[x265] [PATCH] asm: disabled 8bpp AVX & AVX2 primitives having less than 10% speed up over SSE

Deepthi Nandakumar deepthi at multicorewareinc.com
Thu Aug 20 06:32:13 CEST 2015


You have a bad merge issue - the patch doesnt apply again.

On Thu, Aug 20, 2015 at 1:47 AM, <aasaipriya at multicorewareinc.com> wrote:

> # HG changeset patch
> # User Aasaipriya Chandran <aasaipriya at multicorewareinc.com>
> # Date 1440015448 -19800
> #      Thu Aug 20 01:47:28 2015 +0530
> # Node ID 2c00cb2e448dc6612aeb7598e036d5c47676a6cf
> # Parent  cc5ceb11f902094a9755af5a3295e639da620c10
> asm: disabled 8bpp AVX & AVX2 primitives having less than 10% speed up
> over SSE
>
> these primitives are slower than SSE primitives
>
> diff -r cc5ceb11f902 -r 2c00cb2e448d source/common/x86/asm-primitives.cpp
> --- a/source/common/x86/asm-primitives.cpp      Wed Aug 19 16:46:00 2015
> +0530
> +++ b/source/common/x86/asm-primitives.cpp      Thu Aug 20 01:47:28 2015
> +0530
> @@ -2602,25 +2602,6 @@
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_8x8].satd =
> PFX(pixel_satd_8x8_avx);
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_8x32].satd =
> PFX(pixel_satd_8x32_avx);
>
> -        p.pu[LUMA_8x8].satd   = PFX(pixel_satd_8x8_avx);
> -        p.pu[LUMA_16x16].satd = PFX(pixel_satd_16x16_avx);
> -        p.pu[LUMA_32x32].satd = PFX(pixel_satd_32x32_avx);
> -        p.pu[LUMA_64x64].satd = PFX(pixel_satd_64x64_avx);
> -        p.pu[LUMA_8x16].satd  = PFX(pixel_satd_8x16_avx);
> -        p.pu[LUMA_16x32].satd = PFX(pixel_satd_16x32_avx);
> -        p.pu[LUMA_32x16].satd = PFX(pixel_satd_32x16_avx);
> -        p.pu[LUMA_64x32].satd = PFX(pixel_satd_64x32_avx);
> -        p.pu[LUMA_32x64].satd = PFX(pixel_satd_32x64_avx);
> -        p.pu[LUMA_16x12].satd = PFX(pixel_satd_16x12_avx);
> -
> -        p.pu[LUMA_32x24].satd = PFX(pixel_satd_32x24_avx);
> -        p.pu[LUMA_24x32].satd = PFX(pixel_satd_24x32_avx);
> -        p.pu[LUMA_8x32].satd  = PFX(pixel_satd_8x32_avx);
> -        p.pu[LUMA_64x48].satd = PFX(pixel_satd_64x48_avx);
> -        p.pu[LUMA_48x64].satd = PFX(pixel_satd_48x64_avx);
> -        p.pu[LUMA_64x16].satd = PFX(pixel_satd_64x16_avx);
> -        p.pu[LUMA_16x64].satd = PFX(pixel_satd_16x64_avx);
> -
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_8x8].satd =
> PFX(pixel_satd_8x8_avx);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_16x16].satd =
> PFX(pixel_satd_16x16_avx);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_32x32].satd =
> PFX(pixel_satd_32x32_avx);
> @@ -2631,20 +2612,11 @@
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_32x24].satd =
> PFX(pixel_satd_32x24_avx);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_24x32].satd =
> PFX(pixel_satd_24x32_avx);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_8x32].satd =
> PFX(pixel_satd_8x32_avx);
> -
> -        p.cu[BLOCK_64x64].sa8d = PFX(pixel_sa8d_64x64_avx);
> -        p.chroma[X265_CSP_I422].cu[BLOCK_422_8x16].sa8d =
> PFX(pixel_sa8d_8x16_avx);
> -        p.chroma[X265_CSP_I422].cu[BLOCK_422_16x32].sa8d =
> PFX(pixel_sa8d_16x32_avx);
> -        p.chroma[X265_CSP_I422].cu[BLOCK_422_32x64].sa8d =
> PFX(pixel_sa8d_32x64_avx);
> -
>          p.chroma[X265_CSP_I420].cu[BLOCK_420_32x32].sa8d =
> PFX(pixel_sa8d_32x32_avx);
>          p.chroma[X265_CSP_I420].cu[BLOCK_420_16x16].sa8d =
> PFX(pixel_sa8d_16x16_avx);
>          p.chroma[X265_CSP_I420].cu[BLOCK_420_8x8].sa8d =
> PFX(pixel_sa8d_8x8_avx);
> -
>          p.chroma[X265_CSP_I420].cu[BLOCK_420_8x8].sse_pp =
> PFX(pixel_ssd_8x8_avx);
>
> -        p.cu[BLOCK_16x16].var = PFX(pixel_var_16x16_avx);
> -
>          p.pu[LUMA_16x4].sad_x4  = PFX(pixel_sad_x4_16x4_avx);
>
>          p.cu[BLOCK_16x16].copy_ss = PFX(blockcopy_ss_16x16_avx);
> @@ -2693,21 +2665,24 @@
>          p.cu[BLOCK_8x8].sa8d   = PFX(pixel_sa8d_8x8_avx);
>          p.cu[BLOCK_16x16].sa8d = PFX(pixel_sa8d_16x16_avx);
>          p.cu[BLOCK_32x32].sa8d = PFX(pixel_sa8d_32x32_avx);
> +        p.cu[BLOCK_64x64].sa8d = PFX(pixel_sa8d_64x64_avx);
> +        p.chroma[X265_CSP_I420].cu[BLOCK_420_4x4].sa8d =
> PFX(pixel_satd_4x4_avx);
> +        p.chroma[X265_CSP_I422].cu[BLOCK_422_8x16].sa8d =
> PFX(pixel_sa8d_8x16_avx);
> +        p.chroma[X265_CSP_I422].cu[BLOCK_422_16x32].sa8d =
> PFX(pixel_sa8d_16x32_avx);
> +        p.chroma[X265_CSP_I422].cu[BLOCK_422_32x64].sa8d =
> PFX(pixel_sa8d_32x64_avx);
> +        p.pu[LUMA_4x4].satd = p.cu[BLOCK_4x4].sa8d =
> PFX(pixel_satd_4x4_avx);
> +
> +        ALL_LUMA_PU(satd, pixel_satd, avx);
>          p.pu[LUMA_4x4].satd = p.cu[BLOCK_4x4].sa8d =
> PFX(pixel_satd_4x4_avx);
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_8x4].satd =
> PFX(pixel_satd_8x4_avx);
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_4x8].satd =
> PFX(pixel_satd_4x8_avx);
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_4x16].satd =
> PFX(pixel_satd_4x16_avx);
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_4x4].satd =
> PFX(pixel_satd_4x4_avx);
> -        p.pu[LUMA_8x4].satd   = PFX(pixel_satd_8x4_avx);
> -        p.pu[LUMA_4x8].satd   = PFX(pixel_satd_4x8_avx);
> -        p.pu[LUMA_12x16].satd = PFX(pixel_satd_12x16_avx);
> -        p.pu[LUMA_4x16].satd  = PFX(pixel_satd_4x16_avx);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_4x4].satd =
> PFX(pixel_satd_4x4_avx);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_8x4].satd =
> PFX(pixel_satd_8x4_avx);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_4x8].satd =
> PFX(pixel_satd_4x8_avx);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_12x16].satd =
> PFX(pixel_satd_12x16_avx);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_4x16].satd =
> PFX(pixel_satd_4x16_avx);
> -        p.chroma[X265_CSP_I420].cu[BLOCK_420_4x4].sa8d =
> PFX(pixel_satd_4x4_avx);
>
>          p.cu[BLOCK_16x16].sse_pp = PFX(pixel_ssd_16x16_avx);
>          p.cu[BLOCK_32x32].sse_pp = PFX(pixel_ssd_32x32_avx);
> @@ -2716,9 +2691,8 @@
>          p.chroma[X265_CSP_I422].cu[BLOCK_422_16x32].sse_pp =
> PFX(pixel_ssd_16x32_avx);
>          p.chroma[X265_CSP_I422].cu[BLOCK_422_32x64].sse_pp =
> PFX(pixel_ssd_32x64_avx);
>          ASSIGN_SSE_SS(avx);
> -        p.cu[BLOCK_8x8].var   = PFX(pixel_var_8x8_avx);
> -        p.cu[BLOCK_32x32].var = PFX(pixel_var_32x32_avx);
> -        p.cu[BLOCK_64x64].var = PFX(pixel_var_64x64_avx)
> +
> +        LUMA_VAR(avx);
>
>          p.pu[LUMA_12x16].sad_x3 = PFX(pixel_sad_x3_12x16_avx);
>          p.pu[LUMA_16x4].sad_x3  = PFX(pixel_sad_x3_16x4_avx);
> @@ -2777,7 +2751,6 @@
>          p.saoCuOrgE1_2Rows = PFX(saoCuOrgE1_2Rows_avx2);
>          p.saoCuOrgE2[0] = PFX(saoCuOrgE2_avx2);
>          p.saoCuOrgE2[1] = PFX(saoCuOrgE2_32_avx2);
> -        p.saoCuOrgE3[0] = PFX(saoCuOrgE3_avx2);
>          p.saoCuOrgE3[1] = PFX(saoCuOrgE3_32_avx2);
>          p.saoCuOrgB0 = PFX(saoCuOrgB0_avx2);
>          p.sign = PFX(calSign_avx2);
> @@ -2793,34 +2766,24 @@
>          p.cu[BLOCK_16x16].psy_cost_pp = PFX(psyCost_pp_16x16_avx2);
>          p.cu[BLOCK_32x32].psy_cost_pp = PFX(psyCost_pp_32x32_avx2);
>          p.cu[BLOCK_64x64].psy_cost_pp = PFX(psyCost_pp_64x64_avx2);
> -
> -        p.pu[LUMA_8x16].addAvg = PFX(addAvg_8x16_avx2);
> -        p.pu[LUMA_8x32].addAvg = PFX(addAvg_8x32_avx2);
>          p.pu[LUMA_16x4].addAvg = PFX(addAvg_16x4_avx2);
>          p.pu[LUMA_16x8].addAvg = PFX(addAvg_16x8_avx2);
>          p.pu[LUMA_16x12].addAvg = PFX(addAvg_16x12_avx2);
>          p.pu[LUMA_16x16].addAvg = PFX(addAvg_16x16_avx2);
>          p.pu[LUMA_16x32].addAvg = PFX(addAvg_16x32_avx2);
>          p.pu[LUMA_16x64].addAvg = PFX(addAvg_16x64_avx2);
> -
>          p.pu[LUMA_24x32].addAvg = PFX(addAvg_24x32_avx2);
> -
>          p.pu[LUMA_32x8].addAvg = PFX(addAvg_32x8_avx2);
>          p.pu[LUMA_32x16].addAvg = PFX(addAvg_32x16_avx2);
>          p.pu[LUMA_32x24].addAvg = PFX(addAvg_32x24_avx2);
>          p.pu[LUMA_32x32].addAvg = PFX(addAvg_32x32_avx2);
>          p.pu[LUMA_32x64].addAvg = PFX(addAvg_32x64_avx2);
> -
>          p.pu[LUMA_48x64].addAvg = PFX(addAvg_48x64_avx2);
> -
>          p.pu[LUMA_64x16].addAvg = PFX(addAvg_64x16_avx2);
>          p.pu[LUMA_64x32].addAvg = PFX(addAvg_64x32_avx2);
>          p.pu[LUMA_64x48].addAvg = PFX(addAvg_64x48_avx2);
>          p.pu[LUMA_64x64].addAvg = PFX(addAvg_64x64_avx2);
>
> -        p.chroma[X265_CSP_I420].pu[CHROMA_420_8x8].addAvg =
> PFX(addAvg_8x8_avx2);
> -        p.chroma[X265_CSP_I420].pu[CHROMA_420_8x16].addAvg =
> PFX(addAvg_8x16_avx2);
> -        p.chroma[X265_CSP_I420].pu[CHROMA_420_8x32].addAvg =
> PFX(addAvg_8x32_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_16x4].addAvg =
> PFX(addAvg_16x4_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_16x8].addAvg =
> PFX(addAvg_16x8_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_16x12].addAvg =
> PFX(addAvg_16x12_avx2);
> @@ -2831,11 +2794,6 @@
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_32x24].addAvg =
> PFX(addAvg_32x24_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_32x32].addAvg =
> PFX(addAvg_32x32_avx2);
>
> -        p.chroma[X265_CSP_I422].pu[CHROMA_422_8x8].addAvg =
> PFX(addAvg_8x8_avx2);
> -        p.chroma[X265_CSP_I422].pu[CHROMA_422_8x12].addAvg =
> PFX(addAvg_8x12_avx2);
> -        p.chroma[X265_CSP_I422].pu[CHROMA_422_8x16].addAvg =
> PFX(addAvg_8x16_avx2);
> -        p.chroma[X265_CSP_I422].pu[CHROMA_422_8x32].addAvg =
> PFX(addAvg_8x32_avx2);
> -        p.chroma[X265_CSP_I422].pu[CHROMA_422_8x64].addAvg =
> PFX(addAvg_8x64_avx2);
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_16x8].addAvg =
> PFX(addAvg_16x8_avx2);
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_16x16].addAvg =
> PFX(addAvg_16x16_avx2);
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_16x24].addAvg =
> PFX(addAvg_16x24_avx2);
> @@ -2875,7 +2833,6 @@
>          p.pu[LUMA_32x24].pixelavg_pp = PFX(pixel_avg_32x24_avx2);
>          p.pu[LUMA_32x16].pixelavg_pp = PFX(pixel_avg_32x16_avx2);
>          p.pu[LUMA_32x8].pixelavg_pp = PFX(pixel_avg_32x8_avx2);
> -
>          p.pu[LUMA_64x64].pixelavg_pp = PFX(pixel_avg_64x64_avx2);
>          p.pu[LUMA_64x48].pixelavg_pp = PFX(pixel_avg_64x48_avx2);
>          p.pu[LUMA_64x32].pixelavg_pp = PFX(pixel_avg_64x32_avx2);
> @@ -2885,12 +2842,10 @@
>          p.pu[LUMA_16x8].satd  = PFX(pixel_satd_16x8_avx2);
>          p.pu[LUMA_8x16].satd  = PFX(pixel_satd_8x16_avx2);
>          p.pu[LUMA_8x8].satd   = PFX(pixel_satd_8x8_avx2);
> -
>          p.pu[LUMA_16x4].satd  = PFX(pixel_satd_16x4_avx2);
>          p.pu[LUMA_16x12].satd = PFX(pixel_satd_16x12_avx2);
>          p.pu[LUMA_16x32].satd = PFX(pixel_satd_16x32_avx2);
>          p.pu[LUMA_16x64].satd = PFX(pixel_satd_16x64_avx2);
> -
>          p.pu[LUMA_32x8].satd   = PFX(pixel_satd_32x8_avx2);
>          p.pu[LUMA_32x16].satd   = PFX(pixel_satd_32x16_avx2);
>          p.pu[LUMA_32x24].satd   = PFX(pixel_satd_32x24_avx2);
> @@ -2913,9 +2868,6 @@
>          p.pu[LUMA_64x48].sad = PFX(pixel_sad_64x48_avx2);
>          p.pu[LUMA_64x64].sad = PFX(pixel_sad_64x64_avx2);
>
> -        p.pu[LUMA_8x8].sad_x3 = PFX(pixel_sad_x3_8x8_avx2);
> -        p.pu[LUMA_16x8].sad_x4  = PFX(pixel_sad_x4_16x8_avx2);
> -        p.pu[LUMA_16x16].sad_x4 = PFX(pixel_sad_x4_16x16_avx2);
>          p.pu[LUMA_16x32].sad_x4 = PFX(pixel_sad_x4_16x32_avx2);
>          p.pu[LUMA_32x32].sad_x4 = PFX(pixel_sad_x4_32x32_avx2);
>          p.pu[LUMA_32x16].sad_x4 = PFX(pixel_sad_x4_32x16_avx2);
> @@ -2939,7 +2891,6 @@
>          p.cu[BLOCK_16x16].blockfill_s = PFX(blockfill_s_16x16_avx2);
>          p.cu[BLOCK_32x32].blockfill_s = PFX(blockfill_s_32x32_avx2);
>
> -        p.cu[BLOCK_8x8].cpy1Dto2D_shl = PFX(cpy1Dto2D_shl_8_avx2);
>          p.cu[BLOCK_16x16].cpy1Dto2D_shl = PFX(cpy1Dto2D_shl_16_avx2);
>          p.cu[BLOCK_32x32].cpy1Dto2D_shl = PFX(cpy1Dto2D_shl_32_avx2);
>
> @@ -3020,11 +2971,9 @@
>          p.cu[BLOCK_8x8].intra_pred[24] = PFX(intra_pred_ang8_24_avx2);
>          p.cu[BLOCK_8x8].intra_pred[11] = PFX(intra_pred_ang8_11_avx2);
>          p.cu[BLOCK_8x8].intra_pred[13] = PFX(intra_pred_ang8_13_avx2);
> -        p.cu[BLOCK_8x8].intra_pred[21] = PFX(intra_pred_ang8_21_avx2);
>          p.cu[BLOCK_8x8].intra_pred[22] = PFX(intra_pred_ang8_22_avx2);
>          p.cu[BLOCK_8x8].intra_pred[23] = PFX(intra_pred_ang8_23_avx2);
>          p.cu[BLOCK_8x8].intra_pred[14] = PFX(intra_pred_ang8_14_avx2);
> -        p.cu[BLOCK_8x8].intra_pred[15] = PFX(intra_pred_ang8_15_avx2);
>          p.cu[BLOCK_16x16].intra_pred[3] = PFX(intra_pred_ang16_3_avx2);
>          p.cu[BLOCK_16x16].intra_pred[4] = PFX(intra_pred_ang16_4_avx2);
>          p.cu[BLOCK_16x16].intra_pred[5] = PFX(intra_pred_ang16_5_avx2);
> @@ -3044,7 +2993,6 @@
>          p.cu[BLOCK_16x16].intra_pred[33] = PFX(intra_pred_ang16_33_avx2);
>          p.cu[BLOCK_16x16].intra_pred[24] = PFX(intra_pred_ang16_24_avx2);
>          p.cu[BLOCK_16x16].intra_pred[23] = PFX(intra_pred_ang16_23_avx2);
> -        p.cu[BLOCK_16x16].intra_pred[22] = PFX(intra_pred_ang16_22_avx2);
>          p.cu[BLOCK_32x32].intra_pred[4]  = PFX(intra_pred_ang32_4_avx2);
>          p.cu[BLOCK_32x32].intra_pred[5]  = PFX(intra_pred_ang32_5_avx2);
>          p.cu[BLOCK_32x32].intra_pred[6]  = PFX(intra_pred_ang32_6_avx2);
> @@ -3132,7 +3080,6 @@
>          p.pu[LUMA_16x4].luma_hps = PFX(interp_8tap_horiz_ps_16x4_avx2);
>          p.pu[LUMA_16x32].luma_hps = PFX(interp_8tap_horiz_ps_16x32_avx2);
>          p.pu[LUMA_16x64].luma_hps = PFX(interp_8tap_horiz_ps_16x64_avx2);
> -
>          p.pu[LUMA_32x32].luma_hps = PFX(interp_8tap_horiz_ps_32x32_avx2);
>          p.pu[LUMA_32x16].luma_hps = PFX(interp_8tap_horiz_ps_32x16_avx2);
>          p.pu[LUMA_32x24].luma_hps = PFX(interp_8tap_horiz_ps_32x24_avx2);
> @@ -3147,54 +3094,42 @@
>          p.pu[LUMA_24x32].luma_hps = PFX(interp_8tap_horiz_ps_24x32_avx2);
>
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_8x8].filter_hpp =
> PFX(interp_4tap_horiz_pp_8x8_avx2);
> -        p.chroma[X265_CSP_I420].pu[CHROMA_420_4x4].filter_hpp =
> PFX(interp_4tap_horiz_pp_4x4_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_32x32].filter_hpp =
> PFX(interp_4tap_horiz_pp_32x32_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_16x16].filter_hpp =
> PFX(interp_4tap_horiz_pp_16x16_avx2);
> -
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_2x4].filter_hpp =
> PFX(interp_4tap_horiz_pp_2x4_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_2x8].filter_hpp =
> PFX(interp_4tap_horiz_pp_2x8_avx2);
> -
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_4x2].filter_hpp =
> PFX(interp_4tap_horiz_pp_4x2_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_16x4].filter_hpp =
> PFX(interp_4tap_horiz_pp_16x4_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_16x8].filter_hpp =
> PFX(interp_4tap_horiz_pp_16x8_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_16x12].filter_hpp =
> PFX(interp_4tap_horiz_pp_16x12_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_16x32].filter_hpp =
> PFX(interp_4tap_horiz_pp_16x32_avx2);
> -
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_6x8].filter_hpp =
> PFX(interp_4tap_horiz_pp_6x8_avx2);
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_6x16].filter_hpp =
> PFX(interp_4tap_horiz_pp_6x16_avx2);
> -
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_32x16].filter_hpp =
> PFX(interp_4tap_horiz_pp_32x16_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_32x24].filter_hpp =
> PFX(interp_4tap_horiz_pp_32x24_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_32x8].filter_hpp =
> PFX(interp_4tap_horiz_pp_32x8_avx2);
> -
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_8x2].filter_hpp =
> PFX(interp_4tap_horiz_pp_8x2_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_8x4].filter_hpp =
> PFX(interp_4tap_horiz_pp_8x4_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_8x6].filter_hpp =
> PFX(interp_4tap_horiz_pp_8x6_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_8x16].filter_hpp =
> PFX(interp_4tap_horiz_pp_8x16_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_8x32].filter_hpp =
> PFX(interp_4tap_horiz_pp_8x32_avx2);
> -
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_12x16].filter_hpp =
> PFX(interp_4tap_horiz_pp_12x16_avx2);
> -
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_32x32].filter_hps =
> PFX(interp_4tap_horiz_ps_32x32_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_16x16].filter_hps =
> PFX(interp_4tap_horiz_ps_16x16_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_4x4].filter_hps =
> PFX(interp_4tap_horiz_ps_4x4_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_8x8].filter_hps =
> PFX(interp_4tap_horiz_ps_8x8_avx2);
> -
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_4x2].filter_hps =
> PFX(interp_4tap_horiz_ps_4x2_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_4x8].filter_hps =
> PFX(interp_4tap_horiz_ps_4x8_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_4x16].filter_hps =
> PFX(interp_4tap_horiz_ps_4x16_avx2);
> -
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_8x2].filter_hps =
> PFX(interp_4tap_horiz_ps_8x2_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_8x4].filter_hps =
> PFX(interp_4tap_horiz_ps_8x4_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_8x6].filter_hps =
> PFX(interp_4tap_horiz_ps_8x6_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_8x32].filter_hps =
> PFX(interp_4tap_horiz_ps_8x32_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_8x16].filter_hps =
> PFX(interp_4tap_horiz_ps_8x16_avx2);
> -
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_16x32].filter_hps =
> PFX(interp_4tap_horiz_ps_16x32_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_16x12].filter_hps =
> PFX(interp_4tap_horiz_ps_16x12_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_16x8].filter_hps =
> PFX(interp_4tap_horiz_ps_16x8_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_16x4].filter_hps =
> PFX(interp_4tap_horiz_ps_16x4_avx2);
> -
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_24x32].filter_hps =
> PFX(interp_4tap_horiz_ps_24x32_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_32x16].filter_hps =
> PFX(interp_4tap_horiz_ps_32x16_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_32x24].filter_hps =
> PFX(interp_4tap_horiz_ps_32x24_avx2);
> @@ -3205,24 +3140,18 @@
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_24x32].filter_hpp =
> PFX(interp_4tap_horiz_pp_24x32_avx2);
>
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_4x4].filter_vpp =
> PFX(interp_4tap_vert_pp_4x4_avx2);
> -
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_4x4].filter_vpp =
> PFX(interp_4tap_vert_pp_4x4_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_4x16].filter_vpp =
> PFX(interp_4tap_vert_pp_4x16_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_8x8].filter_vpp =
> PFX(interp_4tap_vert_pp_8x8_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_2x4].filter_vpp =
> PFX(interp_4tap_vert_pp_2x4_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_2x8].filter_vpp =
> PFX(interp_4tap_vert_pp_2x8_avx2);
> -        p.chroma[X265_CSP_I420].pu[CHROMA_420_4x2].filter_vpp =
> PFX(interp_4tap_vert_pp_4x2_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_4x8].filter_vpp =
> PFX(interp_4tap_vert_pp_4x8_avx2);
> -        p.chroma[X265_CSP_I420].pu[CHROMA_420_6x8].filter_vpp =
> PFX(interp_4tap_vert_pp_6x8_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_8x2].filter_vpp =
> PFX(interp_4tap_vert_pp_8x2_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_8x4].filter_vpp =
> PFX(interp_4tap_vert_pp_8x4_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_8x6].filter_vpp =
> PFX(interp_4tap_vert_pp_8x6_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_8x16].filter_vpp =
> PFX(interp_4tap_vert_pp_8x16_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_8x32].filter_vpp =
> PFX(interp_4tap_vert_pp_8x32_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_12x16].filter_vpp =
> PFX(interp_4tap_vert_pp_12x16_avx2);
> -        p.chroma[X265_CSP_I420].pu[CHROMA_420_16x12].filter_vpp =
> PFX(interp_4tap_vert_pp_16x12_avx2);
> -        p.chroma[X265_CSP_I420].pu[CHROMA_420_16x32].filter_vpp =
> PFX(interp_4tap_vert_pp_16x32_avx2);
> -        p.chroma[X265_CSP_I420].pu[CHROMA_420_24x32].filter_vpp =
> PFX(interp_4tap_vert_pp_24x32_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_32x8].filter_vpp =
> PFX(interp_4tap_vert_pp_32x8_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_32x16].filter_vpp =
> PFX(interp_4tap_vert_pp_32x16_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_32x24].filter_vpp =
> PFX(interp_4tap_vert_pp_32x24_avx2);
> @@ -3236,8 +3165,6 @@
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_8x4].filter_vps =
> PFX(interp_4tap_vert_ps_8x4_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_8x6].filter_vps =
> PFX(interp_4tap_vert_ps_8x6_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_8x8].filter_vps =
> PFX(interp_4tap_vert_ps_8x8_avx2);
> -        p.chroma[X265_CSP_I420].pu[CHROMA_420_8x16].filter_vps =
> PFX(interp_4tap_vert_ps_8x16_avx2);
> -        p.chroma[X265_CSP_I420].pu[CHROMA_420_8x32].filter_vps =
> PFX(interp_4tap_vert_ps_8x32_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_12x16].filter_vps =
> PFX(interp_4tap_vert_ps_12x16_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_16x4].filter_vps =
> PFX(interp_4tap_vert_ps_16x4_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_16x8].filter_vps =
> PFX(interp_4tap_vert_ps_16x8_avx2);
> @@ -3245,7 +3172,6 @@
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_4x16].filter_vps =
> PFX(interp_4tap_vert_ps_4x16_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_16x16].filter_vps =
> PFX(interp_4tap_vert_ps_16x16_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_16x32].filter_vps =
> PFX(interp_4tap_vert_ps_16x32_avx2);
> -        p.chroma[X265_CSP_I420].pu[CHROMA_420_24x32].filter_vps =
> PFX(interp_4tap_vert_ps_24x32_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_32x32].filter_vps =
> PFX(interp_4tap_vert_ps_32x32_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_32x24].filter_vps =
> PFX(interp_4tap_vert_ps_32x24_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_32x16].filter_vps =
> PFX(interp_4tap_vert_ps_32x16_avx2);
> @@ -3299,7 +3225,6 @@
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_24x32].filter_vss =
> PFX(interp_4tap_vert_ss_24x32_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_32x8].filter_vss =
> PFX(interp_4tap_vert_ss_32x8_avx2);
>          p.chroma[X265_CSP_I420].pu[CHROMA_420_32x16].filter_vss =
> PFX(interp_4tap_vert_ss_32x16_avx2);
> -        p.chroma[X265_CSP_I420].pu[CHROMA_420_32x24].filter_vss =
> PFX(interp_4tap_vert_ss_32x24_avx2);
>
>          //i422 for chroma_vss
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_4x8].filter_vss =
> PFX(interp_4tap_vert_ss_4x8_avx2);
> @@ -3314,11 +3239,7 @@
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_32x32].filter_vss =
> PFX(interp_4tap_vert_ss_32x32_avx2);
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_8x4].filter_vss =
> PFX(interp_4tap_vert_ss_8x4_avx2);
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_32x16].filter_vss =
> PFX(interp_4tap_vert_ss_32x16_avx2);
> -        p.chroma[X265_CSP_I422].pu[CHROMA_422_32x64].filter_vss =
> PFX(interp_4tap_vert_ss_32x64_avx2);
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_16x64].filter_vss =
> PFX(interp_4tap_vert_ss_16x64_avx2);
> -        p.chroma[X265_CSP_I422].pu[CHROMA_422_24x64].filter_vss =
> PFX(interp_4tap_vert_ss_24x64_avx2);
> -        p.chroma[X265_CSP_I422].pu[CHROMA_422_8x64].filter_vss =
> PFX(interp_4tap_vert_ss_8x64_avx2);
> -        p.chroma[X265_CSP_I422].pu[CHROMA_422_32x48].filter_vss =
> PFX(interp_4tap_vert_ss_32x48_avx2);
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_8x12].filter_vss =
> PFX(interp_4tap_vert_ss_8x12_avx2);
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_6x16].filter_vss =
> PFX(interp_4tap_vert_ss_6x16_avx2);
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_2x16].filter_vss =
> PFX(interp_4tap_vert_ss_2x16_avx2);
> @@ -3332,7 +3253,6 @@
>          p.chroma[X265_CSP_I444].pu[LUMA_8x8].filter_vss =
> PFX(interp_4tap_vert_ss_8x8_avx2);
>          p.chroma[X265_CSP_I444].pu[LUMA_16x16].filter_vss =
> PFX(interp_4tap_vert_ss_16x16_avx2);
>          p.chroma[X265_CSP_I444].pu[LUMA_32x32].filter_vss =
> PFX(interp_4tap_vert_ss_32x32_avx2);
> -        p.chroma[X265_CSP_I444].pu[LUMA_64x64].filter_vss =
> PFX(interp_4tap_vert_ss_64x64_avx2);
>          p.chroma[X265_CSP_I444].pu[LUMA_8x4].filter_vss =
> PFX(interp_4tap_vert_ss_8x4_avx2);
>          p.chroma[X265_CSP_I444].pu[LUMA_4x8].filter_vss =
> PFX(interp_4tap_vert_ss_4x8_avx2);
>          p.chroma[X265_CSP_I444].pu[LUMA_16x8].filter_vss =
> PFX(interp_4tap_vert_ss_16x8_avx2);
> @@ -3343,14 +3263,10 @@
>          p.chroma[X265_CSP_I444].pu[LUMA_12x16].filter_vss =
> PFX(interp_4tap_vert_ss_12x16_avx2);
>          p.chroma[X265_CSP_I444].pu[LUMA_16x4].filter_vss =
> PFX(interp_4tap_vert_ss_16x4_avx2);
>          p.chroma[X265_CSP_I444].pu[LUMA_4x16].filter_vss =
> PFX(interp_4tap_vert_ss_4x16_avx2);
> -        p.chroma[X265_CSP_I444].pu[LUMA_32x24].filter_vss =
> PFX(interp_4tap_vert_ss_32x24_avx2);
>          p.chroma[X265_CSP_I444].pu[LUMA_24x32].filter_vss =
> PFX(interp_4tap_vert_ss_24x32_avx2);
>          p.chroma[X265_CSP_I444].pu[LUMA_32x8].filter_vss =
> PFX(interp_4tap_vert_ss_32x8_avx2);
>          p.chroma[X265_CSP_I444].pu[LUMA_8x32].filter_vss =
> PFX(interp_4tap_vert_ss_8x32_avx2);
>          p.chroma[X265_CSP_I444].pu[LUMA_64x32].filter_vss =
> PFX(interp_4tap_vert_ss_64x32_avx2);
> -        p.chroma[X265_CSP_I444].pu[LUMA_32x64].filter_vss =
> PFX(interp_4tap_vert_ss_32x64_avx2);
> -        p.chroma[X265_CSP_I444].pu[LUMA_64x48].filter_vss =
> PFX(interp_4tap_vert_ss_64x48_avx2);
> -        p.chroma[X265_CSP_I444].pu[LUMA_48x64].filter_vss =
> PFX(interp_4tap_vert_ss_48x64_avx2);
>          p.chroma[X265_CSP_I444].pu[LUMA_64x16].filter_vss =
> PFX(interp_4tap_vert_ss_64x16_avx2);
>          p.chroma[X265_CSP_I444].pu[LUMA_16x64].filter_vss =
> PFX(interp_4tap_vert_ss_16x64_avx2);
>
> @@ -3386,17 +3302,15 @@
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_12x32].filter_hpp =
> PFX(interp_4tap_horiz_pp_12x32_avx2);
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_24x64].filter_hpp =
> PFX(interp_4tap_horiz_pp_24x64_avx2);
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_2x16].filter_hpp =
> PFX(interp_4tap_horiz_pp_2x16_avx2);
> -
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_2x16].filter_hpp =
> PFX(interp_4tap_horiz_pp_2x16_avx2);
>
> -        p.chroma[X265_CSP_I422].pu[CHROMA_422_4x4].filter_hpp =
> PFX(interp_4tap_horiz_pp_4x4_avx2);
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_8x4].filter_hpp =
> PFX(interp_4tap_horiz_pp_8x4_avx2);
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_8x8].filter_hpp =
> PFX(interp_4tap_horiz_pp_8x8_avx2);
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_8x16].filter_hpp =
> PFX(interp_4tap_horiz_pp_8x16_avx2);
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_8x32].filter_hpp =
> PFX(interp_4tap_horiz_pp_8x32_avx2);
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_8x64].filter_hpp =
> PFX(interp_4tap_horiz_pp_8x64_avx2);
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_8x12].filter_hpp =
> PFX(interp_4tap_horiz_pp_8x12_avx2);
> -
> +
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_16x8].filter_hpp =
> PFX(interp_4tap_horiz_pp_16x8_avx2);
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_16x16].filter_hpp =
> PFX(interp_4tap_horiz_pp_16x16_avx2);
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_16x32].filter_hpp =
> PFX(interp_4tap_horiz_pp_16x32_avx2);
> @@ -3411,7 +3325,6 @@
>
>          //i444 filters hpp
>
> -        p.chroma[X265_CSP_I444].pu[LUMA_4x4].filter_hpp =
> PFX(interp_4tap_horiz_pp_4x4_avx2);
>          p.chroma[X265_CSP_I444].pu[LUMA_8x8].filter_hpp =
> PFX(interp_4tap_horiz_pp_8x8_avx2);
>          p.chroma[X265_CSP_I444].pu[LUMA_16x16].filter_hpp =
> PFX(interp_4tap_horiz_pp_16x16_avx2);
>          p.chroma[X265_CSP_I444].pu[LUMA_32x32].filter_hpp =
> PFX(interp_4tap_horiz_pp_32x32_avx2);
> @@ -3553,14 +3466,12 @@
>
>          //i422 for chroma_vps
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_4x8].filter_vps =
> PFX(interp_4tap_vert_ps_4x8_avx2);
> -        p.chroma[X265_CSP_I422].pu[CHROMA_422_8x16].filter_vps =
> PFX(interp_4tap_vert_ps_8x16_avx2);
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_16x32].filter_vps =
> PFX(interp_4tap_vert_ps_16x32_avx2);
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_4x4].filter_vps =
> PFX(interp_4tap_vert_ps_4x4_avx2);
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_2x8].filter_vps =
> PFX(interp_4tap_vert_ps_2x8_avx2);
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_8x8].filter_vps =
> PFX(interp_4tap_vert_ps_8x8_avx2);
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_4x16].filter_vps =
> PFX(interp_4tap_vert_ps_4x16_avx2);
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_16x16].filter_vps =
> PFX(interp_4tap_vert_ps_16x16_avx2);
> -        p.chroma[X265_CSP_I422].pu[CHROMA_422_8x32].filter_vps =
> PFX(interp_4tap_vert_ps_8x32_avx2);
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_32x32].filter_vps =
> PFX(interp_4tap_vert_ps_32x32_avx2);
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_8x4].filter_vps =
> PFX(interp_4tap_vert_ps_8x4_avx2);
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_16x8].filter_vps =
> PFX(interp_4tap_vert_ps_16x8_avx2);
> @@ -3585,7 +3496,6 @@
>          p.chroma[X265_CSP_I444].pu[LUMA_8x4].filter_vps =
> PFX(interp_4tap_vert_ps_8x4_avx2);
>          p.chroma[X265_CSP_I444].pu[LUMA_4x8].filter_vps =
> PFX(interp_4tap_vert_ps_4x8_avx2);
>          p.chroma[X265_CSP_I444].pu[LUMA_16x8].filter_vps =
> PFX(interp_4tap_vert_ps_16x8_avx2);
> -        p.chroma[X265_CSP_I444].pu[LUMA_8x16].filter_vps =
> PFX(interp_4tap_vert_ps_8x16_avx2);
>          p.chroma[X265_CSP_I444].pu[LUMA_32x16].filter_vps =
> PFX(interp_4tap_vert_ps_32x16_avx2);
>          p.chroma[X265_CSP_I444].pu[LUMA_16x32].filter_vps =
> PFX(interp_4tap_vert_ps_16x32_avx2);
>          p.chroma[X265_CSP_I444].pu[LUMA_16x12].filter_vps =
> PFX(interp_4tap_vert_ps_16x12_avx2);
> @@ -3593,9 +3503,7 @@
>          p.chroma[X265_CSP_I444].pu[LUMA_16x4].filter_vps =
> PFX(interp_4tap_vert_ps_16x4_avx2);
>          p.chroma[X265_CSP_I444].pu[LUMA_4x16].filter_vps =
> PFX(interp_4tap_vert_ps_4x16_avx2);
>          p.chroma[X265_CSP_I444].pu[LUMA_32x24].filter_vps =
> PFX(interp_4tap_vert_ps_32x24_avx2);
> -        p.chroma[X265_CSP_I444].pu[LUMA_24x32].filter_vps =
> PFX(interp_4tap_vert_ps_24x32_avx2);
>          p.chroma[X265_CSP_I444].pu[LUMA_32x8].filter_vps =
> PFX(interp_4tap_vert_ps_32x8_avx2);
> -        p.chroma[X265_CSP_I444].pu[LUMA_8x32].filter_vps =
> PFX(interp_4tap_vert_ps_8x32_avx2);
>          p.chroma[X265_CSP_I444].pu[LUMA_16x64].filter_vps =
> PFX(interp_4tap_vert_ps_16x64_avx2);
>          p.chroma[X265_CSP_I444].pu[LUMA_32x64].filter_vps =
> PFX(interp_4tap_vert_ps_32x64_avx2);
>          p.chroma[X265_CSP_I444].pu[LUMA_48x64].filter_vps =
> PFX(interp_4tap_vert_ps_48x64_avx2);
> @@ -3607,7 +3515,6 @@
>          //i422 for chroma_vpp
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_4x8].filter_vpp =
> PFX(interp_4tap_vert_pp_4x8_avx2);
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_8x16].filter_vpp =
> PFX(interp_4tap_vert_pp_8x16_avx2);
> -        p.chroma[X265_CSP_I422].pu[CHROMA_422_16x32].filter_vpp =
> PFX(interp_4tap_vert_pp_16x32_avx2);
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_4x4].filter_vpp =
> PFX(interp_4tap_vert_pp_4x4_avx2);
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_2x8].filter_vpp =
> PFX(interp_4tap_vert_pp_2x8_avx2);
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_8x8].filter_vpp =
> PFX(interp_4tap_vert_pp_8x8_avx2);
> @@ -3616,7 +3523,6 @@
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_32x32].filter_vpp =
> PFX(interp_4tap_vert_pp_32x32_avx2);
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_8x4].filter_vpp =
> PFX(interp_4tap_vert_pp_8x4_avx2);
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_32x16].filter_vpp =
> PFX(interp_4tap_vert_pp_32x16_avx2);
> -        p.chroma[X265_CSP_I422].pu[CHROMA_422_16x64].filter_vpp =
> PFX(interp_4tap_vert_pp_16x64_avx2);
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_8x64].filter_vpp =
> PFX(interp_4tap_vert_pp_8x64_avx2);
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_32x64].filter_vpp =
> PFX(interp_4tap_vert_pp_32x64_avx2);
>          p.chroma[X265_CSP_I422].pu[CHROMA_422_32x48].filter_vpp =
> PFX(interp_4tap_vert_pp_32x48_avx2);
> @@ -3635,15 +3541,11 @@
>          p.chroma[X265_CSP_I444].pu[LUMA_4x8].filter_vpp =
> PFX(interp_4tap_vert_pp_4x8_avx2);
>          p.chroma[X265_CSP_I444].pu[LUMA_8x16].filter_vpp =
> PFX(interp_4tap_vert_pp_8x16_avx2);
>          p.chroma[X265_CSP_I444].pu[LUMA_32x16].filter_vpp =
> PFX(interp_4tap_vert_pp_32x16_avx2);
> -        p.chroma[X265_CSP_I444].pu[LUMA_16x32].filter_vpp =
> PFX(interp_4tap_vert_pp_16x32_avx2);
> -        p.chroma[X265_CSP_I444].pu[LUMA_16x12].filter_vpp =
> PFX(interp_4tap_vert_pp_16x12_avx2);
>          p.chroma[X265_CSP_I444].pu[LUMA_12x16].filter_vpp =
> PFX(interp_4tap_vert_pp_12x16_avx2);
>          p.chroma[X265_CSP_I444].pu[LUMA_4x16].filter_vpp =
> PFX(interp_4tap_vert_pp_4x16_avx2);
>          p.chroma[X265_CSP_I444].pu[LUMA_32x24].filter_vpp =
> PFX(interp_4tap_vert_pp_32x24_avx2);
> -        p.chroma[X265_CSP_I444].pu[LUMA_24x32].filter_vpp =
> PFX(interp_4tap_vert_pp_24x32_avx2);
>          p.chroma[X265_CSP_I444].pu[LUMA_32x8].filter_vpp =
> PFX(interp_4tap_vert_pp_32x8_avx2);
>          p.chroma[X265_CSP_I444].pu[LUMA_8x32].filter_vpp =
> PFX(interp_4tap_vert_pp_8x32_avx2);
> -        p.chroma[X265_CSP_I444].pu[LUMA_16x64].filter_vpp =
> PFX(interp_4tap_vert_pp_16x64_avx2);
>          p.chroma[X265_CSP_I444].pu[LUMA_32x64].filter_vpp =
> PFX(interp_4tap_vert_pp_32x64_avx2);
>          p.chroma[X265_CSP_I444].pu[LUMA_48x64].filter_vpp =
> PFX(interp_4tap_vert_pp_48x64_avx2);
>          p.chroma[X265_CSP_I444].pu[LUMA_64x64].filter_vpp =
> PFX(interp_4tap_vert_pp_64x64_avx2);
> @@ -3708,6 +3610,64 @@
>          p.chroma[X265_CSP_I444].pu[LUMA_16x16].filter_vpp =
> PFX(interp_4tap_vert_pp_16x16_avx2);
>          p.chroma[X265_CSP_I444].pu[LUMA_16x8].filter_vpp =
> PFX(interp_4tap_vert_pp_16x8_avx2);
>          p.chroma[X265_CSP_I444].pu[LUMA_16x4].filter_vpp =
> PFX(interp_4tap_vert_pp_16x4_avx2);
> +
> +        p.chroma[X265_CSP_I420].pu[CHROMA_420_4x4].filter_hpp =
> PFX(interp_4tap_horiz_pp_4x4_avx2);
> +        p.chroma[X265_CSP_I444].pu[LUMA_4x4].filter_hpp =
> PFX(interp_4tap_horiz_pp_4x4_avx2);
> +        p.chroma[X265_CSP_I422].pu[CHROMA_422_4x4].filter_hpp =
> PFX(interp_4tap_horiz_pp_4x4_avx2);
> +
> +        p.chroma[X265_CSP_I420].pu[CHROMA_420_24x32].filter_vpp =
> PFX(interp_4tap_vert_pp_24x32_avx2);
> +        p.chroma[X265_CSP_I444].pu[LUMA_24x32].filter_vpp =
> PFX(interp_4tap_vert_pp_24x32_avx2);
> +        p.chroma[X265_CSP_I420].pu[CHROMA_420_6x8].filter_vpp =
> PFX(interp_4tap_vert_pp_6x8_avx2);
> +        p.chroma[X265_CSP_I420].pu[CHROMA_420_16x32].filter_vpp =
> PFX(interp_4tap_vert_pp_16x32_avx2);
> +        p.chroma[X265_CSP_I444].pu[LUMA_16x32].filter_vpp =
> PFX(interp_4tap_vert_pp_16x32_avx2);
> +        p.chroma[X265_CSP_I422].pu[CHROMA_422_16x32].filter_vpp =
> PFX(interp_4tap_vert_pp_16x32_avx2);
> +        p.chroma[X265_CSP_I420].pu[CHROMA_420_16x12].filter_vpp =
> PFX(interp_4tap_vert_pp_16x12_avx2);
> +        p.chroma[X265_CSP_I444].pu[LUMA_16x12].filter_vpp =
> PFX(interp_4tap_vert_pp_16x12_avx2);
> +        p.chroma[X265_CSP_I422].pu[CHROMA_422_16x64].filter_vpp =
> PFX(interp_4tap_vert_pp_16x64_avx2);
> +        p.chroma[X265_CSP_I444].pu[LUMA_16x64].filter_vpp =
> PFX(interp_4tap_vert_pp_16x64_avx2);
> +        p.chroma[X265_CSP_I420].pu[CHROMA_420_4x2].filter_vpp =
> PFX(interp_4tap_vert_pp_4x2_avx2);
> +
> +        p.chroma[X265_CSP_I444].pu[LUMA_64x64].filter_vss =
> PFX(interp_4tap_vert_ss_64x64_avx2);
> +        p.chroma[X265_CSP_I422].pu[CHROMA_422_8x64].filter_vss =
> PFX(interp_4tap_vert_ss_8x64_avx2);
> +        p.chroma[X265_CSP_I444].pu[LUMA_64x48].filter_vss =
> PFX(interp_4tap_vert_ss_64x48_avx2);
> +        p.chroma[X265_CSP_I444].pu[LUMA_48x64].filter_vss =
> PFX(interp_4tap_vert_ss_48x64_avx2);
> +        p.chroma[X265_CSP_I422].pu[CHROMA_422_32x64].filter_vss =
> PFX(interp_4tap_vert_ss_32x64_avx2);
> +        p.chroma[X265_CSP_I444].pu[LUMA_32x64].filter_vss =
> PFX(interp_4tap_vert_ss_32x64_avx2);
> +        p.chroma[X265_CSP_I422].pu[CHROMA_422_32x48].filter_vss =
> PFX(interp_4tap_vert_ss_32x48_avx2);
> +        p.chroma[X265_CSP_I422].pu[CHROMA_422_24x64].filter_vss =
> PFX(interp_4tap_vert_ss_24x64_avx2);
> +        p.chroma[X265_CSP_I420].pu[CHROMA_420_32x24].filter_vss =
> PFX(interp_4tap_vert_ss_32x24_avx2);
> +        p.chroma[X265_CSP_I444].pu[LUMA_32x24].filter_vss =
> PFX(interp_4tap_vert_ss_32x24_avx2);
> +
> +        p.chroma[X265_CSP_I420].pu[CHROMA_420_8x32].filter_vps =
> PFX(interp_4tap_vert_ps_8x32_avx2);
> +        p.chroma[X265_CSP_I422].pu[CHROMA_422_8x32].filter_vps =
> PFX(interp_4tap_vert_ps_8x32_avx2);
> +        p.chroma[X265_CSP_I444].pu[LUMA_8x32].filter_vps =
> PFX(interp_4tap_vert_ps_8x32_avx2);
> +        p.chroma[X265_CSP_I420].pu[CHROMA_420_24x32].filter_vps =
> PFX(interp_4tap_vert_ps_24x32_avx2);
> +        p.chroma[X265_CSP_I444].pu[LUMA_24x32].filter_vps =
> PFX(interp_4tap_vert_ps_24x32_avx2);
> +        p.chroma[X265_CSP_I420].pu[CHROMA_420_8x16].filter_vps =
> PFX(interp_4tap_vert_ps_8x16_avx2);
> +        p.chroma[X265_CSP_I422].pu[CHROMA_422_8x16].filter_vps =
> PFX(interp_4tap_vert_ps_8x16_avx2);
> +        p.chroma[X265_CSP_I444].pu[LUMA_8x16].filter_vps =
> PFX(interp_4tap_vert_ps_8x16_avx2);
> +
> +        p.pu[LUMA_8x8].sad_x3 = PFX(pixel_sad_x3_8x8_avx2);
> +        p.pu[LUMA_16x8].sad_x4 = PFX(pixel_sad_x4_16x8_avx2);
> +        p.pu[LUMA_16x16].sad_x4 = PFX(pixel_sad_x4_16x16_avx2);
> +
> +        p.pu[LUMA_8x16].addAvg = PFX(addAvg_8x16_avx2);
> +        p.pu[LUMA_8x32].addAvg = PFX(addAvg_8x32_avx2);
> +        p.chroma[X265_CSP_I420].pu[CHROMA_420_8x8].addAvg =
> PFX(addAvg_8x8_avx2);
> +        p.chroma[X265_CSP_I420].pu[CHROMA_420_8x16].addAvg =
> PFX(addAvg_8x16_avx2);
> +        p.chroma[X265_CSP_I420].pu[CHROMA_420_8x32].addAvg =
> PFX(addAvg_8x32_avx2);
> +        p.chroma[X265_CSP_I422].pu[CHROMA_422_8x8].addAvg =
> PFX(addAvg_8x8_avx2);
> +        p.chroma[X265_CSP_I422].pu[CHROMA_422_8x12].addAvg =
> PFX(addAvg_8x12_avx2);
> +        p.chroma[X265_CSP_I422].pu[CHROMA_422_8x16].addAvg =
> PFX(addAvg_8x16_avx2);
> +        p.chroma[X265_CSP_I422].pu[CHROMA_422_8x32].addAvg =
> PFX(addAvg_8x32_avx2);
> +        p.chroma[X265_CSP_I422].pu[CHROMA_422_8x64].addAvg =
> PFX(addAvg_8x64_avx2);
> +
> +        p.cu[BLOCK_8x8].cpy1Dto2D_shl = PFX(cpy1Dto2D_shl_8_avx2);
> +        p.saoCuOrgE3[0] = PFX(saoCuOrgE3_avx2);
> +
> +        p.cu[BLOCK_16x16].intra_pred[22] = PFX(intra_pred_ang16_22_avx2);
> +        p.cu[BLOCK_8x8].intra_pred[21] = PFX(intra_pred_ang8_21_avx2);
> +        p.cu[BLOCK_8x8].intra_pred[15] = PFX(intra_pred_ang8_15_avx2);
>  #endif
>      }
>  #endif
> _______________________________________________
> x265-devel mailing list
> x265-devel at videolan.org
> https://mailman.videolan.org/listinfo/x265-devel
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.videolan.org/pipermail/x265-devel/attachments/20150820/72d67a73/attachment-0001.html>


More information about the x265-devel mailing list