GPU Kernel Implementations for BLAKE3

.gitignore

@@ -1,9 +1,17 @@
 run.sh
+profile.sh
 
 **.__pycache__
 
 .clangd
 
 build/
 dist/
-*.egg-info/
+*.egg-info/
+
+*.ncu-rep
+*.nsys-rep
+
+*.ptx
+
+*.o

README.md

@@ -8,6 +8,11 @@ Proposal (from Google Drive): https://docs.google.com/document/d/10B3_nT8xF49vLg
 conda create -n ceg5206 python=3.12
 pip install -r requirements.txt
 
+# clone the cutlass repository
+git clone https://github.com/NVIDIA/cutlass.git ~/cutlass --depth 1
+
+export FLASHASHING_CUDA_ARCH_LIST=<your-arch-here>        # your arch here
+
 # install the cpp source file
 python setup.py install
 
@@ -19,4 +24,15 @@ python benchmark/test_script.py
 
 + 10.3: Implement `SHA256-SIMD` version, still some flaws: only allow short string processing (`len < 55`)
 + 10.2: Consultation with doctor, invite to repo, add `BLAKE3` basic implementations
-+ 10.1: Provide basic template for project, implementing basic sha256 in C++.
++ 10.1: Provide basic template for project, implementing basic sha256 in C++.
+
+# GPU kernel performance on File Compress hashing
+
+> This logs were tested on RTX 4090 Laptop GPU, computation / memory is limited, soon in the latter section we will see the perf on different machine.
+
++ 10.5 - v1 - [commit:4f31a2c55551965a2e5f048565f021c4551554ae]: 1709.34 MiB/s
++ 10.5 - v2 - [commit:2e71051367a9533af36f6c54a46876e20bc0bab5]: 2044.49 MiB/s
++ 10.6 - v3 - [commit:329d5425de9558e9e43216c2b89bed43674b7d83]: 201822.66 MiB/s (False Result)
++ 10.6 - v4 - [commit:9781ad9759d13e0e482162d57c68b65c946f4ff9]: 17069.23 MiB/s
++ 10.6 - v5 - [commit:4be8258f5e82aab4e57e8b70a604ebb9361d8aa0]: 34261.37 MiB/s
++ 10.8 - v6 - [commit:]: 15145.81 MiB/s