[Cherry-pick][Paddle Inference] fix mixed precision diff (#49477)

* disable scale op in amp pass

* Do not insert redundant cast op

* fix fused_fc_elementwise_layernorm kernel diff

* fix fc kerenl diff

paddle/fluid/framework/ir/auto_mixed_precision_pass.cc

@@ -78,11 +78,11 @@ inline bool VarNodeHasDtype(Node* var_node) {
          (type == VarType::VOCAB);
 }
 
-inline bool IsFloatType(VarType::Type type) {
+inline bool IsFP32AndFP64(VarType::Type type) {
   return (type == VarType::FP64) || (type == VarType::FP32);
 }
 
-inline bool IsHalfType(VarType::Type type) {
+inline bool IsFP16AndBFP16(VarType::Type type) {
   return (type == VarType::FP16) || (type == VarType::BF16);
 }
 
@@ -159,26 +159,16 @@ bool OpSupportPrecision(const std::string& op_type,
 // The set of ops that support fp16 calculation and are considered
 // numerically-dangerous, slower and whose effects may also be observed in
 // downstream ops.
+// ref to python/paddle/fluid/contrib/mixed_precision/fp16_lists.py
 void AutoMixedPrecisionPass::SetDefaultBlacklist() const {
   black_list_.insert({
       // numerically-dangerous
-      "acos",
-      "asin",
-      "cosh",
-      "tan",
       "exp",
-      "expm1",
       "square",
       "log",
-      "log2",
-      "log10",
-      "log1p",
-      "logsumexp",
       "mean",
-      "rsqrt",
       "sum",
       "cos_sim",
-      "softmax",
       "softmax_with_cross_entropy",
       "sigmoid_cross_entropy_with_logits",
       "c_softmax_with_cross_entropy",
@@ -272,6 +262,9 @@ void AutoMixedPrecisionPass::ApplyImpl(Graph* graph) const {
   VLOG(4) << "InsertCastOp done";
   RestoreOpOriginType();
   VLOG(4) << "RestoreOpOriginType done";
+  LOG(INFO) << "The number of ops run at low precision ["
+            << op_run_low_precision_.size() << "/" << op_original_type_.size()
+            << "]";
 }
 
 void AutoMixedPrecisionPass::SetOpUniqueType() const {
@@ -315,11 +308,26 @@ void AutoMixedPrecisionPass::ProcessOpWithDtypeAttr() const {
   for (const auto& nodes : all_op_nodes_) {
     for (auto* op_node : nodes) {
       auto op_type = op_node->Op()->Type();
+
+      if (op_node->Op()->HasAttr("in_dtype")) {
+        auto* var_node = op_node->inputs[0];
+        auto* real_var_node = real_vars_[var_node->Var()->Name()];
+        if (IsFP16AndBFP16(real_var_node->Var()->GetDataType())) {
+          op_node->Op()->SetAttr(
+              "in_dtype",
+              static_cast<int>(framework::TransToProtoVarType(low_precision_)));
+          op_node->Op()->Flush();
+          VLOG(4) << "process op with in_dtype attr: " << op_type << " ( "
+                  << static_cast<int>(real_var_node->Var()->GetDataType())
+                  << " --->" << static_cast<int>(low_precision_) << " )";
+        }
+      }
+
       if (op_run_low_precision_.count(op_type) == 0) continue;
 
       if (op_node->Op()->HasAttr("dtype")) {
         auto dtype = op_node->Op()->GetAttrIfExists<int>("dtype");
-        if (IsFloatType(static_cast<VarType::Type>(dtype))) {
+        if (IsFP32AndFP64(static_cast<VarType::Type>(dtype))) {
           op_node->Op()->SetAttr(
               "dtype",
               static_cast<int>(framework::TransToProtoVarType(low_precision_)));
@@ -327,10 +335,9 @@ void AutoMixedPrecisionPass::ProcessOpWithDtypeAttr() const {
           VLOG(4) << "process op with dtype attr: " << op_type << " ( " << dtype
                   << " --->" << static_cast<int>(low_precision_) << " )";
         }
-      }
-      if (op_node->Op()->HasAttr("out_dtype")) {
+      } else if (op_node->Op()->HasAttr("out_dtype")) {
         auto out_dtype = op_node->Op()->GetAttrIfExists<int>("out_dtype");
-        if (IsFloatType(static_cast<VarType::Type>(out_dtype))) {
+        if (IsFP32AndFP64(static_cast<VarType::Type>(out_dtype))) {
           op_node->Op()->SetAttr(
               "out_dtype",
               static_cast<int>(framework::TransToProtoVarType(low_precision_)));
@@ -359,37 +366,55 @@ void AutoMixedPrecisionPass::GetOpPrecision() const {
 
       if (op_node->Op()->HasAttr("dtype")) {
         auto dtype = op_node->Op()->GetAttrIfExists<int>("dtype");
-        support_low_precision = support_low_precision &&
-                                IsFloatType(static_cast<VarType::Type>(dtype));
+        support_low_precision =
+            support_low_precision &&
+            IsFP32AndFP64(static_cast<VarType::Type>(dtype));
       } else if (op_node->Op()->HasAttr("out_dtype")) {
         auto out_dtype = op_node->Op()->GetAttrIfExists<int>("out_dtype");
         support_low_precision =
             support_low_precision &&
-            IsFloatType(static_cast<VarType::Type>(out_dtype));
-      } else {
-        // if op's input var and output var is not dense tensor, the op should
-        // not run at low precision.
-        for (auto* in_var_node : op_node->inputs) {
-          CHECK_EQ(in_var_node->IsVar(), true);
-          auto* real_in_var_node = real_vars_[in_var_node->Var()->Name()];
-          if (real_in_var_node->Var()->Persistable()) continue;
+            IsFP32AndFP64(static_cast<VarType::Type>(out_dtype));
+      }
 
+      // If scale op's "scale" and "bias" attr value exceed the range of fp16
+      // and bf16, it cannot run at low precision.
+      if (GetOpOriginalType(op_node->Op()->Type()) == "scale") {
+        auto scale = op_node->Op()->GetAttrIfExists<float>("scale");
+        auto bias = op_node->Op()->GetAttrIfExists<float>("bias");
+        if (low_precision_ == phi::DataType::FLOAT16) {
           support_low_precision =
               support_low_precision &&
-              (real_in_var_node->Var()->GetType() == VarType::LOD_TENSOR);
-        }
-
-        for (auto* out_var_node : op_node->outputs) {
-          CHECK_EQ(out_var_node->IsVar(), true);
-          auto* real_out_var_node = real_vars_[out_var_node->Var()->Name()];
-          if (real_out_var_node->Var()->Persistable()) continue;
-
+              phi::dtype::isfinite(static_cast<phi::dtype::float16>(scale)) &&
+              phi::dtype::isfinite(static_cast<phi::dtype::float16>(bias));
+        } else if (low_precision_ == phi::DataType::BFLOAT16) {
           support_low_precision =
               support_low_precision &&
-              (real_out_var_node->Var()->GetType() == VarType::LOD_TENSOR);
+              phi::dtype::isfinite(static_cast<phi::dtype::bfloat16>(scale)) &&
+              phi::dtype::isfinite(static_cast<phi::dtype::bfloat16>(bias));
         }
       }
 
+      // if op's input var and output var is not dense tensor, the op should
+      // not run at low precision.
+      for (auto* in_var_node : op_node->inputs) {
+        CHECK_EQ(in_var_node->IsVar(), true);
+        auto* real_in_var_node = real_vars_[in_var_node->Var()->Name()];
+        if (real_in_var_node->Var()->Persistable()) continue;
+
+        support_low_precision =
+            support_low_precision &&
+            (real_in_var_node->Var()->GetType() == VarType::LOD_TENSOR);
+      }
+      for (auto* out_var_node : op_node->outputs) {
+        CHECK_EQ(out_var_node->IsVar(), true);
+        auto* real_out_var_node = real_vars_[out_var_node->Var()->Name()];
+        if (real_out_var_node->Var()->Persistable()) continue;
+
+        support_low_precision =
+            support_low_precision &&
+            (real_out_var_node->Var()->GetType() == VarType::LOD_TENSOR);
+      }
+
       if (support_low_precision) {
         op_run_low_precision_.insert(op_type);
         VLOG(4) << "support precision: " << op_type << " run at low precision";
@@ -439,7 +464,7 @@ void AutoMixedPrecisionPass::UpdateOpPrecision() const {
         }
 
         // when op_1 only support cpu kernel. if op_2's intput var is op_1's
-        // output var, then op_2 should not run half.
+        // output var, then op_2 should not run at low precision.
         if (GetOpOriginalType(op_type) != "feed" &&
             !GpuKernelSupportPrecision(GetOpOriginalType(op_type),
                                        phi::DataType::FLOAT32)) {
@@ -597,7 +622,7 @@ void AutoMixedPrecisionPass::SetVarPrecision() const {
           auto* real_in_var_node = real_vars_[in_var_node->Var()->Name()];
           auto in_var_name = real_in_var_node->Var()->Name();
 
-          if (!IsFloatType(real_in_var_node->Var()->GetDataType())) continue;
+          if (!IsFP32AndFP64(real_in_var_node->Var()->GetDataType())) continue;
           if (!VarNodeHasDtype(real_in_var_node)) continue;
           if (InputVarsNotConvert(op_node, in_var_name)) continue;
 
@@ -616,7 +641,7 @@ void AutoMixedPrecisionPass::SetVarPrecision() const {
           auto* real_out_var_node = real_vars_[out_var_node->Var()->Name()];
           auto out_var_name = real_out_var_node->Var()->Name();
 
-          if (!IsFloatType(real_out_var_node->Var()->GetDataType())) continue;
+          if (!IsFP32AndFP64(real_out_var_node->Var()->GetDataType())) continue;
           if (!VarNodeHasDtype(real_out_var_node)) continue;
           if (OutputVarsNotConvert(op_node, out_var_name)) continue;
 
@@ -656,7 +681,7 @@ void AutoMixedPrecisionPass::ConvertWeightsData() const {
   auto var_names = scope->LocalVarNames();
   for (const auto& var_name : var_names) {
     if (vars_convert_to_low_precision_.count(var_name)) {
-      VLOG(4) << var_name << "'s data type was convert to half";
+      VLOG(4) << var_name << "'s data type was convert to low precision";
 
       auto* var = scope->FindLocalVar(var_name);
       CHECK_EQ(var->IsType<phi::DenseTensor>(), true);
@@ -683,16 +708,18 @@ void AutoMixedPrecisionPass::ConvertWeightsData() const {
           }
         }
       } else if (low_precision_ == phi::DataType::BFLOAT16) {
-        auto* half_data =
+        auto* low_precision_data =
             low_precision_tensor.mutable_data<phi::dtype::bfloat16>(
                 phi::CPUPlace{});
         for (int64_t i = 0; i < origin_tensor->numel(); i++) {
           if (origin_tensor->dtype() == phi::DataType::FLOAT64) {
             auto* origin_data = origin_tensor->data<double>();
-            half_data[i] = static_cast<phi::dtype::bfloat16>(origin_data[i]);
+            low_precision_data[i] =
+                static_cast<phi::dtype::bfloat16>(origin_data[i]);
           } else if (origin_tensor->dtype() == phi::DataType::FLOAT32) {
             auto* origin_data = origin_tensor->data<float>();
-            half_data[i] = static_cast<phi::dtype::bfloat16>(origin_data[i]);
+            low_precision_data[i] =
+                static_cast<phi::dtype::bfloat16>(origin_data[i]);
           }
         }
       }
@@ -732,25 +759,44 @@ void AutoMixedPrecisionPass::InsertCastOp() const {
         VLOG(4) << "process var: " << real_in_var_node->Var()->Name()
                 << " with type " << in_var_type;
 
-        if (IsFloatType(in_var_type) && op_run_low_precision_.count(op_type)) {
-          DoInsertCastOp(subgraphes_[i],
-                         in_var_node,
-                         op_node,
-                         in_var_type,
-                         framework::TransToProtoVarType(low_precision_),
-                         block_desc,
-                         &suffix,
-                         &cache);
-        } else if (IsHalfType(in_var_type) &&
+        if (IsFP32AndFP64(in_var_type) &&
+            op_run_low_precision_.count(op_type)) {
+          auto to_type = framework::TransToProtoVarType(low_precision_);
+          auto* prev_op =
+              in_var_node->inputs.empty() ? nullptr : in_var_node->inputs[0];
+          if (prev_op && GetOpOriginalType(prev_op->Op()->Type()) == "cast") {
+            in_var_node->Var()->SetDataType(to_type);
+            prev_op->Op()->SetAttr("out_dtype", static_cast<int>(to_type));
+            prev_op->Op()->Flush();
+          } else {
+            DoInsertCastOp(subgraphes_[i],
+                           in_var_node,
+                           op_node,
+                           in_var_type,
+                           to_type,
+                           block_desc,
+                           &suffix,
+                           &cache);
+          }
+        } else if (IsFP16AndBFP16(in_var_type) &&
                    op_run_low_precision_.count(op_type) == 0) {
-          DoInsertCastOp(subgraphes_[i],
-                         in_var_node,
-                         op_node,
-                         in_var_type,
-                         VarType::FP32,
-                         block_desc,
-                         &suffix,
-                         &cache);
+          auto to_type = VarType::FP32;
+          auto* prev_op =
+              in_var_node->inputs.empty() ? nullptr : in_var_node->inputs[0];
+          if (prev_op && GetOpOriginalType(prev_op->Op()->Type()) == "cast") {
+            in_var_node->Var()->SetDataType(to_type);
+            prev_op->Op()->SetAttr("out_dtype", static_cast<int>(to_type));
+            prev_op->Op()->Flush();
+          } else {
+            DoInsertCastOp(subgraphes_[i],
+                           in_var_node,
+                           op_node,
+                           in_var_type,
+                           to_type,
+                           block_desc,
+                           &suffix,
+                           &cache);
+          }
         }
       }
 

paddle/fluid/inference/tests/api/gpu_ernie_half_test.cc

@@ -164,7 +164,7 @@ TEST(Ernie_gpu_fp16_no_ir, compare_results) {
     }
     float *result = reinterpret_cast<float *>(output.data.data());
     for (size_t j = 0; j < outputs_size; ++j) {
-      EXPECT_NEAR(ref[i * outputs_size + j], result[j], 5e-2);
+      EXPECT_NEAR(ref[i * outputs_size + j], result[j], 8e-3);
     }
   }
 }
@@ -175,8 +175,6 @@ TEST(Ernie_gpu_fp16_with_ir, compare_results) {
   config.SetModel(FLAGS_infer_model);
   config.EnableUseGpu(512, 0, paddle_infer::PrecisionType::kHalf);
   config.SwitchIrOptim(true);
-  // The fc_fuse_pass has diff, which will be repaired later.
-  config.pass_builder()->DeletePass("fc_fuse_pass");
   // There is a problem with the model itself, which has nothing to do with
   // constant_folding_pass.
   config.pass_builder()->DeletePass("constant_folding_pass");
@@ -206,7 +204,7 @@ TEST(Ernie_gpu_fp16_with_ir, compare_results) {
     }
     float *result = reinterpret_cast<float *>(output.data.data());
     for (size_t j = 0; j < outputs_size; ++j) {
-      EXPECT_NEAR(ref[i * outputs_size + j], result[j], 5e-2);
+      EXPECT_NEAR(ref[i * outputs_size + j], result[j], 2e-2);
     }
   }
 }
@@ -243,7 +241,7 @@ TEST(Ernie_gpu_bf16_no_ir, compare_results) {
     }
     float *result = reinterpret_cast<float *>(output.data.data());
     for (size_t j = 0; j < outputs_size; ++j) {
-      EXPECT_NEAR(ref[i * outputs_size + j], result[j], 7e-2);
+      EXPECT_NEAR(ref[i * outputs_size + j], result[j], 1e-2);
     }
   }
 }
@@ -254,8 +252,6 @@ TEST(Ernie_gpu_bf16_with_ir, compare_results) {
   config.SetModel(FLAGS_infer_model);
   config.EnableUseGpu(512, 0, paddle_infer::PrecisionType::kBf16);
   config.SwitchIrOptim(true);
-  // The fc_fuse_pass has diff, which will be repaired later.
-  config.pass_builder()->DeletePass("fc_fuse_pass");
   // There is a problem with the model itself, which has nothing to do with
   // constant_folding_pass.
   config.pass_builder()->DeletePass("constant_folding_pass");
@@ -285,7 +281,7 @@ TEST(Ernie_gpu_bf16_with_ir, compare_results) {
     }
     float *result = reinterpret_cast<float *>(output.data.data());
     for (size_t j = 0; j < outputs_size; ++j) {
-      EXPECT_NEAR(ref[i * outputs_size + j], result[j], 7e-2);
+      EXPECT_NEAR(ref[i * outputs_size + j], result[j], 5e-3);
     }
   }
 }

paddle/fluid/operators/fused/fused_fc_elementwise_layernorm_op.cu

@@ -223,13 +223,7 @@ __global__ void InplaceAddReluAddLayerNormKernel(const float16* y_data,
 
       // For layer_norm, reduce to calculate mean and std
       sum_i += static_cast<float>(tmp_3);
-#if defined(PADDLE_WITH_CUDA) && __CUDA_ARCH__ >= 530
-      square_sum_i += static_cast<float>(__hmul(tmp_3, tmp_3));
-#elif defined(PADDLE_WITH_CUDA)
       square_sum_i += static_cast<float>(tmp_3) * static_cast<float>(tmp_3);
-#else
-      square_sum_i += static_cast<float>(tmp_3 * tmp_3);
-#endif
     }
     auto pair = BlockReduce(temp_storage)
                     .Reduce(PairForLayerNorm<float>(sum_i, square_sum_i),
@@ -282,9 +276,9 @@ __global__ void InplaceAddReluAddLayerNormKernel(const float16* y_data,
       half tmp_0 = __hdiv(__hsub(save_ptr[save_index], mean_i), std_i);
       half tmp_1 = scale ? __hmul(scale[j], tmp_0) : tmp_0;
 #else
-      half tmp_0 = static_cast<float>(static_cast<float>(save_ptr[save_index]) +
-                                      static_cast<float>(mean_i) /
-                                          static_cast<float>(std_i));
+      half tmp_0 = static_cast<half>(static_cast<float>(save_ptr[save_index]) -
+                                     static_cast<float>(mean_i) /
+                                         static_cast<float>(std_i));
       half tmp_1 = scale ? static_cast<half>(static_cast<float>(scale[j]) *
                                              static_cast<float>(tmp_0))
                          : tmp_0;

paddle/phi/kernels/funcs/fc_functor.cu

@@ -149,7 +149,7 @@ __global__ void bias_relu_v2(const int num,
 #if __CUDA_ARCH__ >= 800
       packed_val = __hmax2(__half2(0, 0), packed_val);
 #elif __CUDA_ARCH__ >= 530
-      packed_val = __hmul2(__hgt2(__half2(0, 0), packed_val), packed_val);
+      packed_val = __hmul2(__hgt2(packed_val, __half2(0, 0)), packed_val);
 #else
       packed_val.x = static_cast<int>(static_cast<float>(packed_val.x) > 0) *
                      static_cast<float>(packed_val.x);