drm/nouveau/gr/gp100-: fix attrib cb setup

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

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

@@ -49,6 +49,7 @@ struct gf100_grctx_func {
 	u32 attrib_nr;
 	u32 alpha_nr_max;
 	u32 alpha_nr;
+	u32 gfxp_nr;
 	/* other patch buffer stuff */
 	void (*patch_ltc)(struct gf100_grctx *);
 	/* floorsweeping */

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

@@ -48,14 +48,17 @@ gp100_grctx_generate_attrib(struct gf100_grctx *info)
 	const struct gf100_grctx_func *grctx = gr->func->grctx;
 	const u32  alpha = grctx->alpha_nr;
 	const u32 attrib = grctx->attrib_nr;
-	const u32 pertpc = 0x20 * (grctx->attrib_nr_max + grctx->alpha_nr_max);
-	const u32   size = roundup(gr->tpc_total * pertpc, 0x80);
 	const int s = 12;
-	const int b = mmio_vram(info, size, (1 << s), false);
 	const int max_batches = 0xffff;
+	u32 size = grctx->alpha_nr_max * gr->tpc_total;
 	u32 ao = 0;
-	u32 bo = ao + grctx->alpha_nr_max * gr->tpc_total;
-	int gpc, ppc, n = 0;
+	u32 bo = ao + size;
+	int gpc, ppc, b, n = 0;
+
+	for (gpc = 0; gpc < gr->gpc_nr; gpc++)
+		size += grctx->attrib_nr_max * gr->ppc_nr[gpc] * gr->ppc_tpc_max;
+	size = ((size * 0x20) + 128) & ~127;
+	b = mmio_vram(info, size, (1 << s), false);
 
 	mmio_refn(info, 0x418810, 0x80000000, s, b);
 	mmio_refn(info, 0x419848, 0x10000000, s, b);
@@ -69,7 +72,7 @@ gp100_grctx_generate_attrib(struct gf100_grctx *info)
 	for (gpc = 0; gpc < gr->gpc_nr; gpc++) {
 		for (ppc = 0; ppc < gr->ppc_nr[gpc]; ppc++, n++) {
 			const u32 as =  alpha * gr->ppc_tpc_nr[gpc][ppc];
-			const u32 bs = attrib * gr->ppc_tpc_nr[gpc][ppc];
+			const u32 bs = attrib * gr->ppc_tpc_max;
 			const u32 u = 0x418ea0 + (n * 0x04);
 			const u32 o = PPC_UNIT(gpc, ppc, 0);
 			if (!(gr->ppc_mask[gpc] & (1 << ppc)))
@@ -77,7 +80,7 @@ gp100_grctx_generate_attrib(struct gf100_grctx *info)
 			mmio_wr32(info, o + 0xc0, bs);
 			mmio_wr32(info, o + 0xf4, bo);
 			mmio_wr32(info, o + 0xf0, bs);
-			bo += grctx->attrib_nr_max * gr->ppc_tpc_nr[gpc][ppc];
+			bo += grctx->attrib_nr_max * gr->ppc_tpc_max;
 			mmio_wr32(info, o + 0xe4, as);
 			mmio_wr32(info, o + 0xf8, ao);
 			ao += grctx->alpha_nr_max * gr->ppc_tpc_nr[gpc][ppc];

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

@@ -43,14 +43,18 @@ gp102_grctx_generate_attrib(struct gf100_grctx *info)
 	const struct gf100_grctx_func *grctx = gr->func->grctx;
 	const u32  alpha = grctx->alpha_nr;
 	const u32 attrib = grctx->attrib_nr;
-	const u32 pertpc = 0x20 * (grctx->attrib_nr_max + grctx->alpha_nr_max);
-	const u32   size = roundup(gr->tpc_total * pertpc, 0x80);
+	const u32   gfxp = grctx->gfxp_nr;
 	const int s = 12;
-	const int b = mmio_vram(info, size, (1 << s), false);
 	const int max_batches = 0xffff;
+	u32 size = grctx->alpha_nr_max * gr->tpc_total;
 	u32 ao = 0;
-	u32 bo = ao + grctx->alpha_nr_max * gr->tpc_total;
-	int gpc, ppc, n = 0;
+	u32 bo = ao + size;
+	int gpc, ppc, b, n = 0;
+
+	for (gpc = 0; gpc < gr->gpc_nr; gpc++)
+		size += grctx->gfxp_nr * gr->ppc_nr[gpc] * gr->ppc_tpc_max;
+	size = ((size * 0x20) + 128) & ~127;
+	b = mmio_vram(info, size, (1 << s), false);
 
 	mmio_refn(info, 0x418810, 0x80000000, s, b);
 	mmio_refn(info, 0x419848, 0x10000000, s, b);
@@ -64,17 +68,18 @@ gp102_grctx_generate_attrib(struct gf100_grctx *info)
 	for (gpc = 0; gpc < gr->gpc_nr; gpc++) {
 		for (ppc = 0; ppc < gr->ppc_nr[gpc]; ppc++, n++) {
 			const u32 as =  alpha * gr->ppc_tpc_nr[gpc][ppc];
-			const u32 bs = attrib * gr->ppc_tpc_nr[gpc][ppc];
+			const u32 bs = attrib * gr->ppc_tpc_max;
+			const u32 gs =   gfxp * gr->ppc_tpc_max;
 			const u32 u = 0x418ea0 + (n * 0x04);
 			const u32 o = PPC_UNIT(gpc, ppc, 0);
 			const u32 p = GPC_UNIT(gpc, 0xc44 + (ppc * 4));
 			if (!(gr->ppc_mask[gpc] & (1 << ppc)))
 				continue;
-			mmio_wr32(info, o + 0xc0, bs);
+			mmio_wr32(info, o + 0xc0, gs);
 			mmio_wr32(info, p, bs);
 			mmio_wr32(info, o + 0xf4, bo);
 			mmio_wr32(info, o + 0xf0, bs);
-			bo += grctx->attrib_nr_max * gr->ppc_tpc_nr[gpc][ppc];
+			bo += gs;
 			mmio_wr32(info, o + 0xe4, as);
 			mmio_wr32(info, o + 0xf8, ao);
 			ao += grctx->alpha_nr_max * gr->ppc_tpc_nr[gpc][ppc];
@@ -97,10 +102,11 @@ gp102_grctx = {
 	.pagepool = gp100_grctx_generate_pagepool,
 	.pagepool_size = 0x20000,
 	.attrib = gp102_grctx_generate_attrib,
-	.attrib_nr_max = 0x5d4,
+	.attrib_nr_max = 0x4b0,
 	.attrib_nr = 0x320,
 	.alpha_nr_max = 0xc00,
 	.alpha_nr = 0x800,
+	.gfxp_nr = 0xba8,
 	.sm_id = gm107_grctx_generate_sm_id,
 	.rop_mapping = gf117_grctx_generate_rop_mapping,
 	.dist_skip_table = gm200_grctx_generate_dist_skip_table,

drivers/gpu/drm/nouveau/nvkm/engine/gr/ctxgp104.c

@@ -32,10 +32,11 @@ gp104_grctx = {
 	.pagepool = gp100_grctx_generate_pagepool,
 	.pagepool_size = 0x20000,
 	.attrib = gp102_grctx_generate_attrib,
-	.attrib_nr_max = 0x5d4,
+	.attrib_nr_max = 0x4b0,
 	.attrib_nr = 0x320,
 	.alpha_nr_max = 0xc00,
 	.alpha_nr = 0x800,
+	.gfxp_nr = 0xba8,
 	.sm_id = gm107_grctx_generate_sm_id,
 	.rop_mapping = gf117_grctx_generate_rop_mapping,
 	.dist_skip_table = gm200_grctx_generate_dist_skip_table,

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

@@ -44,6 +44,7 @@ gp107_grctx = {
 	.attrib_nr = 0x540,
 	.alpha_nr_max = 0xc00,
 	.alpha_nr = 0x800,
+	.gfxp_nr = 0xe94,
 	.sm_id = gm107_grctx_generate_sm_id,
 	.rop_mapping = gf117_grctx_generate_rop_mapping,
 	.dist_skip_table = gm200_grctx_generate_dist_skip_table,

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

@@ -1780,6 +1780,8 @@ gf100_gr_oneinit(struct nvkm_gr *base)
 			if (gr->ppc_tpc_min == 0 ||
 			    gr->ppc_tpc_min > gr->ppc_tpc_nr[i][j])
 				gr->ppc_tpc_min = gr->ppc_tpc_nr[i][j];
+			if (gr->ppc_tpc_max < gr->ppc_tpc_nr[i][j])
+				gr->ppc_tpc_max = gr->ppc_tpc_nr[i][j];
 		}
 	}
 

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

@@ -106,6 +106,7 @@ struct gf100_gr {
 	u8 ppc_tpc_mask[GPC_MAX][4];
 	u8 ppc_tpc_nr[GPC_MAX][4];
 	u8 ppc_tpc_min;
+	u8 ppc_tpc_max;
 
 	u8 screen_tile_row_offset;
 	u8 tile[TPC_MAX];