-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathTASO_02_run_memorysafe.patch
306 lines (303 loc) · 11.6 KB
/
TASO_02_run_memorysafe.patch
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
Index: include/taso/ops.h
IDEA additional info:
Subsystem: com.intellij.openapi.diff.impl.patch.CharsetEP
<+>UTF-8
===================================================================
--- include/taso/ops.h (revision 9ea6c78a6e1584c441c9c205c3b3f73592ff0d51)
+++ include/taso/ops.h (revision 2589ce88b052373161fbcdfe8adc9bc09454974b)
@@ -660,6 +660,7 @@
bool has_loop(void);
float total_cost(void);
float run();
+ float run_memorysafe();
void print_costs(void);
void print_measurements(void);
#ifdef TRT
Index: python/taso/_cython/CCore.pxd
IDEA additional info:
Subsystem: com.intellij.openapi.diff.impl.patch.CharsetEP
<+>UTF-8
===================================================================
--- python/taso/_cython/CCore.pxd (revision 9ea6c78a6e1584c441c9c205c3b3f73592ff0d51)
+++ python/taso/_cython/CCore.pxd (revision 2589ce88b052373161fbcdfe8adc9bc09454974b)
@@ -270,3 +270,4 @@
void print_measurements()
float total_cost()
float run()
+ float run_memorysafe()
Index: python/taso/_cython/core.pyx
IDEA additional info:
Subsystem: com.intellij.openapi.diff.impl.patch.CharsetEP
<+>UTF-8
===================================================================
--- python/taso/_cython/core.pyx (revision 9ea6c78a6e1584c441c9c205c3b3f73592ff0d51)
+++ python/taso/_cython/core.pyx (revision 2589ce88b052373161fbcdfe8adc9bc09454974b)
@@ -175,6 +175,9 @@
def run_time(self):
return self.p_graph.run()
+ def run_time_memorysafe(self):
+ return self.p_graph.run_memorysafe()
+
def cost(self):
return self.p_graph.total_cost()
Index: src/core/ops.cc
IDEA additional info:
Subsystem: com.intellij.openapi.diff.impl.patch.CharsetEP
<+>UTF-8
===================================================================
--- src/core/ops.cc (revision 9ea6c78a6e1584c441c9c205c3b3f73592ff0d51)
+++ src/core/ops.cc (revision 2589ce88b052373161fbcdfe8adc9bc09454974b)
@@ -1362,6 +1362,254 @@
return model->measure_oplist_runtime(opBaseList);
}
+
+float Graph::run_memorysafe(void)
+{
+ std::map<Op, int, OpCompare> todos;
+ std::map<Op, std::set<Edge, EdgeCompare>, OpCompare>::const_iterator it;
+ std::vector<Op> opList;
+ std::vector<OpBase*> opBaseList;
+ for (it = inEdges.begin(); it != inEdges.end(); it++) {
+ int cnt = 0;
+ std::set<Edge, EdgeCompare> inList = it->second;
+ std::set<Edge, EdgeCompare>::const_iterator it2;
+ for (it2 = inList.begin(); it2 != inList.end(); it2++) {
+ if (it2->srcOp.guid > GUID_PRESERVED) cnt ++;
+ }
+ todos[it->first] = cnt;
+ if (todos[it->first] == 0)
+ opList.push_back(it->first);
+ }
+ size_t i = 0;
+ while (i < opList.size()) {
+ Op op = opList[i++];
+ std::set<Edge, EdgeCompare> outList = outEdges[op];
+ std::set<Edge, EdgeCompare> inList = inEdges[op];
+ std::set<Edge, EdgeCompare>::const_iterator it2;
+ assert(inList.size() > 0);
+ OpBase* opPtr = NULL;
+ // Step 1: prepare inputs
+ Tensor inputs[MAX_NUM_INPUTS];
+ if ((op.ptr->type == OP_INPUT) || (op.ptr->type == OP_WEIGHT)) {
+ assert(inList.size() == 1);
+ //Edge e = *inList.begin();
+ //assert(e.srcOp.ptr == NULL); // NoOp's input must not be any Op
+ Tensor t = op.ptr->inputs[0];
+ size_t size = sizeof(DATATYPE);
+ for (int j = 0; j < t.numDim; j++)
+ size *= t.dim[j];
+ if (op.ptr->type == OP_INPUT) {
+ assert(t.data_ptr == NULL);
+ t.data_ptr = (DATATYPE*) model->allocate_memory(size);
+ } else {
+ assert(t.data_ptr != NULL);
+ }
+ inputs[0] = t;
+ } else {
+ for (it2 = inList.begin(); it2 != inList.end(); it2++) {
+ size_t idx2 = 0;
+ for (idx2 = 0; idx2 < opList.size(); idx2++) {
+ if (opList[idx2].guid == it2->srcOp.guid) break;
+ }
+ assert(idx2 < i);
+ assert(inputs[it2->dstIdx].data_ptr == NULL); // No duplicated dstIdxes
+ inputs[it2->dstIdx] = opBaseList[idx2]->outputs[it2->srcIdx];
+ }
+ }
+#ifdef DEADCODE
+ // Step 1: prepare inputs
+ for (it2 = inList.begin(); it2 != inList.end(); it2++) {
+ Edge e = *it2;
+ if (e.srcOp.guid == GUID_INPUT) {
+ Tensor t = op.ptr->inputs[e.dstIdx];
+ t.ptr = (DATATYPE*) model->allocate_memory(sizeof(DATATYPE) * t.size());
+ assert(inputs[e.dstIdx].ptr == NULL); // No duplicated dstIdxes
+ inputs[e.dstIdx] = t;
+ } else if (e.srcOp.guid = GUID_WEIGHT) {
+ Tensor t = op.ptr->inputs[e.dstIdx];
+ t.ptr = (DATATYPE*) model->allocate_memory(sizeof(DATATYPE) * t.size());
+ assert(inputs[e.dstIdx].ptr == NULL); // No duplicated dstIdxes
+ inputs[e.dstIdx] = t;
+ } else {
+ size_t idx2 = 0;
+ for (idx2 = 0; idx2 < opList.size(); idx2++) {
+ if (opList[idx2].guid == e.srcOp.guid) break;
+ }
+ assert(idx2 < i);
+ assert(inputs[e.dstIdx].ptr == NULL); // No duplicated dstIdxes
+ inputs[e.dstIdx] = opBaseList[idx2]->outputs[it2->srcIdx];
+ }
+ }
+#endif
+ // Step 2: create Ops
+ switch (op.ptr->type) {
+ case OP_CONV2D:
+ {
+ Conv2D* conv = (Conv2D*) op.ptr;
+ assert(inList.size() == 2);
+ opPtr = new Conv2D(model, inputs[0], inputs[1],
+ conv->strideH, conv->strideW,
+ conv->padding, conv->activation);
+#ifdef USE_CUDNN
+ ((Conv2D*)opPtr)->fwdAlgo = conv->fwdAlgo;
+#endif
+ break;
+ }
+ case OP_MATMUL:
+ {
+ Matmul* matmul = (Matmul*) op.ptr;
+ assert(inList.size() == 2);
+ opPtr = new Matmul(model, inputs[0], inputs[1], matmul->activation);
+ break;
+ }
+ case OP_RESHAPE:
+ {
+ Reshape* reshape = (Reshape*) op.ptr;
+ assert(inList.size() == 1);
+ std::vector<int> shape;
+ for (int i = 0; i < reshape->outputs[0].numDim; i++)
+ shape.push_back(reshape->outputs[0].dim[i]);
+ opPtr = new Reshape(model, inputs[0], shape);
+ break;
+ }
+ case OP_TRANSPOSE:
+ {
+ Transpose* transpose = (Transpose*) op.ptr;
+ assert(inList.size() == 1);
+ int ndim = inputs[0].numDim, permIdx = transpose->permIdx;
+ std::vector<int> permVec;
+ int permArray[MAX_DIM];
+ for (int i = ndim - 1; i >= 0; i--) {
+ permArray[i] = permIdx % ndim;
+ permIdx = permIdx / ndim;
+ }
+ assert(permIdx == 0);
+ for (int i = 0; i < ndim; i++)
+ for (int j = i + 1; j < ndim; j++)
+ assert(permArray[i] != permArray[j]);
+ for (int i = 0; i < ndim; i++)
+ permVec.push_back(permArray[i]);
+ opPtr = new Transpose(model, inputs[0], permVec, transpose->shuffle);
+ break;
+ }
+ case OP_EW_ADD:
+ case OP_EW_MUL:
+ {
+ //Element* element = (Element*) op.ptr;
+ assert(inList.size() == 2);
+ opPtr = new Element(model, op.ptr->type, inputs[0], inputs[1]);
+ break;
+ }
+ case OP_ENLARGE:
+ {
+ //Enlarge* enlarge = (Enlarge*) op.ptr;
+ assert(inList.size() == 2);
+ opPtr = new Enlarge(model, inputs[0], inputs[1]);
+ break;
+ }
+ case OP_MERGE_GCONV:
+ {
+ MergeGConv* merge = (MergeGConv*) op.ptr;
+ assert(inList.size() == 1);
+ opPtr = new MergeGConv(model, inputs[0], merge->count);
+ break;
+ }
+ case OP_POOL2D_MAX:
+ case OP_POOL2D_AVG:
+ {
+ Pool2D* pool = (Pool2D*) op.ptr;
+ assert(inList.size() == 2);
+ opPtr = new Pool2D(model, inputs[0], inputs[1], pool->type,
+ pool->kernelH, pool->kernelW,
+ pool->strideH, pool->strideW,
+ pool->padding, pool->activation);
+ break;
+ }
+ case OP_RELU:
+ case OP_SIGMOID:
+ case OP_TANH:
+ {
+ Activation* act = (Activation*) op.ptr;
+ assert(inList.size() == 1);
+ opPtr = new Activation(model, inputs[0], act->type, act->inPlace);
+ break;
+ }
+ case OP_BATCHNORM:
+ {
+ assert(inList.size() == 5);
+ opPtr = new BatchNorm(model, inputs[0], inputs[1], inputs[2], inputs[3], inputs[4]);
+ break;
+ }
+ case OP_SPLIT:
+ {
+ Split* split = (Split*) op.ptr;
+ assert(inList.size() == 1);
+ opPtr = new Split(model, inputs[0], split->axis, split->sizes);
+ break;
+ }
+ case OP_INPUT:
+ case OP_WEIGHT:
+ case OP_DROPOUT:
+ {
+ assert(inList.size() == 1);
+ opPtr = new NoOp(model, inputs[0], op.ptr->type);
+ break;
+ }
+ case OP_CONCAT:
+ {
+ Concat* concat = (Concat*) op.ptr;
+ opPtr = new Concat(model, concat->axis, inList.size(), inputs, concat->needCopy);
+ break;
+ }
+ default:
+ printf("op.type = %d\n", op.ptr->type);
+ assert(false);
+ }
+ // Step 3: map new Op
+ opPtr->map();
+ opBaseList.push_back(opPtr);
+ for (it2 = outList.begin(); it2 != outList.end(); it2++) {
+ todos[it2->dstOp] --;
+ //printf("myOp(%zu) dstOp(%zu) dstType(%d) dstTodos(%d)\n",
+ // it2->srcOp.guid, it2->dstOp.guid,
+ // it2->dstOp.ptr->type, todos[it2->dstOp]);
+ if (todos[it2->dstOp] == 0) {
+ opList.push_back(it2->dstOp);
+ }
+ }
+ }
+#ifdef VERBOSE_PRINTS
+ for (int i =0; i < opList.size(); i++) {
+ printf("opList[%d]: guid(%zu) type(%d)\n", i, opList[i].guid,
+ opList[i].ptr->type);
+ }
+ for (it = inEdges.begin(); it != inEdges.end(); it++) {
+ printf("op: guid(%zu) type(%d)\n", it->first.guid, it->first.ptr->type);
+ std::set<Edge, EdgeCompare> inList = it->second;
+ std::set<Edge, EdgeCompare>::const_iterator it2;
+ int cnt = 0;
+ for (it2 = inList.begin(); it2 != inList.end(); it2++) {
+ printf(" inEdge[%d]: srcOp(%zu) srcIdx(%d) dstOp(%zu) dstIdx(%d)\n", cnt++, it2->srcOp.guid, it2->srcIdx, it2->dstOp.guid, it2->dstIdx);
+ }
+ }
+#endif
+
+ assert(opList.size() == inEdges.size());
+ assert(opList.size() == opBaseList.size());
+
+ float result = model->measure_oplist_runtime(opBaseList);
+
+ // Now free GPU memory from the opList
+ for (int i = 0; i < opBaseList.size(); i++) {
+ OpBase* opBase = opBaseList[i];
+ opBase->unmap();
+ free(opBase);
+ opBase = nullptr;
+ }
+
+ return result;
+}
+
void Graph::print_costs(void)
{
float exe_time = 0, flops = 0, mem_acc = 0;