r600g: atomize stencil ref state
[mesa.git] / src / gallium / drivers / r600 / r600_texture.c
index 1fdecabc1fbc331801d431cde08ba4ab62d8fcb0..6de3d6a8645aae6a2678dc295864a73e25bcb427 100644 (file)
@@ -58,10 +58,11 @@ static void r600_copy_from_staging_texture(struct pipe_context *ctx, struct r600
                                  0, &sbox);
 }
 
-unsigned r600_texture_get_offset(struct r600_resource_texture *rtex,
+unsigned r600_texture_get_offset(struct r600_texture *rtex,
                                        unsigned level, unsigned layer)
 {
-       return rtex->offset[level] + layer * rtex->layer_size[level];
+       return rtex->surface.level[level].offset +
+              layer * rtex->surface.level[level].slice_size;
 }
 
 static int r600_init_surface(struct r600_screen *rscreen,
@@ -157,7 +158,7 @@ static int r600_init_surface(struct r600_screen *rscreen,
 }
 
 static int r600_setup_surface(struct pipe_screen *screen,
-                             struct r600_resource_texture *rtex,
+                             struct r600_texture *rtex,
                              unsigned pitch_in_bytes_override)
 {
        struct pipe_resource *ptex = &rtex->resource.b.b;
@@ -182,9 +183,6 @@ static int r600_setup_surface(struct pipe_screen *screen,
                }
        }
        for (i = 0; i <= ptex->last_level; i++) {
-               rtex->offset[i] = rtex->surface.level[i].offset;
-               rtex->layer_size[i] = rtex->surface.level[i].slice_size;
-               rtex->pitch_in_bytes[i] = rtex->surface.level[i].pitch_bytes;
                switch (rtex->surface.level[i].mode) {
                case RADEON_SURF_MODE_LINEAR_ALIGNED:
                        rtex->array_mode[i] = V_038000_ARRAY_LINEAR_ALIGNED;
@@ -208,7 +206,7 @@ static boolean r600_texture_get_handle(struct pipe_screen* screen,
                                        struct pipe_resource *ptex,
                                        struct winsys_handle *whandle)
 {
-       struct r600_resource_texture *rtex = (struct r600_resource_texture*)ptex;
+       struct r600_texture *rtex = (struct r600_texture*)ptex;
        struct r600_resource *resource = &rtex->resource;
        struct radeon_surface *surface = &rtex->surface;
        struct r600_screen *rscreen = (struct r600_screen*)screen;
@@ -223,16 +221,16 @@ static boolean r600_texture_get_handle(struct pipe_screen* screen,
                                       surface->tile_split,
                                       surface->stencil_tile_split,
                                       surface->mtilea,
-                                      rtex->pitch_in_bytes[0]);
+                                      rtex->surface.level[0].pitch_bytes);
 
        return rscreen->ws->buffer_get_handle(resource->buf,
-                                             rtex->pitch_in_bytes[0], whandle);
+                                             rtex->surface.level[0].pitch_bytes, whandle);
 }
 
 static void r600_texture_destroy(struct pipe_screen *screen,
                                 struct pipe_resource *ptex)
 {
-       struct r600_resource_texture *rtex = (struct r600_resource_texture*)ptex;
+       struct r600_texture *rtex = (struct r600_texture*)ptex;
        struct r600_resource *resource = &rtex->resource;
 
        if (rtex->flushed_depth_texture)
@@ -254,21 +252,143 @@ static const struct u_resource_vtbl r600_texture_vtbl =
        NULL                            /* transfer_inline_write */
 };
 
-static struct r600_resource_texture *
+/* The number of samples can be specified independently of the texture. */
+void r600_texture_get_fmask_info(struct r600_screen *rscreen,
+                                struct r600_texture *rtex,
+                                unsigned nr_samples,
+                                struct r600_fmask_info *out)
+{
+       /* FMASK is allocated pretty much like an ordinary texture.
+        * Here we use bpe in the units of bits, not bytes. */
+       struct radeon_surface fmask = rtex->surface;
+
+       switch (nr_samples) {
+       case 2:
+               /* This should be 8,1, but we should set nsamples > 1
+                * for the allocator to treat it as a multisample surface.
+                * Let's set 4,2 then. */
+       case 4:
+               fmask.bpe = 4;
+               fmask.nsamples = 2;
+               break;
+       case 8:
+               fmask.bpe = 8;
+               fmask.nsamples = 4;
+               break;
+       case 16:
+               fmask.bpe = 16;
+               fmask.nsamples = 4;
+               break;
+       default:
+               R600_ERR("Invalid sample count for FMASK allocation.\n");
+               return;
+       }
+
+       /* R600-R700 errata? Anyway, this fixes colorbuffer corruption. */
+       if (rscreen->chip_class <= R700) {
+               fmask.bpe *= 2;
+       }
+
+       if (rscreen->chip_class >= EVERGREEN) {
+               fmask.bankh = nr_samples <= 4 ? 4 : 1;
+       }
+
+       if (rscreen->ws->surface_init(rscreen->ws, &fmask)) {
+               R600_ERR("Got error in surface_init while allocating FMASK.\n");
+               return;
+       }
+       assert(fmask.level[0].mode == RADEON_SURF_MODE_2D);
+
+       out->bank_height = fmask.bankh;
+       out->alignment = MAX2(256, fmask.bo_alignment);
+       out->size = (fmask.bo_size + 7) / 8;
+}
+
+static void r600_texture_allocate_fmask(struct r600_screen *rscreen,
+                                       struct r600_texture *rtex)
+{
+       struct r600_fmask_info fmask;
+
+       r600_texture_get_fmask_info(rscreen, rtex,
+                                   rtex->resource.b.b.nr_samples, &fmask);
+
+       /* Reserve space for FMASK while converting bits back to bytes. */
+       rtex->fmask_bank_height = fmask.bank_height;
+       rtex->fmask_offset = align(rtex->size, fmask.alignment);
+       rtex->fmask_size = fmask.size;
+       rtex->size = rtex->fmask_offset + rtex->fmask_size;
+#if 0
+       printf("FMASK width=%u, height=%i, bits=%u, size=%u\n",
+              fmask.npix_x, fmask.npix_y, fmask.bpe * fmask.nsamples, rtex->fmask_size);
+#endif
+}
+
+void r600_texture_get_cmask_info(struct r600_screen *rscreen,
+                                struct r600_texture *rtex,
+                                struct r600_cmask_info *out)
+{
+       unsigned cmask_tile_width = 8;
+       unsigned cmask_tile_height = 8;
+       unsigned cmask_tile_elements = cmask_tile_width * cmask_tile_height;
+       unsigned element_bits = 4;
+       unsigned cmask_cache_bits = 1024;
+       unsigned num_pipes = rscreen->tiling_info.num_channels;
+       unsigned pipe_interleave_bytes = rscreen->tiling_info.group_bytes;
+
+       unsigned elements_per_macro_tile = (cmask_cache_bits / element_bits) * num_pipes;
+       unsigned pixels_per_macro_tile = elements_per_macro_tile * cmask_tile_elements;
+       unsigned sqrt_pixels_per_macro_tile = sqrt(pixels_per_macro_tile);
+       unsigned macro_tile_width = util_next_power_of_two(sqrt_pixels_per_macro_tile);
+       unsigned macro_tile_height = pixels_per_macro_tile / macro_tile_width;
+
+       unsigned pitch_elements = align(rtex->surface.npix_x, macro_tile_width);
+       unsigned height = align(rtex->surface.npix_y, macro_tile_height);
+
+       unsigned base_align = num_pipes * pipe_interleave_bytes;
+       unsigned slice_bytes =
+               ((pitch_elements * height * element_bits + 7) / 8) / cmask_tile_elements;
+
+       assert(macro_tile_width % 128 == 0);
+       assert(macro_tile_height % 128 == 0);
+
+       out->slice_tile_max = ((pitch_elements * height) / (128*128)) - 1;
+       out->alignment = MAX2(256, base_align);
+       out->size = rtex->surface.array_size * align(slice_bytes, base_align);
+}
+
+static void r600_texture_allocate_cmask(struct r600_screen *rscreen,
+                                       struct r600_texture *rtex)
+{
+       struct r600_cmask_info cmask;
+
+       r600_texture_get_cmask_info(rscreen, rtex, &cmask);
+
+       rtex->cmask_slice_tile_max = cmask.slice_tile_max;
+       rtex->cmask_offset = align(rtex->size, cmask.alignment);
+       rtex->cmask_size = cmask.size;
+       rtex->size = rtex->cmask_offset + rtex->cmask_size;
+#if 0
+       printf("CMASK: macro tile width = %u, macro tile height = %u, "
+              "pitch elements = %u, height = %u, slice tile max = %u\n",
+              macro_tile_width, macro_tile_height, pitch_elements, height,
+              rtex->cmask_slice_tile_max);
+#endif
+}
+
+static struct r600_texture *
 r600_texture_create_object(struct pipe_screen *screen,
                           const struct pipe_resource *base,
-                          unsigned array_mode,
                           unsigned pitch_in_bytes_override,
                           struct pb_buffer *buf,
                           boolean alloc_bo,
                           struct radeon_surface *surface)
 {
-       struct r600_resource_texture *rtex;
+       struct r600_texture *rtex;
        struct r600_resource *resource;
        struct r600_screen *rscreen = (struct r600_screen*)screen;
        int r;
 
-       rtex = CALLOC_STRUCT(r600_resource_texture);
+       rtex = CALLOC_STRUCT(r600_texture);
        if (rtex == NULL)
                return NULL;
 
@@ -278,7 +398,6 @@ r600_texture_create_object(struct pipe_screen *screen,
        pipe_reference_init(&resource->b.b.reference, 1);
        resource->b.b.screen = screen;
        rtex->pitch_override = pitch_in_bytes_override;
-       rtex->real_format = base->format;
 
        /* don't include stencil-only formats which we don't support for rendering */
        rtex->is_depth = util_format_has_depth(util_format_description(rtex->resource.b.b.format));
@@ -291,11 +410,23 @@ r600_texture_create_object(struct pipe_screen *screen,
                return NULL;
        }
 
+       if (base->nr_samples > 1 && !rtex->is_depth && alloc_bo) {
+               r600_texture_allocate_cmask(rscreen, rtex);
+               r600_texture_allocate_fmask(rscreen, rtex);
+       }
+
+       if (!rtex->is_depth && base->nr_samples > 1 &&
+           (!rtex->fmask_size || !rtex->cmask_size)) {
+               FREE(rtex);
+               return NULL;
+       }
+
        /* Now create the backing buffer. */
        if (!buf && alloc_bo) {
                unsigned base_align = rtex->surface.bo_alignment;
+               unsigned usage = R600_TEX_IS_TILED(rtex, 0) ? PIPE_USAGE_STATIC : base->usage;
 
-               if (!r600_init_resource(rscreen, resource, rtex->size, base_align, base->bind, base->usage)) {
+               if (!r600_init_resource(rscreen, resource, rtex->size, base_align, base->bind, usage)) {
                        FREE(rtex);
                        return NULL;
                }
@@ -304,6 +435,13 @@ r600_texture_create_object(struct pipe_screen *screen,
                resource->cs_buf = rscreen->ws->buffer_get_cs_handle(buf);
                resource->domains = RADEON_DOMAIN_GTT | RADEON_DOMAIN_VRAM;
        }
+
+       if (rtex->cmask_size) {
+               /* Initialize the cmask to 0xCC (= compressed state). */
+               char *ptr = rscreen->ws->buffer_map(resource->cs_buf, NULL, PIPE_TRANSFER_WRITE);
+               memset(ptr + rtex->cmask_offset, 0xCC, rtex->cmask_size);
+               rscreen->ws->buffer_unmap(resource->cs_buf);
+       }
        return rtex;
 }
 
@@ -325,6 +463,10 @@ struct pipe_resource *r600_texture_create(struct pipe_screen *screen,
                }
        }
 
+       /* XXX tiling is broken for the 422 formats */
+       if (util_format_description(templ->format)->layout == UTIL_FORMAT_LAYOUT_SUBSAMPLED)
+               array_mode = V_038000_ARRAY_LINEAR_ALIGNED;
+
        r = r600_init_surface(rscreen, &surface, templ, array_mode,
                              templ->flags & R600_RESOURCE_FLAG_TRANSFER,
                              templ->flags & R600_RESOURCE_FLAG_FLUSHED_DEPTH);
@@ -335,7 +477,7 @@ struct pipe_resource *r600_texture_create(struct pipe_screen *screen,
        if (r) {
                return NULL;
        }
-       return (struct pipe_resource *)r600_texture_create_object(screen, templ, array_mode,
+       return (struct pipe_resource *)r600_texture_create_object(screen, templ,
                                                                  0, NULL, TRUE, &surface);
 }
 
@@ -343,6 +485,7 @@ static struct pipe_surface *r600_create_surface(struct pipe_context *pipe,
                                                struct pipe_resource *texture,
                                                const struct pipe_surface *templ)
 {
+       struct r600_texture *rtex = (struct r600_texture*)texture;
        struct r600_surface *surface = CALLOC_STRUCT(r600_surface);
        unsigned level = templ->u.tex.level;
 
@@ -353,8 +496,8 @@ static struct pipe_surface *r600_create_surface(struct pipe_context *pipe,
        pipe_resource_reference(&surface->base.texture, texture);
        surface->base.context = pipe;
        surface->base.format = templ->format;
-       surface->base.width = u_minify(texture->width0, level);
-       surface->base.height = u_minify(texture->height0, level);
+       surface->base.width = rtex->surface.level[level].npix_x;
+       surface->base.height = rtex->surface.level[level].npix_y;
        surface->base.usage = templ->usage;
        surface->base.u = templ->u;
        return &surface->base;
@@ -363,6 +506,9 @@ static struct pipe_surface *r600_create_surface(struct pipe_context *pipe,
 static void r600_surface_destroy(struct pipe_context *pipe,
                                 struct pipe_surface *surface)
 {
+       struct r600_surface *surf = (struct r600_surface*)surface;
+       pipe_resource_reference((struct pipe_resource**)&surf->cb_buffer_fmask, NULL);
+       pipe_resource_reference((struct pipe_resource**)&surf->cb_buffer_cmask, NULL);
        pipe_resource_reference(&surface->texture, NULL);
        FREE(surface);
 }
@@ -405,17 +551,17 @@ struct pipe_resource *r600_texture_from_handle(struct pipe_screen *screen,
        if (r) {
                return NULL;
        }
-       return (struct pipe_resource *)r600_texture_create_object(screen, templ, array_mode,
+       return (struct pipe_resource *)r600_texture_create_object(screen, templ,
                                                                  stride, buf, FALSE, &surface);
 }
 
 bool r600_init_flushed_depth_texture(struct pipe_context *ctx,
                                     struct pipe_resource *texture,
-                                    struct r600_resource_texture **staging)
+                                    struct r600_texture **staging)
 {
-       struct r600_resource_texture *rtex = (struct r600_resource_texture*)texture;
+       struct r600_texture *rtex = (struct r600_texture*)texture;
        struct pipe_resource resource;
-       struct r600_resource_texture **flushed_depth_texture = staging ?
+       struct r600_texture **flushed_depth_texture = staging ?
                        staging : &rtex->flushed_depth_texture;
 
        if (!staging && rtex->flushed_depth_texture)
@@ -429,14 +575,14 @@ bool r600_init_flushed_depth_texture(struct pipe_context *ctx,
        resource.array_size = texture->array_size;
        resource.last_level = texture->last_level;
        resource.nr_samples = texture->nr_samples;
-       resource.usage = staging ? PIPE_USAGE_DYNAMIC : PIPE_USAGE_DEFAULT;
+       resource.usage = staging ? PIPE_USAGE_STAGING : PIPE_USAGE_STATIC;
        resource.bind = texture->bind & ~PIPE_BIND_DEPTH_STENCIL;
        resource.flags = texture->flags | R600_RESOURCE_FLAG_FLUSHED_DEPTH;
 
        if (staging)
                resource.flags |= R600_RESOURCE_FLAG_TRANSFER;
 
-       *flushed_depth_texture = (struct r600_resource_texture *)ctx->screen->resource_create(ctx->screen, &resource);
+       *flushed_depth_texture = (struct r600_texture *)ctx->screen->resource_create(ctx->screen, &resource);
        if (*flushed_depth_texture == NULL) {
                R600_ERR("failed to create temporary texture to hold flushed depth\n");
                return false;
@@ -460,7 +606,7 @@ struct pipe_transfer* r600_texture_get_transfer(struct pipe_context *ctx,
                                                const struct pipe_box *box)
 {
        struct r600_context *rctx = (struct r600_context*)ctx;
-       struct r600_resource_texture *rtex = (struct r600_resource_texture*)texture;
+       struct r600_texture *rtex = (struct r600_texture*)texture;
        struct pipe_resource resource;
        struct r600_transfer *trans;
        boolean use_staging_texture = FALSE;
@@ -506,7 +652,7 @@ struct pipe_transfer* r600_texture_get_transfer(struct pipe_context *ctx,
                */
                /* XXX: when discard is true, no need to read back from depth texture
                */
-               struct r600_resource_texture *staging_depth;
+               struct r600_texture *staging_depth;
 
                if (!r600_init_flushed_depth_texture(ctx, texture, &staging_depth)) {
                        R600_ERR("failed to create temporary texture to hold untiled copy\n");
@@ -515,11 +661,12 @@ struct pipe_transfer* r600_texture_get_transfer(struct pipe_context *ctx,
                        return NULL;
                }
 
-               r600_blit_uncompress_depth(ctx, rtex, staging_depth,
+               r600_blit_decompress_depth(ctx, rtex, staging_depth,
                                           level, level,
-                                          box->z, box->z + box->depth - 1);
+                                          box->z, box->z + box->depth - 1,
+                                          0, 0);
 
-               trans->transfer.stride = staging_depth->pitch_in_bytes[level];
+               trans->transfer.stride = staging_depth->surface.level[level].pitch_bytes;
                trans->offset = r600_texture_get_offset(staging_depth, level, box->z);
                trans->staging = (struct r600_resource*)staging_depth;
                return &trans->transfer;
@@ -555,7 +702,7 @@ struct pipe_transfer* r600_texture_get_transfer(struct pipe_context *ctx,
                }
 
                trans->transfer.stride =
-                       ((struct r600_resource_texture *)trans->staging)->pitch_in_bytes[0];
+                       ((struct r600_texture *)trans->staging)->surface.level[0].pitch_bytes;
                if (usage & PIPE_TRANSFER_READ) {
                        r600_copy_to_staging_texture(ctx, trans);
                        /* Always referenced in the blit. */
@@ -563,8 +710,8 @@ struct pipe_transfer* r600_texture_get_transfer(struct pipe_context *ctx,
                }
                return &trans->transfer;
        }
-       trans->transfer.stride = rtex->pitch_in_bytes[level];
-       trans->transfer.layer_stride = rtex->layer_size[level];
+       trans->transfer.stride = rtex->surface.level[level].pitch_bytes;
+       trans->transfer.layer_stride = rtex->surface.level[level].slice_size;
        trans->offset = r600_texture_get_offset(rtex, level, box->z);
        return &trans->transfer;
 }
@@ -574,7 +721,7 @@ void r600_texture_transfer_destroy(struct pipe_context *ctx,
 {
        struct r600_transfer *rtransfer = (struct r600_transfer*)transfer;
        struct pipe_resource *texture = transfer->resource;
-       struct r600_resource_texture *rtex = (struct r600_resource_texture*)texture;
+       struct r600_texture *rtex = (struct r600_texture*)texture;
 
        if ((transfer->usage & PIPE_TRANSFER_WRITE) && rtransfer->staging) {
                if (rtex->is_depth) {
@@ -600,8 +747,8 @@ void* r600_texture_transfer_map(struct pipe_context *ctx,
        struct r600_context *rctx = (struct r600_context *)ctx;
        struct r600_transfer *rtransfer = (struct r600_transfer*)transfer;
        struct radeon_winsys_cs_handle *buf;
-       struct r600_resource_texture *rtex =
-                       (struct r600_resource_texture*)transfer->resource;
+       struct r600_texture *rtex =
+                       (struct r600_texture*)transfer->resource;
        enum pipe_format format = transfer->resource->format;
        unsigned offset = 0;
        char *map;