[FFmpeg-devel] [PATCH v1 3/3] swscale/la: Add output_lasx.c file.

[Shiyou Yin]

> 2022年9月6日 16:12，Shiyou Yin <yinshiyou-hf at loongson.cn> 写道：
> 
>> 
>> 2022年8月29日 20:30，Andreas Rheinhardt <andreas.rheinhardt at outlook.com <mailto:andreas.rheinhardt at outlook.com>> 写道：
>> 
>> Hao Chen:
>>> ffmpeg -i ~/media/1_h264_1080p_30fps_3Mbps.mp4 -f rawvideo -s 640x480 -pix_fmt
>>> rgb24 -y /dev/null -an
>>> before: 150fps
>>> after: 183fps
>>> 
>>> Signed-off-by: Hao Chen <chenhao at loongson.cn>
>>> ---
>>> libswscale/loongarch/Makefile | 3 +-
>>> libswscale/loongarch/output_lasx.c | 1982 +++++++++++++++++
>>> libswscale/loongarch/swscale_init_loongarch.c | 3 +
>>> libswscale/loongarch/swscale_loongarch.h | 6 +
>>> 4 files changed, 1993 insertions(+), 1 deletion(-)
>>> create mode 100644 libswscale/loongarch/output_lasx.c
>>> 

>>> +static void
>>> +yuv2rgb_2_template_lasx(SwsContext *c, const int16_t *buf[2],
>>> + const int16_t *ubuf[2], const int16_t *vbuf[2],
>>> + const int16_t *abuf[2], uint8_t *dest, int dstW,
>>> + int yalpha, int uvalpha, int y,
>>> + enum AVPixelFormat target, int hasAlpha)
>>> +{
>>> + const int16_t *buf0 = buf[0], *buf1 = buf[1],
>>> + *ubuf0 = ubuf[0], *ubuf1 = ubuf[1],
>>> + *vbuf0 = vbuf[0], *vbuf1 = vbuf[1];
>>> + int yalpha1 = 4096 - yalpha;
>>> + int uvalpha1 = 4096 - uvalpha;
>>> + int i, count = 0;
>>> + int len = dstW - 15;
>>> + int len_count = (dstW + 1) >> 1;
>>> + const void *r, *g, *b;
>>> + int head = YUVRGB_TABLE_HEADROOM;
>>> + __m256i v_yalpha1 = __lasx_xvreplgr2vr_w(yalpha1);
>>> + __m256i v_uvalpha1 = __lasx_xvreplgr2vr_w(uvalpha1);
>>> + __m256i v_yalpha = __lasx_xvreplgr2vr_w(yalpha);
>>> + __m256i v_uvalpha = __lasx_xvreplgr2vr_w(uvalpha);
>>> + __m256i headroom = __lasx_xvreplgr2vr_w(head);
>>> +
>>> + for (i = 0; i < len; i += 16) {
>>> + int Y1, Y2, U, V;
>>> + int i_dex = i << 1;
>>> + int c_dex = count << 1;
>>> + __m256i y0_h, y0_l, y0, u0, v0;
>>> + __m256i y1_h, y1_l, y1, u1, v1;
>>> + __m256i y_l, y_h, u, v;
>>> +
>>> + DUP4_ARG2(__lasx_xvldx, buf0, i_dex, ubuf0, c_dex, vbuf0, c_dex,
>>> + buf1, i_dex, y0, u0, v0, y1);
>>> + DUP2_ARG2(__lasx_xvldx, ubuf1, c_dex, vbuf1, c_dex, u1, v1);
>>> + DUP2_ARG2(__lasx_xvsllwil_w_h, y0, 0, y1, 0, y0_l, y1_l);
>>> + DUP2_ARG1(__lasx_xvexth_w_h, y0, y1, y0_h, y1_h);
>>> + DUP4_ARG1(__lasx_vext2xv_w_h, u0, u1, v0, v1, u0, u1, v0, v1);
>>> + y0_l = __lasx_xvmul_w(y0_l, v_yalpha1);
>>> + y0_h = __lasx_xvmul_w(y0_h, v_yalpha1);
>>> + u0 = __lasx_xvmul_w(u0, v_uvalpha1);
>>> + v0 = __lasx_xvmul_w(v0, v_uvalpha1);
>>> + y_l = __lasx_xvmadd_w(y0_l, v_yalpha, y1_l);
>>> + y_h = __lasx_xvmadd_w(y0_h, v_yalpha, y1_h);
>>> + u = __lasx_xvmadd_w(u0, v_uvalpha, u1);
>>> + v = __lasx_xvmadd_w(v0, v_uvalpha, v1);
>>> + y_l = __lasx_xvsrai_w(y_l, 19);
>>> + y_h = __lasx_xvsrai_w(y_h, 19);
>>> + u = __lasx_xvsrai_w(u, 19);
>>> + v = __lasx_xvsrai_w(v, 19);
>>> + u = __lasx_xvadd_w(u, headroom);
>>> + v = __lasx_xvadd_w(v, headroom);
>>> + WRITE_YUV2RGB(y_l, y_l, u, v, 0, 1, 0, 0);
>>> + WRITE_YUV2RGB(y_l, y_l, u, v, 2, 3, 1, 1);
>>> + WRITE_YUV2RGB(y_h, y_h, u, v, 0, 1, 2, 2);
>>> + WRITE_YUV2RGB(y_h, y_h, u, v, 2, 3, 3, 3);
>>> + WRITE_YUV2RGB(y_l, y_l, u, v, 4, 5, 4, 4);
>>> + WRITE_YUV2RGB(y_l, y_l, u, v, 6, 7, 5, 5);
>>> + WRITE_YUV2RGB(y_h, y_h, u, v, 4, 5, 6, 6);
>>> + WRITE_YUV2RGB(y_h, y_h, u, v, 6, 7, 7, 7);
>>> + }
>>> + if (dstW - i >= 8) {
>>> + int Y1, Y2, U, V;
>>> + int i_dex = i << 1;
>>> + __m256i y0_l, y0, u0, v0;
>>> + __m256i y1_l, y1, u1, v1;
>>> + __m256i y_l, u, v;
>>> +
>>> + y0 = __lasx_xvldx(buf0, i_dex);
>> 
>> 1. Not long ago, I tried to constify the src pointer of several asm
>> functions and noticed that they produced new warnings for loongarch
>> (according to patchwork:
>> https://patchwork.ffmpeg.org/project/ffmpeg/patch/DB6PR0101MB2214178D3E6B8DCA5B86F8198F9A9@DB6PR0101MB2214.eurprd01.prod.exchangelabs.com/),
>> even though I was sure that the code is const-correct. After finding
>> (via https://github.com/opencv/opencv/pull/21833) a toolchain
>> (https://gitee.com/wenux/cross-compiler-la-on-x86) that can build the
>> lasx and lsx code (upstream GCC seems to be lacking lsx and lasx support
>> at the moment; at least, my self-compiled loongarch-GCC did not support
>> lsx and lasx) the issue was clear: lsxintrin.h and lasxintrin.h do not
>> use const at all, even for functions that only read data (I presume the
>> vl in __lsx_vldx stands for "vector load"?).
>> So I sent another iteration
>> https://ffmpeg.org/pipermail/ffmpeg-devel/2022-August/299562.html of
>> that patchset that now added wrappers for __lsx_vldx() and
>> __lasx_xvldx() and cc'ed you and some other developers from loongson to
>> alert you of the issue in the hope that you fix the headers, so that my
>> wrappers wouldn't need to be applied. That didn't work, as my mails
>> could not be delivered to you. So I applied the patchset.
>> 2. You use __lasx_xvldx() to read from a const int16_t. This will give
>> new warnings unless the above issue has been fixed. Has it?
>> 3. I don't know whether it has, as patchwork's fate tests don't work for
>> a few days already. Given that the mails I receive from patchwork when
>> it doesn't like a commit message arrive from "Patchwork
>> <yinshiyou-hf at loongson.cn>" I presume that loongson is now somehow
>> running patchwork, so you should be able to inform the right people to
>> fix it.
>> 4. If you fixed the const-issue, can you please make an updated
>> toolchain with lsx and lasx support enabled available to us?
>> 
>> - Andreas
>> 
> 
> Hi Andreas,
> 
> Sorry for the late reply.
> This issue will be fixed by using const for v1 of __lsx_vldx, and I will update toolchain of LoongArch patchwork runner ASAP.
> Thank you very much for analyzing this problem and giving suggestion. 
> 
> Thanks,
> Shiyou

Hi Andreas,
Has updated patchwork runner environment, you can revert the patchset which add wrappers for __lsx_[x]vldx.

Shiyou