freedreno: use rsc->slice accessor everywhere
[mesa.git] / src / gallium / drivers / freedreno / freedreno_resource.c
index f1e439b713385efe26511b0af39d865d3de4d3eb..fb3d6b65635d87a77137af2d911ceccb83d1bfc9 100644 (file)
  *    Rob Clark <robclark@freedesktop.org>
  */
 
-#include "util/u_format.h"
-#include "util/u_format_rgtc.h"
-#include "util/u_format_zs.h"
+#include "util/format/u_format.h"
+#include "util/format/u_format_rgtc.h"
+#include "util/format/u_format_zs.h"
 #include "util/u_inlines.h"
 #include "util/u_transfer.h"
 #include "util/u_string.h"
 #include "util/u_surface.h"
 #include "util/set.h"
+#include "util/u_drm.h"
 
 #include "freedreno_resource.h"
 #include "freedreno_batch_cache.h"
 #include "freedreno_query_hw.h"
 #include "freedreno_util.h"
 
+#include "drm-uapi/drm_fourcc.h"
 #include <errno.h>
 
 /* XXX this should go away, needed for 'struct winsys_handle' */
 #include "state_tracker/drm_driver.h"
 
+/* A private modifier for now, so we have a way to request tiled but not
+ * compressed.  It would perhaps be good to get real modifiers for the
+ * tiled formats, but would probably need to do some work to figure out
+ * the layout(s) of the tiled modes, and whether they are the same
+ * across generations.
+ */
+#define FD_FORMAT_MOD_QCOM_TILED       fourcc_mod_code(QCOM, 0xffffffff)
+
 /**
  * Go through the entire state and see if the resource is bound
  * anywhere. If it is, mark the relevant state as dirty. This is
@@ -84,6 +94,15 @@ rebind_resource(struct fd_context *ctx, struct pipe_resource *prsc)
                                ctx->dirty_shader[stage] |= FD_DIRTY_SHADER_TEX;
                }
 
+               /* Images */
+               const unsigned num_images = util_last_bit(ctx->shaderimg[stage].enabled_mask);
+               for (unsigned i = 0; i < num_images; i++) {
+                       if (ctx->dirty_shader[stage] & FD_DIRTY_SHADER_IMAGE)
+                               break;
+                       if (ctx->shaderimg[stage].si[i].resource == prsc)
+                               ctx->dirty_shader[stage] |= FD_DIRTY_SHADER_IMAGE;
+               }
+
                /* SSBOs */
                const unsigned num_ssbos = util_last_bit(ctx->shaderbuf[stage].enabled_mask);
                for (unsigned i = 0; i < num_ssbos; i++) {
@@ -125,10 +144,7 @@ do_blit(struct fd_context *ctx, const struct pipe_blit_info *blit, bool fallback
        struct pipe_context *pctx = &ctx->base;
 
        /* TODO size threshold too?? */
-       if (!fallback) {
-               /* do blit on gpu: */
-               pctx->blit(pctx, blit);
-       } else {
+       if (fallback || !fd_blit(pctx, blit)) {
                /* do blit on cpu: */
                util_resource_copy_region(pctx,
                                blit->dst.resource, blit->dst.level, blit->dst.box.x,
@@ -137,9 +153,15 @@ do_blit(struct fd_context *ctx, const struct pipe_blit_info *blit, bool fallback
        }
 }
 
+/**
+ * @rsc: the resource to shadow
+ * @level: the level to discard (if box != NULL, otherwise ignored)
+ * @box: the box to discard (or NULL if none)
+ * @modifier: the modifier for the new buffer state
+ */
 static bool
 fd_try_shadow_resource(struct fd_context *ctx, struct fd_resource *rsc,
-               unsigned level, const struct pipe_box *box)
+               unsigned level, const struct pipe_box *box, uint64_t modifier)
 {
        struct pipe_context *pctx = &ctx->base;
        struct pipe_resource *prsc = &rsc->base;
@@ -161,15 +183,16 @@ fd_try_shadow_resource(struct fd_context *ctx, struct fd_resource *rsc,
        if (prsc->target == PIPE_BUFFER)
                fallback = true;
 
-       bool whole_level = util_texrange_covers_whole_level(prsc, level,
+       bool discard_whole_level = box && util_texrange_covers_whole_level(prsc, level,
                box->x, box->y, box->z, box->width, box->height, box->depth);
 
        /* TODO need to be more clever about current level */
-       if ((prsc->target >= PIPE_TEXTURE_2D) && !whole_level)
+       if ((prsc->target >= PIPE_TEXTURE_2D) && box && !discard_whole_level)
                return false;
 
        struct pipe_resource *pshadow =
-               pctx->screen->resource_create(pctx->screen, prsc);
+               pctx->screen->resource_create_with_modifiers(pctx->screen,
+                               prsc, &modifier, 1);
 
        if (!pshadow)
                return false;
@@ -200,6 +223,10 @@ fd_try_shadow_resource(struct fd_context *ctx, struct fd_resource *rsc,
        /* TODO valid_buffer_range?? */
        swap(rsc->bo,        shadow->bo);
        swap(rsc->write_batch,   shadow->write_batch);
+       swap(rsc->offset, shadow->offset);
+       swap(rsc->ubwc_offset, shadow->ubwc_offset);
+       swap(rsc->ubwc_pitch, shadow->ubwc_pitch);
+       swap(rsc->ubwc_size, shadow->ubwc_size);
        rsc->seqno = p_atomic_inc_return(&ctx->screen->rsc_seqno);
 
        /* at this point, the newly created shadow buffer is not referenced
@@ -232,7 +259,7 @@ fd_try_shadow_resource(struct fd_context *ctx, struct fd_resource *rsc,
 
        /* blit the other levels in their entirety: */
        for (unsigned l = 0; l <= prsc->last_level; l++) {
-               if (l == level)
+               if (box && l == level)
                        continue;
 
                /* just blit whole level: */
@@ -247,7 +274,7 @@ fd_try_shadow_resource(struct fd_context *ctx, struct fd_resource *rsc,
        /* deal w/ current level specially, since we might need to split
         * it up into a couple blits:
         */
-       if (!whole_level) {
+       if (box && !discard_whole_level) {
                set_box(level, level);
 
                switch (prsc->target) {
@@ -285,6 +312,34 @@ fd_try_shadow_resource(struct fd_context *ctx, struct fd_resource *rsc,
        return true;
 }
 
+/**
+ * Uncompress an UBWC compressed buffer "in place".  This works basically
+ * like resource shadowing, creating a new resource, and doing an uncompress
+ * blit, and swapping the state between shadow and original resource so it
+ * appears to the state tracker as if nothing changed.
+ */
+void
+fd_resource_uncompress(struct fd_context *ctx, struct fd_resource *rsc)
+{
+       bool success =
+               fd_try_shadow_resource(ctx, rsc, 0, NULL, FD_FORMAT_MOD_QCOM_TILED);
+
+       /* shadow should not fail in any cases where we need to uncompress: */
+       debug_assert(success);
+
+       /*
+        * TODO what if rsc is used in other contexts, we don't currently
+        * have a good way to rebind_resource() in other contexts.  And an
+        * app that is reading one resource in multiple contexts, isn't
+        * going to expect that the resource is modified.
+        *
+        * Hopefully the edge cases where we need to uncompress are rare
+        * enough that they mostly only show up in deqp.
+        */
+
+       rebind_resource(ctx, &rsc->base);
+}
+
 static struct fd_resource *
 fd_alloc_staging(struct fd_context *ctx, struct fd_resource *rsc,
                unsigned level, const struct pipe_box *box)
@@ -294,8 +349,18 @@ fd_alloc_staging(struct fd_context *ctx, struct fd_resource *rsc,
 
        tmpl.width0  = box->width;
        tmpl.height0 = box->height;
-       tmpl.depth0  = box->depth;
-       tmpl.array_size = 1;
+       /* for array textures, box->depth is the array_size, otherwise
+        * for 3d textures, it is the depth:
+        */
+       if (tmpl.array_size > 1) {
+               if (tmpl.target == PIPE_TEXTURE_CUBE)
+                       tmpl.target = PIPE_TEXTURE_2D_ARRAY;
+               tmpl.array_size = box->depth;
+               tmpl.depth0 = 1;
+       } else {
+               tmpl.array_size = 1;
+               tmpl.depth0 = box->depth;
+       }
        tmpl.last_level = 0;
        tmpl.bind |= PIPE_BIND_LINEAR;
 
@@ -354,7 +419,7 @@ static void fd_resource_transfer_flush_region(struct pipe_context *pctx,
        struct fd_resource *rsc = fd_resource(ptrans->resource);
 
        if (ptrans->resource->target == PIPE_BUFFER)
-               util_range_add(&rsc->valid_buffer_range,
+               util_range_add(&rsc->base, &rsc->valid_buffer_range,
                                           ptrans->box.x + box->x,
                                           ptrans->box.x + box->x + box->width);
 }
@@ -364,7 +429,9 @@ flush_resource(struct fd_context *ctx, struct fd_resource *rsc, unsigned usage)
 {
        struct fd_batch *write_batch = NULL;
 
-       fd_batch_reference(&write_batch, rsc->write_batch);
+       mtx_lock(&ctx->screen->lock);
+       fd_batch_reference_locked(&write_batch, rsc->write_batch);
+       mtx_unlock(&ctx->screen->lock);
 
        if (usage & PIPE_TRANSFER_WRITE) {
                struct fd_batch *batch, *batches[32] = {};
@@ -378,11 +445,11 @@ flush_resource(struct fd_context *ctx, struct fd_resource *rsc, unsigned usage)
                mtx_lock(&ctx->screen->lock);
                batch_mask = rsc->batch_mask;
                foreach_batch(batch, &ctx->screen->batch_cache, batch_mask)
-                       fd_batch_reference(&batches[batch->idx], batch);
+                       fd_batch_reference_locked(&batches[batch->idx], batch);
                mtx_unlock(&ctx->screen->lock);
 
                foreach_batch(batch, &ctx->screen->batch_cache, batch_mask)
-                       fd_batch_flush(batch, false, false);
+                       fd_batch_flush(batch, false);
 
                foreach_batch(batch, &ctx->screen->batch_cache, batch_mask) {
                        fd_batch_sync(batch);
@@ -390,7 +457,7 @@ flush_resource(struct fd_context *ctx, struct fd_resource *rsc, unsigned usage)
                }
                assert(rsc->batch_mask == 0);
        } else if (write_batch) {
-               fd_batch_flush(write_batch, true, false);
+               fd_batch_flush(write_batch, true);
        }
 
        fd_batch_reference(&write_batch, NULL);
@@ -422,7 +489,7 @@ fd_resource_transfer_unmap(struct pipe_context *pctx,
                fd_bo_cpu_fini(rsc->bo);
        }
 
-       util_range_add(&rsc->valid_buffer_range,
+       util_range_add(&rsc->base, &rsc->valid_buffer_range,
                                   ptrans->box.x,
                                   ptrans->box.x + ptrans->box.width);
 
@@ -477,12 +544,14 @@ fd_resource_transfer_map(struct pipe_context *pctx,
 
                staging_rsc = fd_alloc_staging(ctx, rsc, level, box);
                if (staging_rsc) {
+                       struct fd_resource_slice *staging_slice =
+                               fd_resource_slice(staging_rsc, 0);
                        // TODO for PIPE_TRANSFER_READ, need to do untiling blit..
                        trans->staging_prsc = &staging_rsc->base;
                        trans->base.stride = util_format_get_nblocksx(format,
-                               staging_rsc->slices[0].pitch) * staging_rsc->cpp;
+                               staging_slice->pitch) * staging_rsc->cpp;
                        trans->base.layer_stride = staging_rsc->layer_first ?
-                               staging_rsc->layer_size : staging_rsc->slices[0].size0;
+                               staging_rsc->layer_size : staging_slice->size0;
                        trans->staging_box = *box;
                        trans->staging_box.x = 0;
                        trans->staging_box.y = 0;
@@ -490,7 +559,24 @@ fd_resource_transfer_map(struct pipe_context *pctx,
 
                        if (usage & PIPE_TRANSFER_READ) {
                                fd_blit_to_staging(ctx, trans);
-                               fd_bo_cpu_prep(rsc->bo, ctx->pipe, DRM_FREEDRENO_PREP_READ);
+
+                               struct fd_batch *batch = NULL;
+
+                               fd_context_lock(ctx);
+                               fd_batch_reference_locked(&batch, staging_rsc->write_batch);
+                               fd_context_unlock(ctx);
+
+                               /* we can't fd_bo_cpu_prep() until the blit to staging
+                                * is submitted to kernel.. in that case write_batch
+                                * wouldn't be NULL yet:
+                                */
+                               if (batch) {
+                                       fd_batch_sync(batch);
+                                       fd_batch_reference(&batch, NULL);
+                               }
+
+                               fd_bo_cpu_prep(staging_rsc->bo, ctx->pipe,
+                                               DRM_FREEDRENO_PREP_READ);
                        }
 
                        buf = fd_bo_map(staging_rsc->bo);
@@ -527,7 +613,9 @@ fd_resource_transfer_map(struct pipe_context *pctx,
                struct fd_batch *write_batch = NULL;
 
                /* hold a reference, so it doesn't disappear under us: */
-               fd_batch_reference(&write_batch, rsc->write_batch);
+               fd_context_lock(ctx);
+               fd_batch_reference_locked(&write_batch, rsc->write_batch);
+               fd_context_unlock(ctx);
 
                if ((usage & PIPE_TRANSFER_WRITE) && write_batch &&
                                write_batch->back_blit) {
@@ -554,7 +642,8 @@ fd_resource_transfer_map(struct pipe_context *pctx,
                        /* try shadowing only if it avoids a flush, otherwise staging would
                         * be better:
                         */
-                       if (needs_flush && fd_try_shadow_resource(ctx, rsc, level, box)) {
+                       if (needs_flush && fd_try_shadow_resource(ctx, rsc, level,
+                                                       box, DRM_FORMAT_MOD_LINEAR)) {
                                needs_flush = busy = false;
                                rebind_resource(ctx, prsc);
                                ctx->stats.shadow_uploads++;
@@ -573,11 +662,13 @@ fd_resource_transfer_map(struct pipe_context *pctx,
                                 */
                                staging_rsc = fd_alloc_staging(ctx, rsc, level, box);
                                if (staging_rsc) {
+                                       struct fd_resource_slice *staging_slice =
+                                               fd_resource_slice(staging_rsc, 0);
                                        trans->staging_prsc = &staging_rsc->base;
                                        trans->base.stride = util_format_get_nblocksx(format,
-                                               staging_rsc->slices[0].pitch) * staging_rsc->cpp;
+                                               staging_slice->pitch) * staging_rsc->cpp;
                                        trans->base.layer_stride = staging_rsc->layer_first ?
-                                               staging_rsc->layer_size : staging_rsc->slices[0].size0;
+                                               staging_rsc->layer_size : staging_slice->size0;
                                        trans->staging_box = *box;
                                        trans->staging_box.x = 0;
                                        trans->staging_box.y = 0;
@@ -640,11 +731,27 @@ fd_resource_destroy(struct pipe_screen *pscreen,
        fd_bc_invalidate_resource(rsc, true);
        if (rsc->bo)
                fd_bo_del(rsc->bo);
+       if (rsc->scanout)
+               renderonly_scanout_destroy(rsc->scanout, fd_screen(pscreen)->ro);
+
        util_range_destroy(&rsc->valid_buffer_range);
        FREE(rsc);
 }
 
-static boolean
+static uint64_t
+fd_resource_modifier(struct fd_resource *rsc)
+{
+       if (!rsc->tile_mode)
+               return DRM_FORMAT_MOD_LINEAR;
+
+       if (rsc->ubwc_size)
+               return DRM_FORMAT_MOD_QCOM_COMPRESSED;
+
+       /* TODO invent a modifier for tiled but not UBWC buffers: */
+       return DRM_FORMAT_MOD_INVALID;
+}
+
+static bool
 fd_resource_get_handle(struct pipe_screen *pscreen,
                struct pipe_context *pctx,
                struct pipe_resource *prsc,
@@ -653,8 +760,10 @@ fd_resource_get_handle(struct pipe_screen *pscreen,
 {
        struct fd_resource *rsc = fd_resource(prsc);
 
-       return fd_screen_bo_get_handle(pscreen, rsc->bo,
-                       rsc->slices[0].pitch * rsc->cpp, handle);
+       handle->modifier = fd_resource_modifier(rsc);
+
+       return fd_screen_bo_get_handle(pscreen, rsc->bo, rsc->scanout,
+                       fd_resource_slice(rsc, 0)->pitch * rsc->cpp, handle);
 }
 
 static uint32_t
@@ -692,12 +801,12 @@ setup_slices(struct fd_resource *rsc, uint32_t alignment, enum pipe_format forma
                 */
                if (prsc->target == PIPE_TEXTURE_3D && (
                                        level == 1 ||
-                                       (level > 1 && rsc->slices[level - 1].size0 > 0xf000)))
+                                       (level > 1 && fd_resource_slice(rsc, level - 1)->size0 > 0xf000)))
                        slice->size0 = align(blocks * rsc->cpp, alignment);
                else if (level == 0 || rsc->layer_first || alignment == 1)
                        slice->size0 = align(blocks * rsc->cpp, alignment);
                else
-                       slice->size0 = rsc->slices[level - 1].size0;
+                       slice->size0 = fd_resource_slice(rsc, level - 1)->size0;
 
                size += slice->size0 * depth * layers_in_level;
 
@@ -792,15 +901,51 @@ has_depth(enum pipe_format format)
  * Create a new texture object, using the given template info.
  */
 static struct pipe_resource *
-fd_resource_create(struct pipe_screen *pscreen,
-               const struct pipe_resource *tmpl)
+fd_resource_create_with_modifiers(struct pipe_screen *pscreen,
+               const struct pipe_resource *tmpl,
+               const uint64_t *modifiers, int count)
 {
        struct fd_screen *screen = fd_screen(pscreen);
-       struct fd_resource *rsc = CALLOC_STRUCT(fd_resource);
-       struct pipe_resource *prsc = &rsc->base;
+       struct fd_resource *rsc;
+       struct pipe_resource *prsc;
        enum pipe_format format = tmpl->format;
        uint32_t size;
 
+       /* when using kmsro, scanout buffers are allocated on the display device
+        * create_with_modifiers() doesn't give us usage flags, so we have to
+        * assume that all calls with modifiers are scanout-possible
+        */
+       if (screen->ro &&
+               ((tmpl->bind & PIPE_BIND_SCANOUT) ||
+                !(count == 1 && modifiers[0] == DRM_FORMAT_MOD_INVALID))) {
+               struct pipe_resource scanout_templat = *tmpl;
+               struct renderonly_scanout *scanout;
+               struct winsys_handle handle;
+
+               /* apply freedreno alignment requirement */
+               scanout_templat.width0 = align(tmpl->width0, screen->gmem_alignw);
+
+               scanout = renderonly_scanout_for_resource(&scanout_templat,
+                                                                                                 screen->ro, &handle);
+               if (!scanout)
+                       return NULL;
+
+               renderonly_scanout_destroy(scanout, screen->ro);
+
+               assert(handle.type == WINSYS_HANDLE_TYPE_FD);
+               rsc = fd_resource(pscreen->resource_from_handle(pscreen, tmpl,
+                                                                                                               &handle,
+                                                                                                               PIPE_HANDLE_USAGE_FRAMEBUFFER_WRITE));
+               close(handle.handle);
+               if (!rsc)
+                       return NULL;
+
+               return &rsc->base;
+       }
+
+       rsc = CALLOC_STRUCT(fd_resource);
+       prsc = &rsc->base;
+
        DBG("%p: target=%d, format=%s, %ux%ux%u, array_size=%u, last_level=%u, "
                        "nr_samples=%u, usage=%u, bind=%x, flags=%x", prsc,
                        tmpl->target, util_format_name(format),
@@ -818,23 +963,39 @@ fd_resource_create(struct pipe_screen *pscreen,
         PIPE_BIND_LINEAR  | \
         PIPE_BIND_DISPLAY_TARGET)
 
-       if (screen->tile_mode &&
-                       (tmpl->target != PIPE_BUFFER) &&
-                       (tmpl->bind & PIPE_BIND_SAMPLER_VIEW) &&
-                       !(tmpl->bind & LINEAR)) {
-               rsc->tile_mode = screen->tile_mode(tmpl);
-       }
+       bool linear = drm_find_modifier(DRM_FORMAT_MOD_LINEAR, modifiers, count);
+       if (tmpl->bind & LINEAR)
+               linear = true;
+
+       /* Normally, for non-shared buffers, allow buffer compression if
+        * not shared, otherwise only allow if QCOM_COMPRESSED modifier
+        * is requested:
+        *
+        * TODO we should probably also limit tiled in a similar way,
+        * except we don't have a format modifier for tiled.  (We probably
+        * should.)
+        */
+       bool allow_ubwc = drm_find_modifier(DRM_FORMAT_MOD_INVALID, modifiers, count);
+       if (tmpl->bind & PIPE_BIND_SHARED)
+               allow_ubwc = drm_find_modifier(DRM_FORMAT_MOD_QCOM_COMPRESSED, modifiers, count);
+
+       allow_ubwc &= !(fd_mesa_debug & FD_DBG_NOUBWC);
 
        pipe_reference_init(&prsc->reference, 1);
 
        prsc->screen = pscreen;
 
+       if (screen->tile_mode &&
+                       (tmpl->target != PIPE_BUFFER) &&
+                       !linear) {
+               rsc->tile_mode = screen->tile_mode(prsc);
+       }
+
        util_range_init(&rsc->valid_buffer_range);
 
        rsc->internal_format = format;
        rsc->cpp = util_format_get_blocksize(format);
-       prsc->nr_samples = MAX2(1, prsc->nr_samples);
-       rsc->cpp *= prsc->nr_samples;
+       rsc->cpp *= fd_resource_nr_samples(prsc);
 
        assert(rsc->cpp);
 
@@ -866,6 +1027,9 @@ fd_resource_create(struct pipe_screen *pscreen,
 
        size = screen->setup_slices(rsc);
 
+       if (allow_ubwc && screen->fill_ubwc_buffer_sizes && rsc->tile_mode)
+               size += screen->fill_ubwc_buffer_sizes(rsc);
+
        /* special case for hw-query buffer, which we need to allocate before we
         * know the size:
         */
@@ -890,6 +1054,34 @@ fail:
        return NULL;
 }
 
+static struct pipe_resource *
+fd_resource_create(struct pipe_screen *pscreen,
+               const struct pipe_resource *tmpl)
+{
+       const uint64_t mod = DRM_FORMAT_MOD_INVALID;
+       return fd_resource_create_with_modifiers(pscreen, tmpl, &mod, 1);
+}
+
+static bool
+is_supported_modifier(struct pipe_screen *pscreen, enum pipe_format pfmt,
+               uint64_t mod)
+{
+       int count;
+
+       /* Get the count of supported modifiers: */
+       pscreen->query_dmabuf_modifiers(pscreen, pfmt, 0, NULL, NULL, &count);
+
+       /* Get the supported modifiers: */
+       uint64_t modifiers[count];
+       pscreen->query_dmabuf_modifiers(pscreen, pfmt, count, modifiers, NULL, &count);
+
+       for (int i = 0; i < count; i++)
+               if (modifiers[i] == mod)
+                       return true;
+
+       return false;
+}
+
 /**
  * Create a texture from a winsys_handle. The handle is often created in
  * another process by first creating a pipe texture and then calling
@@ -900,8 +1092,9 @@ fd_resource_from_handle(struct pipe_screen *pscreen,
                const struct pipe_resource *tmpl,
                struct winsys_handle *handle, unsigned usage)
 {
+       struct fd_screen *screen = fd_screen(pscreen);
        struct fd_resource *rsc = CALLOC_STRUCT(fd_resource);
-       struct fd_resource_slice *slice = &rsc->slices[0];
+       struct fd_resource_slice *slice = fd_resource_slice(rsc, 0);
        struct pipe_resource *prsc = &rsc->base;
        uint32_t pitchalign = fd_screen(pscreen)->gmem_alignw;
 
@@ -927,9 +1120,9 @@ fd_resource_from_handle(struct pipe_screen *pscreen,
        if (!rsc->bo)
                goto fail;
 
-       prsc->nr_samples = MAX2(1, prsc->nr_samples);
        rsc->internal_format = tmpl->format;
-       rsc->cpp = prsc->nr_samples * util_format_get_blocksize(tmpl->format);
+       rsc->cpp = util_format_get_blocksize(tmpl->format);
+       rsc->cpp *= fd_resource_nr_samples(prsc);
        slice->pitch = handle->stride / rsc->cpp;
        slice->offset = handle->offset;
        slice->size0 = handle->stride * prsc->height0;
@@ -938,8 +1131,29 @@ fd_resource_from_handle(struct pipe_screen *pscreen,
                        (slice->pitch & (pitchalign - 1)))
                goto fail;
 
+       if (handle->modifier == DRM_FORMAT_MOD_QCOM_COMPRESSED) {
+               if (!is_supported_modifier(pscreen, tmpl->format,
+                               DRM_FORMAT_MOD_QCOM_COMPRESSED)) {
+                       DBG("bad modifier: %"PRIx64, handle->modifier);
+                       goto fail;
+               }
+               debug_assert(screen->fill_ubwc_buffer_sizes);
+               screen->fill_ubwc_buffer_sizes(rsc);
+       } else if (handle->modifier &&
+                       (handle->modifier != DRM_FORMAT_MOD_INVALID)) {
+               goto fail;
+       }
+
        assert(rsc->cpp);
 
+       if (screen->ro) {
+               rsc->scanout =
+                       renderonly_create_gpu_import_for_resource(prsc, screen->ro, NULL);
+               /* failure is expected in some cases.. */
+       }
+
+       rsc->valid = true;
+
        return prsc;
 
 fail:
@@ -966,106 +1180,33 @@ fd_render_condition_check(struct pipe_context *pctx)
        return true;
 }
 
-/**
- * Optimal hardware path for blitting pixels.
- * Scaling, format conversion, up- and downsampling (resolve) are allowed.
- */
-static void
-fd_blit(struct pipe_context *pctx, const struct pipe_blit_info *blit_info)
-{
-       struct fd_context *ctx = fd_context(pctx);
-       struct pipe_blit_info info = *blit_info;
-
-       if (info.render_condition_enable && !fd_render_condition_check(pctx))
-               return;
-
-       if (util_try_blit_via_copy_region(pctx, &info)) {
-               return; /* done */
-       }
-
-       if (info.mask & PIPE_MASK_S) {
-               DBG("cannot blit stencil, skipping");
-               info.mask &= ~PIPE_MASK_S;
-       }
-
-       if (!util_blitter_is_blit_supported(ctx->blitter, &info)) {
-               DBG("blit unsupported %s -> %s",
-                               util_format_short_name(info.src.resource->format),
-                               util_format_short_name(info.dst.resource->format));
-               return;
-       }
-
-       if (!(ctx->blit && ctx->blit(ctx, &info)))
-               fd_blitter_blit(ctx, &info);
-}
-
-void
-fd_blitter_pipe_begin(struct fd_context *ctx, bool render_cond, bool discard,
-               enum fd_render_stage stage)
-{
-       fd_fence_ref(ctx->base.screen, &ctx->last_fence, NULL);
-
-       util_blitter_save_fragment_constant_buffer_slot(ctx->blitter,
-                       ctx->constbuf[PIPE_SHADER_FRAGMENT].cb);
-       util_blitter_save_vertex_buffer_slot(ctx->blitter, ctx->vtx.vertexbuf.vb);
-       util_blitter_save_vertex_elements(ctx->blitter, ctx->vtx.vtx);
-       util_blitter_save_vertex_shader(ctx->blitter, ctx->prog.vp);
-       util_blitter_save_so_targets(ctx->blitter, ctx->streamout.num_targets,
-                       ctx->streamout.targets);
-       util_blitter_save_rasterizer(ctx->blitter, ctx->rasterizer);
-       util_blitter_save_viewport(ctx->blitter, &ctx->viewport);
-       util_blitter_save_scissor(ctx->blitter, &ctx->scissor);
-       util_blitter_save_fragment_shader(ctx->blitter, ctx->prog.fp);
-       util_blitter_save_blend(ctx->blitter, ctx->blend);
-       util_blitter_save_depth_stencil_alpha(ctx->blitter, ctx->zsa);
-       util_blitter_save_stencil_ref(ctx->blitter, &ctx->stencil_ref);
-       util_blitter_save_sample_mask(ctx->blitter, ctx->sample_mask);
-       util_blitter_save_framebuffer(ctx->blitter, &ctx->framebuffer);
-       util_blitter_save_fragment_sampler_states(ctx->blitter,
-                       ctx->tex[PIPE_SHADER_FRAGMENT].num_samplers,
-                       (void **)ctx->tex[PIPE_SHADER_FRAGMENT].samplers);
-       util_blitter_save_fragment_sampler_views(ctx->blitter,
-                       ctx->tex[PIPE_SHADER_FRAGMENT].num_textures,
-                       ctx->tex[PIPE_SHADER_FRAGMENT].textures);
-       if (!render_cond)
-               util_blitter_save_render_condition(ctx->blitter,
-                       ctx->cond_query, ctx->cond_cond, ctx->cond_mode);
-
-       if (ctx->batch)
-               fd_batch_set_stage(ctx->batch, stage);
-
-       ctx->in_blit = discard;
-}
-
-void
-fd_blitter_pipe_end(struct fd_context *ctx)
-{
-       if (ctx->batch)
-               fd_batch_set_stage(ctx->batch, FD_STAGE_NULL);
-       ctx->in_blit = false;
-}
-
 static void
 fd_invalidate_resource(struct pipe_context *pctx, struct pipe_resource *prsc)
 {
+       struct fd_context *ctx = fd_context(pctx);
        struct fd_resource *rsc = fd_resource(prsc);
 
        /*
         * TODO I guess we could track that the resource is invalidated and
         * use that as a hint to realloc rather than stall in _transfer_map(),
         * even in the non-DISCARD_WHOLE_RESOURCE case?
+        *
+        * Note: we set dirty bits to trigger invalidate logic fd_draw_vbo
         */
 
        if (rsc->write_batch) {
                struct fd_batch *batch = rsc->write_batch;
                struct pipe_framebuffer_state *pfb = &batch->framebuffer;
 
-               if (pfb->zsbuf && pfb->zsbuf->texture == prsc)
+               if (pfb->zsbuf && pfb->zsbuf->texture == prsc) {
                        batch->resolve &= ~(FD_BUFFER_DEPTH | FD_BUFFER_STENCIL);
+                       ctx->dirty |= FD_DIRTY_ZSA;
+               }
 
                for (unsigned i = 0; i < pfb->nr_cbufs; i++) {
                        if (pfb->cbufs[i] && pfb->cbufs[i]->texture == prsc) {
                                batch->resolve &= ~(PIPE_CLEAR_COLOR0 << i);
+                               ctx->dirty |= FD_DIRTY_FRAMEBUFFER;
                        }
                }
        }
@@ -1113,6 +1254,10 @@ fd_resource_screen_init(struct pipe_screen *pscreen)
        bool fake_rgtc = screen->gpu_id < 400;
 
        pscreen->resource_create = u_transfer_helper_resource_create;
+       /* NOTE: u_transfer_helper does not yet support the _with_modifiers()
+        * variant:
+        */
+       pscreen->resource_create_with_modifiers = fd_resource_create_with_modifiers;
        pscreen->resource_from_handle = fd_resource_from_handle;
        pscreen->resource_get_handle = fd_resource_get_handle;
        pscreen->resource_destroy = u_transfer_helper_resource_destroy;
@@ -1168,6 +1313,13 @@ fd_get_sample_position(struct pipe_context *context,
        pos_out[1] = ptr[sample_index][1] / 16.0f;
 }
 
+static void
+fd_blit_pipe(struct pipe_context *pctx, const struct pipe_blit_info *blit_info)
+{
+       /* wrap fd_blit to return void */
+       fd_blit(pctx, blit_info);
+}
+
 void
 fd_resource_context_init(struct pipe_context *pctx)
 {
@@ -1179,7 +1331,7 @@ fd_resource_context_init(struct pipe_context *pctx)
        pctx->create_surface = fd_create_surface;
        pctx->surface_destroy = fd_surface_destroy;
        pctx->resource_copy_region = fd_resource_copy_region;
-       pctx->blit = fd_blit;
+       pctx->blit = fd_blit_pipe;
        pctx->flush_resource = fd_flush_resource;
        pctx->invalidate_resource = fd_invalidate_resource;
        pctx->get_sample_position = fd_get_sample_position;