diff --git a/src/gl/gl3device.cpp b/src/gl/gl3device.cpp
index db7c610..de7e07c 100644
--- a/src/gl/gl3device.cpp
+++ b/src/gl/gl3device.cpp
@@ -17,12 +17,8 @@
 
 #define PLUGIN_ID 0
 
-extern uint16_t *gIndices;
-extern float *gVertexBuffer;
-extern uint16_t *gIndicesPtr;
 uint16_t *gIndicesIm2DPtr;
 uint16_t *gIndicesIm3DPtr;
-extern float *gVertexBufferPtr;
 float *gVertexBufferIm2DPtr;
 float *gVertexBufferIm3DPtr;
 
@@ -1162,10 +1158,8 @@ showRaster(Raster *raster, uint32 flags)
 		vglWaitVblankStart(GL_FALSE);
 	vglStopRendering();
 	vglStartRendering();
-	gVertexBuffer     = gVertexBufferPtr;
 	gVertexBufferIm2D = gVertexBufferIm2DPtr;
 	gVertexBufferIm3D = gVertexBufferIm3DPtr;
-	gIndices     = gIndicesPtr;
 	gIndicesIm2D = gIndicesIm2DPtr;
 	gIndicesIm3D = gIndicesIm3DPtr;
 #endif
@@ -1417,23 +1411,19 @@ initOpenGL(void)
 
 	resetRenderState();
 
-	gVertexBufferPtr     = (float*)malloc(0x1800000);
 	gVertexBufferIm2DPtr = (float*)malloc(0x500000);
-	gVertexBufferIm3DPtr = (float*)malloc(0x500000);
-	gIndicesPtr       = (uint16_t*)malloc(0x600000);
+	gVertexBufferIm3DPtr = (float*)malloc(0x1800000);
 	gIndicesIm2DPtr   = (uint16_t*)malloc(0x600000);
-	gIndicesIm3DPtr   = (uint16_t*)malloc(0x600000);
+	gIndicesIm3DPtr   = (uint16_t*)malloc(0x1800000);
 
-	gConstIndices = (uint16_t*)malloc(sizeof(uint16_t) * 512);
+	gConstIndices = (uint16_t*)malloc(sizeof(uint16_t) * 0xFFFF);
 	
-	gVertexBuffer     = gVertexBufferPtr;
 	gVertexBufferIm2D = gVertexBufferIm2DPtr;
 	gVertexBufferIm3D = gVertexBufferIm3DPtr;
-	gIndices     = gIndicesPtr;
 	gIndicesIm2D = gIndicesIm2DPtr;
 	gIndicesIm3D = gIndicesIm3DPtr;
 	
-	for (uint16_t i = 0; i < 512; i++) {
+	for (uint16_t i = 0; i < 0xFFFF; i++) {
 		gConstIndices[i] = i;
 	}
 	
diff --git a/src/gl/gl3immed.cpp b/src/gl/gl3immed.cpp
index b0cb76a..829796b 100644
--- a/src/gl/gl3immed.cpp
+++ b/src/gl/gl3immed.cpp
@@ -103,11 +103,6 @@ im2DRenderPrimitive(PrimitiveType primType, void *vertices, int32 numVertices)
 	GLfloat xform[4];
 	Camera *cam;
 	cam = (Camera*)engine->currentCamera;
-
-	memcpy_neon(gVertexBufferIm2D, vertices, numVertices*sizeof(Im2DVertex));
-	vglVertexAttribPointerMapped(0, gVertexBufferIm2D);
-	vglIndexPointerMapped(gConstIndices);
-	gVertexBufferIm2D += numVertices*(sizeof(Im2DVertex)/sizeof(float));
 	
 	xform[0] = 2.0f/cam->frameBuffer->width;
 	xform[1] = -2.0f/cam->frameBuffer->height;
@@ -122,6 +117,11 @@ im2DRenderPrimitive(PrimitiveType primType, void *vertices, int32 numVertices)
 	flushCache();
 	glUniform4fv(currentShader->uniformLocations[u_xform], 1, xform);
 	
+	memcpy_neon(gVertexBufferIm2D, vertices, numVertices*sizeof(Im2DVertex));
+	vglVertexAttribPointerMapped(0, gVertexBufferIm2D);
+	vglIndexPointerMapped(gConstIndices);
+	gVertexBufferIm2D += numVertices*(sizeof(Im2DVertex)/sizeof(float));
+	
 	vglDrawObjects(primTypeMap[primType], numVertices, GL_FALSE);
 	
 }
@@ -135,14 +135,6 @@ im2DRenderIndexedPrimitive(PrimitiveType primType,
 	Camera *cam;
 	cam = (Camera*)engine->currentCamera;
 
-	memcpy_neon(gIndicesIm2D, indices, numIndices * 2);
-	vglIndexPointerMapped(gIndicesIm2D);
-	gIndicesIm2D += numIndices;
-	
-	memcpy_neon(gVertexBufferIm2D, vertices, numVertices*sizeof(Im2DVertex));
-	vglVertexAttribPointerMapped(0, gVertexBufferIm2D);
-	gVertexBufferIm2D += numVertices*(sizeof(Im2DVertex)/sizeof(float));
-
 	xform[0] = 2.0f/cam->frameBuffer->width;
 	xform[1] = -2.0f/cam->frameBuffer->height;
 	xform[2] = -1.0f;
@@ -155,6 +147,14 @@ im2DRenderIndexedPrimitive(PrimitiveType primType,
 
 	flushCache();
 	glUniform4fv(currentShader->uniformLocations[u_xform], 1, xform);
+	
+	memcpy_neon(gIndicesIm2D, indices, numIndices * 2);
+	vglIndexPointerMapped(gIndicesIm2D);
+	gIndicesIm2D += numIndices;
+	
+	memcpy_neon(gVertexBufferIm2D, vertices, numVertices*sizeof(Im2DVertex));
+	vglVertexAttribPointerMapped(0, gVertexBufferIm2D);
+	gVertexBufferIm2D += numVertices*(sizeof(Im2DVertex)/sizeof(float));
 		
 	vglDrawObjects(primTypeMap[primType], numIndices, GL_FALSE);
 }
@@ -205,7 +205,7 @@ void
 im3DTransform(void *vertices, int32 numVertices, Matrix *world, uint32 flags)
 {
 	if(world == nil){
-		Matrix ident;
+		static Matrix ident;
 		ident.setIdentity();
 		world = &ident;
 	}
@@ -236,13 +236,12 @@ im3DRenderPrimitive(PrimitiveType primType)
 void
 im3DRenderIndexedPrimitive(PrimitiveType primType, void *indices, int32 numIndices)
 {
+	flushCache();
 	memcpy_neon(gIndicesIm3D, indices, numIndices * 2);
 	vglIndexPointerMapped(gIndicesIm3D);
 	vglVertexAttribPointerMapped(0, gVertexBufferIm3D);
 	gVertexBufferIm3D += num3DVertices*(sizeof(Im3DVertex)/sizeof(float));
 	gIndicesIm3D += numIndices;
-
-	flushCache();
 	vglDrawObjects(primTypeMap[primType], numIndices, GL_FALSE);
 }
 
diff --git a/src/gl/gl3matfx.cpp b/src/gl/gl3matfx.cpp
index 94d2478..cfc82fe 100644
--- a/src/gl/gl3matfx.cpp
+++ b/src/gl/gl3matfx.cpp
@@ -33,17 +33,18 @@ static int32 u_colorClamp;
 void
 matfxDefaultRender(InstanceDataHeader *header, InstanceData *inst)
 {
-	Material *m;
-	m = inst->material;
+	Material *m = inst->material;
 
 	defaultShader->use();
+	
+	rw::SetRenderState(VERTEXALPHA, inst->vertexAlpha || m->color.alpha != 0xFF);
+
+	flushCache();
 
 	setMaterial(m->color, m->surfaceProps);
 
 	setTexture(0, m->texture);
 
-	rw::SetRenderState(VERTEXALPHA, inst->vertexAlpha || m->color.alpha != 0xFF);
-
 	drawInst(header, inst);
 }
 
@@ -89,6 +90,11 @@ matfxEnvRender(InstanceDataHeader *header, InstanceData *inst, MatFX::Env *env)
 	}
 
 	envShader->use();
+	
+	rw::SetRenderState(VERTEXALPHA, 1);
+	rw::SetRenderState(SRCBLEND, BLENDONE);
+	
+	flushCache();
 
 	setTexture(0, m->texture);
 	setTexture(1, env->tex);
@@ -109,9 +115,6 @@ matfxEnvRender(InstanceDataHeader *header, InstanceData *inst, MatFX::Env *env)
 	else
 		glUniform4fv(U(u_colorClamp), 1, one);
 
-	rw::SetRenderState(VERTEXALPHA, 1);
-	rw::SetRenderState(SRCBLEND, BLENDONE);
-
 	drawInst(header, inst);
 
 	rw::SetRenderState(SRCBLEND, BLENDSRCALPHA);
diff --git a/src/gl/gl3pipe.cpp b/src/gl/gl3pipe.cpp
index 80b75b2..13919cd 100644
--- a/src/gl/gl3pipe.cpp
+++ b/src/gl/gl3pipe.cpp
@@ -12,9 +12,6 @@
 #include "rwgl3.h"
 #include "rwgl3shader.h"
 
-extern float *gVertexBuffer;
-extern uint16_t *gIndices;
-
 namespace rw {
 namespace gl3 {
 
diff --git a/src/gl/gl3render.cpp b/src/gl/gl3render.cpp
index e8716bb..cb4e54e 100644
--- a/src/gl/gl3render.cpp
+++ b/src/gl/gl3render.cpp
@@ -15,11 +15,6 @@
 
 #include "rwgl3impl.h"
 
-extern uint16_t *gIndices;
-extern float *gVertexBuffer;
-extern uint16_t *gIndicesPtr;
-extern float *gVertexBufferPtr;
-
 namespace rw {
 namespace gl3 {
 
@@ -28,13 +23,8 @@ namespace gl3 {
 void
 drawInst_simple(InstanceDataHeader *header, InstanceData *inst)
 {
-	flushCache();
-	memcpy_neon(gIndices, (uint8_t*)header->indexBuffer + inst->offset, inst->numIndex * 2);
-	vglIndexPointerMapped(gIndices);
-	gIndices += inst->numIndex;
-	memcpy_neon(gVertexBuffer, (uint8_t*)header->vertexBuffer, inst->numVertices * header->attribDesc[0].stride);
-	vglVertexAttribPointerMapped(0, gVertexBuffer);
-	gVertexBuffer += (inst->numVertices * header->attribDesc[0].stride) / sizeof(float);
+	vglIndexPointerMapped(header->indexBuffer + inst->offset);
+	vglVertexAttribPointerMapped(0, header->vertexBuffer);
 	vglDrawObjects(header->primType, inst->numIndex, GL_FALSE);
 }
 
@@ -140,12 +130,14 @@ defaultRenderCB(Atomic *atomic, InstanceDataHeader *header)
 	while(n--){
 		m = inst->material;
 
+		rw::SetRenderState(VERTEXALPHA, inst->vertexAlpha || m->color.alpha != 0xFF);
+		
+		flushCache();
+		
 		setMaterial(m->color, m->surfaceProps);
 
 		setTexture(0, m->texture);
 
-		rw::SetRenderState(VERTEXALPHA, inst->vertexAlpha || m->color.alpha != 0xFF);
-
 		drawInst(header, inst);
 		inst++;
 	}
diff --git a/src/gl/gl3shader.cpp b/src/gl/gl3shader.cpp
index 7eb88fb..5cce9d8 100644
--- a/src/gl/gl3shader.cpp
+++ b/src/gl/gl3shader.cpp
@@ -149,12 +149,12 @@ linkprogram(GLint vs, GLint fs, GLuint *program, bool is_2d)
 	
 	int stride = 0;
 	int pos_size = is_2d ? 4 : 3;
-	stride += vglBindPackedAttribLocation(prog, "in_pos"    , pos_size, GL_FLOAT, stride, stride + sizeof(float) * pos_size) * (sizeof(float) * pos_size);
-	stride += vglBindPackedAttribLocation(prog, "in_normal" ,        3, GL_FLOAT, stride, stride + sizeof(float) * 3) * (sizeof(float) * 3);
+	stride += vglBindPackedAttribLocation(prog, "in_pos"    , pos_size, GL_FLOAT,         stride, stride + sizeof(float) * pos_size) * (sizeof(float) * pos_size);
+	stride += vglBindPackedAttribLocation(prog, "in_normal" ,        3, GL_FLOAT,         stride, stride + sizeof(float) * 3) * (sizeof(float) * 3);
 	stride += vglBindPackedAttribLocation(prog, "in_color"  ,        4, GL_UNSIGNED_BYTE, stride, stride + 4) * 4;
-	stride += vglBindPackedAttribLocation(prog, "in_tex0"   ,        2, GL_FLOAT, stride, stride + sizeof(float) * 2) * (sizeof(float) * 2);
-	stride += vglBindPackedAttribLocation(prog, "in_weights",        4, GL_FLOAT, stride, stride + sizeof(float) * 4) * (sizeof(float) * 4);
-	vglBindPackedAttribLocation(prog, "in_indices",        4, GL_UNSIGNED_BYTE, stride, stride + 4);
+	stride += vglBindPackedAttribLocation(prog, "in_tex0"   ,        2, GL_FLOAT,         stride, stride + sizeof(float) * 2) * (sizeof(float) * 2);
+	stride += vglBindPackedAttribLocation(prog, "in_weights",        4, GL_FLOAT,         stride, stride + sizeof(float) * 4) * (sizeof(float) * 4);
+	          vglBindPackedAttribLocation(prog, "in_indices",        4, GL_UNSIGNED_BYTE, stride, stride + 4);
 	
 	glLinkProgram(prog);
 
diff --git a/src/gl/gl3skin.cpp b/src/gl/gl3skin.cpp
index fda79ef..134cb76 100644
--- a/src/gl/gl3skin.cpp
+++ b/src/gl/gl3skin.cpp
@@ -254,17 +254,21 @@ skinRenderCB(Atomic *atomic, InstanceDataHeader *header)
 
 	skinShader->use();
 
-	uploadSkinMatrices(atomic);
+	
 
 	while(n--){
 		m = inst->material;
+		
+		rw::SetRenderState(VERTEXALPHA, inst->vertexAlpha || m->color.alpha != 0xFF);
+		
+		flushCache();
 
 		setMaterial(m->color, m->surfaceProps);
 
 		setTexture(0, m->texture);
-
-		rw::SetRenderState(VERTEXALPHA, inst->vertexAlpha || m->color.alpha != 0xFF);
-
+		
+		uploadSkinMatrices(atomic);
+		
 		drawInst(header, inst);
 		inst++;
 	}
diff --git a/src/gl/wdgl.cpp b/src/gl/wdgl.cpp
index 1f56815..43ae467 100644
--- a/src/gl/wdgl.cpp
+++ b/src/gl/wdgl.cpp
@@ -28,11 +28,6 @@
 
 #define PLUGIN_ID 2
 
-uint16_t *gIndices;
-float *gVertexBuffer;
-uint16_t *gIndicesPtr;
-float *gVertexBufferPtr;
-
 namespace rw {
 namespace wdgl {