Optimize emb_eltwise_layernorm_plugin and support fp16 (#27128)

cmake/cuda.cmake

@@ -107,6 +107,9 @@ function(select_nvcc_arch_flags out_variable)
   elseif(${CUDA_ARCH_NAME} STREQUAL "Maxwell")
     set(cuda_arch_bin "50")
   elseif(${CUDA_ARCH_NAME} STREQUAL "Pascal")
+    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} LESS 10.0)
+      add_definitions("-DSUPPORTS_CUDA_FP16")
+    endif()
     set(cuda_arch_bin "60 61")
   elseif(${CUDA_ARCH_NAME} STREQUAL "Volta")
     if (NOT ${CMAKE_CUDA_COMPILER_VERSION} LESS 10.0)

paddle/fluid/inference/tensorrt/convert/emb_eltwise_layernorm.cc

@@ -80,10 +80,10 @@ class EmbEltwiseLayerNormOpConverter : public OpConverter {
     nvinfer1::ILayer* layer = nullptr;
 
     if (engine_->with_dynamic_shape()) {
-      plugin::DynamicPluginTensorRT* plugin = nullptr;
-      plugin = new plugin::EmbEltwiseLayernormPluginDynamic<float>(
+      auto use_fp16 = engine_->WithFp16();
+      auto plugin = new plugin::EmbEltwiseLayernormPluginDynamic(
           input_embs, bias, scale, emb_sizes, bias_size, scale_size, hidden,
-          eps);
+          eps, use_fp16);
       layer = engine_->AddPluginV2(input_ids.data(), input_num, plugin);
     } else {
       PADDLE_THROW(platform::errors::Fatal(

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

@@ -151,7 +151,7 @@ void trt_ernie(bool with_fp16, std::vector<float> result) {
   run(config, &out_data);         // serialize
   run(*config_deser, &out_data);  // deserialize
   for (size_t i = 0; i < out_data.size(); i++) {
-    EXPECT_NEAR(result[i], out_data[i], 1e-6);
+    EXPECT_NEAR(result[i], out_data[i], 1e-2);
   }
 }
 
@@ -159,13 +159,11 @@ TEST(AnalysisPredictor, no_fp16) {
   std::vector<float> result = {0.597841, 0.219972, 0.182187};
   trt_ernie(false, result);
 }
-
-TEST(AnalysisPredictor, fp16) {
 #ifdef SUPPORTS_CUDA_FP16
-  std::vector<float> result = {0.598336, 0.219558, 0.182106};
+TEST(AnalysisPredictor, fp16) {
+  std::vector<float> result = {0.59923654, 0.21923761, 0.18152587};
   trt_ernie(true, result);
-#endif
 }
-
+#endif  // SUPPORTS_CUDA_FP16
 }  // namespace inference
 }  // namespace paddle