drm/nouveau/gr/gf100-: virtualise sm_id/tpc_nr

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

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

@@ -1079,28 +1079,6 @@ gf100_grctx_generate_unkn(struct gf100_gr *gr)
 {
 }
 
-void
-gf100_grctx_generate_tpcid(struct gf100_gr *gr)
-{
-	struct nvkm_device *device = gr->base.engine.subdev.device;
-	int gpc, tpc, id;
-
-	for (tpc = 0, id = 0; tpc < 4; tpc++) {
-		for (gpc = 0; gpc < gr->gpc_nr; gpc++) {
-			if (tpc < gr->tpc_nr[gpc]) {
-				nvkm_wr32(device, TPC_UNIT(gpc, tpc, 0x698), id);
-				nvkm_wr32(device, TPC_UNIT(gpc, tpc, 0x4e8), id);
-				nvkm_wr32(device, GPC_UNIT(gpc, 0x0c10 + tpc * 4), id);
-				nvkm_wr32(device, TPC_UNIT(gpc, tpc, 0x088), id);
-				id++;
-			}
-
-			nvkm_wr32(device, GPC_UNIT(gpc, 0x0c08), gr->tpc_nr[gpc]);
-			nvkm_wr32(device, GPC_UNIT(gpc, 0x0c8c), gr->tpc_nr[gpc]);
-		}
-	}
-}
-
 void
 gf100_grctx_generate_r406028(struct gf100_gr *gr)
 {
@@ -1230,6 +1208,40 @@ gf100_grctx_generate_r406800(struct gf100_gr *gr)
 	}
 }
 
+void
+gf100_grctx_generate_tpc_nr(struct gf100_gr *gr, int gpc)
+{
+	struct nvkm_device *device = gr->base.engine.subdev.device;
+	nvkm_wr32(device, GPC_UNIT(gpc, 0x0c08), gr->tpc_nr[gpc]);
+	nvkm_wr32(device, GPC_UNIT(gpc, 0x0c8c), gr->tpc_nr[gpc]);
+}
+
+void
+gf100_grctx_generate_sm_id(struct gf100_gr *gr, int gpc, int tpc, int sm)
+{
+	struct nvkm_device *device = gr->base.engine.subdev.device;
+	nvkm_wr32(device, TPC_UNIT(gpc, tpc, 0x698), sm);
+	nvkm_wr32(device, TPC_UNIT(gpc, tpc, 0x4e8), sm);
+	nvkm_wr32(device, GPC_UNIT(gpc, 0x0c10 + tpc * 4), sm);
+	nvkm_wr32(device, TPC_UNIT(gpc, tpc, 0x088), sm);
+}
+
+void
+gf100_grctx_generate_floorsweep(struct gf100_gr *gr)
+{
+	const struct gf100_grctx_func *func = gr->func->grctx;
+	int tpc, gpc, sm;
+
+	for (tpc = 0, sm = 0; tpc < gr->tpc_max; tpc++) {
+		for (gpc = 0; gpc < gr->gpc_nr; gpc++) {
+			if (tpc < gr->tpc_nr[gpc])
+				func->sm_id(gr, gpc, tpc, sm++);
+			if (func->tpc_nr)
+				func->tpc_nr(gr, gpc);
+		}
+	}
+}
+
 void
 gf100_grctx_generate_main(struct gf100_gr *gr, struct gf100_grctx *info)
 {
@@ -1258,7 +1270,7 @@ gf100_grctx_generate_main(struct gf100_gr *gr, struct gf100_grctx *info)
 		grctx->patch_ltc(info);
 	grctx->unkn(gr);
 
-	gf100_grctx_generate_tpcid(gr);
+	gf100_grctx_generate_floorsweep(gr);
 	gf100_grctx_generate_r406028(gr);
 	gf100_grctx_generate_r4060a8(gr);
 	gf100_grctx_generate_r418bb8(gr);
@@ -1410,4 +1422,6 @@ gf100_grctx = {
 	.attrib = gf100_grctx_generate_attrib,
 	.attrib_nr_max = 0x324,
 	.attrib_nr = 0x218,
+	.sm_id = gf100_grctx_generate_sm_id,
+	.tpc_nr = gf100_grctx_generate_tpc_nr,
 };

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

@@ -50,6 +50,9 @@ struct gf100_grctx_func {
 	u32 alpha_nr;
 	/* other patch buffer stuff */
 	void (*patch_ltc)(struct gf100_grctx *);
+	/* floorsweeping */
+	void (*sm_id)(struct gf100_gr *, int gpc, int tpc, int sm);
+	void (*tpc_nr)(struct gf100_gr *, int gpc);
 };
 
 extern const struct gf100_grctx_func gf100_grctx;
@@ -59,11 +62,13 @@ void gf100_grctx_generate_bundle(struct gf100_grctx *);
 void gf100_grctx_generate_pagepool(struct gf100_grctx *);
 void gf100_grctx_generate_attrib(struct gf100_grctx *);
 void gf100_grctx_generate_unkn(struct gf100_gr *);
-void gf100_grctx_generate_tpcid(struct gf100_gr *);
+void gf100_grctx_generate_floorsweep(struct gf100_gr *);
 void gf100_grctx_generate_r406028(struct gf100_gr *);
 void gf100_grctx_generate_r4060a8(struct gf100_gr *);
 void gf100_grctx_generate_r418bb8(struct gf100_gr *);
 void gf100_grctx_generate_r406800(struct gf100_gr *);
+void gf100_grctx_generate_sm_id(struct gf100_gr *, int, int, int);
+void gf100_grctx_generate_tpc_nr(struct gf100_gr *, int);
 
 extern const struct gf100_grctx_func gf108_grctx;
 void gf108_grctx_generate_attrib(struct gf100_grctx *);
@@ -89,6 +94,7 @@ void gk104_grctx_generate_r418bb8(struct gf100_gr *);
 void gm107_grctx_generate_bundle(struct gf100_grctx *);
 void gm107_grctx_generate_pagepool(struct gf100_grctx *);
 void gm107_grctx_generate_attrib(struct gf100_grctx *);
+void gm107_grctx_generate_sm_id(struct gf100_gr *, int, int, int);
 
 extern const struct gf100_grctx_func gk110_grctx;
 extern const struct gf100_grctx_func gk110b_grctx;
@@ -100,7 +106,6 @@ void gm107_grctx_generate_pagepool(struct gf100_grctx *);
 void gm107_grctx_generate_attrib(struct gf100_grctx *);
 
 extern const struct gf100_grctx_func gm200_grctx;
-void gm200_grctx_generate_tpcid(struct gf100_gr *);
 void gm200_grctx_generate_405b60(struct gf100_gr *);
 
 extern const struct gf100_grctx_func gm20b_grctx;

drivers/gpu/drm/nouveau/nvkm/engine/gr/ctxgf104.c

@@ -96,4 +96,6 @@ gf104_grctx = {
 	.attrib = gf100_grctx_generate_attrib,
 	.attrib_nr_max = 0x324,
 	.attrib_nr = 0x218,
+	.sm_id = gf100_grctx_generate_sm_id,
+	.tpc_nr = gf100_grctx_generate_tpc_nr,
 };

drivers/gpu/drm/nouveau/nvkm/engine/gr/ctxgf108.c

@@ -794,4 +794,6 @@ gf108_grctx = {
 	.attrib_nr = 0x218,
 	.alpha_nr_max = 0x324,
 	.alpha_nr = 0x218,
+	.sm_id = gf100_grctx_generate_sm_id,
+	.tpc_nr = gf100_grctx_generate_tpc_nr,
 };

drivers/gpu/drm/nouveau/nvkm/engine/gr/ctxgf110.c

@@ -347,4 +347,6 @@ gf110_grctx = {
 	.attrib = gf100_grctx_generate_attrib,
 	.attrib_nr_max = 0x324,
 	.attrib_nr = 0x218,
+	.sm_id = gf100_grctx_generate_sm_id,
+	.tpc_nr = gf100_grctx_generate_tpc_nr,
 };

drivers/gpu/drm/nouveau/nvkm/engine/gr/ctxgf117.c

@@ -240,7 +240,7 @@ gf117_grctx_generate_main(struct gf100_gr *gr, struct gf100_grctx *info)
 	grctx->attrib(info);
 	grctx->unkn(gr);
 
-	gf100_grctx_generate_tpcid(gr);
+	gf100_grctx_generate_floorsweep(gr);
 	gf100_grctx_generate_r406028(gr);
 	gf100_grctx_generate_r4060a8(gr);
 	gk104_grctx_generate_r418bb8(gr);
@@ -275,4 +275,6 @@ gf117_grctx = {
 	.attrib_nr = 0x218,
 	.alpha_nr_max = 0x7ff,
 	.alpha_nr = 0x324,
+	.sm_id = gf100_grctx_generate_sm_id,
+	.tpc_nr = gf100_grctx_generate_tpc_nr,
 };

drivers/gpu/drm/nouveau/nvkm/engine/gr/ctxgf119.c

@@ -517,4 +517,6 @@ gf119_grctx = {
 	.attrib_nr = 0x218,
 	.alpha_nr_max = 0x324,
 	.alpha_nr = 0x218,
+	.sm_id = gf100_grctx_generate_sm_id,
+	.tpc_nr = gf100_grctx_generate_tpc_nr,
 };

drivers/gpu/drm/nouveau/nvkm/engine/gr/ctxgk104.c

@@ -974,7 +974,7 @@ gk104_grctx_generate_main(struct gf100_gr *gr, struct gf100_grctx *info)
 	grctx->attrib(info);
 	grctx->unkn(gr);
 
-	gf100_grctx_generate_tpcid(gr);
+	gf100_grctx_generate_floorsweep(gr);
 	gf100_grctx_generate_r406028(gr);
 	gk104_grctx_generate_r418bb8(gr);
 	gf100_grctx_generate_r406800(gr);
@@ -1017,4 +1017,6 @@ gk104_grctx = {
 	.alpha_nr_max = 0x7ff,
 	.alpha_nr = 0x648,
 	.patch_ltc = gk104_grctx_generate_patch_ltc,
+	.sm_id = gf100_grctx_generate_sm_id,
+	.tpc_nr = gf100_grctx_generate_tpc_nr,
 };

drivers/gpu/drm/nouveau/nvkm/engine/gr/ctxgk110.c

@@ -831,4 +831,6 @@ gk110_grctx = {
 	.alpha_nr_max = 0x7ff,
 	.alpha_nr = 0x648,
 	.patch_ltc = gk104_grctx_generate_patch_ltc,
+	.sm_id = gf100_grctx_generate_sm_id,
+	.tpc_nr = gf100_grctx_generate_tpc_nr,
 };

drivers/gpu/drm/nouveau/nvkm/engine/gr/ctxgk110b.c

@@ -92,4 +92,6 @@ gk110b_grctx = {
 	.alpha_nr_max = 0x7ff,
 	.alpha_nr = 0x648,
 	.patch_ltc = gk104_grctx_generate_patch_ltc,
+	.sm_id = gf100_grctx_generate_sm_id,
+	.tpc_nr = gf100_grctx_generate_tpc_nr,
 };

drivers/gpu/drm/nouveau/nvkm/engine/gr/ctxgk208.c

@@ -553,4 +553,6 @@ gk208_grctx = {
 	.alpha_nr_max = 0x7ff,
 	.alpha_nr = 0x648,
 	.patch_ltc = gk104_grctx_generate_patch_ltc,
+	.sm_id = gf100_grctx_generate_sm_id,
+	.tpc_nr = gf100_grctx_generate_tpc_nr,
 };

drivers/gpu/drm/nouveau/nvkm/engine/gr/ctxgk20a.c

@@ -42,7 +42,7 @@ gk20a_grctx_generate_main(struct gf100_gr *gr, struct gf100_grctx *info)
 
 	grctx->unkn(gr);
 
-	gf100_grctx_generate_tpcid(gr);
+	gf100_grctx_generate_floorsweep(gr);
 	gf100_grctx_generate_r406028(gr);
 	gk104_grctx_generate_r418bb8(gr);
 	gf100_grctx_generate_r406800(gr);
@@ -82,4 +82,6 @@ gk20a_grctx = {
 	.attrib_nr = 0x240,
 	.alpha_nr_max = 0x648 + (0x648 / 2),
 	.alpha_nr = 0x648,
+	.sm_id = gf100_grctx_generate_sm_id,
+	.tpc_nr = gf100_grctx_generate_tpc_nr,
 };

drivers/gpu/drm/nouveau/nvkm/engine/gr/ctxgm107.c

@@ -930,25 +930,13 @@ gm107_grctx_generate_attrib(struct gf100_grctx *info)
 	}
 }
 
-static void
-gm107_grctx_generate_tpcid(struct gf100_gr *gr)
+void
+gm107_grctx_generate_sm_id(struct gf100_gr *gr, int gpc, int tpc, int sm)
 {
 	struct nvkm_device *device = gr->base.engine.subdev.device;
-	int gpc, tpc, id;
-
-	for (tpc = 0, id = 0; tpc < 4; tpc++) {
-		for (gpc = 0; gpc < gr->gpc_nr; gpc++) {
-			if (tpc < gr->tpc_nr[gpc]) {
-				nvkm_wr32(device, TPC_UNIT(gpc, tpc, 0x698), id);
-				nvkm_wr32(device, GPC_UNIT(gpc, 0x0c10 + tpc * 4), id);
-				nvkm_wr32(device, TPC_UNIT(gpc, tpc, 0x088), id);
-				id++;
-			}
-
-			nvkm_wr32(device, GPC_UNIT(gpc, 0x0c08), gr->tpc_nr[gpc]);
-			nvkm_wr32(device, GPC_UNIT(gpc, 0x0c8c), gr->tpc_nr[gpc]);
-		}
-	}
+	nvkm_wr32(device, TPC_UNIT(gpc, tpc, 0x698), sm);
+	nvkm_wr32(device, GPC_UNIT(gpc, 0x0c10 + tpc * 4), sm);
+	nvkm_wr32(device, TPC_UNIT(gpc, tpc, 0x088), sm);
 }
 
 static void
@@ -972,7 +960,7 @@ gm107_grctx_generate_main(struct gf100_gr *gr, struct gf100_grctx *info)
 	grctx->attrib(info);
 	grctx->unkn(gr);
 
-	gm107_grctx_generate_tpcid(gr);
+	gf100_grctx_generate_floorsweep(gr);
 	gf100_grctx_generate_r406028(gr);
 	gk104_grctx_generate_r418bb8(gr);
 	gf100_grctx_generate_r406800(gr);
@@ -1016,4 +1004,6 @@ gm107_grctx = {
 	.attrib_nr = 0xaa0,
 	.alpha_nr_max = 0x1800,
 	.alpha_nr = 0x1000,
+	.sm_id = gm107_grctx_generate_sm_id,
+	.tpc_nr = gf100_grctx_generate_tpc_nr,
 };

drivers/gpu/drm/nouveau/nvkm/engine/gr/ctxgm200.c

@@ -27,24 +27,6 @@
  * PGRAPH context implementation
  ******************************************************************************/
 
-void
-gm200_grctx_generate_tpcid(struct gf100_gr *gr)
-{
-	struct nvkm_device *device = gr->base.engine.subdev.device;
-	int gpc, tpc, id;
-
-	for (tpc = 0, id = 0; tpc < TPC_MAX_PER_GPC; tpc++) {
-		for (gpc = 0; gpc < gr->gpc_nr; gpc++) {
-			if (tpc < gr->tpc_nr[gpc]) {
-				nvkm_wr32(device, TPC_UNIT(gpc, tpc, 0x698), id);
-				nvkm_wr32(device, GPC_UNIT(gpc, 0x0c10 + tpc * 4), id);
-				nvkm_wr32(device, TPC_UNIT(gpc, tpc, 0x088), id);
-				id++;
-			}
-		}
-	}
-}
-
 void
 gm200_grctx_generate_405b60(struct gf100_gr *gr)
 {
@@ -94,7 +76,7 @@ gm200_grctx_generate_main(struct gf100_gr *gr, struct gf100_grctx *info)
 	grctx->attrib(info);
 	grctx->unkn(gr);
 
-	gm200_grctx_generate_tpcid(gr);
+	gf100_grctx_generate_floorsweep(gr);
 	gf100_grctx_generate_r406028(gr);
 	gk104_grctx_generate_r418bb8(gr);
 
@@ -133,4 +115,5 @@ gm200_grctx = {
 	.attrib_nr = 0x400,
 	.alpha_nr_max = 0x1800,
 	.alpha_nr = 0x1000,
+	.sm_id = gm107_grctx_generate_sm_id,
 };

drivers/gpu/drm/nouveau/nvkm/engine/gr/ctxgm20b.c

@@ -53,7 +53,7 @@ gm20b_grctx_generate_main(struct gf100_gr *gr, struct gf100_grctx *info)
 
 	grctx->unkn(gr);
 
-	gm200_grctx_generate_tpcid(gr);
+	gf100_grctx_generate_floorsweep(gr);
 	gm20b_grctx_generate_r406028(gr);
 	gk104_grctx_generate_r418bb8(gr);
 
@@ -98,4 +98,5 @@ gm20b_grctx = {
 	.attrib_nr = 0x400,
 	.alpha_nr_max = 0xc00,
 	.alpha_nr = 0x800,
+	.sm_id = gm107_grctx_generate_sm_id,
 };

drivers/gpu/drm/nouveau/nvkm/engine/gr/ctxgp100.c

@@ -138,7 +138,7 @@ gp100_grctx_generate_main(struct gf100_gr *gr, struct gf100_grctx *info)
 	grctx->attrib(info);
 	grctx->unkn(gr);
 
-	gm200_grctx_generate_tpcid(gr);
+	gf100_grctx_generate_floorsweep(gr);
 	gf100_grctx_generate_r406028(gr);
 	gk104_grctx_generate_r418bb8(gr);
 
@@ -174,4 +174,5 @@ gp100_grctx = {
 	.attrib_nr = 0x440,
 	.alpha_nr_max = 0xc00,
 	.alpha_nr = 0x800,
+	.sm_id = gm107_grctx_generate_sm_id,
 };

drivers/gpu/drm/nouveau/nvkm/engine/gr/ctxgp102.c

@@ -94,4 +94,5 @@ gp102_grctx = {
 	.attrib_nr = 0x320,
 	.alpha_nr_max = 0xc00,
 	.alpha_nr = 0x800,
+	.sm_id = gm107_grctx_generate_sm_id,
 };

drivers/gpu/drm/nouveau/nvkm/engine/gr/ctxgp107.c

@@ -44,4 +44,5 @@ gp107_grctx = {
 	.attrib_nr = 0x540,
 	.alpha_nr_max = 0xc00,
 	.alpha_nr = 0x800,
+	.sm_id = gm107_grctx_generate_sm_id,
 };

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

@@ -1675,6 +1675,7 @@ gf100_gr_oneinit(struct nvkm_gr *base)
 	gr->gpc_nr = nvkm_rd32(device, 0x409604) & 0x0000001f;
 	for (i = 0; i < gr->gpc_nr; i++) {
 		gr->tpc_nr[i]  = nvkm_rd32(device, GPC_UNIT(i, 0x2608));
+		gr->tpc_max = max(gr->tpc_max, gr->tpc_nr[i]);
 		gr->tpc_total += gr->tpc_nr[i];
 		gr->ppc_nr[i]  = gr->func->ppc_nr;
 		for (j = 0; j < gr->ppc_nr[i]; j++) {

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

@@ -99,6 +99,7 @@ struct gf100_gr {
 	u8 rop_nr;
 	u8 gpc_nr;
 	u8 tpc_nr[GPC_MAX];
+	u8 tpc_max;
 	u8 tpc_total;
 	u8 ppc_nr[GPC_MAX];
 	u8 ppc_mask[GPC_MAX];