Currently avc encoder creates desired number of threads at
the start of every frame and joins after frame is processed.
This change modifies the thread creation part. Now the threads
are created at the start of the sequence. Kept alive through
out the sequence and joined at the end of the sequence. This
helps in reduction of thread creation and deletion overhead.
This change does not effect the encoded bitstream. That is,
encoded output with this change is same as encoded output
without this change.
Bug: 288998933
Test: avcenc -c enc.cfg
Change-Id: I98784169052a5f05c109aaf1de97a5e46d7a773d
'isvce_svc_rc_params_validate' was not being invoked prior to
call to 'isvce_rc_init'. This resulted in an erroneous state
within RC's context wherein the instantaneous estimate for the
texture bits for the frame being processed exceeded INT_MAX.
'isvce_svc_rc_params_validate' has code that detects such a
state and is now being correctly invoked where apprpriate.
Bug = ossfuzz:63175
Test: svc_enc_fuzzer
Before applying the config params (dimensions), the api is validating
the input buffer dimensions. This causes encode fails wrongly. This is
corrected.
The following lines of code have been removed to improve coverage -
[x] Functions -
- isvc_interleaved_copy
- isvc_16bit_interleaved_copy
- isvc_16bit_interleaved_memset
- isvc_iquant_itrans_recon_chroma_4x4_neon
- isvc_iquant_itrans_recon_chroma_4x4_sse42
- isvc_iquant_itrans_recon_4x4_dc_with_res_output_neon
- isvc_iquant_itrans_recon_res_dc_4x4_sse42
- isvc_iquant_itrans_recon_4x4_dc_with_res_accumulate_neon
- isvc_iquant_itrans_recon_res_dc_with_res_acc_4x4_sse42
- isvce_wait_for_thread
[x] Function pointer initialisations for the functions above
Test: svc_enc_fuzzer
For large intra frame intervals or for a choice of bit rate that causes
large bits per frame, it possible that while computing number of bits
available in the gop overflow the datatype bounds. For these
calculations use number_t as this offers wider space
Bug: oss-fuzz-63044
Test: avc_enc_fuzzer
Change-Id: Ifee9130f0021f77905aea67832abc044537088e5
RC uses int32_t to store bitrates and other
parameters internally. For specific magnitudes of
bitrate, framerate, and GOP period, this can
result in signed integer overflow. This is now
detected before calls to RC init.
Note that calls to the 'ISVCE_CMD_CTL_SET_BITRATE' API can also
result in this behaviour but it will be appropriately handled
by 'isvce_svc_frame_params_validate'.
Bug = ossfuzz:63053
Test: svc_enc_fuzzer
If the encoder receives inputs that are empty or frames
that have to be skipped due to frame rate pull down,
do not increment picture count.
Bug: oss-fuzz-62827
Test: ./avc-enc-fuzzer
There were cases where errors in a process thread
executing the last MB row would result in the entropy
job list without a terminnating entry, which would
result in a call to ih264_list_dequeue blocking infinitely.
The enum 'ISVCE_ENTROPY_THREAD_STATES_T' is now used to
monitor the eecution state of the entropy thread and handle
the situation described appropriately.
Bug = ossfuzz:62786
Test: svc_enc_fuzzer
avc-qp(QP) scale is logarithmic and Qstep scale is linear.
The translation between log to linear is carried during
rate control modelling.
Look up tables for these translations are modified as per
the literature
The Lambda tables for intra and inter(P) slices are slightly
modified. Lambda tables for B slices is completely replaced
wih new set.
This change is done in accordance with JVT-O079.doc
This commit does not introduce any new functionality w.r.t previous
commit. But it fixes few things. They are listed below:
1. Guard Bands in header files are fixed
2. Header files contains function definition comments. These are same
as in source file. Maintaining same comment at two locations is
unnecessary. These are removed.
3. Improved consistency and code indentation
4. Removed comments that dont align with implementation
5. Grouped headers of a workspace together
When mb qp delta is not signalled due to cbp being zero,
mb qp delta ctxt is not being reset. This is causing bitstream
decode problems when mb qp is enabled
When the i frame interval is larger than idr frame interval, bitrate
deviations are seen in bitstream.
Forced idr_interval to be a multiple of I frame interval.
When IDR frame is skipped due to VBV constraints, the frame
number should not be reset. Since it was reset earlier, it
was leading to bistream being not decodable in such a case.
Restore frame_num during idr skip.
In frame drop mode, the stream generated is non-compliant.
For now disable frame drop mode.
Bug: oss-fuzz-61202
Bug: oss-fuzz-59598
Test: ./avc_enc_fuzzer
Change-Id: Ifed950082ae2c6e70d6a7da4547b6ffeb44cf759
During bitstream overflow errors, it is possible that update post
enc gets skipped as last row might never have to be entropy coded
as we have already consumed the buffer. This can be fatal for
encoding further frames. Move this update to a location where it
is guaranteed to be updated
Bug: oss-fuzz-62378
Test: ./avc_enc_fuzzer
Change-Id: I279deb1f30c15c057c0b6e909834313a1df4fdac
As bitstream overflow errors are not marked as fatal errors and
these recons are being used for reference, its best to return
the partial encoded buffers for application
Change-Id: I30a8e9907cf42d3bc883ee78b28cbae723bad7ac
During entropy error, even though the bitstream buffer is not written
with additional bytes, the bitstream context state is getting updated.
As the number of bits left in current word is updated, the get_num_bits
used for computing header and texture bits can become negative causing
overflow errors
Change-Id: I2f990071a9935b2ee328dbd3915dfbefccbab4c5
At the end of encoding of a frame, the entropy thread communicates
the encoded bit stream size to rc module for update. After this
update, if rc decides to skip the frame due to vbv overflow, the
bitstream context is reset and frame is marked for skip.
Due to an oversight, if entropy encoding sees an error, then this
update is happening at the end of each row. Now rc has decided to
skip the frame and the context is reset. As the bitstream context is
reset, other threads are unaware of this problem and continue encoding.
This is causing issues.
Restrict the rc update to the thread that entropy code the last row.
Bug: oss-fuzz:59543
Bug: 285891354
Test: avc_enc_fuzzer
Change-Id: If45a5f34abb59ece812733af8f54f72ae5474d03
A 'consumption ratio' is computed fr every MB when
in-frame RC is enabled. This computation can result in
divide-by-zeros in certain circumstances. Such cases
are now appropriately handled.
Bug: ossFuzz:60828
Test: svc_enc_fuzzer
During encoding, if encoder encounters an error, the library returns the
same. If the application ignores it and feeds further input, library
while processing the new input is facing a null dereference.
1. Added a error check in encode API call before processing input
2. Made entropy bitstream buffer overflow errors non-fatal.
Bug: 187499509
Test: POC in bug descriptions
atest VtsHalMediaC2V1_0TargetVideoEncTest
atest -c CtsMediaV2TestCases:CodecEncoderTest
In some cases, s_inp_buf and s_out_buf on stack in ih264e_encode()
can be accessed unininitialized. This is fixed by initializing these
two structures.
Bug: oss-fuzz:57333
Bug: 274906999
Test: avc_enc_fuzzer
The ILP MV struct pointer will be set to NULL for I slices and
for spatial layer ID 0. A NULL check ought to be used in all places
that access this pointer. This was missing in 2 places and has been
added.
Test: svc_enc_fuzzer
RC uses int32_t for storing multiple quantities related to
bits in a given period.
'isvce_svc_frame_params_validate' has been added, which queries
RC API for the relevant quantities and returns with and error
if computations involving those quantities exceed INT32_MAX.
Bug: 274221347
Bug: 274265498
Test: svc_enc_fuzzer
Using incorrect stride during copy from source buffer to intermediate buffer. This is corrected
Bug: 242379731
Bug: 242386193
Test: avcenc -c enc.cfg
Change-Id: I307525906ee6f6df5d8e114ebbafa40d6442662d
(cherry picked from commit 0a87a34fccffe15c2cc80670cf1058e7a3313c6a)
Merged-In: I307525906ee6f6df5d8e114ebbafa40d6442662d
If 'u4_strm_buf_offset < 4' when 'isvce_cabac_flush' is called,
then EPB could not have been inserted into the stream buffer.
BUG = ossfuzz:56816
Test: svc_enc_fuzzer
- Only libavcdec is marked as available to apex modules instead
of marking all decoder libraries to be available to apex modules.
- some formatting changes for consistency with neighboring lines.
Test: Builds
If 'u4_strm_buf_offset == 0' when 'isvce_cabac_flush' is called,
then 'carry' is implicitly 0.
'u4_strm_buf_offset == 0' implies stream buffer is empty.
Invocation of 'isvce_cabac_flush' implies all MB's in a slice
have been encoded.
BUG = ossfuzz:56137
Test: svc_enc_fuzzer
- Add SII flag and SII parameters for the encoder and decoder.
- Encoder: Added support for SII SEI
- Decoder: Added support for SII SEI parsing and exporting
For certain configurations of the fuzzer input, the minimum
output buffer size is insufficient.
The minimum value has been doubled in this commit.
BUG=oss-fuzz:55797
Test: svc_enc_fuzzer
