drm/nouveau/gr/tu102-: use sw_veid_bundle_init from firmware

NVIDIA provided this on Turing, but we kept using the hardcoded version
from Volta (where they didn't).

Switch to the firmware version prior to Ampere.
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Reviewed-by: Lyude Paul <lyude@redhat.com>

drivers/gpu/drm/nouveau/nvkm/engine/gr/ctxgf100.c

@@ -1381,12 +1381,17 @@ gf100_grctx_generate_main(struct gf100_gr_chan *chan)
 	gf100_gr_wait_idle(gr);
 
 	if (grctx->r400088) grctx->r400088(gr, false);
+
 	if (gr->bundle)
 		gf100_gr_icmd(gr, gr->bundle);
 	else
 		gf100_gr_icmd(gr, grctx->icmd);
+
+	if (gr->bundle_veid)
+		gf100_gr_icmd(gr, gr->bundle_veid);
 	if (grctx->sw_veid_bundle_init)
 		gf100_gr_icmd(gr, grctx->sw_veid_bundle_init);
+
 	if (grctx->r400088) grctx->r400088(gr, true);
 
 	nvkm_wr32(device, 0x404154, idle_timeout);

drivers/gpu/drm/nouveau/nvkm/engine/gr/ctxgf100.h

@@ -153,7 +153,6 @@ extern const struct gf100_grctx_func gv100_grctx;
 extern const struct gf100_grctx_func tu102_grctx;
 void gv100_grctx_unkn88c(struct gf100_gr *, bool);
 void gv100_grctx_generate_unkn(struct gf100_gr *);
-extern const struct gf100_gr_init gv100_grctx_init_sw_veid_bundle_init_0[];
 void gv100_grctx_generate_attrib_cb(struct gf100_gr_chan *, u64, u32);
 void gv100_grctx_generate_attrib(struct gf100_gr_chan *);
 void gv100_grctx_generate_rop_mapping(struct gf100_gr *);

drivers/gpu/drm/nouveau/nvkm/engine/gr/ctxgv100.c

@@ -25,7 +25,7 @@
  * PGRAPH context implementation
  ******************************************************************************/
 
-const struct gf100_gr_init
+static const struct gf100_gr_init
 gv100_grctx_init_sw_veid_bundle_init_0[] = {
 	{ 0x00001000, 64, 0x00100000, 0x00000008 },
 	{ 0x00000941, 64, 0x00100000, 0x00000000 },

drivers/gpu/drm/nouveau/nvkm/engine/gr/ctxtu102.c

@@ -51,7 +51,6 @@ tu102_grctx_init_unknown_bundle_init_0[] = {
 
 static const struct gf100_gr_pack
 tu102_grctx_pack_sw_veid_bundle_init[] = {
-	{ gv100_grctx_init_sw_veid_bundle_init_0 },
 	{ tu102_grctx_init_unknown_bundle_init_0 },
 	{}
 };

drivers/gpu/drm/nouveau/nvkm/engine/gr/gf100.c

@@ -2139,6 +2139,7 @@ gf100_gr_dtor(struct nvkm_gr *base)
 	nvkm_blob_dtor(&gr->gpccs.inst);
 	nvkm_blob_dtor(&gr->gpccs.data);
 
+	vfree(gr->bundle_veid);
 	vfree(gr->bundle);
 	vfree(gr->method);
 	vfree(gr->sw_ctx);

drivers/gpu/drm/nouveau/nvkm/engine/gr/gf100.h

@@ -90,6 +90,7 @@ struct gf100_gr {
 	struct gf100_gr_pack *sw_nonctx;
 	struct gf100_gr_pack *sw_ctx;
 	struct gf100_gr_pack *bundle;
+	struct gf100_gr_pack *bundle_veid;
 	struct gf100_gr_pack *method;
 
 	struct gf100_gr_zbc_color zbc_color[NVKM_LTC_MAX_ZBC_COLOR_CNT];
@@ -224,6 +225,7 @@ void gm107_gr_init_shader_exceptions(struct gf100_gr *, int, int);
 void gm107_gr_init_400054(struct gf100_gr *);
 
 int gk20a_gr_init(struct gf100_gr *);
+int gk20a_gr_av_to_init_(struct nvkm_blob *, u8 count, u32 pitch, struct gf100_gr_pack **);
 int gk20a_gr_av_to_init(struct nvkm_blob *, struct gf100_gr_pack **);
 int gk20a_gr_aiv_to_init(struct nvkm_blob *, struct gf100_gr_pack **);
 int gk20a_gr_av_to_method(struct nvkm_blob *, struct gf100_gr_pack **);
@@ -253,6 +255,8 @@ void gv100_gr_init_504430(struct gf100_gr *, int, int);
 void gv100_gr_init_shader_exceptions(struct gf100_gr *, int, int);
 void gv100_gr_trap_mp(struct gf100_gr *, int, int);
 
+int tu102_gr_av_to_init_veid(struct nvkm_blob *, struct gf100_gr_pack **);
+
 #define gf100_gr_chan(p) container_of((p), struct gf100_gr_chan, object)
 #include <core/object.h>
 

drivers/gpu/drm/nouveau/nvkm/engine/gr/gk20a.c

@@ -34,7 +34,7 @@ struct gk20a_fw_av
 };
 
 int
-gk20a_gr_av_to_init(struct nvkm_blob *blob, struct gf100_gr_pack **ppack)
+gk20a_gr_av_to_init_(struct nvkm_blob *blob, u8 count, u32 pitch, struct gf100_gr_pack **ppack)
 {
 	struct gf100_gr_init *init;
 	struct gf100_gr_pack *pack;
@@ -55,14 +55,20 @@ gk20a_gr_av_to_init(struct nvkm_blob *blob, struct gf100_gr_pack **ppack)
 
 		ent->addr = av->addr;
 		ent->data = av->data;
-		ent->count = 1;
-		ent->pitch = 1;
+		ent->count = ((ent->addr & 0xffff) != 0xe100) ? count : 1;
+		ent->pitch = pitch;
 	}
 
 	*ppack = pack;
 	return 0;
 }
 
+int
+gk20a_gr_av_to_init(struct nvkm_blob *blob, struct gf100_gr_pack **ppack)
+{
+	return gk20a_gr_av_to_init_(blob, 1, 1, ppack);
+}
+
 struct gk20a_fw_aiv
 {
 	u32 addr;

drivers/gpu/drm/nouveau/nvkm/engine/gr/tu102.c

@@ -141,6 +141,7 @@ MODULE_FIRMWARE("nvidia/tu102/gr/sw_ctx.bin");
 MODULE_FIRMWARE("nvidia/tu102/gr/sw_nonctx.bin");
 MODULE_FIRMWARE("nvidia/tu102/gr/sw_bundle_init.bin");
 MODULE_FIRMWARE("nvidia/tu102/gr/sw_method_init.bin");
+MODULE_FIRMWARE("nvidia/tu102/gr/sw_veid_bundle_init.bin");
 
 MODULE_FIRMWARE("nvidia/tu104/gr/fecs_bl.bin");
 MODULE_FIRMWARE("nvidia/tu104/gr/fecs_inst.bin");
@@ -154,6 +155,7 @@ MODULE_FIRMWARE("nvidia/tu104/gr/sw_ctx.bin");
 MODULE_FIRMWARE("nvidia/tu104/gr/sw_nonctx.bin");
 MODULE_FIRMWARE("nvidia/tu104/gr/sw_bundle_init.bin");
 MODULE_FIRMWARE("nvidia/tu104/gr/sw_method_init.bin");
+MODULE_FIRMWARE("nvidia/tu104/gr/sw_veid_bundle_init.bin");
 
 MODULE_FIRMWARE("nvidia/tu106/gr/fecs_bl.bin");
 MODULE_FIRMWARE("nvidia/tu106/gr/fecs_inst.bin");
@@ -167,6 +169,7 @@ MODULE_FIRMWARE("nvidia/tu106/gr/sw_ctx.bin");
 MODULE_FIRMWARE("nvidia/tu106/gr/sw_nonctx.bin");
 MODULE_FIRMWARE("nvidia/tu106/gr/sw_bundle_init.bin");
 MODULE_FIRMWARE("nvidia/tu106/gr/sw_method_init.bin");
+MODULE_FIRMWARE("nvidia/tu106/gr/sw_veid_bundle_init.bin");
 
 MODULE_FIRMWARE("nvidia/tu117/gr/fecs_bl.bin");
 MODULE_FIRMWARE("nvidia/tu117/gr/fecs_inst.bin");
@@ -180,6 +183,7 @@ MODULE_FIRMWARE("nvidia/tu117/gr/sw_ctx.bin");
 MODULE_FIRMWARE("nvidia/tu117/gr/sw_nonctx.bin");
 MODULE_FIRMWARE("nvidia/tu117/gr/sw_bundle_init.bin");
 MODULE_FIRMWARE("nvidia/tu117/gr/sw_method_init.bin");
+MODULE_FIRMWARE("nvidia/tu117/gr/sw_veid_bundle_init.bin");
 
 MODULE_FIRMWARE("nvidia/tu116/gr/fecs_bl.bin");
 MODULE_FIRMWARE("nvidia/tu116/gr/fecs_inst.bin");
@@ -193,6 +197,26 @@ MODULE_FIRMWARE("nvidia/tu116/gr/sw_ctx.bin");
 MODULE_FIRMWARE("nvidia/tu116/gr/sw_nonctx.bin");
 MODULE_FIRMWARE("nvidia/tu116/gr/sw_bundle_init.bin");
 MODULE_FIRMWARE("nvidia/tu116/gr/sw_method_init.bin");
+MODULE_FIRMWARE("nvidia/tu116/gr/sw_veid_bundle_init.bin");
+
+int
+tu102_gr_av_to_init_veid(struct nvkm_blob *blob, struct gf100_gr_pack **ppack)
+{
+	return gk20a_gr_av_to_init_(blob, 64, 0x00100000, ppack);
+}
+
+int
+tu102_gr_load(struct gf100_gr *gr, int ver, const struct gf100_gr_fwif *fwif)
+{
+	int ret;
+
+	ret = gm200_gr_load(gr, ver, fwif);
+	if (ret)
+		return ret;
+
+	return gk20a_gr_load_net(gr, "gr/", "sw_veid_bundle_init", ver, tu102_gr_av_to_init_veid,
+				 &gr->bundle_veid);
+}
 
 static const struct gf100_gr_fwif
 tu102_gr_fwif[] = {