ref: 1abd783be7ba902290e0f906eb35679c3ece419b
dir: /src/x86/mc.asm/
; Copyright © 2018, VideoLAN and dav1d authors ; Copyright © 2018, Two Orioles, LLC ; All rights reserved. ; ; Redistribution and use in source and binary forms, with or without ; modification, are permitted provided that the following conditions are met: ; ; 1. Redistributions of source code must retain the above copyright notice, this ; list of conditions and the following disclaimer. ; ; 2. Redistributions in binary form must reproduce the above copyright notice, ; this list of conditions and the following disclaimer in the documentation ; and/or other materials provided with the distribution. ; ; THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ; ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED ; WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE ; DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ; ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES ; (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; ; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ; ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT ; (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS ; SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. %include "config.asm" %include "ext/x86/x86inc.asm" %if ARCH_X86_64 && UNIX64 ; FIXME: Windows SECTION_RODATA 32 subpel_h_shuf4: db 0, 1, 2, 3, 1, 2, 3, 4, 8, 9, 10, 11, 9, 10, 11, 12 db 2, 3, 4, 5, 3, 4, 5, 6, 10, 11, 12, 13, 11, 12, 13, 14 subpel_h_shufA: db 0, 1, 2, 3, 1, 2, 3, 4, 2, 3, 4, 5, 3, 4, 5, 6 subpel_h_shufB: db 4, 5, 6, 7, 5, 6, 7, 8, 6, 7, 8, 9, 7, 8, 9, 10 subpel_h_shufC: db 8, 9, 10, 11, 9, 10, 11, 12, 10, 11, 12, 13, 11, 12, 13, 14 bilin_h_shuf4: db 1, 0, 2, 1, 3, 2, 4, 3, 9, 8, 10, 9, 11, 10, 12, 11 bilin_h_shuf8: db 1, 0, 2, 1, 3, 2, 4, 3, 5, 4, 6, 5, 7, 6, 8, 7 deint_shuf4: db 0, 4, 1, 5, 2, 6, 3, 7, 4, 8, 5, 9, 6, 10, 7, 11 pw_8: times 2 dw 8 pw_26: times 2 dw 26 pw_34: times 2 dw 34 pw_258: times 2 dw 258 pw_512: times 2 dw 512 pw_1024: times 2 dw 1024 pw_2048: times 2 dw 2048 pw_8192: times 2 dw 8192 pd_32: dd 32 pd_512: dd 512 cextern mc_subpel_filters %define subpel_filters (mangle(private_prefix %+ _mc_subpel_filters)-8) %macro BIDIR_JMP_TABLE 1-7 4, 8, 16, 32, 64, 128 %xdefine %1_table (%%table - 2*4) %xdefine %%prefix mangle(private_prefix %+ _%1) %%table: %rep 6 dd %%prefix %+ .w%2 - (%%table - 2*4) %rotate 1 %endrep %endmacro BIDIR_JMP_TABLE avg_avx2 BIDIR_JMP_TABLE w_avg_avx2 BIDIR_JMP_TABLE mask_avx2 BIDIR_JMP_TABLE w_mask_420_avx2 %macro BASE_JMP_TABLE 3-* %xdefine %1_%2_table (%%table - %3) %xdefine %%base %1_%2 %%table: %rep %0 - 2 dw %%base %+ _w%3 - %%base %rotate 1 %endrep %endmacro %xdefine put_avx2 mangle(private_prefix %+ _put_bilin_avx2.put) %xdefine prep_avx2 mangle(private_prefix %+ _prep_bilin_avx2.prep) BASE_JMP_TABLE put, avx2, 2, 4, 8, 16, 32, 64, 128 BASE_JMP_TABLE prep, avx2, 4, 8, 16, 32, 64, 128 %macro HV_JMP_TABLE 5-* %xdefine %%prefix mangle(private_prefix %+ _%1_%2_%3) %xdefine %%base %1_%3 %assign %%types %4 %if %%types & 1 %xdefine %1_%2_h_%3_table (%%h - %5) %%h: %rep %0 - 4 dw %%prefix %+ .h_w%5 - %%base %rotate 1 %endrep %rotate 4 %endif %if %%types & 2 %xdefine %1_%2_v_%3_table (%%v - %5) %%v: %rep %0 - 4 dw %%prefix %+ .v_w%5 - %%base %rotate 1 %endrep %rotate 4 %endif %if %%types & 4 %xdefine %1_%2_hv_%3_table (%%hv - %5) %%hv: %rep %0 - 4 dw %%prefix %+ .hv_w%5 - %%base %rotate 1 %endrep %endif %endmacro HV_JMP_TABLE put, 8tap, avx2, 3, 2, 4, 8, 16, 32, 64, 128 HV_JMP_TABLE prep, 8tap, avx2, 1, 4, 8, 16, 32, 64, 128 HV_JMP_TABLE put, bilin, avx2, 7, 2, 4, 8, 16, 32, 64, 128 HV_JMP_TABLE prep, bilin, avx2, 7, 4, 8, 16, 32, 64, 128 %define table_offset(type, fn) type %+ fn %+ SUFFIX %+ _table - type %+ SUFFIX SECTION .text INIT_XMM avx2 DECLARE_REG_TMP 4, 6, 7 cglobal put_bilin, 4, 8, 8, dst, ds, src, ss, w, h, mxy movifnidn mxyd, r6m ; mx lea t2, [put_avx2] tzcnt wd, wm movifnidn hd, hm test mxyd, mxyd jnz .h mov mxyd, r7m ; my test mxyd, mxyd jnz .v .put: movzx wd, word [t2+wq*2+table_offset(put,)] add wq, t2 lea t1, [ssq*3] lea t2, [dsq*3] jmp wq .put_w2: movzx t0d, word [srcq+ssq*0] movzx t1d, word [srcq+ssq*1] lea srcq, [srcq+ssq*2] mov [dstq+dsq*0], t0w mov [dstq+dsq*1], t1w lea dstq, [dstq+dsq*2] sub hd, 2 jg .put_w2 RET .put_w4: mov t0d, [srcq+ssq*0] mov t1d, [srcq+ssq*1] lea srcq, [srcq+ssq*2] mov [dstq+dsq*0], t0d mov [dstq+dsq*1], t1d lea dstq, [dstq+dsq*2] sub hd, 2 jg .put_w4 RET .put_w8: movq m0, [srcq+ssq*0] movq m1, [srcq+ssq*1] lea srcq, [srcq+ssq*2] movq [dstq+dsq*0], m0 movq [dstq+dsq*1], m1 lea dstq, [dstq+dsq*2] sub hd, 2 jg .put_w8 RET .put_w16: movu m0, [srcq+ssq*0] movu m1, [srcq+ssq*1] movu m2, [srcq+ssq*2] movu m3, [srcq+t1 ] lea srcq, [srcq+ssq*4] mova [dstq+dsq*0], m0 mova [dstq+dsq*1], m1 mova [dstq+dsq*2], m2 mova [dstq+t2 ], m3 lea dstq, [dstq+dsq*4] sub hd, 4 jg .put_w16 RET INIT_YMM avx2 .put_w32: movu m0, [srcq+ssq*0] movu m1, [srcq+ssq*1] movu m2, [srcq+ssq*2] movu m3, [srcq+t1 ] lea srcq, [srcq+ssq*4] mova [dstq+dsq*0], m0 mova [dstq+dsq*1], m1 mova [dstq+dsq*2], m2 mova [dstq+t2 ], m3 lea dstq, [dstq+dsq*4] sub hd, 4 jg .put_w32 RET .put_w64: movu m0, [srcq+ssq*0+32*0] movu m1, [srcq+ssq*0+32*1] movu m2, [srcq+ssq*1+32*0] movu m3, [srcq+ssq*1+32*1] lea srcq, [srcq+ssq*2] mova [dstq+dsq*0+32*0], m0 mova [dstq+dsq*0+32*1], m1 mova [dstq+dsq*1+32*0], m2 mova [dstq+dsq*1+32*1], m3 lea dstq, [dstq+dsq*2] sub hd, 2 jg .put_w64 RET .put_w128: movu m0, [srcq+32*0] movu m1, [srcq+32*1] movu m2, [srcq+32*2] movu m3, [srcq+32*3] add srcq, ssq mova [dstq+32*0], m0 mova [dstq+32*1], m1 mova [dstq+32*2], m2 mova [dstq+32*3], m3 add dstq, dsq dec hd jg .put_w128 RET .h: ; (16 * src[x] + (mx * (src[x + 1] - src[x])) + 8) >> 4 ; = ((16 - mx) * src[x] + mx * src[x + 1] + 8) >> 4 imul mxyd, 0xff01 vbroadcasti128 m4, [bilin_h_shuf8] add mxyd, 16 << 8 movd xm5, mxyd mov mxyd, r7m ; my vpbroadcastw m5, xm5 test mxyd, mxyd jnz .hv movzx wd, word [t2+wq*2+table_offset(put, _bilin_h)] vpbroadcastd m6, [pw_2048] add wq, t2 jmp wq .h_w2: movd xm0, [srcq+ssq*0] pinsrd xm0, [srcq+ssq*1], 1 lea srcq, [srcq+ssq*2] pshufb xm0, xm4 pmaddubsw xm0, xm5 pmulhrsw xm0, xm6 packuswb xm0, xm0 pextrw [dstq+dsq*0], xm0, 0 pextrw [dstq+dsq*1], xm0, 2 lea dstq, [dstq+dsq*2] sub hd, 2 jg .h_w2 RET .h_w4: mova xm4, [bilin_h_shuf4] .h_w4_loop: movq xm0, [srcq+ssq*0] movhps xm0, [srcq+ssq*1] lea srcq, [srcq+ssq*2] pshufb xm0, xm4 pmaddubsw xm0, xm5 pmulhrsw xm0, xm6 packuswb xm0, xm0 movd [dstq+dsq*0], xm0 pextrd [dstq+dsq*1], xm0, 1 lea dstq, [dstq+dsq*2] sub hd, 2 jg .h_w4_loop RET .h_w8: movu xm0, [srcq+ssq*0] movu xm1, [srcq+ssq*1] lea srcq, [srcq+ssq*2] pshufb xm0, xm4 pshufb xm1, xm4 pmaddubsw xm0, xm5 pmaddubsw xm1, xm5 pmulhrsw xm0, xm6 pmulhrsw xm1, xm6 packuswb xm0, xm1 movq [dstq+dsq*0], xm0 movhps [dstq+dsq*1], xm0 lea dstq, [dstq+dsq*2] sub hd, 2 jg .h_w8 RET .h_w16: movu xm0, [srcq+ssq*0+8*0] vinserti128 m0, m0, [srcq+ssq*1+8*0], 1 movu xm1, [srcq+ssq*0+8*1] vinserti128 m1, m1, [srcq+ssq*1+8*1], 1 lea srcq, [srcq+ssq*2] pshufb m0, m4 pshufb m1, m4 pmaddubsw m0, m5 pmaddubsw m1, m5 pmulhrsw m0, m6 pmulhrsw m1, m6 packuswb m0, m1 mova [dstq+dsq*0], xm0 vextracti128 [dstq+dsq*1], m0, 1 lea dstq, [dstq+dsq*2] sub hd, 2 jg .h_w16 RET .h_w32: movu m0, [srcq+8*0] movu m1, [srcq+8*1] add srcq, ssq pshufb m0, m4 pshufb m1, m4 pmaddubsw m0, m5 pmaddubsw m1, m5 pmulhrsw m0, m6 pmulhrsw m1, m6 packuswb m0, m1 mova [dstq], m0 add dstq, dsq dec hd jg .h_w32 RET .h_w64: movu m0, [srcq+8*0] movu m1, [srcq+8*1] movu m2, [srcq+8*4] movu m3, [srcq+8*5] add srcq, ssq pshufb m0, m4 pshufb m1, m4 pshufb m2, m4 pshufb m3, m4 pmaddubsw m0, m5 pmaddubsw m1, m5 pmaddubsw m2, m5 pmaddubsw m3, m5 pmulhrsw m0, m6 pmulhrsw m1, m6 pmulhrsw m2, m6 pmulhrsw m3, m6 packuswb m0, m1 packuswb m2, m3 mova [dstq+32*0], m0 mova [dstq+32*1], m2 add dstq, dsq dec hd jg .h_w64 RET .h_w128: mov t1, -32*3 .h_w128_loop: movu m0, [srcq+t1+32*3+8*0] movu m1, [srcq+t1+32*3+8*1] pshufb m0, m4 pshufb m1, m4 pmaddubsw m0, m5 pmaddubsw m1, m5 pmulhrsw m0, m6 pmulhrsw m1, m6 packuswb m0, m1 mova [dstq+t1+32*3], m0 add t1, 32 jle .h_w128_loop add srcq, ssq add dstq, dsq dec hd jg .h_w128 RET .v: movzx wd, word [t2+wq*2+table_offset(put, _bilin_v)] imul mxyd, 0xff01 vpbroadcastd m7, [pw_2048] add mxyd, 16 << 8 add wq, t2 movd xm6, mxyd vpbroadcastw m6, xm6 jmp wq .v_w2: movd xm0, [srcq+ssq*0] .v_w2_loop: pinsrw xm1, xm0, [srcq+ssq*1], 1 ; 0 1 lea srcq, [srcq+ssq*2] pinsrw xm0, xm1, [srcq+ssq*0], 0 ; 2 1 pshuflw xm1, xm1, q2301 ; 1 0 punpcklbw xm1, xm0, xm1 pmaddubsw xm1, xm6 pmulhrsw xm1, xm7 packuswb xm1, xm1 pextrw [dstq+dsq*0], xm1, 1 pextrw [dstq+dsq*1], xm1, 0 lea dstq, [dstq+dsq*2] sub hd, 2 jg .v_w2_loop RET .v_w4: movd xm0, [srcq+ssq*0] .v_w4_loop: vpbroadcastd xm1, [srcq+ssq*1] lea srcq, [srcq+ssq*2] vpblendd xm2, xm1, xm0, 0x01 ; 0 1 vpbroadcastd xm0, [srcq+ssq*0] vpblendd xm1, xm1, xm0, 0x02 ; 1 2 punpcklbw xm1, xm2 pmaddubsw xm1, xm6 pmulhrsw xm1, xm7 packuswb xm1, xm1 movd [dstq+dsq*0], xm1 pextrd [dstq+dsq*1], xm1, 1 lea dstq, [dstq+dsq*2] sub hd, 2 jg .v_w4_loop RET .v_w8: movq xm0, [srcq+ssq*0] .v_w8_loop: vpbroadcastq xm1, [srcq+ssq*1] lea srcq, [srcq+ssq*2] vpblendd xm2, xm1, xm0, 0x03 ; 0 1 vpbroadcastq xm0, [srcq+ssq*0] vpblendd xm1, xm1, xm0, 0x0c ; 1 2 punpcklbw xm3, xm1, xm2 punpckhbw xm1, xm2 pmaddubsw xm3, xm6 pmaddubsw xm1, xm6 pmulhrsw xm3, xm7 pmulhrsw xm1, xm7 packuswb xm3, xm1 movq [dstq+dsq*0], xm3 movhps [dstq+dsq*1], xm3 lea dstq, [dstq+dsq*2] sub hd, 2 jg .v_w8_loop RET .v_w16: movu xm0, [srcq+ssq*0] .v_w16_loop: vbroadcasti128 m2, [srcq+ssq*1] lea srcq, [srcq+ssq*2] vpblendd m3, m2, m0, 0x0f ; 0 1 vbroadcasti128 m0, [srcq+ssq*0] vpblendd m2, m2, m0, 0xf0 ; 1 2 punpcklbw m1, m2, m3 punpckhbw m2, m3 pmaddubsw m1, m6 pmaddubsw m2, m6 pmulhrsw m1, m7 pmulhrsw m2, m7 packuswb m1, m2 mova [dstq+dsq*0], xm1 vextracti128 [dstq+dsq*1], m1, 1 lea dstq, [dstq+dsq*2] sub hd, 2 jg .v_w16_loop RET .v_w32: %macro PUT_BILIN_V_W32 0 movu m0, [srcq+ssq*0] %%loop: movu m4, [srcq+ssq*1] lea srcq, [srcq+ssq*2] punpcklbw m1, m4, m0 punpckhbw m3, m4, m0 movu m0, [srcq+ssq*0] punpcklbw m2, m0, m4 punpckhbw m4, m0, m4 pmaddubsw m1, m6 pmaddubsw m3, m6 pmaddubsw m2, m6 pmaddubsw m4, m6 pmulhrsw m1, m7 pmulhrsw m3, m7 pmulhrsw m2, m7 pmulhrsw m4, m7 packuswb m1, m3 packuswb m2, m4 mova [dstq+dsq*0], m1 mova [dstq+dsq*1], m2 lea dstq, [dstq+dsq*2] sub hd, 2 jg %%loop %endmacro PUT_BILIN_V_W32 RET .v_w64: movu m0, [srcq+32*0] movu m1, [srcq+32*1] .v_w64_loop: add srcq, ssq movu m3, [srcq+32*0] movu m4, [srcq+32*1] punpcklbw m2, m3, m0 punpckhbw m5, m3, m0 pmaddubsw m2, m6 pmaddubsw m5, m6 mova m0, m3 pmulhrsw m2, m7 pmulhrsw m5, m7 packuswb m2, m5 punpcklbw m3, m4, m1 punpckhbw m5, m4, m1 pmaddubsw m3, m6 pmaddubsw m5, m6 mova m1, m4 pmulhrsw m3, m7 pmulhrsw m5, m7 packuswb m3, m5 mova [dstq+32*0], m2 mova [dstq+32*1], m3 add dstq, dsq dec hd jg .v_w64_loop RET .v_w128: mov t0, dstq mov t1, srcq lea t2d, [hq+(3<<8)] .v_w128_loop: PUT_BILIN_V_W32 mov hb, t2b add t0, 32 add t1, 32 mov dstq, t0 mov srcq, t1 sub t2d, 1<<8 jg .v_w128_loop RET .hv: ; (16 * src[x] + (my * (src[x + src_stride] - src[x])) + 128) >> 8 ; = (src[x] + ((my * (src[x + src_stride] - src[x])) >> 4) + 8) >> 4 movzx wd, word [t2+wq*2+table_offset(put, _bilin_hv)] shl mxyd, 11 ; can't shift by 12 due to signed overflow vpbroadcastd m7, [pw_2048] movd xm6, mxyd add wq, t2 vpbroadcastw m6, xm6 jmp wq .hv_w2: vpbroadcastd xm0, [srcq+ssq*0] pshufb xm0, xm4 pmaddubsw xm0, xm5 .hv_w2_loop: movd xm1, [srcq+ssq*1] lea srcq, [srcq+ssq*2] pinsrd xm1, [srcq+ssq*0], 1 pshufb xm1, xm4 pmaddubsw xm1, xm5 ; 1 _ 2 _ shufps xm2, xm0, xm1, q1032 ; 0 _ 1 _ mova xm0, xm1 psubw xm1, xm2 paddw xm1, xm1 pmulhw xm1, xm6 paddw xm1, xm2 pmulhrsw xm1, xm7 packuswb xm1, xm1 pextrw [dstq+dsq*0], xm1, 0 pextrw [dstq+dsq*1], xm1, 2 lea dstq, [dstq+dsq*2] sub hd, 2 jg .hv_w2_loop RET .hv_w4: mova xm4, [bilin_h_shuf4] movddup xm0, [srcq+ssq*0] pshufb xm0, xm4 pmaddubsw xm0, xm5 .hv_w4_loop: movq xm1, [srcq+ssq*1] lea srcq, [srcq+ssq*2] movhps xm1, [srcq+ssq*0] pshufb xm1, xm4 pmaddubsw xm1, xm5 ; 1 2 shufps xm2, xm0, xm1, q1032 ; 0 1 mova xm0, xm1 psubw xm1, xm2 paddw xm1, xm1 pmulhw xm1, xm6 paddw xm1, xm2 pmulhrsw xm1, xm7 packuswb xm1, xm1 movd [dstq+dsq*0], xm1 pextrd [dstq+dsq*1], xm1, 1 lea dstq, [dstq+dsq*2] sub hd, 2 jg .hv_w4_loop RET .hv_w8: vbroadcasti128 m0, [srcq+ssq*0] pshufb m0, m4 pmaddubsw m0, m5 .hv_w8_loop: movu xm1, [srcq+ssq*1] lea srcq, [srcq+ssq*2] vinserti128 m1, m1, [srcq+ssq*0], 1 pshufb m1, m4 pmaddubsw m1, m5 ; 1 2 vperm2i128 m2, m0, m1, 0x21 ; 0 1 mova m0, m1 psubw m1, m2 paddw m1, m1 pmulhw m1, m6 paddw m1, m2 pmulhrsw m1, m7 vextracti128 xm2, m1, 1 packuswb xm1, xm2 movq [dstq+dsq*0], xm1 movhps [dstq+dsq*1], xm1 lea dstq, [dstq+dsq*2] sub hd, 2 jg .hv_w8_loop RET .hv_w16: movu m0, [srcq+ssq*0+8*0] vinserti128 m0, m0, [srcq+ssq*0+8*1], 1 pshufb m0, m4 pmaddubsw m0, m5 .hv_w16_loop: movu xm2, [srcq+ssq*1+8*0] vinserti128 m2, m2, [srcq+ssq*1+8*1], 1 lea srcq, [srcq+ssq*2] movu xm3, [srcq+ssq*0+8*0] vinserti128 m3, m3, [srcq+ssq*0+8*1], 1 pshufb m2, m4 pshufb m3, m4 pmaddubsw m2, m5 psubw m1, m2, m0 paddw m1, m1 pmulhw m1, m6 paddw m1, m0 pmaddubsw m0, m3, m5 psubw m3, m0, m2 paddw m3, m3 pmulhw m3, m6 paddw m3, m2 pmulhrsw m1, m7 pmulhrsw m3, m7 packuswb m1, m3 vpermq m1, m1, q3120 mova [dstq+dsq*0], xm1 vextracti128 [dstq+dsq*1], m1, 1 lea dstq, [dstq+dsq*2] sub hd, 2 jg .hv_w16_loop RET .hv_w32: %macro PUT_BILIN_HV_W32 0 movu m0, [srcq+8*0] vinserti128 m0, m0, [srcq+8*2], 1 movu m1, [srcq+8*1] vinserti128 m1, m1, [srcq+8*3], 1 pshufb m0, m4 pshufb m1, m4 pmaddubsw m0, m5 pmaddubsw m1, m5 %%loop: add srcq, ssq movu xm2, [srcq+8*1] vinserti128 m2, m2, [srcq+8*3], 1 pshufb m2, m4 pmaddubsw m2, m5 psubw m3, m2, m1 paddw m3, m3 pmulhw m3, m6 paddw m3, m1 mova m1, m2 pmulhrsw m8, m3, m7 ASSERT UNIX64 ; using an additional vector register here movu xm2, [srcq+8*0] vinserti128 m2, m2, [srcq+8*2], 1 pshufb m2, m4 pmaddubsw m2, m5 psubw m3, m2, m0 paddw m3, m3 pmulhw m3, m6 paddw m3, m0 mova m0, m2 pmulhrsw m3, m7 packuswb m3, m8 mova [dstq], m3 add dstq, dsq dec hd jg %%loop %endmacro PUT_BILIN_HV_W32 RET .hv_w64: mov t0, dstq mov t1, srcq lea t2d, [hq+(1<<8)] .hv_w64_loop: PUT_BILIN_HV_W32 mov hb, t2b add t0, 32 add t1, 32 mov dstq, t0 mov srcq, t1 sub t2d, 1<<8 jg .hv_w64_loop RET .hv_w128: mov t0, dstq mov t1, srcq lea t2d, [hq+(3<<8)] .hv_w128_loop: PUT_BILIN_HV_W32 mov hb, t2b add t0, 32 add t1, 32 mov dstq, t0 mov srcq, t1 sub t2d, 1<<8 jg .hv_w128_loop RET DECLARE_REG_TMP 3, 5, 6 cglobal prep_bilin, 3, 7, 7, tmp, src, stride, w, h, mxy, stride3 movifnidn mxyd, r5m ; mx lea t2, [prep_avx2] tzcnt wd, wm movifnidn hd, hm test mxyd, mxyd jnz .h mov mxyd, r6m ; my test mxyd, mxyd jnz .v .prep: movzx wd, word [t2+wq*2+table_offset(prep,)] add wq, t2 lea stride3q, [strideq*3] jmp wq .prep_w4: movd xm0, [srcq+strideq*0] pinsrd xm0, [srcq+strideq*1], 1 pinsrd xm0, [srcq+strideq*2], 2 pinsrd xm0, [srcq+stride3q ], 3 lea srcq, [srcq+strideq*4] pmovzxbw m0, xm0 psllw m0, 4 mova [tmpq], m0 add tmpq, 32 sub hd, 4 jg .prep_w4 RET .prep_w8: movq xm0, [srcq+strideq*0] movhps xm0, [srcq+strideq*1] movq xm1, [srcq+strideq*2] movhps xm1, [srcq+stride3q ] lea srcq, [srcq+strideq*4] pmovzxbw m0, xm0 pmovzxbw m1, xm1 psllw m0, 4 psllw m1, 4 mova [tmpq+32*0], m0 mova [tmpq+32*1], m1 add tmpq, 32*2 sub hd, 4 jg .prep_w8 RET .prep_w16: pmovzxbw m0, [srcq+strideq*0] pmovzxbw m1, [srcq+strideq*1] pmovzxbw m2, [srcq+strideq*2] pmovzxbw m3, [srcq+stride3q ] lea srcq, [srcq+strideq*4] psllw m0, 4 psllw m1, 4 psllw m2, 4 psllw m3, 4 mova [tmpq+32*0], m0 mova [tmpq+32*1], m1 mova [tmpq+32*2], m2 mova [tmpq+32*3], m3 add tmpq, 32*4 sub hd, 4 jg .prep_w16 RET .prep_w32: pmovzxbw m0, [srcq+strideq*0+16*0] pmovzxbw m1, [srcq+strideq*0+16*1] pmovzxbw m2, [srcq+strideq*1+16*0] pmovzxbw m3, [srcq+strideq*1+16*1] lea srcq, [srcq+strideq*2] psllw m0, 4 psllw m1, 4 psllw m2, 4 psllw m3, 4 mova [tmpq+32*0], m0 mova [tmpq+32*1], m1 mova [tmpq+32*2], m2 mova [tmpq+32*3], m3 add tmpq, 32*4 sub hd, 2 jg .prep_w32 RET .prep_w64: pmovzxbw m0, [srcq+16*0] pmovzxbw m1, [srcq+16*1] pmovzxbw m2, [srcq+16*2] pmovzxbw m3, [srcq+16*3] add srcq, strideq psllw m0, 4 psllw m1, 4 psllw m2, 4 psllw m3, 4 mova [tmpq+32*0], m0 mova [tmpq+32*1], m1 mova [tmpq+32*2], m2 mova [tmpq+32*3], m3 add tmpq, 32*4 dec hd jg .prep_w64 RET .prep_w128: pmovzxbw m0, [srcq+16*0] pmovzxbw m1, [srcq+16*1] pmovzxbw m2, [srcq+16*2] pmovzxbw m3, [srcq+16*3] psllw m0, 4 psllw m1, 4 psllw m2, 4 psllw m3, 4 mova [tmpq+32*0], m0 mova [tmpq+32*1], m1 mova [tmpq+32*2], m2 mova [tmpq+32*3], m3 pmovzxbw m0, [srcq+16*4] pmovzxbw m1, [srcq+16*5] pmovzxbw m2, [srcq+16*6] pmovzxbw m3, [srcq+16*7] add tmpq, 32*8 add srcq, strideq psllw m0, 4 psllw m1, 4 psllw m2, 4 psllw m3, 4 mova [tmpq-32*4], m0 mova [tmpq-32*3], m1 mova [tmpq-32*2], m2 mova [tmpq-32*1], m3 dec hd jg .prep_w128 RET .h: ; 16 * src[x] + (mx * (src[x + 1] - src[x])) ; = (16 - mx) * src[x] + mx * src[x + 1] imul mxyd, 0xff01 vbroadcasti128 m4, [bilin_h_shuf8] add mxyd, 16 << 8 movd xm5, mxyd mov mxyd, r6m ; my vpbroadcastw m5, xm5 test mxyd, mxyd jnz .hv movzx wd, word [t2+wq*2+table_offset(prep, _bilin_h)] add wq, t2 lea stride3q, [strideq*3] jmp wq .h_w4: vbroadcasti128 m4, [bilin_h_shuf4] .h_w4_loop: movq xm0, [srcq+strideq*0] movhps xm0, [srcq+strideq*1] movq xm1, [srcq+strideq*2] movhps xm1, [srcq+stride3q ] lea srcq, [srcq+strideq*4] vinserti128 m0, m0, xm1, 1 pshufb m0, m4 pmaddubsw m0, m5 mova [tmpq], m0 add tmpq, 32 sub hd, 4 jg .h_w4_loop RET .h_w8: movu xm0, [srcq+strideq*0] vinserti128 m0, m0, [srcq+strideq*1], 1 movu xm1, [srcq+strideq*2] vinserti128 m1, m1, [srcq+stride3q ], 1 lea srcq, [srcq+strideq*4] pshufb m0, m4 pshufb m1, m4 pmaddubsw m0, m5 pmaddubsw m1, m5 mova [tmpq+32*0], m0 mova [tmpq+32*1], m1 add tmpq, 32*2 sub hd, 4 jg .h_w8 RET .h_w16: movu xm0, [srcq+strideq*0+8*0] vinserti128 m0, m0, [srcq+strideq*0+8*1], 1 movu xm1, [srcq+strideq*1+8*0] vinserti128 m1, m1, [srcq+strideq*1+8*1], 1 movu xm2, [srcq+strideq*2+8*0] vinserti128 m2, m2, [srcq+strideq*2+8*1], 1 movu xm3, [srcq+stride3q +8*0] vinserti128 m3, m3, [srcq+stride3q +8*1], 1 lea srcq, [srcq+strideq*4] pshufb m0, m4 pshufb m1, m4 pshufb m2, m4 pshufb m3, m4 pmaddubsw m0, m5 pmaddubsw m1, m5 pmaddubsw m2, m5 pmaddubsw m3, m5 mova [tmpq+32*0], m0 mova [tmpq+32*1], m1 mova [tmpq+32*2], m2 mova [tmpq+32*3], m3 add tmpq, 32*4 sub hd, 4 jg .h_w16 RET .h_w32: movu xm0, [srcq+strideq*0+8*0] vinserti128 m0, m0, [srcq+strideq*0+8*1], 1 movu xm1, [srcq+strideq*0+8*2] vinserti128 m1, m1, [srcq+strideq*0+8*3], 1 movu xm2, [srcq+strideq*1+8*0] vinserti128 m2, m2, [srcq+strideq*1+8*1], 1 movu xm3, [srcq+strideq*1+8*2] vinserti128 m3, m3, [srcq+strideq*1+8*3], 1 lea srcq, [srcq+strideq*2] pshufb m0, m4 pshufb m1, m4 pshufb m2, m4 pshufb m3, m4 pmaddubsw m0, m5 pmaddubsw m1, m5 pmaddubsw m2, m5 pmaddubsw m3, m5 mova [tmpq+32*0], m0 mova [tmpq+32*1], m1 mova [tmpq+32*2], m2 mova [tmpq+32*3], m3 add tmpq, 32*4 sub hd, 2 jg .h_w32 RET .h_w64: movu xm0, [srcq+8*0] vinserti128 m0, m0, [srcq+8*1], 1 movu xm1, [srcq+8*2] vinserti128 m1, m1, [srcq+8*3], 1 movu xm2, [srcq+8*4] vinserti128 m2, m2, [srcq+8*5], 1 movu xm3, [srcq+8*6] vinserti128 m3, m3, [srcq+8*7], 1 add srcq, strideq pshufb m0, m4 pshufb m1, m4 pshufb m2, m4 pshufb m3, m4 pmaddubsw m0, m5 pmaddubsw m1, m5 pmaddubsw m2, m5 pmaddubsw m3, m5 mova [tmpq+32*0], m0 mova [tmpq+32*1], m1 mova [tmpq+32*2], m2 mova [tmpq+32*3], m3 add tmpq, 32*4 dec hd jg .h_w64 RET .h_w128: movu xm0, [srcq+8*0] vinserti128 m0, m0, [srcq+8*1], 1 movu xm1, [srcq+8*2] vinserti128 m1, m1, [srcq+8*3], 1 movu xm2, [srcq+8*4] vinserti128 m2, m2, [srcq+8*5], 1 movu xm3, [srcq+8*6] vinserti128 m3, m3, [srcq+8*7], 1 pshufb m0, m4 pshufb m1, m4 pshufb m2, m4 pshufb m3, m4 pmaddubsw m0, m5 pmaddubsw m1, m5 pmaddubsw m2, m5 pmaddubsw m3, m5 mova [tmpq+32*0], m0 mova [tmpq+32*1], m1 mova [tmpq+32*2], m2 mova [tmpq+32*3], m3 movu xm0, [srcq+8* 8] vinserti128 m0, m0, [srcq+8* 9], 1 movu xm1, [srcq+8*10] vinserti128 m1, m1, [srcq+8*11], 1 movu xm2, [srcq+8*12] vinserti128 m2, m2, [srcq+8*13], 1 movu xm3, [srcq+8*14] vinserti128 m3, m3, [srcq+8*15], 1 add tmpq, 32*8 add srcq, strideq pshufb m0, m4 pshufb m1, m4 pshufb m2, m4 pshufb m3, m4 pmaddubsw m0, m5 pmaddubsw m1, m5 pmaddubsw m2, m5 pmaddubsw m3, m5 mova [tmpq-32*4], m0 mova [tmpq-32*3], m1 mova [tmpq-32*2], m2 mova [tmpq-32*1], m3 dec hd jg .h_w128 RET .v: movzx wd, word [t2+wq*2+table_offset(prep, _bilin_v)] imul mxyd, 0xff01 add mxyd, 16 << 8 add wq, t2 lea stride3q, [strideq*3] movd xm6, mxyd vpbroadcastw m6, xm6 jmp wq .v_w4: movd xm0, [srcq+strideq*0] .v_w4_loop: vpbroadcastd m1, [srcq+strideq*2] vpbroadcastd xm2, [srcq+strideq*1] vpbroadcastd m3, [srcq+stride3q ] lea srcq, [srcq+strideq*4] vpblendd m1, m1, m0, 0x05 ; 0 2 2 2 vpbroadcastd m0, [srcq+strideq*0] vpblendd m3, m3, m2, 0x0f ; 1 1 3 3 vpblendd m2, m1, m0, 0xa0 ; 0 2 2 4 vpblendd m1, m1, m3, 0xaa ; 0 1 2 3 vpblendd m2, m2, m3, 0x55 ; 1 2 3 4 punpcklbw m2, m1 pmaddubsw m2, m6 mova [tmpq], m2 add tmpq, 32 sub hd, 4 jg .v_w4_loop RET .v_w8: movq xm0, [srcq+strideq*0] .v_w8_loop: vpbroadcastq m1, [srcq+strideq*2] vpbroadcastq m2, [srcq+strideq*1] vpbroadcastq m3, [srcq+stride3q ] lea srcq, [srcq+strideq*4] vpblendd m1, m1, m0, 0x03 ; 0 2 2 2 vpbroadcastq m0, [srcq+strideq*0] vpblendd m3, m3, m2, 0x33 ; 1 3 1 3 vpblendd m2, m1, m3, 0x0f ; 1 3 2 2 vpblendd m1, m1, m3, 0xf0 ; 0 2 1 3 vpblendd m2, m2, m0, 0xc0 ; 1 3 2 4 punpcklbw m3, m2, m1 punpckhbw m2, m1 pmaddubsw m3, m6 pmaddubsw m2, m6 mova [tmpq+32*0], m3 mova [tmpq+32*1], m2 add tmpq, 32*2 sub hd, 4 jg .v_w8_loop RET .v_w16: vbroadcasti128 m0, [srcq+strideq*0] .v_w16_loop: vbroadcasti128 m1, [srcq+strideq*2] vbroadcasti128 m2, [srcq+strideq*1] vbroadcasti128 m3, [srcq+stride3q ] lea srcq, [srcq+strideq*4] shufpd m4, m0, m1, 0x0c ; 0 2 vbroadcasti128 m0, [srcq+strideq*0] shufpd m2, m2, m3, 0x0c ; 1 3 shufpd m1, m1, m0, 0x0c ; 2 4 punpcklbw m3, m2, m4 punpcklbw m5, m1, m2 punpckhbw m1, m2 punpckhbw m2, m4 pmaddubsw m3, m6 pmaddubsw m5, m6 pmaddubsw m2, m6 pmaddubsw m1, m6 mova [tmpq+32*0], m3 mova [tmpq+32*1], m5 mova [tmpq+32*2], m2 mova [tmpq+32*3], m1 add tmpq, 32*4 sub hd, 4 jg .v_w16_loop RET .v_w32: vpermq m0, [srcq+strideq*0], q3120 .v_w32_loop: vpermq m1, [srcq+strideq*1], q3120 vpermq m2, [srcq+strideq*2], q3120 vpermq m3, [srcq+stride3q ], q3120 lea srcq, [srcq+strideq*4] punpcklbw m4, m1, m0 punpckhbw m5, m1, m0 vpermq m0, [srcq+strideq*0], q3120 pmaddubsw m4, m6 pmaddubsw m5, m6 mova [tmpq+32*0], m4 mova [tmpq+32*1], m5 punpcklbw m4, m2, m1 punpckhbw m5, m2, m1 pmaddubsw m4, m6 pmaddubsw m5, m6 mova [tmpq+32*2], m4 mova [tmpq+32*3], m5 add tmpq, 32*8 punpcklbw m4, m3, m2 punpckhbw m5, m3, m2 punpcklbw m1, m0, m3 punpckhbw m2, m0, m3 pmaddubsw m4, m6 pmaddubsw m5, m6 pmaddubsw m1, m6 pmaddubsw m2, m6 mova [tmpq-32*4], m4 mova [tmpq-32*3], m5 mova [tmpq-32*2], m1 mova [tmpq-32*1], m2 sub hd, 4 jg .v_w32_loop RET .v_w64: vpermq m0, [srcq+strideq*0+32*0], q3120 vpermq m1, [srcq+strideq*0+32*1], q3120 .v_w64_loop: vpermq m2, [srcq+strideq*1+32*0], q3120 vpermq m3, [srcq+strideq*1+32*1], q3120 lea srcq, [srcq+strideq*2] punpcklbw m4, m2, m0 punpckhbw m5, m2, m0 pmaddubsw m4, m6 pmaddubsw m5, m6 mova [tmpq+32*0], m4 mova [tmpq+32*1], m5 punpcklbw m4, m3, m1 punpckhbw m5, m3, m1 vpermq m0, [srcq+strideq*0+32*0], q3120 vpermq m1, [srcq+strideq*0+32*1], q3120 pmaddubsw m4, m6 pmaddubsw m5, m6 mova [tmpq+32*2], m4 mova [tmpq+32*3], m5 add tmpq, 32*8 punpcklbw m4, m0, m2 punpckhbw m5, m0, m2 punpcklbw m2, m1, m3 punpckhbw m3, m1, m3 pmaddubsw m4, m6 pmaddubsw m5, m6 pmaddubsw m2, m6 pmaddubsw m3, m6 mova [tmpq-32*4], m4 mova [tmpq-32*3], m5 mova [tmpq-32*2], m2 mova [tmpq-32*1], m3 sub hd, 2 jg .v_w64_loop RET .v_w128: mov t0, tmpq mov t1, srcq lea t2d, [hq+(3<<8)] .v_w128_loop0: vpermq m0, [srcq+strideq*0], q3120 .v_w128_loop: vpermq m1, [srcq+strideq*1], q3120 lea srcq, [srcq+strideq*2] punpcklbw m2, m1, m0 punpckhbw m3, m1, m0 vpermq m0, [srcq+strideq*0], q3120 punpcklbw m4, m0, m1 punpckhbw m5, m0, m1 pmaddubsw m2, m6 pmaddubsw m3, m6 pmaddubsw m4, m6 pmaddubsw m5, m6 mova [tmpq+32*0], m2 mova [tmpq+32*1], m3 mova [tmpq+32*8], m4 mova [tmpq+32*9], m5 add tmpq, 32*16 sub hd, 2 jg .v_w128_loop mov hb, t2b add t0, 64 add t1, 32 mov tmpq, t0 mov srcq, t1 sub t2d, 1<<8 jg .v_w128_loop0 RET .hv: ; (16 * src[x] + (my * (src[x + src_stride] - src[x])) + 8) >> 4 ; = src[x] + (((my * (src[x + src_stride] - src[x])) + 8) >> 4) movzx wd, word [t2+wq*2+table_offset(prep, _bilin_hv)] shl mxyd, 11 movd xm6, mxyd add wq, t2 lea stride3q, [strideq*3] vpbroadcastw m6, xm6 jmp wq .hv_w4: vbroadcasti128 m4, [bilin_h_shuf4] vpbroadcastq m0, [srcq+strideq*0] pshufb m0, m4 pmaddubsw m0, m5 .hv_w4_loop: movq xm1, [srcq+strideq*1] movhps xm1, [srcq+strideq*2] movq xm2, [srcq+stride3q ] lea srcq, [srcq+strideq*4] movhps xm2, [srcq+strideq*0] vinserti128 m1, m1, xm2, 1 pshufb m1, m4 pmaddubsw m1, m5 ; 1 2 3 4 vpblendd m2, m1, m0, 0xc0 vpermq m2, m2, q2103 ; 0 1 2 3 mova m0, m1 psubw m1, m2 pmulhrsw m1, m6 paddw m1, m2 mova [tmpq], m1 add tmpq, 32 sub hd, 4 jg .hv_w4_loop RET .hv_w8: vbroadcasti128 m0, [srcq+strideq*0] pshufb m0, m4 pmaddubsw m0, m5 .hv_w8_loop: movu xm1, [srcq+strideq*1] vinserti128 m1, m1, [srcq+strideq*2], 1 movu xm2, [srcq+stride3q ] lea srcq, [srcq+strideq*4] vinserti128 m2, m2, [srcq+strideq*0], 1 pshufb m1, m4 pshufb m2, m4 pmaddubsw m1, m5 ; 1 2 vperm2i128 m3, m0, m1, 0x21 ; 0 1 pmaddubsw m0, m2, m5 ; 3 4 vperm2i128 m2, m1, m0, 0x21 ; 2 3 psubw m1, m3 pmulhrsw m1, m6 paddw m1, m3 psubw m3, m0, m2 pmulhrsw m3, m6 paddw m3, m2 mova [tmpq+32*0], m1 mova [tmpq+32*1], m3 add tmpq, 32*2 sub hd, 4 jg .hv_w8_loop RET .hv_w16: movu m0, [srcq+strideq*0+8*0] vinserti128 m0, m0, [srcq+strideq*0+8*1], 1 pshufb m0, m4 pmaddubsw m0, m5 .hv_w16_loop: movu xm1, [srcq+strideq*1+8*0] vinserti128 m1, m1, [srcq+strideq*1+8*1], 1 lea srcq, [srcq+strideq*2] movu xm2, [srcq+strideq*0+8*0] vinserti128 m2, m2, [srcq+strideq*0+8*1], 1 pshufb m1, m4 pshufb m2, m4 pmaddubsw m1, m5 psubw m3, m1, m0 pmulhrsw m3, m6 paddw m3, m0 pmaddubsw m0, m2, m5 psubw m2, m0, m1 pmulhrsw m2, m6 paddw m2, m1 mova [tmpq+32*0], m3 mova [tmpq+32*1], m2 add tmpq, 32*2 sub hd, 2 jg .hv_w16_loop RET .hv_w32: movu m0, [srcq+8*0] vinserti128 m0, m0, [srcq+8*1], 1 movu m1, [srcq+8*2] vinserti128 m1, m1, [srcq+8*3], 1 pshufb m0, m4 pshufb m1, m4 pmaddubsw m0, m5 pmaddubsw m1, m5 .hv_w32_loop: add srcq, strideq movu xm2, [srcq+8*0] vinserti128 m2, m2, [srcq+8*1], 1 pshufb m2, m4 pmaddubsw m2, m5 psubw m3, m2, m0 pmulhrsw m3, m6 paddw m3, m0 mova m0, m2 mova [tmpq+ 0], m3 movu xm2, [srcq+8*2] vinserti128 m2, m2, [srcq+8*3], 1 pshufb m2, m4 pmaddubsw m2, m5 psubw m3, m2, m1 pmulhrsw m3, m6 paddw m3, m1 mova m1, m2 mova [tmpq+32], m3 add tmpq, 32*2 dec hd jg .hv_w32_loop RET .hv_w64: mov t0, tmpq mov t1, srcq lea t2d, [hq+(3<<8)] .hv_w64_loop0: movu m0, [srcq+strideq*0+8*0] vinserti128 m0, m0, [srcq+strideq*0+8*1], 1 pshufb m0, m4 pmaddubsw m0, m5 .hv_w64_loop: movu xm1, [srcq+strideq*1+8*0] vinserti128 m1, m1, [srcq+strideq*1+8*1], 1 lea srcq, [srcq+strideq*2] movu xm2, [srcq+strideq*0+8*0] vinserti128 m2, m2, [srcq+strideq*0+8*1], 1 pshufb m1, m4 pshufb m2, m4 pmaddubsw m1, m5 psubw m3, m1, m0 pmulhrsw m3, m6 paddw m3, m0 pmaddubsw m0, m2, m5 psubw m2, m0, m1 pmulhrsw m2, m6 paddw m2, m1 mova [tmpq+32*0], m3 add tmpq, 32*8 mova [tmpq-32*4], m2 sub hd, 2 jg .hv_w64_loop mov hb, t2b add t0, 32 add t1, 16 mov tmpq, t0 mov srcq, t1 sub t2d, 1<<8 jg .hv_w64_loop0 RET .hv_w128: mov t0, tmpq mov t1, srcq lea t2d, [hq+(7<<8)] .hv_w128_loop0: movu m0, [srcq+strideq*0+8*0] vinserti128 m0, m0, [srcq+strideq*0+8*1], 1 pshufb m0, m4 pmaddubsw m0, m5 .hv_w128_loop: movu xm1, [srcq+strideq*1+8*0] vinserti128 m1, m1, [srcq+strideq*1+8*1], 1 lea srcq, [srcq+strideq*2] movu xm2, [srcq+strideq*0+8*0] vinserti128 m2, m2, [srcq+strideq*0+8*1], 1 pshufb m1, m4 pshufb m2, m4 pmaddubsw m1, m5 psubw m3, m1, m0 pmulhrsw m3, m6 paddw m3, m0 pmaddubsw m0, m2, m5 psubw m2, m0, m1 pmulhrsw m2, m6 paddw m2, m1 mova [tmpq+32*0], m3 mova [tmpq+32*8], m2 add tmpq, 32*16 sub hd, 2 jg .hv_w128_loop mov hb, t2b add t0, 32 add t1, 16 mov tmpq, t0 mov srcq, t1 sub t2d, 1<<8 jg .hv_w128_loop0 RET ; int8_t subpel_filters[5][15][8] %assign FILTER_REGULAR (0*15 << 16) | 3*15 %assign FILTER_SMOOTH (1*15 << 16) | 4*15 %assign FILTER_SHARP (2*15 << 16) | 3*15 DECLARE_REG_TMP 7, 8 %macro PUT_8TAP_FN 3 ; type, type_h, type_v cglobal put_8tap_%1 mov t0d, FILTER_%2 mov t1d, FILTER_%3 %ifnidn %1, sharp_smooth ; skip the jump in the last filter jmp mangle(private_prefix %+ _put_8tap %+ SUFFIX) %endif %endmacro PUT_8TAP_FN regular, REGULAR, REGULAR PUT_8TAP_FN regular_sharp, REGULAR, SHARP PUT_8TAP_FN regular_smooth, REGULAR, SMOOTH PUT_8TAP_FN smooth_regular, SMOOTH, REGULAR PUT_8TAP_FN smooth, SMOOTH, SMOOTH PUT_8TAP_FN smooth_sharp, SMOOTH, SHARP PUT_8TAP_FN sharp_regular, SHARP, REGULAR PUT_8TAP_FN sharp, SHARP, SHARP PUT_8TAP_FN sharp_smooth, SHARP, SMOOTH cglobal put_8tap, 4, 9, 16, dst, ds, src, ss, w, h, mx, my, ss3 imul mxd, mxm, 0x010101 add mxd, t0d ; 8tap_h, mx, 4tap_h imul myd, mym, 0x010101 add myd, t1d ; 8tap_v, my, 4tap_v lea r8, [put_avx2] movsxd wq, wm movifnidn hd, hm test mxd, 0xf00 jnz .h test myd, 0xf00 jnz .v tzcnt wd, wd movzx wd, word [r8+wq*2+table_offset(put,)] add wq, r8 lea r6, [ssq*3] lea r7, [dsq*3] jmp wq .h: test myd, 0xf00 jnz .hv vpbroadcastd m5, [pw_34] ; 2 + (8 << 2) cmp wd, 4 jl .h_w2 vbroadcasti128 m6, [subpel_h_shufA] je .h_w4 tzcnt wd, wd vbroadcasti128 m7, [subpel_h_shufB] vbroadcasti128 m8, [subpel_h_shufC] shr mxd, 16 sub srcq, 3 movzx wd, word [r8+wq*2+table_offset(put, _8tap_h)] vpbroadcastd m9, [r8+mxq*8+subpel_filters-put_avx2+0] vpbroadcastd m10, [r8+mxq*8+subpel_filters-put_avx2+4] add wq, r8 jmp wq .h_w2: movzx mxd, mxb dec srcq mova xm4, [subpel_h_shuf4] vpbroadcastd xm3, [r8+mxq*8+subpel_filters-put_avx2+2] .h_w2_loop: movq xm0, [srcq+ssq*0] movhps xm0, [srcq+ssq*1] lea srcq, [srcq+ssq*2] pshufb xm0, xm4 pmaddubsw xm0, xm3 phaddw xm0, xm0 paddw xm0, xm5 psraw xm0, 6 packuswb xm0, xm0 pextrw [dstq+dsq*0], xm0, 0 pextrw [dstq+dsq*1], xm0, 1 lea dstq, [dstq+dsq*2] sub hd, 2 jg .h_w2_loop RET .h_w4: movzx mxd, mxb dec srcq vpbroadcastd xm3, [r8+mxq*8+subpel_filters-put_avx2+2] .h_w4_loop: movq xm0, [srcq+ssq*0] movq xm1, [srcq+ssq*1] lea srcq, [srcq+ssq*2] pshufb xm0, xm6 pshufb xm1, xm6 pmaddubsw xm0, xm3 pmaddubsw xm1, xm3 phaddw xm0, xm1 paddw xm0, xm5 psraw xm0, 6 packuswb xm0, xm0 movd [dstq+dsq*0], xm0 pextrd [dstq+dsq*1], xm0, 1 lea dstq, [dstq+dsq*2] sub hd, 2 jg .h_w4_loop RET .h_w8: %macro PUT_8TAP_H 4 ; dst/src, tmp[1-3] pshufb m%2, m%1, m7 pshufb m%3, m%1, m8 pshufb m%1, m6 pmaddubsw m%4, m%2, m9 pmaddubsw m%2, m10 pmaddubsw m%3, m10 pmaddubsw m%1, m9 paddw m%3, m%4 paddw m%1, m%2 phaddw m%1, m%3 paddw m%1, m5 psraw m%1, 6 %endmacro movu xm0, [srcq+ssq*0] vinserti128 m0, m0, [srcq+ssq*1], 1 lea srcq, [srcq+ssq*2] PUT_8TAP_H 0, 1, 2, 3 vextracti128 xm1, m0, 1 packuswb xm0, xm1 movq [dstq+dsq*0], xm0 movhps [dstq+dsq*1], xm0 lea dstq, [dstq+dsq*2] sub hd, 2 jg .h_w8 RET .h_w16: movu xm0, [srcq+ssq*0+8*0] vinserti128 m0, m0, [srcq+ssq*1+8*0], 1 movu xm1, [srcq+ssq*0+8*1] vinserti128 m1, m1, [srcq+ssq*1+8*1], 1 PUT_8TAP_H 0, 2, 3, 4 lea srcq, [srcq+ssq*2] PUT_8TAP_H 1, 2, 3, 4 packuswb m0, m1 mova [dstq+dsq*0], xm0 vextracti128 [dstq+dsq*1], m0, 1 lea dstq, [dstq+dsq*2] sub hd, 2 jg .h_w16 RET .h_w32: xor r6d, r6d jmp .h_start .h_w64: mov r6, -32*1 jmp .h_start .h_w128: mov r6, -32*3 .h_start: sub srcq, r6 sub dstq, r6 mov r4, r6 .h_loop: movu m0, [srcq+r6+8*0] movu m1, [srcq+r6+8*1] PUT_8TAP_H 0, 2, 3, 4 PUT_8TAP_H 1, 2, 3, 4 packuswb m0, m1 mova [dstq+r6], m0 add r6, 32 jle .h_loop add srcq, ssq add dstq, dsq mov r6, r4 dec hd jg .h_loop RET .v: movzx mxd, myb shr myd, 16 cmp hd, 4 cmovle myd, mxd tzcnt r6d, wd movzx r6d, word [r8+r6*2+table_offset(put, _8tap_v)] vpbroadcastd m7, [pw_512] lea myq, [r8+myq*8+subpel_filters-put_avx2] vpbroadcastw m8, [myq+0] vpbroadcastw m9, [myq+2] vpbroadcastw m10, [myq+4] vpbroadcastw m11, [myq+6] add r6, r8 lea ss3q, [ssq*3] sub srcq, ss3q jmp r6 .v_w2: movd xm2, [srcq+ssq*0] pinsrw xm2, [srcq+ssq*1], 2 pinsrw xm2, [srcq+ssq*2], 4 pinsrw xm2, [srcq+ss3q ], 6 ; 0 1 2 3 lea srcq, [srcq+ssq*4] movd xm3, [srcq+ssq*0] vpbroadcastd xm1, [srcq+ssq*1] vpbroadcastd xm0, [srcq+ssq*2] add srcq, ss3q vpblendd xm3, xm3, xm1, 0x02 ; 4 5 vpblendd xm1, xm1, xm0, 0x02 ; 5 6 palignr xm4, xm3, xm2, 4 ; 1 2 3 4 punpcklbw xm3, xm1 ; 45 56 punpcklbw xm1, xm2, xm4 ; 01 12 punpckhbw xm2, xm4 ; 23 34 .v_w2_loop: pmaddubsw xm5, xm1, xm8 ; a0 b0 mova xm1, xm2 pmaddubsw xm2, xm9 ; a1 b1 paddw xm5, xm2 mova xm2, xm3 pmaddubsw xm3, xm10 ; a2 b2 paddw xm5, xm3 vpbroadcastd xm4, [srcq+ssq*0] vpblendd xm3, xm0, xm4, 0x02 ; 6 7 vpbroadcastd xm0, [srcq+ssq*1] lea srcq, [srcq+ssq*2] vpblendd xm4, xm4, xm0, 0x02 ; 7 8 punpcklbw xm3, xm4 ; 67 78 pmaddubsw xm4, xm3, xm11 ; a3 b3 paddw xm5, xm4 pmulhrsw xm5, xm7 packuswb xm5, xm5 pextrw [dstq+dsq*0], xm5, 0 pextrw [dstq+dsq*1], xm5, 2 lea dstq, [dstq+dsq*2] sub hd, 2 jg .v_w2_loop RET .v_w4: movd xm2, [srcq+ssq*0] pinsrd xm2, [srcq+ssq*1], 1 pinsrd xm2, [srcq+ssq*2], 2 pinsrd xm2, [srcq+ss3q ], 3 ; 0 1 2 3 lea srcq, [srcq+ssq*4] movd xm3, [srcq+ssq*0] vpbroadcastd xm1, [srcq+ssq*1] vpbroadcastd xm0, [srcq+ssq*2] add srcq, ss3q vpblendd xm3, xm3, xm1, 0x02 ; 4 5 vpblendd xm1, xm1, xm0, 0x02 ; 5 6 palignr xm4, xm3, xm2, 4 ; 1 2 3 4 punpcklbw xm3, xm1 ; 45 56 punpcklbw xm1, xm2, xm4 ; 01 12 punpckhbw xm2, xm4 ; 23 34 .v_w4_loop: pmaddubsw xm5, xm1, xm8 ; a0 b0 mova xm1, xm2 pmaddubsw xm2, xm9 ; a1 b1 paddw xm5, xm2 mova xm2, xm3 pmaddubsw xm3, xm10 ; a2 b2 paddw xm5, xm3 vpbroadcastd xm4, [srcq+ssq*0] vpblendd xm3, xm0, xm4, 0x02 ; 6 7 vpbroadcastd xm0, [srcq+ssq*1] lea srcq, [srcq+ssq*2] vpblendd xm4, xm4, xm0, 0x02 ; 7 8 punpcklbw xm3, xm4 ; 67 78 pmaddubsw xm4, xm3, xm11 ; a3 b3 paddw xm5, xm4 pmulhrsw xm5, xm7 packuswb xm5, xm5 movd [dstq+dsq*0], xm5 pextrd [dstq+dsq*1], xm5, 1 lea dstq, [dstq+dsq*2] sub hd, 2 jg .v_w4_loop RET .v_w8: movq xm1, [srcq+ssq*0] vpbroadcastq m4, [srcq+ssq*1] vpbroadcastq m2, [srcq+ssq*2] vpbroadcastq m5, [srcq+ss3q ] lea srcq, [srcq+ssq*4] vpbroadcastq m3, [srcq+ssq*0] vpbroadcastq m6, [srcq+ssq*1] vpbroadcastq m0, [srcq+ssq*2] add srcq, ss3q vpblendd m1, m1, m4, 0x30 vpblendd m4, m4, m2, 0x30 punpcklbw m1, m4 ; 01 12 vpblendd m2, m2, m5, 0x30 vpblendd m5, m5, m3, 0x30 punpcklbw m2, m5 ; 23 34 vpblendd m3, m3, m6, 0x30 vpblendd m6, m6, m0, 0x30 punpcklbw m3, m6 ; 45 56 .v_w8_loop: pmaddubsw m5, m1, m8 ; a0 b0 mova m1, m2 pmaddubsw m2, m9 ; a1 b1 paddw m5, m2 mova m2, m3 pmaddubsw m3, m10 ; a2 b2 paddw m5, m3 vpbroadcastq m4, [srcq+ssq*0] vpblendd m3, m0, m4, 0x30 vpbroadcastq m0, [srcq+ssq*1] lea srcq, [srcq+ssq*2] vpblendd m4, m4, m0, 0x30 punpcklbw m3, m4 ; 67 78 pmaddubsw m4, m3, m11 ; a3 b3 paddw m5, m4 pmulhrsw m5, m7 vextracti128 xm4, m5, 1 packuswb xm5, xm4 movq [dstq+dsq*0], xm5 movhps [dstq+dsq*1], xm5 lea dstq, [dstq+dsq*2] sub hd, 2 jg .v_w8_loop RET .v_w16: .v_w32: .v_w64: .v_w128: lea r6d, [wq-16] mov r4, dstq mov r7, srcq shl r6d, 4 mov r6b, hb .v_w16_loop0: vbroadcasti128 m4, [srcq+ssq*0] vbroadcasti128 m5, [srcq+ssq*1] lea srcq, [srcq+ssq*2] vbroadcasti128 m0, [srcq+ssq*1] vbroadcasti128 m6, [srcq+ssq*0] lea srcq, [srcq+ssq*2] vbroadcasti128 m1, [srcq+ssq*0] vbroadcasti128 m2, [srcq+ssq*1] lea srcq, [srcq+ssq*2] vbroadcasti128 m3, [srcq+ssq*0] shufpd m4, m4, m0, 0x0c shufpd m5, m5, m1, 0x0c punpcklbw m1, m4, m5 ; 01 punpckhbw m4, m5 ; 34 shufpd m6, m6, m2, 0x0c punpcklbw m2, m5, m6 ; 12 punpckhbw m5, m6 ; 45 shufpd m0, m0, m3, 0x0c punpcklbw m3, m6, m0 ; 23 punpckhbw m6, m0 ; 56 .v_w16_loop: vbroadcasti128 m12, [srcq+ssq*1] lea srcq, [srcq+ssq*2] vbroadcasti128 m13, [srcq+ssq*0] pmaddubsw m14, m1, m8 ; a0 pmaddubsw m15, m2, m8 ; b0 mova m1, m3 mova m2, m4 pmaddubsw m3, m9 ; a1 pmaddubsw m4, m9 ; b1 paddw m14, m3 paddw m15, m4 mova m3, m5 mova m4, m6 pmaddubsw m5, m10 ; a2 pmaddubsw m6, m10 ; b2 paddw m14, m5 paddw m15, m6 shufpd m6, m0, m12, 0x0d shufpd m0, m12, m13, 0x0c punpcklbw m5, m6, m0 ; 67 punpckhbw m6, m0 ; 78 pmaddubsw m12, m5, m11 ; a3 pmaddubsw m13, m6, m11 ; b3 paddw m14, m12 paddw m15, m13 pmulhrsw m14, m7 pmulhrsw m15, m7 packuswb m14, m15 vpermq m14, m14, q3120 mova [dstq+dsq*0], xm14 vextracti128 [dstq+dsq*1], m14, 1 lea dstq, [dstq+dsq*2] sub hd, 2 jg .v_w16_loop mov hb, r6b add r4, 16 add r7, 16 mov dstq, r4 mov srcq, r7 sub r6d, 1<<8 jg .v_w16_loop0 RET .hv: cmp wd, 4 jg .hv_w8 movzx mxd, mxb dec srcq vpbroadcastd m7, [r8+mxq*8+subpel_filters-put_avx2+2] movzx mxd, myb shr myd, 16 cmp hd, 4 cmovle myd, mxd vpbroadcastq m0, [r8+myq*8+subpel_filters-put_avx2] lea ss3q, [ssq*3] sub srcq, ss3q punpcklbw m0, m0 psraw m0, 8 ; sign-extend vpbroadcastd m8, [pw_8192] vpbroadcastd m9, [pd_512] pshufd m10, m0, q0000 pshufd m11, m0, q1111 pshufd m12, m0, q2222 pshufd m13, m0, q3333 cmp wd, 4 je .hv_w4 vbroadcasti128 m6, [subpel_h_shuf4] movq xm2, [srcq+ssq*0] movhps xm2, [srcq+ssq*1] movq xm0, [srcq+ssq*2] movhps xm0, [srcq+ss3q ] lea srcq, [srcq+ssq*4] vpbroadcastq m3, [srcq+ssq*0] vpbroadcastq m4, [srcq+ssq*1] vpbroadcastq m1, [srcq+ssq*2] add srcq, ss3q vpblendd m2, m2, m3, 0x30 vpblendd m0, m0, m1, 0x30 vpblendd m2, m2, m4, 0xc0 pshufb m2, m6 pshufb m0, m6 pmaddubsw m2, m7 pmaddubsw m0, m7 phaddw m2, m0 pmulhrsw m2, m8 vextracti128 xm3, m2, 1 palignr xm4, xm3, xm2, 4 punpcklwd xm1, xm2, xm4 ; 01 12 punpckhwd xm2, xm4 ; 23 34 pshufd xm0, xm3, q2121 punpcklwd xm3, xm0 ; 45 56 .hv_w2_loop: pmaddwd xm5, xm1, xm10 ; a0 b0 mova xm1, xm2 pmaddwd xm2, xm11 ; a1 b1 paddd xm5, xm2 mova xm2, xm3 pmaddwd xm3, xm12 ; a2 b2 paddd xm5, xm3 movq xm4, [srcq+ssq*0] movhps xm4, [srcq+ssq*1] lea srcq, [srcq+ssq*2] pshufb xm4, xm6 pmaddubsw xm4, xm7 phaddw xm4, xm4 pmulhrsw xm4, xm8 palignr xm3, xm4, xm0, 12 mova xm0, xm4 punpcklwd xm3, xm0 ; 67 78 pmaddwd xm4, xm3, xm13 ; a3 b3 paddd xm5, xm9 paddd xm5, xm4 psrad xm5, 10 packssdw xm5, xm5 packuswb xm5, xm5 pextrw [dstq+dsq*0], xm5, 0 pextrw [dstq+dsq*1], xm5, 1 lea dstq, [dstq+dsq*2] sub hd, 2 jg .hv_w2_loop RET .hv_w4: mova m6, [subpel_h_shuf4] vpbroadcastq m2, [srcq+ssq*0] vpbroadcastq m4, [srcq+ssq*1] vpbroadcastq m0, [srcq+ssq*2] vpbroadcastq m5, [srcq+ss3q ] lea srcq, [srcq+ssq*4] vpbroadcastq m3, [srcq+ssq*0] vpblendd m2, m2, m4, 0xcc ; 0 1 vpbroadcastq m4, [srcq+ssq*1] vpbroadcastq m1, [srcq+ssq*2] add srcq, ss3q vpblendd m0, m0, m5, 0xcc ; 2 3 vpblendd m3, m3, m4, 0xcc ; 4 5 pshufb m2, m6 pshufb m0, m6 pshufb m3, m6 pshufb m1, m6 pmaddubsw m2, m7 pmaddubsw m0, m7 pmaddubsw m3, m7 pmaddubsw m1, m7 phaddw m2, m0 phaddw m3, m1 pmulhrsw m2, m8 pmulhrsw m3, m8 palignr m4, m3, m2, 4 punpcklwd m1, m2, m4 ; 01 12 punpckhwd m2, m4 ; 23 34 pshufd m0, m3, q2121 punpcklwd m3, m0 ; 45 56 .hv_w4_loop: pmaddwd m5, m1, m10 ; a0 b0 mova m1, m2 pmaddwd m2, m11 ; a1 b1 paddd m5, m2 mova m2, m3 pmaddwd m3, m12 ; a2 b2 paddd m5, m3 vpbroadcastq m4, [srcq+ssq*0] vpbroadcastq m3, [srcq+ssq*1] lea srcq, [srcq+ssq*2] vpblendd m4, m4, m3, 0xcc ; 7 8 pshufb m4, m6 pmaddubsw m4, m7 phaddw m4, m4 pmulhrsw m4, m8 palignr m3, m4, m0, 12 mova m0, m4 punpcklwd m3, m0 ; 67 78 pmaddwd m4, m3, m13 ; a3 b3 paddd m5, m9 paddd m5, m4 psrad m5, 10 vextracti128 xm4, m5, 1 packssdw xm5, xm4 packuswb xm5, xm5 pshuflw xm5, xm5, q3120 movd [dstq+dsq*0], xm5 pextrd [dstq+dsq*1], xm5, 1 lea dstq, [dstq+dsq*2] sub hd, 2 jg .hv_w4_loop RET .hv_w8: shr mxd, 16 sub srcq, 3 vpbroadcastd m10, [r8+mxq*8+subpel_filters-put_avx2+0] vpbroadcastd m11, [r8+mxq*8+subpel_filters-put_avx2+4] movzx mxd, myb shr myd, 16 cmp hd, 4 cmovle myd, mxd vpbroadcastq m0, [r8+myq*8+subpel_filters-put_avx2] lea ss3q, [ssq*3] sub srcq, ss3q punpcklbw m0, m0 psraw m0, 8 ; sign-extend pshufd m12, m0, q0000 pshufd m13, m0, q1111 pshufd m14, m0, q2222 pshufd m15, m0, q3333 lea r6d, [wq-8] mov r4, dstq mov r7, srcq shl r6d, 5 mov r6b, hb .hv_w8_loop0: vbroadcasti128 m7, [subpel_h_shufA] vbroadcasti128 m8, [subpel_h_shufB] vbroadcasti128 m9, [subpel_h_shufC] movu xm4, [srcq+ssq*0] movu xm5, [srcq+ssq*1] lea srcq, [srcq+ssq*2] movu xm6, [srcq+ssq*0] vbroadcasti128 m0, [srcq+ssq*1] lea srcq, [srcq+ssq*2] vpblendd m4, m4, m0, 0xf0 ; 0 3 vinserti128 m5, m5, [srcq+ssq*0], 1 ; 1 4 vinserti128 m6, m6, [srcq+ssq*1], 1 ; 2 5 lea srcq, [srcq+ssq*2] vinserti128 m0, m0, [srcq+ssq*0], 1 ; 3 6 %macro HV_H_W8 4-7 ; src/dst, tmp[1-3], shuf[1-3] pshufb %3, %1, %6 pshufb %4, %1, %7 pshufb %1, %5 pmaddubsw %2, %3, m10 pmaddubsw %4, m11 pmaddubsw %3, m11 pmaddubsw %1, m10 paddw %2, %4 paddw %1, %3 phaddw %1, %2 %endmacro HV_H_W8 m4, m1, m2, m3, m7, m8, m9 HV_H_W8 m5, m1, m2, m3, m7, m8, m9 HV_H_W8 m6, m1, m2, m3, m7, m8, m9 HV_H_W8 m0, m1, m2, m3, m7, m8, m9 vpbroadcastd m7, [pw_8192] vpermq m4, m4, q3120 vpermq m5, m5, q3120 vpermq m6, m6, q3120 pmulhrsw m0, m7 pmulhrsw m4, m7 pmulhrsw m5, m7 pmulhrsw m6, m7 vpermq m7, m0, q3120 punpcklwd m1, m4, m5 ; 01 punpckhwd m4, m5 ; 34 punpcklwd m2, m5, m6 ; 12 punpckhwd m5, m6 ; 45 punpcklwd m3, m6, m7 ; 23 punpckhwd m6, m7 ; 56 .hv_w8_loop: vextracti128 r6m, m0, 1 ; not enough registers movu xm0, [srcq+ssq*1] lea srcq, [srcq+ssq*2] vinserti128 m0, m0, [srcq+ssq*0], 1 ; 7 8 pmaddwd m8, m1, m12 ; a0 pmaddwd m9, m2, m12 ; b0 mova m1, m3 mova m2, m4 pmaddwd m3, m13 ; a1 pmaddwd m4, m13 ; b1 paddd m8, m3 paddd m9, m4 mova m3, m5 mova m4, m6 pmaddwd m5, m14 ; a2 pmaddwd m6, m14 ; b2 paddd m8, m5 paddd m9, m6 vbroadcasti128 m6, [subpel_h_shufB] vbroadcasti128 m7, [subpel_h_shufC] vbroadcasti128 m5, [subpel_h_shufA] HV_H_W8 m0, m5, m6, m7, m5, m6, m7 vpbroadcastd m5, [pw_8192] vpbroadcastd m7, [pd_512] vbroadcasti128 m6, r6m pmulhrsw m0, m5 paddd m8, m7 paddd m9, m7 vpermq m7, m0, q3120 ; 7 8 shufpd m6, m6, m7, 0x04 ; 6 7 punpcklwd m5, m6, m7 ; 67 punpckhwd m6, m7 ; 78 pmaddwd m7, m5, m15 ; a3 paddd m8, m7 pmaddwd m7, m6, m15 ; b3 paddd m7, m9 psrad m8, 10 psrad m7, 10 packssdw m8, m7 vextracti128 xm7, m8, 1 packuswb xm8, xm7 pshufd xm7, xm8, q3120 movq [dstq+dsq*0], xm7 movhps [dstq+dsq*1], xm7 lea dstq, [dstq+dsq*2] sub hd, 2 jg .hv_w8_loop mov hb, r6b add r4, 8 add r7, 8 mov dstq, r4 mov srcq, r7 sub r6d, 1<<8 jg .hv_w8_loop0 RET DECLARE_REG_TMP 6, 7 %macro PREP_8TAP_FN 3 ; type, type_h, type_v cglobal prep_8tap_%1 mov t0d, FILTER_%2 mov t1d, FILTER_%3 %ifnidn %1, sharp_smooth ; skip the jump in the last filter jmp mangle(private_prefix %+ _prep_8tap %+ SUFFIX) %endif %endmacro PREP_8TAP_FN regular, REGULAR, REGULAR PREP_8TAP_FN regular_sharp, REGULAR, SHARP PREP_8TAP_FN regular_smooth, REGULAR, SMOOTH PREP_8TAP_FN smooth_regular, SMOOTH, REGULAR PREP_8TAP_FN smooth, SMOOTH, SMOOTH PREP_8TAP_FN smooth_sharp, SMOOTH, SHARP PREP_8TAP_FN sharp_regular, SHARP, REGULAR PREP_8TAP_FN sharp, SHARP, SHARP PREP_8TAP_FN sharp_smooth, SHARP, SMOOTH cglobal prep_8tap, 3, 8, 16, tmp, src, stride, w, h, mx, my, stride3 imul mxd, mxm, 0x010101 add mxd, t0d ; 8tap_h, mx, 4tap_h imul myd, mym, 0x010101 add myd, t1d ; 8tap_v, my, 4tap_v lea r7, [prep_avx2] movsxd wq, wm movifnidn hd, hm test mxd, 0xf00 jnz .h test myd, 0xf00 jnz .v tzcnt wd, wd movzx wd, word [r7+wq*2+table_offset(prep,)] add wq, r7 lea r6, [strideq*3] jmp wq .h: test myd, 0xf00 jnz .hv vbroadcasti128 m5, [subpel_h_shufA] vpbroadcastd m4, [pw_8192] cmp wd, 4 je .h_w4 tzcnt wd, wd vbroadcasti128 m6, [subpel_h_shufB] vbroadcasti128 m7, [subpel_h_shufC] shr mxd, 16 sub srcq, 3 movzx wd, word [r7+wq*2+table_offset(prep, _8tap_h)] vpbroadcastd m8, [r7+mxq*8+subpel_filters-prep_avx2+0] vpbroadcastd m9, [r7+mxq*8+subpel_filters-prep_avx2+4] add wq, r7 jmp wq .h_w4: movzx mxd, mxb dec srcq vpbroadcastd m3, [r7+mxq*8+subpel_filters-prep_avx2+2] lea stride3q, [strideq*3] .h_w4_loop: movq xm0, [srcq+strideq*0] vpbroadcastq m2, [srcq+strideq*2] movq xm1, [srcq+strideq*1] vpblendd m0, m0, m2, 0xf0 vpbroadcastq m2, [srcq+stride3q ] lea srcq, [srcq+strideq*4] vpblendd m1, m1, m2, 0xf0 pshufb m0, m5 pshufb m1, m5 pmaddubsw m0, m3 pmaddubsw m1, m3 phaddw m0, m1 pmulhrsw m0, m4 mova [tmpq], m0 add tmpq, 32 sub hd, 4 jg .h_w4_loop RET .h_w8: %macro PREP_8TAP_H 0 pshufb m1, m0, m6 pshufb m2, m0, m7 pshufb m0, m5 pmaddubsw m3, m1, m8 pmaddubsw m1, m9 pmaddubsw m2, m9 pmaddubsw m0, m8 paddw m2, m3 paddw m0, m1 phaddw m0, m2 pmulhrsw m0, m4 %endmacro movu xm0, [srcq+strideq*0] vinserti128 m0, m0, [srcq+strideq*1], 1 lea srcq, [srcq+strideq*2] PREP_8TAP_H mova [tmpq], m0 add tmpq, 32 sub hd, 2 jg .h_w8 RET .h_w16: movu xm0, [srcq+strideq*0+8*0] vinserti128 m0, m0, [srcq+strideq*0+8*1], 1 PREP_8TAP_H mova [tmpq+32*0], m0 movu xm0, [srcq+strideq*1+8*0] vinserti128 m0, m0, [srcq+strideq*1+8*1], 1 lea srcq, [srcq+strideq*2] PREP_8TAP_H mova [tmpq+32*1], m0 add tmpq, 64 sub hd, 2 jg .h_w16 RET .h_w32: xor r6d, r6d jmp .h_start .h_w64: mov r6, -32*1 jmp .h_start .h_w128: mov r6, -32*3 .h_start: sub srcq, r6 mov r5, r6 .h_loop: movu xm0, [srcq+r6+8*0] vinserti128 m0, m0, [srcq+r6+8*1], 1 PREP_8TAP_H mova [tmpq+32*0], m0 movu xm0, [srcq+r6+8*2] vinserti128 m0, m0, [srcq+r6+8*3], 1 PREP_8TAP_H mova [tmpq+32*1], m0 add tmpq, 64 add r6, 32 jle .h_loop add srcq, strideq mov r6, r5 dec hd jg .h_loop RET .v: movzx mxd, myb ; Select 4-tap/8-tap filter multipliers. shr myd, 16 ; Note that the code is 8-tap only, having cmp hd, 4 ; a separate 4-tap code path for (4|8|16)x4 cmove myd, mxd ; had a negligible effect on performance. ; TODO: Would a 6-tap code path be worth it? vpbroadcastd m7, [pw_8192] lea myq, [r7+myq*8+subpel_filters-prep_avx2] vpbroadcastw m8, [myq+0] vpbroadcastw m9, [myq+2] vpbroadcastw m10, [myq+4] vpbroadcastw m11, [myq+6] lea stride3q, [strideq*3] sub srcq, stride3q cmp wd, 8 jg .v_w16 je .v_w8 .v_w4: movd xm0, [srcq+strideq*0] vpbroadcastd m1, [srcq+strideq*2] vpbroadcastd xm2, [srcq+strideq*1] vpbroadcastd m3, [srcq+stride3q ] lea srcq, [srcq+strideq*4] vpblendd m1, m1, m0, 0x01 ; 0 2 2 _ 2 _ _ _ vpblendd m3, m3, m2, 0x03 ; 1 1 3 3 3 3 _ _ vpbroadcastd m0, [srcq+strideq*0] vpbroadcastd m2, [srcq+strideq*1] vpblendd m1, m1, m0, 0x68 ; 0 2 2 4 2 4 4 _ vpbroadcastd m0, [srcq+strideq*2] vbroadcasti128 m6, [deint_shuf4] vpblendd m3, m3, m2, 0xc0 ; 1 1 3 3 3 3 5 5 vpblendd m2, m3, m1, 0x55 ; 0 1 2 3 2 3 4 5 vpblendd m3, m3, m1, 0xaa ; 1 2 3 4 3 4 5 _ punpcklbw m1, m2, m3 ; 01 12 23 34 vpblendd m3, m3, m0, 0x80 ; 1 2 3 4 3 4 5 6 punpckhbw m2, m3 ; 23 34 45 56 .v_w4_loop: pinsrd xm0, [srcq+stride3q ], 1 lea srcq, [srcq+strideq*4] vpbroadcastd m3, [srcq+strideq*0] vpbroadcastd m4, [srcq+strideq*1] vpblendd m3, m3, m4, 0x20 ; _ _ 8 _ 8 9 _ _ vpblendd m3, m3, m0, 0x03 ; 6 7 8 _ 8 9 _ _ vpbroadcastd m0, [srcq+strideq*2] vpblendd m3, m3, m0, 0x40 ; 6 7 8 _ 8 9 a _ pshufb m3, m6 ; 67 78 89 9a pmaddubsw m4, m1, m8 vperm2i128 m1, m2, m3, 0x21 ; 45 56 67 78 pmaddubsw m2, m9 paddw m4, m2 mova m2, m3 pmaddubsw m3, m11 paddw m3, m4 pmaddubsw m4, m1, m10 paddw m3, m4 pmulhrsw m3, m7 mova [tmpq], m3 add tmpq, 32 sub hd, 4 jg .v_w4_loop RET .v_w8: movq xm1, [srcq+strideq*0] vpbroadcastq m4, [srcq+strideq*1] vpbroadcastq m2, [srcq+strideq*2] vpbroadcastq m5, [srcq+stride3q ] lea srcq, [srcq+strideq*4] vpbroadcastq m3, [srcq+strideq*0] vpbroadcastq m6, [srcq+strideq*1] vpbroadcastq m0, [srcq+strideq*2] vpblendd m1, m1, m4, 0x30 vpblendd m4, m4, m2, 0x30 punpcklbw m1, m4 ; 01 12 vpblendd m2, m2, m5, 0x30 vpblendd m5, m5, m3, 0x30 punpcklbw m2, m5 ; 23 34 vpblendd m3, m3, m6, 0x30 vpblendd m6, m6, m0, 0x30 punpcklbw m3, m6 ; 45 56 .v_w8_loop: vpbroadcastq m4, [srcq+stride3q ] lea srcq, [srcq+strideq*4] pmaddubsw m5, m2, m9 ; a1 pmaddubsw m6, m2, m8 ; b0 vpblendd m2, m0, m4, 0x30 vpbroadcastq m0, [srcq+strideq*0] vpblendd m4, m4, m0, 0x30 punpcklbw m2, m4 ; 67 78 pmaddubsw m1, m8 ; a0 pmaddubsw m4, m3, m9 ; b1 paddw m5, m1 mova m1, m3 pmaddubsw m3, m10 ; a2 paddw m6, m4 paddw m5, m3 vpbroadcastq m4, [srcq+strideq*1] vpblendd m3, m0, m4, 0x30 vpbroadcastq m0, [srcq+strideq*2] vpblendd m4, m4, m0, 0x30 punpcklbw m3, m4 ; 89 9a pmaddubsw m4, m2, m11 ; a3 paddw m5, m4 pmaddubsw m4, m2, m10 ; b2 paddw m6, m4 pmaddubsw m4, m3, m11 ; b3 paddw m6, m4 pmulhrsw m5, m7 pmulhrsw m6, m7 mova [tmpq+32*0], m5 mova [tmpq+32*1], m6 add tmpq, 32*2 sub hd, 4 jg .v_w8_loop RET .v_w16: lea r6d, [wq-16] mov r5, tmpq mov r7, srcq shl r6d, 4 mov r6b, hb .v_w16_loop0: vbroadcasti128 m4, [srcq+strideq*0] vbroadcasti128 m5, [srcq+strideq*1] lea srcq, [srcq+strideq*2] vbroadcasti128 m0, [srcq+strideq*1] vbroadcasti128 m6, [srcq+strideq*0] lea srcq, [srcq+strideq*2] vbroadcasti128 m1, [srcq+strideq*0] vbroadcasti128 m2, [srcq+strideq*1] lea srcq, [srcq+strideq*2] vbroadcasti128 m3, [srcq+strideq*0] shufpd m4, m4, m0, 0x0c shufpd m5, m5, m1, 0x0c punpcklbw m1, m4, m5 ; 01 punpckhbw m4, m5 ; 34 shufpd m6, m6, m2, 0x0c punpcklbw m2, m5, m6 ; 12 punpckhbw m5, m6 ; 45 shufpd m0, m0, m3, 0x0c punpcklbw m3, m6, m0 ; 23 punpckhbw m6, m0 ; 56 .v_w16_loop: vbroadcasti128 m12, [srcq+strideq*1] lea srcq, [srcq+strideq*2] vbroadcasti128 m13, [srcq+strideq*0] pmaddubsw m14, m1, m8 ; a0 pmaddubsw m15, m2, m8 ; b0 mova m1, m3 mova m2, m4 pmaddubsw m3, m9 ; a1 pmaddubsw m4, m9 ; b1 paddw m14, m3 paddw m15, m4 mova m3, m5 mova m4, m6 pmaddubsw m5, m10 ; a2 pmaddubsw m6, m10 ; b2 paddw m14, m5 paddw m15, m6 shufpd m6, m0, m12, 0x0d shufpd m0, m12, m13, 0x0c punpcklbw m5, m6, m0 ; 67 punpckhbw m6, m0 ; 78 pmaddubsw m12, m5, m11 ; a3 pmaddubsw m13, m6, m11 ; b3 paddw m14, m12 paddw m15, m13 pmulhrsw m14, m7 pmulhrsw m15, m7 mova [tmpq+wq*0], m14 mova [tmpq+wq*2], m15 lea tmpq, [tmpq+wq*4] sub hd, 2 jg .v_w16_loop mov hb, r6b add r5, 32 add r7, 16 mov tmpq, r5 mov srcq, r7 sub r6d, 1<<8 jg .v_w16_loop0 RET .hv: cmp wd, 4 jg .hv_w8 movzx mxd, mxb dec srcq mova m7, [subpel_h_shuf4] vpbroadcastd m8, [r7+mxq*8+subpel_filters-prep_avx2+2] pmovzxbd m9, [deint_shuf4] movzx mxd, myb shr myd, 16 cmp hd, 4 cmove myd, mxd vpbroadcastq m0, [r7+myq*8+subpel_filters-prep_avx2] lea stride3q, [strideq*3] sub srcq, stride3q punpcklbw m0, m0 psraw m0, 8 ; sign-extend vpbroadcastd m10, [pw_8192] vpbroadcastd m11, [pd_32] pshufd m12, m0, q0000 pshufd m13, m0, q1111 pshufd m14, m0, q2222 pshufd m15, m0, q3333 vpbroadcastq m2, [srcq+strideq*0] vpbroadcastq m4, [srcq+strideq*1] vpbroadcastq m0, [srcq+strideq*2] vpbroadcastq m5, [srcq+stride3q ] lea srcq, [srcq+strideq*4] vpbroadcastq m3, [srcq+strideq*0] vpbroadcastq m6, [srcq+strideq*1] vpbroadcastq m1, [srcq+strideq*2] vpblendd m2, m2, m4, 0xcc ; 0 1 vpblendd m0, m0, m5, 0xcc ; 2 3 vpblendd m3, m3, m6, 0xcc ; 4 5 pshufb m2, m7 pshufb m0, m7 pshufb m3, m7 pshufb m1, m7 pmaddubsw m2, m8 pmaddubsw m0, m8 pmaddubsw m3, m8 pmaddubsw m1, m8 phaddw m2, m0 phaddw m3, m1 pmulhrsw m2, m10 pmulhrsw m3, m10 palignr m4, m3, m2, 4 punpcklwd m1, m2, m4 ; 01 12 punpckhwd m2, m4 ; 23 34 pshufd m0, m3, q2121 punpcklwd m3, m0 ; 45 56 .hv_w4_loop: pmaddwd m5, m1, m12 ; a0 b0 pmaddwd m6, m2, m12 ; c0 d0 pmaddwd m2, m13 ; a1 b1 pmaddwd m4, m3, m13 ; c1 d1 mova m1, m3 pmaddwd m3, m14 ; a2 b2 paddd m5, m2 vpbroadcastq m2, [srcq+stride3q ] lea srcq, [srcq+strideq*4] paddd m6, m4 paddd m5, m3 vpbroadcastq m4, [srcq+strideq*0] vpbroadcastq m3, [srcq+strideq*1] vpblendd m2, m2, m4, 0xcc vpbroadcastq m4, [srcq+strideq*2] vpblendd m3, m3, m4, 0xcc pshufb m2, m7 pshufb m3, m7 pmaddubsw m2, m8 pmaddubsw m3, m8 phaddw m2, m3 pmulhrsw m2, m10 palignr m3, m2, m0, 12 mova m0, m2 punpcklwd m2, m3, m0 ; 67 78 punpckhwd m3, m0 ; 89 9a pmaddwd m4, m2, m14 ; c2 d2 paddd m6, m11 paddd m5, m11 paddd m6, m4 pmaddwd m4, m2, m15 ; a3 b3 paddd m5, m4 pmaddwd m4, m3, m15 ; c3 d3 paddd m6, m4 psrad m5, 6 psrad m6, 6 packssdw m5, m6 vpermd m5, m9, m5 mova [tmpq], m5 add tmpq, 32 sub hd, 4 jg .hv_w4_loop RET .hv_w8: shr mxd, 16 sub srcq, 3 vpbroadcastd m10, [r7+mxq*8+subpel_filters-prep_avx2+0] vpbroadcastd m11, [r7+mxq*8+subpel_filters-prep_avx2+4] movzx mxd, myb shr myd, 16 cmp hd, 4 cmove myd, mxd vpbroadcastq m0, [r7+myq*8+subpel_filters-prep_avx2] lea stride3q, [strideq*3] sub srcq, stride3q punpcklbw m0, m0 psraw m0, 8 ; sign-extend pshufd m12, m0, q0000 pshufd m13, m0, q1111 pshufd m14, m0, q2222 pshufd m15, m0, q3333 lea r6d, [wq-8] mov r5, tmpq mov r7, srcq shl r6d, 5 mov r6b, hb .hv_w8_loop0: vbroadcasti128 m7, [subpel_h_shufA] vbroadcasti128 m8, [subpel_h_shufB] vbroadcasti128 m9, [subpel_h_shufC] movu xm4, [srcq+strideq*0] movu xm5, [srcq+strideq*1] lea srcq, [srcq+strideq*2] movu xm6, [srcq+strideq*0] vbroadcasti128 m0, [srcq+strideq*1] lea srcq, [srcq+strideq*2] vpblendd m4, m4, m0, 0xf0 ; 0 3 vinserti128 m5, m5, [srcq+strideq*0], 1 ; 1 4 vinserti128 m6, m6, [srcq+strideq*1], 1 ; 2 5 lea srcq, [srcq+strideq*2] vinserti128 m0, m0, [srcq+strideq*0], 1 ; 3 6 HV_H_W8 m4, m1, m2, m3, m7, m8, m9 HV_H_W8 m5, m1, m2, m3, m7, m8, m9 HV_H_W8 m6, m1, m2, m3, m7, m8, m9 HV_H_W8 m0, m1, m2, m3, m7, m8, m9 vpbroadcastd m7, [pw_8192] vpermq m4, m4, q3120 vpermq m5, m5, q3120 vpermq m6, m6, q3120 pmulhrsw m0, m7 pmulhrsw m4, m7 pmulhrsw m5, m7 pmulhrsw m6, m7 vpermq m7, m0, q3120 punpcklwd m1, m4, m5 ; 01 punpckhwd m4, m5 ; 34 punpcklwd m2, m5, m6 ; 12 punpckhwd m5, m6 ; 45 punpcklwd m3, m6, m7 ; 23 punpckhwd m6, m7 ; 56 .hv_w8_loop: vextracti128 [tmpq], m0, 1 ; not enough registers movu xm0, [srcq+strideq*1] lea srcq, [srcq+strideq*2] vinserti128 m0, m0, [srcq+strideq*0], 1 ; 7 8 pmaddwd m8, m1, m12 ; a0 pmaddwd m9, m2, m12 ; b0 mova m1, m3 mova m2, m4 pmaddwd m3, m13 ; a1 pmaddwd m4, m13 ; b1 paddd m8, m3 paddd m9, m4 mova m3, m5 mova m4, m6 pmaddwd m5, m14 ; a2 pmaddwd m6, m14 ; b2 paddd m8, m5 paddd m9, m6 vbroadcasti128 m6, [subpel_h_shufB] vbroadcasti128 m7, [subpel_h_shufC] vbroadcasti128 m5, [subpel_h_shufA] HV_H_W8 m0, m5, m6, m7, m5, m6, m7 vpbroadcastd m5, [pw_8192] vpbroadcastd m7, [pd_32] vbroadcasti128 m6, [tmpq] pmulhrsw m0, m5 paddd m8, m7 paddd m9, m7 vpermq m7, m0, q3120 ; 7 8 shufpd m6, m6, m7, 0x04 ; 6 7 punpcklwd m5, m6, m7 ; 67 punpckhwd m6, m7 ; 78 pmaddwd m7, m5, m15 ; a3 paddd m8, m7 pmaddwd m7, m6, m15 ; b3 paddd m7, m9 psrad m8, 6 psrad m7, 6 packssdw m8, m7 vpermq m7, m8, q3120 mova [tmpq+wq*0], xm7 vextracti128 [tmpq+wq*2], m7, 1 lea tmpq, [tmpq+wq*4] sub hd, 2 jg .hv_w8_loop mov hb, r6b add r5, 16 add r7, 8 mov tmpq, r5 mov srcq, r7 sub r6d, 1<<8 jg .hv_w8_loop0 RET %macro BIDIR_FN 1 ; op %1 0 lea stride3q, [strideq*3] jmp wq .w4: vextracti128 xm1, m0, 1 movd [dstq ], xm0 pextrd [dstq+strideq*1], xm0, 1 movd [dstq+strideq*2], xm1 pextrd [dstq+stride3q ], xm1, 1 cmp hd, 4 je .ret lea dstq, [dstq+strideq*4] pextrd [dstq ], xm0, 2 pextrd [dstq+strideq*1], xm0, 3 pextrd [dstq+strideq*2], xm1, 2 pextrd [dstq+stride3q ], xm1, 3 cmp hd, 8 je .ret %1 2 lea dstq, [dstq+strideq*4] vextracti128 xm1, m0, 1 movd [dstq ], xm0 pextrd [dstq+strideq*1], xm0, 1 movd [dstq+strideq*2], xm1 pextrd [dstq+stride3q], xm1, 1 lea dstq, [dstq+strideq*4] pextrd [dstq ], xm0, 2 pextrd [dstq+strideq*1], xm0, 3 pextrd [dstq+strideq*2], xm1, 2 pextrd [dstq+stride3q ], xm1, 3 .ret: RET .w8_loop: %1_INC_PTR 2 %1 0 lea dstq, [dstq+strideq*4] .w8: vextracti128 xm1, m0, 1 movq [dstq ], xm0 movq [dstq+strideq*1], xm1 movhps [dstq+strideq*2], xm0 movhps [dstq+stride3q ], xm1 sub hd, 4 jg .w8_loop RET .w16_loop: %1_INC_PTR 4 %1 0 lea dstq, [dstq+strideq*4] .w16: vpermq m0, m0, q3120 mova [dstq ], xm0 vextracti128 [dstq+strideq*1], m0, 1 %1 2 vpermq m0, m0, q3120 mova [dstq+strideq*2], xm0 vextracti128 [dstq+stride3q ], m0, 1 sub hd, 4 jg .w16_loop RET .w32_loop: %1_INC_PTR 4 %1 0 lea dstq, [dstq+strideq*2] .w32: vpermq m0, m0, q3120 mova [dstq], m0 %1 2 vpermq m0, m0, q3120 mova [dstq+strideq*1], m0 sub hd, 2 jg .w32_loop RET .w64_loop: %1_INC_PTR 4 %1 0 add dstq, strideq .w64: vpermq m0, m0, q3120 mova [dstq], m0 %1 2 vpermq m0, m0, q3120 mova [dstq+32], m0 dec hd jg .w64_loop RET .w128_loop: %1 0 add dstq, strideq .w128: vpermq m0, m0, q3120 mova [dstq+0*32], m0 %1 2 vpermq m0, m0, q3120 mova [dstq+1*32], m0 %1_INC_PTR 8 %1 -4 vpermq m0, m0, q3120 mova [dstq+2*32], m0 %1 -2 vpermq m0, m0, q3120 mova [dstq+3*32], m0 dec hd jg .w128_loop RET %endmacro %macro AVG 1 ; src_offset mova m0, [tmp1q+(%1+0)*mmsize] paddw m0, [tmp2q+(%1+0)*mmsize] mova m1, [tmp1q+(%1+1)*mmsize] paddw m1, [tmp2q+(%1+1)*mmsize] pmulhrsw m0, m2 pmulhrsw m1, m2 packuswb m0, m1 %endmacro %macro AVG_INC_PTR 1 add tmp1q, %1*mmsize add tmp2q, %1*mmsize %endmacro cglobal avg, 4, 7, 3, dst, stride, tmp1, tmp2, w, h, stride3 lea r6, [avg_avx2_table] tzcnt wd, wm movifnidn hd, hm movsxd wq, dword [r6+wq*4] vpbroadcastd m2, [pw_1024+r6-avg_avx2_table] add wq, r6 BIDIR_FN AVG %macro W_AVG 1 ; src_offset ; (a * weight + b * (16 - weight) + 128) >> 8 ; = ((a - b) * weight + (b << 4) + 128) >> 8 ; = ((((b - a) * (-weight << 12)) >> 16) + b + 8) >> 4 mova m0, [tmp2q+(%1+0)*mmsize] psubw m2, m0, [tmp1q+(%1+0)*mmsize] mova m1, [tmp2q+(%1+1)*mmsize] psubw m3, m1, [tmp1q+(%1+1)*mmsize] paddw m2, m2 ; compensate for the weight only being half paddw m3, m3 ; of what it should be pmulhw m2, m4 pmulhw m3, m4 paddw m0, m2 paddw m1, m3 pmulhrsw m0, m5 pmulhrsw m1, m5 packuswb m0, m1 %endmacro %define W_AVG_INC_PTR AVG_INC_PTR cglobal w_avg, 4, 7, 6, dst, stride, tmp1, tmp2, w, h, stride3 lea r6, [w_avg_avx2_table] tzcnt wd, wm movifnidn hd, hm vpbroadcastw m0, r6m ; weight movsxd wq, dword [r6+wq*4] pxor m4, m4 psllw m0, 11 ; can't shift by 12, sign bit must be preserved psubw m4, m0 vpbroadcastd m5, [pw_2048+r6-w_avg_avx2_table] add wq, r6 BIDIR_FN W_AVG %macro MASK 1 ; src_offset ; (a * m + b * (64 - m) + 512) >> 10 ; = ((a - b) * m + (b << 6) + 512) >> 10 ; = ((((b - a) * (-m << 10)) >> 16) + b + 8) >> 4 vpermq m3, [maskq+(%1+0)*(mmsize/2)], q3120 mova m0, [tmp2q+(%1+0)*mmsize] psubw m1, m0, [tmp1q+(%1+0)*mmsize] psubb m3, m4, m3 paddw m1, m1 ; (b - a) << 1 paddb m3, m3 punpcklbw m2, m4, m3 ; -m << 9 pmulhw m1, m2 paddw m0, m1 mova m1, [tmp2q+(%1+1)*mmsize] psubw m2, m1, [tmp1q+(%1+1)*mmsize] paddw m2, m2 punpckhbw m3, m4, m3 pmulhw m2, m3 paddw m1, m2 pmulhrsw m0, m5 pmulhrsw m1, m5 packuswb m0, m1 %endmacro %macro MASK_INC_PTR 1 add maskq, %1*mmsize/2 add tmp1q, %1*mmsize add tmp2q, %1*mmsize %endmacro cglobal mask, 4, 8, 6, dst, stride, tmp1, tmp2, w, h, mask, stride3 lea r7, [mask_avx2_table] tzcnt wd, wm movifnidn hd, hm mov maskq, maskmp movsxd wq, dword [r7+wq*4] pxor m4, m4 vpbroadcastd m5, [pw_2048+r7-mask_avx2_table] add wq, r7 BIDIR_FN MASK %macro W_MASK_420 2 ; src_offset, mask_out mova m0, [tmp1q+(%1+0)*mmsize] mova m1, [tmp2q+(%1+0)*mmsize] psubw m1, m0 pabsw m%2, m1 paddw m%2, m6 psrlw m%2, 8 ; (abs(tmp1 - tmp2) + 8) >> 8 psubusw m%2, m7, m%2 ; 64 - min(m, 64) psllw m2, m%2, 10 pmulhw m1, m2 paddw m0, m1 mova m1, [tmp1q+(%1+1)*mmsize] mova m2, [tmp2q+(%1+1)*mmsize] psubw m2, m1 pabsw m3, m2 paddw m3, m6 psrlw m3, 8 psubusw m3, m7, m3 phaddw m%2, m3 psllw m3, 10 pmulhw m2, m3 paddw m1, m2 pmulhrsw m0, m8 pmulhrsw m1, m8 packuswb m0, m1 %endmacro cglobal w_mask_420, 4, 8, 15, dst, stride, tmp1, tmp2, w, h, mask, stride3 lea r7, [w_mask_420_avx2_table] tzcnt wd, wm movifnidn hd, hm mov maskq, maskmp vpbroadcastw m0, r7m ; sign movsxd wq, dword [r7+wq*4] vpbroadcastd m6, [pw_8 +r7-w_mask_420_avx2_table] vpbroadcastd m7, [pw_26 +r7-w_mask_420_avx2_table] ; 64 - 38 vpbroadcastd m8, [pw_2048 +r7-w_mask_420_avx2_table] vpbroadcastd m9, [pw_258 +r7-w_mask_420_avx2_table] ; 64 * 4 + 2 pmovzxbd m10, [deint_shuf4+r7-w_mask_420_avx2_table] psubw m9, m0 add wq, r7 W_MASK_420 0, 4 lea stride3q, [strideq*3] jmp wq .w4: vextracti128 xm1, m0, 1 movd [dstq ], xm0 pextrd [dstq+strideq*1], xm0, 1 movd [dstq+strideq*2], xm1 pextrd [dstq+stride3q ], xm1, 1 cmp hd, 4 je .w4_end lea dstq, [dstq+strideq*4] pextrd [dstq ], xm0, 2 pextrd [dstq+strideq*1], xm0, 3 pextrd [dstq+strideq*2], xm1, 2 pextrd [dstq+stride3q ], xm1, 3 cmp hd, 8 jg .w4_h16 .w4_end: vextracti128 xm0, m4, 1 vpblendd xm1, xm4, xm0, 0x05 vpblendd xm4, xm4, xm0, 0x0a pshufd xm1, xm1, q2301 psubw xm4, xm9, xm4 psubw xm4, xm1 psrlw xm4, 2 packuswb xm4, xm4 movq [maskq], xm4 RET .w4_h16: W_MASK_420 2, 5 lea dstq, [dstq+strideq*4] phaddd m4, m5 vextracti128 xm1, m0, 1 psubw m4, m9, m4 psrlw m4, 2 vpermd m4, m10, m4 vextracti128 xm5, m4, 1 packuswb xm4, xm5 movd [dstq ], xm0 pextrd [dstq+strideq*1], xm0, 1 movd [dstq+strideq*2], xm1 pextrd [dstq+stride3q], xm1, 1 lea dstq, [dstq+strideq*4] pextrd [dstq ], xm0, 2 pextrd [dstq+strideq*1], xm0, 3 pextrd [dstq+strideq*2], xm1, 2 pextrd [dstq+stride3q ], xm1, 3 mova [maskq], xm4 RET .w8_loop: add tmp1q, 2*32 add tmp2q, 2*32 W_MASK_420 0, 4 lea dstq, [dstq+strideq*4] add maskq, 8 .w8: vextracti128 xm2, m4, 1 vextracti128 xm1, m0, 1 psubw xm4, xm9, xm4 psubw xm4, xm2 psrlw xm4, 2 packuswb xm4, xm4 movq [dstq ], xm0 movq [dstq+strideq*1], xm1 movhps [dstq+strideq*2], xm0 movhps [dstq+stride3q ], xm1 movq [maskq], xm4 sub hd, 4 jg .w8_loop RET .w16_loop: add tmp1q, 4*32 add tmp2q, 4*32 W_MASK_420 0, 4 lea dstq, [dstq+strideq*4] add maskq, 16 .w16: vpermq m0, m0, q3120 mova [dstq ], xm0 vextracti128 [dstq+strideq*1], m0, 1 W_MASK_420 2, 5 punpckhqdq m1, m4, m5 punpcklqdq m4, m5 psubw m1, m9, m1 psubw m1, m4 psrlw m1, 2 vpermq m0, m0, q3120 packuswb m1, m1 vpermd m1, m10, m1 mova [dstq+strideq*2], xm0 vextracti128 [dstq+stride3q ], m0, 1 mova [maskq], xm1 sub hd, 4 jg .w16_loop RET .w32_loop: add tmp1q, 4*32 add tmp2q, 4*32 W_MASK_420 0, 4 lea dstq, [dstq+strideq*2] add maskq, 16 .w32: vpermq m0, m0, q3120 mova [dstq], m0 W_MASK_420 2, 5 psubw m4, m9, m4 psubw m4, m5 psrlw m4, 2 vpermq m0, m0, q3120 packuswb m4, m4 vpermd m4, m10, m4 mova [dstq+strideq*1], m0 mova [maskq], xm4 sub hd, 2 jg .w32_loop RET .w64_loop_even: psubw m11, m9, m4 psubw m12, m9, m5 dec hd .w64_loop: add tmp1q, 4*32 add tmp2q, 4*32 W_MASK_420 0, 4 add dstq, strideq .w64: vpermq m0, m0, q3120 mova [dstq], m0 W_MASK_420 2, 5 vpermq m0, m0, q3120 mova [dstq+32], m0 test hd, 1 jz .w64_loop_even psubw m4, m11, m4 psubw m5, m12, m5 psrlw m4, 2 psrlw m5, 2 packuswb m4, m5 vpermd m4, m10, m4 mova [maskq], m4 add maskq, 32 dec hd jg .w64_loop RET .w128_loop_even: psubw m13, m9, m4 psubw m14, m9, m5 dec hd .w128_loop: W_MASK_420 0, 4 add dstq, strideq .w128: vpermq m0, m0, q3120 mova [dstq+0*32], m0 W_MASK_420 2, 5 vpermq m0, m0, q3120 mova [dstq+1*32], m0 add tmp1q, 8*32 add tmp2q, 8*32 test hd, 1 jz .w128_even psubw m4, m11, m4 psubw m5, m12, m5 psrlw m4, 2 psrlw m5, 2 packuswb m4, m5 vpermd m4, m10, m4 mova [maskq], m4 jmp .w128_odd .w128_even: psubw m11, m9, m4 psubw m12, m9, m5 .w128_odd: W_MASK_420 -4, 4 vpermq m0, m0, q3120 mova [dstq+2*32], m0 W_MASK_420 -2, 5 vpermq m0, m0, q3120 mova [dstq+3*32], m0 test hd, 1 jz .w128_loop_even psubw m4, m13, m4 psubw m5, m14, m5 psrlw m4, 2 psrlw m5, 2 packuswb m4, m5 vpermd m4, m10, m4 mova [maskq+32], m4 add maskq, 64 dec hd jg .w128_loop RET %endif ; ARCH_X86_64