fix bug and skip fc_add pass in bf16

PaddlePaddle · Oct 24, 2023 · 88d8b8b · 88d8b8b
1 parent ded5256
commit 88d8b8b
Show file tree

Hide file tree

Showing 5 changed files with 22 additions and 9 deletions.
diff --git a/paddle/fluid/inference/api/paddle_pass_builder.cc b/paddle/fluid/inference/api/paddle_pass_builder.cc
@@ -405,6 +405,7 @@ void CpuPassStrategy::EnableMkldnnQuantizer() {
 
 void CpuPassStrategy::EnableMkldnnBfloat16() {
 #ifdef PADDLE_WITH_DNNL
+  EraseFcAddMkldnnPasses();
   if (!use_mkldnn_bfloat16_) {
     passes_.emplace_back("fc_mkldnn_pass");
     passes_.emplace_back("fc_act_mkldnn_fuse_pass");
@@ -508,6 +509,17 @@ void CpuPassStrategy::EraseFcMkldnnPasses() {
   }
 }
 
+void CpuPassStrategy::EraseFcAddMkldnnPasses() {
+  std::vector<std::string> fc_add_passes_to_erase(
+      {"fc_elementwise_add_mkldnn_fuse_pass"});
+  for (const auto &pass : fc_add_passes_to_erase) {
+    int idx = static_cast<int>(GetPassIndex(pass));
+    if (idx != -1) {
+      passes_.erase(std::begin(passes_) + idx);
+    }
+  }
+}
+
 XpuPassStrategy::XpuPassStrategy() : PassStrategy({}) {
   passes_.assign({
       "delete_assign_op_pass",

diff --git a/paddle/fluid/inference/api/paddle_pass_builder.h b/paddle/fluid/inference/api/paddle_pass_builder.h
@@ -228,6 +228,9 @@ class PD_INFER_DECL CpuPassStrategy : public PassStrategy {
   /// \brief Erase MKLDNN fc passes.
   void EraseFcMkldnnPasses();
 
+  /// \brief Erase MKLDNN fc_elementise_add passes.
+  void EraseFcAddMkldnnPasses();
+
   /// \cond Protected
   bool use_mkldnn_quantizer_{false};
   bool use_mkldnn_bfloat16_{false};

diff --git a/paddle/fluid/operators/mkldnn/fc_mkldnn_op.cc b/paddle/fluid/operators/mkldnn/fc_mkldnn_op.cc
@@ -302,14 +302,10 @@ class FCMKLDNNHandler
         this->dev_ctx_.GetBlob(residual_key));
     if (!memory_p) {
       auto dims = this->fwd_pd_->dst_desc().get_dims();
-      if (phi::funcs::is_int8<T_in>() || phi::funcs::is_bfloat16<T_in>()) {
-        constexpr bool is_int8 = phi::funcs::is_int8<T_in>();
-        auto data_type = dnnl::memory::data_type::bf16;
-        if (is_int8) {
-          data_type = residual->dtype() == phi::DataType::INT8
-                          ? dnnl::memory::data_type::s8
-                          : dnnl::memory::data_type::u8;
-        }
+      if (phi::funcs::is_int8<T_in>()) {
+        auto data_type = residual->dtype() == phi::DataType::INT8
+                             ? dnnl::memory::data_type::s8
+                             : dnnl::memory::data_type::u8;
 
         auto src_0_md =
             dnnl::memory::desc(dims, data_type, dnnl::memory::format_tag::ab);
@@ -603,7 +599,7 @@ class FCMKLDNNKernel : public framework::OpKernel<T_in> {
         fc_args.insert({DNNL_ARG_BIAS, *bias_memory_p});
       }
 
-      if (phi::funcs::is_int8<T_in>() || phi::funcs::is_bfloat16<T_in>()) {
+      if (phi::funcs::is_int8<T_in>()) {
         handler.SetScalesIfNeeded(&fc_args);
       }
 

diff --git a/test/mkldnn/test_fc_add_int8_mkldnn_op.py b/test/mkldnn/test_fc_add_int8_mkldnn_op.py
@@ -14,6 +14,7 @@
 
 import sys
 import unittest
+
 sys.path.append("../legacy_test")
 import numpy as np
 from op_test import OpTest, OpTestTool

diff --git a/test/mkldnn/test_fc_add_mkldnn_op.py b/test/mkldnn/test_fc_add_mkldnn_op.py
@@ -14,6 +14,7 @@
 
 import sys
 import unittest
+
 sys.path.append("../legacy_test")
 
 import numpy as np