diff --git a/GL/draw.c b/GL/draw.c
index 3c1db65..60013ab 100644
--- a/GL/draw.c
+++ b/GL/draw.c
@@ -442,7 +442,7 @@ GL_FORCE_INLINE void transformNormalToEyeSpace(GLfloat* normal) {
     mat_trans_normal3(normal[0], normal[1], normal[2]);
 }
 
-PolyHeader *_glSubmissionTargetHeader(SubmissionTarget* target) {
+GL_FORCE_INLINE PolyHeader *_glSubmissionTargetHeader(SubmissionTarget* target) {
     assert(target->header_offset < target->output->vector.size);
     return aligned_vector_at(&target->output->vector, target->header_offset);
 }
@@ -456,7 +456,7 @@ Vertex* _glSubmissionTargetEnd(SubmissionTarget* target) {
     return _glSubmissionTargetStart(target) + target->count;
 }
 
-static inline void genTriangles(Vertex* output, GLuint count) {
+GL_FORCE_INLINE void genTriangles(Vertex* output, GLuint count) {
     Vertex* it = output + 2;
 
     GLuint i;
@@ -466,17 +466,22 @@ static inline void genTriangles(Vertex* output, GLuint count) {
     }
 }
 
-static inline void genQuads(Vertex* output, GLuint count) {
+GL_FORCE_INLINE void genQuads(Vertex* output, GLuint count) {
+    Vertex* pen = output + 2;
     Vertex* final = output + 3;
-    GLuint i;
-    for(i = 0; i < count; i += 4) {
-        swapVertex((final - 1), final);
+    GLuint i = count >> 2;
+    while(i--) {
+        __asm__("pref @%0" : : "r"(pen + 4));
+
+        swapVertex(pen, final);
         final->flags = GPU_CMD_VERTEX_EOL;
+
+        pen += 4;
         final += 4;
     }
 }
 
-static void genTriangleStrip(Vertex* output, GLuint count) {
+GL_FORCE_INLINE void genTriangleStrip(Vertex* output, GLuint count) {
     output[count - 1].flags = GPU_CMD_VERTEX_EOL;
 }
 
@@ -634,6 +639,8 @@ static void _readPositionData(ReadDiffuseFunc func, const GLuint first, const GL
     uint32_t* flags;
 
     ITERATE(count) {
+        __asm__("pref @%0" : : "r"(vptr + vstride));
+
         func(vptr, out);
         vptr += vstride;
 
@@ -653,6 +660,8 @@ static void _readUVData(ReadUVFunc func, const GLuint first, const GLuint count,
     GLubyte* out = (GLubyte*) output[0].uv;
 
     ITERATE(count) {
+        __asm__("pref @%0" : : "r"(uvptr + uvstride));
+
         func(uvptr, out);
         uvptr += uvstride;
         out += sizeof(Vertex);
@@ -666,6 +675,8 @@ static void _readSTData(ReadUVFunc func, const GLuint first, const GLuint count,
     GLubyte* out = (GLubyte*) extra[0].st;
 
     ITERATE(count) {
+        __asm__("pref @%0" : : "r"(stptr + ststride));
+
         func(stptr, out);
         stptr += ststride;
         out += sizeof(VertexExtra);
@@ -714,6 +725,8 @@ static void _readDiffuseData(ReadDiffuseFunc func, const GLuint first, const GLu
     GLubyte* out = (GLubyte*) output[0].bgra;
 
     ITERATE(count) {
+        __asm__("pref @%0" : : "r"(cptr + cstride));
+
         func(cptr, out);
         cptr += cstride;
         out += sizeof(Vertex);
@@ -874,6 +887,8 @@ static void generateElementsFastPath(
     }
 }
 
+#define likely(x)      __builtin_expect(!!(x), 1)
+
 static void generateArraysFastPath(SubmissionTarget* target, const GLsizei first, const GLuint count) {
     Vertex* start = _glSubmissionTargetStart(target);
 
@@ -1290,7 +1305,6 @@ GL_FORCE_INLINE void submitVertices(GLenum mode, GLsizei first, GLuint count, GL
 
     }
 
-    divide(target);
     push(_glSubmissionTargetHeader(target), GL_FALSE, target->output, 0);
 
     /*
diff --git a/GL/flush.c b/GL/flush.c
index e3d8dc0..23d5c1f 100644
--- a/GL/flush.c
+++ b/GL/flush.c
@@ -42,10 +42,10 @@ void APIENTRY glKosInitConfig(GLdcConfig* config) {
     config->autosort_enabled = GL_FALSE;
     config->fsaa_enabled = GL_FALSE;
 
-    config->initial_op_capacity = 1024;
-    config->initial_pt_capacity = 512;
-    config->initial_tr_capacity = 1024;
-    config->initial_immediate_capacity = 1024;
+    config->initial_op_capacity = 1024 * 3;
+    config->initial_pt_capacity = 512 * 3;
+    config->initial_tr_capacity = 1024 * 3;
+    config->initial_immediate_capacity = 1024 * 3;
     config->internal_palette_format = GL_RGBA4;
 }
 
@@ -86,20 +86,65 @@ void APIENTRY glKosInit() {
     glKosInitEx(&config);
 }
 
+#define likely(x)      __builtin_expect(!!(x), 1)
+#define unlikely(x)    __builtin_expect(!!(x), 0)
+
+GL_FORCE_INLINE bool glIsVertex(const float flags) {
+    return flags == GPU_CMD_VERTEX_EOL || flags == GPU_CMD_VERTEX;
+}
+
+
+GL_FORCE_INLINE void glPerspectiveDivide(void* src, uint32_t n) {
+    TRACE();
+
+    /* Perform perspective divide on each vertex */
+    Vertex* vertex = (Vertex*) src;
+
+    const float h = GetVideoMode()->height;
+
+    while(n--) {
+        __asm__("pref @%0" : : "r"(vertex + 1));
+
+        if(likely(glIsVertex(vertex->flags))) {
+            const float f = MATH_Fast_Invert(vertex->w);
+
+            /* Convert to NDC and apply viewport */
+            vertex->xyz[0] = MATH_fmac(
+                VIEWPORT.hwidth, vertex->xyz[0] * f, VIEWPORT.x_plus_hwidth
+            );
+
+            vertex->xyz[1] = h - MATH_fmac(
+                VIEWPORT.hheight, vertex->xyz[1] * f, VIEWPORT.y_plus_hheight
+            );
+
+            /* Apply depth range */
+            vertex->xyz[2] = MAX(
+                1.0f - MATH_fmac(vertex->xyz[2] * f, 0.5f, 0.5f),
+                PVR_MIN_Z
+            );
+        }
+
+        ++vertex;
+    }
+}
+
 
 void APIENTRY glKosSwapBuffers() {
     TRACE();
 
     SceneBegin();
         SceneListBegin(GPU_LIST_OP_POLY);
+        glPerspectiveDivide(OP_LIST.vector.data, OP_LIST.vector.size);
         SceneListSubmit(OP_LIST.vector.data, OP_LIST.vector.size);
         SceneListFinish();
 
         SceneListBegin(GPU_LIST_PT_POLY);
+        glPerspectiveDivide(PT_LIST.vector.data, PT_LIST.vector.size);
         SceneListSubmit(PT_LIST.vector.data, PT_LIST.vector.size);
         SceneListFinish();
 
         SceneListBegin(GPU_LIST_TR_POLY);
+        glPerspectiveDivide(TR_LIST.vector.data, TR_LIST.vector.size);
         SceneListSubmit(TR_LIST.vector.data, TR_LIST.vector.size);
         SceneListFinish();
     SceneFinish();
diff --git a/GL/private.h b/GL/private.h
index c099360..ded524f 100644
--- a/GL/private.h
+++ b/GL/private.h
@@ -255,7 +255,6 @@ typedef struct {
     AlignedVector* extras;
 } SubmissionTarget;
 
-PolyHeader* _glSubmissionTargetHeader(SubmissionTarget* target);
 Vertex* _glSubmissionTargetStart(SubmissionTarget* target);
 Vertex* _glSubmissionTargetEnd(SubmissionTarget* target);
 
diff --git a/samples/quadmark/main.c b/samples/quadmark/main.c
index eca1261..26ea433 100644
--- a/samples/quadmark/main.c
+++ b/samples/quadmark/main.c
@@ -74,6 +74,8 @@ void setup() {
     glOrtho(0, 640, 0, 480, -100, 100);
     glMatrixMode(GL_PROJECTION);
     glLoadIdentity();
+
+    glDisable(GL_NEARZ_CLIPPING_KOS);
 }
 
 void do_frame() {