diff --git a/drm/nouveau/nvkm/engine/gr/fuc/gpc.fuc b/drm/nouveau/nvkm/engine/gr/fuc/gpc.fuc
index dc60509..4984b00 100644
--- a/drm/nouveau/nvkm/engine/gr/fuc/gpc.fuc
+++ b/drm/nouveau/nvkm/engine/gr/fuc/gpc.fuc
@@ -291,12 +291,13 @@ init:
 // Main program loop, very simple, sleeps until woken up by the interrupt
 // handler, pulls a command from the queue and executes its handler
 //
-main:
-	bset $flags $p0
+wait:
 	sleep $p0
+	bset $flags $p0
+main:
 	mov $r13 #cmd_queue
 	call(queue_get)
-	bra $p1 #main
+	bra $p1 #wait
 
 	// 0x0000-0x0003 are all context transfers
 	cmpu b32 $r14 0x04
diff --git a/drm/nouveau/nvkm/engine/gr/fuc/gpcgf100.fuc3.h b/drm/nouveau/nvkm/engine/gr/fuc/gpcgf100.fuc3.h
index 5f4ddfe..8cb240b 100644
--- a/drm/nouveau/nvkm/engine/gr/fuc/gpcgf100.fuc3.h
+++ b/drm/nouveau/nvkm/engine/gr/fuc/gpcgf100.fuc3.h
@@ -370,9 +370,10 @@ uint32_t gf100_grgpc_code[] = {
 	0xf11f29f0,
 	0xf0080007,
 	0x02d00203,
-/* 0x04bb: main */
+/* 0x04bb: wait */
 	0xf404bd00,
-	0x28f40031,
+	0x31f40028,
+/* 0x04c1: main */
 	0x1cd7f000,
 	0xf43921f4,
 	0xe4b0f401,
@@ -384,10 +385,10 @@ uint32_t gf100_grgpc_code[] = {
 	0x0018fe05,
 	0x05b421f5,
 /* 0x04eb: main_not_ctx_xfer */
-	0x94d30ef4,
+	0x94d90ef4,
 	0xf5f010ef,
 	0x7e21f501,
-	0xc60ef403,
+	0xcc0ef403,
 /* 0x04f8: ih */
 	0x80f900f9,
 	0xf90188fe,
diff --git a/drm/nouveau/nvkm/engine/gr/fuc/gpcgf117.fuc3.h b/drm/nouveau/nvkm/engine/gr/fuc/gpcgf117.fuc3.h
index 03381b1..550d6ba 100644
--- a/drm/nouveau/nvkm/engine/gr/fuc/gpcgf117.fuc3.h
+++ b/drm/nouveau/nvkm/engine/gr/fuc/gpcgf117.fuc3.h
@@ -397,9 +397,10 @@ uint32_t gf117_grgpc_code[] = {
 	0x080007f1,
 	0xd00203f0,
 	0x04bd0002,
-/* 0x0508: main */
-	0xf40031f4,
-	0xd7f00028,
+/* 0x0508: wait */
+	0xf40028f4,
+/* 0x050e: main */
+	0xd7f00031,
 	0x3921f424,
 	0xb0f401f4,
 	0x18f404e4,
@@ -409,13 +410,13 @@ uint32_t gf117_grgpc_code[] = {
 	0xfd01e4b6,
 	0x18fe051e,
 	0x0121f500,
-	0xd30ef406,
+	0xd90ef406,
 /* 0x0538: main_not_ctx_xfer */
 	0xf010ef94,
 	0x21f501f5,
 	0x0ef4037e,
 /* 0x0545: ih */
-	0xf900f9c6,
+	0xf900f9cc,
 	0x0188fe80,
 	0x90f980f9,
 	0xb0f9a0f9,
diff --git a/drm/nouveau/nvkm/engine/gr/fuc/gpcgk104.fuc3.h b/drm/nouveau/nvkm/engine/gr/fuc/gpcgk104.fuc3.h
index 99d9b48..271b59d 100644
--- a/drm/nouveau/nvkm/engine/gr/fuc/gpcgk104.fuc3.h
+++ b/drm/nouveau/nvkm/engine/gr/fuc/gpcgk104.fuc3.h
@@ -397,9 +397,10 @@ uint32_t gk104_grgpc_code[] = {
 	0x080007f1,
 	0xd00203f0,
 	0x04bd0002,
-/* 0x0508: main */
-	0xf40031f4,
-	0xd7f00028,
+/* 0x0508: wait */
+	0xf40028f4,
+/* 0x050e: main */
+	0xd7f00031,
 	0x3921f424,
 	0xb0f401f4,
 	0x18f404e4,
@@ -409,13 +410,13 @@ uint32_t gk104_grgpc_code[] = {
 	0xfd01e4b6,
 	0x18fe051e,
 	0x0121f500,
-	0xd30ef406,
+	0xd90ef406,
 /* 0x0538: main_not_ctx_xfer */
 	0xf010ef94,
 	0x21f501f5,
 	0x0ef4037e,
 /* 0x0545: ih */
-	0xf900f9c6,
+	0xf900f9cc,
 	0x0188fe80,
 	0x90f980f9,
 	0xb0f9a0f9,
diff --git a/drm/nouveau/nvkm/engine/gr/fuc/gpcgk110.fuc3.h b/drm/nouveau/nvkm/engine/gr/fuc/gpcgk110.fuc3.h
index f726769..73b4a32 100644
--- a/drm/nouveau/nvkm/engine/gr/fuc/gpcgk110.fuc3.h
+++ b/drm/nouveau/nvkm/engine/gr/fuc/gpcgk110.fuc3.h
@@ -397,9 +397,10 @@ uint32_t gk110_grgpc_code[] = {
 	0x300007f1,
 	0xd00203f0,
 	0x04bd0002,
-/* 0x0508: main */
-	0xf40031f4,
-	0xd7f00028,
+/* 0x0508: wait */
+	0xf40028f4,
+/* 0x050e: main */
+	0xd7f00031,
 	0x3921f424,
 	0xb0f401f4,
 	0x18f404e4,
@@ -409,13 +410,13 @@ uint32_t gk110_grgpc_code[] = {
 	0xfd01e4b6,
 	0x18fe051e,
 	0x0121f500,
-	0xd30ef406,
+	0xd90ef406,
 /* 0x0538: main_not_ctx_xfer */
 	0xf010ef94,
 	0x21f501f5,
 	0x0ef4037e,
 /* 0x0545: ih */
-	0xf900f9c6,
+	0xf900f9cc,
 	0x0188fe80,
 	0x90f980f9,
 	0xb0f9a0f9,
diff --git a/drm/nouveau/nvkm/engine/gr/fuc/gpcgk208.fuc5.h b/drm/nouveau/nvkm/engine/gr/fuc/gpcgk208.fuc5.h
index 387d1fa..0181698 100644
--- a/drm/nouveau/nvkm/engine/gr/fuc/gpcgk208.fuc5.h
+++ b/drm/nouveau/nvkm/engine/gr/fuc/gpcgk208.fuc5.h
@@ -349,9 +349,10 @@ uint32_t gk208_grgpc_code[] = {
 	0x801f29f0,
 	0xf6023000,
 	0x04bd0002,
-/* 0x0448: main */
-	0xf40031f4,
-	0x240d0028,
+/* 0x0448: wait */
+	0xf40028f4,
+/* 0x044e: main */
+	0x240d0031,
 	0x0000377e,
 	0xb0f401f4,
 	0x18f404e4,
@@ -362,10 +363,10 @@ uint32_t gk208_grgpc_code[] = {
 	0x0018fe05,
 	0x00051f7e,
 /* 0x0477: main_not_ctx_xfer */
-	0x94d40ef4,
+	0x94da0ef4,
 	0xf5f010ef,
 	0x02f87e01,
-	0xc70ef400,
+	0xcd0ef400,
 /* 0x0484: ih */
 	0x80f900f9,
 	0xf90188fe,
diff --git a/drm/nouveau/nvkm/engine/gr/fuc/gpcgm107.fuc5.h b/drm/nouveau/nvkm/engine/gr/fuc/gpcgm107.fuc5.h
index fa9f3c0..eca007f 100644
--- a/drm/nouveau/nvkm/engine/gr/fuc/gpcgm107.fuc5.h
+++ b/drm/nouveau/nvkm/engine/gr/fuc/gpcgm107.fuc5.h
@@ -427,9 +427,10 @@ uint32_t gm107_grgpc_code[] = {
 	0x1f29f024,
 	0x02300080,
 	0xbd0002f6,
-/* 0x0571: main */
-	0x0031f404,
-	0x0d0028f4,
+/* 0x0571: wait */
+	0x0028f404,
+/* 0x0577: main */
+	0x0d0031f4,
 	0x00377e24,
 	0xf401f400,
 	0xf404e4b0,
@@ -439,13 +440,13 @@ uint32_t gm107_grgpc_code[] = {
 	0xfd01e4b6,
 	0x18fe051e,
 	0x06487e00,
-	0xd40ef400,
+	0xda0ef400,
 /* 0x05a0: main_not_ctx_xfer */
 	0xf010ef94,
 	0xf87e01f5,
 	0x0ef40002,
 /* 0x05ad: ih */
-	0xf900f9c7,
+	0xf900f9cd,
 	0x0188fe80,
 	0x90f980f9,
 	0xb0f9a0f9,
diff --git a/drm/nouveau/nvkm/engine/gr/fuc/hub.fuc b/drm/nouveau/nvkm/engine/gr/fuc/hub.fuc
index e3a2fb3..4d416d4 100644
--- a/drm/nouveau/nvkm/engine/gr/fuc/hub.fuc
+++ b/drm/nouveau/nvkm/engine/gr/fuc/hub.fuc
@@ -218,13 +218,14 @@ init:
 // Main program loop, very simple, sleeps until woken up by the interrupt
 // handler, pulls a command from the queue and executes its handler
 //
-main:
+wait:
 	// sleep until we have something to do
-	bset $flags $p0
 	sleep $p0
+	bset $flags $p0
+main:
 	mov $r13 #cmd_queue
 	call(queue_get)
-	bra $p1 #main
+	bra $p1 #wait
 
 	// context switch, requested by GPU?
 	cmpu b32 $r14 0x4001
diff --git a/drm/nouveau/nvkm/engine/gr/fuc/hubgf100.fuc3.h b/drm/nouveau/nvkm/engine/gr/fuc/hubgf100.fuc3.h
index 397921a..8015b40 100644
--- a/drm/nouveau/nvkm/engine/gr/fuc/hubgf100.fuc3.h
+++ b/drm/nouveau/nvkm/engine/gr/fuc/hubgf100.fuc3.h
@@ -584,9 +584,10 @@ uint32_t gf100_grhub_code[] = {
 	0x080007f1,
 	0xd00203f0,
 	0x04bd0001,
-/* 0x0564: main */
-	0xf40031f4,
-	0xd7f00028,
+/* 0x0564: wait */
+	0xf40028f4,
+/* 0x056a: main */
+	0xd7f00031,
 	0x3921f410,
 	0xb1f401f4,
 	0xf54001e4,
@@ -650,7 +651,7 @@ uint32_t gf100_grhub_code[] = {
 	0x170007f1,
 	0xd00203f0,
 	0x04bd0009,
-	0xff080ef5,
+	0xff0e0ef5,
 /* 0x0660: main_not_ctx_switch */
 	0xf401e4b0,
 	0xf2b90d1b,
@@ -675,12 +676,12 @@ uint32_t gf100_grhub_code[] = {
 	0xf501f5f0,
 	0xf5037e21,
 /* 0x06b3: main_done */
-	0xbdfeb50e,
+	0xbdfebb0e,
 	0x1f29f024,
 	0x080007f1,
 	0xd00203f0,
 	0x04bd0002,
-	0xfea00ef5,
+	0xfea60ef5,
 /* 0x06c8: ih */
 	0x80f900f9,
 	0xf90188fe,
diff --git a/drm/nouveau/nvkm/engine/gr/fuc/hubgf117.fuc3.h b/drm/nouveau/nvkm/engine/gr/fuc/hubgf117.fuc3.h
index 50c9716..2af90ec 100644
--- a/drm/nouveau/nvkm/engine/gr/fuc/hubgf117.fuc3.h
+++ b/drm/nouveau/nvkm/engine/gr/fuc/hubgf117.fuc3.h
@@ -584,9 +584,10 @@ uint32_t gf117_grhub_code[] = {
 	0x080007f1,
 	0xd00203f0,
 	0x04bd0001,
-/* 0x0564: main */
-	0xf40031f4,
-	0xd7f00028,
+/* 0x0564: wait */
+	0xf40028f4,
+/* 0x056a: main */
+	0xd7f00031,
 	0x3921f410,
 	0xb1f401f4,
 	0xf54001e4,
@@ -650,7 +651,7 @@ uint32_t gf117_grhub_code[] = {
 	0x170007f1,
 	0xd00203f0,
 	0x04bd0009,
-	0xff080ef5,
+	0xff0e0ef5,
 /* 0x0660: main_not_ctx_switch */
 	0xf401e4b0,
 	0xf2b90d1b,
@@ -675,12 +676,12 @@ uint32_t gf117_grhub_code[] = {
 	0xf501f5f0,
 	0xf5037e21,
 /* 0x06b3: main_done */
-	0xbdfeb50e,
+	0xbdfebb0e,
 	0x1f29f024,
 	0x080007f1,
 	0xd00203f0,
 	0x04bd0002,
-	0xfea00ef5,
+	0xfea60ef5,
 /* 0x06c8: ih */
 	0x80f900f9,
 	0xf90188fe,
diff --git a/drm/nouveau/nvkm/engine/gr/fuc/hubgk104.fuc3.h b/drm/nouveau/nvkm/engine/gr/fuc/hubgk104.fuc3.h
index 125824b..e8b8c1c 100644
--- a/drm/nouveau/nvkm/engine/gr/fuc/hubgk104.fuc3.h
+++ b/drm/nouveau/nvkm/engine/gr/fuc/hubgk104.fuc3.h
@@ -584,9 +584,10 @@ uint32_t gk104_grhub_code[] = {
 	0x080007f1,
 	0xd00203f0,
 	0x04bd0001,
-/* 0x0564: main */
-	0xf40031f4,
-	0xd7f00028,
+/* 0x0564: wait */
+	0xf40028f4,
+/* 0x056a: main */
+	0xd7f00031,
 	0x3921f410,
 	0xb1f401f4,
 	0xf54001e4,
@@ -650,7 +651,7 @@ uint32_t gk104_grhub_code[] = {
 	0x170007f1,
 	0xd00203f0,
 	0x04bd0009,
-	0xff080ef5,
+	0xff0e0ef5,
 /* 0x0660: main_not_ctx_switch */
 	0xf401e4b0,
 	0xf2b90d1b,
@@ -675,12 +676,12 @@ uint32_t gk104_grhub_code[] = {
 	0xf501f5f0,
 	0xf5037e21,
 /* 0x06b3: main_done */
-	0xbdfeb50e,
+	0xbdfebb0e,
 	0x1f29f024,
 	0x080007f1,
 	0xd00203f0,
 	0x04bd0002,
-	0xfea00ef5,
+	0xfea60ef5,
 /* 0x06c8: ih */
 	0x80f900f9,
 	0xf90188fe,
diff --git a/drm/nouveau/nvkm/engine/gr/fuc/hubgk110.fuc3.h b/drm/nouveau/nvkm/engine/gr/fuc/hubgk110.fuc3.h
index 0a1b8c0..f4ed2fb 100644
--- a/drm/nouveau/nvkm/engine/gr/fuc/hubgk110.fuc3.h
+++ b/drm/nouveau/nvkm/engine/gr/fuc/hubgk110.fuc3.h
@@ -584,9 +584,10 @@ uint32_t gk110_grhub_code[] = {
 	0x300007f1,
 	0xd00203f0,
 	0x04bd0001,
-/* 0x0564: main */
-	0xf40031f4,
-	0xd7f00028,
+/* 0x0564: wait */
+	0xf40028f4,
+/* 0x056a: main */
+	0xd7f00031,
 	0x3921f410,
 	0xb1f401f4,
 	0xf54001e4,
@@ -650,7 +651,7 @@ uint32_t gk110_grhub_code[] = {
 	0x170007f1,
 	0xd00203f0,
 	0x04bd0009,
-	0xff080ef5,
+	0xff0e0ef5,
 /* 0x0660: main_not_ctx_switch */
 	0xf401e4b0,
 	0xf2b90d1b,
@@ -675,12 +676,12 @@ uint32_t gk110_grhub_code[] = {
 	0xf501f5f0,
 	0xf5037e21,
 /* 0x06b3: main_done */
-	0xbdfeb50e,
+	0xbdfebb0e,
 	0x1f29f024,
 	0x300007f1,
 	0xd00203f0,
 	0x04bd0002,
-	0xfea00ef5,
+	0xfea60ef5,
 /* 0x06c8: ih */
 	0x80f900f9,
 	0xf90188fe,
diff --git a/drm/nouveau/nvkm/engine/gr/fuc/hubgk208.fuc5.h b/drm/nouveau/nvkm/engine/gr/fuc/hubgk208.fuc5.h
index 16869d0..ed48897 100644
--- a/drm/nouveau/nvkm/engine/gr/fuc/hubgk208.fuc5.h
+++ b/drm/nouveau/nvkm/engine/gr/fuc/hubgk208.fuc5.h
@@ -531,9 +531,10 @@ uint32_t gk208_grhub_code[] = {
 	0x1f19f014,
 	0x02300080,
 	0xbd0001f6,
-/* 0x0491: main */
-	0x0031f404,
-	0x0d0028f4,
+/* 0x0491: wait */
+	0x0028f404,
+/* 0x0497: main */
+	0x0d0031f4,
 	0x00377e10,
 	0xf401f400,
 	0x4001e4b1,
@@ -590,7 +591,7 @@ uint32_t gk208_grhub_code[] = {
 	0x09f60217,
 	0xf504bd00,
 /* 0x056b: main_not_ctx_switch */
-	0xb0ff2a0e,
+	0xb0ff300e,
 	0x1bf401e4,
 	0x7ef2b20c,
 	0xf4000820,
@@ -612,11 +613,11 @@ uint32_t gk208_grhub_code[] = {
 	0x7e01f5f0,
 	0xf50002f8,
 /* 0x05b7: main_done */
-	0xbdfede0e,
+	0xbdfee40e,
 	0x1f29f024,
 	0x02300080,
 	0xbd0002f6,
-	0xcc0ef504,
+	0xd20ef504,
 /* 0x05c9: ih */
 	0xf900f9fe,
 	0x0188fe80,
diff --git a/drm/nouveau/nvkm/engine/gr/fuc/hubgm107.fuc5.h b/drm/nouveau/nvkm/engine/gr/fuc/hubgm107.fuc5.h
index d6343d2..5c90518 100644
--- a/drm/nouveau/nvkm/engine/gr/fuc/hubgm107.fuc5.h
+++ b/drm/nouveau/nvkm/engine/gr/fuc/hubgm107.fuc5.h
@@ -531,9 +531,10 @@ uint32_t gm107_grhub_code[] = {
 	0x1f19f014,
 	0x02300080,
 	0xbd0001f6,
-/* 0x0491: main */
-	0x0031f404,
-	0x0d0028f4,
+/* 0x0491: wait */
+	0x0028f404,
+/* 0x0497: main */
+	0x0d0031f4,
 	0x00377e10,
 	0xf401f400,
 	0x4001e4b1,
@@ -590,7 +591,7 @@ uint32_t gm107_grhub_code[] = {
 	0x09f60217,
 	0xf504bd00,
 /* 0x056b: main_not_ctx_switch */
-	0xb0ff2a0e,
+	0xb0ff300e,
 	0x1bf401e4,
 	0x7ef2b20c,
 	0xf4000820,
@@ -612,11 +613,11 @@ uint32_t gm107_grhub_code[] = {
 	0x7e01f5f0,
 	0xf50002f8,
 /* 0x05b7: main_done */
-	0xbdfede0e,
+	0xbdfee40e,
 	0x1f29f024,
 	0x02300080,
 	0xbd0002f6,
-	0xcc0ef504,
+	0xd20ef504,
 /* 0x05c9: ih */
 	0xf900f9fe,
 	0x0188fe80,