Merge branch 'main' into moflow

FlagOpen · Mar 6, 2024 · fc84f0d · fc84f0d
2 parents e019f7a + a67831d
commit fc84f0d
Show file tree

Hide file tree

Showing 20 changed files with 1,024 additions and 4 deletions.
diff --git a/inference/benchmarks/bertLarge/README.md b/inference/benchmarks/bertLarge/README.md
@@ -77,6 +77,19 @@ bert_reference_results_text_md5.txt
    - IXRT: ixrt-0.8.0+corex.3.2.1
 
 
+####  2.4 沐曦集成电路 C500 
+
+- ##### 硬件环境
+    - 机器、加速卡型号: 曦云®C500 64G
+- ##### 软件环境
+   - OS版本：Ubuntu 20.04.6
+   - OS kernel版本:  5.4.0-26-generic
+   - 加速卡驱动版本：2.2.0
+   - Docker 版本：24.0.7
+   - 推理框架版本：pytorch-2.0.0+mc2.18.0.8-cp38-cp38-linux_x86_64.whl
+
+
+
 ### 4. 运行情况（BERT-Large）
 
 * 指标列表
@@ -103,4 +116,6 @@ bert_reference_results_text_md5.txt
 | tensorrt | fp32   | 32 | 1868.8   | 150.4       | 152.2      | 190.4         | 194.1       | 42.0% | 0.638/0.638 | 16.9/40.0 |
 | kunlunxin_xtcl| W32A16   | 32 |/ | /          | /       | /          | /          | / | 0.638/0.638| /|
 | iluvatar_ixrt| fp16  | 32 |/ | /          | /       | /          | /          | / | 0.599/0.638| /|
+| metax-nocompiler| fp16   | 32 |/ | /          | /       | /          | /          | 27.6% | 0.638/0.638| 4.3/64.0|
+| metax-nocompiler| fp32   | 32 |/ | /          | /       | /          | /          | 28.1% | 0.639/0.638| 6.1/64.0|
 
diff --git a/inference/benchmarks/resnet50/README.md b/inference/benchmarks/resnet50/README.md
@@ -142,4 +142,5 @@ find ./val -name "*JPEG" | wc -l
 | kunlunxin_xtcl | fp32   | 128  | / | /  | / | /      | /      | / | 76.2/76.2 | 4.52/32.0 |
 | kunlunxin_xtcl | fp16   | 256  | / | /  | /   |  / | /  | / | 76.2/76.2 | 4.52/32.0 |
 | zixiao | fp16   | 32*6  | 261.103    | /      | /     |  193.151    | 6342.191  | / | 76.2/76.2 | / |
-
+| metax-nocompiler | fp16      | 256  |/ | /   | / | /   | / | 7.8% | 76.2/76.2 | 3.83/64.0 |
+| metax-nocompiler | fp32   | 256  | /    | /      | /     | /        | /  | 7.7% | 76.2/76.2 | 5.46/64.0 |
diff --git a/inference/benchmarks/swinTransformer/README.md b/inference/benchmarks/swinTransformer/README.md
@@ -84,4 +84,5 @@ find ./val -name "*JPEG" | wc -l
 | ----------- | --------- | ---- | ---- | -------- | ----------- | ---------- | ------------- | ------------ | ----------- | ----------- |
 | tensorrt | fp16      | 512 |1011.7 | 1347.5 | 1511.3 | 1231.7 | 1359.1 | 6.8% | 81.7/83.2 | 19.9/40.0 |
 | tensorrt | fp32   | 256 | 856.9 | 761.5 | 794.3 | 789.2 | 826.4 | 8.2% | 83.2/83.2 | 20.0/40.0 |
-| kunlunxin_xtcl| W32A16   | 256 | 543.745 | / | / | / | / | / | 0.832 | / |
+| kunlunxin_xtcl| W32A16   | 256 | / | / | / | / | / | / | 0.832 | / |
+| metax-nocompiler| fp16   | 512 | / | / | / | / | / | 6.5% | 0.832 |10.6/64.0 |
diff --git a/training/benchmarks/driver/helper.py b/training/benchmarks/driver/helper.py
@@ -88,6 +88,12 @@ def set_seed(self, seed: int, vendor: str = None):
             torch.backends.cudnn.benchmark = getattr(config, "cudnn_benchmark")
             torch.backends.cudnn.deterministic = getattr(
                 config, "cudnn_deterministic")
+        elif lower_vendor == "dcu":
+            import torch
+            torch.manual_seed(seed)
+            torch.cuda.manual_seed(seed)
+            torch.cuda.manual_seed_all(seed)
+            torch.backends.cudnn.benchmark = True
         else:
             # TODO 其他厂商设置seed，在此扩展
             pass

diff --git a/training/dcu/README.md b/training/dcu/README.md
@@ -0,0 +1,66 @@
+# 厂商信息
+
+海光 DCU 系列产品以 GPGPU 架构为基础，兼容通用的“类 CUDA” 环境以及国际主流商业计算软件和人工智能软件，软硬件生态丰富，可广泛应用于大数据处理、人工智能、商业计算等应用领域。
+
+海光 DCU 兼容“类 CUDA” 环境， 软硬件生态丰富，典型应用场景下性能指标达到国际上同类型高端产品的水平。
+
+海光 DCU 主要面向大数据处理、商业计算等计算密集型应用领域，以及人工智能、 泛人工智能类运算加速领域。
+
+# FlagPerf适配验证环境说明
+## 环境配置参考
+  - 硬件
+    - 机器型号：K100 标准机
+    - 加速卡型号: K100 64G
+  - 软件
+    - OS版本：centos 7.6
+    - OS kernel版本: 4.18.0-348.el8.0.2.x86_64
+    - Docker 版本: 24.0.7
+
+## 容器镜像信息
+- 容器构建信息
+  - Dockerfile路径：training/dcu/docker_image/\<framework\>/Dockerfile
+  - 构建后软件安装脚本: training/dcu/docker_image/\<framework\>/\<framework\>_install.sh
+
+- 核心软件信息
+
+  - AI框架&版本
+    - torch: 1.13.1
+
+  - 其它软件版本
+    - dtk: 23.10.1
+
+
+## 加速卡监控采集
+- 加速卡使用信息采集命令
+
+  dcu_monitor.py中79行需要修改为实际source的地址
+
+  ```
+  source /path/of/dtk/env.sh
+  rocm-smi
+  ```
+
+- 监控项示例：
+
+    ```
+    ============================ System Management Interface =============================
+    ======================================================================================
+    DCU     Temp     AvgPwr     Perf     PwrCap     VRAM%      DCU%      Mode     
+    0       53.0C    96.0W      auto     300.0W     0%         0%        Normal   
+    1       53.0C    96.0W      auto     300.0W     0%         0%        Normal   
+    2       54.0C    95.0W      auto     300.0W     0%         0%        Normal   
+    3       55.0C    96.0W      auto     300.0W     0%         0%        Normal   
+    4       54.0C    97.0W      auto     300.0W     0%         0%        Normal   
+    5       54.0C    95.0W      auto     300.0W     0%         0%        Normal   
+    6       55.0C    93.0W      auto     300.0W     0%         0%        Normal   
+    7       54.0C    96.0W      auto     300.0W     0%         0%        Normal   
+    ======================================================================================
+    =================================== End of SMI Log ===================================
+    ```
+
+- 加速卡使用信息采集项说明
+
+|监控项| 日志文件 |
+|---|---|
+|VRAM(%) | dcu_monitor.log |
+|DCU(%) | dcu_monitor.log |