Mercurial > libavcodec.hg
comparison i386/dsputil_mmx.c @ 3568:945caa35ee9a libavcodec
sse and 3dnow implementations of float->int conversion and mdct windowing.
15% faster vorbis.
| author | lorenm |
|---|---|
| date | Thu, 10 Aug 2006 19:06:25 +0000 |
| parents | 97325fecd35a |
| children | c42c03f3b402 |
comparison
equal
deleted
inserted
replaced
| 3567:1f8730f62765 | 3568:945caa35ee9a |
|---|---|
| 2770 ::"memory" | 2770 ::"memory" |
| 2771 ); | 2771 ); |
| 2772 } | 2772 } |
| 2773 } | 2773 } |
| 2774 | 2774 |
| 2775 static void vector_fmul_3dnow(float *dst, const float *src, int len){ | |
| 2776 long i; | |
| 2777 len >>= 1; | |
| 2778 for(i=0; i<len; i++) { | |
| 2779 asm volatile( | |
| 2780 "movq %0, %%mm0 \n\t" | |
| 2781 "pfmul %1, %%mm0 \n\t" | |
| 2782 "movq %%mm0, %0 \n\t" | |
| 2783 :"+m"(dst[i*2]) | |
| 2784 :"m"(src[i*2]) | |
| 2785 :"memory" | |
| 2786 ); | |
| 2787 } | |
| 2788 asm volatile("femms"); | |
| 2789 } | |
| 2790 static void vector_fmul_sse(float *dst, const float *src, int len){ | |
| 2791 long i; | |
| 2792 len >>= 2; | |
| 2793 for(i=0; i<len; i++) { | |
| 2794 asm volatile( | |
| 2795 "movaps %0, %%xmm0 \n\t" | |
| 2796 "mulps %1, %%xmm0 \n\t" | |
| 2797 "movaps %%xmm0, %0 \n\t" | |
| 2798 :"+m"(dst[i*4]) | |
| 2799 :"m"(src[i*4]) | |
| 2800 :"memory" | |
| 2801 ); | |
| 2802 } | |
| 2803 } | |
| 2804 | |
| 2805 static void vector_fmul_reverse_3dnow2(float *dst, const float *src0, const float *src1, int len){ | |
| 2806 long i = len*4-16; | |
| 2807 asm volatile( | |
| 2808 "1: \n\t" | |
| 2809 "pswapd 8(%1), %%mm0 \n\t" | |
| 2810 "pswapd (%1), %%mm1 \n\t" | |
| 2811 "pfmul (%3,%0), %%mm0 \n\t" | |
| 2812 "pfmul 8(%3,%0), %%mm1 \n\t" | |
| 2813 "movq %%mm0, (%2,%0) \n\t" | |
| 2814 "movq %%mm1, 8(%2,%0) \n\t" | |
| 2815 "add $16, %1 \n\t" | |
| 2816 "sub $16, %0 \n\t" | |
| 2817 "jge 1b \n\t" | |
| 2818 :"+r"(i), "+r"(src1) | |
| 2819 :"r"(dst), "r"(src0) | |
| 2820 ); | |
| 2821 asm volatile("femms"); | |
| 2822 } | |
| 2823 static void vector_fmul_reverse_sse2(float *dst, const float *src0, const float *src1, int len){ | |
| 2824 long i = len*4-32; | |
| 2825 asm volatile( | |
| 2826 "1: \n\t" | |
| 2827 "pshufd $0x1b, 16(%1), %%xmm0 \n\t" | |
| 2828 "pshufd $0x1b, (%1), %%xmm1 \n\t" | |
| 2829 "mulps (%3,%0), %%xmm0 \n\t" | |
| 2830 "mulps 16(%3,%0), %%xmm1 \n\t" | |
| 2831 "movaps %%xmm0, (%2,%0) \n\t" | |
| 2832 "movaps %%xmm1, 16(%2,%0) \n\t" | |
| 2833 "add $32, %1 \n\t" | |
| 2834 "sub $32, %0 \n\t" | |
| 2835 "jge 1b \n\t" | |
| 2836 :"+r"(i), "+r"(src1) | |
| 2837 :"r"(dst), "r"(src0) | |
| 2838 ); | |
| 2839 } | |
| 2840 | |
| 2841 static void vector_fmul_add_add_3dnow(float *dst, const float *src0, const float *src1, | |
| 2842 const float *src2, int src3, int len, int step){ | |
| 2843 long i; | |
| 2844 if(step == 2 && src3 == 0){ | |
| 2845 i = (len-4)*4; | |
| 2846 dst += (len-4)*2; | |
| 2847 asm volatile( | |
| 2848 "1: \n\t" | |
| 2849 "movq (%2,%0), %%mm0 \n\t" | |
| 2850 "movq 8(%2,%0), %%mm1 \n\t" | |
| 2851 "pfmul (%3,%0), %%mm0 \n\t" | |
| 2852 "pfmul 8(%3,%0), %%mm1 \n\t" | |
| 2853 "pfadd (%4,%0), %%mm0 \n\t" | |
| 2854 "pfadd 8(%4,%0), %%mm1 \n\t" | |
| 2855 "movd %%mm0, (%1) \n\t" | |
| 2856 "movd %%mm1, 16(%1) \n\t" | |
| 2857 "psrlq $32, %%mm0 \n\t" | |
| 2858 "psrlq $32, %%mm1 \n\t" | |
| 2859 "movd %%mm0, 8(%1) \n\t" | |
| 2860 "movd %%mm1, 24(%1) \n\t" | |
| 2861 "sub $32, %1 \n\t" | |
| 2862 "sub $16, %0 \n\t" | |
| 2863 "jge 1b \n\t" | |
| 2864 :"+r"(i), "+r"(dst) | |
| 2865 :"r"(src0), "r"(src1), "r"(src2) | |
| 2866 :"memory" | |
| 2867 ); | |
| 2868 } | |
| 2869 else if(step == 1 && src3 == 0){ | |
| 2870 for(i=0; i<len; i+=2){ | |
| 2871 asm volatile( | |
| 2872 "movq %1, %%mm0 \n\t" | |
| 2873 "pfmul %2, %%mm0 \n\t" | |
| 2874 "pfadd %3, %%mm0 \n\t" | |
| 2875 "movq %%mm0, %0 \n\t" | |
| 2876 :"=m"(dst[i]) | |
| 2877 :"m"(src0[i]), "m"(src1[i]), "m"(src2[i]) | |
| 2878 ); | |
| 2879 } | |
| 2880 } | |
| 2881 else | |
| 2882 ff_vector_fmul_add_add_c(dst, src0, src1, src2, src3, len, step); | |
| 2883 asm volatile("femms"); | |
| 2884 } | |
| 2885 static void vector_fmul_add_add_sse2(float *dst, const float *src0, const float *src1, | |
| 2886 const float *src2, float src3, int len, int step){ | |
| 2887 long i; | |
| 2888 if(step == 2 && src3 == 0){ | |
| 2889 i = (len-8)*4; | |
| 2890 dst += (len-8)*2; | |
| 2891 asm volatile( | |
| 2892 "1: \n\t" | |
| 2893 "movaps (%2,%0), %%xmm0 \n\t" | |
| 2894 "movaps 16(%2,%0), %%xmm1 \n\t" | |
| 2895 "mulps (%3,%0), %%xmm0 \n\t" | |
| 2896 "mulps 16(%3,%0), %%xmm1 \n\t" | |
| 2897 "addps (%4,%0), %%xmm0 \n\t" | |
| 2898 "addps 16(%4,%0), %%xmm1 \n\t" | |
| 2899 "movd %%xmm0, (%1) \n\t" | |
| 2900 "movd %%xmm1, 32(%1) \n\t" | |
| 2901 "psrldq $4, %%xmm0 \n\t" | |
| 2902 "psrldq $4, %%xmm1 \n\t" | |
| 2903 "movd %%xmm0, 8(%1) \n\t" | |
| 2904 "movd %%xmm1, 40(%1) \n\t" | |
| 2905 "psrldq $4, %%xmm0 \n\t" | |
| 2906 "psrldq $4, %%xmm1 \n\t" | |
| 2907 "movd %%xmm0, 16(%1) \n\t" | |
| 2908 "movd %%xmm1, 48(%1) \n\t" | |
| 2909 "psrldq $4, %%xmm0 \n\t" | |
| 2910 "psrldq $4, %%xmm1 \n\t" | |
| 2911 "movd %%xmm0, 24(%1) \n\t" | |
| 2912 "movd %%xmm1, 56(%1) \n\t" | |
| 2913 "sub $64, %1 \n\t" | |
| 2914 "sub $32, %0 \n\t" | |
| 2915 "jge 1b \n\t" | |
| 2916 :"+r"(i), "+r"(dst) | |
| 2917 :"r"(src0), "r"(src1), "r"(src2) | |
| 2918 :"memory" | |
| 2919 ); | |
| 2920 } | |
| 2921 else if(step == 1 && src3 == 0){ | |
| 2922 for(i=0; i<len; i+=4){ | |
| 2923 asm volatile( | |
| 2924 "movaps %1, %%xmm0 \n\t" | |
| 2925 "mulps %2, %%xmm0 \n\t" | |
| 2926 "addps %3, %%xmm0 \n\t" | |
| 2927 "movaps %%xmm0, %0 \n\t" | |
| 2928 :"=m"(dst[i]) | |
| 2929 :"m"(src0[i]), "m"(src1[i]), "m"(src2[i]) | |
| 2930 ); | |
| 2931 } | |
| 2932 } | |
| 2933 else | |
| 2934 ff_vector_fmul_add_add_c(dst, src0, src1, src2, src3, len, step); | |
| 2935 } | |
| 2936 | |
| 2937 void float_to_int16_3dnow(int16_t *dst, const float *src, int len){ | |
| 2938 // not bit-exact: pf2id uses different rounding than C and SSE | |
| 2939 int i; | |
| 2940 for(i=0; i<len; i+=4) { | |
| 2941 asm volatile( | |
| 2942 "pf2id %1, %%mm0 \n\t" | |
| 2943 "pf2id %2, %%mm1 \n\t" | |
| 2944 "packssdw %%mm1, %%mm0 \n\t" | |
| 2945 "movq %%mm0, %0 \n\t" | |
| 2946 :"=m"(dst[i]) | |
| 2947 :"m"(src[i]), "m"(src[i+2]) | |
| 2948 ); | |
| 2949 } | |
| 2950 asm volatile("femms"); | |
| 2951 } | |
| 2952 void float_to_int16_sse(int16_t *dst, const float *src, int len){ | |
| 2953 int i; | |
| 2954 for(i=0; i<len; i+=4) { | |
| 2955 asm volatile( | |
| 2956 "cvtps2pi %1, %%mm0 \n\t" | |
| 2957 "cvtps2pi %2, %%mm1 \n\t" | |
| 2958 "packssdw %%mm1, %%mm0 \n\t" | |
| 2959 "movq %%mm0, %0 \n\t" | |
| 2960 :"=m"(dst[i]) | |
| 2961 :"m"(src[i]), "m"(src[i+2]) | |
| 2962 ); | |
| 2963 } | |
| 2964 asm volatile("emms"); | |
| 2965 } | |
| 2966 | |
| 2775 #ifdef CONFIG_SNOW_ENCODER | 2967 #ifdef CONFIG_SNOW_ENCODER |
| 2776 extern void ff_snow_horizontal_compose97i_sse2(DWTELEM *b, int width); | 2968 extern void ff_snow_horizontal_compose97i_sse2(DWTELEM *b, int width); |
| 2777 extern void ff_snow_horizontal_compose97i_mmx(DWTELEM *b, int width); | 2969 extern void ff_snow_horizontal_compose97i_mmx(DWTELEM *b, int width); |
| 2778 extern void ff_snow_vertical_compose97i_sse2(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, DWTELEM *b3, DWTELEM *b4, DWTELEM *b5, int width); | 2970 extern void ff_snow_vertical_compose97i_sse2(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, DWTELEM *b3, DWTELEM *b4, DWTELEM *b5, int width); |
| 2779 extern void ff_snow_vertical_compose97i_mmx(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, DWTELEM *b3, DWTELEM *b4, DWTELEM *b5, int width); | 2971 extern void ff_snow_vertical_compose97i_mmx(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, DWTELEM *b3, DWTELEM *b4, DWTELEM *b5, int width); |
| 3197 c->vertical_compose97i = ff_snow_vertical_compose97i_mmx; | 3389 c->vertical_compose97i = ff_snow_vertical_compose97i_mmx; |
| 3198 c->inner_add_yblock = ff_snow_inner_add_yblock_mmx; | 3390 c->inner_add_yblock = ff_snow_inner_add_yblock_mmx; |
| 3199 } | 3391 } |
| 3200 #endif | 3392 #endif |
| 3201 | 3393 |
| 3202 if(mm_flags & MM_SSE) | 3394 if(mm_flags & MM_3DNOW){ |
| 3395 c->vorbis_inverse_coupling = vorbis_inverse_coupling_3dnow; | |
| 3396 c->vector_fmul = vector_fmul_3dnow; | |
| 3397 if(!(avctx->flags & CODEC_FLAG_BITEXACT)) | |
| 3398 c->float_to_int16 = float_to_int16_3dnow; | |
| 3399 } | |
| 3400 if(mm_flags & MM_3DNOWEXT) | |
| 3401 c->vector_fmul_reverse = vector_fmul_reverse_3dnow2; | |
| 3402 if(mm_flags & MM_SSE){ | |
| 3203 c->vorbis_inverse_coupling = vorbis_inverse_coupling_sse; | 3403 c->vorbis_inverse_coupling = vorbis_inverse_coupling_sse; |
| 3204 else if(mm_flags & MM_3DNOW) | 3404 c->vector_fmul = vector_fmul_sse; |
| 3205 c->vorbis_inverse_coupling = vorbis_inverse_coupling_3dnow; | 3405 c->float_to_int16 = float_to_int16_sse; |
| 3406 } | |
| 3407 if(mm_flags & MM_SSE2){ | |
| 3408 c->vector_fmul_reverse = vector_fmul_reverse_sse2; | |
| 3409 c->vector_fmul_add_add = vector_fmul_add_add_sse2; | |
| 3410 } | |
| 3411 if(mm_flags & MM_3DNOW) | |
| 3412 c->vector_fmul_add_add = vector_fmul_add_add_3dnow; // faster than sse2 | |
| 3206 } | 3413 } |
| 3207 | 3414 |
| 3208 #ifdef CONFIG_ENCODERS | 3415 #ifdef CONFIG_ENCODERS |
| 3209 dsputil_init_pix_mmx(c, avctx); | 3416 dsputil_init_pix_mmx(c, avctx); |
| 3210 #endif //CONFIG_ENCODERS | 3417 #endif //CONFIG_ENCODERS |