Martin:


Thank you that point out some problems.
I will push a new patch about transform_luma_neon, after solving the problem.
And another patch about checkasm, which will cover all bitdepths of it. 
And the third patch to pass context as pointer all about checkasm.


------------------ 原始邮件 ------------------
发件人:                                                                                                                        "Martin Storsj�"                                                                                    <martin@martin.st&gt;;
发送时间:&nbsp;2023年4月12日(星期三) 晚上8:53
收件人:&nbsp;"FFmpeg development discussions and patches"<ffmpeg-devel@ffmpeg.org&gt;;
抄送:&nbsp;"徐福隆"<839789740@qq.com&gt;;
主题:&nbsp;Re: [FFmpeg-devel] [PATCH] codec/aarch64/hevc:add transform_luma_neon checkasm



On Sun, 9 Apr 2023, xufuji456 wrote:

&gt; got 56% speed up (run_count=1000, CPU=Cortex A53)
&gt; transform_4x4_luma_neon: 45 transform_4x4_luma_c: 103
&gt; ---
&gt; libavcodec/aarch64/hevcdsp_idct_neon.S&nbsp;&nbsp;&nbsp; | 51 ++++++++++++++++++++++-
&gt; libavcodec/aarch64/hevcdsp_init_aarch64.c |&nbsp; 2 +
&gt; tests/checkasm/hevc_idct.c&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; | 28 +++++++++++++
&gt; 3 files changed, 80 insertions(+), 1 deletion(-)

When sending updates to your patches, please mention somewhere (outside of 
the main git commit message) what has changed. When I got back to looking 
at your patches now, I see a handful of revisions of the same patch with 
no explanation of what has changed.

&gt;
&gt; diff --git a/libavcodec/aarch64/hevcdsp_idct_neon.S b/libavcodec/aarch64/hevcdsp_idct_neon.S
&gt; index 74a96957bf..fc683a6396 100644
&gt; --- a/libavcodec/aarch64/hevcdsp_idct_neon.S
&gt; +++ b/libavcodec/aarch64/hevcdsp_idct_neon.S
&gt; @@ -6,6 +6,7 @@
&gt;&nbsp; * Ported from arm/hevcdsp_idct_neon.S by
&gt;&nbsp; * Copyright (c) 2020 Reimar Döffinger
&gt;&nbsp; * Copyright (c) 2023 J. Dekker <jdek@itanimul.li&gt;
&gt; + * Copyright (c) 2023 xu fulong <839789740@qq.com&gt;
&gt;&nbsp; *
&gt;&nbsp; * This file is part of FFmpeg.
&gt;&nbsp; *
&gt; @@ -656,4 +657,52 @@ idct_dc 16, 8
&gt; idct_dc 16, 10
&gt; 
&gt; idct_dc 32, 8
&gt; -idct_dc 32, 10
&gt; \ No newline at end of file

The patch still does not apply cleanly on git master due to this. Please 
rebase your patch on top of the actual public git master.

&gt; +idct_dc 32, 10
&gt; +
&gt; +.macro tr4_luma_shift r0, r1, r2, r3, shift
&gt; +&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; saddl&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; v0.4s, \r0, \r2&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; // c0 = src0 + src2
&gt; +&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; saddl&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; v1.4s, \r2, \r3&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; // c1 = src2 + src3
&gt; +&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; ssubl&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; v2.4s, \r0, \r3&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; // c2 = src0 - src3
&gt; +&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; smull&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; v3.4s, \r1, v21.4h&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; // c3 = 74 * src1
&gt; +
&gt; +&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; saddl&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; v7.4s, \r0, \r3&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; // src0 + src3
&gt; +&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; ssubw&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; v7.4s, v7.4s, \r2&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; // src0 - src2 + src3
&gt; +&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; mul&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; v7.4s, v7.4s, v18.4s&nbsp;&nbsp;&nbsp; // dst2 = 74 * (src0 - src2 + src3)
&gt; +
&gt; +&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; mul&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; v5.4s, v0.4s, v19.4s&nbsp;&nbsp;&nbsp; // 29 * c0
&gt; +&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; mul&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; v6.4s, v1.4s, v20.4s&nbsp;&nbsp;&nbsp; // 55 * c1
&gt; +&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; add&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; v5.4s, v5.4s, v6.4s&nbsp;&nbsp;&nbsp;&nbsp; // 29 * c0 + 55 * c1
&gt; +&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; add&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; v5.4s, v5.4s, v3.4s&nbsp;&nbsp;&nbsp;&nbsp; // dst0 = 29 * c0 + 55 * c1 + c3
&gt; +
&gt; +&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; mul&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; v1.4s, v1.4s, v19.4s&nbsp;&nbsp;&nbsp; // 29 * c1
&gt; +&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; mul&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; v6.4s, v2.4s, v20.4s&nbsp;&nbsp;&nbsp; // 55 * c2
&gt; +&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; sub&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; v6.4s, v6.4s, v1.4s&nbsp;&nbsp;&nbsp;&nbsp; // 55 * c2 - 29 * c1
&gt; +&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; add&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; v6.4s, v6.4s, v3.4s&nbsp;&nbsp;&nbsp;&nbsp; // dst1 = 55 * c2 - 29 * c1 + c3
&gt; +
&gt; +&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; mul&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; v0.4s, v0.4s, v20.4s&nbsp;&nbsp;&nbsp; // 55 * c0
&gt; +&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; mul&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; v2.4s, v2.4s, v19.4s&nbsp;&nbsp;&nbsp; // 29 * c2
&gt; +&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; add&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; v0.4s, v0.4s, v2.4s&nbsp;&nbsp;&nbsp;&nbsp; // 55 * c0 + 29 * c2
&gt; +&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; sub&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; v0.4s, v0.4s, v3.4s&nbsp;&nbsp;&nbsp;&nbsp; // dst3 = 55 * c0 + 29 * c2 - c3
&gt; +
&gt; +&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; sqrshrn&nbsp;&nbsp;&nbsp;&nbsp; \r0, v5.4s, \shift
&gt; +&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; sqrshrn&nbsp;&nbsp;&nbsp;&nbsp; \r1, v6.4s, \shift
&gt; +&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; sqrshrn&nbsp;&nbsp;&nbsp;&nbsp; \r2, v7.4s, \shift
&gt; +&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; sqrshrn&nbsp;&nbsp;&nbsp;&nbsp; \r3, v0.4s, \shift
&gt; +.endm
&gt; +
&gt; +function ff_hevc_transform_luma_4x4_neon_8, export=1
&gt; +&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; ld1&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; {v28.4h-v31.4h}, [x0]
&gt; +&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; movi&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; v18.4s, #74
&gt; +&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; movi&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; v19.4s, #29
&gt; +&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; movi&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; v20.4s, #55
&gt; +&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; movi&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; v21.4h, #74
&gt; +
&gt; +&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; tr4_luma_shift v28.4h, v29.4h, v30.4h, v31.4h, #7
&gt; +&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; transpose_4x4H v28, v29, v30, v31, v22, v23, v24, v25
&gt; +
&gt; +&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; tr4_luma_shift v28.4h, v29.4h, v30.4h, v31.4h, #12
&gt; +&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; transpose_4x4H v28, v29, v30, v31, v22, v23, v24, v25
&gt; +
&gt; +&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; st1&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; {v28.4h-v31.4h}, [x0]
&gt; +&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; ret
&gt; +endfunc
&gt; diff --git a/libavcodec/aarch64/hevcdsp_init_aarch64.c b/libavcodec/aarch64/hevcdsp_init_aarch64.c
&gt; index a923bae35c..6605a39973 100644
&gt; --- a/libavcodec/aarch64/hevcdsp_init_aarch64.c
&gt; +++ b/libavcodec/aarch64/hevcdsp_init_aarch64.c
&gt; @@ -75,6 +75,7 @@ void ff_hevc_idct_4x4_dc_10_neon(int16_t *coeffs);
&gt; void ff_hevc_idct_8x8_dc_10_neon(int16_t *coeffs);
&gt; void ff_hevc_idct_16x16_dc_10_neon(int16_t *coeffs);
&gt; void ff_hevc_idct_32x32_dc_10_neon(int16_t *coeffs);
&gt; +void ff_hevc_transform_luma_4x4_neon_8(int16_t *coeffs);
&gt; void ff_hevc_sao_band_filter_8x8_8_neon(uint8_t *_dst, const uint8_t *_src,
&gt;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; ptrdiff_t stride_dst, ptrdiff_t stride_src,
&gt;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; const int16_t *sao_offset_val, int sao_left_class,
&gt; @@ -142,6 +143,7 @@ av_cold void ff_hevc_dsp_init_aarch64(HEVCDSPContext *c, const int bit_depth)
&gt;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; c-&gt;idct_dc[1]&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; = ff_hevc_idct_8x8_dc_8_neon;
&gt;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; c-&gt;idct_dc[2]&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; = ff_hevc_idct_16x16_dc_8_neon;
&gt;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; c-&gt;idct_dc[3]&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; = ff_hevc_idct_32x32_dc_8_neon;
&gt; +&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; c-&gt;transform_4x4_luma&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; = ff_hevc_transform_luma_4x4_neon_8;
&gt;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; c-&gt;sao_band_filter[0]&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; =
&gt;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; c-&gt;sao_band_filter[1]&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; =
&gt;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; c-&gt;sao_band_filter[2]&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; =
&gt; diff --git a/tests/checkasm/hevc_idct.c b/tests/checkasm/hevc_idct.c
&gt; index 338b8a23e4..1c2b08d0f8 100644
&gt; --- a/tests/checkasm/hevc_idct.c
&gt; +++ b/tests/checkasm/hevc_idct.c
&gt; @@ -84,6 +84,27 @@ static void check_idct_dc(HEVCDSPContext h, int bit_depth)
&gt;&nbsp;&nbsp;&nbsp;&nbsp; }
&gt; }
&gt; 
&gt; +static void check_transform_luma(HEVCDSPContext h)

Unrelated to your patch, I see that this test file already does this 
elsewhere, but it is uncommon to pass such a context by value here; 
normally we'd pass it as a pointer to the subfunctions. But that's a 
preexisting issue, so stick with the current convention and we can change 
them all in a separate patch.

&gt; +{
&gt; +&nbsp;&nbsp;&nbsp; LOCAL_ALIGNED(32, int16_t, coeffs0, [32 * 32]);
&gt; +&nbsp;&nbsp;&nbsp; LOCAL_ALIGNED(32, int16_t, coeffs1, [32 * 32]);
&gt; +
&gt; +&nbsp;&nbsp;&nbsp; int block_size = 4;
&gt; +&nbsp;&nbsp;&nbsp; int size = block_size * block_size;
&gt; +&nbsp;&nbsp;&nbsp; declare_func_emms(AV_CPU_FLAG_MMXEXT, void, int16_t *coeffs);
&gt; +
&gt; +&nbsp;&nbsp;&nbsp; randomize_buffers(coeffs0, size);
&gt; +&nbsp;&nbsp;&nbsp; memcpy(coeffs1, coeffs0, sizeof(*coeffs0) * size);
&gt; +
&gt; +&nbsp;&nbsp;&nbsp; if (check_func(h.transform_4x4_luma, "hevc_transform_4x4_luma")) {
&gt; +&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; call_ref(coeffs0);
&gt; +&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; call_new(coeffs1);
&gt; +&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; if (memcmp(coeffs0, coeffs1, sizeof(*coeffs0) * size))
&gt; +&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; fail();
&gt; +&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; bench_new(coeffs1);
&gt; +&nbsp;&nbsp;&nbsp; }
&gt; +}
&gt; +
&gt; void checkasm_check_hevc_idct(void)
&gt; {
&gt;&nbsp;&nbsp;&nbsp;&nbsp; int bit_depth;
&gt; @@ -103,4 +124,11 @@ void checkasm_check_hevc_idct(void)
&gt;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; check_idct(h, bit_depth);
&gt;&nbsp;&nbsp;&nbsp;&nbsp; }
&gt;&nbsp;&nbsp;&nbsp;&nbsp; report("idct");
&gt; +
&gt; +&nbsp;&nbsp;&nbsp; bit_depth = 8;
&gt; +&nbsp;&nbsp;&nbsp; HEVCDSPContext h;

This causes warnings:

src/tests/checkasm/hevc_idct.c: In function ‘checkasm_check_hevc_idct’:
src/tests/checkasm/hevc_idct.c:129:5: warning: ISO C90 forbids mixed 
declaration
s and code [-Wdeclaration-after-statement]
&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; HEVCDSPContext h;
&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; ^~~~~~~~~~~~~~

And why not test all bitdepths like the other tests? Even if you might not 
be adding assembly for other bitdepths right now, it's best to make the 
test cover them all from the start.

// Martin
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel

To unsubscribe, visit link above, or email
ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".