ccv_nnc_graph.c

Bug Summary

File:	nnc/ccv_nnc_graph.c
Warning:	line 352, column 7 Branch condition evaluates to a garbage value
Annotated Source Code

Press '?' to see keyboard shortcuts
Show analyzer invocation
clang -cc1 -cc1 -triple x86_64-unknown-linux-gnu -analyze -disable-free -clear-ast-before-backend -disable-llvm-verifier -discard-value-names -main-file-name ccv_nnc_graph.c -analyzer-checker=core -analyzer-checker=apiModeling -analyzer-checker=unix -analyzer-checker=deadcode -analyzer-checker=security.insecureAPI.UncheckedReturn -analyzer-checker=security.insecureAPI.getpw -analyzer-checker=security.insecureAPI.gets -analyzer-checker=security.insecureAPI.mktemp -analyzer-checker=security.insecureAPI.mkstemp -analyzer-checker=security.insecureAPI.vfork -analyzer-checker=nullability.NullPassedToNonnull -analyzer-checker=nullability.NullReturnedFromNonnull -analyzer-output plist -w -setup-static-analyzer -mrelocation-model pic -pic-level 2 -pic-is-pie -mframe-pointer=none -fmath-errno -ffp-contract=on -fno-rounding-math -mconstructor-aliases -funwind-tables=2 -target-cpu x86-64 -target-feature +sse2 -tune-cpu generic -debugger-tuning=gdb -fdebug-compilation-dir=/home/liu/actions-runner/_work/ccv/ccv/lib/nnc -fcoverage-compilation-dir=/home/liu/actions-runner/_work/ccv/ccv/lib/nnc -resource-dir /usr/local/lib/clang/19 -I ../ -I /usr/local/cuda/include -D HAVE_CBLAS -D HAVE_LIBPNG -D HAVE_LIBJPEG -D HAVE_FFTW3 -D HAVE_PTHREAD -D HAVE_LIBLINEAR -D HAVE_TESSERACT -D HAVE_AVCODEC -D HAVE_AVFORMAT -D HAVE_AVUTIL -D HAVE_SWSCALE -D HAVE_SSE2 -D HAVE_GSL -D HAVE_CUDA -D HAVE_CUDNN -D HAVE_NCCL -D USE_SYSTEM_CUB -I /usr/local/include -internal-isystem /usr/local/lib/clang/19/include -internal-isystem /usr/local/include -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/12/../../../../x86_64-linux-gnu/include -internal-externc-isystem /usr/include/x86_64-linux-gnu -internal-externc-isystem /include -internal-externc-isystem /usr/include -O3 -ferror-limit 19 -fgnuc-version=4.2.1 -fskip-odr-check-in-gmf -vectorize-loops -vectorize-slp -analyzer-output=html -faddrsig -D__GCC_HAVE_DWARF2_CFI_ASM=1 -o /home/liu/actions-runner/_work/ccv/ccv/_analyze/2024-12-03-120642-617563-1 -x c ccv_nnc_graph.c
1#include "ccv_nnc.h"
2#include "ccv_nnc_easy.h"
3#include "ccv_nnc_internal.h"
4#include "ccv_internal.h"
5#include "_ccv_nnc_graph.h"
6 
7// MARK - Level-2 API
8 
9ccv_nnc_graph_t* ccv_nnc_graph_new(void)
10{
11	ccv_nnc_graph_t* graph = (ccv_nnc_graph_t*)cccalloccalloc(1, sizeof(ccv_nnc_graph_t));
12	graph->exec_info = ccv_array_new(sizeof(ccv_nnc_graph_exec_info_t), 5, 0);
13	return graph;
14}
15 
16void ccv_nnc_graph_set_sources(ccv_nnc_graph_t* const graph, const ccv_nnc_graph_exec_t* const sources, const int source_size)
17{
18	if (!graph->sources)
19		graph->sources = ccv_array_new(sizeof(ccv_nnc_graph_exec_t), source_size, 0);
20	else
21		ccv_array_clear(graph->sources);
22	int i;
23	for (i = 0; i < source_size; i++)
24		ccv_array_push(graph->sources, sources + i);
25	graph->topsorted = 0;
26}
27 
28ccv_nnc_graph_exec_t* ccv_nnc_graph_sources(const ccv_nnc_graph_t* const graph)
29{
30	return graph->sources ? (ccv_nnc_graph_exec_t*)ccv_array_get(graph->sources, 0)((void*)(((char*)((graph->sources)->data)) + (size_t)(graph
->sources)->rsize * (size_t)(0))) : 0;
31}
32 
33int ccv_nnc_graph_source_size(const ccv_nnc_graph_t* const graph)
34{
35	return graph->sources ? graph->sources->rnum : 0;
36}
37 
38void ccv_nnc_graph_set_destinations(ccv_nnc_graph_t* const graph, const ccv_nnc_graph_exec_t* const destinations, const int destination_size)
39{
40	if (!graph->destinations)
41		graph->destinations = ccv_array_new(sizeof(ccv_nnc_graph_exec_t), destination_size, 0);
42	else
43		ccv_array_clear(graph->sources);
44	int i;
45	for (i = 0; i < destination_size; i++)
46		ccv_array_push(graph->destinations, destinations + i);
47	graph->topsorted = 0;
48}
49 
50ccv_nnc_graph_exec_t* ccv_nnc_graph_destinations(const ccv_nnc_graph_t* const graph)
51{
52	return graph->destinations ? (ccv_nnc_graph_exec_t*)ccv_array_get(graph->destinations, 0)((void*)(((char*)((graph->destinations)->data)) + (size_t
)(graph->destinations)->rsize * (size_t)(0))) : 0;
53}
54 
55int ccv_nnc_graph_destination_size(const ccv_nnc_graph_t* const graph)
56{
57	return graph->destinations ? graph->destinations->rnum : 0;
58}
59 
60void ccv_nnc_graph_exec_set(ccv_nnc_graph_t* const graph, const ccv_nnc_graph_exec_t exec, const ccv_nnc_cmd_t cmd)
61{
62	assert(exec.d < graph->exec_info->rnum)((void) sizeof ((exec.d < graph->exec_info->rnum) ? 1
 : 0), __extension__ ({ if (exec.d < graph->exec_info->
rnum) ; else __assert_fail ("exec.d < graph->exec_info->rnum"
, "ccv_nnc_graph.c", 62, __extension__ __PRETTY_FUNCTION__); }
));
63	assert(exec.graph == graph)((void) sizeof ((exec.graph == graph) ? 1 : 0), __extension__
 ({ if (exec.graph == graph) ; else __assert_fail ("exec.graph == graph"
, "ccv_nnc_graph.c", 63, __extension__ __PRETTY_FUNCTION__); }
));
64	ccv_nnc_graph_exec_info_t* const exec_info = (ccv_nnc_graph_exec_info_t*)ccv_array_get(graph->exec_info, exec.d)((void*)(((char*)((graph->exec_info)->data)) + (size_t)
(graph->exec_info)->rsize * (size_t)(exec.d)));
65	exec_info->cmd = cmd;
66}
67 
68ccv_nnc_cmd_t ccv_nnc_graph_exec_cmd(const ccv_nnc_graph_t* const graph, const ccv_nnc_graph_exec_t exec)
69{
70	assert(exec.d < graph->exec_info->rnum)((void) sizeof ((exec.d < graph->exec_info->rnum) ? 1
 : 0), __extension__ ({ if (exec.d < graph->exec_info->
rnum) ; else __assert_fail ("exec.d < graph->exec_info->rnum"
, "ccv_nnc_graph.c", 70, __extension__ __PRETTY_FUNCTION__); }
));
71	assert(exec.graph == graph)((void) sizeof ((exec.graph == graph) ? 1 : 0), __extension__
 ({ if (exec.graph == graph) ; else __assert_fail ("exec.graph == graph"
, "ccv_nnc_graph.c", 71, __extension__ __PRETTY_FUNCTION__); }
));
72	ccv_nnc_graph_exec_info_t* const exec_info = (ccv_nnc_graph_exec_info_t*)ccv_array_get(graph->exec_info, exec.d)((void*)(((char*)((graph->exec_info)->data)) + (size_t)
(graph->exec_info)->rsize * (size_t)(exec.d)));
73	return exec_info->cmd;
74}
75 
76void ccv_nnc_graph_exec_set_hint(ccv_nnc_graph_t* const graph, const ccv_nnc_graph_exec_t exec, const ccv_nnc_hint_t hint)
77{
78	assert(exec.d < graph->exec_info->rnum)((void) sizeof ((exec.d < graph->exec_info->rnum) ? 1
 : 0), __extension__ ({ if (exec.d < graph->exec_info->
rnum) ; else __assert_fail ("exec.d < graph->exec_info->rnum"
, "ccv_nnc_graph.c", 78, __extension__ __PRETTY_FUNCTION__); }
));
79	assert(exec.graph == graph)((void) sizeof ((exec.graph == graph) ? 1 : 0), __extension__
 ({ if (exec.graph == graph) ; else __assert_fail ("exec.graph == graph"
, "ccv_nnc_graph.c", 79, __extension__ __PRETTY_FUNCTION__); }
));
80	ccv_nnc_graph_exec_info_t* const exec_info = (ccv_nnc_graph_exec_info_t*)ccv_array_get(graph->exec_info, exec.d)((void*)(((char*)((graph->exec_info)->data)) + (size_t)
(graph->exec_info)->rsize * (size_t)(exec.d)));
81	exec_info->hint = hint;
82}
83 
84static int _ccv_nnc_tensor_multiview_level_count(const ccv_nnc_tensor_multiview_t* const mv)
85{
86	if (!CCV_IS_TENSOR_MULTIVIEW(mv)((*(int*)(mv)) & CCV_TENSOR_MULTIVIEW))
87		return 1;
88	const int count = mv->kind + mv->repeat;
89	int i, c = 0;
90	for (i = 0; i < count; i++)
91	{
92		ccv_nnc_tensor_t* tv = CCV_NNC_MULTIVIEW_DATA(mv)((mv)->_heap_data ? (mv)->_heap_data : (mv)->_inline_data
)[i];
93		if (tv == CCV_NNC_TENSOR_PLACEHOLDER((ccv_nnc_tensor_t*)(intptr_t)(0x10)))
94			c = ccv_max(c, 1)({ typeof (c) _a = (c); typeof (1) _b = (1); (_a > _b) ? _a
 : _b; });
95		else
96			c = ccv_max(c, _ccv_nnc_tensor_multiview_level_count((ccv_nnc_tensor_multiview_t*)tv))({ typeof (c) _a = (c); typeof (_ccv_nnc_tensor_multiview_level_count
((ccv_nnc_tensor_multiview_t*)tv)) _b = (_ccv_nnc_tensor_multiview_level_count
((ccv_nnc_tensor_multiview_t*)tv)); (_a > _b) ? _a : _b; }
);
97	}
98	return c + 1;
99}
100 
101static ccv_nnc_graph_tensor_wrap_t* _ccv_nnc_graph_tensor_wrap_new(const ccv_nnc_tensor_multiview_t* const mv)
102{
103	const int level_count = _ccv_nnc_tensor_multiview_level_count(mv);
104	ccv_nnc_graph_tensor_wrap_t* tensor_wrap = (ccv_nnc_graph_tensor_wrap_t*)ccmallocmalloc(sizeof(ccv_nnc_graph_tensor_wrap_t) + sizeof(ccv_nnc_tensor_t*) * (level_count - 1));
105	tensor_wrap->update_required = 0;
106	tensor_wrap->count = level_count;
107	tensor_wrap->index = 0;
108	tensor_wrap->tensors[0] = (ccv_nnc_tensor_t*)mv;
109	return tensor_wrap;
110}
111 
112static void _ccv_nnc_graph_exec_rewind(ccv_nnc_graph_exec_info_t* const info, ccv_nnc_graph_t* const graph)
113{
114	if (!info->tensor_wraps_ref)
115		return;
116	int i;
117	assert(info->tensor_wraps_ref <= graph->tensor_wraps->rnum)((void) sizeof ((info->tensor_wraps_ref <= graph->tensor_wraps
->rnum) ? 1 : 0), __extension__ ({ if (info->tensor_wraps_ref
 <= graph->tensor_wraps->rnum) ; else __assert_fail (
"info->tensor_wraps_ref <= graph->tensor_wraps->rnum"
, "ccv_nnc_graph.c", 117, __extension__ __PRETTY_FUNCTION__);
 }));
118	ccv_nnc_graph_tensor_wrap_array_t* const tensor_wrap_array = *(ccv_nnc_graph_tensor_wrap_array_t**)ccv_array_get(graph->tensor_wraps, info->tensor_wraps_ref - 1)((void*)(((char*)((graph->tensor_wraps)->data)) + (size_t
)(graph->tensor_wraps)->rsize * (size_t)(info->tensor_wraps_ref
 - 1)));;
119	// Rewind from tensor wraps.
120	for (i = 0; i < info->input_size; i++)
121		if (tensor_wrap_array->tensor_wraps[i])
122			info->inputs[i] = tensor_wrap_array->tensor_wraps[i]->tensors[0];
123	const int d = info->input_size;
124	for (i = 0; i < info->output_size; i++)
125		if (tensor_wrap_array->tensor_wraps[d + i])
126			info->outputs[i] = tensor_wrap_array->tensor_wraps[d + i]->tensors[0];
127	const int dd = info->input_size + info->output_size;
128	for (i = 0; i < info->update_size; i++)
129		if (tensor_wrap_array->tensor_wraps[dd + i])
130			info->updates[i] = tensor_wrap_array->tensor_wraps[dd + i]->tensors[0];
131}
132 
133static void _ccv_nnc_graph_tensor_wrap_free(ccv_nnc_graph_tensor_wrap_t* const tensor_wrap)
134{
135	ccfreefree(tensor_wrap);
136}
137 
138ccv_nnc_graph_tensor_wrap_array_t* ccv_nnc_get_tensor_wrap_array(ccv_nnc_graph_t* const graph, const int tensor_wrap_size, int* const tensor_wraps_ref)
139{
140	ccv_nnc_graph_tensor_wrap_array_t** tensor_wrap_array_ref = *tensor_wraps_ref ? (ccv_nnc_graph_tensor_wrap_array_t**)ccv_array_get(graph->tensor_wraps, *tensor_wraps_ref - 1)((void*)(((char*)((graph->tensor_wraps)->data)) + (size_t
)(graph->tensor_wraps)->rsize * (size_t)(*tensor_wraps_ref
 - 1))) : 0;
141	// Otherwise, find an open slot.
142	if (!tensor_wrap_array_ref)
143	{
144		if (!graph->tensor_wraps)
145			graph->tensor_wraps = ccv_array_new(sizeof(ccv_nnc_graph_tensor_wrap_array_t*), 0, 0);
146		ccv_nnc_graph_tensor_wrap_array_t* const tensor_wrap_array = 0;
147		ccv_array_push(graph->tensor_wraps, &tensor_wrap_array);
148		tensor_wrap_array_ref = (ccv_nnc_graph_tensor_wrap_array_t**)ccv_array_get(graph->tensor_wraps, graph->tensor_wraps->rnum - 1)((void*)(((char*)((graph->tensor_wraps)->data)) + (size_t
)(graph->tensor_wraps)->rsize * (size_t)(graph->tensor_wraps
->rnum - 1)));
149		*tensor_wraps_ref = graph->tensor_wraps->rnum;
150	}
151	int i;
152	if (*tensor_wrap_array_ref)
153	{
154		if ((*tensor_wrap_array_ref)->size != tensor_wrap_size)
155			*tensor_wrap_array_ref = (ccv_nnc_graph_tensor_wrap_array_t*)ccreallocrealloc(*tensor_wrap_array_ref, sizeof(ccv_nnc_graph_tensor_wrap_array_t) + sizeof(ccv_nnc_graph_tensor_wrap_t*) * (tensor_wrap_size - 1));
156		for (i = (*tensor_wrap_array_ref)->size; i < tensor_wrap_size; i++)
157			(*tensor_wrap_array_ref)->tensor_wraps[i] = 0;
158	} else
159		*tensor_wrap_array_ref = (ccv_nnc_graph_tensor_wrap_array_t*)cccalloccalloc(sizeof(ccv_nnc_graph_tensor_wrap_array_t) + sizeof(ccv_nnc_graph_tensor_wrap_t*) * (tensor_wrap_size - 1), 1);
160	ccv_nnc_graph_tensor_wrap_array_t* const tensor_wrap_array = *tensor_wrap_array_ref;
161	tensor_wrap_array->size = tensor_wrap_size;
162	return tensor_wrap_array;
163}
164 
165void ccv_nnc_set_tensor_wraps(ccv_nnc_graph_tensor_wrap_t** const tensor_wraps, ccv_nnc_tensor_t* const* const tensors, const int tensor_size)
166{
167	int i;
168	for (i = 0; i < tensor_size; i++)
169		if (tensors[i])
170		{
171			if (CCV_IS_TENSOR_MULTIVIEW(tensors[i])((*(int*)(tensors[i])) & CCV_TENSOR_MULTIVIEW) &&
172				((ccv_nnc_tensor_multiview_t*)tensors[i])->anchor != CCV_NNC_MULTIVIEW_PHI(intptr_t)0x1)
173			{
174				if (!tensor_wraps[i] || tensors[i] != tensor_wraps[i]->tensors[0])
175				{
176					if (tensor_wraps[i])
177						_ccv_nnc_graph_tensor_wrap_free(tensor_wraps[i]);
178					tensor_wraps[i] = _ccv_nnc_graph_tensor_wrap_new((ccv_nnc_tensor_multiview_t*)tensors[i]);
179				}
180			} else {
181				if (tensor_wraps[i])
182					_ccv_nnc_graph_tensor_wrap_free(tensor_wraps[i]);
183				tensor_wraps[i] = 0;
184			}
185		}
186}
187 
188void ccv_nnc_graph_register_tensor_wraps(ccv_nnc_graph_t* graph, const int tensor_wraps_ref_d)
189{
190	ccv_nnc_graph_t* p = graph;
191	const ccv_nnc_graph_tensor_wraps_ref_t tensor_wraps_ref = {
192		.d = tensor_wraps_ref_d,
193		.graph = graph,
194	};
195	do {
196		if (!p->tensor_wraps_refs)
197		{
198			p->tensor_wraps_refs = ccv_array_new(sizeof(ccv_nnc_graph_tensor_wraps_ref_t), 0, 0);
199			ccv_array_push(p->tensor_wraps_refs, &tensor_wraps_ref);
200		} else {
201			int i;
202			int has_tensor_wraps_ref = 0;
203			for (i = 0; !has_tensor_wraps_ref && i < p->tensor_wraps_refs->rnum; i++)
204			{
205				ccv_nnc_graph_tensor_wraps_ref_t* tensor_wraps_ref = (ccv_nnc_graph_tensor_wraps_ref_t*)ccv_array_get(p->tensor_wraps_refs, i)((void*)(((char*)((p->tensor_wraps_refs)->data)) + (size_t
)(p->tensor_wraps_refs)->rsize * (size_t)(i)));
206				has_tensor_wraps_ref = (tensor_wraps_ref->d == tensor_wraps_ref_d && tensor_wraps_ref->graph == graph);
207			}
208			if (!has_tensor_wraps_ref)
209				ccv_array_push(p->tensor_wraps_refs, &tensor_wraps_ref);
210		}
211		p = p->p;
212	} while (p);
213}
214 
215static void _ccv_nnc_graph_redo_tensor_wraps(ccv_nnc_graph_exec_info_t* const info, ccv_nnc_graph_t* const graph)
216{
217	int i;
218	const int has_wrap = ccv_nnc_tensors_have_wraps(info->inputs, info->input_size) ||
219		ccv_nnc_tensors_have_wraps(info->outputs, info->output_size) ||
220		ccv_nnc_tensors_have_wraps(info->updates, info->update_size);
221	if (has_wrap)
222	{
223		const int tensor_wrap_size = info->input_size + info->output_size + info->update_size;
224		ccv_nnc_graph_tensor_wrap_array_t* const tensor_wrap_array = ccv_nnc_get_tensor_wrap_array(graph, tensor_wrap_size, &info->tensor_wraps_ref);
225		ccv_nnc_set_tensor_wraps(tensor_wrap_array->tensor_wraps, info->inputs, info->input_size);
226		const int d = info->input_size;
227		ccv_nnc_set_tensor_wraps(tensor_wrap_array->tensor_wraps + d, info->outputs, info->output_size);
228		const int dd = info->input_size + info->output_size;
229		ccv_nnc_set_tensor_wraps(tensor_wrap_array->tensor_wraps + dd, info->updates, info->update_size);
230	} else if (info->tensor_wraps_ref) {
231		ccv_nnc_graph_tensor_wrap_array_t** tensor_wrap_array_ref = (ccv_nnc_graph_tensor_wrap_array_t**)ccv_array_get(graph->tensor_wraps, info->tensor_wraps_ref - 1)((void*)(((char*)((graph->tensor_wraps)->data)) + (size_t
)(graph->tensor_wraps)->rsize * (size_t)(info->tensor_wraps_ref
 - 1)));
232		ccv_nnc_graph_tensor_wrap_array_t* const tensor_wrap_array = *tensor_wrap_array_ref;
233		if (tensor_wrap_array)
234		{
235			for (i = 0; i < tensor_wrap_array->size; i++)
236				if (tensor_wrap_array->tensor_wraps[i])
237					_ccv_nnc_graph_tensor_wrap_free(tensor_wrap_array->tensor_wraps[i]);
238			ccfreefree(tensor_wrap_array);
239			*tensor_wrap_array_ref = 0;
240			info->tensor_wraps_ref = 0;
241		}
242	}
243}
244 
245static void _ccv_nnc_graph_deregister_tensor_wraps(ccv_nnc_graph_t* graph, const int tensor_wraps_ref_d)
246{
247	ccv_nnc_graph_t* p = graph;
248	do {
249		int i;
250		// Remove from the array.
251		if (p->tensor_wraps_refs)
252			for (i = 0; i < p->tensor_wraps_refs->rnum; i++)
253			{
254				ccv_nnc_graph_tensor_wraps_ref_t* const tensor_wraps_ref = (ccv_nnc_graph_tensor_wraps_ref_t*)ccv_array_get(p->tensor_wraps_refs, i)((void*)(((char*)((p->tensor_wraps_refs)->data)) + (size_t
)(p->tensor_wraps_refs)->rsize * (size_t)(i)));
255				if (tensor_wraps_ref->d == tensor_wraps_ref_d && tensor_wraps_ref->graph == graph)
256				{
257					--p->tensor_wraps_refs->rnum;
258					if (i < p->tensor_wraps_refs->rnum)
259						memcpy(tensor_wraps_ref, tensor_wraps_ref + 1, sizeof(ccv_nnc_graph_exec_t) * (p->tensor_wraps_refs->rnum - i));
260					break;
261				}
262			}
263		p = p->p;
264	} while (p);
265}
266 
267void ccv_nnc_graph_exec_set_io_flags(ccv_nnc_graph_t* const graph, const ccv_nnc_graph_exec_t exec, const int* const input_flags, const int input_flag_size, const int* const output_flags, const int output_flag_size)
268{
269	assert(exec.d < graph->exec_info->rnum)((void) sizeof ((exec.d < graph->exec_info->rnum) ? 1
 : 0), __extension__ ({ if (exec.d < graph->exec_info->
rnum) ; else __assert_fail ("exec.d < graph->exec_info->rnum"
, "ccv_nnc_graph.c", 269, __extension__ __PRETTY_FUNCTION__);
 }));
270	assert(exec.graph == graph)((void) sizeof ((exec.graph == graph) ? 1 : 0), __extension__
 ({ if (exec.graph == graph) ; else __assert_fail ("exec.graph == graph"
, "ccv_nnc_graph.c", 270, __extension__ __PRETTY_FUNCTION__);
 }));
271	ccv_nnc_graph_exec_info_t* const info = (ccv_nnc_graph_exec_info_t*)ccv_array_get(graph->exec_info, exec.d)((void*)(((char*)((graph->exec_info)->data)) + (size_t)
(graph->exec_info)->rsize * (size_t)(exec.d)));
272	assert(input_flag_size <= info->input_size)((void) sizeof ((input_flag_size <= info->input_size) ?
 1 : 0), __extension__ ({ if (input_flag_size <= info->
input_size) ; else __assert_fail ("input_flag_size <= info->input_size"
, "ccv_nnc_graph.c", 272, __extension__ __PRETTY_FUNCTION__);
 }));
273	assert(output_flag_size <= info->output_size)((void) sizeof ((output_flag_size <= info->output_size)
 ? 1 : 0), __extension__ ({ if (output_flag_size <= info->
output_size) ; else __assert_fail ("output_flag_size <= info->output_size"
, "ccv_nnc_graph.c", 273, __extension__ __PRETTY_FUNCTION__);
 }));
274	if (info->input_size + info->output_size == 0)
275		return;
276	if (!info->input_flags)
277	{
278		info->input_flags = (int*)cccalloccalloc(info->input_size + info->output_size, sizeof(int));
279		info->output_flags = info->input_flags + info->input_size;
280	}
281	if (input_flag_size > 0)
282		memcpy(info->input_flags, input_flags, sizeof(int) * input_flag_size);
283	if (output_flag_size > 0)
284		memcpy(info->output_flags, output_flags, sizeof(int) * output_flag_size);
285}
286 
287void ccv_nnc_graph_exec_pair_with(ccv_nnc_graph_t* const graph, const ccv_nnc_graph_exec_t exec, const ccv_nnc_graph_exec_t pair_exec)
288{
289	assert(exec.graph == graph)((void) sizeof ((exec.graph == graph) ? 1 : 0), __extension__
 ({ if (exec.graph == graph) ; else __assert_fail ("exec.graph == graph"
, "ccv_nnc_graph.c", 289, __extension__ __PRETTY_FUNCTION__);
 }));
290	assert(exec.d >= 0)((void) sizeof ((exec.d >= 0) ? 1 : 0), __extension__ ({ if
 (exec.d >= 0) ; else __assert_fail ("exec.d >= 0", "ccv_nnc_graph.c"
, 290, __extension__ __PRETTY_FUNCTION__); }));
291	assert(exec.d < graph->exec_info->rnum)((void) sizeof ((exec.d < graph->exec_info->rnum) ? 1
 : 0), __extension__ ({ if (exec.d < graph->exec_info->
rnum) ; else __assert_fail ("exec.d < graph->exec_info->rnum"
, "ccv_nnc_graph.c", 291, __extension__ __PRETTY_FUNCTION__);
 }));
292	assert(pair_exec.graph == graph || pair_exec.graph == graph->pair)((void) sizeof ((pair_exec.graph == graph || pair_exec.graph ==
 graph->pair) ? 1 : 0), __extension__ ({ if (pair_exec.graph
 == graph || pair_exec.graph == graph->pair) ; else __assert_fail
 ("pair_exec.graph == graph || pair_exec.graph == graph->pair"
, "ccv_nnc_graph.c", 292, __extension__ __PRETTY_FUNCTION__);
 }));
293	assert(pair_exec.d >= 0)((void) sizeof ((pair_exec.d >= 0) ? 1 : 0), __extension__
 ({ if (pair_exec.d >= 0) ; else __assert_fail ("pair_exec.d >= 0"
, "ccv_nnc_graph.c", 293, __extension__ __PRETTY_FUNCTION__);
 }));
294	if (pair_exec.graph == graph)
295		{ assert(pair_exec.d < graph->exec_info->rnum)((void) sizeof ((pair_exec.d < graph->exec_info->rnum
) ? 1 : 0), __extension__ ({ if (pair_exec.d < graph->exec_info
->rnum) ; else __assert_fail ("pair_exec.d < graph->exec_info->rnum"
, "ccv_nnc_graph.c", 295, __extension__ __PRETTY_FUNCTION__);
 })); }
296	else
297		{ assert(pair_exec.d < graph->pair->exec_info->rnum)((void) sizeof ((pair_exec.d < graph->pair->exec_info
->rnum) ? 1 : 0), __extension__ ({ if (pair_exec.d < graph
->pair->exec_info->rnum) ; else __assert_fail ("pair_exec.d < graph->pair->exec_info->rnum"
, "ccv_nnc_graph.c", 297, __extension__ __PRETTY_FUNCTION__);
 })); }
298	ccv_nnc_graph_exec_info_t* const exec_info = (ccv_nnc_graph_exec_info_t*)ccv_array_get(graph->exec_info, exec.d)((void*)(((char*)((graph->exec_info)->data)) + (size_t)
(graph->exec_info)->rsize * (size_t)(exec.d)));
299	exec_info->pair_ref = pair_exec.d + 1;
300}
301 
302static ccv_nnc_tensor_t* _ccv_nnc_any_tensor_from_tensor_multiview(ccv_nnc_tensor_multiview_t* const mv)
303{
304	ccv_nnc_tensor_t* tensor = (ccv_nnc_tensor_t*)mv;
305	while (CCV_IS_TENSOR_MULTIVIEW(tensor)((*(int*)(tensor)) & CCV_TENSOR_MULTIVIEW))
306	{
307		ccv_nnc_tensor_multiview_t* mv = (ccv_nnc_tensor_multiview_t*)tensor;
308		const int count = 0;
309		const int off = mv->kind;
310		const int mod = mv->repeat;
311		// If reached the root.
312		tensor = CCV_NNC_MULTIVIEW_DATA(mv)((mv)->_heap_data ? (mv)->_heap_data : (mv)->_inline_data
)[count >= off ? ((count - off) % mod) + off : count]; // Unwrap.
313	}
314	return tensor;
315}
316 
317void ccv_nnc_graph_exec_set_io(ccv_nnc_graph_t* const graph, const ccv_nnc_graph_exec_t exec, ccv_nnc_tensor_t* const* const inputs, const int input_size, ccv_nnc_tensor_t* const* const outputs, const int output_size)
318{
319	assert(exec.d < graph->exec_info->rnum)((void) sizeof ((exec.d < graph->exec_info->rnum) ? 1
 : 0), __extension__ ({ if (exec.d < graph->exec_info->
rnum) ; else __assert_fail ("exec.d < graph->exec_info->rnum"
, "ccv_nnc_graph.c", 319, __extension__ __PRETTY_FUNCTION__);
 }));
1
Assuming field 'd' is < field 'rnum'→
2
←
Taking true branch→
320	assert(exec.graph == graph)((void) sizeof ((exec.graph == graph) ? 1 : 0), __extension__
 ({ if (exec.graph == graph) ; else __assert_fail ("exec.graph == graph"
, "ccv_nnc_graph.c", 320, __extension__ __PRETTY_FUNCTION__);
 }));
3
←
Assuming 'graph' is equal to field 'graph'→
4
←
Taking true branch→
321	ccv_nnc_graph_exec_info_t* const info = (ccv_nnc_graph_exec_info_t*)ccv_array_get(graph->exec_info, exec.d)((void*)(((char*)((graph->exec_info)->data)) + (size_t)
(graph->exec_info)->rsize * (size_t)(exec.d)));
322	// De-register from the graph if it contains multiview tensors.
323	if (info->tensor_wraps_ref)
5
←
Assuming field 'tensor_wraps_ref' is 0→
6
←
Taking false branch→
324		_ccv_nnc_graph_deregister_tensor_wraps(graph, info->tensor_wraps_ref - 1);
325	// In case it is already executed, rewind.
326	_ccv_nnc_graph_exec_rewind(info, graph);
327	if (input_size == 0 && output_size == 0)
7
←
Assuming 'input_size' is not equal to 0→
328	{
329		if (info->input_size > 0 || info->output_size > 0)
330			ccfreefree(info->inputs);
331		info->inputs = 0;
332		info->outputs = 0;
333		info->input_size = 0;
334		info->output_size = 0;
335		_ccv_nnc_graph_redo_tensor_wraps(info, graph);
336		if (info->tensor_wraps_ref)
337			ccv_nnc_graph_register_tensor_wraps(graph, info->tensor_wraps_ref - 1);
338		return;
339	}
340	if (info->inputs)
8
←
Assuming field 'inputs' is null→
9
←
Taking false branch→
341		info->inputs = (ccv_nnc_tensor_t**)ccreallocrealloc(info->inputs, sizeof(ccv_nnc_tensor_t*) * (input_size + output_size));
342	else
343		info->inputs = (ccv_nnc_tensor_t**)ccmallocmalloc(sizeof(ccv_nnc_tensor_t*) * (input_size + output_size));
10
←
Storing uninitialized value→
344	info->outputs = info->inputs + input_size;
345	if (inputs)
11
←
Assuming 'inputs' is null→
12
←
Taking false branch→
346		memcpy(info->inputs, inputs, sizeof(ccv_nnc_tensor_t*) * input_size);
347	if (outputs)
13
←
Assuming 'outputs' is null→
14
←
Taking false branch→
348		memcpy(info->outputs, outputs, sizeof(ccv_nnc_tensor_t*) * output_size);
349	int i;
350	int tensor_memory = 0, tensor_formats = 0, tensor_datatypes = 0;
351	for (i = 0; i < input_size + output_size; i++)
15
←
The value 0 is assigned to 'i'→
16
←
Assuming the condition is true→
17
←
Loop condition is true.  Entering loop body→
352		if (info->inputs[i])
18
←
Branch condition evaluates to a garbage value
353		{
354			ccv_nnc_tensor_t* const tensor = CCV_IS_TENSOR_MULTIVIEW(info->inputs[i])((*(int*)(info->inputs[i])) & CCV_TENSOR_MULTIVIEW) ? _ccv_nnc_any_tensor_from_tensor_multiview((ccv_nnc_tensor_multiview_t*)info->inputs[i]) : info->inputs[i];
355			tensor_memory |= CCV_TENSOR_GET_MEMORY(tensor->info.type)((tensor->info.type) & 0x3), tensor_formats |= tensor->info.format, tensor_datatypes |= CCV_GET_DATA_TYPE(tensor->info.datatype)((tensor->info.datatype) & 0xFF000);
356		}
357	info->cmd.backend = ccv_nnc_cmd_find_backend(info->cmd, tensor_memory, tensor_formats, tensor_datatypes);
358	info->input_size = input_size;
359	info->output_size = output_size;
360	_ccv_nnc_graph_redo_tensor_wraps(info, graph);
361	// Register again if the tensor wraps exist.
362	if (info->tensor_wraps_ref)
363		ccv_nnc_graph_register_tensor_wraps(graph, info->tensor_wraps_ref - 1);
364	// Free flags.
365	if (info->input_flags)
366	{
367		ccfreefree(info->input_flags);
368		info->input_flags = info->output_flags = 0;
369	}
370}
371 
372void ccv_nnc_graph_exec_add_as_affected(ccv_nnc_graph_t* const graph, const ccv_nnc_graph_exec_t exec, ccv_nnc_tensor_t* const update)
373{
374	assert(CCV_IS_TENSOR_MULTIVIEW(update))((void) sizeof ((((*(int*)(update)) & CCV_TENSOR_MULTIVIEW
)) ? 1 : 0), __extension__ ({ if (((*(int*)(update)) & CCV_TENSOR_MULTIVIEW
)) ; else __assert_fail ("CCV_IS_TENSOR_MULTIVIEW(update)", "ccv_nnc_graph.c"
, 374, __extension__ __PRETTY_FUNCTION__); }));
375	assert(exec.d < graph->exec_info->rnum)((void) sizeof ((exec.d < graph->exec_info->rnum) ? 1
 : 0), __extension__ ({ if (exec.d < graph->exec_info->
rnum) ; else __assert_fail ("exec.d < graph->exec_info->rnum"
, "ccv_nnc_graph.c", 375, __extension__ __PRETTY_FUNCTION__);
 }));
376	assert(exec.graph == graph)((void) sizeof ((exec.graph == graph) ? 1 : 0), __extension__
 ({ if (exec.graph == graph) ; else __assert_fail ("exec.graph == graph"
, "ccv_nnc_graph.c", 376, __extension__ __PRETTY_FUNCTION__);
 }));
377	ccv_nnc_graph_exec_info_t* const info = (ccv_nnc_graph_exec_info_t*)ccv_array_get(graph->exec_info, exec.d)((void*)(((char*)((graph->exec_info)->data)) + (size_t)
(graph->exec_info)->rsize * (size_t)(exec.d)));
378	const int register_tensor_wraps = !info->tensor_wraps_ref;
379	const int update_index = info->update_size;
380	++info->update_size;
381	if (info->updates)
382		info->updates = (ccv_nnc_tensor_t**)ccreallocrealloc(info->updates, sizeof(ccv_nnc_tensor_t*) * info->update_size);
383	else
384		info->updates = (ccv_nnc_tensor_t**)ccmallocmalloc(sizeof(ccv_nnc_tensor_t*) * info->update_size);
385	info->updates[update_index] = update;
386	_ccv_nnc_graph_redo_tensor_wraps(info, graph);
387	if (register_tensor_wraps)
388		ccv_nnc_graph_register_tensor_wraps(graph, info->tensor_wraps_ref - 1);
389}
390 
391ccv_nnc_graph_exec_t ccv_nnc_graph_exec_new(ccv_nnc_graph_t* const graph, const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint, ccv_nnc_tensor_t* const* const inputs, const int input_size, ccv_nnc_tensor_t* const* const outputs, const int output_size)
392{
393	int d = graph->exec_info->rnum;
394	ccv_nnc_graph_exec_info_t info = {
395		.cmd = cmd,
396		.hint = hint,
397		.input_size = input_size,
398		.output_size = output_size,
399	};
400	assert(inputs || input_size == 0)((void) sizeof ((inputs || input_size == 0) ? 1 : 0), __extension__
 ({ if (inputs || input_size == 0) ; else __assert_fail ("inputs || input_size == 0"
, "ccv_nnc_graph.c", 400, __extension__ __PRETTY_FUNCTION__);
 }));
401	assert(outputs || output_size == 0)((void) sizeof ((outputs || output_size == 0) ? 1 : 0), __extension__
 ({ if (outputs || output_size == 0) ; else __assert_fail ("outputs || output_size == 0"
, "ccv_nnc_graph.c", 401, __extension__ __PRETTY_FUNCTION__);
 }));
402	if (input_size > 0 || output_size > 0)
403	{
404		info.inputs = (ccv_nnc_tensor_t**)ccmallocmalloc(sizeof(ccv_nnc_tensor_t*) * (input_size + output_size));
405		info.outputs = info.inputs + input_size;
406		if (inputs)
407			memcpy(info.inputs, inputs, sizeof(ccv_nnc_tensor_t*) * input_size);
408		if (outputs)
409			memcpy(info.outputs, outputs, sizeof(ccv_nnc_tensor_t*) * output_size);
410		info.input_size = input_size;
411		info.output_size = output_size;
412		int i;
413		int tensor_memory = 0, tensor_formats = 0, tensor_datatypes = 0;
414		for (i = 0; i < input_size + output_size; i++)
415			if (info.inputs[i])
416			{
417				ccv_nnc_tensor_t* const tensor = CCV_IS_TENSOR_MULTIVIEW(info.inputs[i])((*(int*)(info.inputs[i])) & CCV_TENSOR_MULTIVIEW) ? _ccv_nnc_any_tensor_from_tensor_multiview((ccv_nnc_tensor_multiview_t*)info.inputs[i]) : info.inputs[i];
418				tensor_memory |= CCV_TENSOR_GET_MEMORY(tensor->info.type)((tensor->info.type) & 0x3), tensor_formats |= tensor->info.format, tensor_datatypes |= CCV_GET_DATA_TYPE(tensor->info.datatype)((tensor->info.datatype) & 0xFF000);
419			}
420		info.cmd.backend = ccv_nnc_cmd_find_backend(info.cmd, tensor_memory, tensor_formats, tensor_datatypes);
421	}
422	_ccv_nnc_graph_redo_tensor_wraps(&info, graph);
423	// Add itself to the graph's wraps array, this will help the run time when we run the graph and do unwrapping.
424	if (info.tensor_wraps_ref)
425		ccv_nnc_graph_register_tensor_wraps(graph, info.tensor_wraps_ref - 1);
426	ccv_array_push(graph->exec_info, &info);
427	return (ccv_nnc_graph_exec_t){
428		.d = d,
429		.graph = graph,
430	};
431}
432 
433void ccv_nnc_graph_add_carry_over(ccv_nnc_graph_t* const graph, const ccv_nnc_tensor_t* const from, const ccv_nnc_tensor_t* const to)
434{
435	ccv_nnc_graph_tensor_carry_over_t carry_over = {
436		.from = _ccv_nnc_graph_tensor_wrap_new((ccv_nnc_tensor_multiview_t*)from),
437		.to = _ccv_nnc_graph_tensor_wrap_new((ccv_nnc_tensor_multiview_t*)to)
438	};
439	if (!graph->carry_overs)
440		graph->carry_overs = ccv_array_new(sizeof(ccv_nnc_graph_tensor_carry_over_t), 0, 0);
441	ccv_array_push(graph->carry_overs, &carry_over);
442}
443 
444int ccv_nnc_graph_exec_concat(ccv_nnc_graph_t* const graph, const ccv_nnc_graph_exec_t source, const ccv_nnc_graph_exec_t destination)
445{
446	assert(graph == source.graph)((void) sizeof ((graph == source.graph) ? 1 : 0), __extension__
 ({ if (graph == source.graph) ; else __assert_fail ("graph == source.graph"
, "ccv_nnc_graph.c", 446, __extension__ __PRETTY_FUNCTION__);
 }));
447	assert(graph == destination.graph)((void) sizeof ((graph == destination.graph) ? 1 : 0), __extension__
 ({ if (graph == destination.graph) ; else __assert_fail ("graph == destination.graph"
, "ccv_nnc_graph.c", 447, __extension__ __PRETTY_FUNCTION__);
 }));
448	assert(source.d < graph->exec_info->rnum)((void) sizeof ((source.d < graph->exec_info->rnum) ?
 1 : 0), __extension__ ({ if (source.d < graph->exec_info
->rnum) ; else __assert_fail ("source.d < graph->exec_info->rnum"
, "ccv_nnc_graph.c", 448, __extension__ __PRETTY_FUNCTION__);
 }));
449	assert(destination.d < graph->exec_info->rnum)((void) sizeof ((destination.d < graph->exec_info->rnum
) ? 1 : 0), __extension__ ({ if (destination.d < graph->
exec_info->rnum) ; else __assert_fail ("destination.d < graph->exec_info->rnum"
, "ccv_nnc_graph.c", 449, __extension__ __PRETTY_FUNCTION__);
 }));
450	ccv_nnc_graph_exec_info_t* src_info = (ccv_nnc_graph_exec_info_t*)ccv_array_get(graph->exec_info, source.d)((void*)(((char*)((graph->exec_info)->data)) + (size_t)
(graph->exec_info)->rsize * (size_t)(source.d)));
451	if (src_info->outgoings == 0)
452		src_info->outgoings = ccv_array_new(sizeof(int32_t), 1, 0);
453	else {
454		int i;
455		// Check if this is already connected, if so, skip.
456		for (i = 0; i < src_info->outgoings->rnum; i++)
457			if (*(int*)ccv_array_get(src_info->outgoings, i)((void*)(((char*)((src_info->outgoings)->data)) + (size_t
)(src_info->outgoings)->rsize * (size_t)(i))) == destination.d)
458				return -1;
459	}
460	ccv_array_push(src_info->outgoings, &destination.d);
461	graph->topsorted = 0;
462	return 0;
463}
464 
465int ccv_nnc_graph_exec_disjoin(ccv_nnc_graph_t* const graph, const ccv_nnc_graph_exec_t source, const ccv_nnc_graph_exec_t destination)
466{
467	assert(graph == source.graph)((void) sizeof ((graph == source.graph) ? 1 : 0), __extension__
 ({ if (graph == source.graph) ; else __assert_fail ("graph == source.graph"
, "ccv_nnc_graph.c", 467, __extension__ __PRETTY_FUNCTION__);
 }));
468	assert(graph == destination.graph)((void) sizeof ((graph == destination.graph) ? 1 : 0), __extension__
 ({ if (graph == destination.graph) ; else __assert_fail ("graph == destination.graph"
, "ccv_nnc_graph.c", 468, __extension__ __PRETTY_FUNCTION__);
 }));
469	assert(source.d < graph->exec_info->rnum)((void) sizeof ((source.d < graph->exec_info->rnum) ?
 1 : 0), __extension__ ({ if (source.d < graph->exec_info
->rnum) ; else __assert_fail ("source.d < graph->exec_info->rnum"
, "ccv_nnc_graph.c", 469, __extension__ __PRETTY_FUNCTION__);
 }));
470	assert(destination.d < graph->exec_info->rnum)((void) sizeof ((destination.d < graph->exec_info->rnum
) ? 1 : 0), __extension__ ({ if (destination.d < graph->
exec_info->rnum) ; else __assert_fail ("destination.d < graph->exec_info->rnum"
, "ccv_nnc_graph.c", 470, __extension__ __PRETTY_FUNCTION__);
 }));
471	ccv_nnc_graph_exec_info_t* src_info = (ccv_nnc_graph_exec_info_t*)ccv_array_get(graph->exec_info, source.d)((void*)(((char*)((graph->exec_info)->data)) + (size_t)
(graph->exec_info)->rsize * (size_t)(source.d)));
472	if (!src_info->outgoings)
473		return -1;
474	int i;
475	// Check if this is already connected, if so, skip.
476	for (i = 0; i < src_info->outgoings->rnum; i++)
477		if (*(int*)ccv_array_get(src_info->outgoings, i)((void*)(((char*)((src_info->outgoings)->data)) + (size_t
)(src_info->outgoings)->rsize * (size_t)(i))) == destination.d)
478		{
479			if (i < src_info->outgoings->rnum - 1)
480				*(int*)ccv_array_get(src_info->outgoings, i)((void*)(((char*)((src_info->outgoings)->data)) + (size_t
)(src_info->outgoings)->rsize * (size_t)(i))) = *(int*)ccv_array_get(src_info->outgoings, src_info->outgoings->rnum - 1)((void*)(((char*)((src_info->outgoings)->data)) + (size_t
)(src_info->outgoings)->rsize * (size_t)(src_info->outgoings
->rnum - 1)));
481			--src_info->outgoings->rnum;
482			graph->topsorted = 0;
483			return 0;
484		}
485	return -1;
486}
487 
488int ccv_nnc_graph_exec_count(const ccv_nnc_graph_t* const graph)
489{
490	return graph->exec_info ? graph->exec_info->rnum : 0;
491}
492 
493void* ccv_nnc_graph_buffer(ccv_nnc_graph_t* const graph, int size)
494{
495	if (graph->buffer_size >= size)
496		return graph->buffer;
497	graph->buffer_size = size;
498	graph->buffer = (graph->buffer) ? ccreallocrealloc(graph->buffer, size) : ccmallocmalloc(size);
499	return graph->buffer;
500}
501 
502void ccv_nnc_graph_topsort(ccv_nnc_graph_t* const graph, int* const exec_cvt, const int exec_cvt_size)
503{
504	if (exec_cvt_size == 0 && graph->exec_info->rnum == 0)
505	{
506		graph->topsorted = 1;
507		return;
508	}
509	assert(exec_cvt_size == graph->exec_info->rnum)((void) sizeof ((exec_cvt_size == graph->exec_info->rnum
) ? 1 : 0), __extension__ ({ if (exec_cvt_size == graph->exec_info
->rnum) ; else __assert_fail ("exec_cvt_size == graph->exec_info->rnum"
, "ccv_nnc_graph.c", 509, __extension__ __PRETTY_FUNCTION__);
 }));
510	assert(graph->sources && graph->sources->rnum)((void) sizeof ((graph->sources && graph->sources
->rnum) ? 1 : 0), __extension__ ({ if (graph->sources &&
 graph->sources->rnum) ; else __assert_fail ("graph->sources && graph->sources->rnum"
, "ccv_nnc_graph.c", 510, __extension__ __PRETTY_FUNCTION__);
 }));
511	assert(graph->destinations && graph->destinations->rnum)((void) sizeof ((graph->destinations && graph->
destinations->rnum) ? 1 : 0), __extension__ ({ if (graph->
destinations && graph->destinations->rnum) ; else
 __assert_fail ("graph->destinations && graph->destinations->rnum"
, "ccv_nnc_graph.c", 511, __extension__ __PRETTY_FUNCTION__);
 }));
512	int i, j;
513	for (i = 0; i < exec_cvt_size; i++)
514		exec_cvt[i] = -1;
515	ccv_array_t* exec_info = ccv_array_new(sizeof(ccv_nnc_graph_exec_info_t), graph->exec_info->rnum, 0);
516	// If there are breakpoints, it is more complicated, we first start to the breakpoints, and then continue from the breakpoints to the destinations.
517	if (graph->breakpoint_size)
518	{
519		ccv_nnc_graph_visit_t* visit = ccv_nnc_graph_visit_new(graph, (ccv_nnc_graph_exec_info_t*)ccv_array_get(graph->exec_info, 0), graph->exec_info->rnum, (ccv_nnc_graph_exec_t*)ccv_array_get(graph->sources, 0), graph->sources->rnum, graph->breakpoints, graph->breakpoint_size, 0)({ ccv_nnc_graph_visit_t* _visit_ = (ccv_nnc_graph_visit_t*)malloc
(sizeof(ccv_nnc_graph_visit_t) + sizeof(_visit_->node[0]) *
 ((graph->exec_info->rnum) - 1)); _visit_->size = 0;
 do { typedef struct { int8_t d; int8_t r; uint16_t c; int32_t
 edges; } ccv_nnc_incoming_t; int _i_, _j_; int _incoming_edges_
 = 0; for (_i_ = 0; _i_ < (graph->exec_info->rnum); _i_
++) _incoming_edges_ += (((ccv_nnc_graph_exec_info_t*)((void*
)(((char*)((graph->exec_info)->data)) + (size_t)(graph->
exec_info)->rsize * (size_t)(0))))[_i_].outgoings) ? ((ccv_nnc_graph_exec_info_t
*)((void*)(((char*)((graph->exec_info)->data)) + (size_t
)(graph->exec_info)->rsize * (size_t)(0))))[_i_].outgoings
->rnum : 0; const int _heap_mem_ = ((graph->exec_info->
rnum) + _incoming_edges_ > 1024); ccv_nnc_incoming_t* _incomings_
; if (_heap_mem_) _incomings_ = (ccv_nnc_incoming_t*)malloc(sizeof
(ccv_nnc_incoming_t) * (graph->exec_info->rnum) + sizeof
(int32_t) * ((graph->exec_info->rnum) * 2 + _incoming_edges_
)); else _incomings_ = (ccv_nnc_incoming_t*)__builtin_alloca (
sizeof(ccv_nnc_incoming_t) * (graph->exec_info->rnum) +
 sizeof(int32_t) * ((graph->exec_info->rnum) * 2 + _incoming_edges_
)); memset(_incomings_, 0, sizeof(ccv_nnc_incoming_t) * (graph
->exec_info->rnum)); int32_t* _exists_[2] = { (int32_t*
)(_incomings_ + (graph->exec_info->rnum)), (int32_t*)(_incomings_
 + (graph->exec_info->rnum)) + (graph->exec_info->
rnum), }; int32_t* const _edges_ = _exists_[1] + (graph->exec_info
->rnum); for (_i_ = 0; _i_ < (graph->sources->rnum
); _i_++) { ((void) sizeof ((((ccv_nnc_graph_exec_t*)((void*)
(((char*)((graph->sources)->data)) + (size_t)(graph->
sources)->rsize * (size_t)(0))))[_i_].graph == graph) ? 1 :
 0), __extension__ ({ if (((ccv_nnc_graph_exec_t*)((void*)(((
char*)((graph->sources)->data)) + (size_t)(graph->sources
)->rsize * (size_t)(0))))[_i_].graph == graph) ; else __assert_fail
 ("((ccv_nnc_graph_exec_t*)((void*)(((char*)((graph->sources)->data)) + (size_t)(graph->sources)->rsize * (size_t)(0))))[_i_].graph == graph"
, "ccv_nnc_graph.c", 519, __extension__ __PRETTY_FUNCTION__);
 })); _incomings_[((ccv_nnc_graph_exec_t*)((void*)(((char*)((
graph->sources)->data)) + (size_t)(graph->sources)->
rsize * (size_t)(0))))[_i_].d].r = 1; _exists_[0][_i_] = ((ccv_nnc_graph_exec_t
*)((void*)(((char*)((graph->sources)->data)) + (size_t)
(graph->sources)->rsize * (size_t)(0))))[_i_].d; } int _exist_size_
[2] = { (graph->sources->rnum), 0, }; int _p_ = 0, _q_ =
 1; while (_exist_size_[_p_] > 0) { _exist_size_[_q_] = 0;
 for (_i_ = 0; _i_ < _exist_size_[_p_]; _i_++) { const int32_t
 _idx_ = _exists_[_p_][_i_]; if (_incomings_[_idx_].r != 1) continue
; _incomings_[_idx_].r = 2; if (((ccv_nnc_graph_exec_info_t*)
((void*)(((char*)((graph->exec_info)->data)) + (size_t)
(graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings
) for (_j_ = 0; _j_ < ((ccv_nnc_graph_exec_info_t*)((void*
)(((char*)((graph->exec_info)->data)) + (size_t)(graph->
exec_info)->rsize * (size_t)(0))))[_idx_].outgoings->rnum
; _j_++) { const int d = *(int*)((void*)(((char*)((((ccv_nnc_graph_exec_info_t
*)((void*)(((char*)((graph->exec_info)->data)) + (size_t
)(graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings
)->data)) + (size_t)(((ccv_nnc_graph_exec_info_t*)((void*)
(((char*)((graph->exec_info)->data)) + (size_t)(graph->
exec_info)->rsize * (size_t)(0))))[_idx_].outgoings)->rsize
 * (size_t)(_j_))); ++_incomings_[d].c; if (_incomings_[d].r !=
 0) continue; _incomings_[d].r = 1; ((void) sizeof ((_exist_size_
[_q_] < (graph->exec_info->rnum)) ? 1 : 0), __extension__
 ({ if (_exist_size_[_q_] < (graph->exec_info->rnum)
) ; else __assert_fail ("_exist_size_[_q_] < (graph->exec_info->rnum)"
, "ccv_nnc_graph.c", 519, __extension__ __PRETTY_FUNCTION__);
 })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[_q_
]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for (
_i_ = 0; _i_ < (graph->sources->rnum); _i_++) { ((void
) sizeof ((((ccv_nnc_graph_exec_t*)((void*)(((char*)((graph->
sources)->data)) + (size_t)(graph->sources)->rsize *
 (size_t)(0))))[_i_].graph == graph) ? 1 : 0), __extension__ (
{ if (((ccv_nnc_graph_exec_t*)((void*)(((char*)((graph->sources
)->data)) + (size_t)(graph->sources)->rsize * (size_t
)(0))))[_i_].graph == graph) ; else __assert_fail ("((ccv_nnc_graph_exec_t*)((void*)(((char*)((graph->sources)->data)) + (size_t)(graph->sources)->rsize * (size_t)(0))))[_i_].graph == graph"
, "ccv_nnc_graph.c", 519, __extension__ __PRETTY_FUNCTION__);
 })); _incomings_[((ccv_nnc_graph_exec_t*)((void*)(((char*)((
graph->sources)->data)) + (size_t)(graph->sources)->
rsize * (size_t)(0))))[_i_].d].r = 3; _exists_[0][_i_] = ((ccv_nnc_graph_exec_t
*)((void*)(((char*)((graph->sources)->data)) + (size_t)
(graph->sources)->rsize * (size_t)(0))))[_i_].d; } _exist_size_
[0] = (graph->sources->rnum); _exist_size_[1] = 0; _p_ =
 0, _q_ = 1; int _bump_ = 1; while (_exist_size_[_p_] > 0)
 { _exist_size_[_q_] = 0; for (_i_ = 0; _i_ < _exist_size_
[_p_]; _i_++) { const int32_t _idx_ = _exists_[_p_][_i_]; if (
_incomings_[_idx_].r != 3) continue; _incomings_[_idx_].r = 4
; if (((ccv_nnc_graph_exec_info_t*)((void*)(((char*)((graph->
exec_info)->data)) + (size_t)(graph->exec_info)->rsize
 * (size_t)(0))))[_idx_].outgoings) for (_j_ = 0; _j_ < ((
ccv_nnc_graph_exec_info_t*)((void*)(((char*)((graph->exec_info
)->data)) + (size_t)(graph->exec_info)->rsize * (size_t
)(0))))[_idx_].outgoings->rnum; _j_++) { const int d = *(int
*)((void*)(((char*)((((ccv_nnc_graph_exec_info_t*)((void*)(((
char*)((graph->exec_info)->data)) + (size_t)(graph->
exec_info)->rsize * (size_t)(0))))[_idx_].outgoings)->data
)) + (size_t)(((ccv_nnc_graph_exec_info_t*)((void*)(((char*)(
(graph->exec_info)->data)) + (size_t)(graph->exec_info
)->rsize * (size_t)(0))))[_idx_].outgoings)->rsize * (size_t
)(_j_))); if (_incomings_[d].edges == 0) { _incomings_[d].edges
 = _bump_; _bump_ += _incomings_[d].c; _incomings_[d].c = 0; }
 _edges_[_incomings_[d].edges - 1 + _incomings_[d].c] = _idx_
; ++_incomings_[d].c; if (_incomings_[d].r != 2) continue; _incomings_
[d].r = 3; ((void) sizeof ((_exist_size_[_q_] < (graph->
exec_info->rnum)) ? 1 : 0), __extension__ ({ if (_exist_size_
[_q_] < (graph->exec_info->rnum)) ; else __assert_fail
 ("_exist_size_[_q_] < (graph->exec_info->rnum)", "ccv_nnc_graph.c"
, 519, __extension__ __PRETTY_FUNCTION__); })); _exists_[_q_]
[_exist_size_[_q_]] = d; ++_exist_size_[_q_]; } } ((_i_) = (_p_
), (_p_) = (_q_), (_q_) = (_i_)); } for (_i_ = 0; _i_ < (graph
->breakpoint_size); _i_++) { ((void) sizeof (((graph->breakpoints
)[_i_].graph == graph) ? 1 : 0), __extension__ ({ if ((graph->
breakpoints)[_i_].graph == graph) ; else __assert_fail ("(graph->breakpoints)[_i_].graph == graph"
, "ccv_nnc_graph.c", 519, __extension__ __PRETTY_FUNCTION__);
 })); _incomings_[(graph->breakpoints)[_i_].d].r = 5; _exists_
[0][_i_] = (graph->breakpoints)[_i_].d; } _exist_size_[0] =
 (graph->breakpoint_size); _exist_size_[1] = 0; _p_ = 0, _q_
 = 1; while (_exist_size_[_p_] > 0) { _exist_size_[_q_] = 0
; for (_i_ = 0; _i_ < _exist_size_[_p_]; _i_++) { const int32_t
 _idx_ = _exists_[_p_][_i_]; if (_incomings_[_idx_].r != 5) continue
; _incomings_[_idx_].r = 6; if (_incomings_[_idx_].edges >
 0) for (_j_ = 0; _j_ < _incomings_[_idx_].c; _j_++) { const
 int d = _edges_[_incomings_[_idx_].edges - 1 + _j_]; if (_incomings_
[d].r != 4) continue; _incomings_[d].r = 5; ((void) sizeof ((
_exist_size_[_q_] < (graph->exec_info->rnum)) ? 1 : 0
), __extension__ ({ if (_exist_size_[_q_] < (graph->exec_info
->rnum)) ; else __assert_fail ("_exist_size_[_q_] < (graph->exec_info->rnum)"
, "ccv_nnc_graph.c", 519, __extension__ __PRETTY_FUNCTION__);
 })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[_q_
]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for (
_i_ = 0; _i_ < (graph->breakpoint_size); _i_++) { ((void
) sizeof (((graph->breakpoints)[_i_].graph == graph) ? 1 :
 0), __extension__ ({ if ((graph->breakpoints)[_i_].graph ==
 graph) ; else __assert_fail ("(graph->breakpoints)[_i_].graph == graph"
, "ccv_nnc_graph.c", 519, __extension__ __PRETTY_FUNCTION__);
 })); _incomings_[(graph->breakpoints)[_i_].d].d = 1; } for
 (_i_ = 0; _i_ < (graph->sources->rnum); _i_++) { ((
void) sizeof ((((ccv_nnc_graph_exec_t*)((void*)(((char*)((graph
->sources)->data)) + (size_t)(graph->sources)->rsize
 * (size_t)(0))))[_i_].graph == graph) ? 1 : 0), __extension__
 ({ if (((ccv_nnc_graph_exec_t*)((void*)(((char*)((graph->
sources)->data)) + (size_t)(graph->sources)->rsize *
 (size_t)(0))))[_i_].graph == graph) ; else __assert_fail ("((ccv_nnc_graph_exec_t*)((void*)(((char*)((graph->sources)->data)) + (size_t)(graph->sources)->rsize * (size_t)(0))))[_i_].graph == graph"
, "ccv_nnc_graph.c", 519, __extension__ __PRETTY_FUNCTION__);
 })); _exists_[0][_i_] = ((ccv_nnc_graph_exec_t*)((void*)(((char
*)((graph->sources)->data)) + (size_t)(graph->sources
)->rsize * (size_t)(0))))[_i_].d; } _p_ = 0; _q_ = 1; _exist_size_
[0] = (graph->sources->rnum); _exist_size_[1] = 0; int _d_
 = 0; while (_exist_size_[_p_] > 0) { _exist_size_[_q_] = 0
; for (_i_ = 0; _i_ < _exist_size_[_p_];) { const int32_t _idx_
 = _exists_[_p_][_i_]; _visit_->node[_visit_->size].index
 = ((_idx_)); _visit_->node[_visit_->size].term = ((_incomings_
[_idx_].d)); ++_visit_->size;; if (_incomings_[_idx_].d) {
 ++_d_; _incomings_[_idx_].r = 7; } if (((ccv_nnc_graph_exec_info_t
*)((void*)(((char*)((graph->exec_info)->data)) + (size_t
)(graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings
) { if (((ccv_nnc_graph_exec_info_t*)((void*)(((char*)((graph
->exec_info)->data)) + (size_t)(graph->exec_info)->
rsize * (size_t)(0))))[_idx_].outgoings->rnum == 1) { const
 int d = *(int*)((void*)(((char*)((((ccv_nnc_graph_exec_info_t
*)((void*)(((char*)((graph->exec_info)->data)) + (size_t
)(graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings
)->data)) + (size_t)(((ccv_nnc_graph_exec_info_t*)((void*)
(((char*)((graph->exec_info)->data)) + (size_t)(graph->
exec_info)->rsize * (size_t)(0))))[_idx_].outgoings)->rsize
 * (size_t)(0))); --_incomings_[d].c; if (_incomings_[d].c ==
 0 && _incomings_[d].r == 6 && _d_ < (graph
->breakpoint_size)) { _exists_[_p_][_i_] = d; continue; } }
 else for (_j_ = 0; _j_ < ((ccv_nnc_graph_exec_info_t*)((void
*)(((char*)((graph->exec_info)->data)) + (size_t)(graph
->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings->
rnum; _j_++) { const int d = *(int*)((void*)(((char*)((((ccv_nnc_graph_exec_info_t
*)((void*)(((char*)((graph->exec_info)->data)) + (size_t
)(graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings
)->data)) + (size_t)(((ccv_nnc_graph_exec_info_t*)((void*)
(((char*)((graph->exec_info)->data)) + (size_t)(graph->
exec_info)->rsize * (size_t)(0))))[_idx_].outgoings)->rsize
 * (size_t)(_j_))); --_incomings_[d].c; if (_incomings_[d].c ==
 0 && _incomings_[d].r == 6 && _d_ < (graph
->breakpoint_size)) { ((void) sizeof ((_exist_size_[_q_] <
 (graph->exec_info->rnum)) ? 1 : 0), __extension__ ({ if
 (_exist_size_[_q_] < (graph->exec_info->rnum)) ; else
 __assert_fail ("_exist_size_[_q_] < (graph->exec_info->rnum)"
, "ccv_nnc_graph.c", 519, __extension__ __PRETTY_FUNCTION__);
 })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[_q_
]; } } } ++_i_; } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_
)); } for (_i_ = 0; _i_ < (graph->breakpoint_size); _i_
++) { ((void) sizeof (((graph->breakpoints)[_i_].graph == graph
) ? 1 : 0), __extension__ ({ if ((graph->breakpoints)[_i_]
.graph == graph) ; else __assert_fail ("(graph->breakpoints)[_i_].graph == graph"
, "ccv_nnc_graph.c", 519, __extension__ __PRETTY_FUNCTION__);
 })); if (_incomings_[(graph->breakpoints)[_i_].d].r == 7)
 continue; if (!(0)) { ((void) sizeof ((_incomings_[(graph->
breakpoints)[_i_].d].c == 0) ? 1 : 0), __extension__ ({ if (_incomings_
[(graph->breakpoints)[_i_].d].c == 0) ; else __assert_fail
 ("_incomings_[(graph->breakpoints)[_i_].d].c == 0", "ccv_nnc_graph.c"
, 519, __extension__ __PRETTY_FUNCTION__); })); } else if (_incomings_
[(graph->breakpoints)[_i_].d].c > 0) continue; _visit_->
node[_visit_->size].index = (((graph->breakpoints)[_i_]
.d)); _visit_->node[_visit_->size].term = ((_incomings_
[(graph->breakpoints)[_i_].d].d)); ++_visit_->size;; } if
 (_heap_mem_) free(_incomings_); } while (0);; ((void) sizeof
 ((_visit_->size <= (graph->exec_info->rnum)) ? 1
 : 0), __extension__ ({ if (_visit_->size <= (graph->
exec_info->rnum)) ; else __assert_fail ("_visit_->size <= (graph->exec_info->rnum)"
, "ccv_nnc_graph.c", 519, __extension__ __PRETTY_FUNCTION__);
 })); _visit_; });
520		for (i = 0; i < graph->breakpoint_size; i++)
521			exec_cvt[graph->breakpoints[i].d] = -2; // Mark this as breakpoints, so we will skip the first round.
522		ccv_nnc_graph_visit_for(visit, (ccv_nnc_graph_exec_info_t*)ccv_array_get(graph->exec_info, 0), node, idx){ int _i_; for (_i_ = 0; _i_ < (visit)->size; _i_++) { const
 int idx __attribute__((unused)) = (visit)->node[_i_].index
; const int _node_unused_ __attribute__((unused)) = (visit)->
node[_i_].term; typeof (((ccv_nnc_graph_exec_info_t*)((void*)
(((char*)((graph->exec_info)->data)) + (size_t)(graph->
exec_info)->rsize * (size_t)(0))))) const node __attribute__
((unused)) = ((ccv_nnc_graph_exec_info_t*)((void*)(((char*)((
graph->exec_info)->data)) + (size_t)(graph->exec_info
)->rsize * (size_t)(0)))) + idx; {
523			assert(!node->pair_ref)((void) sizeof ((!node->pair_ref) ? 1 : 0), __extension__ (
{ if (!node->pair_ref) ; else __assert_fail ("!node->pair_ref"
, "ccv_nnc_graph.c", 523, __extension__ __PRETTY_FUNCTION__);
 })); // If node has a pair ref, we cannot fix it up.
524			if (exec_cvt[idx] == -2) // Skip breakpoint.
525				continue;
526			// Loop over node and push to the array.
527			ccv_array_push(exec_info, node);
528			// Go to its sub-graph to fix exec_idx
529			for (i = 0; i < node->graph_ref_size; i++)
530			{
531				const int graph_ref = CCV_NNC_GRAPH_REF(node)((node)->_heap_graph_ref ? (node)->_heap_graph_ref : (node
)->_inline_graph_ref)[i] - 1;
532				if (graph_ref >= 0)
533				{
534					ccv_nnc_graph_t* const sub_graph = *(ccv_nnc_graph_t**)ccv_array_get(graph->sub_graphs, graph_ref)((void*)(((char*)((graph->sub_graphs)->data)) + (size_t
)(graph->sub_graphs)->rsize * (size_t)(graph_ref)));
535					sub_graph->exec_idx = exec_info->rnum;
536				}
537			}
538			exec_cvt[idx] = exec_info->rnum - 1;
539		} ccv_nnc_graph_visit_endfor} }
540		ccv_nnc_graph_visit_free(visit);
541		graph->breakpoint_offset = exec_info->rnum;
542		visit = ccv_nnc_graph_visit_new(graph, (ccv_nnc_graph_exec_info_t*)ccv_array_get(graph->exec_info, 0), graph->exec_info->rnum, graph->breakpoints, graph->breakpoint_size, (ccv_nnc_graph_exec_t*)ccv_array_get(graph->destinations, 0), graph->destinations->rnum, 0)({ ccv_nnc_graph_visit_t* _visit_ = (ccv_nnc_graph_visit_t*)malloc
(sizeof(ccv_nnc_graph_visit_t) + sizeof(_visit_->node[0]) *
 ((graph->exec_info->rnum) - 1)); _visit_->size = 0;
 do { typedef struct { int8_t d; int8_t r; uint16_t c; int32_t
 edges; } ccv_nnc_incoming_t; int _i_, _j_; int _incoming_edges_
 = 0; for (_i_ = 0; _i_ < (graph->exec_info->rnum); _i_
++) _incoming_edges_ += (((ccv_nnc_graph_exec_info_t*)((void*
)(((char*)((graph->exec_info)->data)) + (size_t)(graph->
exec_info)->rsize * (size_t)(0))))[_i_].outgoings) ? ((ccv_nnc_graph_exec_info_t
*)((void*)(((char*)((graph->exec_info)->data)) + (size_t
)(graph->exec_info)->rsize * (size_t)(0))))[_i_].outgoings
->rnum : 0; const int _heap_mem_ = ((graph->exec_info->
rnum) + _incoming_edges_ > 1024); ccv_nnc_incoming_t* _incomings_
; if (_heap_mem_) _incomings_ = (ccv_nnc_incoming_t*)malloc(sizeof
(ccv_nnc_incoming_t) * (graph->exec_info->rnum) + sizeof
(int32_t) * ((graph->exec_info->rnum) * 2 + _incoming_edges_
)); else _incomings_ = (ccv_nnc_incoming_t*)__builtin_alloca (
sizeof(ccv_nnc_incoming_t) * (graph->exec_info->rnum) +
 sizeof(int32_t) * ((graph->exec_info->rnum) * 2 + _incoming_edges_
)); memset(_incomings_, 0, sizeof(ccv_nnc_incoming_t) * (graph
->exec_info->rnum)); int32_t* _exists_[2] = { (int32_t*
)(_incomings_ + (graph->exec_info->rnum)), (int32_t*)(_incomings_
 + (graph->exec_info->rnum)) + (graph->exec_info->
rnum), }; int32_t* const _edges_ = _exists_[1] + (graph->exec_info
->rnum); for (_i_ = 0; _i_ < (graph->breakpoint_size
); _i_++) { ((void) sizeof (((graph->breakpoints)[_i_].graph
 == graph) ? 1 : 0), __extension__ ({ if ((graph->breakpoints
)[_i_].graph == graph) ; else __assert_fail ("(graph->breakpoints)[_i_].graph == graph"
, "ccv_nnc_graph.c", 542, __extension__ __PRETTY_FUNCTION__);
 })); _incomings_[(graph->breakpoints)[_i_].d].r = 1; _exists_
[0][_i_] = (graph->breakpoints)[_i_].d; } int _exist_size_
[2] = { (graph->breakpoint_size), 0, }; int _p_ = 0, _q_ =
 1; while (_exist_size_[_p_] > 0) { _exist_size_[_q_] = 0;
 for (_i_ = 0; _i_ < _exist_size_[_p_]; _i_++) { const int32_t
 _idx_ = _exists_[_p_][_i_]; if (_incomings_[_idx_].r != 1) continue
; _incomings_[_idx_].r = 2; if (((ccv_nnc_graph_exec_info_t*)
((void*)(((char*)((graph->exec_info)->data)) + (size_t)
(graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings
) for (_j_ = 0; _j_ < ((ccv_nnc_graph_exec_info_t*)((void*
)(((char*)((graph->exec_info)->data)) + (size_t)(graph->
exec_info)->rsize * (size_t)(0))))[_idx_].outgoings->rnum
; _j_++) { const int d = *(int*)((void*)(((char*)((((ccv_nnc_graph_exec_info_t
*)((void*)(((char*)((graph->exec_info)->data)) + (size_t
)(graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings
)->data)) + (size_t)(((ccv_nnc_graph_exec_info_t*)((void*)
(((char*)((graph->exec_info)->data)) + (size_t)(graph->
exec_info)->rsize * (size_t)(0))))[_idx_].outgoings)->rsize
 * (size_t)(_j_))); ++_incomings_[d].c; if (_incomings_[d].r !=
 0) continue; _incomings_[d].r = 1; ((void) sizeof ((_exist_size_
[_q_] < (graph->exec_info->rnum)) ? 1 : 0), __extension__
 ({ if (_exist_size_[_q_] < (graph->exec_info->rnum)
) ; else __assert_fail ("_exist_size_[_q_] < (graph->exec_info->rnum)"
, "ccv_nnc_graph.c", 542, __extension__ __PRETTY_FUNCTION__);
 })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[_q_
]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for (
_i_ = 0; _i_ < (graph->breakpoint_size); _i_++) { ((void
) sizeof (((graph->breakpoints)[_i_].graph == graph) ? 1 :
 0), __extension__ ({ if ((graph->breakpoints)[_i_].graph ==
 graph) ; else __assert_fail ("(graph->breakpoints)[_i_].graph == graph"
, "ccv_nnc_graph.c", 542, __extension__ __PRETTY_FUNCTION__);
 })); _incomings_[(graph->breakpoints)[_i_].d].r = 3; _exists_
[0][_i_] = (graph->breakpoints)[_i_].d; } _exist_size_[0] =
 (graph->breakpoint_size); _exist_size_[1] = 0; _p_ = 0, _q_
 = 1; int _bump_ = 1; while (_exist_size_[_p_] > 0) { _exist_size_
[_q_] = 0; for (_i_ = 0; _i_ < _exist_size_[_p_]; _i_++) {
 const int32_t _idx_ = _exists_[_p_][_i_]; if (_incomings_[_idx_
].r != 3) continue; _incomings_[_idx_].r = 4; if (((ccv_nnc_graph_exec_info_t
*)((void*)(((char*)((graph->exec_info)->data)) + (size_t
)(graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings
) for (_j_ = 0; _j_ < ((ccv_nnc_graph_exec_info_t*)((void*
)(((char*)((graph->exec_info)->data)) + (size_t)(graph->
exec_info)->rsize * (size_t)(0))))[_idx_].outgoings->rnum
; _j_++) { const int d = *(int*)((void*)(((char*)((((ccv_nnc_graph_exec_info_t
*)((void*)(((char*)((graph->exec_info)->data)) + (size_t
)(graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings
)->data)) + (size_t)(((ccv_nnc_graph_exec_info_t*)((void*)
(((char*)((graph->exec_info)->data)) + (size_t)(graph->
exec_info)->rsize * (size_t)(0))))[_idx_].outgoings)->rsize
 * (size_t)(_j_))); if (_incomings_[d].edges == 0) { _incomings_
[d].edges = _bump_; _bump_ += _incomings_[d].c; _incomings_[d
].c = 0; } _edges_[_incomings_[d].edges - 1 + _incomings_[d].
c] = _idx_; ++_incomings_[d].c; if (_incomings_[d].r != 2) continue
; _incomings_[d].r = 3; ((void) sizeof ((_exist_size_[_q_] <
 (graph->exec_info->rnum)) ? 1 : 0), __extension__ ({ if
 (_exist_size_[_q_] < (graph->exec_info->rnum)) ; else
 __assert_fail ("_exist_size_[_q_] < (graph->exec_info->rnum)"
, "ccv_nnc_graph.c", 542, __extension__ __PRETTY_FUNCTION__);
 })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[_q_
]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for (
_i_ = 0; _i_ < (graph->destinations->rnum); _i_++) {
 ((void) sizeof ((((ccv_nnc_graph_exec_t*)((void*)(((char*)((
graph->destinations)->data)) + (size_t)(graph->destinations
)->rsize * (size_t)(0))))[_i_].graph == graph) ? 1 : 0), __extension__
 ({ if (((ccv_nnc_graph_exec_t*)((void*)(((char*)((graph->
destinations)->data)) + (size_t)(graph->destinations)->
rsize * (size_t)(0))))[_i_].graph == graph) ; else __assert_fail
 ("((ccv_nnc_graph_exec_t*)((void*)(((char*)((graph->destinations)->data)) + (size_t)(graph->destinations)->rsize * (size_t)(0))))[_i_].graph == graph"
, "ccv_nnc_graph.c", 542, __extension__ __PRETTY_FUNCTION__);
 })); _incomings_[((ccv_nnc_graph_exec_t*)((void*)(((char*)((
graph->destinations)->data)) + (size_t)(graph->destinations
)->rsize * (size_t)(0))))[_i_].d].r = 5; _exists_[0][_i_] =
 ((ccv_nnc_graph_exec_t*)((void*)(((char*)((graph->destinations
)->data)) + (size_t)(graph->destinations)->rsize * (
size_t)(0))))[_i_].d; } _exist_size_[0] = (graph->destinations
->rnum); _exist_size_[1] = 0; _p_ = 0, _q_ = 1; while (_exist_size_
[_p_] > 0) { _exist_size_[_q_] = 0; for (_i_ = 0; _i_ <
 _exist_size_[_p_]; _i_++) { const int32_t _idx_ = _exists_[_p_
][_i_]; if (_incomings_[_idx_].r != 5) continue; _incomings_[
_idx_].r = 6; if (_incomings_[_idx_].edges > 0) for (_j_ =
 0; _j_ < _incomings_[_idx_].c; _j_++) { const int d = _edges_
[_incomings_[_idx_].edges - 1 + _j_]; if (_incomings_[d].r !=
 4) continue; _incomings_[d].r = 5; ((void) sizeof ((_exist_size_
[_q_] < (graph->exec_info->rnum)) ? 1 : 0), __extension__
 ({ if (_exist_size_[_q_] < (graph->exec_info->rnum)
) ; else __assert_fail ("_exist_size_[_q_] < (graph->exec_info->rnum)"
, "ccv_nnc_graph.c", 542, __extension__ __PRETTY_FUNCTION__);
 })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[_q_
]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for (
_i_ = 0; _i_ < (graph->destinations->rnum); _i_++) {
 ((void) sizeof ((((ccv_nnc_graph_exec_t*)((void*)(((char*)((
graph->destinations)->data)) + (size_t)(graph->destinations
)->rsize * (size_t)(0))))[_i_].graph == graph) ? 1 : 0), __extension__
 ({ if (((ccv_nnc_graph_exec_t*)((void*)(((char*)((graph->
destinations)->data)) + (size_t)(graph->destinations)->
rsize * (size_t)(0))))[_i_].graph == graph) ; else __assert_fail
 ("((ccv_nnc_graph_exec_t*)((void*)(((char*)((graph->destinations)->data)) + (size_t)(graph->destinations)->rsize * (size_t)(0))))[_i_].graph == graph"
, "ccv_nnc_graph.c", 542, __extension__ __PRETTY_FUNCTION__);
 })); _incomings_[((ccv_nnc_graph_exec_t*)((void*)(((char*)((
graph->destinations)->data)) + (size_t)(graph->destinations
)->rsize * (size_t)(0))))[_i_].d].d = 1; } for (_i_ = 0; _i_
 < (graph->breakpoint_size); _i_++) { ((void) sizeof ((
(graph->breakpoints)[_i_].graph == graph) ? 1 : 0), __extension__
 ({ if ((graph->breakpoints)[_i_].graph == graph) ; else __assert_fail
 ("(graph->breakpoints)[_i_].graph == graph", "ccv_nnc_graph.c"
, 542, __extension__ __PRETTY_FUNCTION__); })); _exists_[0][_i_
] = (graph->breakpoints)[_i_].d; } _p_ = 0; _q_ = 1; _exist_size_
[0] = (graph->breakpoint_size); _exist_size_[1] = 0; int _d_
 = 0; while (_exist_size_[_p_] > 0) { _exist_size_[_q_] = 0
; for (_i_ = 0; _i_ < _exist_size_[_p_];) { const int32_t _idx_
 = _exists_[_p_][_i_]; _visit_->node[_visit_->size].index
 = ((_idx_)); _visit_->node[_visit_->size].term = ((_incomings_
[_idx_].d)); ++_visit_->size;; if (_incomings_[_idx_].d) {
 ++_d_; _incomings_[_idx_].r = 7; } if (((ccv_nnc_graph_exec_info_t
*)((void*)(((char*)((graph->exec_info)->data)) + (size_t
)(graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings
) { if (((ccv_nnc_graph_exec_info_t*)((void*)(((char*)((graph
->exec_info)->data)) + (size_t)(graph->exec_info)->
rsize * (size_t)(0))))[_idx_].outgoings->rnum == 1) { const
 int d = *(int*)((void*)(((char*)((((ccv_nnc_graph_exec_info_t
*)((void*)(((char*)((graph->exec_info)->data)) + (size_t
)(graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings
)->data)) + (size_t)(((ccv_nnc_graph_exec_info_t*)((void*)
(((char*)((graph->exec_info)->data)) + (size_t)(graph->
exec_info)->rsize * (size_t)(0))))[_idx_].outgoings)->rsize
 * (size_t)(0))); --_incomings_[d].c; if (_incomings_[d].c ==
 0 && _incomings_[d].r == 6 && _d_ < (graph
->destinations->rnum)) { _exists_[_p_][_i_] = d; continue
; } } else for (_j_ = 0; _j_ < ((ccv_nnc_graph_exec_info_t
*)((void*)(((char*)((graph->exec_info)->data)) + (size_t
)(graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings
->rnum; _j_++) { const int d = *(int*)((void*)(((char*)(((
(ccv_nnc_graph_exec_info_t*)((void*)(((char*)((graph->exec_info
)->data)) + (size_t)(graph->exec_info)->rsize * (size_t
)(0))))[_idx_].outgoings)->data)) + (size_t)(((ccv_nnc_graph_exec_info_t
*)((void*)(((char*)((graph->exec_info)->data)) + (size_t
)(graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings
)->rsize * (size_t)(_j_))); --_incomings_[d].c; if (_incomings_
[d].c == 0 && _incomings_[d].r == 6 && _d_ <
 (graph->destinations->rnum)) { ((void) sizeof ((_exist_size_
[_q_] < (graph->exec_info->rnum)) ? 1 : 0), __extension__
 ({ if (_exist_size_[_q_] < (graph->exec_info->rnum)
) ; else __assert_fail ("_exist_size_[_q_] < (graph->exec_info->rnum)"
, "ccv_nnc_graph.c", 542, __extension__ __PRETTY_FUNCTION__);
 })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[_q_
]; } } } ++_i_; } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_
)); } for (_i_ = 0; _i_ < (graph->destinations->rnum
); _i_++) { ((void) sizeof ((((ccv_nnc_graph_exec_t*)((void*)
(((char*)((graph->destinations)->data)) + (size_t)(graph
->destinations)->rsize * (size_t)(0))))[_i_].graph == graph
) ? 1 : 0), __extension__ ({ if (((ccv_nnc_graph_exec_t*)((void
*)(((char*)((graph->destinations)->data)) + (size_t)(graph
->destinations)->rsize * (size_t)(0))))[_i_].graph == graph
) ; else __assert_fail ("((ccv_nnc_graph_exec_t*)((void*)(((char*)((graph->destinations)->data)) + (size_t)(graph->destinations)->rsize * (size_t)(0))))[_i_].graph == graph"
, "ccv_nnc_graph.c", 542, __extension__ __PRETTY_FUNCTION__);
 })); if (_incomings_[((ccv_nnc_graph_exec_t*)((void*)(((char
*)((graph->destinations)->data)) + (size_t)(graph->destinations
)->rsize * (size_t)(0))))[_i_].d].r == 7) continue; if (!(
0)) { ((void) sizeof ((_incomings_[((ccv_nnc_graph_exec_t*)((
void*)(((char*)((graph->destinations)->data)) + (size_t
)(graph->destinations)->rsize * (size_t)(0))))[_i_].d].
c == 0) ? 1 : 0), __extension__ ({ if (_incomings_[((ccv_nnc_graph_exec_t
*)((void*)(((char*)((graph->destinations)->data)) + (size_t
)(graph->destinations)->rsize * (size_t)(0))))[_i_].d].
c == 0) ; else __assert_fail ("_incomings_[((ccv_nnc_graph_exec_t*)((void*)(((char*)((graph->destinations)->data)) + (size_t)(graph->destinations)->rsize * (size_t)(0))))[_i_].d].c == 0"
, "ccv_nnc_graph.c", 542, __extension__ __PRETTY_FUNCTION__);
 })); } else if (_incomings_[((ccv_nnc_graph_exec_t*)((void*)
(((char*)((graph->destinations)->data)) + (size_t)(graph
->destinations)->rsize * (size_t)(0))))[_i_].d].c > 0
) continue; _visit_->node[_visit_->size].index = ((((ccv_nnc_graph_exec_t
*)((void*)(((char*)((graph->destinations)->data)) + (size_t
)(graph->destinations)->rsize * (size_t)(0))))[_i_].d))
; _visit_->node[_visit_->size].term = ((_incomings_[((ccv_nnc_graph_exec_t
*)((void*)(((char*)((graph->destinations)->data)) + (size_t
)(graph->destinations)->rsize * (size_t)(0))))[_i_].d].
d)); ++_visit_->size;; } if (_heap_mem_) free(_incomings_)
; } while (0);; ((void) sizeof ((_visit_->size <= (graph
->exec_info->rnum)) ? 1 : 0), __extension__ ({ if (_visit_
->size <= (graph->exec_info->rnum)) ; else __assert_fail
 ("_visit_->size <= (graph->exec_info->rnum)", "ccv_nnc_graph.c"
, 542, __extension__ __PRETTY_FUNCTION__); })); _visit_; });
543		ccv_nnc_graph_visit_for(visit, (ccv_nnc_graph_exec_info_t*)ccv_array_get(graph->exec_info, 0), node, idx){ int _i_; for (_i_ = 0; _i_ < (visit)->size; _i_++) { const
 int idx __attribute__((unused)) = (visit)->node[_i_].index
; const int _node_unused_ __attribute__((unused)) = (visit)->
node[_i_].term; typeof (((ccv_nnc_graph_exec_info_t*)((void*)
(((char*)((graph->exec_info)->data)) + (size_t)(graph->
exec_info)->rsize * (size_t)(0))))) const node __attribute__
((unused)) = ((ccv_nnc_graph_exec_info_t*)((void*)(((char*)((
graph->exec_info)->data)) + (size_t)(graph->exec_info
)->rsize * (size_t)(0)))) + idx; {
544			assert(!node->pair_ref)((void) sizeof ((!node->pair_ref) ? 1 : 0), __extension__ (
{ if (!node->pair_ref) ; else __assert_fail ("!node->pair_ref"
, "ccv_nnc_graph.c", 544, __extension__ __PRETTY_FUNCTION__);
 })); // If node has a pair ref, we cannot fix it up.
545			// Loop over node and push to the array.
546			ccv_array_push(exec_info, node);
547			// Go to its sub-graph to fix exec_idx
548			for (i = 0; i < node->graph_ref_size; i++)
549			{
550				const int graph_ref = CCV_NNC_GRAPH_REF(node)((node)->_heap_graph_ref ? (node)->_heap_graph_ref : (node
)->_inline_graph_ref)[i] - 1;
551				if (graph_ref >= 0)
552				{
553					ccv_nnc_graph_t* const sub_graph = *(ccv_nnc_graph_t**)ccv_array_get(graph->sub_graphs, graph_ref)((void*)(((char*)((graph->sub_graphs)->data)) + (size_t
)(graph->sub_graphs)->rsize * (size_t)(graph_ref)));
554					sub_graph->exec_idx = exec_info->rnum;
555				}
556			}
557			exec_cvt[idx] = exec_info->rnum - 1;
558		} ccv_nnc_graph_visit_endfor} }
559		ccv_nnc_graph_visit_free(visit);
560		for (i = 0; i < graph->breakpoint_size; i++)
561			{ assert(exec_cvt[graph->breakpoints[i].d] >= 0)((void) sizeof ((exec_cvt[graph->breakpoints[i].d] >= 0
) ? 1 : 0), __extension__ ({ if (exec_cvt[graph->breakpoints
[i].d] >= 0) ; else __assert_fail ("exec_cvt[graph->breakpoints[i].d] >= 0"
, "ccv_nnc_graph.c", 561, __extension__ __PRETTY_FUNCTION__);
 })); } // All breakpoints should be assigned.
562	} else {
563		ccv_nnc_graph_visit_t* visit = ccv_nnc_graph_visit_new(graph, (ccv_nnc_graph_exec_info_t*)ccv_array_get(graph->exec_info, 0), graph->exec_info->rnum, (ccv_nnc_graph_exec_t*)ccv_array_get(graph->sources, 0), graph->sources->rnum, (ccv_nnc_graph_exec_t*)ccv_array_get(graph->destinations, 0), graph->destinations->rnum, 0)({ ccv_nnc_graph_visit_t* _visit_ = (ccv_nnc_graph_visit_t*)malloc
(sizeof(ccv_nnc_graph_visit_t) + sizeof(_visit_->node[0]) *
 ((graph->exec_info->rnum) - 1)); _visit_->size = 0;
 do { typedef struct { int8_t d; int8_t r; uint16_t c; int32_t
 edges; } ccv_nnc_incoming_t; int _i_, _j_; int _incoming_edges_
 = 0; for (_i_ = 0; _i_ < (graph->exec_info->rnum); _i_
++) _incoming_edges_ += (((ccv_nnc_graph_exec_info_t*)((void*
)(((char*)((graph->exec_info)->data)) + (size_t)(graph->
exec_info)->rsize * (size_t)(0))))[_i_].outgoings) ? ((ccv_nnc_graph_exec_info_t
*)((void*)(((char*)((graph->exec_info)->data)) + (size_t
)(graph->exec_info)->rsize * (size_t)(0))))[_i_].outgoings
->rnum : 0; const int _heap_mem_ = ((graph->exec_info->
rnum) + _incoming_edges_ > 1024); ccv_nnc_incoming_t* _incomings_
; if (_heap_mem_) _incomings_ = (ccv_nnc_incoming_t*)malloc(sizeof
(ccv_nnc_incoming_t) * (graph->exec_info->rnum) + sizeof
(int32_t) * ((graph->exec_info->rnum) * 2 + _incoming_edges_
)); else _incomings_ = (ccv_nnc_incoming_t*)__builtin_alloca (
sizeof(ccv_nnc_incoming_t) * (graph->exec_info->rnum) +
 sizeof(int32_t) * ((graph->exec_info->rnum) * 2 + _incoming_edges_
)); memset(_incomings_, 0, sizeof(ccv_nnc_incoming_t) * (graph
->exec_info->rnum)); int32_t* _exists_[2] = { (int32_t*
)(_incomings_ + (graph->exec_info->rnum)), (int32_t*)(_incomings_
 + (graph->exec_info->rnum)) + (graph->exec_info->
rnum), }; int32_t* const _edges_ = _exists_[1] + (graph->exec_info
->rnum); for (_i_ = 0; _i_ < (graph->sources->rnum
); _i_++) { ((void) sizeof ((((ccv_nnc_graph_exec_t*)((void*)
(((char*)((graph->sources)->data)) + (size_t)(graph->
sources)->rsize * (size_t)(0))))[_i_].graph == graph) ? 1 :
 0), __extension__ ({ if (((ccv_nnc_graph_exec_t*)((void*)(((
char*)((graph->sources)->data)) + (size_t)(graph->sources
)->rsize * (size_t)(0))))[_i_].graph == graph) ; else __assert_fail
 ("((ccv_nnc_graph_exec_t*)((void*)(((char*)((graph->sources)->data)) + (size_t)(graph->sources)->rsize * (size_t)(0))))[_i_].graph == graph"
, "ccv_nnc_graph.c", 563, __extension__ __PRETTY_FUNCTION__);
 })); _incomings_[((ccv_nnc_graph_exec_t*)((void*)(((char*)((
graph->sources)->data)) + (size_t)(graph->sources)->
rsize * (size_t)(0))))[_i_].d].r = 1; _exists_[0][_i_] = ((ccv_nnc_graph_exec_t
*)((void*)(((char*)((graph->sources)->data)) + (size_t)
(graph->sources)->rsize * (size_t)(0))))[_i_].d; } int _exist_size_
[2] = { (graph->sources->rnum), 0, }; int _p_ = 0, _q_ =
 1; while (_exist_size_[_p_] > 0) { _exist_size_[_q_] = 0;
 for (_i_ = 0; _i_ < _exist_size_[_p_]; _i_++) { const int32_t
 _idx_ = _exists_[_p_][_i_]; if (_incomings_[_idx_].r != 1) continue
; _incomings_[_idx_].r = 2; if (((ccv_nnc_graph_exec_info_t*)
((void*)(((char*)((graph->exec_info)->data)) + (size_t)
(graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings
) for (_j_ = 0; _j_ < ((ccv_nnc_graph_exec_info_t*)((void*
)(((char*)((graph->exec_info)->data)) + (size_t)(graph->
exec_info)->rsize * (size_t)(0))))[_idx_].outgoings->rnum
; _j_++) { const int d = *(int*)((void*)(((char*)((((ccv_nnc_graph_exec_info_t
*)((void*)(((char*)((graph->exec_info)->data)) + (size_t
)(graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings
)->data)) + (size_t)(((ccv_nnc_graph_exec_info_t*)((void*)
(((char*)((graph->exec_info)->data)) + (size_t)(graph->
exec_info)->rsize * (size_t)(0))))[_idx_].outgoings)->rsize
 * (size_t)(_j_))); ++_incomings_[d].c; if (_incomings_[d].r !=
 0) continue; _incomings_[d].r = 1; ((void) sizeof ((_exist_size_
[_q_] < (graph->exec_info->rnum)) ? 1 : 0), __extension__
 ({ if (_exist_size_[_q_] < (graph->exec_info->rnum)
) ; else __assert_fail ("_exist_size_[_q_] < (graph->exec_info->rnum)"
, "ccv_nnc_graph.c", 563, __extension__ __PRETTY_FUNCTION__);
 })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[_q_
]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for (
_i_ = 0; _i_ < (graph->sources->rnum); _i_++) { ((void
) sizeof ((((ccv_nnc_graph_exec_t*)((void*)(((char*)((graph->
sources)->data)) + (size_t)(graph->sources)->rsize *
 (size_t)(0))))[_i_].graph == graph) ? 1 : 0), __extension__ (
{ if (((ccv_nnc_graph_exec_t*)((void*)(((char*)((graph->sources
)->data)) + (size_t)(graph->sources)->rsize * (size_t
)(0))))[_i_].graph == graph) ; else __assert_fail ("((ccv_nnc_graph_exec_t*)((void*)(((char*)((graph->sources)->data)) + (size_t)(graph->sources)->rsize * (size_t)(0))))[_i_].graph == graph"
, "ccv_nnc_graph.c", 563, __extension__ __PRETTY_FUNCTION__);
 })); _incomings_[((ccv_nnc_graph_exec_t*)((void*)(((char*)((
graph->sources)->data)) + (size_t)(graph->sources)->
rsize * (size_t)(0))))[_i_].d].r = 3; _exists_[0][_i_] = ((ccv_nnc_graph_exec_t
*)((void*)(((char*)((graph->sources)->data)) + (size_t)
(graph->sources)->rsize * (size_t)(0))))[_i_].d; } _exist_size_
[0] = (graph->sources->rnum); _exist_size_[1] = 0; _p_ =
 0, _q_ = 1; int _bump_ = 1; while (_exist_size_[_p_] > 0)
 { _exist_size_[_q_] = 0; for (_i_ = 0; _i_ < _exist_size_
[_p_]; _i_++) { const int32_t _idx_ = _exists_[_p_][_i_]; if (
_incomings_[_idx_].r != 3) continue; _incomings_[_idx_].r = 4
; if (((ccv_nnc_graph_exec_info_t*)((void*)(((char*)((graph->
exec_info)->data)) + (size_t)(graph->exec_info)->rsize
 * (size_t)(0))))[_idx_].outgoings) for (_j_ = 0; _j_ < ((
ccv_nnc_graph_exec_info_t*)((void*)(((char*)((graph->exec_info
)->data)) + (size_t)(graph->exec_info)->rsize * (size_t
)(0))))[_idx_].outgoings->rnum; _j_++) { const int d = *(int
*)((void*)(((char*)((((ccv_nnc_graph_exec_info_t*)((void*)(((
char*)((graph->exec_info)->data)) + (size_t)(graph->
exec_info)->rsize * (size_t)(0))))[_idx_].outgoings)->data
)) + (size_t)(((ccv_nnc_graph_exec_info_t*)((void*)(((char*)(
(graph->exec_info)->data)) + (size_t)(graph->exec_info
)->rsize * (size_t)(0))))[_idx_].outgoings)->rsize * (size_t
)(_j_))); if (_incomings_[d].edges == 0) { _incomings_[d].edges
 = _bump_; _bump_ += _incomings_[d].c; _incomings_[d].c = 0; }
 _edges_[_incomings_[d].edges - 1 + _incomings_[d].c] = _idx_
; ++_incomings_[d].c; if (_incomings_[d].r != 2) continue; _incomings_
[d].r = 3; ((void) sizeof ((_exist_size_[_q_] < (graph->
exec_info->rnum)) ? 1 : 0), __extension__ ({ if (_exist_size_
[_q_] < (graph->exec_info->rnum)) ; else __assert_fail
 ("_exist_size_[_q_] < (graph->exec_info->rnum)", "ccv_nnc_graph.c"
, 563, __extension__ __PRETTY_FUNCTION__); })); _exists_[_q_]
[_exist_size_[_q_]] = d; ++_exist_size_[_q_]; } } ((_i_) = (_p_
), (_p_) = (_q_), (_q_) = (_i_)); } for (_i_ = 0; _i_ < (graph
->destinations->rnum); _i_++) { ((void) sizeof ((((ccv_nnc_graph_exec_t
*)((void*)(((char*)((graph->destinations)->data)) + (size_t
)(graph->destinations)->rsize * (size_t)(0))))[_i_].graph
 == graph) ? 1 : 0), __extension__ ({ if (((ccv_nnc_graph_exec_t
*)((void*)(((char*)((graph->destinations)->data)) + (size_t
)(graph->destinations)->rsize * (size_t)(0))))[_i_].graph
 == graph) ; else __assert_fail ("((ccv_nnc_graph_exec_t*)((void*)(((char*)((graph->destinations)->data)) + (size_t)(graph->destinations)->rsize * (size_t)(0))))[_i_].graph == graph"
, "ccv_nnc_graph.c", 563, __extension__ __PRETTY_FUNCTION__);
 })); _incomings_[((ccv_nnc_graph_exec_t*)((void*)(((char*)((
graph->destinations)->data)) + (size_t)(graph->destinations
)->rsize * (size_t)(0))))[_i_].d].r = 5; _exists_[0][_i_] =
 ((ccv_nnc_graph_exec_t*)((void*)(((char*)((graph->destinations
)->data)) + (size_t)(graph->destinations)->rsize * (
size_t)(0))))[_i_].d; } _exist_size_[0] = (graph->destinations
->rnum); _exist_size_[1] = 0; _p_ = 0, _q_ = 1; while (_exist_size_
[_p_] > 0) { _exist_size_[_q_] = 0; for (_i_ = 0; _i_ <
 _exist_size_[_p_]; _i_++) { const int32_t _idx_ = _exists_[_p_
][_i_]; if (_incomings_[_idx_].r != 5) continue; _incomings_[
_idx_].r = 6; if (_incomings_[_idx_].edges > 0) for (_j_ =
 0; _j_ < _incomings_[_idx_].c; _j_++) { const int d = _edges_
[_incomings_[_idx_].edges - 1 + _j_]; if (_incomings_[d].r !=
 4) continue; _incomings_[d].r = 5; ((void) sizeof ((_exist_size_
[_q_] < (graph->exec_info->rnum)) ? 1 : 0), __extension__
 ({ if (_exist_size_[_q_] < (graph->exec_info->rnum)
) ; else __assert_fail ("_exist_size_[_q_] < (graph->exec_info->rnum)"
, "ccv_nnc_graph.c", 563, __extension__ __PRETTY_FUNCTION__);
 })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[_q_
]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for (
_i_ = 0; _i_ < (graph->destinations->rnum); _i_++) {
 ((void) sizeof ((((ccv_nnc_graph_exec_t*)((void*)(((char*)((
graph->destinations)->data)) + (size_t)(graph->destinations
)->rsize * (size_t)(0))))[_i_].graph == graph) ? 1 : 0), __extension__
 ({ if (((ccv_nnc_graph_exec_t*)((void*)(((char*)((graph->
destinations)->data)) + (size_t)(graph->destinations)->
rsize * (size_t)(0))))[_i_].graph == graph) ; else __assert_fail
 ("((ccv_nnc_graph_exec_t*)((void*)(((char*)((graph->destinations)->data)) + (size_t)(graph->destinations)->rsize * (size_t)(0))))[_i_].graph == graph"
, "ccv_nnc_graph.c", 563, __extension__ __PRETTY_FUNCTION__);
 })); _incomings_[((ccv_nnc_graph_exec_t*)((void*)(((char*)((
graph->destinations)->data)) + (size_t)(graph->destinations
)->rsize * (size_t)(0))))[_i_].d].d = 1; } for (_i_ = 0; _i_
 < (graph->sources->rnum); _i_++) { ((void) sizeof (
(((ccv_nnc_graph_exec_t*)((void*)(((char*)((graph->sources
)->data)) + (size_t)(graph->sources)->rsize * (size_t
)(0))))[_i_].graph == graph) ? 1 : 0), __extension__ ({ if ((
(ccv_nnc_graph_exec_t*)((void*)(((char*)((graph->sources)->
data)) + (size_t)(graph->sources)->rsize * (size_t)(0))
))[_i_].graph == graph) ; else __assert_fail ("((ccv_nnc_graph_exec_t*)((void*)(((char*)((graph->sources)->data)) + (size_t)(graph->sources)->rsize * (size_t)(0))))[_i_].graph == graph"
, "ccv_nnc_graph.c", 563, __extension__ __PRETTY_FUNCTION__);
 })); _exists_[0][_i_] = ((ccv_nnc_graph_exec_t*)((void*)(((char
*)((graph->sources)->data)) + (size_t)(graph->sources
)->rsize * (size_t)(0))))[_i_].d; } _p_ = 0; _q_ = 1; _exist_size_
[0] = (graph->sources->rnum); _exist_size_[1] = 0; int _d_
 = 0; while (_exist_size_[_p_] > 0) { _exist_size_[_q_] = 0
; for (_i_ = 0; _i_ < _exist_size_[_p_];) { const int32_t _idx_
 = _exists_[_p_][_i_]; _visit_->node[_visit_->size].index
 = ((_idx_)); _visit_->node[_visit_->size].term = ((_incomings_
[_idx_].d)); ++_visit_->size;; if (_incomings_[_idx_].d) {
 ++_d_; _incomings_[_idx_].r = 7; } if (((ccv_nnc_graph_exec_info_t
*)((void*)(((char*)((graph->exec_info)->data)) + (size_t
)(graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings
) { if (((ccv_nnc_graph_exec_info_t*)((void*)(((char*)((graph
->exec_info)->data)) + (size_t)(graph->exec_info)->
rsize * (size_t)(0))))[_idx_].outgoings->rnum == 1) { const
 int d = *(int*)((void*)(((char*)((((ccv_nnc_graph_exec_info_t
*)((void*)(((char*)((graph->exec_info)->data)) + (size_t
)(graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings
)->data)) + (size_t)(((ccv_nnc_graph_exec_info_t*)((void*)
(((char*)((graph->exec_info)->data)) + (size_t)(graph->
exec_info)->rsize * (size_t)(0))))[_idx_].outgoings)->rsize
 * (size_t)(0))); --_incomings_[d].c; if (_incomings_[d].c ==
 0 && _incomings_[d].r == 6 && _d_ < (graph
->destinations->rnum)) { _exists_[_p_][_i_] = d; continue
; } } else for (_j_ = 0; _j_ < ((ccv_nnc_graph_exec_info_t
*)((void*)(((char*)((graph->exec_info)->data)) + (size_t
)(graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings
->rnum; _j_++) { const int d = *(int*)((void*)(((char*)(((
(ccv_nnc_graph_exec_info_t*)((void*)(((char*)((graph->exec_info
)->data)) + (size_t)(graph->exec_info)->rsize * (size_t
)(0))))[_idx_].outgoings)->data)) + (size_t)(((ccv_nnc_graph_exec_info_t
*)((void*)(((char*)((graph->exec_info)->data)) + (size_t
)(graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings
)->rsize * (size_t)(_j_))); --_incomings_[d].c; if (_incomings_
[d].c == 0 && _incomings_[d].r == 6 && _d_ <
 (graph->destinations->rnum)) { ((void) sizeof ((_exist_size_
[_q_] < (graph->exec_info->rnum)) ? 1 : 0), __extension__
 ({ if (_exist_size_[_q_] < (graph->exec_info->rnum)
) ; else __assert_fail ("_exist_size_[_q_] < (graph->exec_info->rnum)"
, "ccv_nnc_graph.c", 563, __extension__ __PRETTY_FUNCTION__);
 })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[_q_
]; } } } ++_i_; } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_
)); } for (_i_ = 0; _i_ < (graph->destinations->rnum
); _i_++) { ((void) sizeof ((((ccv_nnc_graph_exec_t*)((void*)
(((char*)((graph->destinations)->data)) + (size_t)(graph
->destinations)->rsize * (size_t)(0))))[_i_].graph == graph
) ? 1 : 0), __extension__ ({ if (((ccv_nnc_graph_exec_t*)((void
*)(((char*)((graph->destinations)->data)) + (size_t)(graph
->destinations)->rsize * (size_t)(0))))[_i_].graph == graph
) ; else __assert_fail ("((ccv_nnc_graph_exec_t*)((void*)(((char*)((graph->destinations)->data)) + (size_t)(graph->destinations)->rsize * (size_t)(0))))[_i_].graph == graph"
, "ccv_nnc_graph.c", 563, __extension__ __PRETTY_FUNCTION__);
 })); if (_incomings_[((ccv_nnc_graph_exec_t*)((void*)(((char
*)((graph->destinations)->data)) + (size_t)(graph->destinations
)->rsize * (size_t)(0))))[_i_].d].r == 7) continue; if (!(
0)) { ((void) sizeof ((_incomings_[((ccv_nnc_graph_exec_t*)((
void*)(((char*)((graph->destinations)->data)) + (size_t
)(graph->destinations)->rsize * (size_t)(0))))[_i_].d].
c == 0) ? 1 : 0), __extension__ ({ if (_incomings_[((ccv_nnc_graph_exec_t
*)((void*)(((char*)((graph->destinations)->data)) + (size_t
)(graph->destinations)->rsize * (size_t)(0))))[_i_].d].
c == 0) ; else __assert_fail ("_incomings_[((ccv_nnc_graph_exec_t*)((void*)(((char*)((graph->destinations)->data)) + (size_t)(graph->destinations)->rsize * (size_t)(0))))[_i_].d].c == 0"
, "ccv_nnc_graph.c", 563, __extension__ __PRETTY_FUNCTION__);
 })); } else if (_incomings_[((ccv_nnc_graph_exec_t*)((void*)
(((char*)((graph->destinations)->data)) + (size_t)(graph
->destinations)->rsize * (size_t)(0))))[_i_].d].c > 0
) continue; _visit_->node[_visit_->size].index = ((((ccv_nnc_graph_exec_t
*)((void*)(((char*)((graph->destinations)->data)) + (size_t
)(graph->destinations)->rsize * (size_t)(0))))[_i_].d))
; _visit_->node[_visit_->size].term = ((_incomings_[((ccv_nnc_graph_exec_t
*)((void*)(((char*)((graph->destinations)->data)) + (size_t
)(graph->destinations)->rsize * (size_t)(0))))[_i_].d].
d)); ++_visit_->size;; } if (_heap_mem_) free(_incomings_)
; } while (0);; ((void) sizeof ((_visit_->size <= (graph
->exec_info->rnum)) ? 1 : 0), __extension__ ({ if (_visit_
->size <= (graph->exec_info->rnum)) ; else __assert_fail
 ("_visit_->size <= (graph->exec_info->rnum)", "ccv_nnc_graph.c"
, 563, __extension__ __PRETTY_FUNCTION__); })); _visit_; });
564		ccv_nnc_graph_visit_for(visit, (ccv_nnc_graph_exec_info_t*)ccv_array_get(graph->exec_info, 0), node, idx){ int _i_; for (_i_ = 0; _i_ < (visit)->size; _i_++) { const
 int idx __attribute__((unused)) = (visit)->node[_i_].index
; const int _node_unused_ __attribute__((unused)) = (visit)->
node[_i_].term; typeof (((ccv_nnc_graph_exec_info_t*)((void*)
(((char*)((graph->exec_info)->data)) + (size_t)(graph->
exec_info)->rsize * (size_t)(0))))) const node __attribute__
((unused)) = ((ccv_nnc_graph_exec_info_t*)((void*)(((char*)((
graph->exec_info)->data)) + (size_t)(graph->exec_info
)->rsize * (size_t)(0)))) + idx; {
565			assert(!node->pair_ref)((void) sizeof ((!node->pair_ref) ? 1 : 0), __extension__ (
{ if (!node->pair_ref) ; else __assert_fail ("!node->pair_ref"
, "ccv_nnc_graph.c", 565, __extension__ __PRETTY_FUNCTION__);
 })); // If node has a pair ref, we cannot fix it up.
566			// Loop over node and push to the array.
567			ccv_array_push(exec_info, node);
568			// Go to its sub-graph to fix exec_idx
569			for (i = 0; i < node->graph_ref_size; i++)
570			{
571				const int graph_ref = CCV_NNC_GRAPH_REF(node)((node)->_heap_graph_ref ? (node)->_heap_graph_ref : (node
)->_inline_graph_ref)[i] - 1;
572				if (graph_ref >= 0)
573				{
574					ccv_nnc_graph_t* const sub_graph = *(ccv_nnc_graph_t**)ccv_array_get(graph->sub_graphs, graph_ref)((void*)(((char*)((graph->sub_graphs)->data)) + (size_t
)(graph->sub_graphs)->rsize * (size_t)(graph_ref)));
575					sub_graph->exec_idx = exec_info->rnum;
576				}
577			}
578			exec_cvt[idx] = exec_info->rnum - 1;
579		} ccv_nnc_graph_visit_endfor} }
580		ccv_nnc_graph_visit_free(visit);
581	}
582	assert(graph->exec_info->rnum == exec_info->rnum)((void) sizeof ((graph->exec_info->rnum == exec_info->
rnum) ? 1 : 0), __extension__ ({ if (graph->exec_info->
rnum == exec_info->rnum) ; else __assert_fail ("graph->exec_info->rnum == exec_info->rnum"
, "ccv_nnc_graph.c", 582, __extension__ __PRETTY_FUNCTION__);
 }));
583	ccv_array_free(graph->exec_info);
584	graph->exec_info = exec_info;
585	for (i = 0; i < graph->sources->rnum; i++)
586	{
587		ccv_nnc_graph_exec_t* const source = (ccv_nnc_graph_exec_t*)ccv_array_get(graph->sources, i)((void*)(((char*)((graph->sources)->data)) + (size_t)(graph
->sources)->rsize * (size_t)(i)));
588		source->d = exec_cvt[source->d];
589	}
590	for (i = 0; i < graph->destinations->rnum; i++)
591	{
592		ccv_nnc_graph_exec_t* const destination = (ccv_nnc_graph_exec_t*)ccv_array_get(graph->destinations, i)((void*)(((char*)((graph->destinations)->data)) + (size_t
)(graph->destinations)->rsize * (size_t)(i)));
593		destination->d = exec_cvt[destination->d];
594	}
595	// Update all outgoings to reflect the latest.
596	for (i = 0; i < exec_info->rnum; i++)
597	{
598		ccv_nnc_graph_exec_info_t* const info = (ccv_nnc_graph_exec_info_t*)ccv_array_get(exec_info, i)((void*)(((char*)((exec_info)->data)) + (size_t)(exec_info
)->rsize * (size_t)(i)));
599		if (info->outgoings)
600			for (j = 0; j < info->outgoings->rnum; j++)
601				*(int*)ccv_array_get(info->outgoings, j)((void*)(((char*)((info->outgoings)->data)) + (size_t)(
info->outgoings)->rsize * (size_t)(j))) = exec_cvt[*(int*)ccv_array_get(info->outgoings, j)((void*)(((char*)((info->outgoings)->data)) + (size_t)(
info->outgoings)->rsize * (size_t)(j)))];
602	}
603	graph->topsorted = 1;
604}
605 
606typedef struct {
607	int device_id;
608	int exec_idx;
609	ccv_array_t* signal_set;
610	ccv_array_t* command_set; // The set of command executed in this stream. In case there is a tie (on rank). We will check this.
611} ccv_nnc_stream_data_t;
612 
613static void _ccv_nnc_graph_schedule_assign_signals(ccv_array_t* const incoming, ccv_nnc_graph_exec_schedule_t* const node, ccv_array_t* const stream_data, int* const signal_size, ccv_nnc_graph_exec_schedule_t* const exec_info, const int exec_info_size)
614{
615	assert(incoming->rnum > 0)((void) sizeof ((incoming->rnum > 0) ? 1 : 0), __extension__
 ({ if (incoming->rnum > 0) ; else __assert_fail ("incoming->rnum > 0"
, "ccv_nnc_graph.c", 615, __extension__ __PRETTY_FUNCTION__);
 }));
616	int i, j, k;
617	int wait_size = 0, max_wait_size = 0;
618	for (i = 0; i < incoming->rnum; i++)
619	{
620		const int incoming_idx = *(int*)ccv_array_get(incoming, i)((void*)(((char*)((incoming)->data)) + (size_t)(incoming)->
rsize * (size_t)(i)));
621		ccv_nnc_graph_exec_schedule_t* const incoming_exec_info = exec_info + incoming_idx;
622		assert(incoming_exec_info->stream_size > 0)((void) sizeof ((incoming_exec_info->stream_size > 0) ?
 1 : 0), __extension__ ({ if (incoming_exec_info->stream_size
 > 0) ; else __assert_fail ("incoming_exec_info->stream_size > 0"
, "ccv_nnc_graph.c", 622, __extension__ __PRETTY_FUNCTION__);
 }));
623		max_wait_size += incoming_exec_info->stream_size;
624	}
625	int waits[ccv_max(1, max_wait_size)({ typeof (1) _a = (1); typeof (max_wait_size) _b = (max_wait_size
); (_a > _b) ? _a : _b; })];
626	assert(node->stream_size > 0)((void) sizeof ((node->stream_size > 0) ? 1 : 0), __extension__
 ({ if (node->stream_size > 0) ; else __assert_fail ("node->stream_size > 0"
, "ccv_nnc_graph.c", 626, __extension__ __PRETTY_FUNCTION__);
 }));
627	for (i = 0; i < incoming->rnum; i++)
628	{
629		const int incoming_idx = *(int*)ccv_array_get(incoming, i)((void*)(((char*)((incoming)->data)) + (size_t)(incoming)->
rsize * (size_t)(i)));
630		assert(incoming_idx < exec_info_size)((void) sizeof ((incoming_idx < exec_info_size) ? 1 : 0), __extension__
 ({ if (incoming_idx < exec_info_size) ; else __assert_fail
 ("incoming_idx < exec_info_size", "ccv_nnc_graph.c", 630,
 __extension__ __PRETTY_FUNCTION__); }));
631		assert(incoming_idx >= 0)((void) sizeof ((incoming_idx >= 0) ? 1 : 0), __extension__
 ({ if (incoming_idx >= 0) ; else __assert_fail ("incoming_idx >= 0"
, "ccv_nnc_graph.c", 631, __extension__ __PRETTY_FUNCTION__);
 }));
632		ccv_nnc_graph_exec_schedule_t* const incoming_exec_info = exec_info + incoming_idx;
633		assert(incoming_exec_info->stream_size > 0)((void) sizeof ((incoming_exec_info->stream_size > 0) ?
 1 : 0), __extension__ ({ if (incoming_exec_info->stream_size
 > 0) ; else __assert_fail ("incoming_exec_info->stream_size > 0"
, "ccv_nnc_graph.c", 633, __extension__ __PRETTY_FUNCTION__);
 }));
634		int stream_synced = 1;
635		// If the current node's stream is a subset of the incoming node's stream, there
636		// is no need to sync with signal, because we are already synced with the incoming.
637		for (j = 0; stream_synced && j < node->stream_size; j++)
638		{
639			const int s = SCHEDULE_STREAMS(*node)((*node).stream_size <= 1 ? (*node)._inline_streams : (*node
)._heap_streams)[j];
640			assert(s >= 0)((void) sizeof ((s >= 0) ? 1 : 0), __extension__ ({ if (s >=
 0) ; else __assert_fail ("s >= 0", "ccv_nnc_graph.c", 640
, __extension__ __PRETTY_FUNCTION__); }));
641			int flag = 0;
642			for (k = 0; !flag && k < incoming_exec_info->stream_size; k++)
643				flag = (SCHEDULE_STREAMS(*incoming_exec_info)((*incoming_exec_info).stream_size <= 1 ? (*incoming_exec_info
)._inline_streams : (*incoming_exec_info)._heap_streams)[k] == s);
644			stream_synced = flag;
645		}
646		if (stream_synced)
647			continue;
648		// Otherwise, find the streams we need to sync with, and create signals for these.
649		for (j = 0; j < incoming_exec_info->stream_size; j++)
650		{
651			const int s = SCHEDULE_STREAMS(*incoming_exec_info)((*incoming_exec_info).stream_size <= 1 ? (*incoming_exec_info
)._inline_streams : (*incoming_exec_info)._heap_streams)[j];
652			assert(s >= 0)((void) sizeof ((s >= 0) ? 1 : 0), __extension__ ({ if (s >=
 0) ; else __assert_fail ("s >= 0", "ccv_nnc_graph.c", 652
, __extension__ __PRETTY_FUNCTION__); }));
653			int flag = 0;
654			for (k = 0; !flag && k < node->stream_size; k++)
655				flag = (SCHEDULE_STREAMS(*node)((*node).stream_size <= 1 ? (*node)._inline_streams : (*node
)._heap_streams)[k] == s);
656			if (!flag) // Need to have a signal.
657			{
658				if (SCHEDULE_SIGNALS(*incoming_exec_info)((*incoming_exec_info).stream_size <= 1 ? (*incoming_exec_info
)._inline_signals : (*incoming_exec_info)._heap_signals)[j] < 0)
659					SCHEDULE_SIGNALS(*incoming_exec_info)((*incoming_exec_info).stream_size <= 1 ? (*incoming_exec_info
)._inline_signals : (*incoming_exec_info)._heap_signals)[j] = (*signal_size)++;
660				else {
661					int flag = 0;
662					// If any of the stream the current node has already seen this signal, we are good already.
663					for (k = 0; !flag && k < node->stream_size; k++)
664					{
665						assert(SCHEDULE_STREAMS(*node)[k] >= 0)((void) sizeof ((((*node).stream_size <= 1 ? (*node)._inline_streams
 : (*node)._heap_streams)[k] >= 0) ? 1 : 0), __extension__
 ({ if (((*node).stream_size <= 1 ? (*node)._inline_streams
 : (*node)._heap_streams)[k] >= 0) ; else __assert_fail ("SCHEDULE_STREAMS(*node)[k] >= 0"
, "ccv_nnc_graph.c", 665, __extension__ __PRETTY_FUNCTION__);
 }));
666						ccv_nnc_stream_data_t* const data = (ccv_nnc_stream_data_t*)ccv_array_get(stream_data, SCHEDULE_STREAMS(*node)[k])((void*)(((char*)((stream_data)->data)) + (size_t)(stream_data
)->rsize * (size_t)(((*node).stream_size <= 1 ? (*node)
._inline_streams : (*node)._heap_streams)[k])));
667						flag = (data->signal_set && ccv_array_find_int(data->signal_set, SCHEDULE_SIGNALS(*incoming_exec_info)((*incoming_exec_info).stream_size <= 1 ? (*incoming_exec_info
)._inline_signals : (*incoming_exec_info)._heap_signals)[j]));
668					}
669					if (flag)
670						continue;
671				}
672				// Otherwise, we need to wait for this. Currently, our granularity is about wait on all streams.
673				waits[wait_size++] = SCHEDULE_SIGNALS(*incoming_exec_info)((*incoming_exec_info).stream_size <= 1 ? (*incoming_exec_info
)._inline_signals : (*incoming_exec_info)._heap_signals)[j];
674				// All streams on this node have seen this signal.
675				for (k = 0; k < node->stream_size; k++)
676				{
677					ccv_nnc_stream_data_t* const data = (ccv_nnc_stream_data_t*)ccv_array_get(stream_data, SCHEDULE_STREAMS(*node)[k])((void*)(((char*)((stream_data)->data)) + (size_t)(stream_data
)->rsize * (size_t)(((*node).stream_size <= 1 ? (*node)
._inline_streams : (*node)._heap_streams)[k])));
678					if (!data->signal_set)
679						data->signal_set = ccv_array_new(sizeof(int), 0, 0);
680					ccv_array_push(data->signal_set, &SCHEDULE_SIGNALS(*incoming_exec_info)((*incoming_exec_info).stream_size <= 1 ? (*incoming_exec_info
)._inline_signals : (*incoming_exec_info)._heap_signals)[j]);
681				}
682			}
683		}
684	}
685	node->wait_size = wait_size;
686	if (wait_size > 0)
687	{
688		node->waits = node->waits ? ccreallocrealloc(node->waits, sizeof(int) * wait_size) : ccmallocmalloc(sizeof(int) * wait_size);
689		memcpy(node->waits, waits, sizeof(int) * wait_size);
690	}
691}
692 
693typedef struct {
694	int rank;
695	ccv_array_t* outgoings;
696} ccv_nnc_incoming_t;
697 
698static int _ccv_nnc_device_ids_for_stream_data(ccv_nnc_graph_exec_info_t* const node, const int device_id, ccv_array_t* const stream_data, int* const device_ids, const int max_device_id_size)
699{
700	// TODO: I need to re-think whether this is GPU only or not.
701	int device_id_size = ccv_nnc_device_ids_for_io(node->inputs, node->input_size, node->outputs, node->output_size, CCV_TENSOR_GPU_MEMORY, device_ids, max_device_id_size);
702	if (device_id_size == 0)
703	{
704		// If there is a default data, use that device id. Otherwise, use the device id passed in (this will be the default data device id).
705		if (stream_data->rnum > 0)
706		{
707			ccv_nnc_stream_data_t* const default_data = (ccv_nnc_stream_data_t*)ccv_array_get(stream_data, 0)((void*)(((char*)((stream_data)->data)) + (size_t)(stream_data
)->rsize * (size_t)(0)));
708			device_ids[0] = default_data->device_id;
709		} else
710			device_ids[0] = device_id >= 0 ? device_id : 0;
711		device_id_size = 1;
712	}
713	return device_id_size;
714}
715 
716void ccv_nnc_graph_static_schedule_free(ccv_nnc_graph_static_schedule_t* const schedule)
717{
718	int i;
719	ccv_nnc_graph_exec_schedule_t* const schd_info = schedule->exec_info;
720	for (i = 0; i < schedule->exec_info_size; i++)
721	{
722		if (schd_info[i].stream_size > 1)
723			ccfreefree(schd_info[i]._heap_streams);
724		if (schd_info[i].waits)
725			ccfreefree(schd_info[i].waits);
726	}
727	if (schedule->stream_1s)
728		ccfreefree(schedule->stream_1s);
729	if (schedule->waits)
730		ccfreefree(schedule->waits);
731	if (schedule->psort)
732		ccfreefree(schedule->psort);
733	if (schedule->begin)
734		ccv_nnc_stream_signal_free(schedule->begin);
735	if (schedule->end)
736		ccv_nnc_stream_signal_free(schedule->end);
737	ccfreefree(schedule);
738}
739 
740static ccv_nnc_graph_static_schedule_t* _ccv_nnc_graph_static_schedule_new(ccv_nnc_graph_t* const graph, const int stream_type, const int device_id, const int max_stream_count, ccv_nnc_stream_context_t* const stream_context, const ccv_nnc_graph_exec_t* const _sources, const int _source_size, const ccv_nnc_graph_exec_t* const _destinations, const int _destination_size)
741{
742	assert(graph->sources && graph->sources->rnum)((void) sizeof ((graph->sources && graph->sources
->rnum) ? 1 : 0), __extension__ ({ if (graph->sources &&
 graph->sources->rnum) ; else __assert_fail ("graph->sources && graph->sources->rnum"
, "ccv_nnc_graph.c", 742, __extension__ __PRETTY_FUNCTION__);
 }));
743	assert(graph->destinations && graph->destinations->rnum)((void) sizeof ((graph->destinations && graph->
destinations->rnum) ? 1 : 0), __extension__ ({ if (graph->
destinations && graph->destinations->rnum) ; else
 __assert_fail ("graph->destinations && graph->destinations->rnum"
, "ccv_nnc_graph.c", 743, __extension__ __PRETTY_FUNCTION__);
 }));
744	assert(graph->topsorted)((void) sizeof ((graph->topsorted) ? 1 : 0), __extension__
 ({ if (graph->topsorted) ; else __assert_fail ("graph->topsorted"
, "ccv_nnc_graph.c", 744, __extension__ __PRETTY_FUNCTION__);
 })); // Only support this on a topsorted graph.
745	const int exec_info_size = graph->exec_info->rnum;
746	assert(exec_info_size > 0)((void) sizeof ((exec_info_size > 0) ? 1 : 0), __extension__
 ({ if (exec_info_size > 0) ; else __assert_fail ("exec_info_size > 0"
, "ccv_nnc_graph.c", 746, __extension__ __PRETTY_FUNCTION__);
 }));
747	const ccv_nnc_graph_exec_t* const sources = _sources == 0 ? (ccv_nnc_graph_exec_t*)ccv_array_get(graph->sources, 0)((void*)(((char*)((graph->sources)->data)) + (size_t)(graph
->sources)->rsize * (size_t)(0))) : _sources;
748	const int source_size = _sources == 0 ? graph->sources->rnum : _source_size;
749	if (!_sources)
750		{ assert(_source_size == 0)((void) sizeof ((_source_size == 0) ? 1 : 0), __extension__ (
{ if (_source_size == 0) ; else __assert_fail ("_source_size == 0"
, "ccv_nnc_graph.c", 750, __extension__ __PRETTY_FUNCTION__);
 })); }
751	const ccv_nnc_graph_exec_t* const destinations = _destinations == 0 ? (ccv_nnc_graph_exec_t*)ccv_array_get(graph->destinations, 0)((void*)(((char*)((graph->destinations)->data)) + (size_t
)(graph->destinations)->rsize * (size_t)(0))) : _destinations;
752	const int destination_size = _destinations == 0 ? graph->destinations->rnum : _destination_size;
753	if (!_destinations)
754		{ assert(_destination_size == 0)((void) sizeof ((_destination_size == 0) ? 1 : 0), __extension__
 ({ if (_destination_size == 0) ; else __assert_fail ("_destination_size == 0"
, "ccv_nnc_graph.c", 754, __extension__ __PRETTY_FUNCTION__);
 })); }
755	const int root_schedule = (_sources == 0 && _destinations == 0);
756	ccv_nnc_graph_static_schedule_t* const schedule = cccalloccalloc(1, sizeof(ccv_nnc_graph_static_schedule_t) + sizeof(ccv_nnc_graph_exec_schedule_t) * (exec_info_size - 1));
757	schedule->exec_info_size = exec_info_size;
758	ccv_nnc_graph_exec_schedule_t* const schd_info = schedule->exec_info;
759	ccv_nnc_graph_exec_info_t* const exec_info = (ccv_nnc_graph_exec_info_t*)ccv_array_get(graph->exec_info, 0)((void*)(((char*)((graph->exec_info)->data)) + (size_t)
(graph->exec_info)->rsize * (size_t)(0)));
760	ccv_nnc_graph_visit_t* visit = ccv_nnc_graph_visit_new(graph, exec_info, exec_info_size, sources, source_size, destinations, destination_size, 0)({ ccv_nnc_graph_visit_t* _visit_ = (ccv_nnc_graph_visit_t*)malloc
(sizeof(ccv_nnc_graph_visit_t) + sizeof(_visit_->node[0]) *
 ((exec_info_size) - 1)); _visit_->size = 0; do { typedef struct
 { int8_t d; int8_t r; uint16_t c; int32_t edges; } ccv_nnc_incoming_t
; int _i_, _j_; int _incoming_edges_ = 0; for (_i_ = 0; _i_ <
 (exec_info_size); _i_++) _incoming_edges_ += ((exec_info)[_i_
].outgoings) ? (exec_info)[_i_].outgoings->rnum : 0; const
 int _heap_mem_ = ((exec_info_size) + _incoming_edges_ > 1024
); ccv_nnc_incoming_t* _incomings_; if (_heap_mem_) _incomings_
 = (ccv_nnc_incoming_t*)malloc(sizeof(ccv_nnc_incoming_t) * (
exec_info_size) + sizeof(int32_t) * ((exec_info_size) * 2 + _incoming_edges_
)); else _incomings_ = (ccv_nnc_incoming_t*)__builtin_alloca (
sizeof(ccv_nnc_incoming_t) * (exec_info_size) + sizeof(int32_t
) * ((exec_info_size) * 2 + _incoming_edges_)); memset(_incomings_
, 0, sizeof(ccv_nnc_incoming_t) * (exec_info_size)); int32_t*
 _exists_[2] = { (int32_t*)(_incomings_ + (exec_info_size)), (
int32_t*)(_incomings_ + (exec_info_size)) + (exec_info_size),
 }; int32_t* const _edges_ = _exists_[1] + (exec_info_size); for
 (_i_ = 0; _i_ < (source_size); _i_++) { ((void) sizeof ((
(sources)[_i_].graph == graph) ? 1 : 0), __extension__ ({ if (
(sources)[_i_].graph == graph) ; else __assert_fail ("(sources)[_i_].graph == graph"
, "ccv_nnc_graph.c", 760, __extension__ __PRETTY_FUNCTION__);
 })); _incomings_[(sources)[_i_].d].r = 1; _exists_[0][_i_] =
 (sources)[_i_].d; } int _exist_size_[2] = { (source_size), 0
, }; int _p_ = 0, _q_ = 1; while (_exist_size_[_p_] > 0) {
 _exist_size_[_q_] = 0; for (_i_ = 0; _i_ < _exist_size_[_p_
]; _i_++) { const int32_t _idx_ = _exists_[_p_][_i_]; if (_incomings_
[_idx_].r != 1) continue; _incomings_[_idx_].r = 2; if ((exec_info
)[_idx_].outgoings) for (_j_ = 0; _j_ < (exec_info)[_idx_]
.outgoings->rnum; _j_++) { const int d = *(int*)((void*)((
(char*)(((exec_info)[_idx_].outgoings)->data)) + (size_t)(
(exec_info)[_idx_].outgoings)->rsize * (size_t)(_j_))); ++
_incomings_[d].c; if (_incomings_[d].r != 0) continue; _incomings_
[d].r = 1; ((void) sizeof ((_exist_size_[_q_] < (exec_info_size
)) ? 1 : 0), __extension__ ({ if (_exist_size_[_q_] < (exec_info_size
)) ; else __assert_fail ("_exist_size_[_q_] < (exec_info_size)"
, "ccv_nnc_graph.c", 760, __extension__ __PRETTY_FUNCTION__);
 })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[_q_
]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for (
_i_ = 0; _i_ < (source_size); _i_++) { ((void) sizeof (((sources
)[_i_].graph == graph) ? 1 : 0), __extension__ ({ if ((sources
)[_i_].graph == graph) ; else __assert_fail ("(sources)[_i_].graph == graph"
, "ccv_nnc_graph.c", 760, __extension__ __PRETTY_FUNCTION__);
 })); _incomings_[(sources)[_i_].d].r = 3; _exists_[0][_i_] =
 (sources)[_i_].d; } _exist_size_[0] = (source_size); _exist_size_
[1] = 0; _p_ = 0, _q_ = 1; int _bump_ = 1; while (_exist_size_
[_p_] > 0) { _exist_size_[_q_] = 0; for (_i_ = 0; _i_ <
 _exist_size_[_p_]; _i_++) { const int32_t _idx_ = _exists_[_p_
][_i_]; if (_incomings_[_idx_].r != 3) continue; _incomings_[
_idx_].r = 4; if ((exec_info)[_idx_].outgoings) for (_j_ = 0;
 _j_ < (exec_info)[_idx_].outgoings->rnum; _j_++) { const
 int d = *(int*)((void*)(((char*)(((exec_info)[_idx_].outgoings
)->data)) + (size_t)((exec_info)[_idx_].outgoings)->rsize
 * (size_t)(_j_))); if (_incomings_[d].edges == 0) { _incomings_
[d].edges = _bump_; _bump_ += _incomings_[d].c; _incomings_[d
].c = 0; } _edges_[_incomings_[d].edges - 1 + _incomings_[d].
c] = _idx_; ++_incomings_[d].c; if (_incomings_[d].r != 2) continue
; _incomings_[d].r = 3; ((void) sizeof ((_exist_size_[_q_] <
 (exec_info_size)) ? 1 : 0), __extension__ ({ if (_exist_size_
[_q_] < (exec_info_size)) ; else __assert_fail ("_exist_size_[_q_] < (exec_info_size)"
, "ccv_nnc_graph.c", 760, __extension__ __PRETTY_FUNCTION__);
 })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[_q_
]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for (
_i_ = 0; _i_ < (destination_size); _i_++) { ((void) sizeof
 (((destinations)[_i_].graph == graph) ? 1 : 0), __extension__
 ({ if ((destinations)[_i_].graph == graph) ; else __assert_fail
 ("(destinations)[_i_].graph == graph", "ccv_nnc_graph.c", 760
, __extension__ __PRETTY_FUNCTION__); })); _incomings_[(destinations
)[_i_].d].r = 5; _exists_[0][_i_] = (destinations)[_i_].d; } _exist_size_
[0] = (destination_size); _exist_size_[1] = 0; _p_ = 0, _q_ =
 1; while (_exist_size_[_p_] > 0) { _exist_size_[_q_] = 0;
 for (_i_ = 0; _i_ < _exist_size_[_p_]; _i_++) { const int32_t
 _idx_ = _exists_[_p_][_i_]; if (_incomings_[_idx_].r != 5) continue
; _incomings_[_idx_].r = 6; if (_incomings_[_idx_].edges >
 0) for (_j_ = 0; _j_ < _incomings_[_idx_].c; _j_++) { const
 int d = _edges_[_incomings_[_idx_].edges - 1 + _j_]; if (_incomings_
[d].r != 4) continue; _incomings_[d].r = 5; ((void) sizeof ((
_exist_size_[_q_] < (exec_info_size)) ? 1 : 0), __extension__
 ({ if (_exist_size_[_q_] < (exec_info_size)) ; else __assert_fail
 ("_exist_size_[_q_] < (exec_info_size)", "ccv_nnc_graph.c"
, 760, __extension__ __PRETTY_FUNCTION__); })); _exists_[_q_]
[_exist_size_[_q_]] = d; ++_exist_size_[_q_]; } } ((_i_) = (_p_
), (_p_) = (_q_), (_q_) = (_i_)); } for (_i_ = 0; _i_ < (destination_size
); _i_++) { ((void) sizeof (((destinations)[_i_].graph == graph
) ? 1 : 0), __extension__ ({ if ((destinations)[_i_].graph ==
 graph) ; else __assert_fail ("(destinations)[_i_].graph == graph"
, "ccv_nnc_graph.c", 760, __extension__ __PRETTY_FUNCTION__);
 })); _incomings_[(destinations)[_i_].d].d = 1; } for (_i_ = 0
; _i_ < (source_size); _i_++) { ((void) sizeof (((sources)
[_i_].graph == graph) ? 1 : 0), __extension__ ({ if ((sources
)[_i_].graph == graph) ; else __assert_fail ("(sources)[_i_].graph == graph"
, "ccv_nnc_graph.c", 760, __extension__ __PRETTY_FUNCTION__);
 })); _exists_[0][_i_] = (sources)[_i_].d; } _p_ = 0; _q_ = 1
; _exist_size_[0] = (source_size); _exist_size_[1] = 0; int _d_
 = 0; while (_exist_size_[_p_] > 0) { _exist_size_[_q_] = 0
; for (_i_ = 0; _i_ < _exist_size_[_p_];) { const int32_t _idx_
 = _exists_[_p_][_i_]; _visit_->node[_visit_->size].index
 = ((_idx_)); _visit_->node[_visit_->size].term = ((_incomings_
[_idx_].d)); ++_visit_->size;; if (_incomings_[_idx_].d) {
 ++_d_; _incomings_[_idx_].r = 7; } if ((exec_info)[_idx_].outgoings
) { if ((exec_info)[_idx_].outgoings->rnum == 1) { const int
 d = *(int*)((void*)(((char*)(((exec_info)[_idx_].outgoings)->
data)) + (size_t)((exec_info)[_idx_].outgoings)->rsize * (
size_t)(0))); --_incomings_[d].c; if (_incomings_[d].c == 0 &&
 _incomings_[d].r == 6 && _d_ < (destination_size)
) { _exists_[_p_][_i_] = d; continue; } } else for (_j_ = 0; _j_
 < (exec_info)[_idx_].outgoings->rnum; _j_++) { const int
 d = *(int*)((void*)(((char*)(((exec_info)[_idx_].outgoings)->
data)) + (size_t)((exec_info)[_idx_].outgoings)->rsize * (
size_t)(_j_))); --_incomings_[d].c; if (_incomings_[d].c == 0
 && _incomings_[d].r == 6 && _d_ < (destination_size
)) { ((void) sizeof ((_exist_size_[_q_] < (exec_info_size)
) ? 1 : 0), __extension__ ({ if (_exist_size_[_q_] < (exec_info_size
)) ; else __assert_fail ("_exist_size_[_q_] < (exec_info_size)"
, "ccv_nnc_graph.c", 760, __extension__ __PRETTY_FUNCTION__);
 })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[_q_
]; } } } ++_i_; } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_
)); } for (_i_ = 0; _i_ < (destination_size); _i_++) { ((void
) sizeof (((destinations)[_i_].graph == graph) ? 1 : 0), __extension__
 ({ if ((destinations)[_i_].graph == graph) ; else __assert_fail
 ("(destinations)[_i_].graph == graph", "ccv_nnc_graph.c", 760
, __extension__ __PRETTY_FUNCTION__); })); if (_incomings_[(destinations
)[_i_].d].r == 7) continue; if (!(0)) { ((void) sizeof ((_incomings_
[(destinations)[_i_].d].c == 0) ? 1 : 0), __extension__ ({ if
 (_incomings_[(destinations)[_i_].d].c == 0) ; else __assert_fail
 ("_incomings_[(destinations)[_i_].d].c == 0", "ccv_nnc_graph.c"
, 760, __extension__ __PRETTY_FUNCTION__); })); } else if (_incomings_
[(destinations)[_i_].d].c > 0) continue; _visit_->node[
_visit_->size].index = (((destinations)[_i_].d)); _visit_->
node[_visit_->size].term = ((_incomings_[(destinations)[_i_
].d].d)); ++_visit_->size;; } if (_heap_mem_) free(_incomings_
); } while (0);; ((void) sizeof ((_visit_->size <= (exec_info_size
)) ? 1 : 0), __extension__ ({ if (_visit_->size <= (exec_info_size
)) ; else __assert_fail ("_visit_->size <= (exec_info_size)"
, "ccv_nnc_graph.c", 760, __extension__ __PRETTY_FUNCTION__);
 })); _visit_; });
761	if (!root_schedule)
762	{
763		// If this is not a root schedule, we need to do partial topsort.
764		int psort_size = 0;
765		ccv_nnc_graph_visit_for(visit, exec_info, node, idx){ int _i_; for (_i_ = 0; _i_ < (visit)->size; _i_++) { const
 int idx __attribute__((unused)) = (visit)->node[_i_].index
; const int _node_unused_ __attribute__((unused)) = (visit)->
node[_i_].term; typeof ((exec_info)) const node __attribute__
((unused)) = (exec_info) + idx; {
766			++psort_size;
767		} ccv_nnc_graph_visit_endfor} }
768		schedule->psort = (int*)ccmallocmalloc(sizeof(int) * psort_size);
769		schedule->psort_size = psort_size;
770		psort_size = 0;
771		ccv_nnc_graph_visit_for(visit, exec_info, node, idx){ int _i_; for (_i_ = 0; _i_ < (visit)->size; _i_++) { const
 int idx __attribute__((unused)) = (visit)->node[_i_].index
; const int _node_unused_ __attribute__((unused)) = (visit)->
node[_i_].term; typeof ((exec_info)) const node __attribute__
((unused)) = (exec_info) + idx; {
772			schedule->psort[psort_size++] = idx;
773		} ccv_nnc_graph_visit_endfor} }
774	}
775	int i, j, k;
776	// Generate exec dependencies (or, in other words, partial ordering of executions).
777	ccv_sparse_matrix_t* exec_dep = ccv_sparse_matrix_new(exec_info_size, exec_info_size, CCV_32S | CCV_C1, CCV_SPARSE_ROW_MAJOR, 0);
778	int* buf = (int*)ccmallocmalloc(sizeof(int) * exec_info_size * 2);
779	int buf_size;
780#define for_block(x, val) \
781	do { \
782		if (((int32_t*)val)[0] > 0) \
783		{ \
784			buf[buf_size * 2] = x; \
785			buf[buf_size * 2 + 1] = ((int32_t*)val)[0] + 1; \
786			++buf_size; \
787		} \
788	} while (0)
789	for (i = 0; i < exec_info_size; i++)
790		schd_info[i].stream_size = -1;
791	ccv_nnc_graph_visit_for(visit, exec_info, node, idx, term){ int _i_; for (_i_ = 0; _i_ < (visit)->size; _i_++) { const
 int idx __attribute__((unused)) = (visit)->node[_i_].index
; const int term __attribute__((unused)) = (visit)->node[_i_
].term; typeof ((exec_info)) const node __attribute__((unused
)) = (exec_info) + idx; {
792		buf_size = 0; /* save all its parent deps to this buffer */
793		ccv_sparse_matrix_vector_t* vector = ccv_get_sparse_matrix_vector(exec_dep, idx);
794		schd_info[idx].stream_size = 0;
795		if (vector)
796			CCV_SPARSE_VECTOR_FOREACH(exec_dep, vector, for_block)do { switch ((((exec_dep)->type) & 0xFF000)) { case CCV_32S
: { do { int _i_; __attribute__((unused)) const size_t _c_ = (
((exec_dep)->type) & 0xFFF); if ((exec_dep)->type &
 CCV_DENSE_VECTOR) { for (_i_ = 0; _i_ < (vector)->size
; _i_++) { for_block((_i_), ((vector)->data.i32 + (_i_ * _c_
))); } } else { const size_t _idx_size_ = sizeof(ccv_sparse_matrix_index_t
) + ((_ccv_get_data_type_size[(((exec_dep)->type) & 0xFF000
) >> 12] * (((exec_dep)->type) & 0xFFF) + 3) &
 -4); uint8_t* const _vidx_ = (uint8_t*)(vector)->index; for
 (_i_ = 0; _i_ < (vector)->size; _i_++) { ccv_sparse_matrix_index_t
* const _idx_i_ = (ccv_sparse_matrix_index_t*)(_vidx_ + _idx_size_
 * _i_); if (_idx_i_->ifbit <= 1) continue; ccv_numeric_data_t
 _d_ = { .u8 = (uint8_t*)(_idx_i_ + 1) }; for_block((_idx_i_->
i), (_d_.i32 + (0))); } } } while (0); break; } case CCV_32F:
 { do { int _i_; __attribute__((unused)) const size_t _c_ = (
((exec_dep)->type) & 0xFFF); if ((exec_dep)->type &
 CCV_DENSE_VECTOR) { for (_i_ = 0; _i_ < (vector)->size
; _i_++) { for_block((_i_), ((vector)->data.f32 + (_i_ * _c_
))); } } else { const size_t _idx_size_ = sizeof(ccv_sparse_matrix_index_t
) + ((_ccv_get_data_type_size[(((exec_dep)->type) & 0xFF000
) >> 12] * (((exec_dep)->type) & 0xFFF) + 3) &
 -4); uint8_t* const _vidx_ = (uint8_t*)(vector)->index; for
 (_i_ = 0; _i_ < (vector)->size; _i_++) { ccv_sparse_matrix_index_t
* const _idx_i_ = (ccv_sparse_matrix_index_t*)(_vidx_ + _idx_size_
 * _i_); if (_idx_i_->ifbit <= 1) continue; ccv_numeric_data_t
 _d_ = { .u8 = (uint8_t*)(_idx_i_ + 1) }; for_block((_idx_i_->
i), (_d_.f32 + (0))); } } } while (0); break; } case CCV_64S:
 { do { int _i_; __attribute__((unused)) const size_t _c_ = (
((exec_dep)->type) & 0xFFF); if ((exec_dep)->type &
 CCV_DENSE_VECTOR) { for (_i_ = 0; _i_ < (vector)->size
; _i_++) { for_block((_i_), ((vector)->data.i64 + (_i_ * _c_
))); } } else { const size_t _idx_size_ = sizeof(ccv_sparse_matrix_index_t
) + ((_ccv_get_data_type_size[(((exec_dep)->type) & 0xFF000
) >> 12] * (((exec_dep)->type) & 0xFFF) + 3) &
 -4); uint8_t* const _vidx_ = (uint8_t*)(vector)->index; for
 (_i_ = 0; _i_ < (vector)->size; _i_++) { ccv_sparse_matrix_index_t
* const _idx_i_ = (ccv_sparse_matrix_index_t*)(_vidx_ + _idx_size_
 * _i_); if (_idx_i_->ifbit <= 1) continue; ccv_numeric_data_t
 _d_ = { .u8 = (uint8_t*)(_idx_i_ + 1) }; for_block((_idx_i_->
i), (_d_.i64 + (0))); } } } while (0); break; } case CCV_64F:
 { do { int _i_; __attribute__((unused)) const size_t _c_ = (
((exec_dep)->type) & 0xFFF); if ((exec_dep)->type &
 CCV_DENSE_VECTOR) { for (_i_ = 0; _i_ < (vector)->size
; _i_++) { for_block((_i_), ((vector)->data.f64 + (_i_ * _c_
))); } } else { const size_t _idx_size_ = sizeof(ccv_sparse_matrix_index_t
) + ((_ccv_get_data_type_size[(((exec_dep)->type) & 0xFF000
) >> 12] * (((exec_dep)->type) & 0xFFF) + 3) &
 -4); uint8_t* const _vidx_ = (uint8_t*)(vector)->index; for
 (_i_ = 0; _i_ < (vector)->size; _i_++) { ccv_sparse_matrix_index_t
* const _idx_i_ = (ccv_sparse_matrix_index_t*)(_vidx_ + _idx_size_
 * _i_); if (_idx_i_->ifbit <= 1) continue; ccv_numeric_data_t
 _d_ = { .u8 = (uint8_t*)(_idx_i_ + 1) }; for_block((_idx_i_->
i), (_d_.f64 + (0))); } } } while (0); break; } default: { do
 { int _i_; __attribute__((unused)) const size_t _c_ = (((exec_dep
)->type) & 0xFFF); if ((exec_dep)->type & CCV_DENSE_VECTOR
) { for (_i_ = 0; _i_ < (vector)->size; _i_++) { for_block
((_i_), ((vector)->data.u8 + (_i_ * _c_))); } } else { const
 size_t _idx_size_ = sizeof(ccv_sparse_matrix_index_t) + ((_ccv_get_data_type_size
[(((exec_dep)->type) & 0xFF000) >> 12] * (((exec_dep
)->type) & 0xFFF) + 3) & -4); uint8_t* const _vidx_
 = (uint8_t*)(vector)->index; for (_i_ = 0; _i_ < (vector
)->size; _i_++) { ccv_sparse_matrix_index_t* const _idx_i_
 = (ccv_sparse_matrix_index_t*)(_vidx_ + _idx_size_ * _i_); if
 (_idx_i_->ifbit <= 1) continue; ccv_numeric_data_t _d_
 = { .u8 = (uint8_t*)(_idx_i_ + 1) }; for_block((_idx_i_->
i), (_d_.u8 + (0))); } } } while (0); } } } while (0);
797		if (!node->outgoings)
798			continue;
799		for (i = 0; i < node->outgoings->rnum; i++)
800		{
801			int outgoing = *(int*)ccv_array_get(node->outgoings, i)((void*)(((char*)((node->outgoings)->data)) + (size_t)(
node->outgoings)->rsize * (size_t)(i)));
802			const int32_t one = 1;
803			ccv_numeric_data_t cell = ccv_get_sparse_matrix_cell(exec_dep, outgoing, idx);
804			/* If not found, set, if the current node is the destination node, no need
805			 * set itself as parent of subsequent nodes because its terminal nature. */
806			if (!term && (!cell.i32 || cell.i32[0] == 0))
807				ccv_set_sparse_matrix_cell(exec_dep, outgoing, idx, &one);
808			for (j = 0; j < buf_size; j++) /* set with all idx's dependencies as well */
809			{
810				ccv_numeric_data_t cell = ccv_get_sparse_matrix_cell(exec_dep, outgoing, buf[j * 2]);
811				/* If not found, set */
812				if (!cell.i32 || cell.i32[0] == 0)
813					ccv_set_sparse_matrix_cell(exec_dep, outgoing, buf[j * 2], &buf[j * 2 + 1]);
814				else {
815					/* Otherwise, set to the longest one */
816					int32_t dep = ccv_max(cell.i32[0], buf[j * 2 + 1])({ typeof (cell.i32[0]) _a = (cell.i32[0]); typeof (buf[j * 2
 + 1]) _b = (buf[j * 2 + 1]); (_a > _b) ? _a : _b; });
817					ccv_set_sparse_matrix_cell(exec_dep, outgoing, buf[j * 2], &dep);
818				}
819			}
820		}
821	} ccv_nnc_graph_visit_endfor} }
822#undef for_block
823	ccfreefree(buf);
824	// Algorithm to allocate signals and streams for this graph.
825	ccv_array_t* const stream_data = ccv_array_new(sizeof(ccv_nnc_stream_data_t), 0, 0);
826	ccv_array_t** const outgoings = cccalloccalloc(exec_info_size, sizeof(ccv_array_t*));
827	ccv_nnc_incoming_t* const incomings = cccalloccalloc(exec_info_size, sizeof(ccv_nnc_incoming_t));
828	int max_device_id_size = 1;
829	// Filter out outgoing nodes that we will be able to access it afterwards anyway.
830	ccv_nnc_graph_visit_for(visit, exec_info, node, idx){ int _i_; for (_i_ = 0; _i_ < (visit)->size; _i_++) { const
 int idx __attribute__((unused)) = (visit)->node[_i_].index
; const int _node_unused_ __attribute__((unused)) = (visit)->
node[_i_].term; typeof ((exec_info)) const node __attribute__
((unused)) = (exec_info) + idx; {
831		max_device_id_size = ccv_max(node->input_size + node->output_size, max_device_id_size)({ typeof (node->input_size + node->output_size) _a = (
node->input_size + node->output_size); typeof (max_device_id_size
) _b = (max_device_id_size); (_a > _b) ? _a : _b; });
832		if (node->outgoings)
833		{
834			outgoings[idx] = ccv_array_new(sizeof(int), 0, 0);
835			for (i = 0; i < node->outgoings->rnum; i++)
836			{
837				const int di = *(int*)ccv_array_get(node->outgoings, i)((void*)(((char*)((node->outgoings)->data)) + (size_t)(
node->outgoings)->rsize * (size_t)(i)));
838				// Skip if we haven't accessed this exec.
839				if (schd_info[di].stream_size < 0)
840					continue;
841				int flag = 0;
842				for (j = 0; !flag && j < node->outgoings->rnum; j++)
843				{
844					if (j != i)
845					{
846						const int dj = *(int*)ccv_array_get(node->outgoings, j)((void*)(((char*)((node->outgoings)->data)) + (size_t)(
node->outgoings)->rsize * (size_t)(j)));
847						ccv_numeric_data_t cell = ccv_get_sparse_matrix_cell(exec_dep, di, dj);
848						flag = (cell.i32 && cell.i32[0]);
849					}
850				}
851				if (!flag)
852				{
853					ccv_array_push(outgoings[idx], &di);
854					if (!incomings[di].outgoings)
855						incomings[di].outgoings = ccv_array_new(sizeof(int), 1, 0);
856					ccv_array_push(incomings[di].outgoings, &idx);
857				}
858			}
859		}
860	} ccv_nnc_graph_visit_endfor} }
861#define visitor(node, idx, _) \
862	if (node->outgoings) \
863		for (i = 0; i < node->outgoings->rnum; i++) \
864		{ \
865			const int d = *(int*)ccv_array_get(node->outgoings, i)((void*)(((char*)((node->outgoings)->data)) + (size_t)(
node->outgoings)->rsize * (size_t)(i))); \
866			node->rank = ccv_max(incomings[d].rank + 1, node->rank)({ typeof (incomings[d].rank + 1) _a = (incomings[d].rank + 1
); typeof (node->rank) _b = (node->rank); (_a > _b) ?
 _a : _b; }); \
867		}
868	CCV_NNC_GRAPH_VISIT(graph, incomings, exec_info_size, destinations, destination_size, sources, source_size, 0, visitor)do { typedef struct { int8_t d; int8_t r; uint16_t c; int32_t
 edges; } ccv_nnc_incoming_t; int _i_, _j_; int _incoming_edges_
 = 0; for (_i_ = 0; _i_ < (exec_info_size); _i_++) _incoming_edges_
 += ((incomings)[_i_].outgoings) ? (incomings)[_i_].outgoings
->rnum : 0; const int _heap_mem_ = ((exec_info_size) + _incoming_edges_
 > 1024); ccv_nnc_incoming_t* _incomings_; if (_heap_mem_)
 _incomings_ = (ccv_nnc_incoming_t*)malloc(sizeof(ccv_nnc_incoming_t
) * (exec_info_size) + sizeof(int32_t) * ((exec_info_size) * 2
 + _incoming_edges_)); else _incomings_ = (ccv_nnc_incoming_t
*)__builtin_alloca (sizeof(ccv_nnc_incoming_t) * (exec_info_size
) + sizeof(int32_t) * ((exec_info_size) * 2 + _incoming_edges_
)); memset(_incomings_, 0, sizeof(ccv_nnc_incoming_t) * (exec_info_size
)); int32_t* _exists_[2] = { (int32_t*)(_incomings_ + (exec_info_size
)), (int32_t*)(_incomings_ + (exec_info_size)) + (exec_info_size
), }; int32_t* const _edges_ = _exists_[1] + (exec_info_size)
; for (_i_ = 0; _i_ < (destination_size); _i_++) { ((void)
 sizeof (((destinations)[_i_].graph == graph) ? 1 : 0), __extension__
 ({ if ((destinations)[_i_].graph == graph) ; else __assert_fail
 ("(destinations)[_i_].graph == graph", "ccv_nnc_graph.c", 868
, __extension__ __PRETTY_FUNCTION__); })); _incomings_[(destinations
)[_i_].d].r = 1; _exists_[0][_i_] = (destinations)[_i_].d; } int
 _exist_size_[2] = { (destination_size), 0, }; int _p_ = 0, _q_
 = 1; while (_exist_size_[_p_] > 0) { _exist_size_[_q_] = 0
; for (_i_ = 0; _i_ < _exist_size_[_p_]; _i_++) { const int32_t
 _idx_ = _exists_[_p_][_i_]; if (_incomings_[_idx_].r != 1) continue
; _incomings_[_idx_].r = 2; if ((incomings)[_idx_].outgoings)
 for (_j_ = 0; _j_ < (incomings)[_idx_].outgoings->rnum
; _j_++) { const int d = *(int*)((void*)(((char*)(((incomings
)[_idx_].outgoings)->data)) + (size_t)((incomings)[_idx_].
outgoings)->rsize * (size_t)(_j_))); ++_incomings_[d].c; if
 (_incomings_[d].r != 0) continue; _incomings_[d].r = 1; ((void
) sizeof ((_exist_size_[_q_] < (exec_info_size)) ? 1 : 0),
 __extension__ ({ if (_exist_size_[_q_] < (exec_info_size)
) ; else __assert_fail ("_exist_size_[_q_] < (exec_info_size)"
, "ccv_nnc_graph.c", 868, __extension__ __PRETTY_FUNCTION__);
 })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[_q_
]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for (
_i_ = 0; _i_ < (destination_size); _i_++) { ((void) sizeof
 (((destinations)[_i_].graph == graph) ? 1 : 0), __extension__
 ({ if ((destinations)[_i_].graph == graph) ; else __assert_fail
 ("(destinations)[_i_].graph == graph", "ccv_nnc_graph.c", 868
, __extension__ __PRETTY_FUNCTION__); })); _incomings_[(destinations
)[_i_].d].r = 3; _exists_[0][_i_] = (destinations)[_i_].d; } _exist_size_
[0] = (destination_size); _exist_size_[1] = 0; _p_ = 0, _q_ =
 1; int _bump_ = 1; while (_exist_size_[_p_] > 0) { _exist_size_
[_q_] = 0; for (_i_ = 0; _i_ < _exist_size_[_p_]; _i_++) {
 const int32_t _idx_ = _exists_[_p_][_i_]; if (_incomings_[_idx_
].r != 3) continue; _incomings_[_idx_].r = 4; if ((incomings)
[_idx_].outgoings) for (_j_ = 0; _j_ < (incomings)[_idx_].
outgoings->rnum; _j_++) { const int d = *(int*)((void*)(((
char*)(((incomings)[_idx_].outgoings)->data)) + (size_t)((
incomings)[_idx_].outgoings)->rsize * (size_t)(_j_))); if (
_incomings_[d].edges == 0) { _incomings_[d].edges = _bump_; _bump_
 += _incomings_[d].c; _incomings_[d].c = 0; } _edges_[_incomings_
[d].edges - 1 + _incomings_[d].c] = _idx_; ++_incomings_[d].c
; if (_incomings_[d].r != 2) continue; _incomings_[d].r = 3; (
(void) sizeof ((_exist_size_[_q_] < (exec_info_size)) ? 1 :
 0), __extension__ ({ if (_exist_size_[_q_] < (exec_info_size
)) ; else __assert_fail ("_exist_size_[_q_] < (exec_info_size)"
, "ccv_nnc_graph.c", 868, __extension__ __PRETTY_FUNCTION__);
 })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[_q_
]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for (
_i_ = 0; _i_ < (source_size); _i_++) { ((void) sizeof (((sources
)[_i_].graph == graph) ? 1 : 0), __extension__ ({ if ((sources
)[_i_].graph == graph) ; else __assert_fail ("(sources)[_i_].graph == graph"
, "ccv_nnc_graph.c", 868, __extension__ __PRETTY_FUNCTION__);
 })); _incomings_[(sources)[_i_].d].r = 5; _exists_[0][_i_] =
 (sources)[_i_].d; } _exist_size_[0] = (source_size); _exist_size_
[1] = 0; _p_ = 0, _q_ = 1; while (_exist_size_[_p_] > 0) {
 _exist_size_[_q_] = 0; for (_i_ = 0; _i_ < _exist_size_[_p_
]; _i_++) { const int32_t _idx_ = _exists_[_p_][_i_]; if (_incomings_
[_idx_].r != 5) continue; _incomings_[_idx_].r = 6; if (_incomings_
[_idx_].edges > 0) for (_j_ = 0; _j_ < _incomings_[_idx_
].c; _j_++) { const int d = _edges_[_incomings_[_idx_].edges -
 1 + _j_]; if (_incomings_[d].r != 4) continue; _incomings_[d
].r = 5; ((void) sizeof ((_exist_size_[_q_] < (exec_info_size
)) ? 1 : 0), __extension__ ({ if (_exist_size_[_q_] < (exec_info_size
)) ; else __assert_fail ("_exist_size_[_q_] < (exec_info_size)"
, "ccv_nnc_graph.c", 868, __extension__ __PRETTY_FUNCTION__);
 })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[_q_
]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for (
_i_ = 0; _i_ < (source_size); _i_++) { ((void) sizeof (((sources
)[_i_].graph == graph) ? 1 : 0), __extension__ ({ if ((sources
)[_i_].graph == graph) ; else __assert_fail ("(sources)[_i_].graph == graph"
, "ccv_nnc_graph.c", 868, __extension__ __PRETTY_FUNCTION__);
 })); _incomings_[(sources)[_i_].d].d = 1; } for (_i_ = 0; _i_
 < (destination_size); _i_++) { ((void) sizeof (((destinations
)[_i_].graph == graph) ? 1 : 0), __extension__ ({ if ((destinations
)[_i_].graph == graph) ; else __assert_fail ("(destinations)[_i_].graph == graph"
, "ccv_nnc_graph.c", 868, __extension__ __PRETTY_FUNCTION__);
 })); _exists_[0][_i_] = (destinations)[_i_].d; } _p_ = 0; _q_
 = 1; _exist_size_[0] = (destination_size); _exist_size_[1] =
 0; int _d_ = 0; while (_exist_size_[_p_] > 0) { _exist_size_
[_q_] = 0; for (_i_ = 0; _i_ < _exist_size_[_p_];) { const
 int32_t _idx_ = _exists_[_p_][_i_]; visitor(((incomings) + _idx_
), (_idx_), (_incomings_[_idx_].d)); if (_incomings_[_idx_].d
) { ++_d_; _incomings_[_idx_].r = 7; } if ((incomings)[_idx_]
.outgoings) { if ((incomings)[_idx_].outgoings->rnum == 1)
 { const int d = *(int*)((void*)(((char*)(((incomings)[_idx_]
.outgoings)->data)) + (size_t)((incomings)[_idx_].outgoings
)->rsize * (size_t)(0))); --_incomings_[d].c; if (_incomings_
[d].c == 0 && _incomings_[d].r == 6 && _d_ <
 (source_size)) { _exists_[_p_][_i_] = d; continue; } } else for
 (_j_ = 0; _j_ < (incomings)[_idx_].outgoings->rnum; _j_
++) { const int d = *(int*)((void*)(((char*)(((incomings)[_idx_
].outgoings)->data)) + (size_t)((incomings)[_idx_].outgoings
)->rsize * (size_t)(_j_))); --_incomings_[d].c; if (_incomings_
[d].c == 0 && _incomings_[d].r == 6 && _d_ <
 (source_size)) { ((void) sizeof ((_exist_size_[_q_] < (exec_info_size
)) ? 1 : 0), __extension__ ({ if (_exist_size_[_q_] < (exec_info_size
)) ; else __assert_fail ("_exist_size_[_q_] < (exec_info_size)"
, "ccv_nnc_graph.c", 868, __extension__ __PRETTY_FUNCTION__);
 })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[_q_
]; } } } ++_i_; } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_
)); } for (_i_ = 0; _i_ < (source_size); _i_++) { ((void) sizeof
 (((sources)[_i_].graph == graph) ? 1 : 0), __extension__ ({ if
 ((sources)[_i_].graph == graph) ; else __assert_fail ("(sources)[_i_].graph == graph"
, "ccv_nnc_graph.c", 868, __extension__ __PRETTY_FUNCTION__);
 })); if (_incomings_[(sources)[_i_].d].r == 7) continue; if (
!(0)) { ((void) sizeof ((_incomings_[(sources)[_i_].d].c == 0
) ? 1 : 0), __extension__ ({ if (_incomings_[(sources)[_i_].d
].c == 0) ; else __assert_fail ("_incomings_[(sources)[_i_].d].c == 0"
, "ccv_nnc_graph.c", 868, __extension__ __PRETTY_FUNCTION__);
 })); } else if (_incomings_[(sources)[_i_].d].c > 0) continue
; visitor(((incomings) + (sources)[_i_].d), ((sources)[_i_].d
), (_incomings_[(sources)[_i_].d].d)); } if (_heap_mem_) free
(_incomings_); } while (0);;
869#undef visitor
870	int device_ids[max_device_id_size];
871	int outgoing_device_ids[max_device_id_size];
872	int signal_size = 0;
873	ccv_nnc_graph_visit_for(visit, exec_info, node, idx){ int _i_; for (_i_ = 0; _i_ < (visit)->size; _i_++) { const
 int idx __attribute__((unused)) = (visit)->node[_i_].index
; const int _node_unused_ __attribute__((unused)) = (visit)->
node[_i_].term; typeof ((exec_info)) const node __attribute__
((unused)) = (exec_info) + idx; {
874		// Go through the incomings.
875		const int device_id_size = _ccv_nnc_device_ids_for_stream_data(node, device_id, stream_data, device_ids, max_device_id_size);
876		if (schd_info[idx].stream_size == 0)
877		{
878			schd_info[idx].stream_size = device_id_size; // At least at the same size as the device_id_size.
879			if (device_id_size > 1)
880			{
881				schd_info[idx]._heap_streams = (int*)ccmallocmalloc(sizeof(int) * device_id_size * 2);
882				schd_info[idx]._heap_signals = (schd_info[idx]._heap_streams + device_id_size);
883			}
884			for (i = 0; i < device_id_size; i++)
885				SCHEDULE_STREAMS(schd_info[idx])((schd_info[idx]).stream_size <= 1 ? (schd_info[idx])._inline_streams
 : (schd_info[idx])._heap_streams)[i] = -1, SCHEDULE_SIGNALS(schd_info[idx])((schd_info[idx]).stream_size <= 1 ? (schd_info[idx])._inline_signals
 : (schd_info[idx])._heap_signals)[i] = -1;
886		}
887		for (i = 0; i < device_id_size; i++)
888			// Go through until the end to assign streams.
889			if (SCHEDULE_STREAMS(schd_info[idx])((schd_info[idx]).stream_size <= 1 ? (schd_info[idx])._inline_streams
 : (schd_info[idx])._heap_streams)[i] < 0)
890			{
891				int stream_idx = -1;
892				int stream_has_command = 0;
893				// First, find a good stream in stream data (the stream is good if it can be recycled, and it has the same command).
894				// Otherwise, we prefer a usable stream (it doesn't have the command, but it can be recycled).
895				for (j = 0; (stream_idx < 0 || !stream_has_command) && j < stream_data->rnum; j++)
896				{
897					ccv_nnc_stream_data_t* const data = (ccv_nnc_stream_data_t*)ccv_array_get(stream_data, j)((void*)(((char*)((stream_data)->data)) + (size_t)(stream_data
)->rsize * (size_t)(j)));
898					if (data->device_id == device_ids[i])
899					{
900						const ccv_numeric_data_t cell = ccv_get_sparse_matrix_cell(exec_dep, idx, data->exec_idx);
901						// If there is a path to conclude that exec_idx is before idx, then we can reuse
902						// this stream. Otherwise the work in this "empty stream" could still be ongoing,
903						// and we may delay the following work unnecessarily.
904						if (cell.i32 && cell.i32[0] > 0)
905						{
906							if (ccv_array_find_uint(data->command_set, node->cmd.cmd))
907								stream_idx = j, stream_has_command = 1;
908							else if (stream_idx < 0) // Otherwise, only assign the stream idx if it is not assigned yet.
909								stream_idx = j;
910						}
911					}
912				}
913				if (stream_idx < 0)
914				{
915					// Note that the max stream count is a "soft" limit. Even we have different devices, our compute allocation has to be on different streams.
916					if (stream_data->rnum >= max_stream_count && max_stream_count > 0)
917					{
918						// If we are already at out limit, go through again to see if a stream is available, if the stream has command, and also its exec_idx is not preceding this execution.
919						for (j = 0; (stream_idx < 0 || !stream_has_command) && j < stream_data->rnum; j++)
920						{
921							ccv_nnc_stream_data_t* const data = (ccv_nnc_stream_data_t*)ccv_array_get(stream_data, j)((void*)(((char*)((stream_data)->data)) + (size_t)(stream_data
)->rsize * (size_t)(j)));
922							if (data->device_id == device_ids[i])
923							{
924								const ccv_numeric_data_t cell = ccv_get_sparse_matrix_cell(exec_dep, data->exec_idx, idx);
925								// There must be no path from idx to exec_idx otherwise we already have stream_idx. Now we just to verify
926								// there is no path from exec_idx to idx as well.
927								if (!cell.i32 || cell.i32[0] == 0)
928								{
929									if (ccv_array_find_uint(data->command_set, node->cmd.cmd))
930										stream_idx = j, stream_has_command = 1;
931									else if (stream_idx < 0) // Otherwise, only assign the stream idx if it is not assigned yet.
932										stream_idx = j;
933								}
934							}
935						}
936						if (stream_idx >= 0)
937						{
938							// Now need to mark exec_idx is after idx, so we can avoid A -> B -> A deadlock.
939							ccv_nnc_stream_data_t* const data = (ccv_nnc_stream_data_t*)ccv_array_get(stream_data, stream_idx)((void*)(((char*)((stream_data)->data)) + (size_t)(stream_data
)->rsize * (size_t)(stream_idx)));
940							const int32_t one = 1;
941							ccv_set_sparse_matrix_cell(exec_dep, idx, data->exec_idx, &one);
942						}
943					}
944					if (stream_idx < 0)
945					{
946						stream_idx = stream_data->rnum;
947						const ccv_nnc_stream_data_t data = {
948							.device_id = device_ids[i],
949						};
950						ccv_array_push(stream_data, &data);
951					}
952				}
953				assert(stream_idx >= 0)((void) sizeof ((stream_idx >= 0) ? 1 : 0), __extension__ (
{ if (stream_idx >= 0) ; else __assert_fail ("stream_idx >= 0"
, "ccv_nnc_graph.c", 953, __extension__ __PRETTY_FUNCTION__);
 }));
954				ccv_nnc_stream_data_t* const data = (ccv_nnc_stream_data_t*)ccv_array_get(stream_data, stream_idx)((void*)(((char*)((stream_data)->data)) + (size_t)(stream_data
)->rsize * (size_t)(stream_idx)));
955				if (!data->command_set)
956					data->command_set = ccv_array_new(sizeof(uint32_t), 1, 0);
957				SCHEDULE_STREAMS(schd_info[idx])((schd_info[idx]).stream_size <= 1 ? (schd_info[idx])._inline_streams
 : (schd_info[idx])._heap_streams)[i] = stream_idx;
958				ccv_array_add_unique_uint(data->command_set, node->cmd.cmd);
959				// Assign all subsequent node to use this stream.
960				int outgoing_idx = idx;
961				// if we want to enforce the stream count is only 1, we certainly don't want to the greedy approach.
962				// With the greedy approach, the current stream will go all the way down and certainly conflict with
963				// other streams. We'd prefer to interleaving the execution instead in this case.
964				if (max_stream_count != 1)
965					while (outgoings[outgoing_idx] && outgoings[outgoing_idx]->rnum)
966					{
967						int highest_rank = -1;
968						int highest_idx = -1;
969						int stream_n = -1;
970						int stream_has_command = 0;
971						for (j = 0; j < outgoings[outgoing_idx]->rnum; j++)
972						{
973							const int d = *(int*)ccv_array_get(outgoings[outgoing_idx], j)((void*)(((char*)((outgoings[outgoing_idx])->data)) + (size_t
)(outgoings[outgoing_idx])->rsize * (size_t)(j)));
974							// This is not outside of our scope at this point.
975							assert(schd_info[d].stream_size >= 0)((void) sizeof ((schd_info[d].stream_size >= 0) ? 1 : 0), __extension__
 ({ if (schd_info[d].stream_size >= 0) ; else __assert_fail
 ("schd_info[d].stream_size >= 0", "ccv_nnc_graph.c", 975,
 __extension__ __PRETTY_FUNCTION__); }));
976							ccv_nnc_graph_exec_info_t* const outgoing_node = exec_info + d;
977							const int outgoing_device_id_size = _ccv_nnc_device_ids_for_stream_data(outgoing_node, device_id, stream_data, outgoing_device_ids, max_device_id_size);
978							if (schd_info[d].stream_size == 0)
979							{
980								schd_info[d].stream_size = outgoing_device_id_size; // At least at the same size as the device_id_size.
981								if (outgoing_device_id_size > 1)
982								{
983									schd_info[d]._heap_streams = (int*)ccmallocmalloc(sizeof(int) * outgoing_device_id_size * 2);
984									schd_info[d]._heap_signals = (schd_info[d]._heap_streams + outgoing_device_id_size);
985								}
986								for (k = 0; k < outgoing_device_id_size; k++)
987									SCHEDULE_STREAMS(schd_info[d])((schd_info[d]).stream_size <= 1 ? (schd_info[d])._inline_streams
 : (schd_info[d])._heap_streams)[k] = -1, SCHEDULE_SIGNALS(schd_info[d])((schd_info[d]).stream_size <= 1 ? (schd_info[d])._inline_signals
 : (schd_info[d])._heap_signals)[k] = -1;
988							}
989							assert(schd_info[d].stream_size == outgoing_device_id_size)((void) sizeof ((schd_info[d].stream_size == outgoing_device_id_size
) ? 1 : 0), __extension__ ({ if (schd_info[d].stream_size == outgoing_device_id_size
) ; else __assert_fail ("schd_info[d].stream_size == outgoing_device_id_size"
, "ccv_nnc_graph.c", 989, __extension__ __PRETTY_FUNCTION__);
 }));
990							for (k = 0; k < outgoing_device_id_size; k++)
991								// If it should be on the same device and the stream is not assign, potentially.
992								if (outgoing_device_ids[k] == device_ids[i] &&
993									SCHEDULE_STREAMS(schd_info[d])((schd_info[d]).stream_size <= 1 ? (schd_info[d])._inline_streams
 : (schd_info[d])._heap_streams)[k] < 0 &&
994									(incomings[d].rank > highest_rank ||
995									 (incomings[d].rank == highest_rank &&
996									  !stream_has_command && ccv_array_find_uint(data->command_set, outgoing_node->cmd.cmd))))
997								{
998									highest_rank = incomings[d].rank;
999									highest_idx = d;
1000									stream_n = k;
1001									// This is 1 if rank is the same (thus, I must break the tie already), if the rank is not the same, we need to compute this.
1002									stream_has_command = (incomings[d].rank == highest_rank || ccv_array_find_uint(data->command_set, outgoing_node->cmd.cmd));
1003								}
1004						}
1005						if (highest_idx >= 0)
1006						{
1007							outgoing_idx = highest_idx;
1008							ccv_nnc_graph_exec_info_t* const outgoing_node = exec_info + outgoing_idx;
1009							assert(stream_n >= 0)((void) sizeof ((stream_n >= 0) ? 1 : 0), __extension__ ({
 if (stream_n >= 0) ; else __assert_fail ("stream_n >= 0"
, "ccv_nnc_graph.c", 1009, __extension__ __PRETTY_FUNCTION__)
; }));
1010							SCHEDULE_STREAMS(schd_info[outgoing_idx])((schd_info[outgoing_idx]).stream_size <= 1 ? (schd_info[outgoing_idx
])._inline_streams : (schd_info[outgoing_idx])._heap_streams)[stream_n] = stream_idx;
1011							ccv_array_add_unique_uint(data->command_set, outgoing_node->cmd.cmd);
1012						} else
1013							break;
1014					}
1015				data->exec_idx = outgoing_idx;
1016			}
1017	} ccv_nnc_graph_visit_endfor} }
1018	// Go through to assign signals when necessary.
1019	ccv_nnc_graph_visit_for(visit, exec_info, node, idx){ int _i_; for (_i_ = 0; _i_ < (visit)->size; _i_++) { const
 int idx __attribute__((unused)) = (visit)->node[_i_].index
; const int _node_unused_ __attribute__((unused)) = (visit)->
node[_i_].term; typeof ((exec_info)) const node __attribute__
((unused)) = (exec_info) + idx; {
1020		if (incomings[idx].outgoings && incomings[idx].outgoings->rnum)
1021			_ccv_nnc_graph_schedule_assign_signals(incomings[idx].outgoings, schd_info + idx, stream_data, &signal_size, schd_info, exec_info_size);
1022	} ccv_nnc_graph_visit_endfor} }
1023	for (i = 0; i < exec_info_size; i++)
1024		if (outgoings[i])
1025			ccv_array_free(outgoings[i]);
1026	ccfreefree(outgoings);
1027	ccv_matrix_free(exec_dep);
1028	ccv_nnc_stream_data_t* const default_data = (ccv_nnc_stream_data_t*)ccv_array_get(stream_data, 0)((void*)(((char*)((stream_data)->data)) + (size_t)(stream_data
)->rsize * (size_t)(0)));
1029	if (device_id >= 0)
1030	{
1031		// If the default stream (stream 0) is not the same as desired stream, swap with the one that is.
1032		if (default_data->device_id != device_id)
1033		{
1034			int exchange_stream_idx = -1;
1035			// Find the stream idx to exchange.
1036			ccv_nnc_graph_visit_for(visit, exec_info, node, idx){ int _i_; for (_i_ = 0; _i_ < (visit)->size; _i_++) { const
 int idx __attribute__((unused)) = (visit)->node[_i_].index
; const int _node_unused_ __attribute__((unused)) = (visit)->
node[_i_].term; typeof ((exec_info)) const node __attribute__
((unused)) = (exec_info) + idx; {
1037				int flag = 0;
1038				for(i = 0; !flag && i < schd_info[idx].stream_size; i++)
1039				{
1040					const int stream_idx = SCHEDULE_STREAMS(schd_info[idx])((schd_info[idx]).stream_size <= 1 ? (schd_info[idx])._inline_streams
 : (schd_info[idx])._heap_streams)[i];
1041					ccv_nnc_stream_data_t* const data = (ccv_nnc_stream_data_t*)ccv_array_get(stream_data, stream_idx)((void*)(((char*)((stream_data)->data)) + (size_t)(stream_data
)->rsize * (size_t)(stream_idx)));
1042					if (data->device_id == device_id)
1043					{
1044						exchange_stream_idx = stream_idx;
1045						flag = 1;
1046					}
1047				}
1048				if (flag)
1049					break;
1050			} ccv_nnc_graph_visit_endfor} }
1051			assert(exchange_stream_idx >= 0)((void) sizeof ((exchange_stream_idx >= 0) ? 1 : 0), __extension__
 ({ if (exchange_stream_idx >= 0) ; else __assert_fail ("exchange_stream_idx >= 0"
, "ccv_nnc_graph.c", 1051, __extension__ __PRETTY_FUNCTION__)
; }));
1052			ccv_nnc_graph_visit_for(visit, exec_info, node, idx){ int _i_; for (_i_ = 0; _i_ < (visit)->size; _i_++) { const
 int idx __attribute__((unused)) = (visit)->node[_i_].index
; const int _node_unused_ __attribute__((unused)) = (visit)->
node[_i_].term; typeof ((exec_info)) const node __attribute__
((unused)) = (exec_info) + idx; {
1053				for (i = 0; i < schd_info[idx].stream_size; i++)
1054					if (SCHEDULE_STREAMS(schd_info[idx])((schd_info[idx]).stream_size <= 1 ? (schd_info[idx])._inline_streams
 : (schd_info[idx])._heap_streams)[i] == 0)
1055						SCHEDULE_STREAMS(schd_info[idx])((schd_info[idx]).stream_size <= 1 ? (schd_info[idx])._inline_streams
 : (schd_info[idx])._heap_streams)[i] = -1;
1056			} ccv_nnc_graph_visit_endfor} }
1057			ccv_nnc_graph_visit_for(visit, exec_info, node, idx){ int _i_; for (_i_ = 0; _i_ < (visit)->size; _i_++) { const
 int idx __attribute__((unused)) = (visit)->node[_i_].index
; const int _node_unused_ __attribute__((unused)) = (visit)->
node[_i_].term; typeof ((exec_info)) const node __attribute__
((unused)) = (exec_info) + idx; {
1058				for (i = 0; i < schd_info[idx].stream_size; i++)
1059					if (SCHEDULE_STREAMS(schd_info[idx])((schd_info[idx]).stream_size <= 1 ? (schd_info[idx])._inline_streams
 : (schd_info[idx])._heap_streams)[i] == exchange_stream_idx)
1060						SCHEDULE_STREAMS(schd_info[idx])((schd_info[idx]).stream_size <= 1 ? (schd_info[idx])._inline_streams
 : (schd_info[idx])._heap_streams)[i] = 0;
1061			} ccv_nnc_graph_visit_endfor} }
1062			ccv_nnc_graph_visit_for(visit, exec_info, node, idx){ int _i_; for (_i_ = 0; _i_ < (visit)->size; _i_++) { const
 int idx __attribute__((unused)) = (visit)->node[_i_].index
; const int _node_unused_ __attribute__((unused)) = (visit)->
node[_i_].term; typeof ((exec_info)) const node __attribute__
((unused)) = (exec_info) + idx; {
1063				for (i = 0; i < schd_info[idx].stream_size; i++)
1064					if (SCHEDULE_STREAMS(schd_info[idx])((schd_info[idx]).stream_size <= 1 ? (schd_info[idx])._inline_streams
 : (schd_info[idx])._heap_streams)[i] == -1)
1065						SCHEDULE_STREAMS(schd_info[idx])((schd_info[idx]).stream_size <= 1 ? (schd_info[idx])._inline_streams
 : (schd_info[idx])._heap_streams)[i] = exchange_stream_idx;
1066			} ccv_nnc_graph_visit_endfor} }
1067			((ccv_nnc_stream_data_t*)ccv_array_get(stream_data, exchange_stream_idx)((void*)(((char*)((stream_data)->data)) + (size_t)(stream_data
)->rsize * (size_t)(exchange_stream_idx))))->device_id = default_data->device_id;
1068			default_data->device_id = device_id;
1069		}
1070	}
1071	int graph_stream_1_size = 0;
1072	for (i = 0; i < source_size; i++)
1073	{
1074		const int idx = sources[i].d;
1075		// If it has incoming nodes, check whether these are on stream 0.
1076		if (incomings[idx].outgoings && incomings[idx].outgoings->rnum)
1077		{
1078			int flag  = 0;
1079			const ccv_array_t* const incoming = incomings[idx].outgoings;
1080			for (j = 0; !flag && j < incoming->rnum; j++)
1081			{
1082				const int incoming_idx = *(int*)ccv_array_get(incoming, j)((void*)(((char*)((incoming)->data)) + (size_t)(incoming)->
rsize * (size_t)(j)));
1083				for (k = 0; !flag && k < schd_info[incoming_idx].stream_size; k++)
1084					flag = (SCHEDULE_STREAMS(schd_info[incoming_idx])((schd_info[incoming_idx]).stream_size <= 1 ? (schd_info[incoming_idx
])._inline_streams : (schd_info[incoming_idx])._heap_streams)[k] == 0); // If this is the default stream, we already have a good start.
1085			}
1086			if (flag)
1087				continue;
1088		}
1089		for (j = 0; j < schd_info[idx].stream_size; j++)
1090			if (SCHEDULE_STREAMS(schd_info[idx])((schd_info[idx]).stream_size <= 1 ? (schd_info[idx])._inline_streams
 : (schd_info[idx])._heap_streams)[j] != 0) // If this is not the default stream, we need explicit begin signal to start.
1091				++graph_stream_1_size;
1092	}
1093	if (graph_stream_1_size > 0)
1094	{
1095		schedule->stream_1s = ccmallocmalloc(sizeof(int) * graph_stream_1_size);
1096		graph_stream_1_size = 0;
1097		for (i = 0; i < source_size; i++)
1098		{
1099			const int idx = sources[i].d;
1100			// If it has incoming nodes, check whether these are on stream 0.
1101			if (incomings[idx].outgoings && incomings[idx].outgoings->rnum)
1102			{
1103				int flag  = 0;
1104				const ccv_array_t* const incoming = incomings[idx].outgoings;
1105				for (j = 0; !flag && j < incoming->rnum; j++)
1106				{
1107					const int incoming_idx = *(int*)ccv_array_get(incoming, j)((void*)(((char*)((incoming)->data)) + (size_t)(incoming)->
rsize * (size_t)(j)));
1108					for (k = 0; !flag && k < schd_info[incoming_idx].stream_size; k++)
1109						flag = (SCHEDULE_STREAMS(schd_info[incoming_idx])((schd_info[incoming_idx]).stream_size <= 1 ? (schd_info[incoming_idx
])._inline_streams : (schd_info[incoming_idx])._heap_streams)[k] == 0); // If this is the default stream, we already have a good start.
1110				}
1111				if (flag)
1112					continue;
1113			}
1114			for (j = 0; j < schd_info[idx].stream_size; j++)
1115				if (SCHEDULE_STREAMS(schd_info[idx])((schd_info[idx]).stream_size <= 1 ? (schd_info[idx])._inline_streams
 : (schd_info[idx])._heap_streams)[j] != 0) // If this is not the default stream, we need explicit begin signal to start.
1116				{
1117					const int stream_idx = SCHEDULE_STREAMS(schd_info[idx])((schd_info[idx]).stream_size <= 1 ? (schd_info[idx])._inline_streams
 : (schd_info[idx])._heap_streams)[j];
1118					int flag = 0;
1119					for (k = 0; !flag && k < graph_stream_1_size; k++)
1120						flag = (stream_idx == schedule->stream_1s[k]);
1121					if (!flag)
1122						schedule->stream_1s[graph_stream_1_size++] = stream_idx;
1123				}
1124		}
1125		schedule->stream_1_size = graph_stream_1_size;
1126	}
1127	for (i = 0; i < exec_info_size; i++)
1128		if (incomings[i].outgoings)
1129			ccv_array_free(incomings[i].outgoings);
1130	ccfreefree(incomings);
1131	int graph_wait_size = 0;
1132	for (i = 0; i < destination_size; i++)
1133	{
1134		const int idx = destinations[i].d;
1135		for (j = 0; j < schd_info[idx].stream_size; j++)
1136			if (SCHEDULE_STREAMS(schd_info[idx])((schd_info[idx]).stream_size <= 1 ? (schd_info[idx])._inline_streams
 : (schd_info[idx])._heap_streams)[j] != 0) // If this exec_info doesn't end with default stream, we need to wait.
1137				++graph_wait_size;
1138	}
1139	if (graph_wait_size > 0)
1140	{
1141		schedule->waits = ccmallocmalloc(sizeof(int) * graph_wait_size);
1142		graph_wait_size = 0;
1143		for (i = 0; i < destination_size; i++)
1144		{
1145			const int idx = destinations[i].d;
1146			for (j = 0; j < schd_info[idx].stream_size; j++)
1147				if (SCHEDULE_STREAMS(schd_info[idx])((schd_info[idx]).stream_size <= 1 ? (schd_info[idx])._inline_streams
 : (schd_info[idx])._heap_streams)[j] != 0) // If this exec_info doesn't end with default stream, we need to wait.
1148				{
1149					ccv_nnc_stream_data_t* const default_stream_data = (ccv_nnc_stream_data_t*)ccv_array_get(stream_data, 0)((void*)(((char*)((stream_data)->data)) + (size_t)(stream_data
)->rsize * (size_t)(0)));
1150					if (SCHEDULE_SIGNALS(schd_info[idx])((schd_info[idx]).stream_size <= 1 ? (schd_info[idx])._inline_signals
 : (schd_info[idx])._heap_signals)[j] < 0)
1151						SCHEDULE_SIGNALS(schd_info[idx])((schd_info[idx]).stream_size <= 1 ? (schd_info[idx])._inline_signals
 : (schd_info[idx])._heap_signals)[j] = signal_size++;
1152					else if (default_stream_data->signal_set && ccv_array_find_int(default_stream_data->signal_set, SCHEDULE_SIGNALS(schd_info[idx])((schd_info[idx]).stream_size <= 1 ? (schd_info[idx])._inline_signals
 : (schd_info[idx])._heap_signals)[j]))
1153						continue;
1154					schedule->waits[graph_wait_size++] = SCHEDULE_SIGNALS(schd_info[idx])((schd_info[idx]).stream_size <= 1 ? (schd_info[idx])._inline_signals
 : (schd_info[idx])._heap_signals)[j];
1155				}
1156		}
1157		schedule->wait_size = graph_wait_size;
1158	}
1159	for (i = 0; i < stream_data->rnum; i++)
1160	{
1161		ccv_nnc_stream_data_t* const data = (ccv_nnc_stream_data_t*)ccv_array_get(stream_data, i)((void*)(((char*)((stream_data)->data)) + (size_t)(stream_data
)->rsize * (size_t)(i)));
1162		if (data->signal_set)
1163			ccv_array_free(data->signal_set);
1164		assert(data->command_set)((void) sizeof ((data->command_set) ? 1 : 0), __extension__
 ({ if (data->command_set) ; else __assert_fail ("data->command_set"
, "ccv_nnc_graph.c", 1164, __extension__ __PRETTY_FUNCTION__)
; }));
1165		ccv_array_free(data->command_set);
1166	}
1167	// Allocate streams & signals
1168	int default_stream_type = stream_type;
1169	CCV_STREAM_SET_DEVICE_ID(default_stream_type, default_data->device_id)(default_stream_type) = (((default_stream_type) & ~0xfff00
) | (((default_data->device_id) & 0xfff) << 8));
1170	if (root_schedule)
1171	{
1172		assert(!graph->streams)((void) sizeof ((!graph->streams) ? 1 : 0), __extension__ (
{ if (!graph->streams) ; else __assert_fail ("!graph->streams"
, "ccv_nnc_graph.c", 1172, __extension__ __PRETTY_FUNCTION__)
; }));
1173		graph->stream_size = stream_data->rnum;
1174		graph->streams = (ccv_nnc_stream_context_t**)ccmallocmalloc(sizeof(ccv_nnc_stream_context_t*) * graph->stream_size);
1175		graph->block_stream_tasks = (co_routine_t**)cccalloccalloc(graph->stream_size, sizeof(co_routine_t*));
1176		if (stream_context)
1177			graph->streams[0] = stream_context;
1178		for (i = (stream_context ? 1 : 0); i < stream_data->rnum; i++)
1179		{
1180			ccv_nnc_stream_data_t* const data = (ccv_nnc_stream_data_t*)ccv_array_get(stream_data, i)((void*)(((char*)((stream_data)->data)) + (size_t)(stream_data
)->rsize * (size_t)(i)));
1181			int type = stream_type;
1182			CCV_STREAM_SET_DEVICE_ID(type, data->device_id)(type) = (((type) & ~0xfff00) | (((data->device_id) &
 0xfff) << 8));
1183			graph->streams[i] = ccv_nnc_stream_context_new(type);
1184		}
1185		graph->signal_size = signal_size;
1186		graph->signals = (ccv_nnc_stream_signal_t**)cccalloccalloc(signal_size, sizeof(ccv_nnc_stream_signal_t*));
1187		ccv_nnc_graph_visit_for(visit, exec_info, node, idx){ int _i_; for (_i_ = 0; _i_ < (visit)->size; _i_++) { const
 int idx __attribute__((unused)) = (visit)->node[_i_].index
; const int _node_unused_ __attribute__((unused)) = (visit)->
node[_i_].term; typeof ((exec_info)) const node __attribute__
((unused)) = (exec_info) + idx; {
1188			for (i = 0; i < schd_info[idx].stream_size; i++)
1189				if (SCHEDULE_SIGNALS(schd_info[idx])((schd_info[idx]).stream_size <= 1 ? (schd_info[idx])._inline_signals
 : (schd_info[idx])._heap_signals)[i] >= 0)
1190				{
1191					const int signal = SCHEDULE_SIGNALS(schd_info[idx])((schd_info[idx]).stream_size <= 1 ? (schd_info[idx])._inline_signals
 : (schd_info[idx])._heap_signals)[i];
1192					if (!graph->signals[signal])
1193					{
1194						const ccv_nnc_stream_data_t* const data = (ccv_nnc_stream_data_t*)ccv_array_get(stream_data, SCHEDULE_STREAMS(schd_info[idx])[i])((void*)(((char*)((stream_data)->data)) + (size_t)(stream_data
)->rsize * (size_t)(((schd_info[idx]).stream_size <= 1 ?
 (schd_info[idx])._inline_streams : (schd_info[idx])._heap_streams
)[i])));
1195						int type = stream_type;
1196						CCV_STREAM_SET_DEVICE_ID(type, data->device_id)(type) = (((type) & ~0xfff00) | (((data->device_id) &
 0xfff) << 8));
1197						graph->signals[signal] = ccv_nnc_stream_signal_new(type);
1198					}
1199				}
1200		} ccv_nnc_graph_visit_endfor} }
1201	} else {
1202		assert(graph->streams)((void) sizeof ((graph->streams) ? 1 : 0), __extension__ (
{ if (graph->streams) ; else __assert_fail ("graph->streams"
, "ccv_nnc_graph.c", 1202, __extension__ __PRETTY_FUNCTION__)
; }));
1203		assert(graph->stream_size >= stream_data->rnum)((void) sizeof ((graph->stream_size >= stream_data->
rnum) ? 1 : 0), __extension__ ({ if (graph->stream_size >=
 stream_data->rnum) ; else __assert_fail ("graph->stream_size >= stream_data->rnum"
, "ccv_nnc_graph.c", 1203, __extension__ __PRETTY_FUNCTION__)
; }));
1204		// Find streams to proper allocated stream based on the type we need.
1205		int* const stream_idxs = (int*)ccmallocmalloc(sizeof(int) * (stream_data->rnum + signal_size));
1206		uint64_t* const stream_used = (uint64_t*)cccalloccalloc(((graph->stream_size + 63) >> 6) + ((graph->signal_size + 63) >> 6), sizeof(uint64_t));
1207		for (i = 0; i < stream_data->rnum; i++)
1208		{
1209			ccv_nnc_stream_data_t* const data = (ccv_nnc_stream_data_t*)ccv_array_get(stream_data, i)((void*)(((char*)((stream_data)->data)) + (size_t)(stream_data
)->rsize * (size_t)(i)));
1210			int type = stream_type;
1211			CCV_STREAM_SET_DEVICE_ID(type, data->device_id)(type) = (((type) & ~0xfff00) | (((data->device_id) &
 0xfff) << 8));
1212			for (j = 0; j < graph->stream_size; j++)
1213				if (!(stream_used[j >> 6] & ((uint64_t)1 << (j & 63))))
1214				{
1215					const int stream_type = ccv_nnc_stream_context_type(graph->streams[j]);
1216					if (stream_type == type)
1217					{
1218						stream_idxs[i] = j;
1219						stream_used[j >> 6] |= ((uint64_t)1 << (j & 63));
1220						break;
1221					}
1222				}
1223		}
1224		assert(graph->signal_size >= signal_size)((void) sizeof ((graph->signal_size >= signal_size) ? 1
 : 0), __extension__ ({ if (graph->signal_size >= signal_size
) ; else __assert_fail ("graph->signal_size >= signal_size"
, "ccv_nnc_graph.c", 1224, __extension__ __PRETTY_FUNCTION__)
; }));
1225		// Find signals to proper allocated signal based on the type we need.
1226		int* const signal_idxs = stream_idxs + stream_data->rnum;
1227		uint64_t* const signal_used = stream_used + ((graph->stream_size + 63) >> 6);
1228		for (i = 0; i < signal_size; i++)
1229			signal_idxs[i] = -1;
1230		ccv_nnc_graph_visit_for(visit, exec_info, node, idx){ int _i_; for (_i_ = 0; _i_ < (visit)->size; _i_++) { const
 int idx __attribute__((unused)) = (visit)->node[_i_].index
; const int _node_unused_ __attribute__((unused)) = (visit)->
node[_i_].term; typeof ((exec_info)) const node __attribute__
((unused)) = (exec_info) + idx; {
1231			for (i = 0; i < schd_info[idx].stream_size; i++)
1232				if (SCHEDULE_SIGNALS(schd_info[idx])((schd_info[idx]).stream_size <= 1 ? (schd_info[idx])._inline_signals
 : (schd_info[idx])._heap_signals)[i] >= 0)
1233				{
1234					const int signal = SCHEDULE_SIGNALS(schd_info[idx])((schd_info[idx]).stream_size <= 1 ? (schd_info[idx])._inline_signals
 : (schd_info[idx])._heap_signals)[i];
1235					if (signal_idxs[signal] < 0)
1236					{
1237						const ccv_nnc_stream_data_t* const data = (ccv_nnc_stream_data_t*)ccv_array_get(stream_data, SCHEDULE_STREAMS(schd_info[idx])[i])((void*)(((char*)((stream_data)->data)) + (size_t)(stream_data
)->rsize * (size_t)(((schd_info[idx]).stream_size <= 1 ?
 (schd_info[idx])._inline_streams : (schd_info[idx])._heap_streams
)[i])));
1238						int type = stream_type;
1239						CCV_STREAM_SET_DEVICE_ID(type, data->device_id)(type) = (((type) & ~0xfff00) | (((data->device_id) &
 0xfff) << 8));
1240						for (j = 0; j < graph->signal_size; j++)
1241							if (!(signal_used[j >> 6] & ((uint64_t)1 << (j & 63))))
1242							{
1243								const int signal_type = ccv_nnc_stream_signal_type(graph->signals[j]);
1244								if (signal_type == type)
1245								{
1246									signal_idxs[signal] = j;
1247									signal_used[j >> 6] |= ((uint64_t)1 << (j & 63));
1248									break;
1249								}
1250							}
1251					}
1252				}
1253		} ccv_nnc_graph_visit_endfor} }
1254		// Now rebind streams and signals from the schedule.
1255		ccv_nnc_graph_visit_for(visit, exec_info, node, idx){ int _i_; for (_i_ = 0; _i_ < (visit)->size; _i_++) { const
 int idx __attribute__((unused)) = (visit)->node[_i_].index
; const int _node_unused_ __attribute__((unused)) = (visit)->
node[_i_].term; typeof ((exec_info)) const node __attribute__
((unused)) = (exec_info) + idx; {
1256			for (i = 0; i < schd_info[idx].stream_size; i++)
1257			{
1258				SCHEDULE_STREAMS(schd_info[idx])((schd_info[idx]).stream_size <= 1 ? (schd_info[idx])._inline_streams
 : (schd_info[idx])._heap_streams)[i] = stream_idxs[SCHEDULE_STREAMS(schd_info[idx])((schd_info[idx]).stream_size <= 1 ? (schd_info[idx])._inline_streams
 : (schd_info[idx])._heap_streams)[i]];
1259				if (SCHEDULE_SIGNALS(schd_info[idx])((schd_info[idx]).stream_size <= 1 ? (schd_info[idx])._inline_signals
 : (schd_info[idx])._heap_signals)[i] >= 0)
1260					SCHEDULE_SIGNALS(schd_info[idx])((schd_info[idx]).stream_size <= 1 ? (schd_info[idx])._inline_signals
 : (schd_info[idx])._heap_signals)[i] = signal_idxs[SCHEDULE_SIGNALS(schd_info[idx])((schd_info[idx]).stream_size <= 1 ? (schd_info[idx])._inline_signals
 : (schd_info[idx])._heap_signals)[i]];
1261			}
1262			for (i = 0; i < schd_info[idx].wait_size; i++)
1263				schd_info[idx].waits[i] = signal_idxs[schd_info[idx].waits[i]];
1264		} ccv_nnc_graph_visit_endfor} }
1265		for (i = 0; i < schedule->stream_1_size; i++)
1266			schedule->stream_1s[i] = stream_idxs[schedule->stream_1s[i]];
1267		for (i = 0; i < schedule->wait_size; i++)
1268			schedule->waits[i] = signal_idxs[schedule->waits[i]];
1269		// Rebind who is the stream 0 (default stream).
1270		schedule->stream_0 = stream_idxs[0];
1271		ccfreefree(stream_used);
1272		ccfreefree(stream_idxs);
1273	}
1274	assert(graph->streams)((void) sizeof ((graph->streams) ? 1 : 0), __extension__ (
{ if (graph->streams) ; else __assert_fail ("graph->streams"
, "ccv_nnc_graph.c", 1274, __extension__ __PRETTY_FUNCTION__)
; }));
1275	ccv_nnc_graph_visit_free(visit);
1276	for (i = 0; i < signal_size; i++)
1277		{ assert(graph->signals[i])((void) sizeof ((graph->signals[i]) ? 1 : 0), __extension__
 ({ if (graph->signals[i]) ; else __assert_fail ("graph->signals[i]"
, "ccv_nnc_graph.c", 1277, __extension__ __PRETTY_FUNCTION__)
; })); }
1278	if (schedule->stream_1_size)
1279		schedule->begin = ccv_nnc_stream_signal_new(default_stream_type);
1280	schedule->end = ccv_nnc_stream_signal_new(default_stream_type);
1281	// Do this recursively for its sub graphs.
1282	if (graph->sub_graphs)
1283		for (i = 0; i < graph->sub_graphs->rnum; i++)
1284		{
1285			ccv_nnc_graph_t* const sub_graph = *(ccv_nnc_graph_t**)ccv_array_get(graph->sub_graphs, i)((void*)(((char*)((graph->sub_graphs)->data)) + (size_t
)(graph->sub_graphs)->rsize * (size_t)(i)));
1286			if (sub_graph && !sub_graph->default_schedule)
1287			{
1288				const int exec_idx = sub_graph->exec_idx - 1;
1289				assert(schd_info[exec_idx].stream_size == 1)((void) sizeof ((schd_info[exec_idx].stream_size == 1) ? 1 : 0
), __extension__ ({ if (schd_info[exec_idx].stream_size == 1)
 ; else __assert_fail ("schd_info[exec_idx].stream_size == 1"
, "ccv_nnc_graph.c", 1289, __extension__ __PRETTY_FUNCTION__)
; }));
1290				const int stream_idx = SCHEDULE_STREAMS(schd_info[exec_idx])((schd_info[exec_idx]).stream_size <= 1 ? (schd_info[exec_idx
])._inline_streams : (schd_info[exec_idx])._heap_streams)[0];
1291				const int device_id = ((ccv_nnc_stream_data_t*)ccv_array_get(stream_data, stream_idx)((void*)(((char*)((stream_data)->data)) + (size_t)(stream_data
)->rsize * (size_t)(stream_idx))))->device_id;
1292				sub_graph->default_schedule = _ccv_nnc_graph_static_schedule_new(sub_graph, stream_type, device_id, max_stream_count, graph->streams[stream_idx], 0, 0, 0, 0);
1293			}
1294		}
1295	ccv_array_free(stream_data);
1296	return schedule;
1297}
1298void ccv_nnc_graph_set_default_static_schedule(ccv_nnc_graph_t* const graph, const int stream_type, const int max_stream_count)
1299{
1300	assert(graph->p == 0)((void) sizeof ((graph->p == 0) ? 1 : 0), __extension__ ({
 if (graph->p == 0) ; else __assert_fail ("graph->p == 0"
, "ccv_nnc_graph.c", 1300, __extension__ __PRETTY_FUNCTION__)
; }));
1301	if (graph->default_schedule)
1302		ccv_nnc_graph_static_schedule_free(graph->default_schedule);
1303	graph->default_schedule = _ccv_nnc_graph_static_schedule_new(graph, stream_type, -1, max_stream_count, 0, 0, 0, 0, 0);
1304}
1305 
1306ccv_nnc_graph_static_schedule_t* ccv_nnc_graph_static_schedule_new(ccv_nnc_graph_t* const graph, const int stream_type, const int max_stream_count, const ccv_nnc_graph_exec_t* const sources, const int source_size, const ccv_nnc_graph_exec_t* const destinations, const int destination_size)
1307{
1308	assert(graph->p == 0)((void) sizeof ((graph->p == 0) ? 1 : 0), __extension__ ({
 if (graph->p == 0) ; else __assert_fail ("graph->p == 0"
, "ccv_nnc_graph.c", 1308, __extension__ __PRETTY_FUNCTION__)
; }));
1309	return _ccv_nnc_graph_static_schedule_new(graph, stream_type, -1, max_stream_count, 0, sources, source_size, destinations, destination_size);
1310}
1311 
1312ccv_nnc_stream_context_t* ccv_nnc_graph_default_stream(const ccv_nnc_graph_t* const graph)
1313{
1314	if (graph->streams && graph->stream_size > 0)
1315		return graph->streams[0];
1316	return 0;
1317}
1318 
1319static void _ccv_nnc_graph_dot_exec(const int index, const ccv_nnc_graph_exec_info_t* const exec_info, const ccv_nnc_graph_exec_schedule_t* const schd_info, ccv_nnc_stream_context_t** const streams, const int flags, FILE* out)
1320{
1321	if (flags == CCV_NNC_LONG_DOT_GRAPH)
1322		fputc('{', out);
1323	fprintf(out, "node%d", index);
1324	if (flags == CCV_NNC_LONG_DOT_GRAPH)
1325	{
1326		fputs("|Command: ", out);
1327		fputs(ccv_nnc_cmd_name(exec_info->cmd.cmd), out);
1328		if (schd_info)
1329		{
1330			if (schd_info->stream_size > 0)
1331			{
1332				int i, flag = 0;
1333				fputs("|Stream: ", out);
1334				for (i = 0; i < schd_info->stream_size; i++)
1335				{
1336					const int device_id = streams ? CCV_TENSOR_GET_DEVICE_ID(streams[SCHEDULE_STREAMS(*schd_info)[i]]->type)(((streams[((*schd_info).stream_size <= 1 ? (*schd_info)._inline_streams
 : (*schd_info)._heap_streams)[i]]->type) & 0xfff00) >>
 8) : 0;
1337					if (i == 0)
1338						fprintf(out, "%d (d%d)", SCHEDULE_STREAMS(*schd_info)((*schd_info).stream_size <= 1 ? (*schd_info)._inline_streams
 : (*schd_info)._heap_streams)[i], device_id);
1339					else
1340						fprintf(out, ", %d (d%d)", SCHEDULE_STREAMS(*schd_info)((*schd_info).stream_size <= 1 ? (*schd_info)._inline_streams
 : (*schd_info)._heap_streams)[i], device_id);
1341				}
1342				for (i = 0; i < schd_info->stream_size; i++)
1343					if (SCHEDULE_SIGNALS(*schd_info)((*schd_info).stream_size <= 1 ? (*schd_info)._inline_signals
 : (*schd_info)._heap_signals)[i] >= 0)
1344					{
1345						if (!flag)
1346						{
1347							flag = 1;
1348							fprintf(out, "|Signal: %d", SCHEDULE_SIGNALS(*schd_info)((*schd_info).stream_size <= 1 ? (*schd_info)._inline_signals
 : (*schd_info)._heap_signals)[i]);
1349						} else
1350							fprintf(out, ", %d", SCHEDULE_SIGNALS(*schd_info)((*schd_info).stream_size <= 1 ? (*schd_info)._inline_signals
 : (*schd_info)._heap_signals)[i]);
1351					}
1352			}
1353			if (schd_info->wait_size > 0)
1354			{
1355				fputs("|Wait: ", out);
1356				int i;
1357				for (i = 0; i < schd_info->wait_size - 1; i++)
1358					fprintf(out, "%d, ", schd_info->waits[i]);
1359				fprintf(out, "%d", schd_info->waits[schd_info->wait_size - 1]);
1360			}
1361		}
1362		fputc('}', out);
1363	}
1364}
1365 
1366static void _ccv_nnc_graph_dot_tensor(const int index, const ccv_nnc_tensor_t* const tensor, const int zone, const int flags, const int depth, FILE* out)
1367{
1368	// if it has an alias pointer, or, it is a long form.
1369	if (flags == CCV_NNC_LONG_DOT_GRAPH)
1370		fputc('{', out);
1371	const int is_tensor_view = CCV_IS_TENSOR_VIEW(tensor)((*(int*)(tensor)) & CCV_TENSOR_VIEW);
1372	if (is_tensor_view)
1373		fprintf(out, "tensorview%d", index);
1374	else
1375		fprintf(out, "tensor%d", index);
1376	int i;
1377	for (i = 0; i < depth; i++) // Print subscription to denote depth.
1378		fputc('\'', out);
1379	if (CCV_GET_TAPE_ALLOC(tensor->type)((tensor->type) & CCV_TAPE_ALLOC))
1380		fputs(" (t)", out);
1381	if (flags == CCV_NNC_LONG_DOT_GRAPH)
1382	{
1383		const int device_id = CCV_TENSOR_GET_DEVICE_ID(tensor->info.type)(((tensor->info.type) & 0xfff00) >> 8);
1384		fprintf(out, "|d%d|zone%d", device_id, zone);
1385		for (i = 0; i < depth; i++) // Print subscription to denote depth.
1386			fputc('\'', out);
1387		uintptr_t aptr = (uintptr_t)tensor->data.u8;
1388		size_t tensor_size;
1389		if (is_tensor_view)
1390			tensor_size = (size_t)((ccv_nnc_tensor_view_t*)(tensor))->stride[0] * tensor->info.dim[0] * CCV_GET_DATA_TYPE_SIZE(tensor->type)_ccv_get_data_type_size[((tensor->type) & 0xFF000) >>
 12];
1391		else
1392			tensor_size = ccv_nnc_dimension_count(tensor->info.dim) * CCV_GET_DATA_TYPE_SIZE(tensor->type)_ccv_get_data_type_size[((tensor->type) & 0xFF000) >>
 12];
1393		// Print out the range as well.
1394		fprintf(out, "|{%#010x|%#010x}|%d", (uint32_t)aptr, (uint32_t)(aptr + tensor_size - 1), tensor->info.dim[0]);
1395		for (i = 1; i < CCV_NNC_MAX_DIM_ALLOC(12) && tensor->info.dim[i]; i++)
1396			fprintf(out, "x%d", tensor->info.dim[i]);
1397		fputc('}', out);
1398	}
1399}
1400 
1401typedef struct {
1402	int index;
1403	int name;
1404	int zone;
1405	uintptr_t tensor_ref;
1406	uintptr_t start_ptr;
1407	uintptr_t end_ptr;
1408} ccv_nnc_tensor_dot_t;
1409 
1410typedef struct {
1411	ccv_nnc_tensor_dot_t* dots;
1412	int* remap;
1413	int* rename_zone;
1414	int* rename_index;
1415} ccv_nnc_tensor_dot_recovery_t;
1416 
1417// First sort by start_ptr, then sort by tensor ptr (so that we will have the same tensor sorted to one cluster).
1418#define less_than(i1, i2, aux) ((i1).start_ptr < (i2).start_ptr || ((i1).start_ptr == (i2).start_ptr && (i1).tensor_ref < (i2).tensor_ref))
1419static CCV_IMPLEMENT_QSORT(_ccv_nnc_tensor_dot_sort_by_ptr, ccv_nnc_tensor_dot_t, less_than)void _ccv_nnc_tensor_dot_sort_by_ptr(ccv_nnc_tensor_dot_t *array
, size_t total, int aux) { int isort_thresh = 7; ccv_nnc_tensor_dot_t
 t; int sp = 0; struct { ccv_nnc_tensor_dot_t *lb; ccv_nnc_tensor_dot_t
 *ub; } stack[48]; if( total <= 1 ) return; stack[0].lb = array
; stack[0].ub = array + (total - 1); while( sp >= 0 ) { ccv_nnc_tensor_dot_t
* left = stack[sp].lb; ccv_nnc_tensor_dot_t* right = stack[sp
--].ub; for(;;) { int i, n = (int)(right - left) + 1, m; ccv_nnc_tensor_dot_t
* ptr; ccv_nnc_tensor_dot_t* ptr2; if( n <= isort_thresh )
 { insert_sort: for( ptr = left + 1; ptr <= right; ptr++ )
 { for( ptr2 = ptr; ptr2 > left && less_than(ptr2[
0],ptr2[-1], aux); ptr2--) (((t)) = ((ptr2[0])), ((ptr2[0])) =
 ((ptr2[-1])), ((ptr2[-1])) = ((t))); } break; } else { ccv_nnc_tensor_dot_t
* left0; ccv_nnc_tensor_dot_t* left1; ccv_nnc_tensor_dot_t* right0
; ccv_nnc_tensor_dot_t* right1; ccv_nnc_tensor_dot_t* pivot; ccv_nnc_tensor_dot_t
* a; ccv_nnc_tensor_dot_t* b; ccv_nnc_tensor_dot_t* c; int swap_cnt
 = 0; left0 = left; right0 = right; pivot = left + (n/2); if(
 n > 40 ) { int d = n / 8; a = left, b = left + d, c = left
 + 2*d; left = less_than(*a, *b, aux) ? (less_than(*b, *c, aux
) ? b : (less_than(*a, *c, aux) ? c : a)) : (less_than(*c, *b
, aux) ? b : (less_than(*a, *c, aux) ? a : c)); a = pivot - d
, b = pivot, c = pivot + d; pivot = less_than(*a, *b, aux) ? (
less_than(*b, *c, aux) ? b : (less_than(*a, *c, aux) ? c : a)
) : (less_than(*c, *b, aux) ? b : (less_than(*a, *c, aux) ? a
 : c)); a = right - 2*d, b = right - d, c = right; right = less_than
(*a, *b, aux) ? (less_than(*b, *c, aux) ? b : (less_than(*a, *
c, aux) ? c : a)) : (less_than(*c, *b, aux) ? b : (less_than(
*a, *c, aux) ? a : c)); } a = left, b = pivot, c = right; pivot
 = less_than(*a, *b, aux) ? (less_than(*b, *c, aux) ? b : (less_than
(*a, *c, aux) ? c : a)) : (less_than(*c, *b, aux) ? b : (less_than
(*a, *c, aux) ? a : c)); if( pivot != left0 ) { (((t)) = ((*pivot
)), ((*pivot)) = ((*left0)), ((*left0)) = ((t))); pivot = left0
; } left = left1 = left0 + 1; right = right1 = right0; for(;;
) { while( left <= right && !less_than(*pivot, *left
, aux) ) { if( !less_than(*left, *pivot, aux) ) { if( left >
 left1 ) (((t)) = ((*left1)), ((*left1)) = ((*left)), ((*left
)) = ((t))); swap_cnt = 1; left1++; } left++; } while( left <=
 right && !less_than(*right, *pivot, aux) ) { if( !less_than
(*pivot, *right, aux) ) { if( right < right1 ) (((t)) = ((
*right1)), ((*right1)) = ((*right)), ((*right)) = ((t))); swap_cnt
 = 1; right1--; } right--; } if( left > right ) break; (((
t)) = ((*left)), ((*left)) = ((*right)), ((*right)) = ((t)));
 swap_cnt = 1; left++; right--; } if( swap_cnt == 0 ) { left =
 left0, right = right0; goto insert_sort; } n = ({ typeof ((int
)(left1 - left0)) _a = ((int)(left1 - left0)); typeof ((int)(
left - left1)) _b = ((int)(left - left1)); (_a < _b) ? _a :
 _b; }); for( i = 0; i < n; i++ ) (((t)) = ((left0[i])), (
(left0[i])) = ((left[i-n])), ((left[i-n])) = ((t))); n = ({ typeof
 ((int)(right0 - right1)) _a = ((int)(right0 - right1)); typeof
 ((int)(right1 - right)) _b = ((int)(right1 - right)); (_a <
 _b) ? _a : _b; }); for( i = 0; i < n; i++ ) (((t)) = ((left
[i])), ((left[i])) = ((right0[i-n+1])), ((right0[i-n+1])) = (
(t))); n = (int)(left - left1); m = (int)(right1 - right); if
( n > 1 ) { if( m > 1 ) { if( n > m ) { stack[++sp].
lb = left0; stack[sp].ub = left0 + n - 1; left = right0 - m +
 1, right = right0; } else { stack[++sp].lb = right0 - m + 1;
 stack[sp].ub = right0; left = left0, right = left0 + n - 1; }
 } else left = left0, right = left0 + n - 1; } else if( m >
 1 ) left = right0 - m + 1, right = right0; else break; } } }
 }
1420#undef less_than
1421 
1422static int _ccv_nnc_graph_dot_tensor_multiview_count(const ccv_nnc_tensor_multiview_t* const mv)
1423{
1424	if (!CCV_IS_TENSOR_MULTIVIEW(mv)((*(int*)(mv)) & CCV_TENSOR_MULTIVIEW))
1425		return 1;
1426	const int count = mv->kind + mv->repeat;
1427	int i, c = 0;
1428	for (i = 0; i < count; i++)
1429		c += _ccv_nnc_graph_dot_tensor_multiview_count((ccv_nnc_tensor_multiview_t*)CCV_NNC_MULTIVIEW_DATA(mv)((mv)->_heap_data ? (mv)->_heap_data : (mv)->_inline_data
)[i]);
1430	return c;
1431}
1432 
1433static void _ccv_nnc_graph_dot_tensor_multiview_tensor_dots(const ccv_nnc_tensor_multiview_t* const mv, ccv_nnc_tensor_dot_t* const tensor_dots, int* tensor_index)
1434{
1435	const int count = mv->kind + mv->repeat;
1436	int i;
1437	for (i = 0; i < count; i++)
1438		if (CCV_IS_TENSOR_MULTIVIEW(CCV_NNC_MULTIVIEW_DATA(mv)[i])((*(int*)(((mv)->_heap_data ? (mv)->_heap_data : (mv)->
_inline_data)[i])) & CCV_TENSOR_MULTIVIEW))
1439			_ccv_nnc_graph_dot_tensor_multiview_tensor_dots((ccv_nnc_tensor_multiview_t*)CCV_NNC_MULTIVIEW_DATA(mv)((mv)->_heap_data ? (mv)->_heap_data : (mv)->_inline_data
)[i], tensor_dots, tensor_index);
1440		else {
1441			tensor_dots[*tensor_index].name = *tensor_index;
1442			tensor_dots[*tensor_index].start_ptr =  (uintptr_t)CCV_NNC_MULTIVIEW_DATA(mv)((mv)->_heap_data ? (mv)->_heap_data : (mv)->_inline_data
)[i]->data.u8;
1443			// Because tv's pointer will get updated, it is not correct in this case to have one tensor_ref.
1444			tensor_dots[*tensor_index].tensor_ref = tensor_dots[*tensor_index].start_ptr;
1445			const size_t dim_size = ccv_nnc_dimension_count(CCV_NNC_MULTIVIEW_DATA(mv)((mv)->_heap_data ? (mv)->_heap_data : (mv)->_inline_data
)[i]->info.dim) * CCV_GET_DATA_TYPE_SIZE(CCV_NNC_MULTIVIEW_DATA(mv)[i]->type)_ccv_get_data_type_size[((((mv)->_heap_data ? (mv)->_heap_data
 : (mv)->_inline_data)[i]->type) & 0xFF000) >>
 12];
1446			tensor_dots[*tensor_index].end_ptr = tensor_dots[*tensor_index].start_ptr + dim_size - 1;
1447			++(*tensor_index);
1448		}
1449}
1450 
1451static ccv_nnc_tensor_dot_recovery_t _ccv_nnc_graph_tensor_dot_recovery(const ccv_nnc_graph_t* const graph)
1452{
1453	int i, j;
1454	// Recover tensor relationships for all tensors referenced in the graph.
1455	// Most notably, we have to give these indexes, and find if they point to
1456	// the same memory region, and whether they overlap. These information
1457	// are lost since we converted from symbolic form to the execution form.
1458	// and here we do our best to recover because that is easier to understand
1459	// if we want to present the graph visually (also, we don't want to put this
1460	// information into the tensor or execution graph to avoid overhead, thus,
1461	// recovering is the best we can do).
1462	int tensor_count = 0;
1463	for (i = 0; i < graph->exec_info->rnum; i++)
1464	{
1465		ccv_nnc_graph_exec_info_t* exec_info = (ccv_nnc_graph_exec_info_t*)ccv_array_get(graph->exec_info, i)((void*)(((char*)((graph->exec_info)->data)) + (size_t)
(graph->exec_info)->rsize * (size_t)(i)));
1466		for (j = 0; j < exec_info->input_size; j++)
1467			if (exec_info->inputs[j])
1468				tensor_count += CCV_IS_TENSOR_MULTIVIEW(exec_info->inputs[j])((*(int*)(exec_info->inputs[j])) & CCV_TENSOR_MULTIVIEW
) ? _ccv_nnc_graph_dot_tensor_multiview_count((ccv_nnc_tensor_multiview_t*)exec_info->inputs[j]) : 1;
1469		for (j = 0; j < exec_info->output_size; j++)
1470			if (exec_info->outputs[j])
1471				tensor_count += CCV_IS_TENSOR_MULTIVIEW(exec_info->outputs[j])((*(int*)(exec_info->outputs[j])) & CCV_TENSOR_MULTIVIEW
) ? _ccv_nnc_graph_dot_tensor_multiview_count((ccv_nnc_tensor_multiview_t*)exec_info->outputs[j]) : 1;
1472	}
1473	ccv_nnc_tensor_dot_t* tensor_dots = tensor_count > 0 ? (ccv_nnc_tensor_dot_t*)ccmallocmalloc(sizeof(ccv_nnc_tensor_dot_t) * tensor_count) : 0;
1474	int k = 0;
1475	for (i = 0; i < graph->exec_info->rnum; i++)
1476	{
1477		ccv_nnc_graph_exec_info_t* exec_info = (ccv_nnc_graph_exec_info_t*)ccv_array_get(graph->exec_info, i)((void*)(((char*)((graph->exec_info)->data)) + (size_t)
(graph->exec_info)->rsize * (size_t)(i)));
1478		for (j = 0; j < exec_info->input_size; j++)
1479		{
1480			ccv_nnc_tensor_t* tensor = exec_info->inputs[j];
1481			if (!tensor)
1482				continue;
1483			if (CCV_IS_TENSOR_MULTIVIEW(tensor)((*(int*)(tensor)) & CCV_TENSOR_MULTIVIEW))
1484				_ccv_nnc_graph_dot_tensor_multiview_tensor_dots((ccv_nnc_tensor_multiview_t*)tensor, tensor_dots, &k);
1485			else {
1486				tensor_dots[k].name = k;
1487				tensor_dots[k].tensor_ref = (uintptr_t)tensor;
1488				tensor_dots[k].start_ptr = (uintptr_t)tensor->data.u8;
1489				size_t tensor_size;
1490				if (CCV_IS_TENSOR_VIEW(tensor)((*(int*)(tensor)) & CCV_TENSOR_VIEW))
1491					tensor_size = (size_t)((ccv_nnc_tensor_view_t*)(tensor))->stride[0] * tensor->info.dim[0] * CCV_GET_DATA_TYPE_SIZE(tensor->type)_ccv_get_data_type_size[((tensor->type) & 0xFF000) >>
 12];
1492				else
1493					tensor_size = ccv_nnc_dimension_count(tensor->info.dim) * CCV_GET_DATA_TYPE_SIZE(tensor->type)_ccv_get_data_type_size[((tensor->type) & 0xFF000) >>
 12];
1494				tensor_dots[k].end_ptr = tensor_dots[k].start_ptr + tensor_size - 1;
1495				++k;
1496			}
1497		}
1498		for (j = 0; j < exec_info->output_size; j++)
1499		{
1500			ccv_nnc_tensor_t* tensor = exec_info->outputs[j];
1501			if (!tensor)
1502				continue;
1503			if (CCV_IS_TENSOR_MULTIVIEW(tensor)((*(int*)(tensor)) & CCV_TENSOR_MULTIVIEW))
1504				_ccv_nnc_graph_dot_tensor_multiview_tensor_dots((ccv_nnc_tensor_multiview_t*)tensor, tensor_dots, &k);
1505			else {
1506				tensor_dots[k].name = k;
1507				tensor_dots[k].tensor_ref = (uintptr_t)tensor;
1508				tensor_dots[k].start_ptr = (uintptr_t)tensor->data.u8;
1509				size_t tensor_size;
1510				if (CCV_IS_TENSOR_VIEW(tensor)((*(int*)(tensor)) & CCV_TENSOR_VIEW))
1511					tensor_size = (size_t)((ccv_nnc_tensor_view_t*)(tensor))->stride[0] * tensor->info.dim[0] * CCV_GET_DATA_TYPE_SIZE(tensor->type)_ccv_get_data_type_size[((tensor->type) & 0xFF000) >>
 12];
1512				else
1513					tensor_size = ccv_nnc_dimension_count(tensor->info.dim) * CCV_GET_DATA_TYPE_SIZE(tensor->type)_ccv_get_data_type_size[((tensor->type) & 0xFF000) >>
 12];
1514				tensor_dots[k].end_ptr = tensor_dots[k].start_ptr + tensor_size - 1;
1515				++k;
1516			}
1517		}
1518	}
1519	tensor_count = k; // We may over count, now shrink.
1520	// To group overlap memory into one zone, we sort it by start ptr first (secondary by the tensor pointer).
1521	_ccv_nnc_tensor_dot_sort_by_ptr(tensor_dots, tensor_count, 0);
1522	int index = 0, zone = 0;
1523	uintptr_t tensor_ref = tensor_count > 0 ? tensor_dots[0].tensor_ref : 0;
1524	uintptr_t end_ptr = tensor_count > 0 ? tensor_dots[0].end_ptr : 0;
1525	// Then, it is trivial, we go by end ptr. If the next start ptr is still within the end ptr (start ptr <= end ptr),
1526	// they are the same zone.
1527	for (i = 0; i < tensor_count; i++)
1528	{
1529		if (tensor_dots[i].tensor_ref != tensor_ref)
1530		{
1531			tensor_ref = tensor_dots[i].tensor_ref;
1532			++index;
1533		}
1534		if (tensor_dots[i].start_ptr > end_ptr)
1535		{
1536			end_ptr = ccv_max(end_ptr, tensor_dots[i].end_ptr)({ typeof (end_ptr) _a = (end_ptr); typeof (tensor_dots[i].end_ptr
) _b = (tensor_dots[i].end_ptr); (_a > _b) ? _a : _b; });
1537			++zone;
1538		}
1539		tensor_dots[i].index = index;
1540		tensor_dots[i].zone = zone;
1541	}
1542	// We already have index and zone assigned, but the problem is that these are not very human interpretable (because
1543	// it follows the pointer from low to high, not the tensor creation order). The following code renamed both the index
1544	// and the zone so that it is much more understandable.
1545	const int index_count = index + 1;
1546	const int zone_count = zone + 1;
1547	int* remap = (int*)ccmallocmalloc(sizeof(int) * (tensor_count + index_count + zone_count));
1548	int* rename_index = remap + tensor_count;
1549	int* rename_zone = rename_index + index_count;
1550	for (i = 0; i < tensor_count; i++)
1551		remap[tensor_dots[i].name] = i;
1552	for (i = 0; i < index_count; i++)
1553		rename_index[i] = -1;
1554	for (i = 0; i < zone_count; i++)
1555		rename_zone[i] = -1;
1556	index = 0;
1557	zone = 0;
1558	for (i = 0; i < tensor_count; i++)
1559	{
1560		ccv_nnc_tensor_dot_t* tensor_dot = tensor_dots + remap[i];
1561		if (rename_index[tensor_dot->index] == -1)
1562			rename_index[tensor_dot->index] = index++;
1563		if (rename_zone[tensor_dot->zone] == -1)
1564			rename_zone[tensor_dot->zone] = zone++;
1565	}
1566	ccv_nnc_tensor_dot_recovery_t recovery = {
1567		.dots = tensor_dots,
1568		.remap = remap,
1569		.rename_index = rename_index,
1570		.rename_zone = rename_zone,
1571	};
1572	return recovery;
1573}
1574 
1575static void _ccv_nnc_graph_tensor_dot_recovery_free(const ccv_nnc_tensor_dot_recovery_t recovery)
1576{
1577	ccfreefree(recovery.dots);
1578	ccfreefree(recovery.remap);
1579}
1580 
1581static void _ccv_nnc_graph_dot_tensor_multiview_one(const ccv_nnc_tensor_multiview_t* const mv, const ccv_nnc_tensor_dot_recovery_t recovery, const int depth, int* tensor_index, FILE* out)
1582{
1583	const int count = mv->kind + mv->repeat;
1584	int i, j;
1585	fputs("|{", out);
1586	for (i = 0; i < count; i++)
1587		if (CCV_IS_TENSOR_MULTIVIEW(CCV_NNC_MULTIVIEW_DATA(mv)[i])((*(int*)(((mv)->_heap_data ? (mv)->_heap_data : (mv)->
_inline_data)[i])) & CCV_TENSOR_MULTIVIEW))
1588		{
1589			fprintf(out, "{%d", i);
1590			if (mv->kind == CCV_NNC_MULTIVIEW_K0N || (mv->kind == CCV_NNC_MULTIVIEW_K1N && i > 0))
1591				fputc('*', out); // Denotes that we loop on this.
1592			_ccv_nnc_graph_dot_tensor_multiview_one((ccv_nnc_tensor_multiview_t*)CCV_NNC_MULTIVIEW_DATA(mv)((mv)->_heap_data ? (mv)->_heap_data : (mv)->_inline_data
)[i], recovery, depth, tensor_index, out);
1593			if (i == count - 1)
1594				fputc('}', out);
1595			else
1596				fputs("}|", out);
1597		} else {
1598			fprintf(out, "{%d", i);
1599			if (mv->kind == CCV_NNC_MULTIVIEW_K0N || (mv->kind == CCV_NNC_MULTIVIEW_K1N && i > 0))
1600				fputc('*', out); // Denotes that we loop on this.
1601			const ccv_nnc_tensor_dot_t* const tensor_dot = recovery.dots + recovery.remap[*tensor_index];
1602			fprintf(out, "|zone%d", recovery.rename_zone[tensor_dot->zone]);
1603			for (j = 0; j < depth; j++)
1604				fputc('\'', out);
1605			uintptr_t aptr = (uintptr_t)CCV_NNC_MULTIVIEW_DATA(mv)((mv)->_heap_data ? (mv)->_heap_data : (mv)->_inline_data
)[i]->data.u8;
1606			// For the last one, we don't extend to full ainc.
1607			size_t dim_size = ccv_nnc_dimension_count(CCV_NNC_MULTIVIEW_DATA(mv)((mv)->_heap_data ? (mv)->_heap_data : (mv)->_inline_data
)[i]->info.dim) * CCV_GET_DATA_TYPE_SIZE(CCV_NNC_MULTIVIEW_DATA(mv)[i]->type)_ccv_get_data_type_size[((((mv)->_heap_data ? (mv)->_heap_data
 : (mv)->_inline_data)[i]->type) & 0xFF000) >>
 12];
1608			// Print out the range as well.
1609			fprintf(out, "|{%#010x|%#010x}", (uint32_t)aptr, (uint32_t)(aptr + dim_size - 1));
1610			++(*tensor_index);
1611			if (i == count - 1)
1612				fputc('}', out);
1613			else
1614				fputs("}|", out);
1615		}
1616	fputc('}', out);
1617}
1618 
1619static void _ccv_nnc_graph_dot_tensor_multiview(const ccv_nnc_tensor_multiview_t* const mv, const ccv_nnc_tensor_dot_recovery_t recovery, const int flags, const int depth, int* tensor_index, FILE* out)
1620{
1621	// if it has an alias pointer, or, it is a long form.
1622	if (flags == CCV_NNC_LONG_DOT_GRAPH)
1623		fputc('{', out);
1624	const ccv_nnc_tensor_dot_t* const tensor_dot = recovery.dots + recovery.remap[*tensor_index];
1625	fprintf(out, "multiview%d", recovery.rename_index[tensor_dot->index]);
1626	int i;
1627	for (i = 0; i < depth; i++) // Print subscription to denote depth.
1628		fputc('\'', out);
1629	if (CCV_GET_TAPE_ALLOC(mv->type)((mv->type) & CCV_TAPE_ALLOC))
1630		fputs(" (t)", out);
1631	if (flags == CCV_NNC_LONG_DOT_GRAPH)
1632	{
1633		_ccv_nnc_graph_dot_tensor_multiview_one(mv, recovery, depth, tensor_index, out);
1634		const ccv_nnc_tensor_t* root = (ccv_nnc_tensor_t*)mv;
1635		while (CCV_IS_TENSOR_MULTIVIEW(root)((*(int*)(root)) & CCV_TENSOR_MULTIVIEW))
1636			root = CCV_NNC_MULTIVIEW_DATA((ccv_nnc_tensor_multiview_t*)root)(((ccv_nnc_tensor_multiview_t*)root)->_heap_data ? ((ccv_nnc_tensor_multiview_t
*)root)->_heap_data : ((ccv_nnc_tensor_multiview_t*)root)->
_inline_data)[0];
1637		fprintf(out, "|%d", root->info.dim[0]);
1638		for (i = 1; i < CCV_NNC_MAX_DIM_ALLOC(12) && root->info.dim[i]; i++)
1639			fprintf(out, "x%d", root->info.dim[i]);
1640		fputc('}', out);
1641	} else
1642		*tensor_index += _ccv_nnc_graph_dot_tensor_multiview_count(mv);
1643}
1644 
1645static void _ccv_nnc_graph_dot_node(const ccv_nnc_graph_exec_info_t* const exec_info, const ccv_nnc_graph_exec_schedule_t* const schd_info, const int exec_index, ccv_nnc_stream_context_t** const streams, const ccv_nnc_tensor_dot_recovery_t recovery, const int flags, const int depth, FILE* out, int* const tensor_index)
1646{
1647	fprintf(out, "node%d [shape=record,label=\"", exec_index);
1648	_ccv_nnc_graph_dot_exec(exec_index, exec_info, schd_info, streams, flags, out);
1649	int i;
1650	int k = *tensor_index;
1651	if (exec_info->input_size > 0)
1652	{
1653		fputs("|{Input", out);
1654		for (i = 0; i < exec_info->input_size; i++)
1655			if (exec_info->inputs[i])
1656			{
1657				fputc('|', out);
1658				if (CCV_IS_TENSOR_MULTIVIEW(exec_info->inputs[i])((*(int*)(exec_info->inputs[i])) & CCV_TENSOR_MULTIVIEW
))
1659					_ccv_nnc_graph_dot_tensor_multiview((ccv_nnc_tensor_multiview_t*)exec_info->inputs[i], recovery, flags, depth, &k, out);
1660				else {
1661					const ccv_nnc_tensor_dot_t* const tensor_dot = recovery.dots + recovery.remap[k];
1662					_ccv_nnc_graph_dot_tensor(recovery.rename_index[tensor_dot->index], exec_info->inputs[i], recovery.rename_zone[tensor_dot->zone], flags, depth, out);
1663					++k;
1664				}
1665			} else
1666				fputs("|-", out);
1667		fputc('}', out);
1668	}
1669	if (exec_info->output_size > 0)
1670	{
1671		fputs("|{Output", out);
1672		for (i = 0; i < exec_info->output_size; i++)
1673			if (exec_info->outputs[i])
1674			{
1675				fputc('|', out);
1676				if (CCV_IS_TENSOR_MULTIVIEW(exec_info->outputs[i])((*(int*)(exec_info->outputs[i])) & CCV_TENSOR_MULTIVIEW
))
1677					_ccv_nnc_graph_dot_tensor_multiview((ccv_nnc_tensor_multiview_t*)exec_info->outputs[i], recovery, flags, depth, &k, out);
1678				else {
1679					const ccv_nnc_tensor_dot_t* const tensor_dot = recovery.dots + recovery.remap[k];
1680					_ccv_nnc_graph_dot_tensor(recovery.rename_index[tensor_dot->index], exec_info->outputs[i], recovery.rename_zone[tensor_dot->zone], flags, depth, out);
1681					++k;
1682				}
1683			} else
1684				fputs("|-", out);
1685		fputc('}', out);
1686	}
1687	fputs("\"];\n", out);
1688	*tensor_index = k;
1689}
1690 
1691static void _ccv_nnc_graph_dot_while_label(const ccv_nnc_graph_exec_info_t* const exec_info, const int exec_index, const ccv_nnc_tensor_dot_recovery_t recovery, const ccv_nnc_graph_t* const while_graph, const int flags, const int depth, FILE* out, int* tensor_index)
1692{
1693	int i;
1694	fprintf(out, "label=<<b>while%d </b>Command: ", exec_index);
1695	fputs(ccv_nnc_cmd_name(exec_info->cmd.cmd), out);
1696	fputs(">;\n", out);
1697	fprintf(out, "label%d [shape=record,label=\"{", exec_index);
1698	int k = *tensor_index;
1699	if (exec_info->input_size > 0)
1700	{
1701		fputs("{Input|{", out);
1702		for (i = 0; i < exec_info->input_size; i++)
1703		{
1704			if (i > 0)
1705				fputc('|', out);
1706			if (exec_info->inputs[i])
1707			{
1708				if (CCV_IS_TENSOR_MULTIVIEW(exec_info->inputs[i])((*(int*)(exec_info->inputs[i])) & CCV_TENSOR_MULTIVIEW
))
1709					_ccv_nnc_graph_dot_tensor_multiview((ccv_nnc_tensor_multiview_t*)exec_info->inputs[i], recovery, flags, depth, &k, out);
1710				else {
1711					const ccv_nnc_tensor_dot_t* const tensor_dot = recovery.dots + recovery.remap[k];
1712					_ccv_nnc_graph_dot_tensor(recovery.rename_index[tensor_dot->index], exec_info->inputs[i], recovery.rename_zone[tensor_dot->zone], flags, depth, out);
1713					++k;
1714				}
1715			} else
1716				fputc('-', out);
1717		}
1718		fputs("}}", out);
1719	}
1720	if (exec_info->output_size > 0)
1721	{
1722		if (exec_info->input_size > 0)
1723			fputs("|", out);
1724		fputs("{Output|{", out);
1725		for (i = 0; i < exec_info->output_size; i++)
1726		{
1727			if (i > 0)
1728				fputc('|', out);
1729			if (exec_info->outputs[i])
1730			{
1731				if (CCV_IS_TENSOR_MULTIVIEW(exec_info->outputs[i])((*(int*)(exec_info->outputs[i])) & CCV_TENSOR_MULTIVIEW
))
1732					_ccv_nnc_graph_dot_tensor_multiview((ccv_nnc_tensor_multiview_t*)exec_info->outputs[i], recovery, flags, depth, &k, out);
1733				else {
1734					const ccv_nnc_tensor_dot_t* const tensor_dot = recovery.dots + recovery.remap[k];
1735					_ccv_nnc_graph_dot_tensor(recovery.rename_index[tensor_dot->index], exec_info->outputs[i], recovery.rename_zone[tensor_dot->zone], flags, depth, out);
1736					++k;
1737				}
1738			} else
1739				fputc('-', out);
1740		}
1741		fputs("}}", out);
1742	}
1743	fputs("}\"];\n", out);
1744	*tensor_index = k;
1745}
1746 
1747static void _ccv_nnc_graph_dot_case_of_label(const ccv_nnc_graph_exec_info_t* const exec_info, const int exec_index, const ccv_nnc_tensor_dot_recovery_t recovery, const int flags, const int depth, FILE* out, int* tensor_index)
1748{
1749	int i;
1750	fprintf(out, "label=<<b>caseof%d </b>Command: ", exec_index);
1751	fputs(ccv_nnc_cmd_name(exec_info->cmd.cmd), out);
1752	fputs(">;\n", out);
1753	fprintf(out, "label%d [shape=record,label=\"{", exec_index);
1754	int k = *tensor_index;
1755	if (exec_info->input_size > 0)
1756	{
1757		fputs("{Input|{", out);
1758		for (i = 0; i < exec_info->input_size; i++)
1759		{
1760			if (i > 0)
1761				fputc('|', out);
1762			if (exec_info->inputs[i])
1763			{
1764				if (CCV_IS_TENSOR_MULTIVIEW(exec_info->inputs[i])((*(int*)(exec_info->inputs[i])) & CCV_TENSOR_MULTIVIEW
))
1765					_ccv_nnc_graph_dot_tensor_multiview((ccv_nnc_tensor_multiview_t*)exec_info->inputs[i], recovery, flags, depth, &k, out);
1766				else {
1767					const ccv_nnc_tensor_dot_t* const tensor_dot = recovery.dots + recovery.remap[k];
1768					_ccv_nnc_graph_dot_tensor(recovery.rename_index[tensor_dot->index], exec_info->inputs[i], recovery.rename_zone[tensor_dot->zone], flags, depth, out);
1769					++k;
1770				}
1771			} else
1772				fputc('-', out);
1773		}
1774		fputs("}}", out);
1775	}
1776	if (exec_info->output_size > 0)
1777	{
1778		if (exec_info->input_size > 0)
1779			fputs("|", out);
1780		fputs("{Output|{", out);
1781		for (i = 0; i < exec_info->output_size; i++)
1782		{
1783			if (i > 0)
1784				fputc('|', out);
1785			if (exec_info->outputs[i])
1786			{
1787				if (CCV_IS_TENSOR_MULTIVIEW(exec_info->outputs[i])((*(int*)(exec_info->outputs[i])) & CCV_TENSOR_MULTIVIEW
))
1788					_ccv_nnc_graph_dot_tensor_multiview((ccv_nnc_tensor_multiview_t*)exec_info->outputs[i], recovery, flags, depth, &k, out);
1789				else {
1790					const ccv_nnc_tensor_dot_t* const tensor_dot = recovery.dots + recovery.remap[k];
1791					_ccv_nnc_graph_dot_tensor(recovery.rename_index[tensor_dot->index], exec_info->outputs[i], recovery.rename_zone[tensor_dot->zone], flags, depth, out);
1792					++k;
1793				}
1794			} else
1795				fputc('-', out);
1796		}
1797		fputs("}}", out);
1798	}
1799	fputs("}\"];\n", out);
1800	*tensor_index = k;
1801}
1802 
1803static void _ccv_nnc_graph_dot_sub_graphs(const ccv_nnc_graph_exec_info_t* const exec_info, const ccv_nnc_tensor_dot_recovery_t p_recovery, const ccv_array_t* const sub_graphs, const int flags, const int depth, FILE* out, int* tensor_index, int* exec_index)
1804{
1805	if (exec_info->flags & CCV_NNC_GRAPH_EXEC_P_WHILE)
1806	{
1807		fprintf(out, "subgraph cluster%d {\nstyle=\"rounded\";\nnode%d [style=invisible];\n", *exec_index, *exec_index);
1808		const ccv_nnc_graph_t* const while_graph = *(ccv_nnc_graph_t**)ccv_array_get(sub_graphs, CCV_NNC_GRAPH_REF(exec_info)[0] - 1)((void*)(((char*)((sub_graphs)->data)) + (size_t)(sub_graphs
)->rsize * (size_t)(((exec_info)->_heap_graph_ref ? (exec_info
)->_heap_graph_ref : (exec_info)->_inline_graph_ref)[0]
 - 1)));
1809		// Output this node info within this subgraph.
1810		_ccv_nnc_graph_dot_while_label(exec_info, *exec_index, p_recovery, while_graph, flags, depth - 1 /* Label all references to its level above. */, out, tensor_index);
1811	} else if (exec_info->flags & CCV_NNC_GRAPH_EXEC_CASE_OF) {
1812		fprintf(out, "subgraph cluster%d {\nstyle=\"rounded\";\nnode%d [style=invisible];\n", *exec_index, *exec_index);
1813		_ccv_nnc_graph_dot_case_of_label(exec_info, *exec_index, p_recovery, flags, depth - 1 /* Label all references to its level above. */, out, tensor_index);
1814	}
1815	++(*exec_index);
1816	int p;
1817	for (p = 0; p < exec_info->graph_ref_size; p++)
1818	{
1819		if (exec_info->flags & CCV_NNC_GRAPH_EXEC_CASE_OF)
1820		{
1821			fprintf(out, "subgraph cluster%d {\nstyle=\"rounded\";\nnode%d [style=invisible];\nlabel=\"\"\n", *exec_index, *exec_index);
1822			++(*exec_index);
1823		}
1824		const ccv_nnc_graph_t* const graph = *(ccv_nnc_graph_t**)ccv_array_get(sub_graphs, CCV_NNC_GRAPH_REF(exec_info)[p] - 1)((void*)(((char*)((sub_graphs)->data)) + (size_t)(sub_graphs
)->rsize * (size_t)(((exec_info)->_heap_graph_ref ? (exec_info
)->_heap_graph_ref : (exec_info)->_inline_graph_ref)[p]
 - 1)));
1825		const ccv_nnc_graph_static_schedule_t* const schedule = graph->default_schedule;
1826		ccv_nnc_tensor_dot_recovery_t recovery = _ccv_nnc_graph_tensor_dot_recovery(graph);
1827		int i, j;
1828		int k = 0;
1829		int* node_id = (int*)ccmallocmalloc(sizeof(int) * graph->exec_info->rnum);
1830		// Output styles.
1831		for (i = 0; i < graph->exec_info->rnum; i++)
1832		{
1833			node_id[i] = *exec_index;
1834			ccv_nnc_graph_exec_info_t* exec_info = (ccv_nnc_graph_exec_info_t*)ccv_array_get(graph->exec_info, i)((void*)(((char*)((graph->exec_info)->data)) + (size_t)
(graph->exec_info)->rsize * (size_t)(i)));
1835			if (CCV_NNC_GRAPH_REF(exec_info)((exec_info)->_heap_graph_ref ? (exec_info)->_heap_graph_ref
 : (exec_info)->_inline_graph_ref)[0])
1836				_ccv_nnc_graph_dot_sub_graphs(exec_info, recovery, graph->sub_graphs, flags, depth + 1, out, &k, exec_index);
1837			else {
1838				_ccv_nnc_graph_dot_node(exec_info,
1839					schedule ? (i < schedule->exec_info_size ? schedule->exec_info + i : 0) : 0,
1840					*exec_index, graph->streams, recovery, flags, depth, out, &k);
1841				++(*exec_index);
1842			}
1843		}
1844		// Output connections.
1845		for (i = 0; i < graph->exec_info->rnum; i++)
1846		{
1847			ccv_nnc_graph_exec_info_t* exec_info = (ccv_nnc_graph_exec_info_t*)ccv_array_get(graph->exec_info, i)((void*)(((char*)((graph->exec_info)->data)) + (size_t)
(graph->exec_info)->rsize * (size_t)(i)));
1848			if (exec_info->outgoings)
1849				for (j = 0; j < exec_info->outgoings->rnum; j++)
1850				{
1851					const int outgoing_idx = *(int*)ccv_array_get(exec_info->outgoings, j)((void*)(((char*)((exec_info->outgoings)->data)) + (size_t
)(exec_info->outgoings)->rsize * (size_t)(j)));
1852					const ccv_nnc_graph_exec_info_t* const outgoing_info = (ccv_nnc_graph_exec_info_t*)ccv_array_get(graph->exec_info, outgoing_idx)((void*)(((char*)((graph->exec_info)->data)) + (size_t)
(graph->exec_info)->rsize * (size_t)(outgoing_idx)));
1853					// If both are sub-graphs, have both tail and head specified.
1854					if (CCV_NNC_GRAPH_REF(exec_info)((exec_info)->_heap_graph_ref ? (exec_info)->_heap_graph_ref
 : (exec_info)->_inline_graph_ref)[0] && CCV_NNC_GRAPH_REF(outgoing_info)((outgoing_info)->_heap_graph_ref ? (outgoing_info)->_heap_graph_ref
 : (outgoing_info)->_inline_graph_ref)[0])
1855						fprintf(out, "node%d -> node%d [ltail=cluster%d,lhead=cluster%d];\n", node_id[i], node_id[outgoing_idx], node_id[i], node_id[outgoing_idx]);
1856					else if (CCV_NNC_GRAPH_REF(exec_info)((exec_info)->_heap_graph_ref ? (exec_info)->_heap_graph_ref
 : (exec_info)->_inline_graph_ref)[0] && !CCV_NNC_GRAPH_REF(outgoing_info)((outgoing_info)->_heap_graph_ref ? (outgoing_info)->_heap_graph_ref
 : (outgoing_info)->_inline_graph_ref)[0])
1857						fprintf(out, "node%d -> node%d [ltail=cluster%d];\n", node_id[i], node_id[outgoing_idx], node_id[i]);
1858					else if (!CCV_NNC_GRAPH_REF(exec_info)((exec_info)->_heap_graph_ref ? (exec_info)->_heap_graph_ref
 : (exec_info)->_inline_graph_ref)[0] && CCV_NNC_GRAPH_REF(outgoing_info)((outgoing_info)->_heap_graph_ref ? (outgoing_info)->_heap_graph_ref
 : (outgoing_info)->_inline_graph_ref)[0])
1859						fprintf(out, "node%d -> node%d [lhead=cluster%d];\n", node_id[i], node_id[outgoing_idx], node_id[outgoing_idx]);
1860					else
1861						fprintf(out, "node%d -> node%d;\n", node_id[i], node_id[outgoing_idx]);
1862				}
1863		}
1864		fputs("}\n", out);
1865		_ccv_nnc_graph_tensor_dot_recovery_free(recovery);
1866		ccfreefree(node_id);
1867	}
1868	// Extra subgraph cluster.
1869	if (exec_info->flags & CCV_NNC_GRAPH_EXEC_CASE_OF)
1870		fputs("}\n", out);
1871}
1872 
1873void ccv_nnc_graph_dot(const ccv_nnc_graph_t* const graph, const int flags, FILE* out)
1874{
1875	fputs("digraph G {\ncompound=true;\n", out);
1876	ccv_nnc_tensor_dot_recovery_t recovery = _ccv_nnc_graph_tensor_dot_recovery(graph);
1877	int i, j;
1878	int k = 0, c = 0;
1879	int* node_id = (int*)ccmallocmalloc(sizeof(int) * graph->exec_info->rnum);
1880	const ccv_nnc_graph_static_schedule_t* const schedule = graph->default_schedule;
1881	// Output styles.
1882	for (i = 0; i < graph->exec_info->rnum; i++)
1883	{
1884		node_id[i] = c;
1885		ccv_nnc_graph_exec_info_t* exec_info = (ccv_nnc_graph_exec_info_t*)ccv_array_get(graph->exec_info, i)((void*)(((char*)((graph->exec_info)->data)) + (size_t)
(graph->exec_info)->rsize * (size_t)(i)));
1886		if (CCV_NNC_GRAPH_REF(exec_info)((exec_info)->_heap_graph_ref ? (exec_info)->_heap_graph_ref
 : (exec_info)->_inline_graph_ref)[0])
1887			_ccv_nnc_graph_dot_sub_graphs(exec_info, recovery, graph->sub_graphs, flags, 1, out, &k, &c);
1888		else {
1889			_ccv_nnc_graph_dot_node(exec_info,
1890				schedule ? (i < schedule->exec_info_size ? schedule->exec_info + i : 0) : 0,
1891				c, graph->streams, recovery, flags, 0, out, &k);
1892			++c;
1893		}
1894	}
1895	// Output connections.
1896	for (i = 0; i < graph->exec_info->rnum; i++)
1897	{
1898		ccv_nnc_graph_exec_info_t* exec_info = (ccv_nnc_graph_exec_info_t*)ccv_array_get(graph->exec_info, i)((void*)(((char*)((graph->exec_info)->data)) + (size_t)
(graph->exec_info)->rsize * (size_t)(i)));
1899		if (exec_info->outgoings)
1900			for (j = 0; j < exec_info->outgoings->rnum; j++)
1901			{
1902				const int outgoing_idx = *(int*)ccv_array_get(exec_info->outgoings, j)((void*)(((char*)((exec_info->outgoings)->data)) + (size_t
)(exec_info->outgoings)->rsize * (size_t)(j)));
1903				const ccv_nnc_graph_exec_info_t* const outgoing_info = (ccv_nnc_graph_exec_info_t*)ccv_array_get(graph->exec_info, outgoing_idx)((void*)(((char*)((graph->exec_info)->data)) + (size_t)
(graph->exec_info)->rsize * (size_t)(outgoing_idx)));
1904				// If both are sub-graphs, have both tail and head specified.
1905				if (CCV_NNC_GRAPH_REF(exec_info)((exec_info)->_heap_graph_ref ? (exec_info)->_heap_graph_ref
 : (exec_info)->_inline_graph_ref)[0] && CCV_NNC_GRAPH_REF(outgoing_info)((outgoing_info)->_heap_graph_ref ? (outgoing_info)->_heap_graph_ref
 : (outgoing_info)->_inline_graph_ref)[0])
1906					fprintf(out, "node%d -> node%d [ltail=cluster%d,lhead=cluster%d];\n", node_id[i], node_id[outgoing_idx], node_id[i], node_id[outgoing_idx]);
1907				else if (CCV_NNC_GRAPH_REF(exec_info)((exec_info)->_heap_graph_ref ? (exec_info)->_heap_graph_ref
 : (exec_info)->_inline_graph_ref)[0] && !CCV_NNC_GRAPH_REF(outgoing_info)((outgoing_info)->_heap_graph_ref ? (outgoing_info)->_heap_graph_ref
 : (outgoing_info)->_inline_graph_ref)[0])
1908					fprintf(out, "node%d -> node%d [ltail=cluster%d];\n", node_id[i], node_id[outgoing_idx], node_id[i]);
1909				else if (!CCV_NNC_GRAPH_REF(exec_info)((exec_info)->_heap_graph_ref ? (exec_info)->_heap_graph_ref
 : (exec_info)->_inline_graph_ref)[0] && CCV_NNC_GRAPH_REF(outgoing_info)((outgoing_info)->_heap_graph_ref ? (outgoing_info)->_heap_graph_ref
 : (outgoing_info)->_inline_graph_ref)[0])
1910					fprintf(out, "node%d -> node%d [lhead=cluster%d];\n", node_id[i], node_id[outgoing_idx], node_id[outgoing_idx]);
1911				else
1912					fprintf(out, "node%d -> node%d;\n", node_id[i], node_id[outgoing_idx]);
1913			}
1914	}
1915	fputs("}\n", out);
1916	_ccv_nnc_graph_tensor_dot_recovery_free(recovery);
1917	ccfreefree(node_id);
1918}
1919 
1920void ccv_nnc_graph_autotune(ccv_nnc_graph_t* const graph, const size_t max_workspace_size, const int flags, const ccv_nnc_graph_exec_t* const sources, const int source_size, const ccv_nnc_graph_exec_t* const destinations, const int destination_size)
1921{
1922	// exec current node, for synchronous CPU execution, no stream unit.
1923	int i;
1924#define visitor(node, idx, ...) \
1925	do { \
1926		if (node->cmd.cmd == CCV_NNC_NOOP) \
1927			continue; \
1928		if (node->cmd.cmd == CCV_NNC_GRAPH_FORWARD || node->cmd.cmd == CCV_NNC_GRAPH_BACKWARD) \
1929			for (i = 0; i < node->graph_ref_size; i++) \
1930			{ \
1931				ccv_nnc_graph_t* sub_graph = *(ccv_nnc_graph_t**)ccv_array_get(graph->sub_graphs, CCV_NNC_GRAPH_REF(node)[i] - 1)((void*)(((char*)((graph->sub_graphs)->data)) + (size_t
)(graph->sub_graphs)->rsize * (size_t)(((node)->_heap_graph_ref
 ? (node)->_heap_graph_ref : (node)->_inline_graph_ref)
[i] - 1))); \
1932				ccv_nnc_graph_autotune(sub_graph, max_workspace_size, flags, 0, 0, 0, 0); \
1933			} \
1934		else { \
1935			/* Need to unwrap these tensors */ \
1936			for (i = 0; i < node->input_size + node->output_size; i++) \
1937				if (node->inputs[i] && CCV_IS_TENSOR_MULTIVIEW(node->inputs[i])((*(int*)(node->inputs[i])) & CCV_TENSOR_MULTIVIEW)) \
1938					node->inputs[i] = _ccv_nnc_any_tensor_from_tensor_multiview((ccv_nnc_tensor_multiview_t*)node->inputs[i]); \
1939			PRINT(CCV_CLI_VERBOSE, "%s [%d]: [%d] -> [%d]\n", ccv_nnc_cmd_name(node->cmd.cmd), idx, node->input_size, node->output_size)do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
 { printf("%s [%d]: [%d] -> [%d]\n", ccv_nnc_cmd_name(node
->cmd.cmd), idx, node->input_size, node->output_size
); fflush(stdout); } } while (0); \
1940			for (i = 0; i < node->input_size; i++) \
1941			{ \
1942				PRINT(CCV_CLI_VERBOSE, "|-> %d. %p (%p)", i + 1, node->inputs[i], (node->inputs[i] ? node->inputs[i]->data.u8 : 0))do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
 { printf("|-> %d. %p (%p)", i + 1, node->inputs[i], (node
->inputs[i] ? node->inputs[i]->data.u8 : 0)); fflush
(stdout); } } while (0); \
1943				if (node->inputs[i] && CCV_CLI_OUTPUT_LEVEL_IS(CCV_CLI_VERBOSE)(CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) \
1944					ccv_nnc_print_tensor_shape(node->inputs[i]); \
1945				PRINT(CCV_CLI_VERBOSE, "\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
 { printf("\n"); fflush(stdout); } } while (0); \
1946			} \
1947			for (i = 0; i < node->output_size; i++) \
1948			{ \
1949				PRINT(CCV_CLI_VERBOSE, "|<- %d. %p (%p)", i + 1, node->outputs[i], (node->outputs[i] ? node->outputs[i]->data.u8 : 0))do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
 { printf("|<- %d. %p (%p)", i + 1, node->outputs[i], (
node->outputs[i] ? node->outputs[i]->data.u8 : 0)); fflush
(stdout); } } while (0); \
1950				if (node->outputs[i] && CCV_CLI_OUTPUT_LEVEL_IS(CCV_CLI_VERBOSE)(CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) \
1951					ccv_nnc_print_tensor_shape(node->outputs[i]); \
1952				PRINT(CCV_CLI_VERBOSE, "\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
 { printf("\n"); fflush(stdout); } } while (0); \
1953			} \
1954			node->cmd = ccv_nnc_cmd_autotune(node->cmd, max_workspace_size, node->hint, flags, node->inputs, node->input_size, node->outputs, node->output_size, 0); \
1955		} \
1956	} while (0)
1957	const ccv_nnc_graph_exec_t* const graph_sources = sources ? sources : (graph->sources ? (ccv_nnc_graph_exec_t*)ccv_array_get(graph->sources, 0)((void*)(((char*)((graph->sources)->data)) + (size_t)(graph
->sources)->rsize * (size_t)(0))): 0);
1958	const int graph_source_size = source_size ? source_size : (graph->sources ? graph->sources->rnum : 0);
1959	const ccv_nnc_graph_exec_t* const graph_destinations = destinations ? destinations : (graph->destinations ? (ccv_nnc_graph_exec_t*)ccv_array_get(graph->destinations, 0)((void*)(((char*)((graph->destinations)->data)) + (size_t
)(graph->destinations)->rsize * (size_t)(0))) : 0);
1960	const int graph_destination_size = destination_size ? destination_size : (graph->destinations ? graph->destinations->rnum : 0);
1961	CCV_NNC_GRAPH_VISIT(graph, (ccv_nnc_graph_exec_info_t*)ccv_array_get(graph->exec_info, 0), graph->exec_info->rnum, graph_sources, graph_source_size, graph_destinations, graph_destination_size, 0, visitor)do { typedef struct { int8_t d; int8_t r; uint16_t c; int32_t
 edges; } ccv_nnc_incoming_t; int _i_, _j_; int _incoming_edges_
 = 0; for (_i_ = 0; _i_ < (graph->exec_info->rnum); _i_
++) _incoming_edges_ += (((ccv_nnc_graph_exec_info_t*)((void*
)(((char*)((graph->exec_info)->data)) + (size_t)(graph->
exec_info)->rsize * (size_t)(0))))[_i_].outgoings) ? ((ccv_nnc_graph_exec_info_t
*)((void*)(((char*)((graph->exec_info)->data)) + (size_t
)(graph->exec_info)->rsize * (size_t)(0))))[_i_].outgoings
->rnum : 0; const int _heap_mem_ = ((graph->exec_info->
rnum) + _incoming_edges_ > 1024); ccv_nnc_incoming_t* _incomings_
; if (_heap_mem_) _incomings_ = (ccv_nnc_incoming_t*)malloc(sizeof
(ccv_nnc_incoming_t) * (graph->exec_info->rnum) + sizeof
(int32_t) * ((graph->exec_info->rnum) * 2 + _incoming_edges_
)); else _incomings_ = (ccv_nnc_incoming_t*)__builtin_alloca (
sizeof(ccv_nnc_incoming_t) * (graph->exec_info->rnum) +
 sizeof(int32_t) * ((graph->exec_info->rnum) * 2 + _incoming_edges_
)); memset(_incomings_, 0, sizeof(ccv_nnc_incoming_t) * (graph
->exec_info->rnum)); int32_t* _exists_[2] = { (int32_t*
)(_incomings_ + (graph->exec_info->rnum)), (int32_t*)(_incomings_
 + (graph->exec_info->rnum)) + (graph->exec_info->
rnum), }; int32_t* const _edges_ = _exists_[1] + (graph->exec_info
->rnum); for (_i_ = 0; _i_ < (graph_source_size); _i_++
) { ((void) sizeof (((graph_sources)[_i_].graph == graph) ? 1
 : 0), __extension__ ({ if ((graph_sources)[_i_].graph == graph
) ; else __assert_fail ("(graph_sources)[_i_].graph == graph"
, "ccv_nnc_graph.c", 1961, __extension__ __PRETTY_FUNCTION__)
; })); _incomings_[(graph_sources)[_i_].d].r = 1; _exists_[0]
[_i_] = (graph_sources)[_i_].d; } int _exist_size_[2] = { (graph_source_size
), 0, }; int _p_ = 0, _q_ = 1; while (_exist_size_[_p_] > 0
) { _exist_size_[_q_] = 0; for (_i_ = 0; _i_ < _exist_size_
[_p_]; _i_++) { const int32_t _idx_ = _exists_[_p_][_i_]; if (
_incomings_[_idx_].r != 1) continue; _incomings_[_idx_].r = 2
; if (((ccv_nnc_graph_exec_info_t*)((void*)(((char*)((graph->
exec_info)->data)) + (size_t)(graph->exec_info)->rsize
 * (size_t)(0))))[_idx_].outgoings) for (_j_ = 0; _j_ < ((
ccv_nnc_graph_exec_info_t*)((void*)(((char*)((graph->exec_info
)->data)) + (size_t)(graph->exec_info)->rsize * (size_t
)(0))))[_idx_].outgoings->rnum; _j_++) { const int d = *(int
*)((void*)(((char*)((((ccv_nnc_graph_exec_info_t*)((void*)(((
char*)((graph->exec_info)->data)) + (size_t)(graph->
exec_info)->rsize * (size_t)(0))))[_idx_].outgoings)->data
)) + (size_t)(((ccv_nnc_graph_exec_info_t*)((void*)(((char*)(
(graph->exec_info)->data)) + (size_t)(graph->exec_info
)->rsize * (size_t)(0))))[_idx_].outgoings)->rsize * (size_t
)(_j_))); ++_incomings_[d].c; if (_incomings_[d].r != 0) continue
; _incomings_[d].r = 1; ((void) sizeof ((_exist_size_[_q_] <
 (graph->exec_info->rnum)) ? 1 : 0), __extension__ ({ if
 (_exist_size_[_q_] < (graph->exec_info->rnum)) ; else
 __assert_fail ("_exist_size_[_q_] < (graph->exec_info->rnum)"
, "ccv_nnc_graph.c", 1961, __extension__ __PRETTY_FUNCTION__)
; })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[_q_
]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for (
_i_ = 0; _i_ < (graph_source_size); _i_++) { ((void) sizeof
 (((graph_sources)[_i_].graph == graph) ? 1 : 0), __extension__
 ({ if ((graph_sources)[_i_].graph == graph) ; else __assert_fail
 ("(graph_sources)[_i_].graph == graph", "ccv_nnc_graph.c", 1961
, __extension__ __PRETTY_FUNCTION__); })); _incomings_[(graph_sources
)[_i_].d].r = 3; _exists_[0][_i_] = (graph_sources)[_i_].d; }
 _exist_size_[0] = (graph_source_size); _exist_size_[1] = 0; _p_
 = 0, _q_ = 1; int _bump_ = 1; while (_exist_size_[_p_] > 0
) { _exist_size_[_q_] = 0; for (_i_ = 0; _i_ < _exist_size_
[_p_]; _i_++) { const int32_t _idx_ = _exists_[_p_][_i_]; if (
_incomings_[_idx_].r != 3) continue; _incomings_[_idx_].r = 4
; if (((ccv_nnc_graph_exec_info_t*)((void*)(((char*)((graph->
exec_info)->data)) + (size_t)(graph->exec_info)->rsize
 * (size_t)(0))))[_idx_].outgoings) for (_j_ = 0; _j_ < ((
ccv_nnc_graph_exec_info_t*)((void*)(((char*)((graph->exec_info
)->data)) + (size_t)(graph->exec_info)->rsize * (size_t
)(0))))[_idx_].outgoings->rnum; _j_++) { const int d = *(int
*)((void*)(((char*)((((ccv_nnc_graph_exec_info_t*)((void*)(((
char*)((graph->exec_info)->data)) + (size_t)(graph->
exec_info)->rsize * (size_t)(0))))[_idx_].outgoings)->data
)) + (size_t)(((ccv_nnc_graph_exec_info_t*)((void*)(((char*)(
(graph->exec_info)->data)) + (size_t)(graph->exec_info
)->rsize * (size_t)(0))))[_idx_].outgoings)->rsize * (size_t
)(_j_))); if (_incomings_[d].edges == 0) { _incomings_[d].edges
 = _bump_; _bump_ += _incomings_[d].c; _incomings_[d].c = 0; }
 _edges_[_incomings_[d].edges - 1 + _incomings_[d].c] = _idx_
; ++_incomings_[d].c; if (_incomings_[d].r != 2) continue; _incomings_
[d].r = 3; ((void) sizeof ((_exist_size_[_q_] < (graph->
exec_info->rnum)) ? 1 : 0), __extension__ ({ if (_exist_size_
[_q_] < (graph->exec_info->rnum)) ; else __assert_fail
 ("_exist_size_[_q_] < (graph->exec_info->rnum)", "ccv_nnc_graph.c"
, 1961, __extension__ __PRETTY_FUNCTION__); })); _exists_[_q_
][_exist_size_[_q_]] = d; ++_exist_size_[_q_]; } } ((_i_) = (
_p_), (_p_) = (_q_), (_q_) = (_i_)); } for (_i_ = 0; _i_ <
 (graph_destination_size); _i_++) { ((void) sizeof (((graph_destinations
)[_i_].graph == graph) ? 1 : 0), __extension__ ({ if ((graph_destinations
)[_i_].graph == graph) ; else __assert_fail ("(graph_destinations)[_i_].graph == graph"
, "ccv_nnc_graph.c", 1961, __extension__ __PRETTY_FUNCTION__)
; })); _incomings_[(graph_destinations)[_i_].d].r = 5; _exists_
[0][_i_] = (graph_destinations)[_i_].d; } _exist_size_[0] = (
graph_destination_size); _exist_size_[1] = 0; _p_ = 0, _q_ = 1
; while (_exist_size_[_p_] > 0) { _exist_size_[_q_] = 0; for
 (_i_ = 0; _i_ < _exist_size_[_p_]; _i_++) { const int32_t
 _idx_ = _exists_[_p_][_i_]; if (_incomings_[_idx_].r != 5) continue
; _incomings_[_idx_].r = 6; if (_incomings_[_idx_].edges >
 0) for (_j_ = 0; _j_ < _incomings_[_idx_].c; _j_++) { const
 int d = _edges_[_incomings_[_idx_].edges - 1 + _j_]; if (_incomings_
[d].r != 4) continue; _incomings_[d].r = 5; ((void) sizeof ((
_exist_size_[_q_] < (graph->exec_info->rnum)) ? 1 : 0
), __extension__ ({ if (_exist_size_[_q_] < (graph->exec_info
->rnum)) ; else __assert_fail ("_exist_size_[_q_] < (graph->exec_info->rnum)"
, "ccv_nnc_graph.c", 1961, __extension__ __PRETTY_FUNCTION__)
; })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[_q_
]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for (
_i_ = 0; _i_ < (graph_destination_size); _i_++) { ((void) sizeof
 (((graph_destinations)[_i_].graph == graph) ? 1 : 0), __extension__
 ({ if ((graph_destinations)[_i_].graph == graph) ; else __assert_fail
 ("(graph_destinations)[_i_].graph == graph", "ccv_nnc_graph.c"
, 1961, __extension__ __PRETTY_FUNCTION__); })); _incomings_[
(graph_destinations)[_i_].d].d = 1; } for (_i_ = 0; _i_ < (
graph_source_size); _i_++) { ((void) sizeof (((graph_sources)
[_i_].graph == graph) ? 1 : 0), __extension__ ({ if ((graph_sources
)[_i_].graph == graph) ; else __assert_fail ("(graph_sources)[_i_].graph == graph"
, "ccv_nnc_graph.c", 1961, __extension__ __PRETTY_FUNCTION__)
; })); _exists_[0][_i_] = (graph_sources)[_i_].d; } _p_ = 0; _q_
 = 1; _exist_size_[0] = (graph_source_size); _exist_size_[1] =
 0; int _d_ = 0; while (_exist_size_[_p_] > 0) { _exist_size_
[_q_] = 0; for (_i_ = 0; _i_ < _exist_size_[_p_];) { const
 int32_t _idx_ = _exists_[_p_][_i_]; visitor((((ccv_nnc_graph_exec_info_t
*)((void*)(((char*)((graph->exec_info)->data)) + (size_t
)(graph->exec_info)->rsize * (size_t)(0)))) + _idx_), (
_idx_), (_incomings_[_idx_].d)); if (_incomings_[_idx_].d) { ++
_d_; _incomings_[_idx_].r = 7; } if (((ccv_nnc_graph_exec_info_t
*)((void*)(((char*)((graph->exec_info)->data)) + (size_t
)(graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings
) { if (((ccv_nnc_graph_exec_info_t*)((void*)(((char*)((graph
->exec_info)->data)) + (size_t)(graph->exec_info)->
rsize * (size_t)(0))))[_idx_].outgoings->rnum == 1) { const
 int d = *(int*)((void*)(((char*)((((ccv_nnc_graph_exec_info_t
*)((void*)(((char*)((graph->exec_info)->data)) + (size_t
)(graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings
)->data)) + (size_t)(((ccv_nnc_graph_exec_info_t*)((void*)
(((char*)((graph->exec_info)->data)) + (size_t)(graph->
exec_info)->rsize * (size_t)(0))))[_idx_].outgoings)->rsize
 * (size_t)(0))); --_incomings_[d].c; if (_incomings_[d].c ==
 0 && _incomings_[d].r == 6 && _d_ < (graph_destination_size
)) { _exists_[_p_][_i_] = d; continue; } } else for (_j_ = 0;
 _j_ < ((ccv_nnc_graph_exec_info_t*)((void*)(((char*)((graph
->exec_info)->data)) + (size_t)(graph->exec_info)->
rsize * (size_t)(0))))[_idx_].outgoings->rnum; _j_++) { const
 int d = *(int*)((void*)(((char*)((((ccv_nnc_graph_exec_info_t
*)((void*)(((char*)((graph->exec_info)->data)) + (size_t
)(graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings
)->data)) + (size_t)(((ccv_nnc_graph_exec_info_t*)((void*)
(((char*)((graph->exec_info)->data)) + (size_t)(graph->
exec_info)->rsize * (size_t)(0))))[_idx_].outgoings)->rsize
 * (size_t)(_j_))); --_incomings_[d].c; if (_incomings_[d].c ==
 0 && _incomings_[d].r == 6 && _d_ < (graph_destination_size
)) { ((void) sizeof ((_exist_size_[_q_] < (graph->exec_info
->rnum)) ? 1 : 0), __extension__ ({ if (_exist_size_[_q_] <
 (graph->exec_info->rnum)) ; else __assert_fail ("_exist_size_[_q_] < (graph->exec_info->rnum)"
, "ccv_nnc_graph.c", 1961, __extension__ __PRETTY_FUNCTION__)
; })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[_q_
]; } } } ++_i_; } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_
)); } for (_i_ = 0; _i_ < (graph_destination_size); _i_++)
 { ((void) sizeof (((graph_destinations)[_i_].graph == graph)
 ? 1 : 0), __extension__ ({ if ((graph_destinations)[_i_].graph
 == graph) ; else __assert_fail ("(graph_destinations)[_i_].graph == graph"
, "ccv_nnc_graph.c", 1961, __extension__ __PRETTY_FUNCTION__)
; })); if (_incomings_[(graph_destinations)[_i_].d].r == 7) continue
; if (!(0)) { ((void) sizeof ((_incomings_[(graph_destinations
)[_i_].d].c == 0) ? 1 : 0), __extension__ ({ if (_incomings_[
(graph_destinations)[_i_].d].c == 0) ; else __assert_fail ("_incomings_[(graph_destinations)[_i_].d].c == 0"
, "ccv_nnc_graph.c", 1961, __extension__ __PRETTY_FUNCTION__)
; })); } else if (_incomings_[(graph_destinations)[_i_].d].c >
 0) continue; visitor((((ccv_nnc_graph_exec_info_t*)((void*)(
((char*)((graph->exec_info)->data)) + (size_t)(graph->
exec_info)->rsize * (size_t)(0)))) + (graph_destinations)[
_i_].d), ((graph_destinations)[_i_].d), (_incomings_[(graph_destinations
)[_i_].d].d)); } if (_heap_mem_) free(_incomings_); } while (
0);;
1962#undef visitor
1963}
1964 
1965void ccv_nnc_graph_free(ccv_nnc_graph_t* const graph)
1966{
1967	int i, j;
1968	for (i = 0; i < graph->exec_info->rnum; i++)
1969	{
1970		ccv_nnc_graph_exec_info_t *info = (ccv_nnc_graph_exec_info_t*)ccv_array_get(graph->exec_info, i)((void*)(((char*)((graph->exec_info)->data)) + (size_t)
(graph->exec_info)->rsize * (size_t)(i)));
1971		if (info->_heap_graph_ref)
1972			ccfreefree(info->_heap_graph_ref);
1973		ccv_array_t* outgoings = info->outgoings;
1974		if (outgoings)
1975			ccv_array_free(outgoings);
1976		// We allocate inputs & outputs in continuous fashion, therefore, only need to free the input array.
1977		if (info->inputs)
1978			ccfreefree(info->inputs);
1979		if (info->input_flags)
1980			ccfreefree(info->input_flags);
1981		if (info->updates)
1982			ccfreefree(info->updates);
1983		if ((info->flags & CCV_NNC_GRAPH_EXEC_P_WHILE) && info->p_while.inputs)
1984			ccfreefree(info->p_while.inputs);
1985	}
1986	if (graph->tensor_wraps)
1987	{
1988		for (i = 0; i < graph->tensor_wraps->rnum; i++)
1989		{
1990			ccv_nnc_graph_tensor_wrap_array_t* const tensor_wrap_array = *(ccv_nnc_graph_tensor_wrap_array_t**)ccv_array_get(graph->tensor_wraps, i)((void*)(((char*)((graph->tensor_wraps)->data)) + (size_t
)(graph->tensor_wraps)->rsize * (size_t)(i)));
1991			if (tensor_wrap_array)
1992			{
1993				for (j = 0; j < tensor_wrap_array->size; j++)
1994					_ccv_nnc_graph_tensor_wrap_free(tensor_wrap_array->tensor_wraps[j]);
1995				ccfreefree(tensor_wrap_array);
1996			}
1997		}
1998		ccv_array_free(graph->tensor_wraps);
1999	}
2000	if (graph->tensor_wraps_refs)
2001		ccv_array_free(graph->tensor_wraps_refs);
2002	if (graph->breakpoints)
2003		ccfreefree(graph->breakpoints);
2004	if (graph->sources)
2005		ccv_array_free(graph->sources);
2006	if (graph->destinations)
2007		ccv_array_free(graph->destinations);
2008	if (graph->default_schedule)
2009		ccv_nnc_graph_static_schedule_free(graph->default_schedule);
2010	if (graph->streams)
2011	{
2012		// If the graph has parent graph, the default stream is allocated by the parent graph, we need to skip.
2013		if (!graph->p)
2014			ccv_nnc_stream_context_free(graph->streams[0]);
2015		for (i = 1; i < graph->stream_size; i++)
2016			ccv_nnc_stream_context_free(graph->streams[i]);
2017		ccfreefree(graph->streams);
2018	}
2019	if (graph->block_stream_tasks)
2020		ccfreefree(graph->block_stream_tasks);
2021	if (graph->signals)
2022	{
2023		for (i = 0; i < graph->signal_size; i++)
2024			ccv_nnc_stream_signal_free(graph->signals[i]);
2025		ccfreefree(graph->signals);
2026	}
2027	if (graph->carry_overs)
2028	{
2029		for (i = 0; i < graph->carry_overs->rnum; i++)
2030		{
2031			ccv_nnc_graph_tensor_carry_over_t* const carry_over = (ccv_nnc_graph_tensor_carry_over_t*)ccv_array_get(graph->carry_overs, i)((void*)(((char*)((graph->carry_overs)->data)) + (size_t
)(graph->carry_overs)->rsize * (size_t)(i)));
2032			_ccv_nnc_graph_tensor_wrap_free(carry_over->from);
2033			_ccv_nnc_graph_tensor_wrap_free(carry_over->to);
2034		}
2035		ccv_array_free(graph->carry_overs);
2036	}
2037	if (graph->sub_graphs)
2038	{
2039		for (i = 0; i < graph->sub_graphs->rnum; i++)
2040			ccv_nnc_graph_free(*(ccv_nnc_graph_t**)ccv_array_get(graph->sub_graphs, i)((void*)(((char*)((graph->sub_graphs)->data)) + (size_t
)(graph->sub_graphs)->rsize * (size_t)(i))));
2041		ccv_array_free(graph->sub_graphs);
2042	}
2043	ccv_array_free(graph->exec_info);
2044	if (graph->buffer)
2045		ccfreefree(graph->buffer);
2046	ccfreefree(graph);
2047}