#include <fcntl.h>
#include "drm-uapi/drm_fourcc.h"
-#include "state_tracker/winsys_handle.h"
-#include "util/u_format.h"
+#include "frontend/winsys_handle.h"
+#include "util/format/u_format.h"
#include "util/u_memory.h"
#include "util/u_surface.h"
#include "util/u_transfer.h"
#include "pan_resource.h"
#include "pan_util.h"
#include "pan_tiling.h"
+#include "pandecode/decode.h"
+#include "panfrost-quirks.h"
-static void
+/* Wrapper around panfrost_bo_create that handles pandecode */
+
+struct panfrost_bo *
+pan_bo_create(struct panfrost_device *dev, size_t size, uint32_t flags)
+{
+ struct panfrost_bo *bo = panfrost_bo_create(dev, size, flags);
+
+ if (pan_debug & (PAN_DBG_TRACE | PAN_DBG_SYNC)) {
+ if (flags & PAN_BO_INVISIBLE)
+ pandecode_inject_mmap(bo->gpu, NULL, bo->size, NULL);
+ else if (!(flags & PAN_BO_DELAY_MMAP))
+ pandecode_inject_mmap(bo->gpu, bo->cpu, bo->size, NULL);
+ }
+
+ return bo;
+}
+
+void
panfrost_resource_reset_damage(struct panfrost_resource *pres)
{
/* We set the damage extent to the full resource size but keep the
struct winsys_handle *whandle,
unsigned usage)
{
- struct panfrost_screen *screen = pan_screen(pscreen);
+ struct panfrost_device *dev = pan_device(pscreen);
struct panfrost_resource *rsc;
struct pipe_resource *prsc;
pipe_reference_init(&prsc->reference, 1);
prsc->screen = pscreen;
- rsc->bo = panfrost_bo_import(screen, whandle->handle);
+ rsc->bo = panfrost_bo_import(dev, whandle->handle);
+ rsc->internal_format = templat->format;
+ rsc->layout = MALI_TEXTURE_LINEAR;
rsc->slices[0].stride = whandle->stride;
rsc->slices[0].offset = whandle->offset;
rsc->slices[0].initialized = true;
panfrost_resource_reset_damage(rsc);
- if (screen->ro) {
+ if (dev->quirks & IS_BIFROST &&
+ templat->bind & PIPE_BIND_RENDER_TARGET) {
+ unsigned size = panfrost_compute_checksum_size(
+ &rsc->slices[0], templat->width0, templat->height0);
+ rsc->slices[0].checksum_bo = pan_bo_create(dev, size, 0);
+ rsc->checksummed = true;
+ }
+
+ if (dev->ro) {
rsc->scanout =
- renderonly_create_gpu_import_for_resource(prsc, screen->ro, NULL);
+ renderonly_create_gpu_import_for_resource(prsc, dev->ro, NULL);
/* failure is expected in some cases.. */
}
struct winsys_handle *handle,
unsigned usage)
{
- struct panfrost_screen *screen = pan_screen(pscreen);
+ struct panfrost_device *dev = pan_device(pscreen);
struct panfrost_resource *rsrc = (struct panfrost_resource *) pt;
struct renderonly_scanout *scanout = rsrc->scanout;
.flags = DRM_CLOEXEC,
};
- int ret = drmIoctl(screen->ro->kms_fd, DRM_IOCTL_PRIME_HANDLE_TO_FD, &args);
+ int ret = drmIoctl(dev->ro->kms_fd, DRM_IOCTL_PRIME_HANDLE_TO_FD, &args);
if (ret == -1)
return false;
panfrost_create_scanout_res(struct pipe_screen *screen,
const struct pipe_resource *template)
{
- struct panfrost_screen *pscreen = pan_screen(screen);
+ struct panfrost_device *dev = pan_device(screen);
struct pipe_resource scanout_templat = *template;
struct renderonly_scanout *scanout;
struct winsys_handle handle;
struct pipe_resource *res;
scanout = renderonly_scanout_for_resource(&scanout_templat,
- pscreen->ro, &handle);
+ dev->ro, &handle);
if (!scanout)
return NULL;
return res;
}
-/* Computes sizes for checksumming, which is 8 bytes per 16x16 tile */
-
-#define CHECKSUM_TILE_WIDTH 16
-#define CHECKSUM_TILE_HEIGHT 16
-#define CHECKSUM_BYTES_PER_TILE 8
-
-static unsigned
-panfrost_compute_checksum_sizes(
- struct panfrost_slice *slice,
- unsigned width,
- unsigned 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;
-
- slice->checksum_stride = tile_count_x * CHECKSUM_BYTES_PER_TILE;
-
- return slice->checksum_stride * tile_count_y;
-}
-
/* Setup the mip tree given a particular layout, possibly with checksumming */
static void
bool renderable = res->bind &
(PIPE_BIND_RENDER_TARGET | PIPE_BIND_DEPTH_STENCIL);
- bool afbc = pres->layout == PAN_AFBC;
- bool tiled = pres->layout == PAN_TILED;
+ bool afbc = pres->layout == MALI_TEXTURE_AFBC;
+ bool tiled = pres->layout == MALI_TEXTURE_TILED;
bool should_align = renderable || tiled;
/* We don't know how to specify a 2D stride for 3D textures */
/* Compute the would-be stride */
unsigned stride = bytes_per_pixel * effective_width;
+ if (util_format_is_compressed(res->format))
+ stride /= 4;
+
/* ..but cache-line align it for performance */
- if (can_align_stride && pres->layout == PAN_LINEAR)
+ if (can_align_stride && pres->layout == MALI_TEXTURE_LINEAR)
stride = ALIGN_POT(stride, 64);
slice->stride = stride;
unsigned slice_one_size = slice->stride * effective_height;
unsigned slice_full_size = slice_one_size * effective_depth;
+ slice->size0 = slice_one_size;
+
/* Report 2D size for 3D texturing */
if (l == 0)
if (pres->checksummed) {
slice->checksum_offset = offset;
- unsigned size = panfrost_compute_checksum_sizes(
+ unsigned size = panfrost_compute_checksum_size(
slice, width, height);
offset += size;
}
static void
-panfrost_resource_create_bo(struct panfrost_screen *screen, struct panfrost_resource *pres)
+panfrost_resource_create_bo(struct panfrost_device *dev, struct panfrost_resource *pres)
{
struct pipe_resource *res = &pres->base;
* AFBC: Compressed and renderable (so always desirable for non-scanout
* rendertargets). Cheap to sample from. The format is black box, so we
* can't read/write from software.
- */
-
- /* Tiling textures is almost always faster, unless we only use it once */
-
- bool is_texture = (res->bind & PIPE_BIND_SAMPLER_VIEW);
- bool is_2d = res->depth0 == 1 && res->array_size == 1;
- bool is_streaming = (res->usage != PIPE_USAGE_STREAM);
-
- bool should_tile = is_streaming && is_texture && is_2d;
-
- /* Depth/stencil can't be tiled, only linear or AFBC */
- should_tile &= !(res->bind & PIPE_BIND_DEPTH_STENCIL);
+ *
+ * Tiling textures is almost always faster, unless we only use it once.
+ * Only a few types of resources can be tiled, ensure the bind is only
+ * (a combination of) one of the following */
+
+ const unsigned valid_binding =
+ PIPE_BIND_DEPTH_STENCIL |
+ PIPE_BIND_RENDER_TARGET |
+ PIPE_BIND_BLENDABLE |
+ PIPE_BIND_SAMPLER_VIEW |
+ PIPE_BIND_DISPLAY_TARGET;
+
+ unsigned bpp = util_format_get_blocksizebits(res->format);
+ bool is_2d = (res->target == PIPE_TEXTURE_2D) || (res->target == PIPE_TEXTURE_RECT);
+ bool is_sane_bpp = bpp == 8 || bpp == 16 || bpp == 24 || bpp == 32 || bpp == 64 || bpp == 128;
+ bool should_tile = (res->usage != PIPE_USAGE_STREAM);
+ bool must_tile = (res->bind & PIPE_BIND_DEPTH_STENCIL) && (dev->quirks & MIDGARD_SFBD);
+ bool can_tile = is_2d && is_sane_bpp && ((res->bind & ~valid_binding) == 0);
/* FBOs we would like to checksum, if at all possible */
- bool can_checksum = !(res->bind & (PIPE_BIND_SCANOUT | PIPE_BIND_SHARED));
+ bool can_checksum = !(res->bind & ~valid_binding);
bool should_checksum = res->bind & PIPE_BIND_RENDER_TARGET;
pres->checksummed = can_checksum && should_checksum;
/* Set the layout appropriately */
- pres->layout = should_tile ? PAN_TILED : PAN_LINEAR;
+ assert(!(must_tile && !can_tile)); /* must_tile => can_tile */
+ pres->layout = ((can_tile && should_tile) || must_tile) ? MALI_TEXTURE_TILED : MALI_TEXTURE_LINEAR;
size_t bo_size;
/* 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_bo_create(screen, bo_size, PAN_BO_DELAY_MMAP);
+ pres->bo = pan_bo_create(dev, bo_size, PAN_BO_DELAY_MMAP);
}
void
return panfrost_create_scanout_res(screen, template);
struct panfrost_resource *so = rzalloc(screen, struct panfrost_resource);
- struct panfrost_screen *pscreen = (struct panfrost_screen *) screen;
+ struct panfrost_device *dev = pan_device(screen);
so->base = *template;
so->base.screen = screen;
+ so->internal_format = template->format;
pipe_reference_init(&so->base.reference, 1);
util_range_init(&so->valid_buffer_range);
- panfrost_resource_create_bo(pscreen, so);
+ panfrost_resource_create_bo(dev, so);
panfrost_resource_reset_damage(so);
+ if (template->bind & PIPE_BIND_INDEX_BUFFER)
+ so->index_cache = rzalloc(so, struct panfrost_minmax_cache);
+
return (struct pipe_resource *)so;
}
panfrost_resource_destroy(struct pipe_screen *screen,
struct pipe_resource *pt)
{
- struct panfrost_screen *pscreen = pan_screen(screen);
+ struct panfrost_device *dev = pan_device(screen);
struct panfrost_resource *rsrc = (struct panfrost_resource *) pt;
if (rsrc->scanout)
- renderonly_scanout_destroy(rsrc->scanout, pscreen->ro);
+ renderonly_scanout_destroy(rsrc->scanout, dev->ro);
if (rsrc->bo)
panfrost_bo_unreference(rsrc->bo);
+ if (rsrc->slices[0].checksum_bo)
+ panfrost_bo_unreference(rsrc->slices[0].checksum_bo);
+
util_range_destroy(&rsrc->valid_buffer_range);
ralloc_free(rsrc);
}
+
static void *
panfrost_transfer_map(struct pipe_context *pctx,
struct pipe_resource *resource,
const struct pipe_box *box,
struct pipe_transfer **out_transfer)
{
+ struct panfrost_context *ctx = pan_context(pctx);
int bytes_per_pixel = util_format_get_blocksize(resource->format);
struct panfrost_resource *rsrc = pan_resource(resource);
struct panfrost_bo *bo = rsrc->bo;
/* If we haven't already mmaped, now's the time */
panfrost_bo_mmap(bo);
- /* Check if we're bound for rendering and this is a read pixels. If so,
- * we need to flush */
-
- struct panfrost_context *ctx = pan_context(pctx);
- struct pipe_framebuffer_state *fb = &ctx->pipe_framebuffer;
-
- bool is_bound = false;
-
- for (unsigned c = 0; c < fb->nr_cbufs; ++c) {
- /* 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))
- assert(level == 0);
-
- /* TODO: Respect usage flags */
+ if (pan_debug & (PAN_DBG_TRACE | PAN_DBG_SYNC))
+ pandecode_inject_mmap(bo->gpu, bo->cpu, bo->size, NULL);
if (usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE) {
/* If the BO is used by one of the pending batches or if it's
*/
if (panfrost_pending_batches_access_bo(ctx, bo) ||
!panfrost_bo_wait(bo, 0, PAN_BO_ACCESS_RW)) {
- struct panfrost_screen *screen = pan_screen(pctx->screen);
+ struct panfrost_device *dev = pan_device(pctx->screen);
/* We want the BO to be MMAPed. */
uint32_t flags = bo->flags & ~PAN_BO_DELAY_MMAP;
struct panfrost_bo *newbo = NULL;
* doing to it.
*/
if (!(bo->flags & (PAN_BO_IMPORTED | PAN_BO_EXPORTED)))
- newbo = panfrost_bo_create(screen, bo->size,
+ newbo = pan_bo_create(dev, bo->size,
flags);
if (newbo) {
} else if (usage & PIPE_TRANSFER_READ) {
panfrost_flush_batches_accessing_bo(ctx, bo, PAN_BO_ACCESS_WRITE);
panfrost_bo_wait(bo, INT64_MAX, PAN_BO_ACCESS_WRITE);
- } else {
- /* Why are you even mapping?! */
}
}
- if (rsrc->layout != PAN_LINEAR) {
+ if (rsrc->layout != MALI_TEXTURE_LINEAR) {
/* Non-linear resources need to be indirectly mapped */
if (usage & PIPE_TRANSFER_MAP_DIRECTLY)
transfer->base.stride = box->width * bytes_per_pixel;
transfer->base.layer_stride = transfer->base.stride * box->height;
- transfer->map = rzalloc_size(transfer, transfer->base.layer_stride * box->depth);
+ transfer->map = ralloc_size(transfer, transfer->base.layer_stride * box->depth);
assert(box->depth == 1);
if ((usage & PIPE_TRANSFER_READ) && rsrc->slices[level].initialized) {
- if (rsrc->layout == PAN_AFBC) {
+ if (rsrc->layout == MALI_TEXTURE_AFBC) {
DBG("Unimplemented: reads from AFBC");
- } else if (rsrc->layout == PAN_TILED) {
+ } else if (rsrc->layout == MALI_TEXTURE_TILED) {
panfrost_load_tiled_image(
transfer->map,
bo->cpu + rsrc->slices[level].offset,
- box,
+ box->x, box->y, box->width, box->height,
transfer->base.stride,
rsrc->slices[level].stride,
- util_format_get_blocksize(resource->format));
+ resource->format);
}
}
return transfer->map;
} else {
+ /* Direct, persistent writes create holes in time for
+ * caching... I don't know if this is actually possible but we
+ * should still get it right */
+
+ unsigned dpw = PIPE_TRANSFER_MAP_DIRECTLY | PIPE_TRANSFER_WRITE | PIPE_TRANSFER_PERSISTENT;
+
+ if ((usage & dpw) == dpw && rsrc->index_cache)
+ return NULL;
+
transfer->base.stride = rsrc->slices[level].stride;
- transfer->base.layer_stride = rsrc->cubemap_stride;
+ transfer->base.layer_stride = panfrost_get_layer_stride(
+ rsrc->slices, rsrc->base.target == PIPE_TEXTURE_3D,
+ rsrc->cubemap_stride, level);
/* By mapping direct-write, we're implicitly already
* initialized (maybe), so be conservative */
- if ((usage & PIPE_TRANSFER_WRITE) && (usage & PIPE_TRANSFER_MAP_DIRECTLY))
+ if ((usage & PIPE_TRANSFER_WRITE) && (usage & PIPE_TRANSFER_MAP_DIRECTLY)) {
rsrc->slices[level].initialized = true;
+ panfrost_minmax_cache_invalidate(rsrc->index_cache, &transfer->base);
+ }
return bo->cpu
+ rsrc->slices[level].offset
- + transfer->base.box.z * rsrc->cubemap_stride
+ + transfer->base.box.z * transfer->base.layer_stride
+ transfer->base.box.y * rsrc->slices[level].stride
+ transfer->base.box.x * bytes_per_pixel;
}
struct panfrost_bo *bo = prsrc->bo;
if (transfer->usage & PIPE_TRANSFER_WRITE) {
- if (prsrc->layout == PAN_AFBC) {
+ if (prsrc->layout == MALI_TEXTURE_AFBC) {
DBG("Unimplemented: writes to AFBC\n");
- } else if (prsrc->layout == PAN_TILED) {
+ } else if (prsrc->layout == MALI_TEXTURE_TILED) {
assert(transfer->box.depth == 1);
panfrost_store_tiled_image(
bo->cpu + prsrc->slices[transfer->level].offset,
trans->map,
- &transfer->box,
+ transfer->box.x, transfer->box.y,
+ transfer->box.width, transfer->box.height,
prsrc->slices[transfer->level].stride,
transfer->stride,
- util_format_get_blocksize(prsrc->base.format));
+ prsrc->base.format);
}
}
}
transfer->box.x,
transfer->box.x + transfer->box.width);
+ panfrost_minmax_cache_invalidate(prsrc->index_cache, transfer);
+
/* Derefence the resource */
pipe_resource_reference(&transfer->resource, NULL);
}
static enum pipe_format
-panfrost_resource_get_internal_format(struct pipe_resource *prsrc) {
- return prsrc->format;
+panfrost_resource_get_internal_format(struct pipe_resource *rsrc) {
+ struct panfrost_resource *prsrc = (struct panfrost_resource *) rsrc;
+ return prsrc->internal_format;
}
static bool
PIPE_TEX_FILTER_LINEAR);
/* If the blit was successful, flush once more. If it wasn't, well, let
- * the state tracker deal with it. */
+ * the gallium frontend deal with it. */
if (blit_res) {
panfrost_flush_batches_accessing_bo(ctx, rsrc->bo, PAN_BO_ACCESS_WRITE);
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;
+ bool is_3d = rsrc->base.target == PIPE_TEXTURE_3D;
+ return rsrc->bo->gpu + panfrost_texture_offset(rsrc->slices, is_3d, rsrc->cubemap_stride, level, face);
}
/* Given a resource that has already been allocated, hint that it should use a
void
panfrost_resource_hint_layout(
- struct panfrost_screen *screen,
+ struct panfrost_device *dev,
struct panfrost_resource *rsrc,
- enum panfrost_memory_layout layout,
+ enum mali_texture_layout layout,
signed weight)
{
/* Nothing to do, although a sophisticated implementation might store
/* Check if the preferred layout is legal for this buffer */
- if (layout == PAN_AFBC) {
+ if (layout == MALI_TEXTURE_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 we grew in size, reallocate the BO */
if (new_size > rsrc->bo->size) {
panfrost_bo_unreference(rsrc->bo);
- rsrc->bo = panfrost_bo_create(screen, new_size, PAN_BO_DELAY_MMAP);
+ rsrc->bo = pan_bo_create(dev, new_size, PAN_BO_DELAY_MMAP);
}
+
+ /* TODO: If there are textures bound, regenerate their descriptors */
}
static void
};
void
-panfrost_resource_screen_init(struct panfrost_screen *pscreen)
+panfrost_resource_screen_init(struct pipe_screen *pscreen)
{
//pscreen->base.resource_create_with_modifiers =
// panfrost_resource_create_with_modifiers;
- pscreen->base.resource_create = u_transfer_helper_resource_create;
- pscreen->base.resource_destroy = u_transfer_helper_resource_destroy;
- 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,
+ pscreen->resource_create = u_transfer_helper_resource_create;
+ pscreen->resource_destroy = u_transfer_helper_resource_destroy;
+ pscreen->resource_from_handle = panfrost_resource_from_handle;
+ pscreen->resource_get_handle = panfrost_resource_get_handle;
+ pscreen->transfer_helper = u_transfer_helper_create(&transfer_vtbl,
true, false,
true, true);
}