Buffer allocation is based on minimum level required for the
input resolution rather than the input max level.
Number of maximum context sets can be set to 1 to reduce
memory usage.
Added a macro ENC_MIN_PU_SIZE for minimum size of
inter prediction unit supported by encoder.
Changed the maximum constraint on number of MBs for NMB
processing to width in Mbs.
Change-Id: I5a9255e93935d90c13262681aafc772aedf8ae81
Fixed QP getting locked up at particluar values even when
it is supposed to be changing by ensuring that Qp will have
a swing of atleast +-2 from previous value and correcting
the Qp translation tables
Change-Id: Id6f0fa4f282c288b81cfe278f34957a0fa68ff8c
Setting I frame interval to 1 caused encoder to go into
infinite loop. Fixed that by stopping RC update.
Forcing of I and IDR frame did not cause the frames at expected time
due to delayed encoding. This is fixed now.
Change-Id: Ibf549fb335c053ee33de768455b20e46be3e7c1c
This field isn't used at the moment at all. The public field
ive_ctl_set_dimensions_ip_t.u4_strd is kept to avoid breaking
other calling code, although it doesn't have any actual function
any longer.
Change-Id: Icfc88846fa63f209c67b8e1f8e0cc21bc4ddb73a
If horizontal padding is needed, copy the data to avoid having to try
to write into the input buffer.
If vertical padding is needed, we could get away with only copying
the last MB row (as previously was done), but then one extra row of
pixels are needed for intra pred. Thus always copy all of the picture,
for simplicity, if padding is needed.
Change-Id: I6b456a12956120d64e7c6b86d7b571e159f4566b
If the data is read directly from the input buffer, set the stride
values to those of the input buffer. If the data is copied, set the
stride to the value of the local buffer instead.
This fixes handling of cases where the input buffer stride is
(significantly) larger than the maxwidth of the video.
This also effectively makes the parameter
ive_ctl_set_dimensions_ip_t.u4_strd useless - nothing needs to
know the stride until you actually encode a frame, and at that point,
we can either use the stride of the input buffer, or of the local
pu1_y_csc_buf_base where the copied content is stored.
Change-Id: Icde400b4a0867d25855e621e143454e608748aa3
When both luma and chroma are copied to the local buffer (either
due to yuv format conversion, or due to padding), they have got
the same stride, but if chroma is copied while luma is used directly
from the input buffer, they might have different strides.
Therefore add a separate field for chroma stride.
This commit only adds the field, while it still has got the same
value as before.
Change-Id: I0dce97ad4d91cd1d9aba4b4472c6a0de45a314bc
The x18 register is platform specific, and using it in portable assembler
code should be avoided.
In this particular case, using it is not actually needed at all.
Change-Id: I07016fdd3a264ed77eba37abcd1748b51cf7a49e
Some versions of the clang built-in assembler interprets /* as a start
of a multi-line comment, even within lines starting with a @. Therefore,
make sure to terminate these properly.
This fixes warnings about /* within a block comment when building
using the clang built-in assembler (in some versions of clang), and
fixes issues where part of the assembly source was effectively
commented out, when building with this assembler.
Change-Id: I4b520334d92740f8c60def94ee990561b2cb8b7c
This fixes running when built with clang - by pure luck the out of bounds
writes didn't cause any harm when the calling C code was built with gcc.
Change-Id: I4e33dd403b785cbcd24fb80f4db7a86310a32225
binutils gas automatically produces vqmovun when vqshrun is used
with a zero shift.
This is required for building the armv8 assembly with the clang built-in
assembler.
Change-Id: Ie986978e4508a085e4e4c76f8db9a8a8394c4fec
clang only supports the new, unified arm syntax (UAL), while binutils
gas supports both (and defaults to the old syntax). Explicitly declare
that the unified syntax is used, and use the right form of these
instructions that is supported by both.
Change-Id: I9a3940e060c129861c7a542ccf09035dbd1766ae
binutils gas automatically converts cmp with negative values to use
the cmn instruction instead.
This is required for building the armv8 assembly with the clang built-in
assembler.
Change-Id: If6465ea1bec252446c49cf497153e3ef3dad905a
When specifying one specific lane of the vector, the number of
lanes don't need to be specified.
The clang built-in assembler doesn't allow the redundant
declarations, while binutils gas work fine with both forms.
Change-Id: I86077ce0774d4594a1295b6860e4944df87dde2f
If WINDOWS_TIMER is set, windows.h is supposed to be used instead,
and sys/time.h isn't available.
This reduces the amount of duplicated include directives in the test
apps.
Change-Id: I223d80c4ccaed2aab55e3aae66dab903ecacef7a
This function is dangerous since it behaved differently on different
platform implementations. It is not used for any actual purpose, and
not currently necessary.
Change-Id: I9702e9943065235c7a7b29dc6c6eeb36212e7ce1
The ithread_exit wrapper function behaves inconsistently across
platforms; on windows, the argument is a thread handle, denoting
the thread that will be exited. On pthreads, ithread_exit only
terminates the current thread, not any other thread.
These both calls are at the end of each thread function (and the one
with a non-null parameter refers to the handle of the thread itself),
so the thread will exit immediately anyway.
Change-Id: I21e43c56874729452bd80b28ff351e3c5d3f087e
If the environment already indicates that the compiler targets
either x86 or x86_64, we don't need to enforce this by trying to
force the compiler to output a specific bitness.
Change-Id: Ife6e717e90b4da4edd852dcd66ad92dba70939a2
The following defines aren't referenced anywhere in the code base:
_LIB MULTICORE APPLY_CONCEALMENT THREAD_QUAD_CORE DISABLE_NEONINTR
ARMGCC
The DEFAULT_ARCH define isn't used within the encoder at all.
The ANDROID define isn't referenced anywhere either, and if necessary,
the automatically enabled define __ANDROID__ can be used instead.
The defines INSERT_LOGO and LOGO_EN were undefined in the makefiles,
but that is unnecessary since nothing actually defines them. The
decoder x86_64 makefile also undefined LOGO_EN; an architecture
specific makefile shouldn't touch such feature settings, otherwise
there's a risk that different builds for different architectures
behave significantly differently.
Change-Id: I13b86c8bf2feb3a381d904a13f18c3b35f40a575
Added support for CABAC entropy coding.
Added support for B slices.
Fixed an issue in rate control constant QP mode.
Change-Id: Ib759d35e8e943f941aa9b8bbff0362d92c619994
Most of them was benign and didn't risk any data race; they
were used for write-only memory during init.
The array as_itt_memtab in ih264e_get_rate_control_mem_tab
could however cause a real data race, if multiple threads
try to initialize an encoder at the same time, since it
used as temporary storage when mapping between different
structs.
Change-Id: I3ca29f0e3be2fad154becc3ee29867e591975a74
If i4_header_mode is set (without the caller explicitly requesting it),
the next frame to be encoded will be missed altogether (when only
a header is output).
Instead force a header to be generated prepended to the next frame.
If the caller explicitly requests header mode, it will be output
in a separate call.
Change-Id: If475717edf980bccb2076f880c60c9dd7dede904
Each round in the loop uses both argv[i] and argv[i + 1].
This fixes crashes if given an odd number of parameters, 3 or
more.
Change-Id: I9b8be99042eb7d0f69ebd87166d9cc920d15a950
Don't do concatenation with strncat; the length parameter in strncat
is only for how many chars to append at most, not for the full output
buffer size. To safely use strncat, one would have to do
strncat(buf, str, sizeof(buf) - strlen(buf)).
By using snprintf, we guarantee that the buffer is null terminated, and
we don't need to use strnlen at all.
(If compatibility with older MSVC versions that lack snprintf, one
can use _snprintf instead and manually add the null termination.)
Change-Id: I1c2322c7a406ddd5e6551a96c460da60deeffda1
They seem to have been included by accident. Nothing includes
them or refer to the symbols declared in them. (The other functions
declared by the header don't have any matching implementation either.)
Change-Id: I1c77a21226613bda0b029dd4ea5da207aa7cae09
This uses the same mechanism as in ih264e_fmt_conv_420p_to_420sp_a9q
to do an overlapping copy of the last few pixels.
Change-Id: Ic93a00d0683337fd0f6f540119000466685f64bd
Handled non-multiple of 16 dimensions for 420 semiplanar input
Modified test code to remove alignment of width and height
Change-Id: I83ff8165364a863d577fcac81e711b07eec9c004
The function ih264e_get_lvl_idx lacked handling of them
before, which meant that it returned index 0 (corresponding
to level 1.0) if level 5.0 or 5.1 were selected.
Change-Id: I482213e68e11fd091c72ec57a593392aa1a0a4ae
The table values themselves were correct, but the level values
were swapped for 1.1 and 1B.
This fixes encoding using level 1.1, which previously got the
wrong (too small) sizes (corresponding to level 1B).
Change-Id: Ic0acc58eb5841710ffd8be54bf97800fdbb40855
