[iQue] Match libultra/gu (#2402)

* [iQue] Match libultra/gu

* Match sqrtf

4 files changed, 29 insertions, 28 deletions

diff --git a/src/libultra/gu/rotate.c b/src/libultra/gu/rotate.c
index 7d495e98e..dc50aba3c 100644
--- a/src/libultra/gu/rotate.c
+++ b/src/libultra/gu/rotate.c
@@ -1,7 +1,7 @@
 #include "global.h"
 
 void guRotateF(f32 m[4][4], f32 a, f32 x, f32 y, f32 z) {
-    static f32 D_80134D10 = M_PI / 180.0f;
+    static f32 dtor = M_PI / 180.0f;
     f32 sine;
     f32 cosine;
     f32 ab;
@@ -16,14 +16,15 @@ void guRotateF(f32 m[4][4], f32 a, f32 x, f32 y, f32 z) {
 
     guNormalize(&x, &y, &z);
 
-    a *= D_80134D10;
+    a *= dtor;
 
     sine = sinf(a);
     cosine = cosf(a);
 
-    ab = x * y * (1 - cosine);
-    bc = y * z * (1 - cosine);
-    ca = z * x * (1 - cosine);
+    t = 1.0f - cosine;
+    ab = x * y * t;
+    bc = y * z * t;
+    ca = z * x * t;
 
     guMtxIdentF(m);
 
@@ -38,15 +39,15 @@ void guRotateF(f32 m[4][4], f32 a, f32 x, f32 y, f32 z) {
 #endif
 
     t = x * x;
-    m[0][0] = (1 - t) * cosine + t;
+    m[0][0] = t + cosine * (1.0f - t);
     m[2][1] = bc - xs;
     m[1][2] = bc + xs;
     t = y * y;
-    m[1][1] = (1 - t) * cosine + t;
+    m[1][1] = t + cosine * (1.0f - t);
     m[2][0] = ca + ys;
     m[0][2] = ca - ys;
     t = z * z;
-    m[2][2] = (1 - t) * cosine + t;
+    m[2][2] = t + cosine * (1.0f - t);
     m[1][0] = ab - zs;
     m[0][1] = ab + zs;
 }
diff --git a/src/libultra/gu/sqrtf.c b/src/libultra/gu/sqrtf.c
deleted file mode 100644
index 5e5fa233c..000000000
--- a/src/libultra/gu/sqrtf.c
+++ /dev/null
@@ -1,9 +0,0 @@
-#include "global.h"
-
-#ifndef __GNUC__
-#define __builtin_sqrtf sqrtf
-#endif
-
-f32 sqrtf(f32 f) {
-    return __builtin_sqrtf(f);
-}
diff --git a/src/libultra/gu/sqrtf.s b/src/libultra/gu/sqrtf.s
new file mode 100644
index 000000000..aaf502359
--- /dev/null
+++ b/src/libultra/gu/sqrtf.s
@@ -0,0 +1,9 @@
+#include "ultra64/asm.h"
+#include "ultra64/regdef.h"
+
+.text
+
+LEAF(sqrtf)
+    sqrt.s  fv0, fa0
+    j       ra
+END(sqrtf)
diff --git a/src/libultra/gu/us2dex.c b/src/libultra/gu/us2dex.c
index d2e5a2b04..b8708c086 100644
--- a/src/libultra/gu/us2dex.c
+++ b/src/libultra/gu/us2dex.c
@@ -1,25 +1,25 @@
 #include "global.h"
 
 void guS2DInitBg(uObjBg* bg) {
-    u32 size;
-    s32 tmem = (bg->b.imageFmt == G_IM_FMT_CI) ? 0x100 : 0x200;
-    u16 shift = (6 - bg->b.imageSiz);
+    u16 tmem = (bg->b.imageFmt == G_IM_FMT_CI) ? 0x100 : 0x200;
+    u16 shift = 6 - bg->b.imageSiz;
+    u32 tsize;
 
     if (bg->b.imageLoad == G_BGLT_LOADBLOCK) {
         bg->b.tmemW = bg->b.imageW >> shift;
-        bg->b.tmemH = (tmem / bg->b.tmemW) * 4;
-        bg->b.tmemSizeW = bg->b.tmemW * 2;
+        bg->b.tmemH = (tmem / bg->b.tmemW) << 2;
+        bg->b.tmemSizeW = bg->b.tmemW << 1;
         bg->b.tmemSize = bg->b.tmemH * bg->b.tmemSizeW;
         bg->b.tmemLoadSH = (bg->b.tmemSize >> 1) - 1;
-        bg->b.tmemLoadTH = (0x7FF / bg->b.tmemW) + 1;
+        bg->b.tmemLoadTH = GS_CALC_DXT(bg->b.tmemW);
     } else { // G_BGLT_LOADTILE
         bg->b.tmemW = (bg->b.frameW >> shift) + 3;
-        bg->b.tmemH = (tmem / bg->b.tmemW) * 4;
-        bg->b.tmemSizeW = (bg->b.imageW >> shift) * 2;
+        bg->b.tmemH = (tmem / bg->b.tmemW) << 2;
+        bg->b.tmemSizeW = (bg->b.imageW >> shift) << 1;
 
-        size = bg->b.tmemH * bg->b.tmemSizeW;
-        bg->b.tmemSize = (size >> 16);
-        bg->b.tmemLoadSH = (size >> 0) & 0xFFFF;
+        tsize = bg->b.tmemH * bg->b.tmemSizeW;
+        bg->b.tmemSize = tsize >> 16;
+        bg->b.tmemLoadSH = tsize & 0xFFFF;
         bg->b.tmemLoadTH = bg->b.tmemH - 1;
     }
 }