The csa_tables (which always consist of 32 entries of four byte each,
but the type depends upon whether the decoder is fixed or
floating-point) are currently initialized once during decoder
initialization; yet it turns out that this is actually no benefit: The
code used to initialize these tables takes up 153 (fixed point) and 122
(floating point) bytes when compiled with GCC 9.3 with -O3 on x64, so it
is better to just hardcode these tables.
Essentially the same applies to the is_tables: They have a size of 128B
each and the code to initialize them occupies 149 (fixed point) resp.
140 (floating point) bytes. So hardcode them, too.
To make the origin of the tables clear, references to the code used to
create them have been added.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
Commit 1af615683e put initializing
the ff_fft_offsets_lut (which is typically used if FFT_FIXED_32)
behind an ff_thread_once() to make ff_fft_init() thread-safe; yet
there is a second place where said table may be initialized which
is not guarded by this AVOnce: ff_fft_init_mips(). MIPS uses this LUT
even for ordinary floating point FFTs, so that ff_fft_init() is not
thread-safe (on MIPS) for both 32bit fixed-point as well as
floating-point FFTs; e.g. ff_mdct_init() inherits this flaw and
therefore initializing e.g. the AAC decoders is not thread-safe (on
MIPS) despite them having FF_CODEC_CAP_INIT_CLEANUP set.
This commit fixes this by moving the AVOnce to fft_init_table.c and
using it to guard all initializations of ff_fft_offsets_lut.
(It is not that bad in practice, because every entry of
ff_fft_offsets_lut is never read during initialization and is only once
ever written to (namely to its final value); but even these are
conflicting actions which are (by definition) data races and lead to
undefined behaviour.)
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
Test case fate-checkasm-h264pred failed in latest community code.
This patch fixed the bug.
Signed-off-by: Shiyou Yin <yinshiyou-hf@loongson.cn>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
'li.s' is a synthesized instruction, it does not work properly
when compiled with clang on mips, and A segfault occurred.
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Failed fate case: fate-h264-conformance-caba2_sony_e
Clang is more strict in the use of register constraint.
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Clang report following error in aacsbr_mips.c,ac3dsp_mips.c and aacdec_mips.c:
"couldn't allocate output register for constraint 'r'"
Use 'f' constraint for float variable.
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
GCC complains:
warning: listing the stack pointer register ‘$29’ in a clobber
list is deprecated [-Wdeprecated]
Actually stack pointer was restored at the end of the inline assembly
so there is no reason to add it to the clobber list.
Also use $sp insted of $29 to make our intention much more clear.
Signed-off-by: Jiaxun Yang <jiaxun.yang@flygoat.com>
Reviewed-by: Shiyou Yin <yinshiyou-hf@loongson.cn>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Apply optimized functions according to cpuflags.
MSA is usually put after MMI as it's generally faster than MMI.
Signed-off-by: Jiaxun Yang <jiaxun.yang@flygoat.com>
Reviewed-by: Shiyou Yin <yinshiyou-hf@loongson.cn>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
To enable runtime detection for MIPS, we need to refine ffbuild
part to support buildding these feature together.
Firstly, we fixed configure, let it probe native ability of toolchain
to decide wether a feature can to be enabled, also clearly marked
the conflictions between loongson2 & loongson3 and Release 6 & rest.
Secondly, we compile MMI and MSA C sources with their own flags to ensure
their flags won't pollute the whole program and generate illegal code.
Signed-off-by: Jiaxun Yang <jiaxun.yang@flygoat.com>
Reviewed-by: Shiyou Yin <yinshiyou-hf@loongson.cn>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
On other platforms, the functions are named get_cabac_inline_xxx but not
this one. There's also a define.
Signed-off-by: Rosen Penev <rosenp@gmail.com>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Place HAVE_MIPSFPU further up so that functions that use floating point
ASM are defined away. Otherwise compilation failures result when soft
float in enabled on the toolchain.
Signed-off-by: Rosen Penev <rosenp@gmail.com>
Performance of WMV3 decoding has speed up from 3.66x to 5.23x tested on 3A4000.
Reviewed-by: Shiyou Yin <yinshiyou-hf@loongson.cn>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Change the stride argument to ptrdiff_t in the following functions:
ff_put_no_rnd_vc1_chroma_mc8_mmi, ff_put_no_rnd_vc1_chroma_mc4_mmi,
ff_avg_no_rnd_vc1_chroma_mc8_mmi, ff_avg_no_rnd_vc1_chroma_mc4_mmi.
Reviewed-by: Shiyou Yin <yinshiyou-hf@loongson.cn>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Changing details as following:
1. The previous order of parameters are irregular and difficult to
understand. Adjust the order of the parameters according to the
rule: (RTYPE, input registers, input mask/input index/..., output registers).
Most of the existing msa macros follow the rule.
2. Remove the redundant macro SLDI_Bn_0 and use SLDI_Bn instead.
Reviewed-by: Shiyou Yin <yinshiyou-hf@loongson.cn>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
The indentation of code dose not reflect the if block structure in
'apply_ltp_mips', and this will generate a warnning when build with
'-Wall' or '-Wmisleading-indentation'.
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Changing details as following:
1. Remove the local variable 'out_m' in 'CLIP_SH' and store the result in
source vector.
2. Refine the implementation of macro 'CLIP_SH_0_255' and 'CLIP_SW_0_255'.
Performance of VP8 decoding has speed up about 1.1%(from 7.03x to 7.11x).
Performance of H264 decoding has speed up about 0.5%(from 4.35x to 4.37x).
Performance of Theora decoding has speed up about 0.7%(from 5.79x to 5.83x).
3. Remove redundant macro 'CLIP_SH/Wn_0_255_MAX_SATU' and use 'CLIP_SH/Wn_0_255'
instead, because there are no difference in the effect of this two macros.
Reviewed-by: Shiyou Yin <yinshiyou-hf@loongson.cn>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
In function ff_h264_add_pixels4_8_mmi, there is no need to reset '%[ftmp0]'
to 0, because it's value has never changed since the start of the asm block.
This patch remove the redundant 'xor' and set src to zero once it was loaded.
In function ff_h264_idct_add_8_mmi, 'block' is seted to zero twice.
This patch removed the first setting zero operation and move the second one
after the load operation of block.
In function ff_h264_idct8_add_8_mmi, 'block' is seted to zero twice too.
This patch just removed the second setting zero operation.
This patch mainly simplifies the implementation of functions above,
the effect on the performance of whole h264 decoding process is not obvious.
According to the perf data, proportion of ff_h264_idct_add_8_mmi decreased from
0.29% to 0.26% and ff_h264_idct8_add_8_mmi decreased from 0.62% to 0.59% when decoding
H264 format on loongson 3A3000(For reference only , not very stable.).
Reviewed-by: Reimar Döffinger <Reimar.Doeffinger@gmx.de>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Replace STnxm_UB and LDnxm_SH with new macros ST_{H/W/D}{1/2/4/8}.
The old macros are difficult to use because they don't follow the same parameter passing rules.
Changing details as following:
1. remove LD4x4_SH.
2. replace ST2x4_UB with ST_H4.
3. replace ST4x2_UB with ST_W2.
4. replace ST4x4_UB with ST_W4.
5. replace ST4x8_UB with ST_W8.
6. replace ST6x4_UB with ST_W2 and ST_H2.
7. replace ST8x1_UB with ST_D1.
8. replace ST8x2_UB with ST_D2.
9. replace ST8x4_UB with ST_D4.
10. replace ST8x8_UB with ST_D8.
11. replace ST12x4_UB with ST_D4 and ST_W4.
Examples of new macro: ST_H4(in, idx0, idx1, idx2, idx3, pdst, stride)
ST_H4 store four half-word elements in vector 'in' to pdst with stride.
About the macro name:
1) 'ST' means store operation.
2) 'H/W/D' means type of vector element is 'half-word/word/double-word'.
3) Number '1/2/4/8' means how many elements will be stored.
About the macro parameter:
1) 'in0, in1...' 128-bits vector.
2) 'idx0, idx1...' elements index.
3) 'pdst' destination pointer to store to
4) 'stride' stride of each store operation.
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
addi/daddi are deprecated by MIPS for years, and MIPS r6 remove
them.
They should be replace with addiu:
ADDIU performs the same arithmetic operation but
does not trap on overflow.
Reviewed-by: Shiyou Yin <yinshiyou-hf@loongson.cn>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
In function ff_dct_unquantize_mpeg2_intra_mmi,
addr0 shoudn't be changed before storage operation.
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Optimize put_hevc_qpel_h_8 with mmi in the case width=4/8/12/16/24/32/48/64.
This optimization improved HEVC decoding performance 2%(2.39x to 2.44x, tested on loongson 3A3000).
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Optimize put_hevc_qpel_bi_h_8 with mmi in the case width=4/8/12/16/24/32/48/64.
This optimization improved HEVC decoding performance 2.1%(2.34x to 2.39x, tested on loongson 3A3000).
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Optimize put_hevc_epel_bi_hv_8 with mmi in the case width=4/8/12/16/24/32.
This optimization improved HEVC decoding performance 1.7%(2.30x to 2.34x, tested on loongson 3A3000).
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Optimize put_hevc_qpel_uni_hv_8 with mmi in the case width=4/8/12/16/24/32/48/64.
This optimization improved HEVC decoding performance 2.7%(2.24x to 2.30x, tested on loongson 3A3000).
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Optimize put_hevc_qpel_bi_hv_8 with mmi in the case width=4/8/12/16/24/32/48/64.
This optimization improved HEVC decoding performance 11.4%(2.01x to 2.24x, tested on loongson 3A3000).
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Optimize put_hevc_qpel_hv_8 with mmi in the case width=4/8/12/16/24/32/48/64.
This optimization improved HEVC decoding performance 11%(1.81x to 2.01x, tested on loongson 3A3000).
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Optimize put_hevc_pel_bi_pixels_8 with mmi in the case width=8/16/24/32/48/64.
This optimization improved HEVC decoding performance 2%(1.77x to 1.81x, tested on loongson 3A3000).
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
The AV_INPUT_BUFFER_PADDING_SIZE has been increased to 64, but the value is still 32
in function ff_hevc_sao_edge_filter_8_msa. So, use AV_INPUT_BUFFER_PADDING_SIZE directly.
Also, use MAX_PB_SIZE directly instead of 64. Fate tests passed.
Reviewed-by: Derek Buitenhuis <derek.buitenhuis@gmail.com>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Set initialization order of MSA after MMI to make it work on loongson platform(msa is supported by loongson2k、3a4000 etc.).
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Remove invalid operation in the case x and y all equal 0,
this refine made about 2% speedup for H264 decode on loongson platform.
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
This optimization improved h264 decoding performance about 4%(from 74fps to 77fps, tested on loongson 3A3000).
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Combined 1st and 2nd loop into one inline asm in function ff_vc1_inv_trans_8x8_mmi to
reduce memory operation, and made some small optimization in ff_vc1_inv_trans_4x8_mmi.
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Failed case: svq3-watermark
When minimum loop count of following functions are greater than parameter h passed to them, svq3-watermark failed.
1. ff_put_pixels4_8_mmi
2. ff_avg_pixels4_8_mmi
3. ff_put_pixels4_l2_8_mmi
4. ff_avg_pixels4_l2_8_mmi
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Constraint "g" means compiler can store variable in memory or register.
When we use constraint "g" for a variable and this variable was operated by
instruction which only support register operands may lead "invalid operands" error.
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Simplify the usage of intermediate variable addr and remove unused variable all64
in following functions:
1. ff_put_pixels_clamped_mmi
2. ff_put_signed_pixels_clamped_mmi
3. ff_add_pixels_clamped_mmi
This optimization speed up mpeg4 decode about 2% on loongson platform(tested with 3A3000).
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Simplify the usage of intermediate variable addr in following functions:
1. ff_put_pixels4_8_mmi
2. ff_put_pixels8_8_mmi
3. ff_put_pixels16_8_mmi
4. ff_avg_pixels16_8_mmi.
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Failed case: mss2-wmv
In following functions, pmullh was used to multiply two 16-bit data, this will cause data overflow.
1. ff_vc1_inv_trans_8x8_dc_mmi
2. ff_vc1_inv_trans_8x8_mmi
3. ff_vc1_inv_trans_8x4_mmi
4. ff_vc1_inv_trans_4x8_mmi
5. ff_vc1_inv_trans_4x4_mmi
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Optimized memset with mmi in following functions:
1. ff_h264_add_pixels4_8_mmi.
2. ff_h264_idct_add_8_mmi.
3. ff_h264_idct8_add_8_mmi.
This optimization improved h264 decoding performance about 1.3%(tested on loongson 3A3000).
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Reoptimize function ff_put_h264_chroma_mc8_mmi and ff_avg_h264_chroma_mc8_mmi.
Performance of h264 decoding improved about 5%(from 69fps to 73fps, tested on loongson 3A3000).
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Performance of mpeg4 decoding improved about 23%(from 128fps to 158fps, tested on loongson 3A3000).
Reoptimized following functions with mmi.
1. ff_simple_idct_put_8_mmi
2. ff_simple_idct_add_8_mmi
3. ff_simple_idct_8_mmi
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Anton Khirnov
Andreas Rheinhardt
gxw
Shiyou Yin
Jiaxun Yang
Rosen Penev
Linjie Fu
YunQiang Su