From d6a4c0b111d41bbf1effd771605dfb71fd8903cf Mon Sep 17 00:00:00 2001 From: Zdenek Kabelac Date: Mon, 18 Feb 2002 09:40:05 +0000 Subject: [PATCH] * some modifications to allow gcc to compile same code for -fPIC - introduced MOVQ_WONE, MOVQ_WTWO * unrolled put_pixels_clamped_mmx * added .baling where the jump address would be compiled on odd address * added some testing code to disable all these routines (for time measurement) Originally committed as revision 308 to svn://svn.ffmpeg.org/ffmpeg/trunk --- libavcodec/i386/dsputil_mmx.c | 252 +++++++++++++++++++++++----------- 1 file changed, 172 insertions(+), 80 deletions(-) diff --git a/libavcodec/i386/dsputil_mmx.c b/libavcodec/i386/dsputil_mmx.c index c9045d02ad..6b35d47534 100644 --- a/libavcodec/i386/dsputil_mmx.c +++ b/libavcodec/i386/dsputil_mmx.c @@ -40,6 +40,27 @@ static const unsigned long long int mm_wtwo __attribute__ ((aligned(8))) = 0x000 //static const unsigned short mm_wone[4] __attribute__ ((aligned(8))) = { 0x1, 0x1, 0x1, 0x1 }; //static const unsigned short mm_wtwo[4] __attribute__ ((aligned(8))) = { 0x2, 0x2, 0x2, 0x2 }; +#define JUMPALIGN() __asm __volatile (".balign 8"::) +#define MOVQ_ZERO(regd) __asm __volatile ("pxor %%" #regd ", %%" #regd ::) + +#ifndef PIC +#define MOVQ_WONE(regd) __asm __volatile ("movq %0, %%" #regd " \n\t" ::"m"(mm_wone)) +#define MOVQ_WTWO(regd) __asm __volatile ("movq %0, %%" #regd " \n\t" ::"m"(mm_wtwo)) +#else +// for shared library it's better to use this way for accessing constants +// pcmpeqd -> -1 +#define MOVQ_WONE(regd) \ + __asm __volatile ( \ + "pcmpeqd %%" #regd ", %%" #regd " \n\t" \ + "psrlw $15, %%" #regd ::) + +#define MOVQ_WTWO(regd) \ + __asm __volatile ( \ + "pcmpeqd %%" #regd ", %%" #regd " \n\t" \ + "psrlw $15, %%" #regd " \n\t" \ + "psllw $1, %%" #regd ::) +#endif + /***********************************/ /* 3Dnow specific */ @@ -77,7 +98,7 @@ static void get_pixels_mmx(DCTELEM *block, const UINT8 *pixels, int line_size) /* read the pixels */ p = block; pix = pixels; - __asm __volatile("pxor %%mm7, %%mm7":); + MOVQ_ZERO(mm7); for(i=0;i<4;i++) { __asm __volatile( "movq %1, %%mm0\n\t" @@ -104,12 +125,11 @@ static void put_pixels_clamped_mmx(const DCTELEM *block, UINT8 *pixels, int line { const DCTELEM *p; UINT8 *pix; - int i; /* read the pixels */ p = block; pix = pixels; - for(i=0;i<2;i++) { + /* unrolled loop */ __asm __volatile( "movq %3, %%mm0\n\t" "movq 8%3, %%mm1\n\t" @@ -131,7 +151,29 @@ static void put_pixels_clamped_mmx(const DCTELEM *block, UINT8 *pixels, int line :"memory"); pix += line_size*4; p += 32; - } + + // if here would be an exact copy of the code above + // compiler would generate some very strange code + // thus using "r" + __asm __volatile( + "movq (%3), %%mm0\n\t" + "movq 8(%3), %%mm1\n\t" + "movq 16(%3), %%mm2\n\t" + "movq 24(%3), %%mm3\n\t" + "movq 32(%3), %%mm4\n\t" + "movq 40(%3), %%mm5\n\t" + "movq 48(%3), %%mm6\n\t" + "movq 56(%3), %%mm7\n\t" + "packuswb %%mm1, %%mm0\n\t" + "packuswb %%mm3, %%mm2\n\t" + "packuswb %%mm5, %%mm4\n\t" + "packuswb %%mm7, %%mm6\n\t" + "movq %%mm0, (%0)\n\t" + "movq %%mm2, (%0, %1)\n\t" + "movq %%mm4, (%0, %1, 2)\n\t" + "movq %%mm6, (%0, %2)\n\t" + ::"r" (pix), "r" (line_size), "r" (line_size*3), "r"(p) + :"memory"); } static void add_pixels_clamped_mmx(const DCTELEM *block, UINT8 *pixels, int line_size) @@ -143,8 +185,9 @@ static void add_pixels_clamped_mmx(const DCTELEM *block, UINT8 *pixels, int line /* read the pixels */ p = block; pix = pixels; - __asm __volatile("pxor %%mm7, %%mm7":); - for(i=0;i<4;i++) { + MOVQ_ZERO(mm7); + i = 4; + while (i) { __asm __volatile( "movq %2, %%mm0\n\t" "movq 8%2, %%mm1\n\t" @@ -171,19 +214,47 @@ static void add_pixels_clamped_mmx(const DCTELEM *block, UINT8 *pixels, int line :"memory"); pix += line_size*2; p += 16; - } + i--; + }; } static void put_pixels_mmx(UINT8 *block, const UINT8 *pixels, int line_size, int h) { - int dh, hh; + int hh; UINT8 *p; const UINT8 *pix; + p = block; - pix = pixels; + pix = pixels; // 2s +#if 0 + do { + __asm __volatile( + "movq %1, %%mm0\n\t" + "movq %%mm0, %0\n\t" + :"=m"(*p) + :"m"(*pix) + :"memory"); + pix += line_size; + p += line_size; + } while (--h); +#else + // this optimized code is not very usefull + // the above loop is definitely faster + // at least on Celeron 500MHz + hh = h & 3; + while (hh) { + __asm __volatile( + "movq %1, %%mm0\n\t" + "movq %%mm0, %0\n\t" + :"=m"(*p) + :"m"(*pix) + :"memory"); + pix += line_size; + p += line_size; + hh--; + } hh=h>>2; - dh=h&3; - while(hh--) { + while (hh) { __asm __volatile( "movq (%1), %%mm0 \n\t" "movq (%1, %2), %%mm1 \n\t" @@ -195,19 +266,11 @@ static void put_pixels_mmx(UINT8 *block, const UINT8 *pixels, int line_size, int "movq %%mm3, (%0, %3) \n\t" ::"r"(p), "r"(pix), "r"(line_size), "r"(line_size*3) :"memory"); - pix = pix + line_size*4; - p = p + line_size*4; - } - while(dh--) { - __asm __volatile( - "movq %1, %%mm0\n\t" - "movq %%mm0, %0\n\t" - :"=m"(*p) - :"m"(*pix) - :"memory"); - pix = pix + line_size; - p = p + line_size; + pix += line_size*4; + p += line_size*4; + hh--; } +#endif } static void put_pixels_x2_mmx(UINT8 *block, const UINT8 *pixels, int line_size, int h) @@ -216,10 +279,9 @@ static void put_pixels_x2_mmx(UINT8 *block, const UINT8 *pixels, int line_size, const UINT8 *pix; p = block; pix = pixels; - __asm __volatile( - "pxor %%mm7, %%mm7\n\t" - "movq %0, %%mm4\n\t" - ::"m"(mm_wone)); + MOVQ_ZERO(mm7); + MOVQ_WONE(mm4); + JUMPALIGN(); do { __asm __volatile( "movq %1, %%mm0\n\t" @@ -251,10 +313,9 @@ static void put_pixels_y2_mmx(UINT8 *block, const UINT8 *pixels, int line_size, const UINT8 *pix; p = block; pix = pixels; - __asm __volatile( - "pxor %%mm7, %%mm7\n\t" - "movq %0, %%mm4\n\t" - ::"m"(mm_wone)); + MOVQ_ZERO(mm7); + MOVQ_WONE(mm4); + JUMPALIGN(); do { __asm __volatile( "movq %1, %%mm0\n\t" @@ -287,11 +348,10 @@ static void put_pixels_xy2_mmx(UINT8 *block, const UINT8 *pixels, int line_size, UINT8 *p; const UINT8 *pix; p = block; - pix = pixels; - __asm __volatile( - "pxor %%mm7, %%mm7\n\t" - "movq %0, %%mm6\n\t" - ::"m"(mm_wtwo)); + pix = pixels; // 1s + MOVQ_ZERO(mm7); + MOVQ_WTWO(mm6); + JUMPALIGN(); do { __asm __volatile( "movq %1, %%mm0\n\t" @@ -337,7 +397,7 @@ static void put_no_rnd_pixels_x2_mmx( UINT8 *block, const UINT8 *pixels, int const UINT8 *pix; p = block; pix = pixels; - __asm __volatile("pxor %%mm7, %%mm7\n\t":); + MOVQ_ZERO(mm7); do { __asm __volatile( "movq %1, %%mm0\n\t" @@ -368,7 +428,8 @@ static void put_no_rnd_pixels_y2_mmx( UINT8 *block, const UINT8 *pixels, int li const UINT8 *pix; p = block; pix = pixels; - __asm __volatile("pxor %%mm7, %%mm7\n\t":); + MOVQ_ZERO(mm7); + JUMPALIGN(); do { __asm __volatile( "movq %1, %%mm0\n\t" @@ -400,10 +461,9 @@ static void put_no_rnd_pixels_xy2_mmx( UINT8 *block, const UINT8 *pixels, int const UINT8 *pix; p = block; pix = pixels; - __asm __volatile( - "pxor %%mm7, %%mm7\n\t" - "movq %0, %%mm6\n\t" - ::"m"(mm_wone)); + MOVQ_ZERO(mm7); + MOVQ_WONE(mm6); + JUMPALIGN(); do { __asm __volatile( "movq %1, %%mm0\n\t" @@ -449,10 +509,9 @@ static void avg_pixels_mmx(UINT8 *block, const UINT8 *pixels, int line_size, int const UINT8 *pix; p = block; pix = pixels; - __asm __volatile( - "pxor %%mm7, %%mm7\n\t" - "movq %0, %%mm6\n\t" - ::"m"(mm_wone)); + MOVQ_ZERO(mm7); + MOVQ_WONE(mm6); + JUMPALIGN(); do { __asm __volatile( "movq %0, %%mm0\n\t" @@ -486,10 +545,9 @@ static void avg_pixels_x2_mmx( UINT8 *block, const UINT8 *pixels, int line_si const UINT8 *pix; p = block; pix = pixels; - __asm __volatile( - "pxor %%mm7, %%mm7\n\t" - "movq %0, %%mm6\n\t" - ::"m"(mm_wone)); + MOVQ_ZERO(mm7); + MOVQ_WONE(mm6); + JUMPALIGN(); do { __asm __volatile( "movq %1, %%mm1\n\t" @@ -532,10 +590,9 @@ static void avg_pixels_y2_mmx( UINT8 *block, const UINT8 *pixels, int line_si const UINT8 *pix; p = block; pix = pixels; - __asm __volatile( - "pxor %%mm7, %%mm7\n\t" - "movq %0, %%mm6\n\t" - ::"m"(mm_wone)); + MOVQ_ZERO(mm7); + MOVQ_WONE(mm6); + JUMPALIGN(); do { __asm __volatile( "movq %1, %%mm1\n\t" @@ -578,10 +635,10 @@ static void avg_pixels_xy2_mmx( UINT8 *block, const UINT8 *pixels, int line_s const UINT8 *pix; p = block; pix = pixels; - __asm __volatile( - "pxor %%mm7, %%mm7\n\t" - "movq %0, %%mm6\n\t" - ::"m"(mm_wtwo)); + MOVQ_ZERO(mm7); + // this doesn't seem to be used offten - so + // the inside usage of mm_wone is not optimized + MOVQ_WTWO(mm6); do { __asm __volatile( "movq %1, %%mm0\n\t" @@ -638,7 +695,7 @@ static void avg_no_rnd_pixels_mmx( UINT8 *block, const UINT8 *pixels, int line_ const UINT8 *pix; p = block; pix = pixels; - __asm __volatile("pxor %%mm7, %%mm7\n\t":); + MOVQ_ZERO(mm7); do { __asm __volatile( "movq %1, %%mm0\n\t" @@ -669,8 +726,7 @@ static void avg_no_rnd_pixels_x2_mmx( UINT8 *block, const UINT8 *pixels, int const UINT8 *pix; p = block; pix = pixels; - __asm __volatile( - "pxor %%mm7, %%mm7\n\t":); + MOVQ_ZERO(mm7); do { __asm __volatile( "movq %1, %%mm0\n\t" @@ -709,8 +765,7 @@ static void avg_no_rnd_pixels_y2_mmx( UINT8 *block, const UINT8 *pixels, int const UINT8 *pix; p = block; pix = pixels; - __asm __volatile( - "pxor %%mm7, %%mm7\n\t":); + MOVQ_ZERO(mm7); do { __asm __volatile( "movq %1, %%mm0\n\t" @@ -749,10 +804,9 @@ static void avg_no_rnd_pixels_xy2_mmx( UINT8 *block, const UINT8 *pixels, int const UINT8 *pix; p = block; pix = pixels; - __asm __volatile( - "pxor %%mm7, %%mm7\n\t" - "movq %0, %%mm6\n\t" - ::"m"(mm_wone)); + MOVQ_ZERO(mm7); + MOVQ_WONE(mm6); + JUMPALIGN(); do { __asm __volatile( "movq %1, %%mm0\n\t" @@ -806,7 +860,7 @@ static void sub_pixels_mmx( DCTELEM *block, const UINT8 *pixels, int line_size, const UINT8 *pix; p = block; pix = pixels; - __asm __volatile("pxor %%mm7, %%mm7":); + MOVQ_ZERO(mm7); do { __asm __volatile( "movq %0, %%mm0\n\t" @@ -833,10 +887,9 @@ static void sub_pixels_x2_mmx( DCTELEM *block, const UINT8 *pixels, int line_si const UINT8 *pix; p = block; pix = pixels; - __asm __volatile( - "pxor %%mm7, %%mm7\n\t" - "movq %0, %%mm6" - ::"m"(mm_wone)); + MOVQ_ZERO(mm7); + MOVQ_WONE(mm6); + JUMPALIGN(); do { __asm __volatile( "movq %0, %%mm0\n\t" @@ -873,10 +926,8 @@ static void sub_pixels_y2_mmx( DCTELEM *block, const UINT8 *pixels, int line_si const UINT8 *pix; p = block; pix = pixels; - __asm __volatile( - "pxor %%mm7, %%mm7\n\t" - "movq %0, %%mm6" - ::"m"(mm_wone)); + MOVQ_ZERO(mm7); + MOVQ_WONE(mm6); do { __asm __volatile( "movq %0, %%mm0\n\t" @@ -913,10 +964,9 @@ static void sub_pixels_xy2_mmx( DCTELEM *block, const UINT8 *pixels, int line const UINT8 *pix; p = block; pix = pixels; - __asm __volatile( - "pxor %%mm7, %%mm7\n\t" - "movq %0, %%mm6\n\t" - ::"m"(mm_wtwo)); + MOVQ_ZERO(mm7); + MOVQ_WTWO(mm6); + JUMPALIGN(); do { __asm __volatile( "movq %1, %%mm0\n\t" @@ -960,6 +1010,8 @@ static void sub_pixels_xy2_mmx( DCTELEM *block, const UINT8 *pixels, int line } while(--h); } +static void just_return() { return; } + void dsputil_init_mmx(void) { mm_flags = mm_support(); @@ -1053,4 +1105,44 @@ void dsputil_init_mmx(void) ff_idct = simple_idct_mmx; #endif } + +#if 0 + // for speed testing + get_pixels = just_return; + put_pixels_clamped = just_return; + add_pixels_clamped = just_return; + + pix_abs16x16 = just_return; + pix_abs16x16_x2 = just_return; + pix_abs16x16_y2 = just_return; + pix_abs16x16_xy2 = just_return; + + put_pixels_tab[0] = just_return; + put_pixels_tab[1] = just_return; + put_pixels_tab[2] = just_return; + put_pixels_tab[3] = just_return; + + put_no_rnd_pixels_tab[0] = just_return; + put_no_rnd_pixels_tab[1] = just_return; + put_no_rnd_pixels_tab[2] = just_return; + put_no_rnd_pixels_tab[3] = just_return; + + avg_pixels_tab[0] = just_return; + avg_pixels_tab[1] = just_return; + avg_pixels_tab[2] = just_return; + avg_pixels_tab[3] = just_return; + + avg_no_rnd_pixels_tab[0] = just_return; + avg_no_rnd_pixels_tab[1] = just_return; + avg_no_rnd_pixels_tab[2] = just_return; + avg_no_rnd_pixels_tab[3] = just_return; + + sub_pixels_tab[0] = just_return; + sub_pixels_tab[1] = just_return; + sub_pixels_tab[2] = just_return; + sub_pixels_tab[3] = just_return; + + //av_fdct = just_return; + //ff_idct = just_return; +#endif }