[FFmpeg-devel] [PATCH] x86: hevc_mc: remove non necessary moves

[James Almer]

On 02/02/15 2:11 PM, Christophe Gisquet wrote:
> @@ -87,11 +95,22 @@ QPEL_TABLE 12, 4, w, sse4
>  %elif %1 <= 8
>      movdqa            %3, [%2]                                              ; load data from source2
>  %elif %1 <= 12
> +%if avx_enabled

If this is meant for avx2, then the correct check is "cpuflag(avx2)" or "mmsize > 16", 
because if at some point we add an avx version (just to take advantage of the 
non-destructive three operand format compared to sse4), they will have to be replaced.

> +    mova              %3, [%2]
> +%else
>      movdqa            %3, [%2]                                              ; load data from source2
>      movq              %4, [%2+16]                                           ; load data from source2

[...]

> @@ -589,6 +731,89 @@ void ff_hevc_dsp_init_x86(HEVCDSPContext *c, const int bit_depth)
>              c->idct_dc[3] = ff_hevc_idct32x32_dc_8_avx2;
>              if (ARCH_X86_64) {
>                  SAO_BAND_INIT(8, avx2);

Indentation for all the stuff bellow is wrong.

> +            c->put_hevc_epel[7][0][0] = ff_hevc_put_hevc_pel_pixels32_8_avx2;
> +            c->put_hevc_epel[8][0][0] = ff_hevc_put_hevc_pel_pixels48_8_avx2;
> +            c->put_hevc_epel[9][0][0] = ff_hevc_put_hevc_pel_pixels64_8_avx2;
> +
> +            c->put_hevc_qpel[7][0][0] = ff_hevc_put_hevc_pel_pixels32_8_avx2;
> +            c->put_hevc_qpel[8][0][0] = ff_hevc_put_hevc_pel_pixels48_8_avx2;
> +            c->put_hevc_qpel[9][0][0] = ff_hevc_put_hevc_pel_pixels64_8_avx2;
> +
> +            c->put_hevc_epel_uni[7][0][0] = ff_hevc_put_hevc_uni_pel_pixels32_8_avx2;
> +            c->put_hevc_epel_uni[8][0][0] = ff_hevc_put_hevc_uni_pel_pixels48_8_avx2;
> +            c->put_hevc_epel_uni[9][0][0] = ff_hevc_put_hevc_uni_pel_pixels64_8_avx2;
> +
> +            c->put_hevc_qpel_uni[7][0][0] = ff_hevc_put_hevc_uni_pel_pixels32_8_avx2;
> +            c->put_hevc_qpel_uni[8][0][0] = ff_hevc_put_hevc_uni_pel_pixels48_8_avx2;
> +            c->put_hevc_qpel_uni[9][0][0] = ff_hevc_put_hevc_uni_pel_pixels64_8_avx2;
> +
> +            c->put_hevc_qpel_bi[7][0][0] = ff_hevc_put_hevc_bi_pel_pixels32_8_avx2;
> +            c->put_hevc_qpel_bi[8][0][0] = ff_hevc_put_hevc_bi_pel_pixels48_8_avx2;
> +            c->put_hevc_qpel_bi[9][0][0] = ff_hevc_put_hevc_bi_pel_pixels64_8_avx2;
> +
> +            c->put_hevc_epel_bi[7][0][0] = ff_hevc_put_hevc_bi_pel_pixels32_8_avx2;
> +            c->put_hevc_epel_bi[8][0][0] = ff_hevc_put_hevc_bi_pel_pixels48_8_avx2;
> +            c->put_hevc_epel_bi[9][0][0] = ff_hevc_put_hevc_bi_pel_pixels64_8_avx2;
> +
> +            c->put_hevc_epel[7][0][1] = ff_hevc_put_hevc_epel_h32_8_avx2;
> +            c->put_hevc_epel[8][0][1] = ff_hevc_put_hevc_epel_h48_8_avx2;
> +            c->put_hevc_epel[9][0][1] = ff_hevc_put_hevc_epel_h64_8_avx2;
> +
> +            c->put_hevc_epel_uni[7][0][1] = ff_hevc_put_hevc_uni_epel_h32_8_avx2;
> +            c->put_hevc_epel_uni[8][0][1] = ff_hevc_put_hevc_uni_epel_h48_8_avx2;
> +            c->put_hevc_epel_uni[9][0][1] = ff_hevc_put_hevc_uni_epel_h64_8_avx2;
> +
> +            c->put_hevc_epel_bi[7][0][1] = ff_hevc_put_hevc_bi_epel_h32_8_avx2;
> +            c->put_hevc_epel_bi[8][0][1] = ff_hevc_put_hevc_bi_epel_h48_8_avx2;
> +            c->put_hevc_epel_bi[9][0][1] = ff_hevc_put_hevc_bi_epel_h64_8_avx2;
> +
> +            c->put_hevc_epel[7][1][0] = ff_hevc_put_hevc_epel_v32_8_avx2;
> +            c->put_hevc_epel[8][1][0] = ff_hevc_put_hevc_epel_v48_8_avx2;
> +            c->put_hevc_epel[9][1][0] = ff_hevc_put_hevc_epel_v64_8_avx2;
> +
> +            c->put_hevc_epel_uni[7][1][0] = ff_hevc_put_hevc_uni_epel_v32_8_avx2;
> +            c->put_hevc_epel_uni[8][1][0] = ff_hevc_put_hevc_uni_epel_v48_8_avx2;
> +            c->put_hevc_epel_uni[9][1][0] = ff_hevc_put_hevc_uni_epel_v64_8_avx2;
> +
> +            c->put_hevc_epel_bi[7][1][0] = ff_hevc_put_hevc_bi_epel_v32_8_avx2;
> +            c->put_hevc_epel_bi[8][1][0] = ff_hevc_put_hevc_bi_epel_v48_8_avx2;
> +            c->put_hevc_epel_bi[9][1][0] = ff_hevc_put_hevc_bi_epel_v64_8_avx2;
> +
> +            c->put_hevc_epel[7][1][1] = ff_hevc_put_hevc_epel_hv32_8_avx2;
> +            c->put_hevc_epel[8][1][1] = ff_hevc_put_hevc_epel_hv48_8_avx2;
> +            c->put_hevc_epel[9][1][1] = ff_hevc_put_hevc_epel_hv64_8_avx2;
> +
> +            c->put_hevc_epel_uni[7][1][1] = ff_hevc_put_hevc_uni_epel_hv32_8_avx2;
> +            c->put_hevc_epel_uni[8][1][1] = ff_hevc_put_hevc_uni_epel_hv48_8_avx2;
> +            c->put_hevc_epel_uni[9][1][1] = ff_hevc_put_hevc_uni_epel_hv64_8_avx2;
> +
> +            c->put_hevc_epel_bi[7][1][1] = ff_hevc_put_hevc_bi_epel_hv32_8_avx2;
> +            c->put_hevc_epel_bi[8][1][1] = ff_hevc_put_hevc_bi_epel_hv48_8_avx2;
> +            c->put_hevc_epel_bi[9][1][1] = ff_hevc_put_hevc_bi_epel_hv64_8_avx2;
> +
> +            c->put_hevc_qpel[7][0][1] = ff_hevc_put_hevc_qpel_h32_8_avx2;
> +            c->put_hevc_qpel[8][0][1] = ff_hevc_put_hevc_qpel_h48_8_avx2;
> +            c->put_hevc_qpel[9][0][1] = ff_hevc_put_hevc_qpel_h64_8_avx2;
> +
> +            c->put_hevc_qpel[7][1][0] = ff_hevc_put_hevc_qpel_v32_8_avx2;
> +            c->put_hevc_qpel[8][1][0] = ff_hevc_put_hevc_qpel_v48_8_avx2;
> +            c->put_hevc_qpel[9][1][0] = ff_hevc_put_hevc_qpel_v64_8_avx2;
> +
> +            c->put_hevc_qpel_uni[7][0][1] = ff_hevc_put_hevc_uni_qpel_h32_8_avx2;
> +            c->put_hevc_qpel_uni[8][0][1] = ff_hevc_put_hevc_uni_qpel_h48_8_avx2;
> +            c->put_hevc_qpel_uni[9][0][1] = ff_hevc_put_hevc_uni_qpel_h64_8_avx2;
> +
> +            c->put_hevc_qpel_uni[7][1][0] = ff_hevc_put_hevc_uni_qpel_v32_8_avx2;
> +            c->put_hevc_qpel_uni[8][1][0] = ff_hevc_put_hevc_uni_qpel_v48_8_avx2;
> +            c->put_hevc_qpel_uni[9][1][0] = ff_hevc_put_hevc_uni_qpel_v64_8_avx2;
> +
> +            c->put_hevc_qpel_bi[7][0][1] = ff_hevc_put_hevc_bi_qpel_h32_8_avx2;
> +            c->put_hevc_qpel_bi[8][0][1] = ff_hevc_put_hevc_bi_qpel_h48_8_avx2;
> +            c->put_hevc_qpel_bi[9][0][1] = ff_hevc_put_hevc_bi_qpel_h64_8_avx2;
> +
> +            c->put_hevc_qpel_bi[7][1][0] = ff_hevc_put_hevc_bi_qpel_v32_8_avx2;
> +            c->put_hevc_qpel_bi[8][1][0] = ff_hevc_put_hevc_bi_qpel_v48_8_avx2;
> +            c->put_hevc_qpel_bi[9][1][0] = ff_hevc_put_hevc_bi_qpel_v64_8_avx2;
>              }

Mickaël: How much faster than sse4 is this code? It uses lots of perms/inserts/extracts 
that cross ymm reg lanes, which is usually slow.