* 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/set.h"
#include "util/u_drm.h"
+#include "decode/util.h"
+
#include "freedreno_resource.h"
#include "freedreno_batch_cache.h"
#include "freedreno_blitter.h"
#include <errno.h>
/* XXX this should go away, needed for 'struct winsys_handle' */
-#include "state_tracker/drm_driver.h"
+#include "frontend/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
- * called on realloc_bo to ensure the neccessary state is re-
+ * called on realloc_bo to ensure the necessary state is re-
* emitted so the GPU looks at the new backing bo.
*/
static void
-rebind_resource(struct fd_context *ctx, struct pipe_resource *prsc)
+rebind_resource_in_ctx(struct fd_context *ctx, struct fd_resource *rsc)
{
+ struct pipe_resource *prsc = &rsc->base;
+
+ if (ctx->rebind_resource)
+ ctx->rebind_resource(ctx, rsc);
+
/* VBOs */
- for (unsigned i = 0; i < ctx->vtx.vertexbuf.count && !(ctx->dirty & FD_DIRTY_VTXBUF); i++) {
- if (ctx->vtx.vertexbuf.vb[i].buffer.resource == prsc)
- ctx->dirty |= FD_DIRTY_VTXBUF;
+ if (rsc->dirty & FD_DIRTY_VTXBUF) {
+ struct fd_vertexbuf_stateobj *vb = &ctx->vtx.vertexbuf;
+ for (unsigned i = 0; i < vb->count && !(ctx->dirty & FD_DIRTY_VTXBUF); i++) {
+ if (vb->vb[i].buffer.resource == prsc)
+ ctx->dirty |= FD_DIRTY_VTXBUF;
+ }
}
+ const enum fd_dirty_3d_state per_stage_dirty =
+ FD_DIRTY_CONST | FD_DIRTY_TEX | FD_DIRTY_IMAGE | FD_DIRTY_SSBO;
+
+ if (!(rsc->dirty & per_stage_dirty))
+ return;
+
/* per-shader-stage resources: */
for (unsigned stage = 0; stage < PIPE_SHADER_TYPES; stage++) {
/* Constbufs.. note that constbuf[0] is normal uniforms emitted in
* cmdstream rather than by pointer..
*/
- const unsigned num_ubos = util_last_bit(ctx->constbuf[stage].enabled_mask);
- for (unsigned i = 1; i < num_ubos; i++) {
- if (ctx->dirty_shader[stage] & FD_DIRTY_SHADER_CONST)
- break;
- if (ctx->constbuf[stage].cb[i].buffer == prsc)
- ctx->dirty_shader[stage] |= FD_DIRTY_SHADER_CONST;
+ if ((rsc->dirty & FD_DIRTY_CONST) &&
+ !(ctx->dirty_shader[stage] & FD_DIRTY_CONST)) {
+ struct fd_constbuf_stateobj *cb = &ctx->constbuf[stage];
+ const unsigned num_ubos = util_last_bit(cb->enabled_mask);
+ for (unsigned i = 1; i < num_ubos; i++) {
+ if (cb->cb[i].buffer == prsc) {
+ ctx->dirty_shader[stage] |= FD_DIRTY_SHADER_CONST;
+ ctx->dirty |= FD_DIRTY_CONST;
+ break;
+ }
+ }
}
/* Textures */
- for (unsigned i = 0; i < ctx->tex[stage].num_textures; i++) {
- if (ctx->dirty_shader[stage] & FD_DIRTY_SHADER_TEX)
- break;
- if (ctx->tex[stage].textures[i] && (ctx->tex[stage].textures[i]->texture == prsc))
- ctx->dirty_shader[stage] |= FD_DIRTY_SHADER_TEX;
+ if ((rsc->dirty & FD_DIRTY_TEX) &&
+ !(ctx->dirty_shader[stage] & FD_DIRTY_TEX)) {
+ struct fd_texture_stateobj *tex = &ctx->tex[stage];
+ for (unsigned i = 0; i < tex->num_textures; i++) {
+ if (tex->textures[i] && (tex->textures[i]->texture == prsc)) {
+ ctx->dirty_shader[stage] |= FD_DIRTY_SHADER_TEX;
+ ctx->dirty |= FD_DIRTY_TEX;
+ break;
+ }
+ }
+ }
+
+ /* Images */
+ if ((rsc->dirty & FD_DIRTY_IMAGE) &&
+ !(ctx->dirty_shader[stage] & FD_DIRTY_IMAGE)) {
+ struct fd_shaderimg_stateobj *si = &ctx->shaderimg[stage];
+ const unsigned num_images = util_last_bit(si->enabled_mask);
+ for (unsigned i = 0; i < num_images; i++) {
+ if (si->si[i].resource == prsc) {
+ ctx->dirty_shader[stage] |= FD_DIRTY_SHADER_IMAGE;
+ ctx->dirty |= FD_DIRTY_IMAGE;
+ break;
+ }
+ }
}
/* SSBOs */
- const unsigned num_ssbos = util_last_bit(ctx->shaderbuf[stage].enabled_mask);
- for (unsigned i = 0; i < num_ssbos; i++) {
- if (ctx->dirty_shader[stage] & FD_DIRTY_SHADER_SSBO)
- break;
- if (ctx->shaderbuf[stage].sb[i].buffer == prsc)
- ctx->dirty_shader[stage] |= FD_DIRTY_SHADER_SSBO;
+ if ((rsc->dirty & FD_DIRTY_SSBO) &&
+ !(ctx->dirty_shader[stage] & FD_DIRTY_SSBO)) {
+ struct fd_shaderbuf_stateobj *sb = &ctx->shaderbuf[stage];
+ const unsigned num_ssbos = util_last_bit(sb->enabled_mask);
+ for (unsigned i = 0; i < num_ssbos; i++) {
+ if (sb->sb[i].buffer == prsc) {
+ ctx->dirty_shader[stage] |= FD_DIRTY_SHADER_SSBO;
+ ctx->dirty |= FD_DIRTY_SSBO;
+ break;
+ }
+ }
}
}
}
+static void
+rebind_resource(struct fd_resource *rsc)
+{
+ struct fd_screen *screen = fd_screen(rsc->base.screen);
+
+ fd_screen_lock(screen);
+ fd_resource_lock(rsc);
+
+ if (rsc->dirty)
+ list_for_each_entry (struct fd_context, ctx, &screen->context_list, node)
+ rebind_resource_in_ctx(ctx, rsc);
+
+ fd_resource_unlock(rsc);
+ fd_screen_unlock(screen);
+}
+
+static inline void
+fd_resource_set_bo(struct fd_resource *rsc, struct fd_bo *bo)
+{
+ struct fd_screen *screen = fd_screen(rsc->base.screen);
+
+ rsc->bo = bo;
+ rsc->seqno = p_atomic_inc_return(&screen->rsc_seqno);
+}
+
static void
realloc_bo(struct fd_resource *rsc, uint32_t size)
{
if (rsc->bo)
fd_bo_del(rsc->bo);
- rsc->bo = fd_bo_new(screen->dev, size, flags, "%ux%ux%u@%u:%x",
- prsc->width0, prsc->height0, prsc->depth0, rsc->cpp, prsc->bind);
- rsc->seqno = p_atomic_inc_return(&screen->rsc_seqno);
+ struct fd_bo *bo = fd_bo_new(screen->dev, size, flags, "%ux%ux%u@%u:%x",
+ prsc->width0, prsc->height0, prsc->depth0, rsc->layout.cpp, prsc->bind);
+ fd_resource_set_bo(rsc, bo);
+
+ /* Zero out the UBWC area on allocation. This fixes intermittent failures
+ * with UBWC, which I suspect are due to the HW having a hard time
+ * interpreting arbitrary values populating the flags buffer when the BO
+ * was recycled through the bo cache (instead of fresh allocations from
+ * the kernel, which are zeroed). sleep(1) in this spot didn't work
+ * around the issue, but any memset value seems to.
+ */
+ if (rsc->layout.ubwc) {
+ rsc->needs_ubwc_clear = true;
+ }
+
util_range_set_empty(&rsc->valid_buffer_range);
fd_bc_invalidate_resource(rsc, true);
}
}
}
+static void
+flush_resource(struct fd_context *ctx, struct fd_resource *rsc, unsigned usage);
+
+/**
+ * @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;
if (prsc->next)
return false;
+ /* If you have a sequence where there is a single rsc associated
+ * with the current render target, and then you end up shadowing
+ * that same rsc on the 3d pipe (u_blitter), because of how we
+ * swap the new shadow and rsc before the back-blit, you could end
+ * up confusing things into thinking that u_blitter's framebuffer
+ * state is the same as the current framebuffer state, which has
+ * the result of blitting to rsc rather than shadow.
+ *
+ * Normally we wouldn't want to unconditionally trigger a flush,
+ * since that defeats the purpose of shadowing, but this is a
+ * case where we'd have to flush anyways.
+ */
+ if (rsc->write_batch == ctx->batch)
+ flush_resource(ctx, rsc, 0);
+
/* TODO: somehow munge dimensions and format to copy unsupported
* render target format to something that is supported?
*/
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;
* should empty/destroy rsc->batches hashset)
*/
fd_bc_invalidate_resource(rsc, false);
+ rebind_resource(rsc);
- mtx_lock(&ctx->screen->lock);
+ fd_screen_lock(ctx->screen);
/* Swap the backing bo's, so shadow becomes the old buffer,
* blit from shadow to new buffer. From here on out, we
/* TODO valid_buffer_range?? */
swap(rsc->bo, shadow->bo);
swap(rsc->write_batch, shadow->write_batch);
+ swap(rsc->layout, shadow->layout);
rsc->seqno = p_atomic_inc_return(&ctx->screen->rsc_seqno);
/* at this point, the newly created shadow buffer is not referenced
}
swap(rsc->batch_mask, shadow->batch_mask);
- mtx_unlock(&ctx->screen->lock);
+ fd_screen_unlock(ctx->screen);
struct pipe_blit_info blit = {};
blit.dst.resource = prsc;
/* 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: */
set_box(box.height, u_minify(prsc->height0, l));
set_box(box.depth, u_minify(prsc->depth0, l));
- do_blit(ctx, &blit, fallback);
+ for (int i = 0; i < prsc->array_size; i++) {
+ set_box(box.z, i);
+ do_blit(ctx, &blit, fallback);
+ }
}
/* 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) {
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 gallium frontends 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);
+}
+
+/**
+ * Debug helper to hexdump a resource.
+ */
+void
+fd_resource_dump(struct fd_resource *rsc, const char *name)
+{
+ fd_bo_cpu_prep(rsc->bo, NULL, DRM_FREEDRENO_PREP_READ);
+ printf("%s: \n", name);
+ dump_hex(fd_bo_map(rsc->bo), fd_bo_size(rsc->bo));
+}
+
static struct fd_resource *
fd_alloc_staging(struct fd_context *ctx, struct fd_resource *rsc,
unsigned level, const struct pipe_box *box)
* 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 {
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);
}
{
struct fd_batch *write_batch = NULL;
- mtx_lock(&ctx->screen->lock);
+ fd_screen_lock(ctx->screen);
fd_batch_reference_locked(&write_batch, rsc->write_batch);
- mtx_unlock(&ctx->screen->lock);
+ fd_screen_unlock(ctx->screen);
if (usage & PIPE_TRANSFER_WRITE) {
struct fd_batch *batch, *batches[32] = {};
* to iterate the batches which reference this resource. So
* we must first grab references under a lock, then flush.
*/
- mtx_lock(&ctx->screen->lock);
+ fd_screen_lock(ctx->screen);
batch_mask = rsc->batch_mask;
foreach_batch(batch, &ctx->screen->batch_cache, batch_mask)
fd_batch_reference_locked(&batches[batch->idx], batch);
- mtx_unlock(&ctx->screen->lock);
+ fd_screen_unlock(ctx->screen);
foreach_batch(batch, &ctx->screen->batch_cache, batch_mask)
- fd_batch_flush(batch, false, false);
+ fd_batch_flush(batch);
foreach_batch(batch, &ctx->screen->batch_cache, batch_mask) {
- fd_batch_sync(batch);
fd_batch_reference(&batches[batch->idx], NULL);
}
assert(rsc->batch_mask == 0);
} else if (write_batch) {
- fd_batch_flush(write_batch, true, false);
+ fd_batch_flush(write_batch);
}
fd_batch_reference(&write_batch, NULL);
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);
{
struct fd_context *ctx = fd_context(pctx);
struct fd_resource *rsc = fd_resource(prsc);
- struct fd_resource_slice *slice = fd_resource_slice(rsc, level);
struct fd_transfer *trans;
struct pipe_transfer *ptrans;
enum pipe_format format = prsc->format;
DBG("prsc=%p, level=%u, usage=%x, box=%dx%d+%d,%d", prsc, level, usage,
box->width, box->height, box->x, box->y);
+ if ((usage & PIPE_TRANSFER_MAP_DIRECTLY) && rsc->layout.tile_mode) {
+ DBG("CANNOT MAP DIRECTLY!\n");
+ return NULL;
+ }
+
ptrans = slab_alloc(&ctx->transfer_pool);
if (!ptrans)
return NULL;
ptrans->level = level;
ptrans->usage = usage;
ptrans->box = *box;
- ptrans->stride = util_format_get_nblocksx(format, slice->pitch) * rsc->cpp;
- ptrans->layer_stride = rsc->layer_first ? rsc->layer_size : slice->size0;
+ ptrans->stride = fd_resource_pitch(rsc, level);
+ ptrans->layer_stride = fd_resource_layer_stride(rsc, level);
/* we always need a staging texture for tiled buffers:
*
* splitting a batch.. for ex, mid-frame texture uploads to a tiled
* texture.
*/
- if (rsc->tile_mode) {
+ if (rsc->layout.tile_mode) {
struct fd_resource *staging_rsc;
staging_rsc = fd_alloc_staging(ctx, rsc, level, box);
if (staging_rsc) {
// 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;
- trans->base.layer_stride = staging_rsc->layer_first ?
- staging_rsc->layer_size : staging_rsc->slices[0].size0;
+ trans->base.stride = fd_resource_pitch(staging_rsc, 0);
+ trans->base.layer_stride = fd_resource_layer_stride(staging_rsc, 0);
trans->staging_box = *box;
trans->staging_box.x = 0;
trans->staging_box.y = 0;
if (usage & PIPE_TRANSFER_READ) {
fd_blit_to_staging(ctx, trans);
- 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);
}
if (usage & PIPE_TRANSFER_WRITE)
op |= DRM_FREEDRENO_PREP_WRITE;
+ bool needs_flush = pending(rsc, !!(usage & PIPE_TRANSFER_WRITE));
+
if (usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE) {
- realloc_bo(rsc, fd_bo_size(rsc->bo));
- rebind_resource(ctx, prsc);
+ if (needs_flush || fd_resource_busy(rsc, op)) {
+ rebind_resource(rsc);
+ realloc_bo(rsc, fd_bo_size(rsc->bo));
+ }
} else if ((usage & PIPE_TRANSFER_WRITE) &&
prsc->target == PIPE_BUFFER &&
!util_ranges_intersect(&rsc->valid_buffer_range,
/* If the GPU is writing to the resource, or if it is reading from the
* resource and we're trying to write to it, flush the renders.
*/
- bool needs_flush = pending(rsc, !!(usage & PIPE_TRANSFER_WRITE));
- bool busy = needs_flush || (0 != fd_bo_cpu_prep(rsc->bo,
- ctx->pipe, op | DRM_FREEDRENO_PREP_NOSYNC));
+ bool busy = needs_flush || fd_resource_busy(rsc, op);
/* if we need to flush/stall, see if we can make a shadow buffer
* to avoid this:
/* 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++;
} else {
struct fd_resource *staging_rsc;
staging_rsc = fd_alloc_staging(ctx, rsc, level, box);
if (staging_rsc) {
trans->staging_prsc = &staging_rsc->base;
- trans->base.stride = util_format_get_nblocksx(format,
- staging_rsc->slices[0].pitch) * staging_rsc->cpp;
- trans->base.layer_stride = staging_rsc->layer_first ?
- staging_rsc->layer_size : staging_rsc->slices[0].size0;
+ trans->base.stride = fd_resource_pitch(staging_rsc, 0);
+ trans->base.layer_stride =
+ fd_resource_layer_stride(staging_rsc, 0);
trans->staging_box = *box;
trans->staging_box.x = 0;
trans->staging_box.y = 0;
buf = fd_bo_map(rsc->bo);
offset =
box->y / util_format_get_blockheight(format) * ptrans->stride +
- box->x / util_format_get_blockwidth(format) * rsc->cpp +
+ box->x / util_format_get_blockwidth(format) * rsc->layout.cpp +
fd_resource_offset(rsc, level, box->z);
if (usage & PIPE_TRANSFER_WRITE)
renderonly_scanout_destroy(rsc->scanout, fd_screen(pscreen)->ro);
util_range_destroy(&rsc->valid_buffer_range);
+ simple_mtx_destroy(&rsc->lock);
FREE(rsc);
}
static uint64_t
fd_resource_modifier(struct fd_resource *rsc)
{
- if (!rsc->tile_mode)
+ if (!rsc->layout.tile_mode)
return DRM_FORMAT_MOD_LINEAR;
+ if (rsc->layout.ubwc_layer_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,
handle->modifier = fd_resource_modifier(rsc);
return fd_screen_bo_get_handle(pscreen, rsc->bo, rsc->scanout,
- rsc->slices[0].pitch * rsc->cpp, handle);
-}
-
-static uint32_t
-setup_slices(struct fd_resource *rsc, uint32_t alignment, enum pipe_format format)
-{
- struct pipe_resource *prsc = &rsc->base;
- struct fd_screen *screen = fd_screen(prsc->screen);
- enum util_format_layout layout = util_format_description(format)->layout;
- uint32_t pitchalign = screen->gmem_alignw;
- uint32_t level, size = 0;
- uint32_t width = prsc->width0;
- uint32_t height = prsc->height0;
- uint32_t depth = prsc->depth0;
- /* in layer_first layout, the level (slice) contains just one
- * layer (since in fact the layer contains the slices)
- */
- uint32_t layers_in_level = rsc->layer_first ? 1 : prsc->array_size;
-
- for (level = 0; level <= prsc->last_level; level++) {
- struct fd_resource_slice *slice = fd_resource_slice(rsc, level);
- uint32_t blocks;
-
- if (layout == UTIL_FORMAT_LAYOUT_ASTC)
- slice->pitch = width =
- util_align_npot(width, pitchalign * util_format_get_blockwidth(format));
- else
- slice->pitch = width = align(width, pitchalign);
- slice->offset = size;
- blocks = util_format_get_nblocks(format, width, height);
- /* 1d array and 2d array textures must all have the same layer size
- * for each miplevel on a3xx. 3d textures can have different layer
- * sizes for high levels, but the hw auto-sizer is buggy (or at least
- * different than what this code does), so as soon as the layer size
- * range gets into range, we stop reducing it.
- */
- if (prsc->target == PIPE_TEXTURE_3D && (
- level == 1 ||
- (level > 1 && rsc->slices[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;
-
- size += slice->size0 * depth * layers_in_level;
-
- width = u_minify(width, 1);
- height = u_minify(height, 1);
- depth = u_minify(depth, 1);
- }
-
- return size;
-}
-
-static uint32_t
-slice_alignment(enum pipe_texture_target target)
-{
- /* on a3xx, 2d array and 3d textures seem to want their
- * layers aligned to page boundaries:
- */
- switch (target) {
- case PIPE_TEXTURE_3D:
- case PIPE_TEXTURE_1D_ARRAY:
- case PIPE_TEXTURE_2D_ARRAY:
- return 4096;
- default:
- return 1;
- }
-}
-
-/* cross generation texture layout to plug in to screen->setup_slices()..
- * replace with generation specific one as-needed.
- *
- * TODO for a4xx probably can extract out the a4xx specific logic int
- * a small fd4_setup_slices() wrapper that sets up layer_first, and then
- * calls this.
- */
-uint32_t
-fd_setup_slices(struct fd_resource *rsc)
-{
- uint32_t alignment;
-
- alignment = slice_alignment(rsc->base.target);
-
- struct fd_screen *screen = fd_screen(rsc->base.screen);
- if (is_a4xx(screen)) {
- switch (rsc->base.target) {
- case PIPE_TEXTURE_3D:
- rsc->layer_first = false;
- break;
- default:
- rsc->layer_first = true;
- alignment = 1;
- break;
- }
- }
-
- return setup_slices(rsc, alignment, rsc->base.format);
+ fd_resource_pitch(rsc, 0), handle);
}
/* special case to resize query buf after allocated.. */
realloc_bo(rsc, fd_screen(prsc->screen)->setup_slices(rsc));
}
-// TODO common helper?
-static bool
-has_depth(enum pipe_format format)
+static void
+fd_resource_layout_init(struct pipe_resource *prsc)
{
- switch (format) {
- case PIPE_FORMAT_Z16_UNORM:
- case PIPE_FORMAT_Z32_UNORM:
- case PIPE_FORMAT_Z32_FLOAT:
- case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT:
- case PIPE_FORMAT_Z24_UNORM_S8_UINT:
- case PIPE_FORMAT_S8_UINT_Z24_UNORM:
- case PIPE_FORMAT_Z24X8_UNORM:
- case PIPE_FORMAT_X8Z24_UNORM:
- return true;
- default:
- return false;
- }
+ struct fd_resource *rsc = fd_resource(prsc);
+ struct fdl_layout *layout = &rsc->layout;
+
+ layout->format = prsc->format;
+
+ layout->width0 = prsc->width0;
+ layout->height0 = prsc->height0;
+ layout->depth0 = prsc->depth0;
+
+ layout->cpp = util_format_get_blocksize(prsc->format);
+ layout->cpp *= fd_resource_nr_samples(prsc);
+ layout->cpp_shift = ffs(layout->cpp) - 1;
}
/**
- * Create a new texture object, using the given template info.
+ * Helper that allocates a resource and resolves its layout (but doesn't
+ * allocate its bo).
+ *
+ * It returns a pipe_resource (as fd_resource_create_with_modifiers()
+ * would do), and also bo's minimum required size as an output argument.
*/
static struct pipe_resource *
-fd_resource_create_with_modifiers(struct pipe_screen *pscreen,
+fd_resource_allocate_and_resolve(struct pipe_screen *pscreen,
const struct pipe_resource *tmpl,
- const uint64_t *modifiers, int count)
+ const uint64_t *modifiers, int count, uint32_t *psize)
{
struct fd_screen *screen = fd_screen(pscreen);
struct fd_resource *rsc;
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;
-
- 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;
return NULL;
*prsc = *tmpl;
+ fd_resource_layout_init(prsc);
#define LINEAR \
(PIPE_BIND_SCANOUT | \
if (tmpl->bind & LINEAR)
linear = true;
+ if (fd_mesa_debug & FD_DBG_NOTILE)
+ linear = true;
+
/* Normally, for non-shared buffers, allow buffer compression if
* not shared, otherwise only allow if QCOM_COMPRESSED modifier
* is requested:
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->layout.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;
- rsc->cpp = util_format_get_blocksize(format);
- rsc->cpp *= fd_resource_nr_samples(prsc);
-
- assert(rsc->cpp);
-
- // XXX probably need some extra work if we hit rsc shadowing path w/ lrz..
- if ((is_a5xx(screen) || is_a6xx(screen)) &&
- (fd_mesa_debug & FD_DBG_LRZ) && has_depth(format)) {
- const uint32_t flags = DRM_FREEDRENO_GEM_CACHE_WCOMBINE |
- DRM_FREEDRENO_GEM_TYPE_KMEM; /* TODO */
- unsigned lrz_pitch = align(DIV_ROUND_UP(tmpl->width0, 8), 64);
- unsigned lrz_height = DIV_ROUND_UP(tmpl->height0, 8);
-
- /* LRZ buffer is super-sampled: */
- switch (prsc->nr_samples) {
- case 4:
- lrz_pitch *= 2;
- case 2:
- lrz_height *= 2;
- }
+ simple_mtx_init(&rsc->lock, mtx_plain);
- unsigned size = lrz_pitch * lrz_height * 2;
+ rsc->internal_format = format;
- size += 0x1000; /* for GRAS_LRZ_FAST_CLEAR_BUFFER */
+ rsc->layout.ubwc = rsc->layout.tile_mode && is_a6xx(screen) && allow_ubwc;
- rsc->lrz_height = lrz_height;
- rsc->lrz_width = lrz_pitch;
- rsc->lrz_pitch = lrz_pitch;
- rsc->lrz = fd_bo_new(screen->dev, size, flags, "lrz");
+ if (prsc->target == PIPE_BUFFER) {
+ assert(prsc->format == PIPE_FORMAT_R8_UNORM);
+ size = prsc->width0;
+ fdl_layout_buffer(&rsc->layout, size);
+ } else {
+ size = screen->setup_slices(rsc);
}
- 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:
*/
return prsc;
}
- if (rsc->layer_first) {
- rsc->layer_size = align(size, 4096);
- size = rsc->layer_size * prsc->array_size;
+ /* Set the layer size if the (non-a6xx) backend hasn't done so. */
+ if (rsc->layout.layer_first && !rsc->layout.layer_size) {
+ rsc->layout.layer_size = align(size, 4096);
+ size = rsc->layout.layer_size * prsc->array_size;
+ }
+
+ if (fd_mesa_debug & FD_DBG_LAYOUT)
+ fdl_dump_layout(&rsc->layout);
+
+ /* Hand out the resolved size. */
+ if (psize)
+ *psize = size;
+
+ return prsc;
+}
+
+/**
+ * Create a new texture object, using the given template info.
+ */
+static struct pipe_resource *
+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;
+ struct pipe_resource *prsc;
+ 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;
+
+ /* note: alignment is wrong for a6xx */
+ 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;
}
+ prsc = fd_resource_allocate_and_resolve(pscreen, tmpl, modifiers, count, &size);
+ if (!prsc)
+ return NULL;
+ rsc = fd_resource(prsc);
+
realloc_bo(rsc, size);
if (!rsc->bo)
goto fail;
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
{
struct fd_screen *screen = fd_screen(pscreen);
struct fd_resource *rsc = CALLOC_STRUCT(fd_resource);
- struct fd_resource_slice *slice = &rsc->slices[0];
+ struct fdl_slice *slice = fd_resource_slice(rsc, 0);
struct pipe_resource *prsc = &rsc->base;
- uint32_t pitchalign = fd_screen(pscreen)->gmem_alignw;
DBG("target=%d, format=%s, %ux%ux%u, array_size=%u, last_level=%u, "
"nr_samples=%u, usage=%u, bind=%x, flags=%x",
return NULL;
*prsc = *tmpl;
+ fd_resource_layout_init(prsc);
pipe_reference_init(&prsc->reference, 1);
util_range_init(&rsc->valid_buffer_range);
- rsc->bo = fd_screen_bo_from_handle(pscreen, handle);
- if (!rsc->bo)
+ simple_mtx_init(&rsc->lock, mtx_plain);
+
+ struct fd_bo *bo = fd_screen_bo_from_handle(pscreen, handle);
+ if (!bo)
goto fail;
+ fd_resource_set_bo(rsc, bo);
+
rsc->internal_format = tmpl->format;
- rsc->cpp = util_format_get_blocksize(tmpl->format);
- rsc->cpp *= fd_resource_nr_samples(prsc);
- slice->pitch = handle->stride / rsc->cpp;
+ rsc->layout.pitch0 = handle->stride;
slice->offset = handle->offset;
slice->size0 = handle->stride * prsc->height0;
- if ((slice->pitch < align(prsc->width0, pitchalign)) ||
- (slice->pitch & (pitchalign - 1)))
- goto fail;
+ /* use a pitchalign of gmem_alignw pixels, because GMEM resolve for
+ * lower alignments is not implemented (but possible for a6xx at least)
+ *
+ * for UBWC-enabled resources, layout_resource_for_modifier will further
+ * validate the pitch and set the right pitchalign
+ */
+ rsc->layout.pitchalign =
+ fdl_cpp_shift(&rsc->layout) + util_logbase2(screen->gmem_alignw);
- 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);
- 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)) {
+ /* apply the minimum pitchalign (note: actually 4 for a3xx but doesn't matter) */
+ if (is_a6xx(screen) || is_a5xx(screen))
+ rsc->layout.pitchalign = MAX2(rsc->layout.pitchalign, 6);
+ else
+ rsc->layout.pitchalign = MAX2(rsc->layout.pitchalign, 5);
+
+ if (rsc->layout.pitch0 < (prsc->width0 * rsc->layout.cpp) ||
+ fd_resource_pitch(rsc, 0) != rsc->layout.pitch0)
goto fail;
- }
- assert(rsc->cpp);
+ assert(rsc->layout.cpp);
+
+ if (screen->layout_resource_for_modifier(rsc, handle->modifier) < 0)
+ goto fail;
if (screen->ro) {
rsc->scanout =
.get_stencil = fd_resource_get_stencil,
};
+static const uint64_t supported_modifiers[] = {
+ DRM_FORMAT_MOD_LINEAR,
+};
+
+static int
+fd_layout_resource_for_modifier(struct fd_resource *rsc, uint64_t modifier)
+{
+ switch (modifier) {
+ case DRM_FORMAT_MOD_LINEAR:
+ /* The dri gallium frontend will pass DRM_FORMAT_MOD_INVALID to us
+ * when it's called through any of the non-modifier BO create entry
+ * points. Other drivers will determine tiling from the kernel or
+ * other legacy backchannels, but for freedreno it just means
+ * LINEAR. */
+ case DRM_FORMAT_MOD_INVALID:
+ return 0;
+ default:
+ return -1;
+ }
+}
+
+static struct pipe_resource *
+fd_resource_from_memobj(struct pipe_screen *pscreen,
+ const struct pipe_resource *tmpl,
+ struct pipe_memory_object *pmemobj,
+ uint64_t offset)
+{
+ struct fd_screen *screen = fd_screen(pscreen);
+ struct fd_memory_object *memobj = fd_memory_object(pmemobj);
+ struct pipe_resource *prsc;
+ struct fd_resource *rsc;
+ uint32_t size;
+ assert(memobj->bo);
+
+ /* We shouldn't get a scanout buffer here. */
+ assert(!(tmpl->bind & PIPE_BIND_SCANOUT));
+
+ uint64_t modifiers = DRM_FORMAT_MOD_INVALID;
+ if (tmpl->bind & PIPE_BIND_LINEAR) {
+ modifiers = DRM_FORMAT_MOD_LINEAR;
+ } else if (is_a6xx(screen) && tmpl->width0 >= FDL_MIN_UBWC_WIDTH) {
+ modifiers = DRM_FORMAT_MOD_QCOM_COMPRESSED;
+ }
+
+ /* Allocate new pipe resource. */
+ prsc = fd_resource_allocate_and_resolve(pscreen, tmpl, &modifiers, 1, &size);
+ if (!prsc)
+ return NULL;
+ rsc = fd_resource(prsc);
+
+ /* bo's size has to be large enough, otherwise cleanup resource and fail
+ * gracefully.
+ */
+ if (fd_bo_size(memobj->bo) < size) {
+ fd_resource_destroy(pscreen, prsc);
+ return NULL;
+ }
+
+ /* Share the bo with the memory object. */
+ fd_resource_set_bo(rsc, fd_bo_ref(memobj->bo));
+
+ return prsc;
+}
+
+static struct pipe_memory_object *
+fd_memobj_create_from_handle(struct pipe_screen *pscreen,
+ struct winsys_handle *whandle,
+ bool dedicated)
+{
+ struct fd_memory_object *memobj = CALLOC_STRUCT(fd_memory_object);
+ if (!memobj)
+ return NULL;
+
+ struct fd_bo *bo = fd_screen_bo_from_handle(pscreen, whandle);
+ if (!bo) {
+ free(memobj);
+ return NULL;
+ }
+
+ memobj->b.dedicated = dedicated;
+ memobj->bo = bo;
+
+ return &memobj->b;
+}
+
+static void
+fd_memobj_destroy(struct pipe_screen *pscreen,
+ struct pipe_memory_object *pmemobj)
+{
+ struct fd_memory_object *memobj = fd_memory_object(pmemobj);
+
+ assert(memobj->bo);
+ fd_bo_del(memobj->bo);
+
+ free(pmemobj);
+}
+
void
fd_resource_screen_init(struct pipe_screen *pscreen)
{
pscreen->transfer_helper = u_transfer_helper_create(&transfer_vtbl,
true, false, fake_rgtc, true);
- if (!screen->setup_slices)
- screen->setup_slices = fd_setup_slices;
+ if (!screen->layout_resource_for_modifier)
+ screen->layout_resource_for_modifier = fd_layout_resource_for_modifier;
+ if (!screen->supported_modifiers) {
+ screen->supported_modifiers = supported_modifiers;
+ screen->num_supported_modifiers = ARRAY_SIZE(supported_modifiers);
+ }
+
+ /* GL_EXT_memory_object */
+ pscreen->memobj_create_from_handle = fd_memobj_create_from_handle;
+ pscreen->memobj_destroy = fd_memobj_destroy;
+ pscreen->resource_from_memobj = fd_resource_from_memobj;
}
static void