[FFmpeg-devel] [PATCH] avcodec/mips: MSA (MIPS-SIMD-Arch) optimizations for VP9 bilinear functions

[Shivraj Patil]

Hi,

On Mon, Jul 27, 2015 at 7:59 AM, <shivraj.patil at imgtec.com<mailto:shivraj.patil at imgtec.com>> wrote:
From: Shivraj Patil <shivraj.patil at imgtec.com<mailto:shivraj.patil at imgtec.com>>

Signed-off-by: Shivraj Patil <shivraj.patil at imgtec.com<mailto:shivraj.patil at imgtec.com>>
---
 libavcodec/mips/vp9_mc_msa.c       | 2123 ++++++++++++++++++++++++++++++++++++
 libavcodec/mips/vp9dsp_init_mips.c |    2 +
 libavcodec/mips/vp9dsp_mips.h      |   32 +
 3 files changed, 2157 insertions(+)
[..]
+void ff_avg_bilin_4h_msa(uint8_t *dst, ptrdiff_t dst_stride,
+                         const uint8_t *src, ptrdiff_t src_stride,
+                         int height, int mx, int my)
+{
+    const int8_t *filter = vp9_bilinear_filters_msa[mx - 1];
+
+    if (4 == height) {
+        common_hz_2t_and_aver_dst_4x4_msa(src, src_stride, dst, dst_stride,
+                                          filter);
+    } else if (8 == height) {
+        common_hz_2t_and_aver_dst_4x8_msa(src, src_stride, dst, dst_stride,
+                                          filter);
+    }
+}

You're using this construct in various places, how much does it help?

(Otherwise no comments, basically lgtm % the above.)


Shivraj:- For 8 height case, it helps to reduce stalls (perf gain ~20%), as compared to calling 4 height function twice.