SoftGPU: The First gfx1201 Kernel

This is part 12 of SoftGPU. Decoding a SALU subset is not yet a kernel. Developers still need a real dispatch → machine code → memory result path without SoftGPU pretending to run arbitrary hipcc fat binaries.

What SoftGPU claims here

Claim	SoftGPU meaning
Kernel	`tiny_add`: `b[i] = a[i] + 1` for `i32` elements
Machine code	`llvm-mc -mcpu=gfx1201` bytes (`TINY_ADD_TEXT`)
Subset	`softgpu-gfx1201-e2e-tiny-v1` (SMEM/VOP2/GLOBAL plus prior SALU)
Calling convention	SoftGPU sets `s[4:5]=kernarg`, `v0=global_id_x`, EXEC
AQL	Registered `kernel_object` plus SoftGPU kernarg → `softgpu_kernel_success`
Fidelity	Architectural ISA for that subset only

Unregistered kernels still complete as diagnostic_complete_no_execution. That is the part-5 contract, still in force.

AQL KERNEL_DISPATCH
        ↓
Runtime: registered SoftGPU ISA image?
        ├── no  → diagnostic_complete_no_execution
        └── yes → run_code_1d (waves) on SoftGPU allocations
                        ↓
                 softgpu_kernel_success + completion signal 0

Instruction coverage (enumerated)

s_load_b64, s_waitcnt, s_endpgm
v_lshlrev_b32_e32, v_add_nc_u32_e32
global_load_b32, global_store_b32
the earlier SALU subset remains available

Everything else traps.

Differential

ISA tiny_add matches SoftGPU host reference (tiny_add_host_ref) and the Functional IR tiny_add semantics (b[i] = a[i] + 1). Hardware differential against an R9700 is deferred. As of the 16 September 2026 status snapshot that is still not claimed (Phase 12 / v0.9).

Encodings and .text were observed with llvm-mc (Homebrew LLVM 21.1.8), 16 September 2026, cross-checked against LLVM AMDGPUUsage. SoftGPU packaging is hand-maintained tables plus TINY_ADD_TEXT.

cargo test -p softgpu-amd-isa --test phase11_kernel --locked
cargo test -p softgpu-core --test phase11_aql_isa --locked
cargo run --locked -- run-kernel --builtin tiny_add --n 64

Honesty limits

This is not a claim of full gfx1201 or unrestricted HIP hipLaunchKernel success. The calling convention is SoftGPU-defined for this tiny kernel. s_waitcnt remains a SoftGPU no-op under the sequential interpreter.

The next hardware gate is R9700 conformance and profile hardening. Until that evidence exists, do not quote these results as agreement with a particular AMD GPU.

Next: why Rust for a software GPU, and what Rust cannot prove.

Canonical: thanos.github.io.