src/library/framebuffer-cdrop.cpp

   1 #include "framebuffer.hpp"
   2 #include "arch-detect.hpp"
   3 #include <iostream>
   4
   5 namespace framebuffer
   6 {
   7 namespace
   8 {
   9         inline bool ssse3_available()
  10         {
  11                 size_t res = 0;
  12 #ifdef ARCH_IS_I386
  13                 size_t page = 1;
  14                 asm volatile(
  15                         "cpuid\n"
  16                         "\tshr $9,%0\n"
  17                         "\tand $1,%0\n"
  18                         : "=c"(res), "=a"(page) : "a"(page) : "%rbx", "%rdx");
  19 #endif
  20                 return res;
  21         }
  22
  23         const char mask_drop4_8[]  __attribute__ ((aligned (16))) = {
  24                  0,  1,  2,  4,  5,  6,  8,  9, 10, 12, 13, 14, -1, -1, -1, -1,         //0 -> 0
  25                 -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1,  0,  1,  2,  4,         //1 -> 0
  26                  5,  6,  8,  9, 10, 12, 13, 14, -1, -1, -1, -1, -1, -1, -1, -1,         //1 -> 1
  27                 -1, -1, -1, -1, -1, -1, -1, -1,  0,  1,  2,  4,  5,  6,  8,  9,         //2 -> 1
  28                 10, 12, 13, 14, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1,         //2 -> 2
  29                 -1, -1, -1, -1,  0,  1,  2,  4,  5,  6,  8,  9, 10, 12, 13, 14,         //3 -> 2
  30         };
  31         const char mask_drop4s_8[]  __attribute__ ((aligned (16))) = {
  32                  2,  1,  0,  6,  5,  4, 10,  9,  8, 14, 13, 12, -1, -1, -1, -1,         //0 -> 0
  33                 -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1,  2,  1,  0,  6,         //1 -> 0
  34                  5,  4, 10,  9,  8, 14, 13, 12, -1, -1, -1, -1, -1, -1, -1, -1,         //1 -> 1
  35                 -1, -1, -1, -1, -1, -1, -1, -1,  2,  1,  0,  6,  5,  4, 10,  9,         //2 -> 1
  36                  8, 14, 13, 12, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1,         //2 -> 2
  37                 -1, -1, -1, -1,  2,  1,  0,  6,  5,  4, 10,  9,  8, 14, 13, 12,         //3 -> 2
  38         };
  39         const char mask_swap4_8[]  __attribute__ ((aligned (16))) = {
  40                  2,  1,  0,  3,  6,  5,  4,  7, 10,  9,  8, 11, 14, 13, 12, 15,         //0 -> 0
  41         };
  42         const char mask_drop4_16[]  __attribute__ ((aligned (16))) = {
  43                  0,  1,  2,  3,  4,  5,  8,  9, 10, 11, 12, 13, -1, -1, -1, -1,         //0 -> 0
  44                 -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1,  0,  1,  2,  3,         //1 -> 0
  45                  4,  5,  8,  9, 10, 11, 12, 13, -1, -1, -1, -1, -1, -1, -1, -1,         //1 -> 1
  46                 -1, -1, -1, -1, -1, -1, -1, -1,  0,  1,  2,  3,  4,  5,  8,  9,         //2 -> 1
  47                 10, 11, 12, 13, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1,         //2 -> 2
  48                 -1, -1, -1, -1,  0,  1,  2,  3,  4,  5,  8,  9, 10, 11, 12, 13,         //3 -> 2
  49         };
  50         const char mask_drop4s_16[]  __attribute__ ((aligned (16))) = {
  51                  4,  5,  2,  3,  0,  1, 12, 13, 10, 11,  8,  9, -1, -1, -1, -1,         //0 -> 0
  52                 -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1,  4,  5,  2,  3,         //1 -> 0
  53                  0,  1, 12, 13, 10, 11,  8,  9, -1, -1, -1, -1, -1, -1, -1, -1,         //1 -> 1
  54                 -1, -1, -1, -1, -1, -1, -1, -1,  4,  5,  2,  3,  0,  1, 12, 13,         //2 -> 1
  55                 10, 11,  8,  9, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1,         //2 -> 2
  56                 -1, -1, -1, -1,  4,  5,  2,  3,  0,  1, 12, 13, 10, 11,  8,  9,         //3 -> 2
  57         };
  58         const char mask_swap4_16[]  __attribute__ ((aligned (16))) = {
  59                  4,  5,  2,  3,  0,  1,  6,  7, 12, 13, 10, 11,  8,  9, 14, 15,         //0 -> 0
  60         };
  61
  62 #ifdef ARCH_IS_I386
  63         void ssse3_drop(uint8_t* dest, const uint8_t* src, size_t units, const char* masks)
  64         {
  65                 size_t blocks = (units + 11) / 12;
  66                 for(size_t i = 0; i < blocks; i++) {
  67                         asm volatile(
  68                                 "MOVDQA 0(%1),%%xmm0\n"
  69                                 "\tMOVDQA 16(%1),%%xmm1\n"
  70                                 "\tMOVDQA 32(%1),%%xmm3\n"
  71                                 "\tMOVDQA 48(%1),%%xmm5\n"
  72                                 "\tMOVDQA %%xmm1,%%xmm2\n"
  73                                 "\tMOVDQA %%xmm3,%%xmm4\n"
  74                                 "\tPSHUFB 0(%2),%%xmm0\n"
  75                                 "\tPSHUFB 16(%2),%%xmm1\n"
  76                                 "\tPSHUFB 32(%2),%%xmm2\n"
  77                                 "\tPSHUFB 48(%2),%%xmm3\n"
  78                                 "\tPSHUFB 64(%2),%%xmm4\n"
  79                                 "\tPSHUFB 80(%2),%%xmm5\n"
  80                                 "\tPOR %%xmm0,%%xmm1\n"
  81                                 "\tPOR %%xmm2,%%xmm3\n"
  82                                 "\tPOR %%xmm4,%%xmm5\n"
  83                                 "\tMOVDQA %%xmm1,0(%0)\n"
  84                                 "\tMOVDQA %%xmm3,16(%0)\n"
  85                                 "\tMOVDQA %%xmm5,32(%0)\n"
  86                             : : "r" (dest), "r" (src), "r" (masks) : "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
  87                         dest += 48;
  88                         src += 64;
  89                 }
  90         }
  91
  92         void ssse3_swap(uint8_t* dest, const uint8_t* src, size_t units, const char* masks)
  93         {
  94                 size_t blocks = (units + 3) / 4;
  95                 for(size_t i = 0; i < blocks; i++) {
  96                         asm volatile(
  97                                 "MOVDQA (%1),%%xmm0\n"
  98                                 "\tPSHUFB (%2),%%xmm0\n"
  99                                 "\tMOVDQA %%xmm0,(%0)\n"
 100                             : : "r" (dest), "r" (src), "r" (masks) : "xmm0");
 101                         dest += 16;
 102                         src += 16;
 103                 }
 104         }
 105 #endif
 106 }
 107
 108 void copy_drop4(uint8_t* dest, const uint32_t* src, size_t units)
 109 {
 110 #ifdef ARCH_IS_I386
 111         if(ssse3_available()) {
 112                 ssse3_drop(dest, reinterpret_cast<const uint8_t*>(src), units, mask_drop4_8);
 113                 return;
 114         }
 115 #endif
 116         const uint8_t* _src = reinterpret_cast<const uint8_t*>(src);
 117         for(size_t i = 0; i < units; i++) {
 118                 dest[3 * i + 0] = _src[4 * i + 0];
 119                 dest[3 * i + 1] = _src[4 * i + 1];
 120                 dest[3 * i + 2] = _src[4 * i + 2];
 121         }
 122 }
 123
 124 void copy_drop4s(uint8_t* dest, const uint32_t* src, size_t units)
 125 {
 126 #ifdef ARCH_IS_I386
 127         if(ssse3_available()) {
 128                 ssse3_drop(dest, reinterpret_cast<const uint8_t*>(src), units, mask_drop4s_8);
 129                 return;
 130         }
 131 #endif
 132         const uint8_t* _src = reinterpret_cast<const uint8_t*>(src);
 133         for(size_t i = 0; i < units; i++) {
 134                 dest[3 * i + 0] = _src[4 * i + 2];
 135                 dest[3 * i + 1] = _src[4 * i + 1];
 136                 dest[3 * i + 2] = _src[4 * i + 0];
 137         }
 138 }
 139
 140 void copy_swap4(uint8_t* dest, const uint32_t* src, size_t units)
 141 {
 142 #ifdef ARCH_IS_I386
 143         if(ssse3_available()) {
 144                 ssse3_swap(dest, reinterpret_cast<const uint8_t*>(src), units, mask_swap4_8);
 145                 return;
 146         }
 147 #endif
 148         const uint8_t* _src = reinterpret_cast<const uint8_t*>(src);
 149         for(size_t i = 0; i < units; i++) {
 150                 dest[4 * i + 0] = _src[4 * i + 2];
 151                 dest[4 * i + 1] = _src[4 * i + 1];
 152                 dest[4 * i + 2] = _src[4 * i + 0];
 153                 dest[4 * i + 3] = _src[4 * i + 3];
 154         }
 155 }
 156
 157 void copy_drop4(uint16_t* dest, const uint64_t* src, size_t units)
 158 {
 159 #ifdef ARCH_IS_I386
 160         if(ssse3_available()) {
 161                 ssse3_drop(reinterpret_cast<uint8_t*>(dest), reinterpret_cast<const uint8_t*>(src), units,
 162                         mask_drop4_16);
 163                 return;
 164         }
 165 #endif
 166         const uint16_t* _src = reinterpret_cast<const uint16_t*>(src);
 167         for(size_t i = 0; i < units; i++) {
 168                 dest[3 * i + 0] = _src[4 * i + 0];
 169                 dest[3 * i + 1] = _src[4 * i + 1];
 170                 dest[3 * i + 2] = _src[4 * i + 2];
 171         }
 172 }
 173
 174 void copy_drop4s(uint16_t* dest, const uint64_t* src, size_t units)
 175 {
 176 #ifdef ARCH_IS_I386
 177         if(ssse3_available()) {
 178                 ssse3_drop(reinterpret_cast<uint8_t*>(dest), reinterpret_cast<const uint8_t*>(src), units,
 179                         mask_drop4s_16);
 180                 return;
 181         }
 182 #endif
 183         const uint16_t* _src = reinterpret_cast<const uint16_t*>(src);
 184         for(size_t i = 0; i < units; i++) {
 185                 dest[3 * i + 0] = _src[4 * i + 2];
 186                 dest[3 * i + 1] = _src[4 * i + 1];
 187                 dest[3 * i + 2] = _src[4 * i + 0];
 188         }
 189 }
 190
 191 void copy_swap4(uint16_t* dest, const uint64_t* src, size_t units)
 192 {
 193 #ifdef ARCH_IS_I386
 194         if(ssse3_available()) {
 195                 ssse3_swap(reinterpret_cast<uint8_t*>(dest), reinterpret_cast<const uint8_t*>(src), units,
 196                         mask_swap4_16);
 197                 return;
 198         }
 199 #endif
 200         const uint16_t* _src = reinterpret_cast<const uint16_t*>(src);
 201         for(size_t i = 0; i < units; i++) {
 202                 dest[4 * i + 0] = _src[4 * i + 2];
 203                 dest[4 * i + 1] = _src[4 * i + 1];
 204                 dest[4 * i + 2] = _src[4 * i + 0];
 205                 dest[4 * i + 3] = _src[4 * i + 3];
 206         }
 207 }
 208 }