freedreno: use rsc->slice accessor everywhere
[mesa.git] / src / gallium / drivers / freedreno / freedreno_resource.c
index 7108ed18776355031c776c5c15329c4f5c31d9b8..fb3d6b65635d87a77137af2d911ceccb83d1bfc9 100644 (file)
@@ -24,9 +24,9 @@
  *    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"
 /* 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
@@ -86,6 +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++) {
@@ -136,9 +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;
@@ -160,15 +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;
@@ -199,6 +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
@@ -231,7 +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: */
@@ -246,7 +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) {
@@ -284,6 +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)
@@ -297,6 +353,8 @@ fd_alloc_staging(struct fd_context *ctx, struct fd_resource *rsc,
         * 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 {
@@ -361,7 +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);
 }
@@ -391,7 +449,7 @@ flush_resource(struct fd_context *ctx, struct fd_resource *rsc, unsigned usage)
                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);
@@ -399,7 +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);
@@ -431,7 +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);
 
@@ -486,12 +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;
@@ -582,7 +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++;
@@ -601,11 +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;
@@ -681,11 +744,14 @@ 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 boolean
+static bool
 fd_resource_get_handle(struct pipe_screen *pscreen,
                struct pipe_context *pctx,
                struct pipe_resource *prsc,
@@ -697,7 +763,7 @@ fd_resource_get_handle(struct pipe_screen *pscreen,
        handle->modifier = fd_resource_modifier(rsc);
 
        return fd_screen_bo_get_handle(pscreen, rsc->bo, rsc->scanout,
-                       rsc->slices[0].pitch * rsc->cpp, handle);
+                       fd_resource_slice(rsc, 0)->pitch * rsc->cpp, handle);
 }
 
 static uint32_t
@@ -735,12 +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;
 
@@ -856,6 +922,9 @@ fd_resource_create_with_modifiers(struct pipe_screen *pscreen,
                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)
@@ -910,31 +979,18 @@ fd_resource_create_with_modifiers(struct pipe_screen *pscreen,
        if (tmpl->bind & PIPE_BIND_SHARED)
                allow_ubwc = drm_find_modifier(DRM_FORMAT_MOD_QCOM_COMPRESSED, modifiers, count);
 
-       /* TODO turn on UBWC for all internal buffers
-        *
-        * There are still some regressions in deqp with UBWC enabled.  I
-        * think it is mostly related to sampler/image views using a format
-        * that doesn't support compression with a resource created with
-        * a format that does.  We need to track the compression state of
-        * a buffer and do an (in-place, hopefully?) resolve if it is re-
-        * interpreted with a format that does not support compression.
-        *
-        * It is possible (likely?) that we can't do atomic ops on a
-        * compressed buffer as well, so this would also require transition
-        * to a compressed state.
-        */
-       allow_ubwc &= !!(fd_mesa_debug & FD_DBG_UBWC);
+       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(tmpl);
+               rsc->tile_mode = screen->tile_mode(prsc);
        }
 
-       pipe_reference_init(&prsc->reference, 1);
-
-       prsc->screen = pscreen;
-
        util_range_init(&rsc->valid_buffer_range);
 
        rsc->internal_format = format;
@@ -1038,7 +1094,7 @@ fd_resource_from_handle(struct pipe_screen *pscreen,
 {
        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;
 
@@ -1078,7 +1134,7 @@ fd_resource_from_handle(struct pipe_screen *pscreen,
        if (handle->modifier == DRM_FORMAT_MOD_QCOM_COMPRESSED) {
                if (!is_supported_modifier(pscreen, tmpl->format,
                                DRM_FORMAT_MOD_QCOM_COMPRESSED)) {
-                       DBG("bad modifier: %llx", handle->modifier);
+                       DBG("bad modifier: %"PRIx64, handle->modifier);
                        goto fail;
                }
                debug_assert(screen->fill_ubwc_buffer_sizes);