right just improve 20% in ps mode, we may modify algorithm to get more in future
At 2015-03-11 16:15:02,"Divya Manivannan" <[email protected]> wrote: ># HG changeset patch ># User Divya Manivannan <[email protected]> ># Date 1426061646 -19800 ># Wed Mar 11 13:44:06 2015 +0530 ># Node ID e90e9eeda28ece884b4189eedd48051d6dc0b44f ># Parent 97ff5bb38e91610bcbbef706850c777171563e8b >asm: filter_vpp[16x32], filter_vps[16x32] in avx2: 1944c->1870c, 1907c->1628c > >diff -r 97ff5bb38e91 -r e90e9eeda28e source/common/x86/asm-primitives.cpp >--- a/source/common/x86/asm-primitives.cpp Wed Mar 11 12:00:17 2015 +0530 >+++ b/source/common/x86/asm-primitives.cpp Wed Mar 11 13:44:06 2015 +0530 >@@ -1594,6 +1594,7 @@ > p.chroma[X265_CSP_I420].pu[CHROMA_420_16x4].filter_vpp = > x265_interp_4tap_vert_pp_16x4_avx2; > p.chroma[X265_CSP_I420].pu[CHROMA_420_16x8].filter_vpp = > x265_interp_4tap_vert_pp_16x8_avx2; > p.chroma[X265_CSP_I420].pu[CHROMA_420_16x16].filter_vpp = > x265_interp_4tap_vert_pp_16x16_avx2; >+ p.chroma[X265_CSP_I420].pu[CHROMA_420_16x32].filter_vpp = >x265_interp_4tap_vert_pp_16x32_avx2; > p.chroma[X265_CSP_I420].pu[CHROMA_420_32x8].filter_vpp = > x265_interp_4tap_vert_pp_32x8_avx2; > p.chroma[X265_CSP_I420].pu[CHROMA_420_32x16].filter_vpp = > x265_interp_4tap_vert_pp_32x16_avx2; > p.chroma[X265_CSP_I420].pu[CHROMA_420_32x24].filter_vpp = > x265_interp_4tap_vert_pp_32x24_avx2; >@@ -1616,6 +1617,7 @@ > p.chroma[X265_CSP_I420].pu[CHROMA_420_16x8].filter_vps = > x265_interp_4tap_vert_ps_16x8_avx2; > p.chroma[X265_CSP_I420].pu[CHROMA_420_4x16].filter_vps = > x265_interp_4tap_vert_ps_4x16_avx2; > p.chroma[X265_CSP_I420].pu[CHROMA_420_16x16].filter_vps = > x265_interp_4tap_vert_ps_16x16_avx2; >+ p.chroma[X265_CSP_I420].pu[CHROMA_420_16x32].filter_vps = >x265_interp_4tap_vert_ps_16x32_avx2; > p.chroma[X265_CSP_I420].pu[CHROMA_420_32x32].filter_vps = > x265_interp_4tap_vert_ps_32x32_avx2; > p.chroma[X265_CSP_I420].pu[CHROMA_420_32x24].filter_vps = > x265_interp_4tap_vert_ps_32x24_avx2; > p.chroma[X265_CSP_I420].pu[CHROMA_420_32x16].filter_vps = > x265_interp_4tap_vert_ps_32x16_avx2; >diff -r 97ff5bb38e91 -r e90e9eeda28e source/common/x86/ipfilter8.asm >--- a/source/common/x86/ipfilter8.asm Wed Mar 11 12:00:17 2015 +0530 >+++ b/source/common/x86/ipfilter8.asm Wed Mar 11 13:44:06 2015 +0530 >@@ -5905,6 +5905,299 @@ > FILTER_VER_CHROMA_AVX2_16x4 pp > FILTER_VER_CHROMA_AVX2_16x4 ps > >+%macro FILTER_VER_CHROMA_AVX2_16x32 1 >+INIT_YMM avx2 >+%if ARCH_X86_64 == 1 >+cglobal interp_4tap_vert_%1_16x32, 4, 8, 8 >+ mov r4d, r4m >+ shl r4d, 6 >+ >+%ifdef PIC >+ lea r5, [tab_ChromaCoeffVer_32] >+ add r5, r4 >+%else >+ lea r5, [tab_ChromaCoeffVer_32 + r4] >+%endif >+ >+ lea r4, [r1 * 3] >+ sub r0, r1 >+%ifidn %1,pp >+ mova m7, [pw_512] >+%else >+ add r3d, r3d >+ mova m7, [pw_2000] >+%endif >+ lea r6, [r3 * 3] >+ mov r7d, 2 >+.loopH: >+ movu xm0, [r0] >+ vinserti128 m0, m0, [r0 + r1 * 2], 1 >+ movu xm1, [r0 + r1] >+ vinserti128 m1, m1, [r0 + r4], 1 >+ >+ punpcklbw m2, m0, m1 >+ punpckhbw m3, m0, m1 >+ vperm2i128 m4, m2, m3, 0x20 >+ vperm2i128 m2, m2, m3, 0x31 >+ pmaddubsw m4, [r5] >+ pmaddubsw m3, m2, [r5 + mmsize] >+ paddw m4, m3 >+ pmaddubsw m2, [r5] >+ >+ vextracti128 xm0, m0, 1 >+ lea r0, [r0 + r1 * 4] >+ vinserti128 m0, m0, [r0], 1 >+ >+ punpcklbw m5, m1, m0 >+ punpckhbw m3, m1, m0 >+ vperm2i128 m6, m5, m3, 0x20 >+ vperm2i128 m5, m5, m3, 0x31 >+ pmaddubsw m6, [r5] >+ pmaddubsw m3, m5, [r5 + mmsize] >+ paddw m6, m3 >+ pmaddubsw m5, [r5] >+%ifidn %1,pp >+ pmulhrsw m4, m7 ; m4 = word: row 0 >+ pmulhrsw m6, m7 ; m6 = word: row 1 >+ packuswb m4, m6 >+ vpermq m4, m4, 11011000b >+ vextracti128 xm6, m4, 1 >+ movu [r2], xm4 >+ movu [r2 + r3], xm6 >+%else >+ psubw m4, m7 ; m4 = word: row 0 >+ psubw m6, m7 ; m6 = word: row 1 >+ movu [r2], m4 >+ movu [r2 + r3], m6 >+%endif >+ >+ movu xm4, [r0 + r1 * 2] >+ vinserti128 m4, m4, [r0 + r1], 1 >+ vextracti128 xm1, m4, 1 >+ vinserti128 m0, m0, xm1, 0 >+ >+ punpcklbw m6, m0, m4 >+ punpckhbw m1, m0, m4 >+ vperm2i128 m0, m6, m1, 0x20 >+ vperm2i128 m6, m6, m1, 0x31 >+ pmaddubsw m1, m0, [r5 + mmsize] >+ paddw m5, m1 >+ pmaddubsw m0, [r5] >+ pmaddubsw m1, m6, [r5 + mmsize] >+ paddw m2, m1 >+ pmaddubsw m6, [r5] >+ >+%ifidn %1,pp >+ pmulhrsw m2, m7 ; m2 = word: row 2 >+ pmulhrsw m5, m7 ; m5 = word: row 3 >+ packuswb m2, m5 >+ vpermq m2, m2, 11011000b >+ vextracti128 xm5, m2, 1 >+ movu [r2 + r3 * 2], xm2 >+ movu [r2 + r6], xm5 >+%else >+ psubw m2, m7 ; m2 = word: row 2 >+ psubw m5, m7 ; m5 = word: row 3 >+ movu [r2 + r3 * 2], m2 >+ movu [r2 + r6], m5 >+%endif >+ lea r2, [r2 + r3 * 4] >+ >+ movu xm1, [r0 + r4] >+ lea r0, [r0 + r1 * 4] >+ vinserti128 m1, m1, [r0], 1 >+ vinserti128 m4, m4, xm1, 1 >+ >+ punpcklbw m2, m4, m1 >+ punpckhbw m5, m4, m1 >+ vperm2i128 m3, m2, m5, 0x20 >+ vperm2i128 m2, m2, m5, 0x31 >+ pmaddubsw m5, m3, [r5 + mmsize] >+ paddw m6, m5 >+ pmaddubsw m3, [r5] >+ pmaddubsw m5, m2, [r5 + mmsize] >+ paddw m0, m5 >+ pmaddubsw m2, [r5] >+ >+%ifidn %1,pp >+ pmulhrsw m6, m7 ; m6 = word: row 4 >+ pmulhrsw m0, m7 ; m0 = word: row 5 >+ packuswb m6, m0 >+ vpermq m6, m6, 11011000b >+ vextracti128 xm0, m6, 1 >+ movu [r2], xm6 >+ movu [r2 + r3], xm0 >+%else >+ psubw m6, m7 ; m6 = word: row 4 >+ psubw m0, m7 ; m0 = word: row 5 >+ movu [r2], m6 >+ movu [r2 + r3], m0 >+%endif >+ >+ movu xm6, [r0 + r1 * 2] >+ vinserti128 m6, m6, [r0 + r1], 1 >+ vextracti128 xm0, m6, 1 >+ vinserti128 m1, m1, xm0, 0 >+ >+ punpcklbw m4, m1, m6 >+ punpckhbw m5, m1, m6 >+ vperm2i128 m0, m4, m5, 0x20 >+ vperm2i128 m5, m4, m5, 0x31 >+ pmaddubsw m4, m0, [r5 + mmsize] >+ paddw m2, m4 >+ pmaddubsw m0, [r5] >+ pmaddubsw m4, m5, [r5 + mmsize] >+ paddw m3, m4 >+ pmaddubsw m5, [r5] >+ >+%ifidn %1,pp >+ pmulhrsw m3, m7 ; m3 = word: row 6 >+ pmulhrsw m2, m7 ; m2 = word: row 7 >+ packuswb m3, m2 >+ vpermq m3, m3, 11011000b >+ vextracti128 xm2, m3, 1 >+ movu [r2 + r3 * 2], xm3 >+ movu [r2 + r6], xm2 >+%else >+ psubw m3, m7 ; m3 = word: row 6 >+ psubw m2, m7 ; m2 = word: row 7 >+ movu [r2 + r3 * 2], m3 >+ movu [r2 + r6], m2 >+%endif >+ lea r2, [r2 + r3 * 4] >+ >+ movu xm3, [r0 + r4] >+ lea r0, [r0 + r1 * 4] >+ vinserti128 m3, m3, [r0], 1 >+ vinserti128 m6, m6, xm3, 1 >+ >+ punpcklbw m2, m6, m3 >+ punpckhbw m1, m6, m3 >+ vperm2i128 m4, m2, m1, 0x20 >+ vperm2i128 m2, m2, m1, 0x31 >+ pmaddubsw m1, m4, [r5 + mmsize] >+ paddw m5, m1 >+ pmaddubsw m4, [r5] >+ pmaddubsw m1, m2, [r5 + mmsize] >+ paddw m0, m1 >+ pmaddubsw m2, [r5] >+ >+%ifidn %1,pp >+ pmulhrsw m5, m7 ; m5 = word: row 8 >+ pmulhrsw m0, m7 ; m0 = word: row 9 >+ packuswb m5, m0 >+ vpermq m5, m5, 11011000b >+ vextracti128 xm0, m5, 1 >+ movu [r2], xm5 >+ movu [r2 + r3], xm0 >+%else >+ psubw m5, m7 ; m5 = word: row 8 >+ psubw m0, m7 ; m0 = word: row 9 >+ movu [r2], m5 >+ movu [r2 + r3], m0 >+%endif >+ >+ movu xm5, [r0 + r1 * 2] >+ vinserti128 m5, m5, [r0 + r1], 1 >+ vextracti128 xm0, m5, 1 >+ vinserti128 m3, m3, xm0, 0 >+ >+ punpcklbw m1, m3, m5 >+ punpckhbw m0, m3, m5 >+ vperm2i128 m6, m1, m0, 0x20 >+ vperm2i128 m0, m1, m0, 0x31 >+ pmaddubsw m1, m6, [r5 + mmsize] >+ paddw m2, m1 >+ pmaddubsw m6, [r5] >+ pmaddubsw m1, m0, [r5 + mmsize] >+ paddw m4, m1 >+ pmaddubsw m0, [r5] >+ >+%ifidn %1,pp >+ pmulhrsw m4, m7 ; m4 = word: row 10 >+ pmulhrsw m2, m7 ; m2 = word: row 11 >+ packuswb m4, m2 >+ vpermq m4, m4, 11011000b >+ vextracti128 xm2, m4, 1 >+ movu [r2 + r3 * 2], xm4 >+ movu [r2 + r6], xm2 >+%else >+ psubw m4, m7 ; m4 = word: row 10 >+ psubw m2, m7 ; m2 = word: row 11 >+ movu [r2 + r3 * 2], m4 >+ movu [r2 + r6], m2 >+%endif >+ lea r2, [r2 + r3 * 4] >+ >+ movu xm3, [r0 + r4] >+ lea r0, [r0 + r1 * 4] >+ vinserti128 m3, m3, [r0], 1 >+ vinserti128 m5, m5, xm3, 1 >+ >+ punpcklbw m2, m5, m3 >+ punpckhbw m1, m5, m3 >+ vperm2i128 m4, m2, m1, 0x20 >+ vperm2i128 m2, m2, m1, 0x31 >+ pmaddubsw m1, m4, [r5 + mmsize] >+ paddw m0, m1 >+ pmaddubsw m4, [r5] >+ pmaddubsw m1, m2, [r5 + mmsize] >+ paddw m6, m1 >+ pmaddubsw m2, [r5] >+ >+%ifidn %1,pp >+ pmulhrsw m0, m7 ; m0 = word: row 12 >+ pmulhrsw m6, m7 ; m6 = word: row 13 >+ packuswb m0, m6 >+ vpermq m0, m0, 11011000b >+ vextracti128 xm6, m0, 1 >+ movu [r2], xm0 >+ movu [r2 + r3], xm6 >+%else >+ psubw m0, m7 ; m0 = word: row 12 >+ psubw m6, m7 ; m6 = word: row 13 >+ movu [r2], m0 >+ movu [r2 + r3], m6 >+%endif >+ >+ movu xm5, [r0 + r1 * 2] >+ vinserti128 m5, m5, [r0 + r1], 1 >+ vextracti128 xm0, m5, 1 >+ vinserti128 m3, m3, xm0, 0 >+ >+ punpcklbw m1, m3, m5 >+ punpckhbw m0, m3, m5 >+ vperm2i128 m6, m1, m0, 0x20 >+ vperm2i128 m0, m1, m0, 0x31 >+ pmaddubsw m6, [r5 + mmsize] >+ paddw m2, m6 >+ pmaddubsw m0, [r5 + mmsize] >+ paddw m4, m0 >+ >+%ifidn %1,pp >+ pmulhrsw m4, m7 ; m4 = word: row 14 >+ pmulhrsw m2, m7 ; m2 = word: row 15 >+ packuswb m4, m2 >+ vpermq m4, m4, 11011000b >+ vextracti128 xm2, m4, 1 >+ movu [r2 + r3 * 2], xm4 >+ movu [r2 + r6], xm2 >+%else >+ psubw m4, m7 ; m4 = word: row 14 >+ psubw m2, m7 ; m2 = word: row 15 >+ movu [r2 + r3 * 2], m4 >+ movu [r2 + r6], m2 >+%endif >+ lea r2, [r2 + r3 * 4] >+ dec r7d >+ jnz .loopH >+ RET >+%endif >+%endmacro >+ >+FILTER_VER_CHROMA_AVX2_16x32 pp >+FILTER_VER_CHROMA_AVX2_16x32 ps >+ > ;----------------------------------------------------------------------------- > ;void interp_4tap_vert_pp_24x32(pixel *src, intptr_t srcStride, pixel *dst, > intptr_t dstStride, int coeffIdx) > ;----------------------------------------------------------------------------- >_______________________________________________ >x265-devel mailing list >[email protected] >https://mailman.videolan.org/listinfo/x265-devel
_______________________________________________ x265-devel mailing list [email protected] https://mailman.videolan.org/listinfo/x265-devel
