rigayaの日記兼メモ帳

nvenc [info]: createDeviceFeatureList: Success.
nvenc [info]: InitDecoder: Success.
nvenc [info]: InitFilters: Success.
nvenc [info]: SetInputParam: Success.
nvenc [info]: m_pEncodeAPI->nvEncInitializeEncoder: Success.
nvenc [info]: CreateEncoder: Success.
nvenc [info]: AllocateIOBuffers: Success.
nvenc [info]: bitstream: Opened file "F:\temp\sample_movie_1080p.264"
nvenc [info]: bitstream: Added 8 MB output buffer.
nvenc [info]: Output: Initialized bitstream writer.
nvenc [info]: InitOutput: Success.
nvenc [info]: NVEnc (x86) 3.27 (r700) by rigaya, Jan  7 2018 21:47:36 (VC 1900/Win/avx2)
nvenc [info]: OS Version     Windows 10 x64 (16299)
nvenc [info]: CPU            Intel Core i9-7980XE @ 2.60GHz [TB: 4.01GHz] (18C/36T)
nvenc [info]: GPU            #0: GeForce GTX 1060 6GB (10 EU) @ 1708 MHz (388.31)
nvenc [info]: NVENC / CUDA   NVENC API 8.0, CUDA 9.1, schedule mode: auto
nvenc [info]: Input Buffers  CUDA, 36 frames
nvenc [info]: Input Info     auo(yuy2)->nv12 [AVX2], 1920x1080, 30000/1001 fps
nvenc [info]: Vpp Filters    copyHtoD
nvenc [info]: Output Info    H.264/AVC high @ Level auto
nvenc [info]:                1920x1080p 1:1 29.970fps (30000/1001fps)
nvenc [info]: Encoder Preset default
nvenc [info]: Rate Control   CQP  I:20  P:23  B:25
nvenc [info]: Lookahead      off
nvenc [info]: GOP length     300 frames
nvenc [info]: B frames       3 frames
nvenc [info]: Ref frames     3 frames, LTR: off
nvenc [info]: AQ             off
nvenc [info]: Others         mv:auto cabac deblock adapt-transform:auto bdirect:auto 
nvenc [info]:                                                                              
nvenc [info]: Flushed Encoder
nvenc [info]: bitstream: Closing...
nvenc [info]: bitstream: Closed file pointer.
nvenc [info]: bitstream: Closed.
nvenc [info]: encoded 5203 frames, 127.38 fps, 18173.62 kbps, 376.11 MB
nvenc [info]: frame type IDR   18
nvenc [info]: frame type I     18,  total size    5.91 MB
nvenc [info]: frame type P   1301,  total size  157.15 MB
nvenc [info]: frame type B   3884,  total size  213.05 MB
auo [info]: NVEnc エンコード : 0時間 0分42.2秒
nvenc [info]: unknown: Closing...
nvenc [info]: unknown: Close...
nvenc [info]: nvEncDestroyEncoder: success.
nvenc [info]: Closing perf monitor...

int main(int argc, char **argv) {
    alignas(64) int16_t a[32];
    alignas(64) int16_t b[32];

    //適当に初期化
    for (int i = 0; i < 32; i++) {
        a[i] = rand() % INT16_MAX;
        b[i] = rand() % INT16_MAX;
    }
    //ロード
    __m512i z0 = _mm512_load_si512((__m512i *)(a));
    __m512i z1 = _mm512_load_si512((__m512i *)(b));

    //マスクを作るよ
    __mmask32 mBlend = _mm512_cmpgt_epi16_mask(z0, z1);

    //マスクを使うよ
    __m512i z2 = _mm512_mask_mov_epi16(z0, mBlend, z1);

    _mm512_storeu_si512((__m512i*)a, z2);

    //最適化で全体が消されないようになんか出力しとく
    printf("%d\n", a[0]);

    return 0;
}

vmovdqa32 zmm0, ZMMWORD PTR _b$[esp+256]
vmovdqa32 zmm1, ZMMWORD PTR _a$[esp+256]

vpcmpw    k1, zmm1, zmm0, 6
kmovd    eax, k1
kmovd    k1, eax
vmovdqu16 zmm1 {k1}, zmm0

vmovdqu32 zmm0, zmm1
vmovd    eax, xmm0

vmovdqa64    zmm1, ZMMWORD PTR [esp+80]
mov    DWORD PTR [esp], OFFSET FLAT:LC0
vmovdqa64    zmm0, ZMMWORD PTR [esp+16]
vpcmpgtw    k1, zmm0, zmm1
vmovdqu16    zmm0{k1}, zmm1
vpextrw    eax, xmm0, 0