drm/nouveau/gr/gm107: support tpc "strand" ctxsw in gpccs ucode

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>

10 files changed, 1025 insertions, 766 deletions

diff --git a/drivers/gpu/drm/nouveau/nvkm/engine/gr/fuc/gpc.fuc b/drivers/gpu/drm/nouveau/nvkm/engine/gr/fuc/gpc.fuc
index fd45efa6cdf5..194afe910d21 100644
--- a/drivers/gpu/drm/nouveau/nvkm/engine/gr/fuc/gpc.fuc
+++ b/drivers/gpu/drm/nouveau/nvkm/engine/gr/fuc/gpc.fuc
@@ -70,6 +70,43 @@ error:
 	pop $r14
 	ret
 
+#if CHIPSET >= GM107
+tpc_strand_wait:
+	push $r9
+	trace_set(T_STRTPC)
+	tpc_strand_busy:
+		nv_iord($r9, NV_PGRAPH_GPCX_GPCCS_TPC_STATUS, 0)
+		bra b32 $r9 0x0 ne #tpc_strand_busy
+	trace_clr(T_STRTPC)
+	pop $r9
+	ret
+
+#define tpc_strand_wait() call(tpc_strand_wait)
+#define tpc_strand_enable()                                                   /*
+*/	mov $r15 NV_PGRAPH_GPC0_TPCX_STRAND_CMD_ENABLE                        /*
+*/	gpc_wr32(NV_PGRAPH_GPC0_TPCX_STRAND_CMD, $r15)                        /*
+*/	tpc_strand_wait()
+#define tpc_strand_disable()                                                  /*
+*/	mov $r15 NV_PGRAPH_GPC0_TPCX_STRAND_CMD_DISABLE                       /*
+*/	gpc_wr32(NV_PGRAPH_GPC0_TPCX_STRAND_CMD, $r15)                        /*
+*/	tpc_strand_wait()
+#define tpc_strand_seek(p)                                                    /*
+*/	mov $r15 NV_PGRAPH_GPC0_TPCX_STRAND_INDEX_ALL                         /*
+*/	gpc_wr32(NV_PGRAPH_GPC0_TPCX_STRAND_INDEX, $r15)                      /*
+*/	mov $r15 p                                                            /*
+*/	gpc_wr32(NV_PGRAPH_GPC0_TPCX_STRAND_SELECT, $r15)                     /*
+*/	mov $r15 NV_PGRAPH_GPC0_TPCX_STRAND_CMD_SEEK                          /*
+*/	tpc_strand_wait()
+#define tpc_strand_info(m)                                                    /*
+*/	gpc_wr32(NV_PGRAPH_GPC0_TPCX_STRAND_CMD, $r15)                        /*
+*/	mov $r15 m                                                            /*
+*/	gpc_wr32(NV_PGRAPH_GPC0_TPCX_STRAND_DATA, $r15)                       /*
+*/	mov $r15 NV_PGRAPH_GPC0_TPCX_STRAND_CMD_GET_INFO                      /*
+*/	gpc_wr32(NV_PGRAPH_GPC0_TPCX_STRAND_CMD, $r15)                        /*
+*/	tpc_strand_wait()
+#endif
+
+
 // GPC fuc initialisation, executed by triggering ucode start, will
 // fall through to main loop after completion.
 //
@@ -196,8 +233,56 @@ init:
 	// calculate size of strand context data
 	mov b32 $r15 $r2
 	call(strand_ctx_init)
+	add b32 $r2 $r15
 	add b32 $r3 $r15
 
+#if CHIPSET >= GM107
+	// calculate size of tpc strand context data
+	mov $r15 NV_PGRAPH_GPC0_TPCX_STRAND_INDEX_ALL
+	gpc_wr32(NV_PGRAPH_GPC0_TPCX_STRAND_INDEX, $r15)
+	tpc_strand_enable();
+	tpc_strand_seek(0);
+	tpc_strand_info(-1);
+
+	ld b32 $r4 D[$r0 + #tpc_count]
+	mov $r5 NV_PGRAPH_GPC0_TPC0
+	ld b32 $r6 D[$r0 + #gpc_id]
+	shl b32 $r6 15
+	add b32 $r5 $r6
+	tpc_strand_init_tpc_loop:
+		add b32 $r14 $r5 NV_TPC_STRAND_CNT
+		call(nv_rd32)
+		mov b32 $r6 $r15
+		clear b32 $r7
+		tpc_strand_init_idx_loop:
+			add b32 $r14 $r5 NV_TPC_STRAND_INDEX
+			mov b32 $r15 $r7
+			call(nv_wr32)
+			add b32 $r14 $r5 NV_TPC_STRAND_SAVE_SWBASE
+			shr b32 $r15 $r2 8
+			call(nv_wr32)
+			add b32 $r14 $r5 NV_TPC_STRAND_LOAD_SWBASE
+			shr b32 $r15 $r2 8
+			call(nv_wr32)
+			add b32 $r14 $r5 NV_TPC_STRAND_WORDS
+			call(nv_rd32)
+			shr b32 $r15 6
+			add b32 $r15 1
+			shl b32 $r15 8
+			add b32 $r2 $r15
+			add b32 $r3 $r15
+			add b32 $r7 1
+			sub b32 $r6 1
+			bra nz #tpc_strand_init_idx_loop
+		add b32 $r5 NV_PGRAPH_GPC0_TPC0__SIZE
+		sub b32 $r4 1
+		bra nz #tpc_strand_init_tpc_loop
+
+	mov $r15 NV_PGRAPH_GPC0_TPCX_STRAND_INDEX_ALL
+	gpc_wr32(NV_PGRAPH_GPC0_TPCX_STRAND_INDEX, $r15)
+	tpc_strand_disable();
+#endif
+
 	// save context size, and tell HUB we're done
 	nv_iowr(NV_PGRAPH_GPCX_GPCCS_CC_SCRATCH_VAL(1), 0, $r3)
 	clear b32 $r2
@@ -316,6 +401,9 @@ ctx_redswitch:
 ctx_xfer:
 	// set context base address
 	nv_iowr(NV_PGRAPH_GPCX_GPCCS_MEM_BASE, 0, $r15)
+#if CHIPSET >= GM107
+	gpc_wr32(NV_PGRAPH_GPC0_TPCX_STRAND_MEM_BASE, $r15)
+#endif
 	bra not $p1 #ctx_xfer_not_load
 		call(ctx_redswitch)
 	ctx_xfer_not_load:
@@ -328,6 +416,14 @@ ctx_xfer:
 	add b32 $r2 NV_PGRAPH_GPCX_GPCCS_STRAND_CMD_SAVE
 	nv_iowr(NV_PGRAPH_GPCX_GPCCS_STRAND_CMD, 0x3f, $r2)
 
+#if CHIPSET >= GM107
+	tpc_strand_enable();
+	tpc_strand_seek(0);
+	xbit $r15 $flags $p1	// SAVE/LOAD
+	add b32 $r15 NV_PGRAPH_GPC0_TPCX_STRAND_CMD_SAVE
+	gpc_wr32(NV_PGRAPH_GPC0_TPCX_STRAND_CMD, $r15)
+#endif
+
 	// mmio context
 	xbit $r10 $flags $p1	// direction
 	or $r10 2		// first
@@ -372,6 +468,9 @@ ctx_xfer:
 
 	// wait for strands to finish
 	call(strand_wait)
+#if CHIPSET >= GM107
+	tpc_strand_wait()
+#endif
 
 	// if load, or a save without a load following, do some
 	// unknown stuff that's done after finishing a block of
@@ -380,6 +479,9 @@ ctx_xfer:
 	bra not $p2 #ctx_xfer_done
 	ctx_xfer_post:
 		call(strand_post)
+#if CHIPSET >= GM107
+		tpc_strand_disable()
+#endif
 
 	// mark completion in HUB's barrier
 	ctx_xfer_done:
diff --git a/drivers/gpu/drm/nouveau/nvkm/engine/gr/fuc/gpcgf100.fuc3.h b/drivers/gpu/drm/nouveau/nvkm/engine/gr/fuc/gpcgf100.fuc3.h
index 4c0beec5d048..231f696d1e0a 100644
--- a/drivers/gpu/drm/nouveau/nvkm/engine/gr/fuc/gpcgf100.fuc3.h
+++ b/drivers/gpu/drm/nouveau/nvkm/engine/gr/fuc/gpcgf100.fuc3.h
@@ -310,7 +310,7 @@ uint32_t gf100_grgpc_code[] = {
 	0x03f01200,
 	0x0002d000,
 	0x17f104bd,
-	0x10fe04f5,
+	0x10fe04f8,
 	0x0007f100,
 	0x0003f007,
 	0xbd0000d0,
@@ -362,124 +362,124 @@ uint32_t gf100_grgpc_code[] = {
 	0x34b60824,
 	0x022fb908,
 	0x02d321f5,
-	0xf1003fbb,
-	0xf0010007,
-	0x03d00203,
-	0xbd04bd00,
-	0x1f29f024,
-	0x080007f1,
-	0xd00203f0,
-	0x04bd0002,
-/* 0x04b8: main */
-	0xf40031f4,
-	0xd7f00028,
-	0x3921f41c,
-	0xb0f401f4,
-	0x18f404e4,
-	0x0181fe1e,
-	0xbd0627f0,
-	0x0412fd20,
-	0xfd01e4b6,
-	0x18fe051e,
-	0xad21f500,
-	0xd30ef405,
-/* 0x04e8: main_not_ctx_xfer */
-	0xf010ef94,
-	0x21f501f5,
-	0x0ef4037e,
-/* 0x04f5: ih */
-	0xfe80f9c6,
-	0x80f90188,
-	0xa0f990f9,
-	0xd0f9b0f9,
-	0xf0f9e0f9,
-	0xa7f104bd,
-	0xa3f00200,
-	0x00aacf00,
-	0xf404abc4,
-	0xd7f02c0b,
-	0x00e7f11c,
-	0x00e3f01a,
-	0xf100eecf,
-	0xf01900f7,
-	0xffcf00f3,
-	0x0421f400,
-	0xf101e7f0,
-	0xf01d0007,
-	0x0ed00003,
-/* 0x0543: ih_no_fifo */
-	0xf104bd00,
-	0xf0010007,
-	0x0ad00003,
-	0xfc04bd00,
-	0xfce0fcf0,
-	0xfcb0fcd0,
-	0xfc90fca0,
-	0x0088fe80,
-	0x32f480fc,
-/* 0x0567: hub_barrier_done */
-	0xf001f800,
-	0x0e9801f7,
-	0x04febb04,
-	0xf102ffb9,
-	0xf09418e7,
-	0x21f440e3,
-/* 0x057f: ctx_redswitch */
-	0xf000f89d,
-	0x07f120f7,
+	0xbb002fbb,
+	0x07f1003f,
+	0x03f00100,
+	0x0003d002,
+	0x24bd04bd,
+	0xf11f29f0,
+	0xf0080007,
+	0x02d00203,
+/* 0x04bb: main */
+	0xf404bd00,
+	0x28f40031,
+	0x1cd7f000,
+	0xf43921f4,
+	0xe4b0f401,
+	0x1e18f404,
+	0xf00181fe,
+	0x20bd0627,
+	0xb60412fd,
+	0x1efd01e4,
+	0x0018fe05,
+	0x05b021f5,
+/* 0x04eb: main_not_ctx_xfer */
+	0x94d30ef4,
+	0xf5f010ef,
+	0x7e21f501,
+	0xc60ef403,
+/* 0x04f8: ih */
+	0x88fe80f9,
+	0xf980f901,
+	0xf9a0f990,
+	0xf9d0f9b0,
+	0xbdf0f9e0,
+	0x00a7f104,
+	0x00a3f002,
+	0xc400aacf,
+	0x0bf404ab,
+	0x1cd7f02c,
+	0x1a00e7f1,
+	0xcf00e3f0,
+	0xf7f100ee,
+	0xf3f01900,
+	0x00ffcf00,
+	0xf00421f4,
+	0x07f101e7,
+	0x03f01d00,
+	0x000ed000,
+/* 0x0546: ih_no_fifo */
+	0x07f104bd,
+	0x03f00100,
+	0x000ad000,
+	0xf0fc04bd,
+	0xd0fce0fc,
+	0xa0fcb0fc,
+	0x80fc90fc,
+	0xfc0088fe,
+	0x0032f480,
+/* 0x056a: hub_barrier_done */
+	0xf7f001f8,
+	0x040e9801,
+	0xb904febb,
+	0xe7f102ff,
+	0xe3f09418,
+	0x9d21f440,
+/* 0x0582: ctx_redswitch */
+	0xf7f000f8,
+	0x0007f120,
+	0x0103f085,
+	0xbd000fd0,
+	0x08e7f004,
+/* 0x0594: ctx_redswitch_delay */
+	0xf401e2b6,
+	0xf5f1fd1b,
+	0xf5f10800,
+	0x07f10200,
 	0x03f08500,
 	0x000fd001,
-	0xe7f004bd,
-/* 0x0591: ctx_redswitch_delay */
-	0x01e2b608,
-	0xf1fd1bf4,
-	0xf10800f5,
-	0xf10200f5,
-	0xf0850007,
-	0x0fd00103,
-	0xf804bd00,
-/* 0x05ad: ctx_xfer */
-	0x0007f100,
-	0x0203f081,
-	0xbd000fd0,
-	0x0711f404,
-	0x057f21f5,
-/* 0x05c0: ctx_xfer_not_load */
-	0x026a21f5,
-	0x07f124bd,
-	0x03f047fc,
-	0x0002d002,
-	0x2cf004bd,
-	0x0320b601,
-	0x4afc07f1,
+	0x00f804bd,
+/* 0x05b0: ctx_xfer */
+	0x810007f1,
 	0xd00203f0,
-	0x04bd0002,
-	0xf001acf0,
-	0xb7f102a5,
-	0xb3f00000,
-	0x040c9850,
-	0xbb0fc4b6,
-	0x0c9800bc,
-	0x010d9800,
-	0xf500e7f0,
-	0xf0016f21,
+	0x04bd000f,
+	0xf50711f4,
+/* 0x05c3: ctx_xfer_not_load */
+	0xf5058221,
+	0xbd026a21,
+	0xfc07f124,
+	0x0203f047,
+	0xbd0002d0,
+	0x012cf004,
+	0xf10320b6,
+	0xf04afc07,
+	0x02d00203,
+	0xf004bd00,
 	0xa5f001ac,
-	0x00b7f104,
-	0x50b3f040,
+	0x00b7f102,
+	0x50b3f000,
 	0xb6040c98,
 	0xbcbb0fc4,
-	0x010c9800,
-	0x98020d98,
-	0xe7f1060f,
-	0x21f50800,
-	0x21f5016f,
-	0x01f4025e,
-	0x0712f406,
-/* 0x0638: ctx_xfer_post */
-	0x027f21f5,
-/* 0x063c: ctx_xfer_done */
-	0x056721f5,
-	0x000000f8,
+	0x000c9800,
+	0xf0010d98,
+	0x21f500e7,
+	0xacf0016f,
+	0x04a5f001,
+	0x4000b7f1,
+	0x9850b3f0,
+	0xc4b6040c,
+	0x00bcbb0f,
+	0x98010c98,
+	0x0f98020d,
+	0x00e7f106,
+	0x6f21f508,
+	0x5e21f501,
+	0x0601f402,
+/* 0x063b: ctx_xfer_post */
+	0xf50712f4,
+/* 0x063f: ctx_xfer_done */
+	0xf5027f21,
+	0xf8056a21,
 	0x00000000,
 	0x00000000,
 	0x00000000,
diff --git a/drivers/gpu/drm/nouveau/nvkm/engine/gr/fuc/gpcgf117.fuc3.h b/drivers/gpu/drm/nouveau/nvkm/engine/gr/fuc/gpcgf117.fuc3.h
index 3b4b8491b834..64d07df4b8b1 100644
--- a/drivers/gpu/drm/nouveau/nvkm/engine/gr/fuc/gpcgf117.fuc3.h
+++ b/drivers/gpu/drm/nouveau/nvkm/engine/gr/fuc/gpcgf117.fuc3.h
@@ -314,7 +314,7 @@ uint32_t gf117_grgpc_code[] = {
 	0x03f01200,
 	0x0002d000,
 	0x17f104bd,
-	0x10fe053f,
+	0x10fe0542,
 	0x0007f100,
 	0x0003f007,
 	0xbd0000d0,
@@ -387,134 +387,134 @@ uint32_t gf117_grgpc_code[] = {
 	0x0824b601,
 	0xb90834b6,
 	0x21f5022f,
-	0x3fbb02d3,
-	0x0007f100,
-	0x0203f001,
-	0xbd0003d0,
-	0xf024bd04,
-	0x07f11f29,
-	0x03f00800,
-	0x0002d002,
-/* 0x0502: main */
-	0x31f404bd,
-	0x0028f400,
-	0xf424d7f0,
-	0x01f43921,
-	0x04e4b0f4,
-	0xfe1e18f4,
-	0x27f00181,
-	0xfd20bd06,
-	0xe4b60412,
-	0x051efd01,
-	0xf50018fe,
-	0xf405f721,
-/* 0x0532: main_not_ctx_xfer */
-	0xef94d30e,
-	0x01f5f010,
-	0x037e21f5,
-/* 0x053f: ih */
-	0xf9c60ef4,
-	0x0188fe80,
-	0x90f980f9,
-	0xb0f9a0f9,
-	0xe0f9d0f9,
-	0x04bdf0f9,
-	0x0200a7f1,
-	0xcf00a3f0,
-	0xabc400aa,
-	0x2c0bf404,
-	0xf124d7f0,
-	0xf01a00e7,
-	0xeecf00e3,
-	0x00f7f100,
-	0x00f3f019,
-	0xf400ffcf,
-	0xe7f00421,
-	0x0007f101,
-	0x0003f01d,
-	0xbd000ed0,
-/* 0x058d: ih_no_fifo */
-	0x0007f104,
-	0x0003f001,
-	0xbd000ad0,
-	0xfcf0fc04,
-	0xfcd0fce0,
-	0xfca0fcb0,
-	0xfe80fc90,
-	0x80fc0088,
-	0xf80032f4,
-/* 0x05b1: hub_barrier_done */
-	0x01f7f001,
-	0xbb040e98,
-	0xffb904fe,
-	0x18e7f102,
-	0x40e3f094,
-	0xf89d21f4,
-/* 0x05c9: ctx_redswitch */
-	0x20f7f000,
+	0x2fbb02d3,
+	0x003fbb00,
+	0x010007f1,
+	0xd00203f0,
+	0x04bd0003,
+	0x29f024bd,
+	0x0007f11f,
+	0x0203f008,
+	0xbd0002d0,
+/* 0x0505: main */
+	0x0031f404,
+	0xf00028f4,
+	0x21f424d7,
+	0xf401f439,
+	0xf404e4b0,
+	0x81fe1e18,
+	0x0627f001,
+	0x12fd20bd,
+	0x01e4b604,
+	0xfe051efd,
+	0x21f50018,
+	0x0ef405fa,
+/* 0x0535: main_not_ctx_xfer */
+	0x10ef94d3,
+	0xf501f5f0,
+	0xf4037e21,
+/* 0x0542: ih */
+	0x80f9c60e,
+	0xf90188fe,
+	0xf990f980,
+	0xf9b0f9a0,
+	0xf9e0f9d0,
+	0xf104bdf0,
+	0xf00200a7,
+	0xaacf00a3,
+	0x04abc400,
+	0xf02c0bf4,
+	0xe7f124d7,
+	0xe3f01a00,
+	0x00eecf00,
+	0x1900f7f1,
+	0xcf00f3f0,
+	0x21f400ff,
+	0x01e7f004,
+	0x1d0007f1,
+	0xd00003f0,
+	0x04bd000e,
+/* 0x0590: ih_no_fifo */
+	0x010007f1,
+	0xd00003f0,
+	0x04bd000a,
+	0xe0fcf0fc,
+	0xb0fcd0fc,
+	0x90fca0fc,
+	0x88fe80fc,
+	0xf480fc00,
+	0x01f80032,
+/* 0x05b4: hub_barrier_done */
+	0x9801f7f0,
+	0xfebb040e,
+	0x02ffb904,
+	0x9418e7f1,
+	0xf440e3f0,
+	0x00f89d21,
+/* 0x05cc: ctx_redswitch */
+	0xf120f7f0,
+	0xf0850007,
+	0x0fd00103,
+	0xf004bd00,
+/* 0x05de: ctx_redswitch_delay */
+	0xe2b608e7,
+	0xfd1bf401,
+	0x0800f5f1,
+	0x0200f5f1,
 	0x850007f1,
 	0xd00103f0,
 	0x04bd000f,
-/* 0x05db: ctx_redswitch_delay */
-	0xb608e7f0,
-	0x1bf401e2,
-	0x00f5f1fd,
-	0x00f5f108,
-	0x0007f102,
-	0x0103f085,
-	0xbd000fd0,
-/* 0x05f7: ctx_xfer */
-	0xf100f804,
-	0xf0810007,
-	0x0fd00203,
-	0xf404bd00,
-	0x21f50711,
-/* 0x060a: ctx_xfer_not_load */
-	0x21f505c9,
-	0x24bd026a,
-	0x47fc07f1,
-	0xd00203f0,
-	0x04bd0002,
-	0xb6012cf0,
-	0x07f10320,
-	0x03f04afc,
-	0x0002d002,
-	0xacf004bd,
-	0x02a5f001,
-	0x0000b7f1,
-	0x9850b3f0,
-	0xc4b6040c,
-	0x00bcbb0f,
-	0x98000c98,
-	0xe7f0010d,
-	0x6f21f500,
-	0x01acf001,
-	0x4000b7f1,
-	0x9850b3f0,
-	0xc4b6040c,
-	0x00bcbb0f,
-	0x98010c98,
-	0x0f98020d,
-	0x00e7f106,
-	0x6f21f508,
-	0x01acf001,
-	0xf104a5f0,
-	0xf03000b7,
+/* 0x05fa: ctx_xfer */
+	0x07f100f8,
+	0x03f08100,
+	0x000fd002,
+	0x11f404bd,
+	0xcc21f507,
+/* 0x060d: ctx_xfer_not_load */
+	0x6a21f505,
+	0xf124bd02,
+	0xf047fc07,
+	0x02d00203,
+	0xf004bd00,
+	0x20b6012c,
+	0xfc07f103,
+	0x0203f04a,
+	0xbd0002d0,
+	0x01acf004,
+	0xf102a5f0,
+	0xf00000b7,
 	0x0c9850b3,
 	0x0fc4b604,
 	0x9800bcbb,
-	0x0d98020c,
-	0x080f9803,
-	0x0200e7f1,
+	0x0d98000c,
+	0x00e7f001,
 	0x016f21f5,
-	0x025e21f5,
-	0xf40601f4,
-/* 0x06a6: ctx_xfer_post */
-	0x21f50712,
-/* 0x06aa: ctx_xfer_done */
-	0x21f5027f,
-	0x00f805b1,
-	0x00000000,
+	0xf101acf0,
+	0xf04000b7,
+	0x0c9850b3,
+	0x0fc4b604,
+	0x9800bcbb,
+	0x0d98010c,
+	0x060f9802,
+	0x0800e7f1,
+	0x016f21f5,
+	0xf001acf0,
+	0xb7f104a5,
+	0xb3f03000,
+	0x040c9850,
+	0xbb0fc4b6,
+	0x0c9800bc,
+	0x030d9802,
+	0xf1080f98,
+	0xf50200e7,
+	0xf5016f21,
+	0xf4025e21,
+	0x12f40601,
+/* 0x06a9: ctx_xfer_post */
+	0x7f21f507,
+/* 0x06ad: ctx_xfer_done */
+	0xb421f502,
+	0x0000f805,
 	0x00000000,
 	0x00000000,
 	0x00000000,
diff --git a/drivers/gpu/drm/nouveau/nvkm/engine/gr/fuc/gpcgk104.fuc3.h b/drivers/gpu/drm/nouveau/nvkm/engine/gr/fuc/gpcgk104.fuc3.h
index e28b0fd11a74..2f596433c222 100644
--- a/drivers/gpu/drm/nouveau/nvkm/engine/gr/fuc/gpcgk104.fuc3.h
+++ b/drivers/gpu/drm/nouveau/nvkm/engine/gr/fuc/gpcgk104.fuc3.h
@@ -314,7 +314,7 @@ uint32_t gk104_grgpc_code[] = {
 	0x03f01200,
 	0x0002d000,
 	0x17f104bd,
-	0x10fe053f,
+	0x10fe0542,
 	0x0007f100,
 	0x0003f007,
 	0xbd0000d0,
@@ -387,134 +387,134 @@ uint32_t gk104_grgpc_code[] = {
 	0x0824b601,
 	0xb90834b6,
 	0x21f5022f,
-	0x3fbb02d3,
-	0x0007f100,
-	0x0203f001,
-	0xbd0003d0,
-	0xf024bd04,
-	0x07f11f29,
-	0x03f00800,
-	0x0002d002,
-/* 0x0502: main */
-	0x31f404bd,
-	0x0028f400,
-	0xf424d7f0,
-	0x01f43921,
-	0x04e4b0f4,
-	0xfe1e18f4,
-	0x27f00181,
-	0xfd20bd06,
-	0xe4b60412,
-	0x051efd01,
-	0xf50018fe,
-	0xf405f721,
-/* 0x0532: main_not_ctx_xfer */
-	0xef94d30e,
-	0x01f5f010,
-	0x037e21f5,
-/* 0x053f: ih */
-	0xf9c60ef4,
-	0x0188fe80,
-	0x90f980f9,
-	0xb0f9a0f9,
-	0xe0f9d0f9,
-	0x04bdf0f9,
-	0x0200a7f1,
-	0xcf00a3f0,
-	0xabc400aa,
-	0x2c0bf404,
-	0xf124d7f0,
-	0xf01a00e7,
-	0xeecf00e3,
-	0x00f7f100,
-	0x00f3f019,
-	0xf400ffcf,
-	0xe7f00421,
-	0x0007f101,
-	0x0003f01d,
-	0xbd000ed0,
-/* 0x058d: ih_no_fifo */
-	0x0007f104,
-	0x0003f001,
-	0xbd000ad0,
-	0xfcf0fc04,
-	0xfcd0fce0,
-	0xfca0fcb0,
-	0xfe80fc90,
-	0x80fc0088,
-	0xf80032f4,
-/* 0x05b1: hub_barrier_done */
-	0x01f7f001,
-	0xbb040e98,
-	0xffb904fe,
-	0x18e7f102,
-	0x40e3f094,
-	0xf89d21f4,
-/* 0x05c9: ctx_redswitch */
-	0x20f7f000,
+	0x2fbb02d3,
+	0x003fbb00,
+	0x010007f1,
+	0xd00203f0,
+	0x04bd0003,
+	0x29f024bd,
+	0x0007f11f,
+	0x0203f008,
+	0xbd0002d0,
+/* 0x0505: main */
+	0x0031f404,
+	0xf00028f4,
+	0x21f424d7,
+	0xf401f439,
+	0xf404e4b0,
+	0x81fe1e18,
+	0x0627f001,
+	0x12fd20bd,
+	0x01e4b604,
+	0xfe051efd,
+	0x21f50018,
+	0x0ef405fa,
+/* 0x0535: main_not_ctx_xfer */
+	0x10ef94d3,
+	0xf501f5f0,
+	0xf4037e21,
+/* 0x0542: ih */
+	0x80f9c60e,
+	0xf90188fe,
+	0xf990f980,
+	0xf9b0f9a0,
+	0xf9e0f9d0,
+	0xf104bdf0,
+	0xf00200a7,
+	0xaacf00a3,
+	0x04abc400,
+	0xf02c0bf4,
+	0xe7f124d7,
+	0xe3f01a00,
+	0x00eecf00,
+	0x1900f7f1,
+	0xcf00f3f0,
+	0x21f400ff,
+	0x01e7f004,
+	0x1d0007f1,
+	0xd00003f0,
+	0x04bd000e,
+/* 0x0590: ih_no_fifo */
+	0x010007f1,
+	0xd00003f0,
+	0x04bd000a,
+	0xe0fcf0fc,
+	0xb0fcd0fc,
+	0x90fca0fc,
+	0x88fe80fc,
+	0xf480fc00,
+	0x01f80032,
+/* 0x05b4: hub_barrier_done */
+	0x9801f7f0,
+	0xfebb040e,
+	0x02ffb904,
+	0x9418e7f1,
+	0xf440e3f0,
+	0x00f89d21,
+/* 0x05cc: ctx_redswitch */
+	0xf120f7f0,
+	0xf0850007,
+	0x0fd00103,
+	0xf004bd00,
+/* 0x05de: ctx_redswitch_delay */
+	0xe2b608e7,
+	0xfd1bf401,
+	0x0800f5f1,
+	0x0200f5f1,
 	0x850007f1,
 	0xd00103f0,
 	0x04bd000f,
-/* 0x05db: ctx_redswitch_delay */
-	0xb608e7f0,
-	0x1bf401e2,
-	0x00f5f1fd,
-	0x00f5f108,
-	0x0007f102,
-	0x0103f085,
-	0xbd000fd0,
-/* 0x05f7: ctx_xfer */
-	0xf100f804,
-	0xf0810007,
-	0x0fd00203,
-	0xf404bd00,
-	0x21f50711,
-/* 0x060a: ctx_xfer_not_load */
-	0x21f505c9,
-	0x24bd026a,
-	0x47fc07f1,
-	0xd00203f0,
-	0x04bd0002,
-	0xb6012cf0,
-	0x07f10320,
-	0x03f04afc,
-	0x0002d002,
-	0xacf004bd,
-	0x02a5f001,
-	0x0000b7f1,
-	0x9850b3f0,
-	0xc4b6040c,
-	0x00bcbb0f,
-	0x98000c98,
-	0xe7f0010d,
-	0x6f21f500,
-	0x01acf001,
-	0x4000b7f1,
-	0x9850b3f0,
-	0xc4b6040c,
-	0x00bcbb0f,
-	0x98010c98,
-	0x0f98020d,
-	0x00e7f106,
-	0x6f21f508,
-	0x01acf001,
-	0xf104a5f0,
-	0xf03000b7,
+/* 0x05fa: ctx_xfer */
+	0x07f100f8,
+	0x03f08100,
+	0x000fd002,
+	0x11f404bd,
+	0xcc21f507,
+/* 0x060d: ctx_xfer_not_load */
+	0x6a21f505,
+	0xf124bd02,
+	0xf047fc07,
+	0x02d00203,
+	0xf004bd00,
+	0x20b6012c,
+	0xfc07f103,
+	0x0203f04a,
+	0xbd0002d0,
+	0x01acf004,
+	0xf102a5f0,
+	0xf00000b7,
 	0x0c9850b3,
 	0x0fc4b604,
 	0x9800bcbb,
-	0x0d98020c,
-	0x080f9803,
-	0x0200e7f1,
+	0x0d98000c,
+	0x00e7f001,
 	0x016f21f5,
-	0x025e21f5,
-	0xf40601f4,
-/* 0x06a6: ctx_xfer_post */
-	0x21f50712,
-/* 0x06aa: ctx_xfer_done */
-	0x21f5027f,
-	0x00f805b1,
-	0x00000000,
+	0xf101acf0,
+	0xf04000b7,
+	0x0c9850b3,
+	0x0fc4b604,
+	0x9800bcbb,
+	0x0d98010c,
+	0x060f9802,
+	0x0800e7f1,
+	0x016f21f5,
+	0xf001acf0,
+	0xb7f104a5,
+	0xb3f03000,
+	0x040c9850,
+	0xbb0fc4b6,
+	0x0c9800bc,
+	0x030d9802,
+	0xf1080f98,
+	0xf50200e7,
+	0xf5016f21,
+	0xf4025e21,
+	0x12f40601,
+/* 0x06a9: ctx_xfer_post */
+	0x7f21f507,
+/* 0x06ad: ctx_xfer_done */
+	0xb421f502,
+	0x0000f805,
 	0x00000000,
 	0x00000000,
 	0x00000000,
diff --git a/drivers/gpu/drm/nouveau/nvkm/engine/gr/fuc/gpcgk110.fuc3.h b/drivers/gpu/drm/nouveau/nvkm/engine/gr/fuc/gpcgk110.fuc3.h
index 8aae7e8092ca..ee8e54db8fc9 100644
--- a/drivers/gpu/drm/nouveau/nvkm/engine/gr/fuc/gpcgk110.fuc3.h
+++ b/drivers/gpu/drm/nouveau/nvkm/engine/gr/fuc/gpcgk110.fuc3.h
@@ -314,7 +314,7 @@ uint32_t gk110_grgpc_code[] = {
 	0x03f01200,
 	0x0002d000,
 	0x17f104bd,
-	0x10fe053f,
+	0x10fe0542,
 	0x0007f100,
 	0x0003f007,
 	0xbd0000d0,
@@ -387,134 +387,134 @@ uint32_t gk110_grgpc_code[] = {
 	0x0824b601,
 	0xb90834b6,
 	0x21f5022f,
-	0x3fbb02d3,
-	0x0007f100,
-	0x0203f001,
-	0xbd0003d0,
-	0xf024bd04,
-	0x07f11f29,
-	0x03f03000,
-	0x0002d002,
-/* 0x0502: main */
-	0x31f404bd,
-	0x0028f400,
-	0xf424d7f0,
-	0x01f43921,
-	0x04e4b0f4,
-	0xfe1e18f4,
-	0x27f00181,
-	0xfd20bd06,
-	0xe4b60412,
-	0x051efd01,
-	0xf50018fe,
-	0xf405f721,
-/* 0x0532: main_not_ctx_xfer */
-	0xef94d30e,
-	0x01f5f010,
-	0x037e21f5,
-/* 0x053f: ih */
-	0xf9c60ef4,
-	0x0188fe80,
-	0x90f980f9,
-	0xb0f9a0f9,
-	0xe0f9d0f9,
-	0x04bdf0f9,
-	0x0200a7f1,
-	0xcf00a3f0,
-	0xabc400aa,
-	0x2c0bf404,
-	0xf124d7f0,
-	0xf01a00e7,
-	0xeecf00e3,
-	0x00f7f100,
-	0x00f3f019,
-	0xf400ffcf,
-	0xe7f00421,
-	0x0007f101,
-	0x0003f01d,
-	0xbd000ed0,
-/* 0x058d: ih_no_fifo */
-	0x0007f104,
-	0x0003f001,
-	0xbd000ad0,
-	0xfcf0fc04,
-	0xfcd0fce0,
-	0xfca0fcb0,
-	0xfe80fc90,
-	0x80fc0088,
-	0xf80032f4,
-/* 0x05b1: hub_barrier_done */
-	0x01f7f001,
-	0xbb040e98,
-	0xffb904fe,
-	0x18e7f102,
-	0x40e3f094,
-	0xf89d21f4,
-/* 0x05c9: ctx_redswitch */
-	0x20f7f000,
+	0x2fbb02d3,
+	0x003fbb00,
+	0x010007f1,
+	0xd00203f0,
+	0x04bd0003,
+	0x29f024bd,
+	0x0007f11f,
+	0x0203f030,
+	0xbd0002d0,
+/* 0x0505: main */
+	0x0031f404,
+	0xf00028f4,
+	0x21f424d7,
+	0xf401f439,
+	0xf404e4b0,
+	0x81fe1e18,
+	0x0627f001,
+	0x12fd20bd,
+	0x01e4b604,
+	0xfe051efd,
+	0x21f50018,
+	0x0ef405fa,
+/* 0x0535: main_not_ctx_xfer */
+	0x10ef94d3,
+	0xf501f5f0,
+	0xf4037e21,
+/* 0x0542: ih */
+	0x80f9c60e,
+	0xf90188fe,
+	0xf990f980,
+	0xf9b0f9a0,
+	0xf9e0f9d0,
+	0xf104bdf0,
+	0xf00200a7,
+	0xaacf00a3,
+	0x04abc400,
+	0xf02c0bf4,
+	0xe7f124d7,
+	0xe3f01a00,
+	0x00eecf00,
+	0x1900f7f1,
+	0xcf00f3f0,
+	0x21f400ff,
+	0x01e7f004,
+	0x1d0007f1,
+	0xd00003f0,
+	0x04bd000e,
+/* 0x0590: ih_no_fifo */
+	0x010007f1,
+	0xd00003f0,
+	0x04bd000a,
+	0xe0fcf0fc,
+	0xb0fcd0fc,
+	0x90fca0fc,
+	0x88fe80fc,
+	0xf480fc00,
+	0x01f80032,
+/* 0x05b4: hub_barrier_done */
+	0x9801f7f0,
+	0xfebb040e,
+	0x02ffb904,
+	0x9418e7f1,
+	0xf440e3f0,
+	0x00f89d21,
+/* 0x05cc: ctx_redswitch */
+	0xf120f7f0,
+	0xf0850007,
+	0x0fd00103,
+	0xf004bd00,
+/* 0x05de: ctx_redswitch_delay */
+	0xe2b608e7,
+	0xfd1bf401,
+	0x0800f5f1,
+	0x0200f5f1,
 	0x850007f1,
 	0xd00103f0,
 	0x04bd000f,
-/* 0x05db: ctx_redswitch_delay */
-	0xb608e7f0,
-	0x1bf401e2,
-	0x00f5f1fd,
-	0x00f5f108,
-	0x0007f102,
-	0x0103f085,
-	0xbd000fd0,
-/* 0x05f7: ctx_xfer */
-	0xf100f804,
-	0xf0810007,
-	0x0fd00203,
-	0xf404bd00,
-	0x21f50711,
-/* 0x060a: ctx_xfer_not_load */
-	0x21f505c9,
-	0x24bd026a,
-	0x47fc07f1,
-	0xd00203f0,
-	0x04bd0002,
-	0xb6012cf0,
-	0x07f10320,
-	0x03f04afc,
-	0x0002d002,
-	0xacf004bd,
-	0x02a5f001,
-	0x0000b7f1,
-	0x9850b3f0,
-	0xc4b6040c,
-	0x00bcbb0f,
-	0x98000c98,
-	0xe7f0010d,
-	0x6f21f500,
-	0x01acf001,
-	0x4000b7f1,
-	0x9850b3f0,
-	0xc4b6040c,
-	0x00bcbb0f,
-	0x98010c98,
-	0x0f98020d,
-	0x00e7f106,
-	0x6f21f508,
-	0x01acf001,
-	0xf104a5f0,
-	0xf03000b7,
+/* 0x05fa: ctx_xfer */
+	0x07f100f8,
+	0x03f08100,
+	0x000fd002,
+	0x11f404bd,
+	0xcc21f507,
+/* 0x060d: ctx_xfer_not_load */
+	0x6a21f505,
+	0xf124bd02,
+	0xf047fc07,
+	0x02d00203,
+	0xf004bd00,
+	0x20b6012c,
+	0xfc07f103,
+	0x0203f04a,
+	0xbd0002d0,
+	0x01acf004,
+	0xf102a5f0,
+	0xf00000b7,
 	0x0c9850b3,
 	0x0fc4b604,
 	0x9800bcbb,
-	0x0d98020c,
-	0x080f9803,
-	0x0200e7f1,
+	0x0d98000c,
+	0x00e7f001,
 	0x016f21f5,
-	0x025e21f5,
-	0xf40601f4,
-/* 0x06a6: ctx_xfer_post */
-	0x21f50712,
-/* 0x06aa: ctx_xfer_done */
-	0x21f5027f,
-	0x00f805b1,
-	0x00000000,
+	0xf101acf0,
+	0xf04000b7,
+	0x0c9850b3,
+	0x0fc4b604,
+	0x9800bcbb,
+	0x0d98010c,
+	0x060f9802,
+	0x0800e7f1,
+	0x016f21f5,
+	0xf001acf0,
+	0xb7f104a5,
+	0xb3f03000,
+	0x040c9850,
+	0xbb0fc4b6,
+	0x0c9800bc,
+	0x030d9802,
+	0xf1080f98,
+	0xf50200e7,
+	0xf5016f21,
+	0xf4025e21,
+	0x12f40601,
+/* 0x06a9: ctx_xfer_post */
+	0x7f21f507,
+/* 0x06ad: ctx_xfer_done */
+	0xb421f502,
+	0x0000f805,
 	0x00000000,
 	0x00000000,
 	0x00000000,
diff --git a/drivers/gpu/drm/nouveau/nvkm/engine/gr/fuc/gpcgk208.fuc5.h b/drivers/gpu/drm/nouveau/nvkm/engine/gr/fuc/gpcgk208.fuc5.h
index e40c6e43eb86..fbcc342f896f 100644
--- a/drivers/gpu/drm/nouveau/nvkm/engine/gr/fuc/gpcgk208.fuc5.h
+++ b/drivers/gpu/drm/nouveau/nvkm/engine/gr/fuc/gpcgk208.fuc5.h
@@ -276,7 +276,7 @@ uint32_t gk208_grgpc_code[] = {
 	0x02020014,
 	0xf6120040,
 	0x04bd0002,
-	0xfe047e41,
+	0xfe048141,
 	0x00400010,
 	0x0000f607,
 	0x040204bd,
@@ -341,118 +341,118 @@ uint32_t gk208_grgpc_code[] = {
 	0xb60824b6,
 	0x2fb20834,
 	0x0002687e,
-	0x80003fbb,
-	0xf6020100,
-	0x04bd0003,
-	0x29f024bd,
-	0x3000801f,
-	0x0002f602,
-/* 0x0442: main */
-	0x31f404bd,
-	0x0028f400,
-	0x377e240d,
-	0x01f40000,
-	0x04e4b0f4,
-	0xfe1d18f4,
-	0x06020181,
-	0x12fd20bd,
-	0x01e4b604,
-	0xfe051efd,
-	0x157e0018,
-	0x0ef40005,
-/* 0x0471: main_not_ctx_xfer */
-	0x10ef94d4,
-	0x7e01f5f0,
-	0xf40002f8,
-/* 0x047e: ih */
-	0x80f9c70e,
-	0xf90188fe,
-	0xf990f980,
-	0xf9b0f9a0,
-	0xf9e0f9d0,
-	0x4a04bdf0,
-	0xaacf0200,
-	0x04abc400,
-	0x0d1f0bf4,
-	0x1a004e24,
-	0x4f00eecf,
-	0xffcf1900,
-	0x00047e00,
-	0x40010e00,
-	0x0ef61d00,
-/* 0x04bb: ih_no_fifo */
-	0x4004bd00,
-	0x0af60100,
-	0xfc04bd00,
-	0xfce0fcf0,
-	0xfcb0fcd0,
-	0xfc90fca0,
-	0x0088fe80,
-	0x32f480fc,
-/* 0x04db: hub_barrier_done */
-	0x0f01f800,
-	0x040e9801,
-	0xb204febb,
-	0x94188eff,
-	0x008f7e40,
-/* 0x04ef: ctx_redswitch */
-	0x0f00f800,
-	0x85008020,
+	0xbb002fbb,
+	0x0080003f,
+	0x03f60201,
+	0xbd04bd00,
+	0x1f29f024,
+	0x02300080,
+	0xbd0002f6,
+/* 0x0445: main */
+	0x0031f404,
+	0x0d0028f4,
+	0x00377e24,
+	0xf401f400,
+	0xf404e4b0,
+	0x81fe1d18,
+	0xbd060201,
+	0x0412fd20,
+	0xfd01e4b6,
+	0x18fe051e,
+	0x05187e00,
+	0xd40ef400,
+/* 0x0474: main_not_ctx_xfer */
+	0xf010ef94,
+	0xf87e01f5,
+	0x0ef40002,
+/* 0x0481: ih */
+	0xfe80f9c7,
+	0x80f90188,
+	0xa0f990f9,
+	0xd0f9b0f9,
+	0xf0f9e0f9,
+	0x004a04bd,
+	0x00aacf02,
+	0xf404abc4,
+	0x240d1f0b,
+	0xcf1a004e,
+	0x004f00ee,
+	0x00ffcf19,
+	0x0000047e,
+	0x0040010e,
+	0x000ef61d,
+/* 0x04be: ih_no_fifo */
+	0x004004bd,
+	0x000af601,
+	0xf0fc04bd,
+	0xd0fce0fc,
+	0xa0fcb0fc,
+	0x80fc90fc,
+	0xfc0088fe,
+	0x