Cycle times:
inv_txfm_add_16x64_dct_dct_0_8bpc_c: 3973.5
inv_txfm_add_16x64_dct_dct_0_8bpc_ssse3: 185.7
inv_txfm_add_16x64_dct_dct_1_8bpc_c: 37869.1
inv_txfm_add_16x64_dct_dct_1_8bpc_ssse3: 2103.1
inv_txfm_add_16x64_dct_dct_2_8bpc_c: 37822.9
inv_txfm_add_16x64_dct_dct_2_8bpc_ssse3: 2099.1
inv_txfm_add_16x64_dct_dct_3_8bpc_c: 37871.7
inv_txfm_add_16x64_dct_dct_3_8bpc_ssse3: 2663.5
inv_txfm_add_16x64_dct_dct_4_8bpc_c: 38002.9
inv_txfm_add_16x64_dct_dct_4_8bpc_ssse3: 2589.7
inv_txfm_add_32x64_dct_dct_0_8bpc_c: 8319.2
inv_txfm_add_32x64_dct_dct_0_8bpc_ssse3: 376.9
inv_txfm_add_32x64_dct_dct_1_8bpc_c: 85956.8
inv_txfm_add_32x64_dct_dct_1_8bpc_ssse3: 4298.1
inv_txfm_add_32x64_dct_dct_2_8bpc_c: 89906.2
inv_txfm_add_32x64_dct_dct_2_8bpc_ssse3: 4291.3
inv_txfm_add_32x64_dct_dct_3_8bpc_c: 83710.9
inv_txfm_add_32x64_dct_dct_3_8bpc_ssse3: 5589.5
inv_txfm_add_32x64_dct_dct_4_8bpc_c: 87733.5
inv_txfm_add_32x64_dct_dct_4_8bpc_ssse3: 5658.4
i...

This is a workaround so that the AVX2 implementation of deblock can
index the levels array starting from the level type, which causes it
to over-read by up to 3 bytes. This is intended to fix #269.

The exact relative speedup compared to C code is a bit vague and hard
to measure, depending on eactly how many filtered blocks are skipped,
as the NEON version always filters 16 pixels at a time, while the
C code can skip processing individual 4 pixel blocks.

Additionally, the checkasm benchmarking code runs the same function
repeatedly on the same buffer, which can make the filter take
different codepaths on each run, as the function updates the buffer
which will be used as input for the next run.

If tweaking the checkasm test data to try to avoid skipped blocks,
the relative speedups compared to C is between 2x and 5x, while
it is around 1x to 4x with the current checkasm test as such.

Benchmark numbers from a tweaked checkasm that avoids skipped
blocks:

                        Cortex A53     A72     A73
lpf_h_sb_uv_w4_8bpc_c:      2954.7  1399.3  1655.3
lpf_h_sb_uv_w4_8bpc_neon:    895.5   650.8   692.0
lpf_h_sb_uv_w6_8bpc_c:      3879.2  1917.2  2257.7
lpf_h_sb_uv_w6_8bpc_neon:   1125.6   759.5   838.4
lpf_h_sb_y_w4_8bpc_c:       6711.0  3275.5  3913.7
lpf_h_sb_y_w4_8bpc_neon:    1744.0  1342.1  1351.5
lpf_h_sb_y_w8_8bpc_c:      10695.7  6155.8  6638.9
lpf_h_sb_y_w8_8bpc_neon:    2146.5  1560.4  1609.1
lpf_h_sb_y_w16_8bpc_c:     11355.8  6292.0  6995.9
lpf_h_sb_y_w16_8bpc_neon:   2475.4  1949.6  1968.4
lpf_v_sb_uv_w4_8bpc_c:      2639.7  1204.8  1425.9
lpf_v_sb_uv_w4_8bpc_neon:    510.7   351.4   334.7
lpf_v_sb_uv_w6_8bpc_c:      3468.3  1757.1  2021.5
lpf_v_sb_uv_w6_8bpc_neon:    625.0   415.0   397.8
lpf_v_sb_y_w4_8bpc_c:       5428.7  2731.7  3068.5
lpf_v_sb_y_w4_8bpc_neon:    1172.6   792.1   768.0
lpf_v_sb_y_w8_8bpc_c:       8946.1  4412.8  5121.0
lpf_v_sb_y_w8_8bpc_neon:    1565.5  1063.6  1062.7
lpf_v_sb_y_w16_8bpc_c:      8978.9  4411.7  5112.0
lpf_v_sb_y_w16_8bpc_neon:   1775.0  1288.1  1236.7

Martin Storsjö authored 6 years ago and Jean-Baptiste Kempf committed 6 years ago

Relative speedup vs (autovectorized) C code:
                      Cortex A53    A72    A73
selfguided_3x3_8bpc_neon:   2.91   2.12   2.68
selfguided_5x5_8bpc_neon:   3.18   2.65   3.39
selfguided_mix_8bpc_neon:   3.04   2.29   2.98

The relative speedup vs non-vectorized C code is around 2.6-4.6x.

This fixes this compiler warning with MSVC:
../src/msac.c(148): warning C4267: '+=': conversion from 'size_t' to 'unsigned int', possible loss of data

Henrik Gramner authored 6 years ago and Henrik Gramner committed 6 years ago

Also make various minor optimizations/style fixes to the MSAC C functions.

Xuefeng Jiang authored 6 years ago and Henrik Gramner committed 6 years ago

intra_pred_paeth_w4_8bpc_c: 561.6
intra_pred_paeth_w4_8bpc_ssse3: 49.2
intra_pred_paeth_w8_8bpc_c: 1475.8
intra_pred_paeth_w8_8bpc_ssse3: 103.0
intra_pred_paeth_w16_8bpc_c: 4697.8
intra_pred_paeth_w16_8bpc_ssse3: 279.0
intra_pred_paeth_w32_8bpc_c: 13245.1
intra_pred_paeth_w32_8bpc_ssse3: 614.7
intra_pred_paeth_w64_8bpc_c: 32638.9
intra_pred_paeth_w64_8bpc_ssse3: 1477.6

This eases disambiguating these functions when looking at perf
profiles.

The width register has been set to clz(w)-24, not the other way
around. And the 32 bit prep function has got the h parameter in
r4, not in r5.

Martin Storsjö authored 6 years ago and Janne Grunau committed 6 years ago

For cases with indented, nested .if/.macro in asm.S, ident those
by 4 chars.

Some initial assembly files were indented to 4/16 columns, while all
the actual implementation files, starting with src/arm/64/mc.S, have
used 8/24 for indentation.

Xuefeng Jiang authored 6 years ago and Henrik Gramner committed 6 years ago

cfl_ac_444_w4_8bpc_c: 978.2
cfl_ac_444_w4_8bpc_ssse3: 110.4
cfl_ac_444_w8_8bpc_c: 2312.3
cfl_ac_444_w8_8bpc_ssse3: 197.5
cfl_ac_444_w16_8bpc_c: 4081.1
cfl_ac_444_w16_8bpc_ssse3: 274.1
cfl_ac_444_w32_8bpc_c: 9544.3
cfl_ac_444_w32_8bpc_ssse3: 617.1

Port of 65ee1233 for AVX-2
from Kyle Siefring to SSE41, and optimize SSSE3.

---------------------
x86_64:
------------------------------------------
before: cdef_dir_8bpc_ssse3: 110.3
 after: cdef_dir_8bpc_ssse3: 105.9
   new: cdef_dir_8bpc_sse4:   96.4
------------------------------------------

---------------------
x86_32:
------------------------------------------
before: cdef_dir_8bpc_ssse3: 120.6
 after: cdef_dir_8bpc_ssse3: 110.7
   new: cdef_dir_8bpc_sse4:  106.5
------------------------------------------

Port of c204da0f for AVX-2
from Kyle Siefring.

---------------------
x86_64:
------------------------------------------
before: cdef_filter_4x4_8bpc_ssse3: 141.7
 after: cdef_filter_4x4_8bpc_ssse3: 131.6
before: cdef_filter_4x4_8bpc_sse4: 128.3
 after: cdef_filter_4x4_8bpc_sse4: 119.0
------------------------------------------
before: cdef_filter_4x8_8bpc_ssse3: 253.4
 after: cdef_filter_4x8_8bpc_ssse3: 236.1
before: cdef_filter_4x8_8bpc_sse4: 228.5
 after: cdef_filter_4x8_8bpc_sse4: 213.2
------------------------------------------
before: cdef_filter_8x8_8bpc_ssse3: 429.6
 after: cdef_filter_8x8_8bpc_ssse3: 386.9
before: cdef_filter_8x8_8bpc_sse4: 379.9
 after: cdef_filter_8x8_8bpc_sse4: 335.9
------------------------------------------

---------------------
x86_32:
------------------------------------------
before: cdef_filter_4x4_8bpc_ssse3: 184.3
 after: cdef_filter_4x4_8bpc_ssse3: 163.3
before: cdef_filter_4x4_8bpc_sse4: 168.9
 after: cdef_filter_4x4_8bpc_sse4: 146.1
------------------------------------------
before: cdef_filter_4x8_8bpc_ssse3: 335.3
 after: cdef_filter_4x8_8bpc_ssse3: 280.7
before: cdef_filter_4x8_8bpc_sse4: 305.1
 after: cdef_filter_4x8_8bpc_sse4: 257.9
------------------------------------------
before: cdef_filter_8x8_8bpc_ssse3: 579.1
 after: cdef_filter_8x8_8bpc_ssse3: 500.5
before: cdef_filter_8x8_8bpc_sse4: 517.0
 after: cdef_filter_8x8_8bpc_sse4: 455.8
------------------------------------------

Port of dc2ae517 for AVX-2
from Kyle Siefring.

---------------------
x86_64:
------------------------------------------
cdef_filter_4x4_8bpc_ssse3: 141.7
cdef_filter_4x4_8bpc_sse4: 128.3
------------------------------------------
cdef_filter_4x8_8bpc_ssse3: 253.4
cdef_filter_4x8_8bpc_sse4: 228.5
------------------------------------------
cdef_filter_8x8_8bpc_ssse3: 429.6
cdef_filter_8x8_8bpc_sse4: 379.9
------------------------------------------

---------------------
x86_32:
------------------------------------------
cdef_filter_4x4_8bpc_ssse3: 184.3
cdef_filter_4x4_8bpc_sse4: 168.9
------------------------------------------
cdef_filter_4x8_8bpc_ssse3: 335.3
cdef_filter_4x8_8bpc_sse4: 305.1
------------------------------------------
cdef_filter_8x8_8bpc_ssse3: 579.1
cdef_filter_8x8_8bpc_sse4: 517.0
------------------------------------------

Cycle times:
inv_txfm_add_16x32_dct_dct_0_8bpc_c: 2464.6
inv_txfm_add_16x32_dct_dct_0_8bpc_ssse3: 121.6
inv_txfm_add_16x32_dct_dct_1_8bpc_c: 24751.6
inv_txfm_add_16x32_dct_dct_1_8bpc_ssse3: 1101.9
inv_txfm_add_16x32_dct_dct_2_8bpc_c: 24377.0
inv_txfm_add_16x32_dct_dct_2_8bpc_ssse3: 1117.2
inv_txfm_add_16x32_dct_dct_3_8bpc_c: 24155.6
inv_txfm_add_16x32_dct_dct_3_8bpc_ssse3: 2349.3
inv_txfm_add_16x32_dct_dct_4_8bpc_c: 24175.6
inv_txfm_add_16x32_dct_dct_4_8bpc_ssse3: 1642.0
inv_txfm_add_16x32_identity_identity_0_8bpc_c: 10304.7
inv_txfm_add_16x32_identity_identity_0_8bpc_ssse3: 137.7
inv_txfm_add_16x32_identity_identity_1_8bpc_c: 10341.6
inv_txfm_add_16x32_identity_identity_1_8bpc_ssse3: 137.9
inv_txfm_add_16x32_identity_identity_2_8bpc_c: 10299.9
inv_txfm_add_16x32_identity_identity_2_8bpc_ssse3: 253.9
inv_txfm_add_16x32_identity_identity_3_8bpc_c: 10331.4
inv_txfm_add_16x32_identity_identity_3_8bpc_ssse3: 369.7
inv_txfm_add_16x32_identity_identity_4_8bpc_c: 10360.4
inv_txfm_add_16x32_identity_identity_4_8bpc_ssse3: 484.0
inv_txfm_add_32x16_dct_dct_0_8bpc_c: 2288.4
inv_txfm_add_32x16_dct_dct_0_8bpc_ssse3: 142.3
inv_txfm_add_32x16_dct_dct_1_8bpc_c: 23819.9
inv_txfm_add_32x16_dct_dct_1_8bpc_ssse3: 1740.1
inv_txfm_add_32x16_dct_dct_2_8bpc_c: 23755.8
inv_txfm_add_32x16_dct_dct_2_8bpc_ssse3: 1641.4
inv_txfm_add_32x16_dct_dct_3_8bpc_c: 23839.9
inv_txfm_add_32x16_dct_dct_3_8bpc_ssse3: 1559.0
inv_txfm_add_32x16_dct_dct_4_8bpc_c: 23757.7
inv_txfm_add_32x16_dct_dct_4_8bpc_ssse3: 1579.0
inv_txfm_add_32x16_identity_identity_0_8bpc_c: 10381.7
inv_txfm_add_32x16_identity_identity_0_8bpc_ssse3: 126.3
inv_txfm_add_32x16_identity_identity_1_8bpc_c: 10402.5
inv_txfm_add_32x16_identity_identity_1_8bpc_ssse3: 126.5
inv_txfm_add_32x16_identity_identity_2_8bpc_c: 10429.2
inv_txfm_add_32x16_identity_identity_2_8bpc_ssse3: 244.9
inv_txfm_add_32x16_identity_identity_3_8bpc_c: 10382.0
inv_txfm_add_32x16_identity_identity_3_8bpc_ssse3: 491.0
inv_txfm_add_32x16_identity_identity_4_8bpc_c: 10381.0
inv_txfm_add_32x16_identity_identity_4_8bpc_ssse3: 468.0
inv_txfm_add_32x32_dct_dct_0_8bpc_c: 4168.2
inv_txfm_add_32x32_dct_dct_0_8bpc_ssse3: 204.0
inv_txfm_add_32x32_dct_dct_1_8bpc_c: 46306.2
inv_txfm_add_32x32_dct_dct_1_8bpc_ssse3: 2216.0
inv_txfm_add_32x32_dct_dct_2_8bpc_c: 46300.2
inv_txfm_add_32x32_dct_dct_2_8bpc_ssse3: 2194.2
inv_txfm_add_32x32_dct_dct_3_8bpc_c: 46350.1
inv_txfm_add_32x32_dct_dct_3_8bpc_ssse3: 3484.4
inv_txfm_add_32x32_dct_dct_4_8bpc_c: 46318.1
inv_txfm_add_32x32_dct_dct_4_8bpc_ssse3: 3440.9
inv_txfm_add_32x32_identity_identity_0_8bpc_c: 14663.1
inv_txfm_add_32x32_identity_identity_0_8bpc_ssse3: 179.0
inv_txfm_add_32x32_identity_identity_1_8bpc_c: 14737.0
inv_txfm_add_32x32_identity_identity_1_8bpc_ssse3: 179.2
inv_txfm_add_32x32_identity_identity_2_8bpc_c: 14640.4
inv_txfm_add_32x32_identity_identity_2_8bpc_ssse3: 179.1
inv_txfm_add_32x32_identity_identity_3_8bpc_c: 14638.5
inv_txfm_add_32x32_identity_identity_3_8bpc_ssse3: 663.8
inv_txfm_add_32x32_identity_identity_4_8bpc_c: 14635.6
inv_txfm_add_32x32_identity_identity_4_8bpc_ssse3: 663.9

As meson still doesn't allow specifying different cflags between
static and dynamic libraries, this still includes the dllexport
in the static library when built with default_library=both, but
it at least is avoided in static-only builds, and avoids
defining these symbols as dllexport in the callers' translation
units.

Henrik Gramner authored 6 years ago and Henrik Gramner committed 6 years ago

The second shift is constant.

Cycle times:
inv_txfm_add_8x32_dct_dct_0_8bpc_c: 1164.7
inv_txfm_add_8x32_dct_dct_0_8bpc_ssse3: 79.5
inv_txfm_add_8x32_dct_dct_1_8bpc_c: 11291.6
inv_txfm_add_8x32_dct_dct_1_8bpc_ssse3: 508.5
inv_txfm_add_8x32_dct_dct_2_8bpc_c: 10720.4
inv_txfm_add_8x32_dct_dct_2_8bpc_ssse3: 507.9
inv_txfm_add_8x32_dct_dct_3_8bpc_c: 12351.5
inv_txfm_add_8x32_dct_dct_3_8bpc_ssse3: 687.2
inv_txfm_add_8x32_dct_dct_4_8bpc_c: 10402.3
inv_txfm_add_8x32_dct_dct_4_8bpc_ssse3: 687.9
inv_txfm_add_8x32_identity_identity_0_8bpc_c: 3485.0
inv_txfm_add_8x32_identity_identity_0_8bpc_ssse3: 97.7
inv_txfm_add_8x32_identity_identity_1_8bpc_c: 3495.7
inv_txfm_add_8x32_identity_identity_1_8bpc_ssse3: 97.7
inv_txfm_add_8x32_identity_identity_2_8bpc_c: 3503.7
inv_txfm_add_8x32_identity_identity_2_8bpc_ssse3: 97.8
inv_txfm_add_8x32_identity_identity_3_8bpc_c: 3489.5
inv_txfm_add_8x32_identity_identity_3_8bpc_ssse3: 184.4
inv_txfm_add_8x32_identity_identity_4_8bpc_c: 3498.1
inv_txfm_add_8x32_identity_identity_4_8bpc_ssse3: 182.8
inv_txfm_add_32x8_dct_dct_0_8bpc_c: 1220.4
inv_txfm_add_32x8_dct_dct_0_8bpc_ssse3: 65.6
inv_txfm_add_32x8_dct_dct_1_8bpc_c: 11120.7
inv_txfm_add_32x8_dct_dct_1_8bpc_ssse3: 623.8
inv_txfm_add_32x8_dct_dct_2_8bpc_c: 12236.3
inv_txfm_add_32x8_dct_dct_2_8bpc_ssse3: 624.7
inv_txfm_add_32x8_dct_dct_3_8bpc_c: 10866.3
inv_txfm_add_32x8_dct_dct_3_8bpc_ssse3: 694.1
inv_txfm_add_32x8_dct_dct_4_8bpc_c: 10322.8
inv_txfm_add_32x8_dct_dct_4_8bpc_ssse3: 692.5
inv_txfm_add_32x8_identity_identity_0_8bpc_c: 3368.1
inv_txfm_add_32x8_identity_identity_0_8bpc_ssse3: 98.6
inv_txfm_add_32x8_identity_identity_1_8bpc_c: 3381.1
inv_txfm_add_32x8_identity_identity_1_8bpc_ssse3: 98.3
inv_txfm_add_32x8_identity_identity_2_8bpc_c: 3376.6
inv_txfm_add_32x8_identity_identity_2_8bpc_ssse3: 98.3
inv_txfm_add_32x8_identity_identity_3_8bpc_c: 3364.3
inv_txfm_add_32x8_identity_identity_3_8bpc_ssse3: 182.2
inv_txfm_add_32x8_identity_identity_4_8bpc_c: 3390.0
inv_txfm_add_32x8_identity_identity_4_8bpc_ssse3: 182.2

Xuefeng Jiang authored 6 years ago and Henrik Gramner committed 6 years ago

cfl_ac_420_w4_8bpc_c: 1621.0
cfl_ac_420_w4_8bpc_ssse3: 92.5
cfl_ac_420_w8_8bpc_c: 3344.1
cfl_ac_420_w8_8bpc_ssse3: 115.4
cfl_ac_420_w16_8bpc_c: 6024.9
cfl_ac_420_w16_8bpc_ssse3: 187.8
cfl_ac_422_w4_8bpc_c: 1762.5
cfl_ac_422_w4_8bpc_ssse3: 81.4
cfl_ac_422_w8_8bpc_c: 4941.2
cfl_ac_422_w8_8bpc_ssse3: 166.5
cfl_ac_422_w16_8bpc_c: 8261.8
cfl_ac_422_w16_8bpc_ssse3: 272.3

This check was already done in dav1d_parse_obus(), so it's added as an assert
here for extra precaution.

Its previous contents don't need to be preserved.

Janne Grunau authored 6 years ago and Jean-Baptiste Kempf committed 6 years ago

Fixes tests on big endian architectures.

And the API version as the file version.

This optimization is so small 10 runs with a fixed seed were needed
to get some relevant numbers. This has been done for 3x3 case only.

before:
mean=113265.42
stddev=954.392

after:
mean=112654.71
stddev=884.833

This optimization is so tiny we can't even see it in checkasm.
The only actual difference being the removal of a memory load, it has to
be better.

Refs #241, Closes #255.

Janne Grunau authored 6 years ago and Jean-Baptiste Kempf committed 6 years ago

Increments the soname revision number for this behavior change.
Removes the DAV1D_VERSION and DAV1D_VERSION_INT defines and
dav1d_version_vcs() and dav1d_version_int().
Also cleans up the version usage in dav1d CLI.
Refs #241, #255.

---------------------
x86_64:
------------------------------------------
cdef_dir_8bpc_c: 1023.1
cdef_dir_8bpc_ssse3: 110.3
cdef_dir_8bpc_avx2: 71.1
------------------------------------------

---------------------
x86_32:
------------------------------------------
cdef_dir_8bpc_c: 1074.8
cdef_dir_8bpc_ssse3: 120.6
------------------------------------------

Thanks to Ronald for the AVX2 XMM version which was a very good starting
point.