diff --git a/project/jni/glshim/src/gl/gl.c b/project/jni/glshim/src/gl/gl.c
index 5bad5838f..6b0011cb2 100755
--- a/project/jni/glshim/src/gl/gl.c
+++ b/project/jni/glshim/src/gl/gl.c
@@ -30,7 +30,8 @@ GLuint readhack_seq = 0;
 GLuint gl_batch = 0;
 GLuint gl_mergelist = 1;
 int blendhack = 0;
-char gl_version[50];
+int export_blendcolor = 0;
+char glshim_version[50];
 int initialized = 0;
 int noerror = 0;
 
@@ -121,7 +122,7 @@ const GLubyte *glshim_glGetString(GLenum name) {
 		printf("**warning** glGetString(%i) called with bad init\n", name);*/
     switch (name) {
         case GL_VERSION:
-            return (GLubyte *)gl_version;
+            return (GLubyte *)glshim_version;
         case GL_EXTENSIONS:
             return (const GLubyte *)(char *){
                 "GL_EXT_abgr "
@@ -654,8 +655,6 @@ static renderlist_t *arrays_to_renderlist(renderlist_t *list, GLenum mode,
 		    list->tex[i] = copy_gl_pointer_tex(&glstate.vao->pointers.tex_coord[i], 4, skip, count, glstate.vao->pointers.tex_coord[i].buffer);
 		}
 	}
-	
-    end_renderlist(list);
     return list;
 }
 
@@ -672,7 +671,7 @@ static inline bool should_intercept_render(GLenum mode) {
 }
 
 void glshim_glDrawElements(GLenum mode, GLsizei count, GLenum type, const GLvoid *indices) {
-//printf("glDrawElements(0x%04X, %d, 0x%04X, %p), map=%p\n", mode, count, type, indices, (glstate.buffers.elements)?glstate.buffers.elements->data:NULL);
+//printf("glDrawElements(0x%04X, %d, 0x%04X, %p), map=%p\n", mode, count, type, indices, (glstate.vao->elements)?glstate.vao->elements->data:NULL);
     // TODO: split for count > 65535?
     // special check for QUADS and TRIANGLES that need multiple of 4 or 3 vertex...
     if (mode == GL_QUADS) while(count%4) count--;
@@ -914,6 +913,7 @@ void glshim_glDrawArrays(GLenum mode, GLint first, GLsizei count) {
     if (glstate.list.active && (glstate.list.compiling || glstate.gl_batch)) {
         NewStage(glstate.list.active, STAGE_DRAW);
         glstate.list.active = arrays_to_renderlist(glstate.list.active, mode, first, count+first);
+        end_renderlist(list);
         return;
     }
 
@@ -1209,7 +1209,7 @@ void glBegin(GLenum mode) __attribute__((alias("glshim_glBegin")));
 
 void glshim_glEnd() {
     if (!glstate.list.active) return;
-    // check if TEXTUREx is activate and no TexCoord (or texgen), in that cas, create a dummy one base on glstate...
+    // check if TEXTUREx is activate and no TexCoord (or texgen), in that case, create a dummy one base on glstate...
     for (int a=0; a<MAX_TEX; a++)
 		if (glstate.enable.texture_2d[a] && ((glstate.list.active->tex[a]==0) && (!glstate.enable.texgen_s[a])))
 			rlMultiTexCoord4f(glstate.list.active, GL_TEXTURE0+a, glstate.texcoord[a][0], glstate.texcoord[a][1], glstate.texcoord[a][2], glstate.texcoord[a][3]);
@@ -1228,7 +1228,6 @@ void glshim_glEnd() {
 void glEnd() __attribute__((alias("glshim_glEnd")));
 
 void glshim_glNormal3f(GLfloat nx, GLfloat ny, GLfloat nz) {
-    glstate.normal[0] = nx; glstate.normal[1] = ny; glstate.normal[2] = nz;
     if (glstate.list.active) {
         if (glstate.list.active->stage != STAGE_DRAW) {
             if (glstate.list.active->stage != STAGE_DRAW) {
@@ -1246,6 +1245,7 @@ void glshim_glNormal3f(GLfloat nx, GLfloat ny, GLfloat nz) {
         errorGL();
     }
 #endif
+    glstate.normal[0] = nx; glstate.normal[1] = ny; glstate.normal[2] = nz;
 }
 void glNormal3f(GLfloat nx, GLfloat ny, GLfloat nz) __attribute__((alias("glshim_glNormal3f")));
 
@@ -1258,9 +1258,6 @@ void glshim_glVertex4f(GLfloat x, GLfloat y, GLfloat z, GLfloat w) {
 void glVertex4f(GLfloat x, GLfloat y, GLfloat z, GLfloat w) __attribute__((alias("glshim_glVertex4f")));
 
 void glshim_glColor4f(GLfloat red, GLfloat green, GLfloat blue, GLfloat alpha) {
-    // change the state first thing
-    glstate.color[0] = red; glstate.color[1] = green;
-    glstate.color[2] = blue; glstate.color[3] = alpha;
     if (glstate.list.active) {
         if (glstate.list.active->stage != STAGE_DRAW) {
             PUSH_IF_COMPILING(glColor4f);
@@ -1275,19 +1272,22 @@ void glshim_glColor4f(GLfloat red, GLfloat green, GLfloat blue, GLfloat alpha) {
         errorGL();
     }
 #endif
+    // change the state last thing
+    glstate.color[0] = red; glstate.color[1] = green;
+    glstate.color[2] = blue; glstate.color[3] = alpha;
 }
 void glColor4f(GLfloat red, GLfloat green, GLfloat blue, GLfloat alpha) __attribute__((alias("glshim_glColor4f")));
 
 void glshim_glSecondaryColor3f(GLfloat r, GLfloat g, GLfloat b) {
-    // change the state first thing
-    glstate.secondary[0] = r; glstate.secondary[1] = g;
-    glstate.secondary[2] = b;
     if (glstate.list.active) {
         rlSecondary3f(glstate.list.active, r, g, b);
         noerrorShim();
     } else {
         noerrorShim();
     }
+    // change the state last thing
+    glstate.secondary[0] = r; glstate.secondary[1] = g;
+    glstate.secondary[2] = b;
 }
 void glSecondaryColor3f(GLfloat r, GLfloat g, GLfloat b) __attribute__((alias("glshim_glSecondaryColor3f")));
 
@@ -1329,20 +1329,18 @@ void glMaterialf(GLenum face, GLenum pname, const GLfloat param) __attribute__((
 #endif
 
 void glshim_glTexCoord4f(GLfloat s, GLfloat t, GLfloat r, GLfloat q) {
-    glstate.texcoord[0][0] = s; glstate.texcoord[0][1] = t;
-    glstate.texcoord[0][2] = r; glstate.texcoord[0][3] = q;
     if (glstate.list.active) {
         rlTexCoord4f(glstate.list.active, s, t, r, q);
         noerrorShim();
     } else {
         noerrorShim();
     }
+    glstate.texcoord[0][0] = s; glstate.texcoord[0][1] = t;
+    glstate.texcoord[0][2] = r; glstate.texcoord[0][3] = q;
 }
 void glTexCoord4f(GLfloat s, GLfloat t, GLfloat r, GLfloat q) __attribute__((alias("glshim_glTexCoord4f")));
 
 void glshim_glMultiTexCoord4f(GLenum target, GLfloat s, GLfloat t, GLfloat r, GLfloat q) {
-    glstate.texcoord[target-GL_TEXTURE0][0] = s; glstate.texcoord[target-GL_TEXTURE0][1] = t;
-    glstate.texcoord[target-GL_TEXTURE0][2] = r; glstate.texcoord[target-GL_TEXTURE0][3] = q;
 	// TODO, error if target is unsuported texture....
     if (glstate.list.active) {
         rlMultiTexCoord4f(glstate.list.active, target, s, t, r, q);
@@ -1350,6 +1348,8 @@ void glshim_glMultiTexCoord4f(GLenum target, GLfloat s, GLfloat t, GLfloat r, GL
     } else {
         noerrorShim();
     }
+    glstate.texcoord[target-GL_TEXTURE0][0] = s; glstate.texcoord[target-GL_TEXTURE0][1] = t;
+    glstate.texcoord[target-GL_TEXTURE0][2] = r; glstate.texcoord[target-GL_TEXTURE0][3] = q;
 }
 void glMultiTexCoord4f(GLenum target, GLfloat s, GLfloat t, GLfloat r, GLfloat q) __attribute__((alias("glshim_glMultiTexCoord4f")));
 void glMultiTexCoord4fARB(GLenum target, GLfloat s, GLfloat t, GLfloat r, GLfloat q) __attribute__((alias("glshim_glMultiTexCoord4f")));
@@ -1744,10 +1744,17 @@ void glshim_glBlendColor(GLclampf red, GLclampf green, GLclampf blue, GLclampf a
     LOAD_GLES_OES(glBlendColor);
 	if  (gles_glBlendColor)
 		gles_glBlendColor(red, green, blue, alpha);
-	else
-		printf("stub glBlendColor(%f, %f, %f, %f)\n", red, green, blue, alpha);
+	else {
+        static int test = 1;
+        if (test) {
+            printf("stub glBlendColor(%f, %f, %f, %f)\n", red, green, blue, alpha);
+            test = 0;
+        }
+    }
 }
 void glBlendColor(GLclampf red, GLclampf green, GLclampf blue, GLclampf alpha) __attribute__((alias("glshim_glBlendColor")));
+void glBlendColorEXT(GLclampf red, GLclampf green, GLclampf blue, GLclampf alpha) __attribute__((alias("glshim_glBlendColor")));
+void glBlendColorARB(GLclampf red, GLclampf green, GLclampf blue, GLclampf alpha) __attribute__((alias("glshim_glBlendColor")));
 
 void glshim_glBlendFuncSeparate(GLenum sfactorRGB, GLenum dfactorRGB, GLenum sfactorAlpha, GLenum dfactorAlpha)
 {
@@ -1848,6 +1855,15 @@ void glshim_glBlendFunc(GLenum sfactor, GLenum dfactor) {
 }
 void glBlendFunc(GLenum sfactor, GLenum dfactor) __attribute__((alias("glshim_glBlendFunc")));
 
+
+void glshim_glStencilMaskSeparate(GLenum face, GLuint mask) {
+    // fake function..., call it only for front or front_and_back, just ignore back (crappy, I know)
+    if ((face==GL_FRONT) || (face==GL_FRONT_AND_BACK))
+        glshim_glStencilMask(mask);
+}
+void glStencilMaskSeparate(GLenum face, GLuint mask) __attribute__((alias("glshim_glStencilMaskSeparate")));
+
+
 void init_statebatch() {
     memset(&glstate.statebatch, 0, sizeof(statebatch_t));
 }
diff --git a/project/jni/glshim/src/gl/gl.h b/project/jni/glshim/src/gl/gl.h
index d258f7e5f..38c79653b 100755
--- a/project/jni/glshim/src/gl/gl.h
+++ b/project/jni/glshim/src/gl/gl.h
@@ -384,6 +384,8 @@ void glshim_glLoadMatrixf(const GLfloat * m);
 void glshim_glMultMatrixf(const GLfloat * m);
 void glshim_glFogfv(GLenum pname, const GLfloat* params);
 
+void glshim_glStencilMaskSeparate(GLenum face, GLuint mask);
+
 
 void flush();
 void init_batch();
diff --git a/project/jni/glshim/src/gl/list.c b/project/jni/glshim/src/gl/list.c
index 7828860aa..a2760350b 100755
--- a/project/jni/glshim/src/gl/list.c
+++ b/project/jni/glshim/src/gl/list.c
@@ -604,7 +604,7 @@ void adjust_renderlist(renderlist_t *list) {
 }
 
 void end_renderlist(renderlist_t *list) {
-    if (! list->open)
+    if (!list || ! list->open)
         return;
 
     adjust_renderlist(list);
@@ -1112,7 +1112,7 @@ void rlNormal3f(renderlist_t *list, GLfloat x, GLfloat y, GLfloat z) {
         list->normal = alloc_sublist(3, list->cap);
         // catch up
         int i;
-        if (list->len) for (i = 0; i < list->len-1; i++) {
+        if (list->len) for (i = 0; i < list->len; i++) {
             GLfloat *normal = (list->normal + (i * 3));
             memcpy(normal, list->lastNormal, sizeof(GLfloat) * 3);
         }
@@ -1127,7 +1127,7 @@ void rlColor4f(renderlist_t *list, GLfloat r, GLfloat g, GLfloat b, GLfloat a) {
         list->color = alloc_sublist(4, list->cap);
         // catch up
         int i;
-        if (list->len) for (i = 0; i < list->len-1; i++) {
+        if (list->len) for (i = 0; i < list->len; i++) {
             GLfloat *color = (list->color + (i * 4));
             memcpy(color, glstate.color, sizeof(GLfloat) * 4);
         }/* else {
@@ -1145,7 +1145,7 @@ void rlSecondary3f(renderlist_t *list, GLfloat r, GLfloat g, GLfloat b) {
         list->secondary = alloc_sublist(4, list->cap);
         // catch up
         int i;
-        if (list->len) for (i = 0; i < list->len-1; i++) {
+        if (list->len) for (i = 0; i < list->len; i++) {
             GLfloat *secondary = (list->secondary + (i * 4));
             memcpy(secondary, glstate.secondary, sizeof(GLfloat) * 4);
         }
@@ -1250,7 +1250,7 @@ void rlTexCoord4f(renderlist_t *list, GLfloat s, GLfloat t, GLfloat r, GLfloat q
         list->tex[0] = alloc_sublist(4, list->cap);
         // catch up
         GLfloat *tex = list->tex[0];
-        if (list->len) for (int i = 0; i < list->len-1; i++) {
+        if (list->len) for (int i = 0; i < list->len; i++) {
             memcpy(tex, glstate.texcoord[0], sizeof(GLfloat) * 4);
             tex += 4;
         }
@@ -1267,7 +1267,7 @@ void rlMultiTexCoord4f(renderlist_t *list, GLenum target, GLfloat s, GLfloat t,
         list->tex[tmu] = alloc_sublist(4, list->cap);
         // catch up
         GLfloat *tex = list->tex[tmu];
-        if (list->len) for (int i = 0; i < list->len-1; i++) {
+        if (list->len) for (int i = 0; i < list->len; i++) {
             memcpy(tex, glstate.texcoord[tmu], sizeof(GLfloat) * 4);
             tex += 4;
         }
diff --git a/project/jni/glshim/src/gl/pixel.c b/project/jni/glshim/src/gl/pixel.c
index c5fb82251..fb50e8d8c 100755
--- a/project/jni/glshim/src/gl/pixel.c
+++ b/project/jni/glshim/src/gl/pixel.c
@@ -713,12 +713,13 @@ bool pixel_convert(const GLvoid *src, GLvoid **dst,
         }
         return true;
     }
-    if ((src_format == GL_BGRA) && (dst_format == GL_LUMINANCE_ALPHA) && (dst_type == GL_UNSIGNED_BYTE) && ((src_type == GL_UNSIGNED_BYTE)||(src_type == GL_UNSIGNED_INT_8_8_8_8_REV))) {
+    if (((src_format == GL_BGRA)||(src_format == GL_RGBA)) && (dst_format == GL_LUMINANCE_ALPHA) && (dst_type == GL_UNSIGNED_BYTE) && ((src_type == GL_UNSIGNED_BYTE)||(src_type == GL_UNSIGNED_INT_8_8_8_8_REV))) {
         GLuint tmp;
         for (int i = 0; i < height; i++) {
 			for (int j = 0; j < width; j++) {
 				tmp = *(const GLuint*)src_pos;
-				*(GLushort*)dst_pos = (tmp&0x0000ff00) | (tmp&0x000000ff);
+                *(GLushort*)dst_pos = (((((char*)src_pos)[2] + ((char*)src_pos)[1] + ((char*)src_pos)[0])/3)&0xff)<<8 | ((char*)src_pos)[3];
+//				*(GLushort*)dst_pos = (tmp&0x0000ff00) | (tmp&0x000000ff);
 				src_pos += src_stride;
 				dst_pos += dst_stride;
 			}
@@ -727,8 +728,7 @@ bool pixel_convert(const GLvoid *src, GLvoid **dst,
         }
         return true;
     }
-    if ((src_format == GL_BGR) && (dst_format == GL_RGB) && (dst_type == GL_UNSIGNED_BYTE) && ((src_type == GL_UNSIGNED_BYTE)||(src_type == GL_UNSIGNED_INT_8_8_8_8_REV))) {
-        GLuint tmp;
+    if (((src_format == GL_BGR)||(src_format == GL_BGRA)) && (dst_format == GL_RGB) && (dst_type == GL_UNSIGNED_BYTE) && ((src_type == GL_UNSIGNED_BYTE)||(src_type == GL_UNSIGNED_INT_8_8_8_8_REV))) {
         for (int i = 0; i < height; i++) {
 			for (int j = 0; j < width; j++) {
 				((char*)dst_pos)[0] = ((char*)src_pos)[2];
@@ -742,8 +742,21 @@ bool pixel_convert(const GLvoid *src, GLvoid **dst,
         }
         return true;
     }
-    if ((src_format == GL_RGB) && (dst_format == GL_RGB) && (dst_type = GL_UNSIGNED_SHORT_5_6_5) && ((src_type == GL_UNSIGNED_BYTE)||(src_type == GL_UNSIGNED_INT_8_8_8_8_REV))) {
-        GLuint tmp;
+    if ((src_format == GL_RGBA) && (dst_format == GL_RGB) && (dst_type == GL_UNSIGNED_BYTE) && ((src_type == GL_UNSIGNED_BYTE)||(src_type == GL_UNSIGNED_INT_8_8_8_8_REV))) {
+        for (int i = 0; i < height; i++) {
+			for (int j = 0; j < width; j++) {
+				((char*)dst_pos)[0] = ((char*)src_pos)[0];
+				((char*)dst_pos)[1] = ((char*)src_pos)[1];
+				((char*)dst_pos)[2] = ((char*)src_pos)[2];
+				src_pos += src_stride;
+				dst_pos += dst_stride;
+			}
+			if (stride)
+				dst_pos += dst_width;
+        }
+        return true;
+    }
+    if (((src_format == GL_RGB)||(src_format == GL_RGBA)) && (dst_format == GL_RGB) && (dst_type = GL_UNSIGNED_SHORT_5_6_5) && ((src_type == GL_UNSIGNED_BYTE)||(src_type == GL_UNSIGNED_INT_8_8_8_8_REV))) {
         for (int i = 0; i < height; i++) {
 			for (int j = 0; j < width; j++) {
 				*(GLushort*)dst_pos = ((GLushort)(((char*)src_pos)[2]&0xf8)>>(3)) | ((GLushort)(((char*)src_pos)[1]&0xfc)<<(5-2)) | ((GLushort)(((char*)src_pos)[0]&0xf8)<<(11-3));
@@ -756,7 +769,6 @@ bool pixel_convert(const GLvoid *src, GLvoid **dst,
         return true;
     }
     if (((src_format == GL_BGR) || (src_format == GL_BGRA)) && (dst_format == GL_RGB) && (dst_type = GL_UNSIGNED_SHORT_5_6_5) && ((src_type == GL_UNSIGNED_BYTE)||(src_type == GL_UNSIGNED_INT_8_8_8_8_REV))) {
-        GLuint tmp;
         for (int i = 0; i < height; i++) {
 			for (int j = 0; j < width; j++) {
 				*(GLushort*)dst_pos = ((GLushort)(((char*)src_pos)[0]&0xf8)>>(3)) | ((GLushort)(((char*)src_pos)[1]&0xfc)<<(5-2)) | ((GLushort)(((char*)src_pos)[2]&0xf8)<<(11-3));
diff --git a/project/jni/glshim/src/gl/texture.c b/project/jni/glshim/src/gl/texture.c
index b39489b58..83245eed1 100755
--- a/project/jni/glshim/src/gl/texture.c
+++ b/project/jni/glshim/src/gl/texture.c
@@ -104,6 +104,21 @@ void tex_setup_texcoord(GLuint texunit, GLuint len) {
 
 int nolumalpha = 0;
 
+static int is_fake_compressed_rgb(GLenum internalformat)
+{
+    if(internalformat==GL_COMPRESSED_RGB) return 1;
+    if(internalformat==GL_COMPRESSED_RGB_S3TC_DXT1_EXT) return 1;
+    return 0;
+}
+static int is_fake_compressed_rgba(GLenum internalformat)
+{
+    if(internalformat==GL_COMPRESSED_RGBA) return 1;
+    if(internalformat==GL_COMPRESSED_RGBA_S3TC_DXT1_EXT) return 1;
+    if(internalformat==GL_COMPRESSED_RGBA_S3TC_DXT3_EXT) return 1;
+    if(internalformat==GL_COMPRESSED_RGBA_S3TC_DXT5_EXT) return 1;
+    return 0;
+}
+
 static void *swizzle_texture(GLsizei width, GLsizei height,
                              GLenum *format, GLenum *type,
                              GLenum intermediaryformat, GLenum internalformat,
@@ -190,10 +205,10 @@ static void *swizzle_texture(GLsizei width, GLsizei height,
             break;
     }
     // compressed format are not handled here, so mask them....
-    if (intermediaryformat==GL_COMPRESSED_RGB) intermediaryformat=GL_RGB;
-    if (intermediaryformat==GL_COMPRESSED_RGBA) intermediaryformat=GL_RGBA;
-    if (internalformat==GL_COMPRESSED_RGB) internalformat=GL_RGB;
-    if (internalformat==GL_COMPRESSED_RGBA) internalformat=GL_RGBA;
+    if (is_fake_compressed_rgb(intermediaryformat)) intermediaryformat=GL_RGB;
+    if (is_fake_compressed_rgba(intermediaryformat)) intermediaryformat=GL_RGBA;
+    if (is_fake_compressed_rgb(internalformat)) internalformat=GL_RGB;
+    if (is_fake_compressed_rgba(internalformat)) internalformat=GL_RGBA;
     
     if(*format != intermediaryformat || intermediaryformat!=internalformat) {
         dest_format = intermediaryformat;
@@ -332,10 +347,11 @@ GLenum swizzle_internalformat(GLenum *internalformat) {
             ret = GL_COMPRESSED_RGB;
             sret = GL_RGB;
             break;
+        case GL_COMPRESSED_RGBA_S3TC_DXT1_EXT:  // not good...
         case GL_COMPRESSED_RGBA_S3TC_DXT3_EXT:  // not good, but there is no DXT3 compressor
         case GL_COMPRESSED_RGBA_S3TC_DXT5_EXT:
             ret = GL_COMPRESSED_RGBA;
-            sret = GL_RGB;
+            sret = GL_RGBA;
             break;
         default:
             ret = GL_RGBA;
@@ -359,6 +375,7 @@ static int default_tex_mipmap = 0;
 
 static int proxy_width = 0;
 static int proxy_height = 0;
+static GLint proxy_intformat = 0;
 
 void glshim_glTexImage2D(GLenum target, GLint level, GLint internalformat,
                   GLsizei width, GLsizei height, GLint border,
@@ -367,8 +384,9 @@ void glshim_glTexImage2D(GLenum target, GLint level, GLint internalformat,
     //printf("glTexImage2D on target=%s with unpack_row_length(%i), size(%i,%i) and skip(%i,%i), format(internal)=%s(%s), type=%s, data=%08x, level=%i (mipmap_need=%i, mipmap_auto=%i) => texture=%u (streamed=%i), glstate.list.compiling=%d\n", PrintEnum(target), glstate.texture.unpack_row_length, width, height, glstate.texture.unpack_skip_pixels, glstate.texture.unpack_skip_rows, PrintEnum(format), PrintEnum(internalformat), PrintEnum(type), data, level, (glstate.texture.bound[glstate.texture.active])?glstate.texture.bound[glstate.texture.active]->mipmap_need:0, (glstate.texture.bound[glstate.texture.active])?glstate.texture.bound[glstate.texture.active]->mipmap_auto:0, (glstate.texture.bound[glstate.texture.active])?glstate.texture.bound[glstate.texture.active]->texture:0, (glstate.texture.bound[glstate.texture.active])?glstate.texture.bound[glstate.texture.active]->streamed:0, glstate.list.compiling);
     // proxy case
     if (target == GL_PROXY_TEXTURE_2D) {
-        proxy_width = ((width<<level)>(texshrink==8)?8192:2048)?0:width;
-        proxy_height = ((height<<level)>(texshrink==8)?8192:2048)?0:height;
+        proxy_width = ((width<<level)>(texshrink>=8)?8192:2048)?0:width;
+        proxy_height = ((height<<level)>(texshrink>=8)?8192:2048)?0:height;
+        proxy_intformat = swizzle_internalformat(&internalformat);
         return;
     }
     //PUSH_IF_COMPILING(glTexImage2D);
@@ -399,7 +417,7 @@ void glshim_glTexImage2D(GLenum target, GLint level, GLint internalformat,
                 bound->mipmap_need = 1;
      }
      GLenum new_format = swizzle_internalformat(&internalformat);
-     if (bound) {
+     if (bound && (level==0)) {
          bound->orig_internal = internalformat;
          bound->internalformat = new_format;
      }
@@ -821,7 +839,7 @@ void glshim_glTexSubImage2D(GLenum target, GLint level, GLint xoffset, GLint yof
 	    pixels = (GLvoid *)dst;
 	    const GLubyte *src = (GLubyte *)datab;
 	    src += glstate.texture.unpack_skip_pixels * pixelSize + glstate.texture.unpack_skip_rows * imgWidth;
-	    for (int y = 0; y < height; y += 1) {
+	    for (int y = 0; y < height; y ++) {
 		    memcpy(dst, src, width * pixelSize);
 		    src += imgWidth;
 		    dst += width * pixelSize;
@@ -1369,16 +1387,20 @@ void glshim_glGetTexLevelParameteriv(GLenum target, GLint level, GLenum pname, G
             }
 			break;
 		case GL_TEXTURE_INTERNAL_FORMAT:
-            if (bound && bound->compressed)
-                (*params) = bound->format;
-            else {
-                if(bound && ((bound->orig_internal==GL_COMPRESSED_RGB) || (bound->orig_internal==GL_COMPRESSED_RGBA))) {
-                    if(bound->orig_internal==GL_COMPRESSED_RGB)
-                        *(params) = GL_COMPRESSED_RGB_S3TC_DXT1_EXT;
-                    else
-                        *(params) = GL_COMPRESSED_RGBA_S3TC_DXT5_EXT;
-                } else
-                    (*params) = GL_RGBA;
+            if (target==GL_PROXY_TEXTURE_2D)
+				(*params) = proxy_intformat;
+			else {
+                if (bound && bound->compressed)
+                    (*params) = bound->format;
+                else {
+                    if(bound && ((bound->orig_internal==GL_COMPRESSED_RGB) || (bound->orig_internal==GL_COMPRESSED_RGBA))) {
+                        if(bound->orig_internal==GL_COMPRESSED_RGB)
+                            *(params) = GL_COMPRESSED_RGB_S3TC_DXT1_EXT;
+                        else
+                            *(params) = GL_COMPRESSED_RGBA_S3TC_DXT5_EXT;
+                    } else
+                        (*params) = GL_RGBA;
+                }
             }
 			break;
 		case GL_TEXTURE_DEPTH:
@@ -1757,7 +1779,7 @@ void glshim_glCompressedTexImage2D(GLenum target, GLint level, GLenum internalfo
 		errorShim(GL_INVALID_OPERATION);
 	    return;		// no texture bounded...
 	}
-//printf("glCompressedTexImage2D on target=%s with size(%i,%i), internalformat=s, imagesize=%i, upackbuffer=%p\n", PrintEnum(target), width, height, PrintEnum(internalformat), imageSize, glstate.buffers.unpack?glstate.buffers.unpack->data:0);
+    //printf("glCompressedTexImage2D on target=%s with size(%i,%i), internalformat=%s, imagesize=%i, upackbuffer=%p\n", PrintEnum(target), width, height, PrintEnum(internalformat), imageSize, glstate.vao->unpack?glstate.vao->unpack->data:0);
     // hack...
     if (internalformat==GL_RGBA8)
         internalformat = GL_COMPRESSED_RGBA_S3TC_DXT1_EXT;
@@ -1780,6 +1802,9 @@ void glshim_glCompressedTexImage2D(GLenum target, GLint level, GLenum internalfo
     GLvoid *datab = (GLvoid*)data;
     if (unpack)
 		datab += (uintptr_t)unpack->data;
+    
+    GLenum format = GL_RGBA;
+    GLenum type = GL_UNSIGNED_BYTE;
         
     if (isDXTc(internalformat)) {
 		GLvoid *pixels, *half;
@@ -1803,11 +1828,15 @@ void glshim_glCompressedTexImage2D(GLenum target, GLint level, GLenum internalfo
             // automaticaly reduce the pixel size
             half=pixels;
             glstate.texture.bound[glstate.texture.active]->alpha = (internalformat==GL_COMPRESSED_RGB_S3TC_DXT1_EXT)?false:true;
-            glstate.texture.bound[glstate.texture.active]->format = GL_RGBA; //internalformat;
-            glstate.texture.bound[glstate.texture.active]->type = GL_UNSIGNED_SHORT_4_4_4_4;
+            format = (internalformat==GL_COMPRESSED_RGB_S3TC_DXT1_EXT)?GL_RGB:GL_RGBA;
+            glstate.texture.bound[glstate.texture.active]->format = format; //internalformat;
+            type = (internalformat==GL_COMPRESSED_RGB_S3TC_DXT1_EXT)?GL_UNSIGNED_SHORT_5_6_5:GL_UNSIGNED_SHORT_4_4_4_4;
+            glstate.texture.bound[glstate.texture.active]->type = type;
             glstate.texture.bound[glstate.texture.active]->compressed = true;
-            if (pixel_thirdscale(pixels, &half, width, height, GL_RGBA, GL_UNSIGNED_BYTE)) 
-                fact = 1;
+            if (pixel_convert(pixels, &half, width, height, GL_RGBA, GL_UNSIGNED_BYTE, format, type, 0))
+                fact = 0;
+//            if (pixel_thirdscale(pixels, &half, width, height, GL_RGBA, GL_UNSIGNED_BYTE)) 
+//                fact = 1;
             else
                 glstate.texture.bound[glstate.texture.active]->type = GL_UNSIGNED_BYTE;
         } else {
@@ -1818,7 +1847,7 @@ void glshim_glCompressedTexImage2D(GLenum target, GLint level, GLenum internalfo
 		glshim_glGetIntegerv(GL_UNPACK_ALIGNMENT, &oldalign);
 		if (oldalign!=1) 
             glshim_glPixelStorei(GL_UNPACK_ALIGNMENT, 1);
-		glshim_glTexImage2D(target, level, GL_RGBA, width>>fact, height>>fact, border, GL_RGBA, GL_UNSIGNED_SHORT_4_4_4_4, half);
+		glshim_glTexImage2D(target, level, GL_RGBA, width>>fact, height>>fact, border, format, type, half);
 		if (oldalign!=1) 
             glshim_glPixelStorei(GL_UNPACK_ALIGNMENT, oldalign);
 		if (half!=pixels)
@@ -2028,3 +2057,12 @@ void glCopyTexSubImage3DEXT(GLenum target, GLint level, GLint xoffset, GLint yof
 //ARB mapper
 void glActiveTextureARB(GLenum texture) __attribute__((alias("glshim_glActiveTexture")));
 void glClientActiveTextureARB(GLenum texture) __attribute__((alias("glshim_glClientActiveTexture")));
+void glTexSubImage3DARB(GLenum target, GLint level, GLint xoffset, GLint yoffset, GLint zoffset,  GLsizei width, GLsizei height, GLsizei depth, GLenum format, GLenum type, const GLvoid *data) __attribute__((alias("glshim_glTexSubImage3D")));
+void glCompressedTexImage2DARB(GLenum target, GLint level, GLenum internalformat, GLsizei width, GLsizei height, GLint border, GLsizei imageSize, const GLvoid *data) __attribute__((alias("glshim_glCompressedTexImage2D")));
+void glCompressedTexImage1DARB(GLenum target, GLint level, GLenum internalformat, GLsizei width, GLint border, GLsizei imageSize, const GLvoid *data) __attribute__((alias("glshim_glCompressedTexImage1D")));
+void glCompressedTexImage3DARB(GLenum target, GLint level, GLenum internalformat, GLsizei width, GLsizei height, GLsizei depth, GLint border, GLsizei imageSize, const GLvoid *data) __attribute__((alias("glshim_glCompressedTexImage3D")));
+void glCompressedTexSubImage2DARB(GLenum target, GLint level, GLint xoffset, GLint yoffset, GLsizei width, GLsizei height, GLenum format, GLsizei imageSize, const GLvoid *data) __attribute__((alias("glshim_glCompressedTexSubImage2D")));
+void glCompressedTexSubImage1DARB(GLenum target, GLint level, GLint xoffset, GLsizei width, GLenum format, GLsizei imageSize, const GLvoid *data) __attribute__((alias("glshim_glCompressedTexSubImage1D")));
+void glCompressedTexSubImage3DARB(GLenum target, GLint level, GLint xoffset, GLint yoffset, GLint zoffset, GLsizei width, GLsizei height, GLsizei depth, GLenum format, GLsizei imageSize, const GLvoid *data) __attribute__((alias("glshim_glCompressedTexSubImage3D")));
+void glGetCompressedTexImageARB(GLenum target, GLint lod, GLvoid *img) __attribute__((alias("glshim_glGetCompressedTexImage")));
+void glCopyTexSubImage3DARB(GLenum target, GLint level, GLint xoffset, GLint yoffset, GLint zoffset, GLint x, GLint y, GLsizei width, GLsizei height) __attribute__((alias("glshim_glCopyTexSubImage3D")));
diff --git a/project/jni/glshim/src/glx/glx.c b/project/jni/glshim/src/glx/glx.c
index 7777dfde3..eef7a1fee 100755
--- a/project/jni/glshim/src/glx/glx.c
+++ b/project/jni/glshim/src/glx/glx.c
@@ -167,8 +167,9 @@ extern int texstream;
 extern int copytex;
 extern int nolumalpha;
 extern int blendhack;
+extern int export_blendcolor;
 extern int noerror;
-extern char gl_version[50];
+extern char glshim_version[50];
 
 bool g_recyclefbo = false;
 static int  g_width=0, g_height=0;
@@ -430,13 +431,14 @@ static void scan_env() {
     }
 
     env(LIBGL_BLENDHACK, blendhack, "Change Blend GL_SRC_ALPHA, GL_ONE to GL_ONE, GL_ONE");
+    env(LIBGL_BLENDCOLOR, export_blendcolor, "Export a (faked) glBlendColor");
     env(LIBGL_NOERROR, noerror, "glGetError() always return GL_NOERROR");
 
     char *env_version = getenv("LIBGL_VERSION");
     if (env_version) {
-        printf("LIBGL: Overide version string with \"%s\" (should be in the form of \"1.x\")\n", gl_version);
+        printf("LIBGL: Overide version string with \"%s\" (should be in the form of \"1.x\")\n", env_version);
     }
-    snprintf(gl_version, 49, "%s glshim wrapper", (env_version)?env_version:"1.5");
+    snprintf(glshim_version, 49, "%s glshim wrapper", (env_version)?env_version:"1.5");
     
     char cwd[1024];
     if (getcwd(cwd, sizeof(cwd))!= NULL)
diff --git a/project/jni/glshim/src/glx/lookup.c b/project/jni/glshim/src/glx/lookup.c
index 4baad17f3..be29d07f0 100755
--- a/project/jni/glshim/src/glx/lookup.c
+++ b/project/jni/glshim/src/glx/lookup.c
@@ -6,6 +6,8 @@
 #include "../gl/directstate.h"
 //#define DEBUG_ADDRESS
 
+extern int export_blendcolor;
+
 #ifdef DEBUG_ADDRESS
 #define MAP(func_name, func) \
     if(cnt==1) {if ((uint32_t)((void*)func) <0x4000000) printf("glxGetProcAddress %s = %p\n", func_name, (void*)func);} if (strcmp(name, func_name) == 0) return (void *)func;
@@ -309,8 +311,11 @@ void *glXGetProcAddressARB(const char *name) {
     _EX(glArrayElement);
     _EX(glBegin);
     _EX(glBitmap);
-    /*EXT(glBlendColor);
-    ARB(glBlendColor);*/
+    if(export_blendcolor) {
+        _EX(glBlendColor);
+        _EXT(glBlendColor);
+        _ARB(glBlendColor);
+    }
     _EXT(glBlendEquation);
     _ARB(glBlendEquation);
     _EXT(glBlendFunc);
@@ -327,6 +332,8 @@ void *glXGetProcAddressARB(const char *name) {
     _EXT(glBlendFuncSeparatei);
     _ARB(glBlendFuncSeparatei);
 #endif
+    _EX(glStencilMaskSeparate);
+    _EXT(glStencilMaskSeparate);
     _EX(glCallList);
     _EX(glCallLists);
     _EX(glClearDepth);