Use Kaze's folded polynomials

include/config/config_game.h

@@ -9,6 +9,9 @@
 // Disables AA (Improves console performance but causes visible seams between unconnected geometry).
 //#define DISABLE_AA
 
+// Use a folded polynomial instead of a lookup table. This has a speed boots and makes the calculation more precise.
+#define DISABLE_SIN_COS_LOOKUP_TABLE
+
 // Fix annoying glitches (crashes and softlocks)
 #define FIX_ANNOYING_GLITCH
 

include/functions.h

@@ -1978,6 +1978,9 @@ s32 osPfsDeleteFile(OSPfs* pfs, u16 companyCode, u32 gameCode, u8* gameName, u8*
 s32 __osPfsReleasePages(OSPfs* pfs, __OSInode* inode, u8 initialPage, u8 bank, __OSInodeUnit* finalPage);
 void guOrthoF(f32[4][4], f32, f32, f32, f32, f32, f32, f32);
 void guOrtho(Mtx*, f32, f32, f32, f32, f32, f32, f32);
+#ifdef DISABLE_SIN_COS_LOOKUP_TABLE
+f32x2 sincos(s16 angle);
+#endif
 f32 cosf(f32 angle);
 s16 coss(u16 angle);
 void osViSetEvent(OSMesgQueue* mq, OSMesg msg, u32 retraceCount);

include/macros.h

@@ -218,3 +218,11 @@ extern struct GraphicsContext* __gfxCtx;
     } while (0)
 
 #endif
+
+#ifdef DISABLE_SIN_COS_LOOKUP_TABLE
+
+#define ONE 1.0f
+#define SECOND_ORDER_COEFFICIENT 0.0000000010911122665310369f
+#define quasi_cos_2(x) (ONE - SECOND_ORDER_COEFFICIENT * x * x)
+
+#endif

include/ultra64/ultratypes.h

@@ -1,8 +1,16 @@
+#include "config.h"
+
 #ifndef ULTRA64_ULTRATYPES_H
 #define ULTRA64_ULTRATYPES_H
 
 #ifdef _LANGUAGE_C
 
+#ifdef DISABLE_SIN_COS_LOOKUP_TABLE
+typedef _Complex float f32x2;
+#define F32X2_NEW(x, y) __builtin_complex((float) (x), (float) (y))
+#define F32X2_AT(pair, idx)                                                                            \
+    __builtin_choose_expr(idx / (_idx == 0 || idx == 1), __real__(pair), __imag__(pair))
+#endif
 typedef signed char            s8;
 typedef unsigned char          u8;
 typedef signed short int       s16;

src/code/z_lib.c

@@ -1,5 +1,7 @@
 #include "global.h"
 
+#include "config.h"
+
 /**
  * memset: sets `len` bytes to `val` starting at address `dest`.
  *
@@ -23,20 +25,54 @@ void Lib_MemSet(u8* dest, size_t len, u8 val) {
     // clang-format on
 }
 
+#ifdef DISABLE_SIN_COS_LOOKUP_TABLE
+/**
+ * @param angle binang
+ * @return sinecos(sine, cosine)
+ */
+f32x2 sincos(s16 angle) {
+    s32 shifter = (angle ^ (angle << 1)) & 0xC000;
+    f32 x = (f32) (((angle + shifter) << 17) >> 16);
+    float cosx = quasi_cos_2(x);
+    float sinx = sqrtf(1.0f - cosx * cosx);
+
+    if (shifter & 0x4000) {
+        float temp = cosx;
+        cosx = sinx;
+        sinx = temp;
+    }
+    if (angle < 0) {
+        sinx = -sinx;
+    }
+    if (shifter & 0x8000) {
+        cosx = -cosx;
+    }
+        return F32X2_NEW(sinx, cosx);
+}
+#endif
+
 /**
  * @param angle binang
  * @return cos(angle)
  */
 f32 Math_CosS(s16 angle) {
+#ifdef DISABLE_SIN_COS_LOOKUP_TABLE
+    return  __imag__(sincos(angle));
+#else
     return coss(angle) * SHT_MINV;
+#endif
 }
 
 /**
  * @param angle binang
  * @return sin(angle)
  */
 f32 Math_SinS(s16 angle) {
+#ifdef DISABLE_SIN_COS_LOOKUP_TABLE
+    return  __real__(sincos(angle));
+#else
     return sins(angle) * SHT_MINV;
+#endif
 }
 
 /**

src/libultra/gu/coss.c

@@ -1,9 +1,14 @@
 #include "global.h"
 
+#include "config.h"
 /**
  * @param angle binang
  * @return cos(angle)*0x7FFF
  */
 s16 coss(u16 angle) {
+#ifdef DISABLE_SIN_COS_LOOKUP_TABLE
+    return Math_CosS(angle) * 0x7FFF;
+#else
     return sins(angle + 0x4000);
+#endif
 }

src/libultra/gu/sins.c

@@ -1,12 +1,17 @@
 #include "ultra64.h"
 
+#include "config.h"
+#ifndef DISABLE_SIN_COS_LOOKUP_TABLE
 #include "sintable.inc.c"
-
+#endif
 /**
  * @param angle binang
  * @return sin(angle)*0x7FFF
  */
 s16 sins(u16 angle) {
+#ifdef DISABLE_SIN_COS_LOOKUP_TABLE
+return Math_SinS(angle) * 0x7FFF;
+#else
     s16 value;
 
     angle >>= 4;
@@ -22,4 +27,5 @@ s16 sins(u16 angle) {
     } else {
         return value;
     }
+#endif
 }