Add ViT model for FlagPerf (#200)

* Add ViT model

* update the script based on zhiyuan's model

* Update script based on PR review

* Update ViT performance in README.md

inference/benchmarks/vit_l_16/README.md

@@ -83,4 +83,5 @@ find ./val -name "*JPEG" | wc -l
 | ----------- | --------- | ---- | ---- | -------- | ----------- | ---------- | ------------- | ------------ | ----------- | ----------- |
 | tensorrt | fp16    | 64   |1009.7 | 777.8 | 796.7 | 825.8 | 1329.2 | 26.2% | 79.0/79.3 | 35.0/40.0 |
 | tensorrt | fp32   | 32 | 1275.9 | 482.4  | 491.1 | 555.5    | 590.5 | 23.3% | 79.3/79.3 | 35.0/40.0 |
+| kunlunxin_xtcl | W32A16   | 32 | 2118.307 | / | / | 130.006    | 144.914 | 27.9% | 79.3/79.3 | / |
 

inference/configs/host.yaml

@@ -13,4 +13,4 @@ PIP_SOURCE: "https://mirror.baidu.com/pypi/simple"
 CLEAR_CACHES: True
 ACCE_VISIBLE_DEVICE_ENV_NAME: "CUDA_VISIBLE_DEVICES"
 CASES: 
-    "resnet50:pytorch_1.13": "/raid/dataset/ImageNet/imagenet/val"
+    "resnet50:pytorch_1.13": "/raid/dataset/ImageNet/imagenet/val"

inference/configs/vit_l_16/vendor_config/kunlunxin_configurations.yaml

@@ -0,0 +1,5 @@
+compiler: xtcl
+# skip validation(will also skip create_model, export onnx). Assert exist_onnx_path != null
+no_validation: true
+# set a real onnx_path to use exist, or set it to anything but null to avoid export onnx manually(like torch-tensorrt)
+exist_onnx_path: /home/FlagPerf/inference/onnxs/vit_l_16_bs32_pytorch_fp16False.onnx

inference/docker_images/kunlunxin/kunlunxin_analysis.py

@@ -1,23 +1,21 @@
-def analysis_log(logpath):
-    logfile = open(logpath)
-
-    max_usage = 0.0 ## usage_mem
-    max_mem = 0.0 
-    for line in logfile.readlines():
-        '''
-        xpu_smi temp power mem w_mem use_rate
-        '''
-        if "xpu_smi" in line:
-            line = line[:-1]
-            usage = line.split(" ")[4]
-            usage = float(usage)
-            max_usage = max(max_usage, usage)
-            max_mem = line.split(" ")[5]
-            max_mem = float(max_mem)
-
-    return round(max_usage / 1024.0,
-                 2), round(max_mem / 1024.0, 2), eval("32e12"), eval("128e12")
-
-
-if __name__ == "__main__":
-    max1, max2, max2,max4 = analysis_log("/home/zhoujiamin01/workspace/zjm_flag/FlagPerf/inference/result/run20230809192313/resnet50:pytorch_1.13/127.0.0.1_noderank0/kunlunxin_monitor.log")
+def analysis_log(logpath):
+    logfile = open(logpath)
+
+    max_usage = 0.0 ## usage_mem
+    max_mem = 0.0 
+    for line in logfile.readlines():
+        '''
+        xpu_smi temp power mem w_mem use_rate
+        '''
+        if "xpu_smi" in line:
+            line = line[:-1]
+            usage = line.split(" ")[4]
+            usage = float(usage)
+            max_usage = max(max_usage, usage)
+            max_mem = line.split(" ")[5]
+            max_mem = float(max_mem)
+
+    return round(max_usage / 1024.0,
+                 2), round(max_mem / 1024.0, 2), eval("32e12"), eval("128e12")
+
+