diff --git a/src/i965_render.c b/src/i965_render.c
index 96082bb..d6e87a7 100644
--- a/src/i965_render.c
+++ b/src/i965_render.c
@@ -488,16 +488,16 @@ i965_prepare_composite(int op, PicturePtr pSrcPicture,
     vs_offset = ALIGN(next_offset, 64);
     next_offset = vs_offset + sizeof(*vs_state);
 
-    sf_offset = ALIGN(next_offset, 32);
+    sf_offset = ALIGN(next_offset, 64);
     next_offset = sf_offset + sizeof(*sf_state);
 
-    wm_offset = ALIGN(next_offset, 32);
+    wm_offset = ALIGN(next_offset, 64);
     next_offset = wm_offset + sizeof(*wm_state);
 
     wm_scratch_offset = ALIGN(next_offset, 1024);
     next_offset = wm_scratch_offset + PS_SCRATCH_SPACE * PS_MAX_THREADS;
 
-    cc_offset = ALIGN(next_offset, 32);
+    cc_offset = ALIGN(next_offset, 64);
     next_offset = cc_offset + sizeof(*cc_state);
 
     /* keep current sf_kernel, which will send one setup urb entry to
@@ -541,15 +541,15 @@ i965_prepare_composite(int op, PicturePtr pSrcPicture,
     next_offset = sip_kernel_offset + sizeof (sip_kernel_static);
 
     /* needed? */
-    cc_viewport_offset = ALIGN(next_offset, 32);
+    cc_viewport_offset = ALIGN(next_offset, 64);
     next_offset = cc_viewport_offset + sizeof(*cc_viewport);
 
     /* for texture sampler */
-    src_sampler_offset = ALIGN(next_offset, 32);
+    src_sampler_offset = ALIGN(next_offset, 64);
     next_offset = src_sampler_offset + sizeof(*src_sampler_state);
 
     if (pMask) {
-   	mask_sampler_offset = ALIGN(next_offset, 32);
+   	mask_sampler_offset = ALIGN(next_offset, 64);
    	next_offset = mask_sampler_offset + sizeof(*mask_sampler_state);
     }
     /* Align VB to native size of elements, for safety */