ruby-dnn 0.1.0

Sign up to get free protection for your applications and to get access to all the features.
Files changed (127) hide show
  1. checksums.yaml +7 -0
  2. data/.gitignore +8 -0
  3. data/.travis.yml +5 -0
  4. data/CODE_OF_CONDUCT.md +74 -0
  5. data/Gemfile +6 -0
  6. data/LICENSE.txt +21 -0
  7. data/README.md +42 -0
  8. data/Rakefile +10 -0
  9. data/bin/console +14 -0
  10. data/bin/setup +8 -0
  11. data/lib/dnn.rb +14 -0
  12. data/lib/dnn/core/activations.rb +116 -0
  13. data/lib/dnn/core/error.rb +13 -0
  14. data/lib/dnn/core/initializers.rb +46 -0
  15. data/lib/dnn/core/layers.rb +366 -0
  16. data/lib/dnn/core/model.rb +158 -0
  17. data/lib/dnn/core/optimizers.rb +113 -0
  18. data/lib/dnn/core/util.rb +24 -0
  19. data/lib/dnn/core/version.rb +3 -0
  20. data/lib/dnn/ext/cifar10/Makefile +263 -0
  21. data/lib/dnn/ext/cifar10/cifar10_ext.c +52 -0
  22. data/lib/dnn/ext/cifar10/cifar10_ext.o +0 -0
  23. data/lib/dnn/ext/cifar10/cifar10_ext.so +0 -0
  24. data/lib/dnn/ext/cifar10/extconf.rb +3 -0
  25. data/lib/dnn/ext/cifar10/numo/compat.h +23 -0
  26. data/lib/dnn/ext/cifar10/numo/extconf.h +13 -0
  27. data/lib/dnn/ext/cifar10/numo/intern.h +117 -0
  28. data/lib/dnn/ext/cifar10/numo/narray.h +430 -0
  29. data/lib/dnn/ext/cifar10/numo/ndloop.h +94 -0
  30. data/lib/dnn/ext/cifar10/numo/template.h +149 -0
  31. data/lib/dnn/ext/cifar10/numo/types/bit.h +33 -0
  32. data/lib/dnn/ext/cifar10/numo/types/complex.h +409 -0
  33. data/lib/dnn/ext/cifar10/numo/types/complex_macro.h +377 -0
  34. data/lib/dnn/ext/cifar10/numo/types/dcomplex.h +44 -0
  35. data/lib/dnn/ext/cifar10/numo/types/dfloat.h +42 -0
  36. data/lib/dnn/ext/cifar10/numo/types/float_def.h +34 -0
  37. data/lib/dnn/ext/cifar10/numo/types/float_macro.h +186 -0
  38. data/lib/dnn/ext/cifar10/numo/types/int16.h +24 -0
  39. data/lib/dnn/ext/cifar10/numo/types/int32.h +24 -0
  40. data/lib/dnn/ext/cifar10/numo/types/int64.h +24 -0
  41. data/lib/dnn/ext/cifar10/numo/types/int8.h +24 -0
  42. data/lib/dnn/ext/cifar10/numo/types/int_macro.h +41 -0
  43. data/lib/dnn/ext/cifar10/numo/types/real_accum.h +486 -0
  44. data/lib/dnn/ext/cifar10/numo/types/robj_macro.h +75 -0
  45. data/lib/dnn/ext/cifar10/numo/types/robject.h +27 -0
  46. data/lib/dnn/ext/cifar10/numo/types/scomplex.h +44 -0
  47. data/lib/dnn/ext/cifar10/numo/types/sfloat.h +43 -0
  48. data/lib/dnn/ext/cifar10/numo/types/uint16.h +21 -0
  49. data/lib/dnn/ext/cifar10/numo/types/uint32.h +21 -0
  50. data/lib/dnn/ext/cifar10/numo/types/uint64.h +21 -0
  51. data/lib/dnn/ext/cifar10/numo/types/uint8.h +21 -0
  52. data/lib/dnn/ext/cifar10/numo/types/uint_macro.h +32 -0
  53. data/lib/dnn/ext/cifar10/numo/types/xint_macro.h +189 -0
  54. data/lib/dnn/ext/image_io/Makefile +263 -0
  55. data/lib/dnn/ext/image_io/extconf.rb +3 -0
  56. data/lib/dnn/ext/image_io/image_io_ext.c +89 -0
  57. data/lib/dnn/ext/image_io/image_io_ext.so +0 -0
  58. data/lib/dnn/ext/image_io/numo/compat.h +23 -0
  59. data/lib/dnn/ext/image_io/numo/extconf.h +13 -0
  60. data/lib/dnn/ext/image_io/numo/intern.h +117 -0
  61. data/lib/dnn/ext/image_io/numo/narray.h +430 -0
  62. data/lib/dnn/ext/image_io/numo/ndloop.h +94 -0
  63. data/lib/dnn/ext/image_io/numo/template.h +149 -0
  64. data/lib/dnn/ext/image_io/numo/types/bit.h +33 -0
  65. data/lib/dnn/ext/image_io/numo/types/complex.h +409 -0
  66. data/lib/dnn/ext/image_io/numo/types/complex_macro.h +377 -0
  67. data/lib/dnn/ext/image_io/numo/types/dcomplex.h +44 -0
  68. data/lib/dnn/ext/image_io/numo/types/dfloat.h +42 -0
  69. data/lib/dnn/ext/image_io/numo/types/float_def.h +34 -0
  70. data/lib/dnn/ext/image_io/numo/types/float_macro.h +186 -0
  71. data/lib/dnn/ext/image_io/numo/types/int16.h +24 -0
  72. data/lib/dnn/ext/image_io/numo/types/int32.h +24 -0
  73. data/lib/dnn/ext/image_io/numo/types/int64.h +24 -0
  74. data/lib/dnn/ext/image_io/numo/types/int8.h +24 -0
  75. data/lib/dnn/ext/image_io/numo/types/int_macro.h +41 -0
  76. data/lib/dnn/ext/image_io/numo/types/real_accum.h +486 -0
  77. data/lib/dnn/ext/image_io/numo/types/robj_macro.h +75 -0
  78. data/lib/dnn/ext/image_io/numo/types/robject.h +27 -0
  79. data/lib/dnn/ext/image_io/numo/types/scomplex.h +44 -0
  80. data/lib/dnn/ext/image_io/numo/types/sfloat.h +43 -0
  81. data/lib/dnn/ext/image_io/numo/types/uint16.h +21 -0
  82. data/lib/dnn/ext/image_io/numo/types/uint32.h +21 -0
  83. data/lib/dnn/ext/image_io/numo/types/uint64.h +21 -0
  84. data/lib/dnn/ext/image_io/numo/types/uint8.h +21 -0
  85. data/lib/dnn/ext/image_io/numo/types/uint_macro.h +32 -0
  86. data/lib/dnn/ext/image_io/numo/types/xint_macro.h +189 -0
  87. data/lib/dnn/ext/image_io/stb_image.h +7462 -0
  88. data/lib/dnn/ext/image_io/stb_image_write.h +1568 -0
  89. data/lib/dnn/ext/mnist/Makefile +263 -0
  90. data/lib/dnn/ext/mnist/extconf.rb +3 -0
  91. data/lib/dnn/ext/mnist/mnist_ext.c +49 -0
  92. data/lib/dnn/ext/mnist/mnist_ext.o +0 -0
  93. data/lib/dnn/ext/mnist/mnist_ext.so +0 -0
  94. data/lib/dnn/ext/mnist/numo/compat.h +23 -0
  95. data/lib/dnn/ext/mnist/numo/extconf.h +13 -0
  96. data/lib/dnn/ext/mnist/numo/intern.h +117 -0
  97. data/lib/dnn/ext/mnist/numo/narray.h +430 -0
  98. data/lib/dnn/ext/mnist/numo/ndloop.h +94 -0
  99. data/lib/dnn/ext/mnist/numo/template.h +149 -0
  100. data/lib/dnn/ext/mnist/numo/types/bit.h +33 -0
  101. data/lib/dnn/ext/mnist/numo/types/complex.h +409 -0
  102. data/lib/dnn/ext/mnist/numo/types/complex_macro.h +377 -0
  103. data/lib/dnn/ext/mnist/numo/types/dcomplex.h +44 -0
  104. data/lib/dnn/ext/mnist/numo/types/dfloat.h +42 -0
  105. data/lib/dnn/ext/mnist/numo/types/float_def.h +34 -0
  106. data/lib/dnn/ext/mnist/numo/types/float_macro.h +186 -0
  107. data/lib/dnn/ext/mnist/numo/types/int16.h +24 -0
  108. data/lib/dnn/ext/mnist/numo/types/int32.h +24 -0
  109. data/lib/dnn/ext/mnist/numo/types/int64.h +24 -0
  110. data/lib/dnn/ext/mnist/numo/types/int8.h +24 -0
  111. data/lib/dnn/ext/mnist/numo/types/int_macro.h +41 -0
  112. data/lib/dnn/ext/mnist/numo/types/real_accum.h +486 -0
  113. data/lib/dnn/ext/mnist/numo/types/robj_macro.h +75 -0
  114. data/lib/dnn/ext/mnist/numo/types/robject.h +27 -0
  115. data/lib/dnn/ext/mnist/numo/types/scomplex.h +44 -0
  116. data/lib/dnn/ext/mnist/numo/types/sfloat.h +43 -0
  117. data/lib/dnn/ext/mnist/numo/types/uint16.h +21 -0
  118. data/lib/dnn/ext/mnist/numo/types/uint32.h +21 -0
  119. data/lib/dnn/ext/mnist/numo/types/uint64.h +21 -0
  120. data/lib/dnn/ext/mnist/numo/types/uint8.h +21 -0
  121. data/lib/dnn/ext/mnist/numo/types/uint_macro.h +32 -0
  122. data/lib/dnn/ext/mnist/numo/types/xint_macro.h +189 -0
  123. data/lib/dnn/lib/cifar10.rb +26 -0
  124. data/lib/dnn/lib/image_io.rb +33 -0
  125. data/lib/dnn/lib/mnist.rb +61 -0
  126. data/ruby-dnn.gemspec +41 -0
  127. metadata +225 -0
@@ -0,0 +1,34 @@
1
+ #ifndef DBL_EPSILON
2
+ #define DBL_EPSILON 2.2204460492503131e-16
3
+ #endif
4
+ #ifndef FLT_EPSILON
5
+ #define FLT_EPSILON 1.1920928955078125e-07
6
+ #endif
7
+ #ifndef DBL_MAX
8
+ #define DBL_MAX 1.7976931348623157e+308
9
+ #endif
10
+ #ifndef DBL_MAX
11
+ #define DBL_MAX 1.7976931348623157e+308
12
+ #endif
13
+ #ifndef FLT_MIN
14
+ #define FLT_MIN 1.1754943508222875e-38
15
+ #endif
16
+ #ifndef FLT_MAX
17
+ #define FLT_MAX 3.4028234663852886e+38
18
+ #endif
19
+
20
+ #ifndef M_PI_2
21
+ #define M_PI_2 1.57079632679489661923 /* pi/2 */
22
+ #endif
23
+ #ifndef M_LOG2E
24
+ #define M_LOG2E 1.4426950408889634074 /* log_2 e */
25
+ #endif
26
+ #ifndef M_LOG10E
27
+ #define M_LOG10E 0.43429448190325182765 /* log_10 e */
28
+ #endif
29
+ #ifndef M_LN2
30
+ #define M_LN2 0.69314718055994530942 /* log_e 2 */
31
+ #endif
32
+ #ifndef M_LN10
33
+ #define M_LN10 2.30258509299404568402 /* log_e 10 */
34
+ #endif
@@ -0,0 +1,186 @@
1
+ #include "float_def.h"
2
+
3
+ extern double round(double);
4
+ extern double log2(double);
5
+ extern double exp2(double);
6
+ #ifdef HAVE_EXP10
7
+ extern double exp10(double);
8
+ #else
9
+ extern double pow(double, double);
10
+ #endif
11
+
12
+ #define m_zero 0.0
13
+ #define m_one 1.0
14
+
15
+ #define m_num_to_data(x) NUM2DBL(x)
16
+ #define m_data_to_num(x) rb_float_new(x)
17
+
18
+ #define m_from_double(x) (x)
19
+ #define m_from_real(x) (x)
20
+
21
+ #define m_add(x,y) ((x)+(y))
22
+ #define m_sub(x,y) ((x)-(y))
23
+ #define m_mul(x,y) ((x)*(y))
24
+ #define m_div(x,y) ((x)/(y))
25
+ #define m_div_check(x,y) ((y)==0)
26
+ #define m_mod(x,y) fmod(x,y)
27
+ #define m_divmod(x,y,a,b) {a=(x)/(y); b=m_mod(x,y);}
28
+ #define m_pow(x,y) pow(x,y)
29
+ #define m_pow_int(x,y) pow_int(x,y)
30
+
31
+ #define m_abs(x) fabs(x)
32
+ #define m_minus(x) (-(x))
33
+ #define m_reciprocal(x) (1/(x))
34
+ #define m_square(x) ((x)*(x))
35
+ #define m_floor(x) floor(x)
36
+ #define m_round(x) round(x)
37
+ #define m_ceil(x) ceil(x)
38
+ #define m_trunc(x) trunc(x)
39
+ #define m_rint(x) rint(x)
40
+ #define m_sign(x) (((x)==0) ? 0.0:(((x)>0) ? 1.0:(((x)<0) ? -1.0:(x))))
41
+ #define m_copysign(x,y) copysign(x,y)
42
+ #define m_signbit(x) signbit(x)
43
+ #define m_modf(x,y,z) {double d; y=modf(x,&d); z=d;}
44
+
45
+ #define m_eq(x,y) ((x)==(y))
46
+ #define m_ne(x,y) ((x)!=(y))
47
+ #define m_gt(x,y) ((x)>(y))
48
+ #define m_ge(x,y) ((x)>=(y))
49
+ #define m_lt(x,y) ((x)<(y))
50
+ #define m_le(x,y) ((x)<=(y))
51
+
52
+ #define m_isnan(x) isnan(x)
53
+ #define m_isinf(x) isinf(x)
54
+ #define m_isposinf(x) (isinf(x) && signbit(x)==0)
55
+ #define m_isneginf(x) (isinf(x) && signbit(x))
56
+ #define m_isfinite(x) isfinite(x)
57
+
58
+ #define m_mulsum_init INT2FIX(0)
59
+
60
+ #define m_sprintf(s,x) sprintf(s,"%g",x)
61
+
62
+ #define cmp_prnan(a,b) \
63
+ ((qsort_cast(a)==qsort_cast(b)) ? 0 : \
64
+ (qsort_cast(a) > qsort_cast(b)) ? 1 : -1)
65
+
66
+ #define cmp_ignan(a,b) \
67
+ (m_isnan(qsort_cast(a)) ? (m_isnan(qsort_cast(b)) ? 0 : 1) : \
68
+ (m_isnan(qsort_cast(b)) ? -1 : \
69
+ ((qsort_cast(a)==qsort_cast(b)) ? 0 : \
70
+ (qsort_cast(a) > qsort_cast(b)) ? 1 : -1)))
71
+
72
+ #define cmpgt_prnan(a,b) \
73
+ (qsort_cast(a) > qsort_cast(b))
74
+
75
+ #define cmpgt_ignan(a,b) \
76
+ ((m_isnan(qsort_cast(a)) && !m_isnan(qsort_cast(b))) || \
77
+ (qsort_cast(a) > qsort_cast(b)))
78
+
79
+ #define m_sqrt(x) sqrt(x)
80
+ #define m_cbrt(x) cbrt(x)
81
+ #define m_log(x) log(x)
82
+ #define m_log2(x) log2(x)
83
+ #define m_log10(x) log10(x)
84
+ #define m_exp(x) exp(x)
85
+ #define m_exp2(x) exp2(x)
86
+ #ifdef HAVE_EXP10
87
+ #define m_exp10(x) exp10(x)
88
+ #else
89
+ #define m_exp10(x) pow(10, x)
90
+ #endif
91
+ #define m_expm1(x) expm1(x)
92
+ #define m_log1p(x) log1p(x)
93
+
94
+ #define m_sin(x) sin(x)
95
+ #define m_cos(x) cos(x)
96
+ #define m_tan(x) tan(x)
97
+ #define m_asin(x) asin(x)
98
+ #define m_acos(x) acos(x)
99
+ #define m_atan(x) atan(x)
100
+ #define m_sinh(x) sinh(x)
101
+ #define m_cosh(x) cosh(x)
102
+ #define m_tanh(x) tanh(x)
103
+ #define m_asinh(x) asinh(x)
104
+ #define m_acosh(x) acosh(x)
105
+ #define m_atanh(x) atanh(x)
106
+ #define m_atan2(x,y) atan2(x,y)
107
+ #define m_hypot(x,y) hypot(x,y)
108
+ #define m_sinc(x) (sin(x)/(x))
109
+
110
+ #define m_erf(x) erf(x)
111
+ #define m_erfc(x) erfc(x)
112
+ #define m_ldexp(x,y) ldexp(x,y)
113
+ #define m_frexp(x,exp) frexp(x,exp)
114
+
115
+ static inline dtype pow_int(dtype x, int p)
116
+ {
117
+ dtype r=1;
118
+ switch(p) {
119
+ case 0: return 1;
120
+ case 1: return x;
121
+ case 2: return x*x;
122
+ case 3: return x*x*x;
123
+ case 4: x=x*x; return x*x;
124
+ }
125
+ if (p<0) return 1/pow_int(x,-p);
126
+ if (p>64) return pow(x,p);
127
+ while (p) {
128
+ if (p&1) r *= x;
129
+ x *= x;
130
+ p >>= 1;
131
+ }
132
+ return r;
133
+ }
134
+
135
+ static inline dtype f_seq(dtype x, dtype y, double c)
136
+ {
137
+ return x + y * c;
138
+ }
139
+
140
+ static inline dtype f_kahan_sum(size_t n, char *p, ssize_t stride)
141
+ {
142
+ size_t i=n;
143
+ dtype x;
144
+ volatile dtype y=0;
145
+ volatile dtype t,r=0;
146
+
147
+ for (; i--;) {
148
+ x = *(dtype*)p;
149
+ p += stride;
150
+ if (fabs(x) > fabs(y)) {
151
+ dtype z=x; x=y; y=z;
152
+ }
153
+ r += x;
154
+ t = y;
155
+ y += r;
156
+ t = y-t;
157
+ r -= t;
158
+ }
159
+ return y;
160
+ }
161
+
162
+ static inline dtype f_kahan_sum_nan(size_t n, char *p, ssize_t stride)
163
+ {
164
+ size_t i=n;
165
+ dtype x;
166
+ volatile dtype y=0;
167
+ volatile dtype t,r=0;
168
+
169
+ for (; i--;) {
170
+ x = *(dtype*)p;
171
+ p += stride;
172
+ if (!m_isnan(x)) {
173
+ if (fabs(x) > fabs(y)) {
174
+ dtype z=x; x=y; y=z;
175
+ }
176
+ r += x;
177
+ t = y;
178
+ y += r;
179
+ t = y-t;
180
+ r -= t;
181
+ }
182
+ }
183
+ return y;
184
+ }
185
+
186
+ #include "real_accum.h"
@@ -0,0 +1,24 @@
1
+ typedef int16_t dtype;
2
+ typedef int16_t rtype;
3
+ #define cT numo_cInt16
4
+ #define cRT cT
5
+
6
+ #define m_num_to_data(x) ((dtype)NUM2INT(x))
7
+ #define m_data_to_num(x) INT2NUM((int)(x))
8
+ #define m_extract(x) INT2NUM((int)*(dtype*)(x))
9
+ #define m_sprintf(s,x) sprintf(s,"%d",(int)(x))
10
+
11
+ #ifndef INT16_MIN
12
+ #define INT16_MIN (-32767-1)
13
+ #endif
14
+ #ifndef INT16_MAX
15
+ #define INT16_MAX (32767)
16
+ #endif
17
+
18
+ #define DATA_MIN INT16_MIN
19
+ #define DATA_MAX INT16_MAX
20
+
21
+ #define M_MIN m_data_to_num(INT16_MIN)
22
+ #define M_MAX m_data_to_num(INT16_MAX)
23
+
24
+ #include "int_macro.h"
@@ -0,0 +1,24 @@
1
+ typedef int32_t dtype;
2
+ typedef int32_t rtype;
3
+ #define cT numo_cInt32
4
+ #define cRT cT
5
+
6
+ #define m_num_to_data(x) ((dtype)NUM2INT32(x))
7
+ #define m_data_to_num(x) INT322NUM((int32_t)(x))
8
+ #define m_extract(x) INT322NUM((int32_t)*(dtype*)(x))
9
+ #define m_sprintf(s,x) sprintf(s,"%"PRId32,(int32_t)(x))
10
+
11
+ #ifndef INT32_MIN
12
+ #define INT32_MIN (-2147483647-1)
13
+ #endif
14
+ #ifndef INT32_MAX
15
+ #define INT32_MAX (2147483647)
16
+ #endif
17
+
18
+ #define DATA_MIN INT32_MIN
19
+ #define DATA_MAX INT32_MAX
20
+
21
+ #define M_MIN m_data_to_num(INT32_MIN)
22
+ #define M_MAX m_data_to_num(INT32_MAX)
23
+
24
+ #include "int_macro.h"
@@ -0,0 +1,24 @@
1
+ typedef int64_t dtype;
2
+ typedef int64_t rtype;
3
+ #define cT numo_cInt64
4
+ #define cRT cT
5
+
6
+ #define m_num_to_data(x) ((dtype)NUM2INT64(x))
7
+ #define m_data_to_num(x) INT642NUM((int64_t)(x))
8
+ #define m_extract(x) INT642NUM((int64_t)*(dtype*)(x))
9
+ #define m_sprintf(s,x) sprintf(s,"%"PRId64,(int64_t)(x))
10
+
11
+ #ifndef INT64_MIN
12
+ #define INT64_MIN (-9223372036854775807l-1)
13
+ #endif
14
+ #ifndef INT64_MAX
15
+ #define INT64_MAX (9223372036854775807l)
16
+ #endif
17
+
18
+ #define DATA_MIN INT64_MIN
19
+ #define DATA_MAX INT64_MAX
20
+
21
+ #define M_MIN m_data_to_num(INT64_MIN)
22
+ #define M_MAX m_data_to_num(INT64_MAX)
23
+
24
+ #include "int_macro.h"
@@ -0,0 +1,24 @@
1
+ typedef int8_t dtype;
2
+ typedef int8_t rtype;
3
+ #define cT numo_cInt8
4
+ #define cRT cT
5
+
6
+ #define m_num_to_data(x) ((dtype)NUM2INT(x))
7
+ #define m_data_to_num(x) INT2NUM((int)(x))
8
+ #define m_extract(x) INT2NUM((int)*(dtype*)(x))
9
+ #define m_sprintf(s,x) sprintf(s,"%d",(int)(x))
10
+
11
+ #ifndef INT8_MIN
12
+ #define INT8_MIN (-127-1)
13
+ #endif
14
+ #ifndef INT8_MAX
15
+ #define INT8_MAX (127)
16
+ #endif
17
+
18
+ #define DATA_MIN INT8_MIN
19
+ #define DATA_MAX INT8_MAX
20
+
21
+ #define M_MIN INT2FIX(INT8_MIN)
22
+ #define M_MAX INT2FIX(INT8_MAX)
23
+
24
+ #include "int_macro.h"
@@ -0,0 +1,41 @@
1
+ #include "xint_macro.h"
2
+
3
+ #define m_sign(x) (((x)==0) ? 0 : (((x)>0) ? 1 : -1))
4
+
5
+ static inline dtype m_abs(dtype x) {
6
+ if (x==DATA_MIN) {
7
+ rb_raise(nary_eValueError, "cannot convert the minimum integer");
8
+ }
9
+ return (x<0)?-x:x;
10
+ }
11
+
12
+ static inline dtype int_reciprocal(dtype x) {
13
+ switch (x) {
14
+ case 1:
15
+ return 1;
16
+ case -1:
17
+ return -1;
18
+ case 0:
19
+ rb_raise(rb_eZeroDivError, "divided by 0");
20
+ default:
21
+ return 0;
22
+ }
23
+ }
24
+
25
+ static dtype pow_int(dtype x, int p)
26
+ {
27
+ dtype r = m_one;
28
+ switch(p) {
29
+ case 0: return 1;
30
+ case 1: return x;
31
+ case 2: return x*x;
32
+ case 3: return x*x*x;
33
+ }
34
+ if (p<0) return 0;
35
+ while (p) {
36
+ if (p&1) r *= x;
37
+ x *= x;
38
+ p >>= 1;
39
+ }
40
+ return r;
41
+ }
@@ -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
+ }