On 26/05/2024 00:45, James Almer wrote:
> On 5/25/2024 7:31 PM, James Almer wrote:
>> On 5/25/2024 5:57 PM, Lynne via ffmpeg-devel wrote:
>>> The inline asm function had issues running under checkasm.
>>> So I came to finish what I started, and wrote the last part
>>> of LPC computation in assembly.
>>>
>>> autocorr_10_c: 135525.8
>>> autocorr_10_sse2: 50729.8
>>> autocorr_10_fma3: 19007.8
>>> autocorr_30_c: 390100.8
>>> autocorr_30_sse2: 142478.8
>>> autocorr_30_fma3: 50559.8
>>> autocorr_32_c: 407058.3
>>> autocorr_32_sse2: 151633.3
>>> autocorr_32_fma3: 50517.3
>>> ---
>>>   libavcodec/x86/lpc.asm    | 91 +++++++++++++++++++++++++++++++++++++++
>>>   libavcodec/x86/lpc_init.c | 87 ++++---------------------------------
>>>   2 files changed, 100 insertions(+), 78 deletions(-)
>>>
>>> diff --git a/libavcodec/x86/lpc.asm b/libavcodec/x86/lpc.asm
>>> index a585c17ef5..790841b7f4 100644
>>> --- a/libavcodec/x86/lpc.asm
>>> +++ b/libavcodec/x86/lpc.asm
>>> @@ -32,6 +32,8 @@ dec_tab_sse2: times 2 dq -2.0
>>>   dec_tab_scalar: times 2 dq -1.0
>>>   seq_tab_sse2: dq 1.0, 0.0
>>> +autoc_init_tab: times 4 dq 1.0
>>> +
>>>   SECTION .text
>>>   %macro APPLY_WELCH_FN 0
>>> @@ -261,3 +263,92 @@ APPLY_WELCH_FN
>>>   INIT_YMM avx2
>>>   APPLY_WELCH_FN
>>>   %endif
>>> +
>>> +%macro COMPUTE_AUTOCORR_FN 0
>>> +cglobal lpc_compute_autocorr, 4, 7, 8, data, len, lag, autoc, lag_p, 
>>> data_l, len_p
>>
>> Already mentioned, but it should be 3 not 8.
>>
>>> +
>>> +    shl lagd, 3
>>> +    shl lenq, 3
>>> +    xor lag_pq, lag_pq
>>> +
>>> +.lag_l:
>>> +    movaps m8, [autoc_init_tab]
>>
>> m2
>>
>>> +
>>> +    mov len_pq, lag_pq
>>> +
>>> +    lea data_lq, [lag_pq + mmsize - 8]
>>> +    neg data_lq                     ; -j - mmsize
>>> +    add data_lq, dataq              ; data[-j - mmsize]
>>> +.len_l:
>>> +    ; We waste the upper value here on SSE2,
>>> +    ; but we use it on AVX.
>>> +    movupd xm0, [dataq + len_pq]    ; data[i]
>>
>> movsd
>>
>>> +    movupd m1, [data_lq + len_pq]   ; data[i - j]
>>> +
>>> +%if cpuflag(avx)
>>
>> %if mmsize == 32 here and everywhere else.
>>
>>> +    vbroadcastsd m0, xm0
>>
>> This is AVX2. AVX only has memory input argument. So use that and save 
>> the movsd from above for the FMA3 version.
>>
>>> +    vperm2f128 m1, m1, m1, 0x01
>>
>> Aren't you loading 16 extra bytes for no reason if you're just going 
>> to use the upper 16 bytes from the load above?
> 
> Nevermind, this is swapping lanes.
> 
> That aside, these versions are barely better and sometimes worse in all 
> my tests on win64 with GCC with certain seeds.
> For example, seed 4022958484 gives me:
> 
> autocorr_10_c: 21345.6
> autocorr_10_sse2: 16434.6
> autocorr_10_fma3: 24154.6
> autocorr_30_c: 59239.1
> autocorr_30_sse2: 46114.6
> autocorr_30_fma3: 64147.1
> autocorr_32_c: 63022.1
> autocorr_32_sse2: 50040.1
> autocorr_32_fma3: 66594.1
> 
> But seed 2236774811 gives me:
> 
> autocorr_10_c: 37135.3
> autocorr_10_sse2: 26492.3
> autocorr_10_fma3: 32943.3
> autocorr_30_c: 102266.8
> autocorr_30_sse2: 72933.3
> autocorr_30_fma3: 85808.3
> autocorr_32_c: 106537.8
> autocorr_32_sse2: 77623.3
> autocorr_32_fma3: 85844.3
> 
> But if i force len to always be 4999 instead of its value varying 
> depending on seed, i consistently get things like:
> 
> autocorr_10_c: 40447.3
> autocorr_10_sse2: 39526.8
> autocorr_10_fma3: 42955.3
> autocorr_30_c: 111362.3
> autocorr_30_sse2: 111408.3
> autocorr_30_fma3: 116781.8
> autocorr_32_c: 122388.3
> autocorr_32_sse2: 119125.3
> autocorr_32_fma3: 114239.3
> 
> It would help if someone else could confirm this, but overall i don't 
> see any worthwhile gain here. The old inline version, for those seeds 
> where it worked, was somewhat faster.

The metrics given are on Zen 3, with Clang with compiler optimizations 
disabled.
We do not rely on compiler optimizations, and have plenty of assembly 
which turns out to be slower than modern compilers autovectorizing (even 
though we disable tree vectorization on GCC, that does not apply to 
simple loops like this one). On the other hand, we also support ancient 
compilers, and compilers which have no understanding of vectorization at 
all.
To illustrate how different results can look on different arches and 
compilers, and even platforms (you mentioned you tested only on win64):

Zen 3, gcc-9, O2:
autocorr_10_c: 48796.8
autocorr_10_sse2: 39571.8
autocorr_10_fma3: 30272.8
autocorr_30_c: 138499.3
autocorr_30_sse2: 114091.3
autocorr_30_fma3: 82114.3
autocorr_32_c: 146466.8
autocorr_32_sse2: 118400.8
autocorr_32_fma3: 80473.8

Zen 3, gcc-14, O2:
autocorr_10_c: 44981.3
autocorr_10_sse2: 36481.3
autocorr_10_fma3: 18418.8
autocorr_30_c: 129462.8
autocorr_30_sse2: 104175.3
autocorr_30_fma3: 48670.3
autocorr_32_c: 135625.3
autocorr_32_sse2: 109079.8
autocorr_32_fma3: 48670.3

Zen 3, clang-18, O2:
autocorr_10_c: 51872.6
autocorr_10_sse2: 48311.1
autocorr_10_fma3: 30070.1
autocorr_30_c: 145899.6
autocorr_30_sse2: 135793.1
autocorr_30_fma3: 79922.6
autocorr_32_c: 160443.1
autocorr_32_sse2: 147591.1
autocorr_32_fma3: 80075.6

Skylake, gcc-14, O2:
autocorr_10_c: 149251.0
autocorr_10_sse2: 133769.5
autocorr_10_fma3: 72886.0
autocorr_30_c: 396145.0
autocorr_30_sse2: 376618.5
autocorr_30_fma3: 194116.5
autocorr_32_c: 413219.0
autocorr_32_sse2: 400867.5
autocorr_32_fma3: 194117.5

Skylake, clang-18, O2:
autocorr_10_c: 153825.3
autocorr_10_sse2: 133774.3
autocorr_10_fma3: 72883.8
autocorr_30_c: 398339.8
autocorr_30_sse2: 376603.8
autocorr_30_fma3: 194098.8
autocorr_32_c: 432183.3
autocorr_32_sse2: 422583.8
autocorr_32_fma3: 194094.3

<Insert your favorite decade old compiler here>

But again, this is irrelevant, as we do not rely on compilers for 
optimizations. We help them as much as we can, and when they work, its 
nice, but that is in no way reliable, especially to turn down a patch 
like this.