optimize phi::CPUPlace() in paddle/fluid/framework/ir/ (#77198)

* optimize phi::CPUPlace() in paddle/fluid/framework/ir/

* fix

* fix

Author: co63oc

paddle/fluid/framework/ir/attention_lstm_fuse_pass.cc

@@ -297,7 +297,7 @@ void PrepareLSTMWeight(const phi::DenseTensor& W_forget_w0,
   out->Resize(common::make_ddim({D + M, 4 * D}));
   VLOG(3) << "LSTMWeight resized to " << out->dims();
 
-  float* out_data = out->mutable_data<float>(phi::CPUPlace());
+  float* out_data = out->mutable_data<float>(CPUPlace());
   std::array<const float*, 4> tensors{W_forget_w0.data<float>(),
                                       W_input_w0.data<float>(),
                                       W_output_w0.data<float>(),
@@ -342,7 +342,7 @@ void PrepareLSTMBias(const phi::DenseTensor& B_forget,
           B_forget.dims().size()));
   int D = static_cast<int>(B_forget.dims()[0]);
   out->Resize(common::make_ddim({1, 4 * D}));
-  auto* out_data = out->mutable_data<float>(phi::CPUPlace());
+  auto* out_data = out->mutable_data<float>(CPUPlace());
   for (size_t i = 0; i < tensors.size(); i++) {
     memcpy(out_data + D * i, tensors[i], D * sizeof(float));
   }

paddle/fluid/framework/ir/auto_mixed_precision_pass.cc

@@ -948,8 +948,7 @@ void AutoMixedPrecisionPass::ConvertWeightsData() const {
 
       if (low_precision_ == phi::DataType::FLOAT16) {
         auto* low_precision_data =
-            low_precision_tensor.mutable_data<phi::dtype::float16>(
-                phi::CPUPlace{});
+            low_precision_tensor.mutable_data<phi::dtype::float16>(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>();
@@ -963,8 +962,7 @@ void AutoMixedPrecisionPass::ConvertWeightsData() const {
         }
       } else if (low_precision_ == phi::DataType::BFLOAT16) {
         auto* low_precision_data =
-            low_precision_tensor.mutable_data<phi::dtype::bfloat16>(
-                phi::CPUPlace{});
+            low_precision_tensor.mutable_data<phi::dtype::bfloat16>(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>();
@@ -979,7 +977,7 @@ void AutoMixedPrecisionPass::ConvertWeightsData() const {
       }
       origin_tensor->clear();
       paddle::framework::TensorCopySync(
-          low_precision_tensor, phi::CPUPlace{}, origin_tensor);
+          low_precision_tensor, CPUPlace{}, origin_tensor);
     }
   }
 }

paddle/fluid/framework/ir/constant_folding_pass.cc

@@ -173,7 +173,7 @@ void ConstantFoldingPass::ApplyImpl(ir::Graph *graph) const {
         // useless out_node can be removed, not need set it persistable !
         if (out_node->outputs.empty()) remove_nodes.emplace(out_node);
       }
-      op->Run(*local_scope, phi::CPUPlace());
+      op->Run(*local_scope, CPUPlace());
       folded_op_num++;
       for (auto out_node : op_node->outputs) {
         // this out_node is useless, do not set it persistable

paddle/fluid/framework/ir/conv2d_trans_filter_dilations_nxn_to_1x1_pass.cc

@@ -125,7 +125,7 @@ void Conv2dTransFilterDilationsNxNTo1x1Pass::conv2d_dilation_trans(
         scope->Var(new_weights_name)->GetMutable<phi::DenseTensor>();
     new_weights->Resize({weights_shape[0], weights_shape[1], new_kh, new_kw});
     auto* cpu_ctx = static_cast<phi::CPUContext*>(
-        phi::DeviceContextPool::Instance().Get(phi::CPUPlace()));
+        phi::DeviceContextPool::Instance().Get(CPUPlace()));
     if (weights->dtype() == phi::DataType::FLOAT32) {
       auto weights_data = weights->data<float>();
       auto* new_weights_data = cpu_ctx->Alloc<float>(new_weights);

paddle/fluid/framework/ir/cost_model.cc

@@ -226,7 +226,7 @@ CostData CostModel::ProfileMeasure(
   std::string device_lower_case = ToLowerCopy(device);
   if (device_lower_case == "cpu") {
     profiler_state = platform::ProfilerState::kCPU;
-    place = phi::CPUPlace();
+    place = CPUPlace();
   } else if (device_lower_case == "gpu") {
     profiler_state = platform::ProfilerState::kAll;
     place = phi::GPUPlace();

paddle/fluid/framework/ir/delete_cast_op_pass.cc

@@ -692,7 +692,7 @@ int DeleteCastOpPass::ApplyCastLookupTablePass(ir::Graph* graph) const {
     lookup_table_w->Var()->SetDataType(proto::VarType::FP16);
     if (w_tensor->dtype() != phi::DataType::FLOAT16) {
       auto* cpu_ctx = static_cast<phi::CPUContext*>(
-          phi::DeviceContextPool::Instance().Get(phi::CPUPlace()));
+          phi::DeviceContextPool::Instance().Get(CPUPlace()));
       phi::DenseTensor w_fp32_tensor;
       w_fp32_tensor.Resize(w_tensor->dims());
       w_fp32_tensor.set_type(w_tensor->dtype());

paddle/fluid/framework/ir/delete_quant_dequant_filter_op_pass.cc

@@ -128,7 +128,7 @@ void DeleteQuantDequantFilterOpPass::ApplyImpl(ir::Graph* graph) const {
     auto w_dims = weight_tensor->dims();
 
     float* quantized_weight_data =
-        weight_tensor->mutable_data<float>(phi::CPUPlace());
+        weight_tensor->mutable_data<float>(CPUPlace());
 
     // Get weight scale
     if (dequant_type == "fake_channel_wise_quantize_dequantize_abs_max") {

paddle/fluid/framework/ir/embedding_fc_lstm_fuse_pass.cc

@@ -96,8 +96,7 @@ static int BuildFusion(Graph* graph,
     // Multiply embeddings via WeightsX and add bias
     auto embedding_data = embedding_tensor.data<float>();
     auto weightx_data = weightx_tensor.data<float>();
-    auto embeddings_data =
-        embeddings_tensor->mutable_data<float>(phi::CPUPlace());
+    auto embeddings_data = embeddings_tensor->mutable_data<float>(CPUPlace());
 
     // Adding biases to GEMM result to be
     auto* lstm_bias_var = scope->FindVar(bias->Name());

paddle/fluid/framework/ir/fc_fuse_pass.cc

@@ -210,7 +210,7 @@ int FCFusePass::ApplyFCPattern(Graph* graph, bool with_relu) const {
                  w_w * sizeof(float));
         }
         w_tensor->Resize(DDim{weight_dims[0] + 4, weight_dims[1] + 4});
-        auto* weight_data_new = w_tensor->mutable_data<float>(phi::CPUPlace());
+        auto* weight_data_new = w_tensor->mutable_data<float>(CPUPlace());
         for (int i = 0; i < w_h; i++) {
           memcpy(weight_data_new + i * (w_w + 4),
                  weight_data_tmp + i * w_w,

paddle/fluid/framework/ir/fc_gru_fuse_pass.cc

@@ -232,8 +232,7 @@ int FCGRUFusePass::BuildFusion(Graph* graph,
           common::errors::PreconditionNotMet(
               "GRU and FC biases have to have equal number of elements."));
 
-      auto gru_bias_data =
-          gru_bias_tensor->mutable_data<float>(phi::CPUPlace());
+      auto gru_bias_data = gru_bias_tensor->mutable_data<float>(CPUPlace());
       auto* fc_bias_data = fc_bias_tensor->data<float>();
 
       // Recompute GRU bias