diff --git a/GL/matrix.c b/GL/matrix.c
index ebaeec1..1cb3453 100644
--- a/GL/matrix.c
+++ b/GL/matrix.c
@@ -32,18 +32,6 @@ static const Matrix4x4 IDENTITY = {
 
 GLfloat NEAR_PLANE_DISTANCE = 0.0f;
 
-static inline void upload_matrix(Matrix4x4* m) {
-    mat_load((matrix_t*) m);
-}
-
-static inline void multiply_matrix(Matrix4x4* m) {
-    mat_apply((matrix_t*) m);
-}
-
-static inline void download_matrix(Matrix4x4* m) {
-    mat_store((matrix_t*) m);
-}
-
 Matrix4x4* _glGetProjectionMatrix() {
     return (Matrix4x4*) stack_top(&MATRIX_STACKS[1]);
 }
@@ -64,6 +52,8 @@ void _glInitMatrices() {
     memcpy4(NORMAL_MATRIX, IDENTITY, sizeof(Matrix4x4));
     memcpy4(SCREENVIEW_MATRIX, IDENTITY, sizeof(Matrix4x4));
 
+    const VideoMode* vid_mode = GetVideoMode();
+
     glDepthRange(0.0f, 1.0f);
     glViewport(0, 0, vid_mode->width, vid_mode->height);
 }
@@ -143,9 +133,9 @@ void APIENTRY glTranslatef(GLfloat x, GLfloat y, GLfloat z) {
     trn[M13] = y;
     trn[M14] = z;
 
-    upload_matrix(stack_top(MATRIX_STACKS + MATRIX_IDX));
-    multiply_matrix(&trn);
-    download_matrix(stack_top(MATRIX_STACKS + MATRIX_IDX));
+    UploadMatrix4x4(stack_top(MATRIX_STACKS + MATRIX_IDX));
+    MultiplyMatrix4x4(&trn);
+    DownloadMatrix4x4(stack_top(MATRIX_STACKS + MATRIX_IDX));
 
     if(MATRIX_MODE == GL_MODELVIEW) {
         recalculateNormalMatrix();
@@ -165,9 +155,9 @@ void APIENTRY glScalef(GLfloat x, GLfloat y, GLfloat z) {
     scale[M5] = y;
     scale[M10] = z;
 
-    upload_matrix(stack_top(MATRIX_STACKS + MATRIX_IDX));
-    multiply_matrix(&scale);
-    download_matrix(stack_top(MATRIX_STACKS + MATRIX_IDX));
+    UploadMatrix4x4(stack_top(MATRIX_STACKS + MATRIX_IDX));
+    MultiplyMatrix4x4(&scale);
+    DownloadMatrix4x4(stack_top(MATRIX_STACKS + MATRIX_IDX));
 
     if(MATRIX_MODE == GL_MODELVIEW) {
         recalculateNormalMatrix();
@@ -208,9 +198,9 @@ void APIENTRY glRotatef(GLfloat angle, GLfloat x, GLfloat  y, GLfloat z) {
     rotate[M9] = yz * invc - xs;
     rotate[M10] = (z * z) * invc + c;
 
-    upload_matrix(stack_top(MATRIX_STACKS + MATRIX_IDX));
-    multiply_matrix(&rotate);
-    download_matrix(stack_top(MATRIX_STACKS + MATRIX_IDX));
+    UploadMatrix4x4(stack_top(MATRIX_STACKS + MATRIX_IDX));
+    MultiplyMatrix4x4(&rotate);
+    DownloadMatrix4x4(stack_top(MATRIX_STACKS + MATRIX_IDX));
 
     if(MATRIX_MODE == GL_MODELVIEW) {
         recalculateNormalMatrix();
@@ -268,9 +258,9 @@ void APIENTRY glOrtho(GLfloat left, GLfloat right,
     OrthoMatrix[M13] = -(top + bottom) / (top - bottom);
     OrthoMatrix[M14] = -(zfar + znear) / (zfar - znear);
 
-    upload_matrix(stack_top(MATRIX_STACKS + MATRIX_IDX));
-    multiply_matrix(&OrthoMatrix);
-    download_matrix(stack_top(MATRIX_STACKS + MATRIX_IDX));
+    UploadMatrix4x4(stack_top(MATRIX_STACKS + MATRIX_IDX));
+    MultiplyMatrix4x4(&OrthoMatrix);
+    DownloadMatrix4x4(stack_top(MATRIX_STACKS + MATRIX_IDX));
 }
 
 
@@ -282,7 +272,7 @@ void APIENTRY glFrustum(GLfloat left, GLfloat right,
     /* Frustum Matrix */
     static Matrix4x4 FrustumMatrix __attribute__((aligned(32)));
 
-    memset(FrustumMatrix, 0, sizeof(float) * 16);
+    MEMSET(FrustumMatrix, 0, sizeof(float) * 16);
 
     const float near2 = 2.0f * znear;
     const float A = (right + left) / (right - left);
@@ -299,9 +289,9 @@ void APIENTRY glFrustum(GLfloat left, GLfloat right,
     FrustumMatrix[M11] = -1.0f;
     FrustumMatrix[M14] = D;
 
-    upload_matrix(stack_top(MATRIX_STACKS + MATRIX_IDX));
-    multiply_matrix(&FrustumMatrix);
-    download_matrix(stack_top(MATRIX_STACKS + MATRIX_IDX));
+    UploadMatrix4x4(stack_top(MATRIX_STACKS + MATRIX_IDX));
+    MultiplyMatrix4x4(&FrustumMatrix);
+    DownloadMatrix4x4(stack_top(MATRIX_STACKS + MATRIX_IDX));
 }
 
 
@@ -329,9 +319,9 @@ void glMultMatrixf(const GLfloat *m) {
     TEMP[M14] = m[14];
     TEMP[M15] = m[15];
 
-    upload_matrix(stack_top(MATRIX_STACKS + MATRIX_IDX));
-    multiply_matrix((Matrix4x4*) &TEMP);
-    download_matrix(stack_top(MATRIX_STACKS + MATRIX_IDX));
+    UploadMatrix4x4(stack_top(MATRIX_STACKS + MATRIX_IDX));
+    MultiplyMatrix4x4((Matrix4x4*) &TEMP);
+    DownloadMatrix4x4(stack_top(MATRIX_STACKS + MATRIX_IDX));
 
     if(MATRIX_MODE == GL_MODELVIEW) {
         recalculateNormalMatrix();
@@ -396,9 +386,9 @@ void glMultTransposeMatrixf(const GLfloat *m) {
     TEMP[M14] = m[11];
     TEMP[M15] = m[15];
 
-    upload_matrix(stack_top(MATRIX_STACKS + MATRIX_IDX));
-    multiply_matrix(&TEMP);
-    download_matrix(stack_top(MATRIX_STACKS + MATRIX_IDX));
+    UploadMatrix4x4(stack_top(MATRIX_STACKS + MATRIX_IDX));
+    MultiplyMatrix4x4(&TEMP);
+    DownloadMatrix4x4(stack_top(MATRIX_STACKS + MATRIX_IDX));
 
     if(MATRIX_MODE == GL_MODELVIEW) {
         recalculateNormalMatrix();
@@ -407,6 +397,8 @@ void glMultTransposeMatrixf(const GLfloat *m) {
 
 /* Set the GL viewport */
 void APIENTRY glViewport(GLint x, GLint y, GLsizei width, GLsizei height) {
+    const VideoMode* vid_mode = GetVideoMode();
+
     gl_viewport_x1 = x;
     gl_viewport_y1 = y;
     gl_viewport_width = width;
diff --git a/GL/platform.h b/GL/platform.h
index 32e6cf6..374e1eb 100644
--- a/GL/platform.h
+++ b/GL/platform.h
@@ -3,6 +3,8 @@
 #include <stdint.h>
 #include <stdbool.h>
 
+#define MEMSET(dst, v, size) memset((dst), (v), (size))
+
 typedef enum GPUAlpha {
     GPU_ALPHA_DISABLE = 0,
     GPU_ALPHA_ENABLE = 1
@@ -147,6 +149,13 @@ typedef enum GPUTextureEnv {
     GPU_TXRENV_MODULATEALPHA = 3
 } GPUTextureEnv;
 
+typedef struct VideoMode {
+    uint16_t width;
+    uint16_t height;
+} VideoMode;
+
+const VideoMode* GetVideoMode();
+
 /* Duplication of pvr_poly_cxt_t from KOS so that we can
  * compile on non-KOS platforms for testing */
 
diff --git a/GL/platforms/sh4.h b/GL/platforms/sh4.h
index 03a589e..ec69673 100644
--- a/GL/platforms/sh4.h
+++ b/GL/platforms/sh4.h
@@ -9,6 +9,15 @@
 #include "../types.h"
 #include "sh4_math.h"
 
+#define FASTCPY(dst, src, bytes) \
+    (bytes % 32 == 0) ? sq_cpy(dst, src, bytes) : memcpy(dst, src, bytes)
+
+#define FASTCPY4(dst, src, bytes) \
+    (bytes % 32 == 0) ? sq_cpy(dst, src, bytes) : memcpy4(dst, src, bytes)
+
+#define MEMSET4(dst, v, size) memset4((dst), (v), (size))
+#define NORMALIZEVEC3(x, y, z) vec3f_normalize((x), (y), (z))
+
 inline void CompilePolyHeader(PolyHeader* out, const PolyContext* in) {
     pvr_poly_compile((pvr_poly_hdr_t*) out, (pvr_poly_cxt_t*) in);
 }
diff --git a/GL/platforms/x86.h b/GL/platforms/x86.h
index dba039a..e591877 100644
--- a/GL/platforms/x86.h
+++ b/GL/platforms/x86.h
@@ -1,14 +1,20 @@
 #pragma once
 
 #include <math.h>
+#include <memory.h>
+
 #include "../types.h"
 
 #define MATH_Fast_Divide(n, d) (n / d)
 #define MATH_fmac(a, b, c) (a * b + c)
-#define MATH_Fast_Sqrt(x) sqrt((x))
-#define MATH_fsrra(x) (1.0f / sqrt((x)))
+#define MATH_Fast_Sqrt(x) sqrtf((x))
+#define MATH_fsrra(x) (1.0f / sqrtf((x)))
 #define MATH_Fast_Invert(x) (1.0f / (x))
 
+#define FASTCPY(dst, src, bytes) memcpy(dst, src, bytes)
+#define FASTCPY4(dst, src, bytes) memcpy(dst, src, bytes)
+#define MEMSET4(dst, v, size) memset((dst), (v), (size))
+
 inline void CompilePolyHeader(PolyHeader* out, const PolyContext* in) {
     (void) out;
     (void) in;
diff --git a/GL/private.h b/GL/private.h
index b18cb26..9553b1a 100644
--- a/GL/private.h
+++ b/GL/private.h
@@ -18,12 +18,6 @@ extern void* memcpy4 (void *dest, const void *src, size_t count);
 #define GL_FORCE_INLINE static GL_INLINE_DEBUG
 #define _GL_UNUSED(x) (void)(x)
 
-#define FASTCPY(dst, src, bytes) \
-    (bytes % 32 == 0) ? sq_cpy(dst, src, bytes) : memcpy(dst, src, bytes);
-
-#define FASTCPY4(dst, src, bytes) \
-    (bytes % 32 == 0) ? sq_cpy(dst, src, bytes) : memcpy4(dst, src, bytes);
-
 #define _PACK4(v) ((v * 0xF) / 0xFF)
 #define PACK_ARGB4444(a,r,g,b) (_PACK4(a) << 12) | (_PACK4(r) << 8) | (_PACK4(g) << 4) | (_PACK4(b))
 #define PACK_ARGB8888(a,r,g,b) ( ((a & 0xFF) << 24) | ((r & 0xFF) << 16) | ((g & 0xFF) << 8) | (b & 0xFF) )
diff --git a/GL/state.c b/GL/state.c
index 8edfc38..c95391d 100644
--- a/GL/state.c
+++ b/GL/state.c
@@ -557,6 +557,9 @@ void APIENTRY glScissor(GLint x, GLint y, GLsizei width, GLsizei height) {
     PVRTileClipCommand *c = aligned_vector_extend(&_glActivePolyList()->vector, 1);
 
     GLint miny, maxx, maxy;
+
+    const VideoMode* vid_mode = GetVideoMode();
+
     GLsizei gl_scissor_width = MAX( MIN(width, vid_mode->width), 0 );
     GLsizei gl_scissor_height = MAX( MIN(height, vid_mode->height), 0 );
 
diff --git a/GL/texture.c b/GL/texture.c
index 6897b66..ae29b81 100644
--- a/GL/texture.c
+++ b/GL/texture.c
@@ -7,6 +7,7 @@
 #include <string.h>
 
 #include "config.h"
+#include "platform.h"
 #include "../include/glext.h"
 #include "../include/glkos.h"
 
@@ -53,7 +54,7 @@ static TexturePalette* _initTexturePalette() {
     TexturePalette* palette = (TexturePalette*) malloc(sizeof(TexturePalette));
     assert(palette);
 
-    memset4(palette, 0x0, sizeof(TexturePalette));
+    MEMSET4(palette, 0x0, sizeof(TexturePalette));
     palette->bank = -1;
     return palette;
 }
@@ -629,7 +630,7 @@ void APIENTRY glCompressedTexImage2DARB(GLenum target,
     active->data = yalloc_alloc_and_defrag(imageSize);
 
     if(data) {
-        sq_cpy(active->data, data, imageSize);
+        FASTCPY(active->data, data, imageSize);
     }
 }