1function codegen_axb_template (multop, bmult, imult, fmult, dmult, fcmult, dcmult, no_min_max_any_times_monoids)
2%CODEGEN_AXB_TEMPLATE create a function for a semiring with a TxT->T multiplier
3
4% SuiteSparse:GraphBLAS, Timothy A. Davis, (c) 2017-2021, All Rights Reserved.
5% SPDX-License-Identifier: Apache-2.0
6
7fprintf ('\n%-7s', multop) ;
8if (nargin < 8)
9    no_min_max_any_times_monoids = false ;
10end
11
12plusinf32 = 'INFINITY' ;
13neginf32  = '(-INFINITY)' ;
14plusinf64 = '((double) INFINITY)' ;
15neginf64  = '((double) -INFINITY)' ;
16
17% MIN monoid: all are terminal.  None can be done with OpenMP atomic update
18if (~no_min_max_any_times_monoids)
19    add = 'w = GB_IMIN (w, t)' ;
20    addfunc = 'GB_IMIN (w, t)' ;
21    codegen_axb_method ('min', multop, add, addfunc, imult, 'int8_t'  , 'int8_t'  , 'INT8_MAX'  , 'INT8_MIN'  , 0, 0) ;
22    codegen_axb_method ('min', multop, add, addfunc, imult, 'int16_t' , 'int16_t' , 'INT16_MAX' , 'INT16_MIN' , 0, 0) ;
23    codegen_axb_method ('min', multop, add, addfunc, imult, 'int32_t' , 'int32_t' , 'INT32_MAX' , 'INT32_MIN' , 0, 0) ;
24    codegen_axb_method ('min', multop, add, addfunc, imult, 'int64_t' , 'int64_t' , 'INT64_MAX' , 'INT64_MIN' , 0, 0) ;
25    codegen_axb_method ('min', multop, add, addfunc, imult, 'uint8_t' , 'uint8_t' , 'UINT8_MAX' , '0'         , 0, 0) ;
26    codegen_axb_method ('min', multop, add, addfunc, imult, 'uint16_t', 'uint16_t', 'UINT16_MAX', '0'         , 0, 0) ;
27    codegen_axb_method ('min', multop, add, addfunc, imult, 'uint32_t', 'uint32_t', 'UINT32_MAX', '0'         , 0, 0) ;
28    codegen_axb_method ('min', multop, add, addfunc, imult, 'uint64_t', 'uint64_t', 'UINT64_MAX', '0'         , 0, 0) ;
29    add = 'w = fminf (w, t)' ;
30    addfunc = 'fminf (w, t)' ;
31    codegen_axb_method ('min', multop, add, addfunc, fmult, 'float'   , 'float'   , plusinf32   , neginf32    , 0, 0) ;
32    add = 'w = fmin (w, t)' ;
33    addfunc = 'fmin (w, t)' ;
34    codegen_axb_method ('min', multop, add, addfunc, dmult, 'double'  , 'double'  , plusinf64   , neginf64    , 0, 0) ;
35end
36
37% MAX monoid: all are terminal.  None can be done with OpenMP atomic update
38if (~no_min_max_any_times_monoids)
39    add = 'w = GB_IMAX (w, t)' ;
40    addfunc = 'GB_IMAX (w, t)' ;
41    codegen_axb_method ('max', multop, add, addfunc, imult, 'int8_t'  , 'int8_t'  , 'INT8_MIN'  , 'INT8_MAX'  , 0, 0) ;
42    codegen_axb_method ('max', multop, add, addfunc, imult, 'int16_t' , 'int16_t' , 'INT16_MIN' , 'INT16_MAX' , 0, 0) ;
43    codegen_axb_method ('max', multop, add, addfunc, imult, 'int32_t' , 'int32_t' , 'INT32_MIN' , 'INT32_MAX' , 0, 0) ;
44    codegen_axb_method ('max', multop, add, addfunc, imult, 'int64_t' , 'int64_t' , 'INT64_MIN' , 'INT64_MAX' , 0, 0) ;
45    codegen_axb_method ('max', multop, add, addfunc, imult, 'uint8_t' , 'uint8_t' , '0'         , 'UINT8_MAX' , 0, 0) ;
46    codegen_axb_method ('max', multop, add, addfunc, imult, 'uint16_t', 'uint16_t', '0'         , 'UINT16_MAX', 0, 0) ;
47    codegen_axb_method ('max', multop, add, addfunc, imult, 'uint32_t', 'uint32_t', '0'         , 'UINT32_MAX', 0, 0) ;
48    codegen_axb_method ('max', multop, add, addfunc, imult, 'uint64_t', 'uint64_t', '0'         , 'UINT64_MAX', 0, 0) ;
49    % floating-point MAX must use unsigned integer puns for compare-and-swap
50    add = 'w = fmaxf (w, t)' ;
51    addfunc = 'fmaxf (w, t)' ;
52    codegen_axb_method ('max', multop, add, addfunc, fmult, 'float'   , 'float'   , neginf32    , plusinf32   , 0, 0) ;
53    add = 'w = fmax (w, t)' ;
54    addfunc = 'fmax (w, t)' ;
55    codegen_axb_method ('max', multop, add, addfunc, dmult, 'double'  , 'double'  , neginf64    , plusinf64   , 0, 0) ;
56end
57
58% ANY monoid: all are terminal.
59if (~no_min_max_any_times_monoids)
60    add = 'w = t' ;
61    addfunc = 't' ;
62    codegen_axb_method ('any', multop, add, addfunc, imult, 'int8_t'  , 'int8_t'  , '0' , [ ], 0, 0) ;
63    codegen_axb_method ('any', multop, add, addfunc, imult, 'int16_t' , 'int16_t' , '0' , [ ], 0, 0) ;
64    codegen_axb_method ('any', multop, add, addfunc, imult, 'int32_t' , 'int32_t' , '0' , [ ], 0, 0) ;
65    codegen_axb_method ('any', multop, add, addfunc, imult, 'int64_t' , 'int64_t' , '0' , [ ], 0, 0) ;
66    codegen_axb_method ('any', multop, add, addfunc, imult, 'uint8_t' , 'uint8_t' , '0' , [ ], 0, 0) ;
67    codegen_axb_method ('any', multop, add, addfunc, imult, 'uint16_t', 'uint16_t', '0' , [ ], 0, 0) ;
68    codegen_axb_method ('any', multop, add, addfunc, imult, 'uint32_t', 'uint32_t', '0' , [ ], 0, 0) ;
69    codegen_axb_method ('any', multop, add, addfunc, imult, 'uint64_t', 'uint64_t', '0' , [ ], 0, 0) ;
70    codegen_axb_method ('any', multop, add, addfunc, fmult, 'float'   , 'float'   , '0' , [ ], 0, 0) ;
71    codegen_axb_method ('any', multop, add, addfunc, dmult, 'double'  , 'double'  , '0' , [ ], 0, 0) ;
72    % complex case:
73    id = 'GxB_CMPLXF(0,0)' ;
74    codegen_axb_method ('any', multop, add, addfunc, fcmult, 'GxB_FC32_t', 'GxB_FC32_t', id, [ ], 0, 0) ;
75    id = 'GxB_CMPLX(0,0)' ;
76    codegen_axb_method ('any', multop, add, addfunc, dcmult, 'GxB_FC64_t', 'GxB_FC64_t', id, [ ], 0, 0) ;
77end
78
79% PLUS monoid: none are terminal.  All reals can be done with OpenMP atomic update
80add = 'w += t' ;
81addfunc = 'w + t' ;
82codegen_axb_method ('plus', multop, add, addfunc, imult, 'int8_t'  , 'int8_t'  , '0', [ ], 1, 1) ;
83codegen_axb_method ('plus', multop, add, addfunc, imult, 'uint8_t' , 'uint8_t' , '0', [ ], 1, 1) ;
84codegen_axb_method ('plus', multop, add, addfunc, imult, 'int16_t' , 'int16_t' , '0', [ ], 1, 1) ;
85codegen_axb_method ('plus', multop, add, addfunc, imult, 'uint16_t', 'uint16_t', '0', [ ], 1, 1) ;
86codegen_axb_method ('plus', multop, add, addfunc, imult, 'int32_t' , 'int32_t' , '0', [ ], 1, 1) ;
87codegen_axb_method ('plus', multop, add, addfunc, imult, 'uint32_t', 'uint32_t', '0', [ ], 1, 1) ;
88codegen_axb_method ('plus', multop, add, addfunc, imult, 'int64_t' , 'int64_t' , '0', [ ], 1, 1) ;
89codegen_axb_method ('plus', multop, add, addfunc, imult, 'uint64_t', 'uint64_t', '0', [ ], 1, 1) ;
90codegen_axb_method ('plus', multop, add, addfunc, fmult, 'float'   , 'float'   , '0', [ ], 1, 1) ;
91codegen_axb_method ('plus', multop, add, addfunc, dmult, 'double'  , 'double'  , '0', [ ], 1, 1) ;
92% complex types done with two OpenMP atomic updates:
93add = 'w = GB_FC32_add (w, t)' ;
94addfunc = 'GB_FC32_add (w, t)' ;
95id = 'GxB_CMPLXF(0,0)' ;
96codegen_axb_method ('plus', multop, add, addfunc, fcmult, 'GxB_FC32_t', 'GxB_FC32_t', id, [ ], 1, 1) ;
97add = 'w = GB_FC64_add (w, t)' ;
98addfunc = 'GB_FC64_add (w, t)' ;
99id = 'GxB_CMPLX(0,0)' ;
100codegen_axb_method ('plus', multop, add, addfunc, dcmult, 'GxB_FC64_t', 'GxB_FC64_t', id, [ ], 1, 1) ;
101
102% TIMES monoid: integers are terminal, float and double are not.
103if (~no_min_max_any_times_monoids)
104    % All real types can be done with OpenMP atomic update
105    add = 'w *= t' ;
106    addfunc = 'w * t' ;
107    codegen_axb_method ('times', multop, add, addfunc, imult, 'int8_t'  , 'int8_t'  , '1', '0', 1, 1) ;
108    codegen_axb_method ('times', multop, add, addfunc, imult, 'uint8_t' , 'uint8_t' , '1', '0', 1, 1) ;
109    codegen_axb_method ('times', multop, add, addfunc, imult, 'int16_t' , 'int16_t' , '1', '0', 1, 1) ;
110    codegen_axb_method ('times', multop, add, addfunc, imult, 'uint16_t', 'uint16_t', '1', '0', 1, 1) ;
111    codegen_axb_method ('times', multop, add, addfunc, imult, 'int32_t' , 'int32_t' , '1', '0', 1, 1) ;
112    codegen_axb_method ('times', multop, add, addfunc, imult, 'uint32_t', 'uint32_t', '1', '0', 1, 1) ;
113    codegen_axb_method ('times', multop, add, addfunc, imult, 'int64_t' , 'int64_t' , '1', '0', 1, 1) ;
114    codegen_axb_method ('times', multop, add, addfunc, imult, 'uint64_t', 'uint64_t', '1', '0', 1, 1) ;
115    codegen_axb_method ('times', multop, add, addfunc, fmult, 'float'   , 'float'   , '1', [ ], 1, 1) ;
116    codegen_axb_method ('times', multop, add, addfunc, dmult, 'double'  , 'double'  , '1', [ ], 1, 1) ;
117    % complex types cannot be done with OpenMP atomic update:
118    add = 'w = GB_FC32_mul (w, t)' ;
119    addfunc = 'GB_FC32_mul (w, t)' ;
120    id = 'GxB_CMPLXF(1,0)' ;
121    codegen_axb_method ('times', multop, add, addfunc, fcmult, 'GxB_FC32_t', 'GxB_FC32_t', id, [ ], 0, 0) ;
122    add = 'w = GB_FC64_mul (w, t)' ;
123    addfunc = 'GB_FC64_mul (w, t)' ;
124    id = 'GxB_CMPLX(1,0)' ;
125    codegen_axb_method ('times', multop, add, addfunc, dcmult, 'GxB_FC64_t', 'GxB_FC64_t', id, [ ], 0, 0) ;
126end
127
128% boolean monoids: LOR, LAND are terminal; LXOR, EQ are not.
129% For gcc and icc: LOR, LAND, and LXOR can be done as OpenMP atomic updates; EQ cannot.
130% For MS Visual Studio: none can be done with OpenMP atomic updates.
131codegen_axb_method ('lor',  multop, 'w |= t', 'w | t', bmult, 'bool', 'bool', 'false', 'true' , 1, 0) ;
132codegen_axb_method ('land', multop, 'w &= t', 'w & t', bmult, 'bool', 'bool', 'true' , 'false', 1, 0) ;
133codegen_axb_method ('lxor', multop, 'w ^= t', 'w ^ t', bmult, 'bool', 'bool', 'false', [ ]    , 1, 0) ;
134codegen_axb_method ('any' , multop, 'w = t' , 't'    , bmult, 'bool', 'bool', '0'    , [ ]    , 0, 0) ;
135add = 'w = (w == t)' ;
136addfunc = 'w == t' ;
137codegen_axb_method ('eq',   multop, add,      addfunc, bmult, 'bool', 'bool', 'true' , [ ]    , 0, 0) ;
138
139