Skip to content

Commit 5fcd0c3

Browse files
gabyflegabyfle
committed
owl_core preprocessor for cpp compiling.
1 parent e9a1f08 commit 5fcd0c3

File tree

3 files changed

+161
-157
lines changed

3 files changed

+161
-157
lines changed

‎src/owl/core/owl_core.h‎

Lines changed: 135 additions & 123 deletions
Original file line numberDiff line numberDiff line change
@@ -11,220 +11,232 @@
1111
#include "owl_ndarray_contract.h"
1212
#include <stdio.h> // DEBUG
1313

14+
#ifdef __cplusplus
15+
extern "C"
16+
{
17+
#endif
1418

15-
/** Core function declaration **/
19+
/** Core function declaration **/
1620

21+
extern int64_t c_ndarray_numel(struct caml_ba_array *X);
1722

18-
extern int64_t c_ndarray_numel (struct caml_ba_array *X);
23+
extern int64_t c_ndarray_stride_dim(struct caml_ba_array *X, int d);
1924

20-
extern int64_t c_ndarray_stride_dim (struct caml_ba_array *X, int d);
25+
extern int64_t c_ndarray_slice_dim(struct caml_ba_array *X, int d);
2126

22-
extern int64_t c_ndarray_slice_dim (struct caml_ba_array *X, int d);
27+
extern void c_float32_ndarray_transpose(struct slice_pair *sp);
2328

24-
extern void c_float32_ndarray_transpose (struct slice_pair *sp);
29+
extern void c_float64_ndarray_transpose(struct slice_pair *sp);
2530

26-
extern void c_float64_ndarray_transpose (struct slice_pair *sp);
31+
extern void c_complex32_ndarray_transpose(struct slice_pair *sp);
2732

28-
extern void c_complex32_ndarray_transpose (struct slice_pair *sp);
33+
extern void c_complex64_ndarray_transpose(struct slice_pair *sp);
2934

30-
extern void c_complex64_ndarray_transpose (struct slice_pair *sp);
35+
extern void c_float32_ndarray_contract_one(struct contract_pair *sp);
3136

32-
extern void c_float32_ndarray_contract_one (struct contract_pair *sp);
37+
extern void c_float64_ndarray_contract_one(struct contract_pair *sp);
3338

34-
extern void c_float64_ndarray_contract_one (struct contract_pair *sp);
39+
extern void c_complex32_ndarray_contract_one(struct contract_pair *sp);
3540

36-
extern void c_complex32_ndarray_contract_one (struct contract_pair *sp);
41+
extern void c_complex64_ndarray_contract_one(struct contract_pair *sp);
3742

38-
extern void c_complex64_ndarray_contract_one (struct contract_pair *sp);
43+
extern void c_float32_ndarray_contract_two(struct contract_pair *sp);
3944

40-
extern void c_float32_ndarray_contract_two (struct contract_pair *sp);
45+
extern void c_float64_ndarray_contract_two(struct contract_pair *sp);
4146

42-
extern void c_float64_ndarray_contract_two (struct contract_pair *sp);
47+
extern void c_complex32_ndarray_contract_two(struct contract_pair *sp);
4348

44-
extern void c_complex32_ndarray_contract_two (struct contract_pair *sp);
49+
extern void c_complex64_ndarray_contract_two(struct contract_pair *sp);
4550

46-
extern void c_complex64_ndarray_contract_two (struct contract_pair *sp);
51+
extern void c_float32_matrix_swap_rows(float *x, int m, int n, int i, int j);
4752

48-
extern void c_float32_matrix_swap_rows (float *x, int m, int n, int i, int j);
53+
extern void c_float64_matrix_swap_rows(double *x, int m, int n, int i, int j);
4954

50-
extern void c_float64_matrix_swap_rows (double *x, int m, int n, int i, int j);
55+
extern void c_complex32_matrix_swap_rows(_Complex float *x, int m, int n, int i, int j);
5156

52-
extern void c_complex32_matrix_swap_rows (_Complex float *x, int m, int n, int i, int j);
57+
extern void c_complex64_matrix_swap_rows(_Complex double *x, int m, int n, int i, int j);
5358

54-
extern void c_complex64_matrix_swap_rows (_Complex double *x, int m, int n, int i, int j);
59+
extern void c_float32_matrix_swap_cols(float *x, int m, int n, int i, int j);
5560

56-
extern void c_float32_matrix_swap_cols (float *x, int m, int n, int i, int j);
61+
extern void c_float64_matrix_swap_cols(double *x, int m, int n, int i, int j);
5762

58-
extern void c_float64_matrix_swap_cols (double *x, int m, int n, int i, int j);
63+
extern void c_complex32_matrix_swap_cols(_Complex float *x, int m, int n, int i, int j);
5964

60-
extern void c_complex32_matrix_swap_cols (_Complex float *x, int m, int n, int i, int j);
65+
extern void c_complex64_matrix_swap_cols(_Complex double *x, int m, int n, int i, int j);
6166

62-
extern void c_complex64_matrix_swap_cols (_Complex double *x, int m, int n, int i, int j);
67+
extern void c_float32_matrix_transpose(float *x, float *y, int m, int n);
6368

64-
extern void c_float32_matrix_transpose (float *x, float *y, int m, int n);
69+
extern void c_float64_matrix_transpose(double *x, double *y, int m, int n);
6570

66-
extern void c_float64_matrix_transpose (double *x, double *y, int m, int n);
71+
extern void c_complex32_matrix_transpose(_Complex float *x, _Complex float *y, int m, int n);
6772

68-
extern void c_complex32_matrix_transpose (_Complex float *x, _Complex float *y, int m, int n);
73+
extern void c_complex64_matrix_transpose(_Complex double *x, _Complex double *y, int m, int n);
6974

70-
extern void c_complex64_matrix_transpose (_Complex double *x, _Complex double *y, int m, int n);
75+
extern void c_ndarray_stride(struct caml_ba_array *X, int64_t *stride);
7176

72-
extern void c_ndarray_stride (struct caml_ba_array *X, int64_t *stride);
77+
extern void c_ndarray_slice(struct caml_ba_array *X, int64_t *slice);
7378

74-
extern void c_ndarray_slice (struct caml_ba_array *X, int64_t *slice);
79+
extern void c_slicing_stride(struct caml_ba_array *X, int64_t *slice, int64_t *stride);
7580

76-
extern void c_slicing_stride (struct caml_ba_array *X, int64_t *slice, int64_t *stride);
81+
extern void c_slicing_offset(struct caml_ba_array *X, int64_t *slice, int64_t *offset);
7782

78-
extern void c_slicing_offset (struct caml_ba_array *X, int64_t *slice, int64_t *offset);
83+
extern void c_float32_ndarray_get_slice(struct slice_pair *sp);
7984

80-
extern void c_float32_ndarray_get_slice (struct slice_pair *sp);
85+
extern void c_float64_ndarray_get_slice(struct slice_pair *sp);
8186

82-
extern void c_float64_ndarray_get_slice (struct slice_pair *sp);
87+
extern void c_complex32_ndarray_get_slice(struct slice_pair *sp);
8388

84-
extern void c_complex32_ndarray_get_slice (struct slice_pair *sp);
89+
extern void c_complex64_ndarray_get_slice(struct slice_pair *sp);
8590

86-
extern void c_complex64_ndarray_get_slice (struct slice_pair *sp);
91+
extern void c_float32_ndarray_set_slice(struct slice_pair *sp);
8792

88-
extern void c_float32_ndarray_set_slice (struct slice_pair *sp);
93+
extern void c_float64_ndarray_set_slice(struct slice_pair *sp);
8994

90-
extern void c_float64_ndarray_set_slice (struct slice_pair *sp);
95+
extern void c_complex32_ndarray_set_slice(struct slice_pair *sp);
9196

92-
extern void c_complex32_ndarray_set_slice (struct slice_pair *sp);
97+
extern void c_complex64_ndarray_set_slice(struct slice_pair *sp);
9398

94-
extern void c_complex64_ndarray_set_slice (struct slice_pair *sp);
99+
extern void c_float32_ndarray_get_fancy(struct fancy_pair *sp);
95100

96-
extern void c_float32_ndarray_get_fancy (struct fancy_pair *sp);
101+
extern void c_float64_ndarray_get_fancy(struct fancy_pair *sp);
97102

98-
extern void c_float64_ndarray_get_fancy (struct fancy_pair *sp);
103+
extern void c_complex32_ndarray_get_fancy(struct fancy_pair *sp);
99104

100-
extern void c_complex32_ndarray_get_fancy (struct fancy_pair *sp);
105+
extern void c_complex64_ndarray_get_fancy(struct fancy_pair *sp);
101106

102-
extern void c_complex64_ndarray_get_fancy (struct fancy_pair *sp);
107+
extern void c_float32_ndarray_set_fancy(struct fancy_pair *sp);
103108

104-
extern void c_float32_ndarray_set_fancy (struct fancy_pair *sp);
109+
extern void c_float64_ndarray_set_fancy(struct fancy_pair *sp);
105110

106-
extern void c_float64_ndarray_set_fancy (struct fancy_pair *sp);
111+
extern void c_complex32_ndarray_set_fancy(struct fancy_pair *sp);
107112

108-
extern void c_complex32_ndarray_set_fancy (struct fancy_pair *sp);
113+
extern void c_complex64_ndarray_set_fancy(struct fancy_pair *sp);
109114

110-
extern void c_complex64_ndarray_set_fancy (struct fancy_pair *sp);
115+
// compare two numbers (real & complex & int)
111116

117+
#define CEQF(X, Y) ((crealf(X) == crealf(Y)) && (cimagf(X) == cimagf(Y)))
112118

113-
// compare two numbers (real & complex & int)
119+
#define CEQ(X, Y) ((creal(X) == creal(Y)) && (cimag(X) == cimag(Y)))
114120

115-
#define CEQF(X,Y) ((crealf(X) == crealf(Y)) && (cimagf(X) == cimagf(Y)))
121+
#define CNEQF(X, Y) ((crealf(X) != crealf(Y)) || (cimagf(X) != cimagf(Y)))
116122

117-
#define CEQ(X,Y) ((creal(X) == creal(Y)) && (cimag(X) == cimag(Y)))
123+
#define CNEQ(X, Y) ((creal(X) != creal(Y)) || (cimag(X) != cimag(Y)))
118124

119-
#define CNEQF(X,Y) ((crealf(X) != crealf(Y)) || (cimagf(X) != cimagf(Y)))
125+
#define CLTF(X, Y) ((cabsf(X) < cabsf(Y)) || ((cabsf(X) == cabsf(Y)) && (cargf(X) < cargf(Y))))
120126

121-
#define CNEQ(X,Y) ((creal(X) != creal(Y)) || (cimag(X) != cimag(Y)))
127+
#define CGTF(X, Y) ((cabsf(X) > cabsf(Y)) || ((cabsf(X) == cabsf(Y)) && (cargf(X) > cargf(Y))))
122128

123-
#define CLTF(X,Y) ((cabsf(X) < cabsf(Y)) || ((cabsf(X) == cabsf(Y)) && (cargf(X) < cargf(Y))))
129+
#define CLEF(X, Y) !CGTF(X, Y)
124130

125-
#define CGTF(X,Y) ((cabsf(X) > cabsf(Y)) || ((cabsf(X) == cabsf(Y)) && (cargf(X) > cargf(Y))))
131+
#define CGEF(X, Y) !CLTF(X, Y)
126132

127-
#define CLEF(X,Y) !CGTF(X,Y)
133+
#define CLT(X, Y) ((cabs(X) < cabs(Y)) || ((cabs(X) == cabs(Y)) && (carg(X) < carg(Y))))
128134

129-
#define CGEF(X,Y) !CLTF(X,Y)
135+
#define CGT(X, Y) ((cabs(X) > cabs(Y)) || ((cabs(X) == cabs(Y)) && (carg(X) > carg(Y))))
130136

131-
#define CLT(X,Y) ((cabs(X) < cabs(Y)) || ((cabs(X) == cabs(Y)) && (carg(X) < carg(Y))))
137+
#define CLE(X, Y) !CGT(X, Y)
132138

133-
#define CGT(X,Y) ((cabs(X) > cabs(Y)) || ((cabs(X) == cabs(Y)) && (carg(X) > carg(Y))))
139+
#define CGE(X, Y) !CLT(X, Y)
134140

135-
#define CLE(X,Y) !CGT(X,Y)
141+
extern int float32_cmp(const void *a, const void *b);
136142

137-
#define CGE(X,Y) !CLT(X,Y)
143+
extern int float64_cmp(const void *a, const void *b);
138144

139-
extern int float32_cmp (const void * a, const void * b);
145+
extern int complex32_cmp(const void *a, const void *b);
140146

141-
extern int float64_cmp (const void * a, const void * b);
147+
extern int complex64_cmp(const void *a, const void *b);
142148

143-
extern int complex32_cmp (const void * a, const void * b);
149+
extern int int8_cmp(const void *a, const void *b);
144150

145-
extern int complex64_cmp (const void * a, const void * b);
151+
extern int uint8_cmp(const void *a, const void *b);
146152

147-
extern int int8_cmp (const void * a, const void * b);
153+
extern int int16_cmp(const void *a, const void *b);
148154

149-
extern int uint8_cmp (const void * a, const void * b);
155+
extern int uint16_cmp(const void *a, const void *b);
150156

151-
extern int int16_cmp (const void * a, const void * b);
157+
extern int int32_cmp(const void *a, const void *b);
152158

153-
extern int uint16_cmp (const void * a, const void * b);
159+
extern int int64_cmp(const void *a, const void *b);
154160

155-
extern int int32_cmp (const void * a, const void * b);
161+
extern int float32_cmp_r(const void *a, const void *b, const void *z);
156162

157-
extern int int64_cmp (const void * a, const void * b);
163+
extern int float64_cmp_r(const void *a, const void *b, const void *z);
158164

159-
extern int float32_cmp_r (const void * a, const void * b, const void * z);
165+
extern int complex32_cmp_r(const void *a, const void *b, const void *z);
160166

161-
extern int float64_cmp_r (const void * a, const void * b, const void * z);
167+
extern int complex64_cmp_r(const void *a, const void *b, const void *z);
162168

163-
extern int complex32_cmp_r (const void * a, const void * b, const void * z);
169+
extern int int8_cmp_r(const void *a, const void *b, const void *z);
164170

165-
extern int complex64_cmp_r (const void * a, const void * b, const void * z);
171+
extern int uint8_cmp_r(const void *a, const void *b, const void *z);
166172

167-
extern int int8_cmp_r (const void * a, const void * b, const void * z);
173+
extern int int16_cmp_r(const void *a, const void *b, const void *z);
168174

169-
extern int uint8_cmp_r (const void * a, const void * b, const void * z);
175+
extern int uint16_cmp_r(const void *a, const void *b, const void *z);
170176

171-
extern int int16_cmp_r (const void * a, const void * b, const void * z);
177+
extern int int32_cmp_r(const void *a, const void *b, const void *z);
172178

173-
extern int uint16_cmp_r (const void * a, const void * b, const void * z);
179+
extern int int64_cmp_r(const void *a, const void *b, const void *z);
174180

175-
extern int int32_cmp_r (const void * a, const void * b, const void * z);
181+
// acquire CPU cache sizes
176182

177-
extern int int64_cmp_r (const void * a, const void * b, const void * z);
183+
extern void query_cache_sizes(int *l1p, int *l2p, int *l3p);
178184

185+
// copy two double type numbers, for interfacing to foreign functions
186+
OWL_INLINE value cp_two_doubles(double d0, double d1)
187+
{
188+
value res = caml_alloc_small(2 * Double_wosize, Double_array_tag);
189+
Store_double_field(res, 0, d0);
190+
Store_double_field(res, 1, d1);
191+
return res;
192+
}
179193

180-
// acquire CPU cache sizes
181-
182-
extern void query_cache_sizes(int* l1p, int* l2p, int* l3p);
183-
184-
185-
// copy two double type numbers, for interfacing to foreign functions
186-
OWL_INLINE value cp_two_doubles(double d0, double d1) {
187-
value res = caml_alloc_small(2 * Double_wosize, Double_array_tag);
188-
Store_double_field(res, 0, d0);
189-
Store_double_field(res, 1, d1);
190-
return res;
191-
}
192-
193-
// copy x to y with given offset and stride
194-
OWL_INLINE void owl_float32_copy (int N, float* x, int ofsx, int incx, float* y, int ofsy, int incy) {
195-
for (int i = 0; i < N; i++) {
196-
*(y + ofsy) = *(x + ofsx);
197-
ofsx += incx;
198-
ofsy += incy;
194+
// copy x to y with given offset and stride
195+
OWL_INLINE void owl_float32_copy(int N, float *x, int ofsx, int incx, float *y, int ofsy, int incy)
196+
{
197+
for (int i = 0; i < N; i++)
198+
{
199+
*(y + ofsy) = *(x + ofsx);
200+
ofsx += incx;
201+
ofsy += incy;
202+
}
199203
}
200-
}
201204

202-
// copy x to y with given offset and stride
203-
OWL_INLINE void owl_float64_copy (int N, double* x, int ofsx, int incx, double* y, int ofsy, int incy) {
204-
for (int i = 0; i < N; i++) {
205-
*(y + ofsy) = *(x + ofsx);
206-
ofsx += incx;
207-
ofsy += incy;
205+
// copy x to y with given offset and stride
206+
OWL_INLINE void owl_float64_copy(int N, double *x, int ofsx, int incx, double *y, int ofsy, int incy)
207+
{
208+
for (int i = 0; i < N; i++)
209+
{
210+
*(y + ofsy) = *(x + ofsx);
211+
ofsx += incx;
212+
ofsy += incy;
213+
}
208214
}
209-
}
210215

211-
// copy x to y with given offset and stride
212-
OWL_INLINE void owl_complex32_copy (int N, _Complex float* x, int ofsx, int incx, _Complex float* y, int ofsy, int incy) {
213-
for (int i = 0; i < N; i++) {
214-
*(y + ofsy) = *(x + ofsx);
215-
ofsx += incx;
216-
ofsy += incy;
216+
// copy x to y with given offset and stride
217+
OWL_INLINE void owl_complex32_copy(int N, _Complex float *x, int ofsx, int incx, _Complex float *y, int ofsy, int incy)
218+
{
219+
for (int i = 0; i < N; i++)
220+
{
221+
*(y + ofsy) = *(x + ofsx);
222+
ofsx += incx;
223+
ofsy += incy;
224+
}
217225
}
218-
}
219226

220-
// copy x to y with given offset and stride
221-
OWL_INLINE void owl_complex64_copy (int N, _Complex double* x, int ofsx, int incx, _Complex double* y, int ofsy, int incy) {
222-
for (int i = 0; i < N; i++) {
223-
*(y + ofsy) = *(x + ofsx);
224-
ofsx += incx;
225-
ofsy += incy;
227+
// copy x to y with given offset and stride
228+
OWL_INLINE void owl_complex64_copy(int N, _Complex double *x, int ofsx, int incx, _Complex double *y, int ofsy, int incy)
229+
{
230+
for (int i = 0; i < N; i++)
231+
{
232+
*(y + ofsy) = *(x + ofsx);
233+
ofsx += incx;
234+
ofsy += incy;
235+
}
226236
}
227-
}
228237

238+
#ifdef __cplusplus
239+
}
240+
#endif
229241

230-
#endif /* OWL_CORE_H */
242+
#endif /* OWL_CORE_H */

0 commit comments

Comments
 (0)