tree 7fcf5038e6bbec324e2ac968e23582aa544ee56c
parent b7c1a98e50cdfd4c7adc67fb5584c5f182b69c9a
author Albin Bernhardsson <albin.bernhardsson@arm.com> 1713799544 +0000
committer Dawn LUCI CQ <dawn-scoped@luci-project-accounts.iam.gserviceaccount.com> 1713799544 +0000

Vulkan: Implement fast path for writeBuffer

Implement a fast path for writeBuffer in Vulkan where, instead of
writing data to a scratch buffer and then copying the data into the
destination buffer, the data is written directly into the destination
buffer. This reduces bandwidth by avoiding the extra copy and removes
the need to synchronize the transfer command, improving
synchronization.

To take the fast path:
1. The buffer's memory needs to be host visible. In practice this
   means this optimization is for integrated GPUs where memory is
   shared.
2. The buffer must not currently be in use. In practice this means
   the application needs to double buffer its uniform buffers.
3. There must not be any pending GPU writes to the buffer.

If the fast path can't be taken we will fall back to the current
implementation (writing to a scratch buffer and copying the data to
the destination buffer).

In a test Sponza scene, where the application does not manually
suballocate buffers, this reduces frame times by 11% on a Mali-G78
device.

Bug: dawn:774 dawn:851
Change-Id: Ib5f770297d5729df374d0f847a47ade6b0bf24ef
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/162780
Reviewed-by: Austin Eng <enga@chromium.org>
Commit-Queue: Austin Eng <enga@chromium.org>