forked from BachiLi/loma_public
-
Notifications
You must be signed in to change notification settings - Fork 0
/
codegen_openMpi.py
377 lines (345 loc) · 16.6 KB
/
codegen_openMpi.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
import codegen_c
import ir
ir.generate_asdl_file()
import _asdl.loma as loma_ir
import compiler
class OpemMpiCodegenVisitor(codegen_c.CCodegenVisitor):
""" Generates MPI code from loma IR.
See https://ispc.github.io/index.html for more details about ispc.
"""
def __init__(self, func_defs):
super().__init__(func_defs)
def visit_function_def(self, node: loma_ir.FunctionDef):
reverse_diff = False
self.code += f'{codegen_c.type_to_string(node.ret_type)} {node.id}('
for i, arg in enumerate(node.args):
if '_d' in arg.id:
reverse_diff = True
if i > 0:
self.code += ', '
self.code += f'{codegen_c.type_to_string(arg)} {arg.id}'
if node.is_simd:
if len(node.args) > 0:
self.code += ', '
self.code += 'int __total_work'
self.code += ') {\n'
self.byref_args = set([arg.id for arg in node.args if \
arg.i == loma_ir.Out() and (not isinstance(arg.t, loma_ir.Array))])
self.forward_out_args = set([arg.id for arg in node.args if arg.i == loma_ir.Out() and (isinstance(arg.t, loma_ir.Array))])
self.tab_count += 1
if node.is_simd:
self.emit_tabs()
self.code += 'for (int __work_id = 0; __work_id < __total_work; __work_id++) {\n'
self.tab_count += 1
for stmt in node.body:
self.visit_stmt(stmt)
if node.is_simd:
self.tab_count -= 1
self.emit_tabs()
self.code += '}\n'
self.tab_count -= 1
self.code += '}\n'
## openMpi code
if node.is_openMpi and node.id.startswith("d_"):
new_node_id = node.id + '_mpi_worker'
self.code += f'void {new_node_id}()'
self.code += '{\n'
self.code += """
MPI_Init(NULL, NULL);
MPI_Comm parent_comm;
MPI_Comm_get_parent(&parent_comm);
if (parent_comm == MPI_COMM_NULL) {
MPI_Finalize();
return;
}
int world_rank, world_size;
MPI_Comm_rank(MPI_COMM_WORLD, &world_rank);
MPI_Comm_size(MPI_COMM_WORLD, &world_size);
"""
for arg in node.args:
if arg.id in self.byref_args:
self.code += f"{codegen_c.type_to_string(arg.t)} {arg.id};\n"
elif isinstance(arg.t,loma_ir.Array):
self.code += f"int {arg.id}_size;\n"
self.code += f"MPI_Recv(&{arg.id}_size, 1, MPI_INT, 0, 0, parent_comm,MPI_STATUS_IGNORE);\n"
if isinstance(arg.t.t,loma_ir.Int):
self.code += f"int* {arg.id} = (int*)malloc({arg.id}_size * sizeof(int));"
self.code += f"MPI_Recv({arg.id}, {arg.id}_size * sizeof(int), MPI_INT, 0, 0, parent_comm,MPI_STATUS_IGNORE);\n"
else:
if reverse_diff:
self.code += f"float* {arg.id} = (float*)malloc({arg.id}_size * sizeof(float));"
self.code += f"MPI_Recv({arg.id}, {arg.id}_size * sizeof(float), MPI_BYTE, 0, 0, parent_comm,MPI_STATUS_IGNORE);\n"
else:
self.code += f"_dfloat* {arg.id} = (_dfloat*)malloc({arg.id}_size * sizeof(_dfloat));"
self.code += f"MPI_Recv({arg.id}, {arg.id}_size * sizeof(_dfloat), MPI_BYTE, 0, 0, parent_comm,MPI_STATUS_IGNORE);\n"
# self.code += f"""
# for(int g=0;g<{arg.id}_size;g++){{
# printf("Value: %d, Dvalue: %d \\n", {arg.id}[g].val,{arg.id}[g].dval);
# }}
# """
# self.code += f"_dfloat* {arg.id} = (_dfloat*)malloc(4* sizeof(_dfloat));"
# self.code += f"MPI_Recv({arg.id}, sizeof(_dfloat)*4, MPI_BYTE, 0, 0, parent_comm,MPI_STATUS_IGNORE);\n"
elif isinstance(arg.t,loma_ir.Int):
self.code += f"int {arg.id};\n"
self.code += f"MPI_Recv(&{arg.id}, 1, MPI_INT, 0, 0, parent_comm,MPI_STATUS_IGNORE);\n"
elif isinstance(arg.t,loma_ir.Float):
self.code += f"float {arg.id};\n"
self.code += f"MPI_Recv(&{arg.id}, 1, MPI_FLOAT, 0, 0, parent_comm,MPI_STATUS_IGNORE);\n"
else:
self.code += f"{codegen_c.type_to_string(arg.t)} {arg.id};\n"
self.code += f"MPI_Recv(&{arg.id}, sizeof({arg.id}), MPI_BYTE, 0, 0, parent_comm,MPI_STATUS_IGNORE);\n"
temp_string = ''
for i, arg in enumerate(node.args):
if i > 0:
temp_string += ', '
if arg.id in self.byref_args:
temp_string += f"&{arg.id}"
else:
temp_string += f'{arg.id}'
temp_string += ");"
if codegen_c.type_to_string(node.ret_type) == "void":
self.code += f"{node.id}("
self.code += temp_string
for arg in node.args:
if arg.id in self.forward_out_args:
self.code += f"MPI_Send(&{arg.id}_size, 1, MPI_INT, 0, 0, parent_comm);\n"
if reverse_diff:
self.code += f"MPI_Send({arg.id}, {arg.id}_size * sizeof(float), MPI_BYTE, 0, 0, parent_comm);\n"
else:
self.code += f"MPI_Send({arg.id}, {arg.id}_size * sizeof(_dfloat), MPI_BYTE, 0, 0, parent_comm);\n"
else:
self.code += f"{codegen_c.type_to_string(node.ret_type)} out = {node.id}("
self.code += temp_string
self.code += f"""
MPI_Send(&out, sizeof({codegen_c.type_to_string(node.ret_type)}), MPI_BYTE, 0, 0, parent_comm);
"""
for arg in node.args:
if arg.id in self.byref_args:
if isinstance(arg.t,loma_ir.Int):
self.code+=f"""MPI_Send(&{arg.id}, 1, MPI_INT, 0, 0, parent_comm);"""
elif isinstance(arg.t,loma_ir.Float):
self.code += f"MPI_Send(&{arg.id}, 1, MPI_FLOAT, 0, 0, parent_comm);\n"
else:
self.code += f"MPI_Send(&{arg.id}, sizeof({codegen_c.type_to_string(node.ret_type)}), MPI_BYTE, 0, 0, parent_comm);"
self.code += """\nMPI_Finalize();"""
self.tab_count -= 1
self.code += '}\n'
## Write main function
self.code += f"""
int main() {{
{new_node_id}();
return 0;
}}
"""
def create_parent_code(self, node: loma_ir.FunctionDef, output_filename: str):
reverse_diff = False
array_number = 0
self.byref_args = set([arg.id for arg in node.args if \
arg.i == loma_ir.Out() and (not isinstance(arg.t, loma_ir.Array))])
self.forward_out_args = set([arg.id for arg in node.args if arg.i == loma_ir.Out() and (isinstance(arg.t, loma_ir.Array))])
new_node_id = node.id + '_mpi_worker'
self.code += "void mpi_runner("
for i, arg in enumerate(node.args):
if '_d' in arg.id:
reverse_diff = True
if i > 0:
self.code += ', '
if isinstance(arg.t, loma_ir.Array):
self.code += f'{codegen_c.type_to_string(arg)} {arg.id}'
self.code += f',int* {arg.id}_size'
array_number +=1
elif arg.id in self.byref_args:
self.code += f'{codegen_c.type_to_string(arg)} {arg.id}'
else:
self.code += f'{codegen_c.type_to_string(arg)}* {arg.id}'
if len(node.args) > 0:
self.code += ', '
if codegen_c.type_to_string(node.ret_type) != 'void':
self.code += f'{codegen_c.type_to_string(node.ret_type)}* output, '
self.code += 'int __total_work'
self.code += ') {\n'
self.code += f"""
MPI_Init(NULL, NULL);
int world_rank, world_size;
MPI_Comm_rank(MPI_COMM_WORLD, &world_rank);
MPI_Comm_size(MPI_COMM_WORLD, &world_size);
if (world_rank == 0) {{
// This is the parent process
MPI_Comm child_comm;
MPI_Info info;
MPI_Info_create(&info);
// Spawn NUM_CHILDREN child processes
MPI_Comm_spawn("{output_filename}", MPI_ARGV_NULL, __total_work, info, 0, MPI_COMM_SELF, &child_comm, MPI_ERRCODES_IGNORE);
int* array_end = (int*)malloc({array_number} * sizeof(int));
int end = 0;
int array_index = 0;
for (int i = 0; i< {array_number};i++){{
array_end[i] = 0;
}}
for (int i = 0; i < __total_work; i++) {{
array_index = 0;
// Send input to child process
"""
for arg in node.args:
if arg.id in self.byref_args:
continue
if isinstance(arg.t,loma_ir.Array):
self.code += f"MPI_Send(&{arg.id}_size[i], 1, MPI_INT, i, 0, child_comm);\n"
if isinstance(arg.t.t,loma_ir.Int):
self.code += f"""
int* {arg.id}_i = (int*)malloc({arg.id}_size[i] * sizeof(int));
"""
else:
if reverse_diff:
self.code += f"""
float* {arg.id}_i = (float*)malloc({arg.id}_size[i] * sizeof(float));
"""
else:
self.code += f"""
_dfloat* {arg.id}_i = (_dfloat*)malloc({arg.id}_size[i] * sizeof(_dfloat));
"""
self.code += f"""
end = array_end[array_index];
for (int j = 0; j < {arg.id}_size[i]; j++) {{
{arg.id}_i[j] = {arg.id}[end++];
}}
array_end[array_index] = end;
array_index = array_index + 1;
"""
if reverse_diff:
self.code += f"MPI_Send({arg.id}_i, sizeof(float)*{arg.id}_size[i], MPI_BYTE, i, 0, child_comm);\n"
else:
self.code += f"MPI_Send({arg.id}_i, sizeof(_dfloat)*{arg.id}_size[i], MPI_BYTE, i, 0, child_comm);\n"
elif isinstance(arg.t,loma_ir.Int):
self.code += f"MPI_Send(&{arg.id}[i], 1, MPI_INT, i, 0, child_comm);\n"
elif isinstance(arg.t,loma_ir.Float):
self.code += f"MPI_Send(&{arg.id}[i], 1, MPI_FLOAT, i, 0, child_comm);\n"
else:
self.code += f"MPI_Send(&{arg.id}[i], sizeof({arg.id}[i]), MPI_BYTE, i, 0, child_comm);\n"
self.code += "}\n"
tmp_string = ''
for args in node.args:
if args.id in self.byref_args:
tmp_string+= f"{codegen_c.type_to_string(args.t)} _temp_{args.id};"
if isinstance(args.t,loma_ir.Int):
tmp_string += f"MPI_Recv(&_temp_{args.id}, 1, MPI_INT, i, 0, child_comm, MPI_STATUS_IGNORE);\n"
elif isinstance(args.t,loma_ir.Float):
tmp_string += f"MPI_Recv(&_temp_{args.id}, 1, MPI_FLOAT, i, 0, child_comm, MPI_STATUS_IGNORE);\n"
else:
tmp_string += f"MPI_Recv(&_temp_{args.id}, 1, sizeof({codegen_c.type_to_string(node.ret_type)}), i, 0, child_comm, MPI_STATUS_IGNORE);\n"
tmp_string += f"{args.id}[i] = _temp_{args.id};\n"
elif args.id in self.forward_out_args:
tmp_string += f"int {args.id}_size_recv;\n"
tmp_string += f"MPI_Recv(&{args.id}_size_recv, 1, MPI_INT, i, 0, child_comm, MPI_STATUS_IGNORE);\n"
if reverse_diff:
tmp_string += f"float* {args.id}_temp = (float*)malloc({args.id}_size_recv * sizeof(float));"
tmp_string += f"MPI_Recv({args.id}_temp, {args.id}_size_recv* sizeof(float), MPI_BYTE, i, 0, child_comm,MPI_STATUS_IGNORE);\n"
else:
tmp_string += f"_dfloat* {args.id}_temp = (_dfloat*)malloc({args.id}_size_recv * sizeof(_dfloat));"
tmp_string += f"MPI_Recv({args.id}_temp, {args.id}_size_recv* sizeof(_dfloat), MPI_BYTE, i, 0, child_comm,MPI_STATUS_IGNORE);\n"
tmp_string += f"""
rend = rec_end[rec_array_index];
for (int j = 0; j < {args.id}_size_recv; j++) {{
{args.id}[rend] = {args.id}_temp[j];
rend++;
}}
rec_end[rec_array_index] = rend;
rec_array_index = rec_array_index + 1;
"""
if codegen_c.type_to_string(node.ret_type) != "void":
self.code += f"""
// Communicate with children and receive responses
for (int i = 0; i < __total_work; i++) {{
{codegen_c.type_to_string(node.ret_type)} local_output;
MPI_Recv(&local_output, sizeof({codegen_c.type_to_string(node.ret_type)}), MPI_BYTE, i, 0, child_comm, MPI_STATUS_IGNORE);
{tmp_string}
output[i] = local_output;
}}
MPI_Info_free(&info);
}}
MPI_Finalize();
}}"""
else:
self.code += f"""
// Communicate with children and receive responses
int* rec_end = (int*)malloc({array_number} * sizeof(int));
int rend = 0;
int rec_array_index = 0;
for (int i = 0; i< {array_number};i++){{
rec_end[i] = 0;
}}
for (int i = 0; i < __total_work; i++) {{
rec_array_index = 0;
{tmp_string}
}}
"""
self.code += """MPI_Info_free(&info);} MPI_Finalize();}"""
def codegen_openMpiChild(structs : dict[str, loma_ir.Struct],
funcs : dict[str, loma_ir.func],
) -> str:
""" Given loma Structs (structs) and loma functions (funcs),
return a string that represents the equivalent ISPC code.
Parameters:
structs - a dictionary that maps the ID of a Struct to
the corresponding Struct
funcs - a dictionary that maps the ID of a function to
the corresponding func
"""
sorted_structs_list = compiler.topo_sort_structs(structs)
# Definition of structs
code = ''
for s in sorted_structs_list:
code += f'typedef struct {{\n'
for m in s.members:
# Special rule for arrays
if isinstance(m.t, loma_ir.Array) and m.t.static_size is not None:
code += f'\t{codegen_c.type_to_string(m.t.t)} {m.id}[{m.t.static_size}];\n'
else:
code += f'\t{codegen_c.type_to_string(m.t)} {m.id};\n'
code += f'}} {s.id};\n'
# Forward declaration of functions
for f in funcs.values():
code += f'{codegen_c.type_to_string(f.ret_type)} {f.id}('
for i, arg in enumerate(f.args):
if i > 0:
code += ', '
code += f'{codegen_c.type_to_string(arg)} {arg.id}'
if f.is_simd:
if len(f.args) > 0:
code += ', '
code += 'int __total_work'
code += ');\n'
for f in funcs.values():
cg = OpemMpiCodegenVisitor(funcs)
cg.visit_function(f)
code += cg.code
return code
def codegen_openMpiParent(structs : dict[str, loma_ir.Struct],
funcs : dict[str, loma_ir.func],
output_filename: str) -> str:
""" Given loma Structs (structs) and loma functions (funcs),
return a string that represents the equivalent ISPC code.
Parameters:
structs - a dictionary that maps the ID of a Struct to
the corresponding Struct
funcs - a dictionary that maps the ID of a function to
the corresponding func
"""
sorted_structs_list = compiler.topo_sort_structs(structs)
# Definition of structs
code = ''
for s in sorted_structs_list:
code += f'typedef struct {{\n'
for m in s.members:
# Special rule for arrays
if isinstance(m.t, loma_ir.Array) and m.t.static_size is not None:
code += f'\t{codegen_c.type_to_string(m.t.t)} {m.id}[{m.t.static_size}];\n'
else:
code += f'\t{codegen_c.type_to_string(m.t)} {m.id};\n'
code += f'}} {s.id};\n'
to_call_f = [f for f in funcs.values() if f.is_openMpi and f.id.startswith("d_")]
cg = OpemMpiCodegenVisitor(funcs)
if len(to_call_f) == 1:
cg.create_parent_code(to_call_f[0], output_filename)
code += cg.code
return code