With skip block or coeff < zbin it is about twice as fast as C.
If most coeff values are > zbin it is about 10-15x as fast as C.
BUG=webm:1426
Change-Id: I5d3c007b014a372d5ef0882b39bb48983b4131c7
allow the right shift to operate on 64-bits, this matches the rest of
the implementations
previously:
b0f1ae147 vpx_get16x16var_avx2: correct cast order
Change-Id: I632ee5e418f3f9b30e79ecd05588eb172b0783aa
allow the right shift to operate on 64-bits, this matches the rest of
the implementations
missed in:
6acd061aa variance_avx2: sync variance functions with c-code
Change-Id: Icae436b881251ccb9f9ed64fcbf8d358c58a4617
For 8-bit the subtrahend is small enough to fit into uint32_t.
For 10/12-bit apply:
63a37d16f Prevent negative variance
previously:
47b9a0912 Resolve -Wshorten-64-to-32 in highbd variance.
c0241664a Resolve -Wshorten-64-to-32 in variance.
Change-Id: I181c85f0b9a03da37c2e8b89482d48aa3dbc0aee
Avoid unsigned overflow warning:
unsigned integer overflow: 19974 - 32703 cannot be represented in type
'unsigned int'
Change-Id: Ifebee014342e4c6f3b53306c0cad6ae0b465ac12
Backend specific optimization for PPC VSX reads 16 bytes, whereas arm neon /
sse2 only reads <= 8 bytes. Although the extra bytes read are actually never
used, this is not a warrant for groping around. Fixed by allocating more when
building for VSX. This is reported by asan.
Also note - PPC does have assembly that loads 64-bit content from memory - lxsdx
loads one 64-bit doubleword (whereas lxvd2x loads two 64-bit doubleword) from
memory. However, we only have "vec_vsx_ld" builtins that mapped to lxvd2x, no
builtins to lxsdx. The only way to access lxsdx is through inline assembly,
which does not fit well in the origin paradigm.
Refer:
vsx:
vpx_tm_predictor_4x4_vsx @ third_party/libvpx/git_root/vpx_dsp/ppc/intrapred_vsx.c
neon:
vpx_tm_predictor_4x4_neon @ third_party/libvpx/git_root/vpx_dsp/arm/intrapred_neon_asm.asm
sse2:
tm_predictor_4x4 @ third_party/libvpx/git_root/vpx_dsp/x86/intrapred_sse2.asm
BUG=b/63112600
Tested:
asan tests passed.
Change-Id: I5f74b56e35c05b67851de8b5530aece213f2ce9d
Keep optimized code out of the reference implementation. This matches
the style of the other sub calls.
Change-Id: I3da6acd4f2c647b029c420e22ac9410a18259689
0.007% regression on rtc and 0.004% gain on rtc_derf.
1 thread on QVGA,VGA and HD has ~0.2% speed regression while 2 threads has
~0.2% speed gain on Google Pixel.
Change-Id: Ia4a6ec904df670d7001e35e070b01e34149d23dc
Officially the quant structures are 8 elements, with one dc element and
7 repeated ac elements. The low bit depth optimizations take advantage
of this to fill the xmm registers. The high bit depth version manually
duplicates the values.
If all the optimizations were unified, the structure sizes could be
greatly reduced.
Change-Id: Ibd7a0337a7832ce2a1a05ee433c310077e1059ae
Use only valid values for quantize inputs. These were determined by
looping over vp9_init_quantizer and looking for max and min values.
This allows extending the test to the low bit depth functions which were
not designed to handle all possible inputs but only valid inputs.
Change-Id: I94e1d8863a49ac227845b65c6b50130e10e6319e
To fix valgrind issueis with SVC tests.
SVC encoding uses prune_evenmore which is causing uinit value.
Will re-enable later when issue is resolved.
Change-Id: I257ff878cf78197ddd813db056582a4d5fe94f44
When content_state_sb is set to LowVarHighSumdiff, don't reset
it to VeryHighSad. Visually better on clips with strong lighting changes.
Small/negligible change in RTC metrics and speed.
Change-Id: I20c383e3c4cf8d1149de5f9260449c0b7cf7c6aa
When int_pro_motion_estimation is done for superblock in
choose_partitioning, use it to avoid the full_pixel_search
for NEWMV mode, if bsize is >= 32X32.
For speed > 7.
Small/neutral change on RTC metrics.
~1-2% speedup on arm on high motion clip.
Change-Id: I3cfe6833ff4bf75d4afa83eaf058ad45729de85b
Although the low bitdepth functions are identical (excepting the need
for larger intermediate values) they do not pass these tests. This
improves the error output to aid debugging.
Simplify buffer usage with Buffer and removing unnecessarily aligned
variables.
eob is a single element and never written using aligned instructions.
BUG=webm:1426
Change-Id: Ic95789a135cf1e8a3846d85270f2b818f6ec7e35
Reduces memory usage, and speeds up encoding for some difficult clips.
No impact on output or metrics.
Ported from aomedia patch:
https://aomedia-review.googlesource.com/c/14501
Change-Id: I26ec69af8336f9e80da486a1cfbfc89a3596954d
