API tools faq
paste
Advertisement
Guest User

Untitled

a guest
Jun 22nd, 2014
226
0
Never
Add comment
Not a member of Pastebin yet? Sign Up, it unlocks many cool features!
Diff 10.34 KB | None | 0 0
raw download clone embed print report
  1. diff --git a/Makefile b/Makefile
  2. index 38d3910..c17f68e 100644
  3. --- a/Makefile
  4. +++ b/Makefile
  5. @@ -8,13 +8,13 @@ BINDIR        = bin
  6.  OBJDIR         = obj
  7.  VPATH      = src
  8.  
  9. -C_FILES        = main rgb2yuv_wiki rgb2yuv_novell_ch rgb2yuv_tables rgb2yuv_sse rgb2yuv_gstreamer
  10. +C_FILES        = main rgb2yuv_wiki rgb2yuv_novell_ch rgb2yuv_tables rgb2yuv_sse rgb2yuv_sse2 rgb2yuv_gstreamer
  11.  OBJS       = $(addprefix $(OBJDIR)/, $(addsuffix .o, $(C_FILES)))
  12.  
  13.  CFLAGS     = -fmessage-length=0 -fstack-protector -O2 -Wall -D_FORTIFY_SOURCE=2 -funwind-tables -fasynchronous-unwind-tables -g \
  14.               -Wall -Wextra -Wno-unused-parameter -Wformat-y2k -Winit-self -msse4.1 \
  15. -             -Wstrict-prototypes -Winline -Wnested-externs -Wbad-function-cast -Wshadow
  16. -LDFLAGS        = -lrt
  17. +             -Wstrict-prototypes -Winline -Wnested-externs -Wbad-function-cast -Wshadow -std=gnu11
  18. +LDFLAGS        = -lrt -lgomp
  19.  
  20.  #
  21.  # Targets
  22. diff --git a/src/main.c b/src/main.c
  23. index 4689649..7368fab 100644
  24. --- a/src/main.c
  25. +++ b/src/main.c
  26. @@ -3,69 +3,56 @@
  27.  #include "rgb2yuv.h"
  28.  #include <stdlib.h>
  29.  #include <stdio.h>
  30. +#include <omp.h>
  31.  
  32. -#define PIXELS_COUNT   (100 * 1000 * 1000)
  33. +#define PIXELS_COUNT  1080
  34.  
  35.  static uint8_t pixels_wiki[4 * PIXELS_COUNT];
  36.  static uint8_t pixels_novel_ch[4 * PIXELS_COUNT];
  37.  static uint8_t pixels_tables[4 * PIXELS_COUNT];
  38.  static uint8_t pixels_gstreamer[4 * PIXELS_COUNT];
  39.  static uint8_t pixels_see[4 * PIXELS_COUNT];
  40. -
  41. -int main(int argc, char *argv[])
  42. -{
  43. -   struct timespec now;
  44. -   clock_t start, finish;
  45. -
  46. -   clock_gettime(CLOCK_REALTIME, &now);
  47. -   srand(now.tv_nsec);
  48. -
  49. -   // prepare input buffers
  50. -
  51. -   int i;
  52. -   for (i = 0; i < 4 * PIXELS_COUNT; i++)
  53. -   {
  54. -       uint8_t x = ((i % 4) * rand());
  55. -       pixels_wiki[i] = x;
  56. -       pixels_novel_ch[i] = x;
  57. -       pixels_tables[i] = x;
  58. -                pixels_gstreamer[i] = x;
  59. -                pixels_see[i] = x;
  60. -   }
  61. -
  62. -   // run conversations
  63. -
  64. -   start = clock();
  65. -   rgb2yuv_wiki(pixels_wiki, PIXELS_COUNT);
  66. -   finish = clock();
  67. -   printf("rgb2yuv_wiki: %.3f sec\n", (float) (finish - start) / CLOCKS_PER_SEC);
  68. -
  69. -   start = clock();
  70. -   rgb2yuv_novell_ch(pixels_novel_ch, PIXELS_COUNT);
  71. -   finish = clock();
  72. -   printf("rgb2yuv_novell_ch: %.3f sec\n", (float) (finish - start) / CLOCKS_PER_SEC);
  73. -
  74. -   start = clock();
  75. -   rgb2yuv_tables_init();
  76. -   rgb2yuv_tables(pixels_tables, PIXELS_COUNT);
  77. -   finish = clock();
  78. -   printf("rgb2yuv_tables: %.3f sec\n", (float) (finish - start) / CLOCKS_PER_SEC);
  79. -        
  80. -        start = clock();
  81. -        rgb2yuv_gstreamer(pixels_gstreamer, PIXELS_COUNT);
  82. -        finish = clock();
  83. -        printf("rgb2yuv_gstreamer: %.3f sec\n", (float) (finish - start) / CLOCKS_PER_SEC);
  84. -        
  85. -        start = clock();
  86. -        rgb2yuv_sse(pixels_see, PIXELS_COUNT);
  87. -        finish = clock();
  88. -        printf("rgb2yuv_sse: %.3f sec\n", (float) (finish - start) / CLOCKS_PER_SEC);
  89. -
  90. -
  91. -   for (i = 0; i < 100; i++)
  92. -   {
  93. -       printf("%d: %d %d %d %d %d\n", i, pixels_wiki[i], pixels_novel_ch[i], pixels_tables[i], pixels_gstreamer[i], pixels_see[i]);
  94. -   }
  95. -
  96. -   return 0;
  97. +static uint8_t pixels_see2[4 * PIXELS_COUNT];
  98. +
  99. +enum {KB = 1024, MB = KB * 1024, GB = MB * 1024};
  100. +#define Mbench(name, calls, data_len, call) ({\
  101. +  uint64_t i = calls;\
  102. +  double start = omp_get_wtime();\
  103. +  do {\
  104. +    call;\
  105. +  } while(--i);\
  106. +  double time = omp_get_wtime() - start;\
  107. +  fprintf(stderr, "%s: (%lu)%luKB: total %luGB: %fGB/s\n", name, (uint64_t)calls, (uint64_t)data_len/KB, (data_len * calls)/GB, ((double)(data_len * calls)/time)/GB);\
  108. +})
  109. +
  110. +int main(int argc, char * argv[]) {
  111. +  struct timespec now;
  112. +  clock_gettime(CLOCK_REALTIME, &now);
  113. +  srand(now.tv_nsec);
  114. +  // prepare input buffers
  115. +
  116. +  for(uint32_t i = 0; i < 4 * PIXELS_COUNT; i++) {
  117. +    uint8_t x = ((i % 4) * rand());
  118. +    pixels_wiki[i] = x;
  119. +    pixels_novel_ch[i] = x;
  120. +    pixels_tables[i] = x;
  121. +    pixels_gstreamer[i] = x;
  122. +    pixels_see[i] = x;
  123. +    pixels_see2[i] = x;
  124. +  }
  125. +
  126. +  // run conversations
  127. +  Mbench("rgb2yuv_wiki", (1024 * 1024), sizeof(pixels_wiki), rgb2yuv_wiki(pixels_wiki, PIXELS_COUNT));
  128. +  Mbench("rgb2yuv_novell_ch", (1024 * 1024), sizeof(pixels_novel_ch), rgb2yuv_novell_ch(pixels_novel_ch, PIXELS_COUNT));
  129. +  Mbench("rgb2yuv_tables", (1024 * 1024), sizeof(pixels_tables), rgb2yuv_tables(pixels_tables, PIXELS_COUNT));
  130. +  Mbench("rgb2yuv_gstreamer", (1024 * 1024), sizeof(pixels_gstreamer), rgb2yuv_gstreamer(pixels_gstreamer, PIXELS_COUNT));
  131. +  Mbench("rgb2yuv_sse", (1024 * 1024), sizeof(pixels_see), rgb2yuv_sse(pixels_see, PIXELS_COUNT));
  132. +  Mbench("rgb2yuv_sse2", (1024 * 1024), sizeof(pixels_see2), rgb2yuv_sse2(pixels_see2, PIXELS_COUNT));
  133. +
  134. +
  135. +//   for(uint32_t i = 0; i < 100; i++) {
  136. +//     printf("%d: %d %d %d %d %d\n", i, pixels_wiki[i], pixels_novel_ch[i], pixels_tables[i], pixels_gstreamer[i], pixels_see[i]);
  137. +//   }
  138. +
  139. +  return 0;
  140.  }
  141. diff --git a/src/rgb2yuv_sse2.c b/src/rgb2yuv_sse2.c
  142. new file mode 100644
  143. index 0000000..4dc53a8
  144. --- /dev/null
  145. +++ b/src/rgb2yuv_sse2.c
  146. @@ -0,0 +1,130 @@
  147. +/**
  148. + * @file rgb2yuv_sse.c
  149. + *
  150. + * RGB2YUV sse
  151. + */
  152. +
  153. +#include "rgb2yuv.h"
  154. +#include <xmmintrin.h>
  155. +#include <emmintrin.h>
  156. +#include <tmmintrin.h>
  157. +
  158. +typedef struct {
  159. +  __m128i r, g, b;
  160. +} r8g8b8_t;
  161. +
  162. +typedef struct {
  163. +  __m128i y, u, v;
  164. +} y8u8v8_t;
  165. +
  166. +typedef struct {
  167. +  __m128i r_coef, g_coef, b_coef, x_coef;
  168. +} rgb_to_yuv_mat_v8_t;
  169. +
  170. +static inline rgb_to_yuv_mat_v8_t create_vecm(int32_t r_coef, int32_t g_coef, int32_t b_coef, int32_t x_coef) {
  171. +  return (rgb_to_yuv_mat_v8_t) {
  172. +    _mm_set1_epi16(r_coef),
  173. +    _mm_set1_epi16(g_coef),
  174. +    _mm_set1_epi16(b_coef),
  175. +    _mm_set1_epi16(x_coef)
  176. +  };
  177. +}
  178. +
  179. +static inline __m128i op(r8g8b8_t v, rgb_to_yuv_mat_v8_t m) {
  180. +  __m128i a = _mm_mullo_epi16(v.r, m.r_coef);
  181. +  __m128i b = _mm_mullo_epi16(v.g, m.g_coef);
  182. +  __m128i c = _mm_mullo_epi16(v.b, m.b_coef);
  183. +  __m128i ret = _mm_add_epi16(_mm_add_epi16(a, b), _mm_add_epi16(c, m.x_coef));
  184. +  return _mm_srai_epi16(ret, 8);
  185. +}
  186. +
  187. +static inline __m128i r8g8b8_to_y8(r8g8b8_t v) {
  188. +  return op(v, create_vecm(47, 157, 16, 4096));
  189. +}
  190. +static inline __m128i r8g8b8_to_u8(r8g8b8_t v) {
  191. +  return op(v, create_vecm(-26, -87, 112, 32768));
  192. +}
  193. +static inline __m128i r8g8b8_to_v8(r8g8b8_t v) {
  194. +  return op(v, create_vecm(112, -102, -10, 32768));
  195. +}
  196. +
  197. +static inline y8u8v8_t r8g8b8_to_y8u8v8(r8g8b8_t v) {
  198. +  return (y8u8v8_t) {r8g8b8_to_y8(v), r8g8b8_to_u8(v), r8g8b8_to_v8(v)};
  199. +}
  200. +
  201. +// static inline r8g8b8_t p16x2_to_r8g8b8(__m128i a, __m128i b) {
  202. +//   __m128i v_byte1 = _mm_set1_epi32(0x000000ff);
  203. +//   __m128i v_byte3 = _mm_set1_epi32(0x00ff0000);
  204. +//   __m128i r8 = _mm_or_si128(_mm_and_si128(_mm_srli_si128(a, 1), v_byte1), _mm_and_si128(_mm_slli_si128(b, 1), v_byte3));
  205. +//   __m128i g8 = _mm_or_si128(_mm_and_si128(_mm_srli_si128(a, 2), v_byte1), _mm_and_si128(b, v_byte3));
  206. +//   __m128i b8 = _mm_or_si128(_mm_and_si128(_mm_srli_si128(a, 3), v_byte1), _mm_and_si128(_mm_srli_si128(b, 1), v_byte3));
  207. +//   return (r8g8b8_t) {r8, g8, b8};
  208. +// }
  209. +
  210. +static inline r8g8b8_t p16x2_to_r8g8b8(__m128i a, __m128i b) {
  211. +  __m128i shuff_ra  = _mm_setr_epi8(1, 255, 255, 255, 5, 255, 255, 255, 9, 255, 255, 255, 13, 255, 255, 255);
  212. +  __m128i shuff_ga  = _mm_setr_epi8(2, 255, 255, 255, 6, 255, 255, 255, 10, 255, 255, 255, 14, 255, 255, 255);
  213. +  __m128i shuff_ba  = _mm_setr_epi8(3, 255, 255, 255, 7, 255, 255, 255, 11, 255, 255, 255, 15, 255, 255, 255);
  214. +  __m128i shuff_rb  = _mm_setr_epi8(255, 255, 1, 255, 255, 255, 5, 255, 255, 255, 9, 255, 255, 255, 13, 255);
  215. +  __m128i shuff_gb  = _mm_setr_epi8(255, 255, 2, 255, 255, 255, 6, 255, 255, 255, 10, 255, 255, 255, 14, 255);
  216. +  __m128i shuff_bb  = _mm_setr_epi8(255, 255, 3, 255, 255, 255, 7, 255, 255, 255, 11, 255, 255, 255, 15, 255);
  217. +  __m128i r8 = _mm_or_si128(_mm_shuffle_epi8(a, shuff_ra), _mm_shuffle_epi8(b, shuff_rb));
  218. +  __m128i g8 = _mm_or_si128(_mm_shuffle_epi8(a, shuff_ga), _mm_shuffle_epi8(b, shuff_gb));
  219. +  __m128i b8 = _mm_or_si128(_mm_shuffle_epi8(a, shuff_ba), _mm_shuffle_epi8(b, shuff_bb));
  220. +  return (r8g8b8_t) {r8, g8, b8};
  221. +}
  222. +
  223. +static inline __m128i y8u8v8_to_p16a(y8u8v8_t v) {
  224. +  __m128i shuff_ya  = _mm_setr_epi8(255, 0, 255, 255, 255, 4, 255, 255, 255, 8, 255, 255, 255, 12, 255, 255);
  225. +  __m128i shuff_ua  = _mm_setr_epi8(255, 255, 0, 255, 255, 255, 4, 255, 255, 255, 8, 255, 255, 255, 12, 255);
  226. +  __m128i shuff_va  = _mm_setr_epi8(255, 255, 255, 0, 255, 255, 255, 4, 255, 255, 255, 8, 255, 255, 255, 12);
  227. +  return _mm_or_si128(_mm_or_si128(_mm_shuffle_epi8(v.y, shuff_ya), _mm_shuffle_epi8(v.u, shuff_ua)), _mm_shuffle_epi8(v.v, shuff_va));
  228. +}
  229. +
  230. +static inline __m128i y8u8v8_to_p16b(y8u8v8_t v) {
  231. +  __m128i shuff_yb  = _mm_setr_epi8(255, 2, 255, 255, 255, 6, 255, 255, 255, 10, 255, 255, 255, 14, 255, 255);
  232. +  __m128i shuff_ub  = _mm_setr_epi8(255, 255, 2, 255, 255, 255, 6, 255, 255, 255, 10, 255, 255, 255, 14, 255);
  233. +  __m128i shuff_vb  = _mm_setr_epi8(255, 255, 255, 2, 255, 255, 255, 6, 255, 255, 255, 10, 255, 255, 255, 14);
  234. +  return _mm_or_si128(_mm_or_si128(_mm_shuffle_epi8(v.y, shuff_yb), _mm_shuffle_epi8(v.u, shuff_ub)), _mm_shuffle_epi8(v.v, shuff_vb));
  235. +}
  236. +
  237. +void rgb2yuv_sse2(void * pixels, uint64_t len) {
  238. +  __m128i * it = pixels, * end = (__m128i *)((int32_t *)pixels + len);
  239. +  do {
  240. +    {
  241. +    __m128i a = *(it + 0), b = *(it + 1);
  242. +    y8u8v8_t v = r8g8b8_to_y8u8v8(p16x2_to_r8g8b8(a, b));
  243. +    a = y8u8v8_to_p16a(v), b = y8u8v8_to_p16b(v);
  244. +    *(it + 0) = a, *(it + 1) = b;
  245. +    it += 2;
  246. +    }
  247. +//     {
  248. +//     __m128i a = *(it + 0), b = *(it + 1);
  249. +//     y8u8v8_t v = r8g8b8_to_y8u8v8(p16x2_to_r8g8b8(a, b));
  250. +//     a = y8u8v8_to_p16a(v), b = y8u8v8_to_p16b(v);
  251. +//     *(it + 0) = a, *(it + 1) = b;
  252. +//     it += 2;
  253. +//     }
  254. +//     {
  255. +//     __m128i a = *(it + 0), b = *(it + 1);
  256. +//     y8u8v8_t v = r8g8b8_to_y8u8v8(p16x2_to_r8g8b8(a, b));
  257. +//     a = y8u8v8_to_p16a(v), b = y8u8v8_to_p16b(v);
  258. +//     *(it + 0) = a, *(it + 1) = b;
  259. +//     it += 2;
  260. +//     }
  261. +//     {
  262. +//     __m128i a = *(it + 0), b = *(it + 1);
  263. +//     y8u8v8_t v = r8g8b8_to_y8u8v8(p16x2_to_r8g8b8(a, b));
  264. +//     a = y8u8v8_to_p16a(v), b = y8u8v8_to_p16b(v);
  265. +//     *(it + 0) = a, *(it + 1) = b;
  266. +//     it += 2;
  267. +//     }
  268. +//     {
  269. +//     __m128i a = *(it + 0), b = *(it + 1);
  270. +//     y8u8v8_t v = r8g8b8_to_y8u8v8(p16x2_to_r8g8b8(a, b));
  271. +//     a = y8u8v8_to_p16a(v), b = y8u8v8_to_p16b(v);
  272. +//     *(it + 0) = a, *(it + 1) = b;
  273. +//     it += 2;
  274. +//     }
  275. +  } while(it < end);
  276. +}
  277. \ No newline at end of file
Advertisement
Add Comment
Please, Sign In to add comment
Advertisement
Public Pastes
Advertisement
We use cookies for various purposes including analytics. By continuing to use Pastebin, you agree to our use of cookies as described in the Cookies Policy.  OK, I Understand
Not a member of Pastebin yet?
Sign Up, it unlocks many cool features!