_p = handler.template AcquireH0Memory<float>(h0);
+      weight_x_memory_p =
+          handler.template AcquireWeightXMemory<float>(weight_x, origin_mode);
+      weight_h_memory_p =
+          handler.template AcquireWeightHMemory<float>(weight_h, origin_mode);
+    } else if (weight_h->type() ==
+               paddle::framework::proto::VarType_Type_BF16) {
+      h0_memory_p =
+          handler.template AcquireH0Memory<paddle::platform::bfloat16>(h0);
+      weight_x_memory_p =
+          handler.template AcquireWeightXMemory<paddle::platform::bfloat16>(
+              weight_x, origin_mode);
+      weight_h_memory_p =
+          handler.template AcquireWeightHMemory<paddle::platform::bfloat16>(
+              weight_h, origin_mode);
+    } else {
+      h0_memory_p = handler.template AcquireH0Memory<uint8_t>(h0);
+      weight_x_memory_p =
+          handler.template AcquireWeightXMemory<int8_t>(weight_x, origin_mode);
+      weight_h_memory_p =
+          handler.template AcquireWeightHMemory<int8_t>(weight_h, origin_mode);
+    }
+
     auto bias_memory_p = handler.AcquireBiasMemory(bias, origin_mode);
     auto hidden_onednn_memory_p = handler.AcquireOutputMemory();
 

paddle/fluid/operators/fused/mkldnn/fusion_lstm_mkldnn_op.cc

@@ -81,8 +81,11 @@ class LSTMMKLDNNHandler
                                      MKLDNNMemoryFormat::tnc);
       auto h0_md = MKLDNNMemDesc({L, D, N, OC}, MKLDNNGetDataType<T>(),
                                  MKLDNNMemoryFormat::ldnc);
-      auto c0_md = MKLDNNMemDesc({L, D, N, OC}, MKLDNNGetDataType<T>(),
-                                 MKLDNNMemoryFormat::ldnc);
+      auto c0_md = MKLDNNMemDesc(
+          {L, D, N, OC}, MKLDNNGetDataType<float>(),  // Vanilla LSTM and LSTM
+                                                      // with peepoles has c0 as
+                                                      // fp32
+          MKLDNNMemoryFormat::ldnc);
 
       // Create LSTM oneDNN primitive
       const auto direction =
@@ -110,13 +113,14 @@ class LSTMMKLDNNHandler
   // needed
   // PaddlePaddle:  {c, i, f, o}
   // oneDNN:        {i, f, c, o}
-  void ReorderGates(float* weights, int64_t I) {
+  template <typename U>
+  void ReorderGates(U* weights, int64_t I) {
     size_t inner_block_size = this->OC;
     size_t block_size = inner_block_size * this->G;
     for (size_t i = 0; i < (size_t)I; ++i) {
       size_t offset = i * block_size;
 
-      float* base_pos = weights + offset;
+      U* base_pos = weights + offset;
       std::swap_ranges(base_pos, base_pos + inner_block_size,
                        base_pos + inner_block_size);  // c <-> i
       std::swap_ranges(base_pos + inner_block_size,
@@ -125,6 +129,7 @@ class LSTMMKLDNNHandler
     }
   }
 
+  template <typename U>
   std::shared_ptr<dnnl::memory> AcquireWeightXMemory(const Tensor* weight_x) {
     const std::string wx_key = this->memory_key_ + "@weight_x";
     auto memory_p =
@@ -133,13 +138,12 @@ class LSTMMKLDNNHandler
     if (!memory_p) {
       auto user_md =
           MKLDNNMemDesc({1, 1, this->IC, this->G, this->OC},
-                        MKLDNNGetDataType<float>(), MKLDNNMemoryFormat::ldigo);
+                        MKLDNNGetDataType<U>(), MKLDNNMemoryFormat::ldigo);
       auto user_memory = dnnl::memory(user_md, this->engine_);
 
-      auto* weight_x_data =
-          reinterpret_cast<float*>(user_memory.get_data_handle());
-      memcpy(weight_x_data, weight_x->data<float>(),
-             sizeof(float) * this->IC * this->G * this->OC);
+      auto* weight_x_data = reinterpret_cast<U*>(user_memory.get_data_handle());
+      memcpy(weight_x_data, weight_x->data<U>(),
+             sizeof(U) * this->IC * this->G * this->OC);
 
       ReorderGates(weight_x_data, this->IC);
 
@@ -155,6 +159,7 @@ class LSTMMKLDNNHandler
     return memory_p;
   }
 
+  template <typename U>
   std::shared_ptr<dnnl::memory> AcquireWeightHMemory(const Tensor* weight_h) {
     const std::string wh_key = this->memory_key_ + "@weight_h";
     auto memory_p =
@@ -163,13 +168,12 @@ class LSTMMKLDNNHandler
     if (!memory_p) {
       auto user_md =
           MKLDNNMemDesc({1, 1, this->OC, this->G, this->OC},
-                        MKLDNNGetDataType<float>(), MKLDNNMemoryFormat::ldigo);
+                        MKLDNNGetDataType<U>(), MKLDNNMemoryFormat::ldigo);
       auto user_memory = dnnl::memory(user_md, this->engine_);
 
-      auto* weight_h_data =
-          reinterpret_cast<float*>(user_memory.get_data_handle());
-      memcpy(weight_h_data, weight_h->data<float>(),
-             sizeof(float) * this->OC * this->G * this->OC);
+      auto* weight_h_data = reinterpret_cast<U*>(user_memory.get_data_handle());
+      memcpy(weight_h_data, weight_h->data<U>(),
+             sizeof(U) * this->OC * this->G * this->OC);
 
       ReorderGates(weight_h_data, this->OC);
 
@@ -258,8 +262,8 @@ class LSTMMKLDNNHandler
         memset(user_c0_memory.get_data_handle(), 0,
                sizeof(float) * this->N * this->OC);
       }
-      memory_p = std::make_shared<dnnl::memory>(this->fwd_pd_->src_iter_desc(),
-                                                this->engine_);
+      memory_p = std::make_shared<dnnl::memory>(
+          this->fwd_pd_->src_iter_c_desc(), this->engine_);
 
       auto& astream = paddle::platform::MKLDNNDeviceContext::tls().get_stream();
       dnnl::reorder(user_c0_memory, *memory_p, this->attr_)
@@ -275,7 +279,15 @@ template <typename T>
 class FusionLSTMMKLDNNKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
-    RunKernel<float>(ctx);
+    const bool is_bf16 = std::is_same<T, paddle::platform::bfloat16>::value;
+    const bool force_fp32_output = ctx.Attr<bool>("force_fp32_output");
+
+    // BF16 does not support force output
+    if (!is_bf16 && force_fp32_output) {
+      RunKernel<float>(ctx);
+    } else {
+      RunKernel<T>(ctx);
+    }
   }
 
   template <typename Tout = T>
@@ -327,10 +339,29 @@ class FusionLSTMMKLDNNKernel : public framework::OpKernel<T> {
 
     auto input_memory_p =
         handler.AcquireInputMemoryWithReorder(input, is_reverse);
-    auto h0_memory_p = handler.AcquireH0Memory(h0);
     auto c0_memory_p = handler.AcquireC0Memory(c0);
-    auto weight_x_memory_p = handler.AcquireWeightXMemory(weight_x);
-    auto weight_h_memory_p = handler.AcquireWeightHMemory(weight_h);
+
+    std::shared_ptr<dnnl::memory> h0_memory_p, weight_h_memory_p,
+        weight_x_memory_p;
+
+    if (weight_h->type() == paddle::framework::proto::VarType_Type_FP32) {
+      h0_memory_p = handler.template AcquireH0Memory<float>(h0);
+      weight_x_memory_p =
+          handler.template AcquireWeightXMemory<float>(weight_x);
+      weight_h_memory_p =
+          handler.template AcquireWeightHMemory<float>(weight_h);
+    } else if (weight_h->type() ==
+               paddle::framework::proto::VarType_Type_BF16) {
+      h0_memory_p =
+          handler.template AcquireH0Memory<paddle::platform::bfloat16>(h0);
+      weight_x_memory_p =
+          handler.template AcquireWeightXMemory<paddle::platform::bfloat16>(
+              weight_x);
+      weight_h_memory_p =
+          handler.template AcquireWeightHMemory<paddle::platform::bfloat16>(
+              weight_h);
+    }
+
     auto bias_memory_p = handler.AcquireBiasMemory(bias);
     auto hidden_onednn_memory_p = handler.AcquireOutputMemory();
 
@@ -374,4 +405,5 @@ class FusionLSTMMKLDNNKernel : public framework::OpKernel<T> {
 
 namespace ops = paddle::operators;
 REGISTER_OP_KERNEL(fusion_lstm, MKLDNN, paddle::platform::CPUPlace,
-                   ops::FusionLSTMMKLDNNKernel<float>);
+                   ops::FusionLSTMMKLDNNKernel<float>,
+                   ops::FusionLSTMMKLDNNKernel<paddle::platform::bfloat16>);

paddle/fluid/operators/fused/mkldnn/fusion_rnn_mkldnn.h

@@ -179,6 +179,7 @@ class RNNMKLDNNHandler : public platform::MKLDNNHandlerT<T, T_alg> {
   // TODO(grygielski) H0 is for now persistable
   // TODO(jczaja) H0 should be updated each iter and of T type (Fusion pass does
   // not support in yet)
+  template <typename U>
   std::shared_ptr<dnnl::memory> AcquireH0Memory(const Tensor* h0) {
     const std::string h0_key = memory_key_ + "@h0";
     auto memory_p =
@@ -187,17 +188,14 @@ class RNNMKLDNNHandler : public platform::MKLDNNHandlerT<T, T_alg> {
     if (!memory_p) {
       auto user_h0_memory = dnnl::memory();
       if (h0) {
-        user_h0_memory =
-            dnnl::memory({{1, 1, N, OC},
-                          MKLDNNGetDataType<float>(),
-                          MKLDNNMemoryFormat::ldnc},
-                         this->engine_, to_void_cast(h0->data<float>()));
+        user_h0_memory = dnnl::memory(
+            {{1, 1, N, OC}, MKLDNNGetDataType<U>(), MKLDNNMemoryFormat::ldnc},
+            this->engine_, to_void_cast(h0->data<U>()));
       } else {
-        user_h0_memory = dnnl::memory({{1, 1, N, OC},
-                                       MKLDNNGetDataType<float>(),
-                                       MKLDNNMemoryFormat::ldnc},
-                                      this->engine_);
-        memset(user_h0_memory.get_data_handle(), 0, sizeof(float) * N * OC);
+        user_h0_memory = dnnl::memory(
+            {{1, 1, N, OC}, MKLDNNGetDataType<U>(), MKLDNNMemoryFormat::ldnc},
+            this->engine_);
+        memset(user_h0_memory.get_data_handle(), 0, sizeof(U) * N * OC);
       }
       memory_p = std::make_shared<dnnl::memory>(this->fwd_pd_->src_iter_desc(),
                                                 this->engine_);

python/paddle/fluid/tests/unittests/mkldnn/test_fusion_gru_bf16_mkldnn_op.py

@@ -30,6 +30,11 @@ class TestFusionGRUBF16MKLDNNOp(OpTest):
     def set_confs(self):
         self.mkldnn_data_type = False
 
+    def test_check_output(self):
+        for use_seq in {True, False}:
+            self.attrs['use_seq'] = use_seq
+            self.check_output(check_dygraph=False)
+
     def setUp(self):
         self.op_type = "fusion_gru"
         self.lod = [[2, 4, 3]]
@@ -45,6 +50,7 @@ class TestFusionGRUBF16MKLDNNOp(OpTest):
         self.origin_mode = False
         self.use_mkldnn = True
         self.force_fp32_output = False
+        self.weights_dtype = 'fp32'
         self.set_confs()
 
         T = sum(self.lod[0])
@@ -58,6 +64,9 @@ class TestFusionGRUBF16MKLDNNOp(OpTest):
         wx_fp32 = np.random.rand(self.M, 3 * self.D).astype('float32')
         wh_fp32 = np.random.rand(self.D, 3 * self.D).astype('float32')
 
+        wx_bf16 = convert_float_to_uint16(wx_fp32)
+        wh_bf16 = convert_float_to_uint16(wh_fp32)
+
         # bias is fp32 despite other inputs being in bf16
         bias = np.random.rand(
             1, 3 * self.D).astype('float32') if self.with_bias else np.zeros(
@@ -74,20 +83,30 @@ class TestFusionGRUBF16MKLDNNOp(OpTest):
 
         hidden_bf16 = convert_float_to_uint16(hidden)
 
-        self.inputs = {
-            'X': (x_bf16, self.lod),
-            'WeightX': wx_fp32,
-            'WeightH': wh_fp32
-        }
+        if self.weights_dtype == 'bf16':
+            self.inputs = {
+                'X': (x_bf16, self.lod),
+                'WeightX': wx_bf16,
+                'WeightH': wh_bf16
+            }
+        elif self.weights_dtype == 'fp32':
+            self.inputs = {
+                'X': (x_bf16, self.lod),
+                'WeightX': wx_fp32,
+                'WeightH': wh_fp32
+            }
 
         if self.with_bias:
             self.inputs['Bias'] = bias
 
         if self.with_h0:
-            self.inputs['H0'] = h0_bf16
+            if self.weights_dtype == 'bf16':
+                self.inputs['H0'] = h0_bf16
+            elif self.weights_dtype == 'fp32':
+                self.inputs['H0'] = h0_fp32
 
         h0_bf16 = convert_float_to_uint16(h0_fp32)
-        self.outputs = {'Hidden': (hidden_bf16, self.lod)}
+        self.outputs = {'Hidden': (hidden, self.lod)}
 
         self.attrs = {
             'activation': self.act_state,
@@ -109,6 +128,11 @@ class TestFusionGRUINT8MKLDNNOp3(TestFusionGRUBF16MKLDNNOp):
         self.with_bias = False
 
 
+class TestFusionGRUINT8MKLDNNBF16WeightsOp(TestFusionGRUBF16MKLDNNOp):
+    def set_confs(self):
+        self.weights_dtype = 'bf16'
+
+
 if __name__ == "__main__":
     from paddle import enable_static
     enable_static()

python/paddle/fluid/tests/unittests/mkldnn/test_fusion_gru_int8_mkldnn_op.py

@@ -146,4 +146,6 @@ class TestFusionGRUINT8MKLDNNOp5(TestFusionGRUINT8MKLDNNOp):
 
 
 if __name__ == "__main__":
+    from paddle import enable_static
+    enable_static()
     unittest.main()

python/paddle/fluid/tests/unittests/mkldnn/test_fusion_lstm_bf16_mkldnn_op.py

+#   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 struct
+import paddle.fluid.core as core
+from paddle.fluid.tests.unittests.op_test import OpTest, convert_float_to_uint16, convert_uint16_to_float
+from paddle.fluid.tests.unittests.test_fusion_lstm_op import TestFusionLSTMOp, fc, ACTIVATION, fusion_lstm
+from paddle.fluid.tests.unittests.test_fusion_gru_op import fusion_gru
+
+
+@unittest.skipIf(not core.supports_bfloat16(),
+                 "place does not support BF16 evaluation")
+class TestFusionLSTMBF16ONEDNNOp(OpTest):
+    def set_confs(self):
+        self.mkldnn_data_type = False
+
+    def test_check_output(self):
+        for use_seq in {True, False}:
+            self.attrs['use_seq'] = use_seq
+            self.check_output(check_dygraph=False, no_check_set=["Cell"])
+
+    def setUp(self):
+        self.op_type = 'fusion_lstm'
+        self.lod = [[2, 3, 5, 4]]
+        self.M = 8
+        self.D = 16
+        self.has_initial_state = False
+        self.use_peepholes = False
+        self.is_reverse = False
+        self._cpu_only = True
+        self.act_gate = 'sigmoid'
+        self.act_cell = 'tanh'
+        self.act_cand = 'tanh'
+        self.use_mkldnn = True
+        self.force_fp32_output = False
+        self.weights_dtype = 'fp32'
+        self.set_confs()
+
+        T = sum(self.lod[0])
+        bs = len(self.lod[0])
+
+        # fp32 X input for reference implementation and
+        # corressponding bf16 data as input to LSTM oneDNN bf16 kernel
+        x = np.random.normal(size=(T, self.M)).astype('float32')
+
+        x_bf16 = convert_float_to_uint16(x)
+
+        if self.has_initial_state:
+            h0 = np.random.normal(size=(bs, self.D)).astype('float32')
+            c0 = np.random.normal(size=(bs, self.D)).astype('float32')
+        else:
+            h0 = np.zeros((bs, self.D)).astype('float32')
+            c0 = np.zeros((bs, self.D)).astype('float32')
+
+        wh = np.random.normal(size=(self.D, 4 * self.D)).astype('float32')
+
+        h0_bf16 = convert_float_to_uint16(h0)
+
+        if self.use_peepholes:
+            b = np.random.normal(size=(1, 7 * self.D)).astype('float32')
+        else:
+            b = np.random.normal(size=(1, 4 * self.D)).astype('float32')
+        w_b = np.copy(b[:, 0:4 * self.D])
+        w_c = b[:, 4 * self.D:] if self.use_peepholes else None
+
+        wx = np.random.normal(size=(self.M, 4 * self.D)).astype('float32')
+
+        wx_bf16 = convert_float_to_uint16(wx)
+        wh_bf16 = convert_float_to_uint16(wh)
+
+        bx = np.random.normal(size=(1, 4 * self.D)).astype('float32')
+        b[0, 0:4 * self.D] += bx[0, :]
+
+        hidden, c = fusion_lstm(x, self.lod, wx, bx, h0, c0, wh, w_b, w_c,
+                                self.is_reverse, ACTIVATION[self.act_gate],
+                                ACTIVATION[self.act_cell],
+                                ACTIVATION[self.act_cand])
+
+        hidden = hidden.astype('float32')
+        hidden_bf16 = convert_float_to_uint16(hidden)
+
+        if self.weights_dtype == 'bf16':
+            self.inputs = {
+                'X': (x_bf16, self.lod),
+                'WeightX': wx_bf16,
+                'WeightH': wh_bf16,
+                'Bias': b
+            }
+        elif self.weights_dtype == 'fp32':
+            self.inputs = {
+                'X': (x_bf16, self.lod),
+                'WeightX': wx,
+                'WeightH': wh,
+                'Bias': b
+            }
+
+        if self.has_initial_state:
+            if self.weights_dtype == 'bf16':
+                self.inputs['H0'] = h0_bf16
+            elif self.weights_dtype == 'fp32':
+                self.inputs['H0'] = h0
+
+            self.inputs['C0'] = c0
+
+        self.outputs = {
+            'Hidden': (hidden, self.lod),
+            'Cell': (c, self.lod),
+        }
+
+        self.attrs = {
+            'use_peepholes': self.use_peepholes,
+            'is_reverse': self.is_reverse,
+            'gate_activation': self.act_gate,
+            'cell_activation': self.act_cell,
+            'candidate_activation': self.act_cand,
+            'force_fp32_output': self.force_fp32_output,
+            'use_mkldnn': self.use_mkldnn
+        }
+
+
+class TestFusionLSTMBF16ONEDNNPeepholesOp(TestFusionLSTMBF16ONEDNNOp):
+    def set_confs(self):
+        self.use_peepholes = True
+
+
+class TestFusionLSTMBF16ONEDNNInitializedStateOp(TestFusionLSTMBF16ONEDNNOp):
+    def set_confs(self):
+        self.has_initial_state = True
+
+
+class TestFusionLSTMBF16ONEDNNReverseOp(TestFusionLSTMBF16ONEDNNOp):
+    def set_confs(self):
+        self.is_reverse = True
+
+
+class TestFusionLSTMBF16ONEDNNBF16WeightsOp(TestFusionLSTMBF16ONEDNNOp):
+    def set_confs(self):
+        self.weights_dtype = 'bf16'
+
+
+if __name__ == "__main__":
+    from paddle import enable_static
+    enable_static()
+    unittest.main()

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

@@ -235,6 +235,19 @@ def convert_float_to_uint16(float_list, data_format="NCHW"):
     return new_output
 
 
+def copy_bits_from_uint16_to_float(i):
+    i = np.uint32(i) << 16
+    return struct.unpack('<f', struct.pack('<I', i))[0]
+
+
+def convert_uint16_to_float(uint16_list):
+    new_output = []
+    for x in np.nditer(uint16_list):
+        new_output.append(np.float32(copy_bits_from_uint16_to_float(x)))
+
+    return np.reshape(new_output, uint16_list.shape).view(np.float32)
+
+
 class OpTest(unittest.TestCase):
     @classmethod
     def setUpClass(cls):
@@ -1143,8 +1156,14 @@ class OpTest(unittest.TestCase):
                 idx = find_actual(out_name, fetch_list)
                 actual = outs[idx]
                 actual_t = np.array(actual)
+
                 expect = self.outputs[out_name]
                 expect_t = expect[0] if isinstance(expect, tuple) else expect
+
+                if actual_t.dtype == np.uint16 and expect_t.dtype == np.float32:
+                    actual_t = convert_uint16_to_float(actual_t)
+                    atol = 0.03
+
                 self.assertTrue(
                     np.allclose(
                         actual_t, expect_t, atol=atol, equal_nan=equal_nan),

tools/static_mode_white_list.py

@@ -602,8 +602,10 @@ STATIC_MODE_TESTING_LIST = [
     'test_nearest_interp_mkldnn_op',
     'test_bilinear_interp_mkldnn_op',
     'test_fusion_gru_int8_mkldnn_op',
+    'test_fusion_gru_bf16_mkldnn_op',
     'test_fusion_gru_mkldnn_op',
     'test_fusion_lstm_mkldnn_op',
+    'test_fusion_lstm_bf16_mkldnn_op',
     'test_gaussian_random_mkldnn_op',
     'test_lrn_mkldnn_op',
     'test_matmul_mkldnn_op',