[Bug 773073] New: audioconvert: endian conversion optimization

[GStreamer (GNOME Bugzilla)]

https://bugzilla.gnome.org/show_bug.cgi?id=773073

            Bug ID: 773073
           Summary: audioconvert: endian conversion optimization
    Classification: Platform
           Product: GStreamer
           Version: 1.8.3
                OS: Linux
            Status: NEW
          Severity: normal
          Priority: Normal
         Component: gst-plugins-base
          Assignee: gstreamer-bugs at lists.freedesktop.org
          Reporter: brain at jikos.cz
        QA Contact: gstreamer-bugs at lists.freedesktop.org
     GNOME version: ---

Endian conversion in the audioconvert plugin is highly inefficient.
On Cortex A7 @ 500MHz it takes additional 7-8% CPU to convert S24LE to S24BE,
similar results can be observed for 16 and 32 bits. This is unacceptably high.

The current implementation performs 24->32 unpack, 24-bit quantization and
32->24 bit pack. It requires to allocate additional buffers as it cannot run in
place:

0:00:00.264383167   406   0x304320 LOG          audio-converter
audio-converter.c:322:gst_audio_converter_update_config: new rate 0 -> 0
0:00:00.265950834   406   0x304320 INFO         audio-converter
audio-converter.c:863:gst_audio_converter_new: unitsizes: 6 -> 6
0:00:00.266781125   406   0x304320 INFO         audio-converter
audio-converter.c:566:chain_unpack: unpack format S24LE to S32LE
0:00:00.267911417   406   0x304320 INFO         audio-converter
audio-converter.c:623:chain_mix: mix format S32LE, passthrough 1, in_channels
2, out_channels 2
0:00:00.268835292   406   0x304320 INFO         audio-converter
audio-converter.c:674:chain_quantize: depth in 32, out 24
0:00:00.269497709   406   0x304320 INFO         audio-converter
audio-converter.c:686:chain_quantize: using no dither and noise shaping
0:00:00.270300750   406   0x304320 INFO         audio-converter
audio-converter.c:698:chain_quantize: quantize to 24 bits, dither 0, ns 0
0:00:00.271145542   406   0x304320 INFO         audio-converter
audio-converter.c:721:chain_pack: pack format S32LE to S24BE
0:00:00.271963250   406   0x304320 INFO         audio-converter
audio-converter.c:884:gst_audio_converter_new: do full conversion
0:00:00.272734250   406   0x304320 LOG          audio-converter
audio-converter.c:747:setup_allocators: chain 0x3016c8: 1 1
0:00:00.273480709   406   0x304320 LOG          audio-converter
audio-converter.c:747:setup_allocators: chain 0x301688: 1 0
0:00:00.277083625   406   0x304320 INFO            audioconvert
gstaudioconvert.c:224:gst_audio_convert_get_unit_size:<audioconvert0> unit_size
= 6
0:00:00.278164584   406   0x304320 INFO            audioconvert
gstaudioconvert.c:224:gst_audio_convert_get_unit_size:<audioconvert0> unit_size
= 6

Similar behaviour can be observed with 16 and 32 bits.

On the other hand with SIMD instructions the conversion can be done very
efficiently. E.g. only 3 NEON instructions are needed to convert eight 24-bit
values in place:
VLD3.8
VSWP d0,d2
VST3.8 

or to convert sixteen 16-bit values in place:
VLD2.16
VSWP d0,d1
VST2.16

The attempt here is to discuss the best way how to integrate these instructions
into the audioconvert module, still keeping the portability. Also without SIMD
instructions the code needs to be optimized (e.g. to perform the conversion in
place).
Eventually the discussion should result into a patch.

-- 
You are receiving this mail because:
You are the QA Contact for the bug.
You are the assignee for the bug.