diff --git a/src/vid_voodoo_banshee.c b/src/vid_voodoo_banshee.c
index 15ac43e..769d9de 100644
--- a/src/vid_voodoo_banshee.c
+++ b/src/vid_voodoo_banshee.c
@@ -1163,6 +1163,8 @@ static void banshee_cmd_write(banshee_t *banshee, uint32_t addr, uint32_t val)
                 voodoo->cmdfifo_size = val;
                 voodoo->cmdfifo_end = voodoo->cmdfifo_base + (((voodoo->cmdfifo_size & 0xff) + 1) << 12);
                 voodoo->cmdfifo_enabled = val & 0x100;
+                if (!voodoo->cmdfifo_enabled)
+                        voodoo->cmdfifo_in_sub = 0; /*Not sure exactly when this should be reset*/
 //                pclog("cmdfifo_base=%08x  cmdfifo_end=%08x\n", voodoo->cmdfifo_base, voodoo->cmdfifo_end);
                 break;
                 
diff --git a/src/vid_voodoo_common.h b/src/vid_voodoo_common.h
index cf62410..d8ca533 100644
--- a/src/vid_voodoo_common.h
+++ b/src/vid_voodoo_common.h
@@ -301,7 +301,8 @@ typedef struct voodoo_t
         volatile int params_read_idx[4], params_write_idx;
 
         uint32_t cmdfifo_base, cmdfifo_end, cmdfifo_size;
-        int cmdfifo_rp;
+        int cmdfifo_rp, cmdfifo_ret_addr;
+        int cmdfifo_in_sub;
         volatile int cmdfifo_depth_rd, cmdfifo_depth_wr;
         volatile int cmdfifo_enabled;
         uint32_t cmdfifo_amin, cmdfifo_amax;
diff --git a/src/vid_voodoo_fifo.c b/src/vid_voodoo_fifo.c
index 472df3f..59861f3 100644
--- a/src/vid_voodoo_fifo.c
+++ b/src/vid_voodoo_fifo.c
@@ -113,15 +113,19 @@ static uint32_t cmdfifo_get(voodoo_t *voodoo)
 {
         uint32_t val;
 
-        while (voodoo->cmdfifo_depth_rd == voodoo->cmdfifo_depth_wr)
+        if (!voodoo->cmdfifo_in_sub)
         {
-                thread_wait_event(voodoo->wake_fifo_thread, -1);
-                thread_reset_event(voodoo->wake_fifo_thread);
+                while (voodoo->cmdfifo_depth_rd == voodoo->cmdfifo_depth_wr)
+                {
+                        thread_wait_event(voodoo->wake_fifo_thread, -1);
+                        thread_reset_event(voodoo->wake_fifo_thread);
+                }
         }
 
         val = *(uint32_t *)&voodoo->fb_mem[voodoo->cmdfifo_rp & voodoo->fb_mask];
 
-        voodoo->cmdfifo_depth_rd++;
+        if (!voodoo->cmdfifo_in_sub)
+                voodoo->cmdfifo_depth_rd++;
         voodoo->cmdfifo_rp += 4;
 
 //        pclog("  CMDFIFO get %08x\n", val);
@@ -242,7 +246,7 @@ void voodoo_fifo_thread(void *param)
                         voodoo->time += end_time - start_time;
                 }
 
-                while (voodoo->cmdfifo_enabled && voodoo->cmdfifo_depth_rd != voodoo->cmdfifo_depth_wr)
+                while (voodoo->cmdfifo_enabled && (voodoo->cmdfifo_depth_rd != voodoo->cmdfifo_depth_wr || voodoo->cmdfifo_in_sub))
                 {
                         uint64_t start_time = timer_read();
                         uint64_t end_time;
@@ -265,6 +269,18 @@ void voodoo_fifo_thread(void *param)
                                         case 0: /*NOP*/
                                         break;
 
+                                        case 1: /*JSR*/
+//                                        pclog("JSR %08x\n", (header >> 4) & 0xfffffc);
+                                        voodoo->cmdfifo_ret_addr = voodoo->cmdfifo_rp;
+                                        voodoo->cmdfifo_rp = (header >> 4) & 0xfffffc;
+                                        voodoo->cmdfifo_in_sub = 1;
+                                        break;
+
+                                        case 2: /*RET*/
+                                        voodoo->cmdfifo_rp = voodoo->cmdfifo_ret_addr;
+                                        voodoo->cmdfifo_in_sub = 0;
+                                        break;
+
                                         case 3: /*JMP local frame buffer*/
                                         voodoo->cmdfifo_rp = (header >> 4) & 0xfffffc;
 //                                        pclog("JMP to %08x %04x\n", voodoo->cmdfifo_rp, header);