nv50,nvc0: fix texture layer issues

9 files changed, 78 insertions, 62 deletions

diff --git a/src/gallium/drivers/nvc0/nvc0_3d.xml.h b/src/gallium/drivers/nvc0/nvc0_3d.xml.h
index 73a605f94e..5857f7cee9 100644
--- a/src/gallium/drivers/nvc0/nvc0_3d.xml.h
+++ b/src/gallium/drivers/nvc0/nvc0_3d.xml.h
@@ -154,8 +154,8 @@ WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
 
 #define NVC0_3D_LOCAL_SIZE_LOW					0x0000079c
 
-#define NVC0_3D_RT(i0)					       (0x00000800 + 0x20*(i0))
-#define NVC0_3D_RT__ESIZE					0x00000020
+#define NVC0_3D_RT(i0)					       (0x00000800 + 0x40*(i0))
+#define NVC0_3D_RT__ESIZE					0x00000040
 #define NVC0_3D_RT__LEN						0x00000008
 
 #define NVC0_3D_RT_ADDRESS_HIGH(i0)			       (0x00000800 + 0x40*(i0))
@@ -169,11 +169,13 @@ WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
 #define NVC0_3D_RT_FORMAT(i0)				       (0x00000810 + 0x40*(i0))
 
 #define NVC0_3D_RT_TILE_MODE(i0)			       (0x00000814 + 0x40*(i0))
-#define NVC0_3D_RT_TILE_MODE_UNK0				0x00000001
+#define NVC0_3D_RT_TILE_MODE_X					0x00000001
 #define NVC0_3D_RT_TILE_MODE_Y__MASK				0x00000070
 #define NVC0_3D_RT_TILE_MODE_Y__SHIFT				4
 #define NVC0_3D_RT_TILE_MODE_Z__MASK				0x00000700
 #define NVC0_3D_RT_TILE_MODE_Z__SHIFT				8
+#define NVC0_3D_RT_TILE_MODE_LINEAR				0x00001000
+#define NVC0_3D_RT_TILE_MODE_UNK16				0x00010000
 
 #define NVC0_3D_RT_ARRAY_MODE(i0)			       (0x00000818 + 0x40*(i0))
 #define NVC0_3D_RT_ARRAY_MODE_LAYERS__MASK			0x0000ffff
@@ -182,6 +184,10 @@ WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
 
 #define NVC0_3D_RT_LAYER_STRIDE(i0)			       (0x0000081c + 0x40*(i0))
 
+#define NVC0_3D_RT_BASE_LAYER(i0)			       (0x00000820 + 0x40*(i0))
+
+#define NVC0_3D_RT_UNK14(i0)				       (0x00000824 + 0x40*(i0))
+
 #define NVC0_3D_VIEWPORT_SCALE_X(i0)			       (0x00000a00 + 0x20*(i0))
 #define NVC0_3D_VIEWPORT_SCALE_X__ESIZE				0x00000020
 #define NVC0_3D_VIEWPORT_SCALE_X__LEN				0x00000010
@@ -795,8 +801,10 @@ WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
 
 #define NVC0_3D_POLYGON_OFFSET_UNITS				0x000015bc
 
-#define NVC0_3D_GP_BUILTIN_RESULT_EN				0x000015cc
-#define NVC0_3D_GP_BUILTIN_RESULT_EN_LAYER			0x00010000
+#define NVC0_3D_LAYER						0x000015cc
+#define NVC0_3D_LAYER_IDX__MASK					0x0000ffff
+#define NVC0_3D_LAYER_IDX__SHIFT				0
+#define NVC0_3D_LAYER_USE_GP					0x00010000
 
 #define NVC0_3D_MULTISAMPLE_MODE				0x000015d0
 #define NVC0_3D_MULTISAMPLE_MODE_1X				0x00000000
@@ -917,6 +925,8 @@ WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
 #define NVC0_3D_POLYGON_STIPPLE_PATTERN__ESIZE			0x00000004
 #define NVC0_3D_POLYGON_STIPPLE_PATTERN__LEN			0x00000020
 
+#define NVC0_3D_ZETA_BASE_LAYER					0x0000179c
+
 #define NVC0_3D_STRMOUT_UNK1780(i0)			       (0x00001780 + 0x4*(i0))
 #define NVC0_3D_STRMOUT_UNK1780__ESIZE				0x00000004
 #define NVC0_3D_STRMOUT_UNK1780__LEN				0x00000004
diff --git a/src/gallium/drivers/nvc0/nvc0_miptree.c b/src/gallium/drivers/nvc0/nvc0_miptree.c
index ea3ed9e022..db9117c3ff 100644
--- a/src/gallium/drivers/nvc0/nvc0_miptree.c
+++ b/src/gallium/drivers/nvc0/nvc0_miptree.c
@@ -57,17 +57,25 @@ get_tile_dims(unsigned nx, unsigned ny, unsigned nz)
    return tile_mode | 0x100;
 }
 
-static INLINE unsigned
-calc_zslice_offset(uint32_t tile_mode, unsigned z, unsigned pitch, unsigned nbh)
+uint32_t
+nvc0_miptree_zslice_offset(struct nvc0_miptree *mt, unsigned l, unsigned z)
 {
-   unsigned tile_h = NVC0_TILE_HEIGHT(tile_mode);
-   unsigned tile_d_shift = NVC0_TILE_DIM_SHIFT(tile_mode, 2);
+   unsigned nblocksy; /* height of texture level aligned to tile height */
+
+   unsigned stride_2d; /* to next slice within a 3D tile */
+   unsigned stride_3d; /* to slice in the next (in z direction !) 3D tile */
+
+   unsigned tile_d_shift = NVC0_TILE_DIM_SHIFT(mt->level[l].tile_mode, 2);
    unsigned tile_d = 1 << tile_d_shift;
 
-   /* stride_2d == to next slice within this volume tile */
-   /* stride_3d == size (in bytes) of a volume tile */
-   unsigned stride_2d = tile_h * NVC0_TILE_PITCH(tile_mode);
-   unsigned stride_3d = tile_d * align(nbh, tile_h) * pitch;
+   nblocksy = util_format_get_nblocksy(mt->base.base.format,
+                                       u_minify(mt->base.base.height0, l));
+
+   nblocksy = align(nblocksy, NVC0_TILE_HEIGHT(mt->level[l].tile_mode));
+
+   stride_2d = NVC0_TILE_SIZE_2D(mt->level[l].tile_mode);
+
+   stride_3d = (nblocksy * mt->level[l].pitch) << tile_d_shift;
 
    return (z & (tile_d - 1)) * stride_2d + (z >> tile_d_shift) * stride_3d;
 }
@@ -298,21 +306,6 @@ nvc0_miptree_surface_new(struct pipe_context *pipe,
    ps->width = ns->width;
    ps->height = ns->height;
 
-   if (mt->layout_3d) {
-      unsigned zslice = ps->u.tex.first_layer;
-
-      /* TODO: re-layout the texture to use only depth 1 tiles in this case: */
-      if (ns->depth > 1 && (zslice & (NVC0_TILE_DEPTH(lvl->tile_mode) - 1)))
-         NOUVEAU_ERR("Creating unsupported 3D surface of slices [%u:%u].\n",
-                     zslice, ps->u.tex.last_layer);
-
-      ns->offset += calc_zslice_offset(lvl->tile_mode, zslice, lvl->pitch,
-                                       util_format_get_nblocksy(pt->format,
-                                                                ns->height));
-   } else {
-      ns->offset += mt->layer_stride * ps->u.tex.first_layer;
-   }
-
    return ps;
 }
 
diff --git a/src/gallium/drivers/nvc0/nvc0_resource.h b/src/gallium/drivers/nvc0/nvc0_resource.h
index 7821db51b7..f1c445b515 100644
--- a/src/gallium/drivers/nvc0/nvc0_resource.h
+++ b/src/gallium/drivers/nvc0/nvc0_resource.h
@@ -69,4 +69,7 @@ nvc0_miptree_surface_new(struct pipe_context *,
 void
 nvc0_miptree_surface_del(struct pipe_context *, struct pipe_surface *);
 
+uint32_t
+nvc0_miptree_zslice_offset(struct nvc0_miptree *, unsigned l, unsigned z);
+
 #endif
diff --git a/src/gallium/drivers/nvc0/nvc0_screen.c b/src/gallium/drivers/nvc0/nvc0_screen.c
index 923bb83e8a..bf4e796adf 100644
--- a/src/gallium/drivers/nvc0/nvc0_screen.c
+++ b/src/gallium/drivers/nvc0/nvc0_screen.c
@@ -582,7 +582,7 @@ nvc0_screen_create(struct pipe_winsys *ws, struct nouveau_device *dev)
    OUT_RING  (chan, 1);
    BEGIN_RING(chan, RING_3D(GP_SELECT), 1);
    OUT_RING  (chan, 0x40);
-   BEGIN_RING(chan, RING_3D(GP_BUILTIN_RESULT_EN), 1);
+   BEGIN_RING(chan, RING_3D(LAYER), 1);
    OUT_RING  (chan, 0);
    BEGIN_RING(chan, RING_3D(TEP_SELECT), 1);
    OUT_RING  (chan, 0x30);
diff --git a/src/gallium/drivers/nvc0/nvc0_state_validate.c b/src/gallium/drivers/nvc0/nvc0_state_validate.c
index ab8119a3bb..6fd880829e 100644
--- a/src/gallium/drivers/nvc0/nvc0_state_validate.c
+++ b/src/gallium/drivers/nvc0/nvc0_state_validate.c
@@ -76,7 +76,7 @@ nvc0_validate_fb(struct nvc0_context *nvc0)
         struct nouveau_bo *bo = mt->base.bo;
         uint32_t offset = sf->offset;
 
-        BEGIN_RING(chan, RING_3D(RT_ADDRESS_HIGH(i)), 8);
+        BEGIN_RING(chan, RING_3D(RT_ADDRESS_HIGH(i)), 9);
         OUT_RELOCh(chan, bo, offset, NOUVEAU_BO_VRAM | NOUVEAU_BO_RDWR);
         OUT_RELOCl(chan, bo, offset, NOUVEAU_BO_VRAM | NOUVEAU_BO_RDWR);
         OUT_RING  (chan, sf->width);
@@ -84,8 +84,9 @@ nvc0_validate_fb(struct nvc0_context *nvc0)
         OUT_RING  (chan, nvc0_format_table[sf->base.format].rt);
         OUT_RING  (chan, (mt->layout_3d << 16) |
                    mt->level[sf->base.u.tex.level].tile_mode);
-        OUT_RING  (chan, sf->depth);
+        OUT_RING  (chan, sf->base.u.tex.first_layer + sf->depth);
         OUT_RING  (chan, mt->layer_stride >> 2);
+        OUT_RING  (chan, sf->base.u.tex.first_layer);
 
         if (mt->base.status & NOUVEAU_BUFFER_STATUS_GPU_READING)
            serialize = TRUE;
@@ -115,7 +116,10 @@ nvc0_validate_fb(struct nvc0_context *nvc0)
         BEGIN_RING(chan, RING_3D(ZETA_HORIZ), 3);
         OUT_RING  (chan, sf->width);
         OUT_RING  (chan, sf->height);
-        OUT_RING  (chan, (unk << 16) | sf->depth);
+        OUT_RING  (chan, (unk << 16) |
+                   (sf->base.u.tex.first_layer + sf->depth));
+        BEGIN_RING(chan, RING_3D(ZETA_BASE_LAYER), 1);
+        OUT_RING  (chan, sf->base.u.tex.first_layer);
 
         if (mt->base.status & NOUVEAU_BUFFER_STATUS_GPU_READING)
            serialize = TRUE;
diff --git a/src/gallium/drivers/nvc0/nvc0_stateobj.h b/src/gallium/drivers/nvc0/nvc0_stateobj.h
index 57566128ab..c0c77a2ee1 100644
--- a/src/gallium/drivers/nvc0/nvc0_stateobj.h
+++ b/src/gallium/drivers/nvc0/nvc0_stateobj.h
@@ -34,7 +34,7 @@ nvc0_tsc_entry(void *hwcso)
 struct nvc0_tic_entry {
    struct pipe_sampler_view pipe;
    int id;
-   uint32_t tic[8];
+   uint32_t tic[8]; /* tic[1] (low 32 bit of address) is used for offset */
 };
 
 static INLINE struct nvc0_tic_entry *
diff --git a/src/gallium/drivers/nvc0/nvc0_surface.c b/src/gallium/drivers/nvc0/nvc0_surface.c
index a4b2b94812..67991a0b5a 100644
--- a/src/gallium/drivers/nvc0/nvc0_surface.c
+++ b/src/gallium/drivers/nvc0/nvc0_surface.c
@@ -91,14 +91,18 @@ nvc0_2d_texture_set(struct nouveau_channel *chan, int dst,
 
    width = u_minify(mt->base.base.width0, level);
    height = u_minify(mt->base.base.height0, level);
+   depth = u_minify(mt->base.base.depth0, level);
+
+   /* layer has to be < depth, and depth > tile depth / 2 */
 
-   offset = mt->level[level].offset;
    if (!mt->layout_3d) {
       offset += mt->layer_stride * layer;
+      layer = 0;
       depth = 1;
+   } else
+   if (!dst) {
+      offset += nvc0_miptree_zslice_offset(mt, level, layer);
       layer = 0;
-   } else {
-      depth = u_minify(mt->base.base.depth0, level);
    }
 
    if (!(bo->tile_flags & NOUVEAU_BO_TILE_LAYOUT_MASK)) {
@@ -233,15 +237,17 @@ nvc0_clear_render_target(struct pipe_context *pipe,
 
 	BEGIN_RING(chan, RING_3D(RT_CONTROL), 1);
 	OUT_RING  (chan, 1);
-	BEGIN_RING(chan, RING_3D(RT_ADDRESS_HIGH(0)), 8);
+	BEGIN_RING(chan, RING_3D(RT_ADDRESS_HIGH(0)), 9);
 	OUT_RELOCh(chan, bo, sf->offset, NOUVEAU_BO_VRAM | NOUVEAU_BO_WR);
 	OUT_RELOCl(chan, bo, sf->offset, NOUVEAU_BO_VRAM | NOUVEAU_BO_WR);
 	OUT_RING  (chan, sf->width);
 	OUT_RING  (chan, sf->height);
 	OUT_RING  (chan, nvc0_format_table[dst->format].rt);
-	OUT_RING  (chan, mt->level[sf->base.u.tex.level].tile_mode);
-	OUT_RING  (chan, 1);
-	OUT_RING  (chan, 0);
+	OUT_RING  (chan, (mt->layout_3d << 16) |
+              mt->level[sf->base.u.tex.level].tile_mode);
+	OUT_RING  (chan, dst->u.tex.first_layer + sf->depth);
+	OUT_RING  (chan, mt->layer_stride >> 2);
+	OUT_RING  (chan, dst->u.tex.first_layer);
 
 	BEGIN_RING(chan, RING_3D(CLIP_RECT_HORIZ(0)), 2);
 	OUT_RING  (chan, ((dstx + width) << 16) | dstx);
@@ -272,6 +278,7 @@ nvc0_clear_depth_stencil(struct pipe_context *pipe,
 	struct nvc0_surface *sf = nvc0_surface(dst);
 	struct nouveau_bo *bo = mt->base.bo;
 	uint32_t mode = 0;
+	int unk = mt->base.base.target == PIPE_TEXTURE_2D;
 
 	if (clear_flags & PIPE_CLEAR_DEPTH) {
 		BEGIN_RING(chan, RING_3D(CLEAR_DEPTH), 1);
@@ -293,13 +300,15 @@ nvc0_clear_depth_stencil(struct pipe_context *pipe,
 	OUT_RELOCl(chan, bo, sf->offset, NOUVEAU_BO_VRAM | NOUVEAU_BO_WR);
 	OUT_RING  (chan, nvc0_format_table[dst->format].rt);
 	OUT_RING  (chan, mt->level[sf->base.u.tex.level].tile_mode);
-	OUT_RING  (chan, 0);
+	OUT_RING  (chan, mt->layer_stride >> 2);
 	BEGIN_RING(chan, RING_3D(ZETA_ENABLE), 1);
 	OUT_RING  (chan, 1);
 	BEGIN_RING(chan, RING_3D(ZETA_HORIZ), 3);
 	OUT_RING  (chan, sf->width);
 	OUT_RING  (chan, sf->height);
-	OUT_RING  (chan, (1 << 16) | 1);
+	OUT_RING  (chan, (unk << 16) | (dst->u.tex.first_layer + sf->depth));
+	BEGIN_RING(chan, RING_3D(ZETA_BASE_LAYER), 1);
+	OUT_RING  (chan, dst->u.tex.first_layer);
 
 	BEGIN_RING(chan, RING_3D(CLIP_RECT_HORIZ(0)), 2);
 	OUT_RING  (chan, ((dstx + width) << 16) | dstx);
diff --git a/src/gallium/drivers/nvc0/nvc0_tex.c b/src/gallium/drivers/nvc0/nvc0_tex.c
index 6822e597b3..c52c76597a 100644
--- a/src/gallium/drivers/nvc0/nvc0_tex.c
+++ b/src/gallium/drivers/nvc0/nvc0_tex.c
@@ -89,7 +89,7 @@ nvc0_create_sampler_view(struct pipe_context *pipe,
       (swz[2] << NV50_TIC_0_MAPB__SHIFT) |
       (swz[3] << NV50_TIC_0_MAPA__SHIFT);
 
-   /* tic[1] = mt->base.bo->offset; */
+   tic[1] = /* mt->base.bo->offset; */ 0;
    tic[2] = /* mt->base.bo->offset >> 32 */ 0;
 
    tic[2] |= 0x10001000 | /* NV50_TIC_2_NO_BORDER */ 0x40000000;
@@ -106,6 +106,13 @@ nvc0_create_sampler_view(struct pipe_context *pipe,
 
    depth = MAX2(mt->base.base.array_size, mt->base.base.depth0);
 
+   if (mt->base.base.target == PIPE_TEXTURE_1D_ARRAY ||
+       mt->base.base.target == PIPE_TEXTURE_2D_ARRAY) {
+      /* there doesn't seem to be a base layer field in TIC */
+      tic[1] = view->pipe.u.tex.first_layer * mt->layer_stride;
+      depth = view->pipe.u.tex.last_layer - view->pipe.u.tex.first_layer + 1;
+   }
+
    switch (mt->base.base.target) {
    case PIPE_TEXTURE_1D:
       tic[2] |= NV50_TIC_2_TARGET_1D;
@@ -177,6 +184,8 @@ nvc0_validate_tic(struct nvc0_context *nvc0, int s)
       res = &nvc0_miptree(tic->pipe.texture)->base;
 
       if (tic->id < 0) {
+         uint32_t offset = tic->tic[1];
+
          tic->id = nvc0_screen_tic_alloc(nvc0->screen, tic);
 
          MARK_RING (chan, 9 + 8, 4);
@@ -190,8 +199,8 @@ nvc0_validate_tic(struct nvc0_context *nvc0, int s)
          OUT_RING  (chan, 0x100111);
          BEGIN_RING_NI(chan, RING_MF(DATA), 8);
          OUT_RING  (chan, tic->tic[0]);
-         OUT_RELOCl(chan, res->bo, 0, NOUVEAU_BO_VRAM | NOUVEAU_BO_RD);
-         OUT_RELOC (chan, res->bo, 0, NOUVEAU_BO_VRAM | NOUVEAU_BO_RD |
+         OUT_RELOCl(chan, res->bo, offset, NOUVEAU_BO_VRAM | NOUVEAU_BO_RD);
+         OUT_RELOC (chan, res->bo, offset, NOUVEAU_BO_VRAM | NOUVEAU_BO_RD |
                     NOUVEAU_BO_HIGH | NOUVEAU_BO_OR, tic->tic[2], tic->tic[2]);
          OUT_RINGp (chan, &tic->tic[3], 5);
 
diff --git a/src/gallium/drivers/nvc0/nvc0_transfer.c b/src/gallium/drivers/nvc0/nvc0_transfer.c
index a38bdb8f0a..bc642669e5 100644
--- a/src/gallium/drivers/nvc0/nvc0_transfer.c
+++ b/src/gallium/drivers/nvc0/nvc0_transfer.c
@@ -243,35 +243,23 @@ nvc0_miptree_transfer_new(struct pipe_context *pctx,
    struct nvc0_miptree_level *lvl = &mt->level[level];
    struct nvc0_transfer *tx;
    uint32_t size;
-   uint32_t w, h, d, z, layer, box_h, box_y;
+   uint32_t w, h, d, z, layer;
    int ret;
 
    tx = CALLOC_STRUCT(nvc0_transfer);
    if (!tx)
       return NULL;
 
-   box_y = box->y;
-   box_h = box->height;
-
    if (mt->layout_3d) {
       z = box->z;
       d = u_minify(res->depth0, level);
       layer = 0;
-      tx->nlayers = box->depth;
    } else {
       z = 0;
       d = 1;
-      if (res->target == PIPE_TEXTURE_1D ||
-          res->target == PIPE_TEXTURE_1D_ARRAY) {
-         box_y = 0;
-         box_h = 1;
-         layer = box->y;
-         tx->nlayers = box->height;
-      } else {
-         layer = box->z;
-         tx->nlayers = box->depth;
-      }
+      layer = box->z;
    }
+   tx->nlayers = box->depth;
 
    pipe_resource_reference(&tx->base.resource, res);
 
@@ -280,7 +268,7 @@ nvc0_miptree_transfer_new(struct pipe_context *pctx,
    tx->base.box = *box;
 
    tx->nblocksx = util_format_get_nblocksx(res->format, box->width);
-   tx->nblocksy = util_format_get_nblocksy(res->format, box_h);
+   tx->nblocksy = util_format_get_nblocksy(res->format, box->height);
 
    tx->base.stride = tx->nblocksx * util_format_get_blocksize(res->format);
    tx->base.layer_stride = tx->nblocksy * tx->base.stride;
@@ -294,7 +282,7 @@ nvc0_miptree_transfer_new(struct pipe_context *pctx,
    tx->rect[0].base = lvl->offset + layer * mt->layer_stride;
    tx->rect[0].tile_mode = lvl->tile_mode;
    tx->rect[0].x = util_format_get_nblocksx(res->format, box->x);
-   tx->rect[0].y = util_format_get_nblocksy(res->format, box_y);
+   tx->rect[0].y = util_format_get_nblocksy(res->format, box->y);
    tx->rect[0].z = z;
    tx->rect[0].width = util_format_get_nblocksx(res->format, w);
    tx->rect[0].height = util_format_get_nblocksy(res->format, h);