Merge branch 'develop' of https://github.com/PaddlePaddle/FastDeploy …

…into release/1.0.5

FastDeploy.cmake.in

@@ -321,6 +321,8 @@ if(ENABLE_VISION)
     find_library(CVCUDA_LIB cvcuda ${CMAKE_CURRENT_LIST_DIR}/third_libs/install/cvcuda/lib NO_DEFAULT_PATH)
     find_library(NVCV_TYPES_LIB nvcv_types ${CMAKE_CURRENT_LIST_DIR}/third_libs/install/cvcuda/lib NO_DEFAULT_PATH)
     list(APPEND FASTDEPLOY_LIBS ${CVCUDA_LIB} ${NVCV_TYPES_LIB})
+    list(APPEND FASTDEPLOY_INCS ${CMAKE_CURRENT_LIST_DIR}/third_libs/install/cvcuda/include NO_DEFAULT_PATH)
+    add_definitions(-DENABLE_CVCUDA)
   endif()
 
 endif()

benchmark/cpp/benchmark_gpu_trt.sh

@@ -8,39 +8,39 @@ if [ ! "$1" = "$CONFIG_PATH" ]; then
 fi
 
 # PaddleClas
-./benchmark_ppcls --model PPLCNet_x1_0_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH
-./benchmark_ppcls --model PPLCNetV2_base_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH
-./benchmark_ppcls --model EfficientNetB7_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH
-./benchmark_ppcls --model EfficientNetB0_small_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH
-./benchmark_ppcls --model GhostNet_x0_5_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH
-./benchmark_ppcls --model GhostNet_x1_3_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH
-./benchmark_ppcls --model GhostNet_x1_3_ssld_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH
-./benchmark_ppcls --model MobileNetV1_x0_25_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH
-./benchmark_ppcls --model MobileNetV1_ssld_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH
-./benchmark_ppcls --model MobileNetV2_x0_25_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH
-./benchmark_ppcls --model MobileNetV2_ssld_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH
-./benchmark_ppcls --model MobileNetV3_small_x0_35_ssld_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH
-./benchmark_ppcls --model MobileNetV3_large_x1_0_ssld_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH
-./benchmark_ppcls --model ShuffleNetV2_x0_25_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH
-./benchmark_ppcls --model ShuffleNetV2_x2_0_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH
-./benchmark_ppcls --model SqueezeNet1_1_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH
-./benchmark_ppcls --model InceptionV3_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH
-./benchmark_ppcls --model ResNet50_vd_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH
-./benchmark_ppcls --model PPHGNet_tiny_ssld_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH
-./benchmark_ppcls --model PPHGNet_base_ssld_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH
-./benchmark_ppcls --model ResNet50_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH
-./benchmark_ppcls --model EfficientNetB0_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH
-./benchmark_ppcls --model MobileNetV2_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH
-./benchmark_ppcls --model MobileNetV3_small_x1_0_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH
-./benchmark_ppcls --model ViT_large_patch16_224_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH
-./benchmark_ppcls --model ResNeXt50_32x4d_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH
-./benchmark_ppcls --model DenseNet121_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH
-./benchmark_ppcls --model PPHGNet_small_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH
-./benchmark_ppcls --model person_exists_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH
+./benchmark_ppcls --model PPLCNet_x1_0_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH --trt_shape "1,3,224,224:1,3,224,224:1,3,224,224" --input_name "x"
+./benchmark_ppcls --model PPLCNetV2_base_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH --trt_shape "1,3,224,224:1,3,224,224:1,3,224,224" --input_name "x"
+./benchmark_ppcls --model EfficientNetB7_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH --trt_shape "1,3,600,600:1,3,600,600:1,3,600,600" --input_name "x"
+./benchmark_ppcls --model EfficientNetB0_small_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH --trt_shape "1,3,224,224:1,3,224,224:1,3,224,224" --input_name "inputs"
+./benchmark_ppcls --model GhostNet_x0_5_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH --trt_shape "1,3,224,224:1,3,224,224:1,3,224,224" --input_name "inputs"
+./benchmark_ppcls --model GhostNet_x1_3_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH --trt_shape "1,3,224,224:1,3,224,224:1,3,224,224" --input_name "x"
+./benchmark_ppcls --model GhostNet_x1_3_ssld_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH --trt_shape "1,3,224,224:1,3,224,224:1,3,224,224" --input_name "inputs"
+./benchmark_ppcls --model MobileNetV1_x0_25_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH --trt_shape "1,3,224,224:1,3,224,224:1,3,224,224" --input_name "inputs"
+./benchmark_ppcls --model MobileNetV1_ssld_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH --trt_shape "1,3,224,224:1,3,224,224:1,3,224,224" --input_name "inputs"
+./benchmark_ppcls --model MobileNetV2_x0_25_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH --trt_shape "1,3,224,224:1,3,224,224:1,3,224,224" --input_name "inputs"
+./benchmark_ppcls --model MobileNetV2_ssld_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH --trt_shape "1,3,224,224:1,3,224,224:1,3,224,224" --input_name "inputs"
+./benchmark_ppcls --model MobileNetV3_small_x0_35_ssld_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH --trt_shape "1,3,224,224:1,3,224,224:1,3,224,224" --input_name "x"
+./benchmark_ppcls --model MobileNetV3_large_x1_0_ssld_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH --trt_shape "1,3,224,224:1,3,224,224:1,3,224,224" --input_name "inputs"
+./benchmark_ppcls --model ShuffleNetV2_x0_25_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH --trt_shape "1,3,224,224:1,3,224,224:1,3,224,224" --input_name "inputs"
+./benchmark_ppcls --model ShuffleNetV2_x2_0_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH --trt_shape "1,3,224,224:1,3,224,224:1,3,224,224" --input_name "inputs"
+./benchmark_ppcls --model SqueezeNet1_1_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH --trt_shape "1,3,224,224:1,3,224,224:1,3,224,224" --input_name "inputs"
+./benchmark_ppcls --model InceptionV3_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH --trt_shape "1,3,299,299:1,3,299,299:1,3,299,299" --input_name "x"
+./benchmark_ppcls --model ResNet50_vd_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH --trt_shape "1,3,224,224:1,3,224,224:1,3,224,224" --input_name "inputs"
+./benchmark_ppcls --model PPHGNet_tiny_ssld_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH --trt_shape "1,3,224,224:1,3,224,224:1,3,224,224" --input_name "x"
+./benchmark_ppcls --model PPHGNet_base_ssld_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH --trt_shape "1,3,224,224:1,3,224,224:1,3,224,224" --input_name "x"
+./benchmark_ppcls --model ResNet50_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH --trt_shape "1,3,224,224:1,3,224,224:1,3,224,224" --input_name "inputs"
+./benchmark_ppcls --model EfficientNetB0_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH --trt_shape "1,3,224,224:1,3,224,224:1,3,224,224" --input_name "x"
+./benchmark_ppcls --model MobileNetV2_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH --trt_shape "1,3,224,224:1,3,224,224:1,3,224,224" --input_name "inputs"
+./benchmark_ppcls --model MobileNetV3_small_x1_0_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH --trt_shape "1,3,224,224:1,3,224,224:1,3,224,224" --input_name "inputs"
+./benchmark_ppcls --model ViT_large_patch16_224_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH --trt_shape "1,3,224,224:1,3,224,224:1,3,224,224" --input_name "inputs"
+./benchmark_ppcls --model ResNeXt50_32x4d_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH --trt_shape "1,3,224,224:1,3,224,224:1,3,224,224" --input_name "inputs"
+./benchmark_ppcls --model DenseNet121_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH --trt_shape "1,3,224,224:1,3,224,224:1,3,224,224" --input_name "inputs"
+./benchmark_ppcls --model PPHGNet_small_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH --trt_shape "1,3,224,224:1,3,224,224:1,3,224,224" --input_name "x"
+./benchmark_ppcls --model person_exists_infer --image ILSVRC2012_val_00000010.jpeg --config_path $CONFIG_PATH --trt_shape "1,3,224,224:1,3,224,224:1,3,224,224" --input_name "x"
 
 # PaddleOCR
 ./benchmark_ppocr_det --model ch_PP-OCRv3_det_infer --image 12.jpg --config_path $CONFIG_PATH --trt_shape 1,3,64,64:1,3,640,640:1,3,960,960
-./benchmark_ppocr_cls --model ch_ppocr_mobile_v2.0_cls_infer --image rec_img.jpg --config_path $CONFIG_PATH --trt_shape 1,3,48,10:4,3,48,320:8,3,48,1024
+./benchmark_ppocr_cls --model ch_ppocr_mobile_v2.0_cls_infer --image rec_img.jpg --config_path $CONFIG_PATH --trt_shape 1,3,48,192:1,3,48,192:1,3,48,192
 ./benchmark_ppocr_rec --model ch_PP-OCRv3_rec_infer --image rec_img.jpg --rec_label_file ppocr_keys_v1.txt --config_path $CONFIG_PATH --trt_shape 1,3,48,10:4,3,48,320:8,3,48,2304
 ./benchmark_ppocr_det --model ch_PP-OCRv2_det_infer --image 12.jpg --config_path $CONFIG_PATH --trt_shape 1,3,64,64:1,3,640,640:1,3,960,960
 ./benchmark_ppocr_rec --model ch_PP-OCRv2_rec_infer --image rec_img.jpg --rec_label_file ppocr_keys_v1.txt --config_path $CONFIG_PATH --trt_shape 1,3,32,10:4,3,32,320:8,3,32,2304

benchmark/cpp/benchmark_picodet.cc

@@ -42,6 +42,16 @@ int main(int argc, char* argv[]) {
   auto model_file = FLAGS_model + sep + model_name;
   auto params_file = FLAGS_model + sep + params_name;
   auto config_file = FLAGS_model + sep + config_name;
+  if (config_info["backend"] == "paddle_trt") {
+    option.paddle_infer_option.collect_trt_shape = true;
+  }
+  if (config_info["backend"] == "paddle_trt" ||
+      config_info["backend"] == "trt") {
+    option.trt_option.SetShape("image", {1, 3, 640, 640}, {1, 3, 640, 640},
+                               {1, 3, 640, 640});
+    option.trt_option.SetShape("scale_factor", {1, 2}, {1, 2},
+                               {1, 2});
+  }
   auto model_picodet = vision::detection::PicoDet(
       model_file, params_file, config_file, option, model_format);
   if (FLAGS_no_nms) {

benchmark/cpp/benchmark_ppcls.cc

@@ -11,14 +11,22 @@
 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 // See the License for the specific language governing permissions and
 // limitations under the License.
-
+ 
 #include "flags.h"
 #include "macros.h"
 #include "option.h"
 
 namespace vision = fastdeploy::vision;
 namespace benchmark = fastdeploy::benchmark;
 
+DEFINE_string(trt_shape, "1,3,224,224:1,3,224,224:1,3,224,224",
+              "Set min/opt/max shape for trt/paddle_trt backend."
+              "eg:--trt_shape 1,3,224,224:1,3,224,224:1,3,224,224");
+
+DEFINE_string(input_name, "x",
+              "Set input name for trt/paddle_trt backend."
+              "eg:--input_names x");
+
 int main(int argc, char* argv[]) {
 #if defined(ENABLE_BENCHMARK) && defined(ENABLE_VISION)
   // Initialization
@@ -44,6 +52,17 @@ int main(int argc, char* argv[]) {
   auto model_file = FLAGS_model + sep + model_name;
   auto params_file = FLAGS_model + sep + params_name;
   auto config_file = FLAGS_model + sep + config_name;
+  if (config_info["backend"] == "paddle_trt") {
+    option.paddle_infer_option.collect_trt_shape = true;
+  }
+  if (config_info["backend"] == "paddle_trt" ||
+      config_info["backend"] == "trt") {
+    std::vector<std::vector<int32_t>> trt_shapes =
+        benchmark::ResultManager::GetInputShapes(FLAGS_trt_shape);
+    option.trt_option.SetShape(FLAGS_input_name, trt_shapes[0], 
+                               trt_shapes[1], trt_shapes[2]);
+  }
+
   auto model_ppcls = vision::classification::PaddleClasModel(
       model_file, params_file, config_file, option, model_format);
   vision::ClassifyResult res;

benchmark/cpp/benchmark_ppyoloe.cc

@@ -41,6 +41,16 @@ int main(int argc, char* argv[]) {
   auto model_file = FLAGS_model + sep + model_name;
   auto params_file = FLAGS_model + sep + params_name;
   auto config_file = FLAGS_model + sep + config_name;
+  if (config_info["backend"] == "paddle_trt") {
+    option.paddle_infer_option.collect_trt_shape = true;
+  }
+  if (config_info["backend"] == "paddle_trt" ||
+      config_info["backend"] == "trt") {
+    option.trt_option.SetShape("image", {1, 3, 640, 640}, {1, 3, 640, 640},
+                               {1, 3, 640, 640});
+    option.trt_option.SetShape("scale_factor", {1, 2}, {1, 2},
+                               {1, 2});
+  }
   auto model_ppyoloe = vision::detection::PPYOLOE(
       model_file, params_file, config_file, option, model_format);
   if (FLAGS_no_nms) {

benchmark/cpp/benchmark_ppyolov5.cc

@@ -41,6 +41,16 @@ int main(int argc, char* argv[]) {
   auto model_file = FLAGS_model + sep + model_name;
   auto params_file = FLAGS_model + sep + params_name;
   auto config_file = FLAGS_model + sep + config_name;
+  if (config_info["backend"] == "paddle_trt") {
+    option.paddle_infer_option.collect_trt_shape = true;
+  }
+  if (config_info["backend"] == "paddle_trt" ||
+      config_info["backend"] == "trt") {
+    option.trt_option.SetShape("image", {1, 3, 640, 640}, {1, 3, 640, 640},
+                               {1, 3, 640, 640});
+    option.trt_option.SetShape("scale_factor", {1, 2}, {1, 2},
+                               {1, 2});
+  }
   auto model_ppyolov5 = vision::detection::PaddleYOLOv5(
       model_file, params_file, config_file, option, model_format);
   vision::DetectionResult res;

benchmark/cpp/benchmark_ppyolov6.cc

@@ -41,6 +41,16 @@ int main(int argc, char* argv[]) {
   auto model_file = FLAGS_model + sep + model_name;
   auto params_file = FLAGS_model + sep + params_name;
   auto config_file = FLAGS_model + sep + config_name;
+  if (config_info["backend"] == "paddle_trt") {
+    option.paddle_infer_option.collect_trt_shape = true;
+  }
+  if (config_info["backend"] == "paddle_trt" ||
+      config_info["backend"] == "trt") {
+    option.trt_option.SetShape("image", {1, 3, 640, 640}, {1, 3, 640, 640},
+                               {1, 3, 640, 640});
+    option.trt_option.SetShape("scale_factor", {1, 2}, {1, 2},
+                               {1, 2});
+  }
   auto model_ppyolov6 = vision::detection::PaddleYOLOv6(
       model_file, params_file, config_file, option, model_format);
   vision::DetectionResult res;

benchmark/cpp/benchmark_ppyolov7.cc

@@ -41,6 +41,16 @@ int main(int argc, char* argv[]) {
   auto model_file = FLAGS_model + sep + model_name;
   auto params_file = FLAGS_model + sep + params_name;
   auto config_file = FLAGS_model + sep + config_name;
+  if (config_info["backend"] == "paddle_trt") {
+    option.paddle_infer_option.collect_trt_shape = true;
+  }
+  if (config_info["backend"] == "paddle_trt" ||
+      config_info["backend"] == "trt") {
+    option.trt_option.SetShape("image", {1, 3, 640, 640}, {1, 3, 640, 640},
+                               {1, 3, 640, 640});
+    option.trt_option.SetShape("scale_factor", {1, 2}, {1, 2},
+                               {1, 2});
+  }
   auto model_ppyolov7 = vision::detection::PaddleYOLOv7(
       model_file, params_file, config_file, option, model_format);
   vision::DetectionResult res;

benchmark/cpp/benchmark_ppyolov8.cc

@@ -41,6 +41,16 @@ int main(int argc, char* argv[]) {
   auto model_file = FLAGS_model + sep + model_name;
   auto params_file = FLAGS_model + sep + params_name;
   auto config_file = FLAGS_model + sep + config_name;
+  if (config_info["backend"] == "paddle_trt") {
+    option.paddle_infer_option.collect_trt_shape = true;
+  }
+  if (config_info["backend"] == "paddle_trt" ||
+      config_info["backend"] == "trt") {
+    option.trt_option.SetShape("image", {1, 3, 640, 640}, {1, 3, 640, 640},
+                               {1, 3, 640, 640});
+    option.trt_option.SetShape("scale_factor", {1, 2}, {1, 2},
+                               {1, 2});
+  }
   auto model_ppyolov8 = vision::detection::PaddleYOLOv8(
       model_file, params_file, config_file, option, model_format);
   vision::DetectionResult res;

benchmark/cpp/benchmark_ppyolox.cc

@@ -41,6 +41,16 @@ int main(int argc, char* argv[]) {
   auto model_file = FLAGS_model + sep + model_name;
   auto params_file = FLAGS_model + sep + params_name;
   auto config_file = FLAGS_model + sep + config_name;
+  if (config_info["backend"] == "paddle_trt") {
+    option.paddle_infer_option.collect_trt_shape = true;
+  }
+  if (config_info["backend"] == "paddle_trt" ||
+      config_info["backend"] == "trt") {
+    option.trt_option.SetShape("image", {1, 3, 640, 640}, {1, 3, 640, 640},
+                               {1, 3, 640, 640});
+    option.trt_option.SetShape("scale_factor", {1, 2}, {1, 2},
+                               {1, 2});
+  }
   auto model_ppyolox = vision::detection::PaddleYOLOX(
       model_file, params_file, config_file, option, model_format);
   vision::DetectionResult res;

benchmark/cpp/benchmark_x86.sh

@@ -62,9 +62,9 @@ fi
 ./benchmark_ppseg --model PP_LiteSeg_B_STDC2_cityscapes_with_argmax_infer --image cityscapes_demo.png --config_path $CONFIG_PATH
 ./benchmark_ppseg --model FCN_HRNet_W18_cityscapes_with_argmax_infer --image cityscapes_demo.png --config_path $CONFIG_PATH
 ./benchmark_ppseg --model SegFormer_B0-cityscapes-with-argmax --image cityscapes_demo.png --config_path $CONFIG_PATH
-./benchmark_ppseg --model Deeplabv3_ResNet101_OS8_cityscapes_with_argmax_infer --image cityscapes_demo.png --config_path $CONFIG_PATH
+./benchmark_ppseg --model Deeplabv3_ResNet101_OS8_cityscapes_with_argmax_infer --image cityscapes_demo.png --warmup 10 --repeat 50 --config_path $CONFIG_PATH
 ./benchmark_ppseg --model Unet_cityscapes_with_argmax_infer --image cityscapes_demo.png --config_path $CONFIG_PATH
-./benchmark_ppseg --model PP_HumanSegV1_Server_with_argmax_infer --image portrait_heng.jpg --config_path $CONFIG_PATH
-./benchmark_ppmatting --model PP-Matting-512 --image matting_input.jpg --config_path $CONFIG_PATH
-./benchmark_ppmatting --model PPHumanMatting --image matting_input.jpg --config_path $CONFIG_PATH
+./benchmark_ppseg --model PP_HumanSegV1_Server_with_argmax_infer --image portrait_heng.jpg --warmup 10 --repeat 50 --config_path $CONFIG_PATH
+./benchmark_ppmatting --model PP-Matting-512 --image matting_input.jpg --warmup 10 --repeat 50 --config_path $CONFIG_PATH
+./benchmark_ppmatting --model PPHumanMatting --image matting_input.jpg --warmup 10 --repeat 50 --config_path $CONFIG_PATH
 ./benchmark_ppmatting --model PPModnet_MobileNetV2 --image matting_input.jpg --config_path $CONFIG_PATH

benchmark/cpp/config/config.gpu.ort.fp32.e2e.txt

@@ -1,8 +1,8 @@
 device: gpu
 device_id: 3
 cpu_thread_nums: 1
-warmup: 20
-repeat: 100
+warmup: 200
+repeat: 1000
 backend: ort
 profile_mode: end2end
 include_h2d_d2h: false

benchmark/cpp/config/config.gpu.ort.fp32.txt

@@ -1,8 +1,8 @@
 device: gpu
 device_id: 3
 cpu_thread_nums: 1
-warmup: 20
-repeat: 100
+warmup: 200
+repeat: 1000
 backend: ort
 profile_mode: runtime
 include_h2d_d2h: false

benchmark/cpp/config/config.gpu.paddle.fp32.e2e.txt

@@ -1,8 +1,8 @@
 device: gpu
 device_id: 3
 cpu_thread_nums: 1
-warmup: 20
-repeat: 100
+warmup: 200
+repeat: 1000
 backend: paddle
 profile_mode: end2end
 include_h2d_d2h: false

benchmark/cpp/config/config.gpu.paddle.fp32.txt

@@ -1,8 +1,8 @@
 device: gpu
 device_id: 3
 cpu_thread_nums: 1
-warmup: 20
-repeat: 100
+warmup: 200
+repeat: 1000
 backend: paddle
 profile_mode: runtime
 include_h2d_d2h: false

benchmark/cpp/config/config.gpu.paddle_trt.fp16.h2d.txt

@@ -0,0 +1,14 @@
+device: gpu
+device_id: 3
+cpu_thread_nums: 1
+warmup: 200
+repeat: 1000
+backend: paddle_trt
+profile_mode: runtime
+include_h2d_d2h: true
+use_fp16: true
+collect_memory_info: false
+sampling_interval: 1
+precision_compare: false
+xpu_l3_cache: 0
+result_path: benchmark_gpu_paddle_trt_fp16_h2d.txt