cumo 0.1.0

data/ext/cumo/include/cumo/types/complex_macro_kernel.h

@@ -0,0 +1,186 @@
+#ifndef CUMO_COMPLEX_MACRO_KERNEL_H
+#define CUMO_COMPLEX_MACRO_KERNEL_H
+
+#include "float_def_kernel.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_kernel.h"
+
+__host__ __device__ static inline dtype c_from_scomplex(scomplex x) {
+    dtype z;
+    REAL(z) = REAL(x);
+    IMAG(z) = IMAG(x);
+    return z;
+}
+
+__host__ __device__ 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_sint(x)   c_new(x,0)
+#define m_from_int32(x)  c_new(x,0)
+#define m_from_int64(x)  c_new(x,0)
+#define m_from_uint32(x) c_new(x,0)
+#define m_from_uint64(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 0
+#define m_mulsum_init 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);    \
+        }}
+
+__host__ __device__ static inline dtype f_seq(dtype x, dtype y, double c)
+{
+    return c_add(x,c_mul_r(y,c));
+}
+
+/* --------- thrust ----------------- */
+#include "cumo/cuda/cumo_thrust_complex.hpp"
+
+struct cumo_thrust_plus : public thrust::binary_function<dtype, dtype, dtype>
+{
+    __host__ __device__ dtype operator()(dtype x, dtype y) { return m_add(x,y); }
+};
+
+struct cumo_thrust_multiplies : public thrust::binary_function<dtype, dtype, dtype>
+{
+    __host__ __device__ dtype operator()(dtype x, dtype y) { return m_mul(x,y); }
+};
+
+struct cumo_thrust_multiplies_mulsum_nan : public thrust::binary_function<dtype, dtype, dtype>
+{
+    __host__ __device__ dtype operator()(dtype x, dtype y) {
+        if (not_nan(x) && not_nan(y)) {
+            return m_mul(x, y);
+        } else {
+            return m_zero;
+        }
+    }
+};
+
+struct cumo_thrust_square : public thrust::unary_function<dtype, dtype>
+{
+    __host__ __device__ rtype operator()(const dtype& x) const { return c_abs_square(x); }
+};
+
+#endif // CUMO_COMPLEX_MACRO_KERNEL_H

data/ext/cumo/include/cumo/types/dcomplex.h

@@ -0,0 +1,46 @@
+typedef dcomplex dtype;
+typedef double rtype;
+#define cT  cumo_cDComplex
+#define cRT cumo_cDFloat
+#define mTM cumo_mDComplexMath
+
+#include "complex_macro.h"
+#include "cublas_v2.h"
+#include "cumo/cuda/cublas.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/ext/cumo/include/cumo/types/dcomplex_kernel.h

@@ -0,0 +1,13 @@
+#ifndef CUMO_DCOMPLEX_KERNEL_H
+#define CUMO_DCOMPLEX_KERNEL_H
+
+typedef dcomplex dtype;
+typedef double rtype;
+
+#include "complex_macro_kernel.h"
+
+__device__ 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;
+}
+
+#endif // CUMO_DCOMPLEX_KERNEL_H

data/ext/cumo/include/cumo/types/dfloat.h

@@ -0,0 +1,47 @@
+typedef double dtype;
+typedef double rtype;
+#define cT  cumo_cDFloat
+#define cRT cumo_cDFloat
+#define mTM cumo_mDFloatMath
+
+#include "float_macro.h"
+#include "cublas_v2.h"
+#include "cumo/cuda/cublas.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 cumo_dfloat_new_dim0(0.0/0.0)
+#define m_max_init cumo_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)
+
+#define DATA_MIN DBL_MIN
+#define DATA_MAX DBL_MAX

data/ext/cumo/include/cumo/types/dfloat_kernel.h

@@ -0,0 +1,14 @@
+#ifndef CUMO_DFLOAT_KERNEL_H
+#define CUMO_DFLOAT_KERNEL_H
+
+typedef double dtype;
+typedef double rtype;
+
+#include "float_macro_kernel.h"
+
+#define m_nearly_eq(x,y) (fabs(x-y)<=(fabs(x)+fabs(y))*DBL_EPSILON*2)
+
+#define DATA_MIN DBL_MIN
+#define DATA_MAX DBL_MAX
+
+#endif // CUMO_DFLOAT_KERNEL_H

data/ext/cumo/include/cumo/types/float_def.h

@@ -0,0 +1,34 @@
+#ifndef DBL_EPSILON
+#define DBL_EPSILON 2.2204460492503131e-16
+#endif
+#ifndef FLT_EPSILON
+#define FLT_EPSILON 1.1920928955078125e-07
+#endif
+#ifndef DBL_MAX
+#define DBL_MAX 1.7976931348623157e+308
+#endif
+#ifndef DBL_MAX
+#define DBL_MAX 1.7976931348623157e+308
+#endif
+#ifndef FLT_MIN
+#define FLT_MIN 1.1754943508222875e-38
+#endif
+#ifndef FLT_MAX
+#define FLT_MAX 3.4028234663852886e+38
+#endif
+
+#ifndef M_PI_2
+#define M_PI_2         1.57079632679489661923  /* pi/2 */
+#endif
+#ifndef M_LOG2E
+#define M_LOG2E        1.4426950408889634074   /* log_2 e */
+#endif
+#ifndef M_LOG10E
+#define M_LOG10E       0.43429448190325182765  /* log_10 e */
+#endif
+#ifndef M_LN2
+#define M_LN2          0.69314718055994530942  /* log_e 2 */
+#endif
+#ifndef M_LN10
+#define M_LN10         2.30258509299404568402  /* log_e 10 */
+#endif

data/ext/cumo/include/cumo/types/float_def_kernel.h

@@ -0,0 +1,39 @@
+#ifndef CUMO_FLOAT_DEF_KERNEL_H
+#define CUMO_FLOAT_DEF_KERNEL_H
+
+#ifndef DBL_EPSILON
+#define DBL_EPSILON 2.2204460492503131e-16
+#endif
+#ifndef FLT_EPSILON
+#define FLT_EPSILON 1.1920928955078125e-07
+#endif
+#ifndef DBL_MAX
+#define DBL_MAX 1.7976931348623157e+308
+#endif
+#ifndef DBL_MAX
+#define DBL_MAX 1.7976931348623157e+308
+#endif
+#ifndef FLT_MIN
+#define FLT_MIN 1.1754943508222875e-38
+#endif
+#ifndef FLT_MAX
+#define FLT_MAX 3.4028234663852886e+38
+#endif
+
+#ifndef M_PI_2
+#define M_PI_2         1.57079632679489661923  /* pi/2 */
+#endif
+#ifndef M_LOG2E
+#define M_LOG2E        1.4426950408889634074   /* log_2 e */
+#endif
+#ifndef M_LOG10E
+#define M_LOG10E       0.43429448190325182765  /* log_10 e */
+#endif
+#ifndef M_LN2
+#define M_LN2          0.69314718055994530942  /* log_e 2 */
+#endif
+#ifndef M_LN10
+#define M_LN10         2.30258509299404568402  /* log_e 10 */
+#endif
+
+#endif // CUMO_FLOAT_DEF_KERNEL_H

data/ext/cumo/include/cumo/types/float_macro.h

@@ -0,0 +1,191 @@
+#include "float_def.h"
+
+extern double round(double);
+extern double log2(double);
+extern double exp2(double);
+#ifdef HAVE_EXP10
+extern double exp10(double);
+#else
+extern double pow(double, double);
+#endif
+
+#define m_zero 0.0
+#define m_one  1.0
+
+#define m_num_to_data(x) NUM2DBL(x)
+#define m_data_to_num(x) rb_float_new(x)
+
+#define m_from_double(x) (x)
+#define m_from_real(x) (x)
+#define m_from_sint(x) (x)
+#define m_from_int32(x) (x)
+#define m_from_int64(x) (x)
+#define m_from_uint32(x) (x)
+#define m_from_uint64(x) (x)
+
+#define m_add(x,y) ((x)+(y))
+#define m_sub(x,y) ((x)-(y))
+#define m_mul(x,y) ((x)*(y))
+#define m_div(x,y) ((x)/(y))
+#define m_div_check(x,y) ((y)==0)
+#define m_mod(x,y) fmod(x,y)
+#define m_divmod(x,y,a,b) {a=(x)/(y); b=m_mod(x,y);}
+#define m_pow(x,y) pow(x,y)
+#define m_pow_int(x,y) pow_int(x,y)
+
+#define m_abs(x)     fabs(x)
+#define m_minus(x)   (-(x))
+#define m_reciprocal(x) (1/(x))
+#define m_square(x)  ((x)*(x))
+#define m_floor(x)   floor(x)
+#define m_round(x)   round(x)
+#define m_ceil(x)    ceil(x)
+#define m_trunc(x)   trunc(x)
+#define m_rint(x)    rint(x)
+#define m_sign(x)    (((x)==0) ? 0.0:(((x)>0) ? 1.0:(((x)<0) ? -1.0:(x))))
+#define m_copysign(x,y) copysign(x,y)
+#define m_signbit(x) signbit(x)
+#define m_modf(x,y,z) {double d; y=modf(x,&d); z=d;}
+
+#define m_eq(x,y) ((x)==(y))
+#define m_ne(x,y) ((x)!=(y))
+#define m_gt(x,y) ((x)>(y))
+#define m_ge(x,y) ((x)>=(y))
+#define m_lt(x,y) ((x)<(y))
+#define m_le(x,y) ((x)<=(y))
+
+#define m_isnan(x) isnan(x)
+#define m_isinf(x) isinf(x)
+#define m_isposinf(x) (isinf(x) && signbit(x)==0)
+#define m_isneginf(x) (isinf(x) && signbit(x))
+#define m_isfinite(x) isfinite(x)
+
+#define m_mulsum_init INT2FIX(0)
+
+#define m_sprintf(s,x) sprintf(s,"%g",x)
+
+#define cmp_prnan(a,b)                         \
+    ((qsort_cast(a)==qsort_cast(b)) ? 0 :      \
+     (qsort_cast(a) > qsort_cast(b)) ? 1 : -1)
+
+#define cmp_ignan(a,b)                                                  \
+    (m_isnan(qsort_cast(a)) ? (m_isnan(qsort_cast(b)) ? 0 : 1) :        \
+     (m_isnan(qsort_cast(b)) ? -1 :                                     \
+      ((qsort_cast(a)==qsort_cast(b)) ? 0 :                             \
+       (qsort_cast(a) > qsort_cast(b)) ? 1 : -1)))
+
+#define cmpgt_prnan(a,b)                        \
+    (qsort_cast(a) > qsort_cast(b))
+
+#define cmpgt_ignan(a,b)                                      \
+    ((m_isnan(qsort_cast(a)) && !m_isnan(qsort_cast(b))) ||   \
+     (qsort_cast(a) > qsort_cast(b)))
+
+#define m_sqrt(x)    sqrt(x)
+#define m_cbrt(x)    cbrt(x)
+#define m_log(x)     log(x)
+#define m_log2(x)    log2(x)
+#define m_log10(x)   log10(x)
+#define m_exp(x)     exp(x)
+#define m_exp2(x)    exp2(x)
+#ifdef HAVE_EXP10
+#define m_exp10(x)   exp10(x)
+#else
+#define m_exp10(x)   pow(10, x)
+#endif
+#define m_expm1(x)   expm1(x)
+#define m_log1p(x)   log1p(x)
+
+#define m_sin(x)     sin(x)
+#define m_cos(x)     cos(x)
+#define m_tan(x)     tan(x)
+#define m_asin(x)    asin(x)
+#define m_acos(x)    acos(x)
+#define m_atan(x)    atan(x)
+#define m_sinh(x)    sinh(x)
+#define m_cosh(x)    cosh(x)
+#define m_tanh(x)    tanh(x)
+#define m_asinh(x)   asinh(x)
+#define m_acosh(x)   acosh(x)
+#define m_atanh(x)   atanh(x)
+#define m_atan2(x,y) atan2(x,y)
+#define m_hypot(x,y) hypot(x,y)
+#define m_sinc(x)    (sin(x)/(x))
+
+#define m_erf(x)     erf(x)
+#define m_erfc(x)    erfc(x)
+#define m_ldexp(x,y) ldexp(x,y)
+#define m_frexp(x,exp) frexp(x,exp)
+
+static inline dtype pow_int(dtype x, int p)
+{
+    dtype r=1;
+    switch(p) {
+    case 0: return 1;
+    case 1: return x;
+    case 2: return x*x;
+    case 3: return x*x*x;
+    case 4: x=x*x; return x*x;
+    }
+    if (p<0)  return 1/pow_int(x,-p);
+    if (p>64) return pow(x,p);
+    while (p) {
+        if (p&1) r *= x;
+        x *= x;
+        p >>= 1;
+    }
+    return r;
+}
+
+static inline dtype f_seq(dtype x, dtype y, double c)
+{
+    return x + y * c;
+}
+
+static inline dtype f_kahan_sum(size_t n, char *p, ssize_t stride)
+{
+    size_t i=n;
+    dtype x;
+    volatile dtype y=0;
+    volatile dtype t,r=0;
+
+    for (; i--;) {
+        x = *(dtype*)p;
+        p += stride;
+        if (fabs(x) > fabs(y)) {
+            dtype z=x; x=y; y=z;
+        }
+        r += x;
+        t = y;
+        y += r;
+        t = y-t;
+        r -= t;
+    }
+    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=0;
+    volatile dtype t,r=0;
+
+    for (; i--;) {
+        x = *(dtype*)p;
+        p += stride;
+        if (!m_isnan(x)) {
+            if (fabs(x) > fabs(y)) {
+                dtype z=x; x=y; y=z;
+            }
+            r += x;
+            t = y;
+            y += r;
+            t = y-t;
+            r -= t;
+        }
+    }
+    return y;
+}
+
+#include "real_accum.h"