Philip Turner
Philip Turner
Some comments mentioned that MPSGraph convolutions used Winograd. Could there be documentation of exactly how much FLOPS? Anything over 8.49 TFLOPS is not physically possible. https://github.com/geohot/tinygrad/blob/master/extra/gemm/metal_conv.py#L11-L14
We could use Accelerate to scale the vector here, similarly to how `add` and `exp` use Accelerate. https://github.com/ggerganov/ggml/blob/2992df03010bb6afe399f13378f20ed45b0758c8/src/ggml.c#L3250-L3277 https://developer.apple.com/documentation/accelerate/1450020-vdsp_vsmul
In collaboration with @liuliu, I have been developing a new kernel library for GEMM and attention operations. Stable Diffusion/NNC is the primary use case, but I hope to integrate into...
I am working on an API to export the raw data from ARHeadsetKit. It takes all geometry and color data present at one point in time, then reorders it into...
Add an experimental API for serializing and exporting ARHeadsetKit's internal data. It only supports a geometry + color format now, but is extendable to the [lidar-scanning-app](https://github.com/philipturner/lidar-scanning-app) format (geometry only). Example...
NDArrayMatrixMultiplyA16 does not contain simd async copy instructions, although the kernel for A14 does. Starting with AGX3 (A15), there are some new instructions used for GEMM and Conv. I haven't...
The website seems out of date regarding recent commits. Could it be brought up to date?
I've been working on making GROMACS faster on the Apple GPU, and I recently [came across a need to work with VkFFT](https://gitlab.com/gromacs/gromacs/-/merge_requests/3162). The software package was using clFFT, which couldn't...
macOS Ventura deprecated `sprintf` in favor of `snprintf`, a more secure alternative. This is creating vast amounts of errors in GROMACS, so I'm making a PR to suppress the warning....
Congratulations on the name change! That reminded me, perhaps we should support OpenCL 2.0 on Android. I should also definitely make Metal runnable on iOS, perhaps we introduce both at...