tree 300b403694be52c688a2f40de5a06571c510ad5f
parent 14fca80f683e8f33768f245d4141cbb96e4f1801
author Nicolas Capens <capn@google.com> 1634331910 -0400
committer Nicolas Capens <nicolascapens@google.com> 1634618873 +0000

Avoid recompiling identical SPIR-V binaries

We previously took the serial ID from the unoptimized SPIR-V module, for
use as part of the Routine cache keys. If a pipeline is created with
specialization constants, we assumed them to produce unique optimized
SPIR-V binaries, regardless of their actual values, and obtained a new
serial ID each time. If a shader module was used with identical
specialization data, we would redundantly JIT-compile the same routine if
no other state used by the pipeline stage was different.

This change takes advantage of the fact that the pipeline cache doesn't
add a new entry (of optimized SPIR-V) for an already optimized module
unless the values of the specialization constants are different and
thus would produce a distinct binary (assuming said specialization
constants affect the code). A new serial ID is now associated with every
new SPIR-V binary. For a (pipeline) cache hit, the existing entry's
ID is used as part of the lookup key for the routine cache.

Note an application has to use a VkPipelineCache object to take
advantage of this.

When there's no pipeline cache, we still inherit the ID from the
unoptimized binary, to avoid Reactor routine recompiles.

Bug: b/172839674
Change-Id: I18b937800ea021235081bba0d827436265816414
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/57988
Kokoro-Result: kokoro <noreply+kokoro@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Tested-by: Nicolas Capens <nicolascapens@google.com>