[Know-It-All-Brother]

Part of the Ampere accuracy series (#2040). It depends on #2041 for the test: the commits before this PR's own are #2041's and drop out once it lands, so this PR's own change is the last commit.

What and why

I2F from a 64-bit integer (cvt.rn.f32.s64, opcode 0x312) computed float(hi) * 2^32 + float(lo), which rounds twice. For a negative value whose high word is -1 or -2, the two big terms nearly cancel and the low word's rounding error is left behind, so results were off by 1 ULP and up to 2^30 ULP. On a Jetson Orin, the 64-bit value with words 0xfffffffe:0xaaaaaaaa converts to 0xcfaaaaab; we gave 0xcfaaaaaa. The instruction's rounding field was ignored too.

AI2f64 now rounds once, with integer operations on the two 32-bit halves, since the IR has no 64-bit integers here. It takes the magnitude (carrying into the high word), finds the leading one, shifts it to the top keeping the dropped bits, and rounds to nearest even, or by the instruction's .RM / .RP / .RZ. Then it builds the float's bits and puts the sign back. Integer operations are exact under Metal's fast math. Destination F32 is still the only one supported, as before. It costs about 33 integer operations per 64-bit I2F, a rare instruction (the compiler's 32-bit division sequence is one place it comes up).

Reference entry: "I2F from a 64-bit integer rounds once" in docs/reference/shaders.md.

How to verify

DMSH_TEST_XLATE=dm/gfx/tools/dmsh_xlate_bin/dmsh_xlate AMPERE_ORIN_ONLY=i2f_s64 python3 dm/gfx/tests/test_dmsh_ampere_orin.py

i2f_s64: 16 of 1,024 lanes differed on main (up to 1,073,741,824 ULP), and 0 differ here. The test takes it out of KNOWN.

The fixture only has the signed, round-to-nearest form, so the unsigned source and the directed roundings haven't run against the Orin. I checked a model of exactly these steps against an exact reference on 200,000 random values (signed and unsigned, all four roundings, including high words 0, -1, -2 and 0x80000000), and they all matched.

This overlaps #1800 only by position: #1800 changes the 32-bit I2F roundings in AI2f, and this replaces AI2f64 a few dozen lines further down.

Checked

• Tests: test_dmsh_ampere_orin.py; every other test_dmsh_* gives the same result as on main (test_dmsh_ampere_mem.py, test_dmsh_meshstage.py and test_dmsh_pairs.py fail on main the same way); test_nsp_chain.py, test_xlate_anycpu.py, test_docs.py.
• Translator: the dlls are rebuilt from this tree with .NET 10; tool_lock.py --update-lock and dmsh_chain.py --header were run (chain 6e3c16385059).
• Devices and OS: MacBook Pro M5 Max, macOS 27.
• regress.py: not run. My only local Switch 2 build's shader package comes from an older shader split than main's map, so main's chain can't build it. I'll add a report once I have a fresh build.
• Games run: none.

Docs

• docs/reference/shaders.md has the entry, and python3 dm/tools/docs_symptoms.py was run
• A chain tool changed: dmsh_chain.py --header was run, and dmsh_chain.h, the translator dlls and the lock are part of the change
• No game files, keys or data extracted from a game