fix performance regression

PaddlePaddle · Dec 2, 2022 · 5d82a37 · 5d82a37
1 parent 03c86be
commit 5d82a37
Show file tree

Hide file tree

Showing 4 changed files with 9 additions and 4 deletions.
diff --git a/paddle/fluid/framework/ir/mkldnn/compute_propagate_scales_mkldnn_pass.cc b/paddle/fluid/framework/ir/mkldnn/compute_propagate_scales_mkldnn_pass.cc
@@ -319,7 +319,7 @@ void ComputePropagateScalesMkldnnPass::ComputeWeightScales(
     ir::Graph* graph, Scope* scope, StringPairMap* var_quant_scales) const {
   ComputeVarScales(graph,
                    scope,
-                   {"conv2d", "depthwise_conv2d"},
+                   {"conv2d", "depthwise_conv2d", "fused_conv2d"},
                    "Filter",
                    1,
                    var_quant_scales);
@@ -446,7 +446,7 @@ void ComputePropagateScalesMkldnnPass::UpdateReluOutputScales(
     if (op->Type() == "relu") {
       is_unsigned = true;
     } else {
-      if (op->Type() == "conv2d") {
+      if (op->Type() == "conv2d" || op->Type() == "fused_conv2d") {
         act_name = "fuse_activation";
         output_name = "Output";
       } else if (op->Type() == "fc") {

diff --git a/paddle/fluid/framework/ir/mkldnn/cpu_quantize_placement_pass.cc b/paddle/fluid/framework/ir/mkldnn/cpu_quantize_placement_pass.cc
@@ -60,7 +60,6 @@ void CPUQuantizePlacementPass::ApplyImpl(ir::Graph* graph) const {
   auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
                      Graph* g) {
     GET_IR_NODE_FROM_SUBGRAPH(op, op, quantize_placement_pattern);
-
     if (std::find(excluded_ids_list.begin(),
                   excluded_ids_list.end(),
                   op->id()) != excluded_ids_list.end()) {
@@ -70,7 +69,6 @@ void CPUQuantizePlacementPass::ApplyImpl(ir::Graph* graph) const {
     if (op->Op()->GetAttrIfExists<int>("skip_quant") == 1) {
       return;
     }
-
     op->Op()->SetAttr("mkldnn_data_type", std::string("int8"));
   };
   gpd(graph, handler);

diff --git a/paddle/fluid/inference/api/mkldnn_quantizer_config.cc b/paddle/fluid/inference/api/mkldnn_quantizer_config.cc
@@ -26,6 +26,12 @@ MkldnnQuantizerConfig::MkldnnQuantizerConfig() {
   rules_["conv2d"]["ResidualData"] = ScaleAlgo::KL;
   rules_["conv2d"]["Output"] = ScaleAlgo::KL;
 
+  rules_["fused_conv2d"]["Input"] = ScaleAlgo::KL;
+  rules_["fused_conv2d"]["Filter"] = ScaleAlgo::MAX_CH;
+  rules_["fused_conv2d"]["Bias"] = ScaleAlgo::NONE;  // do not compute scale
+  rules_["fused_conv2d"]["ResidualData"] = ScaleAlgo::KL;
+  rules_["fused_conv2d"]["Output"] = ScaleAlgo::KL;
+
   rules_["pool2d"]["X"] = ScaleAlgo::KL;
   rules_["pool2d"]["Out"] = ScaleAlgo::KL;
 

diff --git a/paddle/fluid/inference/api/paddle_analysis_config.h b/paddle/fluid/inference/api/paddle_analysis_config.h
@@ -1151,6 +1151,7 @@ struct PD_INFER_DECL AnalysisConfig {
       "concat",
       "conv2d",
       "depthwise_conv2d",
+      "fused_conv2d",
       "elementwise_add",
       "elementwise_mul",
       "fc",