git.sesse.net Git - casparcg/blob - common/utility/memory.cpp

   1 #include "../stdafx.h"\r
   2 \r
   3 #include "memory.h"\r
   4 \r
   5 #include <intrin.h>\r
   6 \r
   7 #include <tbb/parallel_for.h>\r
   8 #include <tbb/blocked_range.h>\r
   9 \r
  10 namespace caspar { namespace common {\r
  11         \r
  12 void* memcpy_SSE2(void* dest, const void* source, size_t num)\r
  13 {       \r
  14         __asm\r
  15         {\r
  16                 mov esi, source;    \r
  17                 mov edi, dest;   \r
  18  \r
  19                 mov ebx, num; \r
  20                 shr ebx, 7;      \r
  21   \r
  22                 cpy:\r
  23                         prefetchnta [esi+80h];\r
  24  \r
  25                         movdqa xmm0, [esi+00h];\r
  26                         movdqa xmm1, [esi+10h];\r
  27                         movdqa xmm2, [esi+20h];\r
  28                         movdqa xmm3, [esi+30h];\r
  29           \r
  30                         movntdq [edi+00h], xmm0;\r
  31                         movntdq [edi+10h], xmm1;\r
  32                         movntdq [edi+20h], xmm2;\r
  33                         movntdq [edi+30h], xmm3;\r
  34                         \r
  35                         prefetchnta [esi+0C0h];\r
  36                         \r
  37                         movdqa xmm4, [esi+40h];\r
  38                         movdqa xmm5, [esi+50h];\r
  39                         movdqa xmm6, [esi+60h];\r
  40                         movdqa xmm7, [esi+70h];\r
  41           \r
  42                         movntdq [edi+40h], xmm4;\r
  43                         movntdq [edi+50h], xmm5;\r
  44                         movntdq [edi+60h], xmm6;\r
  45                         movntdq [edi+70h], xmm7;\r
  46  \r
  47                         lea edi, [edi+80h];\r
  48                         lea esi, [esi+80h];\r
  49                         dec ebx;\r
  50  \r
  51                 jnz cpy;\r
  52         }\r
  53         return dest;\r
  54 }\r
  55 \r
  56 void* aligned_memcpy(void* dest, const void* source, size_t num)\r
  57 {       \r
  58         if(num < 128)\r
  59                 return memcpy(dest, source, num);\r
  60 \r
  61         tbb::parallel_for(tbb::blocked_range<size_t>(0, num/128), [&](const tbb::blocked_range<size_t>& r)\r
  62         {\r
  63                 memcpy_SSE2(reinterpret_cast<char*>(dest) + r.begin()*128, reinterpret_cast<const char*>(source) + r.begin()*128, r.size()*128);\r
  64         }, tbb::affinity_partitioner());\r
  65 \r
  66         return dest;\r
  67 }\r
  68 \r
  69 void* clear(void* dest, size_t size)\r
  70 {\r
  71         tbb::parallel_for(tbb::blocked_range<size_t>(0, size/16), [&](const tbb::blocked_range<size_t>& r)\r
  72         {\r
  73                 __m128i val = _mm_setzero_si128();\r
  74                 __m128i* ptr = reinterpret_cast<__m128i*>(dest)+r.begin();\r
  75                 __m128i* end = ptr + r.size();\r
  76 \r
  77                 while(ptr != end)       \r
  78                         _mm_stream_si128(ptr++, val);\r
  79         });     \r
  80         return dest;\r
  81 }\r
  82 \r
  83 }}