drm/nouveau/gr/gv100-: port smid mapping code from nvgpu

Essentially ripped verbatim from NVGPU, comments and all, and adapted to
nvkm's structs and style.

- maybe fixes an nvgpu bug though, a small tweak was needed to match RM

v2:
- remove unnecessary WARN_ON
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Reviewed-by: Lyude Paul <lyude@redhat.com>

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

@@ -157,6 +157,9 @@ static void
 gv100_grctx_generate_sm_id(struct gf100_gr *gr, int gpc, int tpc, int sm)
 {
 	struct nvkm_device *device = gr->base.engine.subdev.device;
+
+	tpc = gv100_gr_nonpes_aware_tpc(gr, gpc, tpc);
+
 	nvkm_wr32(device, TPC_UNIT(gpc, tpc, 0x608), sm);
 	nvkm_wr32(device, GPC_UNIT(gpc, 0x0c10 + tpc * 4), sm);
 	nvkm_wr32(device, TPC_UNIT(gpc, tpc, 0x088), sm);

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

@@ -34,6 +34,9 @@ static void
 tu102_grctx_generate_sm_id(struct gf100_gr *gr, int gpc, int tpc, int sm)
 {
 	struct nvkm_device *device = gr->base.engine.subdev.device;
+
+	tpc = gv100_gr_nonpes_aware_tpc(gr, gpc, tpc);
+
 	nvkm_wr32(device, TPC_UNIT(gpc, tpc, 0x608), sm);
 	nvkm_wr32(device, TPC_UNIT(gpc, tpc, 0x088), sm);
 }

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

@@ -1889,10 +1889,11 @@ gf100_gr_init_ctxctl(struct gf100_gr *gr)
 	return ret;
 }
 
-void
+int
 gf100_gr_oneinit_sm_id(struct gf100_gr *gr)
 {
 	int tpc, gpc;
+
 	for (tpc = 0; tpc < gr->tpc_max; tpc++) {
 		for (gpc = 0; gpc < gr->gpc_nr; gpc++) {
 			if (tpc < gr->tpc_nr[gpc]) {
@@ -1902,6 +1903,8 @@ gf100_gr_oneinit_sm_id(struct gf100_gr *gr)
 			}
 		}
 	}
+
+	return 0;
 }
 
 void
@@ -2019,6 +2022,8 @@ gf100_gr_oneinit(struct nvkm_gr *base)
 			if (gr->ppc_tpc_max < gr->ppc_tpc_nr[i][j])
 				gr->ppc_tpc_max = gr->ppc_tpc_nr[i][j];
 		}
+
+		gr->ppc_total += gr->ppc_nr[i];
 	}
 
 	/* Allocate global context buffers. */
@@ -2046,8 +2051,8 @@ gf100_gr_oneinit(struct nvkm_gr *base)
 
 	memset(gr->tile, 0xff, sizeof(gr->tile));
 	gr->func->oneinit_tiles(gr);
-	gr->func->oneinit_sm_id(gr);
-	return 0;
+
+	return gr->func->oneinit_sm_id(gr);
 }
 
 static int

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

@@ -107,6 +107,7 @@ struct gf100_gr {
 	u8 ppc_tpc_nr[GPC_MAX][4];
 	u8 ppc_tpc_min;
 	u8 ppc_tpc_max;
+	u8 ppc_total;
 
 	struct nvkm_memory *pagepool;
 	struct nvkm_memory *bundle_cb;
@@ -141,7 +142,7 @@ struct gf100_gr_func_zbc {
 
 struct gf100_gr_func {
 	void (*oneinit_tiles)(struct gf100_gr *);
-	void (*oneinit_sm_id)(struct gf100_gr *);
+	int (*oneinit_sm_id)(struct gf100_gr *);
 	int (*init)(struct gf100_gr *);
 	void (*init_419bd8)(struct gf100_gr *);
 	void (*init_gpc_mmu)(struct gf100_gr *);
@@ -191,7 +192,7 @@ struct gf100_gr_func {
 
 int gf100_gr_rops(struct gf100_gr *);
 void gf100_gr_oneinit_tiles(struct gf100_gr *);
-void gf100_gr_oneinit_sm_id(struct gf100_gr *);
+int gf100_gr_oneinit_sm_id(struct gf100_gr *);
 int gf100_gr_init(struct gf100_gr *);
 void gf100_gr_init_vsc_stream_master(struct gf100_gr *);
 void gf100_gr_init_zcull(struct gf100_gr *);
@@ -228,7 +229,7 @@ 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 **);
 
 void gm200_gr_oneinit_tiles(struct gf100_gr *);
-void gm200_gr_oneinit_sm_id(struct gf100_gr *);
+int gm200_gr_oneinit_sm_id(struct gf100_gr *);
 int gm200_gr_rops(struct gf100_gr *);
 void gm200_gr_init_num_active_ltcs(struct gf100_gr *);
 void gm200_gr_init_ds_hww_esr_2(struct gf100_gr *);
@@ -245,6 +246,8 @@ extern const struct gf100_gr_func_zbc gp102_gr_zbc;
 
 extern const struct gf100_gr_func gp107_gr;
 
+int gv100_gr_oneinit_sm_id(struct gf100_gr *);
+u32 gv100_gr_nonpes_aware_tpc(struct gf100_gr *gr, u32 gpc, u32 tpc);
 void gv100_gr_init_419bd8(struct gf100_gr *);
 void gv100_gr_init_504430(struct gf100_gr *, int, int);
 void gv100_gr_init_shader_exceptions(struct gf100_gr *, int, int);

drivers/gpu/drm/nouveau/nvkm/engine/gr/gm200.c

@@ -148,11 +148,11 @@ gm200_gr_tile_map_2_8[] = {
 	0, 1, 1, 0, 0, 1, 1, 0,
 };
 
-void
+int
 gm200_gr_oneinit_sm_id(struct gf100_gr *gr)
 {
 	/*XXX: There's a different algorithm here I've not yet figured out. */
-	gf100_gr_oneinit_sm_id(gr);
+	return gf100_gr_oneinit_sm_id(gr);
 }
 
 void

drivers/gpu/drm/nouveau/nvkm/engine/gr/gv100.c

@@ -85,10 +85,202 @@ gv100_gr_init_419bd8(struct gf100_gr *gr)
 	nvkm_mask(device, 0x419bd8, 0x00000700, 0x00000000);
 }
 
+u32
+gv100_gr_nonpes_aware_tpc(struct gf100_gr *gr, u32 gpc, u32 tpc)
+{
+	u32 pes, temp, tpc_new = 0;
+
+	for (pes = 0; pes < gr->ppc_nr[gpc]; pes++) {
+		if (gr->ppc_tpc_mask[gpc][pes] & BIT(tpc))
+			break;
+
+		tpc_new += gr->ppc_tpc_nr[gpc][pes];
+	}
+
+	temp = (BIT(tpc) - 1) & gr->ppc_tpc_mask[gpc][pes];
+	temp = hweight32(temp);
+	return tpc_new + temp;
+}
+
+static int
+gv100_gr_scg_estimate_perf(struct gf100_gr *gr, unsigned long *gpc_tpc_mask,
+			   u32 disable_gpc, u32 disable_tpc, int *perf)
+{
+	const u32 scale_factor = 512UL;		/* Use fx23.9 */
+	const u32 pix_scale = 1024*1024UL;	/* Pix perf in [29:20] */
+	const u32 world_scale = 1024UL;		/* World performance in [19:10] */
+	const u32 tpc_scale = 1;		/* TPC balancing in [9:0] */
+	u32 scg_num_pes = 0;
+	u32 min_scg_gpc_pix_perf = scale_factor; /* Init perf as maximum */
+	u32 average_tpcs = 0; /* Average of # of TPCs per GPC */
+	u32 deviation; /* absolute diff between TPC# and average_tpcs, averaged across GPCs */
+	u32 norm_tpc_deviation;	/* deviation/max_tpc_per_gpc */
+	u32 tpc_balance;
+	u32 scg_gpc_pix_perf;
+	u32 scg_world_perf;
+	u32 gpc;
+	u32 pes;
+	int diff;
+	bool tpc_removed_gpc = false;
+	bool tpc_removed_pes = false;
+	u32 max_tpc_gpc = 0;
+	u32 num_tpc_mask;
+	u32 *num_tpc_gpc;
+	int ret = -EINVAL;
+
+	if (!(num_tpc_gpc = kcalloc(gr->gpc_nr, sizeof(*num_tpc_gpc), GFP_KERNEL)))
+		return -ENOMEM;
+
+	/* Calculate pix-perf-reduction-rate per GPC and find bottleneck TPC */
+	for (gpc = 0; gpc < gr->gpc_nr; gpc++) {
+		num_tpc_mask = gpc_tpc_mask[gpc];
+
+		if ((gpc == disable_gpc) && num_tpc_mask & BIT(disable_tpc)) {
+			/* Safety check if a TPC is removed twice */
+			if (WARN_ON(tpc_removed_gpc))
+				goto done;
+
+			/* Remove logical TPC from set */
+			num_tpc_mask &= ~BIT(disable_tpc);
+			tpc_removed_gpc = true;
+		}
+
+		/* track balancing of tpcs across gpcs */
+		num_tpc_gpc[gpc] = hweight32(num_tpc_mask);
+		average_tpcs += num_tpc_gpc[gpc];
+
+		/* save the maximum numer of gpcs */
+		max_tpc_gpc = num_tpc_gpc[gpc] > max_tpc_gpc ? num_tpc_gpc[gpc] : max_tpc_gpc;
+
+		/*
+		 * Calculate ratio between TPC count and post-FS and post-SCG
+		 *
+		 * ratio represents relative throughput of the GPC
+		 */
+		scg_gpc_pix_perf = scale_factor * num_tpc_gpc[gpc] / gr->tpc_nr[gpc];
+		if (min_scg_gpc_pix_perf > scg_gpc_pix_perf)
+			min_scg_gpc_pix_perf = scg_gpc_pix_perf;
+
+		/* Calculate # of surviving PES */
+		for (pes = 0; pes < gr->ppc_nr[gpc]; pes++) {
+			/* Count the number of TPC on the set */
+			num_tpc_mask = gr->ppc_tpc_mask[gpc][pes] & gpc_tpc_mask[gpc];
+
+			if ((gpc == disable_gpc) && (num_tpc_mask & BIT(disable_tpc))) {
+				if (WARN_ON(tpc_removed_pes))
+					goto done;
+
+				num_tpc_mask &= ~BIT(disable_tpc);
+				tpc_removed_pes = true;
+			}
+
+			if (hweight32(num_tpc_mask))
+				scg_num_pes++;
+		}
+	}
+
+	if (WARN_ON(!tpc_removed_gpc || !tpc_removed_pes))
+		goto done;
+
+	if (max_tpc_gpc == 0) {
+		*perf = 0;
+		goto done_ok;
+	}
+
+	/* Now calculate perf */
+	scg_world_perf = (scale_factor * scg_num_pes) / gr->ppc_total;
+	deviation = 0;
+	average_tpcs = scale_factor * average_tpcs / gr->gpc_nr;
+	for (gpc = 0; gpc < gr->gpc_nr; gpc++) {
+		diff = average_tpcs - scale_factor * num_tpc_gpc[gpc];
+		if (diff < 0)
+			diff = -diff;
+
+		deviation += diff;
+	}
+
+	deviation /= gr->gpc_nr;
+
+	norm_tpc_deviation = deviation / max_tpc_gpc;
+
+	tpc_balance = scale_factor - norm_tpc_deviation;
+
+	if ((tpc_balance > scale_factor)          ||
+	    (scg_world_perf > scale_factor)       ||
+	    (min_scg_gpc_pix_perf > scale_factor) ||
+	    (norm_tpc_deviation > scale_factor)) {
+		WARN_ON(1);
+		goto done;
+	}
+
+	*perf = (pix_scale * min_scg_gpc_pix_perf) +
+		(world_scale * scg_world_perf) +
+		(tpc_scale * tpc_balance);
+done_ok:
+	ret = 0;
+done:
+	kfree(num_tpc_gpc);
+	return ret;
+}
+
+int
+gv100_gr_oneinit_sm_id(struct gf100_gr *gr)
+{
+	unsigned long *gpc_tpc_mask;
+	u32 *tpc_table, *gpc_table;
+	u32 gpc, tpc, pes, gtpc;
+	int perf, maxperf, ret = 0;
+
+	gpc_tpc_mask = kcalloc(gr->gpc_nr, sizeof(*gpc_tpc_mask), GFP_KERNEL);
+	gpc_table = kcalloc(gr->tpc_total, sizeof(*gpc_table), GFP_KERNEL);
+	tpc_table = kcalloc(gr->tpc_total, sizeof(*tpc_table), GFP_KERNEL);
+	if (!gpc_table || !tpc_table || !gpc_tpc_mask) {
+		ret = -ENOMEM;
+		goto done;
+	}
+
+	for (gpc = 0; gpc < gr->gpc_nr; gpc++) {
+		for (pes = 0; pes < gr->ppc_nr[gpc]; pes++)
+			gpc_tpc_mask[gpc] |= gr->ppc_tpc_mask[gpc][pes];
+	}
+
+	for (gtpc = 0; gtpc < gr->tpc_total; gtpc++) {
+		for (maxperf = -1, gpc = 0; gpc < gr->gpc_nr; gpc++) {
+			for_each_set_bit(tpc, &gpc_tpc_mask[gpc], gr->tpc_nr[gpc]) {
+				ret = gv100_gr_scg_estimate_perf(gr, gpc_tpc_mask, gpc, tpc, &perf);
+				if (ret)
+					goto done;
+
+				/* nvgpu does ">=" here, but this gets us RM's numbers. */
+				if (perf > maxperf) {
+					maxperf = perf;
+					gpc_table[gtpc] = gpc;
+					tpc_table[gtpc] = tpc;
+				}
+			}
+		}
+
+		gpc_tpc_mask[gpc_table[gtpc]] &= ~BIT(tpc_table[gtpc]);
+	}
+
+	/*TODO: build table for sm_per_tpc != 1, don't use yet, but might need later? */
+	for (gtpc = 0; gtpc < gr->tpc_total; gtpc++) {
+		gr->sm[gtpc].gpc = gpc_table[gtpc];
+		gr->sm[gtpc].tpc = tpc_table[gtpc];
+		gr->sm_nr++;
+	}
+
+done:
+	kfree(gpc_table);
+	kfree(tpc_table);
+	kfree(gpc_tpc_mask);
+	return ret;
+}
+
 static const struct gf100_gr_func
 gv100_gr = {
 	.oneinit_tiles = gm200_gr_oneinit_tiles,
-	.oneinit_sm_id = gm200_gr_oneinit_sm_id,
+	.oneinit_sm_id = gv100_gr_oneinit_sm_id,
 	.init = gf100_gr_init,
 	.init_419bd8 = gv100_gr_init_419bd8,
 	.init_gpc_mmu = gm200_gr_init_gpc_mmu,

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

@@ -40,8 +40,9 @@ tu102_gr_init_fs(struct gf100_gr *gr)
 	gk104_grctx_generate_gpc_tpc_nr(gr);
 
 	for (sm = 0; sm < gr->sm_nr; sm++) {
-		nvkm_wr32(device, GPC_UNIT(gr->sm[sm].gpc, 0x0c10 +
-					   gr->sm[sm].tpc * 4), sm);
+		int tpc = gv100_gr_nonpes_aware_tpc(gr, gr->sm[sm].gpc, gr->sm[sm].tpc);
+
+		nvkm_wr32(device, GPC_UNIT(gr->sm[sm].gpc, 0x0c10 + tpc * 4), sm);
 	}
 
 	gm200_grctx_generate_dist_skip_table(gr);
@@ -93,7 +94,7 @@ tu102_gr_init_gpc_mmu(struct gf100_gr *gr)
 static const struct gf100_gr_func
 tu102_gr = {
 	.oneinit_tiles = gm200_gr_oneinit_tiles,
-	.oneinit_sm_id = gm200_gr_oneinit_sm_id,
+	.oneinit_sm_id = gv100_gr_oneinit_sm_id,
 	.init = gf100_gr_init,
 	.init_419bd8 = gv100_gr_init_419bd8,
 	.init_gpc_mmu = tu102_gr_init_gpc_mmu,