ruby-dnn 0.1.0

data/lib/dnn/ext/cifar10/numo/types/complex_macro.h

@@ -0,0 +1,377 @@
+#include "float_def.h"
+
+extern double round(double);
+extern double log2(double);
+extern double exp2(double);
+extern double exp10(double);
+
+#define r_abs(x)   fabs(x)
+#define r_sqrt(x)  sqrt(x)
+#define r_exp(x)   exp(x)
+#define r_log(x)   log(x)
+#define r_sin(x)   sin(x)
+#define r_cos(x)   cos(x)
+#define r_sinh(x)  sinh(x)
+#define r_cosh(x)  cosh(x)
+#define r_tanh(x)  tanh(x)
+#define r_atan2(y,x)  atan2(y,x)
+#define r_hypot(x,y)  hypot(x,y)
+
+#include "complex.h"
+
+static inline dtype c_from_scomplex(scomplex x) {
+    dtype z;
+    REAL(z) = REAL(x);
+    IMAG(z) = IMAG(x);
+    return z;
+}
+
+static inline dtype c_from_dcomplex(dcomplex x) {
+    dtype z;
+    REAL(z) = REAL(x);
+    IMAG(z) = IMAG(x);
+    return z;
+}
+
+/* --------------------------- */
+
+#define m_zero c_zero()
+#define m_one  c_one()
+
+#define m_num_to_data(x) NUM2COMP(x)
+#define m_data_to_num(x) COMP2NUM(x)
+
+#define m_from_double(x) c_new(x,0)
+#define m_from_real(x)   c_new(x,0)
+#define m_from_scomplex(x) c_from_scomplex(x)
+#define m_from_dcomplex(x) c_from_dcomplex(x)
+
+#define m_extract(x) COMP2NUM(*(dtype*)x)
+
+#define m_real(x)  REAL(x)
+#define m_imag(x)  IMAG(x)
+#define m_set_real(x,y)  c_set_real(x,y)
+#define m_set_imag(x,y)  c_set_imag(x,y)
+
+#define m_add(x,y) c_add(x,y)
+#define m_sub(x,y) c_sub(x,y)
+#define m_mul(x,y) c_mul(x,y)
+#define m_div(x,y) c_div(x,y)
+#define m_mod(x,y) c_mod(x,y)
+#define m_pow(x,y) c_pow(x,y)
+#define m_pow_int(x,y) c_pow_int(x,y)
+
+#define m_abs(x)   c_abs(x)
+#define m_minus(x) c_minus(x)
+#define m_reciprocal(x) c_reciprocal(x)
+#define m_square(x) c_square(x)
+#define m_floor(x) c_new(floor(REAL(x)),floor(IMAG(x)))
+#define m_round(x) c_new(round(REAL(x)),round(IMAG(x)))
+#define m_ceil(x)  c_new(ceil(REAL(x)),ceil(IMAG(x)))
+#define m_trunc(x) c_new(trunc(REAL(x)),trunc(IMAG(x)))
+#define m_rint(x)  c_new(rint(REAL(x)),rint(IMAG(x)))
+#define m_sign(x)  c_new( \
+ ((REAL(x)==0) ? 0.0:((REAL(x)>0) ? 1.0:((REAL(x)<0) ? -1.0:REAL(x)))), \
+ ((IMAG(x)==0) ? 0.0:((IMAG(x)>0) ? 1.0:((IMAG(x)<0) ? -1.0:IMAG(x)))))
+#define m_copysign(x,y) c_new(copysign(REAL(x),REAL(y)),copysign(IMAG(x),IMAG(y)))
+
+#define m_im(x)    c_im(x)
+#define m_conj(x)  c_new(REAL(x),-IMAG(x))
+#define m_arg(x)   atan2(IMAG(x),REAL(x))
+
+#define m_eq(x,y) c_eq(x,y)
+#define m_ne(x,y) c_ne(x,y)
+#define m_nearly_eq(x,y) c_nearly_eq(x,y)
+
+#define m_isnan(x)    c_isnan(x)
+#define m_isinf(x)    c_isinf(x)
+#define m_isposinf(x) c_isposinf(x)
+#define m_isneginf(x) c_isneginf(x)
+#define m_isfinite(x) c_isfinite(x)
+
+#define m_sprintf(s,x) sprintf(s,"%g%+gi",REAL(x),IMAG(x))
+
+#define m_sqrt(x)    c_sqrt(x)
+#define m_cbrt(x)    c_cbrt(x)
+#define m_log(x)     c_log(x)
+#define m_log2(x)    c_log2(x)
+#define m_log10(x)   c_log10(x)
+#define m_exp(x)     c_exp(x)
+#define m_exp2(x)    c_exp2(x)
+#define m_exp10(x)   c_exp10(x)
+#define m_sin(x)     c_sin(x)
+#define m_cos(x)     c_cos(x)
+#define m_tan(x)     c_tan(x)
+#define m_asin(x)    c_asin(x)
+#define m_acos(x)    c_acos(x)
+#define m_atan(x)    c_atan(x)
+#define m_sinh(x)    c_sinh(x)
+#define m_cosh(x)    c_cosh(x)
+#define m_tanh(x)    c_tanh(x)
+#define m_asinh(x)   c_asinh(x)
+#define m_acosh(x)   c_acosh(x)
+#define m_atanh(x)   c_atanh(x)
+#define m_hypot(x,y) c_hypot(x,y)
+#define m_sinc(x)    c_div(c_sin(x),x)
+
+#define m_sum_init INT2FIX(0)
+#define m_mulsum_init INT2FIX(0)
+
+#define not_nan(x) (REAL(x)==REAL(x) && IMAG(x)==IMAG(x))
+
+#define m_mulsum(x,y,z) {z = m_add(m_mul(x,y),z);}
+#define m_mulsum_nan(x,y,z) {          \
+        if(not_nan(x) && not_nan(y)) { \
+            z = m_add(m_mul(x,y),z);   \
+        }}
+
+#define m_cumsum(x,y) {(x)=m_add(x,y);}
+#define m_cumsum_nan(x,y) {      \
+        if (!not_nan(x)) {       \
+            (x) = (y);           \
+        } else if (not_nan(y)) { \
+            (x) = m_add(x,y);    \
+        }}
+
+#define m_cumprod(x,y) {(x)=m_mul(x,y);}
+#define m_cumprod_nan(x,y) {     \
+        if (!not_nan(x)) {       \
+            (x) = (y);           \
+        } else if (not_nan(y)) { \
+            (x) = m_mul(x,y);    \
+        }}
+
+static inline dtype f_sum(size_t n, char *p, ssize_t stride)
+{
+    size_t i=n;
+    dtype x,y;
+
+    y = c_zero();
+    for (; i--;) {
+        x = *(dtype*)p;
+        y = c_add(x,y);
+        p += stride;
+    }
+    return y;
+}
+
+static inline dtype f_sum_nan(size_t n, char *p, ssize_t stride)
+{
+    size_t i=n;
+    dtype x,y;
+
+    y = c_zero();
+    for (; i--;) {
+        x = *(dtype*)p;
+        if (not_nan(x)) {
+            y = c_add(x,y);
+        }
+        p += stride;
+    }
+    return y;
+}
+
+static inline dtype f_kahan_sum(size_t n, char *p, ssize_t stride)
+{
+    size_t i=n;
+    dtype x;
+    volatile dtype y,t,r;
+
+    y = c_zero();
+    r = c_zero();
+    for (; i--;) {
+        x = *(dtype*)p;
+        if (fabs(REAL(x)) > fabs(REAL(y))) {
+            double z=REAL(x); REAL(x)=REAL(y); REAL(y)=z;
+        }
+        if (fabs(IMAG(x)) > fabs(IMAG(y))) {
+            double z=IMAG(x); IMAG(x)=IMAG(y); IMAG(y)=z;
+        }
+        r = c_add(x, r);
+        t = y;
+        y = c_add(r, y);
+        t = c_sub(y, t);
+        r = c_sub(r, t);
+        p += stride;
+    }
+    return y;
+}
+
+static inline dtype f_kahan_sum_nan(size_t n, char *p, ssize_t stride)
+{
+    size_t i=n;
+    dtype x;
+    volatile dtype y,t,r;
+
+    y = c_zero();
+    r = c_zero();
+    for (; i--;) {
+        x = *(dtype*)p;
+        if (not_nan(x)) {
+            if (fabs(REAL(x)) > fabs(REAL(y))) {
+                double z=REAL(x); REAL(x)=REAL(y); REAL(y)=z;
+            }
+            if (fabs(IMAG(x)) > fabs(IMAG(y))) {
+                double z=IMAG(x); IMAG(x)=IMAG(y); IMAG(y)=z;
+            }
+            r = c_add(x, r);
+            t = y;
+            y = c_add(r, y);
+            t = c_sub(y, t);
+            r = c_sub(r, t);
+        }
+        p += stride;
+    }
+    return y;
+}
+
+static inline dtype f_prod(size_t n, char *p, ssize_t stride)
+{
+    size_t i=n;
+    dtype x,y;
+
+    y = c_one();
+    for (; i--;) {
+        x = *(dtype*)p;
+        y = c_mul(x,y);
+        p += stride;
+    }
+    return y;
+}
+
+static inline dtype f_prod_nan(size_t n, char *p, ssize_t stride)
+{
+    size_t i=n;
+    dtype x,y;
+
+    y = c_one();
+    for (; i--;) {
+        x = *(dtype*)p;
+        if (not_nan(x)) {
+            y = c_mul(x,y);
+        }
+        p += stride;
+    }
+    return y;
+}
+
+static inline dtype f_mean(size_t n, char *p, ssize_t stride)
+{
+    size_t i=n;
+    size_t count=0;
+    dtype x,y;
+
+    y = c_zero();
+    for (; i--;) {
+        x = *(dtype*)p;
+        y = c_add(x,y);
+        count++;
+        p += stride;
+    }
+    return c_div_r(y,count);
+}
+
+static inline dtype f_mean_nan(size_t n, char *p, ssize_t stride)
+{
+    size_t i=n;
+    size_t count=0;
+    dtype x,y;
+
+    y = c_zero();
+    for (; i--;) {
+        x = *(dtype*)p;
+        if (not_nan(x)) {
+            y = c_add(x,y);
+            count++;
+        }
+        p += stride;
+    }
+    return c_div_r(y,count);
+}
+
+static inline rtype f_var(size_t n, char *p, ssize_t stride)
+{
+    size_t i=n;
+    size_t count=0;
+    dtype x,m;
+    rtype y=0;
+
+    m = f_mean(n,p,stride);
+
+    for (; i--;) {
+        x = *(dtype*)p;
+        y += c_abs_square(c_sub(x,m));
+        count++;
+        p += stride;
+    }
+    return y/(count-1);
+}
+
+static inline rtype f_var_nan(size_t n, char *p, ssize_t stride)
+{
+    size_t i=n;
+    size_t count=0;
+    dtype x,m;
+    rtype y=0;
+
+    m = f_mean_nan(n,p,stride);
+
+    for (; i--;) {
+        x = *(dtype*)p;
+        if (not_nan(x)) {
+            y += c_abs_square(c_sub(x,m));
+            count++;
+        }
+        p += stride;
+    }
+    return y/(count-1);
+}
+
+static inline rtype f_stddev(size_t n, char *p, ssize_t stride)
+{
+    return r_sqrt(f_var(n,p,stride));
+}
+
+static inline rtype f_stddev_nan(size_t n, char *p, ssize_t stride)
+{
+    return r_sqrt(f_var_nan(n,p,stride));
+}
+
+static inline rtype f_rms(size_t n, char *p, ssize_t stride)
+{
+    size_t i=n;
+    size_t count=0;
+    dtype x;
+    rtype y=0;
+
+    for (; i--;) {
+        x = *(dtype*)p;
+        y += c_abs_square(x);
+        count++;
+        p += stride;
+    }
+    return r_sqrt(y/count);
+}
+
+static inline rtype f_rms_nan(size_t n, char *p, ssize_t stride)
+{
+    size_t i=n;
+    size_t count=0;
+    dtype x;
+    rtype y=0;
+
+    for (; i--;) {
+        x = *(dtype*)p;
+        if (not_nan(x)) {
+            y += c_abs_square(x);
+            count++;
+        }
+        p += stride;
+    }
+    return r_sqrt(y/count);
+}
+
+static inline dtype f_seq(dtype x, dtype y, double c)
+{
+    return c_add(x,c_mul_r(y,c));
+}

data/lib/dnn/ext/cifar10/numo/types/dcomplex.h

@@ -0,0 +1,44 @@
+typedef dcomplex dtype;
+typedef double rtype;
+#define cT  numo_cDComplex
+#define cRT numo_cDFloat
+#define mTM numo_mDComplexMath
+
+#include "complex_macro.h"
+
+static inline bool c_nearly_eq(dtype x, dtype y) {
+    return c_abs(c_sub(x,y)) <= (c_abs(x)+c_abs(y))*DBL_EPSILON*2;
+}
+
+#ifdef SFMT_H
+/* generates a random number on [0,1)-real-interval */
+inline static dtype m_rand(dtype max)
+{
+    dtype z;
+    REAL(z) = genrand_res53_mix() * REAL(max);
+    IMAG(z) = genrand_res53_mix() * IMAG(max);
+    return z;
+}
+
+/* generates random numbers from the normal distribution
+   using Box-Muller Transformation.
+ */
+inline static void m_rand_norm(dtype mu, rtype sigma, dtype *a0)
+{
+    rtype x1, x2, w;
+    do {
+	x1 = genrand_res53_mix();
+	x1 = x1*2-1;
+	x2 = genrand_res53_mix();
+	x2 = x2*2-1;
+	w = x1 * x1 + x2 * x2;
+    } while (w>=1);
+    w = sqrt( (-2*log(w)) / w );
+    REAL(*a0) = x1*w * sigma + REAL(mu);
+    IMAG(*a0) = x2*w * sigma + IMAG(mu);
+}
+#endif
+
+#define M_EPSILON rb_float_new(2.2204460492503131e-16)
+#define M_MIN     rb_float_new(2.2250738585072014e-308)
+#define M_MAX     rb_float_new(1.7976931348623157e+308)

data/lib/dnn/ext/cifar10/numo/types/dfloat.h

@@ -0,0 +1,42 @@
+typedef double dtype;
+typedef double rtype;
+#define cT  numo_cDFloat
+#define cRT numo_cDFloat
+#define mTM numo_mDFloatMath
+
+#include "float_macro.h"
+
+#ifdef SFMT_H
+/* generates a random number on [0,1)-real-interval */
+inline static dtype m_rand(dtype max)
+{
+    return genrand_res53_mix() * max;
+}
+
+/* generates random numbers from the normal distribution
+   using Box-Muller Transformation.
+ */
+inline static void m_rand_norm(dtype mu, dtype sigma, dtype *a0, dtype *a1)
+{
+    dtype x1, x2, w;
+    do {
+	x1 = genrand_res53_mix();
+	x1 = x1*2-1;
+	x2 = genrand_res53_mix();
+	x2 = x2*2-1;
+	w = x1 * x1 + x2 * x2;
+    } while (w>=1);
+    w = sqrt( (-2*log(w)) / w );
+    if (a0) {*a0 = x1*w * sigma + mu;}
+    if (a1) {*a1 = x2*w * sigma + mu;}
+}
+#endif
+
+#define m_min_init numo_dfloat_new_dim0(0.0/0.0)
+#define m_max_init numo_dfloat_new_dim0(0.0/0.0)
+#define m_extract(x) rb_float_new(*(double*)x)
+#define m_nearly_eq(x,y) (fabs(x-y)<=(fabs(x)+fabs(y))*DBL_EPSILON*2)
+
+#define M_EPSILON rb_float_new(2.2204460492503131e-16)
+#define M_MIN     rb_float_new(2.2250738585072014e-308)
+#define M_MAX     rb_float_new(1.7976931348623157e+308)