virgl: better support for PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE

When the resource to be mapped is busy and the backing storage can
be discarded, reallocate the backing storage to avoid waiting.

In this new path, we allocate a new buffer, emit a state change,
write, and add the transfer to the queue .  In the
PIPE_TRANSFER_DISCARD_RANGE path, we suballocate a staging buffer,
write, and emit a copy_transfer (which may allocate, memcpy, and
blit internally).  The win might not always be clear.  But another
win comes from that the new path clears res->valid_buffer_range and
does not clear res->clean_mask.  This makes it much more preferable
in scenarios such as

  access = enough_space ? GL_MAP_UNSYNCHRONIZED_BIT :
                          GL_MAP_INVALIDATE_BUFFER_BIT;
  glMapBufferRange(..., GL_MAP_WRITE_BIT | access);
  memcpy(...); // append new data
  glUnmapBuffer(...);

Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
Reviewed-by: Alexandros Frantzis <alexandros.frantzis@collabora.com>
This commit is contained in:
Chia-I Wu 2019-05-09 21:44:33 -07:00
parent 9975a0a84c
commit 1fece5fa5f
4 changed files with 92 additions and 25 deletions

View File

@ -47,6 +47,13 @@ static void *virgl_buffer_transfer_map(struct pipe_context *ctx,
map_type = virgl_resource_transfer_prepare(vctx, trans);
switch (map_type) {
case VIRGL_TRANSFER_MAP_REALLOC:
if (!virgl_resource_realloc(vctx, vbuf)) {
map_addr = NULL;
break;
}
vs->vws->resource_reference(vs->vws, &trans->hw_res, vbuf->hw_res);
/* fall through */
case VIRGL_TRANSFER_MAP_HW_RES:
trans->hw_res_map = vs->vws->resource_map(vs->vws, vbuf->hw_res);
if (trans->hw_res_map)

View File

@ -87,13 +87,9 @@ virgl_resource_transfer_prepare(struct virgl_context *vctx,
struct virgl_winsys *vws = vs->vws;
struct virgl_resource *res = virgl_resource(xfer->base.resource);
enum virgl_transfer_map_type map_type = VIRGL_TRANSFER_MAP_HW_RES;
bool unsynchronized = xfer->base.usage & PIPE_TRANSFER_UNSYNCHRONIZED;
bool discard = xfer->base.usage & (PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE |
PIPE_TRANSFER_DISCARD_RANGE);
bool flush;
bool readback;
bool wait;
bool copy_transfer;
/* there is no way to map the host storage currently */
if (xfer->base.usage & PIPE_TRANSFER_MAP_DIRECTLY)
@ -110,19 +106,10 @@ virgl_resource_transfer_prepare(struct virgl_context *vctx,
flush = virgl_res_needs_flush(vctx, xfer);
readback = virgl_res_needs_readback(vctx, res, xfer->base.usage,
xfer->base.level);
/* Check if we should perform a copy transfer through the transfer_uploader. */
copy_transfer = discard &&
!readback &&
!unsynchronized &&
vctx->transfer_uploader &&
!vctx->transfer_uploader_in_use &&
(flush || vws->resource_is_busy(vws, res->hw_res));
/* We need to wait for all cmdbufs, current or previous, that access the
* resource to finish unless synchronization is disabled.
*/
wait = !unsynchronized;
wait = !(xfer->base.usage & PIPE_TRANSFER_UNSYNCHRONIZED);
/* When the transfer range consists of only uninitialized data, we can
* assume the GPU is not accessing the range and readback is unnecessary.
@ -135,17 +122,42 @@ virgl_resource_transfer_prepare(struct virgl_context *vctx,
flush = false;
readback = false;
wait = false;
copy_transfer = false;
}
/* When performing a copy transfer there is no need wait for the target
* resource. There is normally no need to flush either, unless the amount of
* memory we are using for staging resources starts growing, in which case
* we want to flush to keep our memory consumption in check.
/* When the resource is busy but its content can be discarded, we can
* replace its HW resource or use a staging buffer to avoid waiting.
*/
if (copy_transfer) {
flush = (vctx->queued_staging_res_size > VIRGL_QUEUED_STAGING_RES_SIZE_LIMIT);
wait = false;
if (wait && (xfer->base.usage & (PIPE_TRANSFER_DISCARD_RANGE |
PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE))) {
const bool can_realloc =
(xfer->base.usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE) &&
virgl_can_rebind_resource(vctx, &res->u.b);
const bool can_staging = vctx->transfer_uploader &&
!vctx->transfer_uploader_in_use;
/* discard implies no readback */
assert(!readback);
if (can_realloc || can_staging) {
/* Both map types have some costs. Do them only when the resource is
* (or will be) busy for real. Otherwise, set wait to false.
*/
wait = (flush || vws->resource_is_busy(vws, res->hw_res));
if (wait) {
map_type = (can_realloc) ?
VIRGL_TRANSFER_MAP_REALLOC :
VIRGL_TRANSFER_MAP_STAGING;
wait = false;
/* There is normally no need to flush either, unless the amount of
* memory we are using for staging resources starts growing, in
* which case we want to flush to keep our memory consumption in
* check.
*/
flush = (vctx->queued_staging_res_size >
VIRGL_QUEUED_STAGING_RES_SIZE_LIMIT);
}
}
}
/* readback has some implications */
@ -185,9 +197,6 @@ virgl_resource_transfer_prepare(struct virgl_context *vctx,
if (wait)
vws->resource_wait(vws, res->hw_res);
if (copy_transfer)
map_type = VIRGL_TRANSFER_MAP_STAGING;
return map_type;
}
@ -598,3 +607,39 @@ void *virgl_transfer_uploader_map(struct virgl_context *vctx,
return map_addr;
}
bool
virgl_resource_realloc(struct virgl_context *vctx, struct virgl_resource *res)
{
struct virgl_screen *vs = virgl_screen(vctx->base.screen);
const struct pipe_resource *templ = &res->u.b;
unsigned vbind;
struct virgl_hw_res *hw_res;
vbind = pipe_to_virgl_bind(vs, templ->bind, templ->flags);
hw_res = vs->vws->resource_create(vs->vws,
templ->target,
templ->format,
vbind,
templ->width0,
templ->height0,
templ->depth0,
templ->array_size,
templ->last_level,
templ->nr_samples,
res->metadata.total_size);
if (!hw_res)
return false;
vs->vws->resource_reference(vs->vws, &res->hw_res, NULL);
res->hw_res = hw_res;
util_range_set_empty(&res->valid_buffer_range);
/* count toward the staging resource size limit */
vctx->queued_staging_res_size += res->metadata.total_size;
virgl_rebind_resource(vctx, &res->u.b);
return true;
}

View File

@ -73,10 +73,14 @@ struct virgl_resource {
enum virgl_transfer_map_type {
VIRGL_TRANSFER_MAP_ERROR = -1,
VIRGL_TRANSFER_MAP_HW_RES,
/* Map a range of a staging buffer. The updated contents should be transferred
* with a copy transfer.
*/
VIRGL_TRANSFER_MAP_STAGING,
/* Reallocate the underlying virgl_hw_res. */
VIRGL_TRANSFER_MAP_REALLOC,
};
struct virgl_transfer {
@ -187,4 +191,8 @@ void virgl_resource_dirty(struct virgl_resource *res, uint32_t level);
void *virgl_transfer_uploader_map(struct virgl_context *vctx,
struct virgl_transfer *vtransfer);
bool
virgl_resource_realloc(struct virgl_context *vctx,
struct virgl_resource *res);
#endif

View File

@ -134,6 +134,13 @@ static void *texture_transfer_map_plain(struct pipe_context *ctx,
map_type = virgl_resource_transfer_prepare(vctx, trans);
switch (map_type) {
case VIRGL_TRANSFER_MAP_REALLOC:
if (!virgl_resource_realloc(vctx, vtex)) {
map_addr = NULL;
break;
}
vws->resource_reference(vws, &trans->hw_res, vtex->hw_res);
/* fall through */
case VIRGL_TRANSFER_MAP_HW_RES:
trans->hw_res_map = vws->resource_map(vws, vtex->hw_res);
if (trans->hw_res_map)