tm_predictor_4x4 is implemented with SSE2 using XMM registers.
Speed up by ~25% in ./test_intra_pred_speed.
Change-Id: I25074b78d476a2cb17f81cf654bdfd80df2070e0
Added optimization of the 8 bit assembly quantizer routines. This makes
these functions up to 100% faster, depending on encoding parameters.
This patch maskes the encoder faster in both the high bitdepth and 8bit
configurations. In the high bitdepth configuration, it effects profile 0
only.
Based on my profiling using 1080p input the net gain is between 1-3% for
the 8 bit config, and around 2.5-4.5% for the high bitdepth config,
depending on target bitrate. The difference between the 8 bit and high
bitdepth configurations for the same encoder run is reduced by 1% in all
cases I have profiled.
Change-Id: I86714a6b7364da20cd468cd784247009663a5140
This experiment allows using full above/right edges for all transform
sizes whenever available (for d45/d63), and adds bottom/left edges for
d207.
See issue 1043.
Change-Id: I5cf7f345e783e8539bb6b6d2c9972fb1d6d0a78b
When configured with high bit detpth enabled, the 8bit quantize
function stopped using optimised code. This made 8bit content
decode slowly. This commit re-enables the SSSE3 optimisations.
Change-Id: I194b505dd3f4c494e5c5e53e020f5d94534b16b5
When configured with high bit detpth enabled, the 8bit quantize
function stopped using optimised code. This made 8bit content
decode slowly. This commit re-enables the SSE2 optimisation
(but not the SSSE3 optimisation).
Change-Id: Id015fe3c1c44580a4bff3f4bd985170f2806a9d9
This is more a proof of concept than anything else. The problem here
isn't so much how to code it, but rather where to place the resulting
code. All intrapred DSP code lives in vpx_dsp, so do we want the vp10
specific intra pred functions to live there, or in vp10/?
See issue 1015.
Change-Id: I675f7badcc8e18fd99a9553910ecf3ddf81f0a05
I've added a few new functions (d45e, d63e, he, ve) to cover the
filtered h/v 4x4 predictors that are vp8-specific, the "correct"
d45 with the correctly filtered bottom-right pixel (as opposed to
the unfiltered version in vp9), and the "broken" d63 with weirdly
filtered bottom-right pixels (which is correctly filtered in vp9).
There may be a minor performance impact on all systems because we
have to do an extra copy of the Above pixel array to incorporate
the topleft pixel in the same array (thus fitting the vpx_dsp API).
In addition, armv6 will have a more serious performance impact b/c
I removed the armv6/vp8-specific assembly. I'm not sure anyone
cares...
Change-Id: I7f9e5ebee11d8e21aca2cd517a69eefc181b2e86
When configured with high bitdepth enabled, the 8bit transform
stopped using optimised code. This made 8bit content decode slowly.
Change-Id: I67d91f9b212921d5320f949fc0a0d3f32f90c0ea
fixes build on windows x64; previously 'heightq' i.e., the 64-bit register
was accessed when only the 32-bit value was needed. given this is from a
stack variable the upper bits were undefined.
+ bump register/xmm counts; users of SETUP_LOCAL_VARS touch xmm13 in
64-bit builds and filter_block1d16_v* uses one extra temp variable
Change-Id: I9c768c0b2047481d1d3b11c2e16b2f8de6eb0d80
For reading, this makes the operation branchless, although it still
requires two shifts. For writing, this makes the operation as fast
as writing an unsigned value, branchlessly. This is also how other
codecs typically code signed, non-arithcoded bitstream elements.
See issue 1039.
Change-Id: I6a8182cc88a16842fb431688c38f6b52d7f24ead
This is based on the original patch optimized for 32bit
platforms by Tamar/Ilya and now uses the x86inc style asm.
The assembly was also modified to support 64bit platforms.
Change-Id: Ice12f249bbbc162a7427e3d23fbf0cbe4135aff2
* changes:
Only build ssse3 filter functions on 64 bit
Clean up unused function warnings in vp8 encoder
Clean up unused function warnings in vp8 onyx_if.c
These were lost in the great sub pixel variance move of
6a82f0d7fb
Not having these functions caused a ~10% performance regression in
some realtime vp8 encodes.
Change-Id: I50658483d9198391806b27899f2c0d309233c4b5
* changes:
Only use .text sections for aout
Use newer x86inc.asm
Use .text instead of .rodata on macho
Copy PIC handling code from x86_abi_support
Set 'private_extern' visibility for macho targets
Avoid 'amdnop' when building with nasm
Catch all elf formats
Expand PIC default to macho64 and respect CONFIG_PIC from libvpx
Use libvpx defines to set name mangling rules
Customize x86inc.asm for libvpx
Rename updated version of x86inc.asm
Use "private_prefix" instead of "program_name" and make vpx the default
prefix.
Change-Id: I4883a99b2aee8e5dc9f2c16a2e6f4b5d6e4de458
and FUN_CONV_2D macros. The predict lut now handles
this case. The encoder now calls vpx_scaled_2d() instead
of vpx_convolve8() for scaling.
Change-Id: Ia1c8af8a31e4cb4887a587143108cb45835f7df7
This reverts commit a5e97d874b.
Additionally:
Revert "vpx_convolve_copy_sse2: fix win64"
This reverts commit 22a8474fe7.
This change performs poorly on various x86_64 devices affecting
performance by 1-3% at 1080P. Performance on chromebook like devices was
mixed neutral to slightly negative, so there should be minimal change
there.
Change-Id: I95831233b4b84ee96369baa192a2d4cc7639658c
This commit clears all the vp9_ prefix use case in vpx_dsp. It gets
the vp9 folder ready to branch out vp10.
Change-Id: I2906eec179ee792b4af8c9b4161313653050e931
This commit clears the function naming convention in vpx_dsp. It
replaces vp9_ prefix of global functions with vpx_ prefix. It also
removes the vp9_ prefix from static functions.
Change-Id: I6394359a63b71a51dda01342eec6a3cc08dfeedf
Add a guard to exclud dspr2 inverse transform files from vpx_dsp
make file, when high bit-depth is turned on. This fixes the jenkins
nightly build.
Change-Id: Ibacd86563af1ec4810c550905b3fa0397baeeafc
This commit moves the module inverse transform functions from vp9
to vpx_dsp folder. The hybrid transform wrapper functions stay in
the vp9 folder, since it involves codec-specific data structures.
Change-Id: Ib066367c953d3d024c73ba65157bbd70a95c9ef8
This function suffers from a couple problems in small core(tablets):
-The load of the next iteration is blocked by the store of previous iteration
-4k aliasing (between future store and older loads)
-current small core machine are in-order machine and because of it the store will spin the rehabQ until the load is finished
fixed by:
- prefetching 2 lines ahead
- unroll copy of 2 rows of block
- pre-load all xmm regiters before the loop, final stores after the loop
The function is optimized by:
copy_convolve_sse2 64x64 - 16%
copy_convolve_sse2 32x32 - 52%
copy_convolve_sse2 16x16 - 6%
copy_convolve_sse2 8x8 - 2.5%
copy_convolve_sse2 4x4 - 2.7%
credit goes to Tom Craver(tom.r.craver@intel.com) and Ilya Albrekht(ilya.albrekht@intel.com)
Change-Id: I63d3428799c50b2bf7b5677c8268bacb9fc29671
