pipe_reference_init(&prsc->reference, 1);
prsc->screen = pscreen;
- rsc->bo = panfrost_drm_import_bo(screen, whandle->handle);
- rsc->slices[0].stride = whandle->stride;
- rsc->slices[0].initialized = true;
-
- if (screen->ro) {
- rsc->scanout =
- renderonly_create_gpu_import_for_resource(prsc, screen->ro, NULL);
- /* failure is expected in some cases.. */
- }
+ rsc->bo = panfrost_drm_import_bo(screen, whandle->handle);
+ rsc->slices[0].stride = whandle->stride;
+ rsc->slices[0].initialized = true;
+
+ if (screen->ro) {
+ rsc->scanout =
+ renderonly_create_gpu_import_for_resource(prsc, screen->ro, NULL);
+ /* failure is expected in some cases.. */
+ }
return prsc;
}
-static boolean
+static bool
panfrost_resource_get_handle(struct pipe_screen *pscreen,
struct pipe_context *ctx,
struct pipe_resource *pt,
handle->modifier = DRM_FORMAT_MOD_INVALID;
- if (handle->type == WINSYS_HANDLE_TYPE_SHARED) {
- return FALSE;
- } else if (handle->type == WINSYS_HANDLE_TYPE_KMS) {
- if (renderonly_get_handle(scanout, handle))
- return TRUE;
+ if (handle->type == WINSYS_HANDLE_TYPE_SHARED) {
+ return false;
+ } else if (handle->type == WINSYS_HANDLE_TYPE_KMS) {
+ if (renderonly_get_handle(scanout, handle))
+ return true;
- handle->handle = rsrc->bo->gem_handle;
- handle->stride = rsrc->slices[0].stride;
- return TRUE;
- } else if (handle->type == WINSYS_HANDLE_TYPE_FD) {
+ handle->handle = rsrc->bo->gem_handle;
+ handle->stride = rsrc->slices[0].stride;
+ return TRUE;
+ } else if (handle->type == WINSYS_HANDLE_TYPE_FD) {
if (scanout) {
struct drm_prime_handle args = {
.handle = scanout->handle,
int ret = drmIoctl(screen->ro->kms_fd, DRM_IOCTL_PRIME_HANDLE_TO_FD, &args);
if (ret == -1)
- return FALSE;
+ return false;
handle->stride = scanout->stride;
handle->handle = args.fd;
- return TRUE;
+ return true;
} else {
int fd = panfrost_drm_export_bo(screen, rsrc->bo);
if (fd < 0)
- return FALSE;
+ return false;
handle->handle = fd;
handle->stride = rsrc->slices[0].stride;
- return TRUE;
- }
- }
+ return true;
+ }
+ }
- return FALSE;
+ return false;
}
static void
struct pipe_resource *res;
scanout = renderonly_scanout_for_resource(&scanout_templat,
- pscreen->ro, &handle);
+ pscreen->ro, &handle);
if (!scanout)
return NULL;
struct panfrost_resource *pres = pan_resource(res);
pres->scanout = scanout;
- pscreen->display_target = pres;
return res;
}
static unsigned
panfrost_compute_checksum_sizes(
- struct panfrost_slice *slice,
- unsigned width,
- unsigned height)
+ struct panfrost_slice *slice,
+ unsigned width,
+ unsigned height)
{
- unsigned aligned_width = ALIGN(width, CHECKSUM_TILE_WIDTH);
- unsigned aligned_height = ALIGN(height, CHECKSUM_TILE_HEIGHT);
+ unsigned aligned_width = ALIGN_POT(width, CHECKSUM_TILE_WIDTH);
+ unsigned aligned_height = ALIGN_POT(height, CHECKSUM_TILE_HEIGHT);
unsigned tile_count_x = aligned_width / CHECKSUM_TILE_WIDTH;
unsigned tile_count_y = aligned_height / CHECKSUM_TILE_HEIGHT;
* makes code a lot simpler */
bool renderable = res->bind &
- (PIPE_BIND_RENDER_TARGET | PIPE_BIND_DEPTH_STENCIL);
+ (PIPE_BIND_RENDER_TARGET | PIPE_BIND_DEPTH_STENCIL);
bool afbc = pres->layout == PAN_AFBC;
bool tiled = pres->layout == PAN_TILED;
bool should_align = renderable || tiled;
unsigned effective_depth = depth;
if (should_align) {
- effective_width = ALIGN(effective_width, 16);
- effective_height = ALIGN(effective_height, 16);
+ effective_width = ALIGN_POT(effective_width, 16);
+ effective_height = ALIGN_POT(effective_height, 16);
/* We don't need to align depth */
}
+ /* Align levels to cache-line as a performance improvement for
+ * linear/tiled and as a requirement for AFBC */
+
+ offset = ALIGN_POT(offset, 64);
+
slice->offset = offset;
/* Compute the would-be stride */
/* ..but cache-line align it for performance */
if (can_align_stride && pres->layout == PAN_LINEAR)
- stride = ALIGN(stride, 64);
+ stride = ALIGN_POT(stride, 64);
slice->stride = stride;
slice->checksum_offset = offset;
unsigned size = panfrost_compute_checksum_sizes(
- slice, width, height);
+ slice, width, height);
offset += size;
}
if (res->target != PIPE_TEXTURE_3D) {
/* Arrays and cubemaps have the entire miptree duplicated */
- pres->cubemap_stride = ALIGN(offset, 64);
- *bo_size = ALIGN(pres->cubemap_stride * res->array_size, 4096);
+ pres->cubemap_stride = ALIGN_POT(offset, 64);
+ *bo_size = ALIGN_POT(pres->cubemap_stride * res->array_size, 4096);
} else {
/* 3D strides across the 2D layers */
assert(res->array_size == 1);
pres->cubemap_stride = size_2d;
- *bo_size = ALIGN(offset, 4096);
+ *bo_size = ALIGN_POT(offset, 4096);
}
}
static void
panfrost_resource_create_bo(struct panfrost_screen *screen, struct panfrost_resource *pres)
{
- struct pipe_resource *res = &pres->base;
+ struct pipe_resource *res = &pres->base;
/* Based on the usage, figure out what storing will be used. There are
* various tradeoffs:
*
* Tiled: Not compressed, but cache-optimized. Expensive to write into
* (due to software tiling), but cheap to sample from. Ideal for most
- * textures.
+ * textures.
*
* AFBC: Compressed and renderable (so always desirable for non-scanout
* rendertargets). Cheap to sample from. The format is black box, so we
size_t bo_size;
panfrost_setup_slices(pres, &bo_size);
- pres->bo = panfrost_drm_create_bo(screen, bo_size, 0);
+
+ /* We create a BO immediately but don't bother mapping, since we don't
+ * care to map e.g. FBOs which the CPU probably won't touch */
+ pres->bo = panfrost_drm_create_bo(screen, bo_size, PAN_ALLOCATE_DELAY_MMAP);
}
static struct pipe_resource *
{
/* Make sure we're familiar */
switch (template->target) {
- case PIPE_BUFFER:
- case PIPE_TEXTURE_1D:
- case PIPE_TEXTURE_2D:
- case PIPE_TEXTURE_3D:
- case PIPE_TEXTURE_CUBE:
- case PIPE_TEXTURE_RECT:
- case PIPE_TEXTURE_2D_ARRAY:
- break;
- default:
- DBG("Unknown texture target %d\n", template->target);
- assert(0);
+ case PIPE_BUFFER:
+ case PIPE_TEXTURE_1D:
+ case PIPE_TEXTURE_2D:
+ case PIPE_TEXTURE_3D:
+ case PIPE_TEXTURE_CUBE:
+ case PIPE_TEXTURE_RECT:
+ case PIPE_TEXTURE_2D_ARRAY:
+ break;
+ default:
+ DBG("Unknown texture target %d\n", template->target);
+ assert(0);
}
if (template->bind &
/* When the reference count goes to zero, we need to cleanup */
if (pipe_reference(&bo->reference, NULL))
- panfrost_drm_release_bo(pan_screen(screen), bo);
+ panfrost_drm_release_bo(pan_screen(screen), bo, true);
}
static void
struct panfrost_screen *pscreen = pan_screen(screen);
struct panfrost_resource *rsrc = (struct panfrost_resource *) pt;
- if (rsrc->scanout)
- renderonly_scanout_destroy(rsrc->scanout, pscreen->ro);
+ if (rsrc->scanout)
+ renderonly_scanout_destroy(rsrc->scanout, pscreen->ro);
- if (rsrc->bo)
+ if (rsrc->bo)
panfrost_bo_unreference(screen, rsrc->bo);
util_range_destroy(&rsrc->valid_buffer_range);
- ralloc_free(rsrc);
+ ralloc_free(rsrc);
}
static void *
*out_transfer = &transfer->base;
+ /* If we haven't already mmaped, now's the time */
+
+ if (!bo->cpu) {
+ struct panfrost_screen *screen = pan_screen(pctx->screen);
+ panfrost_drm_mmap_bo(screen, bo);
+ }
+
/* Check if we're bound for rendering and this is a read pixels. If so,
* we need to flush */
bool is_bound = false;
for (unsigned c = 0; c < fb->nr_cbufs; ++c) {
- is_bound |= fb->cbufs[c]->texture == resource;
+ /* If cbufs is NULL, we're definitely not bound here */
+
+ if (fb->cbufs[c])
+ is_bound |= fb->cbufs[c]->texture == resource;
}
if (is_bound && (usage & PIPE_TRANSFER_READ)) {
/* TODO: reallocate */
//printf("debug: Missed reallocate\n");
} else if ((usage & PIPE_TRANSFER_WRITE)
- && resource->target == PIPE_BUFFER
- && !util_ranges_intersect(&rsrc->valid_buffer_range, box->x, box->x + box->width)) {
+ && resource->target == PIPE_BUFFER
+ && !util_ranges_intersect(&rsrc->valid_buffer_range, box->x, box->x + box->width)) {
/* No flush for writes to uninitialized */
} else if (!(usage & PIPE_TRANSFER_UNSYNCHRONIZED)) {
if (usage & PIPE_TRANSFER_WRITE) {
DBG("Unimplemented: reads from AFBC");
} else if (rsrc->layout == PAN_TILED) {
panfrost_load_tiled_image(
- transfer->map,
- bo->cpu + rsrc->slices[level].offset,
- box,
- transfer->base.stride,
- rsrc->slices[level].stride,
- util_format_get_blocksize(resource->format));
+ transfer->map,
+ bo->cpu + rsrc->slices[level].offset,
+ box,
+ transfer->base.stride,
+ rsrc->slices[level].stride,
+ util_format_get_blocksize(resource->format));
}
}
rsrc->slices[level].initialized = true;
return bo->cpu
- + rsrc->slices[level].offset
- + transfer->base.box.z * rsrc->cubemap_stride
- + transfer->base.box.y * rsrc->slices[level].stride
- + transfer->base.box.x * bytes_per_pixel;
+ + rsrc->slices[level].offset
+ + transfer->base.box.z * rsrc->cubemap_stride
+ + transfer->base.box.y * rsrc->slices[level].stride
+ + transfer->base.box.x * bytes_per_pixel;
}
}
struct panfrost_gtransfer *trans = pan_transfer(transfer);
struct panfrost_resource *prsrc = (struct panfrost_resource *) transfer->resource;
+ /* Mark whatever we wrote as written */
+ if (transfer->usage & PIPE_TRANSFER_WRITE)
+ prsrc->slices[transfer->level].initialized = true;
+
if (trans->map) {
struct panfrost_bo *bo = prsrc->bo;
if (transfer->usage & PIPE_TRANSFER_WRITE) {
- unsigned level = transfer->level;
- prsrc->slices[level].initialized = true;
-
if (prsrc->layout == PAN_AFBC) {
DBG("Unimplemented: writes to AFBC\n");
} else if (prsrc->layout == PAN_TILED) {
assert(transfer->box.depth == 1);
panfrost_store_tiled_image(
- bo->cpu + prsrc->slices[level].offset,
- trans->map,
- &transfer->box,
- prsrc->slices[level].stride,
- transfer->stride,
- util_format_get_blocksize(prsrc->base.format));
+ bo->cpu + prsrc->slices[transfer->level].offset,
+ trans->map,
+ &transfer->box,
+ prsrc->slices[transfer->level].stride,
+ transfer->stride,
+ util_format_get_blocksize(prsrc->base.format));
}
}
}
- util_range_add(&prsrc->valid_buffer_range,
- transfer->box.x,
- transfer->box.x + transfer->box.width);
+ util_range_add(&prsrc->valid_buffer_range,
+ transfer->box.x,
+ transfer->box.x + transfer->box.width);
/* Derefence the resource */
pipe_resource_reference(&transfer->resource, NULL);
static void
panfrost_transfer_flush_region(struct pipe_context *pctx,
- struct pipe_transfer *transfer,
- const struct pipe_box *box)
+ struct pipe_transfer *transfer,
+ const struct pipe_box *box)
{
- struct panfrost_resource *rsc = pan_resource(transfer->resource);
-
- if (transfer->resource->target == PIPE_BUFFER) {
- util_range_add(&rsc->valid_buffer_range,
- transfer->box.x + box->x,
- transfer->box.x + box->x + box->width);
- }
-}
+ struct panfrost_resource *rsc = pan_resource(transfer->resource);
-static struct pb_slab *
-panfrost_slab_alloc(void *priv, unsigned heap, unsigned entry_size, unsigned group_index)
-{
- struct panfrost_screen *screen = (struct panfrost_screen *) priv;
- struct panfrost_memory *mem = rzalloc(screen, struct panfrost_memory);
-
- size_t slab_size = (1 << (MAX_SLAB_ENTRY_SIZE + 1));
-
- mem->slab.num_entries = slab_size / entry_size;
- mem->slab.num_free = mem->slab.num_entries;
-
- LIST_INITHEAD(&mem->slab.free);
- for (unsigned i = 0; i < mem->slab.num_entries; ++i) {
- /* Create a slab entry */
- struct panfrost_memory_entry *entry = rzalloc(mem, struct panfrost_memory_entry);
- entry->offset = entry_size * i;
-
- entry->base.slab = &mem->slab;
- entry->base.group_index = group_index;
-
- LIST_ADDTAIL(&entry->base.head, &mem->slab.free);
+ if (transfer->resource->target == PIPE_BUFFER) {
+ util_range_add(&rsc->valid_buffer_range,
+ transfer->box.x + box->x,
+ transfer->box.x + box->x + box->width);
+ } else {
+ unsigned level = transfer->level;
+ rsc->slices[level].initialized = true;
}
-
- /* Actually allocate the memory from kernel-space. Mapped, same_va, no
- * special flags */
-
- panfrost_drm_allocate_slab(screen, mem, slab_size / 4096, true, 0, 0, 0);
-
- return &mem->slab;
-}
-
-static bool
-panfrost_slab_can_reclaim(void *priv, struct pb_slab_entry *entry)
-{
- struct panfrost_memory_entry *p_entry = (struct panfrost_memory_entry *) entry;
- return p_entry->freed;
-}
-
-static void
-panfrost_slab_free(void *priv, struct pb_slab *slab)
-{
- struct panfrost_memory *mem = (struct panfrost_memory *) slab;
- struct panfrost_screen *screen = (struct panfrost_screen *) priv;
-
- panfrost_drm_free_slab(screen, mem);
- ralloc_free(mem);
}
static void
}
static enum pipe_format
-panfrost_resource_get_internal_format(struct pipe_resource *prsrc)
-{
+panfrost_resource_get_internal_format(struct pipe_resource *prsrc) {
return prsrc->format;
}
-static boolean
+static bool
panfrost_generate_mipmap(
- struct pipe_context *pctx,
- struct pipe_resource *prsrc,
- enum pipe_format format,
- unsigned base_level,
- unsigned last_level,
- unsigned first_layer,
- unsigned last_layer)
+ struct pipe_context *pctx,
+ struct pipe_resource *prsrc,
+ enum pipe_format format,
+ unsigned base_level,
+ unsigned last_level,
+ unsigned first_layer,
+ unsigned last_layer)
{
struct panfrost_context *ctx = pan_context(pctx);
struct panfrost_resource *rsrc = pan_resource(prsrc);
/* We've flushed the original buffer if needed, now trigger a blit */
bool blit_res = util_gen_mipmap(
- pctx, prsrc, format,
- base_level, last_level,
- first_layer, last_layer,
- PIPE_TEX_FILTER_LINEAR);
+ pctx, prsrc, format,
+ base_level, last_level,
+ first_layer, last_layer,
+ PIPE_TEX_FILTER_LINEAR);
/* If the blit was successful, flush once more. If it wasn't, well, let
* the state tracker deal with it. */
mali_ptr
panfrost_get_texture_address(
- struct panfrost_resource *rsrc,
- unsigned level, unsigned face)
+ struct panfrost_resource *rsrc,
+ unsigned level, unsigned face)
{
unsigned level_offset = rsrc->slices[level].offset;
unsigned face_offset = face * rsrc->cubemap_stride;
return rsrc->bo->gpu + level_offset + face_offset;
}
+/* Given a resource that has already been allocated, hint that it should use a
+ * given layout. These are suggestions, not commands; it is perfectly legal to
+ * stub out this function, but there will be performance implications. */
+
+void
+panfrost_resource_hint_layout(
+ struct panfrost_screen *screen,
+ struct panfrost_resource *rsrc,
+ enum panfrost_memory_layout layout,
+ signed weight)
+{
+ /* Nothing to do, although a sophisticated implementation might store
+ * the hint */
+
+ if (rsrc->layout == layout)
+ return;
+
+ /* We don't use the weight yet, but we should check that it's positive
+ * (semantically meaning that we should choose the given `layout`) */
+
+ if (weight <= 0)
+ return;
+
+ /* Check if the preferred layout is legal for this buffer */
+
+ if (layout == PAN_AFBC) {
+ bool can_afbc = panfrost_format_supports_afbc(rsrc->base.format);
+ bool is_scanout = rsrc->base.bind &
+ (PIPE_BIND_DISPLAY_TARGET | PIPE_BIND_SCANOUT | PIPE_BIND_SHARED);
+
+ if (!can_afbc || is_scanout)
+ return;
+ }
+
+ /* Simple heuristic so far: if the resource is uninitialized, switch to
+ * the hinted layout. If it is initialized, keep the original layout.
+ * This misses some cases where it would be beneficial to switch and
+ * blit. */
+
+ bool is_initialized = false;
+
+ for (unsigned i = 0; i < MAX_MIP_LEVELS; ++i)
+ is_initialized |= rsrc->slices[i].initialized;
+
+ if (is_initialized)
+ return;
+
+ /* We're uninitialized, so do a layout switch. Reinitialize slices. */
+
+ size_t new_size;
+ rsrc->layout = layout;
+ panfrost_setup_slices(rsrc, &new_size);
+
+ /* If we grew in size, reallocate the BO */
+ if (new_size > rsrc->bo->size) {
+ panfrost_drm_release_bo(screen, rsrc->bo, true);
+ rsrc->bo = panfrost_drm_create_bo(screen, new_size, PAN_ALLOCATE_DELAY_MMAP);
+ }
+}
+
static void
panfrost_resource_set_stencil(struct pipe_resource *prsrc,
struct pipe_resource *stencil)
pscreen->base.resource_from_handle = panfrost_resource_from_handle;
pscreen->base.resource_get_handle = panfrost_resource_get_handle;
pscreen->base.transfer_helper = u_transfer_helper_create(&transfer_vtbl,
- true, false,
- true, true);
-
- pb_slabs_init(&pscreen->slabs,
- MIN_SLAB_ENTRY_SIZE,
- MAX_SLAB_ENTRY_SIZE,
-
- 3, /* Number of heaps */
-
- pscreen,
-
- panfrost_slab_can_reclaim,
- panfrost_slab_alloc,
- panfrost_slab_free);
-}
-
-void
-panfrost_resource_screen_deinit(struct panfrost_screen *pscreen)
-{
- pb_slabs_deinit(&pscreen->slabs);
+ true, false,
+ true, true);
}
void