Even though the DMA buffer itself does not move the ISR copies from a movable
buffer into the static commit buffer. To ensure this copying yields consistent
data it must not be interrupted by this ISR..

Also bump the commit buffer size to 2k, this should reduce the overhead
considerably because many clips are smaller than that (especially on
swcodec/speex).

Change-Id: I4e1ab83074f31fc91b51a58baa1df55ce659ac73

Change-Id: If3cbcd7557797684f1b7fea8e1bb134777680dee

The voice engine can now request more voice data during decoding, it does
not require the entire clip to be available before start of decoding anymore.
Therefore the commit buffer does not need to hold an entire voice clip anymore,
and can be made greatly smaller.

Change-Id: I3eca9026448e725b9b8d0dae1efca0ad185371da

This unifies the talk.c for all possible voice payload. .talk clips are placed
onto the same unified clip cache, along with normal clips. This allows for more
effecient memory usage.

The cache handling makes a slight difference between normal clips and .talk
ones: .talk clips can be cached multiple and are always freed first.The extra
logic to avoid loading multiple copies of .talks is not necessary because the
will be freed first anyway.

Change-Id: I88d056a0a613b129f5875f50fdb757b58bac0a42

This unifies the talk.c for all targets. The only separation is left is
TALK_PROGRESSIVE_LOAD: When this is defined the talk buffer will not be
initially prefilled. This is useful for super slow storage or when the buffer
is not large enough to prefill it with useful clips (the prefill code could
be made smarter too).

The buffer size can be adjusted. By default lowmem uses 100k while
other targets load the entire file. The bigger the more clips can be cached
but with diminishing returns.

Change-Id: Ife38fb101c53093117e6638c40c65f7d177a31d4

Previously the clip cache of TALK_PARTIAL_LOAD reserved space N clips, each slot
was as big as the maximum sized clip which was necessary to replace clips
in-memory in MRU-style.

The cache management now uses buflib to allocate and free each clip, using the
clip's real size. This allows the clip cache to be much more compact, because
no space is wasted for the max. sized clip. This makes use of buflib's ability
to easily manage differently-sized memory chunks by moving them to make free
space.

As an example: for english.voice TALK_PARTIAL_LOAD allocated 288k in advance.
for just 64 clips. With this patch ~70 clips can be stored in a 100k buffer.
This, the memory usage is cut by 2/3 and almost optimal (there's still the
buflib per-alloc cookie overhead).

As a result the TALK_PARTIAL_LOAD buffer is restricted to 100k which still
allows for more clips than previously, on average.

Change-Id: I257654071e9a95770cd6db2c2765f020befce412

This engine includes voicefile, memory usage and cache
hits/misses for TALK_PARTIAL_LOAD.

Change-Id: I331981ddda39ea30c57b4b74504accb3c556c3b9

Change-Id: I3a04760d550efab7f011a917597ef29c039b05bd

This is necessary because when voice is active audio is disabled. But only
audio was able to shrink it's buffer to let other memory allocs succeed.
talk needs to be able to do this too when it owns the audio buffer exclusively.

Change-Id: Idea8ab90da7169f977c0c766cccb42c4fe6d6e81

Change-Id: I420155d7f01ca0ea301c0678ac7245d251d365b0