Age | Commit message (Collapse) | Author |
|
|
|
This CL resolves the mismatch between C and intrinsic implementation
of vpx_hadamard_32x32 function. The mismatch was due to integer
overflow during the addition operation in the intrinsic functions.
Specifically, the addition in the intrinsic function was performed
at the 16-bit level, while the calculation of a0 + a1 resulted in
a 17-bit value.
This code change addresses the problem by performing
the addition at the 32-bit level (with sign extension) in both SSE2
and AVX2, and then converting the results back to the 16-bit level
after a right shift.
STATS_CHANGED
Change-Id: I576ca64e3b9ebb31d143fcd2da64322790bc5853
|
|
impace -> impact
taget -> target
prediciton -> prediction
addtion -> addition
the the -> the
Bug: webm:1803
Change-Id: I759c9d930a037ca69662164fcd6be160ed707d77
|
|
* changes:
Optimize Neon implementation of vpx_int_pro_row
Optimize Neon implementation of vpx_int_pro_col
|
|
Double the number of accumulator registers to remove the bottleneck.
Also peel the first loop iteration.
Change-Id: I6a90680369f9c33cdfe14ea547ac1569ec3f50de
|
|
Use widening pairwise addition instructions to halve the number of
additions required.
Change-Id: I0307a3b65e50d2b1ae582938bc5df9c2b21df734
|
|
cl.exe targeting AArch64 with optimizations enabled
produces invalid code for clip_pixel() when the return type is uint8_t.
See:
https://developercommunity.visualstudio.com/t/Misoptimization-for-ARM64-in-VS-2022-17/10363361
Bug: b/277255076
Bug: webm:1788
Change-Id: Ia3647698effd34f1cf196cd33fa4a8cab9fa53d6
|
|
cl.exe targeting AArch64 with optimizations enabled
will fail with an internal compiler error.
See:
https://developercommunity.visualstudio.com/t/Compiler-crash-C1001-when-building-a-for/10346110
Bug: b/277255076
Bug: webm:1788
Change-Id: I55caf34e910dab47a7775f07280677cdfe606f5b
|
|
|
|
This CL refactors the code related to convolve function.
Furthermore, improved the AVX2 intrinsic to compute
convolve vertical for w = 4 case, and convolve horiz for
w = 16 case.
Please note the module level scaling w.r.t C function
(timer based) for existing (AVX2) and new AVX2 intrinsics:
Block Scaling
Size AVX2 AVX2
(existing) (New)
4x4 5.34x 5.91x
4x8 7.10x 7.79x
16x8 23.52x 25.63x
16x16 29.47x 30.22x
16x32 33.42x 33.44x
This is a bit exact change.
Change-Id: If130183bc12faab9ca2bcec0ceeaa8d0af05e413
|
|
* changes:
Add 2D-specific Neon horizontal convolution functions
Refactor standard bitdepth Neon convolution functions
|
|
2D 8-tap convolution filtering is performed in two passes -
horizontal and vertical. The horizontal pass must produce enough
input data for the subsequent vertical pass - 3 rows above and 4 rows
below, in addition to the actual block height.
At present, all Neon horizontal convolution algorithms process 4 rows
at a time, but this means we end up doing at least 1 row too much
work in the 2D first pass case where we need h + 7, not h + 8 rows of
output.
This patch adds additional dot-product (SDOT and USDOT) Neon paths
that process h + 7 rows of data exactly, saving the work of the
unnecessary extra row. It is impractical to take a similar approach
for the Armv8.0 MLA paths since we have to transpose the data block
both before and after calling the convolution helper functions.
vpx_convolve_neon performance impact: we observe a speedup of ~9% for
smaller (and wider) blocks, and a speedup of 0-3% for larger blocks.
This is to be expected since the proportion of redundant work
decreases as the block height increases.
Change-Id: Ie77ad1848707d2d48bb8851345a469aae9d097e1
|
|
1) Use #define constant instead of magic numbers for right shifts.
2) Move saturating narrow into helper functions that return 4-element
result vectors.
3) Use mem_neon.h helpers for load/store sequences in Armv8.0 paths.
4) Tidy up: assert conditions and some longer variable names.
5) Prefer != 0 to > 0 where possible for loop termination conditions.
Change-Id: Idfcac43ca38faf729dca07b8cc8f7f45ad264d24
|
|
The module level scaling w.r.t C function (timer based) for
existing (SSE2) and new AVX2 intrinsics:
If ref_padding = 0
Block Scaling
size SSE2 AVX2
8x4 3.24x 3.24x
8x8 4.22x 4.90x
8x16 5.91x 5.93x
16x8 1.63x 3.52x
16x16 1.53x 4.19x
16x32 1.38x 4.82x
32x16 1.28x 3.08x
32x32 1.45x 3.13x
32x64 1.38x 3.04x
64x32 1.39x 2.12x
64x64 1.46x 2.24x
If ref_padding = 8
Block Scaling
size SSE2 AVX2
8x4 3.20x 3.21x
8x8 4.61x 4.83x
8x16 5.50x 6.45x
16x8 1.56x 3.35x
16x16 1.53x 4.19x
16x32 1.37x 4.83x
32x16 1.28x 3.07x
32x32 1.46x 3.29x
32x64 1.38x 3.22x
64x32 1.38x 2.14x
64x64 1.38x 2.12x
This is a bit-exact change.
Change-Id: I72c5d155f64d0c630bc8c3aef21dc8bbd045d9e6
|
|
|
|
|
|
Bug: webm:1793
Change-Id: Ib2e3bd3c52632cdd4410cb2c54d69750e64e5201
|
|
|
|
Added AVX2 intrinsic optimization for the following functions
1. vpx_idct16x16_256_add
2. vpx_idct32x32_1024_add
3. vpx_idct32x32_135_add
The module level scaling w.r.t C function (timer based) for
existing (SSE2) and new AVX2 intrinsics:
Scaling
Function Name SSE2 AVX2
vpx_idct32x32_1024_add 3.62x 7.49x
vpx_idct32x32_135_add 4.85x 9.41x
vpx_idct16x16_256_add 4.82x 7.70x
This is a bit-exact change.
Change-Id: Id9dda933aa1f5093bb6b35ac3b8a41846afca9d2
|
|
Change-Id: I8badedc2ad07d60896e45de28b707ad9f6c4d499
|
|
* changes:
Fix mismatched param names in vpx_dsp/x86/sad4d_avx2.c
Fix mismatched param names in vpx_dsp/arm/highbd_sad4d_neon.c
Fix mismatched param names in vpx_dsp/arm/sad4d_neon.c
Fix mismatched param names in vpx_dsp/arm/highbd_avg_neon.c
Fix clang warning on const-qualification of parameters
|
|
Change-Id: I226215a2ff8798b72abe0c2caf3d18875595caa5
|
|
Change-Id: Ia4918eb0bac3b28b27e1ef205b9171680b2eb9a4
|
|
Change-Id: If621944684cf9bb9f353db5961ed8b4b4ae38f24
|
|
Change-Id: Ibf00a6e1029284e637b10ef01ac9b31ffadc74ca
|
|
Change-Id: I900a0a48dde5fcb262157b191ac536e18269feb3
|
|
This allows AArch64 to be correctly detected when building with Visual
Studio (cl.exe) and fixes a crash in vp9_diamond_search_sad_neon.c.
There are still test failures, however.
Microsoft's compiler doesn't define __ARM_FEATURE_*. To use those paths
we may need to rely on _M_ARM64_EXTENSION.
Bug: webm:1788
Bug: b/277255076
Change-Id: I4d26f5f84dbd0cbcd1cdf0d7d932ebcf109febe5
|
|
Bug: webm:1793
Change-Id: If1a46fe183cd18e05b5538b1eba098e420b745ec
|
|
Add Neon implementations of high bitdepth downsampling SAD4D
functions for all block sizes.
Also add corresponding unit tests.
Change-Id: Ib0c2f852e269cbd6cbb8f4dfb54349654abb0adb
|
|
Add Neon implementations of standard bitdepth downsampling SAD4D
functions for all block sizes.
Also add corresponding unit tests.
Change-Id: Ieb77661ea2bbe357529862a5fb54956e34e8d758
|
|
Add Neon implementations of high bitdepth downsampling SAD functions
for all block sizes.
Also add corresponding unit tests.
Change-Id: I56ea656e9bb5f8b2aedfdc4637c9ab4e1951b31b
|
|
Add Neon implementations of standard bitdepth downsampling SAD
functions for all block sizes.
Also add corresponding unit tests.
Change-Id: Ibda734c270278d947673ffcc29ef17a2f4970b01
|
|
|
|
|
|
Introduced AVX2 intrinsic to compute FDCT for block size
16x16 case. This is a bit-exact change.
Please check the module level scaling w.r.t C function (timer based)
for existing (SSE2) and new AVX2 intrinsics:
Scaling
SSE2 AVX2
3.88x 5.95x
Change-Id: I02299c3746fcb52d808e2a75d30aa62652c816dc
|
|
Added a speed feature to skip every other row
in SAD computation during motion search.
Instruction Count BD-Rate Loss(%)
cpu Resolution Reduction(%) avg.psnr ovr.psnr ssim
0 LOWRES2 0.958 0.0204 0.0095 0.0275
0 MIDRES2 1.891 -0.0636 0.0032 0.0247
0 HDRES2 2.869 0.0434 0.0345 0.0686
0 Average 1.905 0.0000 0.0157 0.0403
STATS_CHANGED
Change-Id: I1a8692757ed0cbcb2259729b3ecfb0436cdf49ce
|
|
|
|
|
|
Bug: webm:1793
Change-Id: Icba4ad242dcd0cad736b9a203829361c5bd1ca3f
|
|
into main
|
|
Add a widening 4D reduction function operating on uint16x8_t vectors
and use it to optimize the final reduction in Armv8.0 Neon standard
bitdepth 16xh, 32xh and 64h SAD4D computations.
Also simplify the Armv8.0 Neon version of the sad64xhx4d_neon helper
function since VP9 block sizes are not large enough to require
widening to 32-bit accumulators before the final reduction.
Change-Id: I32b0a283d7688d8cdf21791add9476ed24c66a28
|
|
Add a 4D reduction function operating on uint16x8_t vectors and use
it to optimize the final reduction in standard bitdepth 4xh and 8xh
SAD4D computations. Similar 4D reduction optimizations have already
been implemented for all other standard bitdepth block sizes, and all
high bitdepth block sizes.[1]
[1] https://chromium-review.googlesource.com/c/webm/libvpx/+/4224681
Change-Id: I0aa0b6e0f70449776f316879cafc4b830e86ea51
|
|
Added AVX2 intrinsic optimization for the following functions
1. vpx_variance8x4
2. vpx_variance8x8
3. vpx_variance8x16
This is a bit-exact change.
Instruction Count
cpu Resolution Reduction(%)
0 LOWRES2 0.698
0 MIDRES2 0.577
0 HDRES2 0.469
0 Average 0.582
Change-Id: Iae8fdf9344fd012cda4955ed140633141d60ba86
|
|
The shift instructions have marginally worse performance on some
micro-architectures, and the vget_{low,high} instructions are
unnecessary.
This commit improves performance of the d135 predictors by 1.5% geomean
averaged across a range of compilers and micro-architectures.
Change-Id: Ied4c3eecc12fc973841696459d868ce403ed4e6c
|
|
Use sum_neon.h helpers for horizontal reductions in Neon DC predictors,
enabling use of dedicated Neon reduction instructions on AArch64. Some
of the surrounding code is also optimized to remove redundant broadcast
instructions in the dc_store helpers.
Performance is largely unchanged on both the standard as well as the
high bit-depth predictors. The main improvement appears to be the 16x16
standard-bitdepth dc predictor, which improves by 10-15% when
benchmarked on Neoverse N1.
Change-Id: Ibfcc6ecf4b1b2f87ce1e1f63c314d0cc35a0c76f
|
|
* changes:
Avoid LD2/ST2 instructions in highbd v predictors in Neon
Avoid interleaving loads/stores in Neon for highbd dc predictor
Avoid LD2/ST2 instructions in vpx_dc_predictor_32x32_neon
|
|
For these block sizes there is no need to widen to 32-bits until the
final reduction, so use a single vabaq instead of vabd + vpadalq.
Change-Id: I9c19d620f7bb8b3a6b0bedd37789c03bb628b563
|
|
The interleaving load/store instructions (LD2/LD3/LD4 and ST2/ST3/ST4)
are useful if we are dealing with interleaved data (e.g. real/imag
components of complex numbers), but for simply loading or storing larger
quantities of data it is preferable to simply use the normal load/store
instructions.
This patch replaces such occurrences in the two larger block sizes:
vpx_highbd_v_predictor_16x16_neon and vpx_highbd_v_predictor_32x32_neon.
Change-Id: Ie4ffa298a2466ceaf893566fd0aefe3f66f439e4
|
|
The interleaving load/store instructions (LD2/LD3/LD4 and ST2/ST3/ST4)
are useful if we are dealing with interleaved data (e.g. real/imag
components of complex numbers), but for simply loading or storing larger
quantities of data it is preferable to simply use two or more of the
normal load/store instructions.
This patch replaces such occurrences in the two larger block sizes:
vpx_highbd_dc_predictor_16x16_neon, vpx_highbd_dc_predictor_32x32_neon,
and related helper functions.
Speedups over the original Neon code (higher is better):
Microarch. | Compiler | Block | Speedup
Neoverse N1 | LLVM 15 | 16x16 | 1.25
Neoverse N1 | LLVM 15 | 32x32 | 1.13
Neoverse N1 | GCC 12 | 16x16 | 1.56
Neoverse N1 | GCC 12 | 32x32 | 1.52
Neoverse V1 | LLVM 15 | 16x16 | 1.63
Neoverse V1 | LLVM 15 | 32x32 | 1.08
Neoverse V1 | GCC 12 | 16x16 | 1.59
Neoverse V1 | GCC 12 | 32x32 | 1.37
Change-Id: If5ec220aba9dd19785454eabb0f3d6affec0cc8b
|
|
The LD2 and ST2 instructions are useful if we are dealing with
interleaved data (e.g. real/imag components of complex numbers), but for
simply loading or storing larger quantities of data it is preferable to
simply use two of the normal load/store instructions.
This patch replaces such occurrences in vpx_dc_predictor_32x32_neon and
related functions.
With Clang-15 this speeds up this function by 10-30% depending on the
micro-architecture being benchmarked on. With GCC-12 this speeds up the
function by 40-60% depending on the micro-architecture being benchmarked
on.
Change-Id: I670dc37908aa238f360104efd74d6c2108ecf945
|