[opus] Initial implementation of ch.mapping 253/3

[Mark Harris]

On Tue, May 30, 2017 at 3:13 PM, Drew Allen <bitllama at google.com> wrote:
> Hello all,
>
> Attached is the initial proposed implementation for ch.mapping 253/3, based
> on the IETF proposal:
> https://tools.ietf.org/html/draft-ietf-codec-ambisonics-03
>
> A brief overview of the patch, as it is slightly lengthy:
> After discussion with Jean-Marc, we determined that ch.253/3 will need the
> demixing matrix as part of the encoder/decoder struct stack and thus will
> require a new Encoder/Decoder structure. For this, I've proposed
> OpusPEnc/OpusPDec (for "projection"), which wraps OpusMSEnc/OpusMSDec aside
> from the addition of the mixing matrix structure. The functionality also
> wraps the functionality of the multistream API aside from matrix transforms
> during encode/decode.
>
> We have included 3 mixing/demixing matrix pairs that are sufficient for
> coding/decoding 1st/2nd/3rd order content respecitvely.
>
> I've modified the multistream api slightly, first to include some
> supplemental API extensions to allow wrapping and moved the declaration of
> the MSEnc/MSDec to opus_private so that the projection API can compile.
>
> Some basic tests to ensure matrix multiplication and encoding/decoding using
> projection has been included.
>
> I am of course happy and eager to take any and all
> suggestions/comments/questions/recommendations regarding the patch,
> including naming conventions/style/etc.
>
> Looking forward to your feedback,
>
> Cheers,
> Drew


Hi Drew,

I ran into some issues with this new test.

Building the test for floating point produces a clang compiler warning:

  CC       tests/test_opus_projection.o
tests/test_opus_projection.c:137:9: warning: using floating point
absolute value function
      'fabs' when argument is of integer type [-Wabsolute-value]
    if (ABS16(a[i] - b[i]) > tolerance)
        ^
./celt/arch.h:194:26: note: expanded from macro 'ABS16'
#define ABS16(x) ((float)fabs(x))
                         ^
tests/test_opus_projection.c:137:9: note: use function 'abs' instead
./celt/arch.h:194:26: note: expanded from macro 'ABS16'
#define ABS16(x) ((float)fabs(x))
                         ^

Also the test fails to compile with -std=c89:

tests/test_opus_projection.c:57:1: error: C++ style comments are not
allowed in ISO C90
 // static const MappingMatrix test_matrix = {
 ^
tests/test_opus_projection.c:57:1: error: (this will be reported only
once per input file)
tests/test_opus_projection.c: In function 'main':
tests/test_opus_projection.c:403:1: warning: control reaches end of
non-void function [-Wreturn-type]
 }
 ^

Running the test, an issue occurs in test_creation_arguments().
opus_projection_ambisonics_encoder_create() is intentionally called
with invalid arguments (initially, channels=0), which causes an error
to be returned and streams and coupled_streams are left uninitialized.
These uninitialized values are passed to
opus_projection_decoder_create(), which does not check that the
arguments are not too large before using them to attempt allocation of
a ridiculous amount of memory.  Results may vary depending on the
uninitialized values of streams and coupled_streams.  Valgrind reports
several uses of uninitialized values such as:

 Conditional jump or move depends on uninitialised value(s)
    at 0x100067B72: opus_multistream_decoder_get_size
(opus_multistream_decoder.c:47)
    by 0x10006B098: opus_projection_decoder_get_size
(opus_projection_decoder.c:54)
    by 0x10006B136: opus_projection_decoder_create
(opus_projection_decoder.c:68)
    by 0x10000140D: test_creation_arguments (test_opus_projection.c:273)
    by 0x100001D06: main (test_opus_projection.c:396)
  Uninitialised value was created by a stack allocation
    at 0x100001330: test_creation_arguments (test_opus_projection.c:252)

Additionally, the demixing_matrix appears to be used uninitialized
when order_plus_one is 1 (or greater than 4):

 Conditional jump or move depends on uninitialised value(s)
    at 0x10006A6DF: opus_projection_encoder_init_impl
(opus_projection_encoder.c:125)
    by 0x10006A9CA: opus_projection_ambisonics_encoder_init
(opus_projection_encoder.c:196)
    by 0x10006AC7B: opus_projection_ambisonics_encoder_create
(opus_projection_encoder.c:261)
    by 0x1000013D9: test_creation_arguments (test_opus_projection.c:268)
    by 0x100001D06: main (test_opus_projection.c:396)
  Uninitialised value was created by a heap allocation
    at 0x10010FE71: malloc (vg_replace_malloc.c:302)
    by 0x10006AB84: opus_alloc (os_support.h:49)
    by 0x10006AC1F: opus_projection_ambisonics_encoder_create
(opus_projection_encoder.c:252)
    by 0x1000013D9: test_creation_arguments (test_opus_projection.c:268)
    by 0x100001D06: main (test_opus_projection.c:396)

Another issue occurs in mapping_matrix_dot_short().  It attempts to
read beyond the end of the input buffer and may cause a segmentation
fault.  It appears that the input buffer contains only 2 channels of
480 samples each, but 16 channels of 960 samples each are being read
from the buffer.  Valgrind reports the following:

 Invalid read of size 2
    at 0x10006B9D8: mapping_matrix_dot_short (mapping_matrix.c:98)
    by 0x10006BABD: mapping_matrix_multiply_short (mapping_matrix.c:119)
    by 0x10006AD5B: opus_projection_encode (opus_projection_encoder.c:283)
    by 0x100001BCB: test_encode_decode (test_opus_projection.c:374)
    by 0x100001D2D: main (test_opus_projection.c:400)
  Address 0x10490b000 is not stack'd, malloc'd or (recently) free'd

 Process terminating with default action of signal 11 (SIGSEGV)
  Access not within mapped region at address 0x10490B000
    at 0x10006B9D8: mapping_matrix_dot_short (mapping_matrix.c:98)
    by 0x10006BABD: mapping_matrix_multiply_short (mapping_matrix.c:119)
    by 0x10006AD5B: opus_projection_encode (opus_projection_encoder.c:283)
    by 0x100001BCB: test_encode_decode (test_opus_projection.c:374)
    by 0x100001D2D: main (test_opus_projection.c:400)

 - Mark