Skip to content

masa-cirkovic: Implemented CUDA matrix multiplication #18

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
masa-byte wants to merge 1 commit into
base: main
Choose a base branch
from
Open

masa-cirkovic: Implemented CUDA matrix multiplication #18

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
9 changes: 8 additions & 1 deletion CMakeLists.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,15 +2,22 @@ cmake_minimum_required(VERSION 3.18)
project(app LANGUAGES CXX CUDA)

# Set C++ and CUDA standards
set(CMAKE_CXX_STANDARD 17)
set(CMAKE_CXX_STANDARD 20)
set(CMAKE_CXX_STANDARD_REQUIRED ON)

# Added for running on Puhti
set(CMAKE_CUDA_ARCHITECTURES 70)
set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} --generate-code=arch=compute_70,code=sm_70")

# Find CUDA package
find_package(CUDAToolkit REQUIRED)
if (NOT CUDAToolkit_FOUND)
message(FATAL_ERROR "CUDA Toolkit not found. Please install it or set the CUDAToolkit_DIR variable.")
endif ()

# Add definitions
add_definitions(-DSOURCE_DIR="${CMAKE_SOURCE_DIR}")

# Add executable
add_executable(app main.cu)

Expand Down
26 changes: 26 additions & 0 deletions README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -205,3 +205,29 @@ git push origin student-name

Good luck, and enjoy accelerating matrix multiplication with CUDA!

### Results
In order to get more objective results, all of these are an average across 5 runs (naive CPU, blocked CPU, parallel CPU, naive CUDA, tiled CUDA, speedups).
Results for naive CPU, blocked CPU, and parallel CPU I copied from the previous assignment (meaning they were not re-run).
I tested with tile sizes: 16, 32, and 64. I found that there were no significant differences regarding speedup, possibly due to matrix dimensions being small. True speed up differences and memory management come to light only when matrices are big enough to account for the overhead.

I provided a shell script **run_experiments.sh** which will run all 9 cases 5 times and save that to **results.txt**. Shell script is in the root folder, and results file will be created in a new directory called *results*.
I provided **results-assignment2.csv** file with results from the previous assignment in csv format. Next to them we have **results-assignment3.csv** for the results of this assignment. Both results are provided in order to do a join and calculate Tiled CUDA Speedup vs. Parallel CPU. Both are inside of *results* directory.

Lastly, I provided a **results.ipynb** notebook to perform the join of two csv files, calculate Tiled CUDA Speedup vs. Parallel CPU, and finally calculate average metrics. This notebook is also located in the *results* directory.

Tiled CUDA being faster than Naive CUDA is not surprising, as it makes use of shared memory on GPUs, which is much faster than global memory and reduced the overhead of fetching data. Tiled CUDA vs Parallel CPU was a bit surprising, as Tiled CUDA performs worse. That just highlights the bottlenecks of data transfer between CPU and GPU. For cases 6, 7, and 9, which have the biggest matrices, Tiled CUDA still performs worse but only slightly, which again suggest that true speedup power can be seen only with large enough matrices.

I ran these tests on Puhti, and commands that I used are given in the file **puhti_commands**. CMake file had to be configured accordingly (it was giving some errors about platform mismatch).

| Test Case | Dimensions (m × n × p) | Naive CPU (s) | Blocked CPU (s) | Parallel CPU (s) | Naive CUDA (s) | Tiled CUDA (s) | Tiled CUDA Speedup (vs. Naive CUDA) (x) | Tiled CUDA Speedup (vs. Parallel CPU) (x) |
|-----------|------------------------|---------------|-----------------|------------------|----------------|----------------|-------------------------------------|---------------------------------------|
| 0 | 64 × 64 × 64 | 0.001820 | 0.003150 | 0.001971 | 0.147885 | 0.052384 | 2.819076 | 0.037568 |
| 1 | 128 × 64 × 128 | 0.007134 | 0.012213 | 0.006315 | 0.179840 | 0.061050 | 2.946002 | 0.103452 |
| 2 | 100 x 128 x 56 | 0.004972 | 0.007658 | 0.004033 | 0.155040 | 0.053658 | 2.890278 | 0.075254 |
| 3 | 128 x 64 x 128 | 0.007388 | 0.011238 | 0.006139 | 0.182496 | 0.061440 | 2.970274 | 0.099949 |
| 4 | 32 x 128 x 32 | 0.000872 | 0.001411 | 0.001214 | 0.144858 | 0.048442 | 2.992014 | 0.025087 |
| 5 | 200 x 100 x 256 | 0.035313 | 0.056534 | 0.028929 | 0.281901 | 0.102822 | 2.757510 | 0.283063 |
| 6 | 256 × 256 × 256 | 0.118653 | 0.178597 | 0.098928 | 0.311059 | 0.123488 | 2.522300 | 0.810078 |
| 7 | 256 × 300 × 256 | 0.138863 | 0.207555 | 0.118014 | 0.299462 | 0.132384 | 2.265972 | 0.893724 |
| 8 | 64 x 128 x 64 | 0.003496 | 0.005769 | 0.003319 | 0.138234 | 0.050682 | 2.729630 | 0.065424 |
| 9 | 256 x 256 x 257 | 0.120380 | 0.179924 | 0.107247 | 0.294950 | 0.114016 | 2.587976 | 0.941563 |
Loading