2.2 KiB
Native F2I immediate regression
This game-free fixture constructs its instructions from independently chosen registers, float values and bit patterns. It calls the production Ampere lowerer, Maxwell decoder and conversion implementations, then emits real IR and SPIR-V. No game, dump, cache, driver or GPU is used.
The immediate contains all 32 source bits. Bits 62 and 63 of the native
instruction belong to that literal; they are not the register source's abs/neg
modifiers. The test checks literal preservation, signedness, destination width,
four rounding modes, FTZ, ordinary/inverted predicates and @!PT skipping. It
also checks existing register/cbuf/uniform source forms and explicitly rejects
immediate F16, F64, 64-bit output, non-NTZ and unknown modes. The accepted U16
form remains FTZ/nearest only.
CC=clang CXX=clang++ cmake -S tests/f2i_immediate -B build-f2i \
-DREPO_ROOT="$PWD" -DSPIRV_HEADERS_ROOT=/path/to/SPIRV-Headers
cmake --build build-f2i --parallel 2
build-f2i/f2i_immediate build-f2i/modules
for module in build-f2i/modules/*.spv; do
spirv-val --target-env vulkan1.3 --uniform-buffer-standard-layout "$module"
done
The unmodified 232eecb lowerer fails the first authored immediate case with
unsupported native opcode. The fixture tests the conversion path and its IR
controls; SPIR-V validation does not establish numerical results on a GPU or a
fix for any particular visible game symptom. In particular, it does not change
or independently certify existing register-path saturation at numeric limits.
Encoding reference: Mesa NAK, pinned sm70_encode.rs
(OpF2I, encode_alu_base and encode_alu_imm). The 0x105 base plus immediate
form 4 in bits 9–11 gives 0x905. Inputs or outputs wider than 32 bits use a
different base, 0x111. NVIDIA PTX conversion documentation
describes rounding, FTZ, saturation and NaN results; it is not a SASS encoding
specification.