cumo 0.1.0

Sign up to get free protection for your applications and to get access to all the features.
Files changed (266) hide show
  1. checksums.yaml +7 -0
  2. data/.gitignore +27 -0
  3. data/.travis.yml +5 -0
  4. data/3rd_party/mkmf-cu/.gitignore +36 -0
  5. data/3rd_party/mkmf-cu/Gemfile +3 -0
  6. data/3rd_party/mkmf-cu/LICENSE +21 -0
  7. data/3rd_party/mkmf-cu/README.md +36 -0
  8. data/3rd_party/mkmf-cu/Rakefile +11 -0
  9. data/3rd_party/mkmf-cu/bin/mkmf-cu-nvcc +4 -0
  10. data/3rd_party/mkmf-cu/lib/mkmf-cu.rb +32 -0
  11. data/3rd_party/mkmf-cu/lib/mkmf-cu/cli.rb +80 -0
  12. data/3rd_party/mkmf-cu/lib/mkmf-cu/nvcc.rb +157 -0
  13. data/3rd_party/mkmf-cu/mkmf-cu.gemspec +16 -0
  14. data/3rd_party/mkmf-cu/test/test_mkmf-cu.rb +67 -0
  15. data/CODE_OF_CONDUCT.md +46 -0
  16. data/Gemfile +8 -0
  17. data/LICENSE.txt +82 -0
  18. data/README.md +252 -0
  19. data/Rakefile +43 -0
  20. data/bench/broadcast_fp32.rb +138 -0
  21. data/bench/cumo_bench.rb +193 -0
  22. data/bench/numo_bench.rb +138 -0
  23. data/bench/reduction_fp32.rb +117 -0
  24. data/bin/console +14 -0
  25. data/bin/setup +8 -0
  26. data/cumo.gemspec +32 -0
  27. data/ext/cumo/cuda/cublas.c +278 -0
  28. data/ext/cumo/cuda/driver.c +421 -0
  29. data/ext/cumo/cuda/memory_pool.cpp +185 -0
  30. data/ext/cumo/cuda/memory_pool_impl.cpp +308 -0
  31. data/ext/cumo/cuda/memory_pool_impl.hpp +370 -0
  32. data/ext/cumo/cuda/memory_pool_impl_test.cpp +554 -0
  33. data/ext/cumo/cuda/nvrtc.c +207 -0
  34. data/ext/cumo/cuda/runtime.c +167 -0
  35. data/ext/cumo/cumo.c +148 -0
  36. data/ext/cumo/depend.erb +58 -0
  37. data/ext/cumo/extconf.rb +179 -0
  38. data/ext/cumo/include/cumo.h +25 -0
  39. data/ext/cumo/include/cumo/compat.h +23 -0
  40. data/ext/cumo/include/cumo/cuda/cublas.h +153 -0
  41. data/ext/cumo/include/cumo/cuda/cumo_thrust.hpp +187 -0
  42. data/ext/cumo/include/cumo/cuda/cumo_thrust_complex.hpp +79 -0
  43. data/ext/cumo/include/cumo/cuda/driver.h +22 -0
  44. data/ext/cumo/include/cumo/cuda/memory_pool.h +28 -0
  45. data/ext/cumo/include/cumo/cuda/nvrtc.h +22 -0
  46. data/ext/cumo/include/cumo/cuda/runtime.h +40 -0
  47. data/ext/cumo/include/cumo/indexer.h +238 -0
  48. data/ext/cumo/include/cumo/intern.h +142 -0
  49. data/ext/cumo/include/cumo/intern_fwd.h +38 -0
  50. data/ext/cumo/include/cumo/intern_kernel.h +6 -0
  51. data/ext/cumo/include/cumo/narray.h +429 -0
  52. data/ext/cumo/include/cumo/narray_kernel.h +149 -0
  53. data/ext/cumo/include/cumo/ndloop.h +95 -0
  54. data/ext/cumo/include/cumo/reduce_kernel.h +126 -0
  55. data/ext/cumo/include/cumo/template.h +158 -0
  56. data/ext/cumo/include/cumo/template_kernel.h +77 -0
  57. data/ext/cumo/include/cumo/types/bit.h +40 -0
  58. data/ext/cumo/include/cumo/types/bit_kernel.h +34 -0
  59. data/ext/cumo/include/cumo/types/complex.h +402 -0
  60. data/ext/cumo/include/cumo/types/complex_kernel.h +414 -0
  61. data/ext/cumo/include/cumo/types/complex_macro.h +382 -0
  62. data/ext/cumo/include/cumo/types/complex_macro_kernel.h +186 -0
  63. data/ext/cumo/include/cumo/types/dcomplex.h +46 -0
  64. data/ext/cumo/include/cumo/types/dcomplex_kernel.h +13 -0
  65. data/ext/cumo/include/cumo/types/dfloat.h +47 -0
  66. data/ext/cumo/include/cumo/types/dfloat_kernel.h +14 -0
  67. data/ext/cumo/include/cumo/types/float_def.h +34 -0
  68. data/ext/cumo/include/cumo/types/float_def_kernel.h +39 -0
  69. data/ext/cumo/include/cumo/types/float_macro.h +191 -0
  70. data/ext/cumo/include/cumo/types/float_macro_kernel.h +158 -0
  71. data/ext/cumo/include/cumo/types/int16.h +24 -0
  72. data/ext/cumo/include/cumo/types/int16_kernel.h +23 -0
  73. data/ext/cumo/include/cumo/types/int32.h +24 -0
  74. data/ext/cumo/include/cumo/types/int32_kernel.h +19 -0
  75. data/ext/cumo/include/cumo/types/int64.h +24 -0
  76. data/ext/cumo/include/cumo/types/int64_kernel.h +19 -0
  77. data/ext/cumo/include/cumo/types/int8.h +24 -0
  78. data/ext/cumo/include/cumo/types/int8_kernel.h +19 -0
  79. data/ext/cumo/include/cumo/types/int_macro.h +67 -0
  80. data/ext/cumo/include/cumo/types/int_macro_kernel.h +48 -0
  81. data/ext/cumo/include/cumo/types/real_accum.h +486 -0
  82. data/ext/cumo/include/cumo/types/real_accum_kernel.h +101 -0
  83. data/ext/cumo/include/cumo/types/robj_macro.h +80 -0
  84. data/ext/cumo/include/cumo/types/robj_macro_kernel.h +0 -0
  85. data/ext/cumo/include/cumo/types/robject.h +27 -0
  86. data/ext/cumo/include/cumo/types/robject_kernel.h +7 -0
  87. data/ext/cumo/include/cumo/types/scomplex.h +46 -0
  88. data/ext/cumo/include/cumo/types/scomplex_kernel.h +13 -0
  89. data/ext/cumo/include/cumo/types/sfloat.h +48 -0
  90. data/ext/cumo/include/cumo/types/sfloat_kernel.h +14 -0
  91. data/ext/cumo/include/cumo/types/uint16.h +25 -0
  92. data/ext/cumo/include/cumo/types/uint16_kernel.h +20 -0
  93. data/ext/cumo/include/cumo/types/uint32.h +25 -0
  94. data/ext/cumo/include/cumo/types/uint32_kernel.h +20 -0
  95. data/ext/cumo/include/cumo/types/uint64.h +25 -0
  96. data/ext/cumo/include/cumo/types/uint64_kernel.h +20 -0
  97. data/ext/cumo/include/cumo/types/uint8.h +25 -0
  98. data/ext/cumo/include/cumo/types/uint8_kernel.h +20 -0
  99. data/ext/cumo/include/cumo/types/uint_macro.h +58 -0
  100. data/ext/cumo/include/cumo/types/uint_macro_kernel.h +38 -0
  101. data/ext/cumo/include/cumo/types/xint_macro.h +169 -0
  102. data/ext/cumo/include/cumo/types/xint_macro_kernel.h +88 -0
  103. data/ext/cumo/narray/SFMT-params.h +97 -0
  104. data/ext/cumo/narray/SFMT-params19937.h +46 -0
  105. data/ext/cumo/narray/SFMT.c +620 -0
  106. data/ext/cumo/narray/SFMT.h +167 -0
  107. data/ext/cumo/narray/array.c +638 -0
  108. data/ext/cumo/narray/data.c +961 -0
  109. data/ext/cumo/narray/gen/cogen.rb +56 -0
  110. data/ext/cumo/narray/gen/cogen_kernel.rb +58 -0
  111. data/ext/cumo/narray/gen/def/bit.rb +37 -0
  112. data/ext/cumo/narray/gen/def/dcomplex.rb +39 -0
  113. data/ext/cumo/narray/gen/def/dfloat.rb +37 -0
  114. data/ext/cumo/narray/gen/def/int16.rb +36 -0
  115. data/ext/cumo/narray/gen/def/int32.rb +36 -0
  116. data/ext/cumo/narray/gen/def/int64.rb +36 -0
  117. data/ext/cumo/narray/gen/def/int8.rb +36 -0
  118. data/ext/cumo/narray/gen/def/robject.rb +37 -0
  119. data/ext/cumo/narray/gen/def/scomplex.rb +39 -0
  120. data/ext/cumo/narray/gen/def/sfloat.rb +37 -0
  121. data/ext/cumo/narray/gen/def/uint16.rb +36 -0
  122. data/ext/cumo/narray/gen/def/uint32.rb +36 -0
  123. data/ext/cumo/narray/gen/def/uint64.rb +36 -0
  124. data/ext/cumo/narray/gen/def/uint8.rb +36 -0
  125. data/ext/cumo/narray/gen/erbpp2.rb +346 -0
  126. data/ext/cumo/narray/gen/narray_def.rb +268 -0
  127. data/ext/cumo/narray/gen/spec.rb +425 -0
  128. data/ext/cumo/narray/gen/tmpl/accum.c +86 -0
  129. data/ext/cumo/narray/gen/tmpl/accum_binary.c +121 -0
  130. data/ext/cumo/narray/gen/tmpl/accum_binary_kernel.cu +61 -0
  131. data/ext/cumo/narray/gen/tmpl/accum_index.c +119 -0
  132. data/ext/cumo/narray/gen/tmpl/accum_index_kernel.cu +66 -0
  133. data/ext/cumo/narray/gen/tmpl/accum_kernel.cu +12 -0
  134. data/ext/cumo/narray/gen/tmpl/alloc_func.c +107 -0
  135. data/ext/cumo/narray/gen/tmpl/allocate.c +37 -0
  136. data/ext/cumo/narray/gen/tmpl/aref.c +66 -0
  137. data/ext/cumo/narray/gen/tmpl/aref_cpu.c +50 -0
  138. data/ext/cumo/narray/gen/tmpl/aset.c +56 -0
  139. data/ext/cumo/narray/gen/tmpl/binary.c +162 -0
  140. data/ext/cumo/narray/gen/tmpl/binary2.c +70 -0
  141. data/ext/cumo/narray/gen/tmpl/binary2_kernel.cu +15 -0
  142. data/ext/cumo/narray/gen/tmpl/binary_kernel.cu +31 -0
  143. data/ext/cumo/narray/gen/tmpl/binary_s.c +45 -0
  144. data/ext/cumo/narray/gen/tmpl/binary_s_kernel.cu +15 -0
  145. data/ext/cumo/narray/gen/tmpl/bincount.c +181 -0
  146. data/ext/cumo/narray/gen/tmpl/cast.c +44 -0
  147. data/ext/cumo/narray/gen/tmpl/cast_array.c +13 -0
  148. data/ext/cumo/narray/gen/tmpl/class.c +9 -0
  149. data/ext/cumo/narray/gen/tmpl/class_kernel.cu +6 -0
  150. data/ext/cumo/narray/gen/tmpl/clip.c +121 -0
  151. data/ext/cumo/narray/gen/tmpl/coerce_cast.c +10 -0
  152. data/ext/cumo/narray/gen/tmpl/complex_accum_kernel.cu +129 -0
  153. data/ext/cumo/narray/gen/tmpl/cond_binary.c +68 -0
  154. data/ext/cumo/narray/gen/tmpl/cond_binary_kernel.cu +18 -0
  155. data/ext/cumo/narray/gen/tmpl/cond_unary.c +46 -0
  156. data/ext/cumo/narray/gen/tmpl/cum.c +50 -0
  157. data/ext/cumo/narray/gen/tmpl/each.c +47 -0
  158. data/ext/cumo/narray/gen/tmpl/each_with_index.c +70 -0
  159. data/ext/cumo/narray/gen/tmpl/ewcomp.c +79 -0
  160. data/ext/cumo/narray/gen/tmpl/ewcomp_kernel.cu +19 -0
  161. data/ext/cumo/narray/gen/tmpl/extract.c +22 -0
  162. data/ext/cumo/narray/gen/tmpl/extract_cpu.c +26 -0
  163. data/ext/cumo/narray/gen/tmpl/extract_data.c +53 -0
  164. data/ext/cumo/narray/gen/tmpl/eye.c +105 -0
  165. data/ext/cumo/narray/gen/tmpl/eye_kernel.cu +19 -0
  166. data/ext/cumo/narray/gen/tmpl/fill.c +52 -0
  167. data/ext/cumo/narray/gen/tmpl/fill_kernel.cu +29 -0
  168. data/ext/cumo/narray/gen/tmpl/float_accum_kernel.cu +106 -0
  169. data/ext/cumo/narray/gen/tmpl/format.c +62 -0
  170. data/ext/cumo/narray/gen/tmpl/format_to_a.c +49 -0
  171. data/ext/cumo/narray/gen/tmpl/frexp.c +38 -0
  172. data/ext/cumo/narray/gen/tmpl/gemm.c +203 -0
  173. data/ext/cumo/narray/gen/tmpl/init_class.c +20 -0
  174. data/ext/cumo/narray/gen/tmpl/init_module.c +12 -0
  175. data/ext/cumo/narray/gen/tmpl/inspect.c +21 -0
  176. data/ext/cumo/narray/gen/tmpl/lib.c +50 -0
  177. data/ext/cumo/narray/gen/tmpl/lib_kernel.cu +24 -0
  178. data/ext/cumo/narray/gen/tmpl/logseq.c +102 -0
  179. data/ext/cumo/narray/gen/tmpl/logseq_kernel.cu +31 -0
  180. data/ext/cumo/narray/gen/tmpl/map_with_index.c +98 -0
  181. data/ext/cumo/narray/gen/tmpl/median.c +66 -0
  182. data/ext/cumo/narray/gen/tmpl/minmax.c +47 -0
  183. data/ext/cumo/narray/gen/tmpl/module.c +9 -0
  184. data/ext/cumo/narray/gen/tmpl/module_kernel.cu +1 -0
  185. data/ext/cumo/narray/gen/tmpl/new_dim0.c +15 -0
  186. data/ext/cumo/narray/gen/tmpl/new_dim0_kernel.cu +8 -0
  187. data/ext/cumo/narray/gen/tmpl/poly.c +50 -0
  188. data/ext/cumo/narray/gen/tmpl/pow.c +97 -0
  189. data/ext/cumo/narray/gen/tmpl/pow_kernel.cu +29 -0
  190. data/ext/cumo/narray/gen/tmpl/powint.c +17 -0
  191. data/ext/cumo/narray/gen/tmpl/qsort.c +212 -0
  192. data/ext/cumo/narray/gen/tmpl/rand.c +168 -0
  193. data/ext/cumo/narray/gen/tmpl/rand_norm.c +121 -0
  194. data/ext/cumo/narray/gen/tmpl/real_accum_kernel.cu +75 -0
  195. data/ext/cumo/narray/gen/tmpl/seq.c +112 -0
  196. data/ext/cumo/narray/gen/tmpl/seq_kernel.cu +43 -0
  197. data/ext/cumo/narray/gen/tmpl/set2.c +57 -0
  198. data/ext/cumo/narray/gen/tmpl/sort.c +48 -0
  199. data/ext/cumo/narray/gen/tmpl/sort_index.c +111 -0
  200. data/ext/cumo/narray/gen/tmpl/store.c +41 -0
  201. data/ext/cumo/narray/gen/tmpl/store_array.c +187 -0
  202. data/ext/cumo/narray/gen/tmpl/store_array_kernel.cu +58 -0
  203. data/ext/cumo/narray/gen/tmpl/store_bit.c +86 -0
  204. data/ext/cumo/narray/gen/tmpl/store_bit_kernel.cu +66 -0
  205. data/ext/cumo/narray/gen/tmpl/store_from.c +81 -0
  206. data/ext/cumo/narray/gen/tmpl/store_from_kernel.cu +58 -0
  207. data/ext/cumo/narray/gen/tmpl/store_kernel.cu +3 -0
  208. data/ext/cumo/narray/gen/tmpl/store_numeric.c +9 -0
  209. data/ext/cumo/narray/gen/tmpl/to_a.c +43 -0
  210. data/ext/cumo/narray/gen/tmpl/unary.c +132 -0
  211. data/ext/cumo/narray/gen/tmpl/unary2.c +60 -0
  212. data/ext/cumo/narray/gen/tmpl/unary_kernel.cu +72 -0
  213. data/ext/cumo/narray/gen/tmpl/unary_ret2.c +34 -0
  214. data/ext/cumo/narray/gen/tmpl/unary_s.c +86 -0
  215. data/ext/cumo/narray/gen/tmpl/unary_s_kernel.cu +58 -0
  216. data/ext/cumo/narray/gen/tmpl_bit/allocate.c +24 -0
  217. data/ext/cumo/narray/gen/tmpl_bit/aref.c +54 -0
  218. data/ext/cumo/narray/gen/tmpl_bit/aref_cpu.c +57 -0
  219. data/ext/cumo/narray/gen/tmpl_bit/aset.c +56 -0
  220. data/ext/cumo/narray/gen/tmpl_bit/binary.c +98 -0
  221. data/ext/cumo/narray/gen/tmpl_bit/bit_count.c +64 -0
  222. data/ext/cumo/narray/gen/tmpl_bit/bit_count_cpu.c +88 -0
  223. data/ext/cumo/narray/gen/tmpl_bit/bit_count_kernel.cu +76 -0
  224. data/ext/cumo/narray/gen/tmpl_bit/bit_reduce.c +133 -0
  225. data/ext/cumo/narray/gen/tmpl_bit/each.c +48 -0
  226. data/ext/cumo/narray/gen/tmpl_bit/each_with_index.c +70 -0
  227. data/ext/cumo/narray/gen/tmpl_bit/extract.c +30 -0
  228. data/ext/cumo/narray/gen/tmpl_bit/extract_cpu.c +29 -0
  229. data/ext/cumo/narray/gen/tmpl_bit/fill.c +69 -0
  230. data/ext/cumo/narray/gen/tmpl_bit/format.c +64 -0
  231. data/ext/cumo/narray/gen/tmpl_bit/format_to_a.c +51 -0
  232. data/ext/cumo/narray/gen/tmpl_bit/inspect.c +21 -0
  233. data/ext/cumo/narray/gen/tmpl_bit/mask.c +136 -0
  234. data/ext/cumo/narray/gen/tmpl_bit/none_p.c +14 -0
  235. data/ext/cumo/narray/gen/tmpl_bit/store_array.c +108 -0
  236. data/ext/cumo/narray/gen/tmpl_bit/store_bit.c +70 -0
  237. data/ext/cumo/narray/gen/tmpl_bit/store_from.c +60 -0
  238. data/ext/cumo/narray/gen/tmpl_bit/to_a.c +47 -0
  239. data/ext/cumo/narray/gen/tmpl_bit/unary.c +81 -0
  240. data/ext/cumo/narray/gen/tmpl_bit/where.c +90 -0
  241. data/ext/cumo/narray/gen/tmpl_bit/where2.c +95 -0
  242. data/ext/cumo/narray/index.c +880 -0
  243. data/ext/cumo/narray/kwargs.c +153 -0
  244. data/ext/cumo/narray/math.c +142 -0
  245. data/ext/cumo/narray/narray.c +1948 -0
  246. data/ext/cumo/narray/ndloop.c +2105 -0
  247. data/ext/cumo/narray/rand.c +45 -0
  248. data/ext/cumo/narray/step.c +474 -0
  249. data/ext/cumo/narray/struct.c +886 -0
  250. data/lib/cumo.rb +3 -0
  251. data/lib/cumo/cuda.rb +11 -0
  252. data/lib/cumo/cuda/compile_error.rb +36 -0
  253. data/lib/cumo/cuda/compiler.rb +161 -0
  254. data/lib/cumo/cuda/device.rb +47 -0
  255. data/lib/cumo/cuda/link_state.rb +31 -0
  256. data/lib/cumo/cuda/module.rb +40 -0
  257. data/lib/cumo/cuda/nvrtc_program.rb +27 -0
  258. data/lib/cumo/linalg.rb +12 -0
  259. data/lib/cumo/narray.rb +2 -0
  260. data/lib/cumo/narray/extra.rb +1278 -0
  261. data/lib/erbpp.rb +294 -0
  262. data/lib/erbpp/line_number.rb +137 -0
  263. data/lib/erbpp/narray_def.rb +381 -0
  264. data/numo-narray-version +1 -0
  265. data/run.gdb +7 -0
  266. metadata +353 -0
@@ -0,0 +1,486 @@
1
+ #define not_nan(x) ((x)==(x))
2
+
3
+ #define m_mulsum(x,y,z) {z = m_add(m_mul(x,y),z);}
4
+ #define m_mulsum_nan(x,y,z) { \
5
+ if(not_nan(x) && not_nan(y)) { \
6
+ z = m_add(m_mul(x,y),z); \
7
+ }}
8
+
9
+ #define m_cumsum(x,y) {(x)=m_add(x,y);}
10
+ #define m_cumsum_nan(x,y) { \
11
+ if (!not_nan(x)) { \
12
+ (x) = (y); \
13
+ } else if (not_nan(y)) { \
14
+ (x) = m_add(x,y); \
15
+ }}
16
+
17
+ #define m_cumprod(x,y) {(x)=m_mul(x,y);}
18
+ #define m_cumprod_nan(x,y) { \
19
+ if (!not_nan(x)) { \
20
+ (x) = (y); \
21
+ } else if (not_nan(y)) { \
22
+ (x) = m_mul(x,y); \
23
+ }}
24
+
25
+ static inline dtype f_sum(size_t n, char *p, ssize_t stride)
26
+ {
27
+ size_t i=n;
28
+ dtype x,y=m_zero;
29
+
30
+ for (; i--;) {
31
+ x = *(dtype*)p;
32
+ y = m_add(x,y);
33
+ p += stride;
34
+ }
35
+ return y;
36
+ }
37
+
38
+ static inline dtype f_sum_nan(size_t n, char *p, ssize_t stride)
39
+ {
40
+ size_t i=n;
41
+ dtype x,y=m_zero;
42
+
43
+ for (; i--;) {
44
+ x = *(dtype*)p;
45
+ p += stride;
46
+ if (not_nan(x)) {
47
+ y = m_add(x,y);
48
+ }
49
+ }
50
+ return y;
51
+ }
52
+
53
+
54
+ static inline dtype f_prod(size_t n, char *p, ssize_t stride)
55
+ {
56
+ size_t i=n;
57
+ dtype x,y=m_one;
58
+
59
+ for (; i--;) {
60
+ x = *(dtype*)p;
61
+ p += stride;
62
+ y = m_mul(x,y);
63
+ }
64
+ return y;
65
+ }
66
+
67
+ static inline dtype f_prod_nan(size_t n, char *p, ssize_t stride)
68
+ {
69
+ size_t i=n;
70
+ dtype x,y=m_one;
71
+
72
+ for (; i--;) {
73
+ x = *(dtype*)p;
74
+ p += stride;
75
+ if (not_nan(x)) {
76
+ y = m_mul(x,y);
77
+ }
78
+ }
79
+ return y;
80
+ }
81
+
82
+ static inline dtype f_mean(size_t n, char *p, ssize_t stride)
83
+ {
84
+ size_t i=n;
85
+ size_t count=0;
86
+ dtype x,y=m_zero;
87
+
88
+ for (; i--;) {
89
+ x = *(dtype*)p;
90
+ p += stride;
91
+ y = m_add(x,y);
92
+ count++;
93
+ }
94
+ return m_div(y,m_from_real(count));
95
+ }
96
+
97
+ static inline dtype f_mean_nan(size_t n, char *p, ssize_t stride)
98
+ {
99
+ size_t i=n;
100
+ size_t count=0;
101
+ dtype x,y=m_zero;
102
+
103
+ for (; i--;) {
104
+ x = *(dtype*)p;
105
+ p += stride;
106
+ if (not_nan(x)) {
107
+ y = m_add(x,y);
108
+ count++;
109
+ }
110
+ }
111
+ return m_div(y,m_from_real(count));
112
+ }
113
+
114
+ static inline dtype f_var(size_t n, char *p, ssize_t stride)
115
+ {
116
+ size_t i=n;
117
+ size_t count=0;
118
+ dtype x,y=m_zero;
119
+ dtype a,m;
120
+
121
+ m = f_mean(n,p,stride);
122
+
123
+ for (; i--;) {
124
+ x = *(dtype*)p;
125
+ p += stride;
126
+ a = m_abs(m_sub(x,m));
127
+ y = m_add(y,m_square(a));
128
+ count++;
129
+ }
130
+ return m_div(y,m_from_real(count-1));
131
+ }
132
+
133
+ static inline dtype f_var_nan(size_t n, char *p, ssize_t stride)
134
+ {
135
+ size_t i=n;
136
+ size_t count=0;
137
+ dtype x,y=m_zero;
138
+ dtype a,m;
139
+
140
+ m = f_mean_nan(n,p,stride);
141
+
142
+ for (; i--;) {
143
+ x = *(dtype*)p;
144
+ p += stride;
145
+ if (not_nan(x)) {
146
+ a = m_abs(m_sub(x,m));
147
+ y = m_add(y,m_square(a));
148
+ count++;
149
+ }
150
+ }
151
+ return m_div(y,m_from_real(count-1));
152
+ }
153
+
154
+ static inline dtype f_stddev(size_t n, char *p, ssize_t stride)
155
+ {
156
+ return m_sqrt(f_var(n,p,stride));
157
+ }
158
+
159
+ static inline dtype f_stddev_nan(size_t n, char *p, ssize_t stride)
160
+ {
161
+ return m_sqrt(f_var_nan(n,p,stride));
162
+ }
163
+
164
+ static inline dtype f_rms(size_t n, char *p, ssize_t stride)
165
+ {
166
+ size_t i=n;
167
+ size_t count=0;
168
+ dtype x,y=m_zero;
169
+
170
+ for (; i--;) {
171
+ x = *(dtype*)p;
172
+ p += stride;
173
+ y = m_add(y,m_square(m_abs(x)));
174
+ count++;
175
+ }
176
+ return m_sqrt(m_div(y,m_from_real(count)));
177
+ }
178
+
179
+ static inline dtype f_rms_nan(size_t n, char *p, ssize_t stride)
180
+ {
181
+ size_t i=n;
182
+ size_t count=0;
183
+ dtype x,y=m_zero;
184
+
185
+ for (; i--;) {
186
+ x = *(dtype*)p;
187
+ p += stride;
188
+ if (not_nan(x)) {
189
+ y = m_add(y,m_square(m_abs(x)));
190
+ count++;
191
+ }
192
+ }
193
+ return m_sqrt(m_div(y,m_from_real(count)));
194
+ }
195
+
196
+ // ---------------------------------------------------------
197
+
198
+ static inline dtype f_min_nan(size_t n, char *p, ssize_t stride)
199
+ {
200
+ dtype x,y;
201
+ size_t i=n;
202
+
203
+ y = *(dtype*)p;
204
+ p += stride;
205
+ if (!not_nan(y)) {return y;}
206
+ for (i--; i--;) {
207
+ x = *(dtype*)p;
208
+ p += stride;
209
+ if (!not_nan(x)) {return x;}
210
+ if (m_lt(x,y)) {
211
+ y = x;
212
+ }
213
+ }
214
+ return y;
215
+ }
216
+
217
+ static inline dtype f_min(size_t n, char *p, ssize_t stride)
218
+ {
219
+ dtype x,y=m_zero;
220
+ size_t i=n;
221
+
222
+ for (; i--; ) {
223
+ y = *(dtype*)p;
224
+ p += stride;
225
+ if (not_nan(y)) {
226
+ for (; i--;) {
227
+ x = *(dtype*)p;
228
+ p += stride;
229
+ if (m_lt(x,y)) {
230
+ y = x;
231
+ }
232
+ }
233
+ break;
234
+ }
235
+ }
236
+ return y;
237
+ }
238
+
239
+ static inline dtype f_max_nan(size_t n, char *p, ssize_t stride)
240
+ {
241
+ dtype x,y;
242
+ size_t i=n;
243
+
244
+ y = *(dtype*)p;
245
+ p += stride;
246
+ if (!not_nan(y)) {return y;}
247
+ for (i--; i--;) {
248
+ x = *(dtype*)p;
249
+ p += stride;
250
+ if (!not_nan(x)) {return x;}
251
+ if (m_gt(x,y)) {
252
+ y = x;
253
+ }
254
+ }
255
+ return y;
256
+ }
257
+
258
+ static inline dtype f_max(size_t n, char *p, ssize_t stride)
259
+ {
260
+ dtype x,y=m_zero;
261
+ size_t i=n;
262
+
263
+ for (; i--; ) {
264
+ y = *(dtype*)p;
265
+ p += stride;
266
+ if (not_nan(y)) {
267
+ for (; i--;) {
268
+ x = *(dtype*)p;
269
+ p += stride;
270
+ if (m_gt(x,y)) {
271
+ y = x;
272
+ }
273
+ }
274
+ break;
275
+ }
276
+ }
277
+ return y;
278
+ }
279
+
280
+ static inline size_t f_min_index_nan(size_t n, char *p, ssize_t stride)
281
+ {
282
+ dtype x, y;
283
+ size_t i, j=0;
284
+
285
+ y = *(dtype*)p;
286
+ p += stride;
287
+ if (!not_nan(y)) {return j;}
288
+ for (i=1; i<n; i++) {
289
+ x = *(dtype*)p;
290
+ p += stride;
291
+ if (!not_nan(x)) {return i;}
292
+ if (m_lt(x,y)) {
293
+ y = x;
294
+ j = i;
295
+ }
296
+ }
297
+ return j;
298
+ }
299
+
300
+ static inline size_t f_min_index(size_t n, char *p, ssize_t stride)
301
+ {
302
+ dtype x, y;
303
+ size_t i, j=0;
304
+
305
+ for (i=0; i<n; i++) {
306
+ y = *(dtype*)p;
307
+ p += stride;
308
+ if (not_nan(y)) {
309
+ j = i; i++;
310
+ for (; i<n; i++) {
311
+ x = *(dtype*)p;
312
+ p += stride;
313
+ if (m_lt(x,y)) {
314
+ y = x;
315
+ j = i;
316
+ }
317
+ }
318
+ break;
319
+ }
320
+ }
321
+ return j;
322
+ }
323
+
324
+ static inline size_t f_max_index_nan(size_t n, char *p, ssize_t stride)
325
+ {
326
+ dtype x, y;
327
+ size_t i, j=0;
328
+
329
+ y = *(dtype*)p;
330
+ p += stride;
331
+ if (!not_nan(y)) {return j;}
332
+ for (i=1; i<n; i++) {
333
+ x = *(dtype*)p;
334
+ p += stride;
335
+ if (!not_nan(x)) {return i;}
336
+ if (m_gt(x,y)) {
337
+ y = x;
338
+ j = i;
339
+ }
340
+ }
341
+ return j;
342
+ }
343
+
344
+ static inline size_t f_max_index(size_t n, char *p, ssize_t stride)
345
+ {
346
+ dtype x, y;
347
+ size_t i, j=0;
348
+
349
+ for (i=0; i<n; i++) {
350
+ y = *(dtype*)p;
351
+ p += stride;
352
+ if (not_nan(y)) {
353
+ j = i; i++;
354
+ for (; i<n; i++) {
355
+ x = *(dtype*)p;
356
+ p += stride;
357
+ if (m_gt(x,y)) {
358
+ y = x;
359
+ j = i;
360
+ }
361
+ }
362
+ break;
363
+ }
364
+ }
365
+ return j;
366
+ }
367
+
368
+ static inline void
369
+ f_minmax_nan(size_t n, char *p, ssize_t stride, dtype *amin, dtype *amax)
370
+ {
371
+ dtype x,min,max;
372
+ size_t i=n;
373
+
374
+ min = max = *(dtype*)p;
375
+ p += stride;
376
+ if (!not_nan(min)) {
377
+ *amin = *amax = min;
378
+ return;
379
+ }
380
+ for (i--; i--;) {
381
+ x = *(dtype*)p;
382
+ p += stride;
383
+ if (!not_nan(x)) {
384
+ *amin = *amax = x;
385
+ return;
386
+ }
387
+ if (m_lt(x,min)) {
388
+ min = x;
389
+ }
390
+ if (m_gt(x,max)) {
391
+ max = x;
392
+ }
393
+ }
394
+ *amin = min;
395
+ *amax = max;
396
+ return;
397
+ }
398
+
399
+ static inline dtype f_ptp_nan(size_t n, char *p, ssize_t stride)
400
+ {
401
+ dtype min,max;
402
+ f_minmax_nan(n,p,stride,&min,&max);
403
+ return m_sub(max,min);
404
+ }
405
+
406
+ static inline void
407
+ f_minmax(size_t n, char *p, ssize_t stride, dtype *amin, dtype *amax)
408
+ {
409
+ dtype x,min,max;
410
+ size_t i=n;
411
+
412
+ min = max = m_zero;
413
+ for (; i--; ) {
414
+ min = *(dtype*)p;
415
+ p += stride;
416
+ if (not_nan(min)) {
417
+ max = min;
418
+ for (; i--;) {
419
+ x = *(dtype*)p;
420
+ p += stride;
421
+ if (m_lt(x,min)) {
422
+ min = x;
423
+ }
424
+ if (m_gt(x,max)) {
425
+ max = x;
426
+ }
427
+ }
428
+ break;
429
+ }
430
+ }
431
+ *amin = min;
432
+ *amax = max;
433
+ return;
434
+ }
435
+
436
+ static inline dtype f_ptp(size_t n, char *p, ssize_t stride)
437
+ {
438
+ dtype min,max;
439
+ f_minmax(n,p,stride,&min,&max);
440
+ return m_sub(max,min);
441
+ }
442
+
443
+
444
+ static inline dtype f_maximum(dtype x, dtype y)
445
+ {
446
+ if (m_ge(x,y)) {
447
+ return x;
448
+ }
449
+ if (not_nan(y)) {
450
+ return y;
451
+ }
452
+ return x;
453
+ }
454
+
455
+ static inline dtype f_maximum_nan(dtype x, dtype y)
456
+ {
457
+ if (m_ge(x,y)) {
458
+ return x;
459
+ }
460
+ if (!not_nan(x)) {
461
+ return x;
462
+ }
463
+ return y;
464
+ }
465
+
466
+ static inline dtype f_minimum(dtype x, dtype y)
467
+ {
468
+ if (m_le(x,y)) {
469
+ return x;
470
+ }
471
+ if (not_nan(y)) {
472
+ return y;
473
+ }
474
+ return x;
475
+ }
476
+
477
+ static inline dtype f_minimum_nan(dtype x, dtype y)
478
+ {
479
+ if (m_le(x,y)) {
480
+ return x;
481
+ }
482
+ if (!not_nan(x)) {
483
+ return x;
484
+ }
485
+ return y;
486
+ }