diff options
Diffstat (limited to 'src/libffmpeg/libavcodec/i386/dsputil_mmx.c')
| -rw-r--r-- | src/libffmpeg/libavcodec/i386/dsputil_mmx.c | 211 |
1 files changed, 155 insertions, 56 deletions
diff --git a/src/libffmpeg/libavcodec/i386/dsputil_mmx.c b/src/libffmpeg/libavcodec/i386/dsputil_mmx.c index 02558604b..4336e4bde 100644 --- a/src/libffmpeg/libavcodec/i386/dsputil_mmx.c +++ b/src/libffmpeg/libavcodec/i386/dsputil_mmx.c @@ -343,7 +343,7 @@ static void add_pixels_clamped_mmx(const DCTELEM *block, UINT8 *pixels, int line } while (--i); } -static void put_pixels_mmx(UINT8 *block, const UINT8 *pixels, int line_size, int h) +static void put_pixels8_mmx(UINT8 *block, const UINT8 *pixels, int line_size, int h) { __asm __volatile( "lea (%3, %3), %%eax \n\t" @@ -369,6 +369,40 @@ static void put_pixels_mmx(UINT8 *block, const UINT8 *pixels, int line_size, int ); } +static void put_pixels16_mmx(UINT8 *block, const UINT8 *pixels, int line_size, int h) +{ + __asm __volatile( + "lea (%3, %3), %%eax \n\t" + ".balign 8 \n\t" + "1: \n\t" + "movq (%1), %%mm0 \n\t" + "movq 8(%1), %%mm4 \n\t" + "movq (%1, %3), %%mm1 \n\t" + "movq 8(%1, %3), %%mm5 \n\t" + "movq %%mm0, (%2) \n\t" + "movq %%mm4, 8(%2) \n\t" + "movq %%mm1, (%2, %3) \n\t" + "movq %%mm5, 8(%2, %3) \n\t" + "addl %%eax, %1 \n\t" + "addl %%eax, %2 \n\t" + "movq (%1), %%mm0 \n\t" + "movq 8(%1), %%mm4 \n\t" + "movq (%1, %3), %%mm1 \n\t" + "movq 8(%1, %3), %%mm5 \n\t" + "movq %%mm0, (%2) \n\t" + "movq %%mm4, 8(%2) \n\t" + "movq %%mm1, (%2, %3) \n\t" + "movq %%mm5, 8(%2, %3) \n\t" + "addl %%eax, %1 \n\t" + "addl %%eax, %2 \n\t" + "subl $4, %0 \n\t" + "jnz 1b \n\t" + : "+g"(h), "+r" (pixels), "+r" (block) + : "r"(line_size) + : "%eax", "memory" + ); +} + static void clear_blocks_mmx(DCTELEM *blocks) { __asm __volatile( @@ -393,19 +427,19 @@ static void just_return() { return; } void dsputil_init_mmx(void) { mm_flags = mm_support(); -#if 1 - printf("libavcodec: CPU flags:"); +#if 0 + fprintf(stderr, "libavcodec: CPU flags:"); if (mm_flags & MM_MMX) - printf(" mmx"); + fprintf(stderr, " mmx"); if (mm_flags & MM_MMXEXT) - printf(" mmxext"); + fprintf(stderr, " mmxext"); if (mm_flags & MM_3DNOW) - printf(" 3dnow"); + fprintf(stderr, " 3dnow"); if (mm_flags & MM_SSE) - printf(" sse"); + fprintf(stderr, " sse"); if (mm_flags & MM_SSE2) - printf(" sse2"); - printf("\n"); + fprintf(stderr, " sse2"); + fprintf(stderr, "\n"); #endif if (mm_flags & MM_MMX) { @@ -424,27 +458,45 @@ void dsputil_init_mmx(void) pix_abs8x8_y2 = pix_abs8x8_y2_mmx; pix_abs8x8_xy2= pix_abs8x8_xy2_mmx; - av_fdct = fdct_mmx; - - put_pixels_tab[0] = put_pixels_mmx; - put_pixels_tab[1] = put_pixels_x2_mmx; - put_pixels_tab[2] = put_pixels_y2_mmx; - put_pixels_tab[3] = put_pixels_xy2_mmx; - - put_no_rnd_pixels_tab[0] = put_pixels_mmx; - put_no_rnd_pixels_tab[1] = put_no_rnd_pixels_x2_mmx; - put_no_rnd_pixels_tab[2] = put_no_rnd_pixels_y2_mmx; - put_no_rnd_pixels_tab[3] = put_no_rnd_pixels_xy2_mmx; - - avg_pixels_tab[0] = avg_pixels_mmx; - avg_pixels_tab[1] = avg_pixels_x2_mmx; - avg_pixels_tab[2] = avg_pixels_y2_mmx; - avg_pixels_tab[3] = avg_pixels_xy2_mmx; - - avg_no_rnd_pixels_tab[0] = avg_no_rnd_pixels_mmx; - avg_no_rnd_pixels_tab[1] = avg_no_rnd_pixels_x2_mmx; - avg_no_rnd_pixels_tab[2] = avg_no_rnd_pixels_y2_mmx; - avg_no_rnd_pixels_tab[3] = avg_no_rnd_pixels_xy2_mmx; + put_pixels_tab[0][0] = put_pixels16_mmx; + put_pixels_tab[0][1] = put_pixels16_x2_mmx; + put_pixels_tab[0][2] = put_pixels16_y2_mmx; + put_pixels_tab[0][3] = put_pixels16_xy2_mmx; + + put_no_rnd_pixels_tab[0][0] = put_pixels16_mmx; + put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_mmx; + put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_mmx; + put_no_rnd_pixels_tab[0][3] = put_no_rnd_pixels16_xy2_mmx; + + avg_pixels_tab[0][0] = avg_pixels16_mmx; + avg_pixels_tab[0][1] = avg_pixels16_x2_mmx; + avg_pixels_tab[0][2] = avg_pixels16_y2_mmx; + avg_pixels_tab[0][3] = avg_pixels16_xy2_mmx; + + avg_no_rnd_pixels_tab[0][0] = avg_no_rnd_pixels16_mmx; + avg_no_rnd_pixels_tab[0][1] = avg_no_rnd_pixels16_x2_mmx; + avg_no_rnd_pixels_tab[0][2] = avg_no_rnd_pixels16_y2_mmx; + avg_no_rnd_pixels_tab[0][3] = avg_no_rnd_pixels16_xy2_mmx; + + put_pixels_tab[1][0] = put_pixels8_mmx; + put_pixels_tab[1][1] = put_pixels8_x2_mmx; + put_pixels_tab[1][2] = put_pixels8_y2_mmx; + put_pixels_tab[1][3] = put_pixels8_xy2_mmx; + + put_no_rnd_pixels_tab[1][0] = put_pixels8_mmx; + put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_mmx; + put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_mmx; + put_no_rnd_pixels_tab[1][3] = put_no_rnd_pixels8_xy2_mmx; + + avg_pixels_tab[1][0] = avg_pixels8_mmx; + avg_pixels_tab[1][1] = avg_pixels8_x2_mmx; + avg_pixels_tab[1][2] = avg_pixels8_y2_mmx; + avg_pixels_tab[1][3] = avg_pixels8_xy2_mmx; + + avg_no_rnd_pixels_tab[1][0] = avg_no_rnd_pixels8_mmx; + avg_no_rnd_pixels_tab[1][1] = avg_no_rnd_pixels8_x2_mmx; + avg_no_rnd_pixels_tab[1][2] = avg_no_rnd_pixels8_y2_mmx; + avg_no_rnd_pixels_tab[1][3] = avg_no_rnd_pixels8_xy2_mmx; if (mm_flags & MM_MMXEXT) { pix_abs16x16 = pix_abs16x16_mmx2; @@ -457,25 +509,45 @@ void dsputil_init_mmx(void) pix_abs8x8_y2 = pix_abs8x8_y2_mmx2; pix_abs8x8_xy2= pix_abs8x8_xy2_mmx2; - put_pixels_tab[1] = put_pixels_x2_mmx2; - put_pixels_tab[2] = put_pixels_y2_mmx2; - put_no_rnd_pixels_tab[1] = put_no_rnd_pixels_x2_mmx2; - put_no_rnd_pixels_tab[2] = put_no_rnd_pixels_y2_mmx2; - - avg_pixels_tab[0] = avg_pixels_mmx2; - avg_pixels_tab[1] = avg_pixels_x2_mmx2; - avg_pixels_tab[2] = avg_pixels_y2_mmx2; - avg_pixels_tab[3] = avg_pixels_xy2_mmx2; + put_pixels_tab[0][1] = put_pixels16_x2_mmx2; + put_pixels_tab[0][2] = put_pixels16_y2_mmx2; + put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_mmx2; + put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_mmx2; + + avg_pixels_tab[0][0] = avg_pixels16_mmx2; + avg_pixels_tab[0][1] = avg_pixels16_x2_mmx2; + avg_pixels_tab[0][2] = avg_pixels16_y2_mmx2; + avg_pixels_tab[0][3] = avg_pixels16_xy2_mmx2; + + put_pixels_tab[1][1] = put_pixels8_x2_mmx2; + put_pixels_tab[1][2] = put_pixels8_y2_mmx2; + put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_mmx2; + put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_mmx2; + + avg_pixels_tab[1][0] = avg_pixels8_mmx2; + avg_pixels_tab[1][1] = avg_pixels8_x2_mmx2; + avg_pixels_tab[1][2] = avg_pixels8_y2_mmx2; + avg_pixels_tab[1][3] = avg_pixels8_xy2_mmx2; } else if (mm_flags & MM_3DNOW) { - put_pixels_tab[1] = put_pixels_x2_3dnow; - put_pixels_tab[2] = put_pixels_y2_3dnow; - put_no_rnd_pixels_tab[1] = put_no_rnd_pixels_x2_3dnow; - put_no_rnd_pixels_tab[2] = put_no_rnd_pixels_y2_3dnow; - - avg_pixels_tab[0] = avg_pixels_3dnow; - avg_pixels_tab[1] = avg_pixels_x2_3dnow; - avg_pixels_tab[2] = avg_pixels_y2_3dnow; - avg_pixels_tab[3] = avg_pixels_xy2_3dnow; + put_pixels_tab[0][1] = put_pixels16_x2_3dnow; + put_pixels_tab[0][2] = put_pixels16_y2_3dnow; + put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_3dnow; + put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_3dnow; + + avg_pixels_tab[0][0] = avg_pixels16_3dnow; + avg_pixels_tab[0][1] = avg_pixels16_x2_3dnow; + avg_pixels_tab[0][2] = avg_pixels16_y2_3dnow; + avg_pixels_tab[0][3] = avg_pixels16_xy2_3dnow; + + put_pixels_tab[1][1] = put_pixels8_x2_3dnow; + put_pixels_tab[1][2] = put_pixels8_y2_3dnow; + put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_3dnow; + put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_3dnow; + + avg_pixels_tab[1][0] = avg_pixels8_3dnow; + avg_pixels_tab[1][1] = avg_pixels8_x2_3dnow; + avg_pixels_tab[1][2] = avg_pixels8_y2_3dnow; + avg_pixels_tab[1][3] = avg_pixels8_xy2_3dnow; } /* idct */ @@ -526,27 +598,54 @@ void dsputil_init_mmx(void) #endif } +void gen_idct_put(UINT8 *dest, int line_size, DCTELEM *block); + +/** + * this will send coeff matrixes which would have different results for the 16383 type MMX vs C IDCTs to the C IDCT + */ +void bit_exact_idct_put(UINT8 *dest, int line_size, INT16 *block){ + if( block[0]>1022 && block[1]==0 && block[4 ]==0 && block[5 ]==0 + && block[8]==0 && block[9]==0 && block[12]==0 && block[13]==0){ + int16_t tmp[64]; + int i; + + for(i=0; i<64; i++) + tmp[i]= block[i]; + for(i=0; i<64; i++) + block[i]= tmp[block_permute_op(i)]; + + simple_idct_put(dest, line_size, block); + } + else + gen_idct_put(dest, line_size, block); +} + /* remove any non bit exact operation (testing purpose). NOTE that this function should be kept as small as possible because it is always difficult to test automatically non bit exact cases. */ void dsputil_set_bit_exact_mmx(void) { if (mm_flags & MM_MMX) { - if (mm_flags & MM_MMXEXT) { - put_no_rnd_pixels_tab[1] = put_no_rnd_pixels_x2_mmx; - put_no_rnd_pixels_tab[2] = put_no_rnd_pixels_y2_mmx; - avg_pixels_tab[3] = avg_pixels_xy2_mmx; + + /* MMX2 & 3DNOW */ + put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_mmx; + put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_mmx; + avg_pixels_tab[0][3] = avg_pixels16_xy2_mmx; + put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_mmx; + put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_mmx; + avg_pixels_tab[1][3] = avg_pixels8_xy2_mmx; + if (mm_flags & MM_MMXEXT) { pix_abs16x16_x2 = pix_abs16x16_x2_mmx; pix_abs16x16_y2 = pix_abs16x16_y2_mmx; pix_abs16x16_xy2 = pix_abs16x16_xy2_mmx; pix_abs8x8_x2 = pix_abs8x8_x2_mmx; pix_abs8x8_y2 = pix_abs8x8_y2_mmx; pix_abs8x8_xy2= pix_abs8x8_xy2_mmx; - } else if (mm_flags & MM_3DNOW) { - put_no_rnd_pixels_tab[1] = put_no_rnd_pixels_x2_mmx; - put_no_rnd_pixels_tab[2] = put_no_rnd_pixels_y2_mmx; - avg_pixels_tab[3] = avg_pixels_xy2_mmx; } +#ifdef SIMPLE_IDCT + if(ff_idct_put==gen_idct_put && ff_idct == simple_idct_mmx) + ff_idct_put= bit_exact_idct_put; +#endif } } |