carray 3.0.2 → 3.0.4
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- checksums.yaml +4 -4
- data/CHANGELOG.md +568 -0
- data/README.md +6 -6
- data/carray.gemspec +1 -1
- data/ext/ca_axis_descriptor.h +27 -24
- data/ext/ca_axis_dispatch.c +220 -0
- data/ext/ca_axis_group.c +201 -123
- data/ext/ca_binop_dispatch.c +46 -52
- data/ext/ca_binop_dispatch.h +1 -1
- data/ext/ca_categorical_iterator.c +249 -183
- data/ext/ca_compare.c +7 -12
- data/ext/ca_for_buffer.h +21 -68
- data/ext/ca_group_iter.c +38 -23
- data/ext/ca_kernel_iterator.c +1765 -2076
- data/ext/ca_kernel_iterator.h +584 -566
- data/ext/ca_monop_dispatch.c +86 -63
- data/ext/ca_monop_dispatch.h +5 -0
- data/ext/ca_obj_array.c +223 -56
- data/ext/ca_obj_bincmp.c +172 -138
- data/ext/ca_obj_binop.c +116 -230
- data/ext/ca_obj_bitarray.c +40 -26
- data/ext/ca_obj_bitfield.c +100 -68
- data/ext/ca_obj_block.c +19 -9
- data/ext/ca_obj_byte_swap.c +45 -46
- data/ext/ca_obj_const_string.c +269 -44
- data/ext/ca_obj_face.c +612 -89
- data/ext/ca_obj_face.h +111 -47
- data/ext/ca_obj_fake.c +36 -79
- data/ext/ca_obj_farray.c +7 -3
- data/ext/ca_obj_field.c +33 -8
- data/ext/ca_obj_fixlen_string.c +23 -24
- data/ext/ca_obj_grid.c +541 -210
- data/ext/ca_obj_meld.c +331 -111
- data/ext/ca_obj_moncmp.c +80 -65
- data/ext/ca_obj_monop.c +177 -149
- data/ext/ca_obj_object.c +664 -248
- data/ext/ca_obj_record.c +63 -23
- data/ext/ca_obj_reduce.c +18 -16
- data/ext/ca_obj_refer.c +140 -58
- data/ext/ca_obj_remap.c +85 -24
- data/ext/ca_obj_repeat.c +24 -9
- data/ext/ca_obj_roll.c +121 -16
- data/ext/ca_obj_select.c +61 -58
- data/ext/ca_obj_select_axis.c +73 -107
- data/ext/ca_obj_shift.c +32 -102
- data/ext/ca_obj_stack.c +372 -122
- data/ext/ca_obj_stride.c +348 -140
- data/ext/ca_obj_string.c +88 -22
- data/ext/ca_obj_tile.c +38 -21
- data/ext/ca_obj_time.c +45 -31
- data/ext/ca_obj_timedelta.c +43 -29
- data/ext/ca_obj_transpose.c +17 -6
- data/ext/ca_obj_triop.c +125 -141
- data/ext/ca_obj_window.c +330 -44
- data/ext/ca_op_ipower.c +24 -11
- data/ext/ca_op_powi.h +57 -30
- data/ext/ca_pow5_table.h +656 -0
- data/ext/ca_sort_kernels.h +70 -0
- data/ext/ca_sweep_engine.c +104 -81
- data/ext/ca_sweep_engine.h +7 -20
- data/ext/ca_triop_dispatch.c +4 -1
- data/ext/ca_triop_dispatch.h +5 -2
- data/ext/caframe_csv_reader.c +542 -0
- data/ext/carray.h +241 -92
- data/ext/carray_access.c +608 -276
- data/ext/carray_address_basis.c +494 -97
- data/ext/carray_attribute.c +129 -4
- data/ext/carray_bincount.c +39 -7
- data/ext/carray_broadcast.c +56 -19
- data/ext/carray_build_flags.h +1 -1
- data/ext/carray_call_cfunc.c +108 -28
- data/ext/carray_cast.c +1576 -223
- data/ext/carray_cast_func.rb +77 -24
- data/ext/carray_class.c +4 -3
- data/ext/carray_conversion.c +208 -75
- data/ext/carray_copy.c +82 -24
- data/ext/carray_core.c +646 -162
- data/ext/carray_count.c +171 -88
- data/ext/carray_element.c +94 -60
- data/ext/carray_factorize.c +670 -324
- data/ext/carray_generate.c +260 -83
- data/ext/carray_histogram.c +209 -106
- data/ext/carray_hold.c +14 -16
- data/ext/carray_index_classifier.c +422 -243
- data/ext/carray_internal.h +692 -3
- data/ext/carray_kernels_bincmp.c +7 -0
- data/ext/carray_kernels_binop.c +487 -120
- data/ext/carray_kernels_init.c +653 -50
- data/ext/carray_kernels_moncmp.c +15 -8
- data/ext/carray_kernels_monop.c +99 -43
- data/ext/carray_kernels_reduce_aggregate.c +3925 -6112
- data/ext/carray_kernels_reduce_boolean.c +112 -36
- data/ext/carray_kernels_reduce_cumulative.c +2127 -6903
- data/ext/carray_kernels_reduce_extreme.c +3032 -2892
- data/ext/carray_kernels_reduce_variance.c +439 -500
- data/ext/carray_kernels_scan.c +1097 -488
- data/ext/carray_kernels_search.c +3787 -2176
- data/ext/carray_kernels_sort.c +1200 -920
- data/ext/carray_kernels_triop.c +479 -7
- data/ext/carray_key_windows.c +216 -0
- data/ext/carray_lazy.c +967 -153
- data/ext/carray_loop.c +81 -65
- data/ext/carray_mask.c +910 -324
- data/ext/carray_math_kernel.h +1 -0
- data/ext/carray_median_percentile.c +299 -100
- data/ext/carray_memory_view.c +872 -758
- data/ext/carray_operator.c +1171 -364
- data/ext/carray_order.c +505 -133
- data/ext/carray_partition.c +112 -65
- data/ext/carray_random.c +327 -239
- data/ext/carray_round.c +273 -0
- data/ext/carray_scatter.c +346 -78
- data/ext/carray_segment.c +257 -0
- data/ext/carray_slab.c +249 -86
- data/ext/carray_sort.c +224 -227
- data/ext/carray_sort_kernel.c +8 -1
- data/ext/carray_struct.c +118 -39
- data/ext/carray_test.c +193 -69
- data/ext/carray_time_text.c +1182 -0
- data/ext/carray_undef.c +4 -0
- data/ext/carray_utils.c +579 -113
- data/ext/extconf.rb +27 -2
- data/ext/mk_call_cfunc.rb +19 -4
- data/ext/mk_pow5_table.rb +35 -0
- data/ext/mkkernel.rb +1202 -1014
- data/ext/ruby_carray.c +17 -5
- data/ext/version.h +4 -4
- data/ext/version.rb +4 -2
- data/lib/carray/attribute.rb +160 -17
- data/lib/carray/autoload_carray.rb +23 -27
- data/lib/carray/autoload_method_extension.rb +20 -5
- data/lib/carray/axis_group.rb +360 -63
- data/lib/carray/basics.rb +162 -28
- data/lib/carray/bincount_nd.rb +18 -2
- data/lib/carray/block_iterator.rb +231 -58
- data/lib/carray/boolean_reduce.rb +34 -42
- data/lib/carray/categorical.rb +242 -52
- data/lib/carray/categorical_iterator.rb +437 -827
- data/lib/carray/complex.rb +41 -34
- data/lib/carray/conditional.rb +150 -14
- data/lib/carray/const_string.rb +59 -4
- data/lib/carray/construct.rb +15 -11
- data/lib/carray/data_type_extension.rb +1 -0
- data/lib/carray/distance_limit.rb +77 -0
- data/lib/carray/fixlen_string.rb +7 -1
- data/lib/carray/frame/column_verbs.rb +70 -0
- data/lib/carray/frame/concat.rb +157 -6
- data/lib/carray/frame/convert.rb +82 -7
- data/lib/carray/frame/csv_parser.rb +688 -87
- data/lib/carray/frame/describe.rb +80 -0
- data/lib/carray/frame/frame.rb +369 -45
- data/lib/carray/frame/group.rb +631 -90
- data/lib/carray/frame/io.rb +506 -79
- data/lib/carray/frame/join.rb +411 -62
- data/lib/carray/frame/records.rb +116 -19
- data/lib/carray/frame/reshape.rb +455 -0
- data/lib/carray/frame/sort.rb +30 -18
- data/lib/carray/frame/verbs.rb +868 -61
- data/lib/carray/frame/windows.rb +117 -0
- data/lib/carray/frame.rb +4 -0
- data/lib/carray/fuse_source.rb +118 -2
- data/lib/carray/fusion.rb +257 -42
- data/lib/carray/histogram.rb +6 -1
- data/lib/carray/inspect.rb +54 -13
- data/lib/carray/iterator.rb +99 -9
- data/lib/carray/key_windows.rb +620 -0
- data/lib/carray/lazy.rb +216 -193
- data/lib/carray/mask_gap_fill.rb +112 -92
- data/lib/carray/math.rb +32 -19
- data/lib/carray/meld_reduce.rb +52 -21
- data/lib/carray/methods/align_addr.rb +8 -4
- data/lib/carray/methods/bin.rb +26 -18
- data/lib/carray/methods/bincount.rb +18 -2
- data/lib/carray/methods/bit_string.rb +15 -4
- data/lib/carray/methods/choose.rb +16 -3
- data/lib/carray/methods/composition.rb +18 -10
- data/lib/carray/methods/factorize.rb +2 -2
- data/lib/carray/methods/gather_nd.rb +25 -5
- data/lib/carray/methods/insert_block.rb +9 -14
- data/lib/carray/methods/is_between.rb +59 -0
- data/lib/carray/methods/is_in.rb +140 -10
- data/lib/carray/methods/join.rb +2 -7
- data/lib/carray/methods/lag.rb +49 -0
- data/lib/carray/methods/locate_addr.rb +182 -45
- data/lib/carray/methods/mode.rb +18 -1
- data/lib/carray/methods/nunique.rb +11 -6
- data/lib/carray/methods/repeat.rb +1 -20
- data/lib/carray/methods/resize.rb +21 -23
- data/lib/carray/methods/snap.rb +34 -12
- data/lib/carray/methods/unique.rb +11 -2
- data/lib/carray/methods/value_counts.rb +4 -11
- data/lib/carray/record_order.rb +208 -0
- data/lib/carray/rng.rb +3 -0
- data/lib/carray/runtime.rb +65 -17
- data/lib/carray/segment_iterator.rb +1221 -0
- data/lib/carray/serialize.rb +214 -30
- data/lib/carray/slab_iterator.rb +101 -27
- data/lib/carray/stack.rb +10 -4
- data/lib/carray/string.rb +105 -5
- data/lib/carray/string_operation_extension.rb +12 -4
- data/lib/carray/struct.rb +137 -58
- data/lib/carray/struct_builder.rb +124 -17
- data/lib/carray/time.rb +1222 -186
- data/lib/carray/window_iterator.rb +242 -81
- data/lib/carray.rb +9 -2
- data/yard-stubs/ca_obj_array.rb +4 -4
- data/yard-stubs/ca_obj_bitarray.rb +7 -0
- data/yard-stubs/ca_obj_bitfield.rb +23 -12
- data/yard-stubs/ca_obj_fake.rb +1 -1
- data/yard-stubs/ca_obj_field.rb +5 -1
- data/yard-stubs/ca_obj_grid.rb +21 -10
- data/yard-stubs/ca_obj_refer.rb +5 -4
- data/yard-stubs/ca_obj_roll.rb +2 -3
- data/yard-stubs/ca_obj_shift.rb +5 -1
- data/yard-stubs/ca_obj_stride.rb +10 -7
- data/yard-stubs/ca_obj_tile.rb +1 -1
- data/yard-stubs/ca_obj_transpose.rb +1 -1
- data/yard-stubs/ca_obj_window.rb +9 -5
- data/yard-stubs/carray_access.rb +11 -5
- data/yard-stubs/carray_attribute.rb +0 -5
- data/yard-stubs/carray_broadcast.rb +1 -1
- data/yard-stubs/carray_cast.rb +88 -43
- data/yard-stubs/carray_conversion.rb +15 -2
- data/yard-stubs/carray_copy.rb +12 -5
- data/yard-stubs/carray_core.rb +3 -3
- data/yard-stubs/carray_count.rb +9 -0
- data/yard-stubs/carray_element.rb +10 -4
- data/yard-stubs/carray_generate.rb +21 -1
- data/yard-stubs/carray_lazy.rb +5 -1
- data/yard-stubs/carray_loop.rb +4 -1
- data/yard-stubs/carray_mask.rb +60 -12
- data/yard-stubs/carray_math.rb +52 -14
- data/yard-stubs/carray_median_percentile.rb +15 -8
- data/yard-stubs/carray_memory_view.rb +40 -23
- data/yard-stubs/carray_order.rb +61 -29
- data/yard-stubs/carray_random.rb +13 -4
- data/yard-stubs/carray_scatter.rb +21 -9
- data/yard-stubs/carray_segment.rb +72 -0
- data/yard-stubs/carray_slab.rb +10 -2
- data/yard-stubs/carray_sort.rb +49 -8
- data/yard-stubs/carray_test.rb +15 -3
- data/yard-stubs/carray_utils.rb +0 -26
- metadata +19 -5
- data/ext/ca_for_each_element.h +0 -239
- data/ext/carray_kernels_map.c +0 -3467
- data/lib/carray/arrow_tensor.rb +0 -401
- data/lib/carray/table.rb +0 -108
data/ext/carray_kernels_init.c
CHANGED
|
@@ -3,10 +3,9 @@
|
|
|
3
3
|
|
|
4
4
|
void Init_carray_kernels_reduce_aggregate (void);
|
|
5
5
|
void Init_carray_kernels_reduce_extreme (void);
|
|
6
|
-
void Init_carray_kernels_reduce_cumulative (void);
|
|
7
6
|
void Init_carray_kernels_reduce_boolean (void);
|
|
8
7
|
void Init_carray_kernels_reduce_variance (void);
|
|
9
|
-
void
|
|
8
|
+
void Init_carray_kernels_reduce_cumulative (void);
|
|
10
9
|
void Init_carray_kernels_scan (void);
|
|
11
10
|
void Init_carray_kernels_sort (void);
|
|
12
11
|
void Init_carray_kernels_search (void);
|
|
@@ -28,6 +27,262 @@ typedef struct {
|
|
|
28
27
|
} ca_kernel_body_t;
|
|
29
28
|
|
|
30
29
|
static const ca_kernel_body_t ca_kernel_bodies[] = {
|
|
30
|
+
{ "binop", "ipow", "float32",
|
|
31
|
+
"(#3) = op_powi_<type>((#1), (#2));" },
|
|
32
|
+
{ "binop", "ipow", "float64",
|
|
33
|
+
"(#3) = op_powi_<type>((#1), (#2));" },
|
|
34
|
+
{ "binop", "ipow", "cmplx64",
|
|
35
|
+
"(#3) = op_powi_<type>((#1), (#2));" },
|
|
36
|
+
{ "binop", "ipow", "cmplx128",
|
|
37
|
+
"(#3) = op_powi_<type>((#1), (#2));" },
|
|
38
|
+
{ "monop", "cast_int8", "boolean",
|
|
39
|
+
"(#2) = (int8_t)(#1);" },
|
|
40
|
+
{ "monop", "cast_uint8", "boolean",
|
|
41
|
+
"(#2) = (uint8_t)(#1);" },
|
|
42
|
+
{ "monop", "cast_int16", "boolean",
|
|
43
|
+
"(#2) = (int16_t)(#1);" },
|
|
44
|
+
{ "monop", "cast_uint16", "boolean",
|
|
45
|
+
"(#2) = (uint16_t)(#1);" },
|
|
46
|
+
{ "monop", "cast_int32", "boolean",
|
|
47
|
+
"(#2) = (int32_t)(#1);" },
|
|
48
|
+
{ "monop", "cast_uint32", "boolean",
|
|
49
|
+
"(#2) = (uint32_t)(#1);" },
|
|
50
|
+
{ "monop", "cast_int64", "boolean",
|
|
51
|
+
"(#2) = (int64_t)(#1);" },
|
|
52
|
+
{ "monop", "cast_uint64", "boolean",
|
|
53
|
+
"(#2) = (uint64_t)(#1);" },
|
|
54
|
+
{ "monop", "cast_float32", "boolean",
|
|
55
|
+
"(#2) = (float)(#1);" },
|
|
56
|
+
{ "monop", "cast_float64", "boolean",
|
|
57
|
+
"(#2) = (double)(#1);" },
|
|
58
|
+
{ "monop", "cast_cmplx64", "boolean",
|
|
59
|
+
"(#2) = (cmplx64_t)(#1);" },
|
|
60
|
+
{ "monop", "cast_cmplx128", "boolean",
|
|
61
|
+
"(#2) = (cmplx128_t)(#1);" },
|
|
62
|
+
{ "monop", "cast_int8", "int8",
|
|
63
|
+
"(#2) = (int8_t)(#1);" },
|
|
64
|
+
{ "monop", "cast_uint8", "int8",
|
|
65
|
+
"(#2) = (uint8_t)(#1);" },
|
|
66
|
+
{ "monop", "cast_int16", "int8",
|
|
67
|
+
"(#2) = (int16_t)(#1);" },
|
|
68
|
+
{ "monop", "cast_uint16", "int8",
|
|
69
|
+
"(#2) = (uint16_t)(#1);" },
|
|
70
|
+
{ "monop", "cast_int32", "int8",
|
|
71
|
+
"(#2) = (int32_t)(#1);" },
|
|
72
|
+
{ "monop", "cast_uint32", "int8",
|
|
73
|
+
"(#2) = (uint32_t)(#1);" },
|
|
74
|
+
{ "monop", "cast_int64", "int8",
|
|
75
|
+
"(#2) = (int64_t)(#1);" },
|
|
76
|
+
{ "monop", "cast_uint64", "int8",
|
|
77
|
+
"(#2) = (uint64_t)(#1);" },
|
|
78
|
+
{ "monop", "cast_float32", "int8",
|
|
79
|
+
"(#2) = (float)(#1);" },
|
|
80
|
+
{ "monop", "cast_float64", "int8",
|
|
81
|
+
"(#2) = (double)(#1);" },
|
|
82
|
+
{ "monop", "cast_cmplx64", "int8",
|
|
83
|
+
"(#2) = (cmplx64_t)(#1);" },
|
|
84
|
+
{ "monop", "cast_cmplx128", "int8",
|
|
85
|
+
"(#2) = (cmplx128_t)(#1);" },
|
|
86
|
+
{ "monop", "cast_int8", "uint8",
|
|
87
|
+
"(#2) = (int8_t)(#1);" },
|
|
88
|
+
{ "monop", "cast_uint8", "uint8",
|
|
89
|
+
"(#2) = (uint8_t)(#1);" },
|
|
90
|
+
{ "monop", "cast_int16", "uint8",
|
|
91
|
+
"(#2) = (int16_t)(#1);" },
|
|
92
|
+
{ "monop", "cast_uint16", "uint8",
|
|
93
|
+
"(#2) = (uint16_t)(#1);" },
|
|
94
|
+
{ "monop", "cast_int32", "uint8",
|
|
95
|
+
"(#2) = (int32_t)(#1);" },
|
|
96
|
+
{ "monop", "cast_uint32", "uint8",
|
|
97
|
+
"(#2) = (uint32_t)(#1);" },
|
|
98
|
+
{ "monop", "cast_int64", "uint8",
|
|
99
|
+
"(#2) = (int64_t)(#1);" },
|
|
100
|
+
{ "monop", "cast_uint64", "uint8",
|
|
101
|
+
"(#2) = (uint64_t)(#1);" },
|
|
102
|
+
{ "monop", "cast_float32", "uint8",
|
|
103
|
+
"(#2) = (float)(#1);" },
|
|
104
|
+
{ "monop", "cast_float64", "uint8",
|
|
105
|
+
"(#2) = (double)(#1);" },
|
|
106
|
+
{ "monop", "cast_cmplx64", "uint8",
|
|
107
|
+
"(#2) = (cmplx64_t)(#1);" },
|
|
108
|
+
{ "monop", "cast_cmplx128", "uint8",
|
|
109
|
+
"(#2) = (cmplx128_t)(#1);" },
|
|
110
|
+
{ "monop", "cast_int8", "int16",
|
|
111
|
+
"(#2) = (int8_t)(#1);" },
|
|
112
|
+
{ "monop", "cast_uint8", "int16",
|
|
113
|
+
"(#2) = (uint8_t)(#1);" },
|
|
114
|
+
{ "monop", "cast_int16", "int16",
|
|
115
|
+
"(#2) = (int16_t)(#1);" },
|
|
116
|
+
{ "monop", "cast_uint16", "int16",
|
|
117
|
+
"(#2) = (uint16_t)(#1);" },
|
|
118
|
+
{ "monop", "cast_int32", "int16",
|
|
119
|
+
"(#2) = (int32_t)(#1);" },
|
|
120
|
+
{ "monop", "cast_uint32", "int16",
|
|
121
|
+
"(#2) = (uint32_t)(#1);" },
|
|
122
|
+
{ "monop", "cast_int64", "int16",
|
|
123
|
+
"(#2) = (int64_t)(#1);" },
|
|
124
|
+
{ "monop", "cast_uint64", "int16",
|
|
125
|
+
"(#2) = (uint64_t)(#1);" },
|
|
126
|
+
{ "monop", "cast_float32", "int16",
|
|
127
|
+
"(#2) = (float)(#1);" },
|
|
128
|
+
{ "monop", "cast_float64", "int16",
|
|
129
|
+
"(#2) = (double)(#1);" },
|
|
130
|
+
{ "monop", "cast_cmplx64", "int16",
|
|
131
|
+
"(#2) = (cmplx64_t)(#1);" },
|
|
132
|
+
{ "monop", "cast_cmplx128", "int16",
|
|
133
|
+
"(#2) = (cmplx128_t)(#1);" },
|
|
134
|
+
{ "monop", "cast_int8", "uint16",
|
|
135
|
+
"(#2) = (int8_t)(#1);" },
|
|
136
|
+
{ "monop", "cast_uint8", "uint16",
|
|
137
|
+
"(#2) = (uint8_t)(#1);" },
|
|
138
|
+
{ "monop", "cast_int16", "uint16",
|
|
139
|
+
"(#2) = (int16_t)(#1);" },
|
|
140
|
+
{ "monop", "cast_uint16", "uint16",
|
|
141
|
+
"(#2) = (uint16_t)(#1);" },
|
|
142
|
+
{ "monop", "cast_int32", "uint16",
|
|
143
|
+
"(#2) = (int32_t)(#1);" },
|
|
144
|
+
{ "monop", "cast_uint32", "uint16",
|
|
145
|
+
"(#2) = (uint32_t)(#1);" },
|
|
146
|
+
{ "monop", "cast_int64", "uint16",
|
|
147
|
+
"(#2) = (int64_t)(#1);" },
|
|
148
|
+
{ "monop", "cast_uint64", "uint16",
|
|
149
|
+
"(#2) = (uint64_t)(#1);" },
|
|
150
|
+
{ "monop", "cast_float32", "uint16",
|
|
151
|
+
"(#2) = (float)(#1);" },
|
|
152
|
+
{ "monop", "cast_float64", "uint16",
|
|
153
|
+
"(#2) = (double)(#1);" },
|
|
154
|
+
{ "monop", "cast_cmplx64", "uint16",
|
|
155
|
+
"(#2) = (cmplx64_t)(#1);" },
|
|
156
|
+
{ "monop", "cast_cmplx128", "uint16",
|
|
157
|
+
"(#2) = (cmplx128_t)(#1);" },
|
|
158
|
+
{ "monop", "cast_int8", "int32",
|
|
159
|
+
"(#2) = (int8_t)(#1);" },
|
|
160
|
+
{ "monop", "cast_uint8", "int32",
|
|
161
|
+
"(#2) = (uint8_t)(#1);" },
|
|
162
|
+
{ "monop", "cast_int16", "int32",
|
|
163
|
+
"(#2) = (int16_t)(#1);" },
|
|
164
|
+
{ "monop", "cast_uint16", "int32",
|
|
165
|
+
"(#2) = (uint16_t)(#1);" },
|
|
166
|
+
{ "monop", "cast_int32", "int32",
|
|
167
|
+
"(#2) = (int32_t)(#1);" },
|
|
168
|
+
{ "monop", "cast_uint32", "int32",
|
|
169
|
+
"(#2) = (uint32_t)(#1);" },
|
|
170
|
+
{ "monop", "cast_int64", "int32",
|
|
171
|
+
"(#2) = (int64_t)(#1);" },
|
|
172
|
+
{ "monop", "cast_uint64", "int32",
|
|
173
|
+
"(#2) = (uint64_t)(#1);" },
|
|
174
|
+
{ "monop", "cast_float32", "int32",
|
|
175
|
+
"(#2) = (float)(#1);" },
|
|
176
|
+
{ "monop", "cast_float64", "int32",
|
|
177
|
+
"(#2) = (double)(#1);" },
|
|
178
|
+
{ "monop", "cast_cmplx64", "int32",
|
|
179
|
+
"(#2) = (cmplx64_t)(#1);" },
|
|
180
|
+
{ "monop", "cast_cmplx128", "int32",
|
|
181
|
+
"(#2) = (cmplx128_t)(#1);" },
|
|
182
|
+
{ "monop", "cast_int8", "uint32",
|
|
183
|
+
"(#2) = (int8_t)(#1);" },
|
|
184
|
+
{ "monop", "cast_uint8", "uint32",
|
|
185
|
+
"(#2) = (uint8_t)(#1);" },
|
|
186
|
+
{ "monop", "cast_int16", "uint32",
|
|
187
|
+
"(#2) = (int16_t)(#1);" },
|
|
188
|
+
{ "monop", "cast_uint16", "uint32",
|
|
189
|
+
"(#2) = (uint16_t)(#1);" },
|
|
190
|
+
{ "monop", "cast_int32", "uint32",
|
|
191
|
+
"(#2) = (int32_t)(#1);" },
|
|
192
|
+
{ "monop", "cast_uint32", "uint32",
|
|
193
|
+
"(#2) = (uint32_t)(#1);" },
|
|
194
|
+
{ "monop", "cast_int64", "uint32",
|
|
195
|
+
"(#2) = (int64_t)(#1);" },
|
|
196
|
+
{ "monop", "cast_uint64", "uint32",
|
|
197
|
+
"(#2) = (uint64_t)(#1);" },
|
|
198
|
+
{ "monop", "cast_float32", "uint32",
|
|
199
|
+
"(#2) = (float)(#1);" },
|
|
200
|
+
{ "monop", "cast_float64", "uint32",
|
|
201
|
+
"(#2) = (double)(#1);" },
|
|
202
|
+
{ "monop", "cast_cmplx64", "uint32",
|
|
203
|
+
"(#2) = (cmplx64_t)(#1);" },
|
|
204
|
+
{ "monop", "cast_cmplx128", "uint32",
|
|
205
|
+
"(#2) = (cmplx128_t)(#1);" },
|
|
206
|
+
{ "monop", "cast_int8", "int64",
|
|
207
|
+
"(#2) = (int8_t)(#1);" },
|
|
208
|
+
{ "monop", "cast_uint8", "int64",
|
|
209
|
+
"(#2) = (uint8_t)(#1);" },
|
|
210
|
+
{ "monop", "cast_int16", "int64",
|
|
211
|
+
"(#2) = (int16_t)(#1);" },
|
|
212
|
+
{ "monop", "cast_uint16", "int64",
|
|
213
|
+
"(#2) = (uint16_t)(#1);" },
|
|
214
|
+
{ "monop", "cast_int32", "int64",
|
|
215
|
+
"(#2) = (int32_t)(#1);" },
|
|
216
|
+
{ "monop", "cast_uint32", "int64",
|
|
217
|
+
"(#2) = (uint32_t)(#1);" },
|
|
218
|
+
{ "monop", "cast_int64", "int64",
|
|
219
|
+
"(#2) = (int64_t)(#1);" },
|
|
220
|
+
{ "monop", "cast_uint64", "int64",
|
|
221
|
+
"(#2) = (uint64_t)(#1);" },
|
|
222
|
+
{ "monop", "cast_float32", "int64",
|
|
223
|
+
"(#2) = (float)(double)(#1);" },
|
|
224
|
+
{ "monop", "cast_float64", "int64",
|
|
225
|
+
"(#2) = (double)(#1);" },
|
|
226
|
+
{ "monop", "cast_cmplx64", "int64",
|
|
227
|
+
"(#2) = (cmplx64_t)(double)(#1);" },
|
|
228
|
+
{ "monop", "cast_cmplx128", "int64",
|
|
229
|
+
"(#2) = (cmplx128_t)(#1);" },
|
|
230
|
+
{ "monop", "cast_int8", "uint64",
|
|
231
|
+
"(#2) = (int8_t)(#1);" },
|
|
232
|
+
{ "monop", "cast_uint8", "uint64",
|
|
233
|
+
"(#2) = (uint8_t)(#1);" },
|
|
234
|
+
{ "monop", "cast_int16", "uint64",
|
|
235
|
+
"(#2) = (int16_t)(#1);" },
|
|
236
|
+
{ "monop", "cast_uint16", "uint64",
|
|
237
|
+
"(#2) = (uint16_t)(#1);" },
|
|
238
|
+
{ "monop", "cast_int32", "uint64",
|
|
239
|
+
"(#2) = (int32_t)(#1);" },
|
|
240
|
+
{ "monop", "cast_uint32", "uint64",
|
|
241
|
+
"(#2) = (uint32_t)(#1);" },
|
|
242
|
+
{ "monop", "cast_int64", "uint64",
|
|
243
|
+
"(#2) = (int64_t)(#1);" },
|
|
244
|
+
{ "monop", "cast_uint64", "uint64",
|
|
245
|
+
"(#2) = (uint64_t)(#1);" },
|
|
246
|
+
{ "monop", "cast_float32", "uint64",
|
|
247
|
+
"(#2) = (float)(double)(#1);" },
|
|
248
|
+
{ "monop", "cast_float64", "uint64",
|
|
249
|
+
"(#2) = (double)(#1);" },
|
|
250
|
+
{ "monop", "cast_cmplx64", "uint64",
|
|
251
|
+
"(#2) = (cmplx64_t)(double)(#1);" },
|
|
252
|
+
{ "monop", "cast_cmplx128", "uint64",
|
|
253
|
+
"(#2) = (cmplx128_t)(#1);" },
|
|
254
|
+
{ "monop", "cast_float32", "float32",
|
|
255
|
+
"(#2) = (float)(#1);" },
|
|
256
|
+
{ "monop", "cast_float64", "float32",
|
|
257
|
+
"(#2) = (double)(#1);" },
|
|
258
|
+
{ "monop", "cast_cmplx64", "float32",
|
|
259
|
+
"(#2) = (cmplx64_t)(#1);" },
|
|
260
|
+
{ "monop", "cast_cmplx128", "float32",
|
|
261
|
+
"(#2) = (cmplx128_t)(#1);" },
|
|
262
|
+
{ "monop", "cast_float32", "float64",
|
|
263
|
+
"(#2) = (float)(#1);" },
|
|
264
|
+
{ "monop", "cast_float64", "float64",
|
|
265
|
+
"(#2) = (double)(#1);" },
|
|
266
|
+
{ "monop", "cast_cmplx64", "float64",
|
|
267
|
+
"(#2) = (cmplx64_t)(#1);" },
|
|
268
|
+
{ "monop", "cast_cmplx128", "float64",
|
|
269
|
+
"(#2) = (cmplx128_t)(#1);" },
|
|
270
|
+
{ "monop", "cast_float32", "cmplx64",
|
|
271
|
+
"(#2) = (float)(#1);" },
|
|
272
|
+
{ "monop", "cast_float64", "cmplx64",
|
|
273
|
+
"(#2) = (double)(#1);" },
|
|
274
|
+
{ "monop", "cast_cmplx64", "cmplx64",
|
|
275
|
+
"(#2) = (cmplx64_t)(#1);" },
|
|
276
|
+
{ "monop", "cast_cmplx128", "cmplx64",
|
|
277
|
+
"(#2) = (cmplx128_t)(#1);" },
|
|
278
|
+
{ "monop", "cast_float32", "cmplx128",
|
|
279
|
+
"(#2) = (float)(#1);" },
|
|
280
|
+
{ "monop", "cast_float64", "cmplx128",
|
|
281
|
+
"(#2) = (double)(#1);" },
|
|
282
|
+
{ "monop", "cast_cmplx64", "cmplx128",
|
|
283
|
+
"(#2) = (cmplx64_t)(#1);" },
|
|
284
|
+
{ "monop", "cast_cmplx128", "cmplx128",
|
|
285
|
+
"(#2) = (cmplx128_t)(#1);" },
|
|
31
286
|
{ "monop", "zero", "int8",
|
|
32
287
|
"(#2) = 0;" },
|
|
33
288
|
{ "monop", "zero", "uint8",
|
|
@@ -313,7 +568,7 @@ static const ca_kernel_body_t ca_kernel_bodies[] = {
|
|
|
313
568
|
{ "monop", "imag_i", "cmplx128",
|
|
314
569
|
"(#2) = cimag(#1);" },
|
|
315
570
|
{ "monop", "arg_i", "float32",
|
|
316
|
-
"(#2) =
|
|
571
|
+
"(#2) = carg((cmplx128_t)(#1));" },
|
|
317
572
|
{ "monop", "arg_i", "float64",
|
|
318
573
|
"(#2) = carg((cmplx128_t)(#1));" },
|
|
319
574
|
{ "monop", "arg_i", "cmplx64",
|
|
@@ -387,23 +642,23 @@ static const ca_kernel_body_t ca_kernel_bodies[] = {
|
|
|
387
642
|
{ "monop", "round", "uint64",
|
|
388
643
|
"(#2) = (#1);" },
|
|
389
644
|
{ "monop", "round", "float32",
|
|
390
|
-
"(#2) = (
|
|
645
|
+
"(#2) = round(#1);" },
|
|
391
646
|
{ "monop", "round", "float64",
|
|
392
|
-
"(#2) = (
|
|
647
|
+
"(#2) = round(#1);" },
|
|
393
648
|
{ "monop", "rcp", "int8",
|
|
394
|
-
"if ((#1)==0) {ca_zerodiv();}; (#2) = 1
|
|
649
|
+
"if ((#1)==0) {ca_zerodiv();}; (#2) = ((#1)==1) ? 1 : ((#1) < 0) ? -1 : 0;" },
|
|
395
650
|
{ "monop", "rcp", "uint8",
|
|
396
651
|
"if ((#1)==0) {ca_zerodiv();}; (#2) = 1/(#1);" },
|
|
397
652
|
{ "monop", "rcp", "int16",
|
|
398
|
-
"if ((#1)==0) {ca_zerodiv();}; (#2) = 1
|
|
653
|
+
"if ((#1)==0) {ca_zerodiv();}; (#2) = ((#1)==1) ? 1 : ((#1) < 0) ? -1 : 0;" },
|
|
399
654
|
{ "monop", "rcp", "uint16",
|
|
400
655
|
"if ((#1)==0) {ca_zerodiv();}; (#2) = 1/(#1);" },
|
|
401
656
|
{ "monop", "rcp", "int32",
|
|
402
|
-
"if ((#1)==0) {ca_zerodiv();}; (#2) = 1
|
|
657
|
+
"if ((#1)==0) {ca_zerodiv();}; (#2) = ((#1)==1) ? 1 : ((#1) < 0) ? -1 : 0;" },
|
|
403
658
|
{ "monop", "rcp", "uint32",
|
|
404
659
|
"if ((#1)==0) {ca_zerodiv();}; (#2) = 1/(#1);" },
|
|
405
660
|
{ "monop", "rcp", "int64",
|
|
406
|
-
"if ((#1)==0) {ca_zerodiv();}; (#2) = 1
|
|
661
|
+
"if ((#1)==0) {ca_zerodiv();}; (#2) = ((#1)==1) ? 1 : ((#1) < 0) ? -1 : 0;" },
|
|
407
662
|
{ "monop", "rcp", "uint64",
|
|
408
663
|
"if ((#1)==0) {ca_zerodiv();}; (#2) = 1/(#1);" },
|
|
409
664
|
{ "monop", "rcp", "float32",
|
|
@@ -695,9 +950,9 @@ static const ca_kernel_body_t ca_kernel_bodies[] = {
|
|
|
695
950
|
{ "binop", "maximum", "uint64",
|
|
696
951
|
"(#3) = (#1) > (#2) ? (#1) : (#2);" },
|
|
697
952
|
{ "binop", "maximum", "float32",
|
|
698
|
-
"(#3) = isnan(#1) ? (#1) :
|
|
953
|
+
"(#3) = isnan(#1) ? (#1) : isnan(#2) ? (#2) : ((#1) > (#2) || ((#1) == (#2) && !signbit(#1))) ? (#1) : (#2);" },
|
|
699
954
|
{ "binop", "maximum", "float64",
|
|
700
|
-
"(#3) = isnan(#1) ? (#1) :
|
|
955
|
+
"(#3) = isnan(#1) ? (#1) : isnan(#2) ? (#2) : ((#1) > (#2) || ((#1) == (#2) && !signbit(#1))) ? (#1) : (#2);" },
|
|
701
956
|
{ "binop", "minimum", "int8",
|
|
702
957
|
"(#3) = (#1) < (#2) ? (#1) : (#2);" },
|
|
703
958
|
{ "binop", "minimum", "uint8",
|
|
@@ -715,9 +970,9 @@ static const ca_kernel_body_t ca_kernel_bodies[] = {
|
|
|
715
970
|
{ "binop", "minimum", "uint64",
|
|
716
971
|
"(#3) = (#1) < (#2) ? (#1) : (#2);" },
|
|
717
972
|
{ "binop", "minimum", "float32",
|
|
718
|
-
"(#3) = isnan(#1) ? (#1) :
|
|
973
|
+
"(#3) = isnan(#1) ? (#1) : isnan(#2) ? (#2) : ((#1) < (#2) || ((#1) == (#2) && signbit(#1))) ? (#1) : (#2);" },
|
|
719
974
|
{ "binop", "minimum", "float64",
|
|
720
|
-
"(#3) = isnan(#1) ? (#1) :
|
|
975
|
+
"(#3) = isnan(#1) ? (#1) : isnan(#2) ? (#2) : ((#1) < (#2) || ((#1) == (#2) && signbit(#1))) ? (#1) : (#2);" },
|
|
721
976
|
{ "binop", "add", "int8",
|
|
722
977
|
"(#3) = (#1) + (#2);" },
|
|
723
978
|
{ "binop", "add", "uint8",
|
|
@@ -791,19 +1046,19 @@ static const ca_kernel_body_t ca_kernel_bodies[] = {
|
|
|
791
1046
|
{ "binop", "mul", "cmplx128",
|
|
792
1047
|
"(#3) = (#1) * (#2);" },
|
|
793
1048
|
{ "binop", "div", "int8",
|
|
794
|
-
"\n if ((#2)==0) {ca_zerodiv();};\n {\n <type> _q = (#1) / (#2);\n if ( (#1) % (#2) != 0 && (((#1) < 0) != ((#2) < 0)) ) { _q -= 1; }\n (#3) = _q;\n }\n " },
|
|
1049
|
+
"\n if ((#2)==0) {ca_zerodiv();};\n if ((#2)==-1) { (#3) = (<type>) (0 - (uint64_t) (#1)); }\n else {\n <type> _q = (#1) / (#2);\n if ( (#1) % (#2) != 0 && (((#1) < 0) != ((#2) < 0)) ) { _q -= 1; }\n (#3) = _q;\n }\n " },
|
|
795
1050
|
{ "binop", "div", "uint8",
|
|
796
1051
|
"if ((#2)==0) {ca_zerodiv();}; (#3) = (#1) / (#2);" },
|
|
797
1052
|
{ "binop", "div", "int16",
|
|
798
|
-
"\n if ((#2)==0) {ca_zerodiv();};\n {\n <type> _q = (#1) / (#2);\n if ( (#1) % (#2) != 0 && (((#1) < 0) != ((#2) < 0)) ) { _q -= 1; }\n (#3) = _q;\n }\n " },
|
|
1053
|
+
"\n if ((#2)==0) {ca_zerodiv();};\n if ((#2)==-1) { (#3) = (<type>) (0 - (uint64_t) (#1)); }\n else {\n <type> _q = (#1) / (#2);\n if ( (#1) % (#2) != 0 && (((#1) < 0) != ((#2) < 0)) ) { _q -= 1; }\n (#3) = _q;\n }\n " },
|
|
799
1054
|
{ "binop", "div", "uint16",
|
|
800
1055
|
"if ((#2)==0) {ca_zerodiv();}; (#3) = (#1) / (#2);" },
|
|
801
1056
|
{ "binop", "div", "int32",
|
|
802
|
-
"\n if ((#2)==0) {ca_zerodiv();};\n {\n <type> _q = (#1) / (#2);\n if ( (#1) % (#2) != 0 && (((#1) < 0) != ((#2) < 0)) ) { _q -= 1; }\n (#3) = _q;\n }\n " },
|
|
1057
|
+
"\n if ((#2)==0) {ca_zerodiv();};\n if ((#2)==-1) { (#3) = (<type>) (0 - (uint64_t) (#1)); }\n else {\n <type> _q = (#1) / (#2);\n if ( (#1) % (#2) != 0 && (((#1) < 0) != ((#2) < 0)) ) { _q -= 1; }\n (#3) = _q;\n }\n " },
|
|
803
1058
|
{ "binop", "div", "uint32",
|
|
804
1059
|
"if ((#2)==0) {ca_zerodiv();}; (#3) = (#1) / (#2);" },
|
|
805
1060
|
{ "binop", "div", "int64",
|
|
806
|
-
"\n if ((#2)==0) {ca_zerodiv();};\n {\n <type> _q = (#1) / (#2);\n if ( (#1) % (#2) != 0 && (((#1) < 0) != ((#2) < 0)) ) { _q -= 1; }\n (#3) = _q;\n }\n " },
|
|
1061
|
+
"\n if ((#2)==0) {ca_zerodiv();};\n if ((#2)==-1) { (#3) = (<type>) (0 - (uint64_t) (#1)); }\n else {\n <type> _q = (#1) / (#2);\n if ( (#1) % (#2) != 0 && (((#1) < 0) != ((#2) < 0)) ) { _q -= 1; }\n (#3) = _q;\n }\n " },
|
|
807
1062
|
{ "binop", "div", "uint64",
|
|
808
1063
|
"if ((#2)==0) {ca_zerodiv();}; (#3) = (#1) / (#2);" },
|
|
809
1064
|
{ "binop", "div", "float32",
|
|
@@ -815,19 +1070,19 @@ static const ca_kernel_body_t ca_kernel_bodies[] = {
|
|
|
815
1070
|
{ "binop", "div", "cmplx128",
|
|
816
1071
|
"(#3) = (#1) / (#2);" },
|
|
817
1072
|
{ "binop", "rcp_mul", "int8",
|
|
818
|
-
"if ((#1)==0) {ca_zerodiv();}
|
|
1073
|
+
"\n if ((#1)==0) {ca_zerodiv();};\n if ((#1)==-1) { (#3) = (<type>) (0 - (uint64_t) (#2)); }\n else {\n <type> _q = (#2) / (#1);\n if ( (#2) % (#1) != 0 && (((#2) < 0) != ((#1) < 0)) ) { _q -= 1; }\n (#3) = _q;\n }\n " },
|
|
819
1074
|
{ "binop", "rcp_mul", "uint8",
|
|
820
1075
|
"if ((#1)==0) {ca_zerodiv();}; (#3) = (#2) / (#1);" },
|
|
821
1076
|
{ "binop", "rcp_mul", "int16",
|
|
822
|
-
"if ((#1)==0) {ca_zerodiv();}
|
|
1077
|
+
"\n if ((#1)==0) {ca_zerodiv();};\n if ((#1)==-1) { (#3) = (<type>) (0 - (uint64_t) (#2)); }\n else {\n <type> _q = (#2) / (#1);\n if ( (#2) % (#1) != 0 && (((#2) < 0) != ((#1) < 0)) ) { _q -= 1; }\n (#3) = _q;\n }\n " },
|
|
823
1078
|
{ "binop", "rcp_mul", "uint16",
|
|
824
1079
|
"if ((#1)==0) {ca_zerodiv();}; (#3) = (#2) / (#1);" },
|
|
825
1080
|
{ "binop", "rcp_mul", "int32",
|
|
826
|
-
"if ((#1)==0) {ca_zerodiv();}
|
|
1081
|
+
"\n if ((#1)==0) {ca_zerodiv();};\n if ((#1)==-1) { (#3) = (<type>) (0 - (uint64_t) (#2)); }\n else {\n <type> _q = (#2) / (#1);\n if ( (#2) % (#1) != 0 && (((#2) < 0) != ((#1) < 0)) ) { _q -= 1; }\n (#3) = _q;\n }\n " },
|
|
827
1082
|
{ "binop", "rcp_mul", "uint32",
|
|
828
1083
|
"if ((#1)==0) {ca_zerodiv();}; (#3) = (#2) / (#1);" },
|
|
829
1084
|
{ "binop", "rcp_mul", "int64",
|
|
830
|
-
"if ((#1)==0) {ca_zerodiv();}
|
|
1085
|
+
"\n if ((#1)==0) {ca_zerodiv();};\n if ((#1)==-1) { (#3) = (<type>) (0 - (uint64_t) (#2)); }\n else {\n <type> _q = (#2) / (#1);\n if ( (#2) % (#1) != 0 && (((#2) < 0) != ((#1) < 0)) ) { _q -= 1; }\n (#3) = _q;\n }\n " },
|
|
831
1086
|
{ "binop", "rcp_mul", "uint64",
|
|
832
1087
|
"if ((#1)==0) {ca_zerodiv();}; (#3) = (#2) / (#1);" },
|
|
833
1088
|
{ "binop", "rcp_mul", "float32",
|
|
@@ -839,19 +1094,19 @@ static const ca_kernel_body_t ca_kernel_bodies[] = {
|
|
|
839
1094
|
{ "binop", "rcp_mul", "cmplx128",
|
|
840
1095
|
"(#3) = (#2) / (#1);" },
|
|
841
1096
|
{ "binop", "mod", "int8",
|
|
842
|
-
"\n if ((#2)==0) {ca_zerodiv();};\n {\n <type> _r = (#1) % (#2);\n if ( _r != 0 && ((_r < 0) != ((#2) < 0)) ) { _r += (#2); }\n (#3) = _r;\n }\n " },
|
|
1097
|
+
"\n if ((#2)==0) {ca_zerodiv();};\n if ((#2)==-1) { (#3) = 0; }\n else {\n <type> _r = (#1) % (#2);\n if ( _r != 0 && ((_r < 0) != ((#2) < 0)) ) { _r += (#2); }\n (#3) = _r;\n }\n " },
|
|
843
1098
|
{ "binop", "mod", "uint8",
|
|
844
1099
|
"if ((#2)==0) {ca_zerodiv();}; (#3) = (#1) % (#2);" },
|
|
845
1100
|
{ "binop", "mod", "int16",
|
|
846
|
-
"\n if ((#2)==0) {ca_zerodiv();};\n {\n <type> _r = (#1) % (#2);\n if ( _r != 0 && ((_r < 0) != ((#2) < 0)) ) { _r += (#2); }\n (#3) = _r;\n }\n " },
|
|
1101
|
+
"\n if ((#2)==0) {ca_zerodiv();};\n if ((#2)==-1) { (#3) = 0; }\n else {\n <type> _r = (#1) % (#2);\n if ( _r != 0 && ((_r < 0) != ((#2) < 0)) ) { _r += (#2); }\n (#3) = _r;\n }\n " },
|
|
847
1102
|
{ "binop", "mod", "uint16",
|
|
848
1103
|
"if ((#2)==0) {ca_zerodiv();}; (#3) = (#1) % (#2);" },
|
|
849
1104
|
{ "binop", "mod", "int32",
|
|
850
|
-
"\n if ((#2)==0) {ca_zerodiv();};\n {\n <type> _r = (#1) % (#2);\n if ( _r != 0 && ((_r < 0) != ((#2) < 0)) ) { _r += (#2); }\n (#3) = _r;\n }\n " },
|
|
1105
|
+
"\n if ((#2)==0) {ca_zerodiv();};\n if ((#2)==-1) { (#3) = 0; }\n else {\n <type> _r = (#1) % (#2);\n if ( _r != 0 && ((_r < 0) != ((#2) < 0)) ) { _r += (#2); }\n (#3) = _r;\n }\n " },
|
|
851
1106
|
{ "binop", "mod", "uint32",
|
|
852
1107
|
"if ((#2)==0) {ca_zerodiv();}; (#3) = (#1) % (#2);" },
|
|
853
1108
|
{ "binop", "mod", "int64",
|
|
854
|
-
"\n if ((#2)==0) {ca_zerodiv();};\n {\n <type> _r = (#1) % (#2);\n if ( _r != 0 && ((_r < 0) != ((#2) < 0)) ) { _r += (#2); }\n (#3) = _r;\n }\n " },
|
|
1109
|
+
"\n if ((#2)==0) {ca_zerodiv();};\n if ((#2)==-1) { (#3) = 0; }\n else {\n <type> _r = (#1) % (#2);\n if ( _r != 0 && ((_r < 0) != ((#2) < 0)) ) { _r += (#2); }\n (#3) = _r;\n }\n " },
|
|
855
1110
|
{ "binop", "mod", "uint64",
|
|
856
1111
|
"if ((#2)==0) {ca_zerodiv();}; (#3) = (#1) % (#2);" },
|
|
857
1112
|
{ "binop", "mod", "float32",
|
|
@@ -859,19 +1114,19 @@ static const ca_kernel_body_t ca_kernel_bodies[] = {
|
|
|
859
1114
|
{ "binop", "mod", "float64",
|
|
860
1115
|
"\n {\n <type> _r = fmod((#1), (#2));\n if ( _r != 0 ) {\n if ( (_r < 0) != ((#2) < 0) ) { _r += (#2); }\n }\n else { _r = copysign((<type>) 0, (#2)); }\n (#3) = _r;\n }\n " },
|
|
861
1116
|
{ "binop", "fmod", "int8",
|
|
862
|
-
"if ((#2)==0) {ca_zerodiv();}; (#3) = (#1) % (#2);" },
|
|
1117
|
+
"if ((#2)==0) {ca_zerodiv();}; (#3) = ((#2)==-1) ? 0 : (#1) % (#2);" },
|
|
863
1118
|
{ "binop", "fmod", "uint8",
|
|
864
1119
|
"if ((#2)==0) {ca_zerodiv();}; (#3) = (#1) % (#2);" },
|
|
865
1120
|
{ "binop", "fmod", "int16",
|
|
866
|
-
"if ((#2)==0) {ca_zerodiv();}; (#3) = (#1) % (#2);" },
|
|
1121
|
+
"if ((#2)==0) {ca_zerodiv();}; (#3) = ((#2)==-1) ? 0 : (#1) % (#2);" },
|
|
867
1122
|
{ "binop", "fmod", "uint16",
|
|
868
1123
|
"if ((#2)==0) {ca_zerodiv();}; (#3) = (#1) % (#2);" },
|
|
869
1124
|
{ "binop", "fmod", "int32",
|
|
870
|
-
"if ((#2)==0) {ca_zerodiv();}; (#3) = (#1) % (#2);" },
|
|
1125
|
+
"if ((#2)==0) {ca_zerodiv();}; (#3) = ((#2)==-1) ? 0 : (#1) % (#2);" },
|
|
871
1126
|
{ "binop", "fmod", "uint32",
|
|
872
1127
|
"if ((#2)==0) {ca_zerodiv();}; (#3) = (#1) % (#2);" },
|
|
873
1128
|
{ "binop", "fmod", "int64",
|
|
874
|
-
"if ((#2)==0) {ca_zerodiv();}; (#3) = (#1) % (#2);" },
|
|
1129
|
+
"if ((#2)==0) {ca_zerodiv();}; (#3) = ((#2)==-1) ? 0 : (#1) % (#2);" },
|
|
875
1130
|
{ "binop", "fmod", "uint64",
|
|
876
1131
|
"if ((#2)==0) {ca_zerodiv();}; (#3) = (#1) % (#2);" },
|
|
877
1132
|
{ "binop", "fmod", "float32",
|
|
@@ -933,37 +1188,37 @@ static const ca_kernel_body_t ca_kernel_bodies[] = {
|
|
|
933
1188
|
{ "binop", "bit_xor_i", "boolean",
|
|
934
1189
|
"(#3) = ((#1) != (#2)) ? 1 : 0;" },
|
|
935
1190
|
{ "binop", "bit_lshift", "int8",
|
|
936
|
-
"(#
|
|
1191
|
+
"\n {\n int64_t _x = (int64_t) (#1), _n = (int64_t) (#2), _r;\n const int64_t _w = (int64_t) (8 * sizeof(<type>));\n if ( _n >= _w ) _r = 0;\n else if ( _n <= -_w ) _r = _x < 0 ? -1 : 0;\n else if ( _n < 0 ) _r = _x >> -_n;\n else _r = (int64_t) ((uint64_t) _x << _n);\n (#3) = (<type>) _r;\n }\n " },
|
|
937
1192
|
{ "binop", "bit_lshift", "uint8",
|
|
938
|
-
"(#3) = (#1) << (#2);" },
|
|
1193
|
+
"(#3) = ((uint64_t) (#2) >= 8 * sizeof(<type>)) ? 0 : (<type>) ((uint64_t) (#1) << (#2));" },
|
|
939
1194
|
{ "binop", "bit_lshift", "int16",
|
|
940
|
-
"(#
|
|
1195
|
+
"\n {\n int64_t _x = (int64_t) (#1), _n = (int64_t) (#2), _r;\n const int64_t _w = (int64_t) (8 * sizeof(<type>));\n if ( _n >= _w ) _r = 0;\n else if ( _n <= -_w ) _r = _x < 0 ? -1 : 0;\n else if ( _n < 0 ) _r = _x >> -_n;\n else _r = (int64_t) ((uint64_t) _x << _n);\n (#3) = (<type>) _r;\n }\n " },
|
|
941
1196
|
{ "binop", "bit_lshift", "uint16",
|
|
942
|
-
"(#3) = (#1) << (#2);" },
|
|
1197
|
+
"(#3) = ((uint64_t) (#2) >= 8 * sizeof(<type>)) ? 0 : (<type>) ((uint64_t) (#1) << (#2));" },
|
|
943
1198
|
{ "binop", "bit_lshift", "int32",
|
|
944
|
-
"(#
|
|
1199
|
+
"\n {\n int64_t _x = (int64_t) (#1), _n = (int64_t) (#2), _r;\n const int64_t _w = (int64_t) (8 * sizeof(<type>));\n if ( _n >= _w ) _r = 0;\n else if ( _n <= -_w ) _r = _x < 0 ? -1 : 0;\n else if ( _n < 0 ) _r = _x >> -_n;\n else _r = (int64_t) ((uint64_t) _x << _n);\n (#3) = (<type>) _r;\n }\n " },
|
|
945
1200
|
{ "binop", "bit_lshift", "uint32",
|
|
946
|
-
"(#3) = (#1) << (#2);" },
|
|
1201
|
+
"(#3) = ((uint64_t) (#2) >= 8 * sizeof(<type>)) ? 0 : (<type>) ((uint64_t) (#1) << (#2));" },
|
|
947
1202
|
{ "binop", "bit_lshift", "int64",
|
|
948
|
-
"(#
|
|
1203
|
+
"\n {\n int64_t _x = (int64_t) (#1), _n = (int64_t) (#2), _r;\n const int64_t _w = (int64_t) (8 * sizeof(<type>));\n if ( _n >= _w ) _r = 0;\n else if ( _n <= -_w ) _r = _x < 0 ? -1 : 0;\n else if ( _n < 0 ) _r = _x >> -_n;\n else _r = (int64_t) ((uint64_t) _x << _n);\n (#3) = (<type>) _r;\n }\n " },
|
|
949
1204
|
{ "binop", "bit_lshift", "uint64",
|
|
950
|
-
"(#3) = (#1) << (#2);" },
|
|
1205
|
+
"(#3) = ((uint64_t) (#2) >= 8 * sizeof(<type>)) ? 0 : (<type>) ((uint64_t) (#1) << (#2));" },
|
|
951
1206
|
{ "binop", "bit_rshift", "int8",
|
|
952
|
-
"(#
|
|
1207
|
+
"\n {\n int64_t _x = (int64_t) (#1), _n = (int64_t) (#2), _r;\n const int64_t _w = (int64_t) (8 * sizeof(<type>));\n if ( _n >= _w ) _r = _x < 0 ? -1 : 0;\n else if ( _n <= -_w ) _r = 0;\n else if ( _n < 0 ) _r = (int64_t) ((uint64_t) _x << -_n);\n else _r = _x >> _n;\n (#3) = (<type>) _r;\n }\n " },
|
|
953
1208
|
{ "binop", "bit_rshift", "uint8",
|
|
954
|
-
"(#3) = (#1) >> (#2);" },
|
|
1209
|
+
"(#3) = ((uint64_t) (#2) >= 8 * sizeof(<type>)) ? 0 : (<type>) ((uint64_t) (#1) >> (#2));" },
|
|
955
1210
|
{ "binop", "bit_rshift", "int16",
|
|
956
|
-
"(#
|
|
1211
|
+
"\n {\n int64_t _x = (int64_t) (#1), _n = (int64_t) (#2), _r;\n const int64_t _w = (int64_t) (8 * sizeof(<type>));\n if ( _n >= _w ) _r = _x < 0 ? -1 : 0;\n else if ( _n <= -_w ) _r = 0;\n else if ( _n < 0 ) _r = (int64_t) ((uint64_t) _x << -_n);\n else _r = _x >> _n;\n (#3) = (<type>) _r;\n }\n " },
|
|
957
1212
|
{ "binop", "bit_rshift", "uint16",
|
|
958
|
-
"(#3) = (#1) >> (#2);" },
|
|
1213
|
+
"(#3) = ((uint64_t) (#2) >= 8 * sizeof(<type>)) ? 0 : (<type>) ((uint64_t) (#1) >> (#2));" },
|
|
959
1214
|
{ "binop", "bit_rshift", "int32",
|
|
960
|
-
"(#
|
|
1215
|
+
"\n {\n int64_t _x = (int64_t) (#1), _n = (int64_t) (#2), _r;\n const int64_t _w = (int64_t) (8 * sizeof(<type>));\n if ( _n >= _w ) _r = _x < 0 ? -1 : 0;\n else if ( _n <= -_w ) _r = 0;\n else if ( _n < 0 ) _r = (int64_t) ((uint64_t) _x << -_n);\n else _r = _x >> _n;\n (#3) = (<type>) _r;\n }\n " },
|
|
961
1216
|
{ "binop", "bit_rshift", "uint32",
|
|
962
|
-
"(#3) = (#1) >> (#2);" },
|
|
1217
|
+
"(#3) = ((uint64_t) (#2) >= 8 * sizeof(<type>)) ? 0 : (<type>) ((uint64_t) (#1) >> (#2));" },
|
|
963
1218
|
{ "binop", "bit_rshift", "int64",
|
|
964
|
-
"(#
|
|
1219
|
+
"\n {\n int64_t _x = (int64_t) (#1), _n = (int64_t) (#2), _r;\n const int64_t _w = (int64_t) (8 * sizeof(<type>));\n if ( _n >= _w ) _r = _x < 0 ? -1 : 0;\n else if ( _n <= -_w ) _r = 0;\n else if ( _n < 0 ) _r = (int64_t) ((uint64_t) _x << -_n);\n else _r = _x >> _n;\n (#3) = (<type>) _r;\n }\n " },
|
|
965
1220
|
{ "binop", "bit_rshift", "uint64",
|
|
966
|
-
"(#3) = (#1) >> (#2);" },
|
|
1221
|
+
"(#3) = ((uint64_t) (#2) >= 8 * sizeof(<type>)) ? 0 : (<type>) ((uint64_t) (#1) >> (#2));" },
|
|
967
1222
|
{ "binop", "and", "boolean",
|
|
968
1223
|
"(#3) = (#1) && (#2);" },
|
|
969
1224
|
{ "binop", "or", "boolean",
|
|
@@ -987,7 +1242,7 @@ static const ca_kernel_body_t ca_kernel_bodies[] = {
|
|
|
987
1242
|
{ "triop", "fma", "uint64",
|
|
988
1243
|
"(#4) = (#1) * (#2) + (#3);" },
|
|
989
1244
|
{ "triop", "fma", "float32",
|
|
990
|
-
"(#4) =
|
|
1245
|
+
"(#4) = fmaf((#1), (#2), (#3));" },
|
|
991
1246
|
{ "triop", "fma", "float64",
|
|
992
1247
|
"(#4) = fma((#1), (#2), (#3));" },
|
|
993
1248
|
{ "triop", "fma", "cmplx64",
|
|
@@ -1011,7 +1266,7 @@ static const ca_kernel_body_t ca_kernel_bodies[] = {
|
|
|
1011
1266
|
{ "triop", "fms", "uint64",
|
|
1012
1267
|
"(#4) = (#1) * (#2) - (#3);" },
|
|
1013
1268
|
{ "triop", "fms", "float32",
|
|
1014
|
-
"(#4) =
|
|
1269
|
+
"(#4) = fmaf((#1), (#2), -(#3));" },
|
|
1015
1270
|
{ "triop", "fms", "float64",
|
|
1016
1271
|
"(#4) = fma((#1), (#2), -(#3));" },
|
|
1017
1272
|
{ "triop", "fms", "cmplx64",
|
|
@@ -1038,6 +1293,32 @@ static const ca_kernel_body_t ca_kernel_bodies[] = {
|
|
|
1038
1293
|
"(#4) = isnan(#1) ? (#1) : fmin(fmax((#1), (#2)), (#3));" },
|
|
1039
1294
|
{ "triop", "clip", "float64",
|
|
1040
1295
|
"(#4) = isnan(#1) ? (#1) : fmin(fmax((#1), (#2)), (#3));" },
|
|
1296
|
+
{ "triop", "select", "int8",
|
|
1297
|
+
"(#4) = (#1) ? (#2) : (#3);" },
|
|
1298
|
+
{ "triop", "select", "uint8",
|
|
1299
|
+
"(#4) = (#1) ? (#2) : (#3);" },
|
|
1300
|
+
{ "triop", "select", "int16",
|
|
1301
|
+
"(#4) = (#1) ? (#2) : (#3);" },
|
|
1302
|
+
{ "triop", "select", "uint16",
|
|
1303
|
+
"(#4) = (#1) ? (#2) : (#3);" },
|
|
1304
|
+
{ "triop", "select", "int32",
|
|
1305
|
+
"(#4) = (#1) ? (#2) : (#3);" },
|
|
1306
|
+
{ "triop", "select", "uint32",
|
|
1307
|
+
"(#4) = (#1) ? (#2) : (#3);" },
|
|
1308
|
+
{ "triop", "select", "int64",
|
|
1309
|
+
"(#4) = (#1) ? (#2) : (#3);" },
|
|
1310
|
+
{ "triop", "select", "uint64",
|
|
1311
|
+
"(#4) = (#1) ? (#2) : (#3);" },
|
|
1312
|
+
{ "triop", "select", "float32",
|
|
1313
|
+
"(#4) = (#1) ? (#2) : (#3);" },
|
|
1314
|
+
{ "triop", "select", "float64",
|
|
1315
|
+
"(#4) = (#1) ? (#2) : (#3);" },
|
|
1316
|
+
{ "triop", "select", "boolean",
|
|
1317
|
+
"(#4) = (#1) ? (#2) : (#3);" },
|
|
1318
|
+
{ "triop", "select", "cmplx64",
|
|
1319
|
+
"(#4) = (#1) ? (#2) : (#3);" },
|
|
1320
|
+
{ "triop", "select", "cmplx128",
|
|
1321
|
+
"(#4) = (#1) ? (#2) : (#3);" },
|
|
1041
1322
|
{ "binop", "power", "int8",
|
|
1042
1323
|
"(#3) = op_powi_<type>((#1), (#2));" },
|
|
1043
1324
|
{ "binop", "power", "uint8",
|
|
@@ -1082,9 +1363,319 @@ static const ca_kernel_body_t ca_kernel_bodies[] = {
|
|
|
1082
1363
|
"(#3) = fmaxf((#1), (#2)) + log1pf(expf(-fabsf((#1) - (#2))));" },
|
|
1083
1364
|
{ "binop", "logaddexp", "float64",
|
|
1084
1365
|
"(#3) = fmax((#1), (#2)) + log1p(exp(-fabs((#1) - (#2))));" },
|
|
1366
|
+
{ "moncmp", "is_nan", "int8",
|
|
1367
|
+
"(#2) = 0;" },
|
|
1368
|
+
{ "moncmp", "is_nan", "uint8",
|
|
1369
|
+
"(#2) = 0;" },
|
|
1370
|
+
{ "moncmp", "is_nan", "int16",
|
|
1371
|
+
"(#2) = 0;" },
|
|
1372
|
+
{ "moncmp", "is_nan", "uint16",
|
|
1373
|
+
"(#2) = 0;" },
|
|
1374
|
+
{ "moncmp", "is_nan", "int32",
|
|
1375
|
+
"(#2) = 0;" },
|
|
1376
|
+
{ "moncmp", "is_nan", "uint32",
|
|
1377
|
+
"(#2) = 0;" },
|
|
1378
|
+
{ "moncmp", "is_nan", "int64",
|
|
1379
|
+
"(#2) = 0;" },
|
|
1380
|
+
{ "moncmp", "is_nan", "uint64",
|
|
1381
|
+
"(#2) = 0;" },
|
|
1382
|
+
{ "moncmp", "is_nan", "float32",
|
|
1383
|
+
"(#2) = isnan(#1);" },
|
|
1384
|
+
{ "moncmp", "is_nan", "float64",
|
|
1385
|
+
"(#2) = isnan(#1);" },
|
|
1386
|
+
{ "moncmp", "is_nan", "cmplx64",
|
|
1387
|
+
"(#2) = (isnan(creal(#1)) || isnan(cimag(#1)));" },
|
|
1388
|
+
{ "moncmp", "is_nan", "cmplx128",
|
|
1389
|
+
"(#2) = (isnan(creal(#1)) || isnan(cimag(#1)));" },
|
|
1390
|
+
{ "moncmp", "is_inf", "int8",
|
|
1391
|
+
"(#2) = 0;" },
|
|
1392
|
+
{ "moncmp", "is_inf", "uint8",
|
|
1393
|
+
"(#2) = 0;" },
|
|
1394
|
+
{ "moncmp", "is_inf", "int16",
|
|
1395
|
+
"(#2) = 0;" },
|
|
1396
|
+
{ "moncmp", "is_inf", "uint16",
|
|
1397
|
+
"(#2) = 0;" },
|
|
1398
|
+
{ "moncmp", "is_inf", "int32",
|
|
1399
|
+
"(#2) = 0;" },
|
|
1400
|
+
{ "moncmp", "is_inf", "uint32",
|
|
1401
|
+
"(#2) = 0;" },
|
|
1402
|
+
{ "moncmp", "is_inf", "int64",
|
|
1403
|
+
"(#2) = 0;" },
|
|
1404
|
+
{ "moncmp", "is_inf", "uint64",
|
|
1405
|
+
"(#2) = 0;" },
|
|
1406
|
+
{ "moncmp", "is_inf", "float32",
|
|
1407
|
+
"(#2) = isinf(#1);" },
|
|
1408
|
+
{ "moncmp", "is_inf", "float64",
|
|
1409
|
+
"(#2) = isinf(#1);" },
|
|
1410
|
+
{ "moncmp", "is_inf", "cmplx64",
|
|
1411
|
+
"(#2) = (isinf(creal(#1)) || isinf(cimag(#1)));" },
|
|
1412
|
+
{ "moncmp", "is_inf", "cmplx128",
|
|
1413
|
+
"(#2) = (isinf(creal(#1)) || isinf(cimag(#1)));" },
|
|
1414
|
+
{ "moncmp", "is_finite", "int8",
|
|
1415
|
+
"(#2) = 1;" },
|
|
1416
|
+
{ "moncmp", "is_finite", "uint8",
|
|
1417
|
+
"(#2) = 1;" },
|
|
1418
|
+
{ "moncmp", "is_finite", "int16",
|
|
1419
|
+
"(#2) = 1;" },
|
|
1420
|
+
{ "moncmp", "is_finite", "uint16",
|
|
1421
|
+
"(#2) = 1;" },
|
|
1422
|
+
{ "moncmp", "is_finite", "int32",
|
|
1423
|
+
"(#2) = 1;" },
|
|
1424
|
+
{ "moncmp", "is_finite", "uint32",
|
|
1425
|
+
"(#2) = 1;" },
|
|
1426
|
+
{ "moncmp", "is_finite", "int64",
|
|
1427
|
+
"(#2) = 1;" },
|
|
1428
|
+
{ "moncmp", "is_finite", "uint64",
|
|
1429
|
+
"(#2) = 1;" },
|
|
1430
|
+
{ "moncmp", "is_finite", "float32",
|
|
1431
|
+
"(#2) = isfinite(#1);" },
|
|
1432
|
+
{ "moncmp", "is_finite", "float64",
|
|
1433
|
+
"(#2) = isfinite(#1);" },
|
|
1434
|
+
{ "moncmp", "is_finite", "cmplx64",
|
|
1435
|
+
"(#2) = (isfinite(creal(#1)) && isfinite(cimag(#1)));" },
|
|
1436
|
+
{ "moncmp", "is_finite", "cmplx128",
|
|
1437
|
+
"(#2) = (isfinite(creal(#1)) && isfinite(cimag(#1)));" },
|
|
1438
|
+
{ "moncmp", "is_invalid", "int8",
|
|
1439
|
+
"(#2) = 0;" },
|
|
1440
|
+
{ "moncmp", "is_invalid", "uint8",
|
|
1441
|
+
"(#2) = 0;" },
|
|
1442
|
+
{ "moncmp", "is_invalid", "int16",
|
|
1443
|
+
"(#2) = 0;" },
|
|
1444
|
+
{ "moncmp", "is_invalid", "uint16",
|
|
1445
|
+
"(#2) = 0;" },
|
|
1446
|
+
{ "moncmp", "is_invalid", "int32",
|
|
1447
|
+
"(#2) = 0;" },
|
|
1448
|
+
{ "moncmp", "is_invalid", "uint32",
|
|
1449
|
+
"(#2) = 0;" },
|
|
1450
|
+
{ "moncmp", "is_invalid", "int64",
|
|
1451
|
+
"(#2) = 0;" },
|
|
1452
|
+
{ "moncmp", "is_invalid", "uint64",
|
|
1453
|
+
"(#2) = 0;" },
|
|
1454
|
+
{ "moncmp", "is_invalid", "float32",
|
|
1455
|
+
"(#2) = (!isfinite(#1));" },
|
|
1456
|
+
{ "moncmp", "is_invalid", "float64",
|
|
1457
|
+
"(#2) = (!isfinite(#1));" },
|
|
1458
|
+
{ "moncmp", "is_invalid", "boolean",
|
|
1459
|
+
"(#2) = 0;" },
|
|
1460
|
+
{ "moncmp", "is_invalid", "cmplx64",
|
|
1461
|
+
"(#2) = (!isfinite(creal(#1)) || !isfinite(cimag(#1)));" },
|
|
1462
|
+
{ "moncmp", "is_invalid", "cmplx128",
|
|
1463
|
+
"(#2) = (!isfinite(creal(#1)) || !isfinite(cimag(#1)));" },
|
|
1464
|
+
{ "moncmp", "signbit", "int8",
|
|
1465
|
+
"(#2) = ((#1) < 0);" },
|
|
1466
|
+
{ "moncmp", "signbit", "uint8",
|
|
1467
|
+
"(#2) = 0;" },
|
|
1468
|
+
{ "moncmp", "signbit", "int16",
|
|
1469
|
+
"(#2) = ((#1) < 0);" },
|
|
1470
|
+
{ "moncmp", "signbit", "uint16",
|
|
1471
|
+
"(#2) = 0;" },
|
|
1472
|
+
{ "moncmp", "signbit", "int32",
|
|
1473
|
+
"(#2) = ((#1) < 0);" },
|
|
1474
|
+
{ "moncmp", "signbit", "uint32",
|
|
1475
|
+
"(#2) = 0;" },
|
|
1476
|
+
{ "moncmp", "signbit", "int64",
|
|
1477
|
+
"(#2) = ((#1) < 0);" },
|
|
1478
|
+
{ "moncmp", "signbit", "uint64",
|
|
1479
|
+
"(#2) = 0;" },
|
|
1480
|
+
{ "moncmp", "signbit", "float32",
|
|
1481
|
+
"(#2) = signbit(#1);" },
|
|
1482
|
+
{ "moncmp", "signbit", "float64",
|
|
1483
|
+
"(#2) = signbit(#1);" },
|
|
1484
|
+
{ "bincmp", "feq", "float32",
|
|
1485
|
+
"\n <type> f1a = fabs((float64_t) #1);\n <type> f2a = fabs((float64_t) #2);\n <type> fmax = (f1a > f2a) ? f1a : f2a;\n (#3) = ( fabs(((float64_t) #1)-( (float64_t) #2)) <= fmax * <epsilon> ) ? 1 : 0;\n " },
|
|
1486
|
+
{ "bincmp", "feq", "float64",
|
|
1487
|
+
"\n <type> f1a = fabs((float64_t) #1);\n <type> f2a = fabs((float64_t) #2);\n <type> fmax = (f1a > f2a) ? f1a : f2a;\n (#3) = ( fabs(((float64_t) #1)-( (float64_t) #2)) <= fmax * <epsilon> ) ? 1 : 0;\n " },
|
|
1488
|
+
{ "bincmp", "eq", "int8",
|
|
1489
|
+
"(#3) = ( (#1) == (#2) );" },
|
|
1490
|
+
{ "bincmp", "eq", "uint8",
|
|
1491
|
+
"(#3) = ( (#1) == (#2) );" },
|
|
1492
|
+
{ "bincmp", "eq", "int16",
|
|
1493
|
+
"(#3) = ( (#1) == (#2) );" },
|
|
1494
|
+
{ "bincmp", "eq", "uint16",
|
|
1495
|
+
"(#3) = ( (#1) == (#2) );" },
|
|
1496
|
+
{ "bincmp", "eq", "int32",
|
|
1497
|
+
"(#3) = ( (#1) == (#2) );" },
|
|
1498
|
+
{ "bincmp", "eq", "uint32",
|
|
1499
|
+
"(#3) = ( (#1) == (#2) );" },
|
|
1500
|
+
{ "bincmp", "eq", "int64",
|
|
1501
|
+
"(#3) = ( (#1) == (#2) );" },
|
|
1502
|
+
{ "bincmp", "eq", "uint64",
|
|
1503
|
+
"(#3) = ( (#1) == (#2) );" },
|
|
1504
|
+
{ "bincmp", "eq", "float32",
|
|
1505
|
+
"(#3) = ( (#1) == (#2) );" },
|
|
1506
|
+
{ "bincmp", "eq", "float64",
|
|
1507
|
+
"(#3) = ( (#1) == (#2) );" },
|
|
1508
|
+
{ "bincmp", "eq", "boolean",
|
|
1509
|
+
"(#3) = ( (#1) == (#2) );" },
|
|
1510
|
+
{ "bincmp", "eq", "cmplx64",
|
|
1511
|
+
"(#3) = (#1) == (#2);" },
|
|
1512
|
+
{ "bincmp", "eq", "cmplx128",
|
|
1513
|
+
"(#3) = (#1) == (#2);" },
|
|
1514
|
+
{ "bincmp", "ne", "int8",
|
|
1515
|
+
"(#3) = ( (#1) != (#2) );" },
|
|
1516
|
+
{ "bincmp", "ne", "uint8",
|
|
1517
|
+
"(#3) = ( (#1) != (#2) );" },
|
|
1518
|
+
{ "bincmp", "ne", "int16",
|
|
1519
|
+
"(#3) = ( (#1) != (#2) );" },
|
|
1520
|
+
{ "bincmp", "ne", "uint16",
|
|
1521
|
+
"(#3) = ( (#1) != (#2) );" },
|
|
1522
|
+
{ "bincmp", "ne", "int32",
|
|
1523
|
+
"(#3) = ( (#1) != (#2) );" },
|
|
1524
|
+
{ "bincmp", "ne", "uint32",
|
|
1525
|
+
"(#3) = ( (#1) != (#2) );" },
|
|
1526
|
+
{ "bincmp", "ne", "int64",
|
|
1527
|
+
"(#3) = ( (#1) != (#2) );" },
|
|
1528
|
+
{ "bincmp", "ne", "uint64",
|
|
1529
|
+
"(#3) = ( (#1) != (#2) );" },
|
|
1530
|
+
{ "bincmp", "ne", "float32",
|
|
1531
|
+
"(#3) = ( (#1) != (#2) );" },
|
|
1532
|
+
{ "bincmp", "ne", "float64",
|
|
1533
|
+
"(#3) = ( (#1) != (#2) );" },
|
|
1534
|
+
{ "bincmp", "ne", "boolean",
|
|
1535
|
+
"(#3) = ( (#1) != (#2) );" },
|
|
1536
|
+
{ "bincmp", "ne", "cmplx64",
|
|
1537
|
+
"(#3) = (#1) != (#2);" },
|
|
1538
|
+
{ "bincmp", "ne", "cmplx128",
|
|
1539
|
+
"(#3) = (#1) != (#2);" },
|
|
1540
|
+
{ "bincmp", "gt", "int8",
|
|
1541
|
+
"(#3) = ( (#1) > (#2) );" },
|
|
1542
|
+
{ "bincmp", "gt", "uint8",
|
|
1543
|
+
"(#3) = ( (#1) > (#2) );" },
|
|
1544
|
+
{ "bincmp", "gt", "int16",
|
|
1545
|
+
"(#3) = ( (#1) > (#2) );" },
|
|
1546
|
+
{ "bincmp", "gt", "uint16",
|
|
1547
|
+
"(#3) = ( (#1) > (#2) );" },
|
|
1548
|
+
{ "bincmp", "gt", "int32",
|
|
1549
|
+
"(#3) = ( (#1) > (#2) );" },
|
|
1550
|
+
{ "bincmp", "gt", "uint32",
|
|
1551
|
+
"(#3) = ( (#1) > (#2) );" },
|
|
1552
|
+
{ "bincmp", "gt", "int64",
|
|
1553
|
+
"(#3) = ( (#1) > (#2) );" },
|
|
1554
|
+
{ "bincmp", "gt", "uint64",
|
|
1555
|
+
"(#3) = ( (#1) > (#2) );" },
|
|
1556
|
+
{ "bincmp", "gt", "float32",
|
|
1557
|
+
"(#3) = ( (#1) > (#2) );" },
|
|
1558
|
+
{ "bincmp", "gt", "float64",
|
|
1559
|
+
"(#3) = ( (#1) > (#2) );" },
|
|
1560
|
+
{ "bincmp", "gt", "boolean",
|
|
1561
|
+
"(#3) = ( (#1) > (#2) );" },
|
|
1562
|
+
{ "bincmp", "lt", "int8",
|
|
1563
|
+
"(#3) = ( (#1) < (#2) );" },
|
|
1564
|
+
{ "bincmp", "lt", "uint8",
|
|
1565
|
+
"(#3) = ( (#1) < (#2) );" },
|
|
1566
|
+
{ "bincmp", "lt", "int16",
|
|
1567
|
+
"(#3) = ( (#1) < (#2) );" },
|
|
1568
|
+
{ "bincmp", "lt", "uint16",
|
|
1569
|
+
"(#3) = ( (#1) < (#2) );" },
|
|
1570
|
+
{ "bincmp", "lt", "int32",
|
|
1571
|
+
"(#3) = ( (#1) < (#2) );" },
|
|
1572
|
+
{ "bincmp", "lt", "uint32",
|
|
1573
|
+
"(#3) = ( (#1) < (#2) );" },
|
|
1574
|
+
{ "bincmp", "lt", "int64",
|
|
1575
|
+
"(#3) = ( (#1) < (#2) );" },
|
|
1576
|
+
{ "bincmp", "lt", "uint64",
|
|
1577
|
+
"(#3) = ( (#1) < (#2) );" },
|
|
1578
|
+
{ "bincmp", "lt", "float32",
|
|
1579
|
+
"(#3) = ( (#1) < (#2) );" },
|
|
1580
|
+
{ "bincmp", "lt", "float64",
|
|
1581
|
+
"(#3) = ( (#1) < (#2) );" },
|
|
1582
|
+
{ "bincmp", "lt", "boolean",
|
|
1583
|
+
"(#3) = ( (#1) < (#2) );" },
|
|
1584
|
+
{ "bincmp", "ge", "int8",
|
|
1585
|
+
"(#3) = ( (#1) >= (#2) );" },
|
|
1586
|
+
{ "bincmp", "ge", "uint8",
|
|
1587
|
+
"(#3) = ( (#1) >= (#2) );" },
|
|
1588
|
+
{ "bincmp", "ge", "int16",
|
|
1589
|
+
"(#3) = ( (#1) >= (#2) );" },
|
|
1590
|
+
{ "bincmp", "ge", "uint16",
|
|
1591
|
+
"(#3) = ( (#1) >= (#2) );" },
|
|
1592
|
+
{ "bincmp", "ge", "int32",
|
|
1593
|
+
"(#3) = ( (#1) >= (#2) );" },
|
|
1594
|
+
{ "bincmp", "ge", "uint32",
|
|
1595
|
+
"(#3) = ( (#1) >= (#2) );" },
|
|
1596
|
+
{ "bincmp", "ge", "int64",
|
|
1597
|
+
"(#3) = ( (#1) >= (#2) );" },
|
|
1598
|
+
{ "bincmp", "ge", "uint64",
|
|
1599
|
+
"(#3) = ( (#1) >= (#2) );" },
|
|
1600
|
+
{ "bincmp", "ge", "float32",
|
|
1601
|
+
"(#3) = ( (#1) >= (#2) );" },
|
|
1602
|
+
{ "bincmp", "ge", "float64",
|
|
1603
|
+
"(#3) = ( (#1) >= (#2) );" },
|
|
1604
|
+
{ "bincmp", "ge", "boolean",
|
|
1605
|
+
"(#3) = ( (#1) >= (#2) );" },
|
|
1606
|
+
{ "bincmp", "le", "int8",
|
|
1607
|
+
"(#3) = ( (#1) <= (#2) );" },
|
|
1608
|
+
{ "bincmp", "le", "uint8",
|
|
1609
|
+
"(#3) = ( (#1) <= (#2) );" },
|
|
1610
|
+
{ "bincmp", "le", "int16",
|
|
1611
|
+
"(#3) = ( (#1) <= (#2) );" },
|
|
1612
|
+
{ "bincmp", "le", "uint16",
|
|
1613
|
+
"(#3) = ( (#1) <= (#2) );" },
|
|
1614
|
+
{ "bincmp", "le", "int32",
|
|
1615
|
+
"(#3) = ( (#1) <= (#2) );" },
|
|
1616
|
+
{ "bincmp", "le", "uint32",
|
|
1617
|
+
"(#3) = ( (#1) <= (#2) );" },
|
|
1618
|
+
{ "bincmp", "le", "int64",
|
|
1619
|
+
"(#3) = ( (#1) <= (#2) );" },
|
|
1620
|
+
{ "bincmp", "le", "uint64",
|
|
1621
|
+
"(#3) = ( (#1) <= (#2) );" },
|
|
1622
|
+
{ "bincmp", "le", "float32",
|
|
1623
|
+
"(#3) = ( (#1) <= (#2) );" },
|
|
1624
|
+
{ "bincmp", "le", "float64",
|
|
1625
|
+
"(#3) = ( (#1) <= (#2) );" },
|
|
1626
|
+
{ "bincmp", "le", "boolean",
|
|
1627
|
+
"(#3) = ( (#1) <= (#2) );" },
|
|
1628
|
+
{ "bincmp", "is_close", "int8",
|
|
1629
|
+
"(#3) = (fabs((double)(#1) - (double)(#2)) <= tol) ? 1 : 0;" },
|
|
1630
|
+
{ "bincmp", "is_close", "uint8",
|
|
1631
|
+
"(#3) = (fabs((double)(#1) - (double)(#2)) <= tol) ? 1 : 0;" },
|
|
1632
|
+
{ "bincmp", "is_close", "int16",
|
|
1633
|
+
"(#3) = (fabs((double)(#1) - (double)(#2)) <= tol) ? 1 : 0;" },
|
|
1634
|
+
{ "bincmp", "is_close", "uint16",
|
|
1635
|
+
"(#3) = (fabs((double)(#1) - (double)(#2)) <= tol) ? 1 : 0;" },
|
|
1636
|
+
{ "bincmp", "is_close", "int32",
|
|
1637
|
+
"(#3) = (fabs((double)(#1) - (double)(#2)) <= tol) ? 1 : 0;" },
|
|
1638
|
+
{ "bincmp", "is_close", "uint32",
|
|
1639
|
+
"(#3) = (fabs((double)(#1) - (double)(#2)) <= tol) ? 1 : 0;" },
|
|
1640
|
+
{ "bincmp", "is_close", "int64",
|
|
1641
|
+
"(#3) = (fabs((double)(#1) - (double)(#2)) <= tol) ? 1 : 0;" },
|
|
1642
|
+
{ "bincmp", "is_close", "uint64",
|
|
1643
|
+
"(#3) = (fabs((double)(#1) - (double)(#2)) <= tol) ? 1 : 0;" },
|
|
1644
|
+
{ "bincmp", "is_close", "float32",
|
|
1645
|
+
"(#3) = (fabs((#1) - (#2)) <= tol) ? 1 : 0;" },
|
|
1646
|
+
{ "bincmp", "is_close", "float64",
|
|
1647
|
+
"(#3) = (fabs((#1) - (#2)) <= tol) ? 1 : 0;" },
|
|
1648
|
+
{ "bincmp", "is_close", "cmplx64",
|
|
1649
|
+
"(#3) = (cabs((#1) - (#2)) <= tol) ? 1 : 0;" },
|
|
1650
|
+
{ "bincmp", "is_close", "cmplx128",
|
|
1651
|
+
"(#3) = (cabs((#1) - (#2)) <= tol) ? 1 : 0;" },
|
|
1652
|
+
{ "bincmp", "is_equiv", "int8",
|
|
1653
|
+
"\n double a = (double)(#1), b = (double)(#2);\n double aa = fabs(a), ab = fabs(b);\n double m = (aa > ab) ? aa : ab;\n (#3) = ( (#1) == (#2) || (m > 0 && fabs(a - b) / m <= tol) ) ? 1 : 0;\n " },
|
|
1654
|
+
{ "bincmp", "is_equiv", "uint8",
|
|
1655
|
+
"\n double a = (double)(#1), b = (double)(#2);\n double aa = fabs(a), ab = fabs(b);\n double m = (aa > ab) ? aa : ab;\n (#3) = ( (#1) == (#2) || (m > 0 && fabs(a - b) / m <= tol) ) ? 1 : 0;\n " },
|
|
1656
|
+
{ "bincmp", "is_equiv", "int16",
|
|
1657
|
+
"\n double a = (double)(#1), b = (double)(#2);\n double aa = fabs(a), ab = fabs(b);\n double m = (aa > ab) ? aa : ab;\n (#3) = ( (#1) == (#2) || (m > 0 && fabs(a - b) / m <= tol) ) ? 1 : 0;\n " },
|
|
1658
|
+
{ "bincmp", "is_equiv", "uint16",
|
|
1659
|
+
"\n double a = (double)(#1), b = (double)(#2);\n double aa = fabs(a), ab = fabs(b);\n double m = (aa > ab) ? aa : ab;\n (#3) = ( (#1) == (#2) || (m > 0 && fabs(a - b) / m <= tol) ) ? 1 : 0;\n " },
|
|
1660
|
+
{ "bincmp", "is_equiv", "int32",
|
|
1661
|
+
"\n double a = (double)(#1), b = (double)(#2);\n double aa = fabs(a), ab = fabs(b);\n double m = (aa > ab) ? aa : ab;\n (#3) = ( (#1) == (#2) || (m > 0 && fabs(a - b) / m <= tol) ) ? 1 : 0;\n " },
|
|
1662
|
+
{ "bincmp", "is_equiv", "uint32",
|
|
1663
|
+
"\n double a = (double)(#1), b = (double)(#2);\n double aa = fabs(a), ab = fabs(b);\n double m = (aa > ab) ? aa : ab;\n (#3) = ( (#1) == (#2) || (m > 0 && fabs(a - b) / m <= tol) ) ? 1 : 0;\n " },
|
|
1664
|
+
{ "bincmp", "is_equiv", "int64",
|
|
1665
|
+
"\n double a = (double)(#1), b = (double)(#2);\n double aa = fabs(a), ab = fabs(b);\n double m = (aa > ab) ? aa : ab;\n (#3) = ( (#1) == (#2) || (m > 0 && fabs(a - b) / m <= tol) ) ? 1 : 0;\n " },
|
|
1666
|
+
{ "bincmp", "is_equiv", "uint64",
|
|
1667
|
+
"\n double a = (double)(#1), b = (double)(#2);\n double aa = fabs(a), ab = fabs(b);\n double m = (aa > ab) ? aa : ab;\n (#3) = ( (#1) == (#2) || (m > 0 && fabs(a - b) / m <= tol) ) ? 1 : 0;\n " },
|
|
1668
|
+
{ "bincmp", "is_equiv", "float32",
|
|
1669
|
+
"\n <type> aa = fabs((#1)), ab = fabs((#2));\n <type> m = (aa > ab) ? aa : ab;\n (#3) = ( (#1) == (#2) || (m > 0 && fabs((#1) - (#2)) / m <= tol) ) ? 1 : 0;\n " },
|
|
1670
|
+
{ "bincmp", "is_equiv", "float64",
|
|
1671
|
+
"\n <type> aa = fabs((#1)), ab = fabs((#2));\n <type> m = (aa > ab) ? aa : ab;\n (#3) = ( (#1) == (#2) || (m > 0 && fabs((#1) - (#2)) / m <= tol) ) ? 1 : 0;\n " },
|
|
1672
|
+
{ "bincmp", "is_equiv", "cmplx64",
|
|
1673
|
+
"\n double aa = cabs((#1)), ab = cabs((#2));\n double m = (aa > ab) ? aa : ab;\n (#3) = ( m > 0 && cabs((#1) - (#2)) / m <= tol ) ? 1 : 0;\n " },
|
|
1674
|
+
{ "bincmp", "is_equiv", "cmplx128",
|
|
1675
|
+
"\n double aa = cabs((#1)), ab = cabs((#2));\n double m = (aa > ab) ? aa : ab;\n (#3) = ( m > 0 && cabs((#1) - (#2)) / m <= tol ) ? 1 : 0;\n " },
|
|
1085
1676
|
};
|
|
1086
1677
|
|
|
1087
|
-
static const int ca_kernel_bodies_count =
|
|
1678
|
+
static const int ca_kernel_bodies_count = 823;
|
|
1088
1679
|
|
|
1089
1680
|
/* CArray.__kernel_body__(kind, name, data_type) -> String, or nil where
|
|
1090
1681
|
this operation has no body at that data type. */
|
|
@@ -1109,15 +1700,25 @@ rb_ca_s_kernel_body (VALUE klass, VALUE rkind, VALUE rname, VALUE rtype)
|
|
|
1109
1700
|
return Qnil;
|
|
1110
1701
|
}
|
|
1111
1702
|
|
|
1703
|
+
/* CArray.__kernel_helpers__ -> String: the C a body above may call
|
|
1704
|
+
beyond the standard library, for the caller to put ahead of it. */
|
|
1705
|
+
static const char ca_kernel_helpers[] =
|
|
1706
|
+
"\n/* x raised to a non-negative power, by binary exponentiation. The power\n is unsigned so that the magnitude of INT64_MIN is representable. */\n#define op_powi_magnitude(type) \\\nstatic inline type \\\nop_powi_mag_## type (type x, uint64_t p) \\\n{ \\\n type r=1; \\\n\\\n switch(p) { \\\n case 2: return x*x; \\\n case 3: return x*x*x; \\\n case 0: return 1; \\\n case 1: return x; \\\n } \\\n while (p) { \\\n if ( p & 1 ) r *= x; \\\n x *= x; \\\n p >>= 1; \\\n } \\\n return r; \\\n}\n\n/* An integer to a negative power is 1/x^|p| in integer division, which is\n decided by x alone: 0 has no reciprocal, 1 and -1 keep their magnitude,\n anything larger truncates to 0. Answering it that way, rather than by\n computing x^|p|, keeps a power that overflows -- 2 ** -64 -- from wrapping\n to a zero divisor. */\n#define op_powi(type) \\\nop_powi_magnitude(type) \\\nstatic inline type \\\nop_powi_## type (type x, int64_t p) \\\n{ \\\n if ( p < 0 ) { \\\n if ( x == 0 ) ca_zerodiv(); \\\n if ( x == 1 ) return 1; \\\n if ( x == (type) -1 ) return ( p % 2 ) ? x : 1; \\\n return 0; \\\n } \\\n return op_powi_mag_## type(x, (uint64_t) p); \\\n}\n\n/* An unsigned x is never -1; (type) -1 is its largest value. */\n#define op_powi_u(type) \\\nop_powi_magnitude(type) \\\nstatic inline type \\\nop_powi_## type (type x, int64_t p) \\\n{ \\\n if ( p < 0 ) { \\\n if ( x == 0 ) ca_zerodiv(); \\\n return ( x == 1 ) ? 1 : 0; \\\n } \\\n return op_powi_mag_## type(x, (uint64_t) p); \\\n}\n\n#define op_powi_fc(type) \\\nop_powi_magnitude(type) \\\nstatic inline type \\\nop_powi_## type (type x, int64_t p) \\\n{ \\\n if ( p < 0 ) { \\\n return 1 / op_powi_mag_## type(x, (uint64_t) 0 - (uint64_t) p); \\\n } \\\n return op_powi_mag_## type(x, (uint64_t) p); \\\n}\n\nop_powi(int8_t)\nop_powi_u(uint8_t)\nop_powi(int16_t)\nop_powi_u(uint16_t)\nop_powi(int32_t)\nop_powi_u(uint32_t)\nop_powi(int64_t)\nop_powi_u(uint64_t)\nop_powi_fc(float32_t)\nop_powi_fc(float64_t)\n\n";
|
|
1707
|
+
|
|
1708
|
+
static VALUE
|
|
1709
|
+
rb_ca_s_kernel_helpers (VALUE klass)
|
|
1710
|
+
{
|
|
1711
|
+
return rb_str_new_cstr(ca_kernel_helpers);
|
|
1712
|
+
}
|
|
1713
|
+
|
|
1112
1714
|
void
|
|
1113
1715
|
Init_carray_kernels (void)
|
|
1114
1716
|
{
|
|
1115
1717
|
Init_carray_kernels_reduce_aggregate();
|
|
1116
1718
|
Init_carray_kernels_reduce_extreme();
|
|
1117
|
-
Init_carray_kernels_reduce_cumulative();
|
|
1118
1719
|
Init_carray_kernels_reduce_boolean();
|
|
1119
1720
|
Init_carray_kernels_reduce_variance();
|
|
1120
|
-
|
|
1721
|
+
Init_carray_kernels_reduce_cumulative();
|
|
1121
1722
|
Init_carray_kernels_scan();
|
|
1122
1723
|
Init_carray_kernels_sort();
|
|
1123
1724
|
Init_carray_kernels_search();
|
|
@@ -1128,4 +1729,6 @@ Init_carray_kernels (void)
|
|
|
1128
1729
|
Init_carray_kernels_bincmp();
|
|
1129
1730
|
rb_define_singleton_method(rb_cCArray, "__kernel_body__",
|
|
1130
1731
|
rb_ca_s_kernel_body, 3);
|
|
1732
|
+
rb_define_singleton_method(rb_cCArray, "__kernel_helpers__",
|
|
1733
|
+
rb_ca_s_kernel_helpers, 0);
|
|
1131
1734
|
}
|