diff --git a/driver/tests/demo/kernel.bin b/driver/tests/demo/kernel.bin
index 2b8280da..aea7c061 100755
Binary files a/driver/tests/demo/kernel.bin and b/driver/tests/demo/kernel.bin differ
diff --git a/driver/tests/demo/kernel.elf b/driver/tests/demo/kernel.elf
index 852f887b..d91b509c 100755
Binary files a/driver/tests/demo/kernel.elf and b/driver/tests/demo/kernel.elf differ
diff --git a/runtime/include/vx_spawn.h b/runtime/include/vx_spawn.h
index 5acbfecc..d2f2cfd4 100644
--- a/runtime/include/vx_spawn.h
+++ b/runtime/include/vx_spawn.h
@@ -12,25 +12,6 @@ typedef void (*func_t)(void *);
 
 void vx_spawn_warps(int num_warps, int num_threads, func_t func_ptr , void * args);
 
-struct context_t {
-  uint32_t num_groups[3];
-  uint32_t global_offset[3];
-  uint32_t local_size[3];  
-  char * printf_buffer;
-  uint32_t *printf_buffer_position;
-  uint32_t printf_buffer_capacity;  
-  uint32_t work_dim;
-};
-
-/* The default work-group function prototype as generated by Workgroup.cc. */
-typedef void (*vx_pocl_workgroup_func) (const void * /* args */,
-				     const struct context_t * /* context */,
-				     uint32_t /* group_x */,
-				     uint32_t /* group_y */,
-				     uint32_t /* group_z */);
-
-void pocl_spawn(struct context_t * ctx, vx_pocl_workgroup_func pfn, const void * args);
-
 #ifdef __cplusplus
 }
 #endif
diff --git a/runtime/src/vx_spawn.c b/runtime/src/vx_spawn.c
index 3461da34..a215d1b3 100644
--- a/runtime/src/vx_spawn.c
+++ b/runtime/src/vx_spawn.c
@@ -6,16 +6,20 @@
 extern "C" {
 #endif
 
-func_t global_function_pointer;
-void * global_argument_struct;
-int global_num_threads;
+typedef struct {
+	func_t function;
+	void * arguments;
+	int    nthreads;
+} spawn_t;
+
+spawn_t* g_spawn = NULL;
 
 void spawn_warp_runonce() {
 	// active all threads
-	vx_tmc(global_num_threads);
+	vx_tmc(g_spawn->nthreads);
 
 	// call user routine
-	global_function_pointer(global_argument_struct);
+	g_spawn->function(g_spawn->arguments);
 
 	// resume single-thread execution on exit
 	int wid = vx_warp_id();
@@ -23,55 +27,16 @@ void spawn_warp_runonce() {
 	vx_tmc(tmask);
 }
 
-void vx_spawn_warps(int numWarps, int numThreads, func_t func_ptr, void * args) {
-	global_function_pointer = func_ptr;
-	global_argument_struct  = args;
-	global_num_threads      = numThreads;
-	if (numWarps > 1) {
-		vx_wspawn(numWarps, (unsigned)spawn_warp_runonce);
+void vx_spawn_warps(int num_warps, int num_threads, func_t func_ptr , void * args) {
+	spawn_t spawn = { func_ptr, args, num_threads };
+	g_spawn = &spawn;
+
+	if (num_warps > 1) {
+		vx_wspawn(num_warps, (unsigned)spawn_warp_runonce);
 	}
 	spawn_warp_runonce();
 }
 
-int                    pocl_threads;
-struct context_t *     pocl_ctx;
-vx_pocl_workgroup_func pocl_pfn;
-const void *           pocl_args;
-
-void pocl_spawn_warp_runonce() {
-	// active all threads
-	vx_tmc(pocl_threads);
-
-	int x = vx_thread_id();
-	int y = vx_warp_gid();
-
-	// call kernel routine
-	(pocl_pfn)(pocl_args, pocl_ctx, x, y, 0);
-
-	// resume single-thread execution on exit
-	int wid = vx_warp_id();
-	unsigned tmask = (0 == wid) ? 0x1 : 0x0;
-	vx_tmc(tmask);
-}
-
-void pocl_spawn(struct context_t * ctx, vx_pocl_workgroup_func pfn, const void * args) {
-	if (ctx->num_groups[2] > 1)	{
-		printf("ERROR: pocl_spawn doesn't support Z dimension yet!\n");
-		return;
-	}
-
-	pocl_threads = ctx->num_groups[0];
-	pocl_ctx     = ctx;
-	pocl_pfn     = pfn;
-	pocl_args    = args;
-
-	if (ctx->num_groups[1] > 1)	{
-		vx_wspawn(ctx->num_groups[1], (unsigned)&pocl_spawn_warp_runonce);
-	}
-
-	pocl_spawn_warp_runonce();
-}
-
 #ifdef __cplusplus
 }
 #endif
\ No newline at end of file