yading@10: ;****************************************************************************** yading@10: ;* yading@10: ;* Copyright (c) 2000-2001 Fabrice Bellard yading@10: ;* Copyright (c) Nick Kurshev yading@10: ;* Copyright (c) 2002 Michael Niedermayer yading@10: ;* Copyright (c) 2002 Zdenek Kabelac yading@10: ;* Copyright (c) 2013 Daniel Kang yading@10: ;* yading@10: ;* MMX optimized hpel functions yading@10: ;* yading@10: ;* This file is part of FFmpeg. yading@10: ;* yading@10: ;* FFmpeg 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: ;* FFmpeg 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 FFmpeg; 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: cextern pb_1 yading@10: yading@10: SECTION_TEXT yading@10: yading@10: ; put_pixels8_x2(uint8_t *block, const uint8_t *pixels, ptrdiff_t line_size, int h) yading@10: %macro PUT_PIXELS8_X2 0 yading@10: cglobal put_pixels8_x2, 4,5 yading@10: lea r4, [r2*2] yading@10: .loop: yading@10: mova m0, [r1] yading@10: mova m1, [r1+r2] yading@10: PAVGB m0, [r1+1] yading@10: PAVGB m1, [r1+r2+1] yading@10: mova [r0], m0 yading@10: mova [r0+r2], m1 yading@10: add r1, r4 yading@10: add r0, r4 yading@10: mova m0, [r1] yading@10: mova m1, [r1+r2] yading@10: PAVGB m0, [r1+1] yading@10: PAVGB m1, [r1+r2+1] yading@10: add r1, r4 yading@10: mova [r0], m0 yading@10: mova [r0+r2], m1 yading@10: add r0, r4 yading@10: sub r3d, 4 yading@10: jne .loop yading@10: REP_RET yading@10: %endmacro yading@10: yading@10: INIT_MMX mmxext yading@10: PUT_PIXELS8_X2 yading@10: INIT_MMX 3dnow yading@10: PUT_PIXELS8_X2 yading@10: yading@10: yading@10: ; put_pixels16_x2(uint8_t *block, const uint8_t *pixels, ptrdiff_t line_size, int h) yading@10: %macro PUT_PIXELS_16 0 yading@10: cglobal put_pixels16_x2, 4,5 yading@10: lea r4, [r2*2] yading@10: .loop: yading@10: mova m0, [r1] yading@10: mova m1, [r1+r2] yading@10: mova m2, [r1+8] yading@10: mova m3, [r1+r2+8] yading@10: PAVGB m0, [r1+1] yading@10: PAVGB m1, [r1+r2+1] yading@10: PAVGB m2, [r1+9] yading@10: PAVGB m3, [r1+r2+9] yading@10: mova [r0], m0 yading@10: mova [r0+r2], m1 yading@10: mova [r0+8], m2 yading@10: mova [r0+r2+8], m3 yading@10: add r1, r4 yading@10: add r0, r4 yading@10: mova m0, [r1] yading@10: mova m1, [r1+r2] yading@10: mova m2, [r1+8] yading@10: mova m3, [r1+r2+8] yading@10: PAVGB m0, [r1+1] yading@10: PAVGB m1, [r1+r2+1] yading@10: PAVGB m2, [r1+9] yading@10: PAVGB m3, [r1+r2+9] yading@10: add r1, r4 yading@10: mova [r0], m0 yading@10: mova [r0+r2], m1 yading@10: mova [r0+8], m2 yading@10: mova [r0+r2+8], m3 yading@10: add r0, r4 yading@10: sub r3d, 4 yading@10: jne .loop yading@10: REP_RET yading@10: %endmacro yading@10: yading@10: INIT_MMX mmxext yading@10: PUT_PIXELS_16 yading@10: INIT_MMX 3dnow yading@10: PUT_PIXELS_16 yading@10: yading@10: yading@10: ; put_no_rnd_pixels8_x2(uint8_t *block, const uint8_t *pixels, ptrdiff_t line_size, int h) yading@10: %macro PUT_NO_RND_PIXELS8_X2 0 yading@10: cglobal put_no_rnd_pixels8_x2, 4,5 yading@10: mova m6, [pb_1] yading@10: lea r4, [r2*2] yading@10: .loop: yading@10: mova m0, [r1] yading@10: mova m2, [r1+r2] yading@10: mova m1, [r1+1] yading@10: mova m3, [r1+r2+1] yading@10: add r1, r4 yading@10: psubusb m0, m6 yading@10: psubusb m2, m6 yading@10: PAVGB m0, m1 yading@10: PAVGB m2, m3 yading@10: mova [r0], m0 yading@10: mova [r0+r2], m2 yading@10: mova m0, [r1] yading@10: mova m1, [r1+1] yading@10: mova m2, [r1+r2] yading@10: mova m3, [r1+r2+1] yading@10: add r0, r4 yading@10: add r1, r4 yading@10: psubusb m0, m6 yading@10: psubusb m2, m6 yading@10: PAVGB m0, m1 yading@10: PAVGB m2, m3 yading@10: mova [r0], m0 yading@10: mova [r0+r2], m2 yading@10: add r0, r4 yading@10: sub r3d, 4 yading@10: jne .loop yading@10: REP_RET yading@10: %endmacro yading@10: yading@10: INIT_MMX mmxext yading@10: PUT_NO_RND_PIXELS8_X2 yading@10: INIT_MMX 3dnow yading@10: PUT_NO_RND_PIXELS8_X2 yading@10: yading@10: yading@10: ; put_no_rnd_pixels8_x2_exact(uint8_t *block, const uint8_t *pixels, ptrdiff_t line_size, int h) yading@10: %macro PUT_NO_RND_PIXELS8_X2_EXACT 0 yading@10: cglobal put_no_rnd_pixels8_x2_exact, 4,5 yading@10: lea r4, [r2*3] yading@10: pcmpeqb m6, m6 yading@10: .loop: yading@10: mova m0, [r1] yading@10: mova m2, [r1+r2] yading@10: mova m1, [r1+1] yading@10: mova m3, [r1+r2+1] yading@10: pxor m0, m6 yading@10: pxor m2, m6 yading@10: pxor m1, m6 yading@10: pxor m3, m6 yading@10: PAVGB m0, m1 yading@10: PAVGB m2, m3 yading@10: pxor m0, m6 yading@10: pxor m2, m6 yading@10: mova [r0], m0 yading@10: mova [r0+r2], m2 yading@10: mova m0, [r1+r2*2] yading@10: mova m1, [r1+r2*2+1] yading@10: mova m2, [r1+r4] yading@10: mova m3, [r1+r4+1] yading@10: pxor m0, m6 yading@10: pxor m1, m6 yading@10: pxor m2, m6 yading@10: pxor m3, m6 yading@10: PAVGB m0, m1 yading@10: PAVGB m2, m3 yading@10: pxor m0, m6 yading@10: pxor m2, m6 yading@10: mova [r0+r2*2], m0 yading@10: mova [r0+r4], m2 yading@10: lea r1, [r1+r2*4] yading@10: lea r0, [r0+r2*4] yading@10: sub r3d, 4 yading@10: jg .loop yading@10: REP_RET yading@10: %endmacro yading@10: yading@10: INIT_MMX mmxext yading@10: PUT_NO_RND_PIXELS8_X2_EXACT yading@10: INIT_MMX 3dnow yading@10: PUT_NO_RND_PIXELS8_X2_EXACT yading@10: yading@10: yading@10: ; put_pixels8_y2(uint8_t *block, const uint8_t *pixels, ptrdiff_t line_size, int h) yading@10: %macro PUT_PIXELS8_Y2 0 yading@10: cglobal put_pixels8_y2, 4,5 yading@10: lea r4, [r2*2] yading@10: mova m0, [r1] yading@10: sub r0, r2 yading@10: .loop: yading@10: mova m1, [r1+r2] yading@10: mova m2, [r1+r4] yading@10: add r1, r4 yading@10: PAVGB m0, m1 yading@10: PAVGB m1, m2 yading@10: mova [r0+r2], m0 yading@10: mova [r0+r4], m1 yading@10: mova m1, [r1+r2] yading@10: mova m0, [r1+r4] yading@10: add r0, r4 yading@10: add r1, r4 yading@10: PAVGB m2, m1 yading@10: PAVGB m1, m0 yading@10: mova [r0+r2], m2 yading@10: mova [r0+r4], m1 yading@10: add r0, r4 yading@10: sub r3d, 4 yading@10: jne .loop yading@10: REP_RET yading@10: %endmacro yading@10: yading@10: INIT_MMX mmxext yading@10: PUT_PIXELS8_Y2 yading@10: INIT_MMX 3dnow yading@10: PUT_PIXELS8_Y2 yading@10: yading@10: yading@10: ; put_no_rnd_pixels8_y2(uint8_t *block, const uint8_t *pixels, ptrdiff_t line_size, int h) yading@10: %macro PUT_NO_RND_PIXELS8_Y2 0 yading@10: cglobal put_no_rnd_pixels8_y2, 4,5 yading@10: mova m6, [pb_1] yading@10: lea r4, [r2+r2] yading@10: mova m0, [r1] yading@10: sub r0, r2 yading@10: .loop: yading@10: mova m1, [r1+r2] yading@10: mova m2, [r1+r4] yading@10: add r1, r4 yading@10: psubusb m1, m6 yading@10: PAVGB m0, m1 yading@10: PAVGB m1, m2 yading@10: mova [r0+r2], m0 yading@10: mova [r0+r4], m1 yading@10: mova m1, [r1+r2] yading@10: mova m0, [r1+r4] yading@10: add r0, r4 yading@10: add r1, r4 yading@10: psubusb m1, m6 yading@10: PAVGB m2, m1 yading@10: PAVGB m1, m0 yading@10: mova [r0+r2], m2 yading@10: mova [r0+r4], m1 yading@10: add r0, r4 yading@10: sub r3d, 4 yading@10: jne .loop yading@10: REP_RET yading@10: %endmacro yading@10: yading@10: INIT_MMX mmxext yading@10: PUT_NO_RND_PIXELS8_Y2 yading@10: INIT_MMX 3dnow yading@10: PUT_NO_RND_PIXELS8_Y2 yading@10: yading@10: yading@10: ; put_no_rnd_pixels8_y2_exact(uint8_t *block, const uint8_t *pixels, ptrdiff_t line_size, int h) yading@10: %macro PUT_NO_RND_PIXELS8_Y2_EXACT 0 yading@10: cglobal put_no_rnd_pixels8_y2_exact, 4,5 yading@10: lea r4, [r2*3] yading@10: mova m0, [r1] yading@10: pcmpeqb m6, m6 yading@10: add r1, r2 yading@10: pxor m0, m6 yading@10: .loop: yading@10: mova m1, [r1] yading@10: mova m2, [r1+r2] yading@10: pxor m1, m6 yading@10: pxor m2, m6 yading@10: PAVGB m0, m1 yading@10: PAVGB m1, m2 yading@10: pxor m0, m6 yading@10: pxor m1, m6 yading@10: mova [r0], m0 yading@10: mova [r0+r2], m1 yading@10: mova m1, [r1+r2*2] yading@10: mova m0, [r1+r4] yading@10: pxor m1, m6 yading@10: pxor m0, m6 yading@10: PAVGB m2, m1 yading@10: PAVGB m1, m0 yading@10: pxor m2, m6 yading@10: pxor m1, m6 yading@10: mova [r0+r2*2], m2 yading@10: mova [r0+r4], m1 yading@10: lea r1, [r1+r2*4] yading@10: lea r0, [r0+r2*4] yading@10: sub r3d, 4 yading@10: jg .loop yading@10: REP_RET yading@10: %endmacro yading@10: yading@10: INIT_MMX mmxext yading@10: PUT_NO_RND_PIXELS8_Y2_EXACT yading@10: INIT_MMX 3dnow yading@10: PUT_NO_RND_PIXELS8_Y2_EXACT yading@10: yading@10: yading@10: ; avg_pixels8(uint8_t *block, const uint8_t *pixels, ptrdiff_t line_size, int h) yading@10: %macro AVG_PIXELS8 0 yading@10: cglobal avg_pixels8, 4,5 yading@10: lea r4, [r2*2] yading@10: .loop: yading@10: mova m0, [r0] yading@10: mova m1, [r0+r2] yading@10: PAVGB m0, [r1] yading@10: PAVGB m1, [r1+r2] yading@10: mova [r0], m0 yading@10: mova [r0+r2], m1 yading@10: add r1, r4 yading@10: add r0, r4 yading@10: mova m0, [r0] yading@10: mova m1, [r0+r2] yading@10: PAVGB m0, [r1] yading@10: PAVGB m1, [r1+r2] yading@10: add r1, r4 yading@10: mova [r0], m0 yading@10: mova [r0+r2], m1 yading@10: add r0, r4 yading@10: sub r3d, 4 yading@10: jne .loop yading@10: REP_RET yading@10: %endmacro yading@10: yading@10: INIT_MMX 3dnow yading@10: AVG_PIXELS8 yading@10: yading@10: yading@10: ; avg_pixels8_x2(uint8_t *block, const uint8_t *pixels, ptrdiff_t line_size, int h) yading@10: %macro AVG_PIXELS8_X2 0 yading@10: cglobal avg_pixels8_x2, 4,5 yading@10: lea r4, [r2*2] yading@10: .loop: yading@10: mova m0, [r1] yading@10: mova m2, [r1+r2] yading@10: PAVGB m0, [r1+1] yading@10: PAVGB m2, [r1+r2+1] yading@10: PAVGB m0, [r0] yading@10: PAVGB m2, [r0+r2] yading@10: add r1, r4 yading@10: mova [r0], m0 yading@10: mova [r0+r2], m2 yading@10: mova m0, [r1] yading@10: mova m2, [r1+r2] yading@10: PAVGB m0, [r1+1] yading@10: PAVGB m2, [r1+r2+1] yading@10: add r0, r4 yading@10: add r1, r4 yading@10: PAVGB m0, [r0] yading@10: PAVGB m2, [r0+r2] yading@10: mova [r0], m0 yading@10: mova [r0+r2], m2 yading@10: add r0, r4 yading@10: sub r3d, 4 yading@10: jne .loop yading@10: REP_RET yading@10: %endmacro yading@10: yading@10: INIT_MMX mmxext yading@10: AVG_PIXELS8_X2 yading@10: INIT_MMX 3dnow yading@10: AVG_PIXELS8_X2 yading@10: yading@10: yading@10: ; avg_pixels8_y2(uint8_t *block, const uint8_t *pixels, ptrdiff_t line_size, int h) yading@10: %macro AVG_PIXELS8_Y2 0 yading@10: cglobal avg_pixels8_y2, 4,5 yading@10: lea r4, [r2*2] yading@10: mova m0, [r1] yading@10: sub r0, r2 yading@10: .loop: yading@10: mova m1, [r1+r2] yading@10: mova m2, [r1+r4] yading@10: add r1, r4 yading@10: PAVGB m0, m1 yading@10: PAVGB m1, m2 yading@10: mova m3, [r0+r2] yading@10: mova m4, [r0+r4] yading@10: PAVGB m0, m3 yading@10: PAVGB m1, m4 yading@10: mova [r0+r2], m0 yading@10: mova [r0+r4], m1 yading@10: mova m1, [r1+r2] yading@10: mova m0, [r1+r4] yading@10: PAVGB m2, m1 yading@10: PAVGB m1, m0 yading@10: add r0, r4 yading@10: add r1, r4 yading@10: mova m3, [r0+r2] yading@10: mova m4, [r0+r4] yading@10: PAVGB m2, m3 yading@10: PAVGB m1, m4 yading@10: mova [r0+r2], m2 yading@10: mova [r0+r4], m1 yading@10: add r0, r4 yading@10: sub r3d, 4 yading@10: jne .loop yading@10: REP_RET yading@10: %endmacro yading@10: yading@10: INIT_MMX mmxext yading@10: AVG_PIXELS8_Y2 yading@10: INIT_MMX 3dnow yading@10: AVG_PIXELS8_Y2 yading@10: yading@10: yading@10: ; avg_pixels8_xy2(uint8_t *block, const uint8_t *pixels, ptrdiff_t line_size, int h) yading@10: %macro AVG_PIXELS8_XY2 0 yading@10: cglobal avg_pixels8_xy2, 4,5 yading@10: mova m6, [pb_1] yading@10: lea r4, [r2*2] yading@10: mova m0, [r1] yading@10: pavgb m0, [r1+1] yading@10: .loop: yading@10: mova m2, [r1+r4] yading@10: mova m1, [r1+r2] yading@10: psubusb m2, m6 yading@10: pavgb m1, [r1+r2+1] yading@10: pavgb m2, [r1+r4+1] yading@10: add r1, r4 yading@10: pavgb m0, m1 yading@10: pavgb m1, m2 yading@10: pavgb m0, [r0] yading@10: pavgb m1, [r0+r2] yading@10: mova [r0], m0 yading@10: mova [r0+r2], m1 yading@10: mova m1, [r1+r2] yading@10: mova m0, [r1+r4] yading@10: pavgb m1, [r1+r2+1] yading@10: pavgb m0, [r1+r4+1] yading@10: add r0, r4 yading@10: add r1, r4 yading@10: pavgb m2, m1 yading@10: pavgb m1, m0 yading@10: pavgb m2, [r0] yading@10: pavgb m1, [r0+r2] yading@10: mova [r0], m2 yading@10: mova [r0+r2], m1 yading@10: add r0, r4 yading@10: sub r3d, 4 yading@10: jne .loop yading@10: REP_RET yading@10: %endmacro yading@10: yading@10: INIT_MMX mmxext yading@10: AVG_PIXELS8_XY2 yading@10: INIT_MMX 3dnow yading@10: AVG_PIXELS8_XY2