From b28c2f924771b8c8edb0d7ccec73913958e56a59 Mon Sep 17 00:00:00 2001
From: Giulio Eulisse <10544+ktf@users.noreply.github.com>
Date: Sun, 11 Oct 2026 18:34:45 +0200
Subject: [PATCH] GPU: give Metal its own spellings for two math helpers

GPUCA_CHOICE routes Metal down the OpenCL arm, and three of those spellings do
not exist in MSL. One of the three needs no Metal case: MSL has no nan() in any
form, but it does define NAN, and so does OpenCL -- a constant expression of
type float representing a quiet NaN -- so the shared arm uses that instead of
nan(0u).

remainder() does not exist either, and __builtin_remainderf is not a way around
it: that compiles, then fails to link against an undefined remainderf. MSL does
have fmod, which is exact, so Remainderf reduces with that and nudges the result
into [-|y|/2, |y|/2], rounding the quotient of a tie to even. It agrees with
remainderf bit for bit over six million samples, from the TwoPI wrapping its
only caller does up to |x| of 1e30, ties swept exhaustively for a power-of-two
divisor. Reducing as x - y * rint(x / y) would have been shorter, but x / y
rounds in float, so rint picks the wrong multiple outside a narrow range around
the divisor and the result is wrong almost everywhere else.

MSL's sincos returns the sine and takes the cosine by thread reference rather
than by pointer, and cannot write through the generic reference SinCos is given,
so the result goes via a local.

Host, CUDA, HIP and cling keep the GPUCA_CHOICE arms they had; OpenCL swaps
nan(0u) for NAN.
---
 GPU/Common/GPUCommonMath.h | 23 +++++++++++++++++++++--
 1 file changed, 21 insertions(+), 2 deletions(-)

diff --git a/GPU/Common/GPUCommonMath.h b/GPU/Common/GPUCommonMath.h
index 057d3afca0945..65ff3092bc313 100644
--- a/GPU/Common/GPUCommonMath.h
+++ b/GPU/Common/GPUCommonMath.h
@@ -105,7 +105,7 @@ class GPUCommonMath
   GPUd() constexpr static bool Finite(float x);
   GPUd() constexpr static bool IsNaN(float x);
 #ifndef __FAST_MATH__
-  GPUd() constexpr static float QuietNaN() { return GPUCA_CHOICE(std::numeric_limits<float>::quiet_NaN(), __builtin_nanf(""), nan(0u)); }
+  GPUd() constexpr static float QuietNaN() { return GPUCA_CHOICE(std::numeric_limits<float>::quiet_NaN(), __builtin_nanf(""), NAN); }
 #endif
   GPUd() constexpr static uint32_t Clz(uint32_t val);
   GPUd() constexpr static uint32_t Ctz(uint32_t val);
@@ -245,7 +245,21 @@ GPUdi() float2 GPUCommonMath::MakeFloat2(float x, float y)
 }
 
 GPUdi() constexpr float GPUCommonMath::Modf(float x, float y) { return GPUCA_CHOICE(fmodf(x, y), fmodf(x, y), fmod(x, y)); }
-GPUhdi() float GPUCommonMath::Remainderf(float x, float y) { return GPUCA_CHOICE(std::remainderf(x, y), remainderf(x, y), remainder(x, y)); }
+GPUhdi() float GPUCommonMath::Remainderf(float x, float y)
+{
+#ifdef __METAL__ // MSL has no remainder(), so reduce with fmod, which is exact
+  const float r = fmod(x, y), h = 0.5f * fabs(y), a = fabs(r);
+  if (a > h) {
+    return r - copysign(y, r);
+  }
+  if (a == h && fmod((x - r) / y, 2.f) != 0.f) { // the quotient of a tie is rounded to even
+    return r - copysign(y, r);
+  }
+  return r;
+#else
+  return GPUCA_CHOICE(std::remainderf(x, y), remainderf(x, y), remainder(x, y));
+#endif
+}
 
 GPUdi() uint32_t GPUCommonMath::Float2UIntReint(const float& x)
 {
@@ -302,6 +316,11 @@ GPUhdi() void GPUCommonMath::SinCos(float x, float& s, float& c)
     __sincosf(x, &s, &c);
 #elif !defined(GPUCA_GPUCODE_DEVICE) && (defined(__GNU_SOURCE__) || defined(_GNU_SOURCE) || defined(GPUCA_GPUCODE))
     sincosf(x, &s, &c);
+#elif defined(__METAL__) // MSL's sincos returns sin and takes cos by thread
+                         // reference, so it cannot write through a generic one
+    float metalCos;
+    s = sincos(x, metalCos);
+    c = metalCos;
 #else
     GPUCA_CHOICE((void)((s = sinf(x)) + (c = cosf(x))), sincosf(x, &s, &c), s = sincos(x, &c));
 #endif