* When opening a capture file, a format is now available to allow
easy import from another format without a completely different
interface. Only rdc files can be replayed, but any other file can
load and access structured data through the same interface.
* The replay initialisation and capture writing interfaces also use the
RDCFile instead of passing filenames or Serialisers around directly.
Driver initialisation parameters are now entirely private, and don't
need to be exposed - any agnostic metadata like thumbnail, driver, etc
are all accessed via the RDCFile container itself.
* Callstack resolution is now part of the container file, not the
back-end via way of its Serialiser.
* Importers/Exporters to other non-RDC formats are registered in a
similar way to replay/remote drivers.
* It is also then possible to construct an RDC file from thin air, by
creating an empty RDCFile container and filling it with data, then
requesting it to be written to disk.
* Since the initial states will all be written directly to the file
serialiser without intermediate copies, there will be no ability to
seek backwards to fixup chunk lengths.
* Instead all chunks will write an upper bound size estimate first, then
any padding is written at the end if they came in under the estimate.
* The new system contains the ability to export serialised data to a
structured form in memory - and conversion back to serialised bytes.
* This will allow offline transformations/visualisation of capture files
as well as more rich representations of API calls in the UI.
* Likewise it enables a number of optimisations such as the ability to
write straight from mapped API memory to disk via a compressor,
without any intermediate copies.
* The commits following this add a new refactored serialisation system.
As a balance between readability of individual commits and ability to
bisect/compile at interim steps, the individual API back-ends are
disabled in this commit, so that after the serialisation system is
committed the project compiles again as soon as possible.
* However this is of course not much use in itself as then the code
can't function usefully. The drivers are re-enabled as each one is
updated to the new system, so at least the codebase should stay
compiling, even if each individual API may not yet be functioning.
* Normally this wasn't a problem but in a few cases we were doing lots
of push_back calls and this caused horrible resize-by-1 behaviour.
* So instead we double the available count (if that's larger than just
satisfying the request).
* Hopefully these can be restored at some point, when the features are
implemented. For now where possible we remove options that are just
unavailable always, and selectively disable others when they may or
may not be available based on what API the capture uses.
* The DrawcallDescription children is mutable, so we don't need a
separate struct for that purpose in D3D11/GL where there's no side
data to carry along with (where draws are late-bound when submitted).
* This will be used in future to record the exact entry point the application
called, instead of just the function that we serialised (so calls aren't
falsely promoted to DSA variants, etc).
* Where the same enum value is used in multiple cases this then picks
the GL version. This is especially important when the same value is
used for radically different enum values.
* The new serialiser system will not handle the container file itself,
so we've separated it out to deal in sections and stream I/O to read
from and write to sections.
* We also add the ability to choose the compression scheme for a section
instead of always using LZ4. LZ4 will still be used on the fly for
fast write performance, but then RDC files can be re-compressed for
savings offline.
* This I/O will form the basis of the new serialiser - it will simply
read to or write from one of these I/O streams. Then that stream can
come from a file, go to a memory buffer, or go through a compressor
or decompressor transparently.
* It also allows a unified way of writing over sockets instead of
needing special socket helper functions.
* With this commit, the code isn't used aside from in tests.
* Note that while this is public and uses std::string, because it's a
template with specialisations in a .inl the string never crosses a
module boundary - each including module has its own implementation.
* This will be used as part of the upcoming serialisation refactor.
* Some POD structs are still given ToStr implementations as we haven't
yet switched over the serialisation system to expect all structs to
have serialise functions.
* We have some special handling to allow SWIG wrapping of these types:
SDFile owns the chunks and buffers within, and each object owns its
children. Copying is disallowed except from SWIG where we assume the
wrapper is handling lifetime management for its objects externally.
* Previously we would convert from python to C++ arrays immediately by
copying, and vice-versa convert TO python immediately by creating a
new python list by copying.
* This however behaves rather poorly in common situations, e.g.:
> foo.bar.append(5)
Would not append 5 to foo.bar, but copy foo.bar to a temporary, append
5 to it, then destroy it leaving foo.bar untouched.
* Instead we leave the C++ array type as a pointer for as long as we can
and instead implement the python sequence API as extensions/slots that
work in-place on the original array.
* Since these types are more prevalent than originally designed, it
makes more sense to remove the namespace for ease of typing/naming.
* Also add a specialised type 'bytebuf' for an array of bytes.
* This makes mapping easier to SWIG since there's no special casing for
namespaced arrays. Especially so for nested cases like
rdctype::array<rdctype::str> -> rdcarray<rdcstr>
baldurk