yading@10: ;***************************************************************************** yading@10: ;* MMX/SSE2/AVX-optimized 10-bit H.264 intra prediction code yading@10: ;***************************************************************************** yading@10: ;* Copyright (C) 2005-2011 x264 project yading@10: ;* yading@10: ;* Authors: Daniel Kang yading@10: ;* yading@10: ;* This file is part of Libav. yading@10: ;* yading@10: ;* Libav is free software; you can redistribute it and/or yading@10: ;* modify it under the terms of the GNU Lesser General Public yading@10: ;* License as published by the Free Software Foundation; either yading@10: ;* version 2.1 of the License, or (at your option) any later version. yading@10: ;* yading@10: ;* Libav is distributed in the hope that it will be useful, yading@10: ;* but WITHOUT ANY WARRANTY; without even the implied warranty of yading@10: ;* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU yading@10: ;* Lesser General Public License for more details. yading@10: ;* yading@10: ;* You should have received a copy of the GNU Lesser General Public yading@10: ;* License along with Libav; if not, write to the Free Software yading@10: ;* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA yading@10: ;****************************************************************************** yading@10: yading@10: %include "libavutil/x86/x86util.asm" yading@10: yading@10: SECTION_RODATA yading@10: yading@10: cextern pw_16 yading@10: cextern pw_8 yading@10: cextern pw_4 yading@10: cextern pw_2 yading@10: cextern pw_1 yading@10: yading@10: pw_m32101234: dw -3, -2, -1, 0, 1, 2, 3, 4 yading@10: pw_m3: times 8 dw -3 yading@10: pw_pixel_max: times 8 dw ((1 << 10)-1) yading@10: pw_512: times 8 dw 512 yading@10: pd_17: times 4 dd 17 yading@10: pd_16: times 4 dd 16 yading@10: yading@10: SECTION .text yading@10: yading@10: ; dest, left, right, src yading@10: ; output: %1 = (t[n-1] + t[n]*2 + t[n+1] + 2) >> 2 yading@10: %macro PRED4x4_LOWPASS 4 yading@10: paddw %2, %3 yading@10: psrlw %2, 1 yading@10: pavgw %1, %4, %2 yading@10: %endmacro yading@10: yading@10: ;----------------------------------------------------------------------------- yading@10: ; void pred4x4_down_right(pixel *src, const pixel *topright, int stride) yading@10: ;----------------------------------------------------------------------------- yading@10: %macro PRED4x4_DR 0 yading@10: cglobal pred4x4_down_right_10, 3, 3 yading@10: sub r0, r2 yading@10: lea r1, [r0+r2*2] yading@10: movhps m1, [r1-8] yading@10: movhps m2, [r0+r2*1-8] yading@10: movhps m4, [r0-8] yading@10: punpckhwd m2, m4 yading@10: movq m3, [r0] yading@10: punpckhdq m1, m2 yading@10: PALIGNR m3, m1, 10, m1 yading@10: movhps m4, [r1+r2*1-8] yading@10: PALIGNR m0, m3, m4, 14, m4 yading@10: movhps m4, [r1+r2*2-8] yading@10: PALIGNR m2, m0, m4, 14, m4 yading@10: PRED4x4_LOWPASS m0, m2, m3, m0 yading@10: movq [r1+r2*2], m0 yading@10: psrldq m0, 2 yading@10: movq [r1+r2*1], m0 yading@10: psrldq m0, 2 yading@10: movq [r0+r2*2], m0 yading@10: psrldq m0, 2 yading@10: movq [r0+r2*1], m0 yading@10: RET yading@10: %endmacro yading@10: yading@10: INIT_XMM sse2 yading@10: PRED4x4_DR yading@10: INIT_XMM ssse3 yading@10: PRED4x4_DR yading@10: %if HAVE_AVX_EXTERNAL yading@10: INIT_XMM avx yading@10: PRED4x4_DR yading@10: %endif yading@10: yading@10: ;----------------------------------------------------------------------------- yading@10: ; void pred4x4_vertical_right(pixel *src, const pixel *topright, int stride) yading@10: ;----------------------------------------------------------------------------- yading@10: %macro PRED4x4_VR 0 yading@10: cglobal pred4x4_vertical_right_10, 3, 3, 6 yading@10: sub r0, r2 yading@10: lea r1, [r0+r2*2] yading@10: movq m5, [r0] ; ........t3t2t1t0 yading@10: movhps m1, [r0-8] yading@10: PALIGNR m0, m5, m1, 14, m1 ; ......t3t2t1t0lt yading@10: pavgw m5, m0 yading@10: movhps m1, [r0+r2*1-8] yading@10: PALIGNR m0, m1, 14, m1 ; ....t3t2t1t0ltl0 yading@10: movhps m2, [r0+r2*2-8] yading@10: PALIGNR m1, m0, m2, 14, m2 ; ..t3t2t1t0ltl0l1 yading@10: movhps m3, [r1+r2*1-8] yading@10: PALIGNR m2, m1, m3, 14, m3 ; t3t2t1t0ltl0l1l2 yading@10: PRED4x4_LOWPASS m1, m0, m2, m1 yading@10: pslldq m0, m1, 12 yading@10: psrldq m1, 4 yading@10: movq [r0+r2*1], m5 yading@10: movq [r0+r2*2], m1 yading@10: PALIGNR m5, m0, 14, m2 yading@10: pslldq m0, 2 yading@10: movq [r1+r2*1], m5 yading@10: PALIGNR m1, m0, 14, m0 yading@10: movq [r1+r2*2], m1 yading@10: RET yading@10: %endmacro yading@10: yading@10: INIT_XMM sse2 yading@10: PRED4x4_VR yading@10: INIT_XMM ssse3 yading@10: PRED4x4_VR yading@10: %if HAVE_AVX_EXTERNAL yading@10: INIT_XMM avx yading@10: PRED4x4_VR yading@10: %endif yading@10: yading@10: ;----------------------------------------------------------------------------- yading@10: ; void pred4x4_horizontal_down(pixel *src, const pixel *topright, int stride) yading@10: ;----------------------------------------------------------------------------- yading@10: %macro PRED4x4_HD 0 yading@10: cglobal pred4x4_horizontal_down_10, 3, 3 yading@10: sub r0, r2 yading@10: lea r1, [r0+r2*2] yading@10: movq m0, [r0-8] ; lt .. yading@10: movhps m0, [r0] yading@10: pslldq m0, 2 ; t2 t1 t0 lt .. .. .. .. yading@10: movq m1, [r1+r2*2-8] ; l3 yading@10: movq m3, [r1+r2*1-8] yading@10: punpcklwd m1, m3 ; l2 l3 yading@10: movq m2, [r0+r2*2-8] ; l1 yading@10: movq m3, [r0+r2*1-8] yading@10: punpcklwd m2, m3 ; l0 l1 yading@10: punpckhdq m1, m2 ; l0 l1 l2 l3 yading@10: punpckhqdq m1, m0 ; t2 t1 t0 lt l0 l1 l2 l3 yading@10: psrldq m0, m1, 4 ; .. .. t2 t1 t0 lt l0 l1 yading@10: psrldq m3, m1, 2 ; .. t2 t1 t0 lt l0 l1 l2 yading@10: pavgw m5, m1, m3 yading@10: PRED4x4_LOWPASS m3, m1, m0, m3 yading@10: punpcklwd m5, m3 yading@10: psrldq m3, 8 yading@10: PALIGNR m3, m5, 12, m4 yading@10: movq [r1+r2*2], m5 yading@10: movhps [r0+r2*2], m5 yading@10: psrldq m5, 4 yading@10: movq [r1+r2*1], m5 yading@10: movq [r0+r2*1], m3 yading@10: RET yading@10: %endmacro yading@10: yading@10: INIT_XMM sse2 yading@10: PRED4x4_HD yading@10: INIT_XMM ssse3 yading@10: PRED4x4_HD yading@10: %if HAVE_AVX_EXTERNAL yading@10: INIT_XMM avx yading@10: PRED4x4_HD yading@10: %endif yading@10: yading@10: ;----------------------------------------------------------------------------- yading@10: ; void pred4x4_dc(pixel *src, const pixel *topright, int stride) yading@10: ;----------------------------------------------------------------------------- yading@10: %macro HADDD 2 ; sum junk yading@10: %if mmsize == 16 yading@10: movhlps %2, %1 yading@10: paddd %1, %2 yading@10: pshuflw %2, %1, 0xE yading@10: paddd %1, %2 yading@10: %else yading@10: pshufw %2, %1, 0xE yading@10: paddd %1, %2 yading@10: %endif yading@10: %endmacro yading@10: yading@10: %macro HADDW 2 yading@10: pmaddwd %1, [pw_1] yading@10: HADDD %1, %2 yading@10: %endmacro yading@10: yading@10: INIT_MMX mmxext yading@10: cglobal pred4x4_dc_10, 3, 3 yading@10: sub r0, r2 yading@10: lea r1, [r0+r2*2] yading@10: movq m2, [r0+r2*1-8] yading@10: paddw m2, [r0+r2*2-8] yading@10: paddw m2, [r1+r2*1-8] yading@10: paddw m2, [r1+r2*2-8] yading@10: psrlq m2, 48 yading@10: movq m0, [r0] yading@10: HADDW m0, m1 yading@10: paddw m0, [pw_4] yading@10: paddw m0, m2 yading@10: psrlw m0, 3 yading@10: SPLATW m0, m0, 0 yading@10: movq [r0+r2*1], m0 yading@10: movq [r0+r2*2], m0 yading@10: movq [r1+r2*1], m0 yading@10: movq [r1+r2*2], m0 yading@10: RET yading@10: yading@10: ;----------------------------------------------------------------------------- yading@10: ; void pred4x4_down_left(pixel *src, const pixel *topright, int stride) yading@10: ;----------------------------------------------------------------------------- yading@10: %macro PRED4x4_DL 0 yading@10: cglobal pred4x4_down_left_10, 3, 3 yading@10: sub r0, r2 yading@10: movq m0, [r0] yading@10: movhps m0, [r1] yading@10: psrldq m2, m0, 2 yading@10: pslldq m3, m0, 2 yading@10: pshufhw m2, m2, 10100100b yading@10: PRED4x4_LOWPASS m0, m3, m2, m0 yading@10: lea r1, [r0+r2*2] yading@10: movhps [r1+r2*2], m0 yading@10: psrldq m0, 2 yading@10: movq [r0+r2*1], m0 yading@10: psrldq m0, 2 yading@10: movq [r0+r2*2], m0 yading@10: psrldq m0, 2 yading@10: movq [r1+r2*1], m0 yading@10: RET yading@10: %endmacro yading@10: yading@10: INIT_XMM sse2 yading@10: PRED4x4_DL yading@10: %if HAVE_AVX_EXTERNAL yading@10: INIT_XMM avx yading@10: PRED4x4_DL yading@10: %endif yading@10: yading@10: ;----------------------------------------------------------------------------- yading@10: ; void pred4x4_vertical_left(pixel *src, const pixel *topright, int stride) yading@10: ;----------------------------------------------------------------------------- yading@10: %macro PRED4x4_VL 0 yading@10: cglobal pred4x4_vertical_left_10, 3, 3 yading@10: sub r0, r2 yading@10: movu m1, [r0] yading@10: movhps m1, [r1] yading@10: psrldq m0, m1, 2 yading@10: psrldq m2, m1, 4 yading@10: pavgw m4, m0, m1 yading@10: PRED4x4_LOWPASS m0, m1, m2, m0 yading@10: lea r1, [r0+r2*2] yading@10: movq [r0+r2*1], m4 yading@10: movq [r0+r2*2], m0 yading@10: psrldq m4, 2 yading@10: psrldq m0, 2 yading@10: movq [r1+r2*1], m4 yading@10: movq [r1+r2*2], m0 yading@10: RET yading@10: %endmacro yading@10: yading@10: INIT_XMM sse2 yading@10: PRED4x4_VL yading@10: %if HAVE_AVX_EXTERNAL yading@10: INIT_XMM avx yading@10: PRED4x4_VL yading@10: %endif yading@10: yading@10: ;----------------------------------------------------------------------------- yading@10: ; void pred4x4_horizontal_up(pixel *src, const pixel *topright, int stride) yading@10: ;----------------------------------------------------------------------------- yading@10: INIT_MMX mmxext yading@10: cglobal pred4x4_horizontal_up_10, 3, 3 yading@10: sub r0, r2 yading@10: lea r1, [r0+r2*2] yading@10: movq m0, [r0+r2*1-8] yading@10: punpckhwd m0, [r0+r2*2-8] yading@10: movq m1, [r1+r2*1-8] yading@10: punpckhwd m1, [r1+r2*2-8] yading@10: punpckhdq m0, m1 yading@10: pshufw m1, m1, 0xFF yading@10: movq [r1+r2*2], m1 yading@10: movd [r1+r2*1+4], m1 yading@10: pshufw m2, m0, 11111001b yading@10: movq m1, m2 yading@10: pavgw m2, m0 yading@10: yading@10: pshufw m5, m0, 11111110b yading@10: PRED4x4_LOWPASS m1, m0, m5, m1 yading@10: movq m6, m2 yading@10: punpcklwd m6, m1 yading@10: movq [r0+r2*1], m6 yading@10: psrlq m2, 16 yading@10: psrlq m1, 16 yading@10: punpcklwd m2, m1 yading@10: movq [r0+r2*2], m2 yading@10: psrlq m2, 32 yading@10: movd [r1+r2*1], m2 yading@10: RET yading@10: yading@10: yading@10: yading@10: ;----------------------------------------------------------------------------- yading@10: ; void pred8x8_vertical(pixel *src, int stride) yading@10: ;----------------------------------------------------------------------------- yading@10: INIT_XMM sse2 yading@10: cglobal pred8x8_vertical_10, 2, 2 yading@10: sub r0, r1 yading@10: mova m0, [r0] yading@10: %rep 3 yading@10: mova [r0+r1*1], m0 yading@10: mova [r0+r1*2], m0 yading@10: lea r0, [r0+r1*2] yading@10: %endrep yading@10: mova [r0+r1*1], m0 yading@10: mova [r0+r1*2], m0 yading@10: RET yading@10: yading@10: ;----------------------------------------------------------------------------- yading@10: ; void pred8x8_horizontal(pixel *src, int stride) yading@10: ;----------------------------------------------------------------------------- yading@10: INIT_XMM sse2 yading@10: cglobal pred8x8_horizontal_10, 2, 3 yading@10: mov r2d, 4 yading@10: .loop: yading@10: movq m0, [r0+r1*0-8] yading@10: movq m1, [r0+r1*1-8] yading@10: pshuflw m0, m0, 0xff yading@10: pshuflw m1, m1, 0xff yading@10: punpcklqdq m0, m0 yading@10: punpcklqdq m1, m1 yading@10: mova [r0+r1*0], m0 yading@10: mova [r0+r1*1], m1 yading@10: lea r0, [r0+r1*2] yading@10: dec r2d yading@10: jg .loop yading@10: REP_RET yading@10: yading@10: ;----------------------------------------------------------------------------- yading@10: ; void predict_8x8_dc(pixel *src, int stride) yading@10: ;----------------------------------------------------------------------------- yading@10: %macro MOV8 2-3 yading@10: ; sort of a hack, but it works yading@10: %if mmsize==8 yading@10: movq [%1+0], %2 yading@10: movq [%1+8], %3 yading@10: %else yading@10: movdqa [%1], %2 yading@10: %endif yading@10: %endmacro yading@10: yading@10: %macro PRED8x8_DC 1 yading@10: cglobal pred8x8_dc_10, 2, 6 yading@10: sub r0, r1 yading@10: pxor m4, m4 yading@10: movq m0, [r0+0] yading@10: movq m1, [r0+8] yading@10: %if mmsize==16 yading@10: punpcklwd m0, m1 yading@10: movhlps m1, m0 yading@10: paddw m0, m1 yading@10: %else yading@10: pshufw m2, m0, 00001110b yading@10: pshufw m3, m1, 00001110b yading@10: paddw m0, m2 yading@10: paddw m1, m3 yading@10: punpcklwd m0, m1 yading@10: %endif yading@10: %1 m2, m0, 00001110b yading@10: paddw m0, m2 yading@10: yading@10: lea r5, [r1*3] yading@10: lea r4, [r0+r1*4] yading@10: movzx r2d, word [r0+r1*1-2] yading@10: movzx r3d, word [r0+r1*2-2] yading@10: add r2d, r3d yading@10: movzx r3d, word [r0+r5*1-2] yading@10: add r2d, r3d yading@10: movzx r3d, word [r4-2] yading@10: add r2d, r3d yading@10: movd m2, r2d ; s2 yading@10: yading@10: movzx r2d, word [r4+r1*1-2] yading@10: movzx r3d, word [r4+r1*2-2] yading@10: add r2d, r3d yading@10: movzx r3d, word [r4+r5*1-2] yading@10: add r2d, r3d yading@10: movzx r3d, word [r4+r1*4-2] yading@10: add r2d, r3d yading@10: movd m3, r2d ; s3 yading@10: yading@10: punpcklwd m2, m3 yading@10: punpckldq m0, m2 ; s0, s1, s2, s3 yading@10: %1 m3, m0, 11110110b ; s2, s1, s3, s3 yading@10: %1 m0, m0, 01110100b ; s0, s1, s3, s1 yading@10: paddw m0, m3 yading@10: psrlw m0, 2 yading@10: pavgw m0, m4 ; s0+s2, s1, s3, s1+s3 yading@10: %if mmsize==16 yading@10: punpcklwd m0, m0 yading@10: pshufd m3, m0, 11111010b yading@10: punpckldq m0, m0 yading@10: SWAP 0,1 yading@10: %else yading@10: pshufw m1, m0, 0x00 yading@10: pshufw m2, m0, 0x55 yading@10: pshufw m3, m0, 0xaa yading@10: pshufw m4, m0, 0xff yading@10: %endif yading@10: MOV8 r0+r1*1, m1, m2 yading@10: MOV8 r0+r1*2, m1, m2 yading@10: MOV8 r0+r5*1, m1, m2 yading@10: MOV8 r0+r1*4, m1, m2 yading@10: MOV8 r4+r1*1, m3, m4 yading@10: MOV8 r4+r1*2, m3, m4 yading@10: MOV8 r4+r5*1, m3, m4 yading@10: MOV8 r4+r1*4, m3, m4 yading@10: RET yading@10: %endmacro yading@10: yading@10: INIT_MMX mmxext yading@10: PRED8x8_DC pshufw yading@10: INIT_XMM sse2 yading@10: PRED8x8_DC pshuflw yading@10: yading@10: ;----------------------------------------------------------------------------- yading@10: ; void pred8x8_top_dc(pixel *src, int stride) yading@10: ;----------------------------------------------------------------------------- yading@10: INIT_XMM sse2 yading@10: cglobal pred8x8_top_dc_10, 2, 4 yading@10: sub r0, r1 yading@10: mova m0, [r0] yading@10: pshuflw m1, m0, 0x4e yading@10: pshufhw m1, m1, 0x4e yading@10: paddw m0, m1 yading@10: pshuflw m1, m0, 0xb1 yading@10: pshufhw m1, m1, 0xb1 yading@10: paddw m0, m1 yading@10: lea r2, [r1*3] yading@10: lea r3, [r0+r1*4] yading@10: paddw m0, [pw_2] yading@10: psrlw m0, 2 yading@10: mova [r0+r1*1], m0 yading@10: mova [r0+r1*2], m0 yading@10: mova [r0+r2*1], m0 yading@10: mova [r0+r1*4], m0 yading@10: mova [r3+r1*1], m0 yading@10: mova [r3+r1*2], m0 yading@10: mova [r3+r2*1], m0 yading@10: mova [r3+r1*4], m0 yading@10: RET yading@10: yading@10: ;----------------------------------------------------------------------------- yading@10: ; void pred8x8_plane(pixel *src, int stride) yading@10: ;----------------------------------------------------------------------------- yading@10: INIT_XMM sse2 yading@10: cglobal pred8x8_plane_10, 2, 7, 7 yading@10: sub r0, r1 yading@10: lea r2, [r1*3] yading@10: lea r3, [r0+r1*4] yading@10: mova m2, [r0] yading@10: pmaddwd m2, [pw_m32101234] yading@10: HADDD m2, m1 yading@10: movd m0, [r0-4] yading@10: psrld m0, 14 yading@10: psubw m2, m0 ; H yading@10: movd m0, [r3+r1*4-4] yading@10: movd m1, [r0+12] yading@10: paddw m0, m1 yading@10: psllw m0, 4 ; 16*(src[7*stride-1] + src[-stride+7]) yading@10: movzx r4d, word [r3+r1*1-2] ; src[4*stride-1] yading@10: movzx r5d, word [r0+r2*1-2] ; src[2*stride-1] yading@10: sub r4d, r5d yading@10: movzx r6d, word [r3+r1*2-2] ; src[5*stride-1] yading@10: movzx r5d, word [r0+r1*2-2] ; src[1*stride-1] yading@10: sub r6d, r5d yading@10: lea r4d, [r4+r6*2] yading@10: movzx r5d, word [r3+r2*1-2] ; src[6*stride-1] yading@10: movzx r6d, word [r0+r1*1-2] ; src[0*stride-1] yading@10: sub r5d, r6d yading@10: lea r5d, [r5*3] yading@10: add r4d, r5d yading@10: movzx r6d, word [r3+r1*4-2] ; src[7*stride-1] yading@10: movzx r5d, word [r0+r1*0-2] ; src[ -stride-1] yading@10: sub r6d, r5d yading@10: lea r4d, [r4+r6*4] yading@10: movd m3, r4d ; V yading@10: punpckldq m2, m3 yading@10: pmaddwd m2, [pd_17] yading@10: paddd m2, [pd_16] yading@10: psrad m2, 5 ; b, c yading@10: yading@10: mova m3, [pw_pixel_max] yading@10: pxor m1, m1 yading@10: SPLATW m0, m0, 1 yading@10: SPLATW m4, m2, 2 yading@10: SPLATW m2, m2, 0 yading@10: pmullw m2, [pw_m32101234] ; b yading@10: pmullw m5, m4, [pw_m3] ; c yading@10: paddw m5, [pw_16] yading@10: mov r2d, 8 yading@10: add r0, r1 yading@10: .loop: yading@10: paddsw m6, m2, m5 yading@10: paddsw m6, m0 yading@10: psraw m6, 5 yading@10: CLIPW m6, m1, m3 yading@10: mova [r0], m6 yading@10: paddw m5, m4 yading@10: add r0, r1 yading@10: dec r2d yading@10: jg .loop yading@10: REP_RET yading@10: yading@10: yading@10: ;----------------------------------------------------------------------------- yading@10: ; void pred8x8l_128_dc(pixel *src, int has_topleft, int has_topright, int stride) yading@10: ;----------------------------------------------------------------------------- yading@10: %macro PRED8x8L_128_DC 0 yading@10: cglobal pred8x8l_128_dc_10, 4, 4 yading@10: mova m0, [pw_512] ; (1<<(BIT_DEPTH-1)) yading@10: lea r1, [r3*3] yading@10: lea r2, [r0+r3*4] yading@10: MOV8 r0+r3*0, m0, m0 yading@10: MOV8 r0+r3*1, m0, m0 yading@10: MOV8 r0+r3*2, m0, m0 yading@10: MOV8 r0+r1*1, m0, m0 yading@10: MOV8 r2+r3*0, m0, m0 yading@10: MOV8 r2+r3*1, m0, m0 yading@10: MOV8 r2+r3*2, m0, m0 yading@10: MOV8 r2+r1*1, m0, m0 yading@10: RET yading@10: %endmacro yading@10: yading@10: INIT_MMX mmxext yading@10: PRED8x8L_128_DC yading@10: INIT_XMM sse2 yading@10: PRED8x8L_128_DC yading@10: yading@10: ;----------------------------------------------------------------------------- yading@10: ; void pred8x8l_top_dc(pixel *src, int has_topleft, int has_topright, int stride) yading@10: ;----------------------------------------------------------------------------- yading@10: %macro PRED8x8L_TOP_DC 0 yading@10: cglobal pred8x8l_top_dc_10, 4, 4, 6 yading@10: sub r0, r3 yading@10: mova m0, [r0] yading@10: shr r1d, 14 yading@10: shr r2d, 13 yading@10: neg r1 yading@10: pslldq m1, m0, 2 yading@10: psrldq m2, m0, 2 yading@10: pinsrw m1, [r0+r1], 0 yading@10: pinsrw m2, [r0+r2+14], 7 yading@10: lea r1, [r3*3] yading@10: lea r2, [r0+r3*4] yading@10: PRED4x4_LOWPASS m0, m2, m1, m0 yading@10: HADDW m0, m1 yading@10: paddw m0, [pw_4] yading@10: psrlw m0, 3 yading@10: SPLATW m0, m0, 0 yading@10: mova [r0+r3*1], m0 yading@10: mova [r0+r3*2], m0 yading@10: mova [r0+r1*1], m0 yading@10: mova [r0+r3*4], m0 yading@10: mova [r2+r3*1], m0 yading@10: mova [r2+r3*2], m0 yading@10: mova [r2+r1*1], m0 yading@10: mova [r2+r3*4], m0 yading@10: RET yading@10: %endmacro yading@10: yading@10: INIT_XMM sse2 yading@10: PRED8x8L_TOP_DC yading@10: %if HAVE_AVX_EXTERNAL yading@10: INIT_XMM avx yading@10: PRED8x8L_TOP_DC yading@10: %endif yading@10: yading@10: ;----------------------------------------------------------------------------- yading@10: ;void pred8x8l_dc(pixel *src, int has_topleft, int has_topright, int stride) yading@10: ;----------------------------------------------------------------------------- yading@10: ;TODO: see if scalar is faster yading@10: %macro PRED8x8L_DC 0 yading@10: cglobal pred8x8l_dc_10, 4, 6, 6 yading@10: sub r0, r3 yading@10: lea r4, [r0+r3*4] yading@10: lea r5, [r3*3] yading@10: mova m0, [r0+r3*2-16] yading@10: punpckhwd m0, [r0+r3*1-16] yading@10: mova m1, [r4+r3*0-16] yading@10: punpckhwd m1, [r0+r5*1-16] yading@10: punpckhdq m1, m0 yading@10: mova m2, [r4+r3*2-16] yading@10: punpckhwd m2, [r4+r3*1-16] yading@10: mova m3, [r4+r3*4-16] yading@10: punpckhwd m3, [r4+r5*1-16] yading@10: punpckhdq m3, m2 yading@10: punpckhqdq m3, m1 yading@10: mova m0, [r0] yading@10: shr r1d, 14 yading@10: shr r2d, 13 yading@10: neg r1 yading@10: pslldq m1, m0, 2 yading@10: psrldq m2, m0, 2 yading@10: pinsrw m1, [r0+r1], 0 yading@10: pinsrw m2, [r0+r2+14], 7 yading@10: not r1 yading@10: and r1, r3 yading@10: pslldq m4, m3, 2 yading@10: psrldq m5, m3, 2 yading@10: pshuflw m4, m4, 11100101b yading@10: pinsrw m5, [r0+r1-2], 7 yading@10: PRED4x4_LOWPASS m3, m4, m5, m3 yading@10: PRED4x4_LOWPASS m0, m2, m1, m0 yading@10: paddw m0, m3 yading@10: HADDW m0, m1 yading@10: paddw m0, [pw_8] yading@10: psrlw m0, 4 yading@10: SPLATW m0, m0 yading@10: mova [r0+r3*1], m0 yading@10: mova [r0+r3*2], m0 yading@10: mova [r0+r5*1], m0 yading@10: mova [r0+r3*4], m0 yading@10: mova [r4+r3*1], m0 yading@10: mova [r4+r3*2], m0 yading@10: mova [r4+r5*1], m0 yading@10: mova [r4+r3*4], m0 yading@10: RET yading@10: %endmacro yading@10: yading@10: INIT_XMM sse2 yading@10: PRED8x8L_DC yading@10: %if HAVE_AVX_EXTERNAL yading@10: INIT_XMM avx yading@10: PRED8x8L_DC yading@10: %endif yading@10: yading@10: ;----------------------------------------------------------------------------- yading@10: ; void pred8x8l_vertical(pixel *src, int has_topleft, int has_topright, int stride) yading@10: ;----------------------------------------------------------------------------- yading@10: %macro PRED8x8L_VERTICAL 0 yading@10: cglobal pred8x8l_vertical_10, 4, 4, 6 yading@10: sub r0, r3 yading@10: mova m0, [r0] yading@10: shr r1d, 14 yading@10: shr r2d, 13 yading@10: neg r1 yading@10: pslldq m1, m0, 2 yading@10: psrldq m2, m0, 2 yading@10: pinsrw m1, [r0+r1], 0 yading@10: pinsrw m2, [r0+r2+14], 7 yading@10: lea r1, [r3*3] yading@10: lea r2, [r0+r3*4] yading@10: PRED4x4_LOWPASS m0, m2, m1, m0 yading@10: mova [r0+r3*1], m0 yading@10: mova [r0+r3*2], m0 yading@10: mova [r0+r1*1], m0 yading@10: mova [r0+r3*4], m0 yading@10: mova [r2+r3*1], m0 yading@10: mova [r2+r3*2], m0 yading@10: mova [r2+r1*1], m0 yading@10: mova [r2+r3*4], m0 yading@10: RET yading@10: %endmacro yading@10: yading@10: INIT_XMM sse2 yading@10: PRED8x8L_VERTICAL yading@10: %if HAVE_AVX_EXTERNAL yading@10: INIT_XMM avx yading@10: PRED8x8L_VERTICAL yading@10: %endif yading@10: yading@10: ;----------------------------------------------------------------------------- yading@10: ; void pred8x8l_horizontal(uint8_t *src, int has_topleft, int has_topright, int stride) yading@10: ;----------------------------------------------------------------------------- yading@10: %macro PRED8x8L_HORIZONTAL 0 yading@10: cglobal pred8x8l_horizontal_10, 4, 4, 5 yading@10: mova m0, [r0-16] yading@10: shr r1d, 14 yading@10: dec r1 yading@10: and r1, r3 yading@10: sub r1, r3 yading@10: punpckhwd m0, [r0+r1-16] yading@10: mova m1, [r0+r3*2-16] yading@10: punpckhwd m1, [r0+r3*1-16] yading@10: lea r2, [r0+r3*4] yading@10: lea r1, [r3*3] yading@10: punpckhdq m1, m0 yading@10: mova m2, [r2+r3*0-16] yading@10: punpckhwd m2, [r0+r1-16] yading@10: mova m3, [r2+r3*2-16] yading@10: punpckhwd m3, [r2+r3*1-16] yading@10: punpckhdq m3, m2 yading@10: punpckhqdq m3, m1 yading@10: PALIGNR m4, m3, [r2+r1-16], 14, m0 yading@10: pslldq m0, m4, 2 yading@10: pshuflw m0, m0, 11100101b yading@10: PRED4x4_LOWPASS m4, m3, m0, m4 yading@10: punpckhwd m3, m4, m4 yading@10: punpcklwd m4, m4 yading@10: pshufd m0, m3, 0xff yading@10: pshufd m1, m3, 0xaa yading@10: pshufd m2, m3, 0x55 yading@10: pshufd m3, m3, 0x00 yading@10: mova [r0+r3*0], m0 yading@10: mova [r0+r3*1], m1 yading@10: mova [r0+r3*2], m2 yading@10: mova [r0+r1*1], m3 yading@10: pshufd m0, m4, 0xff yading@10: pshufd m1, m4, 0xaa yading@10: pshufd m2, m4, 0x55 yading@10: pshufd m3, m4, 0x00 yading@10: mova [r2+r3*0], m0 yading@10: mova [r2+r3*1], m1 yading@10: mova [r2+r3*2], m2 yading@10: mova [r2+r1*1], m3 yading@10: RET yading@10: %endmacro yading@10: yading@10: INIT_XMM sse2 yading@10: PRED8x8L_HORIZONTAL yading@10: INIT_XMM ssse3 yading@10: PRED8x8L_HORIZONTAL yading@10: %if HAVE_AVX_EXTERNAL yading@10: INIT_XMM avx yading@10: PRED8x8L_HORIZONTAL yading@10: %endif yading@10: yading@10: ;----------------------------------------------------------------------------- yading@10: ;void pred8x8l_down_left(pixel *src, int has_topleft, int has_topright, int stride) yading@10: ;----------------------------------------------------------------------------- yading@10: %macro PRED8x8L_DOWN_LEFT 0 yading@10: cglobal pred8x8l_down_left_10, 4, 4, 7 yading@10: sub r0, r3 yading@10: mova m3, [r0] yading@10: shr r1d, 14 yading@10: neg r1 yading@10: shr r2d, 13 yading@10: pslldq m1, m3, 2 yading@10: psrldq m2, m3, 2 yading@10: pinsrw m1, [r0+r1], 0 yading@10: pinsrw m2, [r0+r2+14], 7 yading@10: PRED4x4_LOWPASS m6, m2, m1, m3 yading@10: jz .fix_tr ; flags from shr r2d yading@10: mova m1, [r0+16] yading@10: psrldq m5, m1, 2 yading@10: PALIGNR m2, m1, m3, 14, m3 yading@10: pshufhw m5, m5, 10100100b yading@10: PRED4x4_LOWPASS m1, m2, m5, m1 yading@10: .do_topright: yading@10: lea r1, [r3*3] yading@10: psrldq m5, m1, 14 yading@10: lea r2, [r0+r3*4] yading@10: PALIGNR m2, m1, m6, 2, m0 yading@10: PALIGNR m3, m1, m6, 14, m0 yading@10: PALIGNR m5, m1, 2, m0 yading@10: pslldq m4, m6, 2 yading@10: PRED4x4_LOWPASS m6, m4, m2, m6 yading@10: PRED4x4_LOWPASS m1, m3, m5, m1 yading@10: mova [r2+r3*4], m1 yading@10: PALIGNR m1, m6, 14, m2 yading@10: pslldq m6, 2 yading@10: mova [r2+r1*1], m1 yading@10: PALIGNR m1, m6, 14, m2 yading@10: pslldq m6, 2 yading@10: mova [r2+r3*2], m1 yading@10: PALIGNR m1, m6, 14, m2 yading@10: pslldq m6, 2 yading@10: mova [r2+r3*1], m1 yading@10: PALIGNR m1, m6, 14, m2 yading@10: pslldq m6, 2 yading@10: mova [r0+r3*4], m1 yading@10: PALIGNR m1, m6, 14, m2 yading@10: pslldq m6, 2 yading@10: mova [r0+r1*1], m1 yading@10: PALIGNR m1, m6, 14, m2 yading@10: pslldq m6, 2 yading@10: mova [r0+r3*2], m1 yading@10: PALIGNR m1, m6, 14, m6 yading@10: mova [r0+r3*1], m1 yading@10: RET yading@10: .fix_tr: yading@10: punpckhwd m3, m3 yading@10: pshufd m1, m3, 0xFF yading@10: jmp .do_topright yading@10: %endmacro yading@10: yading@10: INIT_XMM sse2 yading@10: PRED8x8L_DOWN_LEFT yading@10: INIT_XMM ssse3 yading@10: PRED8x8L_DOWN_LEFT yading@10: %if HAVE_AVX_EXTERNAL yading@10: INIT_XMM avx yading@10: PRED8x8L_DOWN_LEFT yading@10: %endif yading@10: yading@10: ;----------------------------------------------------------------------------- yading@10: ;void pred8x8l_down_right(pixel *src, int has_topleft, int has_topright, int stride) yading@10: ;----------------------------------------------------------------------------- yading@10: %macro PRED8x8L_DOWN_RIGHT 0 yading@10: ; standard forbids this when has_topleft is false yading@10: ; no need to check yading@10: cglobal pred8x8l_down_right_10, 4, 5, 8 yading@10: sub r0, r3 yading@10: lea r4, [r0+r3*4] yading@10: lea r1, [r3*3] yading@10: mova m0, [r0+r3*1-16] yading@10: punpckhwd m0, [r0+r3*0-16] yading@10: mova m1, [r0+r1*1-16] yading@10: punpckhwd m1, [r0+r3*2-16] yading@10: punpckhdq m1, m0 yading@10: mova m2, [r4+r3*1-16] yading@10: punpckhwd m2, [r4+r3*0-16] yading@10: mova m3, [r4+r1*1-16] yading@10: punpckhwd m3, [r4+r3*2-16] yading@10: punpckhdq m3, m2 yading@10: punpckhqdq m3, m1 yading@10: mova m0, [r4+r3*4-16] yading@10: mova m1, [r0] yading@10: PALIGNR m4, m3, m0, 14, m0 yading@10: PALIGNR m1, m3, 2, m2 yading@10: pslldq m0, m4, 2 yading@10: pshuflw m0, m0, 11100101b yading@10: PRED4x4_LOWPASS m6, m1, m4, m3 yading@10: PRED4x4_LOWPASS m4, m3, m0, m4 yading@10: mova m3, [r0] yading@10: shr r2d, 13 yading@10: pslldq m1, m3, 2 yading@10: psrldq m2, m3, 2 yading@10: pinsrw m1, [r0-2], 0 yading@10: pinsrw m2, [r0+r2+14], 7 yading@10: PRED4x4_LOWPASS m3, m2, m1, m3 yading@10: PALIGNR m2, m3, m6, 2, m0 yading@10: PALIGNR m5, m3, m6, 14, m0 yading@10: psrldq m7, m3, 2 yading@10: PRED4x4_LOWPASS m6, m4, m2, m6 yading@10: PRED4x4_LOWPASS m3, m5, m7, m3 yading@10: mova [r4+r3*4], m6 yading@10: PALIGNR m3, m6, 14, m2 yading@10: pslldq m6, 2 yading@10: mova [r0+r3*1], m3 yading@10: PALIGNR m3, m6, 14, m2 yading@10: pslldq m6, 2 yading@10: mova [r0+r3*2], m3 yading@10: PALIGNR m3, m6, 14, m2 yading@10: pslldq m6, 2 yading@10: mova [r0+r1*1], m3 yading@10: PALIGNR m3, m6, 14, m2 yading@10: pslldq m6, 2 yading@10: mova [r0+r3*4], m3 yading@10: PALIGNR m3, m6, 14, m2 yading@10: pslldq m6, 2 yading@10: mova [r4+r3*1], m3 yading@10: PALIGNR m3, m6, 14, m2 yading@10: pslldq m6, 2 yading@10: mova [r4+r3*2], m3 yading@10: PALIGNR m3, m6, 14, m6 yading@10: mova [r4+r1*1], m3 yading@10: RET yading@10: %endmacro yading@10: yading@10: INIT_XMM sse2 yading@10: PRED8x8L_DOWN_RIGHT yading@10: INIT_XMM ssse3 yading@10: PRED8x8L_DOWN_RIGHT yading@10: %if HAVE_AVX_EXTERNAL yading@10: INIT_XMM avx yading@10: PRED8x8L_DOWN_RIGHT yading@10: %endif yading@10: yading@10: ;----------------------------------------------------------------------------- yading@10: ; void pred8x8l_vertical_right(pixel *src, int has_topleft, int has_topright, int stride) yading@10: ;----------------------------------------------------------------------------- yading@10: %macro PRED8x8L_VERTICAL_RIGHT 0 yading@10: ; likewise with 8x8l_down_right yading@10: cglobal pred8x8l_vertical_right_10, 4, 5, 7 yading@10: sub r0, r3 yading@10: lea r4, [r0+r3*4] yading@10: lea r1, [r3*3] yading@10: mova m0, [r0+r3*1-16] yading@10: punpckhwd m0, [r0+r3*0-16] yading@10: mova m1, [r0+r1*1-16] yading@10: punpckhwd m1, [r0+r3*2-16] yading@10: punpckhdq m1, m0 yading@10: mova m2, [r4+r3*1-16] yading@10: punpckhwd m2, [r4+r3*0-16] yading@10: mova m3, [r4+r1*1-16] yading@10: punpckhwd m3, [r4+r3*2-16] yading@10: punpckhdq m3, m2 yading@10: punpckhqdq m3, m1 yading@10: mova m0, [r4+r3*4-16] yading@10: mova m1, [r0] yading@10: PALIGNR m4, m3, m0, 14, m0 yading@10: PALIGNR m1, m3, 2, m2 yading@10: PRED4x4_LOWPASS m3, m1, m4, m3 yading@10: mova m2, [r0] yading@10: shr r2d, 13 yading@10: pslldq m1, m2, 2 yading@10: psrldq m5, m2, 2 yading@10: pinsrw m1, [r0-2], 0 yading@10: pinsrw m5, [r0+r2+14], 7 yading@10: PRED4x4_LOWPASS m2, m5, m1, m2 yading@10: PALIGNR m6, m2, m3, 12, m1 yading@10: PALIGNR m5, m2, m3, 14, m0 yading@10: PRED4x4_LOWPASS m0, m6, m2, m5 yading@10: pavgw m2, m5 yading@10: mova [r0+r3*2], m0 yading@10: mova [r0+r3*1], m2 yading@10: pslldq m6, m3, 4 yading@10: pslldq m1, m3, 2 yading@10: PRED4x4_LOWPASS m1, m3, m6, m1 yading@10: PALIGNR m2, m1, 14, m4 yading@10: mova [r0+r1*1], m2 yading@10: pslldq m1, 2 yading@10: PALIGNR m0, m1, 14, m3 yading@10: mova [r0+r3*4], m0 yading@10: pslldq m1, 2 yading@10: PALIGNR m2, m1, 14, m4 yading@10: mova [r4+r3*1], m2 yading@10: pslldq m1, 2 yading@10: PALIGNR m0, m1, 14, m3 yading@10: mova [r4+r3*2], m0 yading@10: pslldq m1, 2 yading@10: PALIGNR m2, m1, 14, m4 yading@10: mova [r4+r1*1], m2 yading@10: pslldq m1, 2 yading@10: PALIGNR m0, m1, 14, m1 yading@10: mova [r4+r3*4], m0 yading@10: RET yading@10: %endmacro yading@10: yading@10: INIT_XMM sse2 yading@10: PRED8x8L_VERTICAL_RIGHT yading@10: INIT_XMM ssse3 yading@10: PRED8x8L_VERTICAL_RIGHT yading@10: %if HAVE_AVX_EXTERNAL yading@10: INIT_XMM avx yading@10: PRED8x8L_VERTICAL_RIGHT yading@10: %endif yading@10: yading@10: ;----------------------------------------------------------------------------- yading@10: ; void pred8x8l_horizontal_up(pixel *src, int has_topleft, int has_topright, int stride) yading@10: ;----------------------------------------------------------------------------- yading@10: %macro PRED8x8L_HORIZONTAL_UP 0 yading@10: cglobal pred8x8l_horizontal_up_10, 4, 4, 6 yading@10: mova m0, [r0+r3*0-16] yading@10: punpckhwd m0, [r0+r3*1-16] yading@10: shr r1d, 14 yading@10: dec r1 yading@10: and r1, r3 yading@10: sub r1, r3 yading@10: mova m4, [r0+r1*1-16] yading@10: lea r1, [r3*3] yading@10: lea r2, [r0+r3*4] yading@10: mova m1, [r0+r3*2-16] yading@10: punpckhwd m1, [r0+r1*1-16] yading@10: punpckhdq m0, m1 yading@10: mova m2, [r2+r3*0-16] yading@10: punpckhwd m2, [r2+r3*1-16] yading@10: mova m3, [r2+r3*2-16] yading@10: punpckhwd m3, [r2+r1*1-16] yading@10: punpckhdq m2, m3 yading@10: punpckhqdq m0, m2 yading@10: PALIGNR m1, m0, m4, 14, m4 yading@10: psrldq m2, m0, 2 yading@10: pshufhw m2, m2, 10100100b yading@10: PRED4x4_LOWPASS m0, m1, m2, m0 yading@10: psrldq m1, m0, 2 yading@10: psrldq m2, m0, 4 yading@10: pshufhw m1, m1, 10100100b yading@10: pshufhw m2, m2, 01010100b yading@10: pavgw m4, m0, m1 yading@10: PRED4x4_LOWPASS m1, m2, m0, m1 yading@10: punpckhwd m5, m4, m1 yading@10: punpcklwd m4, m1 yading@10: mova [r2+r3*0], m5 yading@10: mova [r0+r3*0], m4 yading@10: pshufd m0, m5, 11111001b yading@10: pshufd m1, m5, 11111110b yading@10: pshufd m2, m5, 11111111b yading@10: mova [r2+r3*1], m0 yading@10: mova [r2+r3*2], m1 yading@10: mova [r2+r1*1], m2 yading@10: PALIGNR m2, m5, m4, 4, m0 yading@10: PALIGNR m3, m5, m4, 8, m1 yading@10: PALIGNR m5, m5, m4, 12, m4 yading@10: mova [r0+r3*1], m2 yading@10: mova [r0+r3*2], m3 yading@10: mova [r0+r1*1], m5 yading@10: RET yading@10: %endmacro yading@10: yading@10: INIT_XMM sse2 yading@10: PRED8x8L_HORIZONTAL_UP yading@10: INIT_XMM ssse3 yading@10: PRED8x8L_HORIZONTAL_UP yading@10: %if HAVE_AVX_EXTERNAL yading@10: INIT_XMM avx yading@10: PRED8x8L_HORIZONTAL_UP yading@10: %endif yading@10: yading@10: yading@10: ;----------------------------------------------------------------------------- yading@10: ; void pred16x16_vertical(pixel *src, int stride) yading@10: ;----------------------------------------------------------------------------- yading@10: %macro MOV16 3-5 yading@10: mova [%1+ 0], %2 yading@10: mova [%1+mmsize], %3 yading@10: %if mmsize==8 yading@10: mova [%1+ 16], %4 yading@10: mova [%1+ 24], %5 yading@10: %endif yading@10: %endmacro yading@10: yading@10: %macro PRED16x16_VERTICAL 0 yading@10: cglobal pred16x16_vertical_10, 2, 3 yading@10: sub r0, r1 yading@10: mov r2d, 8 yading@10: mova m0, [r0+ 0] yading@10: mova m1, [r0+mmsize] yading@10: %if mmsize==8 yading@10: mova m2, [r0+16] yading@10: mova m3, [r0+24] yading@10: %endif yading@10: .loop: yading@10: MOV16 r0+r1*1, m0, m1, m2, m3 yading@10: MOV16 r0+r1*2, m0, m1, m2, m3 yading@10: lea r0, [r0+r1*2] yading@10: dec r2d yading@10: jg .loop yading@10: REP_RET yading@10: %endmacro yading@10: yading@10: INIT_MMX mmxext yading@10: PRED16x16_VERTICAL yading@10: INIT_XMM sse2 yading@10: PRED16x16_VERTICAL yading@10: yading@10: ;----------------------------------------------------------------------------- yading@10: ; void pred16x16_horizontal(pixel *src, int stride) yading@10: ;----------------------------------------------------------------------------- yading@10: %macro PRED16x16_HORIZONTAL 0 yading@10: cglobal pred16x16_horizontal_10, 2, 3 yading@10: mov r2d, 8 yading@10: .vloop: yading@10: movd m0, [r0+r1*0-4] yading@10: movd m1, [r0+r1*1-4] yading@10: SPLATW m0, m0, 1 yading@10: SPLATW m1, m1, 1 yading@10: MOV16 r0+r1*0, m0, m0, m0, m0 yading@10: MOV16 r0+r1*1, m1, m1, m1, m1 yading@10: lea r0, [r0+r1*2] yading@10: dec r2d yading@10: jg .vloop yading@10: REP_RET yading@10: %endmacro yading@10: yading@10: INIT_MMX mmxext yading@10: PRED16x16_HORIZONTAL yading@10: INIT_XMM sse2 yading@10: PRED16x16_HORIZONTAL yading@10: yading@10: ;----------------------------------------------------------------------------- yading@10: ; void pred16x16_dc(pixel *src, int stride) yading@10: ;----------------------------------------------------------------------------- yading@10: %macro PRED16x16_DC 0 yading@10: cglobal pred16x16_dc_10, 2, 6 yading@10: mov r5, r0 yading@10: sub r0, r1 yading@10: mova m0, [r0+0] yading@10: paddw m0, [r0+mmsize] yading@10: %if mmsize==8 yading@10: paddw m0, [r0+16] yading@10: paddw m0, [r0+24] yading@10: %endif yading@10: HADDW m0, m2 yading@10: yading@10: lea r0, [r0+r1-2] yading@10: movzx r3d, word [r0] yading@10: movzx r4d, word [r0+r1] yading@10: %rep 7 yading@10: lea r0, [r0+r1*2] yading@10: movzx r2d, word [r0] yading@10: add r3d, r2d yading@10: movzx r2d, word [r0+r1] yading@10: add r4d, r2d yading@10: %endrep yading@10: lea r3d, [r3+r4+16] yading@10: yading@10: movd m1, r3d yading@10: paddw m0, m1 yading@10: psrlw m0, 5 yading@10: SPLATW m0, m0 yading@10: mov r3d, 8 yading@10: .loop: yading@10: MOV16 r5+r1*0, m0, m0, m0, m0 yading@10: MOV16 r5+r1*1, m0, m0, m0, m0 yading@10: lea r5, [r5+r1*2] yading@10: dec r3d yading@10: jg .loop yading@10: REP_RET yading@10: %endmacro yading@10: yading@10: INIT_MMX mmxext yading@10: PRED16x16_DC yading@10: INIT_XMM sse2 yading@10: PRED16x16_DC yading@10: yading@10: ;----------------------------------------------------------------------------- yading@10: ; void pred16x16_top_dc(pixel *src, int stride) yading@10: ;----------------------------------------------------------------------------- yading@10: %macro PRED16x16_TOP_DC 0 yading@10: cglobal pred16x16_top_dc_10, 2, 3 yading@10: sub r0, r1 yading@10: mova m0, [r0+0] yading@10: paddw m0, [r0+mmsize] yading@10: %if mmsize==8 yading@10: paddw m0, [r0+16] yading@10: paddw m0, [r0+24] yading@10: %endif yading@10: HADDW m0, m2 yading@10: yading@10: SPLATW m0, m0 yading@10: paddw m0, [pw_8] yading@10: psrlw m0, 4 yading@10: mov r2d, 8 yading@10: .loop: yading@10: MOV16 r0+r1*1, m0, m0, m0, m0 yading@10: MOV16 r0+r1*2, m0, m0, m0, m0 yading@10: lea r0, [r0+r1*2] yading@10: dec r2d yading@10: jg .loop yading@10: REP_RET yading@10: %endmacro yading@10: yading@10: INIT_MMX mmxext yading@10: PRED16x16_TOP_DC yading@10: INIT_XMM sse2 yading@10: PRED16x16_TOP_DC yading@10: yading@10: ;----------------------------------------------------------------------------- yading@10: ; void pred16x16_left_dc(pixel *src, int stride) yading@10: ;----------------------------------------------------------------------------- yading@10: %macro PRED16x16_LEFT_DC 0 yading@10: cglobal pred16x16_left_dc_10, 2, 6 yading@10: mov r5, r0 yading@10: yading@10: sub r0, 2 yading@10: movzx r3d, word [r0] yading@10: movzx r4d, word [r0+r1] yading@10: %rep 7 yading@10: lea r0, [r0+r1*2] yading@10: movzx r2d, word [r0] yading@10: add r3d, r2d yading@10: movzx r2d, word [r0+r1] yading@10: add r4d, r2d yading@10: %endrep yading@10: lea r3d, [r3+r4+8] yading@10: shr r3d, 4 yading@10: yading@10: movd m0, r3d yading@10: SPLATW m0, m0 yading@10: mov r3d, 8 yading@10: .loop: yading@10: MOV16 r5+r1*0, m0, m0, m0, m0 yading@10: MOV16 r5+r1*1, m0, m0, m0, m0 yading@10: lea r5, [r5+r1*2] yading@10: dec r3d yading@10: jg .loop yading@10: REP_RET yading@10: %endmacro yading@10: yading@10: INIT_MMX mmxext yading@10: PRED16x16_LEFT_DC yading@10: INIT_XMM sse2 yading@10: PRED16x16_LEFT_DC yading@10: yading@10: ;----------------------------------------------------------------------------- yading@10: ; void pred16x16_128_dc(pixel *src, int stride) yading@10: ;----------------------------------------------------------------------------- yading@10: %macro PRED16x16_128_DC 0 yading@10: cglobal pred16x16_128_dc_10, 2,3 yading@10: mova m0, [pw_512] yading@10: mov r2d, 8 yading@10: .loop: yading@10: MOV16 r0+r1*0, m0, m0, m0, m0 yading@10: MOV16 r0+r1*1, m0, m0, m0, m0 yading@10: lea r0, [r0+r1*2] yading@10: dec r2d yading@10: jg .loop yading@10: REP_RET yading@10: %endmacro yading@10: yading@10: INIT_MMX mmxext yading@10: PRED16x16_128_DC yading@10: INIT_XMM sse2 yading@10: PRED16x16_128_DC