mirror of
https://git.triforceproject.com/triforce/source
synced 2026-11-22 17:19:58 -04:00
The 0.6 crashes after warps and area changes fixed (no write-back into memory the capture leaves out), invisible objects fixed (occlusion gate), DLSS frame generation flicker fixed, achievements, the renderer built with a play profile, faster translated game code, helper threads judged on settled scenes, memory, shader preparation, audio effects, launcher, controls and Linux fixes.
63 lines
3.6 KiB
Python
63 lines
3.6 KiB
Python
"""(Triforce, 11-22, the performance audit's LIFT-05, after Source A's) The single-precision FRECPS and FRSQRTS steps as
|
|
one fused instruction where the build has FMA3.
|
|
|
|
The kit's fast path for the Newton-Raphson steps, recomp_fpx_recps32 (2 - a*b) and recomp_fpx_rsqrts32 ((3 - a*b) / 2),
|
|
has two forms: on an Arm host one fused multiply-add, the step rounded once as the guest rounds it; on x86-64 the step in
|
|
binary64 (two widenings, a multiply, a subtraction, a narrowing) and a test for a binary32 midpoint, which sends the call
|
|
to the exact body, because x86-64's baseline has no FMA. The game is built for x86-64-v3 where the CPU has it, and there
|
|
one FMA instruction is that same correctly rounded step. The kit already took this step for the fused multiply-add itself
|
|
(a64_rtgen.fma32_native, recomp_fpx_fma32); this gives the two steps the same guard. The tests that decide whether the
|
|
native result is kept (a normal result; under FZ no subnormal operand) are unchanged, and halving a normal result is
|
|
exact.
|
|
|
|
The kit's files stay as they are (third_party/ is not edited): install(), called from nx2/lift.py, wraps the kit's
|
|
a64_rtgen.runtime_h. Each function's head must be in its text exactly once, so a refreshed kit that changed one stops the
|
|
lift with a message, not the game.
|
|
|
|
Switches: A64_NO_FMA_STEPS=1 at lift time (in the lift key, an A64_ name) leaves the kit's text for the two steps. The
|
|
build define -DA64_FPX_NO_FMA32, the kit's own for the fused multiply-add, now returns all three to the binary64 form.
|
|
Nothing changes in a build without FMA3 (x86-64-v2). tests/test_fma_steps.py runs the guest instructions both ways."""
|
|
import os
|
|
|
|
import a64_rtgen
|
|
|
|
GUARD_OLD = "#if RECOMP_FPX_HOST == 1\n"
|
|
GUARD_NEW = "#if RECOMP_FPX_HOST == 1 || (RECOMP_FPX_HOST == 2 && defined(__FMA__) && !defined(A64_FPX_NO_FMA32))\n"
|
|
# (each function's head line through its #if line, and the line after it, so the form that follows is the one known here)
|
|
RECPS32_OLD = ("static RECOMP_INLINE int recomp_fpx_recps32(uint64_t a, uint64_t b, uint32_t fz, uint64_t* v) {\n" + GUARD_OLD +
|
|
" const uint64_t t = recomp_fpx_fb(__builtin_fmaf(-recomp_fpx_f(a), recomp_fpx_f(b), 2.0f));\n")
|
|
RSQRTS32_OLD = ("static RECOMP_INLINE int recomp_fpx_rsqrts32(uint64_t a, uint64_t b, uint32_t fz, uint64_t* v) {\n" + GUARD_OLD +
|
|
" /* Halving a normal result is exact, and a normal result is all that is kept. */\n"
|
|
" const uint64_t t = recomp_fpx_fb(__builtin_fmaf(-recomp_fpx_f(a), recomp_fpx_f(b), 3.0f) * 0.5f);\n")
|
|
RECPS32_NEW = RECPS32_OLD.replace(GUARD_OLD, GUARD_NEW)
|
|
RSQRTS32_NEW = RSQRTS32_OLD.replace(GUARD_OLD, GUARD_NEW)
|
|
|
|
|
|
def patch(h):
|
|
"""The kit's runtime header with the two steps' fused form on x86-64 with FMA3; the kit's text with A64_NO_FMA_STEPS=1"""
|
|
if os.environ.get("A64_NO_FMA_STEPS"):
|
|
return h
|
|
for name, old, new in (("recomp_fpx_recps32", RECPS32_OLD, RECPS32_NEW), ("recomp_fpx_rsqrts32", RSQRTS32_OLD, RSQRTS32_NEW)):
|
|
if h.count(old) != 1: # (the kit changed: the lift must not go on)
|
|
raise SystemExit("fma_steps: the kit's %s changed (its head is %d times in the runtime header): update "
|
|
"nx2/fma_steps.py" % (name, h.count(old)))
|
|
h = h.replace(old, new)
|
|
return h
|
|
|
|
|
|
_installed = False
|
|
|
|
|
|
def install():
|
|
"""patch() around the kit's a64_rtgen.runtime_h, which its write() calls by name. Once"""
|
|
global _installed
|
|
if _installed:
|
|
return
|
|
_installed = True
|
|
base = a64_rtgen.runtime_h
|
|
|
|
def runtime_h():
|
|
return patch(base())
|
|
|
|
a64_rtgen.runtime_h = runtime_h
|