For Intel Macs, it runs on all devices supporting macOS Ventura. This includes the A13 Bionic and all other Apple GPUs that support the Metal 3 feature set. Since MetalFloat64 requires function pointers and Metal dynamic libraries, it only runs on Apple6 family or newer GPUs. The vectorized function variants minimize binary size and amortize the overhead of function calls. If possible, clients should operate on multiple floats at a time (e.g.

To address this problem, larger functions invoke function calls into the Metal dynamic library. Some FP64 operations can be fully inlined for maximum ALU performance, but others are so large they would bloat the client executable. The source code compiles into a Metal dynamic library and a header, which other applications can utilize for double-precision arithmetic. For Intel GPUs, FP64 emulation will often make GPU slower than multicore CPU, even when 90% of the compute workload is FP32. For AMD GPUs with hardware double precision, OpenCL provides better performance for FP64-heavy workloads. MetalFloat64 runs fastest on Apple silicon, although it also runs on Intel Macs. This is the same throughput ratio as native FP64 on recent NVIDIA GPUs. Based on theoretical estimates, additions and multiplications will have 1/32-1/64 the throughput of their 32-bit counterparts. Still in the planning stage - this is not a finished library!Įmulating double-precision arithmetic on Apple GPUs, with full IEEE compliance.

Metal gpus