r600g: do fine-grained depth texture flushing
[mesa.git] / src / gallium / drivers / r600 / r600_texture.c
index 923efce59e64356426ea3a82291e27decccb6bdf..c34df2daa71301e508f3cfb66155c51f2020899f 100644 (file)
@@ -37,7 +37,7 @@ static void r600_copy_to_staging_texture(struct pipe_context *ctx, struct r600_t
        struct pipe_transfer *transfer = (struct pipe_transfer*)rtransfer;
        struct pipe_resource *texture = transfer->resource;
 
-       ctx->resource_copy_region(ctx, &rtransfer->staging->b.b.b,
+       ctx->resource_copy_region(ctx, &rtransfer->staging->b.b,
                                0, 0, 0, 0, texture, transfer->level,
                                &transfer->box);
 }
@@ -50,14 +50,11 @@ static void r600_copy_from_staging_texture(struct pipe_context *ctx, struct r600
        struct pipe_resource *texture = transfer->resource;
        struct pipe_box sbox;
 
-       sbox.x = sbox.y = sbox.z = 0;
-       sbox.width = transfer->box.width;
-       sbox.height = transfer->box.height;
-       /* XXX that might be wrong */
-       sbox.depth = 1;
+       u_box_origin_2d(transfer->box.width, transfer->box.height, &sbox);
+
        ctx->resource_copy_region(ctx, texture, transfer->level,
                                  transfer->box.x, transfer->box.y, transfer->box.z,
-                                 &rtransfer->staging->b.b.b,
+                                 &rtransfer->staging->b.b,
                                  0, &sbox);
 }
 
@@ -66,7 +63,7 @@ unsigned r600_texture_get_offset(struct r600_resource_texture *rtex,
 {
        unsigned offset = rtex->offset[level];
 
-       switch (rtex->resource.b.b.b.target) {
+       switch (rtex->resource.b.b.target) {
        case PIPE_TEXTURE_3D:
        case PIPE_TEXTURE_CUBE:
        default:
@@ -163,7 +160,7 @@ static unsigned r600_texture_get_nblocksx(struct pipe_screen *screen,
                                          struct r600_resource_texture *rtex,
                                          unsigned level)
 {
-       struct pipe_resource *ptex = &rtex->resource.b.b.b;
+       struct pipe_resource *ptex = &rtex->resource.b.b;
        unsigned nblocksx, block_align, width;
        unsigned blocksize = util_format_get_blocksize(rtex->real_format);
 
@@ -183,7 +180,7 @@ static unsigned r600_texture_get_nblocksy(struct pipe_screen *screen,
                                          struct r600_resource_texture *rtex,
                                          unsigned level)
 {
-       struct pipe_resource *ptex = &rtex->resource.b.b.b;
+       struct pipe_resource *ptex = &rtex->resource.b.b;
        unsigned height, tile_height;
 
        height = mip_minify(ptex->height0, level);
@@ -208,7 +205,7 @@ static void r600_texture_set_array_mode(struct pipe_screen *screen,
                                        struct r600_resource_texture *rtex,
                                        unsigned level, unsigned array_mode)
 {
-       struct pipe_resource *ptex = &rtex->resource.b.b.b;
+       struct pipe_resource *ptex = &rtex->resource.b.b;
 
        switch (array_mode) {
        case V_0280A0_ARRAY_LINEAR_GENERAL:
@@ -237,7 +234,7 @@ static void r600_texture_set_array_mode(struct pipe_screen *screen,
 
 static int r600_init_surface(struct radeon_surface *surface,
                             const struct pipe_resource *ptex,
-                            unsigned array_mode)
+                            unsigned array_mode, bool is_transfer)
 {
        surface->npix_x = ptex->width0;
        surface->npix_y = ptex->height0;
@@ -298,7 +295,8 @@ static int r600_init_surface(struct radeon_surface *surface,
        if (ptex->bind & PIPE_BIND_SCANOUT) {
                surface->flags |= RADEON_SURF_SCANOUT;
        }
-       if (util_format_is_depth_and_stencil(ptex->format)) {
+       if ((ptex->bind & PIPE_BIND_DEPTH_STENCIL) &&
+                       util_format_is_depth_and_stencil(ptex->format) && !is_transfer) {
                surface->flags |= RADEON_SURF_ZBUFFER;
                surface->flags |= RADEON_SURF_SBUFFER;
        }
@@ -311,16 +309,11 @@ static int r600_setup_surface(struct pipe_screen *screen,
                              unsigned array_mode,
                              unsigned pitch_in_bytes_override)
 {
-       struct pipe_resource *ptex = &rtex->resource.b.b.b;
+       struct pipe_resource *ptex = &rtex->resource.b.b;
        struct r600_screen *rscreen = (struct r600_screen*)screen;
        unsigned i;
        int r;
 
-       if (util_format_is_depth_or_stencil(rtex->real_format)) {
-               rtex->surface.flags |= RADEON_SURF_ZBUFFER;
-               rtex->surface.flags |= RADEON_SURF_SBUFFER;
-       }
-
        r = rscreen->ws->surface_init(rscreen->ws, &rtex->surface);
        if (r) {
                return r;
@@ -364,7 +357,7 @@ static void r600_setup_miptree(struct pipe_screen *screen,
                               struct r600_resource_texture *rtex,
                               unsigned array_mode)
 {
-       struct pipe_resource *ptex = &rtex->resource.b.b.b;
+       struct pipe_resource *ptex = &rtex->resource.b.b;
        enum chip_class chipc = ((struct r600_screen*)screen)->chip_class;
        unsigned size, layer_size, i, offset;
        unsigned nblocksx, nblocksy;
@@ -447,8 +440,21 @@ static boolean r600_texture_get_handle(struct pipe_screen* screen,
 {
        struct r600_resource_texture *rtex = (struct r600_resource_texture*)ptex;
        struct r600_resource *resource = &rtex->resource;
+       struct radeon_surface *surface = &rtex->surface;
        struct r600_screen *rscreen = (struct r600_screen*)screen;
 
+       rscreen->ws->buffer_set_tiling(resource->buf,
+                                      NULL,
+                                      surface->level[0].mode >= RADEON_SURF_MODE_1D ?
+                                      RADEON_LAYOUT_TILED : RADEON_LAYOUT_LINEAR,
+                                      surface->level[0].mode >= RADEON_SURF_MODE_2D ?
+                                      RADEON_LAYOUT_TILED : RADEON_LAYOUT_LINEAR,
+                                      surface->bankw, surface->bankh,
+                                      surface->tile_split,
+                                      surface->stencil_tile_split,
+                                      surface->mtilea,
+                                      rtex->pitch_in_bytes[0]);
+
        return rscreen->ws->buffer_get_handle(resource->buf,
                                              rtex->pitch_in_bytes[0], whandle);
 }
@@ -501,15 +507,15 @@ r600_texture_create_object(struct pipe_screen *screen,
                return NULL;
 
        resource = &rtex->resource;
-       resource->b.b.b = *base;
-       resource->b.b.vtbl = &r600_texture_vtbl;
-       pipe_reference_init(&resource->b.b.b.reference, 1);
-       resource->b.b.b.screen = screen;
+       resource->b.b = *base;
+       resource->b.vtbl = &r600_texture_vtbl;
+       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;
 
        /* We must split depth and stencil into two separate buffers on Evergreen. */
-       if (!(base->flags & R600_RESOURCE_FLAG_TRANSFER) &&
+       if ((base->bind & PIPE_BIND_DEPTH_STENCIL) &&
            ((struct r600_screen*)screen)->chip_class >= EVERGREEN &&
            util_format_is_depth_and_stencil(base->format) &&
            !rscreen->use_surface_alloc) {
@@ -553,13 +559,15 @@ r600_texture_create_object(struct pipe_screen *screen,
        }
 
        /* only mark depth textures the HW can hit as depth textures */
-       if (util_format_is_depth_or_stencil(rtex->real_format) && permit_hardware_blit(screen, base))
+       if (util_format_is_depth_or_stencil(rtex->real_format) &&
+                       permit_hardware_blit(screen, base))
                rtex->is_depth = true;
 
        r600_setup_miptree(screen, rtex, array_mode);
        if (rscreen->use_surface_alloc) {
                rtex->surface = *surface;
-               r = r600_setup_surface(screen, rtex, array_mode, pitch_in_bytes_override);
+               r = r600_setup_surface(screen, rtex, array_mode,
+                                      pitch_in_bytes_override);
                if (r) {
                        FREE(rtex);
                        return NULL;
@@ -573,7 +581,7 @@ r600_texture_create_object(struct pipe_screen *screen,
                stencil_align = r600_get_base_alignment(screen, rtex->stencil->real_format, array_mode);
                stencil_offset = align(rtex->size, stencil_align);
 
-               for (unsigned i = 0; i <= rtex->stencil->resource.b.b.b.last_level; i++)
+               for (unsigned i = 0; i <= rtex->stencil->resource.b.b.last_level; i++)
                        rtex->stencil->offset[i] += stencil_offset;
 
                rtex->size = stencil_offset + rtex->stencil->size;
@@ -581,7 +589,7 @@ r600_texture_create_object(struct pipe_screen *screen,
 
        /* Now create the backing buffer. */
        if (!buf && alloc_bo) {
-               struct pipe_resource *ptex = &rtex->resource.b.b.b;
+               struct pipe_resource *ptex = &rtex->resource.b.b;
                unsigned base_align = r600_get_base_alignment(screen, ptex->format, array_mode);
 
                if (rscreen->use_surface_alloc) {
@@ -629,7 +637,8 @@ struct pipe_resource *r600_texture_create(struct pipe_screen *screen,
                }
        }
 
-       r = r600_init_surface(&surface, templ, array_mode);
+       r = r600_init_surface(&surface, templ, array_mode,
+                             templ->flags & R600_RESOURCE_FLAG_TRANSFER);
        if (r) {
                return NULL;
        }
@@ -710,7 +719,7 @@ struct pipe_resource *r600_texture_from_handle(struct pipe_screen *screen,
        else
                array_mode = 0;
 
-       r = r600_init_surface(&surface, templ, array_mode);
+       r = r600_init_surface(&surface, templ, array_mode, 0);
        if (r) {
                return NULL;
        }
@@ -718,14 +727,17 @@ struct pipe_resource *r600_texture_from_handle(struct pipe_screen *screen,
                                                                  stride, 0, buf, FALSE, &surface);
 }
 
-int r600_texture_depth_flush(struct pipe_context *ctx,
-                            struct pipe_resource *texture, boolean just_create)
+void r600_init_flushed_depth_texture(struct pipe_context *ctx,
+                                    struct pipe_resource *texture,
+                                    struct r600_resource_texture **staging)
 {
        struct r600_resource_texture *rtex = (struct r600_resource_texture*)texture;
        struct pipe_resource resource;
+       struct r600_resource_texture **flushed_depth_texture = staging ?
+                       staging : &rtex->flushed_depth_texture;
 
-       if (rtex->flushed_depth_texture)
-               goto out;
+       if (!staging && rtex->flushed_depth_texture)
+               return; /* it's ready */
 
        resource.target = texture->target;
        resource.format = texture->format;
@@ -735,25 +747,50 @@ int r600_texture_depth_flush(struct pipe_context *ctx,
        resource.array_size = texture->array_size;
        resource.last_level = texture->last_level;
        resource.nr_samples = texture->nr_samples;
-       resource.usage = PIPE_USAGE_DYNAMIC;
-       resource.bind = texture->bind | PIPE_BIND_DEPTH_STENCIL;
-       resource.flags = R600_RESOURCE_FLAG_TRANSFER | texture->flags;
-
-       rtex->flushed_depth_texture = (struct r600_resource_texture *)ctx->screen->resource_create(ctx->screen, &resource);
-       if (rtex->flushed_depth_texture == NULL) {
-               R600_ERR("failed to create temporary texture to hold untiled copy\n");
-               return -ENOMEM;
+       resource.usage = staging ? PIPE_USAGE_DYNAMIC : PIPE_USAGE_DEFAULT;
+       resource.bind = texture->bind & ~PIPE_BIND_DEPTH_STENCIL;
+       resource.flags = texture->flags;
+
+       if (staging)
+               resource.flags |= R600_RESOURCE_FLAG_TRANSFER;
+       else
+               rtex->dirty_db_mask = (1 << (resource.last_level+1)) - 1;
+
+       *flushed_depth_texture = (struct r600_resource_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;
        }
 
-       ((struct r600_resource_texture *)rtex->flushed_depth_texture)->is_flushing_texture = TRUE;
-out:
-       if (just_create)
-               return 0;
+       (*flushed_depth_texture)->is_flushing_texture = TRUE;
 
-       /* XXX: only do this if the depth texture has actually changed:
-        */
-       r600_blit_uncompress_depth(ctx, rtex);
-       return 0;
+}
+
+void r600_texture_depth_flush(struct pipe_context *ctx,
+                             struct pipe_resource *texture,
+                             struct r600_resource_texture **staging,
+                             unsigned first_level, unsigned last_level,
+                             unsigned first_layer, unsigned last_layer)
+{
+       struct r600_resource_texture *rtex = (struct r600_resource_texture*)texture;
+
+       r600_init_flushed_depth_texture(ctx, texture, staging);
+
+       if (staging) {
+               if (!*staging)
+                       return; /* error */
+
+               r600_blit_uncompress_depth(ctx, rtex, *staging,
+                                          first_level, last_level,
+                                          first_layer, last_layer);
+       } else {
+               if (!rtex->flushed_depth_texture)
+                       return; /* error */
+
+               r600_blit_uncompress_depth(ctx, rtex, NULL,
+                                          first_level, last_level,
+                                          first_layer, last_layer);
+       }
 }
 
 /* Needs adjustment for pixelformat:
@@ -761,7 +798,7 @@ out:
 static INLINE unsigned u_box_volume( const struct pipe_box *box )
 {
        return box->width * box->depth * box->height;
-};
+}
 
 struct pipe_transfer* r600_texture_get_transfer(struct pipe_context *ctx,
                                                struct pipe_resource *texture,
@@ -773,7 +810,6 @@ struct pipe_transfer* r600_texture_get_transfer(struct pipe_context *ctx,
        struct r600_resource_texture *rtex = (struct r600_resource_texture*)texture;
        struct pipe_resource resource;
        struct r600_transfer *trans;
-       int r;
        boolean use_staging_texture = FALSE;
 
        /* We cannot map a tiled texture directly because the data is
@@ -783,8 +819,9 @@ struct pipe_transfer* r600_texture_get_transfer(struct pipe_context *ctx,
         * the CPU is much happier reading out of cached system memory
         * than uncached VRAM.
         */
-       if (R600_TEX_IS_TILED(rtex, level))
+       if (R600_TEX_IS_TILED(rtex, level)) {
                use_staging_texture = TRUE;
+       }
 
        if ((usage & PIPE_TRANSFER_READ) && u_box_volume(box) > 1024)
                use_staging_texture = TRUE;
@@ -792,15 +829,18 @@ struct pipe_transfer* r600_texture_get_transfer(struct pipe_context *ctx,
        /* Use a staging texture for uploads if the underlying BO is busy. */
        if (!(usage & PIPE_TRANSFER_READ) &&
            (rctx->ws->cs_is_buffer_referenced(rctx->cs, rtex->resource.cs_buf, RADEON_USAGE_READWRITE) ||
-            rctx->ws->buffer_is_busy(rtex->resource.buf, RADEON_USAGE_READWRITE)))
+            rctx->ws->buffer_is_busy(rtex->resource.buf, RADEON_USAGE_READWRITE))) {
                use_staging_texture = TRUE;
+       }
 
        if (!permit_hardware_blit(ctx->screen, texture) ||
-               (texture->flags & R600_RESOURCE_FLAG_TRANSFER))
+               (texture->flags & R600_RESOURCE_FLAG_TRANSFER)) {
                use_staging_texture = FALSE;
+       }
 
-       if (use_staging_texture && (usage & PIPE_TRANSFER_MAP_DIRECTLY))
+       if (use_staging_texture && (usage & PIPE_TRANSFER_MAP_DIRECTLY)) {
                return NULL;
+       }
 
        trans = CALLOC_STRUCT(r600_transfer);
        if (trans == NULL)
@@ -814,15 +854,20 @@ struct pipe_transfer* r600_texture_get_transfer(struct pipe_context *ctx,
                */
                /* XXX: when discard is true, no need to read back from depth texture
                */
-               r = r600_texture_depth_flush(ctx, texture, FALSE);
-               if (r < 0) {
+               struct r600_resource_texture *staging_depth;
+
+               r600_texture_depth_flush(ctx, texture, &staging_depth,
+                                        level, level,
+                                        box->z, box->z + box->depth - 1);
+               if (!staging_depth) {
                        R600_ERR("failed to create temporary texture to hold untiled copy\n");
                        pipe_resource_reference(&trans->transfer.resource, NULL);
                        FREE(trans);
                        return NULL;
                }
-               trans->transfer.stride = rtex->flushed_depth_texture->pitch_in_bytes[level];
-               trans->offset = r600_texture_get_offset(rtex->flushed_depth_texture, level, box->z);
+               trans->transfer.stride = staging_depth->pitch_in_bytes[level];
+               trans->offset = r600_texture_get_offset(staging_depth, level, box->z);
+               trans->staging = (struct r600_resource*)staging_depth;
                return &trans->transfer;
        } else if (use_staging_texture) {
                resource.target = PIPE_TEXTURE_2D;
@@ -877,17 +922,24 @@ void r600_texture_transfer_destroy(struct pipe_context *ctx,
        struct pipe_resource *texture = transfer->resource;
        struct r600_resource_texture *rtex = (struct r600_resource_texture*)texture;
 
-       if (rtransfer->staging) {
+       if (rtex->is_depth) {
+               if ((transfer->usage & PIPE_TRANSFER_WRITE) && rtransfer->staging) {
+                       struct pipe_box sbox;
+
+                       u_box_origin_2d(texture->width0, texture->height0, &sbox);
+
+                       ctx->resource_copy_region(ctx, texture, 0, 0, 0, 0,
+                                                 &rtransfer->staging->b.b, 0,
+                                                 &sbox);
+               }
+       } else if (rtransfer->staging) {
                if (transfer->usage & PIPE_TRANSFER_WRITE) {
                        r600_copy_from_staging_texture(ctx, rtransfer);
                }
-               pipe_resource_reference((struct pipe_resource**)&rtransfer->staging, NULL);
        }
 
-       if (rtex->is_depth && !rtex->is_flushing_texture) {
-               if ((transfer->usage & PIPE_TRANSFER_WRITE) && rtex->flushed_depth_texture)
-                       r600_blit_push_depth(ctx, rtex);
-       }
+       if (rtransfer->staging)
+               pipe_resource_reference((struct pipe_resource**)&rtransfer->staging, NULL);
 
        pipe_resource_reference(&transfer->resource, NULL);
        FREE(transfer);
@@ -898,25 +950,27 @@ 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 pb_buffer *buf;
+       struct radeon_winsys_cs_handle *buf;
+       struct r600_resource_texture *rtex =
+                       (struct r600_resource_texture*)transfer->resource;
        enum pipe_format format = transfer->resource->format;
        unsigned offset = 0;
        char *map;
 
+       if ((transfer->resource->bind & PIPE_BIND_GLOBAL) && transfer->resource->target == PIPE_BUFFER) {
+               return r600_compute_global_transfer_map(ctx, transfer);
+       }
+
        if (rtransfer->staging) {
-               buf = ((struct r600_resource *)rtransfer->staging)->buf;
+               buf = ((struct r600_resource *)rtransfer->staging)->cs_buf;
        } else {
-               struct r600_resource_texture *rtex = (struct r600_resource_texture*)transfer->resource;
-
-               if (rtex->flushed_depth_texture)
-                       buf = ((struct r600_resource *)rtex->flushed_depth_texture)->buf;
-               else
-                       buf = ((struct r600_resource *)transfer->resource)->buf;
+               buf = ((struct r600_resource *)transfer->resource)->cs_buf;
+       }
 
+       if (rtex->is_depth || !rtransfer->staging)
                offset = rtransfer->offset +
                        transfer->box.y / util_format_get_blockheight(format) * transfer->stride +
                        transfer->box.x / util_format_get_blockwidth(format) * util_format_get_blocksize(format);
-       }
 
        if (!(map = rctx->ws->buffer_map(buf, rctx->cs, transfer->usage))) {
                return NULL;
@@ -930,18 +984,16 @@ void r600_texture_transfer_unmap(struct pipe_context *ctx,
 {
        struct r600_transfer *rtransfer = (struct r600_transfer*)transfer;
        struct r600_context *rctx = (struct r600_context*)ctx;
-       struct pb_buffer *buf;
+       struct radeon_winsys_cs_handle *buf;
+
+       if ((transfer->resource->bind & PIPE_BIND_GLOBAL) && transfer->resource->target == PIPE_BUFFER) {
+               return r600_compute_global_transfer_unmap(ctx, transfer);
+       }
 
        if (rtransfer->staging) {
-               buf = ((struct r600_resource *)rtransfer->staging)->buf;
+               buf = ((struct r600_resource *)rtransfer->staging)->cs_buf;
        } else {
-               struct r600_resource_texture *rtex = (struct r600_resource_texture*)transfer->resource;
-
-               if (rtex->flushed_depth_texture) {
-                       buf = ((struct r600_resource *)rtex->flushed_depth_texture)->buf;
-               } else {
-                       buf = ((struct r600_resource *)transfer->resource)->buf;
-               }
+               buf = ((struct r600_resource *)transfer->resource)->cs_buf;
        }
        rctx->ws->buffer_unmap(buf);
 }
@@ -1046,6 +1098,8 @@ uint32_t r600_translate_texformat(struct pipe_screen *screen,
                case PIPE_FORMAT_Z32_FLOAT:
                        result = FMT_32_FLOAT;
                        goto out_word4;
+               case PIPE_FORMAT_X32_S8X24_UINT:
+                       word4 |= S_038010_NUM_FORMAT_ALL(V_038010_SQ_NUM_FORMAT_INT);
                case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT:
                        result = FMT_X24_8_32_FLOAT;
                        goto out_word4;
@@ -1135,6 +1189,21 @@ uint32_t r600_translate_texformat(struct pipe_screen *screen,
                }
        }
 
+       if (desc->layout == UTIL_FORMAT_LAYOUT_SUBSAMPLED) {
+               switch (format) {
+               case PIPE_FORMAT_R8G8_B8G8_UNORM:
+               case PIPE_FORMAT_G8R8_B8R8_UNORM:
+                       result = FMT_GB_GR;
+                       goto out_word4;
+               case PIPE_FORMAT_G8R8_G8B8_UNORM:
+               case PIPE_FORMAT_R8G8_R8B8_UNORM:
+                       result = FMT_BG_RG;
+                       goto out_word4;
+               default:
+                       goto out_unknown;
+               }
+       }
+
        if (format == PIPE_FORMAT_R9G9B9E5_FLOAT) {
                result = FMT_5_9_9_9_SHAREDEXP;
                goto out_word4;