diff options
| author | Steinar H. Gunderson <sgunderson@bigfoot.com> | 2019-03-18 23:52:11 +0100 |
|---|---|---|
| committer | Xiang, Haihao <haihao.xiang@intel.com> | 2019-04-29 10:25:31 +0800 |
| commit | f6573cb75a1c6d57c35f22a8b5d8ce134a9d3d86 (patch) | |
| tree | 13d6355f46c2857e6a1a4d491b2b7e7aa8540a04 | |
| parent | 881e67a49b14631cb24a3880a28af9e5e5bacb96 (diff) | |
| download | libva-intel-driver-f6573cb75a1c6d57c35f22a8b5d8ce134a9d3d86.tar.gz | |
i965_encoder: Speed up i965_MapBuffer for JPEG encoding somewhat.
Searching for the EOI marker byte-by-byte turned out to notable in
profiles when encoding large amounts of JPEGs (~15% of a core at 480 fps
of 1080p images). Using memmem() will typically give us an AVX-optimized
version of at least finding the 0xFF character, which is much more
efficient. It seems to speed up this part by about 3–4x in practice,
taking it largely off the profiles.
Signed-off-by: Steinar H. Gunderson <steinar+nageru@gunderson.no>
| -rw-r--r-- | src/i965_drv_video.c | 12 |
1 files changed, 6 insertions, 6 deletions
diff --git a/src/i965_drv_video.c b/src/i965_drv_video.c index e1b688ae..4aa8da7b 100644 --- a/src/i965_drv_video.c +++ b/src/i965_drv_video.c @@ -3116,12 +3116,12 @@ i965_MapBuffer(VADriverContextP ctx, } if (coded_buffer_segment->codec == CODEC_JPEG) { - for (i = 0; i < obj_buffer->size_element - header_offset - 1 - 0x1000; i++) { - if ((buffer[i] == 0xFF) && (buffer[i + 1] == 0xD9)) { - break; - } - } - coded_buffer_segment->base.size = i + 2; + int len = obj_buffer->size_element - header_offset - 1 - 0x1000; + unsigned char *end_of_file_marker = memmem(buffer, len, "\xff\xd9", 2); + if (end_of_file_marker == NULL) + coded_buffer_segment->base.size = len + 2; + else + coded_buffer_segment->base.size = (end_of_file_marker - buffer) + 2; } else if (coded_buffer_segment->codec != CODEC_VP8) { /* vp8 coded buffer size can be told by vp8 internal statistics buffer, so it don't need to traversal the coded buffer */ |