ccv_cnnp_model.c

Bug Summary

File:	nnc/ccv_cnnp_model.c
Warning:	line 2442, column 1 Assigned value is garbage or undefined
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_cnnp_model.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/2024-11-19-223635-166075-1 -x c ccv_cnnp_model.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_cnnp_model.h"
6#include "_ccv_nnc_graph.h"
7 
8// MARK - Level-5 API
9 
10ccv_cnnp_model_io_t ccv_cnnp_model_apply(ccv_cnnp_model_t* const model, const ccv_cnnp_model_io_t* const inputs, const int input_size)
11{
12	if (!model->io)
13		model->io = ccv_array_new(sizeof(ccv_cnnp_model_io_t), 1, 0);
14	ccv_cnnp_model_io_t model_io = ccmallocmalloc(sizeof(struct ccv_cnnp_model_io_s) + sizeof(ccv_nnc_tensor_symbol_t) * model->output_size);
15	model_io->param_ref = 0;
16	model_io->param_sel = 0;
17	model_io->visit = 0;
18	model_io->model = model;
19	model_io->dependencies = 0;
20	model_io->dependents = 0;
21	model_io->outgoings = 0;
22	model_io->outputs = (ccv_nnc_tensor_symbol_t*)(model_io + 1);
23	ccv_array_push(model->io, &model_io);
24	if (input_size > 0)
25	{
26		model_io->incomings = ccv_array_new(sizeof(ccv_cnnp_model_io_t), input_size, 0);
27		ccv_array_resize(model_io->incomings, input_size);
28		int i;
29		memcpy(ccv_array_get(model_io->incomings, 0)((void*)(((char*)((model_io->incomings)->data)) + (size_t
)(model_io->incomings)->rsize * (size_t)(0))), inputs, sizeof(ccv_cnnp_model_io_t) * input_size);
30		for (i = 0; i < input_size; i++)
31		{
32			if (!inputs[i]->outgoings)
33				inputs[i]->outgoings = ccv_array_new(sizeof(ccv_cnnp_model_io_t), 1, 0);
34			ccv_array_push(inputs[i]->outgoings, &model_io);
35		}
36	} else {
37		model_io->incomings = 0;
38	}
39	return model_io;
40}
41 
42void ccv_cnnp_model_add_dependencies(ccv_cnnp_model_io_t model_io, const ccv_cnnp_model_io_t* const dependencies, const int dependency_size)
43{
44	assert(dependency_size > 0)((void) sizeof ((dependency_size > 0) ? 1 : 0), __extension__
 ({ if (dependency_size > 0) ; else __assert_fail ("dependency_size > 0"
, "ccv_cnnp_model.c", 44, __extension__ __PRETTY_FUNCTION__);
 }));
45	if (!model_io->dependencies)
46		model_io->dependencies = ccv_array_new(sizeof(ccv_cnnp_model_io_t), dependency_size, 0);
47	int i, j;
48	for (i = 0; i < dependency_size; i++)
49	{
50		int flag = 0;
51		// Check if it is already exist or not.
52		for (j = 0; !flag && j < model_io->dependencies->rnum; j++)
53			if (*(ccv_cnnp_model_io_t*)ccv_array_get(model_io->dependencies, j)((void*)(((char*)((model_io->dependencies)->data)) + (size_t
)(model_io->dependencies)->rsize * (size_t)(j))) == dependencies[i])
54				flag = 1;
55		if (flag)
56			continue;
57		ccv_array_push(model_io->dependencies, dependencies + i);
58		++dependencies[i]->dependents;
59	}
60}
61 
62int ccv_cnnp_model_output_size(const ccv_cnnp_model_t* const model)
63{
64	return model->output_size;
65}
66 
67int ccv_cnnp_model_is_trainable(const ccv_cnnp_model_t* const model)
68{
69	// If the model is compiled, it is default to 1 unless it is not.
70	if (model->compiled_data)
71		return model->is_trainable >= 0 ? model->is_trainable : 1;
72	return model->is_trainable;
73}
74 
75ccv_cnnp_model_io_t ccv_cnnp_model_parameters(ccv_cnnp_model_t* const model, const int selector, const int index)
76{
77	if (!model->io)
78		model->io = ccv_array_new(sizeof(ccv_cnnp_model_io_t), 1, 0);
79	ccv_cnnp_model_io_t model_io = ccmallocmalloc(sizeof(struct ccv_cnnp_model_io_s));
80	model_io->param_ref = index >= 0 ? index + 1 : ALL_PARAMETERS-1;
81	model_io->param_sel = selector >= 0 ? selector + 1 : ALL_PARAMETERS-1;
82	model_io->visit = 0;
83	model_io->model = model;
84	model_io->outputs = 0;
85	model_io->dependencies = 0;
86	model_io->dependents = 0;
87	model_io->incomings = 0;
88	model_io->outgoings = 0;
89	ccv_array_push(model->io, &model_io);
90	return model_io;
91}
92 
93void ccv_cnnp_model_notify_hook(ccv_cnnp_model_t* const model, ccv_cnnp_model_notify_f func, void* const context)
94{
95	model->notify_hook.func = func;
96	model->notify_hook.context = context;
97}
98 
99void ccv_cnnp_model_notify(const ccv_cnnp_model_t* const model, const int tag, void* const payload)
100{
101	if (model->notify_hook.func)
102		model->notify_hook.func(model, tag, payload, model->notify_hook.context);
103	if (model->isa->notify)
104		model->isa->notify(model, tag, payload);
105}
106 
107static int _ccv_nnc_array_dedup_graph_exec_symbols(ccv_nnc_graph_exec_symbol_t* const graph_exec_symbols, int graph_exec_symbol_size)
108{
109	int i, j;
110	for (i = 0; i < graph_exec_symbol_size; i++)
111	{
112		ccv_nnc_graph_exec_symbol_t* const graph_exec_symbol = graph_exec_symbols + i;
113		// Check whether this tensor symbol has any duplicate.
114		for (j = i + 1; j < graph_exec_symbol_size;)
115		{
116			ccv_nnc_graph_exec_symbol_t* const other_symbol = graph_exec_symbols + j;
117			// If there is a same tensor symbol, remove it.
118			if (other_symbol->d == graph_exec_symbol->d && other_symbol->graph == graph_exec_symbol->graph)
119			{
120				if (j + 1 < graph_exec_symbol_size)
121					*other_symbol = graph_exec_symbols[graph_exec_symbol_size - 1];
122				--graph_exec_symbol_size;
123				continue;
124			}
125			++j;
126		}
127	}
128	return graph_exec_symbol_size;
129}
130 
131void ccv_cnnp_model_add_to_array(void* const context, const ccv_nnc_tensor_symbol_t symbol, const int is_trainable)
132{
133	ccv_cnnp_model_add_to_array_context_t* const add_to_array_context = (ccv_cnnp_model_add_to_array_context_t*)context;
134	ccv_cnnp_model_t* const model = add_to_array_context->sequence->model;
135	int i;
136	if (add_to_array_context->add_parameter_indices && !model->parameter_indices)
137		model->parameter_indices = ccv_array_new(sizeof(int), 0, 0);
138	for (i = 0; i < add_to_array_context->symbols->rnum; i++)
139	{
140		const ccv_nnc_tensor_symbol_t other_symbol = *(ccv_nnc_tensor_symbol_t*)ccv_array_get(add_to_array_context->symbols, i)((void*)(((char*)((add_to_array_context->symbols)->data
)) + (size_t)(add_to_array_context->symbols)->rsize * (
size_t)(i)));
141		if (other_symbol.d == symbol.d && other_symbol.graph == symbol.graph)
142		{
143			// Only add to parameter_indices if it is trainable.
144			if (add_to_array_context->add_parameter_indices)
145				ccv_array_add_unique_int(model->parameter_indices, i);
146			// Found it, return, don't add it.
147			return;
148		}
149	}
150	// Only add to parameter_indices if it is trainable.
151	if (add_to_array_context->add_parameter_indices)
152		ccv_array_push(model->parameter_indices, &add_to_array_context->symbols->rnum);
153	// This is a new one, no need to add_unique_int, it is unique.
154	ccv_array_push(add_to_array_context->symbols, &symbol);
155	if (add_to_array_context->trainables)
156		ccv_array_push(add_to_array_context->trainables, &is_trainable);
157	char id[2048];
158	id[0] = add_to_array_context->prefix;
159	id[1] = '-';
160	int total_len = 2;
161	for (i = 0; i < add_to_array_context->sequence->sequences->rnum; i++)
162	{
163		const ccv_cnnp_model_name_t* const name = (ccv_cnnp_model_name_t*)ccv_array_get(add_to_array_context->sequence->sequences, i)((void*)(((char*)((add_to_array_context->sequence->sequences
)->data)) + (size_t)(add_to_array_context->sequence->
sequences)->rsize * (size_t)(i)));
164		int len;
165		if (name->name && name->name[0] != '\0')
166			len = snprintf(id + total_len, 2048 - total_len, "%s-%d-", name->name, name->sequence);
167		else
168			len = snprintf(id + total_len, 2048 - total_len, "%d-", name->sequence);
169		total_len += len;
170		if (total_len >= 2047)
171			break;
172	}
173	if (total_len < 2047)
174		total_len += snprintf(id + total_len, 2048 - total_len, "%d", add_to_array_context->sequence->it);
175	assert(total_len < 2048)((void) sizeof ((total_len < 2048) ? 1 : 0), __extension__
 ({ if (total_len < 2048) ; else __assert_fail ("total_len < 2048"
, "ccv_cnnp_model.c", 175, __extension__ __PRETTY_FUNCTION__)
; }));
176	char *heap_id = (char*)ccmallocmalloc(total_len + 1);
177	memcpy(heap_id, id, total_len + 1);
178	ccv_array_push(add_to_array_context->ids, &heap_id);
179	++add_to_array_context->sequence->it;
180}
181 
182static void _ccv_cnnp_compiled_data_init(ccv_cnnp_compiled_data_t* const compiled_data, const int output_size, ccv_array_t* const gradient_checkpoints)
183{
184	compiled_data->f = compiled_data->fits + output_size;
185	compiled_data->xpu_alloc.mp_hdr = -1;
186	compiled_data->xpu_alloc.freed = kh_init(dy_str)kh_init_dy_str();
187	compiled_data->xpu_alloc.allocd = kh_init(dy_alloc)kh_init_dy_alloc();
188	compiled_data->gradient_checkpoints = gradient_checkpoints;
189}
190 
191static void _ccv_cnnp_model_compile(ccv_cnnp_model_t* const model, const ccv_nnc_tensor_param_t* const inputs, const int input_size, const ccv_nnc_cmd_t loss)
192{
193	assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if
 (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c"
, 193, __extension__ __PRETTY_FUNCTION__); }));
194	model->inputs = ccmallocmalloc(sizeof(ccv_nnc_tensor_symbol_t) * input_size);
195	int i;
196	for (i = 0; i < input_size; i++)
197		model->inputs[i] = ccv_nnc_tensor_symbol_new(model->graph, inputs[i], 0);
198	ccv_array_t* const parameters = ccv_array_new(sizeof(ccv_nnc_tensor_symbol_t), 0, 0);
199	ccv_array_t* const parameter_ids = ccv_array_new(sizeof(char*), 0, 0);
200	ccv_array_t* const parameter_trainables = ccv_array_new(sizeof(int), 0, 0);
201	ccv_cnnp_model_sequence_t model_sequence = {
202		.bank = kh_init(ccv_cnnp_model_name_bank)kh_init_ccv_cnnp_model_name_bank()
203	};
204	ccv_cnnp_model_add_to_array_context_t add_to_parameter_context = {
205		.add_parameter_indices = 1,
206		.prefix = 't',
207		.sequence = &model_sequence,
208		.symbols = parameters,
209		.ids = parameter_ids,
210		.trainables = parameter_trainables,
211	};
212	ccv_array_t* const internals = ccv_array_new(sizeof(ccv_nnc_tensor_symbol_t), 0, 0);
213	ccv_array_t* const internal_ids = ccv_array_new(sizeof(char*), 0, 0);
214	ccv_cnnp_model_add_to_array_context_t add_to_output_context = {
215		.add_parameter_indices = 0,
216		.prefix = 'r',
217		.sequence = &model_sequence,
218		.symbols = internals,
219		.ids = internal_ids,
220		.trainables = 0,
221	};
222	ccv_cnnp_model_build_data_t build_data = {
223		.is_trainable = model->is_trainable >= 0 ? model->is_trainable : 1,
224		.model_sequence = &model_sequence,
225		.add_to_array = ccv_cnnp_model_add_to_array,
226		.parameters = parameters,
227		.context = {
228			.add_to_parameter = &add_to_parameter_context,
229			.add_to_output = &add_to_output_context,
230		},
231		.gradient_checkpoints = 0,
232	};
233	model->data = &build_data;
234	ccv_cnnp_model_build(model, model->graph, model->inputs, input_size, 0, 0);
235	for (i = 0; i < model->output_size; i++)
236	{
237		const ccv_nnc_tensor_symbol_t output = model->outputs[i];
238		const ccv_nnc_tensor_symbol_t alias_to = ccv_nnc_tensor_symbol_alias_to(model->graph, output);
239		if (alias_to.d == CCV_NNC_NO_TENSOR_SYMBOL)
240			continue;
241		// If output is an alias, insert data transform regardless for result correctness (we cannot bind an alias). You can check ccv_nnc_tensor_bind_symbol method
242		// to see that we can correctly bind a tensor which from it, has aliases, but we cannot bind an alias tensor correctly (this is expected, sort of, to be
243		// honest, because we cannot handle cases of alias is part of the original tensor but bind differently).
244		const ccv_nnc_tensor_param_t output_params = ccv_nnc_tensor_symbol_params(model->graph, output);
245		model->outputs[i] = ccv_nnc_tensor_symbol_new(model->graph, output_params, 0);
246		ccv_nnc_graph_exec_symbol_t make_contiguous = ccv_nnc_graph_exec_symbol_new(model->graph, CMD_FORMAT_TRANSFORM_FORWARD()ccv_nnc_cmd(CCV_NNC_FORMAT_TRANSFORM_FORWARD, 0, ccv_nnc_cmd_auto
, 0), &output, 1, model->outputs + i, 1, "contiguous");
247		ccv_nnc_graph_exec_symbol_set_flags(model->graph, make_contiguous, CCV_NNC_GRAPH_EXEC_DISABLE_OPT);
248	}
249	model->data = 0;
250	kh_destroy(ccv_cnnp_model_name_bank, model_sequence.bank)kh_destroy_ccv_cnnp_model_name_bank(model_sequence.bank);
251	if (model_sequence.sequences)
252		ccv_array_free(model_sequence.sequences);
253	// Check if there are parameters that are not trainables. If there are, we will allocate uint64 bitmap to record that.
254	int not_trainables = 0;
255	// Assert no parameter is alias.
256	for (i = 0; i < parameters->rnum; i++)
257	{
258		const ccv_nnc_tensor_symbol_t parameter = *(ccv_nnc_tensor_symbol_t*)ccv_array_get(parameters, i)((void*)(((char*)((parameters)->data)) + (size_t)(parameters
)->rsize * (size_t)(i)));
259		const ccv_nnc_tensor_symbol_t alias_to = ccv_nnc_tensor_symbol_alias_to(parameter.graph, parameter);
260		assert(alias_to.graph == 0)((void) sizeof ((alias_to.graph == 0) ? 1 : 0), __extension__
 ({ if (alias_to.graph == 0) ; else __assert_fail ("alias_to.graph == 0"
, "ccv_cnnp_model.c", 260, __extension__ __PRETTY_FUNCTION__)
; })); // Cannot find the one alias to.
261		if (*(int*)ccv_array_get(parameter_trainables, i)((void*)(((char*)((parameter_trainables)->data)) + (size_t
)(parameter_trainables)->rsize * (size_t)(i))) == 0)
262			not_trainables = 1;
263	}
264	assert(parameters->rnum == parameter_trainables->rnum)((void) sizeof ((parameters->rnum == parameter_trainables->
rnum) ? 1 : 0), __extension__ ({ if (parameters->rnum == parameter_trainables
->rnum) ; else __assert_fail ("parameters->rnum == parameter_trainables->rnum"
, "ccv_cnnp_model.c", 264, __extension__ __PRETTY_FUNCTION__)
; }));
265	uint64_t* parameter_flags = 0;
266	if (not_trainables)
267	{
268		parameter_flags = (uint64_t*)cccalloccalloc(((parameters->rnum + 63) >> 6), sizeof(uint64_t));
269		for (i = 0; i < parameter_trainables->rnum; i++)
270			if (*(int*)ccv_array_get(parameter_trainables, i)((void*)(((char*)((parameter_trainables)->data)) + (size_t
)(parameter_trainables)->rsize * (size_t)(i))))
271				parameter_flags[i >> 6] |= ((uint64_t)1 << (i & 63));
272	}
273	ccv_array_free(parameter_trainables);
274	// Assert no internal is alias.
275	for (i = 0; i < internals->rnum; i++)
276	{
277		const ccv_nnc_tensor_symbol_t internal = *(ccv_nnc_tensor_symbol_t*)ccv_array_get(internals, i)((void*)(((char*)((internals)->data)) + (size_t)(internals
)->rsize * (size_t)(i)));
278		const ccv_nnc_tensor_symbol_t alias_to = ccv_nnc_tensor_symbol_alias_to(internal.graph, internal);
279		assert(alias_to.graph == 0)((void) sizeof ((alias_to.graph == 0) ? 1 : 0), __extension__
 ({ if (alias_to.graph == 0) ; else __assert_fail ("alias_to.graph == 0"
, "ccv_cnnp_model.c", 279, __extension__ __PRETTY_FUNCTION__)
; })); // Cannot find the one alias to.
280	}
281	const int output_size = model->output_size;
282	ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
283	const int parameters_rnum = parameters->rnum;
284	if (input_size > 0)
285	{
286		ccv_array_resize(parameters, parameters_rnum + input_size);
287		memcpy(ccv_array_get(parameters, parameters_rnum)((void*)(((char*)((parameters)->data)) + (size_t)(parameters
)->rsize * (size_t)(parameters_rnum))), model->inputs, input_size * sizeof(ccv_nnc_tensor_symbol_t));
288	}
289	ccv_nnc_symbolic_graph_simplify(model->graph,
290		SYMBOLIC_GRAPH_PASSES(CCV_NNC_SIMPLIFY_COMMON_SUBEXPRESSION_ELIMINATION,(const int []){CCV_NNC_SIMPLIFY_COMMON_SUBEXPRESSION_ELIMINATION
, CCV_NNC_SIMPLIFY_DATA_TRANSFER_OPT, CCV_NNC_SIMPLIFY_OPS_FUSION
, CCV_NNC_SIMPLIFY_GRAPH_PRUNING}, (1 +1 +1 +1 +1 +0 +0 +0 +0
 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1)
291			CCV_NNC_SIMPLIFY_DATA_TRANSFER_OPT,(const int []){CCV_NNC_SIMPLIFY_COMMON_SUBEXPRESSION_ELIMINATION
, CCV_NNC_SIMPLIFY_DATA_TRANSFER_OPT, CCV_NNC_SIMPLIFY_OPS_FUSION
, CCV_NNC_SIMPLIFY_GRAPH_PRUNING}, (1 +1 +1 +1 +1 +0 +0 +0 +0
 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1)
292			CCV_NNC_SIMPLIFY_OPS_FUSION,(const int []){CCV_NNC_SIMPLIFY_COMMON_SUBEXPRESSION_ELIMINATION
, CCV_NNC_SIMPLIFY_DATA_TRANSFER_OPT, CCV_NNC_SIMPLIFY_OPS_FUSION
, CCV_NNC_SIMPLIFY_GRAPH_PRUNING}, (1 +1 +1 +1 +1 +0 +0 +0 +0
 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1)
293			CCV_NNC_SIMPLIFY_GRAPH_PRUNING)(const int []){CCV_NNC_SIMPLIFY_COMMON_SUBEXPRESSION_ELIMINATION
, CCV_NNC_SIMPLIFY_DATA_TRANSFER_OPT, CCV_NNC_SIMPLIFY_OPS_FUSION
, CCV_NNC_SIMPLIFY_GRAPH_PRUNING}, (1 +1 +1 +1 +1 +0 +0 +0 +0
 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1),
294		ccv_array_get(parameters, 0)((void*)(((char*)((parameters)->data)) + (size_t)(parameters
)->rsize * (size_t)(0))), parameters_rnum + input_size,
295		model->outputs, output_size,
296		SYMBOLIC_GRAPH_SOURCES(model->graph)ccv_nnc_symbolic_graph_sources(model->graph), ccv_nnc_symbolic_graph_source_size
(model->graph), SYMBOLIC_GRAPH_DESTINATIONS(model->graph)ccv_nnc_symbolic_graph_destinations(model->graph), ccv_nnc_symbolic_graph_destination_size
(model->graph));
297	ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
298	// Size it down.
299	parameters->rnum = parameters_rnum;
300	ccv_cnnp_compiled_data_t* compiled_data = model->compiled_data = cccalloccalloc(1, sizeof(ccv_cnnp_compiled_data_t) + sizeof(ccv_nnc_tensor_symbol_t) * (output_size * 2 - 1));
301	_ccv_cnnp_compiled_data_init(compiled_data, output_size, build_data.gradient_checkpoints);
302	const int evaluate_to_size = compiled_data->evaluate.to_size = ccv_nnc_symbolic_graph_destination_size(model->graph);
303	assert(evaluate_to_size > 0)((void) sizeof ((evaluate_to_size > 0) ? 1 : 0), __extension__
 ({ if (evaluate_to_size > 0) ; else __assert_fail ("evaluate_to_size > 0"
, "ccv_cnnp_model.c", 303, __extension__ __PRETTY_FUNCTION__)
; }));
304	compiled_data->evaluate.tos = ccmallocmalloc(sizeof(ccv_nnc_graph_exec_symbol_t) * evaluate_to_size);
305	memcpy(compiled_data->evaluate.tos, ccv_nnc_symbolic_graph_destinations(model->graph), sizeof(ccv_nnc_graph_exec_symbol_t) * evaluate_to_size);
306	compiled_data->loss = loss;
307	if (loss.cmd == CCV_NNC_NOOP)
308	{
309		// If no loss function provided, there is no fits.
310		for (i = 0; i < output_size; i++)
311		{
312			compiled_data->fits[i] = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
313			const ccv_nnc_tensor_symbol_t alias_to = ccv_nnc_tensor_symbol_alias_to(model->graph, model->outputs[i]);
314			if (alias_to.d < 0)
315				compiled_data->f[i] = model->outputs[i];
316			else { // We cannot differentiate against an alias, therefore, we have to verify this output is full, and we can diff against the original.
317				int ofs[CCV_NNC_MAX_DIM_ALLOC(12)];
318				int inc[CCV_NNC_MAX_DIM_ALLOC(12)];
319				ccv_nnc_tensor_symbol_alias_params(model->graph, model->outputs[i], ofs, inc);
320				int j;
321				for (j = 0; j < CCV_NNC_MAX_DIM_ALLOC(12); j++)
322					{ assert(ofs[j] == 0)((void) sizeof ((ofs[j] == 0) ? 1 : 0), __extension__ ({ if (
ofs[j] == 0) ; else __assert_fail ("ofs[j] == 0", "ccv_cnnp_model.c"
, 322, __extension__ __PRETTY_FUNCTION__); })); } // There is no ofs.
323				compiled_data->f[i] = alias_to; // Unfortunately, I cannot assert the size yet.
324			}
325		}
326	} else {
327		for (i = 0; i < output_size; i++)
328		{
329			const ccv_nnc_tensor_param_t info = ccv_nnc_tensor_symbol_params(model->graph, model->outputs[i]);
330			const ccv_nnc_tensor_symbol_t fit = compiled_data->fits[i] = ccv_nnc_tensor_symbol_new(model->graph, info, 0);
331			compiled_data->f[i] = ccv_nnc_tensor_symbol_new(model->graph, ccv_nnc_tensor_auto, 0);
332			ccv_nnc_graph_exec_symbol_new(model->graph, loss, TENSOR_SYMBOL_LIST(model->outputs[i], fit)(const ccv_nnc_tensor_symbol_t []){model->outputs[i], fit}
, (1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(compiled_data->f[i])(const ccv_nnc_tensor_symbol_t []){compiled_data->f[i]}, (
1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
 +0 +0 -1), 0);
333		}
334	}
335	ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
336	ccv_nnc_symbolic_graph_simplify(model->graph,
337		SYMBOLIC_GRAPH_PASSES(CCV_NNC_SIMPLIFY_OPS_FUSION)(const int []){CCV_NNC_SIMPLIFY_OPS_FUSION}, (1 +1 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), // Only do Ops fusion, in this way, we can fuse the loss function.
338		0, 0, // No need to provide binds at this point.
339		compiled_data->f, model->output_size,
340		SYMBOLIC_GRAPH_SOURCES(model->graph)ccv_nnc_symbolic_graph_sources(model->graph), ccv_nnc_symbolic_graph_source_size
(model->graph), SYMBOLIC_GRAPH_DESTINATIONS(model->graph)ccv_nnc_symbolic_graph_destinations(model->graph), ccv_nnc_symbolic_graph_destination_size
(model->graph));
341	ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
342	// If inputs are from GPU, stream type is GPU.
343	compiled_data->parameters = parameters;
344	compiled_data->parameter_flags = parameter_flags;
345	compiled_data->internals = internals;
346	compiled_data->ids.parameters = parameter_ids;
347	compiled_data->ids.internals = internal_ids;
348	ccv_cnnp_model_gradient_checkpoints_cleanup_after_build(compiled_data, model->graph);
349}
350 
351static void _ccv_cnnp_graph_push_graph_exec_symbol(void* context, const ccv_nnc_graph_exec_symbol_t symbol, const ccv_nnc_cmd_t cmd, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, const ccv_nnc_tensor_symbol_t* const outputs, const int output_size, const char* const name)
352{
353	ccv_array_t* const stack = (ccv_array_t*)context;
354	ccv_array_push(stack, &symbol.d);
355}
356 
357static void _ccv_nnc_tensor_symbol_reinit(const ccv_nnc_symbolic_graph_t* const src_graph, ccv_nnc_symbolic_graph_t* const dest_graph, const int src_index, const int dest_index)
358{
359	const ccv_nnc_tensor_symbol_t src_symbol = {
360		.d = src_index,
361		.graph = src_graph
362	};
363	const ccv_nnc_tensor_symbol_t dest_symbol = {
364		.d = dest_index,
365		.graph = dest_graph
366	};
367	const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(src_graph, src_symbol);
368	ccv_nnc_tensor_symbol_set(dest_graph, dest_symbol, params);
369	int ofs[CCV_NNC_MAX_DIM_ALLOC(12)];
370	int inc[CCV_NNC_MAX_DIM_ALLOC(12)];
371	if (0 == ccv_nnc_tensor_symbol_alias_params(src_graph, src_symbol, ofs, inc))
372		ccv_nnc_tensor_symbol_alias_set(dest_graph, dest_symbol, ofs, inc);
373}
374 
375static int _ccv_nnc_tensor_symbol_check_dim(const ccv_nnc_symbolic_graph_t* const src_graph, ccv_nnc_symbolic_graph_t* const dest_graph, const int src_index, const int dest_index)
376{
377	const ccv_nnc_tensor_symbol_t src_symbol = {
378		.d = src_index,
379		.graph = src_graph
380	};
381	const ccv_nnc_tensor_param_t src_params = ccv_nnc_tensor_symbol_params(src_graph, src_symbol);
382	const ccv_nnc_tensor_symbol_t dest_symbol = {
383		.d = dest_index,
384		.graph = dest_graph
385	};
386	const ccv_nnc_tensor_param_t dest_params = ccv_nnc_tensor_symbol_params(dest_graph, dest_symbol);
387	return memcmp(src_params.dim, dest_params.dim, sizeof(src_params.dim)) == 0;
388}
389 
390static void _ccv_cnnp_model_gradient_init(ccv_cnnp_model_t* const model, const int gradient_mode, const uint64_t disable_outgrad, ccv_nnc_tensor_t* const* const fits, const int fit_size);
391static void _ccv_cnnp_compiled_data_graph_free(ccv_cnnp_compiled_data_t* const compiled_data);
392 
393typedef struct {
394	int parallel_count;
395	ccv_nnc_symbolic_graph_t* graph;
396	ccv_nnc_graph_exec_arena_t* graph_exec_arena;
397} ccv_nnc_graph_exec_update_t;
398 
399static void _ccv_cnnp_cmd_update_for_execs(void* const context, const ccv_nnc_graph_exec_symbol_t symbol, const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint)
400{
401	ccv_nnc_graph_exec_update_t* const graph_exec_update = (ccv_nnc_graph_exec_update_t*)context;
402	ccv_nnc_graph_exec_arena_t* const graph_exec_arena = graph_exec_update->graph_exec_arena;
403	ccv_nnc_graph_exec_t graph_exec = ccv_nnc_graph_exec_from_symbol(graph_exec_arena, symbol);
404	ccv_nnc_graph_exec_set(graph_exec.graph, graph_exec, cmd);
405	ccv_nnc_graph_exec_set_hint(graph_exec.graph, graph_exec, hint);
406	const ccv_nnc_symbolic_graph_t* const graph = graph_exec_update->graph;
407	const int parallel_count = graph_exec_update->parallel_count;
408	int i;
409	for (i = 1; i < parallel_count; i++)
410	{
411		const ccv_nnc_graph_exec_t copy = ccv_nnc_graph_exec_from_symbol(graph_exec_arena, ccv_nnc_graph_exec_symbol_copy(graph, symbol, i));
412		if (!CCV_NO_GRAPH_EXEC(copy)((copy).graph == 0))
413		{
414			ccv_nnc_graph_exec_set(copy.graph, copy, cmd);
415			ccv_nnc_graph_exec_set_hint(copy.graph, copy, hint);
416		}
417	}
418}
419 
420void ccv_cnnp_model_absorb(ccv_cnnp_model_t* const model, ccv_cnnp_model_t* const init, const ccv_nnc_tensor_param_t* const inputs, const int input_size)
421{
422	assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if
 (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c"
, 422, __extension__ __PRETTY_FUNCTION__); }));
423	assert(model->compiled_data)((void) sizeof ((model->compiled_data) ? 1 : 0), __extension__
 ({ if (model->compiled_data) ; else __assert_fail ("model->compiled_data"
, "ccv_cnnp_model.c", 423, __extension__ __PRETTY_FUNCTION__)
; }));
424	assert(!init->graph)((void) sizeof ((!init->graph) ? 1 : 0), __extension__ ({ if
 (!init->graph) ; else __assert_fail ("!init->graph", "ccv_cnnp_model.c"
, 424, __extension__ __PRETTY_FUNCTION__); }));
425	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
426	init->graph = ccv_nnc_symbolic_graph_new();
427	ccv_array_t* const stack = ccv_array_new(sizeof(int), 0, 0);
428	ccv_nnc_graph_exec_symbol_new_hook(init->graph, _ccv_cnnp_graph_push_graph_exec_symbol, stack, 0);
429	_ccv_cnnp_model_compile(init, inputs, input_size, compiled_data->loss);
430	init->parallel_count = model->parallel_count;
431	init->memory_compression = model->memory_compression;
432	init->memory_reduction = model->memory_reduction;
433	init->gradient_checkpointing = model->gradient_checkpointing;
434	init->compiled_data->stream_type = model->compiled_data->stream_type;
435	init->compiled_data->minimize.minimizer = model->compiled_data->minimize.minimizer;
436	init->compiled_data->minimize.max_saved_aux_size = model->compiled_data->minimize.max_saved_aux_size;
437	if (model->compiled_data->gradient_mode != CCV_CNNP_COMPILED_DATA_GRADIENT_NONE)
438		_ccv_cnnp_model_gradient_init(init, model->compiled_data->gradient_mode, model->compiled_data->disable_outgrad, 0, 0);
439	ccv_nnc_graph_exec_symbol_new_hook(init->graph, 0, 0, 0);
440	ccv_nnc_symbolic_graph_tensor_auto(init->graph, TRAVERSE_FULL0,0,0,0);
441	int i, j;
442	// Verify parameters, internals and saved_aux in both graph has the same dimensionality.
443	for (i = 0; i < compiled_data->parameters->rnum; i++)
444	{
445		const int d = ((ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->parameters, i)((void*)(((char*)((compiled_data->parameters)->data)) +
 (size_t)(compiled_data->parameters)->rsize * (size_t)(
i))))->d;
446		assert(_ccv_nnc_tensor_symbol_check_dim(model->graph, init->graph, d, d))((void) sizeof ((_ccv_nnc_tensor_symbol_check_dim(model->graph
, init->graph, d, d)) ? 1 : 0), __extension__ ({ if (_ccv_nnc_tensor_symbol_check_dim
(model->graph, init->graph, d, d)) ; else __assert_fail
 ("_ccv_nnc_tensor_symbol_check_dim(model->graph, init->graph, d, d)"
, "ccv_cnnp_model.c", 446, __extension__ __PRETTY_FUNCTION__)
; }));
447	}
448	for (i = 0; i < compiled_data->internals->rnum; i++)
449	{
450		const int d = ((ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->internals, i)((void*)(((char*)((compiled_data->internals)->data)) + (
size_t)(compiled_data->internals)->rsize * (size_t)(i))
))->d;
451		assert(_ccv_nnc_tensor_symbol_check_dim(model->graph, init->graph, d, d))((void) sizeof ((_ccv_nnc_tensor_symbol_check_dim(model->graph
, init->graph, d, d)) ? 1 : 0), __extension__ ({ if (_ccv_nnc_tensor_symbol_check_dim
(model->graph, init->graph, d, d)) ; else __assert_fail
 ("_ccv_nnc_tensor_symbol_check_dim(model->graph, init->graph, d, d)"
, "ccv_cnnp_model.c", 451, __extension__ __PRETTY_FUNCTION__)
; }));
452	}
453	// Update inputs.
454	assert(model->input_size == init->input_size)((void) sizeof ((model->input_size == init->input_size)
 ? 1 : 0), __extension__ ({ if (model->input_size == init->
input_size) ; else __assert_fail ("model->input_size == init->input_size"
, "ccv_cnnp_model.c", 454, __extension__ __PRETTY_FUNCTION__)
; }));
455	for (i = 0; i < model->input_size; i++)
456		if (model->inputs[i].d >= 0)
457		{
458			assert(init->inputs[i].d >= 0)((void) sizeof ((init->inputs[i].d >= 0) ? 1 : 0), __extension__
 ({ if (init->inputs[i].d >= 0) ; else __assert_fail ("init->inputs[i].d >= 0"
, "ccv_cnnp_model.c", 458, __extension__ __PRETTY_FUNCTION__)
; }));
459			_ccv_nnc_tensor_symbol_reinit(init->graph, model->graph, init->inputs[i].d, model->inputs[i].d);
460		}
461	// Update outputs.
462	assert(model->output_size == init->output_size)((void) sizeof ((model->output_size == init->output_size
) ? 1 : 0), __extension__ ({ if (model->output_size == init
->output_size) ; else __assert_fail ("model->output_size == init->output_size"
, "ccv_cnnp_model.c", 462, __extension__ __PRETTY_FUNCTION__)
; }));
463	for (i = 0; i < model->output_size; i++)
464	{
465		if (model->outputs[i].d >= 0)
466		{
467			assert(init->outputs[i].d >= 0)((void) sizeof ((init->outputs[i].d >= 0) ? 1 : 0), __extension__
 ({ if (init->outputs[i].d >= 0) ; else __assert_fail (
"init->outputs[i].d >= 0", "ccv_cnnp_model.c", 467, __extension__
 __PRETTY_FUNCTION__); }));
468			_ccv_nnc_tensor_symbol_reinit(init->graph, model->graph, init->outputs[i].d, model->outputs[i].d);
469		}
470		if (model->outputs[i].d != model->compiled_data->f[i].d)
471		{
472			assert(init->outputs[i].d != init->compiled_data->f[i].d)((void) sizeof ((init->outputs[i].d != init->compiled_data
->f[i].d) ? 1 : 0), __extension__ ({ if (init->outputs[
i].d != init->compiled_data->f[i].d) ; else __assert_fail
 ("init->outputs[i].d != init->compiled_data->f[i].d"
, "ccv_cnnp_model.c", 472, __extension__ __PRETTY_FUNCTION__)
; }));
473			if (model->compiled_data->f[i].d >= 0)
474			{
475				assert(init->compiled_data->f[i].d >= 0)((void) sizeof ((init->compiled_data->f[i].d >= 0) ?
 1 : 0), __extension__ ({ if (init->compiled_data->f[i]
.d >= 0) ; else __assert_fail ("init->compiled_data->f[i].d >= 0"
, "ccv_cnnp_model.c", 475, __extension__ __PRETTY_FUNCTION__)
; }));
476				_ccv_nnc_tensor_symbol_reinit(init->graph, model->graph, init->compiled_data->f[i].d, model->compiled_data->f[i].d);
477			}
478		}
479	}
480	// Go through the graph to set tensor on matching symbols
481	for (i = 0; i < stack->rnum; i++)
482	{
483		const int d = *(int*)ccv_array_get(stack, i)((void*)(((char*)((stack)->data)) + (size_t)(stack)->rsize
 * (size_t)(i)));
484		// If exceed range, skip.
485		if (d >= ccv_nnc_graph_exec_symbol_count(init->graph) ||
486			d >= ccv_nnc_graph_exec_symbol_count(model->graph))
487			continue;
488		const ccv_nnc_graph_exec_symbol_t src_symbol = {
489			.d = d,
490			.graph = init->graph
491		};
492		const ccv_nnc_graph_exec_symbol_t dest_symbol = {
493			.d = d,
494			.graph = model->graph
495		};
496		const ccv_nnc_cmd_t src_cmd = ccv_nnc_graph_exec_symbol_cmd(init->graph, src_symbol);
497		const ccv_nnc_cmd_t dest_cmd = ccv_nnc_graph_exec_symbol_cmd(model->graph, dest_symbol);
498		// If the name doesn't match, skip.
499		if (dest_cmd.cmd != src_cmd.cmd && src_cmd.cmd != CCV_NNC_NOOP)
500			continue;
501		// Now get all the inputs and outputs, if matches, set them.
502		const int* src_inputs;
503		int src_input_size;
504		const int* src_outputs;
505		int src_output_size;
506		ccv_nnc_graph_exec_symbol_io(init->graph, src_symbol, &src_inputs, &src_input_size, &src_outputs, &src_output_size);
507		const int* dest_inputs;
508		int dest_input_size;
509		const int* dest_outputs;
510		int dest_output_size;
511		ccv_nnc_graph_exec_symbol_io(model->graph, dest_symbol, &dest_inputs, &dest_input_size, &dest_outputs, &dest_output_size);
512		// We may have unmatched input / output size because this is the minimizer and it has
513		// different saved_aux (for example, when we shrunk with CMD_NOOP).
514		if (src_input_size != dest_input_size)
515			continue;
516		if (src_output_size != dest_output_size)
517			continue;
518		ccv_nnc_graph_exec_symbol_set(model->graph, dest_symbol, src_cmd);
519		// There may be mismatches of the source tensor symbols and destination tensor symbols. The reason is because
520		// we may later passed-in the minimizer, therefore, we may allocate tensors for minimizer later in the original
521		// graph whereas in the newly created graph, it is streamlined (the minimizer exists from the beginning). That
522		// will make the order of tensor symbols creation different, therefore, exact which tensor is which wrong as
523		// well. However, set a new minimizer won't change the exec symbol ordering, because we never create new exec
524		// symbols after gradient init step. Changing a new minimizer just updated that exec symbols setting, it is not
525		// a new exec symbol.
526		for (j = 0; j < src_input_size; j++)
527			if (src_inputs[j] >= 0)
528				_ccv_nnc_tensor_symbol_reinit(init->graph, model->graph, src_inputs[j], dest_inputs[j]);
529		for (j = 0; j < src_output_size; j++)
530			if (src_outputs[j] >= 0)
531				_ccv_nnc_tensor_symbol_reinit(init->graph, model->graph, src_outputs[j], dest_outputs[j]);
532	}
533	ccv_array_free(stack);
534	// After this, we get all tensors in the model graph resolved through tensor_auto.
535	ccv_nnc_symbolic_graph_tensor_auto(model->graph, TRAVERSE_FULL0,0,0,0);
536	// Verify symbols we get matches.
537	const int parameter_size = compiled_data->parameters->rnum;
538	for (i = 0; i < parameter_size; i++)
539		{ assert(((ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->parameters, i))->d == ((ccv_nnc_tensor_symbol_t*)ccv_array_get(init->compiled_data->parameters, i))->d)((void) sizeof ((((ccv_nnc_tensor_symbol_t*)((void*)(((char*)
((compiled_data->parameters)->data)) + (size_t)(compiled_data
->parameters)->rsize * (size_t)(i))))->d == ((ccv_nnc_tensor_symbol_t
*)((void*)(((char*)((init->compiled_data->parameters)->
data)) + (size_t)(init->compiled_data->parameters)->
rsize * (size_t)(i))))->d) ? 1 : 0), __extension__ ({ if (
((ccv_nnc_tensor_symbol_t*)((void*)(((char*)((compiled_data->
parameters)->data)) + (size_t)(compiled_data->parameters
)->rsize * (size_t)(i))))->d == ((ccv_nnc_tensor_symbol_t
*)((void*)(((char*)((init->compiled_data->parameters)->
data)) + (size_t)(init->compiled_data->parameters)->
rsize * (size_t)(i))))->d) ; else __assert_fail ("((ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->parameters, i))->d == ((ccv_nnc_tensor_symbol_t*)ccv_array_get(init->compiled_data->parameters, i))->d"
, "ccv_cnnp_model.c", 539, __extension__ __PRETTY_FUNCTION__)
; })); }
540	const int internal_size = compiled_data->internals->rnum;
541	for (i = 0; i < internal_size; i++)
542		{ assert(((ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->internals, i))->d == ((ccv_nnc_tensor_symbol_t*)ccv_array_get(init->compiled_data->internals, i))->d)((void) sizeof ((((ccv_nnc_tensor_symbol_t*)((void*)(((char*)
((compiled_data->internals)->data)) + (size_t)(compiled_data
->internals)->rsize * (size_t)(i))))->d == ((ccv_nnc_tensor_symbol_t
*)((void*)(((char*)((init->compiled_data->internals)->
data)) + (size_t)(init->compiled_data->internals)->rsize
 * (size_t)(i))))->d) ? 1 : 0), __extension__ ({ if (((ccv_nnc_tensor_symbol_t
*)((void*)(((char*)((compiled_data->internals)->data)) +
 (size_t)(compiled_data->internals)->rsize * (size_t)(i
))))->d == ((ccv_nnc_tensor_symbol_t*)((void*)(((char*)((init
->compiled_data->internals)->data)) + (size_t)(init->
compiled_data->internals)->rsize * (size_t)(i))))->d
) ; else __assert_fail ("((ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->internals, i))->d == ((ccv_nnc_tensor_symbol_t*)ccv_array_get(init->compiled_data->internals, i))->d"
, "ccv_cnnp_model.c", 542, __extension__ __PRETTY_FUNCTION__)
; })); }
543	// Go through compiled data.
544	if (compiled_data->tensor_arena)
545	{
546		const int flag = ccv_nnc_tensor_arena_reinit(compiled_data->tensor_arena, model->graph);
547		if (flag == 0 && compiled_data->graph_exec_arena)
548		{
549			ccv_nnc_graph_exec_reinit(compiled_data->graph_exec_arena, compiled_data->graph, model->graph);
550			// Since we will reinit, if we previously set is_test, we need to set it again.
551			if (compiled_data->is_test)
552			{
553				const int parallel_count = ccv_max(model->parallel_count, 1)({ typeof (model->parallel_count) _a = (model->parallel_count
); typeof (1) _b = (1); (_a > _b) ? _a : _b; });
554				ccv_nnc_graph_exec_update_t update = {
555					.parallel_count = parallel_count,
556					.graph = model->graph,
557					.graph_exec_arena = compiled_data->graph_exec_arena,
558				};
559				ccv_cnnp_model_set_is_test(model, 1, _ccv_cnnp_cmd_update_for_execs, &update);
560			}
561		} else
562			// Free-up tensor arena & graph exec arena.
563			_ccv_cnnp_compiled_data_graph_free(compiled_data);
564	}
565	// There are other compiled graphs, for accum and apply gradients.
566	// However, the main conclusion is, these absorb operations shouldn't impact parameters.
567	// Thus, it won't impact the shape of gradients (only outgrad). Since for outgrad, we
568	// don't allocate ourselves, it is not a concern. For normal gradients, the shape cannot
569	// be changed otherwise parameters' shape will be meaningless. The same goes to internals.
570	// That is why we don't update these compiled graphs at all this point.
571	// Free the model, we've already "absorbed" it.
572	ccv_cnnp_model_free(init);
573}
574 
575void ccv_cnnp_model_compile(ccv_cnnp_model_t* const model, const ccv_nnc_tensor_param_t* const inputs, const int input_size, const ccv_nnc_cmd_t minimizer, const ccv_nnc_cmd_t loss)
576{
577	assert(input_size == model->input_size || model->input_size == 0)((void) sizeof ((input_size == model->input_size || model->
input_size == 0) ? 1 : 0), __extension__ ({ if (input_size ==
 model->input_size || model->input_size == 0) ; else __assert_fail
 ("input_size == model->input_size || model->input_size == 0"
, "ccv_cnnp_model.c", 577, __extension__ __PRETTY_FUNCTION__)
; }));
578	if (model->input_size == 0)
579		model->input_size = input_size;
580	if (!model->graph) // The graph is not compiled yet.
581	{
582		model->graph = ccv_nnc_symbolic_graph_new();
583		_ccv_cnnp_model_compile(model, inputs, input_size, loss);
584		assert(model->compiled_data)((void) sizeof ((model->compiled_data) ? 1 : 0), __extension__
 ({ if (model->compiled_data) ; else __assert_fail ("model->compiled_data"
, "ccv_cnnp_model.c", 584, __extension__ __PRETTY_FUNCTION__)
; }));
585		int i, flag = 0;
586		for (i = 0; !flag && i < input_size; i++)
587			flag = (CCV_TENSOR_GET_MEMORY(inputs[i].type)((inputs[i].type) & 0x3) == CCV_TENSOR_GPU_MEMORY);
588		// If inputs are from GPU, stream type is GPU.
589		model->compiled_data->stream_type = flag ? CCV_STREAM_CONTEXT_GPU : CCV_STREAM_CONTEXT_CPU;
590		model->compiled_data->minimize.minimizer = minimizer;
591		model->compiled_data->minimize.max_saved_aux_size = ccv_nnc_minimizer_saved_aux_size(minimizer);
592	} else {
593		// Now, finally fill in this part. If the graph is already compiled, we make a copy of the model.
594		// And then absorb the "new model" to the old one.
595		ccv_cnnp_model_t* const init = ccv_cnnp_model_copy(model, model->is_trainable);
596		ccv_cnnp_model_absorb(model, init, inputs, input_size);
597		// Reset minimizer.
598		ccv_cnnp_model_set_minimizer(model, minimizer, 1, 0, 0);
599	}
600}
601 
602ccv_cnnp_model_t* ccv_cnnp_model_copy(const ccv_cnnp_model_t* const model, const int is_trainable)
603{
604	ccv_cnnp_model_t* const new_model = _ccv_cnnp_model_copy(model, 0);
605	new_model->is_trainable = is_trainable;
606	return new_model;
607}
608 
609void ccv_cnnp_model_tensor_auto(ccv_cnnp_model_t* const model, ccv_nnc_tensor_param_t* const outputs, const int output_size)
610{
611	assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if
 (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c"
, 611, __extension__ __PRETTY_FUNCTION__); }));
612	assert(output_size == model->output_size)((void) sizeof ((output_size == model->output_size) ? 1 : 0
), __extension__ ({ if (output_size == model->output_size)
 ; else __assert_fail ("output_size == model->output_size"
, "ccv_cnnp_model.c", 612, __extension__ __PRETTY_FUNCTION__)
; }));
613	ccv_nnc_symbolic_graph_t* const graph = model->graph;
614	ccv_nnc_symbolic_graph_tensor_auto(graph, TRAVERSE_FULL0,0,0,0);
615	int i;
616	for (i = 0; i < output_size; i++)
617	{
618		assert(model->outputs[i].d != CCV_NNC_NO_TENSOR_SYMBOL)((void) sizeof ((model->outputs[i].d != CCV_NNC_NO_TENSOR_SYMBOL
) ? 1 : 0), __extension__ ({ if (model->outputs[i].d != CCV_NNC_NO_TENSOR_SYMBOL
) ; else __assert_fail ("model->outputs[i].d != CCV_NNC_NO_TENSOR_SYMBOL"
, "ccv_cnnp_model.c", 618, __extension__ __PRETTY_FUNCTION__)
; }));
619		outputs[i] = ccv_nnc_tensor_symbol_params(graph, model->outputs[i]);
620	}
621}
622 
623void ccv_cnnp_model_set_workspace_size(ccv_cnnp_model_t* const model, size_t workspace_size)
624{
625	if (workspace_size == model->workspace_size)
626		return;
627	model->workspace_size = workspace_size;
628	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
629	if (compiled_data && compiled_data->graph)
630		ccv_nnc_graph_autotune(compiled_data->graph, workspace_size, 0, TRAVERSE_FULL0,0,0,0);
631}
632 
633size_t ccv_cnnp_model_workspace_size(ccv_cnnp_model_t* const model)
634{
635	return model->workspace_size;
636}
637 
638void ccv_cnnp_model_set_data_parallel(ccv_cnnp_model_t* const model, const int parallel)
639{
640	if (parallel == 0)
641		model->parallel_count = ccv_nnc_device_count(CCV_STREAM_CONTEXT_GPU);
642	else
643		model->parallel_count = parallel;
644	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
645	if (compiled_data)
646		{ assert(!compiled_data->graph)((void) sizeof ((!compiled_data->graph) ? 1 : 0), __extension__
 ({ if (!compiled_data->graph) ; else __assert_fail ("!compiled_data->graph"
, "ccv_cnnp_model.c", 646, __extension__ __PRETTY_FUNCTION__)
; })); }
647}
648 
649void ccv_cnnp_model_set_max_concurrency(ccv_cnnp_model_t* const model, const int max_stream_count)
650{
651	model->max_stream_count = max_stream_count;
652	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
653	if (compiled_data)
654		{ assert(!compiled_data->graph)((void) sizeof ((!compiled_data->graph) ? 1 : 0), __extension__
 ({ if (!compiled_data->graph) ; else __assert_fail ("!compiled_data->graph"
, "ccv_cnnp_model.c", 654, __extension__ __PRETTY_FUNCTION__)
; })); }
655}
656 
657void ccv_cnnp_model_set_memory_compression(ccv_cnnp_model_t* const model, const int memory_compression)
658{
659	model->memory_compression = memory_compression;
660	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
661	if (compiled_data)
662		{ assert(!compiled_data->graph)((void) sizeof ((!compiled_data->graph) ? 1 : 0), __extension__
 ({ if (!compiled_data->graph) ; else __assert_fail ("!compiled_data->graph"
, "ccv_cnnp_model.c", 662, __extension__ __PRETTY_FUNCTION__)
; })); }
663}
664 
665void ccv_cnnp_model_set_memory_reduction(ccv_cnnp_model_t* const model, const int memory_reduction)
666{
667	model->memory_reduction = memory_reduction;
668	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
669	if (compiled_data)
670		{ assert(!compiled_data->graph)((void) sizeof ((!compiled_data->graph) ? 1 : 0), __extension__
 ({ if (!compiled_data->graph) ; else __assert_fail ("!compiled_data->graph"
, "ccv_cnnp_model.c", 670, __extension__ __PRETTY_FUNCTION__)
; })); }
671}
672 
673void ccv_cnnp_model_set_gradient_checkpointing(ccv_cnnp_model_t* const model, const int gradient_checkpointing)
674{
675	model->gradient_checkpointing = gradient_checkpointing;
676}
677 
678int ccv_cnnp_model_gradient_checkpointing(ccv_cnnp_model_t* const model)
679{
680	return model->gradient_checkpointing;
681}
682 
683typedef struct {
684	int parallel_count;
685	ccv_nnc_symbolic_graph_t* graph;
686	ccv_cnnp_compiled_data_t* compiled_data;
687	ccv_nnc_tensor_arena_t* tensor_arena;
688} ccv_nnc_tensor_init_states_t;
689 
690static int _ccv_cnnp_any_to_init(const ccv_cnnp_compiled_data_t* const compiled_data)
691{
692	int i;
693	const uint32_t* const init_v = CCV_NNC_INIT_V(compiled_data->tensors_init.v)((uint32_t*)((uintptr_t)(compiled_data->tensors_init.v) &
 ~(uintptr_t)1));
694	for (i = 0; i < compiled_data->parameters->rnum; i++)
695	{
696		const int d = ((ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->parameters, i)((void*)(((char*)((compiled_data->parameters)->data)) +
 (size_t)(compiled_data->parameters)->rsize * (size_t)(
i))))->d;
697		if (!(init_v[d >> 5] & (1u << (d & 0x1f))))
698			return 1;
699	}
700	for (i = 0; i < compiled_data->internals->rnum; i++)
701	{
702		const int d = ((ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->internals, i)((void*)(((char*)((compiled_data->internals)->data)) + (
size_t)(compiled_data->internals)->rsize * (size_t)(i))
))->d;
703		if (!(init_v[d >> 5] & (1u << (d & 0x1f))))
704			return 1;
705	}
706	return 0;
707}
708 
709static void _ccv_cnnp_init_states_for_tensors(void* const context, const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint, const int flags, ccv_nnc_tensor_t* const input, const ccv_nnc_tensor_symbol_t output_symbol)
710{
711	ccv_nnc_tensor_init_states_t* const tensor_init_states = (ccv_nnc_tensor_init_states_t*)context;
712	ccv_nnc_tensor_arena_t* const tensor_arena = tensor_init_states->tensor_arena;
713	ccv_nnc_tensor_t* const output_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, output_symbol);
714	if (!output_tensor)
715		return;
716	const int d = output_symbol.d;
717	assert(d < tensor_init_states->compiled_data->tensors_init.size)((void) sizeof ((d < tensor_init_states->compiled_data->
tensors_init.size) ? 1 : 0), __extension__ ({ if (d < tensor_init_states
->compiled_data->tensors_init.size) ; else __assert_fail
 ("d < tensor_init_states->compiled_data->tensors_init.size"
, "ccv_cnnp_model.c", 717, __extension__ __PRETTY_FUNCTION__)
; }));
718  uint32_t* const init_v = CCV_NNC_INIT_V(tensor_init_states->compiled_data->tensors_init.v)((uint32_t*)((uintptr_t)(tensor_init_states->compiled_data
->tensors_init.v) & ~(uintptr_t)1));
719	if (init_v[d >> 5] & (1u << (d & 0x1f)))
720		return;
721	init_v[d >> 5] |= (1u << (d & 0x1f));
722	ccv_nnc_cmd_exec(cmd, hint, flags, &input, input ? 1 : 0, &output_tensor, 1, 0);
723	const ccv_nnc_symbolic_graph_t* const graph = tensor_init_states->graph;
724	const int parallel_count = tensor_init_states->parallel_count;
725	int i;
726	for (i = 1; i < parallel_count; i++)
727	{
728		ccv_nnc_tensor_t* const copy = ccv_nnc_tensor_from_symbol(tensor_arena, ccv_nnc_tensor_symbol_copy(graph, output_symbol, i));
729		if (copy)
730			ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD()ccv_nnc_cmd(CCV_NNC_DATA_TRANSFER_FORWARD, 0, ccv_nnc_cmd_auto
, 0), ccv_nnc_no_hint, 0, &output_tensor, 1, &copy, 1, 0);
731	}
732}
733 
734// This method can only handle cases we added new tensors and exec, never delete. This invariant is true because
735// we setup everything (including calling simplify method) in ccv_cnnp_model_compile method, before this rewind setup.
736static void _ccv_cnnp_model_rewind_graph(ccv_cnnp_model_t* const model)
737{
738	assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if
 (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c"
, 738, __extension__ __PRETTY_FUNCTION__); }));
739	assert(model->compiled_data)((void) sizeof ((model->compiled_data) ? 1 : 0), __extension__
 ({ if (model->compiled_data) ; else __assert_fail ("model->compiled_data"
, "ccv_cnnp_model.c", 739, __extension__ __PRETTY_FUNCTION__)
; }));
740	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
741	assert(compiled_data->rewindables)((void) sizeof ((compiled_data->rewindables) ? 1 : 0), __extension__
 ({ if (compiled_data->rewindables) ; else __assert_fail (
"compiled_data->rewindables", "ccv_cnnp_model.c", 741, __extension__
 __PRETTY_FUNCTION__); }));
742	int i;
743	for (i = 0; i < compiled_data->rewindables->rnum; i++)
744	{
745		const ccv_cnnp_rewind_symbol_t* const rewind_symbol = (ccv_cnnp_rewind_symbol_t*)ccv_array_get(compiled_data->rewindables, i)((void*)(((char*)((compiled_data->rewindables)->data)) +
 (size_t)(compiled_data->rewindables)->rsize * (size_t)
(i)));
746		if (rewind_symbol->type == CCV_CNNP_REWIND_GRAPH_EXEC)
747			ccv_nnc_graph_exec_symbol_free(model->graph, rewind_symbol->graph_exec);
748		else if (rewind_symbol->type == CCV_CNNP_REWIND_TENSOR)
749			ccv_nnc_tensor_symbol_free(model->graph, rewind_symbol->tensor);
750	}
751	ccv_array_clear(compiled_data->rewindables);
752	ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
753}
754 
755static void _ccv_cnnp_model_tensor_symbol_new_hook(void* context, const ccv_nnc_tensor_symbol_t symbol, const ccv_nnc_tensor_param_t info, const char* const name)
756{
757	const ccv_cnnp_rewind_symbol_t rewind_symbol = {
758		.type = CCV_CNNP_REWIND_TENSOR,
759		.tensor = symbol
760	};
761	ccv_array_t* const rewind_symbols = (ccv_array_t*)context;
762	ccv_array_push(rewind_symbols, &rewind_symbol);
763}
764 
765static void _ccv_cnnp_model_tensor_symbol_alias_new_hook(void* context, const ccv_nnc_tensor_symbol_t symbol, const ccv_nnc_tensor_symbol_t from_symbol, const int ofs[CCV_NNC_MAX_DIM_ALLOC(12)], const int inc[CCV_NNC_MAX_DIM_ALLOC(12)], const ccv_nnc_tensor_param_t info, const char* const name)
766{
767	const ccv_cnnp_rewind_symbol_t rewind_symbol = {
768		.type = CCV_CNNP_REWIND_TENSOR,
769		.tensor = symbol
770	};
771	ccv_array_t* const rewind_symbols = (ccv_array_t*)context;
772	ccv_array_push(rewind_symbols, &rewind_symbol);
773}
774 
775static void _ccv_cnnp_model_graph_exec_symbol_new_hook(void* context, const ccv_nnc_graph_exec_symbol_t symbol, const ccv_nnc_cmd_t cmd, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, const ccv_nnc_tensor_symbol_t* const outputs, const int output_size, const char* const name)
776{
777	const ccv_cnnp_rewind_symbol_t rewind_symbol = {
778		.type = CCV_CNNP_REWIND_GRAPH_EXEC,
779		.graph_exec = symbol
780	};
781	ccv_array_t* const rewind_symbols = (ccv_array_t*)context;
782	ccv_array_push(rewind_symbols, &rewind_symbol);
783}
784 
785static void _ccv_cnnp_model_graph_symbol_exec_set_for_graph_exec_arena(const ccv_nnc_graph_exec_arena_t* const graph_exec_arena, const int parallel_count, const ccv_nnc_graph_exec_symbol_t exec_symbol, const ccv_nnc_cmd_t cmd, ccv_nnc_symbolic_graph_t* const symbolic_graph)
786{
787	ccv_nnc_graph_exec_t const update_exec = ccv_nnc_graph_exec_from_symbol(graph_exec_arena, exec_symbol);
788	if (!CCV_NO_GRAPH_EXEC(update_exec)((update_exec).graph == 0))
789		ccv_nnc_graph_exec_set(update_exec.graph, update_exec, cmd);
790	int i;
791	for (i = 1; i < parallel_count; i++)
792	{
793		ccv_nnc_graph_exec_symbol_t copy_symbol = ccv_nnc_graph_exec_symbol_copy(symbolic_graph, exec_symbol, i);
794		const ccv_nnc_graph_exec_t copy = ccv_nnc_graph_exec_from_symbol(graph_exec_arena, copy_symbol);
795		if (!CCV_NO_GRAPH_EXEC(copy)((copy).graph == 0))
796			ccv_nnc_graph_exec_set(copy.graph, copy, cmd);
797	}
798}
799 
800static void _ccv_cnnp_model_graph_exec_symbol_set(ccv_nnc_symbolic_graph_t* const symbolic_graph, ccv_cnnp_compiled_data_t* const compiled_data, const int parallel_count, const ccv_nnc_graph_exec_symbol_t exec_symbol, const ccv_nnc_cmd_t cmd)
801{
802	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 802, __extension__ __PRETTY_FUNCTION__); }));
803	assert(symbolic_graph)((void) sizeof ((symbolic_graph) ? 1 : 0), __extension__ ({ if
 (symbolic_graph) ; else __assert_fail ("symbolic_graph", "ccv_cnnp_model.c"
, 803, __extension__ __PRETTY_FUNCTION__); }));
804	ccv_nnc_graph_exec_symbol_set(symbolic_graph, exec_symbol, cmd);
805	int i;
806	for (i = 1; i < parallel_count; i++)
807	{
808		ccv_nnc_graph_exec_symbol_t copy_symbol = ccv_nnc_graph_exec_symbol_copy(symbolic_graph, exec_symbol, i);
809		if (copy_symbol.graph)
810			ccv_nnc_graph_exec_symbol_set(symbolic_graph, copy_symbol, cmd);
811	}
812	ccv_nnc_graph_exec_arena_t* const graph_exec_arena = compiled_data->graph_exec_arena;
813	if (graph_exec_arena)
814		_ccv_cnnp_model_graph_symbol_exec_set_for_graph_exec_arena(graph_exec_arena, parallel_count, exec_symbol, cmd, symbolic_graph);
815	// Skip backward graph exec arena because it is for a specific accum symbolic graph, not the main graph (model->graph)
816	ccv_nnc_graph_exec_arena_t* const gradient_graph_exec_arena = compiled_data->apply_gradients.graph_exec_arena;
817	if (gradient_graph_exec_arena)
818		_ccv_cnnp_model_graph_symbol_exec_set_for_graph_exec_arena(gradient_graph_exec_arena, parallel_count, exec_symbol, cmd, symbolic_graph);
819}
820 
821static int _ccv_cnnp_set_minimizer_for_parameter(ccv_nnc_symbolic_graph_t* const graph, ccv_cnnp_compiled_data_t* const compiled_data, ccv_nnc_graph_exec_symbol_t* const update_nodes, ccv_nnc_tensor_symbol_t* const updated_parameters, ccv_nnc_tensor_symbol_map_t* const saved_aux, const int parallel_count, const ccv_nnc_cmd_t minimizer, const int saved_aux_size, const int max_saved_aux_size, const int parameter_indice)
822{
823	int this_parameter_flag = 0;
824	if (update_nodes[parameter_indice].d == CCV_NNC_NO_TENSOR_SYMBOL)
825		return this_parameter_flag;
826	const ccv_nnc_cmd_t old_minimizer = ccv_nnc_graph_exec_symbol_cmd(graph, update_nodes[parameter_indice]);
827	int j, k;
828	// For no-op, we can preserve previous saved_aux_size.
829	if (old_minimizer.cmd != minimizer.cmd && minimizer.cmd != CCV_NNC_NOOP)
830	{
831		// If the old minimizer is a noop, then the old_saved_aux_size should be whatever its previous
832		// saved_aux_size is, otherwise we will reinit the saved_aux repeatedly if you switch between
833		// noop and a minimizer. We don't want that because we do that in high-level frameworks to
834		// make sure some model parameters don't update if we don't want them to.
835		int old_saved_aux_size;
836		if (old_minimizer.cmd == CCV_NNC_NOOP)
837		{
838			int input_size;
839			ccv_nnc_graph_exec_symbol_io(graph, update_nodes[parameter_indice], 0, &input_size, 0, 0);
840			if (input_size < 2) // This is not legit.
841				old_saved_aux_size = ccv_nnc_minimizer_saved_aux_size(old_minimizer);
842			else // See ccv_nnc_minimizer_saved_aux_size, the saved_aux is inputs excluding gradients and parameters.
843				old_saved_aux_size = input_size - 2;
844		} else
845			old_saved_aux_size = ccv_nnc_minimizer_saved_aux_size(old_minimizer);
846		if (old_saved_aux_size != saved_aux_size)
847		{
848			this_parameter_flag = 1;
849			if (saved_aux_size > old_saved_aux_size)
850			{
851				// Allocate new tensor symbols.
852				const ccv_nnc_tensor_param_t info = ccv_nnc_tensor_symbol_params(graph, updated_parameters[parameter_indice]);
853				for (j = old_saved_aux_size; j < saved_aux_size; j++)
854				{
855					saved_aux[parameter_indice * max_saved_aux_size + j].source = ccv_nnc_tensor_symbol_new(graph, info, 0);
856					saved_aux[parameter_indice * max_saved_aux_size + j].destination = ccv_nnc_tensor_symbol_new(graph, info, 0);
857					const int device_id = CCV_TENSOR_GET_DEVICE_ID(info.type)(((info.type) & 0xfff00) >> 8);
858					for (k = 1; k < parallel_count; k++)
859					{
860						ccv_nnc_tensor_param_t dev_info = info;
861						if (k != device_id)
862							CCV_TENSOR_SET_DEVICE_ID(dev_info.type, k)(dev_info.type) = (((dev_info.type) & ~0xfff00) | (((k) &
 0xfff) << 8));
863						else
864							CCV_TENSOR_SET_DEVICE_ID(dev_info.type, 0)(dev_info.type) = (((dev_info.type) & ~0xfff00) | (((0) &
 0xfff) << 8));
865						const ccv_nnc_tensor_symbol_t src_copy = ccv_nnc_tensor_symbol_new(graph, dev_info, 0);
866						const ccv_nnc_tensor_symbol_t dest_copy = ccv_nnc_tensor_symbol_new(graph, dev_info, 0);
867						ccv_nnc_tensor_symbol_set_copy(graph, saved_aux[parameter_indice * max_saved_aux_size + j].source, k, src_copy);
868						ccv_nnc_tensor_symbol_set_copy(graph, saved_aux[parameter_indice * max_saved_aux_size + j].destination, k, dest_copy);
869					}
870				}
871			} else {
872				for (j = saved_aux_size; j < old_saved_aux_size; j++)
873				{
874					for (k = 1; k < parallel_count; k++)
875					{
876						const ccv_nnc_tensor_symbol_t src_copy = ccv_nnc_tensor_symbol_copy(graph, saved_aux[parameter_indice * max_saved_aux_size + j].source, k);
877						if (src_copy.d >= 0)
878						{
879							ccv_nnc_tensor_symbol_free(graph, src_copy);
880							ccv_nnc_tensor_symbol_set_copy(graph, saved_aux[parameter_indice * max_saved_aux_size + j].source, k, NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
});
881						}
882						const ccv_nnc_tensor_symbol_t dest_copy = ccv_nnc_tensor_symbol_copy(graph, saved_aux[parameter_indice * max_saved_aux_size + j].destination, k);
883						if (dest_copy.d >= 0)
884						{
885							ccv_nnc_tensor_symbol_free(graph, dest_copy);
886							ccv_nnc_tensor_symbol_set_copy(graph, saved_aux[parameter_indice * max_saved_aux_size + j].destination, k, NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
});
887						}
888					}
889					ccv_nnc_tensor_symbol_free(graph, saved_aux[parameter_indice * max_saved_aux_size + j].source);
890					ccv_nnc_tensor_symbol_free(graph, saved_aux[parameter_indice * max_saved_aux_size + j].destination);
891					saved_aux[parameter_indice * max_saved_aux_size + j].source = saved_aux[parameter_indice * max_saved_aux_size + j].destination = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
892				}
893			}
894		}
895	}
896	_ccv_cnnp_model_graph_exec_symbol_set(graph, compiled_data, parallel_count, update_nodes[parameter_indice], minimizer);
897	if (this_parameter_flag)
898	{
899		ccv_nnc_tensor_symbol_t update_inputs[saved_aux_size + 2];
900		ccv_nnc_tensor_symbol_t update_outputs[saved_aux_size + 1];
901		const int* inputs = 0;
902		int input_size = 0;
903		ccv_nnc_graph_exec_symbol_io(graph, update_nodes[parameter_indice], &inputs, &input_size, 0, 0);
904		assert(input_size >= 1)((void) sizeof ((input_size >= 1) ? 1 : 0), __extension__ (
{ if (input_size >= 1) ; else __assert_fail ("input_size >= 1"
, "ccv_cnnp_model.c", 904, __extension__ __PRETTY_FUNCTION__)
; }));
905		update_inputs[0].d = inputs[0];
906		update_inputs[0].graph = graph;
907		update_inputs[1].d = inputs[1];
908		update_inputs[1].graph = graph;
909		update_outputs[0] = updated_parameters[parameter_indice];
910		for (j = 0; j < saved_aux_size; j++)
911		{
912			update_inputs[j + 2] = saved_aux[parameter_indice * max_saved_aux_size + j].source;
913			update_outputs[j + 1] = saved_aux[parameter_indice * max_saved_aux_size + j].destination;
914		}
915		ccv_nnc_graph_exec_symbol_set_io(graph, update_nodes[parameter_indice], update_inputs, saved_aux_size + 2, update_outputs, saved_aux_size + 1);
916		for (k = 1; k < parallel_count; k++)
917		{
918			const ccv_nnc_graph_exec_symbol_t copy = ccv_nnc_graph_exec_symbol_copy(graph, update_nodes[parameter_indice], k);
919			assert(copy.d >= 0)((void) sizeof ((copy.d >= 0) ? 1 : 0), __extension__ ({ if
 (copy.d >= 0) ; else __assert_fail ("copy.d >= 0", "ccv_cnnp_model.c"
, 919, __extension__ __PRETTY_FUNCTION__); }));
920			ccv_nnc_graph_exec_symbol_io(graph, copy, &inputs, &input_size, 0, 0);
921			assert(input_size >= 1)((void) sizeof ((input_size >= 1) ? 1 : 0), __extension__ (
{ if (input_size >= 1) ; else __assert_fail ("input_size >= 1"
, "ccv_cnnp_model.c", 921, __extension__ __PRETTY_FUNCTION__)
; }));
922			update_inputs[0].d = inputs[0];
923			update_inputs[0].graph = graph;
924			update_inputs[1].d = inputs[1];
925			update_inputs[1].graph = graph;
926			update_outputs[0] = ccv_nnc_tensor_symbol_copy(graph, updated_parameters[parameter_indice], k);
927			for (j = 0; j < saved_aux_size; j++)
928			{
929				update_inputs[j + 2] = ccv_nnc_tensor_symbol_copy(graph, saved_aux[parameter_indice * max_saved_aux_size + j].source, k);
930				update_outputs[j + 1] = ccv_nnc_tensor_symbol_copy(graph, saved_aux[parameter_indice * max_saved_aux_size + j].destination, k);
931			}
932			ccv_nnc_graph_exec_symbol_set_io(graph, copy, update_inputs, saved_aux_size + 2, update_outputs, saved_aux_size + 1);
933		}
934	}
935	return this_parameter_flag;
936}
937 
938typedef struct {
939	int parameter_size;
940	ccv_nnc_cmd_t minimizer;
941	ccv_cnnp_model_io_t parameters[1];
942} ccv_cnnp_set_minimizer_for_parameter_t;
943 
944static int _ccv_cnnp_apply_parameters_with_minimizer(ccv_cnnp_model_t* const model)
945{
946	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
947	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 947, __extension__ __PRETTY_FUNCTION__); }));
948	const int max_saved_aux_size = compiled_data->minimize.max_saved_aux_size;
949	// We update all parameters, at this point, we have one minimizer.
950	const int parameter_size = compiled_data->parameters->rnum;
951	ccv_nnc_graph_exec_symbol_t* const update_nodes = compiled_data->update_nodes;
952	ccv_nnc_symbolic_graph_t* const symbolic_graph = model->graph;
953	assert(symbolic_graph)((void) sizeof ((symbolic_graph) ? 1 : 0), __extension__ ({ if
 (symbolic_graph) ; else __assert_fail ("symbolic_graph", "ccv_cnnp_model.c"
, 953, __extension__ __PRETTY_FUNCTION__); }));
954	const int parallel_count = ccv_max(model->parallel_count, 1)({ typeof (model->parallel_count) _a = (model->parallel_count
); typeof (1) _b = (1); (_a > _b) ? _a : _b; });
955	ccv_array_t* const parameters = compiled_data->minimize.parameters;
956	ccv_array_t* const parameter_indices = ccv_array_new(sizeof(int), 0, 0);
957	int i, j, flag = 0;
958	for (i = 0; i < parameters->rnum; i++)
959	{
960		ccv_cnnp_set_minimizer_for_parameter_t* const set_minimizer_for_parameter = *(ccv_cnnp_set_minimizer_for_parameter_t**)ccv_array_get(parameters, i)((void*)(((char*)((parameters)->data)) + (size_t)(parameters
)->rsize * (size_t)(i)));
961		for (j = 0; j < set_minimizer_for_parameter->parameter_size; j++)
962		{
963			const int param_sel = set_minimizer_for_parameter->parameters[j]->param_sel > 0 ? set_minimizer_for_parameter->parameters[j]->param_sel - 1 : set_minimizer_for_parameter->parameters[j]->param_sel;
964			assert(set_minimizer_for_parameter->parameters[j]->param_sel != 0)((void) sizeof ((set_minimizer_for_parameter->parameters[j
]->param_sel != 0) ? 1 : 0), __extension__ ({ if (set_minimizer_for_parameter
->parameters[j]->param_sel != 0) ; else __assert_fail (
"set_minimizer_for_parameter->parameters[j]->param_sel != 0"
, "ccv_cnnp_model.c", 964, __extension__ __PRETTY_FUNCTION__)
; }));
965			const int old_rnum = parameter_indices->rnum;
966			ccv_cnnp_model_add_to_parameter_indices(set_minimizer_for_parameter->parameters[j]->model, param_sel, parameter_indices);
967			const int param_ref = set_minimizer_for_parameter->parameters[j]->param_ref > 0 ? set_minimizer_for_parameter->parameters[j]->param_ref - 1 : set_minimizer_for_parameter->parameters[j]->param_ref;
968			assert(set_minimizer_for_parameter->parameters[j]->param_ref != 0)((void) sizeof ((set_minimizer_for_parameter->parameters[j
]->param_ref != 0) ? 1 : 0), __extension__ ({ if (set_minimizer_for_parameter
->parameters[j]->param_ref != 0) ; else __assert_fail (
"set_minimizer_for_parameter->parameters[j]->param_ref != 0"
, "ccv_cnnp_model.c", 968, __extension__ __PRETTY_FUNCTION__)
; }));
969			if (param_ref >= 0)
970			{
971				assert(param_ref + old_rnum < parameter_indices->rnum)((void) sizeof ((param_ref + old_rnum < parameter_indices->
rnum) ? 1 : 0), __extension__ ({ if (param_ref + old_rnum <
 parameter_indices->rnum) ; else __assert_fail ("param_ref + old_rnum < parameter_indices->rnum"
, "ccv_cnnp_model.c", 971, __extension__ __PRETTY_FUNCTION__)
; }));
972				*(int*)ccv_array_get(parameter_indices, old_rnum)((void*)(((char*)((parameter_indices)->data)) + (size_t)(parameter_indices
)->rsize * (size_t)(old_rnum))) = *(int*)ccv_array_get(parameter_indices, param_ref + old_rnum)((void*)(((char*)((parameter_indices)->data)) + (size_t)(parameter_indices
)->rsize * (size_t)(param_ref + old_rnum)));
973				parameter_indices->rnum = old_rnum + 1;
974			}
975		}
976		const int saved_aux_size = ccv_nnc_minimizer_saved_aux_size(set_minimizer_for_parameter->minimizer);
977		// We may have duplicated indices, but that is OK, we will set it twice.
978		for (j = 0; j < parameter_indices->rnum; j++)
979		{
980			const int d = *(int*)ccv_array_get(parameter_indices, j)((void*)(((char*)((parameter_indices)->data)) + (size_t)(parameter_indices
)->rsize * (size_t)(j)));
981			assert(d <= parameter_size)((void) sizeof ((d <= parameter_size) ? 1 : 0), __extension__
 ({ if (d <= parameter_size) ; else __assert_fail ("d <= parameter_size"
, "ccv_cnnp_model.c", 981, __extension__ __PRETTY_FUNCTION__)
; }));
982			if (_ccv_cnnp_set_minimizer_for_parameter(symbolic_graph, compiled_data, update_nodes, compiled_data->updated_parameters, compiled_data->saved_aux, parallel_count, set_minimizer_for_parameter->minimizer, saved_aux_size, max_saved_aux_size, d))
983				flag = 1;
984		}
985		ccv_array_clear(parameter_indices);
986	}
987	ccv_array_free(parameter_indices);
988	return flag;
989}
990 
991static void _ccv_cnnp_scatter_saved_aux(ccv_nnc_tensor_symbol_map_t* const saved_aux, const int parameter_size, const int old_saved_aux_size, const int new_saved_aux_size)
992{
993	if (new_saved_aux_size == old_saved_aux_size)
994		return;
995	assert(new_saved_aux_size > old_saved_aux_size)((void) sizeof ((new_saved_aux_size > old_saved_aux_size) ?
 1 : 0), __extension__ ({ if (new_saved_aux_size > old_saved_aux_size
) ; else __assert_fail ("new_saved_aux_size > old_saved_aux_size"
, "ccv_cnnp_model.c", 995, __extension__ __PRETTY_FUNCTION__)
; }));
996	int i, j;
997	for (i = parameter_size - 1; i >= 0; i--)
998	{
999		for (j = new_saved_aux_size - 1; j >= old_saved_aux_size; j--)
1000			saved_aux[i * new_saved_aux_size + j].source = saved_aux[i * new_saved_aux_size + j].destination = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
1001		for (j = old_saved_aux_size - 1; j >= 0; j--)
1002			saved_aux[i * new_saved_aux_size + j] = saved_aux[i * old_saved_aux_size + j];
1003	}
1004}
1005 
1006static void _ccv_cnnp_model_set_rewindables(ccv_cnnp_model_t* const model)
1007{
1008	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1009	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 1009, __extension__ __PRETTY_FUNCTION__); }));
1010	if (!compiled_data->rewindables)
1011		compiled_data->rewindables = ccv_array_new(sizeof(ccv_cnnp_rewind_symbol_t), 0, 0);
1012	ccv_nnc_tensor_symbol_new_hook(model->graph, _ccv_cnnp_model_tensor_symbol_new_hook, compiled_data->rewindables, 0);
1013	ccv_nnc_tensor_symbol_alias_new_hook(model->graph, _ccv_cnnp_model_tensor_symbol_alias_new_hook, compiled_data->rewindables, 0);
1014	ccv_nnc_graph_exec_symbol_new_hook(model->graph, _ccv_cnnp_model_graph_exec_symbol_new_hook, compiled_data->rewindables, 0);
1015}
1016 
1017static void _ccv_cnnp_model_gradient_init(ccv_cnnp_model_t* const model, const int gradient_mode, const uint64_t disable_outgrad, ccv_nnc_tensor_t* const* const fits, const int fit_size)
1018{
1019	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1020	assert(compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_NONE)((void) sizeof ((compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_NONE
) ? 1 : 0), __extension__ ({ if (compiled_data->gradient_mode
 == CCV_CNNP_COMPILED_DATA_GRADIENT_NONE) ; else __assert_fail
 ("compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_NONE"
, "ccv_cnnp_model.c", 1020, __extension__ __PRETTY_FUNCTION__
); }));
1021	assert(gradient_mode != CCV_CNNP_COMPILED_DATA_GRADIENT_NONE)((void) sizeof ((gradient_mode != CCV_CNNP_COMPILED_DATA_GRADIENT_NONE
) ? 1 : 0), __extension__ ({ if (gradient_mode != CCV_CNNP_COMPILED_DATA_GRADIENT_NONE
) ; else __assert_fail ("gradient_mode != CCV_CNNP_COMPILED_DATA_GRADIENT_NONE"
, "ccv_cnnp_model.c", 1021, __extension__ __PRETTY_FUNCTION__
); }));
1022	const int evaluate_to_size = compiled_data->evaluate.to_size;
1023	assert(evaluate_to_size > 0)((void) sizeof ((evaluate_to_size > 0) ? 1 : 0), __extension__
 ({ if (evaluate_to_size > 0) ; else __assert_fail ("evaluate_to_size > 0"
, "ccv_cnnp_model.c", 1023, __extension__ __PRETTY_FUNCTION__
); }));
1024	const int parallel_count = ccv_max(model->parallel_count, 1)({ typeof (model->parallel_count) _a = (model->parallel_count
); typeof (1) _b = (1); (_a > _b) ? _a : _b; });
1025	compiled_data->evaluate.tos = ccreallocrealloc(compiled_data->evaluate.tos, sizeof(ccv_nnc_graph_exec_symbol_t) * evaluate_to_size * parallel_count + sizeof(ccv_nnc_graph_exec_t) * evaluate_to_size * parallel_count);
1026	compiled_data->evaluate.to_ops = (ccv_nnc_graph_exec_t*)(compiled_data->evaluate.tos + evaluate_to_size * parallel_count);
1027	int i, j;
1028	const int output_size = model->output_size;
1029	assert(!fits || fit_size == output_size * parallel_count)((void) sizeof ((!fits || fit_size == output_size * parallel_count
) ? 1 : 0), __extension__ ({ if (!fits || fit_size == output_size
 * parallel_count) ; else __assert_fail ("!fits || fit_size == output_size * parallel_count"
, "ccv_cnnp_model.c", 1029, __extension__ __PRETTY_FUNCTION__
); }));
1030	if (fits)
1031		for (i = 0; i < output_size; i++)
1032			ccv_nnc_tensor_symbol_set(model->graph, compiled_data->fits[i], fits[i]->info);
1033	const int max_saved_aux_size = compiled_data->minimize.max_saved_aux_size;
1034	const int parameter_size = compiled_data->parameters->rnum;
1035	compiled_data->updated_parameters = (ccv_nnc_tensor_symbol_t*)ccmallocmalloc(sizeof(ccv_nnc_tensor_symbol_t) * parameter_size + sizeof(ccv_nnc_graph_exec_symbol_t) * parameter_size + sizeof(ccv_nnc_tensor_symbol_map_t) * max_saved_aux_size * parameter_size);
1036	compiled_data->update_nodes = (ccv_nnc_graph_exec_symbol_t*)(compiled_data->updated_parameters + parameter_size);
1037	compiled_data->saved_aux = (ccv_nnc_tensor_symbol_map_t*)(compiled_data->update_nodes + parameter_size);
1038	int parameter_size_maybe_more = parameter_size;
1039	compiled_data->disable_outgrad = disable_outgrad;
1040	int outgrad_size;
1041	if (gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES || model->input_size == 0)
1042		outgrad_size = 0;
1043	else if (disable_outgrad == CCV_CNNP_DISABLE_OUTGRAD_NONE) // Compute minimize with gradients including inputs.
1044		outgrad_size = model->input_size;
1045	else {
1046		assert(disable_outgrad != CCV_CNNP_DISABLE_OUTGRAD_ALL)((void) sizeof ((disable_outgrad != CCV_CNNP_DISABLE_OUTGRAD_ALL
) ? 1 : 0), __extension__ ({ if (disable_outgrad != CCV_CNNP_DISABLE_OUTGRAD_ALL
) ; else __assert_fail ("disable_outgrad != CCV_CNNP_DISABLE_OUTGRAD_ALL"
, "ccv_cnnp_model.c", 1046, __extension__ __PRETTY_FUNCTION__
); })); // If it is disable all, gradient mode won't be this.
1047		outgrad_size = 0;
1048		for (i = 0; i < model->input_size; i++)
1049			if (!(disable_outgrad & ((uint64_t)1 << i)))
1050				++outgrad_size;
1051	}
1052	compiled_data->outgrad_size = outgrad_size;
1053	parameter_size_maybe_more += outgrad_size;
1054	compiled_data->gradients = (ccv_nnc_tensor_symbol_t*)ccmallocmalloc(sizeof(ccv_nnc_tensor_symbol_t) * parameter_size_maybe_more + sizeof(ccv_nnc_graph_exec_symbol_t) * parameter_size_maybe_more * parallel_count);
1055	compiled_data->outgrads = parameter_size_maybe_more > parameter_size ? compiled_data->gradients + parameter_size : 0;
1056	compiled_data->backward.tos = (ccv_nnc_graph_exec_symbol_t*)(compiled_data->gradients + parameter_size_maybe_more);
1057	compiled_data->backward.to_size = parameter_size_maybe_more;
1058	ccv_nnc_tensor_symbol_t* parameters = (ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->parameters, 0)((void*)(((char*)((compiled_data->parameters)->data)) +
 (size_t)(compiled_data->parameters)->rsize * (size_t)(
0)));
1059	if (compiled_data->parameter_flags)
1060	{
1061		parameters = (ccv_nnc_tensor_symbol_t*)ccmallocmalloc(sizeof(ccv_nnc_tensor_symbol_t) * parameter_size);
1062		for (i = 0; i < parameter_size; i++)
1063			if (compiled_data->parameter_flags[i >> 6] & ((uint64_t)1 << (i & 63)))
1064				parameters[i] = *(ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->parameters, i)((void*)(((char*)((compiled_data->parameters)->data)) +
 (size_t)(compiled_data->parameters)->rsize * (size_t)(
i)));
1065			else
1066				parameters[i] = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
1067	}
1068	if (gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES || model->input_size == 0)
1069		ccv_nnc_symbolic_graph_minimize(model->graph, compiled_data->minimize.minimizer, compiled_data->f, output_size, parameters, parameter_size, 0, 0, SYMBOLIC_GRAPH_SOURCES(model->graph)ccv_nnc_symbolic_graph_sources(model->graph), ccv_nnc_symbolic_graph_source_size
(model->graph), SYMBOLIC_GRAPH_DESTINATIONS(model->graph)ccv_nnc_symbolic_graph_destinations(model->graph), ccv_nnc_symbolic_graph_destination_size
(model->graph), compiled_data->gradients, compiled_data->updated_parameters, compiled_data->saved_aux, compiled_data->update_nodes);
1070	else if (disable_outgrad == CCV_CNNP_DISABLE_OUTGRAD_NONE) // Compute minimize with gradients including inputs.
1071		ccv_nnc_symbolic_graph_minimize(model->graph, compiled_data->minimize.minimizer, compiled_data->f, output_size, parameters, parameter_size, model->inputs, model->input_size, SYMBOLIC_GRAPH_SOURCES(model->graph)ccv_nnc_symbolic_graph_sources(model->graph), ccv_nnc_symbolic_graph_source_size
(model->graph), SYMBOLIC_GRAPH_DESTINATIONS(model->graph)ccv_nnc_symbolic_graph_destinations(model->graph), ccv_nnc_symbolic_graph_destination_size
(model->graph), compiled_data->gradients, compiled_data->updated_parameters, compiled_data->saved_aux, compiled_data->update_nodes);
1072	else { // Compute minimize with gradients including selected inputs.
1073		assert(model->input_size > 0)((void) sizeof ((model->input_size > 0) ? 1 : 0), __extension__
 ({ if (model->input_size > 0) ; else __assert_fail ("model->input_size > 0"
, "ccv_cnnp_model.c", 1073, __extension__ __PRETTY_FUNCTION__
); }));
1074		assert(disable_outgrad != CCV_CNNP_DISABLE_OUTGRAD_ALL)((void) sizeof ((disable_outgrad != CCV_CNNP_DISABLE_OUTGRAD_ALL
) ? 1 : 0), __extension__ ({ if (disable_outgrad != CCV_CNNP_DISABLE_OUTGRAD_ALL
) ; else __assert_fail ("disable_outgrad != CCV_CNNP_DISABLE_OUTGRAD_ALL"
, "ccv_cnnp_model.c", 1074, __extension__ __PRETTY_FUNCTION__
); })); // If it is disable all, gradient mode won't be this.
1075		assert(outgrad_size > 0)((void) sizeof ((outgrad_size > 0) ? 1 : 0), __extension__
 ({ if (outgrad_size > 0) ; else __assert_fail ("outgrad_size > 0"
, "ccv_cnnp_model.c", 1075, __extension__ __PRETTY_FUNCTION__
); }));
1076		ccv_nnc_tensor_symbol_t outgrads[outgrad_size];
1077		j = 0;
1078		for (i = 0; i < model->input_size; i++)
1079			if (!(disable_outgrad & ((uint64_t)1 << i)))
1080				outgrads[j++] = model->inputs[i];
1081		ccv_nnc_symbolic_graph_minimize(model->graph, compiled_data->minimize.minimizer, compiled_data->f, output_size, parameters, parameter_size, outgrads, outgrad_size, SYMBOLIC_GRAPH_SOURCES(model->graph)ccv_nnc_symbolic_graph_sources(model->graph), ccv_nnc_symbolic_graph_source_size
(model->graph), SYMBOLIC_GRAPH_DESTINATIONS(model->graph)ccv_nnc_symbolic_graph_destinations(model->graph), ccv_nnc_symbolic_graph_destination_size
(model->graph), compiled_data->gradients, compiled_data->updated_parameters, compiled_data->saved_aux, compiled_data->update_nodes);
1082	}
1083	if (compiled_data->parameter_flags)
1084		ccfreefree(parameters);
1085	_ccv_cnnp_scatter_saved_aux(compiled_data->saved_aux, parameter_size, ccv_nnc_minimizer_saved_aux_size(compiled_data->minimize.minimizer), compiled_data->minimize.max_saved_aux_size);
1086	if (compiled_data->minimize.parameters)
1087		_ccv_cnnp_apply_parameters_with_minimizer(model);
1088	// Go through gradient checkpoints to generate tensor inputs for backward pass just before executing the backward pass.
1089	ccv_cnnp_model_apply_gradient_checkpoints(compiled_data, model->graph);
1090	for (i = 0; i < output_size; i++)
1091	{
1092		const ccv_nnc_tensor_symbol_t df = ccv_nnc_tensor_symbol_for_backward(model->graph, compiled_data->f[i]);
1093		// Init this to 1 so we can backprop.
1094		ccv_nnc_tensor_symbol_set_flags(model->graph, df, CCV_NNC_TENSOR_SYMBOL_INIT_ONES);
1095	}
1096	compiled_data->backward.to_size = 0;
1097	for (i = 0; i < parameter_size_maybe_more; i++)
1098		if (compiled_data->gradients[i].d != CCV_NNC_NO_TENSOR_SYMBOL)
1099			compiled_data->backward.tos[compiled_data->backward.to_size++] = ccv_nnc_graph_exec_symbol_for_backward(model->graph, compiled_data->gradients[i]);
1100	ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS);
1101	ccv_nnc_symbolic_graph_set_destinations(model->graph, compiled_data->update_nodes, parameter_size);
1102	for (i = 0; i < parameter_size_maybe_more - parameter_size; i++)
1103	{
1104		if (compiled_data->outgrads[i].d < 0) // When we go through input, we might find zero-length inputs, and for these, we cannot have any outgrads.
1105			continue;
1106		const ccv_nnc_graph_exec_symbol_t outgrad = ccv_nnc_graph_exec_symbol_for_backward(model->graph, compiled_data->outgrads[i]);
1107		const int* tos;
1108		int to_size;
1109		ccv_nnc_graph_exec_symbol_to(model->graph, outgrad, &tos, &to_size);
1110		if (to_size == 0) // If this is the end (no minimizers afterwards). We need to attach this as a destination. Otherwise this is covered in update_nodes.
1111		{
1112			const ccv_nnc_graph_exec_symbol_t* destinations = ccv_nnc_symbolic_graph_destinations(model->graph);
1113			const int destination_count = ccv_nnc_symbolic_graph_destination_size(model->graph);
1114			int flag = 0;
1115			const int outgrad_destination_start = ccv_max(0, destination_count - i)({ typeof (0) _a = (0); typeof (destination_count - i) _b = (
destination_count - i); (_a > _b) ? _a : _b; });
1116			for (j = i - 1; !flag && j >= 0; j--)
1117				if (j + outgrad_destination_start < destination_count)
1118					flag = (destinations[j + outgrad_destination_start].d == outgrad.d);
1119			if (!flag) // Only if we cannot find it, we add it.
1120				ccv_nnc_symbolic_graph_add_destination(model->graph, outgrad);
1121		}
1122	}
1123	if (parallel_count > 1)
1124	{
1125		ccv_nnc_symbolic_graph_data_parallel(model->graph, parallel_count,
1126			0, 0,
1127			compiled_data->gradients, parameter_size /* No need to deal with outgrads, we don't allreduce outgrads */,
1128			compiled_data->gradients /* We only care about gradients before allreduce, thus, update our current pointers */,
1129			0, 0, 0,
1130			CCV_NNC_PARALLEL_REDUCE_OP_SUM,
1131			SYMBOLIC_GRAPH_SOURCES(model->graph)ccv_nnc_symbolic_graph_sources(model->graph), ccv_nnc_symbolic_graph_source_size
(model->graph), SYMBOLIC_GRAPH_DESTINATIONS(model->graph)ccv_nnc_symbolic_graph_destinations(model->graph), ccv_nnc_symbolic_graph_destination_size
(model->graph));
1132		ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
1133		for (i = 0; i < evaluate_to_size; i++)
1134			for (j = 1; j < parallel_count; j++)
1135			{
1136				const ccv_nnc_graph_exec_symbol_t copy = ccv_nnc_graph_exec_symbol_copy(model->graph, compiled_data->evaluate.tos[i], j);
1137				if (copy.d != CCV_NNC_NO_GRAPH_EXEC_SYMBOL)
1138					compiled_data->evaluate.tos[compiled_data->evaluate.to_size++] = copy;
1139			}
1140		const int backward_to_size = compiled_data->backward.to_size;
1141		for (i = 0; i < backward_to_size; i++)
1142			for (j = 1; j < parallel_count; j++)
1143			{
1144				const ccv_nnc_graph_exec_symbol_t copy = ccv_nnc_graph_exec_symbol_copy(model->graph, compiled_data->backward.tos[i], j);
1145				if (copy.d != CCV_NNC_NO_GRAPH_EXEC_SYMBOL)
1146					compiled_data->backward.tos[compiled_data->backward.to_size++] = copy;
1147			}
1148	}
1149	// Only use memory compression if we are in gradient parameter mode.
1150	if (gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES || gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES_AND_INPUTS)
1151	{
1152		if (model->memory_compression)
1153			ccv_nnc_symbolic_graph_memory_compression(model->graph, SYMBOLIC_GRAPH_SOURCES(model->graph)ccv_nnc_symbolic_graph_sources(model->graph), ccv_nnc_symbolic_graph_source_size
(model->graph), SYMBOLIC_GRAPH_DESTINATIONS(model->graph)ccv_nnc_symbolic_graph_destinations(model->graph), ccv_nnc_symbolic_graph_destination_size
(model->graph));
1154		if (model->memory_reduction)
1155			ccv_nnc_symbolic_graph_memory_reduction(model->graph, SYMBOLIC_GRAPH_SOURCES(model->graph)ccv_nnc_symbolic_graph_sources(model->graph), ccv_nnc_symbolic_graph_source_size
(model->graph), SYMBOLIC_GRAPH_DESTINATIONS(model->graph)ccv_nnc_symbolic_graph_destinations(model->graph), ccv_nnc_symbolic_graph_destination_size
(model->graph));
1156	}
1157	compiled_data->backward.to_size = _ccv_nnc_array_dedup_graph_exec_symbols(compiled_data->backward.tos, compiled_data->backward.to_size);
1158	compiled_data->gradient_mode = gradient_mode;
1159}
1160 
1161void ccv_cnnp_model_tensors_init_0(const ccv_cnnp_model_t* const model, ccv_cnnp_compiled_data_t* const compiled_data)
1162{
1163	assert(!compiled_data->tensors.parameters)((void) sizeof ((!compiled_data->tensors.parameters) ? 1 :
 0), __extension__ ({ if (!compiled_data->tensors.parameters
) ; else __assert_fail ("!compiled_data->tensors.parameters"
, "ccv_cnnp_model.c", 1163, __extension__ __PRETTY_FUNCTION__
); }));
1164	const int parameter_size = compiled_data->parameters->rnum;
1165	const int parallel_count = ccv_max(model->parallel_count, 1)({ typeof (model->parallel_count) _a = (model->parallel_count
); typeof (1) _b = (1); (_a > _b) ? _a : _b; });
1166	const int internal_size = compiled_data->internals->rnum;
1167	compiled_data->tensors_init.size = ccv_nnc_tensor_symbol_count(model->graph);
1168	compiled_data->tensors_init.v = cccalloccalloc(((compiled_data->tensors_init.size + 31) >> 5), sizeof(uint32_t));
1169	compiled_data->tensors.parameters = (ccv_nnc_tensor_t**)cccalloccalloc((parameter_size + internal_size) * parallel_count, sizeof(ccv_nnc_tensor_t*));
1170	compiled_data->tensors.internals = compiled_data->tensors.parameters + parameter_size * parallel_count;
1171}
1172 
1173int ccv_cnnp_model_tensors_any_to_alloc(const ccv_cnnp_model_t* const model, ccv_cnnp_compiled_data_t* const compiled_data)
1174{
1175	int i, j;
1176	const int parameter_size = compiled_data->parameters->rnum;
1177	const int parallel_count = ccv_max(model->parallel_count, 1)({ typeof (model->parallel_count) _a = (model->parallel_count
); typeof (1) _b = (1); (_a > _b) ? _a : _b; });
1178	const int internal_size = compiled_data->internals->rnum;
1179	for (i = 0; i < parameter_size; i++)
1180	{
1181		// parameters has to be allocated all together.
1182		if (compiled_data->tensors.parameters[i])
1183		{
1184			for (j = 1; j < parallel_count; j++)
1185				{ assert(compiled_data->tensors.parameters[i + j * parameter_size])((void) sizeof ((compiled_data->tensors.parameters[i + j *
 parameter_size]) ? 1 : 0), __extension__ ({ if (compiled_data
->tensors.parameters[i + j * parameter_size]) ; else __assert_fail
 ("compiled_data->tensors.parameters[i + j * parameter_size]"
, "ccv_cnnp_model.c", 1185, __extension__ __PRETTY_FUNCTION__
); })); }
1186			continue;
1187		}
1188		return 1;
1189	}
1190	for (i = 0; i < internal_size; i++)
1191	{
1192		if (!compiled_data->tensors.internals[i])
1193			return 1;
1194		for (j = 1; j < parallel_count; j++)
1195			if (!compiled_data->tensors.internals[i + j * internal_size])
1196				return 1;
1197	}
1198	return 0;
1199}
1200 
1201void ccv_cnnp_model_tensors_init_1(const ccv_cnnp_model_t* const model, ccv_cnnp_compiled_data_t* const compiled_data)
1202{
1203	int i, j;
1204	const int parameter_size = compiled_data->parameters->rnum;
1205	const int parallel_count = ccv_max(model->parallel_count, 1)({ typeof (model->parallel_count) _a = (model->parallel_count
); typeof (1) _b = (1); (_a > _b) ? _a : _b; });
1206	const int internal_size = compiled_data->internals->rnum;
1207	for (i = 0; i < parameter_size; i++)
1208	{
1209		// parameters has to be allocated all together.
1210		if (compiled_data->tensors.parameters[i])
1211		{
1212			for (j = 1; j < parallel_count; j++)
1213				{ assert(compiled_data->tensors.parameters[i + j * parameter_size])((void) sizeof ((compiled_data->tensors.parameters[i + j *
 parameter_size]) ? 1 : 0), __extension__ ({ if (compiled_data
->tensors.parameters[i + j * parameter_size]) ; else __assert_fail
 ("compiled_data->tensors.parameters[i + j * parameter_size]"
, "ccv_cnnp_model.c", 1213, __extension__ __PRETTY_FUNCTION__
); })); }
1214			continue;
1215		}
1216		const ccv_nnc_tensor_symbol_t parameter = *(ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->parameters, i)((void*)(((char*)((compiled_data->parameters)->data)) +
 (size_t)(compiled_data->parameters)->rsize * (size_t)(
i)));
1217		ccv_nnc_tensor_param_t info = ccv_nnc_tensor_symbol_params(parameter.graph, parameter);
1218		if (CCV_TENSOR_GET_DEVICE(info.type)((info.type) & 0xfff00) == CCV_COMPUTE_DEVICE_ANY)
1219			CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff
) << 8));
1220		const int device_id = CCV_TENSOR_GET_DEVICE_ID(info.type)(((info.type) & 0xfff00) >> 8);
1221		compiled_data->tensors.parameters[i] = ccv_nnc_tensor_new(0, info, 0);
1222		for (j = 1; j < parallel_count; j++)
1223		{
1224			if (j != device_id)
1225				CCV_TENSOR_SET_DEVICE_ID(info.type, j)(info.type) = (((info.type) & ~0xfff00) | (((j) & 0xfff
) << 8));
1226			else
1227				CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff
) << 8));
1228			compiled_data->tensors.parameters[i + j * parameter_size] = ccv_nnc_tensor_new(0, info, 0);
1229		}
1230	}
1231	const uint32_t* const init_v = CCV_NNC_INIT_V(compiled_data->tensors_init.v)((uint32_t*)((uintptr_t)(compiled_data->tensors_init.v) &
 ~(uintptr_t)1));
1232	for (i = 0; i < internal_size; i++)
1233	{
1234		const ccv_nnc_tensor_symbol_t retained = *(ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->internals, i)((void*)(((char*)((compiled_data->internals)->data)) + (
size_t)(compiled_data->internals)->rsize * (size_t)(i))
);
1235		const int d = retained.d;
1236		if (init_v[d >> 5] & (1u << (d & 0x1f)))
1237			continue;
1238		ccv_nnc_tensor_param_t info = ccv_nnc_tensor_symbol_params(retained.graph, retained);
1239		if (CCV_TENSOR_GET_DEVICE(info.type)((info.type) & 0xfff00) == CCV_COMPUTE_DEVICE_ANY)
1240			CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff
) << 8));
1241		const int device_id = CCV_TENSOR_GET_DEVICE_ID(info.type)(((info.type) & 0xfff00) >> 8);
1242		if (!compiled_data->tensors.internals[i])
1243			compiled_data->tensors.internals[i] = ccv_nnc_tensor_new(0, info, 0);
1244		for (j = 1; j < parallel_count; j++)
1245		{
1246			if (j != device_id)
1247				CCV_TENSOR_SET_DEVICE_ID(info.type, j)(info.type) = (((info.type) & ~0xfff00) | (((j) & 0xfff
) << 8));
1248			else
1249				CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff
) << 8));
1250			if (!compiled_data->tensors.internals[i + j * internal_size])
1251				compiled_data->tensors.internals[i + j * internal_size] = ccv_nnc_tensor_new(0, info, 0);
1252		}
1253	}
1254	compiled_data->tensors_init.v = CCV_NNC_INIT_V(compiled_data->tensors_init.v)((uint32_t*)((uintptr_t)(compiled_data->tensors_init.v) &
 ~(uintptr_t)1)); // Remove 1 if any.
1255}
1256 
1257static void _ccv_cnnp_model_tensors_init(const ccv_cnnp_model_t* const model, ccv_cnnp_compiled_data_t* const compiled_data)
1258{
1259	ccv_cnnp_model_tensors_init_0(model, compiled_data);
1260	ccv_cnnp_model_tensors_init_1(model, compiled_data);
1261}
1262 
1263static void _ccv_cnnp_model_copy_tensors(const uint32_t* const tensors_init, const ccv_nnc_tensor_symbol_t* const tensor_symbols, ccv_nnc_tensor_t* const* const tensors, const int tensor_size, const int parallel_count)
1264{
1265	assert(parallel_count > 0)((void) sizeof ((parallel_count > 0) ? 1 : 0), __extension__
 ({ if (parallel_count > 0) ; else __assert_fail ("parallel_count > 0"
, "ccv_cnnp_model.c", 1265, __extension__ __PRETTY_FUNCTION__
); }));
1266	int i, j;
1267	for (i = 0; i < tensor_size; i++)
1268	{
1269		if (!tensors[i])
1270			continue;
1271		const int d = tensor_symbols[i].d;
1272		if (!(tensors_init[d >> 5] & (1u << (d & 0x1f))))
1273			continue;
1274		for (j = 1; j < parallel_count; j++)
1275			if (tensors[i + j * tensor_size])
1276			{
1277				ccv_nnc_tensor_t* const input = CCV_NNC_TENSOR(tensors[i])((ccv_nnc_tensor_t*)((uintptr_t)(tensors[i]) & ~(uintptr_t
)1));
1278				ccv_nnc_tensor_t* const output = CCV_NNC_TENSOR(tensors[i + j * tensor_size])((ccv_nnc_tensor_t*)((uintptr_t)(tensors[i + j * tensor_size]
) & ~(uintptr_t)1));
1279				ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD()ccv_nnc_cmd(CCV_NNC_DATA_TRANSFER_FORWARD, 0, ccv_nnc_cmd_auto
, 0), ccv_nnc_no_hint, 0, &input, 1, &output, 1, 0);
1280			}
1281	}
1282}
1283 
1284static void _ccv_cnnp_model_remove_nocopies(const ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const tensor_symbols, ccv_nnc_tensor_t** const tensors, const int tensor_size, const int parallel_count)
1285{
1286	assert(parallel_count > 0)((void) sizeof ((parallel_count > 0) ? 1 : 0), __extension__
 ({ if (parallel_count > 0) ; else __assert_fail ("parallel_count > 0"
, "ccv_cnnp_model.c", 1286, __extension__ __PRETTY_FUNCTION__
); }));
1287	int i, j;
1288	for (i = 0; i < tensor_size; i++)
1289	{
1290		const ccv_nnc_tensor_symbol_t tensor_symbol = tensor_symbols[i];
1291		for (j = 1; j < parallel_count; j++)
1292		{
1293			const ccv_nnc_tensor_symbol_t copy = ccv_nnc_tensor_symbol_copy(graph, tensor_symbol, j);
1294			ccv_nnc_tensor_t* copy_tensor = tensors[i + j * tensor_size];
1295			if (copy_tensor && copy.d == CCV_NNC_NO_TENSOR_SYMBOL)
1296			{ // We shouldn't allocate this, free it up.
1297				ccv_nnc_tensor_free(tensors[i + j * tensor_size]);
1298				tensors[i + j * tensor_size] = 0;
1299			}
1300		}
1301	}
1302}
1303 
1304static void _ccv_cnnp_model_bind_tensors(const ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const tensor_symbols, ccv_nnc_tensor_t* const* const tensors, const int tensor_size, const int parallel_count, ccv_array_t* const tensor_binds)
1305{
1306	assert(parallel_count > 0)((void) sizeof ((parallel_count > 0) ? 1 : 0), __extension__
 ({ if (parallel_count > 0) ; else __assert_fail ("parallel_count > 0"
, "ccv_cnnp_model.c", 1306, __extension__ __PRETTY_FUNCTION__
); }));
1307	int i, j;
1308	for (i = 0; i < tensor_size; i++)
1309	{
1310		ccv_nnc_tensor_symbol_t tensor_symbol = tensor_symbols[i];
1311		if (tensor_symbol.d == CCV_NNC_NO_TENSOR_SYMBOL)
1312			continue;
1313		if (graph)
1314		{
1315			const ccv_nnc_tensor_symbol_t alias_to = ccv_nnc_tensor_symbol_alias_to(graph, tensor_symbol);
1316			if (alias_to.d != CCV_NNC_NO_TENSOR_SYMBOL)
1317				tensor_symbol = alias_to;
1318		}
1319		ccv_nnc_tensor_t* const tensor = CCV_NNC_TENSOR(tensors[i])((ccv_nnc_tensor_t*)((uintptr_t)(tensors[i]) & ~(uintptr_t
)1));
1320		if (tensor && tensor_symbol.d != CCV_NNC_NO_TENSOR_SYMBOL)
1321		{
1322			const ccv_nnc_tensor_bind_t retained_bind = {
1323				.symbol = tensor_symbol,
1324				.tensor = tensor
1325			};
1326			ccv_array_push(tensor_binds, &retained_bind);
1327		}
1328		for (j = 1; j < parallel_count; j++)
1329		{
1330			const ccv_nnc_tensor_symbol_t copy = ccv_nnc_tensor_symbol_copy(graph, tensor_symbol, j);
1331			ccv_nnc_tensor_t* copy_tensor = tensors[i + j * tensor_size];
1332			if (copy_tensor && copy.d != CCV_NNC_NO_TENSOR_SYMBOL)
1333			{
1334				const ccv_nnc_tensor_bind_t bind = {
1335					.symbol = copy,
1336					.tensor = tensors[i + j * tensor_size]
1337				};
1338				ccv_array_push(tensor_binds, &bind);
1339			}
1340		}
1341	}
1342}
1343 
1344static void _ccv_cnnp_compiled_data_graph_free(ccv_cnnp_compiled_data_t* const compiled_data)
1345{
1346	if (compiled_data->graph)
1347		ccv_nnc_graph_free(compiled_data->graph);
1348	compiled_data->graph = 0;
1349	compiled_data->is_test = 0;
1350	if (compiled_data->tensor_arena)
1351		ccv_nnc_tensor_arena_free(compiled_data->tensor_arena);
1352	compiled_data->tensor_arena = 0;
1353	if (compiled_data->graph_exec_arena)
1354		ccv_nnc_graph_exec_arena_free(compiled_data->graph_exec_arena);
1355	compiled_data->graph_exec_arena = 0;
1356	if (compiled_data->backward.from_ops)
1357		ccfreefree(compiled_data->backward.from_ops);
1358	compiled_data->backward.from_ops = 0;
1359	if (compiled_data->evaluate.schedule)
1360		ccv_nnc_graph_static_schedule_free(compiled_data->evaluate.schedule);
1361	compiled_data->evaluate.schedule = 0;
1362	if (compiled_data->backward.schedule)
1363		ccv_nnc_graph_static_schedule_free(compiled_data->backward.schedule);
1364	compiled_data->backward.schedule = 0;
1365}
1366 
1367static void _ccv_cnnp_compiled_data_gradient_free(ccv_cnnp_compiled_data_t* const compiled_data)
1368{
1369	if (compiled_data->gradients)
1370		ccfreefree(compiled_data->gradients);
1371	compiled_data->gradients = 0;
1372	if (compiled_data->updated_parameters)
1373		ccfreefree(compiled_data->updated_parameters);
1374	compiled_data->updated_parameters = 0;
1375	compiled_data->update_nodes = 0;
1376	compiled_data->saved_aux = 0;
1377}
1378 
1379static void _ccv_cnnp_compiled_data_backward_free(ccv_cnnp_compiled_data_t* const compiled_data)
1380{
1381	if (compiled_data->backward.gradients)
1382		ccfreefree(compiled_data->backward.gradients);
1383	compiled_data->backward.gradients = 0;
1384	if (compiled_data->backward.accum)
1385		ccv_nnc_graph_free(compiled_data->backward.accum);
1386	compiled_data->backward.accum = 0;
1387	if (compiled_data->backward.tensor_arena)
1388		ccv_nnc_tensor_arena_free(compiled_data->backward.tensor_arena);
1389	compiled_data->backward.tensor_arena = 0;
1390	if (compiled_data->backward.graph_exec_arena)
1391		ccv_nnc_graph_exec_arena_free(compiled_data->backward.graph_exec_arena);
1392	compiled_data->backward.graph_exec_arena = 0;
1393}
1394 
1395static void _ccv_cnnp_compiled_data_apply_gradients_free(ccv_cnnp_compiled_data_t* const compiled_data)
1396{
1397	if (compiled_data->apply_gradients.graph)
1398		ccv_nnc_graph_free(compiled_data->apply_gradients.graph);
1399	compiled_data->apply_gradients.graph = 0;
1400	if (compiled_data->apply_gradients.tensor_arena)
1401		ccv_nnc_tensor_arena_free(compiled_data->apply_gradients.tensor_arena);
1402	compiled_data->apply_gradients.tensor_arena = 0;
1403	if (compiled_data->apply_gradients.graph_exec_arena)
1404		ccv_nnc_graph_exec_arena_free(compiled_data->apply_gradients.graph_exec_arena);
1405	compiled_data->apply_gradients.graph_exec_arena = 0;
1406}
1407 
1408// Compile the graph to run ccv_cnnp_model_fit
1409static void _ccv_cnnp_model_fit_jit(ccv_cnnp_model_t* const model, ccv_nnc_tensor_t* const* const inputs, const int input_size, ccv_nnc_tensor_t* const* const fits, const int fit_size, ccv_nnc_tensor_t* const* const outputs, const int output_size)
1410{
1411	int i, j;
1412	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1413	assert(!compiled_data->graph || compiled_data->graph_mode != CCV_CNNP_MODEL_GRAPH_FIT_MODE)((void) sizeof ((!compiled_data->graph || compiled_data->
graph_mode != CCV_CNNP_MODEL_GRAPH_FIT_MODE) ? 1 : 0), __extension__
 ({ if (!compiled_data->graph || compiled_data->graph_mode
 != CCV_CNNP_MODEL_GRAPH_FIT_MODE) ; else __assert_fail ("!compiled_data->graph || compiled_data->graph_mode != CCV_CNNP_MODEL_GRAPH_FIT_MODE"
, "ccv_cnnp_model.c", 1413, __extension__ __PRETTY_FUNCTION__
); }));
1414	compiled_data->graph_mode = CCV_CNNP_MODEL_GRAPH_FIT_MODE;
1415	const int parallel_count = ccv_max(model->parallel_count, 1)({ typeof (model->parallel_count) _a = (model->parallel_count
); typeof (1) _b = (1); (_a > _b) ? _a : _b; });
1416	assert(output_size == model->output_size * parallel_count)((void) sizeof ((output_size == model->output_size * parallel_count
) ? 1 : 0), __extension__ ({ if (output_size == model->output_size
 * parallel_count) ; else __assert_fail ("output_size == model->output_size * parallel_count"
, "ccv_cnnp_model.c", 1416, __extension__ __PRETTY_FUNCTION__
); }));
1417	assert(!fits || output_size == fit_size)((void) sizeof ((!fits || output_size == fit_size) ? 1 : 0), __extension__
 ({ if (!fits || output_size == fit_size) ; else __assert_fail
 ("!fits || output_size == fit_size", "ccv_cnnp_model.c", 1417
, __extension__ __PRETTY_FUNCTION__); }));
1418	assert(output_size > 0)((void) sizeof ((output_size > 0) ? 1 : 0), __extension__ (
{ if (output_size > 0) ; else __assert_fail ("output_size > 0"
, "ccv_cnnp_model.c", 1418, __extension__ __PRETTY_FUNCTION__
); }));
1419	if (compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_NONE)
1420	{
1421		_ccv_cnnp_model_set_rewindables(model);
1422		_ccv_cnnp_model_gradient_init(model, CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES, CCV_CNNP_DISABLE_OUTGRAD_ALL, fits, fit_size);
1423	} else if (compiled_data->gradient_mode != CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES) {
1424		_ccv_cnnp_model_rewind_graph(model);
1425		_ccv_cnnp_compiled_data_gradient_free(compiled_data);
1426		compiled_data->gradient_mode = CCV_CNNP_COMPILED_DATA_GRADIENT_NONE;
1427		_ccv_cnnp_model_gradient_init(model, CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES, CCV_CNNP_DISABLE_OUTGRAD_ALL, fits, fit_size);
1428	}
1429	const int tensors_init = !!compiled_data->tensors_init.v;
1430	if (!tensors_init)
1431		_ccv_cnnp_model_tensors_init(model, compiled_data);
1432	else if ((uintptr_t)compiled_data->tensors_init.v & (uintptr_t)1)
1433	// Check if it is not fully allocated, if it is not, init_1.
1434		ccv_cnnp_model_tensors_init_1(model, compiled_data);
1435	ccv_array_t* const tensor_binds = ccv_array_new(sizeof(ccv_nnc_tensor_bind_t), 0, 0);
1436	assert((input_size % parallel_count) == 0)((void) sizeof (((input_size % parallel_count) == 0) ? 1 : 0)
, __extension__ ({ if ((input_size % parallel_count) == 0) ; else
 __assert_fail ("(input_size % parallel_count) == 0", "ccv_cnnp_model.c"
, 1436, __extension__ __PRETTY_FUNCTION__); }));
1437	assert((output_size % parallel_count) == 0)((void) sizeof (((output_size % parallel_count) == 0) ? 1 : 0
), __extension__ ({ if ((output_size % parallel_count) == 0) ;
 else __assert_fail ("(output_size % parallel_count) == 0", "ccv_cnnp_model.c"
, 1437, __extension__ __PRETTY_FUNCTION__); }));
1438	assert((fit_size % parallel_count) == 0)((void) sizeof (((fit_size % parallel_count) == 0) ? 1 : 0), __extension__
 ({ if ((fit_size % parallel_count) == 0) ; else __assert_fail
 ("(fit_size % parallel_count) == 0", "ccv_cnnp_model.c", 1438
, __extension__ __PRETTY_FUNCTION__); }));
1439	const int input_size_per_p = input_size / parallel_count;
1440	_ccv_cnnp_model_bind_tensors(model->graph, model->inputs, inputs, input_size_per_p, parallel_count, tensor_binds);
1441	const int output_size_per_p = output_size / parallel_count;
1442	_ccv_cnnp_model_bind_tensors(model->graph, model->outputs, outputs, output_size_per_p, parallel_count, tensor_binds);
1443	const int fit_size_per_p = fit_size / parallel_count;
1444	_ccv_cnnp_model_bind_tensors(model->graph, compiled_data->fits, fits, fit_size_per_p, parallel_count, tensor_binds);
1445	const int parameter_size = compiled_data->parameters->rnum;
1446	_ccv_cnnp_model_bind_tensors(model->graph, (ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->parameters, 0)((void*)(((char*)((compiled_data->parameters)->data)) +
 (size_t)(compiled_data->parameters)->rsize * (size_t)(
0))), compiled_data->tensors.parameters, parameter_size, parallel_count, tensor_binds);
1447	_ccv_cnnp_model_bind_tensors(model->graph, compiled_data->updated_parameters, compiled_data->tensors.parameters, parameter_size, parallel_count, tensor_binds);
1448	const int internal_size = compiled_data->internals->rnum;
1449	_ccv_cnnp_model_remove_nocopies(model->graph, (ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->internals, 0)((void*)(((char*)((compiled_data->internals)->data)) + (
size_t)(compiled_data->internals)->rsize * (size_t)(0))
), compiled_data->tensors.internals, internal_size, parallel_count);
1450	_ccv_cnnp_model_bind_tensors(model->graph, (ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->internals, 0)((void*)(((char*)((compiled_data->internals)->data)) + (
size_t)(compiled_data->internals)->rsize * (size_t)(0))
), compiled_data->tensors.internals, internal_size, parallel_count, tensor_binds);
1451	ccv_nnc_symbolic_graph_compile(model->graph, compiled_data->compile_params, (ccv_nnc_tensor_bind_t*)ccv_array_get(tensor_binds, 0)((void*)(((char*)((tensor_binds)->data)) + (size_t)(tensor_binds
)->rsize * (size_t)(0))), tensor_binds->rnum, 0, 0, SYMBOLIC_GRAPH_SOURCES(model->graph)ccv_nnc_symbolic_graph_sources(model->graph), ccv_nnc_symbolic_graph_source_size
(model->graph), SYMBOLIC_GRAPH_DESTINATIONS(model->graph)ccv_nnc_symbolic_graph_destinations(model->graph), ccv_nnc_symbolic_graph_destination_size
(model->graph), &compiled_data->graph, &compiled_data->tensor_arena, &compiled_data->graph_exec_arena);
1452	ccv_array_free(tensor_binds);
1453	const uint32_t* const init_v = CCV_NNC_INIT_V(compiled_data->tensors_init.v)((uint32_t*)((uintptr_t)(compiled_data->tensors_init.v) &
 ~(uintptr_t)1));
1454	if (tensors_init && parallel_count > 1)
1455		_ccv_cnnp_model_copy_tensors(init_v, (ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->parameters, 0)((void*)(((char*)((compiled_data->parameters)->data)) +
 (size_t)(compiled_data->parameters)->rsize * (size_t)(
0))), compiled_data->tensors.parameters, compiled_data->parameters->rnum, parallel_count);
1456	// If tensor is not init'ed, we need to init states first.
1457	if (_ccv_cnnp_any_to_init(compiled_data))
1458	{
1459		ccv_nnc_tensor_init_states_t tensor_init_states = {
1460			.parallel_count = parallel_count,
1461			.graph = model->graph,
1462			.compiled_data = compiled_data,
1463			.tensor_arena = compiled_data->tensor_arena
1464		};
1465		ccv_cnnp_model_init_states(model, model->graph, _ccv_cnnp_init_states_for_tensors, &tensor_init_states);
1466	}
1467	compiled_data->is_test = 0;
1468	const int saved_aux_size = ccv_nnc_minimizer_saved_aux_size(compiled_data->minimize.minimizer);
1469	// No need to set because it is default to training mode.
1470	// ccv_cnnp_model_set_is_test(model, 0, _ccv_cnnp_cmd_update_for_execs, &update);
1471	for (i = 0; i < saved_aux_size * parameter_size; i++)
1472	{
1473		if (compiled_data->saved_aux[i].source.d == CCV_NNC_NO_TENSOR_SYMBOL)
1474			continue;
1475		ccv_nnc_tensor_t* const tensor = ccv_nnc_tensor_from_symbol(compiled_data->tensor_arena, compiled_data->saved_aux[i].source);
1476		ccv_nnc_cmd_exec(CMD_SET_FORWARD(0)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={0,}}}, 0), ccv_nnc_no_hint, 0, 0, 0, &tensor, 1, 0);
1477		for (j = 1; j < parallel_count; j++)
1478		{
1479			ccv_nnc_tensor_t* const copy = ccv_nnc_tensor_from_symbol(compiled_data->tensor_arena, ccv_nnc_tensor_symbol_copy(model->graph, compiled_data->saved_aux[i].source, j));
1480			if (copy)
1481				ccv_nnc_cmd_exec(CMD_SET_FORWARD(0)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={0,}}}, 0), ccv_nnc_no_hint, 0, 0, 0, &copy, 1, 0);
1482		}
1483	}
1484	const int evaluate_to_size = compiled_data->evaluate.to_size;
1485	compiled_data->evaluate.to_op_size = 0;
1486	for (i = 0; i < evaluate_to_size; i++)
1487	{
1488		ccv_nnc_graph_exec_t const to = ccv_nnc_graph_exec_from_symbol(compiled_data->graph_exec_arena, compiled_data->evaluate.tos[i]);
1489		if (to.graph)
1490			compiled_data->evaluate.to_ops[compiled_data->evaluate.to_op_size++] = to;
1491	}
1492	ccv_nnc_graph_set_default_static_schedule(compiled_data->graph, compiled_data->stream_type, model->max_stream_count);
1493	ccv_nnc_graph_autotune(compiled_data->graph, model->workspace_size, 0, TRAVERSE_FULL0,0,0,0);
1494}
1495 
1496ccv_nnc_stream_context_t* ccv_cnnp_model_default_stream(const ccv_cnnp_model_t* const model)
1497{
1498	const ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1499	if (!compiled_data || !compiled_data->graph)
1500		return 0;
1501	return ccv_nnc_graph_default_stream(compiled_data->graph);
1502}
1503 
1504uint64_t ccv_cnnp_model_memory_size(const ccv_cnnp_model_t* const model)
1505{
1506	const ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1507	if (!compiled_data || !compiled_data->tensor_arena)
1508		return 0;
1509	return ccv_nnc_tensor_arena_size(compiled_data->tensor_arena);
1510}
1511 
1512static void _ccv_cnnp_bind_tensors_to_arena(ccv_nnc_tensor_arena_t* const tensor_arena, const ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const tensor_symbols, ccv_nnc_tensor_t* const* const tensors, const int tensor_size, const int parallel_count)
1513{
1514	int i, j;
1515	for (i = 0; i < tensor_size; i++)
1516	{
1517		ccv_nnc_tensor_symbol_t tensor_symbol = tensor_symbols[i];
1518		if (tensor_symbol.d == CCV_NNC_NO_TENSOR_SYMBOL)
1519			continue;
1520		if (graph)
1521		{
1522			const ccv_nnc_tensor_symbol_t alias_to = ccv_nnc_tensor_symbol_alias_to(graph, tensor_symbol);
1523			if (alias_to.d != CCV_NNC_NO_TENSOR_SYMBOL)
1524				tensor_symbol = alias_to;
1525		}
1526		ccv_nnc_tensor_bind_symbol(tensor_arena, tensor_symbol, tensors[i]);
1527		for (j = 1; j < parallel_count; j++)
1528		{
1529			const ccv_nnc_tensor_symbol_t copy = ccv_nnc_tensor_symbol_copy(graph, tensor_symbol, j);
1530			if (copy.d != CCV_NNC_NO_TENSOR_SYMBOL)
1531				ccv_nnc_tensor_bind_symbol(tensor_arena, copy, tensors[i + tensor_size * j]);
1532		}
1533	}
1534}
1535 
1536void ccv_cnnp_model_fit(ccv_cnnp_model_t* const model, ccv_nnc_tensor_t* const* const inputs, const int input_size, ccv_nnc_tensor_t* const* const fits, const int fit_size, ccv_nnc_tensor_t* const* const outputs, const int output_size, ccv_nnc_tensor_tape_t* const tensor_tape, ccv_nnc_stream_context_t* const stream_context)
1537{
1538	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1539	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 1539, __extension__ __PRETTY_FUNCTION__); }));
1540	const int parallel_count = ccv_max(model->parallel_count, 1)({ typeof (model->parallel_count) _a = (model->parallel_count
); typeof (1) _b = (1); (_a > _b) ? _a : _b; });
1541	assert(output_size == model->output_size * parallel_count)((void) sizeof ((output_size == model->output_size * parallel_count
) ? 1 : 0), __extension__ ({ if (output_size == model->output_size
 * parallel_count) ; else __assert_fail ("output_size == model->output_size * parallel_count"
, "ccv_cnnp_model.c", 1541, __extension__ __PRETTY_FUNCTION__
); }));
1542	assert(input_size == model->input_size * parallel_count)((void) sizeof ((input_size == model->input_size * parallel_count
) ? 1 : 0), __extension__ ({ if (input_size == model->input_size
 * parallel_count) ; else __assert_fail ("input_size == model->input_size * parallel_count"
, "ccv_cnnp_model.c", 1542, __extension__ __PRETTY_FUNCTION__
); }));
1543	assert(!fits || fit_size == output_size)((void) sizeof ((!fits || fit_size == output_size) ? 1 : 0), __extension__
 ({ if (!fits || fit_size == output_size) ; else __assert_fail
 ("!fits || fit_size == output_size", "ccv_cnnp_model.c", 1543
, __extension__ __PRETTY_FUNCTION__); }));
1544	assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if
 (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c"
, 1544, __extension__ __PRETTY_FUNCTION__); }));
1545	if (!compiled_data->graph || compiled_data->graph_mode != CCV_CNNP_MODEL_GRAPH_FIT_MODE)
1546	{
1547		_ccv_cnnp_compiled_data_graph_free(compiled_data);
1548		_ccv_cnnp_compiled_data_backward_free(compiled_data);
1549		_ccv_cnnp_compiled_data_apply_gradients_free(compiled_data);
1550		// Compile the symbolic graph down only when needed.
1551		_ccv_cnnp_model_fit_jit(model, inputs, input_size, fits, fit_size, outputs, output_size);
1552	} else {
1553		assert((input_size % parallel_count) == 0)((void) sizeof (((input_size % parallel_count) == 0) ? 1 : 0)
, __extension__ ({ if ((input_size % parallel_count) == 0) ; else
 __assert_fail ("(input_size % parallel_count) == 0", "ccv_cnnp_model.c"
, 1553, __extension__ __PRETTY_FUNCTION__); }));
1554		assert((output_size % parallel_count) == 0)((void) sizeof (((output_size % parallel_count) == 0) ? 1 : 0
), __extension__ ({ if ((output_size % parallel_count) == 0) ;
 else __assert_fail ("(output_size % parallel_count) == 0", "ccv_cnnp_model.c"
, 1554, __extension__ __PRETTY_FUNCTION__); }));
1555		assert((fit_size % parallel_count) == 0)((void) sizeof (((fit_size % parallel_count) == 0) ? 1 : 0), __extension__
 ({ if ((fit_size % parallel_count) == 0) ; else __assert_fail
 ("(fit_size % parallel_count) == 0", "ccv_cnnp_model.c", 1555
, __extension__ __PRETTY_FUNCTION__); }));
1556		const int input_size_per_p = input_size / parallel_count;
1557		_ccv_cnnp_bind_tensors_to_arena(compiled_data->tensor_arena, model->graph, model->inputs, inputs, input_size_per_p, parallel_count);
1558		const int output_size_per_p = output_size / parallel_count;
1559		_ccv_cnnp_bind_tensors_to_arena(compiled_data->tensor_arena, model->graph, model->outputs, outputs, output_size_per_p, parallel_count);
1560		const int fit_size_per_p = fit_size / parallel_count;
1561		_ccv_cnnp_bind_tensors_to_arena(compiled_data->tensor_arena, model->graph, compiled_data->fits, fits, fit_size_per_p, parallel_count);
1562	}
1563	if (compiled_data->is_test)
1564	{
1565		compiled_data->is_test = 0;
1566		ccv_nnc_graph_exec_update_t update = {
1567			.parallel_count = parallel_count,
1568			.graph = model->graph,
1569			.graph_exec_arena = compiled_data->graph_exec_arena,
1570		};
1571		ccv_cnnp_model_set_is_test(model, 0, _ccv_cnnp_cmd_update_for_execs, &update);
1572	}
1573	ccv_nnc_graph_run_with_schedule(compiled_data->graph, 0, 0, tensor_tape, stream_context);
1574}
1575 
1576// Compile the graph to run ccv_cnnp_model_evaluate with require_grad = false (MULTISTAGE_MODE_NO_GRAD).
1577static void _ccv_cnnp_model_multistage_no_grad_jit(ccv_cnnp_model_t* const model, ccv_nnc_tensor_t* const* const inputs, const int input_size, ccv_nnc_tensor_t* const* const outputs, const int output_size)
1578{
1579	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1580	compiled_data->graph_mode = CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE_NO_GRAD;
1581	const int parallel_count = ccv_max(model->parallel_count, 1)({ typeof (model->parallel_count) _a = (model->parallel_count
); typeof (1) _b = (1); (_a > _b) ? _a : _b; });
1582	assert(output_size == model->output_size * parallel_count)((void) sizeof ((output_size == model->output_size * parallel_count
) ? 1 : 0), __extension__ ({ if (output_size == model->output_size
 * parallel_count) ; else __assert_fail ("output_size == model->output_size * parallel_count"
, "ccv_cnnp_model.c", 1582, __extension__ __PRETTY_FUNCTION__
); }));
1583	assert(output_size > 0)((void) sizeof ((output_size > 0) ? 1 : 0), __extension__ (
{ if (output_size > 0) ; else __assert_fail ("output_size > 0"
, "ccv_cnnp_model.c", 1583, __extension__ __PRETTY_FUNCTION__
); }));
1584	// If the gradient is not initialized, continue to setup parallel process. We don't init gradient here, but rather,
1585	// we setup proper rewindables so the graph can be rewinded to previous state before we run data parallel.
1586	if (parallel_count > 1 && compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_NONE)
1587	{
1588		const int evaluate_to_size = compiled_data->evaluate.to_size;
1589		compiled_data->evaluate.tos = ccreallocrealloc(compiled_data->evaluate.tos, sizeof(ccv_nnc_graph_exec_symbol_t) * evaluate_to_size * parallel_count + sizeof(ccv_nnc_graph_exec_t) * evaluate_to_size * parallel_count);
1590		_ccv_cnnp_model_set_rewindables(model);
1591		ccv_nnc_symbolic_graph_data_parallel(model->graph, parallel_count,
1592			0, 0,
1593			0, 0, 0,
1594			0, 0, 0,
1595			CCV_NNC_PARALLEL_REDUCE_OP_SUM,
1596			SYMBOLIC_GRAPH_SOURCES(model->graph)ccv_nnc_symbolic_graph_sources(model->graph), ccv_nnc_symbolic_graph_source_size
(model->graph), SYMBOLIC_GRAPH_DESTINATIONS(model->graph)ccv_nnc_symbolic_graph_destinations(model->graph), ccv_nnc_symbolic_graph_destination_size
(model->graph));
1597		ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
1598		int i, j;
1599		for (i = 0; i < evaluate_to_size; i++)
1600			for (j = 1; j < parallel_count; j++)
1601			{
1602				const ccv_nnc_graph_exec_symbol_t copy = ccv_nnc_graph_exec_symbol_copy(model->graph, compiled_data->evaluate.tos[i], j);
1603				if (copy.d != CCV_NNC_NO_GRAPH_EXEC_SYMBOL)
1604					compiled_data->evaluate.tos[compiled_data->evaluate.to_size++] = copy;
1605			}
1606	}
1607	const int tensors_init = !!compiled_data->tensors_init.v;
1608	if (!tensors_init)
1609		_ccv_cnnp_model_tensors_init(model, compiled_data);
1610	else if ((uintptr_t)compiled_data->tensors_init.v & (uintptr_t)1)
1611	// Check if it is not fully allocated, if it is not, init_1.
1612		ccv_cnnp_model_tensors_init_1(model, compiled_data);
1613	ccv_array_t* const tensor_binds = ccv_array_new(sizeof(ccv_nnc_tensor_bind_t), 0, 0);
1614	assert((input_size % parallel_count) == 0)((void) sizeof (((input_size % parallel_count) == 0) ? 1 : 0)
, __extension__ ({ if ((input_size % parallel_count) == 0) ; else
 __assert_fail ("(input_size % parallel_count) == 0", "ccv_cnnp_model.c"
, 1614, __extension__ __PRETTY_FUNCTION__); }));
1615	assert((output_size % parallel_count) == 0)((void) sizeof (((output_size % parallel_count) == 0) ? 1 : 0
), __extension__ ({ if ((output_size % parallel_count) == 0) ;
 else __assert_fail ("(output_size % parallel_count) == 0", "ccv_cnnp_model.c"
, 1615, __extension__ __PRETTY_FUNCTION__); }));
1616	const int input_size_per_p = input_size / parallel_count;
1617	_ccv_cnnp_model_bind_tensors(model->graph, model->inputs, inputs, input_size_per_p, parallel_count, tensor_binds);
1618	const int output_size_per_p = output_size / parallel_count;
1619	_ccv_cnnp_model_bind_tensors(model->graph, model->outputs, outputs, output_size_per_p, parallel_count, tensor_binds);
1620	const int parameter_size = compiled_data->parameters->rnum;
1621	_ccv_cnnp_model_bind_tensors(model->graph, (ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->parameters, 0)((void*)(((char*)((compiled_data->parameters)->data)) +
 (size_t)(compiled_data->parameters)->rsize * (size_t)(
0))), compiled_data->tensors.parameters, parameter_size, parallel_count, tensor_binds);
1622	const int internal_size = compiled_data->internals->rnum;
1623	_ccv_cnnp_model_remove_nocopies(model->graph, (ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->internals, 0)((void*)(((char*)((compiled_data->internals)->data)) + (
size_t)(compiled_data->internals)->rsize * (size_t)(0))
), compiled_data->tensors.internals, internal_size, parallel_count);
1624	_ccv_cnnp_model_bind_tensors(model->graph, (ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->internals, 0)((void*)(((char*)((compiled_data->internals)->data)) + (
size_t)(compiled_data->internals)->rsize * (size_t)(0))
), compiled_data->tensors.internals, internal_size, parallel_count, tensor_binds);
1625	// If we generated gradient for the graph, only compile part of the graph because the rest is irrelevant for evaluation.
1626	ccv_nnc_symbolic_graph_compile(model->graph, compiled_data->compile_params, (ccv_nnc_tensor_bind_t*)ccv_array_get(tensor_binds, 0)((void*)(((char*)((tensor_binds)->data)) + (size_t)(tensor_binds
)->rsize * (size_t)(0))), tensor_binds->rnum, 0, 0, SYMBOLIC_GRAPH_SOURCES(model->graph)ccv_nnc_symbolic_graph_sources(model->graph), ccv_nnc_symbolic_graph_source_size
(model->graph), compiled_data->evaluate.tos, compiled_data->evaluate.to_size, &compiled_data->graph, &compiled_data->tensor_arena, &compiled_data->graph_exec_arena);
1627	ccv_array_free(tensor_binds);
1628	const uint32_t* const init_v = CCV_NNC_INIT_V(compiled_data->tensors_init.v)((uint32_t*)((uintptr_t)(compiled_data->tensors_init.v) &
 ~(uintptr_t)1));
1629	// If tensor is not init'ed, we need to init states first.
1630	if (tensors_init && parallel_count > 1)
1631		_ccv_cnnp_model_copy_tensors(init_v, (ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->parameters, 0)((void*)(((char*)((compiled_data->parameters)->data)) +
 (size_t)(compiled_data->parameters)->rsize * (size_t)(
0))), compiled_data->tensors.parameters, compiled_data->parameters->rnum, parallel_count);
1632	if (_ccv_cnnp_any_to_init(compiled_data))
1633	{
1634		ccv_nnc_tensor_init_states_t tensor_init_states = {
1635			.parallel_count = parallel_count,
1636			.graph = model->graph,
1637			.compiled_data = compiled_data,
1638			.tensor_arena = compiled_data->tensor_arena
1639		};
1640		ccv_cnnp_model_init_states(model, model->graph, _ccv_cnnp_init_states_for_tensors, &tensor_init_states);
1641	}
1642	compiled_data->is_test = 1;
1643	ccv_nnc_graph_exec_update_t update = {
1644		.parallel_count = parallel_count,
1645		.graph = model->graph,
1646		.graph_exec_arena = compiled_data->graph_exec_arena,
1647	};
1648	ccv_cnnp_model_set_is_test(model, 1, _ccv_cnnp_cmd_update_for_execs, &update);
1649	ccv_nnc_graph_set_default_static_schedule(compiled_data->graph, compiled_data->stream_type, model->max_stream_count);
1650	ccv_nnc_graph_autotune(compiled_data->graph, model->workspace_size, 0, TRAVERSE_FULL0,0,0,0);
1651}
1652 
1653static void _ccv_cnnp_model_gradient_tensors_init(const ccv_cnnp_model_t* const model, ccv_cnnp_compiled_data_t* const compiled_data)
1654{
1655	assert(!compiled_data->tensors.gradients)((void) sizeof ((!compiled_data->tensors.gradients) ? 1 : 0
), __extension__ ({ if (!compiled_data->tensors.gradients)
 ; else __assert_fail ("!compiled_data->tensors.gradients"
, "ccv_cnnp_model.c", 1655, __extension__ __PRETTY_FUNCTION__
); }));
1656	const int parameter_size = compiled_data->parameters->rnum;
1657	const int parallel_count = ccv_max(model->parallel_count, 1)({ typeof (model->parallel_count) _a = (model->parallel_count
); typeof (1) _b = (1); (_a > _b) ? _a : _b; });
1658	compiled_data->tensors.gradients = (ccv_nnc_tensor_t**)ccmallocmalloc(sizeof(ccv_nnc_tensor_t*) * parameter_size * 2 * parallel_count);
1659	compiled_data->tensors.accum_gradients = compiled_data->tensors.gradients + parameter_size * parallel_count;
1660	int i, j;
1661	for (i = 0; i < parameter_size; i++)
1662	{
1663		if (compiled_data->parameter_flags && !(compiled_data->parameter_flags[i >> 6] & ((uint64_t)1 << (i & 63))))
1664		{
1665			compiled_data->tensors.gradients[i] = 0;
1666			compiled_data->tensors.accum_gradients[i] = 0;
1667			for (j = 1; j < parallel_count; j++)
1668			{
1669				compiled_data->tensors.gradients[i + j * parameter_size] = 0;
1670				compiled_data->tensors.accum_gradients[i + j * parameter_size] = 0;
1671			}
1672			continue;
1673		}
1674		const ccv_nnc_tensor_symbol_t parameter = *(ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->parameters, i)((void*)(((char*)((compiled_data->parameters)->data)) +
 (size_t)(compiled_data->parameters)->rsize * (size_t)(
i)));
1675		ccv_nnc_tensor_param_t info = ccv_nnc_tensor_symbol_params(parameter.graph, parameter);
1676		if (CCV_TENSOR_GET_DEVICE(info.type)((info.type) & 0xfff00) == CCV_COMPUTE_DEVICE_ANY)
1677			CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff
) << 8));
1678		const int device_id = CCV_TENSOR_GET_DEVICE_ID(info.type)(((info.type) & 0xfff00) >> 8);
1679		compiled_data->tensors.gradients[i] = ccv_nnc_tensor_new(0, info, 0);
1680		compiled_data->tensors.accum_gradients[i] = 0; // delay the accumulated gradient allocation until when we need it.
1681		for (j = 1; j < parallel_count; j++)
1682		{
1683			if (j != device_id)
1684				CCV_TENSOR_SET_DEVICE_ID(info.type, j)(info.type) = (((info.type) & ~0xfff00) | (((j) & 0xfff
) << 8));
1685			else
1686				CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff
) << 8));
1687			compiled_data->tensors.gradients[i + j * parameter_size] = ccv_nnc_tensor_new(0, info, 0);
1688			compiled_data->tensors.accum_gradients[i + j * parameter_size] = 0;
1689		}
1690	}
1691}
1692 
1693static int _ccv_cnnp_is_disable_outgrad_all(const uint64_t disable_outgrad, const int input_size)
1694{
1695	if (disable_outgrad == CCV_CNNP_DISABLE_OUTGRAD_ALL)
1696		return 1;
1697	if (disable_outgrad == CCV_CNNP_DISABLE_OUTGRAD_NONE)
1698		return 0;
1699	int i;
1700	for (i = 0; i < input_size; i++)
1701		if (!(disable_outgrad & ((uint64_t)1 << i)))
1702			return 0;
1703	return 1;
1704}
1705 
1706// Compile the graph to run ccv_cnnp_model_evaluate with requires_grad = true (MULTISTAGE_MODE).
1707// Particularly, this method compiles the evaluation and backprop graph (the main graph).
1708static void _ccv_cnnp_model_multistage_jit_0(ccv_cnnp_model_t* const model, const uint64_t disable_outgrad, const int is_test, ccv_nnc_tensor_t* const* const inputs, const int input_size, ccv_nnc_tensor_t* const* const outputs, const int output_size)
1709{
1710	int i, j;
1711	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1712	const int target_gradient_mode = _ccv_cnnp_is_disable_outgrad_all(disable_outgrad, model->input_size) ? CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES : CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES_AND_INPUTS;
1713	assert(!compiled_data->graph || compiled_data->graph_mode != CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE || compiled_data->gradient_mode != target_gradient_mode)((void) sizeof ((!compiled_data->graph || compiled_data->
graph_mode != CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE || compiled_data
->gradient_mode != target_gradient_mode) ? 1 : 0), __extension__
 ({ if (!compiled_data->graph || compiled_data->graph_mode
 != CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE || compiled_data->
gradient_mode != target_gradient_mode) ; else __assert_fail (
"!compiled_data->graph || compiled_data->graph_mode != CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE || compiled_data->gradient_mode != target_gradient_mode"
, "ccv_cnnp_model.c", 1713, __extension__ __PRETTY_FUNCTION__
); }));
1714	compiled_data->graph_mode = CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE;
1715	const int parallel_count = ccv_max(model->parallel_count, 1)({ typeof (model->parallel_count) _a = (model->parallel_count
); typeof (1) _b = (1); (_a > _b) ? _a : _b; });
1716	assert(output_size == model->output_size * parallel_count)((void) sizeof ((output_size == model->output_size * parallel_count
) ? 1 : 0), __extension__ ({ if (output_size == model->output_size
 * parallel_count) ; else __assert_fail ("output_size == model->output_size * parallel_count"
, "ccv_cnnp_model.c", 1716, __extension__ __PRETTY_FUNCTION__
); }));
1717	assert(output_size > 0)((void) sizeof ((output_size > 0) ? 1 : 0), __extension__ (
{ if (output_size > 0) ; else __assert_fail ("output_size > 0"
, "ccv_cnnp_model.c", 1717, __extension__ __PRETTY_FUNCTION__
); }));
1718	// There shouldn't be a loss function if we evaluate with multistage jit.
1719	assert(compiled_data->loss.cmd == CCV_NNC_NOOP)((void) sizeof ((compiled_data->loss.cmd == CCV_NNC_NOOP) ?
 1 : 0), __extension__ ({ if (compiled_data->loss.cmd == CCV_NNC_NOOP
) ; else __assert_fail ("compiled_data->loss.cmd == CCV_NNC_NOOP"
, "ccv_cnnp_model.c", 1719, __extension__ __PRETTY_FUNCTION__
); }));
1720	if (compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_NONE)
1721	{
1722		_ccv_cnnp_model_set_rewindables(model);
1723		_ccv_cnnp_model_gradient_init(model, target_gradient_mode, disable_outgrad, 0, 0); // The type of outputs and fits should be the same. We only use type here.
1724	} else if (compiled_data->gradient_mode != target_gradient_mode) {
1725		_ccv_cnnp_model_rewind_graph(model);
1726		_ccv_cnnp_compiled_data_gradient_free(compiled_data);
1727		compiled_data->gradient_mode = CCV_CNNP_COMPILED_DATA_GRADIENT_NONE;
1728		_ccv_cnnp_model_gradient_init(model, target_gradient_mode, disable_outgrad, 0, 0); // The type of outputs and fits should be the same. We only use type here.
1729	}
1730	const int tensors_init = !!compiled_data->tensors_init.v;
1731	if (!tensors_init)
1732		_ccv_cnnp_model_tensors_init(model, compiled_data);
1733	else if ((uintptr_t)compiled_data->tensors_init.v & (uintptr_t)1)
1734	// Check if it is not fully allocated, if it is not, init_1.
1735		ccv_cnnp_model_tensors_init_1(model, compiled_data);
1736	ccv_array_t* const tensor_binds = ccv_array_new(sizeof(ccv_nnc_tensor_bind_t), 0, 0);
1737	assert((input_size % parallel_count) == 0)((void) sizeof (((input_size % parallel_count) == 0) ? 1 : 0)
, __extension__ ({ if ((input_size % parallel_count) == 0) ; else
 __assert_fail ("(input_size % parallel_count) == 0", "ccv_cnnp_model.c"
, 1737, __extension__ __PRETTY_FUNCTION__); }));
1738	assert((output_size % parallel_count) == 0)((void) sizeof (((output_size % parallel_count) == 0) ? 1 : 0
), __extension__ ({ if ((output_size % parallel_count) == 0) ;
 else __assert_fail ("(output_size % parallel_count) == 0", "ccv_cnnp_model.c"
, 1738, __extension__ __PRETTY_FUNCTION__); }));
1739	const int input_size_per_p = input_size / parallel_count;
1740	_ccv_cnnp_model_bind_tensors(model->graph, model->inputs, inputs, input_size_per_p, parallel_count, tensor_binds);
1741	const int output_size_per_p = output_size / parallel_count;
1742	_ccv_cnnp_model_bind_tensors(model->graph, model->outputs, outputs, output_size_per_p, parallel_count, tensor_binds);
1743	const int parameter_size = compiled_data->parameters->rnum;
1744	_ccv_cnnp_model_bind_tensors(model->graph, (ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->parameters, 0)((void*)(((char*)((compiled_data->parameters)->data)) +
 (size_t)(compiled_data->parameters)->rsize * (size_t)(
0))), compiled_data->tensors.parameters, parameter_size, parallel_count, tensor_binds);
1745	const int internal_size = compiled_data->internals->rnum;
1746	_ccv_cnnp_model_remove_nocopies(model->graph, (ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->internals, 0)((void*)(((char*)((compiled_data->internals)->data)) + (
size_t)(compiled_data->internals)->rsize * (size_t)(0))
), compiled_data->tensors.internals, internal_size, parallel_count);
1747	_ccv_cnnp_model_bind_tensors(model->graph, (ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->internals, 0)((void*)(((char*)((compiled_data->internals)->data)) + (
size_t)(compiled_data->internals)->rsize * (size_t)(0))
), compiled_data->tensors.internals, internal_size, parallel_count, tensor_binds);
1748	if (!compiled_data->tensors.gradients)
1749		_ccv_cnnp_model_gradient_tensors_init(model, compiled_data);
1750	_ccv_cnnp_model_bind_tensors(model->graph, compiled_data->gradients, compiled_data->tensors.gradients, parameter_size, parallel_count, tensor_binds);
1751	if (compiled_data->backward.to_size > 0)
1752		ccv_nnc_symbolic_graph_compile(model->graph, compiled_data->compile_params, (ccv_nnc_tensor_bind_t*)ccv_array_get(tensor_binds, 0)((void*)(((char*)((tensor_binds)->data)) + (size_t)(tensor_binds
)->rsize * (size_t)(0))), tensor_binds->rnum, 0, 0, SYMBOLIC_GRAPH_SOURCES(model->graph)ccv_nnc_symbolic_graph_sources(model->graph), ccv_nnc_symbolic_graph_source_size
(model->graph), compiled_data->backward.tos, compiled_data->backward.to_size, &compiled_data->graph, &compiled_data->tensor_arena, &compiled_data->graph_exec_arena);
1753	else
1754		ccv_nnc_symbolic_graph_compile(model->graph, compiled_data->compile_params, (ccv_nnc_tensor_bind_t*)ccv_array_get(tensor_binds, 0)((void*)(((char*)((tensor_binds)->data)) + (size_t)(tensor_binds
)->rsize * (size_t)(0))), tensor_binds->rnum, 0, 0, SYMBOLIC_GRAPH_SOURCES(model->graph)ccv_nnc_symbolic_graph_sources(model->graph), ccv_nnc_symbolic_graph_source_size
(model->graph), compiled_data->evaluate.tos, compiled_data->evaluate.to_size, &compiled_data->graph, &compiled_data->tensor_arena, &compiled_data->graph_exec_arena);
1755	ccv_array_free(tensor_binds);
1756	const uint32_t* const init_v = CCV_NNC_INIT_V(compiled_data->tensors_init.v)((uint32_t*)((uintptr_t)(compiled_data->tensors_init.v) &
 ~(uintptr_t)1));
1757	if (tensors_init && parallel_count > 1)
1758		_ccv_cnnp_model_copy_tensors(init_v, (ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->parameters, 0)((void*)(((char*)((compiled_data->parameters)->data)) +
 (size_t)(compiled_data->parameters)->rsize * (size_t)(
0))), compiled_data->tensors.parameters, compiled_data->parameters->rnum, parallel_count);
1759	// If tensor is not init'ed, we need to init states first.
1760	if (_ccv_cnnp_any_to_init(compiled_data))
1761	{
1762		ccv_nnc_tensor_init_states_t tensor_init_states = {
1763			.parallel_count = parallel_count,
1764			.graph = model->graph,
1765			.compiled_data = compiled_data,
1766			.tensor_arena = compiled_data->tensor_arena
1767		};
1768		ccv_cnnp_model_init_states(model, model->graph, _ccv_cnnp_init_states_for_tensors, &tensor_init_states);
1769	}
1770	compiled_data->is_test = is_test;
1771	ccv_nnc_graph_exec_update_t update = {
1772		.parallel_count = parallel_count,
1773		.graph = model->graph,
1774		.graph_exec_arena = compiled_data->graph_exec_arena,
1775	};
1776	ccv_cnnp_model_set_is_test(model, is_test, _ccv_cnnp_cmd_update_for_execs, &update);
1777	const int evaluate_to_size = compiled_data->evaluate.to_size;
1778	compiled_data->evaluate.to_op_size = 0;
1779	ccv_array_t* const backward_from = ccv_array_new(sizeof(int), 0, 0);
1780	for (i = 0; i < evaluate_to_size; i++)
1781	{
1782		ccv_nnc_graph_exec_t const to_op = ccv_nnc_graph_exec_from_symbol(compiled_data->graph_exec_arena, compiled_data->evaluate.tos[i]);
1783		if (to_op.graph)
1784			compiled_data->evaluate.to_ops[compiled_data->evaluate.to_op_size++] = to_op;
1785		const int* tos;
1786		int to_size;
1787		ccv_nnc_graph_exec_symbol_to(model->graph, compiled_data->evaluate.tos[i], &tos, &to_size);
1788		for (j = 0; j < to_size; j++)
1789		{
1790			ccv_nnc_graph_exec_t const to_op = ccv_nnc_graph_exec_from_symbol(compiled_data->graph_exec_arena, (ccv_nnc_graph_exec_symbol_t){
1791				.d = tos[j],
1792				.graph = model->graph
1793			});
1794			if (to_op.graph)
1795				ccv_array_add_unique_int(backward_from, to_op.d);
1796		}
1797	}
1798	assert(backward_from->rnum > 0)((void) sizeof ((backward_from->rnum > 0) ? 1 : 0), __extension__
 ({ if (backward_from->rnum > 0) ; else __assert_fail (
"backward_from->rnum > 0", "ccv_cnnp_model.c", 1798, __extension__
 __PRETTY_FUNCTION__); }));
1799	compiled_data->backward.from_op_size = backward_from->rnum;
1800	compiled_data->backward.from_ops = (ccv_nnc_graph_exec_t*)ccmallocmalloc(sizeof(ccv_nnc_graph_exec_t) * backward_from->rnum);
1801	for (i = 0; i < backward_from->rnum; i++)
1802		compiled_data->backward.from_ops[i] = (ccv_nnc_graph_exec_t){
1803			.d = *(int*)ccv_array_get(backward_from, i)((void*)(((char*)((backward_from)->data)) + (size_t)(backward_from
)->rsize * (size_t)(i))),
1804			.graph = compiled_data->graph,
1805		};
1806	// If there are any set node (to set some tensors to 0) inserted through backward pass, these won't be executed if we just do sources -> evaluate.to_ops, backward.from_ops -> destinations. We need this logic to find out these nodes and explicitly adding them to backward.from_ops.
1807	ccv_nnc_graph_exec_info_t* const exec_info = (ccv_nnc_graph_exec_info_t*)ccv_array_get(compiled_data->graph->exec_info, 0)((void*)(((char*)((compiled_data->graph->exec_info)->
data)) + (size_t)(compiled_data->graph->exec_info)->
rsize * (size_t)(0)));
1808	const int exec_info_size = compiled_data->graph->exec_info->rnum;
1809	uint32_t* const visited = cccalloccalloc((exec_info_size + 31) >> 5, sizeof(uint32_t));
1810	const ccv_nnc_graph_exec_t* const sources = (ccv_nnc_graph_exec_t*)ccv_array_get(compiled_data->graph->sources, 0)((void*)(((char*)((compiled_data->graph->sources)->data
)) + (size_t)(compiled_data->graph->sources)->rsize *
 (size_t)(0)));
1811	const int source_size = compiled_data->graph->sources->rnum;
1812	ccv_nnc_graph_visit_t* visit = ccv_nnc_graph_visit_new(compiled_data->graph, exec_info, exec_info_size, sources, source_size, compiled_data->evaluate.to_ops, compiled_data->evaluate.to_op_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 == compiled_data->graph) ? 1 : 0), __extension__
 ({ if ((sources)[_i_].graph == compiled_data->graph) ; else
 __assert_fail ("(sources)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1812, __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_cnnp_model.c", 1812, __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 == compiled_data->graph) ? 1 : 0), __extension__
 ({ if ((sources)[_i_].graph == compiled_data->graph) ; else
 __assert_fail ("(sources)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1812, __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_cnnp_model.c", 1812, __extension__ __PRETTY_FUNCTION__
); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[
_q_]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for
 (_i_ = 0; _i_ < (compiled_data->evaluate.to_op_size); _i_
++) { ((void) sizeof (((compiled_data->evaluate.to_ops)[_i_
].graph == compiled_data->graph) ? 1 : 0), __extension__ (
{ if ((compiled_data->evaluate.to_ops)[_i_].graph == compiled_data
->graph) ; else __assert_fail ("(compiled_data->evaluate.to_ops)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1812, __extension__ __PRETTY_FUNCTION__
); })); _incomings_[(compiled_data->evaluate.to_ops)[_i_].
d].r = 5; _exists_[0][_i_] = (compiled_data->evaluate.to_ops
)[_i_].d; } _exist_size_[0] = (compiled_data->evaluate.to_op_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_cnnp_model.c", 1812, __extension__ __PRETTY_FUNCTION__
); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[
_q_]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for
 (_i_ = 0; _i_ < (compiled_data->evaluate.to_op_size); _i_
++) { ((void) sizeof (((compiled_data->evaluate.to_ops)[_i_
].graph == compiled_data->graph) ? 1 : 0), __extension__ (
{ if ((compiled_data->evaluate.to_ops)[_i_].graph == compiled_data
->graph) ; else __assert_fail ("(compiled_data->evaluate.to_ops)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1812, __extension__ __PRETTY_FUNCTION__
); })); _incomings_[(compiled_data->evaluate.to_ops)[_i_].
d].d = 1; } for (_i_ = 0; _i_ < (source_size); _i_++) { ((
void) sizeof (((sources)[_i_].graph == compiled_data->graph
) ? 1 : 0), __extension__ ({ if ((sources)[_i_].graph == compiled_data
->graph) ; else __assert_fail ("(sources)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1812, __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_ <
 (compiled_data->evaluate.to_op_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_ < (compiled_data->evaluate.to_op_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_cnnp_model.c", 1812, __extension__ __PRETTY_FUNCTION__
); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[
_q_]; } } } ++_i_; } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (
_i_)); } for (_i_ = 0; _i_ < (compiled_data->evaluate.to_op_size
); _i_++) { ((void) sizeof (((compiled_data->evaluate.to_ops
)[_i_].graph == compiled_data->graph) ? 1 : 0), __extension__
 ({ if ((compiled_data->evaluate.to_ops)[_i_].graph == compiled_data
->graph) ; else __assert_fail ("(compiled_data->evaluate.to_ops)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1812, __extension__ __PRETTY_FUNCTION__
); })); if (_incomings_[(compiled_data->evaluate.to_ops)[_i_
].d].r == 7) continue; if (!(0)) { ((void) sizeof ((_incomings_
[(compiled_data->evaluate.to_ops)[_i_].d].c == 0) ? 1 : 0)
, __extension__ ({ if (_incomings_[(compiled_data->evaluate
.to_ops)[_i_].d].c == 0) ; else __assert_fail ("_incomings_[(compiled_data->evaluate.to_ops)[_i_].d].c == 0"
, "ccv_cnnp_model.c", 1812, __extension__ __PRETTY_FUNCTION__
); })); } else if (_incomings_[(compiled_data->evaluate.to_ops
)[_i_].d].c > 0) continue; _visit_->node[_visit_->size
].index = (((compiled_data->evaluate.to_ops)[_i_].d)); _visit_
->node[_visit_->size].term = ((_incomings_[(compiled_data
->evaluate.to_ops)[_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_cnnp_model.c"
, 1812, __extension__ __PRETTY_FUNCTION__); })); _visit_; });
1813	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; {
1814		visited[(idx >> 5)] |= (1u << (idx & 31));
1815	} ccv_nnc_graph_visit_endfor} }
1816	ccv_nnc_graph_visit_free(visit);
1817	const ccv_nnc_graph_exec_t* const destinations = (ccv_nnc_graph_exec_t*)ccv_array_get(compiled_data->graph->destinations, 0)((void*)(((char*)((compiled_data->graph->destinations)->
data)) + (size_t)(compiled_data->graph->destinations)->
rsize * (size_t)(0)));
1818	const int destination_size = compiled_data->graph->destinations->rnum;
1819	visit = ccv_nnc_graph_visit_new(compiled_data->graph, exec_info, exec_info_size, compiled_data->backward.from_ops, compiled_data->backward.from_op_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_ < (compiled_data->backward.from_op_size)
; _i_++) { ((void) sizeof (((compiled_data->backward.from_ops
)[_i_].graph == compiled_data->graph) ? 1 : 0), __extension__
 ({ if ((compiled_data->backward.from_ops)[_i_].graph == compiled_data
->graph) ; else __assert_fail ("(compiled_data->backward.from_ops)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1819, __extension__ __PRETTY_FUNCTION__
); })); _incomings_[(compiled_data->backward.from_ops)[_i_
].d].r = 1; _exists_[0][_i_] = (compiled_data->backward.from_ops
)[_i_].d; } int _exist_size_[2] = { (compiled_data->backward
.from_op_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_cnnp_model.c", 1819, __extension__ __PRETTY_FUNCTION__
); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[
_q_]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for
 (_i_ = 0; _i_ < (compiled_data->backward.from_op_size)
; _i_++) { ((void) sizeof (((compiled_data->backward.from_ops
)[_i_].graph == compiled_data->graph) ? 1 : 0), __extension__
 ({ if ((compiled_data->backward.from_ops)[_i_].graph == compiled_data
->graph) ; else __assert_fail ("(compiled_data->backward.from_ops)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1819, __extension__ __PRETTY_FUNCTION__
); })); _incomings_[(compiled_data->backward.from_ops)[_i_
].d].r = 3; _exists_[0][_i_] = (compiled_data->backward.from_ops
)[_i_].d; } _exist_size_[0] = (compiled_data->backward.from_op_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_cnnp_model.c"
, 1819, __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 == compiled_data->graph) ? 1 : 0), __extension__
 ({ if ((destinations)[_i_].graph == compiled_data->graph)
 ; else __assert_fail ("(destinations)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1819, __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_cnnp_model.c", 1819, __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 == compiled_data->graph) ? 1 :
 0), __extension__ ({ if ((destinations)[_i_].graph == compiled_data
->graph) ; else __assert_fail ("(destinations)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1819, __extension__ __PRETTY_FUNCTION__
); })); _incomings_[(destinations)[_i_].d].d = 1; } for (_i_ =
 0; _i_ < (compiled_data->backward.from_op_size); _i_++
) { ((void) sizeof (((compiled_data->backward.from_ops)[_i_
].graph == compiled_data->graph) ? 1 : 0), __extension__ (
{ if ((compiled_data->backward.from_ops)[_i_].graph == compiled_data
->graph) ; else __assert_fail ("(compiled_data->backward.from_ops)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1819, __extension__ __PRETTY_FUNCTION__
); })); _exists_[0][_i_] = (compiled_data->backward.from_ops
)[_i_].d; } _p_ = 0; _q_ = 1; _exist_size_[0] = (compiled_data
->backward.from_op_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_cnnp_model.c", 1819, __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 == compiled_data->
graph) ? 1 : 0), __extension__ ({ if ((destinations)[_i_].graph
 == compiled_data->graph) ; else __assert_fail ("(destinations)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1819, __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_cnnp_model.c", 1819, __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_cnnp_model.c", 1819, __extension__ __PRETTY_FUNCTION__
); })); _visit_; });
1820	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; {
1821		visited[(idx >> 5)] |= (1u << (idx & 31));
1822	} ccv_nnc_graph_visit_endfor} }
1823	ccv_nnc_graph_visit_free(visit);
1824	visit = ccv_nnc_graph_visit_new(compiled_data->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 == compiled_data->graph) ? 1 : 0), __extension__
 ({ if ((sources)[_i_].graph == compiled_data->graph) ; else
 __assert_fail ("(sources)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1824, __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_cnnp_model.c", 1824, __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 == compiled_data->graph) ? 1 : 0), __extension__
 ({ if ((sources)[_i_].graph == compiled_data->graph) ; else
 __assert_fail ("(sources)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1824, __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_cnnp_model.c", 1824, __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 == compiled_data->graph) ? 1 :
 0), __extension__ ({ if ((destinations)[_i_].graph == compiled_data
->graph) ; else __assert_fail ("(destinations)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1824, __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_cnnp_model.c", 1824, __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 == compiled_data->graph) ? 1 :
 0), __extension__ ({ if ((destinations)[_i_].graph == compiled_data
->graph) ; else __assert_fail ("(destinations)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1824, __extension__ __PRETTY_FUNCTION__
); })); _incomings_[(destinations)[_i_].d].d = 1; } for (_i_ =
 0; _i_ < (source_size); _i_++) { ((void) sizeof (((sources
)[_i_].graph == compiled_data->graph) ? 1 : 0), __extension__
 ({ if ((sources)[_i_].graph == compiled_data->graph) ; else
 __assert_fail ("(sources)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1824, __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_cnnp_model.c", 1824, __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 == compiled_data->
graph) ? 1 : 0), __extension__ ({ if ((destinations)[_i_].graph
 == compiled_data->graph) ; else __assert_fail ("(destinations)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1824, __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_cnnp_model.c", 1824, __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_cnnp_model.c", 1824, __extension__ __PRETTY_FUNCTION__
); })); _visit_; });
1825	// Find any missing nodes to be added as source. Right now, these are only set nodes.
1826	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; {
1827		if (!(visited[(idx >> 5)] & (1u << (idx & 31))))
1828		{
1829			assert(exec_info[idx].cmd.cmd == CCV_NNC_SET_FORWARD)((void) sizeof ((exec_info[idx].cmd.cmd == CCV_NNC_SET_FORWARD
) ? 1 : 0), __extension__ ({ if (exec_info[idx].cmd.cmd == CCV_NNC_SET_FORWARD
) ; else __assert_fail ("exec_info[idx].cmd.cmd == CCV_NNC_SET_FORWARD"
, "ccv_cnnp_model.c", 1829, __extension__ __PRETTY_FUNCTION__
); }));
1830			if (exec_info[idx].cmd.info.blas.a[0] == 0) // Special-casing for empty out the tensor set function, not for the set grad to 1 one.
1831				ccv_array_add_unique_int(backward_from, idx);
1832		}
1833	} ccv_nnc_graph_visit_endfor} }
1834	ccv_nnc_graph_visit_free(visit);
1835	ccfreefree(visited);
1836	if (backward_from->rnum != compiled_data->backward.from_op_size) // If it doesn't match, need to redo this.
1837	{
1838		compiled_data->backward.from_op_size = backward_from->rnum;
1839		compiled_data->backward.from_ops = (ccv_nnc_graph_exec_t*)ccreallocrealloc(compiled_data->backward.from_ops, sizeof(ccv_nnc_graph_exec_t) * backward_from->rnum);
1840		for (i = 0; i < backward_from->rnum; i++)
1841			compiled_data->backward.from_ops[i] = (ccv_nnc_graph_exec_t){
1842				.d = *(int*)ccv_array_get(backward_from, i)((void*)(((char*)((backward_from)->data)) + (size_t)(backward_from
)->rsize * (size_t)(i))),
1843				.graph = compiled_data->graph,
1844			};
1845	}
1846	ccv_array_free(backward_from);
1847	ccv_nnc_graph_set_default_static_schedule(compiled_data->graph, compiled_data->stream_type, model->max_stream_count);
1848	ccv_nnc_graph_autotune(compiled_data->graph, model->workspace_size, 0, TRAVERSE_FULL0,0,0,0);
1849}
1850 
1851void ccv_cnnp_model_dry_run(ccv_cnnp_model_t* const model, const ccv_cnnp_evaluate_param_t params, ccv_nnc_tensor_t* const* const inputs, const int input_size, ccv_nnc_tensor_t* const* const outputs, const int output_size)
1852{
1853	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1854	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 1854, __extension__ __PRETTY_FUNCTION__); }));
1855	const int parallel_count = ccv_max(model->parallel_count, 1)({ typeof (model->parallel_count) _a = (model->parallel_count
); typeof (1) _b = (1); (_a > _b) ? _a : _b; });
1856	assert(output_size == model->output_size * parallel_count)((void) sizeof ((output_size == model->output_size * parallel_count
) ? 1 : 0), __extension__ ({ if (output_size == model->output_size
 * parallel_count) ; else __assert_fail ("output_size == model->output_size * parallel_count"
, "ccv_cnnp_model.c", 1856, __extension__ __PRETTY_FUNCTION__
); }));
1857	assert(input_size == model->input_size * parallel_count)((void) sizeof ((input_size == model->input_size * parallel_count
) ? 1 : 0), __extension__ ({ if (input_size == model->input_size
 * parallel_count) ; else __assert_fail ("input_size == model->input_size * parallel_count"
, "ccv_cnnp_model.c", 1857, __extension__ __PRETTY_FUNCTION__
); }));
1858	assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if
 (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c"
, 1858, __extension__ __PRETTY_FUNCTION__); }));
1859	const int target_gradient_mode = _ccv_cnnp_is_disable_outgrad_all(params.disable_outgrad, model->input_size) ? CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES : CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES_AND_INPUTS;
1860	const int mode_mismatch = (params.requires_grad && (compiled_data->graph_mode != CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE || compiled_data->gradient_mode != target_gradient_mode || compiled_data->disable_outgrad != params.disable_outgrad));
1861	if (!compiled_data->graph || mode_mismatch)
1862	{
1863		_ccv_cnnp_compiled_data_graph_free(compiled_data);
1864		if (mode_mismatch) // If mode mismatch, we need to redo the backward as well (no need to redo apply_gradients, it doesn't require target_gradient_mode or disable_outgrad.
1865			_ccv_cnnp_compiled_data_backward_free(compiled_data);
1866		if (params.requires_grad)
1867			_ccv_cnnp_model_multistage_jit_0(model, params.disable_outgrad, params.is_test, inputs, input_size, outputs, output_size);
1868		else
1869			_ccv_cnnp_model_multistage_no_grad_jit(model, inputs, input_size, outputs, output_size);
1870	} else {
1871		ccv_nnc_tensor_arena_clear_bindings(compiled_data->tensor_arena);
1872		assert((input_size % parallel_count) == 0)((void) sizeof (((input_size % parallel_count) == 0) ? 1 : 0)
, __extension__ ({ if ((input_size % parallel_count) == 0) ; else
 __assert_fail ("(input_size % parallel_count) == 0", "ccv_cnnp_model.c"
, 1872, __extension__ __PRETTY_FUNCTION__); }));
1873		const int input_size_per_p = input_size / parallel_count;
1874		_ccv_cnnp_bind_tensors_to_arena(compiled_data->tensor_arena, model->graph, model->inputs, inputs, input_size_per_p, parallel_count);
1875		assert((output_size % parallel_count) == 0)((void) sizeof (((output_size % parallel_count) == 0) ? 1 : 0
), __extension__ ({ if ((output_size % parallel_count) == 0) ;
 else __assert_fail ("(output_size % parallel_count) == 0", "ccv_cnnp_model.c"
, 1875, __extension__ __PRETTY_FUNCTION__); }));
1876		const int output_size_per_p = output_size / parallel_count;
1877		_ccv_cnnp_bind_tensors_to_arena(compiled_data->tensor_arena, model->graph, model->outputs, outputs, output_size_per_p, parallel_count);
1878	}
1879	if (compiled_data->is_test != params.is_test)
1880	{
1881		compiled_data->is_test = params.is_test;
1882		ccv_nnc_graph_exec_update_t update = {
1883			.parallel_count = parallel_count,
1884			.graph = model->graph,
1885			.graph_exec_arena = compiled_data->graph_exec_arena,
1886		};
1887		ccv_cnnp_model_set_is_test(model, params.is_test, _ccv_cnnp_cmd_update_for_execs, &update);
1888	}
1889}
1890 
1891void ccv_cnnp_model_evaluate(ccv_cnnp_model_t* const model, const ccv_cnnp_evaluate_param_t params, ccv_nnc_tensor_t* const* const inputs, const int input_size, ccv_nnc_tensor_t* const* const outputs, const int output_size, ccv_nnc_tensor_tape_t* const tensor_tape, ccv_nnc_stream_context_t* const stream_context)
1892{
1893	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1894	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 1894, __extension__ __PRETTY_FUNCTION__); }));
1895	ccv_cnnp_model_dry_run(model, params, inputs, input_size, outputs, output_size);
1896	if (compiled_data->graph_mode == CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE_NO_GRAD)
1897		ccv_nnc_graph_run_with_schedule(compiled_data->graph, 0, 0, tensor_tape, stream_context);
1898	else {
1899		if (!compiled_data->evaluate.schedule)
1900			compiled_data->evaluate.schedule = ccv_nnc_graph_static_schedule_new(compiled_data->graph, compiled_data->stream_type, model->max_stream_count, 0, 0, compiled_data->evaluate.to_ops, compiled_data->evaluate.to_op_size);
1901		ccv_nnc_graph_run_with_schedule(compiled_data->graph, 0, compiled_data->evaluate.schedule, tensor_tape, stream_context);
1902	}
1903}
1904 
1905// Compile the graph to run ccv_cnnp_model_backward after ccv_cnnp_model_evaluate with requires_grad = true (MULTISTAGE_MODE).
1906// Particularly, this method compiles the accumulator graph.
1907static void _ccv_cnnp_model_multistage_jit_1(ccv_cnnp_model_t* const model)
1908{
1909	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1910	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 1910, __extension__ __PRETTY_FUNCTION__); }));
1911	assert(compiled_data->graph_mode == CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE)((void) sizeof ((compiled_data->graph_mode == CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE
) ? 1 : 0), __extension__ ({ if (compiled_data->graph_mode
 == CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE) ; else __assert_fail
 ("compiled_data->graph_mode == CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE"
, "ccv_cnnp_model.c", 1911, __extension__ __PRETTY_FUNCTION__
); }));
1912	ccv_nnc_symbolic_graph_t* accum = ccv_nnc_symbolic_graph_new();
1913	const int parallel_count = ccv_max(model->parallel_count, 1)({ typeof (model->parallel_count) _a = (model->parallel_count
); typeof (1) _b = (1); (_a > _b) ? _a : _b; });
1914	const int parameter_size = compiled_data->parameters->rnum;
1915	int i, j;
1916	compiled_data->backward.gradients = (ccv_nnc_tensor_symbol_t*)ccmallocmalloc(sizeof(ccv_nnc_tensor_symbol_t) * parameter_size * parallel_count * 3);
1917	compiled_data->backward.accum_gradients = compiled_data->backward.gradients + parameter_size * parallel_count;
1918	compiled_data->backward.updated_accum_gradients = compiled_data->backward.accum_gradients + parameter_size * parallel_count;
1919	for (i = 0; i < parameter_size; i++)
1920		for (j = 0; j < parallel_count; j++)
1921			if (compiled_data->tensors.gradients[i + j * parameter_size])
1922			{
1923				const ccv_nnc_tensor_param_t info = compiled_data->tensors.gradients[i + j * parameter_size]->info;
1924				// Now, the old gradient is the accumulated gradient, getting new gradient tensor setup so we can collect them.
1925				compiled_data->tensors.accum_gradients[i + j * parameter_size] = compiled_data->tensors.gradients[i + j * parameter_size];
1926				compiled_data->tensors.gradients[i + j * parameter_size] = ccv_nnc_tensor_new(0, info, 0);
1927				ccv_nnc_tensor_symbol_t inputs[2];
1928				inputs[0] = compiled_data->backward.accum_gradients[i + j * parameter_size] = ccv_nnc_tensor_symbol_new(accum, info, 0);
1929				inputs[1] = compiled_data->backward.gradients[i + j * parameter_size] = ccv_nnc_tensor_symbol_new(accum, info, 0);
1930				ccv_nnc_tensor_symbol_t output = compiled_data->backward.updated_accum_gradients[i + j * parameter_size] = ccv_nnc_tensor_symbol_new(accum, info, 0);
1931				ccv_nnc_graph_exec_symbol_new(accum, CMD_EWSUM_FORWARD()ccv_nnc_cmd(CCV_NNC_EWSUM_FORWARD, 0, ccv_nnc_cmd_auto, 0), inputs, 2, &output, 1, 0);
1932			} else {
1933				compiled_data->backward.accum_gradients[i + j * parameter_size] = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
1934				compiled_data->backward.gradients[i + j * parameter_size] = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
1935				compiled_data->backward.updated_accum_gradients[i + j * parameter_size] = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
1936			}
1937	ccv_nnc_graph_exec_symbol_autogen(accum, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
1938	if (ccv_nnc_symbolic_graph_source_size(accum) == 0)
1939	{
1940		ccv_nnc_symbolic_graph_free(accum);
1941		// Create empty graph.
1942		compiled_data->backward.accum = ccv_nnc_graph_new();
1943		ccv_nnc_graph_topsort(compiled_data->backward.accum, 0, 0);
1944		return;
1945	}
1946	ccv_array_t* const tensor_binds = ccv_array_new(sizeof(ccv_nnc_tensor_bind_t), 0, 0);
1947	_ccv_cnnp_model_bind_tensors(accum, compiled_data->backward.accum_gradients, compiled_data->tensors.accum_gradients, parameter_size * parallel_count, 1, tensor_binds);
1948	_ccv_cnnp_model_bind_tensors(accum, compiled_data->backward.gradients, compiled_data->tensors.gradients, parameter_size * parallel_count, 1, tensor_binds);
1949	_ccv_cnnp_model_bind_tensors(accum, compiled_data->backward.updated_accum_gradients, compiled_data->tensors.accum_gradients, parameter_size * parallel_count, 1, tensor_binds);
1950	ccv_nnc_symbolic_graph_compile(accum, compiled_data->compile_params, (ccv_nnc_tensor_bind_t*)ccv_array_get(tensor_binds, 0)((void*)(((char*)((tensor_binds)->data)) + (size_t)(tensor_binds
)->rsize * (size_t)(0))), tensor_binds->rnum, 0, 0, SYMBOLIC_GRAPH_SOURCES(accum)ccv_nnc_symbolic_graph_sources(accum), ccv_nnc_symbolic_graph_source_size
(accum), SYMBOLIC_GRAPH_DESTINATIONS(accum)ccv_nnc_symbolic_graph_destinations(accum), ccv_nnc_symbolic_graph_destination_size
(accum), &compiled_data->backward.accum, &compiled_data->backward.tensor_arena, &compiled_data->backward.graph_exec_arena);
1951	ccv_nnc_symbolic_graph_free(accum);
1952	ccv_array_free(tensor_binds);
1953	ccv_nnc_graph_set_default_static_schedule(compiled_data->backward.accum, compiled_data->stream_type, model->max_stream_count);
1954}
1955 
1956void ccv_cnnp_model_backward(ccv_cnnp_model_t* const model, ccv_nnc_tensor_t* const* const ingrads, const int ingrad_size, ccv_nnc_tensor_t* const* const outgrads, const int outgrad_size, ccv_nnc_tensor_tape_t* const tensor_tape, ccv_nnc_stream_context_t* const stream_context)
1957{
1958	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1959	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 1959, __extension__ __PRETTY_FUNCTION__); }));
1960	assert(compiled_data->graph_mode == CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE)((void) sizeof ((compiled_data->graph_mode == CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE
) ? 1 : 0), __extension__ ({ if (compiled_data->graph_mode
 == CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE) ; else __assert_fail
 ("compiled_data->graph_mode == CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE"
, "ccv_cnnp_model.c", 1960, __extension__ __PRETTY_FUNCTION__
); }));
1961	const int parallel_count = ccv_max(model->parallel_count, 1)({ typeof (model->parallel_count) _a = (model->parallel_count
); typeof (1) _b = (1); (_a > _b) ? _a : _b; });
1962	assert(ingrad_size == 0 || ingrad_size == model->output_size * parallel_count)((void) sizeof ((ingrad_size == 0 || ingrad_size == model->
output_size * parallel_count) ? 1 : 0), __extension__ ({ if (
ingrad_size == 0 || ingrad_size == model->output_size * parallel_count
) ; else __assert_fail ("ingrad_size == 0 || ingrad_size == model->output_size * parallel_count"
, "ccv_cnnp_model.c", 1962, __extension__ __PRETTY_FUNCTION__
); }));
1963	if (outgrad_size > 0)
1964		{ assert(outgrad_size == compiled_data->outgrad_size * parallel_count)((void) sizeof ((outgrad_size == compiled_data->outgrad_size
 * parallel_count) ? 1 : 0), __extension__ ({ if (outgrad_size
 == compiled_data->outgrad_size * parallel_count) ; else __assert_fail
 ("outgrad_size == compiled_data->outgrad_size * parallel_count"
, "ccv_cnnp_model.c", 1964, __extension__ __PRETTY_FUNCTION__
); })); }
1965	assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if
 (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c"
, 1965, __extension__ __PRETTY_FUNCTION__); }));
1966	assert(compiled_data->graph)((void) sizeof ((compiled_data->graph) ? 1 : 0), __extension__
 ({ if (compiled_data->graph) ; else __assert_fail ("compiled_data->graph"
, "ccv_cnnp_model.c", 1966, __extension__ __PRETTY_FUNCTION__
); }));
1967	const int parameter_size = compiled_data->parameters->rnum;
1968	// If we need to accumulate the gradients now, do jit on accumulator.
1969	if (compiled_data->backward.count > 0)
1970	{
1971		if (!compiled_data->backward.accum)
1972			_ccv_cnnp_model_multistage_jit_1(model);
1973		else if (compiled_data->backward.count == 1) {
1974			//  On this round, we need to switch accumulated gradients with gradients (so we can do accumulation properly).
1975			int i;
1976			for (i = 0; i < parameter_size * parallel_count; i++)
1977			{
1978				ccv_nnc_tensor_t* tensor;
1979				CCV_SWAP(compiled_data->tensors.accum_gradients[i], compiled_data->tensors.gradients[i], tensor)((tensor) = (compiled_data->tensors.accum_gradients[i]), (
compiled_data->tensors.accum_gradients[i]) = (compiled_data
->tensors.gradients[i]), (compiled_data->tensors.gradients
[i]) = (tensor));
1980			}
1981			if (compiled_data->backward.tensor_arena)
1982			{
1983				ccv_nnc_tensor_arena_clear_bindings(compiled_data->backward.tensor_arena);
1984				// Do rebind in case we messed up the binding (we switch accum_gradients and gradients).
1985				_ccv_cnnp_bind_tensors_to_arena(compiled_data->backward.tensor_arena, 0, compiled_data->backward.gradients, compiled_data->tensors.gradients, parameter_size * parallel_count, 1);
1986				_ccv_cnnp_bind_tensors_to_arena(compiled_data->backward.tensor_arena, 0, compiled_data->backward.accum_gradients, compiled_data->tensors.accum_gradients, parameter_size * parallel_count, 1);
1987				_ccv_cnnp_bind_tensors_to_arena(compiled_data->backward.tensor_arena, 0, compiled_data->backward.updated_accum_gradients, compiled_data->tensors.accum_gradients, parameter_size * parallel_count, 1);
1988			}
1989		}
1990	}
1991	const int ingrad_size_per_p = model->output_size;
1992	const int outgrad_size_per_p = compiled_data->outgrad_size;
1993	int i, j;
1994	for (i = 0; i < ingrad_size_per_p; i++)
1995	{
1996		const ccv_nnc_tensor_symbol_t ingrad = ccv_nnc_tensor_symbol_for_backward(model->graph, compiled_data->f[i]);
1997		if (!ingrad_size || !ingrads || ingrads[i] == 0)
1998		{
1999			// Set it to 1 if it is not specified.
2000			ccv_nnc_tensor_t* const ingrad_tensor = ccv_nnc_tensor_from_symbol(compiled_data->tensor_arena, ingrad);
2001			if (ingrad_tensor)
2002				ccv_nnc_cmd_exec(CMD_SET_FORWARD(1)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={1,}}}, 0), ccv_nnc_no_hint, 0, 0, 0, TENSOR_LIST(ingrad_tensor)(ccv_nnc_tensor_t* []){ingrad_tensor}, (1 +1 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_context);
2003			for (j = 1; j < parallel_count; j++)
2004			{
2005				ccv_nnc_tensor_t* const ingrad_tensor = ccv_nnc_tensor_from_symbol(compiled_data->tensor_arena, ccv_nnc_tensor_symbol_copy(model->graph, ingrad, j));
2006				if (ingrad_tensor)
2007					ccv_nnc_cmd_exec(CMD_SET_FORWARD(1)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={1,}}}, 0), ccv_nnc_no_hint, 0, 0, 0, TENSOR_LIST(ingrad_tensor)(ccv_nnc_tensor_t* []){ingrad_tensor}, (1 +1 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_context);
2008			}
2009		} else {
2010			// Make sure the length matches, in case it is an alias.
2011			assert(ccv_nnc_tensor_count(ingrads[i]->info) == ccv_nnc_tensor_count(ccv_nnc_tensor_symbol_params(model->graph, ingrad)))((void) sizeof ((ccv_nnc_tensor_count(ingrads[i]->info) ==
 ccv_nnc_tensor_count(ccv_nnc_tensor_symbol_params(model->
graph, ingrad))) ? 1 : 0), __extension__ ({ if (ccv_nnc_tensor_count
(ingrads[i]->info) == ccv_nnc_tensor_count(ccv_nnc_tensor_symbol_params
(model->graph, ingrad))) ; else __assert_fail ("ccv_nnc_tensor_count(ingrads[i]->info) == ccv_nnc_tensor_count(ccv_nnc_tensor_symbol_params(model->graph, ingrad))"
, "ccv_cnnp_model.c", 2011, __extension__ __PRETTY_FUNCTION__
); }));
2012			ccv_nnc_tensor_bind_symbol(compiled_data->tensor_arena, ingrad, ingrads[i]);
2013			for (j = 1; j < parallel_count; j++)
2014				ccv_nnc_tensor_bind_symbol(compiled_data->tensor_arena, ccv_nnc_tensor_symbol_copy(model->graph, ingrad, j), ingrads[i + ingrad_size_per_p * j]);
2015		}
2016	}
2017	if (outgrad_size > 0)
2018	{
2019		assert(compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES_AND_INPUTS && "shouldn't pass disable_outgrad to ccv_cnnp_model_evaluate before if you plan to compute outgrad")((void) sizeof ((compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES_AND_INPUTS
 && "shouldn't pass disable_outgrad to ccv_cnnp_model_evaluate before if you plan to compute outgrad"
) ? 1 : 0), __extension__ ({ if (compiled_data->gradient_mode
 == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES_AND_INPUTS &&
 "shouldn't pass disable_outgrad to ccv_cnnp_model_evaluate before if you plan to compute outgrad"
) ; else __assert_fail ("compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES_AND_INPUTS && \"shouldn't pass disable_outgrad to ccv_cnnp_model_evaluate before if you plan to compute outgrad\""
, "ccv_cnnp_model.c", 2019, __extension__ __PRETTY_FUNCTION__
); }));
2020		for (i = 0; i < outgrad_size_per_p; i++)
2021			if (outgrads[i])
2022			{
2023				const ccv_nnc_tensor_symbol_t outgrad = compiled_data->outgrads[i];
2024				ccv_nnc_tensor_bind_symbol(compiled_data->tensor_arena, outgrad, outgrads[i]);
2025				for (j = 1; j < parallel_count; j++)
2026					ccv_nnc_tensor_bind_symbol(compiled_data->tensor_arena, ccv_nnc_tensor_symbol_copy(model->graph, outgrad, j), outgrads[i + outgrad_size_per_p * j]);
2027			}
2028	} else {
2029		assert(compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES ||((void) sizeof ((compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES
 || compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES_AND_INPUTS
) ? 1 : 0), __extension__ ({ if (compiled_data->gradient_mode
 == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES || compiled_data
->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES_AND_INPUTS
) ; else __assert_fail ("compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES || compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES_AND_INPUTS"
, "ccv_cnnp_model.c", 2030, __extension__ __PRETTY_FUNCTION__
); }))
2030			compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES_AND_INPUTS)((void) sizeof ((compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES
 || compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES_AND_INPUTS
) ? 1 : 0), __extension__ ({ if (compiled_data->gradient_mode
 == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES || compiled_data
->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES_AND_INPUTS
) ; else __assert_fail ("compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES || compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES_AND_INPUTS"
, "ccv_cnnp_model.c", 2030, __extension__ __PRETTY_FUNCTION__
); }));
2031	}
2032	// We need to rebind here because in ccv_cnnp_evaluate, we clear bindings, that will reset all bindings for the gradients.
2033	// For parameters and internals these are fine because when we clear bindings, it restores to original bindings, which are these
2034	// parameters and internals. The same cannot be said for gradients due to the accum_gradients switching.
2035	_ccv_cnnp_bind_tensors_to_arena(compiled_data->tensor_arena, model->graph, compiled_data->gradients, compiled_data->tensors.gradients, parameter_size, parallel_count);
2036	if (!compiled_data->backward.schedule)
2037		compiled_data->backward.schedule = ccv_nnc_graph_static_schedule_new(compiled_data->graph, compiled_data->stream_type, model->max_stream_count, compiled_data->backward.from_ops, compiled_data->backward.from_op_size, 0, 0);
2038	// Run the backward pass.
2039	ccv_nnc_graph_run_with_schedule(compiled_data->graph, 0, compiled_data->backward.schedule, tensor_tape, stream_context);
2040	// If we need to run accumulation round, do that now.
2041	if (compiled_data->backward.count > 0)
2042		ccv_nnc_graph_run_with_schedule(compiled_data->backward.accum, 0, 0, 0, stream_context);
2043	// Update the count, this determines whether we need to accumulate or not.
2044	++compiled_data->backward.count;
2045}
2046 
2047// Compile the graph to run ccv_cnnp_model_apply_gradients after ccv_cnnp_model_backward (MULTISTAGE_MODE).
2048// Particularly, this method compiles the parameter update graph.
2049static void _ccv_cnnp_model_multistage_jit_2(ccv_cnnp_model_t* const model)
2050{
2051	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2052	assert(compiled_data->graph_mode == CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE)((void) sizeof ((compiled_data->graph_mode == CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE
) ? 1 : 0), __extension__ ({ if (compiled_data->graph_mode
 == CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE) ; else __assert_fail
 ("compiled_data->graph_mode == CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE"
, "ccv_cnnp_model.c", 2052, __extension__ __PRETTY_FUNCTION__
); }));
2053	const int parallel_count = ccv_max(model->parallel_count, 1)({ typeof (model->parallel_count) _a = (model->parallel_count
); typeof (1) _b = (1); (_a > _b) ? _a : _b; });
2054	const int parameter_size = compiled_data->parameters->rnum;
2055	ccv_array_t* const tensor_binds = ccv_array_new(sizeof(ccv_nnc_tensor_bind_t), 0, 0);
2056	_ccv_cnnp_model_bind_tensors(model->graph, (ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->parameters, 0)((void*)(((char*)((compiled_data->parameters)->data)) +
 (size_t)(compiled_data->parameters)->rsize * (size_t)(
0))), compiled_data->tensors.parameters, parameter_size, parallel_count, tensor_binds);
2057	_ccv_cnnp_model_bind_tensors(model->graph, compiled_data->updated_parameters, compiled_data->tensors.parameters, parameter_size, parallel_count, tensor_binds);
2058	// Bind accumulated gradients.
2059	if (compiled_data->backward.count > 1)
2060		_ccv_cnnp_model_bind_tensors(model->graph, compiled_data->gradients, compiled_data->tensors.accum_gradients, parameter_size, parallel_count, tensor_binds);
2061	else
2062		_ccv_cnnp_model_bind_tensors(model->graph, compiled_data->gradients, compiled_data->tensors.gradients, parameter_size, parallel_count, tensor_binds);
2063	ccv_array_t* const apply_gradients_from = ccv_array_new(sizeof(int), 0, 0);
2064	int i, j;
2065	for (i = 0; i < compiled_data->backward.to_size; i++)
2066	{
2067		const int* tos;
2068		int to_size;
2069		ccv_nnc_graph_exec_symbol_to(model->graph, compiled_data->backward.tos[i], &tos, &to_size);
2070		for (j = 0; j < to_size; j++)
2071		{
2072			// Check if this is already show up in the backward graph, if that is the case, it won't be in the apply
2073			// gradients graph.
2074			const ccv_nnc_graph_exec_t exec = ccv_nnc_graph_exec_from_symbol(compiled_data->graph_exec_arena, (ccv_nnc_graph_exec_symbol_t){
2075				.d = tos[j],
2076				.graph = model->graph,
2077			});
2078			if (!exec.graph)
2079				ccv_array_add_unique_int(apply_gradients_from, tos[j]);
2080		}
2081	}
2082	const int from_size = apply_gradients_from->rnum;
2083	if (from_size == 0)
2084	{
2085		ccv_array_free(apply_gradients_from);
2086		ccv_array_free(tensor_binds);
2087		return;
2088	}
2089	ccv_nnc_graph_exec_symbol_t* const froms = (ccv_nnc_graph_exec_symbol_t*)ccmallocmalloc(sizeof(ccv_nnc_graph_exec_symbol_t) * from_size);
2090	for (i = 0; i < from_size; i++)
2091		froms[i] = (ccv_nnc_graph_exec_symbol_t){
2092			.d = *(int*)ccv_array_get(apply_gradients_from, i)((void*)(((char*)((apply_gradients_from)->data)) + (size_t
)(apply_gradients_from)->rsize * (size_t)(i))),
2093			.graph = model->graph
2094		};
2095	ccv_array_free(apply_gradients_from);
2096	// It can only ends with updates on the parameters.
2097	ccv_array_t* const tos = ccv_array_new(sizeof(ccv_nnc_graph_exec_symbol_t), parameter_size * parallel_count, 0);
2098	for (i = 0;  i < parameter_size; i++)
2099	{
2100		if (compiled_data->update_nodes[i].d == CCV_NNC_NO_TENSOR_SYMBOL)
2101			continue;
2102		ccv_array_push(tos, &compiled_data->update_nodes[i]);
2103		for (j = 1; j < parallel_count; j++)
2104		{
2105			const ccv_nnc_graph_exec_symbol_t copy = ccv_nnc_graph_exec_symbol_copy(model->graph, compiled_data->update_nodes[i], j);
2106			ccv_array_push(tos, &copy);
2107		}
2108	}
2109	ccv_nnc_symbolic_graph_compile(model->graph, compiled_data->compile_params, (ccv_nnc_tensor_bind_t*)ccv_array_get(tensor_binds, 0)((void*)(((char*)((tensor_binds)->data)) + (size_t)(tensor_binds
)->rsize * (size_t)(0))), tensor_binds->rnum, 0, 0, froms, from_size, (ccv_nnc_graph_exec_symbol_t*)ccv_array_get(tos, 0)((void*)(((char*)((tos)->data)) + (size_t)(tos)->rsize *
 (size_t)(0))), tos->rnum, &compiled_data->apply_gradients.graph, &compiled_data->apply_gradients.tensor_arena, &compiled_data->apply_gradients.graph_exec_arena);
2110	ccv_array_free(tos);
2111	ccv_array_free(tensor_binds);
2112	ccfreefree(froms);
2113	const int max_saved_aux_size = compiled_data->minimize.max_saved_aux_size;
2114	for (i = 0; i < max_saved_aux_size * parameter_size; i++)
2115	{
2116		// Skip on no tensor.
2117		if (compiled_data->saved_aux[i].source.d == CCV_NNC_NO_TENSOR_SYMBOL)
2118			continue;
2119		ccv_nnc_tensor_t* const tensor = ccv_nnc_tensor_from_symbol(compiled_data->apply_gradients.tensor_arena, compiled_data->saved_aux[i].source);
2120		ccv_nnc_cmd_exec(CMD_SET_FORWARD(0)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={0,}}}, 0), ccv_nnc_no_hint, 0, 0, 0, &tensor, 1, 0);
2121		for (j = 1; j < parallel_count; j++)
2122		{
2123			ccv_nnc_tensor_t* const copy = ccv_nnc_tensor_from_symbol(compiled_data->apply_gradients.tensor_arena, ccv_nnc_tensor_symbol_copy(model->graph, compiled_data->saved_aux[i].source, j));
2124			if (copy)
2125				ccv_nnc_cmd_exec(CMD_SET_FORWARD(0)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={0,}}}, 0), ccv_nnc_no_hint, 0, 0, 0, &copy, 1, 0);
2126		}
2127	}
2128	ccv_nnc_graph_set_default_static_schedule(compiled_data->apply_gradients.graph, compiled_data->stream_type, model->max_stream_count);
2129}
2130 
2131void ccv_cnnp_model_apply_gradients(ccv_cnnp_model_t* const model, ccv_nnc_stream_context_t* const stream_context)
2132{
2133	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2134	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 2134, __extension__ __PRETTY_FUNCTION__); }));
2135	assert(compiled_data->graph_mode == CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE)((void) sizeof ((compiled_data->graph_mode == CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE
) ? 1 : 0), __extension__ ({ if (compiled_data->graph_mode
 == CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE) ; else __assert_fail
 ("compiled_data->graph_mode == CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE"
, "ccv_cnnp_model.c", 2135, __extension__ __PRETTY_FUNCTION__
); }));
2136	const int parallel_count = ccv_max(model->parallel_count, 1)({ typeof (model->parallel_count) _a = (model->parallel_count
); typeof (1) _b = (1); (_a > _b) ? _a : _b; });
2137	assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if
 (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c"
, 2137, __extension__ __PRETTY_FUNCTION__); }));
2138	assert(compiled_data->graph)((void) sizeof ((compiled_data->graph) ? 1 : 0), __extension__
 ({ if (compiled_data->graph) ; else __assert_fail ("compiled_data->graph"
, "ccv_cnnp_model.c", 2138, __extension__ __PRETTY_FUNCTION__
); }));
2139	// Skip if there is no backward pass.
2140	if (compiled_data->backward.count <= 0)
2141		return;
2142	// Skip if there is no parameters.
2143	if (compiled_data->parameters->rnum == 0)
2144	{
2145		compiled_data->backward.count = 0;
2146		return;
2147	}
2148	if (!compiled_data->apply_gradients.graph)
2149		_ccv_cnnp_model_multistage_jit_2(model);
2150	else {
2151		const int parameter_size = compiled_data->parameters->rnum;
2152		ccv_nnc_tensor_arena_clear_bindings(compiled_data->apply_gradients.tensor_arena);
2153		// Change to bind accum_gradients if we do gradient accumulation (run backward more than once).
2154		if (compiled_data->backward.count > 1)
2155			_ccv_cnnp_bind_tensors_to_arena(compiled_data->apply_gradients.tensor_arena, model->graph, compiled_data->gradients, compiled_data->tensors.accum_gradients, parameter_size, parallel_count);
2156		else
2157			_ccv_cnnp_bind_tensors_to_arena(compiled_data->apply_gradients.tensor_arena, model->graph, compiled_data->gradients, compiled_data->tensors.gradients, parameter_size, parallel_count);
2158	}
2159	if (compiled_data->apply_gradients.graph)
2160		ccv_nnc_graph_run_with_schedule(compiled_data->apply_gradients.graph, 0, 0, 0, stream_context);
2161	// Reset backward count to 0.
2162	compiled_data->backward.count = 0;
2163}
2164 
2165void ccv_cnnp_model_set_parameter(ccv_cnnp_model_t* const model, const ccv_cnnp_model_io_t parameter, const ccv_nnc_tensor_t* const tensor)
2166{
2167	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2168	const int param_sel = parameter->param_sel > 0 ? parameter->param_sel - 1 : parameter->param_sel;
2169	assert(parameter->param_sel != 0)((void) sizeof ((parameter->param_sel != 0) ? 1 : 0), __extension__
 ({ if (parameter->param_sel != 0) ; else __assert_fail ("parameter->param_sel != 0"
, "ccv_cnnp_model.c", 2169, __extension__ __PRETTY_FUNCTION__
); }));
2170	const int tensors_init = !!compiled_data->tensors_init.v;
2171	if (!tensors_init)
2172		_ccv_cnnp_model_tensors_init(model, compiled_data);
2173	else if ((uintptr_t)compiled_data->tensors_init.v & (uintptr_t)1)
2174	// Check if it is not fully allocated, if it is not, init_1.
2175		ccv_cnnp_model_tensors_init_1(model, compiled_data);
2176	ccv_array_t* const parameter_indices = ccv_array_new(sizeof(int), 0, 0);
2177	ccv_cnnp_model_add_to_parameter_indices(parameter->model, param_sel, parameter_indices);
2178	const int param_ref = parameter->param_ref > 0 ? parameter->param_ref - 1 : parameter->param_ref;
2179	if (param_ref < 0)
2180		{ assert(parameter_indices->rnum == 1)((void) sizeof ((parameter_indices->rnum == 1) ? 1 : 0), __extension__
 ({ if (parameter_indices->rnum == 1) ; else __assert_fail
 ("parameter_indices->rnum == 1", "ccv_cnnp_model.c", 2180
, __extension__ __PRETTY_FUNCTION__); })); }
2181	else
2182		{ assert(param_ref < parameter_indices->rnum)((void) sizeof ((param_ref < parameter_indices->rnum) ?
 1 : 0), __extension__ ({ if (param_ref < parameter_indices
->rnum) ; else __assert_fail ("param_ref < parameter_indices->rnum"
, "ccv_cnnp_model.c", 2182, __extension__ __PRETTY_FUNCTION__
); })); }
2183	const int d = *(int*)ccv_array_get(parameter_indices, param_ref >= 0 ? param_ref : 0)((void*)(((char*)((parameter_indices)->data)) + (size_t)(parameter_indices
)->rsize * (size_t)(param_ref >= 0 ? param_ref : 0)));
2184	ccv_array_free(parameter_indices);
2185	const int parameter_size = compiled_data->parameters->rnum;
2186	assert(d >= 0)((void) sizeof ((d >= 0) ? 1 : 0), __extension__ ({ if (d >=
 0) ; else __assert_fail ("d >= 0", "ccv_cnnp_model.c", 2186
, __extension__ __PRETTY_FUNCTION__); }));
2187	assert(d < parameter_size)((void) sizeof ((d < parameter_size) ? 1 : 0), __extension__
 ({ if (d < parameter_size) ; else __assert_fail ("d < parameter_size"
, "ccv_cnnp_model.c", 2187, __extension__ __PRETTY_FUNCTION__
); }));
2188	const int parallel_count = ccv_max(model->parallel_count, 1)({ typeof (model->parallel_count) _a = (model->parallel_count
); typeof (1) _b = (1); (_a > _b) ? _a : _b; });
2189	ccv_nnc_tensor_t* const dest = CCV_NNC_TENSOR(compiled_data->tensors.parameters[d])((ccv_nnc_tensor_t*)((uintptr_t)(compiled_data->tensors.parameters
[d]) & ~(uintptr_t)1));
2190	assert(dest)((void) sizeof ((dest) ? 1 : 0), __extension__ ({ if (dest) ;
 else __assert_fail ("dest", "ccv_cnnp_model.c", 2190, __extension__
 __PRETTY_FUNCTION__); }));
2191	ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD()ccv_nnc_cmd(CCV_NNC_DATA_TRANSFER_FORWARD, 0, ccv_nnc_cmd_auto
, 0), ccv_nnc_no_hint, 0, TENSOR_LIST((ccv_nnc_tensor_t*)tensor)(ccv_nnc_tensor_t* []){(ccv_nnc_tensor_t*)tensor}, (1 +1 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1
), TENSOR_LIST(dest)(ccv_nnc_tensor_t* []){dest}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), 0);
2192	int i;
2193	for (i = 1; i < parallel_count; i++)
2194	{
2195		ccv_nnc_tensor_t* const copy_tensor = CCV_NNC_TENSOR(compiled_data->tensors.parameters[d + i * parameter_size])((ccv_nnc_tensor_t*)((uintptr_t)(compiled_data->tensors.parameters
[d + i * parameter_size]) & ~(uintptr_t)1));
2196		if (copy_tensor)
2197			ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD()ccv_nnc_cmd(CCV_NNC_DATA_TRANSFER_FORWARD, 0, ccv_nnc_cmd_auto
, 0), ccv_nnc_no_hint, 0, TENSOR_LIST(dest)(ccv_nnc_tensor_t* []){dest}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_LIST(copy_tensor)(ccv_nnc_tensor_t* []){copy_tensor}, (1 +1 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), 0);
2198	}
2199	// Mark this symbol as init'ed.
2200	const int s = ((ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->parameters, d)((void*)(((char*)((compiled_data->parameters)->data)) +
 (size_t)(compiled_data->parameters)->rsize * (size_t)(
d))))->d;
2201  uint32_t* const init_v = CCV_NNC_INIT_V(compiled_data->tensors_init.v)((uint32_t*)((uintptr_t)(compiled_data->tensors_init.v) &
 ~(uintptr_t)1));
2202	init_v[s >> 5] |= (1u << (s & 0x1f));
2203}
2204 
2205void ccv_cnnp_model_parameter_copy(ccv_cnnp_model_t* const model, const ccv_cnnp_model_io_t parameter, ccv_nnc_tensor_t* const tensor)
2206{
2207	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2208	const int param_sel = parameter->param_sel > 0 ? parameter->param_sel - 1 : parameter->param_sel;
2209	assert(parameter->param_sel != 0)((void) sizeof ((parameter->param_sel != 0) ? 1 : 0), __extension__
 ({ if (parameter->param_sel != 0) ; else __assert_fail ("parameter->param_sel != 0"
, "ccv_cnnp_model.c", 2209, __extension__ __PRETTY_FUNCTION__
); }));
2210	assert(compiled_data->tensors.parameters)((void) sizeof ((compiled_data->tensors.parameters) ? 1 : 0
), __extension__ ({ if (compiled_data->tensors.parameters)
 ; else __assert_fail ("compiled_data->tensors.parameters"
, "ccv_cnnp_model.c", 2210, __extension__ __PRETTY_FUNCTION__
); }));
2211	ccv_array_t* const parameter_indices = ccv_array_new(sizeof(int), 0, 0);
2212	ccv_cnnp_model_add_to_parameter_indices(parameter->model, param_sel, parameter_indices);
2213	const int param_ref = parameter->param_ref > 0 ? parameter->param_ref - 1 : parameter->param_ref;
2214	if (param_ref < 0)
2215		{ assert(parameter_indices->rnum == 1)((void) sizeof ((parameter_indices->rnum == 1) ? 1 : 0), __extension__
 ({ if (parameter_indices->rnum == 1) ; else __assert_fail
 ("parameter_indices->rnum == 1", "ccv_cnnp_model.c", 2215
, __extension__ __PRETTY_FUNCTION__); })); }
2216	else
2217		{ assert(param_ref < parameter_indices->rnum)((void) sizeof ((param_ref < parameter_indices->rnum) ?
 1 : 0), __extension__ ({ if (param_ref < parameter_indices
->rnum) ; else __assert_fail ("param_ref < parameter_indices->rnum"
, "ccv_cnnp_model.c", 2217, __extension__ __PRETTY_FUNCTION__
); })); }
2218	const int d = *(int*)ccv_array_get(parameter_indices, param_ref >= 0 ? param_ref : 0)((void*)(((char*)((parameter_indices)->data)) + (size_t)(parameter_indices
)->rsize * (size_t)(param_ref >= 0 ? param_ref : 0)));
2219	ccv_array_free(parameter_indices);
2220	const int parameter_size = compiled_data->parameters->rnum;
2221	assert(d >= 0)((void) sizeof ((d >= 0) ? 1 : 0), __extension__ ({ if (d >=
 0) ; else __assert_fail ("d >= 0", "ccv_cnnp_model.c", 2221
, __extension__ __PRETTY_FUNCTION__); }));
2222	assert(d < parameter_size)((void) sizeof ((d < parameter_size) ? 1 : 0), __extension__
 ({ if (d < parameter_size) ; else __assert_fail ("d < parameter_size"
, "ccv_cnnp_model.c", 2222, __extension__ __PRETTY_FUNCTION__
); }));
2223	// We don't need to consider parallel_count, every parameter on each device is identical.
2224	ccv_nnc_tensor_t* const src = CCV_NNC_TENSOR(compiled_data->tensors.parameters[d])((ccv_nnc_tensor_t*)((uintptr_t)(compiled_data->tensors.parameters
[d]) & ~(uintptr_t)1));
2225	assert(src)((void) sizeof ((src) ? 1 : 0), __extension__ ({ if (src) ; else
 __assert_fail ("src", "ccv_cnnp_model.c", 2225, __extension__
 __PRETTY_FUNCTION__); }));
2226	ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD()ccv_nnc_cmd(CCV_NNC_DATA_TRANSFER_FORWARD, 0, ccv_nnc_cmd_auto
, 0), ccv_nnc_no_hint, 0, TENSOR_LIST(src)(ccv_nnc_tensor_t* []){src}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0
 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_LIST(tensor)(ccv_nnc_tensor_t* []){tensor}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0
 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), 0);
2227}
2228 
2229ccv_nnc_tensor_param_t ccv_cnnp_model_parameter_tensor_params(ccv_cnnp_model_t* const model, const ccv_cnnp_model_io_t parameter)
2230{
2231	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2232	const int param_sel = parameter->param_sel > 0 ? parameter->param_sel - 1 : parameter->param_sel;
2233	assert(parameter->param_sel != 0)((void) sizeof ((parameter->param_sel != 0) ? 1 : 0), __extension__
 ({ if (parameter->param_sel != 0) ; else __assert_fail ("parameter->param_sel != 0"
, "ccv_cnnp_model.c", 2233, __extension__ __PRETTY_FUNCTION__
); }));
2234	assert(compiled_data->tensors.parameters)((void) sizeof ((compiled_data->tensors.parameters) ? 1 : 0
), __extension__ ({ if (compiled_data->tensors.parameters)
 ; else __assert_fail ("compiled_data->tensors.parameters"
, "ccv_cnnp_model.c", 2234, __extension__ __PRETTY_FUNCTION__
); }));
2235	ccv_array_t* const parameter_indices = ccv_array_new(sizeof(int), 0, 0);
2236	ccv_cnnp_model_add_to_parameter_indices(parameter->model, param_sel, parameter_indices);
2237	const int param_ref = parameter->param_ref > 0 ? parameter->param_ref - 1 : parameter->param_ref;
2238	if (param_ref < 0)
2239		{ assert(parameter_indices->rnum == 1)((void) sizeof ((parameter_indices->rnum == 1) ? 1 : 0), __extension__
 ({ if (parameter_indices->rnum == 1) ; else __assert_fail
 ("parameter_indices->rnum == 1", "ccv_cnnp_model.c", 2239
, __extension__ __PRETTY_FUNCTION__); })); }
2240	else
2241		{ assert(param_ref < parameter_indices->rnum)((void) sizeof ((param_ref < parameter_indices->rnum) ?
 1 : 0), __extension__ ({ if (param_ref < parameter_indices
->rnum) ; else __assert_fail ("param_ref < parameter_indices->rnum"
, "ccv_cnnp_model.c", 2241, __extension__ __PRETTY_FUNCTION__
); })); }
2242	const int d = *(int*)ccv_array_get(parameter_indices, param_ref >= 0 ? param_ref : 0)((void*)(((char*)((parameter_indices)->data)) + (size_t)(parameter_indices
)->rsize * (size_t)(param_ref >= 0 ? param_ref : 0)));
2243	ccv_array_free(parameter_indices);
2244	const int parameter_size = compiled_data->parameters->rnum;
2245	assert(d >= 0)((void) sizeof ((d >= 0) ? 1 : 0), __extension__ ({ if (d >=
 0) ; else __assert_fail ("d >= 0", "ccv_cnnp_model.c", 2245
, __extension__ __PRETTY_FUNCTION__); }));
2246	assert(d < parameter_size)((void) sizeof ((d < parameter_size) ? 1 : 0), __extension__
 ({ if (d < parameter_size) ; else __assert_fail ("d < parameter_size"
, "ccv_cnnp_model.c", 2246, __extension__ __PRETTY_FUNCTION__
); }));
2247	// We don't need to consider parallel_count, every parameter on each device is identical.
2248	ccv_nnc_tensor_t* const tensor = CCV_NNC_TENSOR(compiled_data->tensors.parameters[d])((ccv_nnc_tensor_t*)((uintptr_t)(compiled_data->tensors.parameters
[d]) & ~(uintptr_t)1));
2249	assert(tensor)((void) sizeof ((tensor) ? 1 : 0), __extension__ ({ if (tensor
) ; else __assert_fail ("tensor", "ccv_cnnp_model.c", 2249, __extension__
 __PRETTY_FUNCTION__); }));
2250	return tensor->info;
2251}
2252 
2253const char* ccv_cnnp_model_parameter_name(ccv_cnnp_model_t* const model, const ccv_cnnp_model_io_t parameter)
2254{
2255	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2256	const int param_sel = parameter->param_sel > 0 ? parameter->param_sel - 1 : parameter->param_sel;
2257	assert(parameter->param_sel != 0)((void) sizeof ((parameter->param_sel != 0) ? 1 : 0), __extension__
 ({ if (parameter->param_sel != 0) ; else __assert_fail ("parameter->param_sel != 0"
, "ccv_cnnp_model.c", 2257, __extension__ __PRETTY_FUNCTION__
); }));
2258	ccv_array_t* const parameter_indices = ccv_array_new(sizeof(int), 0, 0);
2259	ccv_cnnp_model_add_to_parameter_indices(parameter->model, param_sel, parameter_indices);
2260	const int param_ref = parameter->param_ref > 0 ? parameter->param_ref - 1 : parameter->param_ref;
2261	if (param_ref < 0)
2262		{ assert(parameter_indices->rnum == 1)((void) sizeof ((parameter_indices->rnum == 1) ? 1 : 0), __extension__
 ({ if (parameter_indices->rnum == 1) ; else __assert_fail
 ("parameter_indices->rnum == 1", "ccv_cnnp_model.c", 2262
, __extension__ __PRETTY_FUNCTION__); })); }
2263	else
2264		{ assert(param_ref < parameter_indices->rnum)((void) sizeof ((param_ref < parameter_indices->rnum) ?
 1 : 0), __extension__ ({ if (param_ref < parameter_indices
->rnum) ; else __assert_fail ("param_ref < parameter_indices->rnum"
, "ccv_cnnp_model.c", 2264, __extension__ __PRETTY_FUNCTION__
); })); }
2265	const int d = *(int*)ccv_array_get(parameter_indices, param_ref >= 0 ? param_ref : 0)((void*)(((char*)((parameter_indices)->data)) + (size_t)(parameter_indices
)->rsize * (size_t)(param_ref >= 0 ? param_ref : 0)));
2266	ccv_array_free(parameter_indices);
2267	const int parameter_size = compiled_data->parameters->rnum;
2268	assert(d >= 0)((void) sizeof ((d >= 0) ? 1 : 0), __extension__ ({ if (d >=
 0) ; else __assert_fail ("d >= 0", "ccv_cnnp_model.c", 2268
, __extension__ __PRETTY_FUNCTION__); }));
2269	assert(d < parameter_size)((void) sizeof ((d < parameter_size) ? 1 : 0), __extension__
 ({ if (d < parameter_size) ; else __assert_fail ("d < parameter_size"
, "ccv_cnnp_model.c", 2269, __extension__ __PRETTY_FUNCTION__
); }));
2270	return *(char**)ccv_array_get(compiled_data->ids.parameters, d)((void*)(((char*)((compiled_data->ids.parameters)->data
)) + (size_t)(compiled_data->ids.parameters)->rsize * (
size_t)(d)));
2271}
2272 
2273int ccv_cnnp_model_parameter_count(ccv_cnnp_model_t* const model)
2274{
2275	assert(model->compiled_data)((void) sizeof ((model->compiled_data) ? 1 : 0), __extension__
 ({ if (model->compiled_data) ; else __assert_fail ("model->compiled_data"
, "ccv_cnnp_model.c", 2275, __extension__ __PRETTY_FUNCTION__
); }));
2276	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2277	return compiled_data->parameters->rnum;
2278}
2279 
2280ccv_cnnp_model_io_t ccv_cnnp_model_parameter_first(ccv_cnnp_model_t* const model, ccv_cnnp_model_parameters_filter_f first, void* const context)
2281{
2282	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2283	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 2283, __extension__ __PRETTY_FUNCTION__); }));
2284	const int parameter_size = compiled_data->parameters->rnum;
2285	int i;
2286	for (i = 0; i < parameter_size; i++)
2287	{
2288		const char* const name = *(char**)ccv_array_get(compiled_data->ids.parameters, i)((void*)(((char*)((compiled_data->ids.parameters)->data
)) + (size_t)(compiled_data->ids.parameters)->rsize * (
size_t)(i)));
2289		if (first(model, name, context))
2290			return ccv_cnnp_model_parameters(model, -1, i);
2291	}
2292	return 0;
2293}
2294 
2295ccv_array_t* ccv_cnnp_model_parameters_filter(ccv_cnnp_model_t* const model, ccv_cnnp_model_parameters_filter_f filter, void* const context)
2296{
2297	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2298	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 2298, __extension__ __PRETTY_FUNCTION__); }));
2299	ccv_array_t* const parameters = ccv_array_new(sizeof(ccv_cnnp_model_io_t), 0, 0);
2300	const int parameter_size = compiled_data->parameters->rnum;
2301	int i;
2302	for (i = 0; i < parameter_size; i++)
2303	{
2304		const char* const name = *(char**)ccv_array_get(compiled_data->ids.parameters, i)((void*)(((char*)((compiled_data->ids.parameters)->data
)) + (size_t)(compiled_data->ids.parameters)->rsize * (
size_t)(i)));
2305		if (filter(model, name, context))
2306		{
2307			ccv_cnnp_model_io_t parameter = ccv_cnnp_model_parameters(model, -1, i);
2308			ccv_array_push(parameters, &parameter);
2309		}
2310	}
2311	return parameters;
2312 
2313}
2314 
2315CCV_WARN_UNUSED(ccv_cnnp_model_io_t)ccv_cnnp_model_io_t __attribute__((warn_unused_result)) ccv_cnnp_model_parameter_first_uninit(ccv_cnnp_model_t* const model)
2316{
2317	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2318	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 2318, __extension__ __PRETTY_FUNCTION__); }));
2319	const int tensors_init = !!compiled_data->tensors_init.v;
2320	if (!tensors_init) // If nothing initialized, we return parameter 0.
2321		return ccv_cnnp_model_parameters(model, -1, 0);
2322	const int parameter_size = compiled_data->parameters->rnum;
2323	int i;
2324	const uint32_t* const init_v = CCV_NNC_INIT_V(compiled_data->tensors_init.v)((uint32_t*)((uintptr_t)(compiled_data->tensors_init.v) &
 ~(uintptr_t)1));
2325	for (i = 0; i < parameter_size; i++)
2326	{
2327		const int d = ((ccv_nnc_tensor_symbol_t*)ccv_array_get(compiled_data->parameters, i)((void*)(((char*)((compiled_data->parameters)->data)) +
 (size_t)(compiled_data->parameters)->rsize * (size_t)(
i))))->d;
2328		if (!(init_v[d >> 5] & (1u << (d & 0x1f))))
2329			return ccv_cnnp_model_parameters(model, -1, i);
2330	}
2331	return 0;
2332}
2333 
2334static ccv_array_t* _ccv_cnnp_model_parameter_indices(const ccv_cnnp_model_t* const model, const ccv_cnnp_model_io_t parameters, int* const param_ref)
2335{
2336	const int to_param_sel = parameters->param_sel > 0 ? parameters->param_sel - 1 : parameters->param_sel;
2337	assert(parameters->param_sel != 0)((void) sizeof ((parameters->param_sel != 0) ? 1 : 0), __extension__
 ({ if (parameters->param_sel != 0) ; else __assert_fail (
"parameters->param_sel != 0", "ccv_cnnp_model.c", 2337, __extension__
 __PRETTY_FUNCTION__); }));
2338	ccv_array_t* const to_parameter_indices = ccv_array_new(sizeof(int), 0, 0);
2339	ccv_cnnp_model_add_to_parameter_indices(parameters->model, to_param_sel, to_parameter_indices);
2340	*param_ref = parameters->param_ref > 0 ? parameters->param_ref - 1 : parameters->param_ref;
2341	return to_parameter_indices;
2342}
2343 
2344static void _ccv_cnnp_model_to_parameter_indices_and_from_parameter_indices(ccv_cnnp_model_t* const model, const ccv_cnnp_model_io_t parameters, const ccv_cnnp_model_t* const from_model, const ccv_cnnp_model_io_t from_parameters, ccv_array_t** const parameter_indices, int* const param_ref, ccv_array_t** const from_parameter_indices, int* const from_param_ref, const int only_init_0)
2345{
2346	// If the model is not compiled yet. Compile them now.
2347	if (!model->graph)
2348	{
2349		model->graph = ccv_nnc_symbolic_graph_new();
2350		assert(from_model->compiled_data)((void) sizeof ((from_model->compiled_data) ? 1 : 0), __extension__
 ({ if (from_model->compiled_data) ; else __assert_fail ("from_model->compiled_data"
, "ccv_cnnp_model.c", 2350, __extension__ __PRETTY_FUNCTION__
); }));
2351		const int input_size = from_model->input_size;
2352		ccv_nnc_tensor_param_t input_params[input_size];
2353		int i;
2354		for (i = 0; i < input_size; i++)
2355			input_params[i] = ccv_nnc_tensor_symbol_params(from_model->graph, from_model->inputs[i]);
2356		_ccv_cnnp_model_compile(model, input_params, input_size, from_model->compiled_data->loss);
2357		model->parallel_count = from_model->parallel_count;
2358		model->memory_compression = from_model->memory_compression;
2359		model->memory_reduction = from_model->memory_reduction;
2360		model->gradient_checkpointing = from_model->gradient_checkpointing;
2361		model->compiled_data->stream_type = from_model->compiled_data->stream_type;
2362		model->compiled_data->minimize.minimizer = from_model->compiled_data->minimize.minimizer;
2363		model->compiled_data->minimize.max_saved_aux_size = from_model->compiled_data->minimize.max_saved_aux_size;
2364	}
2365	ccv_cnnp_compiled_data_t* const to_compiled_data = model->compiled_data;
2366	assert(to_compiled_data)((void) sizeof ((to_compiled_data) ? 1 : 0), __extension__ ({
 if (to_compiled_data) ; else __assert_fail ("to_compiled_data"
, "ccv_cnnp_model.c", 2366, __extension__ __PRETTY_FUNCTION__
); }));
2367	const int to_tensors_init = !!to_compiled_data->tensors_init.v;
2368	if (!to_tensors_init)
2369	{
2370		if (only_init_0)
2371			ccv_cnnp_model_tensors_init_0(model, to_compiled_data);
2372		else
2373			_ccv_cnnp_model_tensors_init(model, to_compiled_data);
2374	} else if (!only_init_0 && (uintptr_t)to_compiled_data->tensors_init.v & (uintptr_t)1)
2375		// Check if it is not fully allocated, if it is not, init_1.
2376			ccv_cnnp_model_tensors_init_1(model, to_compiled_data);
2377	assert(to_compiled_data->tensors.parameters)((void) sizeof ((to_compiled_data->tensors.parameters) ? 1
 : 0), __extension__ ({ if (to_compiled_data->tensors.parameters
) ; else __assert_fail ("to_compiled_data->tensors.parameters"
, "ccv_cnnp_model.c", 2377, __extension__ __PRETTY_FUNCTION__
); }));
2378	*parameter_indices = _ccv_cnnp_model_parameter_indices(model, parameters, param_ref);
2379	*from_parameter_indices = _ccv_cnnp_model_parameter_indices(from_model, from_parameters, from_param_ref);
2380	if (*from_param_ref < 0 && *param_ref >= 0)
2381		{ assert((*from_parameter_indices)->rnum == 1)((void) sizeof (((*from_parameter_indices)->rnum == 1) ? 1
 : 0), __extension__ ({ if ((*from_parameter_indices)->rnum
 == 1) ; else __assert_fail ("(*from_parameter_indices)->rnum == 1"
, "ccv_cnnp_model.c", 2381, __extension__ __PRETTY_FUNCTION__
); })); }
2382	else if (*from_param_ref >= 0)
2383		{ assert(*from_param_ref < (*from_parameter_indices)->rnum)((void) sizeof ((*from_param_ref < (*from_parameter_indices
)->rnum) ? 1 : 0), __extension__ ({ if (*from_param_ref <
 (*from_parameter_indices)->rnum) ; else __assert_fail ("*from_param_ref < (*from_parameter_indices)->rnum"
, "ccv_cnnp_model.c", 2383, __extension__ __PRETTY_FUNCTION__
); })); }
2384	if (*param_ref < 0 && *from_param_ref >= 0)
2385		{ assert((*parameter_indices)->rnum == 1)((void) sizeof (((*parameter_indices)->rnum == 1) ? 1 : 0)
, __extension__ ({ if ((*parameter_indices)->rnum == 1) ; else
 __assert_fail ("(*parameter_indices)->rnum == 1", "ccv_cnnp_model.c"
, 2385, __extension__ __PRETTY_FUNCTION__); })); }
2386	else if (*param_ref >= 0)
2387		{ assert(*param_ref < (*parameter_indices)->rnum)((void) sizeof ((*param_ref < (*parameter_indices)->rnum
) ? 1 : 0), __extension__ ({ if (*param_ref < (*parameter_indices
)->rnum) ; else __assert_fail ("*param_ref < (*parameter_indices)->rnum"
, "ccv_cnnp_model.c", 2387, __extension__ __PRETTY_FUNCTION__
); })); }
2388}
2389 
2390void ccv_cnnp_model_set_parameters(ccv_cnnp_model_t* const model, const ccv_cnnp_model_io_t parameters, const ccv_cnnp_model_t* const from_model, const ccv_cnnp_model_io_t from_parameters)
2391{
2392	ccv_array_t* to_parameter_indices;
2393	int to_param_ref;
2394	ccv_array_t* from_parameter_indices;
2395	int from_param_ref;
2396	_ccv_cnnp_model_to_parameter_indices_and_from_parameter_indices(model, parameters, from_model, from_parameters, &to_parameter_indices, &to_param_ref, &from_parameter_indices, &from_param_ref, 0);
2397	// Should be exactly the same tensor.
2398	if (to_param_ref < 0 && from_param_ref < 0)
2399		{ assert(from_parameter_indices->rnum == to_parameter_indices->rnum)((void) sizeof ((from_parameter_indices->rnum == to_parameter_indices
->rnum) ? 1 : 0), __extension__ ({ if (from_parameter_indices
->rnum == to_parameter_indices->rnum) ; else __assert_fail
 ("from_parameter_indices->rnum == to_parameter_indices->rnum"
, "ccv_cnnp_model.c", 2399, __extension__ __PRETTY_FUNCTION__
); })); }
2400	// To models.
2401	ccv_cnnp_compiled_data_t* const to_compiled_data = model->compiled_data;
2402	assert(to_compiled_data)((void) sizeof ((to_compiled_data) ? 1 : 0), __extension__ ({
 if (to_compiled_data) ; else __assert_fail ("to_compiled_data"
, "ccv_cnnp_model.c", 2402, __extension__ __PRETTY_FUNCTION__
); }));
2403	// From models.
2404	const ccv_cnnp_compiled_data_t* const from_compiled_data = from_model->compiled_data;
2405	const int parallel_count = ccv_max(model->parallel_count, 1)({ typeof (model->parallel_count) _a = (model->parallel_count
); typeof (1) _b = (1); (_a > _b) ? _a : _b; });
2406	const int to_parameter_size = to_compiled_data->parameters->rnum;
2407	const int rnum = (to_param_ref < 0 && from_param_ref < 0) ? from_parameter_indices->rnum : 1;
2408	int i, j;
2409	const uint32_t* const from_init_v = CCV_NNC_INIT_V(from_compiled_data->tensors_init.v)((uint32_t*)((uintptr_t)(from_compiled_data->tensors_init.
v) & ~(uintptr_t)1));
2410  uint32_t* const to_init_v = CCV_NNC_INIT_V(to_compiled_data->tensors_init.v)((uint32_t*)((uintptr_t)(to_compiled_data->tensors_init.v)
 & ~(uintptr_t)1));
2411	for (i = 0; i < rnum; i++)
2412	{
2413		const int src_d = *(int*)ccv_array_get(from_parameter_indices,from_param_ref >= 0 ? from_param_ref : i)((void*)(((char*)((from_parameter_indices)->data)) + (size_t
)(from_parameter_indices)->rsize * (size_t)(from_param_ref
 >= 0 ? from_param_ref : i)));
2414		assert(src_d >= 0)((void) sizeof ((src_d >= 0) ? 1 : 0), __extension__ ({ if
 (src_d >= 0) ; else __assert_fail ("src_d >= 0", "ccv_cnnp_model.c"
, 2414, __extension__ __PRETTY_FUNCTION__); }));
2415		assert(src_d < from_compiled_data->parameters->rnum)((void) sizeof ((src_d < from_compiled_data->parameters
->rnum) ? 1 : 0), __extension__ ({ if (src_d < from_compiled_data
->parameters->rnum) ; else __assert_fail ("src_d < from_compiled_data->parameters->rnum"
, "ccv_cnnp_model.c", 2415, __extension__ __PRETTY_FUNCTION__
); }));
2416		const int s = ((ccv_nnc_tensor_symbol_t*)ccv_array_get(from_compiled_data->parameters, src_d)((void*)(((char*)((from_compiled_data->parameters)->data
)) + (size_t)(from_compiled_data->parameters)->rsize * (
size_t)(src_d))))->d;
2417		// If the original is not init'ed. We cannot copy from.
2418		if (!(from_init_v[s >> 5] & (1u << (s & 0x1f))))
2419			continue;
2420		const int dest_d = *(int*)ccv_array_get(to_parameter_indices, to_param_ref >= 0 ? to_param_ref : i)((void*)(((char*)((to_parameter_indices)->data)) + (size_t
)(to_parameter_indices)->rsize * (size_t)(to_param_ref >=
 0 ? to_param_ref : i)));
2421		assert(dest_d >= 0)((void) sizeof ((dest_d >= 0) ? 1 : 0), __extension__ ({ if
 (dest_d >= 0) ; else __assert_fail ("dest_d >= 0", "ccv_cnnp_model.c"
, 2421, __extension__ __PRETTY_FUNCTION__); }));
2422		assert(dest_d < to_compiled_data->parameters->rnum)((void) sizeof ((dest_d < to_compiled_data->parameters->
rnum) ? 1 : 0), __extension__ ({ if (dest_d < to_compiled_data
->parameters->rnum) ; else __assert_fail ("dest_d < to_compiled_data->parameters->rnum"
, "ccv_cnnp_model.c", 2422, __extension__ __PRETTY_FUNCTION__
); }));
2423		ccv_nnc_tensor_t* const src = CCV_NNC_TENSOR(from_compiled_data->tensors.parameters[src_d])((ccv_nnc_tensor_t*)((uintptr_t)(from_compiled_data->tensors
.parameters[src_d]) & ~(uintptr_t)1));
2424		assert(src)((void) sizeof ((src) ? 1 : 0), __extension__ ({ if (src) ; else
 __assert_fail ("src", "ccv_cnnp_model.c", 2424, __extension__
 __PRETTY_FUNCTION__); }));
2425		ccv_nnc_tensor_t* const dest = CCV_NNC_TENSOR(to_compiled_data->tensors.parameters[dest_d])((ccv_nnc_tensor_t*)((uintptr_t)(to_compiled_data->tensors
.parameters[dest_d]) & ~(uintptr_t)1));
2426		assert(dest)((void) sizeof ((dest) ? 1 : 0), __extension__ ({ if (dest) ;
 else __assert_fail ("dest", "ccv_cnnp_model.c", 2426, __extension__
 __PRETTY_FUNCTION__); }));
2427		ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD()ccv_nnc_cmd(CCV_NNC_DATA_TRANSFER_FORWARD, 0, ccv_nnc_cmd_auto
, 0), ccv_nnc_no_hint, 0, TENSOR_LIST(src)(ccv_nnc_tensor_t* []){src}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0
 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_LIST(dest)(ccv_nnc_tensor_t* []){dest}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), 0);
2428		for (j = 1; j < parallel_count; j++)
2429		{
2430			ccv_nnc_tensor_t* const copy_tensor = CCV_NNC_TENSOR(to_compiled_data->tensors.parameters[dest_d + j * to_parameter_size])((ccv_nnc_tensor_t*)((uintptr_t)(to_compiled_data->tensors
.parameters[dest_d + j * to_parameter_size]) & ~(uintptr_t
)1));
2431			if (copy_tensor)
2432				ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD()ccv_nnc_cmd(CCV_NNC_DATA_TRANSFER_FORWARD, 0, ccv_nnc_cmd_auto
, 0), ccv_nnc_no_hint, 0, TENSOR_LIST(dest)(ccv_nnc_tensor_t* []){dest}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_LIST(copy_tensor)(ccv_nnc_tensor_t* []){copy_tensor}, (1 +1 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), 0);
2433		}
2434		// Mark this symbol as init'ed.
2435		const int d = ((ccv_nnc_tensor_symbol_t*)ccv_array_get(to_compiled_data->parameters, dest_d)((void*)(((char*)((to_compiled_data->parameters)->data)
) + (size_t)(to_compiled_data->parameters)->rsize * (size_t
)(dest_d))))->d;
2436		to_init_v[d >> 5] |= (1u << (d & 0x1f));
2437	}
2438	ccv_array_free(to_parameter_indices);
2439	ccv_array_free(from_parameter_indices);
2440}
2441 
2442KHASH_MAP_INIT_STR(ccv_cnnp_parameter_id, int)typedef struct kh_ccv_cnnp_parameter_id_s { khint_t n_buckets
, size, n_occupied, upper_bound; khint32_t *flags; kh_cstr_t *
keys; int *vals; } kh_ccv_cnnp_parameter_id_t; static inline __attribute__
 ((__unused__)) kh_ccv_cnnp_parameter_id_t *kh_init_ccv_cnnp_parameter_id
(void) { return (kh_ccv_cnnp_parameter_id_t*)calloc(1,sizeof(
kh_ccv_cnnp_parameter_id_t)); } static inline __attribute__ (
(__unused__)) void kh_destroy_ccv_cnnp_parameter_id(kh_ccv_cnnp_parameter_id_t
 *h) { if (h) { free((void *)h->keys); free(h->flags); free
((void *)h->vals); free(h); } } static inline __attribute__
 ((__unused__)) void kh_clear_ccv_cnnp_parameter_id(kh_ccv_cnnp_parameter_id_t
 *h) { if (h && h->flags) { memset(h->flags, 0xaa
, ((h->n_buckets) < 16? 1 : (h->n_buckets)>>4)
 * sizeof(khint32_t)); h->size = h->n_occupied = 0; } }
 static inline __attribute__ ((__unused__)) khint_t kh_get_ccv_cnnp_parameter_id
(const kh_ccv_cnnp_parameter_id_t *h, kh_cstr_t key) { if (h->
n_buckets) { khint_t k, i, last, mask, step = 0; mask = h->
n_buckets - 1; k = __ac_X31_hash_string(key); i = k & mask
; last = i; while (!((h->flags[i>>4]>>((i&
0xfU)<<1))&2) && (((h->flags[i>>4]
>>((i&0xfU)<<1))&1) || !(strcmp(h->keys
[i], key) == 0))) { i = (i + (++step)) & mask; if (i == last
) return h->n_buckets; } return ((h->flags[i>>4]>>
((i&0xfU)<<1))&3)? h->n_buckets : i; } else return
 0; } static inline __attribute__ ((__unused__)) int kh_resize_ccv_cnnp_parameter_id
(kh_ccv_cnnp_parameter_id_t *h, khint_t new_n_buckets) { khint32_t
 *new_flags = 0; khint_t j = 1; { (--(new_n_buckets), (new_n_buckets
)|=(new_n_buckets)>>1, (new_n_buckets)|=(new_n_buckets)
>>2, (new_n_buckets)|=(new_n_buckets)>>4, (new_n_buckets
)|=(new_n_buckets)>>8, (new_n_buckets)|=(new_n_buckets)
>>16, ++(new_n_buckets)); if (new_n_buckets < 4) new_n_buckets
 = 4; if (h->size >= (khint_t)(new_n_buckets * __ac_HASH_UPPER
 + 0.5)) j = 0; else { new_flags = (khint32_t*)malloc(((new_n_buckets
) < 16? 1 : (new_n_buckets)>>4) * sizeof(khint32_t))
; if (!new_flags) return -1; memset(new_flags, 0xaa, ((new_n_buckets
) < 16? 1 : (new_n_buckets)>>4) * sizeof(khint32_t))
; if (h->n_buckets < new_n_buckets) { kh_cstr_t *new_keys
 = (kh_cstr_t*)realloc((void *)h->keys,new_n_buckets * sizeof
(kh_cstr_t)); if (!new_keys) { free(new_flags); return -1; } h
->keys = new_keys; if (1) { int *new_vals = (int*)realloc(
(void *)h->vals,new_n_buckets * sizeof(int)); if (!new_vals
) { free(new_flags); return -1; } h->vals = new_vals; } } }
 } if (j) { for (j = 0; j != h->n_buckets; ++j) { if (((h->
flags[j>>4]>>((j&0xfU)<<1))&3) == 0
) { kh_cstr_t key = h->keys[j]; int val; khint_t new_mask;
 new_mask = new_n_buckets - 1; if (1) val = h->vals[j]; (h
->flags[j>>4]|=1ul<<((j&0xfU)<<1)); while
 (1) { khint_t k, i, step = 0; k = __ac_X31_hash_string(key);
 i = k & new_mask; while (!((new_flags[i>>4]>>
((i&0xfU)<<1))&2)) i = (i + (++step)) & new_mask
; (new_flags[i>>4]&=~(2ul<<((i&0xfU)<<
1))); if (i < h->n_buckets && ((h->flags[i>>
4]>>((i&0xfU)<<1))&3) == 0) { { kh_cstr_t
 tmp = h->keys[i]; h->keys[i] = key; key = tmp; } if (1
) { int tmp = h->vals[i]; h->vals[i] = val; val = tmp; }
 (h->flags[i>>4]|=1ul<<((i&0xfU)<<1)
); } else { h->keys[i] = key; if (1) h->vals[i] = val; break
; } } } } if (h->n_buckets > new_n_buckets) { h->keys
 = (kh_cstr_t*)realloc((void *)h->keys,new_n_buckets * sizeof
(kh_cstr_t)); if (1) h->vals = (int*)realloc((void *)h->
vals,new_n_buckets * sizeof(int)); } free(h->flags); h->
flags = new_flags; h->n_buckets = new_n_buckets; h->n_occupied
 = h->size; h->upper_bound = (khint_t)(h->n_buckets *
 __ac_HASH_UPPER + 0.5); } return 0; } static inline __attribute__
 ((__unused__)) khint_t kh_put_ccv_cnnp_parameter_id(kh_ccv_cnnp_parameter_id_t
 *h, kh_cstr_t key, int *ret) { khint_t x; if (h->n_occupied
 >= h->upper_bound) { if (h->n_buckets > (h->size
<<1)) { if (kh_resize_ccv_cnnp_parameter_id(h, h->n_buckets
 - 1) < 0) { *ret = -1; return h->n_buckets; } } else if
 (kh_resize_ccv_cnnp_parameter_id(h, h->n_buckets + 1) <
 0) { *ret = -1; return h->n_buckets; } } { khint_t k, i, site
, last, mask = h->n_buckets - 1, step = 0; x = site = h->
n_buckets; k = __ac_X31_hash_string(key); i = k & mask; if
 (((h->flags[i>>4]>>((i&0xfU)<<1))&
2)) x = i; else { last = i; while (!((h->flags[i>>4]
>>((i&0xfU)<<1))&2) && (((h->flags
[i>>4]>>((i&0xfU)<<1))&1) || !(strcmp
(h->keys[i], key) == 0))) { if (((h->flags[i>>4]>>
((i&0xfU)<<1))&1)) site = i; i = (i + (++step))
 & mask; if (i == last) { x = site; break; } } if (x == h
->n_buckets) { if (((h->flags[i>>4]>>((i&
0xfU)<<1))&2) && site != h->n_buckets) x
 = site; else x = i; } } } if (((h->flags[x>>4]>>
((x&0xfU)<<1))&2)) { h->keys[x] = key; (h->
flags[x>>4]&=~(3ul<<((x&0xfU)<<1)))
; ++h->size; ++h->n_occupied; *ret = 1; } else if (((h->
flags[x>>4]>>((x&0xfU)<<1))&1)) { h
->keys[x] = key; (h->flags[x>>4]&=~(3ul<<
((x&0xfU)<<1))); ++h->size; *ret = 2; } else *ret
 = 0; return x; } static inline __attribute__ ((__unused__)) void
 kh_del_ccv_cnnp_parameter_id(kh_ccv_cnnp_parameter_id_t *h, khint_t
 x) { if (x != h->n_buckets && !((h->flags[x>>
4]>>((x&0xfU)<<1))&3)) { (h->flags[x>>
4]|=1ul<<((x&0xfU)<<1)); --h->size; } }
27
←
Taking true branch→
28
←
Taking false branch→
29
←
Calling 'kh_resize_ccv_cnnp_parameter_id'→
30
←
Taking true branch→
31
←
Assuming the condition is false→
32
←
Taking false branch→
33
←
'?' condition is true→
34
←
Assuming 'new_flags' is non-null→
35
←
Taking false branch→
36
←
'?' condition is true→
37
←
Taking true branch→
38
←
Storing uninitialized value→
39
←
Assuming 'new_keys' is non-null→
40
←
Taking false branch→
41
←
Taking true branch→
42
←
Assuming 'new_vals' is non-null→
43
←
Taking false branch→
44
←
Taking true branch→
45
←
Loop condition is false. Execution continues on line 2442→
46
←
Taking false branch→
47
←
Returning from 'kh_resize_ccv_cnnp_parameter_id'→
48
←
Taking false branch→
49
←
Assuming the condition is true→
50
←
Taking true branch→
51
←
Taking true branch→
57
←
Assuming field 'n_occupied' is >= field 'upper_bound'→
58
←
Taking true branch→
59
←
Taking true branch→
60
←
Calling 'kh_resize_ccv_cnnp_parameter_id'→
61
←
Taking false branch→
62
←
Assuming the condition is false→
63
←
Taking false branch→
64
←
'?' condition is true→
65
←
Assuming 'new_flags' is non-null→
66
←
Taking false branch→
67
←
'?' condition is true→
68
←
Taking false branch→
69
←
Taking true branch→
70
←
Loop condition is true.  Entering loop body→
71
←
Assuming the condition is false→
72
←
Taking false branch→
73
←
The value 1 is assigned to 'j'→
74
←
Loop condition is true.  Entering loop body→
75
←
Assuming the condition is true→
76
←
Taking true branch→
77
←
Assigned value is garbage or undefined
2443 
2444void ccv_cnnp_model_share_parameters(ccv_cnnp_model_t* const model, const ccv_cnnp_model_io_t parameters, const ccv_cnnp_model_t* const from_model, const ccv_cnnp_model_io_t from_parameters, ccv_cnnp_model_parameters_renamer_f renamer, void* const context)
2445{
2446	ccv_array_t* to_parameter_indices;
2447	int to_param_ref;
2448	ccv_array_t* from_parameter_indices;
2449	int from_param_ref;
2450	_ccv_cnnp_model_to_parameter_indices_and_from_parameter_indices(model, parameters, from_model, from_parameters, &to_parameter_indices, &to_param_ref, &from_parameter_indices, &from_param_ref, 1);
2451	// Should be exactly the same tensor.
2452	if (renamer == 0 && to_param_ref < 0 && from_param_ref < 0)
1
Assuming 'renamer' is not equal to null→
2453		{ assert(from_parameter_indices->rnum == to_parameter_indices->rnum)((void) sizeof ((from_parameter_indices->rnum == to_parameter_indices
->rnum) ? 1 : 0), __extension__ ({ if (from_parameter_indices
->rnum == to_parameter_indices->rnum) ; else __assert_fail
 ("from_parameter_indices->rnum == to_parameter_indices->rnum"
, "ccv_cnnp_model.c", 2453, __extension__ __PRETTY_FUNCTION__
); })); }
2454	// To models.
2455	ccv_cnnp_compiled_data_t* const to_compiled_data = model->compiled_data;
2456	assert(to_compiled_data)((void) sizeof ((to_compiled_data) ? 1 : 0), __extension__ ({
 if (to_compiled_data) ; else __assert_fail ("to_compiled_data"
, "ccv_cnnp_model.c", 2456, __extension__ __PRETTY_FUNCTION__
); }));
2
←
Assuming 'to_compiled_data' is non-null→
3
←
Taking true branch→
2457	// From models.
2458	const ccv_cnnp_compiled_data_t* const from_compiled_data = from_model->compiled_data;
2459	const int parallel_count = ccv_max(model->parallel_count, 1)({ typeof (model->parallel_count) _a = (model->parallel_count
); typeof (1) _b = (1); (_a > _b) ? _a : _b; });
4
←
Assuming '_a' is <= '_b'→
5
←
'?' condition is false→
2460	assert(parallel_count == ccv_max(from_model->parallel_count, 1))((void) sizeof ((parallel_count == ({ typeof (from_model->
parallel_count) _a = (from_model->parallel_count); typeof (
1) _b = (1); (_a > _b) ? _a : _b; })) ? 1 : 0), __extension__
 ({ if (parallel_count == ({ typeof (from_model->parallel_count
) _a = (from_model->parallel_count); typeof (1) _b = (1); (
_a > _b) ? _a : _b; })) ; else __assert_fail ("parallel_count == ccv_max(from_model->parallel_count, 1)"
, "ccv_cnnp_model.c", 2460, __extension__ __PRETTY_FUNCTION__
); })); // Should have the same parallel count can share parameters.
6
←
Assuming '_a' is <= '_b'→
7
←
'?' condition is false→
8
←
Taking true branch→
2461	const int from_parameter_size = from_compiled_data->parameters->rnum;
2462	const int to_parameter_size = to_compiled_data->parameters->rnum;
2463	const int rnum = (to_param_ref < 0 && from_param_ref < 0) ? to_parameter_indices->rnum : 1;
9
←
Assuming 'to_param_ref' is >= 0→
2464	int i, j;
2465	khash_t(ccv_cnnp_parameter_id)kh_ccv_cnnp_parameter_id_t* id_map = 0;
2466	char* updated_name = 0;
2467	const uint32_t* const from_init_v = CCV_NNC_INIT_V(from_compiled_data->tensors_init.v)((uint32_t*)((uintptr_t)(from_compiled_data->tensors_init.
v) & ~(uintptr_t)1));
2468  uint32_t* const to_init_v = CCV_NNC_INIT_V(to_compiled_data->tensors_init.v)((uint32_t*)((uintptr_t)(to_compiled_data->tensors_init.v)
 & ~(uintptr_t)1));
2469	for (i = 0; i < rnum; i++)
2470	{
2471		int src_d = (from_param_ref >= 0 ? from_param_ref : i) < from_parameter_indices->rnum ? *(int*)ccv_array_get(from_parameter_indices,from_param_ref >= 0 ? from_param_ref : i)((void*)(((char*)((from_parameter_indices)->data)) + (size_t
)(from_parameter_indices)->rsize * (size_t)(from_param_ref
 >= 0 ? from_param_ref : i))) : from_parameter_size;
10
←
Assuming 'from_param_ref' is < 0→
11
←
'?' condition is false→
12
←
Assuming the condition is false→
13
←
'?' condition is false→
2472		// Need to figure out how to use the renamer here.
2473		const int dest_d = *(int*)ccv_array_get(to_parameter_indices, to_param_ref >= 0 ? to_param_ref : i)((void*)(((char*)((to_parameter_indices)->data)) + (size_t
)(to_parameter_indices)->rsize * (size_t)(to_param_ref >=
 0 ? to_param_ref : i)));
14
←
'?' condition is true→
2474		assert(dest_d >= 0)((void) sizeof ((dest_d >= 0) ? 1 : 0), __extension__ ({ if
 (dest_d >= 0) ; else __assert_fail ("dest_d >= 0", "ccv_cnnp_model.c"
, 2474, __extension__ __PRETTY_FUNCTION__); }));
15
←
Assuming 'dest_d' is >= 0→
16
←
Taking true branch→
2475		assert(dest_d < to_parameter_size)((void) sizeof ((dest_d < to_parameter_size) ? 1 : 0), __extension__
 ({ if (dest_d < to_parameter_size) ; else __assert_fail (
"dest_d < to_parameter_size", "ccv_cnnp_model.c", 2475, __extension__
 __PRETTY_FUNCTION__); }));
17
←
Assuming 'dest_d' is < 'to_parameter_size'→
18
←
Taking true branch→
2476		if (renamer18.1
'renamer' is non-null
)
2477		{
2478			const char* const src_name = (src_d18.2
'src_d' is >= 'from_parameter_size'
 < from_parameter_size && src_d >= 0) ? *(char**)ccv_array_get(from_compiled_data->ids.parameters, src_d)((void*)(((char*)((from_compiled_data->ids.parameters)->
data)) + (size_t)(from_compiled_data->ids.parameters)->
rsize * (size_t)(src_d))) : 0;
2479			const char* const dest_name = *(char**)ccv_array_get(to_compiled_data->ids.parameters, dest_d)((void*)(((char*)((to_compiled_data->ids.parameters)->data
)) + (size_t)(to_compiled_data->ids.parameters)->rsize *
 (size_t)(dest_d)));
2480			if (!updated_name18.3
'updated_name' is null
)
19
←
Taking true branch→
2481				updated_name = (char*)ccmallocmalloc(1024);
2482			const size_t src_name_len = src_name19.1
'src_name' is equal to null
 == 0 ? 0 : ccv_min(strnlen(src_name, 1023), 1023)({ typeof (strnlen(src_name, 1023)) _a = (strnlen(src_name, 1023
)); typeof (1023) _b = (1023); (_a < _b) ? _a : _b; });
20
←
'?' condition is true→
2483			if (src_name_len20.1
'src_name_len' is <= 0
 > 0)
21
←
Taking false branch→
2484				memcpy(updated_name, src_name, src_name_len);
2485			updated_name[src_name_len] = 0;
2486			if (renamer(context, dest_name, updated_name, 1024) != 0)
22
←
Assuming the condition is false→
2487				continue; // Skip this.
2488			if (src_name22.1
'src_name' is equal to null
 != 0 && memcmp(updated_name, src_name, src_name_len) == 0 && strnlen(updated_name, 1023) == src_name_len)
2489			{
2490				// Nothing changed.
2491			} else {
2492				if (!id_map22.2
'id_map' is null
)
23
←
Taking true branch→
2493				{
2494					id_map = kh_init(ccv_cnnp_parameter_id)kh_init_ccv_cnnp_parameter_id();
2495					for (j = 0; j < from_parameter_size; j++)
24
←
Assuming 'j' is < 'from_parameter_size'→
25
←
Loop condition is true.  Entering loop body→
54
←
Assuming 'j' is < 'from_parameter_size'→
55
←
Loop condition is true.  Entering loop body→
2496					{
2497						int ret;
2498						const khiter_t k = kh_put(ccv_cnnp_parameter_id, id_map, *(char**)ccv_array_get(from_compiled_data->ids.parameters, j), &ret)kh_put_ccv_cnnp_parameter_id(id_map, *(char**)((void*)(((char
*)((from_compiled_data->ids.parameters)->data)) + (size_t
)(from_compiled_data->ids.parameters)->rsize * (size_t)
(j))), &ret);
26
←
Calling 'kh_put_ccv_cnnp_parameter_id'→
52
←
Returning from 'kh_put_ccv_cnnp_parameter_id'→
56
←
Calling 'kh_put_ccv_cnnp_parameter_id'→
2499						assert(ret != 0)((void) sizeof ((ret != 0) ? 1 : 0), __extension__ ({ if (ret
 != 0) ; else __assert_fail ("ret != 0", "ccv_cnnp_model.c", 2499
, __extension__ __PRETTY_FUNCTION__); }));
53
←
Taking true branch→
2500						kh_val(id_map, k)((id_map)->vals[k]) = j;
2501					}
2502				}
2503				const khiter_t k = kh_get(ccv_cnnp_parameter_id, id_map, updated_name)kh_get_ccv_cnnp_parameter_id(id_map, updated_name);
2504				if (k == kh_end(id_map)((id_map)->n_buckets)) // Cannot find the name, skip.
2505					continue;
2506				src_d = kh_val(id_map, k)((id_map)->vals[k]);
2507				assert(src_d >= 0)((void) sizeof ((src_d >= 0) ? 1 : 0), __extension__ ({ if
 (src_d >= 0) ; else __assert_fail ("src_d >= 0", "ccv_cnnp_model.c"
, 2507, __extension__ __PRETTY_FUNCTION__); }));
2508				assert(src_d < from_parameter_size)((void) sizeof ((src_d < from_parameter_size) ? 1 : 0), __extension__
 ({ if (src_d < from_parameter_size) ; else __assert_fail (
"src_d < from_parameter_size", "ccv_cnnp_model.c", 2508, __extension__
 __PRETTY_FUNCTION__); }));
2509			}
2510		}
2511		assert(src_d >= 0)((void) sizeof ((src_d >= 0) ? 1 : 0), __extension__ ({ if
 (src_d >= 0) ; else __assert_fail ("src_d >= 0", "ccv_cnnp_model.c"
, 2511, __extension__ __PRETTY_FUNCTION__); }));
2512		assert(src_d < from_parameter_size)((void) sizeof ((src_d < from_parameter_size) ? 1 : 0), __extension__
 ({ if (src_d < from_parameter_size) ; else __assert_fail (
"src_d < from_parameter_size", "ccv_cnnp_model.c", 2512, __extension__
 __PRETTY_FUNCTION__); }));
2513		const int s = ((ccv_nnc_tensor_symbol_t*)ccv_array_get(from_compiled_data->parameters, src_d)((void*)(((char*)((from_compiled_data->parameters)->data
)) + (size_t)(from_compiled_data->parameters)->rsize * (
size_t)(src_d))))->d;
2514		// If the original is not init'ed. We cannot share from.
2515		if (!(from_init_v[s >> 5] & (1u << (s & 0x1f))))
2516			continue;
2517		for (j = 0; j < parallel_count; j++)
2518		{
2519			ccv_nnc_tensor_t* const src = CCV_NNC_TENSOR(from_compiled_data->tensors.parameters[src_d + j * from_parameter_size])((ccv_nnc_tensor_t*)((uintptr_t)(from_compiled_data->tensors
.parameters[src_d + j * from_parameter_size]) & ~(uintptr_t
)1));
2520			assert(src)((void) sizeof ((src) ? 1 : 0), __extension__ ({ if (src) ; else
 __assert_fail ("src", "ccv_cnnp_model.c", 2520, __extension__
 __PRETTY_FUNCTION__); }));
2521			ccv_nnc_tensor_t* const dest = to_compiled_data->tensors.parameters[dest_d + j * to_parameter_size];
2522			if (dest && !((uintptr_t)dest & (uintptr_t)1))
2523				ccv_nnc_tensor_free(dest);
2524			to_compiled_data->tensors.parameters[dest_d + j * to_parameter_size] = (ccv_nnc_tensor_t*)((uintptr_t)src | (uintptr_t)1);
2525		}
2526		// Mark this symbol as init'ed.
2527		const int d = ((ccv_nnc_tensor_symbol_t*)ccv_array_get(to_compiled_data->parameters, dest_d)((void*)(((char*)((to_compiled_data->parameters)->data)
) + (size_t)(to_compiled_data->parameters)->rsize * (size_t
)(dest_d))))->d;
2528		to_init_v[d >> 5] |= (1u << (d & 0x1f));
2529	}
2530	ccv_array_free(to_parameter_indices);
2531	ccv_array_free(from_parameter_indices);
2532	if (id_map)
2533		kh_destroy(ccv_cnnp_parameter_id, id_map)kh_destroy_ccv_cnnp_parameter_id(id_map);
2534	if (updated_name)
2535		ccfreefree(updated_name);
2536	// Mark it as incomplete so we will call init_1.
2537	if (ccv_cnnp_model_tensors_any_to_alloc(model, to_compiled_data))
2538		to_compiled_data->tensors_init.v = (uint32_t*)((uintptr_t)to_compiled_data->tensors_init.v | (uintptr_t)1);
2539	else // Remove the flag.
2540		to_compiled_data->tensors_init.v = CCV_NNC_INIT_V(to_compiled_data->tensors_init.v)((uint32_t*)((uintptr_t)(to_compiled_data->tensors_init.v)
 & ~(uintptr_t)1));
2541}
2542 
2543ccv_nnc_stream_context_t* ccv_cnnp_compiled_data_get_stream(ccv_cnnp_compiled_data_t* const compiled_data, const int type)
2544{
2545	if (!compiled_data->stream_map)
2546		compiled_data->stream_map = kh_init(stream_map)kh_init_stream_map();
2547	int ret = 0;
2548	khiter_t k = kh_put(stream_map, compiled_data->stream_map, type, &ret)kh_put_stream_map(compiled_data->stream_map, type, &ret
);
2549	assert(ret >= 0)((void) sizeof ((ret >= 0) ? 1 : 0), __extension__ ({ if (
ret >= 0) ; else __assert_fail ("ret >= 0", "ccv_cnnp_model.c"
, 2549, __extension__ __PRETTY_FUNCTION__); }));
2550	ccv_nnc_stream_context_t* stream = kh_val(compiled_data->stream_map, k)((compiled_data->stream_map)->vals[k]);
2551	// If ret == 0, the key already exist, we can return directly, otherwise, create and return.
2552	if (ret != 0)
2553	{
2554		stream = ccv_nnc_stream_context_new(type);
2555		kh_val(compiled_data->stream_map, k)((compiled_data->stream_map)->vals[k]) = stream;
2556	}
2557	return stream;
2558}
2559 
2560void ccv_cnnp_model_parameters_zip_map(ccv_cnnp_model_t* const model, const ccv_cnnp_model_io_t parameters, const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint, const int flags, ccv_nnc_tensor_t* const* const aux_ins, const int aux_in_size, ccv_nnc_tensor_t* const* const aux_outs, const int aux_out_size, ccv_nnc_stream_context_t* const stream_context, const ccv_cnnp_model_t* const from_model, const ccv_cnnp_model_io_t from_parameters)
2561{
2562	ccv_array_t* to_parameter_indices;
2563	int to_param_ref;
2564	ccv_array_t* from_parameter_indices;
2565	int from_param_ref;
2566	_ccv_cnnp_model_to_parameter_indices_and_from_parameter_indices(model, parameters, from_model, from_parameters, &to_parameter_indices, &to_param_ref, &from_parameter_indices, &from_param_ref, 0);
2567	// Should be exactly the same tensor.
2568	if (to_param_ref < 0 && from_param_ref < 0)
2569		{ assert(from_parameter_indices->rnum == to_parameter_indices->rnum)((void) sizeof ((from_parameter_indices->rnum == to_parameter_indices
->rnum) ? 1 : 0), __extension__ ({ if (from_parameter_indices
->rnum == to_parameter_indices->rnum) ; else __assert_fail
 ("from_parameter_indices->rnum == to_parameter_indices->rnum"
, "ccv_cnnp_model.c", 2569, __extension__ __PRETTY_FUNCTION__
); })); }
2570	// To models.
2571	ccv_cnnp_compiled_data_t* const to_compiled_data = model->compiled_data;
2572	assert(to_compiled_data)((void) sizeof ((to_compiled_data) ? 1 : 0), __extension__ ({
 if (to_compiled_data) ; else __assert_fail ("to_compiled_data"
, "ccv_cnnp_model.c", 2572, __extension__ __PRETTY_FUNCTION__
); }));
2573	// From models.
2574	const ccv_cnnp_compiled_data_t* const from_compiled_data = from_model->compiled_data;
2575	const int parallel_count = ccv_max(model->parallel_count, 1)({ typeof (model->parallel_count) _a = (model->parallel_count
); typeof (1) _b = (1); (_a > _b) ? _a : _b; });
2576	const int to_parameter_size = to_compiled_data->parameters->rnum;
2577	const int rnum = (to_param_ref < 0 && from_param_ref < 0) ? from_parameter_indices->rnum : 1;
2578	assert(aux_in_size >= 0)((void) sizeof ((aux_in_size >= 0) ? 1 : 0), __extension__
 ({ if (aux_in_size >= 0) ; else __assert_fail ("aux_in_size >= 0"
, "ccv_cnnp_model.c", 2578, __extension__ __PRETTY_FUNCTION__
); }));
2579	assert(aux_out_size >= 0)((void) sizeof ((aux_out_size >= 0) ? 1 : 0), __extension__
 ({ if (aux_out_size >= 0) ; else __assert_fail ("aux_out_size >= 0"
, "ccv_cnnp_model.c", 2579, __extension__ __PRETTY_FUNCTION__
); }));
2580	int i, j;
2581	ccv_nnc_tensor_t* inputs[aux_in_size + 2];
2582	ccv_nnc_tensor_t* outputs[aux_out_size + 1];
2583	for (i = 0; i < aux_in_size; i++)
2584		inputs[i + 2] = aux_ins[i];
2585	for (i = 0; i < aux_out_size; i++)
2586		outputs[i + 1] = aux_outs[i];
2587	const uint32_t* const from_init_v = CCV_NNC_INIT_V(from_compiled_data->tensors_init.v)((uint32_t*)((uintptr_t)(from_compiled_data->tensors_init.
v) & ~(uintptr_t)1));
2588  uint32_t* const to_init_v = CCV_NNC_INIT_V(to_compiled_data->tensors_init.v)((uint32_t*)((uintptr_t)(to_compiled_data->tensors_init.v)
 & ~(uintptr_t)1));
2589	for (i = 0; i < rnum; i++)
2590	{
2591		const int src_d = *(int*)ccv_array_get(from_parameter_indices,from_param_ref >= 0 ? from_param_ref : i)((void*)(((char*)((from_parameter_indices)->data)) + (size_t
)(from_parameter_indices)->rsize * (size_t)(from_param_ref
 >= 0 ? from_param_ref : i)));
2592		assert(src_d >= 0)((void) sizeof ((src_d >= 0) ? 1 : 0), __extension__ ({ if
 (src_d >= 0) ; else __assert_fail ("src_d >= 0", "ccv_cnnp_model.c"
, 2592, __extension__ __PRETTY_FUNCTION__); }));
2593		assert(src_d < from_compiled_data->parameters->rnum)((void) sizeof ((src_d < from_compiled_data->parameters
->rnum) ? 1 : 0), __extension__ ({ if (src_d < from_compiled_data
->parameters->rnum) ; else __assert_fail ("src_d < from_compiled_data->parameters->rnum"
, "ccv_cnnp_model.c", 2593, __extension__ __PRETTY_FUNCTION__
); }));
2594		const int s = ((ccv_nnc_tensor_symbol_t*)ccv_array_get(from_compiled_data->parameters, src_d)((void*)(((char*)((from_compiled_data->parameters)->data
)) + (size_t)(from_compiled_data->parameters)->rsize * (
size_t)(src_d))))->d;
2595		// If the original is not init'ed. We cannot copy from.
2596		if (!(from_init_v[s >> 5] & (1u << (s & 0x1f))))
2597			continue;
2598		const int dest_d = *(int*)ccv_array_get(to_parameter_indices, to_param_ref >= 0 ? to_param_ref : i)((void*)(((char*)((to_parameter_indices)->data)) + (size_t
)(to_parameter_indices)->rsize * (size_t)(to_param_ref >=
 0 ? to_param_ref : i)));
2599		assert(dest_d >= 0)((void) sizeof ((dest_d >= 0) ? 1 : 0), __extension__ ({ if
 (dest_d >= 0) ; else __assert_fail ("dest_d >= 0", "ccv_cnnp_model.c"
, 2599, __extension__ __PRETTY_FUNCTION__); }));
2600		assert(dest_d < to_compiled_data->parameters->rnum)((void) sizeof ((dest_d < to_compiled_data->parameters->
rnum) ? 1 : 0), __extension__ ({ if (dest_d < to_compiled_data
->parameters->rnum) ; else __assert_fail ("dest_d < to_compiled_data->parameters->rnum"
, "ccv_cnnp_model.c", 2600, __extension__ __PRETTY_FUNCTION__
); }));
2601		if (parallel_count > 1)
2602		{
2603			ccv_nnc_stream_context_t* streams[parallel_count];
2604			ccv_nnc_stream_signal_t* signal;
2605			if (stream_context)
2606				signal = ccv_nnc_stream_context_emit_signal_new(stream_context);
2607			for (j = 0; j < parallel_count; j++)
2608			{
2609				ccv_nnc_tensor_t* const src = CCV_NNC_TENSOR(from_compiled_data->tensors.parameters[src_d + j * to_parameter_size])((ccv_nnc_tensor_t*)((uintptr_t)(from_compiled_data->tensors
.parameters[src_d + j * to_parameter_size]) & ~(uintptr_t
)1));
2610				ccv_nnc_tensor_t* const dest = CCV_NNC_TENSOR(to_compiled_data->tensors.parameters[dest_d + j * to_parameter_size])((ccv_nnc_tensor_t*)((uintptr_t)(to_compiled_data->tensors
.parameters[dest_d + j * to_parameter_size]) & ~(uintptr_t
)1));
2611				if (!dest || !src)
2612				{
2613					streams[j] = 0;
2614					continue;
2615				}
2616				// At the moment, can only handle them on the same device.
2617				assert(CCV_TENSOR_GET_MEMORY(src->info.type) == CCV_TENSOR_GET_MEMORY(dest->info.type))((void) sizeof ((((src->info.type) & 0x3) == ((dest->
info.type) & 0x3)) ? 1 : 0), __extension__ ({ if (((src->
info.type) & 0x3) == ((dest->info.type) & 0x3)) ; else
 __assert_fail ("CCV_TENSOR_GET_MEMORY(src->info.type) == CCV_TENSOR_GET_MEMORY(dest->info.type)"
, "ccv_cnnp_model.c", 2617, __extension__ __PRETTY_FUNCTION__
); }));
2618				assert(CCV_TENSOR_GET_DEVICE_ID(src->info.type) == CCV_TENSOR_GET_DEVICE_ID(dest->info.type))((void) sizeof (((((src->info.type) & 0xfff00) >>
 8) == (((dest->info.type) & 0xfff00) >> 8)) ? 1
 : 0), __extension__ ({ if ((((src->info.type) & 0xfff00
) >> 8) == (((dest->info.type) & 0xfff00) >>
 8)) ; else __assert_fail ("CCV_TENSOR_GET_DEVICE_ID(src->info.type) == CCV_TENSOR_GET_DEVICE_ID(dest->info.type)"
, "ccv_cnnp_model.c", 2618, __extension__ __PRETTY_FUNCTION__
); }));
2619				const int stream_type = CCV_TENSOR_GET_MEMORY(src->info.type)((src->info.type) & 0x3) == CCV_TENSOR_GPU_MEMORY ? CCV_STREAM_CONTEXT_GPU : CCV_STREAM_CONTEXT_CPU;
2620				const int device_id = CCV_TENSOR_GET_DEVICE_ID(src->info.type)(((src->info.type) & 0xfff00) >> 8);
2621				int type = stream_type;
2622				CCV_STREAM_SET_DEVICE_ID(type, device_id)(type) = (((type) & ~0xfff00) | (((device_id) & 0xfff
) << 8));
2623				ccv_nnc_stream_context_t* const stream_0 = ccv_cnnp_compiled_data_get_stream(to_compiled_data, type);
2624				// Wait signal to finish.
2625				if (stream_context)
2626					ccv_nnc_stream_context_wait_signal(stream_0, signal);
2627				inputs[0] = outputs[0] = dest;
2628				inputs[1] = src;
2629				ccv_nnc_cmd_exec(cmd, hint, flags, inputs, aux_in_size + 2, outputs, aux_out_size + 1, stream_0);
2630				if (stream_context)
2631				{
2632					ccv_nnc_stream_signal_t* const signal = ccv_nnc_stream_context_emit_signal_new(stream_0);
2633					ccv_nnc_stream_context_wait_signal(stream_context, signal);
2634				}
2635				streams[j] = stream_0;
2636			}
2637			// If this should be blocking, blocking it.
2638			if (!stream_context)
2639				for (j = 0; j < parallel_count; j++)
2640					if (streams[j])
2641						ccv_nnc_stream_context_wait(streams[j]);
2642		} else {
2643			ccv_nnc_tensor_t* const src = CCV_NNC_TENSOR(from_compiled_data->tensors.parameters[src_d])((ccv_nnc_tensor_t*)((uintptr_t)(from_compiled_data->tensors
.parameters[src_d]) & ~(uintptr_t)1));
2644			assert(src)((void) sizeof ((src) ? 1 : 0), __extension__ ({ if (src) ; else
 __assert_fail ("src", "ccv_cnnp_model.c", 2644, __extension__
 __PRETTY_FUNCTION__); }));
2645			ccv_nnc_tensor_t* const dest = CCV_NNC_TENSOR(to_compiled_data->tensors.parameters[dest_d])((ccv_nnc_tensor_t*)((uintptr_t)(to_compiled_data->tensors
.parameters[dest_d]) & ~(uintptr_t)1));
2646			assert(dest)((void) sizeof ((dest) ? 1 : 0), __extension__ ({ if (dest) ;
 else __assert_fail ("dest", "ccv_cnnp_model.c", 2646, __extension__
 __PRETTY_FUNCTION__); }));
2647			inputs[0] = outputs[0] = dest;
2648			inputs[1] = src;
2649			ccv_nnc_cmd_exec(cmd, hint, flags, inputs, aux_in_size + 2, outputs, aux_out_size + 1, stream_context);
2650		}
2651		// Mark this symbol as init'ed.
2652		const int d = ((ccv_nnc_tensor_symbol_t*)ccv_array_get(to_compiled_data->parameters, dest_d)((void*)(((char*)((to_compiled_data->parameters)->data)
) + (size_t)(to_compiled_data->parameters)->rsize * (size_t
)(dest_d))))->d;
2653		to_init_v[d >> 5] |= (1u << (d & 0x1f));
2654	}
2655	ccv_array_free(to_parameter_indices);
2656	ccv_array_free(from_parameter_indices);
2657}
2658 
2659void ccv_cnnp_model_parameters_map(ccv_cnnp_model_t* const model, const ccv_cnnp_model_io_t parameters, const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint, const int flags, ccv_nnc_tensor_t* const* const aux_ins, const int aux_in_size, ccv_nnc_tensor_t* const* const aux_outs, const int aux_out_size, ccv_nnc_stream_context_t* const stream_context)
2660{
2661	int to_param_ref;
2662	ccv_array_t* const to_parameter_indices = _ccv_cnnp_model_parameter_indices(model, parameters, &to_param_ref);
2663	// To models.
2664	ccv_cnnp_compiled_data_t* const to_compiled_data = model->compiled_data;
2665	assert(to_compiled_data)((void) sizeof ((to_compiled_data) ? 1 : 0), __extension__ ({
 if (to_compiled_data) ; else __assert_fail ("to_compiled_data"
, "ccv_cnnp_model.c", 2665, __extension__ __PRETTY_FUNCTION__
); }));
2666	// Tensor has to be inited already.
2667	assert(!!to_compiled_data->tensors_init.v)((void) sizeof ((!!to_compiled_data->tensors_init.v) ? 1 :
 0), __extension__ ({ if (!!to_compiled_data->tensors_init
.v) ; else __assert_fail ("!!to_compiled_data->tensors_init.v"
, "ccv_cnnp_model.c", 2667, __extension__ __PRETTY_FUNCTION__
); }));
2668	assert(to_compiled_data->tensors.parameters)((void) sizeof ((to_compiled_data->tensors.parameters) ? 1
 : 0), __extension__ ({ if (to_compiled_data->tensors.parameters
) ; else __assert_fail ("to_compiled_data->tensors.parameters"
, "ccv_cnnp_model.c", 2668, __extension__ __PRETTY_FUNCTION__
); }));
2669	// From models.
2670	const int parallel_count = ccv_max(model->parallel_count, 1)({ typeof (model->parallel_count) _a = (model->parallel_count
); typeof (1) _b = (1); (_a > _b) ? _a : _b; });
2671	const int to_parameter_size = to_compiled_data->parameters->rnum;
2672	const int rnum = (to_param_ref < 0) ? to_parameter_indices->rnum : 1;
2673	assert(aux_in_size >= 0)((void) sizeof ((aux_in_size >= 0) ? 1 : 0), __extension__
 ({ if (aux_in_size >= 0) ; else __assert_fail ("aux_in_size >= 0"
, "ccv_cnnp_model.c", 2673, __extension__ __PRETTY_FUNCTION__
); }));
2674	assert(aux_out_size >= 0)((void) sizeof ((aux_out_size >= 0) ? 1 : 0), __extension__
 ({ if (aux_out_size >= 0) ; else __assert_fail ("aux_out_size >= 0"
, "ccv_cnnp_model.c", 2674, __extension__ __PRETTY_FUNCTION__
); }));
2675	int i, j;
2676	ccv_nnc_tensor_t* inputs[aux_in_size + 1];
2677	ccv_nnc_tensor_t* outputs[aux_out_size + 1];
2678	for (i = 0; i < aux_in_size; i++)
2679		inputs[i + 1] = aux_ins[i];
2680	for (i = 0; i < aux_out_size; i++)
2681		outputs[i + 1] = aux_outs[i];
2682	for (i = 0; i < rnum; i++)
2683	{
2684		const int dest_d = *(int*)ccv_array_get(to_parameter_indices, to_param_ref >= 0 ? to_param_ref : i)((void*)(((char*)((to_parameter_indices)->data)) + (size_t
)(to_parameter_indices)->rsize * (size_t)(to_param_ref >=
 0 ? to_param_ref : i)));
2685		assert(dest_d >= 0)((void) sizeof ((dest_d >= 0) ? 1 : 0), __extension__ ({ if
 (dest_d >= 0) ; else __assert_fail ("dest_d >= 0", "ccv_cnnp_model.c"
, 2685, __extension__ __PRETTY_FUNCTION__); }));
2686		assert(dest_d < to_compiled_data->parameters->rnum)((void) sizeof ((dest_d < to_compiled_data->parameters->
rnum) ? 1 : 0), __extension__ ({ if (dest_d < to_compiled_data
->parameters->rnum) ; else __assert_fail ("dest_d < to_compiled_data->parameters->rnum"
, "ccv_cnnp_model.c", 2686, __extension__ __PRETTY_FUNCTION__
); }));
2687		if (parallel_count > 1)
2688		{
2689			ccv_nnc_stream_context_t* streams[parallel_count];
2690			ccv_nnc_stream_signal_t* signal;
2691			if (stream_context)
2692				signal = ccv_nnc_stream_context_emit_signal_new(stream_context);
2693			for (j = 0; j < parallel_count; j++)
2694			{
2695				ccv_nnc_tensor_t* const dest = CCV_NNC_TENSOR(to_compiled_data->tensors.parameters[dest_d + j * to_parameter_size])((ccv_nnc_tensor_t*)((uintptr_t)(to_compiled_data->tensors
.parameters[dest_d + j * to_parameter_size]) & ~(uintptr_t
)1));
2696				if (!dest)
2697				{
2698					streams[j] = 0;
2699					continue;
2700				}
2701				const int stream_type = CCV_TENSOR_GET_MEMORY(dest->info.type)((dest->info.type) & 0x3) == CCV_TENSOR_GPU_MEMORY ? CCV_STREAM_CONTEXT_GPU : CCV_STREAM_CONTEXT_CPU;
2702				const int device_id = CCV_TENSOR_GET_DEVICE_ID(dest->info.type)(((dest->info.type) & 0xfff00) >> 8);
2703				int type = stream_type;
2704				CCV_STREAM_SET_DEVICE_ID(type, device_id)(type) = (((type) & ~0xfff00) | (((device_id) & 0xfff
) << 8));
2705				ccv_nnc_stream_context_t* const stream_0 = ccv_cnnp_compiled_data_get_stream(to_compiled_data, type);
2706				// Wait signal to finish.
2707				if (stream_context)
2708					ccv_nnc_stream_context_wait_signal(stream_0, signal);
2709				inputs[0] = outputs[0] = dest;
2710				ccv_nnc_cmd_exec(cmd, hint, flags, inputs, aux_in_size + 1, outputs, aux_out_size + 1, stream_0);
2711				if (stream_context)
2712				{
2713					ccv_nnc_stream_signal_t* const signal = ccv_nnc_stream_context_emit_signal_new(stream_0);
2714					ccv_nnc_stream_context_wait_signal(stream_context, signal);
2715				}
2716				streams[j] = stream_0;
2717			}
2718			// If this should be blocking, blocking it.
2719			if (!stream_context)
2720				for (j = 0; j < parallel_count; j++)
2721					if (streams[j])
2722						ccv_nnc_stream_context_wait(streams[j]);
2723		} else {
2724			ccv_nnc_tensor_t* const dest = CCV_NNC_TENSOR(to_compiled_data->tensors.parameters[dest_d])((ccv_nnc_tensor_t*)((uintptr_t)(to_compiled_data->tensors
.parameters[dest_d]) & ~(uintptr_t)1));
2725			assert(dest)((void) sizeof ((dest) ? 1 : 0), __extension__ ({ if (dest) ;
 else __assert_fail ("dest", "ccv_cnnp_model.c", 2725, __extension__
 __PRETTY_FUNCTION__); }));
2726			inputs[0] = outputs[0] = dest;
2727			ccv_nnc_cmd_exec(cmd, hint, flags, inputs, aux_in_size + 1, outputs, aux_out_size + 1, stream_context);
2728		}
2729		// No need to mark this symbol as init'ed, it is already.
2730	}
2731	ccv_array_free(to_parameter_indices);
2732}
2733 
2734void ccv_cnnp_model_parameter_gradients_map(ccv_cnnp_model_t* const model, const ccv_cnnp_model_io_t parameters, const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint, const int flags, ccv_nnc_tensor_t* const* const aux_ins, const int aux_in_size, ccv_nnc_tensor_t* const* const aux_outs, const int aux_out_size, ccv_nnc_stream_context_t* const stream_context)
2735{
2736	int to_param_ref;
2737	ccv_array_t* const to_parameter_indices = _ccv_cnnp_model_parameter_indices(model, parameters, &to_param_ref);
2738	// To models.
2739	ccv_cnnp_compiled_data_t* const to_compiled_data = model->compiled_data;
2740	assert(to_compiled_data)((void) sizeof ((to_compiled_data) ? 1 : 0), __extension__ ({
 if (to_compiled_data) ; else __assert_fail ("to_compiled_data"
, "ccv_cnnp_model.c", 2740, __extension__ __PRETTY_FUNCTION__
); }));
2741	// Tensor has to be inited already.
2742	assert(!!to_compiled_data->tensors_init.v)((void) sizeof ((!!to_compiled_data->tensors_init.v) ? 1 :
 0), __extension__ ({ if (!!to_compiled_data->tensors_init
.v) ; else __assert_fail ("!!to_compiled_data->tensors_init.v"
, "ccv_cnnp_model.c", 2742, __extension__ __PRETTY_FUNCTION__
); }));
2743	ccv_nnc_tensor_t** tensor_gradients;
2744	if (to_compiled_data->backward.count > 1)
2745		tensor_gradients = to_compiled_data->tensors.accum_gradients;
2746	else
2747		tensor_gradients = to_compiled_data->tensors.gradients;
2748	assert(tensor_gradients)((void) sizeof ((tensor_gradients) ? 1 : 0), __extension__ ({
 if (tensor_gradients) ; else __assert_fail ("tensor_gradients"
, "ccv_cnnp_model.c", 2748, __extension__ __PRETTY_FUNCTION__
); }));
2749	// From models.
2750	const int parallel_count = ccv_max(model->parallel_count, 1)({ typeof (model->parallel_count) _a = (model->parallel_count
); typeof (1) _b = (1); (_a > _b) ? _a : _b; });
2751	const int to_parameter_size = to_compiled_data->parameters->rnum;
2752	const int rnum = (to_param_ref < 0) ? to_parameter_indices->rnum : 1;
2753	assert(aux_in_size >= 0)((void) sizeof ((aux_in_size >= 0) ? 1 : 0), __extension__
 ({ if (aux_in_size >= 0) ; else __assert_fail ("aux_in_size >= 0"
, "ccv_cnnp_model.c", 2753, __extension__ __PRETTY_FUNCTION__
); }));
2754	assert(aux_out_size >= 0)((void) sizeof ((aux_out_size >= 0) ? 1 : 0), __extension__
 ({ if (aux_out_size >= 0) ; else __assert_fail ("aux_out_size >= 0"
, "ccv_cnnp_model.c", 2754, __extension__ __PRETTY_FUNCTION__
); }));
2755	int i, j;
2756	ccv_nnc_tensor_t* inputs[aux_in_size + 1];
2757	ccv_nnc_tensor_t* outputs[aux_out_size + 1];
2758	for (i = 0; i < aux_in_size; i++)
2759		inputs[i + 1] = aux_ins[i];
2760	for (i = 0; i < aux_out_size; i++)
2761		outputs[i + 1] = aux_outs[i];
2762	for (i = 0; i < rnum; i++)
2763	{
2764		const int dest_d = *(int*)ccv_array_get(to_parameter_indices, to_param_ref >= 0 ? to_param_ref : i)((void*)(((char*)((to_parameter_indices)->data)) + (size_t
)(to_parameter_indices)->rsize * (size_t)(to_param_ref >=
 0 ? to_param_ref : i)));
2765		assert(dest_d >= 0)((void) sizeof ((dest_d >= 0) ? 1 : 0), __extension__ ({ if
 (dest_d >= 0) ; else __assert_fail ("dest_d >= 0", "ccv_cnnp_model.c"
, 2765, __extension__ __PRETTY_FUNCTION__); }));
2766		assert(dest_d < to_compiled_data->parameters->rnum)((void) sizeof ((dest_d < to_compiled_data->parameters->
rnum) ? 1 : 0), __extension__ ({ if (dest_d < to_compiled_data
->parameters->rnum) ; else __assert_fail ("dest_d < to_compiled_data->parameters->rnum"
, "ccv_cnnp_model.c", 2766, __extension__ __PRETTY_FUNCTION__
); }));
2767		if (parallel_count > 1)
2768		{
2769			ccv_nnc_stream_context_t* streams[parallel_count];
2770			ccv_nnc_stream_signal_t* signal;
2771			if (stream_context)
2772				signal = ccv_nnc_stream_context_emit_signal_new(stream_context);
2773			for (j = 0; j < parallel_count; j++)
2774			{
2775				ccv_nnc_tensor_t* const dest = tensor_gradients[dest_d + j * to_parameter_size];
2776				if (!dest)
2777				{
2778					streams[j] = 0;
2779					continue;
2780				}
2781				const int stream_type = CCV_TENSOR_GET_MEMORY(dest->info.type)((dest->info.type) & 0x3) == CCV_TENSOR_GPU_MEMORY ? CCV_STREAM_CONTEXT_GPU : CCV_STREAM_CONTEXT_CPU;
2782				const int device_id = CCV_TENSOR_GET_DEVICE_ID(dest->info.type)(((dest->info.type) & 0xfff00) >> 8);
2783				int type = stream_type;
2784				CCV_STREAM_SET_DEVICE_ID(type, device_id)(type) = (((type) & ~0xfff00) | (((device_id) & 0xfff
) << 8));
2785				ccv_nnc_stream_context_t* const stream_0 = ccv_cnnp_compiled_data_get_stream(to_compiled_data, type);
2786				// Wait signal to finish.
2787				if (stream_context)
2788					ccv_nnc_stream_context_wait_signal(stream_0, signal);
2789				inputs[0] = outputs[0] = dest;
2790				ccv_nnc_cmd_exec(cmd, hint, flags, inputs, aux_in_size + 1, outputs, aux_out_size + 1, stream_0);
2791				if (stream_context)
2792				{
2793					ccv_nnc_stream_signal_t* const signal = ccv_nnc_stream_context_emit_signal_new(stream_0);
2794					ccv_nnc_stream_context_wait_signal(stream_context, signal);
2795				}
2796				streams[j] = stream_0;
2797			}
2798			// If this should be blocking, blocking it.
2799			if (!stream_context)
2800				for (j = 0; j < parallel_count; j++)
2801					if (streams[j])
2802						ccv_nnc_stream_context_wait(streams[j]);
2803		} else {
2804			ccv_nnc_tensor_t* const dest = tensor_gradients[dest_d];
2805			if (!dest)
2806				continue;
2807			assert(dest)((void) sizeof ((dest) ? 1 : 0), __extension__ ({ if (dest) ;
 else __assert_fail ("dest", "ccv_cnnp_model.c", 2807, __extension__
 __PRETTY_FUNCTION__); }));
2808			inputs[0] = outputs[0] = dest;
2809			ccv_nnc_cmd_exec(cmd, hint, flags, inputs, aux_in_size + 1, outputs, aux_out_size + 1, stream_context);
2810		}
2811		// No need to mark this symbol as init'ed, it is already.
2812	}
2813	ccv_array_free(to_parameter_indices);
2814}
2815 
2816ccv_nnc_cmd_t ccv_cnnp_model_minimizer(ccv_cnnp_model_t* const model)
2817{
2818	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2819	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 2819, __extension__ __PRETTY_FUNCTION__); }));
2820	return compiled_data->minimize.minimizer;
2821}
2822 
2823void ccv_cnnp_model_set_minimizer(ccv_cnnp_model_t* const model, const ccv_nnc_cmd_t minimizer, const int reset, const ccv_cnnp_model_io_t* const set_parameters, const int set_parameter_size)
2824{
2825	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2826	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 2826, __extension__ __PRETTY_FUNCTION__); }));
2827	const int parameter_size = compiled_data->parameters->rnum;
2828	if (parameter_size == 0)
2829		return;
2830	if (reset)
2831		{ assert(set_parameters == 0 && set_parameter_size == 0)((void) sizeof ((set_parameters == 0 && set_parameter_size
 == 0) ? 1 : 0), __extension__ ({ if (set_parameters == 0 &&
 set_parameter_size == 0) ; else __assert_fail ("set_parameters == 0 && set_parameter_size == 0"
, "ccv_cnnp_model.c", 2831, __extension__ __PRETTY_FUNCTION__
); })); }
2832	const int old_max_saved_aux_size = compiled_data->minimize.max_saved_aux_size;
2833	const int saved_aux_size = ccv_nnc_minimizer_saved_aux_size(minimizer);
2834	if (saved_aux_size > compiled_data->minimize.max_saved_aux_size)
2835		compiled_data->minimize.max_saved_aux_size = saved_aux_size;
2836	const int max_saved_aux_size = compiled_data->minimize.max_saved_aux_size;
2837	// We update all parameters, at this point, we have one minimizer.
2838	if (set_parameters == 0 || set_parameter_size == 0)
2839		compiled_data->minimize.minimizer = minimizer;
2840	int i;
2841	if (set_parameters && set_parameter_size)
2842	{
2843		// I need to save what's the minimizer along with this.
2844		if (!compiled_data->minimize.parameters)
2845			compiled_data->minimize.parameters = ccv_array_new(sizeof(ccv_cnnp_set_minimizer_for_parameter_t*), 1, 0);
2846		ccv_cnnp_set_minimizer_for_parameter_t* const set_minimizer_for_parameter = ccmallocmalloc(sizeof(ccv_cnnp_set_minimizer_for_parameter_t) + (set_parameter_size - 1) * sizeof(ccv_cnnp_model_io_t));
2847		set_minimizer_for_parameter->minimizer = minimizer;
2848		set_minimizer_for_parameter->parameter_size = set_parameter_size;
2849		memcpy(set_minimizer_for_parameter->parameters, set_parameters, sizeof(ccv_cnnp_model_io_t) * set_parameter_size);
2850		ccv_array_push(compiled_data->minimize.parameters, &set_minimizer_for_parameter);
2851	}
2852	// If reset is true, clear the parameters array.
2853	if (reset && compiled_data->minimize.parameters)
2854	{
2855		for (i = 0; i < compiled_data->minimize.parameters->rnum; i++)
2856			ccfreefree(*(ccv_cnnp_set_minimizer_for_parameter_t**)ccv_array_get(compiled_data->minimize.parameters, i)((void*)(((char*)((compiled_data->minimize.parameters)->
data)) + (size_t)(compiled_data->minimize.parameters)->
rsize * (size_t)(i))));
2857		ccv_array_clear(compiled_data->minimize.parameters);
2858	}
2859	if (!compiled_data->update_nodes)
2860		return;
2861	ccv_nnc_symbolic_graph_t* const symbolic_graph = model->graph;
2862	assert(symbolic_graph)((void) sizeof ((symbolic_graph) ? 1 : 0), __extension__ ({ if
 (symbolic_graph) ; else __assert_fail ("symbolic_graph", "ccv_cnnp_model.c"
, 2862, __extension__ __PRETTY_FUNCTION__); }));
2863	if (saved_aux_size > old_max_saved_aux_size)
2864	{
2865		assert(compiled_data->updated_parameters)((void) sizeof ((compiled_data->updated_parameters) ? 1 : 0
), __extension__ ({ if (compiled_data->updated_parameters)
 ; else __assert_fail ("compiled_data->updated_parameters"
, "ccv_cnnp_model.c", 2865, __extension__ __PRETTY_FUNCTION__
); }));
2866		// Reallocate first, move them around later.
2867		compiled_data->updated_parameters = (ccv_nnc_tensor_symbol_t*)ccreallocrealloc(compiled_data->updated_parameters, sizeof(ccv_nnc_tensor_symbol_t) * parameter_size + sizeof(ccv_nnc_graph_exec_symbol_t) * parameter_size + sizeof(ccv_nnc_tensor_symbol_map_t) * saved_aux_size * parameter_size);
2868		compiled_data->update_nodes = (ccv_nnc_graph_exec_symbol_t*)(compiled_data->updated_parameters + parameter_size);
2869		compiled_data->saved_aux = (ccv_nnc_tensor_symbol_map_t*)(compiled_data->update_nodes + parameter_size);
2870		// We need to do this from back to front because saved_aux_size > old_saved_aux_size, it could overlap.
2871		_ccv_cnnp_scatter_saved_aux(compiled_data->saved_aux, parameter_size, old_max_saved_aux_size, saved_aux_size);
2872	}
2873	int flag = 0;
2874	const int parallel_count = ccv_max(model->parallel_count, 1)({ typeof (model->parallel_count) _a = (model->parallel_count
); typeof (1) _b = (1); (_a > _b) ? _a : _b; });
2875	if (set_parameters && set_parameter_size)
2876	{
2877		ccv_array_t* const parameter_indices = ccv_array_new(sizeof(int), 0, 0);
2878		for (i = 0; i < set_parameter_size; i++)
2879		{
2880			const int param_sel = set_parameters[i]->param_sel > 0 ? set_parameters[i]->param_sel - 1 : set_parameters[i]->param_sel;
2881			assert(set_parameters[i]->param_sel != 0)((void) sizeof ((set_parameters[i]->param_sel != 0) ? 1 : 0
), __extension__ ({ if (set_parameters[i]->param_sel != 0)
 ; else __assert_fail ("set_parameters[i]->param_sel != 0"
, "ccv_cnnp_model.c", 2881, __extension__ __PRETTY_FUNCTION__
); }));
2882			const int old_rnum = parameter_indices->rnum;
2883			ccv_cnnp_model_add_to_parameter_indices(set_parameters[i]->model, param_sel, parameter_indices);
2884			const int param_ref = set_parameters[i]->param_ref > 0 ? set_parameters[i]->param_ref - 1 : set_parameters[i]->param_ref;
2885			assert(set_parameters[i]->param_ref != 0)((void) sizeof ((set_parameters[i]->param_ref != 0) ? 1 : 0
), __extension__ ({ if (set_parameters[i]->param_ref != 0)
 ; else __assert_fail ("set_parameters[i]->param_ref != 0"
, "ccv_cnnp_model.c", 2885, __extension__ __PRETTY_FUNCTION__
); }));
2886			if (param_ref >= 0)
2887			{
2888				assert(param_ref + old_rnum < parameter_indices->rnum)((void) sizeof ((param_ref + old_rnum < parameter_indices->
rnum) ? 1 : 0), __extension__ ({ if (param_ref + old_rnum <
 parameter_indices->rnum) ; else __assert_fail ("param_ref + old_rnum < parameter_indices->rnum"
, "ccv_cnnp_model.c", 2888, __extension__ __PRETTY_FUNCTION__
); }));
2889				*(int*)ccv_array_get(parameter_indices, old_rnum)((void*)(((char*)((parameter_indices)->data)) + (size_t)(parameter_indices
)->rsize * (size_t)(old_rnum))) = *(int*)ccv_array_get(parameter_indices, param_ref + old_rnum)((void*)(((char*)((parameter_indices)->data)) + (size_t)(parameter_indices
)->rsize * (size_t)(param_ref + old_rnum)));
2890				parameter_indices->rnum = old_rnum + 1;
2891			}
2892		}
2893		// We may have duplicated indices, but that is OK, we will set it twice.
2894		for (i = 0; i < parameter_indices->rnum; i++)
2895		{
2896			const int d = *(int*)ccv_array_get(parameter_indices, i)((void*)(((char*)((parameter_indices)->data)) + (size_t)(parameter_indices
)->rsize * (size_t)(i)));
2897			if (_ccv_cnnp_set_minimizer_for_parameter(symbolic_graph, compiled_data, compiled_data->update_nodes, compiled_data->updated_parameters, compiled_data->saved_aux, parallel_count, minimizer, saved_aux_size, max_saved_aux_size, d))
2898				flag = 1;
2899		}
2900		ccv_array_free(parameter_indices);
2901	} else {
2902		for (i = 0; i < parameter_size; i++)
2903			if (_ccv_cnnp_set_minimizer_for_parameter(symbolic_graph, compiled_data, compiled_data->update_nodes, compiled_data->updated_parameters, compiled_data->saved_aux, parallel_count, minimizer, saved_aux_size, max_saved_aux_size, i))
2904				flag = 1;
2905		if (compiled_data->minimize.parameters)
2906			if (_ccv_cnnp_apply_parameters_with_minimizer(model))
2907				flag = 1;
2908	}
2909	if (flag)
2910	{
2911		// If saved_aux_size doesn't match, we need to remove / add new saved_aux to the graph. But first, free up apply gradients graph.
2912		if (compiled_data->graph_mode == CCV_CNNP_MODEL_GRAPH_FIT_MODE)
2913			_ccv_cnnp_compiled_data_graph_free(compiled_data);
2914		_ccv_cnnp_compiled_data_apply_gradients_free(compiled_data);
2915	}
2916}
2917 
2918void ccv_cnnp_model_set_compile_params(ccv_cnnp_model_t* const model, const ccv_nnc_symbolic_graph_compile_param_t compile_params)
2919{
2920	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2921	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 2921, __extension__ __PRETTY_FUNCTION__); }));
2922	compiled_data->compile_params = compile_params;
2923}
2924 
2925void ccv_cnnp_model_dot(const ccv_cnnp_model_t* const model, const int flags, FILE** const outs, const int out_size)
2926{
2927	if (model->graph && out_size > 0)
2928		ccv_nnc_symbolic_graph_dot(model->graph, flags, outs[0]);
2929	if (model->compiled_data && model->compiled_data->graph && out_size > 1)
2930		ccv_nnc_graph_dot(model->compiled_data->graph, flags, outs[1]);
2931	if (model->compiled_data && model->compiled_data->backward.accum && out_size > 2)
2932		ccv_nnc_graph_dot(model->compiled_data->backward.accum, flags, outs[2]);
2933	if (model->compiled_data && model->compiled_data->apply_gradients.graph && out_size > 3)
2934		ccv_nnc_graph_dot(model->compiled_data->apply_gradients.graph, flags, outs[3]);
2935}
2936 
2937void ccv_cnnp_model_format(const ccv_cnnp_model_t* const model, const ccv_nnc_symbolic_graph_format_f format_fn, void* const context)
2938{
2939	if (model->graph)
2940		ccv_nnc_symbolic_graph_format(model->graph, 0, 0, 0, 0, format_fn, context);
2941}
2942 
2943static void _ccv_cnnp_compiled_data_free(const ccv_cnnp_model_t* const model, ccv_cnnp_compiled_data_t* const compiled_data)
2944{
2945	int i;
2946	const int parameter_size = compiled_data->parameters->rnum;
2947	ccv_array_free(compiled_data->parameters);
2948	if (compiled_data->parameter_flags)
2949		ccfreefree(compiled_data->parameter_flags);
2950	const int internal_size = compiled_data->internals->rnum;
2951	ccv_array_free(compiled_data->internals);
2952	assert(compiled_data->ids.parameters->rnum == parameter_size)((void) sizeof ((compiled_data->ids.parameters->rnum ==
 parameter_size) ? 1 : 0), __extension__ ({ if (compiled_data
->ids.parameters->rnum == parameter_size) ; else __assert_fail
 ("compiled_data->ids.parameters->rnum == parameter_size"
, "ccv_cnnp_model.c", 2952, __extension__ __PRETTY_FUNCTION__
); }));
2953	assert(compiled_data->ids.internals->rnum == internal_size)((void) sizeof ((compiled_data->ids.internals->rnum == internal_size
) ? 1 : 0), __extension__ ({ if (compiled_data->ids.internals
->rnum == internal_size) ; else __assert_fail ("compiled_data->ids.internals->rnum == internal_size"
, "ccv_cnnp_model.c", 2953, __extension__ __PRETTY_FUNCTION__
); }));
2954	for (i = 0; i < parameter_size; i++)
2955		ccfreefree(*(char**)ccv_array_get(compiled_data->ids.parameters, i)((void*)(((char*)((compiled_data->ids.parameters)->data
)) + (size_t)(compiled_data->ids.parameters)->rsize * (
size_t)(i))));
2956	ccv_array_free(compiled_data->ids.parameters);
2957	for (i = 0; i < internal_size; i++)
2958		ccfreefree(*(char**)ccv_array_get(compiled_data->ids.internals, i)((void*)(((char*)((compiled_data->ids.internals)->data)
) + (size_t)(compiled_data->ids.internals)->rsize * (size_t
)(i))));
2959	ccv_array_free(compiled_data->ids.internals);
2960	const int parallel_count = ccv_max(model->parallel_count, 1)({ typeof (model->parallel_count) _a = (model->parallel_count
); typeof (1) _b = (1); (_a > _b) ? _a : _b; });
2961	if (compiled_data->tensors.parameters)
2962	{
2963		for (i = 0; i < parameter_size * parallel_count; i++)
2964			// If it is not marked as not belonging, we can free it.
2965			if (!((uintptr_t)compiled_data->tensors.parameters[i] & (uintptr_t)1))
2966				if (compiled_data->tensors.parameters[i])
2967					ccv_nnc_tensor_free(compiled_data->tensors.parameters[i]);
2968		for (i = 0; i < internal_size * parallel_count; i++)
2969			if (compiled_data->tensors.internals[i])
2970				ccv_nnc_tensor_free(compiled_data->tensors.internals[i]);
2971		ccfreefree(compiled_data->tensors.parameters);
2972	}
2973	if (compiled_data->tensors.gradients)
2974	{
2975		for (i = 0; i < parameter_size * parallel_count; i++)
2976		{
2977			if (compiled_data->tensors.gradients[i])
2978				ccv_nnc_tensor_free(compiled_data->tensors.gradients[i]);
2979			if (compiled_data->tensors.accum_gradients[i])
2980				ccv_nnc_tensor_free(compiled_data->tensors.accum_gradients[i]);
2981		}
2982		ccfreefree(compiled_data->tensors.gradients);
2983	}
2984	if (compiled_data->minimize.parameters)
2985	{
2986		for (i = 0; i < compiled_data->minimize.parameters->rnum; i++)
2987			ccfreefree(*(ccv_cnnp_set_minimizer_for_parameter_t**)ccv_array_get(compiled_data->minimize.parameters, i)((void*)(((char*)((compiled_data->minimize.parameters)->
data)) + (size_t)(compiled_data->minimize.parameters)->
rsize * (size_t)(i))));
2988		ccv_array_free(compiled_data->minimize.parameters);
2989	}
2990	if (compiled_data->rewindables)
2991		ccv_array_free(compiled_data->rewindables);
2992	if (compiled_data->tensors_init.v)
2993		ccfreefree(CCV_NNC_INIT_V(compiled_data->tensors_init.v)((uint32_t*)((uintptr_t)(compiled_data->tensors_init.v) &
 ~(uintptr_t)1)));
2994	if (compiled_data->evaluate.tos)
2995		ccfreefree(compiled_data->evaluate.tos);
2996	compiled_data->evaluate.tos = 0;
2997	if (compiled_data->stream_map)
2998	{
2999		khiter_t k;
3000		for (k = kh_begin(compiled_data->stream_map)(khint_t)(0); k != kh_end(compiled_data->stream_map)((compiled_data->stream_map)->n_buckets); ++k)
3001		{
3002			if (!kh_exist(compiled_data->stream_map, k)(!(((compiled_data->stream_map)->flags[(k)>>4]>>
(((k)&0xfU)<<1))&3)))
3003				continue;
3004			ccv_nnc_stream_context_t* const stream = kh_val(compiled_data->stream_map, k)((compiled_data->stream_map)->vals[k]);
3005			ccv_nnc_stream_context_free(stream);
3006		}
3007		kh_destroy(stream_map, compiled_data->stream_map)kh_destroy_stream_map(compiled_data->stream_map);
3008	}
3009	_ccv_cnnp_compiled_data_graph_free(compiled_data);
3010	_ccv_cnnp_compiled_data_gradient_free(compiled_data);
3011	_ccv_cnnp_compiled_data_backward_free(compiled_data);
3012	_ccv_cnnp_compiled_data_apply_gradients_free(compiled_data);
3013	if (compiled_data->gradient_checkpoints)
3014	{
3015		for (i = 0; i < compiled_data->gradient_checkpoints->rnum; i++)
3016		{
3017			ccv_cnnp_model_gradient_checkpoint_t* const checkpoint = (ccv_cnnp_model_gradient_checkpoint_t*)ccv_array_get(compiled_data->gradient_checkpoints, i)((void*)(((char*)((compiled_data->gradient_checkpoints)->
data)) + (size_t)(compiled_data->gradient_checkpoints)->
rsize * (size_t)(i)));
3018			assert(checkpoint->inputs)((void) sizeof ((checkpoint->inputs) ? 1 : 0), __extension__
 ({ if (checkpoint->inputs) ; else __assert_fail ("checkpoint->inputs"
, "ccv_cnnp_model.c", 3018, __extension__ __PRETTY_FUNCTION__
); }));
3019			ccfreefree(checkpoint->inputs);
3020			ccv_array_free(checkpoint->tensor_symbols);
3021		}
3022		ccv_array_free(compiled_data->gradient_checkpoints);
3023	}
3024	ccv_nnc_xpu_alloc_destroy(&compiled_data->xpu_alloc);
3025	ccfreefree(compiled_data);
3026}
3027 
3028void ccv_cnnp_model_free(ccv_cnnp_model_t* const model)
3029{
3030	if (model->isa->deinit)
3031		model->isa->deinit(model);
3032	if (model->io)
3033	{
3034		int i;
3035		for (i = 0; i < model->io->rnum; i++)
3036		{
3037			ccv_cnnp_model_io_t model_io = *(ccv_cnnp_model_io_t*)ccv_array_get(model->io, i)((void*)(((char*)((model->io)->data)) + (size_t)(model->
io)->rsize * (size_t)(i)));
3038			if (model_io->outgoings)
3039				ccv_array_free(model_io->outgoings);
3040			if (model_io->incomings)
3041				ccv_array_free(model_io->incomings);
3042			if (model_io->dependencies)
3043				ccv_array_free(model_io->dependencies);
3044			ccfreefree(model_io);
3045		}
3046		ccv_array_free(model->io);
3047	}
3048	if (model->parameter_indices)
3049		ccv_array_free(model->parameter_indices);
3050	if (model->inputs)
3051		ccfreefree(model->inputs);
3052	if (model->graph)
3053		ccv_nnc_symbolic_graph_free(model->graph);
3054	if (model->compiled_data)
3055		_ccv_cnnp_compiled_data_free(model, model->compiled_data);
3056	if (model->name)
3057		ccfreefree(model->name);
3058	ccfreefree(model);
3059}
3060 
3061void ccv_cnnp_model_cancel(ccv_cnnp_model_t* const model)
3062{
3063	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
3064	if (!compiled_data)
3065		return;
3066	if (compiled_data->graph)
3067		ccv_nnc_graph_cancel(compiled_data->graph);
3068	if (compiled_data->apply_gradients.graph)
3069		ccv_nnc_graph_cancel(compiled_data->apply_gradients.graph);
3070}