Adding Two Vectors
Task: Add two std::vector<float>s; since result of each iteration is independent of the previous (causal independence) - they can be computed in SIMD.
Results¶
My M1 Mac (with 8 Gigs of memory) started to tap out as I approached 2e8 elements. And the results were pretty inconsistent.
| Benchmark | Time (ns) | CPU (ns) | Iterations |
|---|---|---|---|
| BM_CPU/100000000 | 1.2876e+10 | 1.2803e+10 | 1 |
| BM_CPU/200000000 | 2.5761e+10 | 2.5589e+10 | 1 |
| BM_Metal/100000000 | 1.3022e+10 | 1.2469e+10 | 1 |
| BM_Metal/200000000 | 2.6355e+10 | 2.5305e+10 | 1 |
So I reran the program on a stronger machine - M3 Pro with 18Gig of memory:
| Benchmark | Time | CPU | Iterations |
|---|---|---|---|
| BM_CPU/100000000 | 9903958917 ns | 9868660000 ns | 1 |
| BM_CPU/200000000 | 1.9523e+10 ns | 1.9491e+10 ns | 1 |
| BM_Metal/100000000 | 9591659916 ns | 9466587000 ns | 1 |
| BM_Metal/200000000 | 2.0187e+10 ns | 1.9418e+10 ns | 1 |
The difference is still not clear. My guess is that since I'm using floats here, the program itself has become memory-bound. I can only confirm this by profiling the GPU (perhaps using Xcode Instruments) but I'm yet to learn how to do that.
What exactly is a data buffer?¶
We created a data buffer - this is to load the data into the GPU's memory.
Ig it's similar to prefetching - but I'll have to check on that.
Now when you use MTL::ResourceStorageModeShared, the buffer is shared between the CPU and the GPU. Therefore, there's no need to copy data between them. See other paradigms.