AR30 is optimized with 3 techniques
1. vpmulhuw is used to replicate 8 bits to 10 bits.
2. Two channels are processed at a time. R and B, and A and G.
3. vpshufb is used to shift and mask 2 channels of R and B
Red Blue
With the 8 bit value in the upper bits, vpmulhuw by (1024+4) will produce a 10
bit value in the low 10 bits of each 16 bit value. This is whats wanted for the
blue channel. The red needs to be shifted 4 left, so multiply by (1024+4)*16 for
red.
Alpha Green
Alpha and Green are already in the high bits so vpand can zero out the other
bits, keeping just 2 upper bits of alpha and 8 bit green. The same multiplier
could be used for Green - (1024+4) putting the 10 bit green in the lsb. Alpha
would be a simple multiplier to shift it into position. It wants a gap of 10
above the green. Green is 10 bits, so there are 6 bits in the low short. 4
more are needed, so a multiplier of 4 gets the 2 bits into the upper 16 bits,
and then a shift of 4 is a multiply of 16, so (4*16) = 64. Then shift the
result left 10 to position the A and G channels.
Bug: libyuv:751
Test: ARGBToAR30_Opt
Change-Id: Ie4f20dce18203bae7b75acb1fd5232db8a8a4f11
Reviewed-on: https://chromium-review.googlesource.com/820046
Commit-Queue: Frank Barchard <fbarchard@chromium.org>
Reviewed-by: Cheng Wang <wangcheng@google.com>
Port ARGBToAR30Row_AVX2 to ARGBToAR30Row_SSE2 using same instructions
but xmm registers and doing half as many pixels per loop.
Bug: libyuv:751
Test: LibYUVConvertTest.ARGBToAR30_Opt
Change-Id: Id644e54639133d1caf28ea3cd11ff6ab6891a673
Reviewed-on: https://chromium-review.googlesource.com/817918
Commit-Queue: Frank Barchard <fbarchard@chromium.org>
Reviewed-by: richard winterton <rrwinterton@gmail.com>
Reduce allocations of row buffers to 1 alloc/free.
Do 2 rows at a time to avoid converting U and V planes twice.
Bug: libyuv:715
Test: LibYUVConvertTest.H010ToAR30_Opt
Change-Id: I2f3a03b4875df5e3b969112a78a1a0b28399fa2f
Reviewed-on: https://chromium-review.googlesource.com/816021
Reviewed-by: Cheng Wang <wangcheng@google.com>
H010ToAR30 uses Convert16To8Row_SSSE3 to convert 10 bit YUV to 8 bit.
Then standard YUV conversion can be used. This improves performance
on low end CPUs.
Future CL will by pass this conversion allowing for 10 bit YUV source,
but the function will be useful as a utility for YUV conversions.
Bug: libyuv:559, libyuv:751
Test: out/Release/libyuv_unittest --gtest_filter=*H010ToAR30* --libyuv_width=1280 --libyuv_height=720 --libyuv_repeat=999 --libyuv_flags=-1 --libyuv_cpu_info=-1
Change-Id: I9b3ef22d88a5fd861de4cf1900b4c6e8fd24d0af
Reviewed-on: https://chromium-review.googlesource.com/792334
Commit-Queue: Frank Barchard <fbarchard@chromium.org>
Reviewed-by: Frank Barchard <fbarchard@chromium.org>
This allows the linker to move the variables from the .data section to
the .rodata section.
Bug: libyuv:254
Test: out/Release/libyuv_unittest --gtest_filter=* --libyuv_width=1280 --libyuv_height=720 --libyuv_repeat=999 --libyuv_flags=-1 --libyuv_cpu_info=-1
Change-Id: I6998570f1af4337d7b80313d9e18e36aa20d6ec0
Reviewed-on: https://chromium-review.googlesource.com/777033
Commit-Queue: Frank Barchard <fbarchard@chromium.org>
Reviewed-by: Frank Barchard <fbarchard@chromium.org>
This version of the H010ToAR30 provides a 3 step conversion
Convert16To8Row_AVX2
H420ToARGB_AVX2
ARGBToAR30_AVX2
Low level function added to convert 16 bit to 8 bit using multiply
to adjust 10 bit or other bit depths and then save the upper 16 bits.
Bug: libyuv:751
Test: LibYUVPlanarTest.Convert16To8Row_Opt unittest added
Change-Id: I9cc576fda8afa1003cb961d03e0e656e0b478f03
Reviewed-on: https://chromium-review.googlesource.com/783554
Commit-Queue: Frank Barchard <fbarchard@chromium.org>
Reviewed-by: richard winterton <rrwinterton@gmail.com>
The unittest compares the results of 8 and 16 bit scaling
and expects them to be the same. This CL makes the 16 bit
scaling filter logic match.
Bug: libyuv:749
Test: LibYUVScaleTest.DISABLED_ScaleDownBy4_Linear_16
Change-Id: Ifb3ca4d770ef38f9f16abe9b9aeb843b779bf371
Reviewed-on: https://chromium-review.googlesource.com/772370
Reviewed-by: Weiyong Yao <braveyao@chromium.org>
Commit-Queue: Frank Barchard <fbarchard@google.com>
When converting from lsb 10 bit formats to msb, the values
need to be shifted to the top 10 bits. Using a multiply
allows the different numbers of bits to be copied:
// 128 = 9 bits
// 64 = 10 bits
// 16 = 12 bits
// 1 = 16 bits
Bug: libyuv:751
Test: LibYUVPlanarTest.MultiplyRow_16_Opt
Change-Id: I9cf226053a164baa14155215cb175065b1c4f169
Reviewed-on: https://chromium-review.googlesource.com/762951
Reviewed-by: richard winterton <rrwinterton@gmail.com>
Reviewed-by: Frank Barchard <fbarchard@google.com>
Commit-Queue: Frank Barchard <fbarchard@google.com>
clang does not require -msse2 or -msse for inline, except
the "x" parameter. So change this to "m" for 32 bit. 64 bit
requires sse2 so use "x" for 64 bit.
gcc requires -msse for xmm registers in clobber list.
Reduce compiler requirement from -msse2 to -msse for enabling
assembly.
Bug: libyuv:754, libyuv:757
Test: CC=clang CXX=clang++ CFLAGS="-m32" CXXFLAGS="-m32 -mno-sse -O2" make -f linux.mk
Change-Id: I86df72cfee80b7d349561c1fd7c97ad360767255
Reviewed-on: https://chromium-review.googlesource.com/759303
Reviewed-by: richard winterton <rrwinterton@gmail.com>
Reviewed-by: Frank Barchard <fbarchard@google.com>
Commit-Queue: Frank Barchard <fbarchard@google.com>
cleanup to remove ifdefs around functions affected by
a clang bug.
gn gen out/Release "--args=is_debug=false target_os=\"android\" target_cpu=\"mips64el\" mips_arch_variant=\"r6\" mips_use_msa=true is_component_build=true is_clang=true"
ninja -v -C out/Release libyuv_unittest
Bug: libyuv:634
Test: build for mips with clang
Change-Id: I278b368dbb2fe89082240e280267d0a27a214c78
Reviewed-on: https://chromium-review.googlesource.com/757980
Reviewed-by: Frank Barchard <fbarchard@google.com>
Commit-Queue: Frank Barchard <fbarchard@google.com>
Key instruction sets added for each microarchitecture:
AVX512BW, AVX512VL, AVX512DQ - skylake server or later
AVX512_VBMI, AVX512_IFMA - cannon lake or later
AVX512_BITALG, AVX512_VBMI2, AVX512_VPOPCNTDQ, AVX512_VNNI, GFNI, VAES, VPCLMULQDQ - ice lake or later
Bug: libyuv:752
Test: ~/intelsde/sde -icl -- out/Release/libyuv_unittest --gtest_filter=*Cpu*
Change-Id: I9ee28904c90009d66721b9f805a440c5fc2da122
Reviewed-on: https://chromium-review.googlesource.com/755617
Reviewed-by: Frank Barchard <fbarchard@google.com>
Reviewed-by: richard winterton <rrwinterton@gmail.com>
vmovd is an AVX instruction. This will crash on an older CPU with
only SSSE3 but not AVX. Use movd instead.
Bug: libyuv:753
Test: ~/intelsde/sde -mrm -- out/Release/libyuv_unittest --gtest_filter=LibYUVCompareTest.BenchmarkHammingDistance_Opt
Change-Id: I1fb0026039d5f83d124f5d03fed7dc0d2d723e49
Reviewed-on: https://chromium-review.googlesource.com/756200
Reviewed-by: Cheng Wang <wangcheng@google.com>
Reviewed-by: richard winterton <rrwinterton@gmail.com>
H010 is 10 bit planar format with 10 bits in lower bits.
P010 is 10 bit biplanar format with 10 bits in upper bits.
This function weaves the U and V channels and shifts the bits
into the upper bits.
Bug: libyuv:751
Test: LibYUVPlanarTest.MergeUV10Row_Opt
Change-Id: I4a0bac0ef1ff95aa1b8d68261ec8e8e86f2d1fbf
Reviewed-on: https://chromium-review.googlesource.com/752692
Reviewed-by: Cheng Wang <wangcheng@google.com>
Reviewed-by: Frank Barchard <fbarchard@google.com>
Commit-Queue: Frank Barchard <fbarchard@google.com>
Bug: libyuv:701
Test: objdump to confirm code gen
Change-Id: Ibdcb2cc6bc9bf14b4ccb874c49fc9ff664650e1a
Reviewed-on: https://chromium-review.googlesource.com/745390
Reviewed-by: Frank Barchard <fbarchard@google.com>
Reviewed-by: richard winterton <rrwinterton@gmail.com>
popcnt has a fake dependency on the destination.
This assembly avoids the dependency by using a different
register for each popcnt.
Bug: libyuv:701
Test: LIBYUV_DISABLE_SSSE3=1 out/Release/libyuv_unittest --gtest_filter=*Ham*Opt --libyuv_width=1280 --libyuv_height=720 --libyuv_repeat=9999 --libyuv_flags=-1 --libyuv_cpu_info=-1
Change-Id: Ie1d202e2613b7fa8a3c02acd433940e92c80eafa
Reviewed-on: https://chromium-review.googlesource.com/731826
Reviewed-by: Cheng Wang <wangcheng@google.com>
Reviewed-by: Frank Barchard <fbarchard@google.com>
Commit-Queue: Frank Barchard <fbarchard@google.com>
mingw gcc sets the macro _M_IX86 which is normally only set
by Visual C and clangcl which are Visual C style source code
style for assembly, but gcc is not Visual C compatible.
Add _MSC_VER to most ifdefs to detect that its really Visual C
or clangcl and not mingw gcc so the gcc source code will be used.
Bug: libyuv:744
Test: CXXFLAGS=-m32 CXX=~/prebuilts/gcc/linux-x86/host/x86_64-w64-mingw32-4.8/bin/x86_64-w64-mingw32-g++ make -f linux.mk
Change-Id: I3431aa486eb769b145faa8d5eb75ed639f9d6f5e
Reviewed-on: https://chromium-review.googlesource.com/722319
Reviewed-by: Cheng Wang <wangcheng@google.com>
Commit-Queue: Frank Barchard <fbarchard@google.com>
Bug: libyuv:701
Test: HammingDistance unittest with large size
Change-Id: Id41a2c27eb8922d03b3a21dab32fa2e7b015ba38
Reviewed-on: https://chromium-review.googlesource.com/708335
Reviewed-by: Cheng Wang <wangcheng@google.com>
Commit-Queue: Frank Barchard <fbarchard@google.com>
This reverts commit ec75df5894845b8d6b1341885a78db1de83decd8.
Reason for revert: <INSERT REASONING HERE>
Original change's description:
> ComputeHammingDistance reduce SIMD loop to 1 call when possible.
>
> 32 bit x86 has high overhead due to -fpic. So this reduces the
> number of calls by 1.
>
> TBR=kjellander@chromium.org
> Bug: libyuv:701
> Test: BenchmarkHammingDistance
> Change-Id: I7f557ef047920db65eab362a5f93abbd274ca051
> Reviewed-on: https://chromium-review.googlesource.com/701755
> Reviewed-by: Frank Barchard <fbarchard@google.com>
> Reviewed-by: Cheng Wang <wangcheng@google.com>
TBR=rrwinterton@gmail.com,fbarchard@google.com,wangcheng@google.com
Change-Id: Ia61e8558a8f083c14be5f51e0e141550b6f2b5c1
No-Presubmit: true
No-Tree-Checks: true
No-Try: true
Bug: libyuv:701
Reviewed-on: https://chromium-review.googlesource.com/707823
Reviewed-by: Frank Barchard <fbarchard@google.com>
Commit-Queue: Frank Barchard <fbarchard@google.com>
32 bit x86 has high overhead due to -fpic. So this reduces the
number of calls by 1.
TBR=kjellander@chromium.org
Bug: libyuv:701
Test: BenchmarkHammingDistance
Change-Id: I7f557ef047920db65eab362a5f93abbd274ca051
Reviewed-on: https://chromium-review.googlesource.com/701755
Reviewed-by: Frank Barchard <fbarchard@google.com>
Reviewed-by: Cheng Wang <wangcheng@google.com>
If length of HammingDistance was not a multiple of 4,
the result was incorrect. The old tests did not catch this
so a new test is done to count 1s.
Bug: libyuv:740
Test: LibYUVCompareTest.TestHammingDistance
Change-Id: I93db5437821c597f1f162ac263d4a594bb83231f
Reviewed-on: https://chromium-review.googlesource.com/699614
Reviewed-by: richard winterton <rrwinterton@gmail.com>
Reviewed-by: Cheng Wang <wangcheng@google.com>
Reviewed-by: Frank Barchard <fbarchard@google.com>
Commit-Queue: Frank Barchard <fbarchard@google.com>
Under cache thrashing circumstances, ldp/stp perform better than
ld1/st1 on QC820/QC821 CPUs. Same performance when hitting cache.
Bug: libyuv:738
Test: LibYUVPlanarTest.TestCopySamples_Opt (445 ms)
Change-Id: Ib6a0a5d5e6a1b7ef667b9bb2edb39d681cf3614c
Reviewed-on: https://chromium-review.googlesource.com/691281
Commit-Queue: Frank Barchard <fbarchard@google.com>
Reviewed-by: Cheng Wang <wangcheng@google.com>
Full color test is the slowest of the unittests, and not catching any
additional bugs at the moment. Step thru range of 0 to 255 in steps of
5 to speed up the test. 255 is 3 * 5 * 17, so any of those primes would
hit 0 and 255 exactly.
Was LibYUVColorTest.TestFullYUV (896 ms)
Now LibYUVColorTest.TestFullYUV (212 ms)
TBR=kjellander@chromium.org
Bug: libyuv:736
Test: LibYUVColorTest.TestFullYUV
Change-Id: I5b55fb07ada0dc7bdc3c3c20569d36bf09bb3804
Reviewed-on: https://chromium-review.googlesource.com/672064
Commit-Queue: Frank Barchard <fbarchard@google.com>
Reviewed-by: Frank Barchard <fbarchard@google.com>
Use ld2 to load even and odd pixels into different registers
and hadd to half add them to each other.
Previously used paired and shift.
TBR=kjellander@chromium.org
BUG=libyuv:723
TEST=ScaleDownBy2_Linear
Change-Id: I3ec72bcf7d4c746837217496c301eb4e4ad963cf
Reviewed-on: https://chromium-review.googlesource.com/644113
Reviewed-by: Cheng Wang <wangcheng@google.com>
urhadd is a rounded average. Linear filter wants to average
horizontally, so use ld2 to separate even and odd pixels.
TBR=jkellander@chromium.org
BUG=None
TEST=LibYUVScaleTest.*ScaleDownBy2*
Change-Id: Id667288a030e72ce8e1c1d6719b69c555c0db063
Reviewed-on: https://chromium-review.googlesource.com/642448
Commit-Queue: Frank Barchard <fbarchard@google.com>
Reviewed-by: Cheng Wang <wangcheng@google.com>
Roughly. instead of 4 loads and 8 multiples, use 1 load and 2 multiples
4 times over. The original code, as with the C code from clang and gcc,
did all the loads, then all the math, then the store. The new code
does a load, then the math, then the next load, etc.
This schedules better on current arm 64 cpus.
Number of registers also reduced, reusing the same registers.
HiSilicon ARM A73:
Now
TestGaussRow_Opt (890 ms)
TestGaussCol_Opt (571 ms)
Was
TestGaussRow_Opt (1061 ms)
TestGaussCol_Opt (595 ms)
Qualcomm 821 (Pixel):
Now
TestGaussRow_Opt (571 ms)
TestGaussCol_Opt (474 ms)
Was
TestGaussRow_Opt (751 ms)
TestGaussCol_Opt (520 ms)
TBR=kjellander@chromium.org
BUG=libyuv:719
TEST=LibYUVPlanarTest.TestGaussRow_Opt
Reviewed-on: https://chromium-review.googlesource.com/627478
Reviewed-by: Cheng Wang <wangcheng@google.com>
Reviewed-by: Frank Barchard <fbarchard@google.com>
Change-Id: I5ec81191d460801f0d4a89f0384f89925ff036de
Reviewed-on: https://chromium-review.googlesource.com/634448
Commit-Queue: Frank Barchard <fbarchard@google.com>
