drm/nve0/graph: fix fuc, and enable acceleration on all known chipsets

Also adds GK106 to chipsets known by ucode.
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>

drivers/gpu/drm/nouveau/core/engine/graph/fuc/gpcnve0.fuc

@@ -57,6 +57,11 @@ chipsets:
 .b16 #nve4_gpc_mmio_tail
 .b16 #nve4_tpc_mmio_head
 .b16 #nve4_tpc_mmio_tail
+.b8  0xe6 0 0 0
+.b16 #nve4_gpc_mmio_head
+.b16 #nve4_gpc_mmio_tail
+.b16 #nve4_tpc_mmio_head
+.b16 #nve4_tpc_mmio_tail
 .b8  0 0 0 0
 
 // GPC mmio lists

drivers/gpu/drm/nouveau/core/engine/graph/fuc/gpcnve0.fuc.h

@@ -34,13 +34,16 @@ uint32_t nve0_grgpc_data[] = {
 	0x00000000,
 /* 0x0064: chipsets */
 	0x000000e4,
-	0x01040080,
-	0x014c0104,
+	0x0110008c,
+	0x01580110,
 	0x000000e7,
-	0x01040080,
-	0x014c0104,
+	0x0110008c,
+	0x01580110,
+	0x000000e6,
+	0x0110008c,
+	0x01580110,
 	0x00000000,
-/* 0x0080: nve4_gpc_mmio_head */
+/* 0x008c: nve4_gpc_mmio_head */
 	0x00000380,
 	0x04000400,
 	0x0800040c,
@@ -74,8 +77,8 @@ uint32_t nve0_grgpc_data[] = {
 	0x14003100,
 	0x000031d0,
 	0x040031e0,
-/* 0x0104: nve4_gpc_mmio_tail */
-/* 0x0104: nve4_tpc_mmio_head */
+/* 0x0110: nve4_gpc_mmio_tail */
+/* 0x0110: nve4_tpc_mmio_head */
 	0x00000048,
 	0x00000064,
 	0x00000088,

drivers/gpu/drm/nouveau/core/engine/graph/fuc/hubnve0.fuc

@@ -44,6 +44,9 @@ chipsets:
 .b8  0xe7 0 0 0
 .b16 #nve4_hub_mmio_head
 .b16 #nve4_hub_mmio_tail
+.b8  0xe6 0 0 0
+.b16 #nve4_hub_mmio_head
+.b16 #nve4_hub_mmio_tail
 .b8  0 0 0 0
 
 nve4_hub_mmio_head:
@@ -680,6 +683,16 @@ ctx_mmio_exec:
 //		on load it means: "a save preceeded this load"
 //
 ctx_xfer:
+	// according to mwk, some kind of wait for idle
+	mov $r15 0xc00
+	shl b32 $r15 6
+	mov $r14 4
+	iowr I[$r15 + 0x200] $r14
+	ctx_xfer_idle:
+		iord $r14 I[$r15 + 0x000]
+		and $r14 0x2000
+		bra ne #ctx_xfer_idle
+
 	bra not $p1 #ctx_xfer_pre
 	bra $p2 #ctx_xfer_pre_load
 	ctx_xfer_pre:

drivers/gpu/drm/nouveau/core/engine/graph/fuc/hubnve0.fuc.h

@@ -30,11 +30,13 @@ uint32_t nve0_grhub_data[] = {
 	0x00000000,
 /* 0x005c: chipsets */
 	0x000000e4,
-	0x013c0070,
+	0x01440078,
 	0x000000e7,
-	0x013c0070,
+	0x01440078,
+	0x000000e6,
+	0x01440078,
 	0x00000000,
-/* 0x0070: nve4_hub_mmio_head */
+/* 0x0078: nve4_hub_mmio_head */
 	0x0417e91c,
 	0x04400204,
 	0x18404010,
@@ -86,9 +88,7 @@ uint32_t nve0_grhub_data[] = {
 	0x00408840,
 	0x08408900,
 	0x00408980,
-/* 0x013c: nve4_hub_mmio_tail */
-	0x00000000,
-	0x00000000,
+/* 0x0144: nve4_hub_mmio_tail */
 	0x00000000,
 	0x00000000,
 	0x00000000,
@@ -781,77 +781,78 @@ uint32_t nve0_grhub_code[] = {
 	0x0613f002,
 	0xf80601fa,
 /* 0x07fb: ctx_xfer */
-	0xf400f803,
-	0x02f40611,
-/* 0x0801: ctx_xfer_pre */
-	0x10f7f00d,
-	0x067221f5,
-/* 0x080b: ctx_xfer_pre_load */
-	0xf01c11f4,
-	0x21f502f7,
-	0x21f50631,
-	0x21f50640,
-	0xf4bd0652,
-	0x063121f5,
-	0x069221f5,
-/* 0x0824: ctx_xfer_exec */
-	0xf1160198,
-	0xb6041427,
-	0x20d00624,
-	0x00e7f100,
-	0x41e3f0a5,
-	0xf4021fb9,
-	0xe0b68d21,
-	0x01fcf004,
-	0xb6022cf0,
-	0xf2fd0124,
-	0x8d21f405,
-	0x4afc17f1,
-	0xf00213f0,
-	0x12d00c27,
-	0x0721f500,
-	0xfc27f102,
-	0x0223f047,
-	0xf00020d0,
-	0x20b6012c,
-	0x0012d003,
-	0xf001acf0,
-	0xb7f006a5,
-	0x140c9800,
-	0xf0150d98,
-	0x21f500e7,
-	0xa7f0015c,
-	0x0321f508,
-	0x0721f501,
-	0x2201f402,
-	0xf40ca7f0,
-	0x17f1c921,
-	0x14b60a10,
-	0x0527f006,
-/* 0x08ab: ctx_xfer_post_save_wait */
-	0xcf0012d0,
-	0x22fd0012,
-	0xfa1bf405,
-/* 0x08b7: ctx_xfer_post */
-	0xf02e02f4,
-	0x21f502f7,
-	0xf4bd0631,
-	0x067221f5,
-	0x022621f5,
-	0x064021f5,
-	0x21f5f4bd,
-	0x11f40631,
-	0x80019810,
-	0xf40511fd,
-	0x21f5070b,
-/* 0x08e2: ctx_xfer_no_post_mmio */
-/* 0x08e2: ctx_xfer_done */
-	0x00f807b1,
-	0x00000000,
-	0x00000000,
-	0x00000000,
-	0x00000000,
-	0x00000000,
-	0x00000000,
+	0xf100f803,
+	0xb60c00f7,
+	0xe7f006f4,
+	0x80fed004,
+/* 0x0808: ctx_xfer_idle */
+	0xf100fecf,
+	0xf42000e4,
+	0x11f4f91b,
+	0x0d02f406,
+/* 0x0818: ctx_xfer_pre */
+	0xf510f7f0,
+	0xf4067221,
+/* 0x0822: ctx_xfer_pre_load */
+	0xf7f01c11,
+	0x3121f502,
+	0x4021f506,
+	0x5221f506,
+	0xf5f4bd06,
+	0xf5063121,
+/* 0x083b: ctx_xfer_exec */
+	0x98069221,
+	0x27f11601,
+	0x24b60414,
+	0x0020d006,
+	0xa500e7f1,
+	0xb941e3f0,
+	0x21f4021f,
+	0x04e0b68d,
+	0xf001fcf0,
+	0x24b6022c,
+	0x05f2fd01,
+	0xf18d21f4,
+	0xf04afc17,
+	0x27f00213,
+	0x0012d00c,
+	0x020721f5,
+	0x47fc27f1,
+	0xd00223f0,
+	0x2cf00020,
+	0x0320b601,
+	0xf00012d0,
+	0xa5f001ac,
+	0x00b7f006,
+	0x98140c98,
+	0xe7f0150d,
+	0x5c21f500,
+	0x08a7f001,
+	0x010321f5,
+	0x020721f5,
+	0xf02201f4,
+	0x21f40ca7,
+	0x1017f1c9,
+	0x0614b60a,
+	0xd00527f0,
+/* 0x08c2: ctx_xfer_post_save_wait */
+	0x12cf0012,
+	0x0522fd00,
+	0xf4fa1bf4,
+/* 0x08ce: ctx_xfer_post */
+	0xf7f02e02,
+	0x3121f502,
+	0xf5f4bd06,
+	0xf5067221,
+	0xf5022621,
+	0xbd064021,
+	0x3121f5f4,
+	0x1011f406,
+	0xfd800198,
+	0x0bf40511,
+	0xb121f507,
+/* 0x08f9: ctx_xfer_no_post_mmio */
+/* 0x08f9: ctx_xfer_done */
+	0x0000f807,
 	0x00000000,
 };

drivers/gpu/drm/nouveau/core/engine/graph/nve0.c

@@ -203,7 +203,7 @@ nve0_graph_ctor(struct nouveau_object *parent, struct nouveau_object *engine,
 	struct nvc0_graph_priv *priv;
 	int ret, i;
 
-	ret = nouveau_graph_create(parent, engine, oclass, false, &priv);
+	ret = nouveau_graph_create(parent, engine, oclass, true, &priv);
 	*pobject = nv_object(priv);
 	if (ret)
 		return ret;