[x265] [PATCH] asm: Re-enabling asm primitives
aasaipriya at multicorewareinc.com
aasaipriya at multicorewareinc.com
Fri Sep 11 08:06:50 CEST 2015
# HG changeset patch
# User Aasaipriya Chandran <aasaipriya at multicorewareinc.com>
# Date 1441951115 -19800
# Fri Sep 11 11:28:35 2015 +0530
# Branch stable
# Node ID eb9939c00cb1e87813bf246266e7c11f647d5009
# Parent 1900d52fae07f645be4e9f0bb76e26876be3e3bc
asm: Re-enabling asm primitives
Based on the performance comparison between
case 1: disabling primitives less than 3% and less than 10%
case 2: re - enabling all the primitives
case 3: disabling primitives less than 3% alone
using 2k videos ( BasketballDrive_1920x1080_50y4m, Kimono1_1920x1080_24y4m, ParkScene_1920x1080_24y4m ) and 4k videos (Coastguard_4ky4m, sintel_4k_600fy4m, Traffic_4096x2048_30py4m) on presets (ultrafast, medium , veryslow) ,
the performance gain seems to be flat or low in both case 1 and 3 compared to case 2. So its basically safe to enable all the primitives.
diff -r 1900d52fae07 -r eb9939c00cb1 source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp Mon Sep 07 11:47:54 2015 +0530
+++ b/source/common/x86/asm-primitives.cpp Fri Sep 11 11:28:35 2015 +0530
@@ -1185,6 +1185,7 @@
p.chroma[X265_CSP_I422].pu[CHROMA_422_8x32].satd = PFX(pixel_satd_8x32_avx);
p.chroma[X265_CSP_I422].pu[CHROMA_422_8x4].satd = PFX(pixel_satd_8x4_avx);
+ ALL_LUMA_PU(satd, pixel_satd, avx);
p.chroma[X265_CSP_I420].pu[CHROMA_420_8x8].satd = PFX(pixel_satd_8x8_avx);
p.chroma[X265_CSP_I420].pu[CHROMA_420_8x4].satd = PFX(pixel_satd_8x4_avx);
p.chroma[X265_CSP_I420].pu[CHROMA_420_8x16].satd = PFX(pixel_satd_8x16_avx);
@@ -1193,10 +1194,15 @@
p.chroma[X265_CSP_I420].pu[CHROMA_420_24x32].satd = PFX(pixel_satd_24x32_avx);
p.chroma[X265_CSP_I420].pu[CHROMA_420_4x16].satd = PFX(pixel_satd_4x16_avx);
p.chroma[X265_CSP_I420].pu[CHROMA_420_4x8].satd = PFX(pixel_satd_4x8_avx);
-
+#if X265_DEPTH <= 10
+ ASSIGN_SA8D(avx);
+#endif
p.chroma[X265_CSP_I420].cu[BLOCK_420_8x8].sa8d = PFX(pixel_sa8d_8x8_avx);
p.chroma[X265_CSP_I420].cu[BLOCK_420_16x16].sa8d = PFX(pixel_sa8d_16x16_avx);
p.chroma[X265_CSP_I420].cu[BLOCK_420_32x32].sa8d = PFX(pixel_sa8d_32x32_avx);
+ LUMA_VAR(avx);
+ p.ssim_4x4x2_core = PFX(pixel_ssim_4x4x2_core_avx);
+ p.ssim_end_4 = PFX(pixel_ssim_end4_avx);
// copy_pp primitives
// 16 x N
@@ -1293,20 +1299,6 @@
p.pu[LUMA_64x32].copy_pp = (copy_pp_t)PFX(blockcopy_ss_64x32_avx);
p.pu[LUMA_64x48].copy_pp = (copy_pp_t)PFX(blockcopy_ss_64x48_avx);
p.pu[LUMA_64x64].copy_pp = (copy_pp_t)PFX(blockcopy_ss_64x64_avx);
-
- /* The following primitives have been disabled since performance compared to SSE is negligible/negative */
-#if 0
- ALL_LUMA_PU(satd, pixel_satd, avx);
-
- p.ssim_4x4x2_core = PFX(pixel_ssim_4x4x2_core_avx);
- p.ssim_end_4 = PFX(pixel_ssim_end4_avx);
-
- LUMA_VAR(avx);
-
- #if X265_DEPTH <= 10
- ASSIGN_SA8D(avx);
- #endif
-#endif
}
if (cpuMask & X265_CPU_XOP)
{
@@ -1422,8 +1414,11 @@
p.pu[LUMA_64x64].pixelavg_pp = PFX(pixel_avg_64x64_avx2);
p.pu[LUMA_48x64].pixelavg_pp = PFX(pixel_avg_48x64_avx2);
+ p.pu[LUMA_8x4].addAvg = PFX(addAvg_8x4_avx2);
+ p.pu[LUMA_8x8].addAvg = PFX(addAvg_8x8_avx2);
p.pu[LUMA_8x16].addAvg = PFX(addAvg_8x16_avx2);
p.pu[LUMA_8x32].addAvg = PFX(addAvg_8x32_avx2);
+ p.pu[LUMA_12x16].addAvg = PFX(addAvg_12x16_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);
@@ -1443,10 +1438,12 @@
p.pu[LUMA_64x64].addAvg = PFX(addAvg_64x64_avx2);
p.chroma[X265_CSP_I420].pu[CHROMA_420_8x2].addAvg = PFX(addAvg_8x2_avx2);
+ p.chroma[X265_CSP_I420].pu[CHROMA_420_8x4].addAvg = PFX(addAvg_8x4_avx2);
p.chroma[X265_CSP_I420].pu[CHROMA_420_8x6].addAvg = PFX(addAvg_8x6_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_12x16].addAvg = PFX(addAvg_12x16_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);
@@ -1460,15 +1457,18 @@
p.chroma[X265_CSP_I422].pu[CHROMA_422_8x16].addAvg = PFX(addAvg_8x16_avx2);
p.chroma[X265_CSP_I422].pu[CHROMA_422_16x32].addAvg = PFX(addAvg_16x32_avx2);
p.chroma[X265_CSP_I422].pu[CHROMA_422_32x64].addAvg = PFX(addAvg_32x64_avx2);
+ p.chroma[X265_CSP_I422].pu[CHROMA_422_8x8].addAvg = PFX(addAvg_8x8_avx2);
p.chroma[X265_CSP_I422].pu[CHROMA_422_16x16].addAvg = PFX(addAvg_16x16_avx2);
p.chroma[X265_CSP_I422].pu[CHROMA_422_8x32].addAvg = PFX(addAvg_8x32_avx2);
p.chroma[X265_CSP_I422].pu[CHROMA_422_32x32].addAvg = PFX(addAvg_32x32_avx2);
p.chroma[X265_CSP_I422].pu[CHROMA_422_16x64].addAvg = PFX(addAvg_16x64_avx2);
p.chroma[X265_CSP_I422].pu[CHROMA_422_8x12].addAvg = PFX(addAvg_8x12_avx2);
+ p.chroma[X265_CSP_I422].pu[CHROMA_422_8x4].addAvg = PFX(addAvg_8x4_avx2);
p.chroma[X265_CSP_I422].pu[CHROMA_422_16x24].addAvg = PFX(addAvg_16x24_avx2);
p.chroma[X265_CSP_I422].pu[CHROMA_422_16x8].addAvg = PFX(addAvg_16x8_avx2);
p.chroma[X265_CSP_I422].pu[CHROMA_422_8x64].addAvg = PFX(addAvg_8x64_avx2);
p.chroma[X265_CSP_I422].pu[CHROMA_422_24x64].addAvg = PFX(addAvg_24x64_avx2);
+ p.chroma[X265_CSP_I422].pu[CHROMA_422_12x32].addAvg = PFX(addAvg_12x32_avx2);
p.chroma[X265_CSP_I422].pu[CHROMA_422_32x16].addAvg = PFX(addAvg_32x16_avx2);
p.chroma[X265_CSP_I422].pu[CHROMA_422_32x48].addAvg = PFX(addAvg_32x48_avx2);
@@ -1565,17 +1565,11 @@
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].count_nonzero = PFX(count_nonzero_8x8_avx2);
- p.cu[BLOCK_16x16].count_nonzero = PFX(count_nonzero_16x16_avx2);
- p.cu[BLOCK_32x32].count_nonzero = PFX(count_nonzero_32x32_avx2);
-
- p.cu[BLOCK_16x16].cpy1Dto2D_shl = PFX(cpy1Dto2D_shl_16_avx2);
- p.cu[BLOCK_32x32].cpy1Dto2D_shl = PFX(cpy1Dto2D_shl_32_avx2);
-
- p.cu[BLOCK_8x8].cpy1Dto2D_shr = PFX(cpy1Dto2D_shr_8_avx2);
- p.cu[BLOCK_16x16].cpy1Dto2D_shr = PFX(cpy1Dto2D_shr_16_avx2);
- p.cu[BLOCK_32x32].cpy1Dto2D_shr = PFX(cpy1Dto2D_shr_32_avx2);
-
+ ALL_LUMA_TU(count_nonzero, count_nonzero, avx2);
+ ALL_LUMA_TU_S(cpy1Dto2D_shl, cpy1Dto2D_shl_, avx2);
+ ALL_LUMA_TU_S(cpy1Dto2D_shr, cpy1Dto2D_shr_, avx2);
+
+ p.cu[BLOCK_8x8].copy_cnt = PFX(copy_cnt_8_avx2);
p.cu[BLOCK_16x16].copy_cnt = PFX(copy_cnt_16_avx2);
p.cu[BLOCK_32x32].copy_cnt = PFX(copy_cnt_32_avx2);
@@ -1772,29 +1766,48 @@
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_8x2].filter_hps = PFX(interp_4tap_horiz_ps_8x2_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_16x16].filter_hps = PFX(interp_4tap_horiz_ps_16x16_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_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_16x4].filter_hps = PFX(interp_4tap_horiz_ps_16x4_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_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);
p.chroma[X265_CSP_I420].pu[CHROMA_420_32x8].filter_hps = PFX(interp_4tap_horiz_ps_32x8_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_12x16].filter_hps = PFX(interp_4tap_horiz_ps_12x16_avx2);
+ p.chroma[X265_CSP_I420].pu[CHROMA_420_6x8].filter_hps = PFX(interp_4tap_horiz_ps_6x8_avx2);
p.chroma[X265_CSP_I422].pu[CHROMA_422_8x8].filter_hps = PFX(interp_4tap_horiz_ps_8x8_avx2);
p.chroma[X265_CSP_I422].pu[CHROMA_422_8x16].filter_hps = PFX(interp_4tap_horiz_ps_8x16_avx2);
p.chroma[X265_CSP_I422].pu[CHROMA_422_8x32].filter_hps = PFX(interp_4tap_horiz_ps_8x32_avx2);
p.chroma[X265_CSP_I422].pu[CHROMA_422_8x12].filter_hps = PFX(interp_4tap_horiz_ps_8x12_avx2);
+ p.chroma[X265_CSP_I422].pu[CHROMA_422_8x64].filter_hps = PFX(interp_4tap_horiz_ps_8x64_avx2);
p.chroma[X265_CSP_I422].pu[CHROMA_422_8x4].filter_hps = PFX(interp_4tap_horiz_ps_8x4_avx2);
+ p.chroma[X265_CSP_I422].pu[CHROMA_422_16x32].filter_hps = PFX(interp_4tap_horiz_ps_16x32_avx2);
+ p.chroma[X265_CSP_I422].pu[CHROMA_422_16x16].filter_hps = PFX(interp_4tap_horiz_ps_16x16_avx2);
p.chroma[X265_CSP_I422].pu[CHROMA_422_16x64].filter_hps = PFX(interp_4tap_horiz_ps_16x64_avx2);
p.chroma[X265_CSP_I422].pu[CHROMA_422_16x24].filter_hps = PFX(interp_4tap_horiz_ps_16x24_avx2);
p.chroma[X265_CSP_I422].pu[CHROMA_422_16x8].filter_hps = PFX(interp_4tap_horiz_ps_16x8_avx2);
p.chroma[X265_CSP_I422].pu[CHROMA_422_32x64].filter_hps = PFX(interp_4tap_horiz_ps_32x64_avx2);
+ p.chroma[X265_CSP_I422].pu[CHROMA_422_32x32].filter_hps = PFX(interp_4tap_horiz_ps_32x32_avx2);
+ p.chroma[X265_CSP_I422].pu[CHROMA_422_32x48].filter_hps = PFX(interp_4tap_horiz_ps_32x48_avx2);
+ p.chroma[X265_CSP_I422].pu[CHROMA_422_32x16].filter_hps = PFX(interp_4tap_horiz_ps_32x16_avx2);
+ p.chroma[X265_CSP_I422].pu[CHROMA_422_12x32].filter_hps = PFX(interp_4tap_horiz_ps_12x32_avx2);
p.chroma[X265_CSP_I422].pu[CHROMA_422_24x64].filter_hps = PFX(interp_4tap_horiz_ps_24x64_avx2);
+ p.chroma[X265_CSP_I422].pu[CHROMA_422_6x16].filter_hps = PFX(interp_4tap_horiz_ps_6x16_avx2);
p.chroma[X265_CSP_I444].pu[LUMA_8x8].filter_hps = PFX(interp_4tap_horiz_ps_8x8_avx2);
p.chroma[X265_CSP_I444].pu[LUMA_8x4].filter_hps = PFX(interp_4tap_horiz_ps_8x4_avx2);
p.chroma[X265_CSP_I444].pu[LUMA_8x16].filter_hps = PFX(interp_4tap_horiz_ps_8x16_avx2);
p.chroma[X265_CSP_I444].pu[LUMA_8x32].filter_hps = PFX(interp_4tap_horiz_ps_8x32_avx2);
+ p.chroma[X265_CSP_I444].pu[LUMA_16x16].filter_hps = PFX(interp_4tap_horiz_ps_16x16_avx2);
p.chroma[X265_CSP_I444].pu[LUMA_16x8].filter_hps = PFX(interp_4tap_horiz_ps_16x8_avx2);
+ p.chroma[X265_CSP_I444].pu[LUMA_16x32].filter_hps = PFX(interp_4tap_horiz_ps_16x32_avx2);
p.chroma[X265_CSP_I444].pu[LUMA_16x12].filter_hps = PFX(interp_4tap_horiz_ps_16x12_avx2);
p.chroma[X265_CSP_I444].pu[LUMA_16x4].filter_hps = PFX(interp_4tap_horiz_ps_16x4_avx2);
p.chroma[X265_CSP_I444].pu[LUMA_16x64].filter_hps = PFX(interp_4tap_horiz_ps_16x64_avx2);
+ p.chroma[X265_CSP_I444].pu[LUMA_32x32].filter_hps = PFX(interp_4tap_horiz_ps_32x32_avx2);
+ p.chroma[X265_CSP_I444].pu[LUMA_32x16].filter_hps = PFX(interp_4tap_horiz_ps_32x16_avx2);
p.chroma[X265_CSP_I444].pu[LUMA_32x64].filter_hps = PFX(interp_4tap_horiz_ps_32x64_avx2);
p.chroma[X265_CSP_I444].pu[LUMA_32x24].filter_hps = PFX(interp_4tap_horiz_ps_32x24_avx2);
p.chroma[X265_CSP_I444].pu[LUMA_32x8].filter_hps = PFX(interp_4tap_horiz_ps_32x8_avx2);
@@ -1803,6 +1816,8 @@
p.chroma[X265_CSP_I444].pu[LUMA_64x32].filter_hps = PFX(interp_4tap_horiz_ps_64x32_avx2);
p.chroma[X265_CSP_I444].pu[LUMA_64x16].filter_hps = PFX(interp_4tap_horiz_ps_64x16_avx2);
p.chroma[X265_CSP_I444].pu[LUMA_48x64].filter_hps = PFX(interp_4tap_horiz_ps_48x64_avx2);
+ p.chroma[X265_CSP_I444].pu[LUMA_24x32].filter_hps = PFX(interp_4tap_horiz_ps_24x32_avx2);
+ p.chroma[X265_CSP_I444].pu[LUMA_12x16].filter_hps = PFX(interp_4tap_horiz_ps_12x16_avx2);
p.chroma[X265_CSP_I420].pu[CHROMA_420_6x8].filter_hpp = PFX(interp_4tap_horiz_pp_6x8_avx2);
p.chroma[X265_CSP_I420].pu[CHROMA_420_8x2].filter_hpp = PFX(interp_4tap_horiz_pp_8x2_avx2);
@@ -1820,6 +1835,7 @@
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_32x32].filter_hpp = PFX(interp_4tap_horiz_pp_32x32_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_24x32].filter_hpp = PFX(interp_4tap_horiz_pp_24x32_avx2);
p.chroma[X265_CSP_I422].pu[CHROMA_422_6x16].filter_hpp = PFX(interp_4tap_horiz_pp_6x16_avx2);
p.chroma[X265_CSP_I422].pu[CHROMA_422_8x4].filter_hpp = PFX(interp_4tap_horiz_pp_8x4_avx2);
@@ -1837,6 +1853,7 @@
p.chroma[X265_CSP_I422].pu[CHROMA_422_32x32].filter_hpp = PFX(interp_4tap_horiz_pp_32x32_avx2);
p.chroma[X265_CSP_I422].pu[CHROMA_422_32x48].filter_hpp = PFX(interp_4tap_horiz_pp_32x48_avx2);
p.chroma[X265_CSP_I422].pu[CHROMA_422_32x64].filter_hpp = PFX(interp_4tap_horiz_pp_32x64_avx2);
+ 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_I444].pu[LUMA_8x4].filter_hpp = PFX(interp_4tap_horiz_pp_8x4_avx2);
p.chroma[X265_CSP_I444].pu[LUMA_8x8].filter_hpp = PFX(interp_4tap_horiz_pp_8x8_avx2);
@@ -1853,12 +1870,14 @@
p.chroma[X265_CSP_I444].pu[LUMA_32x24].filter_hpp = PFX(interp_4tap_horiz_pp_32x24_avx2);
p.chroma[X265_CSP_I444].pu[LUMA_32x32].filter_hpp = PFX(interp_4tap_horiz_pp_32x32_avx2);
p.chroma[X265_CSP_I444].pu[LUMA_32x64].filter_hpp = PFX(interp_4tap_horiz_pp_32x64_avx2);
+ p.chroma[X265_CSP_I444].pu[LUMA_12x16].filter_hpp = PFX(interp_4tap_horiz_pp_12x16_avx2);
p.chroma[X265_CSP_I444].pu[LUMA_24x32].filter_hpp = PFX(interp_4tap_horiz_pp_24x32_avx2);
p.chroma[X265_CSP_I444].pu[LUMA_64x16].filter_hpp = PFX(interp_4tap_horiz_pp_64x16_avx2);
p.chroma[X265_CSP_I444].pu[LUMA_64x32].filter_hpp = PFX(interp_4tap_horiz_pp_64x32_avx2);
p.chroma[X265_CSP_I444].pu[LUMA_64x48].filter_hpp = PFX(interp_4tap_horiz_pp_64x48_avx2);
p.chroma[X265_CSP_I444].pu[LUMA_64x64].filter_hpp = PFX(interp_4tap_horiz_pp_64x64_avx2);
p.chroma[X265_CSP_I444].pu[LUMA_48x64].filter_hpp = PFX(interp_4tap_horiz_pp_48x64_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_4x2].filter_vps = PFX(interp_4tap_vert_ps_4x2_avx2);
p.chroma[X265_CSP_I420].pu[CHROMA_420_4x2].filter_vsp = PFX(interp_4tap_vert_sp_4x2_avx2);
@@ -1939,6 +1958,7 @@
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_8x64].filter_vps = PFX(interp_4tap_vert_ps_8x64_avx2);
p.chroma[X265_CSP_I422].pu[CHROMA_422_8x64].filter_vsp = PFX(interp_4tap_vert_sp_8x64_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_4x4].filter_vpp = PFX(interp_4tap_vert_pp_4x4_avx2);
p.chroma[X265_CSP_I444].pu[LUMA_4x4].filter_vps = PFX(interp_4tap_vert_ps_4x4_avx2);
p.chroma[X265_CSP_I444].pu[LUMA_4x4].filter_vsp = PFX(interp_4tap_vert_sp_4x4_avx2);
@@ -1968,6 +1988,7 @@
p.chroma[X265_CSP_I444].pu[LUMA_8x32].filter_vsp = PFX(interp_4tap_vert_sp_8x32_avx2);
p.chroma[X265_CSP_I444].pu[LUMA_8x32].filter_vss = PFX(interp_4tap_vert_ss_8x32_avx2);
+
p.chroma[X265_CSP_I420].pu[CHROMA_420_6x8].filter_vss = PFX(interp_4tap_vert_ss_6x8_avx2);
p.chroma[X265_CSP_I420].pu[CHROMA_420_6x8].filter_vsp = PFX(interp_4tap_vert_sp_6x8_avx2);
p.chroma[X265_CSP_I420].pu[CHROMA_420_6x8].filter_vps = PFX(interp_4tap_vert_ps_6x8_avx2);
@@ -2016,6 +2037,10 @@
p.chroma[X265_CSP_I420].pu[CHROMA_420_32x16].filter_vsp = PFX(interp_4tap_vert_sp_32x16_avx2);
p.chroma[X265_CSP_I420].pu[CHROMA_420_32x24].filter_vsp = PFX(interp_4tap_vert_sp_32x24_avx2);
p.chroma[X265_CSP_I420].pu[CHROMA_420_32x32].filter_vsp = PFX(interp_4tap_vert_sp_32x32_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_6x16].filter_vsp = PFX(interp_4tap_vert_sp_6x16_avx2);
+ p.chroma[X265_CSP_I422].pu[CHROMA_422_6x16].filter_vps = PFX(interp_4tap_vert_ps_6x16_avx2);
+ p.chroma[X265_CSP_I422].pu[CHROMA_422_6x16].filter_vpp = PFX(interp_4tap_vert_pp_6x16_avx2);
p.chroma[X265_CSP_I422].pu[CHROMA_422_12x32].filter_vpp = PFX(interp_4tap_vert_pp_12x32_avx2);
p.chroma[X265_CSP_I422].pu[CHROMA_422_12x32].filter_vps = PFX(interp_4tap_vert_ps_12x32_avx2);
p.chroma[X265_CSP_I422].pu[CHROMA_422_12x32].filter_vss = PFX(interp_4tap_vert_ss_12x32_avx2);
@@ -2132,56 +2157,6 @@
p.chroma[X265_CSP_I444].pu[LUMA_64x32].filter_vsp = PFX(interp_4tap_vert_sp_64x32_avx2);
p.chroma[X265_CSP_I444].pu[LUMA_64x48].filter_vsp = PFX(interp_4tap_vert_sp_64x48_avx2);
p.chroma[X265_CSP_I444].pu[LUMA_64x64].filter_vsp = PFX(interp_4tap_vert_sp_64x64_avx2);
-
- /* The following primitives have been disabled since performance compared to SSE is negligible/negative */
-#if 0
- 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_6x16].filter_vps = PFX(interp_4tap_vert_ps_6x16_avx2);
-
- p.chroma[X265_CSP_I420].pu[CHROMA_420_12x16].filter_hpp = PFX(interp_4tap_horiz_pp_12x16_avx2);
- p.chroma[X265_CSP_I444].pu[LUMA_12x16].filter_hpp = PFX(interp_4tap_horiz_pp_12x16_avx2);
- p.chroma[X265_CSP_I422].pu[CHROMA_422_12x32].filter_hpp = PFX(interp_4tap_horiz_pp_12x32_avx2);
-
- p.cu[BLOCK_4x4].cpy1Dto2D_shl = PFX(cpy1Dto2D_shl_4_avx2);
- p.cu[BLOCK_8x8].cpy1Dto2D_shl = PFX(cpy1Dto2D_shl_8_avx2);
- p.cu[BLOCK_4x4].cpy1Dto2D_shr = PFX(cpy1Dto2D_shr_4_avx2);
- p.cu[BLOCK_4x4].count_nonzero = PFX(count_nonzero_4x4_avx2);
- p.cu[BLOCK_8x8].copy_cnt = PFX(copy_cnt_8_avx2);
-
- p.chroma[X265_CSP_I444].pu[LUMA_12x16].filter_hps = PFX(interp_4tap_horiz_ps_12x16_avx2);
- p.chroma[X265_CSP_I420].pu[CHROMA_420_12x16].filter_hps = PFX(interp_4tap_horiz_ps_12x16_avx2);
- p.chroma[X265_CSP_I422].pu[CHROMA_422_12x32].filter_hps = PFX(interp_4tap_horiz_ps_12x32_avx2);
- p.chroma[X265_CSP_I422].pu[CHROMA_422_6x16].filter_hps = PFX(interp_4tap_horiz_ps_6x16_avx2);
- p.chroma[X265_CSP_I420].pu[CHROMA_420_6x8].filter_hps = PFX(interp_4tap_horiz_ps_6x8_avx2);
- p.chroma[X265_CSP_I444].pu[LUMA_16x32].filter_hps = PFX(interp_4tap_horiz_ps_16x32_avx2);
- p.chroma[X265_CSP_I420].pu[CHROMA_420_16x32].filter_hps = PFX(interp_4tap_horiz_ps_16x32_avx2);
- p.chroma[X265_CSP_I422].pu[CHROMA_422_16x32].filter_hps = PFX(interp_4tap_horiz_ps_16x32_avx2);
- p.chroma[X265_CSP_I420].pu[CHROMA_420_32x32].filter_hps = PFX(interp_4tap_horiz_ps_32x32_avx2);
- p.chroma[X265_CSP_I422].pu[CHROMA_422_32x32].filter_hps = PFX(interp_4tap_horiz_ps_32x32_avx2);
- p.chroma[X265_CSP_I444].pu[LUMA_32x32].filter_hps = PFX(interp_4tap_horiz_ps_32x32_avx2);
- p.chroma[X265_CSP_I420].pu[CHROMA_420_24x32].filter_hps = PFX(interp_4tap_horiz_ps_24x32_avx2);
- p.chroma[X265_CSP_I444].pu[LUMA_24x32].filter_hps = PFX(interp_4tap_horiz_ps_24x32_avx2);
- p.chroma[X265_CSP_I422].pu[CHROMA_422_32x48].filter_hps = PFX(interp_4tap_horiz_ps_32x48_avx2);
- p.chroma[X265_CSP_I422].pu[CHROMA_422_8x64].filter_hps = PFX(interp_4tap_horiz_ps_8x64_avx2);
- p.chroma[X265_CSP_I420].pu[CHROMA_420_32x16].filter_hps = PFX(interp_4tap_horiz_ps_32x16_avx2);
- p.chroma[X265_CSP_I422].pu[CHROMA_422_32x16].filter_hps = PFX(interp_4tap_horiz_ps_32x16_avx2);
- p.chroma[X265_CSP_I444].pu[LUMA_32x16].filter_hps = PFX(interp_4tap_horiz_ps_32x16_avx2);
- p.chroma[X265_CSP_I444].pu[LUMA_16x16].filter_hps = PFX(interp_4tap_horiz_ps_16x16_avx2);
- p.chroma[X265_CSP_I422].pu[CHROMA_422_16x16].filter_hps = PFX(interp_4tap_horiz_ps_16x16_avx2);
- p.chroma[X265_CSP_I420].pu[CHROMA_420_16x16].filter_hps = PFX(interp_4tap_horiz_ps_16x16_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_6x16].filter_vsp = PFX(interp_4tap_vert_sp_6x16_avx2);
- p.chroma[X265_CSP_I422].pu[CHROMA_422_6x16].filter_vpp = PFX(interp_4tap_vert_pp_6x16_avx2);
-
- p.pu[LUMA_8x4].addAvg = PFX(addAvg_8x4_avx2);
- p.pu[LUMA_8x8].addAvg = PFX(addAvg_8x8_avx2);
- p.pu[LUMA_12x16].addAvg = PFX(addAvg_12x16_avx2);
- p.chroma[X265_CSP_I420].pu[CHROMA_420_8x4].addAvg = PFX(addAvg_8x4_avx2);
- p.chroma[X265_CSP_I420].pu[CHROMA_420_12x16].addAvg = PFX(addAvg_12x16_avx2);
- p.chroma[X265_CSP_I422].pu[CHROMA_422_8x8].addAvg = PFX(addAvg_8x8_avx2);
- p.chroma[X265_CSP_I422].pu[CHROMA_422_8x4].addAvg = PFX(addAvg_8x4_avx2);
- p.chroma[X265_CSP_I422].pu[CHROMA_422_12x32].addAvg = PFX(addAvg_12x32_avx2);
-#endif
#endif
p.frameInitLowres = PFX(frame_init_lowres_core_avx2);
@@ -2585,6 +2560,7 @@
}
if (cpuMask & X265_CPU_AVX)
{
+ p.pu[LUMA_4x4].satd = p.cu[BLOCK_4x4].sa8d = PFX(pixel_satd_4x4_avx);
p.chroma[X265_CSP_I422].pu[CHROMA_422_16x24].satd = PFX(pixel_satd_16x24_avx);
p.chroma[X265_CSP_I422].pu[CHROMA_422_32x48].satd = PFX(pixel_satd_32x48_avx);
p.chroma[X265_CSP_I422].pu[CHROMA_422_24x64].satd = PFX(pixel_satd_24x64_avx);
@@ -2597,25 +2573,55 @@
p.chroma[X265_CSP_I422].pu[CHROMA_422_16x16].satd = PFX(pixel_satd_16x16_avx);
p.chroma[X265_CSP_I422].pu[CHROMA_422_32x32].satd = PFX(pixel_satd_32x32_avx);
p.chroma[X265_CSP_I422].pu[CHROMA_422_16x64].satd = PFX(pixel_satd_16x64_avx);
+ p.chroma[X265_CSP_I422].pu[CHROMA_422_16x8].satd = PFX(pixel_satd_16x8_avx);
p.chroma[X265_CSP_I422].pu[CHROMA_422_32x16].satd = PFX(pixel_satd_32x16_avx);
+ p.chroma[X265_CSP_I422].pu[CHROMA_422_8x4].satd = PFX(pixel_satd_8x4_avx);
p.chroma[X265_CSP_I422].pu[CHROMA_422_8x16].satd = PFX(pixel_satd_8x16_avx);
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.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);
+ ALL_LUMA_PU(satd, pixel_satd, avx);
+ p.chroma[X265_CSP_I420].pu[CHROMA_420_4x4].satd = PFX(pixel_satd_4x4_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);
+ 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_16x8].satd = PFX(pixel_satd_16x8_avx);
p.chroma[X265_CSP_I420].pu[CHROMA_420_8x16].satd = PFX(pixel_satd_8x16_avx);
p.chroma[X265_CSP_I420].pu[CHROMA_420_32x16].satd = PFX(pixel_satd_32x16_avx);
p.chroma[X265_CSP_I420].pu[CHROMA_420_16x32].satd = PFX(pixel_satd_16x32_avx);
p.chroma[X265_CSP_I420].pu[CHROMA_420_16x12].satd = PFX(pixel_satd_16x12_avx);
+ p.chroma[X265_CSP_I420].pu[CHROMA_420_12x16].satd = PFX(pixel_satd_12x16_avx);
+ p.chroma[X265_CSP_I420].pu[CHROMA_420_16x4].satd = PFX(pixel_satd_16x4_avx);
+ p.chroma[X265_CSP_I420].pu[CHROMA_420_4x16].satd = PFX(pixel_satd_4x16_avx);
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_32x8].satd = PFX(pixel_satd_32x8_avx);
p.chroma[X265_CSP_I420].pu[CHROMA_420_8x32].satd = PFX(pixel_satd_8x32_avx);
+ ASSIGN_SA8D(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_4x4].sa8d = PFX(pixel_satd_4x4_avx);
+ ASSIGN_SSE_PP(avx);
p.chroma[X265_CSP_I420].cu[BLOCK_420_8x8].sse_pp = PFX(pixel_ssd_8x8_avx);
+ ASSIGN_SSE_SS(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);
+ HEVC_SAD_X3(avx);
+
+ p.pu[LUMA_12x16].sad_x4 = PFX(pixel_sad_x4_12x16_avx);
p.pu[LUMA_16x4].sad_x4 = PFX(pixel_sad_x4_16x4_avx);
+ HEVC_SAD_X4(avx);
+
+ p.ssim_4x4x2_core = PFX(pixel_ssim_4x4x2_core_avx);
+ p.ssim_end_4 = PFX(pixel_ssim_end4_avx);
+
p.cu[BLOCK_16x16].copy_ss = PFX(blockcopy_ss_16x16_avx);
p.cu[BLOCK_32x32].copy_ss = PFX(blockcopy_ss_32x32_avx);
p.cu[BLOCK_64x64].copy_ss = PFX(blockcopy_ss_64x64_avx);
@@ -2650,71 +2656,7 @@
p.pu[LUMA_48x64].copy_pp = PFX(blockcopy_pp_48x64_avx);
- /* The following primitives have been disabled since performance compared to SSE4.2 is negligible/negative */
-#if 0
- p.chroma[X265_CSP_I420].pu[CHROMA_420_16x4].satd = PFX(pixel_satd_16x4_avx);
- p.pu[LUMA_16x4].satd = PFX(pixel_satd_16x4_avx);
- p.pu[LUMA_16x8].satd = PFX(pixel_satd_16x8_avx);
- p.chroma[X265_CSP_I420].pu[CHROMA_420_16x8].satd = PFX(pixel_satd_16x8_avx);
- p.chroma[X265_CSP_I422].pu[CHROMA_422_16x8].satd = PFX(pixel_satd_16x8_avx);
- p.chroma[X265_CSP_I420].pu[CHROMA_420_32x8].satd = PFX(pixel_satd_32x8_avx);
- p.pu[LUMA_32x8].satd = PFX(pixel_satd_32x8_avx);
- 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.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.cu[BLOCK_16x16].sse_pp = PFX(pixel_ssd_16x16_avx);
- p.cu[BLOCK_32x32].sse_pp = PFX(pixel_ssd_32x32_avx);
- p.cu[BLOCK_8x8].sse_pp = PFX(pixel_ssd_8x8_avx);
- p.chroma[X265_CSP_I422].cu[BLOCK_422_8x16].sse_pp = PFX(pixel_ssd_8x16_avx);
- 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);
-
- 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);
- HEVC_SAD_X3(avx);
-
- p.pu[LUMA_32x8].sad_x4 = PFX(pixel_sad_x4_32x8_avx);
- p.pu[LUMA_32x16].sad_x4 = PFX(pixel_sad_x4_32x16_avx);
- p.pu[LUMA_32x24].sad_x4 = PFX(pixel_sad_x4_32x24_avx);
- p.pu[LUMA_32x32].sad_x4 = PFX(pixel_sad_x4_32x32_avx);
- p.pu[LUMA_32x64].sad_x4 = PFX(pixel_sad_x4_32x64_avx);
- p.pu[LUMA_16x8].sad_x4 = PFX(pixel_sad_x4_16x8_avx);
- p.pu[LUMA_16x16].sad_x4 = PFX(pixel_sad_x4_16x16_avx);
- p.pu[LUMA_16x32].sad_x4 = PFX(pixel_sad_x4_16x32_avx);
- p.pu[LUMA_16x12].sad_x4 = PFX(pixel_sad_x4_16x12_avx);
- p.pu[LUMA_12x16].sad_x4 = PFX(pixel_sad_x4_12x16_avx);
- p.pu[LUMA_16x64].sad_x4 = PFX(pixel_sad_x4_16x64_avx);
- p.pu[LUMA_24x32].sad_x4 = PFX(pixel_sad_x4_24x32_avx);
- p.pu[LUMA_48x64].sad_x4 = PFX(pixel_sad_x4_48x64_avx);
- p.pu[LUMA_64x16].sad_x4 = PFX(pixel_sad_x4_64x16_avx);
- p.pu[LUMA_64x32].sad_x4 = PFX(pixel_sad_x4_64x32_avx);
- p.pu[LUMA_64x48].sad_x4 = PFX(pixel_sad_x4_64x48_avx);
- p.pu[LUMA_64x64].sad_x4 = PFX(pixel_sad_x4_64x64_avx)
-
- p.ssim_4x4x2_core = PFX(pixel_ssim_4x4x2_core_avx);
- p.ssim_end_4 = PFX(pixel_ssim_end4_avx);
p.frameInitLowres = PFX(frame_init_lowres_core_avx);
-#endif
}
if (cpuMask & X265_CPU_XOP)
{
@@ -2746,6 +2688,7 @@
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);
@@ -2761,23 +2704,43 @@
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_8x4].addAvg = PFX(addAvg_8x4_avx2);
+ p.pu[LUMA_8x8].addAvg = PFX(addAvg_8x8_avx2);
+ p.pu[LUMA_8x16].addAvg = PFX(addAvg_8x16_avx2);
+ p.pu[LUMA_8x32].addAvg = PFX(addAvg_8x32_avx2);
+
+ p.pu[LUMA_12x16].addAvg = PFX(addAvg_12x16_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_8x2].addAvg = PFX(addAvg_8x2_avx2);
+ p.chroma[X265_CSP_I420].pu[CHROMA_420_8x4].addAvg = PFX(addAvg_8x4_avx2);
+ p.chroma[X265_CSP_I420].pu[CHROMA_420_8x6].addAvg = PFX(addAvg_8x6_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_12x16].addAvg = PFX(addAvg_12x16_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);
@@ -2787,6 +2750,14 @@
p.chroma[X265_CSP_I420].pu[CHROMA_420_32x16].addAvg = PFX(addAvg_32x16_avx2);
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_8x4].addAvg = PFX(addAvg_8x4_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_12x32].addAvg = PFX(addAvg_12x32_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);
@@ -2821,11 +2792,19 @@
p.chroma[X265_CSP_I422].cu[BLOCK_422_16x32].sub_ps = PFX(pixel_sub_ps_16x32_avx2);
p.chroma[X265_CSP_I422].cu[BLOCK_422_32x64].sub_ps = PFX(pixel_sub_ps_32x64_avx2);
+ p.pu[LUMA_16x4].pixelavg_pp = PFX(pixel_avg_16x4_avx2);
+ p.pu[LUMA_16x8].pixelavg_pp = PFX(pixel_avg_16x8_avx2);
+ p.pu[LUMA_16x12].pixelavg_pp = PFX(pixel_avg_16x12_avx2);
+ p.pu[LUMA_16x16].pixelavg_pp = PFX(pixel_avg_16x16_avx2);
+ p.pu[LUMA_16x32].pixelavg_pp = PFX(pixel_avg_16x32_avx2);
+ p.pu[LUMA_16x64].pixelavg_pp = PFX(pixel_avg_16x64_avx2);
+
p.pu[LUMA_32x64].pixelavg_pp = PFX(pixel_avg_32x64_avx2);
p.pu[LUMA_32x32].pixelavg_pp = PFX(pixel_avg_32x32_avx2);
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);
@@ -2835,10 +2814,12 @@
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);
@@ -2860,6 +2841,15 @@
p.pu[LUMA_64x32].sad = PFX(pixel_sad_64x32_avx2);
p.pu[LUMA_64x48].sad = PFX(pixel_sad_64x48_avx2);
p.pu[LUMA_64x64].sad = PFX(pixel_sad_64x64_avx2);
+
+ p.pu[LUMA_8x4].sad_x3 = PFX(pixel_sad_x3_8x4_avx2);
+ p.pu[LUMA_8x8].sad_x3 = PFX(pixel_sad_x3_8x8_avx2);
+ p.pu[LUMA_8x16].sad_x3 = PFX(pixel_sad_x3_8x16_avx2);
+
+ p.pu[LUMA_8x8].sad_x4 = PFX(pixel_sad_x4_8x8_avx2);
+ p.pu[LUMA_16x8].sad_x4 = PFX(pixel_sad_x4_16x8_avx2);
+ p.pu[LUMA_16x12].sad_x4 = PFX(pixel_sad_x4_16x12_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);
@@ -2882,11 +2872,9 @@
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_16x16].cpy1Dto2D_shl = PFX(cpy1Dto2D_shl_16_avx2);
- p.cu[BLOCK_32x32].cpy1Dto2D_shl = PFX(cpy1Dto2D_shl_32_avx2);
- p.cu[BLOCK_8x8].cpy1Dto2D_shr = PFX(cpy1Dto2D_shr_8_avx2);
- p.cu[BLOCK_16x16].cpy1Dto2D_shr = PFX(cpy1Dto2D_shr_16_avx2);
- p.cu[BLOCK_32x32].cpy1Dto2D_shr = PFX(cpy1Dto2D_shr_32_avx2);
+
+ ALL_LUMA_TU_S(cpy1Dto2D_shl, cpy1Dto2D_shl_, avx2);
+ ALL_LUMA_TU_S(cpy1Dto2D_shr, cpy1Dto2D_shr_, avx2);
p.cu[BLOCK_8x8].cpy2Dto1D_shl = PFX(cpy2Dto1D_shl_8_avx2);
p.cu[BLOCK_16x16].cpy2Dto1D_shl = PFX(cpy2Dto1D_shl_16_avx2);
@@ -2896,10 +2884,7 @@
p.cu[BLOCK_16x16].cpy2Dto1D_shr = PFX(cpy2Dto1D_shr_16_avx2);
p.cu[BLOCK_32x32].cpy2Dto1D_shr = PFX(cpy2Dto1D_shr_32_avx2);
- p.cu[BLOCK_8x8].count_nonzero = PFX(count_nonzero_8x8_avx2);
- p.cu[BLOCK_16x16].count_nonzero = PFX(count_nonzero_16x16_avx2);
- p.cu[BLOCK_32x32].count_nonzero = PFX(count_nonzero_32x32_avx2);
-
+ ALL_LUMA_TU(count_nonzero, count_nonzero, avx2);
p.denoiseDct = PFX(denoise_dct_avx2);
p.quant = PFX(quant_avx2);
p.nquant = PFX(nquant_avx2);
@@ -2961,9 +2946,13 @@
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[20] = PFX(intra_pred_ang8_20_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_8x8].intra_pred[16] = PFX(intra_pred_ang8_16_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);
@@ -2973,6 +2962,7 @@
p.cu[BLOCK_16x16].intra_pred[9] = PFX(intra_pred_ang16_9_avx2);
p.cu[BLOCK_16x16].intra_pred[12] = PFX(intra_pred_ang16_12_avx2);
p.cu[BLOCK_16x16].intra_pred[11] = PFX(intra_pred_ang16_11_avx2);
+ p.cu[BLOCK_16x16].intra_pred[13] = PFX(intra_pred_ang16_13_avx2);
p.cu[BLOCK_16x16].intra_pred[25] = PFX(intra_pred_ang16_25_avx2);
p.cu[BLOCK_16x16].intra_pred[28] = PFX(intra_pred_ang16_28_avx2);
p.cu[BLOCK_16x16].intra_pred[27] = PFX(intra_pred_ang16_27_avx2);
@@ -2983,6 +2973,7 @@
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[34] = PFX(intra_pred_ang32_34_avx2);
p.cu[BLOCK_32x32].intra_pred[2] = PFX(intra_pred_ang32_2_avx2);
p.cu[BLOCK_32x32].intra_pred[26] = PFX(intra_pred_ang32_26_avx2);
@@ -3004,6 +2995,11 @@
// all_angs primitives
p.cu[BLOCK_4x4].intra_pred_allangs = PFX(all_angs_pred_4x4_avx2);
+ // copy_sp primitives
+ p.cu[BLOCK_16x16].copy_sp = PFX(blockcopy_sp_16x16_avx2);
+ p.chroma[X265_CSP_I420].cu[BLOCK_420_16x16].copy_sp = PFX(blockcopy_sp_16x16_avx2);
+ p.chroma[X265_CSP_I422].cu[BLOCK_422_16x32].copy_sp = PFX(blockcopy_sp_16x32_avx2);
+
p.cu[BLOCK_32x32].copy_sp = PFX(blockcopy_sp_32x32_avx2);
p.chroma[X265_CSP_I420].cu[BLOCK_420_32x32].copy_sp = PFX(blockcopy_sp_32x32_avx2);
p.chroma[X265_CSP_I422].cu[BLOCK_422_32x64].copy_sp = PFX(blockcopy_sp_32x64_avx2);
@@ -3015,6 +3011,12 @@
p.chroma[X265_CSP_I420].cu[CHROMA_420_16x16].copy_ps = PFX(blockcopy_ps_16x16_avx2);
p.chroma[X265_CSP_I422].cu[CHROMA_422_16x32].copy_ps = PFX(blockcopy_ps_16x32_avx2);
+ p.cu[BLOCK_32x32].copy_ps = PFX(blockcopy_ps_32x32_avx2);
+ p.chroma[X265_CSP_I420].cu[CHROMA_420_32x32].copy_ps = PFX(blockcopy_ps_32x32_avx2);
+ p.chroma[X265_CSP_I422].cu[CHROMA_422_32x64].copy_ps = PFX(blockcopy_ps_32x64_avx2);
+
+ p.cu[BLOCK_64x64].copy_ps = PFX(blockcopy_ps_64x64_avx2);
+
ALL_LUMA_TU_S(dct, dct, avx2);
ALL_LUMA_TU_S(idct, idct, avx2);
ALL_LUMA_CU_S(transpose, transpose, avx2);
@@ -3065,6 +3067,7 @@
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);
@@ -3077,43 +3080,59 @@
p.pu[LUMA_64x16].luma_hps = PFX(interp_8tap_horiz_ps_64x16_avx2);
p.pu[LUMA_12x16].luma_hps = PFX(interp_8tap_horiz_ps_12x16_avx2);
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_4x8].filter_hpp = PFX(interp_4tap_horiz_pp_4x8_avx2);
+ p.chroma[X265_CSP_I420].pu[CHROMA_420_4x16].filter_hpp = PFX(interp_4tap_horiz_pp_4x16_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);
@@ -3122,19 +3141,29 @@
p.chroma[X265_CSP_I420].pu[CHROMA_420_2x8].filter_hps = PFX(interp_4tap_horiz_ps_2x8_avx2);
p.chroma[X265_CSP_I420].pu[CHROMA_420_6x8].filter_hps = PFX(interp_4tap_horiz_ps_6x8_avx2);
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_16x4].filter_vpp = PFX(interp_4tap_vert_pp_16x4_avx2);
+ p.chroma[X265_CSP_I420].pu[CHROMA_420_16x8].filter_vpp = PFX(interp_4tap_vert_pp_16x8_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_16x16].filter_vpp = PFX(interp_4tap_vert_pp_16x16_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);
@@ -3142,12 +3171,16 @@
p.chroma[X265_CSP_I420].pu[CHROMA_420_2x4].filter_vps = PFX(interp_4tap_vert_ps_2x4_avx2);
p.chroma[X265_CSP_I420].pu[CHROMA_420_2x8].filter_vps = PFX(interp_4tap_vert_ps_2x8_avx2);
+ p.chroma[X265_CSP_I420].pu[CHROMA_420_4x2].filter_vps = PFX(interp_4tap_vert_ps_4x2_avx2);
p.chroma[X265_CSP_I420].pu[CHROMA_420_4x4].filter_vps = PFX(interp_4tap_vert_ps_4x4_avx2);
p.chroma[X265_CSP_I420].pu[CHROMA_420_4x8].filter_vps = PFX(interp_4tap_vert_ps_4x8_avx2);
+ p.chroma[X265_CSP_I420].pu[CHROMA_420_6x8].filter_vps = PFX(interp_4tap_vert_ps_6x8_avx2);
p.chroma[X265_CSP_I420].pu[CHROMA_420_8x2].filter_vps = PFX(interp_4tap_vert_ps_8x2_avx2);
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);
@@ -3155,6 +3188,7 @@
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);
@@ -3208,6 +3242,7 @@
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);
@@ -3222,7 +3257,11 @@
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);
@@ -3236,6 +3275,7 @@
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);
@@ -3246,12 +3286,17 @@
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);
+
p.pu[LUMA_16x16].luma_hvpp = PFX(interp_8tap_hv_pp_16x16_avx2);
ALL_LUMA_PU_T(luma_hvpp, interp_8tap_hv_pp_cpu);
@@ -3284,28 +3329,43 @@
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_4x8].filter_hpp = PFX(interp_4tap_horiz_pp_4x8_avx2);
+ p.chroma[X265_CSP_I422].pu[CHROMA_422_4x16].filter_hpp = PFX(interp_4tap_horiz_pp_4x16_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);
p.chroma[X265_CSP_I422].pu[CHROMA_422_16x64].filter_hpp = PFX(interp_4tap_horiz_pp_16x64_avx2);
p.chroma[X265_CSP_I422].pu[CHROMA_422_16x24].filter_hpp = PFX(interp_4tap_horiz_pp_16x24_avx2);
+
p.chroma[X265_CSP_I422].pu[CHROMA_422_32x16].filter_hpp = PFX(interp_4tap_horiz_pp_32x16_avx2);
p.chroma[X265_CSP_I422].pu[CHROMA_422_32x32].filter_hpp = PFX(interp_4tap_horiz_pp_32x32_avx2);
p.chroma[X265_CSP_I422].pu[CHROMA_422_32x64].filter_hpp = PFX(interp_4tap_horiz_pp_32x64_avx2);
p.chroma[X265_CSP_I422].pu[CHROMA_422_32x48].filter_hpp = PFX(interp_4tap_horiz_pp_32x48_avx2);
+
p.chroma[X265_CSP_I422].pu[CHROMA_422_2x8].filter_hpp = PFX(interp_4tap_horiz_pp_2x8_avx2);
//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);
+
+ p.chroma[X265_CSP_I444].pu[LUMA_4x8].filter_hpp = PFX(interp_4tap_horiz_pp_4x8_avx2);
+ p.chroma[X265_CSP_I444].pu[LUMA_4x16].filter_hpp = PFX(interp_4tap_horiz_pp_4x16_avx2);
+
p.chroma[X265_CSP_I444].pu[LUMA_8x4].filter_hpp = PFX(interp_4tap_horiz_pp_8x4_avx2);
p.chroma[X265_CSP_I444].pu[LUMA_8x16].filter_hpp = PFX(interp_4tap_horiz_pp_8x16_avx2);
p.chroma[X265_CSP_I444].pu[LUMA_8x32].filter_hpp = PFX(interp_4tap_horiz_pp_8x32_avx2);
@@ -3444,12 +3504,14 @@
//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);
@@ -3474,6 +3536,7 @@
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);
@@ -3481,7 +3544,9 @@
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);
@@ -3493,20 +3558,25 @@
//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);
p.chroma[X265_CSP_I422].pu[CHROMA_422_4x16].filter_vpp = PFX(interp_4tap_vert_pp_4x16_avx2);
+ p.chroma[X265_CSP_I422].pu[CHROMA_422_16x16].filter_vpp = PFX(interp_4tap_vert_pp_16x16_avx2);
p.chroma[X265_CSP_I422].pu[CHROMA_422_8x32].filter_vpp = PFX(interp_4tap_vert_pp_8x32_avx2);
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_16x8].filter_vpp = PFX(interp_4tap_vert_pp_16x8_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);
p.chroma[X265_CSP_I422].pu[CHROMA_422_12x32].filter_vpp = PFX(interp_4tap_vert_pp_12x32_avx2);
p.chroma[X265_CSP_I422].pu[CHROMA_422_8x12].filter_vpp = PFX(interp_4tap_vert_pp_8x12_avx2);
p.chroma[X265_CSP_I422].pu[CHROMA_422_2x4].filter_vpp = PFX(interp_4tap_vert_pp_2x4_avx2);
+ p.chroma[X265_CSP_I422].pu[CHROMA_422_16x24].filter_vpp = PFX(interp_4tap_vert_pp_16x24_avx2);
p.chroma[X265_CSP_I422].pu[CHROMA_422_2x16].filter_vpp = PFX(interp_4tap_vert_pp_2x16_avx2);
p.chroma[X265_CSP_I422].pu[CHROMA_422_4x32].filter_vpp = PFX(interp_4tap_vert_pp_4x32_avx2);
p.chroma[X265_CSP_I422].pu[CHROMA_422_24x64].filter_vpp = PFX(interp_4tap_vert_pp_24x64_avx2);
@@ -3514,16 +3584,23 @@
//i444 for chroma_vpp
p.chroma[X265_CSP_I444].pu[LUMA_4x4].filter_vpp = PFX(interp_4tap_vert_pp_4x4_avx2);
p.chroma[X265_CSP_I444].pu[LUMA_8x8].filter_vpp = PFX(interp_4tap_vert_pp_8x8_avx2);
+ p.chroma[X265_CSP_I444].pu[LUMA_16x16].filter_vpp = PFX(interp_4tap_vert_pp_16x16_avx2);
p.chroma[X265_CSP_I444].pu[LUMA_32x32].filter_vpp = PFX(interp_4tap_vert_pp_32x32_avx2);
p.chroma[X265_CSP_I444].pu[LUMA_8x4].filter_vpp = PFX(interp_4tap_vert_pp_8x4_avx2);
p.chroma[X265_CSP_I444].pu[LUMA_4x8].filter_vpp = PFX(interp_4tap_vert_pp_4x8_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_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_16x4].filter_vpp = PFX(interp_4tap_vert_pp_16x4_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);
@@ -3535,118 +3612,11 @@
if (cpuMask & X265_CPU_BMI2)
p.scanPosLast = PFX(scanPosLast_avx2_bmi2);
-
p.cu[BLOCK_32x32].copy_ps = PFX(blockcopy_ps_32x32_avx2);
p.chroma[X265_CSP_I420].cu[CHROMA_420_32x32].copy_ps = PFX(blockcopy_ps_32x32_avx2);
p.chroma[X265_CSP_I422].cu[CHROMA_422_32x64].copy_ps = PFX(blockcopy_ps_32x64_avx2);
p.cu[BLOCK_64x64].copy_ps = PFX(blockcopy_ps_64x64_avx2);
- /* The following primitives have been disabled since performance compared to SSE is negligible/negative */
-#if 0
- p.pu[LUMA_8x4].addAvg = PFX(addAvg_8x4_avx2);
- p.pu[LUMA_8x8].addAvg = PFX(addAvg_8x8_avx2);
- p.pu[LUMA_12x16].addAvg = PFX(addAvg_12x16_avx2);
- p.chroma[X265_CSP_I420].pu[CHROMA_420_8x2].addAvg = PFX(addAvg_8x2_avx2);
- p.chroma[X265_CSP_I420].pu[CHROMA_420_8x4].addAvg = PFX(addAvg_8x4_avx2);
- p.chroma[X265_CSP_I420].pu[CHROMA_420_8x6].addAvg = PFX(addAvg_8x6_avx2);
- p.chroma[X265_CSP_I420].pu[CHROMA_420_12x16].addAvg = PFX(addAvg_12x16_avx2);
- p.chroma[X265_CSP_I422].pu[CHROMA_422_8x4].addAvg = PFX(addAvg_8x4_avx2);
- p.chroma[X265_CSP_I422].pu[CHROMA_422_12x32].addAvg = PFX(addAvg_12x32_avx2);
- p.pu[LUMA_16x4].pixelavg_pp = PFX(pixel_avg_16x4_avx2);
- p.pu[LUMA_16x8].pixelavg_pp = PFX(pixel_avg_16x8_avx2);
- p.pu[LUMA_16x12].pixelavg_pp = PFX(pixel_avg_16x12_avx2);
- p.pu[LUMA_16x16].pixelavg_pp = PFX(pixel_avg_16x16_avx2);
- p.pu[LUMA_16x32].pixelavg_pp = PFX(pixel_avg_16x32_avx2);
- p.pu[LUMA_16x64].pixelavg_pp = PFX(pixel_avg_16x64_avx2);
- p.pu[LUMA_8x4].sad_x3 = PFX(pixel_sad_x3_8x4_avx2);
- p.pu[LUMA_8x16].sad_x3 = PFX(pixel_sad_x3_8x16_avx2);
- p.pu[LUMA_8x8].sad_x4 = PFX(pixel_sad_x4_8x8_avx2);
- p.pu[LUMA_16x12].sad_x4 = PFX(pixel_sad_x4_16x12_avx2);
- p.cu[BLOCK_4x4].cpy1Dto2D_shl = PFX(cpy1Dto2D_shl_4_avx2);
- p.cu[BLOCK_4x4].cpy1Dto2D_shr = PFX(cpy1Dto2D_shr_4_avx2);
- p.cu[BLOCK_4x4].count_nonzero = PFX(count_nonzero_4x4_avx2);
- p.cu[BLOCK_8x8].intra_pred[20] = PFX(intra_pred_ang8_20_avx2);
- p.cu[BLOCK_8x8].intra_pred[16] = PFX(intra_pred_ang8_16_avx2);
- p.cu[BLOCK_16x16].intra_pred[13] = PFX(intra_pred_ang16_13_avx2);
- p.cu[BLOCK_16x16].copy_sp = PFX(blockcopy_sp_16x16_avx2);
- p.chroma[X265_CSP_I420].cu[BLOCK_420_16x16].copy_sp = PFX(blockcopy_sp_16x16_avx2);
- p.chroma[X265_CSP_I422].cu[BLOCK_422_16x32].copy_sp = PFX(blockcopy_sp_16x32_avx2);
- p.chroma[X265_CSP_I420].pu[CHROMA_420_4x8].filter_hpp = PFX(interp_4tap_horiz_pp_4x8_avx2);
- p.chroma[X265_CSP_I420].pu[CHROMA_420_4x16].filter_hpp = PFX(interp_4tap_horiz_pp_4x16_avx2);
- p.chroma[X265_CSP_I420].pu[CHROMA_420_16x4].filter_vpp = PFX(interp_4tap_vert_pp_16x4_avx2);
- p.chroma[X265_CSP_I420].pu[CHROMA_420_16x8].filter_vpp = PFX(interp_4tap_vert_pp_16x8_avx2);
- p.chroma[X265_CSP_I420].pu[CHROMA_420_16x16].filter_vpp = PFX(interp_4tap_vert_pp_16x16_avx2);
- p.chroma[X265_CSP_I420].pu[CHROMA_420_4x2].filter_vps = PFX(interp_4tap_vert_ps_4x2_avx2);
- p.chroma[X265_CSP_I420].pu[CHROMA_420_6x8].filter_vps = PFX(interp_4tap_vert_ps_6x8_avx2);
- p.chroma[X265_CSP_I422].pu[CHROMA_422_4x8].filter_hpp = PFX(interp_4tap_horiz_pp_4x8_avx2);
- p.chroma[X265_CSP_I422].pu[CHROMA_422_4x16].filter_hpp = PFX(interp_4tap_horiz_pp_4x16_avx2);
- p.chroma[X265_CSP_I444].pu[LUMA_4x8].filter_hpp = PFX(interp_4tap_horiz_pp_4x8_avx2);
- p.chroma[X265_CSP_I444].pu[LUMA_4x16].filter_hpp = PFX(interp_4tap_horiz_pp_4x16_avx2);
- p.chroma[X265_CSP_I422].pu[CHROMA_422_16x16].filter_vpp = PFX(interp_4tap_vert_pp_16x16_avx2);
- p.chroma[X265_CSP_I422].pu[CHROMA_422_16x8].filter_vpp = PFX(interp_4tap_vert_pp_16x8_avx2);
- p.chroma[X265_CSP_I422].pu[CHROMA_422_16x24].filter_vpp = PFX(interp_4tap_vert_pp_16x24_avx2);
- 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
}
More information about the x265-devel
mailing list