[FFmpeg-cvslog] arm: vp9itxfm: Optimize 16x16 and 32x32 idct dc by unrolling

[Martin Storsjö]

ffmpeg | branch: master | Martin Storsjö <martin at martin.st> | Wed Jan  4 13:08:51 2017 +0200| [758302e4bc14e93989e7feb1135ec3f807c3310d] | committer: Martin Storsjö

arm: vp9itxfm: Optimize 16x16 and 32x32 idct dc by unrolling

This work is sponsored by, and copyright, Google.

Before:                            Cortex A7      A8      A9     A53
vp9_inv_dct_dct_16x16_sub1_add_neon:   273.0   189.5   211.7   235.8
vp9_inv_dct_dct_32x32_sub1_add_neon:   752.0   459.2   862.2   553.9
After:
vp9_inv_dct_dct_16x16_sub1_add_neon:   226.5   145.0   225.1   171.8
vp9_inv_dct_dct_32x32_sub1_add_neon:   721.2   415.7   727.6   475.0

This is cherrypicked from libav commit
a76bf8cf1277ef6feb1580b578f5e6ca327e713c.

Signed-off-by: Martin Storsjö <martin at martin.st>

> http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=758302e4bc14e93989e7feb1135ec3f807c3310d
---

 libavcodec/arm/vp9itxfm_neon.S | 54 ++++++++++++++++++++++++++++--------------
 1 file changed, 36 insertions(+), 18 deletions(-)

diff --git a/libavcodec/arm/vp9itxfm_neon.S b/libavcodec/arm/vp9itxfm_neon.S
index 78fdae6..dee2f05 100644
--- a/libavcodec/arm/vp9itxfm_neon.S
+++ b/libavcodec/arm/vp9itxfm_neon.S
@@ -542,16 +542,23 @@ function idct16x16_dc_add_neon
 
         vrshr.s16       q8,  q8,  #6
 
+        mov             r3,  r0
         mov             r12, #16
 1:
         @ Loop to add the constant from q8 into all 16x16 outputs
-        vld1.8          {q3},  [r0,:128]
-        vaddw.u8        q10, q8,  d6
-        vaddw.u8        q11, q8,  d7
-        vqmovun.s16     d6,  q10
-        vqmovun.s16     d7,  q11
-        vst1.8          {q3},  [r0,:128], r1
-        subs            r12, r12, #1
+        subs            r12, r12, #2
+        vld1.8          {q2},  [r0,:128], r1
+        vaddw.u8        q10, q8,  d4
+        vld1.8          {q3},  [r0,:128], r1
+        vaddw.u8        q11, q8,  d5
+        vaddw.u8        q12, q8,  d6
+        vaddw.u8        q13, q8,  d7
+        vqmovun.s16     d4,  q10
+        vqmovun.s16     d5,  q11
+        vqmovun.s16     d6,  q12
+        vst1.8          {q2},  [r3,:128], r1
+        vqmovun.s16     d7,  q13
+        vst1.8          {q3},  [r3,:128], r1
         bne             1b
 
         bx              lr
@@ -1147,20 +1154,31 @@ function idct32x32_dc_add_neon
 
         vrshr.s16       q8,  q8,  #6
 
+        mov             r3,  r0
         mov             r12, #32
 1:
         @ Loop to add the constant from q8 into all 32x32 outputs
-        vld1.8          {q2-q3},  [r0,:128]
-        vaddw.u8        q10, q8,  d4
-        vaddw.u8        q11, q8,  d5
-        vaddw.u8        q12, q8,  d6
-        vaddw.u8        q13, q8,  d7
-        vqmovun.s16     d4,  q10
-        vqmovun.s16     d5,  q11
-        vqmovun.s16     d6,  q12
-        vqmovun.s16     d7,  q13
-        vst1.8          {q2-q3},  [r0,:128], r1
-        subs            r12, r12, #1
+        subs            r12, r12, #2
+        vld1.8          {q0-q1},  [r0,:128], r1
+        vaddw.u8        q9,  q8,  d0
+        vaddw.u8        q10, q8,  d1
+        vld1.8          {q2-q3},  [r0,:128], r1
+        vaddw.u8        q11, q8,  d2
+        vaddw.u8        q12, q8,  d3
+        vaddw.u8        q13, q8,  d4
+        vaddw.u8        q14, q8,  d5
+        vaddw.u8        q15, q8,  d6
+        vqmovun.s16     d0,  q9
+        vaddw.u8        q9,  q8,  d7
+        vqmovun.s16     d1,  q10
+        vqmovun.s16     d2,  q11
+        vqmovun.s16     d3,  q12
+        vqmovun.s16     d4,  q13
+        vqmovun.s16     d5,  q14
+        vst1.8          {q0-q1},  [r3,:128], r1
+        vqmovun.s16     d6,  q15
+        vqmovun.s16     d7,  q9
+        vst1.8          {q2-q3},  [r3,:128], r1
         bne             1b
 
         bx              lr