1 ;*****************************************************************************
2 ;* mc-a.asm: h264 encoder library
3 ;*****************************************************************************
4 ;* Copyright (C) 2003-2008 x264 project
6 ;* Authors: Loren Merritt <lorenm@u.washington.edu>
7 ;* Laurent Aimar <fenrir@via.ecp.fr>
8 ;* Fiona Glaser <fiona@x264.com>
9 ;* Min Chen <chenm001.163.com>
11 ;* This program is free software; you can redistribute it and/or modify
12 ;* it under the terms of the GNU General Public License as published by
13 ;* the Free Software Foundation; either version 2 of the License, or
14 ;* (at your option) any later version.
16 ;* This program is distributed in the hope that it will be useful,
17 ;* but WITHOUT ANY WARRANTY; without even the implied warranty of
18 ;* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
19 ;* GNU General Public License for more details.
21 ;* You should have received a copy of the GNU General Public License
22 ;* along with this program; if not, write to the Free Software
23 ;* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02111, USA.
24 ;*****************************************************************************
38 ;=============================================================================
40 ;=============================================================================
41 ; implicit bipred only:
42 ; assumes log2_denom = 5, offset = 0, weight1 + weight2 = 64
98 %macro BIWEIGHT_START_MMX 0
100 SPLATW m4, m4 ; weight_dst
101 mova m5, [pw_64 GLOBAL]
102 psubw m5, m4 ; weight_src
103 mova m6, [pw_32 GLOBAL] ; rounding
107 %macro BIWEIGHT_SSSE3 2
116 %macro BIWEIGHT_START_SSSE3 0
117 movzx t6d, byte r6m ; FIXME x86_64
123 mova m6, [pw_32 GLOBAL]
124 SPLATW m5, m5 ; weight_dst,src
127 %macro BIWEIGHT_ROW 4
134 BIWEIGHT [%2+mmsize/2], [%3+mmsize/2]
140 ;-----------------------------------------------------------------------------
141 ; int x264_pixel_avg_weight_w16_mmxext( uint8_t *dst, int, uint8_t *src1, int, uint8_t *src2, int, int i_weight )
142 ;-----------------------------------------------------------------------------
144 cglobal x264_pixel_avg_weight_w%2_%1, 0,0
147 %if %2==8 && mmsize==16
150 BIWEIGHT [t2+t3], [t4+t5]
157 BIWEIGHT_ROW t0+x, t2+x, t4+x, %2
158 BIWEIGHT_ROW t0+x+t1, t2+x+t3, t4+x+t5, %2
170 %define BIWEIGHT BIWEIGHT_MMX
171 %define BIWEIGHT_START BIWEIGHT_START_MMX
175 AVG_WEIGHT mmxext, 16
177 %define x264_pixel_avg_weight_w4_sse2 x264_pixel_avg_weight_w4_mmxext
180 %define BIWEIGHT BIWEIGHT_SSSE3
181 %define BIWEIGHT_START BIWEIGHT_START_SSSE3
190 ;=============================================================================
192 ;=============================================================================
194 ;-----------------------------------------------------------------------------
195 ; void x264_pixel_avg_4x4_mmxext( uint8_t *dst, int dst_stride,
196 ; uint8_t *src1, int src1_stride, uint8_t *src2, int src2_stride, int weight );
197 ;-----------------------------------------------------------------------------
199 cglobal x264_pixel_avg_%1x%2_%3,0,0
202 jne x264_pixel_avg_weight_w%1_%3
203 %if mmsize == 16 && %1 == 16
205 jz x264_pixel_avg_w%1_sse2
207 jmp x264_pixel_avg_w%1_mmxext
210 ;-----------------------------------------------------------------------------
211 ; void x264_pixel_avg_w4_mmxext( uint8_t *dst, int dst_stride,
212 ; uint8_t *src1, int src1_stride, uint8_t *src2, int src2_stride,
213 ; int height, int weight );
214 ;-----------------------------------------------------------------------------
238 AVG_FUNC x264_pixel_avg_w4_mmxext, movd, movd
243 AVG_FUNC x264_pixel_avg_w8_mmxext, movq, movq
248 cglobal x264_pixel_avg_w16_mmxext
268 AVG_FUNC x264_pixel_avg_w16_sse2, movdqu, movdqa
286 ;=============================================================================
288 ;=============================================================================
290 ;-----------------------------------------------------------------------------
291 ; void x264_pixel_avg2_w4_mmxext( uint8_t *dst, int dst_stride,
292 ; uint8_t *src1, int src_stride,
293 ; uint8_t *src2, int height );
294 ;-----------------------------------------------------------------------------
296 cglobal x264_pixel_avg2_w%1_mmxext, 6,7
317 cglobal x264_pixel_avg2_w%1_mmxext, 6,7
343 cglobal x264_pixel_avg2_w20_mmxext, 6,7
355 pavgb mm2, [r2+r4+16]
358 pavgb mm5, [r2+r6+16]
371 cglobal x264_pixel_avg2_w16_sse2, 6,7
390 cglobal x264_pixel_avg2_w20_%1, 6,7
398 %ifidn %1, sse2_misalign
407 pavgb mm4, [r2+r4+16]
408 pavgb mm5, [r2+r6+16]
421 AVG2_W20 sse2_misalign
423 ; Cacheline split code for processors with high latencies for loads
424 ; split over cache lines. See sad-a.asm for a more detailed explanation.
425 ; This particular instance is complicated by the fact that src1 and src2
426 ; can have different alignments. For simplicity and code size, only the
427 ; MMX cacheline workaround is used. As a result, in the case of SSE2
428 ; pixel_avg, the cacheline check functions calls the SSE2 version if there
429 ; is no cacheline split, and the MMX workaround if there is.
434 movd %1, [sw_64 GLOBAL]
439 %macro AVG_CACHELINE_CHECK 3 ; width, cacheline, instruction set
440 cglobal x264_pixel_avg2_w%1_cache%2_%3, 0,0
442 and eax, 0x1f|(%2>>1)
443 cmp eax, (32-%1)|(%2>>1)
444 jle x264_pixel_avg2_w%1_%3
445 ;w12 isn't needed because w16 is just as fast if there's no cacheline split
447 jmp x264_pixel_avg2_w16_cache_mmxext
449 jmp x264_pixel_avg2_w%1_cache_mmxext
453 %macro AVG_CACHELINE_START 0
454 %assign stack_offset 0
465 %macro AVG_CACHELINE_LOOP 2
468 movq mm2, [r2+r4+8+%1]
480 x264_pixel_avg2_w8_cache_mmxext:
482 AVG_CACHELINE_LOOP 0, movq
489 x264_pixel_avg2_w16_cache_mmxext:
491 AVG_CACHELINE_LOOP 0, movq
492 AVG_CACHELINE_LOOP 8, movq
499 x264_pixel_avg2_w20_cache_mmxext:
501 AVG_CACHELINE_LOOP 0, movq
502 AVG_CACHELINE_LOOP 8, movq
503 AVG_CACHELINE_LOOP 16, movd
511 AVG_CACHELINE_CHECK 8, 32, mmxext
512 AVG_CACHELINE_CHECK 12, 32, mmxext
513 AVG_CACHELINE_CHECK 16, 32, mmxext
514 AVG_CACHELINE_CHECK 20, 32, mmxext
515 AVG_CACHELINE_CHECK 16, 64, mmxext
516 AVG_CACHELINE_CHECK 20, 64, mmxext
519 AVG_CACHELINE_CHECK 8, 64, mmxext
520 AVG_CACHELINE_CHECK 12, 64, mmxext
521 AVG_CACHELINE_CHECK 16, 64, sse2
522 AVG_CACHELINE_CHECK 20, 64, sse2
524 ;=============================================================================
526 ;=============================================================================
540 ;-----------------------------------------------------------------------------
541 ; void x264_mc_copy_w4_mmx( uint8_t *dst, int i_dst_stride,
542 ; uint8_t *src, int i_src_stride, int i_height )
543 ;-----------------------------------------------------------------------------
544 cglobal x264_mc_copy_w4_mmx, 4,6
549 COPY4 movd, movd, r4, r5
553 COPY4 movd, movd, r4, r5
556 cglobal x264_mc_copy_w8_mmx, 5,7
560 COPY4 movq, movq, r5, r6
567 cglobal x264_mc_copy_w16_mmx, 5,7
576 movq mm5, [r2+r3*2+8]
584 movq [r0+r1*2+8], mm5
594 %macro COPY_W16_SSE2 2
599 COPY4 movdqa, %2, r5, r6
607 COPY_W16_SSE2 x264_mc_copy_w16_sse2, movdqu
608 ; cacheline split with mmx has too much overhead; the speed benefit is near-zero.
609 ; but with SSE3 the overhead is zero, so there's no reason not to include it.
610 COPY_W16_SSE2 x264_mc_copy_w16_sse3, lddqu
611 COPY_W16_SSE2 x264_mc_copy_w16_aligned_sse2, movdqa
615 ;=============================================================================
617 ;=============================================================================
618 ; FIXME assumes 64 byte cachelines
620 ;-----------------------------------------------------------------------------
621 ; void x264_prefetch_fenc_mmxext( uint8_t *pix_y, int stride_y,
622 ; uint8_t *pix_uv, int stride_uv, int mb_x )
623 ;-----------------------------------------------------------------------------
625 cglobal x264_prefetch_fenc_mmxext, 5,5
629 lea r0, [r0+rax*4+64]
644 cglobal x264_prefetch_fenc_mmxext
668 ;-----------------------------------------------------------------------------
669 ; void x264_prefetch_ref_mmxext( uint8_t *pix, int stride, int parity )
670 ;-----------------------------------------------------------------------------
671 cglobal x264_prefetch_ref_mmxext, 3,3
689 ;=============================================================================
691 ;=============================================================================
701 %macro MC_CHROMA_START 0
713 add r2, t0 ; src += (dx>>3) + (dy>>3) * src_stride
716 ;-----------------------------------------------------------------------------
717 ; void x264_mc_chroma_mmxext( uint8_t *dst, int dst_stride,
718 ; uint8_t *src, int src_stride,
720 ; int width, int height )
721 ;-----------------------------------------------------------------------------
723 cglobal x264_mc_chroma_%1, 0,6
726 jle x264_mc_chroma_mmxext %+ .skip_prologue
738 SPLATW m5, m5 ; m5 = dx
739 SPLATW m6, m6 ; m6 = dy
741 mova m4, [pw_8 GLOBAL]
743 psubw m4, m5 ; m4 = 8-dx
744 psubw m0, m6 ; m0 = 8-dy
747 pmullw m5, m0 ; m5 = dx*(8-dy) = cB
748 pmullw m7, m6 ; m7 = dx*dy = cD
749 pmullw m6, m4 ; m6 = (8-dx)*dy = cC
750 pmullw m4, m0 ; m4 = (8-dx)*(8-dy) = cA
765 punpcklbw m1, m3 ; 00 px1 | 00 px2 | 00 px3 | 00 px4
767 pmullw m1, m6 ; 2nd line * cC
768 pmullw m0, m4 ; 1st line * cA
769 paddw m0, m1 ; m0 <- result
776 paddw m0, [pw_32 GLOBAL]
778 pmullw m2, m5 ; line * cB
779 pmullw m1, m7 ; line * cD
784 packuswb m0, m3 ; 00 00 00 00 px1 px2 px3 px4
788 add r0, r1 ; dst_stride
794 jnz .finish ; width != 8 so assume 4
796 lea r0, [r10+4] ; dst
797 lea r2, [r11+4] ; src
803 mov r4d, r7m ; height
812 mov r5, r3 ; pel_offset = dx ? 1 : src_stride
818 mova m5, [pw_8 GLOBAL]
820 mova m7, [pw_4 GLOBAL]
877 %endmacro ; MC_CHROMA
885 cglobal x264_mc_chroma_ssse3, 0,6
895 imul r5d, t0d ; (x*255+8)*y
896 imul r4d, t0d ; (x*255+8)*(8-y)
899 mova m5, [pw_32 GLOBAL]
914 punpcklbw m3, [r2+r3+1]
940 mova m5, [pw_32 GLOBAL]
981 ; mc_chroma 1d ssse3 is negligibly faster, and definitely not worth the extra code size