src/imageeffects_sse.cpp

   1 /*
   2   Copyright (c) 2006 Paolo Capriotti <p.capriotti@sns.it>
   3             (c) 2006 Maurizio Monge <maurizio.monge@kdemail.net>
   4
   5   This program is free software; you can redistribute it and/or modify
   6   it under the terms of the GNU General Public License as published by
   7   the Free Software Foundation; either version 2 of the License, or
   8   (at your option) any later version.
   9 */
  10
  11 #include <math.h>
  12 #include <inttypes.h>
  13 #include <xmmintrin.h>
  14 #include <QImage>
  15
  16
  17 namespace ImageEffects {
  18
  19 union vec4i
  20 {
  21   uint16_t i[8];
  22   __m128i v;
  23 };
  24
  25 static inline void blur_sse_near(void* pixels, __m128i& state, __m128i alpha)
  26 {
  27   uint64_t z1 = 0ULL;
  28   uint64_t z2 = 0ULL;
  29   uint64_t z3 = 0ULL;
  30
  31   asm(
  32     "movq %[ppix], %[pixels]\n"
  33     "punpcklbw %[pixels], %[aux1]\n"           // unpack two pixels setting their bytes
  34                                                // as the most significant in the corr. word
  35     "psrlw $1, %[aux1]\n"                      // shift right by 1, i.e. shift the colour
  36                                                // bytes left by 7
  37     "psubw %[state], %[aux1]\n"                // - state
  38     "pmulhw %[alpha], %[aux1]\n"               // * alpha, and take the 16 most significant bits
  39     "psllw $1, %[aux1]\n"                      // shift left (we trade 1 bit for performance, here)
  40     "paddw %[aux1], %[state]\n"                // sum result to state
  41     "movdqa %[state], %[aux2]\n"                 // copy state to the aux2 register
  42     "psrlw $7, %[aux2]\n"                      // shift right by 7: this is the new pixel value
  43     "packuswb %[aux2], %[aux2]\n"              // pack pixels as 8 bits
  44     "movq %[aux2], %[ppix]\n"
  45   : [state] "+x"(state)
  46   , [ppix] "+m"(*(uint64_t*)pixels)
  47   , [aux1] "+x"(z1)
  48   , [aux2] "+x"(z2)
  49   , [pixels] "+x"(z3)
  50   : [alpha] "x"(alpha)
  51   );
  52 }
  53
  54 static inline void blur_sse_sep(void* pixel1, void* pixel2, __m128i& state, __m128i alpha)
  55 {
  56   uint64_t z1 = 0ULL;
  57   uint64_t z2 = 0ULL;
  58   uint64_t z3 = 0ULL;
  59   uint64_t z4 = 0ULL;
  60
  61   asm(
  62     "movd %[ppix1], %[pixels]\n"             // load the first pixel
  63     "movd %[ppix2], %[tmp]\n"                // load the second pixel in [tmp]
  64     "pslldq $4, %[tmp]\n"                      // shift left the second pixel
  65     "paddd %[tmp], %[pixels]\n"                // now both pixel are packed in [pixels]
  66
  67     "punpcklbw %[pixels], %[aux1]\n"           // unpack two pixels setting their bytes
  68                                                // as the most significant in the corr. word
  69     "psrlw $1, %[aux1]\n"                      // shift right by 1, i.e. shift the colour
  70                                                // bytes left by 7
  71     "psubw %[state], %[aux1]\n"                // - state
  72     "pmulhw %[alpha], %[aux1]\n"               // * alpha, and take the 16 most significant bits
  73     "psllw $1, %[aux1]\n"                      // shift left (we trade 1 bit for performance, here)
  74     "paddw %[aux1], %[state]\n"                // sum result to state
  75     "movdqa %[state], %[aux2]\n"                 // copy state to the aux2 register
  76     "psrlw $7, %[aux2]\n"                      // shift right by 7: this is the new pixel value
  77     "packuswb %[aux2], %[aux2]\n"              // pack pixels as 8 bits
  78
  79     "movd %[aux2], %[ppix1]\n"
  80     "psrldq $4, %[aux2]\n"
  81     "movd %[aux2], %[ppix2]\n"
  82   : [state] "+x"(state)
  83   , [ppix1] "+m"(*(uint32_t*)pixel1)
  84   , [ppix2] "+m"(*(uint32_t*)pixel2)
  85   , [aux1] "+x"(z1)
  86   , [aux2] "+x"(z2)
  87   , [tmp]  "+x"(z3)
  88   , [pixels] "+x"(z4)
  89   : [alpha] "x"(alpha)
  90   );
  91 }
  92
  93 void expblur_sse( QImage &img, int radius )
  94 {
  95   if(radius<1)
  96     return;
  97
  98   /* Calculate the alpha such that 90% of
  99      the kernel is within the radius.
 100      (Kernel extends to infinity)
 101   */
 102   uint16_t alpha = (uint16_t)((1<<15)*(1.0f-expf(-2.3f/(radius+1.f))));
 103
 104   vec4i a;
 105   QRgb *ptr = (QRgb *)img.bits();
 106   int h = img.height();
 107   int w = img.width();
 108   int hw = (img.height()-1)*img.width();
 109   for(int i=0;i<8;i++)
 110     a.i[i] = alpha;
 111
 112   for(int row=0;row<h-1;row+=2)
 113   {
 114     vec4i z;
 115     uint8_t *cptr = (uint8_t*)(ptr+row*w);
 116     for(int i=0;i<4;i++)
 117       z.i[i] = cptr[i]<<7;
 118     for(int i=0;i<4;i++)
 119       z.i[4+i] = cptr[w*4+i]<<7;
 120
 121     for(int index=1; index<w; index++)
 122       blur_sse_sep(&cptr[index*4], &cptr[(index+w)*4], z.v, a.v);
 123
 124     for(int index=w-2; index>=0; index--)
 125       blur_sse_sep(&cptr[index*4], &cptr[(index+w)*4] , z.v, a.v);
 126   }
 127
 128   if(h & 1)
 129   {
 130     vec4i z;
 131     int dummy;
 132     uint8_t *cptr = (uint8_t*)(ptr+(h-1)*w);
 133     for(int i=0;i<4;i++)
 134       z.i[i] = cptr[i]<<7;
 135
 136     for(int index=1; index<w; index++)
 137       blur_sse_sep(&cptr[index*4], &dummy, z.v, a.v);
 138
 139     for(int index=w-2; index>=0; index--)
 140       blur_sse_sep(&cptr[index*4], &dummy, z.v, a.v);
 141   }
 142
 143   for(int col=0;col<w-1;col+=2)
 144   {
 145     vec4i z;
 146     uint8_t *cptr = (uint8_t*)(ptr+col);
 147
 148     for(int i=0;i<8;i++)
 149       z.i[i] = cptr[i]<<7;
 150
 151     for(int index=w; index<hw; index+=w)
 152       blur_sse_near(&cptr[index*4], z.v, a.v);
 153
 154     for(int index=hw-2*w; index>=0; index-=w)
 155       blur_sse_near(&cptr[index*4], z.v, a.v);
 156   }
 157
 158   if(w & 1)
 159   {
 160     vec4i z;
 161     int dummy;
 162     uint8_t *cptr = (uint8_t*)(ptr+w-1);
 163
 164     for(int i=0;i<4;i++)
 165       z.i[i] = cptr[i]<<7;
 166
 167     for(int index=w; index<hw; index+=w)
 168       blur_sse_sep(&cptr[index*4], &dummy, z.v, a.v);
 169
 170     for(int index=hw-w; index>=0; index-=w)
 171       blur_sse_sep(&cptr[index*4], &dummy, z.v, a.v);
 172   }
 173
 174   return;
 175 }
 176
 177 } //end namespace ImageEffects