https://github.com/nielsouvrard/matrix-multiplication-m1-gpu-accelerated
Metal API to accelerate training of neural networks
https://github.com/nielsouvrard/matrix-multiplication-m1-gpu-accelerated
apple-silicon matrix-multiplication metal
Last synced: 3 months ago
JSON representation
Metal API to accelerate training of neural networks
- Host: GitHub
- URL: https://github.com/nielsouvrard/matrix-multiplication-m1-gpu-accelerated
- Owner: NielsOuvrard
- Created: 2024-10-31T17:57:48.000Z (7 months ago)
- Default Branch: main
- Last Pushed: 2024-11-04T04:00:34.000Z (7 months ago)
- Last Synced: 2025-01-08T12:14:21.509Z (4 months ago)
- Topics: apple-silicon, matrix-multiplication, metal
- Language: Objective-C
- Homepage:
- Size: 51.8 KB
- Stars: 1
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE/LICENSE.txt
Awesome Lists containing this project
README
# Matrix Multiplication Metal Apple Sicicone Accelerated
Use Metal to find GPUs and perform calculations on them.
My goal is to use the GPU to calculate matrices multiplications.
I used the sample code from the Apple Developer "Performing Calculations on a GPU" to understand how to use the Apple Metal framework to perform calculations on the GPU.
```c
int multiplied[SIZE_Y_A][SIZE_X_B];
for (int i = 0; i < SIZE_Y_A; i++) {
for (int j = 0; j < SIZE_X_B; j++) {
int tmp[SIZE_Y_B];
for (int k = 0; k < SIZE_Y_B; k++) {
tmp[k] = b[k][j];
}
// Create buffers to hold data
[adder prepareListInt:a[i] :tmp :SIZE_X_A];
// Send a command to the GPU to perform the calculation.
[adder sendComputeCommand:SIZE_X_A :&(multiplied[i][j])];
}
}
```
This is a time-consuming operation, and the GPU can perform it much faster than the CPU.
It need:
n \* n² = n³ multiplications
(n - 1) \* n = n² additions
Every value of the matrix `multiplied` is calculated by the GPU in parallel, using this code:
```c
kernel void multiply_ints_arrays(device const int* inA,
device const int* inB,
device int* result,
uint index [[thread_position_in_grid]])
{
result[index] = inA[index] * inB[index];
}
```
First version:
```shell
./MetalComputeBasic 0.15s user 5.77s system 84% cpu 7.011 total
```
[MTLDevice]: https://developer.apple.com/documentation/metal/mtldevice
[MTLCreateSystemDefaultDevice]: https://developer.apple.com/documentation/metal/1433401-mtlcreatesystemdefaultdevice
[MTLResource]: https://developer.apple.com/documentation/metal/mtlresource
[MTLBuffer]: https://developer.apple.com/documentation/metal/mtlbuffer
[MTLResourceStorageModeShared]: https://developer.apple.com/documentation/metal/mtlresourceoptions/mtlresourcestoragemodeshared
[MTLComputePipelineState]: https://developer.apple.com/documentation/metal/mtlcomputepipelinestate
[maxTotalThreadsPerThreadgroup]: https://developer.apple.com/documentation/metal/mtlcomputepipelinestate/1414927-maxtotalthreadsperthreadgroup
[status]: https://developer.apple.com/documentation/metal/mtlcommandbuffer/1443048-status
[addCompletedHandler]: https://developer.apple.com/documentation/metal/mtlcommandbuffer/1442997-addcompletedhandler
[MTLLibrary]: https://developer.apple.com/documentation/metal/mtllibrary
[MTLFunction]: https://developer.apple.com/documentation/metal/mtlfunction
[HelloTriangle]: https://developer.apple.com/documentation/metal