From 6be3e6185de95143de6640d150c803cf59ced18a Mon Sep 17 00:00:00 2001
From: Luke Benstead <kazade@gmail.com>
Date: Wed, 14 Sep 2022 19:20:24 +0100
Subject: [PATCH] Faster inversion

---
 GL/platforms/sh4.c      |  9 +++++++--
 GL/texture.c            | 10 ++++++----
 samples/quadmark/main.c |  5 +++++
 3 files changed, 18 insertions(+), 6 deletions(-)

diff --git a/GL/platforms/sh4.c b/GL/platforms/sh4.c
index 7974f61..6fb3093 100644
--- a/GL/platforms/sh4.c
+++ b/GL/platforms/sh4.c
@@ -47,8 +47,13 @@ void SceneListBegin(GPUList list) {
     pvr_list_begin(list);
 }
 
+GL_FORCE_INLINE float _glFastInvert(float x) {
+    const float sgn = (x > 0) - (x < 0);
+    return sgn * MATH_fsrra(x * x);
+}
+
 GL_FORCE_INLINE void _glPerspectiveDivideVertex(Vertex* vertex, const float h) {
-    const float f = MATH_Fast_Invert(vertex->w);
+    const float f = _glFastInvert(vertex->w);
 
     /* Convert to NDC and apply viewport */
     vertex->xyz[0] = __builtin_fmaf(
@@ -65,7 +70,7 @@ GL_FORCE_INLINE void _glPerspectiveDivideVertex(Vertex* vertex, const float h) {
     avoid a divide by zero.
     */
     if(unlikely(vertex->w == 1.0f)) {
-        vertex->xyz[2] = MATH_Fast_Invert(1.0001f + vertex->xyz[2]);
+        vertex->xyz[2] = _glFastInvert(1.0001f + vertex->xyz[2]);
     } else {
         vertex->xyz[2] = f;
     }
diff --git a/GL/texture.c b/GL/texture.c
index 2c93377..2adaf94 100644
--- a/GL/texture.c
+++ b/GL/texture.c
@@ -161,9 +161,11 @@ static void GPUTextureTwiddle8PPP(void* src, void* dst, uint32_t w, uint32_t h)
     for(y = 0; y < h; y += 2) {
         yout = y;
         for(x = 0; x < w; x++) {
-            vtex[TWIDOUT((yout & mask) / 2, x & mask) +
-                 (x / min + yout / min)*min * min / 2] =
-                     pixels[y * w + x] | (pixels[(y + 1) * w + x] << 8);
+            int32_t idx = TWIDOUT((yout & mask) / 2, x & mask) +
+                 (x / min + yout / min)*min * min / 2;
+
+            gl_assert(idx < (w * h));
+            vtex[idx] = pixels[y * w + x] | (pixels[(y + 1) * w + x] << 8);
         }
     }
 }
@@ -1719,7 +1721,7 @@ GLAPI void APIENTRY glColorTableEXT(GLenum target, GLenum internalFormat, GLsize
         sharedPaletteUsed = GL_TRUE;
     }
 
-    for (GLbyte i = 1; i < MAX_GLDC_SHARED_PALETTES; ++i) {
+    for (GLubyte i = 1; i < MAX_GLDC_SHARED_PALETTES; ++i) {
         if (target == GL_SHARED_TEXTURE_PALETTE_0_KOS + i) {
             palette = SHARED_PALETTES[i];
             sharedPaletteUsed = GL_TRUE;
diff --git a/samples/quadmark/main.c b/samples/quadmark/main.c
index 26ea433..4f83bde 100644
--- a/samples/quadmark/main.c
+++ b/samples/quadmark/main.c
@@ -68,6 +68,7 @@ int check_start() {
 }
 
 void setup() {
+    //PVR needs to warm up for a frame, or results will be low
     glKosInit();
     glMatrixMode(GL_MODELVIEW);
     glLoadIdentity();
@@ -76,6 +77,10 @@ void setup() {
     glLoadIdentity();
 
     glDisable(GL_NEARZ_CLIPPING_KOS);
+
+    pvr_wait_ready();
+    pvr_scene_begin();
+    pvr_scene_finish();
 }
 
 void do_frame() {