Update README.md

README.md

@@ -8,6 +8,39 @@ export LD_LIBRARY_PATH=/opt/llvm/llvm-14.x-install/lib:$LD_LIBRARY_PATH
 make run
 ```
 
-#### TODOS:
-- reduce redundant data copy
-- optimize matrix multiplication
+#### Reproductivity
+*Step 0: import data*  
+`MNIST` dataset needs to be imported to this repo and the file struture looks like:
+```
+`-- MNIST
+    |-- train
+    |   |-- 0
+    |   |   |-- aaaa.jpg
+    |   |   |-- bbbb.jpg
+    |   |   |   ...
+    |   |   `-- zzzz.jpg
+    ...
+    `-- test
+        |-- 0
+        |   |-- AAAA.jpg
+        |   |-- BBBB.jpg
+        |   |   ...
+        |   `-- ZZZZ.jpg
+        ...
+```
+
+*Step 1: create executable*  
+For OpenMP CPU, OpenMP GPU and cuDNN implemented CNN, we type `make omp-cpu`, `make`, and `make cudnn` to generate `main-omp-cpu`, `main` and `main-cudnn`, accordingly.  
+
+*Step 2: run executable*  
+The usage for executable is: `./main <training images> <testing images> <batches> <epochs> <devices>`.  
+Among all of the test cases, we use full MNIST dataset and 100 epochs, then, for instance, for experiment of OpenMP GPU version with batch size of 1k, we type:
+```
+./main 60000 10000 60 100 1
+```
+For rest of the tests, with batch sizes of 1k, 2k, 4k, 10k, 15k, 30k, 60k, we change the `<batches>` to be 60, 30, 15, 6, 4, 2, and 1.
+
+*Step 3: run nvprof*  
+```
+nvprof ./main 60000 10000 60 100 1
+```