“i think of getting pounded when i see that [name]” – Satisfied Customer
IMPORTANT: IN ORDER TO SQUEEZE AS MUCH PERFORMANCE FOR SWITCH 1 AND 2 EMULATORS, DEVELOPMENT HAS FULLY SHIFTED TO CREATING A NEW ARM RECOMPILER FROM THE GROUND UP. IF YOU ARE A COMPILER DEVELOPER PLEASE GIVE US YOUR SUPPORT
- Translate SM86 to SPIR-V to Vulkan.
- Add
mimallocfor host allocator. - Create a custom pool / slab allocator for Horizon OS.
- Create a JIT code cache memory allocator for Ballistic.
- Create a JIT metadata manager.
- Integrate Ballistic into Pound.
- ARM64: share of the 3856 A64 encodings in Ballistic's decoder table that its x86 tier-1 compiler can run (54 run, 201 decoded only, 3601 not started). Ballistic commit
9db2adab.
GPU status details - updated 2026-10-13.
The decoded prototype covers 32 SM86 families: 31 in the straight-line SPIR-V path
and all 32 in the bounded branch/loop path. At the earlier integer
emitter checkpoint e6e19f3f (22-family scope),
301,056 complete straight-line lane states pass on RTX 3080; 112,640 bounded states pass
on each of RTX 3080 and software Vulkan. These synthetic checks are not an
overall GPU completion percentage or evidence that games render.
New FP32 comparison/min-max support: FSETP and FMNMX pass 4,718,592 native SM86 result comparisons and 62,464 full Vulkan lane states per backend.
FP32 addition/multiplication: FADD and FMUL pass 12,582,912 native result checks and 77,824 complete Vulkan states per backend, including all four rounding modes, FTZ and saturation.
Fused FP32 multiply-add: FFMA passes 29,360,128 native result checks and 45,056 complete Vulkan states per backend. The new checks also correct a rounding-dependent underflow boundary shared by FMUL and FFMA. FP32 / 32-bit integer conversions: F2I and I2F pass 10,289,152 native result checks and 45,056 complete Vulkan states per backend. Signed/unsigned results, four rounding modes and F2I FTZ/NaN handling are covered.
FP32 integral rounding: FRND passes 2,752,512 native results and 45,056 complete Vulkan states per backend, including nearest-even, floor, ceil, truncation, FTZ, abs/neg and signed-zero behavior.
I2FP conversion encoding: signed/unsigned 32-bit to FP32 with RN/RZ passes 1,474,560 native results and 45,056 complete Vulkan states per backend. Other I2FP sizes/modes reject explicitly.
Scalar FP32 to FP16 conversion: F2F.F16.F32 passes 2,752,512 native results and 45,056 complete Vulkan states per backend. Four rounding modes, FTZ, subnormals, overflow and NaNs are covered. Scalar FP16 to FP32 widening: F2F.F32.F16 passes 12,582,912 native comparisons covering every binary16 bit pattern and 45,056 complete Vulkan states per backend. Register sources, FTZ and abs/neg are verified. FP16 arithmetic, packed conversions and graphics rendering remain unimplemented.
The optional PoundGpuCompute tool now runs the supported subset through Vulkan, including resumable bounded state. Its separate runner checks pass 37,888 full states across 58 dispatches per backend. Reusable sessions and compiled kernels avoid rebuilding the device/pipeline for each dispatch; another 63,168 states pass per backend while alternating 3 kernels in one session. These remain compute-only development tools.
GPU development details cover the software DMA, tiled transfers, inline uploads, mappings, queues, cooperative scheduling and semaphore prototypes. Guest OS submission and production rendering remain unimplemented. CPU status above is a separate Ballistic snapshot and is not refreshed by this GPU update.
Regenerate GPU visuals with python tools/status/generate_gpu_status.py.
Regenerate the CPU snapshot with python tools/status/generate_status.py --ballistic <path-to-ballistic-checkout>.
Pound actively supports Linux and Windows with the Clang compiler only. I cannot run Pound on Windows, so the Windows builds will fall behind Linux builds. I would appreciate any Windows specific contributions greatly.