diff --git a/libavcodec/x86/h264_intrapred.asm b/libavcodec/x86/h264_intrapred.asm index d01b134635..1a87cfdfbf 100644 --- a/libavcodec/x86/h264_intrapred.asm +++ b/libavcodec/x86/h264_intrapred.asm @@ -1724,6 +1724,119 @@ INIT_MMX PRED8x8L_DOWN_RIGHT ssse3 %endif +;----------------------------------------------------------------------------- +; void pred8x8l_vertical_right(uint8_t *src, int has_topleft, int has_topright, int stride) +;----------------------------------------------------------------------------- +%ifdef CONFIG_GPL +INIT_MMX +%define PALIGNR PALIGNR_MMX +cglobal pred8x8l_vertical_right_mmxext, 4,5 + sub r0, r3 + lea r4, [r0+r3*2] + movq mm0, [r0+r3*1-8] + punpckhbw mm0, [r0+r3*0-8] + movq mm1, [r4+r3*1-8] + punpckhbw mm1, [r0+r3*2-8] + mov r4, r0 + punpckhwd mm1, mm0 + lea r0, [r0+r3*4] + movq mm2, [r0+r3*1-8] + punpckhbw mm2, [r0+r3*0-8] + lea r0, [r0+r3*2] + movq mm3, [r0+r3*1-8] + punpckhbw mm3, [r0+r3*0-8] + punpckhwd mm3, mm2 + punpckhdq mm3, mm1 + lea r0, [r0+r3*2] + movq mm0, [r0+r3*0-8] + movq mm1, [r4] + mov r0, r4 + movq mm4, mm3 + movq mm2, mm3 + PALIGNR mm4, mm0, 7, mm0 + PALIGNR mm1, mm2, 1, mm2 + test r1, r1 + jz .fix_lt_1 + jmp .do_left +.fix_lt_1: + movq mm5, mm3 + pxor mm5, mm4 + psrlq mm5, 56 + psllq mm5, 48 + pxor mm1, mm5 + jmp .do_left +.fix_lt_2: + movq mm5, mm3 + pxor mm5, mm2 + psllq mm5, 56 + psrlq mm5, 56 + pxor mm2, mm5 + test r2, r2 + jnz .do_top +.fix_tr_1: + movq mm5, mm3 + pxor mm5, mm1 + psrlq mm5, 56 + psllq mm5, 56 + pxor mm1, mm5 + jmp .do_top +.do_left: + movq mm0, mm4 + PRED4x4_LOWPASS mm2, mm1, mm4, mm3, mm5 + movq mm7, mm2 + movq mm0, [r0-8] + movq mm3, [r0] + movq mm1, [r0+8] + movq mm2, mm3 + movq mm4, mm3 + PALIGNR mm2, mm0, 7, mm0 + PALIGNR mm1, mm4, 1, mm4 + test r1, r1 + jz .fix_lt_2 + test r2, r2 + jz .fix_tr_1 +.do_top + PRED4x4_LOWPASS mm6, mm2, mm1, mm3, mm5 + lea r1, [r0+r3*2] + movq mm2, mm6 + movq mm3, mm6 + PALIGNR mm3, mm7, 7, mm0 + PALIGNR mm6, mm7, 6, mm1 + movq mm4, mm3 + pavgb mm3, mm2 + lea r2, [r1+r3*2] + PRED4x4_LOWPASS mm0, mm6, mm2, mm4, mm5 + movq [r0+r3*1], mm3 + movq [r0+r3*2], mm0 + movq mm5, mm0 + movq mm6, mm3 + movq mm1, mm7 + movq mm2, mm1 + psllq mm2, 8 + movq mm3, mm1 + psllq mm3, 16 + lea r4, [r2+r3*2] + PRED4x4_LOWPASS mm0, mm1, mm3, mm2, mm4 + PALIGNR mm6, mm0, 7, mm2 + movq [r1+r3*1], mm6 + psllq mm0, 8 + PALIGNR mm5, mm0, 7, mm1 + movq [r1+r3*2], mm5 + psllq mm0, 8 + PALIGNR mm6, mm0, 7, mm2 + movq [r2+r3*1], mm6 + psllq mm0, 8 + PALIGNR mm5, mm0, 7, mm1 + movq [r2+r3*2], mm5 + psllq mm0, 8 + PALIGNR mm6, mm0, 7, mm2 + movq [r4+r3*1], mm6 + psllq mm0, 8 + PALIGNR mm5, mm0, 7, mm1 + movq [r4+r3*2], mm5 + RET +%endif + ;----------------------------------------------------------------------------- ; void pred4x4_dc_mmxext(uint8_t *src, const uint8_t *topright, int stride) ;----------------------------------------------------------------------------- diff --git a/libavcodec/x86/h264_intrapred_init.c b/libavcodec/x86/h264_intrapred_init.c index 5068a81215..360837173b 100644 --- a/libavcodec/x86/h264_intrapred_init.c +++ b/libavcodec/x86/h264_intrapred_init.c @@ -72,6 +72,7 @@ void ff_pred8x8l_down_left_ssse3 (uint8_t *src, int has_topleft, int has_topri void ff_pred8x8l_down_right_mmxext (uint8_t *src, int has_topleft, int has_topright, int stride); void ff_pred8x8l_down_right_sse2 (uint8_t *src, int has_topleft, int has_topright, int stride); void ff_pred8x8l_down_right_ssse3 (uint8_t *src, int has_topleft, int has_topright, int stride); +void ff_pred8x8l_vertical_right_mmxext(uint8_t *src, int has_topleft, int has_topright, int stride); void ff_pred4x4_dc_mmxext (uint8_t *src, const uint8_t *topright, int stride); void ff_pred4x4_down_left_mmxext (uint8_t *src, const uint8_t *topright, int stride); void ff_pred4x4_tm_vp8_mmx (uint8_t *src, const uint8_t *topright, int stride); @@ -115,6 +116,7 @@ void ff_h264_pred_init_x86(H264PredContext *h, int codec_id) h->pred8x8l [HOR_PRED ] = ff_pred8x8l_horizontal_mmxext; h->pred8x8l [VERT_PRED ] = ff_pred8x8l_vertical_mmxext; h->pred8x8l [DIAG_DOWN_RIGHT_PRED] = ff_pred8x8l_down_right_mmxext; + h->pred8x8l [VERT_RIGHT_PRED ] = ff_pred8x8l_vertical_right_mmxext; #endif h->pred4x4 [DC_PRED ] = ff_pred4x4_dc_mmxext; #if CONFIG_GPL