Commit Graph

600 Commits

Author SHA1 Message Date
Yi Luo
8440cc4817 Merge "Improve idct32x32_1024_add SSSE3 intrinsics performance" 2017-03-15 02:32:52 +00:00
Linfeng Zhang
c756eb01c8 Fix overflow issue in 32x32 idct NEON intrinsics
Similar issue as Change bc1c18e.

The PartialIDctTest.ResultsMatch test on vpx_idct32x32_135_add_neon()
in high bit-depth mode exposes 16-bit overflow in final stage of pass
2, when changing the test number from 1,000 to 1,000,000.

Change to use saturating add/sub for vpx_idct32x32_34_add_neon(),
vpx_idct32x32_135_add_neon and vpx_idct32x32_1024_add_neon() in high
bit-depth mode.

Change-Id: Iaec0e9aeab41a3fdb4e170d7e9b3ad1fda922f6f
2017-03-14 16:59:14 -07:00
Yi Luo
fedcf83f33 Improve idct32x32_1024_add SSSE3 intrinsics performance
- Function level speed improves ~12%.

Change-Id: I9b7dbddabf08c7d0f6b25264e6074d5ccbe39290
2017-03-14 14:04:08 -07:00
Linfeng Zhang
b0bfcc368c Merge "Add vpx_highbd_idct32x32_135_add_c()" 2017-03-13 18:49:01 +00:00
James Zern
48fca113d1 inv_txfm_ssse3,butterfly: fix win32 abi compatibility
only the first 3 parameters can be aligned to 16 as required by __m128i,
make them all pointers for consistency.

since:
07c48ccfe Improve idct32x32_34_add SSSE3 intrinsics performance

BUG=webm:1384

Change-Id: I0324f701e723a27cb470036a180693ba8829d01d
2017-03-10 19:57:17 -08:00
Yi Luo
327add990f Improve idct32x32_135_add SSSE3 intrinsics performance
- Split the inv txfm into three parts to avoid stack spillover.
- Function level speed improves ~12%.
- Use function and macro to remove some repeated code.

Change-Id: I14f5f072334fd766808cb52bf648df792e7379ee
2017-03-09 16:17:54 -08:00
Linfeng Zhang
77311e0dff Update vpx_idct32x32_1024_add_neon()
Most are cosmetics changes.
Speed has no change with clang 3.8, and about 5% faster with gcc 4.8.4

Tried the strategy used in 8x8 and 16x16 (which operations' orders are
similar to the C code), though speed gets better with gcc, it's worse
with clang.

Tried to remove store_in_output(), but speed gets worse.

Change-Id: I93c8d284e90836f98962bb23d63a454cd40f776e
2017-03-08 12:39:04 -08:00
Linfeng Zhang
48f5886605 Add vpx_highbd_idct32x32_135_add_c()
When eob is less than or equal to 135 for high-bitdepth 32x32 idct,
call this function.

BUG=webm:1301

Change-Id: I8a5864f5c076e449c984e602946547a7b09c9fe6
2017-03-08 10:46:33 -08:00
Linfeng Zhang
c4e5c54d69 cosmetics,dsp/arm/: vpx_idct32x32_{34,135}_add_neon()
No speed changes and disassembly is almost identical.

Change-Id: Id07996237d2607ca6004da5906b7d288b8307e1f
2017-03-08 08:58:32 -08:00
Linfeng Zhang
3cf5c213f1 cosmetics,dsp/arm/: rename a variable
Rename cospi_6_26_14_18N to cospi_6_26N_14_18N for consistency.

Change-Id: I00498b43bb612b368219a489b3adaa41729bf31a
2017-03-08 08:55:41 -08:00
Yi Luo
07c48ccfe0 Improve idct32x32_34_add SSSE3 intrinsics performance
- Split the transform into first half and second half.
- Reschedule the instructions to avoid stack spillover.
- Function level speed improves ~16%.

Change-Id: I166889840d23aa8a273eca00f6fbdae8b4566f35
2017-03-01 11:14:48 -08:00
James Zern
47d6f16a04 get_prob(): rationalize int types
promote the unsigned int calculation to uint64_t rather than int64_t for
type consistency

Change-Id: Ic34dee1dc707d9faf6a3ae250bfe39b60bef3438
2017-02-24 15:36:52 -08:00
Jerome Jiang
b1dcaf7f1e Merge "Fix segmentation fault caused by denoiser working with spatial SVC." 2017-02-22 04:44:55 +00:00
Yi Luo
6036a0d24f Following SSSE3 intrinsics functions also work for HBD
- vpx_idct8x8_12_add_ssse3
  vpx_idct8x8_64_add_ssse3
  vpx_idct32x32_34_add_ssse3
  vpx_idct32x32_135_add_ssse3
  vpx_idct32x32_1024_add_ssse3
- turn on unit tests.

Change-Id: I788b2b3b2074a6f3ab6a0e6f469c1327a123eff7
2017-02-21 12:37:53 -08:00
Jerome Jiang
0d1e5a21c4 Fix segmentation fault caused by denoiser working with spatial SVC.
Re-enable the affected test.
BUG=webm:1374

Change-Id: I98cd49403927123546d1d0056660b98c9cb8babb
2017-02-21 09:38:28 -08:00
Yi Luo
1f8e8e5bf1 Fix idct8x8 SSSE3 SingleExtremeCoeff unit tests
- In SSSE3 optimization, 16-bit addition and subtraction would
  overflow when input coefficient is 16-bit signed extreme values.
- Function-level speed becomes slower (unit ms):
  idct8x8_64: 284 -> 294
  idct8x8_12: 145 -> 158.

BUG=webm:1332

Change-Id: I1e4bf9d30a6d4112b8cac5823729565bf145e40b
2017-02-17 14:05:05 -08:00
James Zern
3e7025022e Merge "Add vpx_highbd_idct16x16_10_add_neon()" 2017-02-17 20:29:37 +00:00
Yi Luo
f62dcc9c33 Replace idct32x32_1024_add_ssse3 assembly with intrinsics
- Encoding/decoding test, BQTerrace_1920x1080_60.y4m, on
  i7-6700, no obvious user-level speed performance downgrade.
- Passed unit tests.

Change-Id: I20688e0dd3731021ec8fb4404734336f1a426bfc
2017-02-16 16:10:40 -08:00
Johann Koenig
a9b81da575 Merge "block error avx2: use tran_low_t" 2017-02-16 23:51:14 +00:00
Linfeng Zhang
0620081731 Add vpx_highbd_idct16x16_10_add_neon()
BUG=webm:1301

Change-Id: If686c8144764c4162458f0bc4bb1bbf6555c48ab
2017-02-16 15:13:50 -08:00
James Zern
0f014c97e5 Merge "Fix mips vpx_post_proc_down_and_across_mb_row_msa function" 2017-02-16 23:02:10 +00:00
Johann Koenig
06a82af0de Merge "correct bitdepth_conversion_sse2.h header guard" 2017-02-16 21:41:28 +00:00
Johann
6c2d732bf4 correct bitdepth_conversion_sse2.h header guard
Change-Id: Ic4ffd861608e67fe59bcb3a86010ce3ef11a5519
2017-02-16 12:43:33 -08:00
Yi Luo
1cb44945fb Merge "Add idct32x32_135_add SSSE3 intrinsics" 2017-02-16 20:43:29 +00:00
Johann
2104454607 block error avx2: use tran_low_t
Change-Id: Ic5f3a1f569d6f82afeaf4fcd7235374bb460db3c
2017-02-16 12:39:02 -08:00
Yi Luo
72a43e2378 Add idct32x32_135_add SSSE3 intrinsics
- Replace the corresponding assembly code.
- No user level speed performance degrade.
- Unit tests passed.

Change-Id: Idd0c5a4bad4976f1617c34100cb46e75e3b961e5
2017-02-16 11:29:34 -08:00
Johann
4682130b60 quantize_fp highbd ssse3: use tran_low_t for coeff
Change-Id: Iebade0efc0efbb0a80a0f3adbef4962e3a2f25e8
2017-02-16 07:40:56 -08:00
Johann
44600442dc bitdepth conversion: really use num elements
The previous implementation confused bit/bytes/elements. It was using
'32' as the multiplier but that was mistakenly adopted because a 32x32
transform embedded the stride.

Change-Id: Ieeb867a332416b9a40580b5e7c9b20088e9e691a
2017-02-16 15:02:48 +00:00
Kaustubh Raste
fddf66b741 Fix mips vpx_post_proc_down_and_across_mb_row_msa function
Added fix to handle non-multiple of 16 cols case for size 16

Change-Id: If3a6d772d112077c5e0a9be9e612e1148f04338c
2017-02-16 13:17:00 +05:30
Johann Koenig
b63e88e506 Merge "Use 'packssdw' for loading tran_low_t values" 2017-02-16 02:41:00 +00:00
Linfeng Zhang
106c342659 cosmetics,dsp/inv_txfm.c: reorder functions
Change-Id: Ie0f7689ebe230c68eadb22a32b14838c1a7543a6
2017-02-15 11:40:35 -08:00
Linfeng Zhang
81914ce68a Add vpx_highbd_idct16x16_38_add_neon()
BUG=webm:1301

Change-Id: Ic6cd8c1e63e1b7a997cbed221e20fff4c599e0fe
2017-02-15 09:12:02 -08:00
Linfeng Zhang
e07e74fb0f Add vpx_highbd_idct16x16_38_add_c()
When eob is less than or equal to 38 for high-bitdepth 16x16 idct,
call this function.

BUG=webm:1301

Change-Id: I09167f89d29c401f9c36710b0fd2d02644052060
2017-02-14 17:25:52 -08:00
Johann
327a02d77e Use 'packssdw' for loading tran_low_t values
This matches bitdepth_conversion_sse2.asm and produces substantially
better assembly. The old way had lots of 'movzwl' and 'shl' and storing
back to memory before loading into an xmm register.

Change-Id: Ib33e35354dfd691a4f8b1e39f4dbcbb14cd5302b
2017-02-14 22:39:49 +00:00
Linfeng Zhang
429e652809 Replace 14 with DCT_CONST_BITS in idct NEON functions' shifts
Change-Id: I2a39a3bb87516b04d273bc1c0f4a634e3fb6f0f6
2017-02-14 13:08:41 -08:00
clang-format
4b402746ca apply clang-format
Change-Id: I75e4a9e0b37bd4586f26c8d6c1fa27f3f6ff1bce
2017-02-14 12:45:52 -08:00
Yi Luo
c1a90dc160 Merge "Replace idct32x32_34_add_ssse3 assembly with intrinsics" 2017-02-14 20:13:27 +00:00
Yi Luo
bd86de1ac8 Replace idct32x32_34_add_ssse3 assembly with intrinsics
- No user-level speed performance change.
- Pass unit tests.

Change-Id: Idfc598e00f354265e41f6b3219f4734216c115c6
2017-02-14 10:38:36 -08:00
Linfeng Zhang
de9ae32b93 Merge "Add vpx_highbd_idct16x16_256_add_neon()" 2017-02-14 01:15:34 +00:00
Linfeng Zhang
5ad4159ebb Add vpx_highbd_idct16x16_256_add_neon()
BUG=webm:1301

Change-Id: I6bb755552a39bdd26eef3f449601f6a9766c65ec
2017-02-13 15:50:33 -08:00
Johann
5ecde212a8 fdct8x8 highbd neon: use tran_low_t for output
Change-Id: I100c4a1955d80bec4d28e82796b3e7f57e84d0ba
2017-02-13 22:16:14 +00:00
Linfeng Zhang
016933ad48 Add vpx_highbd_idct{16x16,32x32}_1_add_neon()
and update vpx_highbd_idct8x8_1_add_neon()

BUG=webm:1301

Change-Id: I18d1a0cbe98ba822d5194c1b4e13a4c29c5c75f4
2017-02-13 10:25:22 -08:00
James Zern
91f87e7513 Merge "Add vpx_idct16x16_38_add_neon()" 2017-02-11 03:42:36 +00:00
Linfeng Zhang
bc1c18e18c Add vpx_idct16x16_38_add_neon()
The RunQuantCheck() test on it exposes 16-bit overflow in stage 7 of
pass 2. Change to use saturating add/sub for both
vpx_idct16x16_38_add_neon() and vpx_idct16x16_256_add_neon() for high
bitdepth.

Change-Id: Ibf4c107a887553a52852cc582e28d38a5a5a2712
2017-02-08 12:15:22 -08:00
Yi Luo
ac04d11abc Replace idct8x8_12_add_ssse3 assembly code with intrinsics
- Performance achieves the same as assembly.
- Unit tests pass.

Change-Id: I6eacfbbd826b3946c724d78fbef7948af6406ccd
2017-02-08 10:07:45 -08:00
Linfeng Zhang
cf76ee2cb7 Add vpx_idct16x16_38_add_c()
When eob is less than or equal to 38 for 16x16 idct, call this function.

Change-Id: Ief6f3fb16a49ace3c92cebf4e220bf5bf52a6087
2017-02-07 09:40:51 -08:00
Linfeng Zhang
66695533a8 Merge "Update 16x16 8-bit idct NEON intrinsics" 2017-02-07 16:52:40 +00:00
Johann
641fda79bb highbd x86: consolidate tran_low_t conversions
Create new helper files specifically for converting tran_low_t types.

Change-Id: I7c4c458ef910f3b3d10a3cfbf9df4de7682fd905
2017-02-06 10:43:26 -08:00
Jingning Han
bb40844e32 Merge "Add SSSE3 intrinsic 8x8 inverse 2D-DCT" 2017-02-02 22:18:32 +00:00
Kaustubh Raste
5b10674b5c Merge "Add mips msa sum_squares_2d_i16 function" 2017-02-02 08:09:21 +00:00