fix translation for fake quantize ops (#65717)

kangguangli · web-flow · commit 170a1c7e0a96 · 2024-07-05T11:12:33.000+08:00
diff --git a/paddle/fluid/eager/auto_code_generator/eager_generator.cc b/paddle/fluid/eager/auto_code_generator/eager_generator.cc
@@ -1956,12 +1956,14 @@ static std::pair<std::string, std::string> GenerateForwardFunctionContents(
         } else {
           const char* FWD_OUT_TENSOR_TEMPLATE =
               "  egr::EagerUtils::GetOutput(outs[\"%s\"][0], %s);\n"
-              "  paddle::Tensor& %s = *%s;\n";
+              "  paddle::Tensor& %s = *%s;\n"
+              "  (void)%s; // To avoid error: unused variable\n";
           out_tensor_str = paddle::string::Sprintf(FWD_OUT_TENSOR_TEMPLATE,
                                                    output_name,
                                                    output_var_args_name,
                                                    output_varname,
-                                                   output_var_args_name);
+                                                   output_var_args_name,
+                                                   output_varname);
         }
       } else {
         if (!forward_inplace_map.empty() &&
diff --git a/paddle/fluid/ir_adaptor/translator/op_translator.cc b/paddle/fluid/ir_adaptor/translator/op_translator.cc
@@ -3208,32 +3208,20 @@ struct FusedElemwiseAddActivationGradOpTranscriber
   }
 };
 
-struct FakeQuantizeMovingAverageAbsMaxOpTranscriber : public OpTranscriber {
+// a more general version for fake quantize ops
+// if one has a more special property, then don't use this
+struct FakeQuantizeOpTranscriber : public OpTranscriber {
   void HandleNonexistentAttribute(pir::IrContext* ctx,
                                   pir::AttributeMap* attribute_map,
                                   const OpAttributeInfo& info) override {
     if (info.name == "round_type") {
       (*attribute_map)[info.name] = pir::Int32Attribute::get(ctx, 1);
-    }
-  }
-};
-
-struct FakeChannelWiseDequantizeMaxAbsOpTranscriber : public OpTranscriber {
-  void HandleNonexistentAttribute(pir::IrContext* ctx,
-                                  pir::AttributeMap* attribute_map,
-                                  const OpAttributeInfo& info) override {
-    if (info.name == "quant_axis") {
-      (*attribute_map)[info.name] = pir::Int32Attribute::get(ctx, 0);
-    }
-  }
-};
-
-struct FakeQuantizeRangeAbsMaxOpTranscriber : public OpTranscriber {
-  void HandleNonexistentAttribute(pir::IrContext* ctx,
-                                  pir::AttributeMap* attribute_map,
-                                  const OpAttributeInfo& info) override {
-    if (info.name == "round_type") {
+    } else if (info.name == "x_num_col_dims") {
       (*attribute_map)[info.name] = pir::Int32Attribute::get(ctx, 1);
+    } else if (info.name == "round_type") {
+      (*attribute_map)[info.name] = pir::Int32Attribute::get(ctx, 1);
+    } else if (info.name == "quant_axis") {
+      (*attribute_map)[info.name] = pir::Int32Attribute::get(ctx, 0);
     }
   }
 };
@@ -3615,11 +3603,12 @@ OpTranslator::OpTranslator() {
   special_handlers["fused_elemwise_add_activation_grad"] =
       FusedElemwiseAddActivationGradOpTranscriber();
   special_handlers["fake_quantize_moving_average_abs_max"] =
-      FakeQuantizeMovingAverageAbsMaxOpTranscriber();
+      FakeQuantizeOpTranscriber();
   special_handlers["fake_channel_wise_dequantize_max_abs"] =
-      FakeChannelWiseDequantizeMaxAbsOpTranscriber();
-  special_handlers["fake_quantize_range_abs_max"] =
-      FakeQuantizeRangeAbsMaxOpTranscriber();
+      FakeQuantizeOpTranscriber();
+  special_handlers["fake_quantize_range_abs_max"] = FakeQuantizeOpTranscriber();
+  special_handlers["fake_quantize_dequantize_moving_average_abs_max"] =
+      FakeQuantizeOpTranscriber();
   special_handlers["grad_add"] = GradAddOpTranscriber();
   special_handlers["increment"] = IncrementOpTranscriber();
   special_handlers["lookup_table_v2"] = EmbeddingOpTranscriber();
diff --git a/paddle/phi/ops/yaml/ops.yaml b/paddle/phi/ops/yaml/ops.yaml
@@ -1653,6 +1653,7 @@
     data_type : x
   optional : in_accum, in_state, out_state, out_accum
   backward : fake_quantize_dequantize_moving_average_abs_max_grad
+  inplace: (in_scale -> out_scale)
 
 - op : fake_quantize_moving_average_abs_max
   args : (Tensor x, Tensor in_scale, Tensor in_accum, Tensor in_state, float moving_rate = 0.9, int bit_length = 8, bool is_test = false, int round_type = 1)
@@ -1663,6 +1664,7 @@
     func : fake_quantize_moving_average_abs_max
     data_type : x
   optional : in_accum, in_state, out_state, out_accum
+  inplace: (in_scale -> out_scale)
 
 - op : fake_quantize_range_abs_max
   args : (Tensor x, Tensor in_scale, Tensor iter, int window_size = 10000,  int bit_length = 8, bool is_test = false, int round_type = 1)
@@ -1673,6 +1675,7 @@
     func : fake_quantize_range_abs_max
     data_type : x
   optional : iter, out_scales
+  inplace: (in_scale -> out_scale)
 
 - op : fft_c2c
   args : (Tensor x, int64_t[] axes, str normalization, bool forward)
diff --git a/test/cpp/inference/api/analyzer_quant_image_classification_tester.cc b/test/cpp/inference/api/analyzer_quant_image_classification_tester.cc
@@ -29,6 +29,10 @@ void SetConfig(AnalysisConfig *cfg, std::string model_path) {
   cfg->DisableGpu();
   cfg->SwitchIrOptim(true);
   cfg->SetCpuMathLibraryNumThreads(FLAGS_cpu_num_threads);
+  cfg->EnableNewIR();
+  cfg->EnableNewExecutor();
+  cfg->SetOptimizationLevel(3);
+
   if (FLAGS_enable_mkldnn) cfg->EnableMKLDNN();
 }
 
