Commit Graph

51 Commits

Author SHA1 Message Date
Christian Duvivier
5b1dc1515f Fix a bunch of TODO from vp9_short_idct32x32_add_neon.
- full ASM version, no more C gateway file.
- integrate combine-add with last step of 2nd pass.
- remove a few push/pop pairs.
- some instruction reordering to hide latency.

Change-Id: Ic9d9933c908b65d1bf7ba8fd47b524cda808c9c6
2013-09-25 21:15:19 -07:00
Johann
a6a00fc6a3 Use lowercase instruction in assembly
The iOS compiler does not recognize BLE:
bad instruction `BLE idct32_transpose_pair_loop'

Change-Id: I7426694c66bc31caf939a2d5000968da1222c15b
2013-09-20 16:11:05 -07:00
hkuang
23e1a29fc7 Speed up iht8x8 by rearranging instructions.
Speed improves from 282% to 302% faster based on assembly-perf.

Change-Id: I08c5c1a542d43361611198f750b725e4303d19e2
2013-09-16 14:23:26 -07:00
hkuang
86fb12b600 Merge "Add neon optimize iht8x8 which is 282% faster than C." 2013-09-12 15:42:44 -07:00
hkuang
182366c736 Add neon optimize iht8x8 which is 282% faster than C.
Change-Id: I963dd4a6e8671957403ccbb9a16ea7de703e3530
2013-09-12 11:49:05 -07:00
Christian Duvivier
6a501462f8 First draft of vp9_short_idct32x32_add_neon.
Lots of TODO which will be taken care in upcoming changes. As is,
about 6x faster than C version.

Change-Id: Ie2557b72fd2d8edca376dbf400a4d173aa5e63e0
2013-09-11 15:19:38 -07:00
hkuang
fc5ec206a7 Speed up idct16x16 by rearrange instructions.
Speed improve from 376% to 400% faster base on assembly-perf.

Change-Id: If0b2eccc39d5793dc101ce9feb7fcadf88396ea2
2013-09-09 18:00:13 -07:00
hkuang
01c4e04424 Speed up idct8x8 by rearrange instructions.
Speed improve from 264% ~ 270% to 280% ~ 300% base on assembly-perf.

Change-Id: I3e2cc818ec14b432204ff43732f39b6438db685d
2013-09-04 15:57:22 -07:00
hkuang
3b8614a8f6 Add neon optimize vp9_short_iht4x4_add.
Change-Id: I42c497b68ae1ee645b59c9968ad805db0a43e37e
2013-09-04 12:37:58 -07:00
hkuang
3a679e56b2 Add neon optimize vp9_short_idct16x16_1_add.
Change-Id: Ib9354c1d975d03e8081df20d50b6a77dfe2dc7e5
2013-08-27 14:00:27 -07:00
hkuang
36e9b82080 Add neon optimize vp9_short_idct8x8_1_add.
Change-Id: I0b15d5e3b0eb97abb9ab5ec08e88b61f8723aaf4
2013-08-26 16:28:57 -07:00
hkuang
69384f4fad Add neon optimize vp9_short_idct4x4_1_add.
Change-Id: I6ecb5c4a1a472feb8e84e9f3352b536d5e28a4a5
2013-08-26 15:55:16 -07:00
hkuang
b85367a608 Merge "Optimise idct4x4: rearrange the instructions a bit to improve instruction scheduling." 2013-08-23 10:08:43 -07:00
hkuang
4082bf9d7c Add neon optimize vp9_short_idct10_16x16_add.
vp9_short_idct10_16x16_add is used to handle the block that only have valid data
at top left 4x4 block. All the other datas are 0. So we could cut many
unnecessary calculations in order to save instructions.

Change-Id: I6e30a3fee1ece5af7f258532416d0bfddd1143f0
2013-08-22 15:53:22 -07:00
hkuang
610642c130 Optimise idct4x4: rearrange the instructions a bit
to improve instruction scheduling.

Change-Id: I5ea881a6e419f9e8ed4b3b619406403b4de24134
2013-08-22 11:02:22 -07:00
hkuang
37cda6dc4c Add neon optimize vp9_short_idct10_8x8_add.
vp9_short_idct10_8x8_add is used to handle the block that only have valid data
at top left 4x4 block. All the other datas are 0. So we could cut several
unnecessary calculations in order to save instructions.

Change-Id: I34fda95e29082b789aded97c2df193991c2d9195
2013-08-20 11:51:07 -07:00
Johann
d514b778c4 Merge "Reduce the instructions of idct8x8. Also add the saving and restoring of D registers." 2013-08-16 11:30:21 -07:00
Johann
65aa89af1a Merge "Reduce instructions of idct4x4." 2013-08-16 11:28:35 -07:00
Frank Galligan
bdc785e976 Merge "vp9: neon: optimise vp9_wide_mbfilter_neon" 2013-08-16 11:16:48 -07:00
hkuang
df0715204c Reduce instructions of idct4x4.
Change-Id: Ia26a2526804e7e2f656b0051618a615fca8fc79d
2013-08-16 10:54:56 -07:00
hkuang
60ecd60c9a Reduce the instructions of idct8x8. Also add the
saving and restoring of D registers.

Change-Id: Id3630c90fcb160ef939fef55411342608af5f990
2013-08-16 10:32:12 -07:00
Mans Rullgard
4fa93bcef4 vp9: neon: use aligned stores in convolve functions
The destination is block-aligned so it is safe to use aligned
stores.

Change-Id: I38261e4fa40bc60e6472edffece59e372908da7e
2013-08-16 14:25:08 +01:00
Johann
a9aa7d07d0 Merge "vp9: neon: add vp9_convolve_avg_neon" 2013-08-15 14:55:15 -07:00
Johann
63e140eaa7 Merge "vp9: neon: add vp9_convolve_copy_neon" 2013-08-15 14:55:08 -07:00
Mans Rullgard
67e53716e0 vp9: neon: optimise vp9_wide_mbfilter_neon
Break up long dependency chains to improve instruction scheduling.

Change-Id: I0e0cb66943df24af920767bb4167b25c38af9630
2013-08-15 19:07:22 +01:00
hkuang
39f42c8713 Merge "Add neon optimize vp9_short_idct16x16_add." 2013-08-14 14:16:20 -07:00
hkuang
cf6beea661 Add neon optimize vp9_short_idct16x16_add.
Change-Id: I27134b9a5cace2bdad53534562c91d829b48838d
2013-08-14 13:52:16 -07:00
Mans Rullgard
0f1deccf86 vp9: neon: add vp9_convolve_avg_neon
Change-Id: I33cff9ac4f2234558f6f87729f9b2e88a33fbf58
2013-08-14 16:27:55 +01:00
Mans Rullgard
635ba269be vp9: neon: add vp9_convolve_copy_neon
Change-Id: I15adbbda15d1842e9f15f21878a5ffbb75c3c0c9
2013-08-14 16:27:55 +01:00
Johann
4417c04531 Merge "vp9: neon: optimise convolve8_vert functions" 2013-08-12 17:54:47 -07:00
Mans Rullgard
ad7021dd6c vp9: neon: optimise convolve8_vert functions
Invert loops to operate vertically in the inner loop.  This allows
removing redundant loads.

Also add preloading of data.

Change-Id: I4fa85c0ab1735bcb1dd6ea58937efac949172bdc
2013-08-12 15:37:48 +01:00
Mans Rullgard
b84dc949c8 vp9: neon: optimise convolve8_horiz functions
Each iteration of the horizontal loop reuses 7 of the 11 source
values.  Loading only the 4 new values saves some time.

Also add preload for source data.

Overall 4% faster on Chromebook.

Change-Id: I8f69e749f2b7f79e9734620dcee51dbfcd716b44
2013-08-11 16:21:55 +01:00
Christian Duvivier
78182538d6 Neon version of vp9_short_idct4x4_add.
Change-Id: Idec4cae0cb9b3a29835fd2750d354c1393d47aa4
2013-08-06 18:41:27 -07:00
Mans Rullgard
355cb14dc7 vp9: neon: convolve: replace some insns with simpler equivalents
Change-Id: I5d6906772e6e6adf68d7f0fd5b8b5207a64a3a37
2013-08-02 08:11:28 -07:00
Mans Rullgard
2003468df8 vp9: neon: convolve: simplify branching to C fallbacks
Change-Id: Ic7cacd02d6dc9243ad8fc85082c5618a9d1e66dc
2013-08-02 08:11:25 -07:00
Mans Rullgard
5e2e78d024 vp9: neon: optimise loads in horiz convolve functions
Loading to single lanes in multiple registers is expensive since
it requires a read and write of each register which saturates
the register file access.  Loading to single registers followed
by a separate transpose reduces this pressure.

Change-Id: I4cc35887ddbca80e5e635b50d2b1d158de9668ee
2013-08-02 08:11:08 -07:00
Mans Rullgard
d85ae87183 vp9: neon: add vp9_mb_lpf_* functions
Change-Id: I13e0880df234f15abc4cc7c57fe84488d5d46a75
2013-08-02 08:10:50 -07:00
hkuang
588b4daf54 Fix some format error and code error in neon code.
Change-Id: I748dee8938dfb19f417f24eed005f3d216f83a82
2013-07-26 14:14:57 -07:00
Frank Galligan
e88db77892 Merge "Speedup loopfilter neon code." 2013-07-22 17:39:42 -07:00
Frank Galligan
5af6bf6c43 Speedup loopfilter neon code.
Try and cut down the cycle count by rearranging the instructions
so there are less stalls.

Change-Id: Ic1383335ee0f05e656477d9ee9c179ec231285d5
2013-07-22 17:00:01 -07:00
hkuang
97dbee00dd Merge "Add neon optimize vp9_short_idct8x8_add." 2013-07-19 08:28:39 -07:00
hkuang
d757de744c Add neon optimize vp9_short_idct8x8_add.
Change-Id: Ic32acf3e2939c6d12d9c2bf192a5f5da59705fda
2013-07-18 16:40:41 -07:00
Frank Galligan
7fd5d8e6a4 Fix horz loopfilter loops
If count was greater than 1 the src pointer would be off on
the second loop.

Change-Id: I8e09037e68dc4ae92076a8067f7b6dacbbef8263
2013-07-18 09:44:15 -07:00
Johann
59dc4e9cdd vp9_convolve8_neon placeholder
Call the individually optimized horizontal and vertical functions. This
implementation abuses the temp buffer.

This will be replaced with a custom optimized function.

Over 2x speedup.

Change-Id: I5b908d2a73d264e9810d6022bbff73207a3055dd
2013-07-17 08:39:27 -07:00
Johann
90ebfe621f Merge "vp9_convolve8_[horiz|vert]_avg" 2013-07-16 09:42:52 -07:00
Frank Galligan
f4f60f6005 Neon: Update mbfilter if all vectors follow one branch.
Change the mbfilter Neon code from executing both branches if all
vectors follow only one branch.

The code is about 5% faster when executing only one branch and about
1% slower when executing both branches.

-PS5: Remove local stack space from mbfilter.

Change-Id: I6a23f9b318a9f4568a2718b4c9348db988fe2182
2013-07-15 13:08:28 -07:00
Johann
a15bebfc0a vp9_convolve8_[horiz|vert]_avg
Super basic conversion from the other implementations. Any changes to
one should be trivial to copy over keep in sync.

Change-Id: I1720b4128e0aba4b2779e3761f6494f8a09d3ea8
2013-07-12 16:21:33 -07:00
Johann
158c80cbb0 convolve8 optimizations for neon
Independent horizontal and vertical implementations.

Requires that blocks be built from 4x4 and [xy]_step_q4 == 16

6-10% improvement. CIF improved the least.

Change-Id: I137f5ceae4440adc0960bf88e4453e55a618bcda
2013-07-11 11:08:19 -07:00
hkuang
c9b25dcae4 Add neon optimize vp9_dc_only_idct_add.
Change-Id: Iae84ab945cc9662a0ddd839aa2b9ca59f2ae5423
2013-07-11 10:30:47 -07:00
Frank Galligan
198fa6d0a0 Add Neon horizontal and vertical vp9_mbloop_filter
- The vp9 mbfilter C code will branch on flat and mask. This CL
  will perform both branches and combine the data. A later CL will
  perform a check to see if all patch will take one branch.
- These functions are about 1.75 times faster than the C code on
  Nexus 7.

PS #3
- Changed all functions to dub limit, blimit, and thresh from
  vld {dx[]}, freeing up r4-r6.
- Changed code to use vbif to reduce one instruction and free
  up a d register.

Change-Id: I028dae0e434dc9891c3677bdb182e201ffb04777
2013-07-09 12:40:05 -07:00