ccv_cnnp_model.c

Bug Summary

File:	nnc/ccv_cnnp_model.c
Warning:	line 2463, column 13 Array access (via field 'vals') results in a null pointer dereference
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/2024-11-04-165540-38187-1 -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 
7// MARK - Level-5 API
8 
9ccv_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)
10{
11	if (!model->io)
12		model->io = ccv_array_new(sizeof(ccv_cnnp_model_io_t), 1, 0);
13	ccv_cnnp_model_io_t model_io = ccmallocmalloc(sizeof(struct ccv_cnnp_model_io_s) + sizeof(ccv_nnc_tensor_symbol_t) * model->output_size);
14	model_io->param_ref = 0;
15	model_io->param_sel = 0;
16	model_io->visit = 0;
17	model_io->model = model;
18	model_io->dependencies = 0;
19	model_io->dependents = 0;
20	model_io->outgoings = 0;
21	model_io->outputs = (ccv_nnc_tensor_symbol_t*)(model_io + 1);
22	ccv_array_push(model->io, &model_io);
23	if (input_size > 0)
24	{
25		model_io->incomings = ccv_array_new(sizeof(ccv_cnnp_model_io_t), input_size, 0);
26		ccv_array_resize(model_io->incomings, input_size);
27		int i;
28		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);
29		for (i = 0; i < input_size; i++)
30		{
31			if (!inputs[i]->outgoings)
32				inputs[i]->outgoings = ccv_array_new(sizeof(ccv_cnnp_model_io_t), 1, 0);
33			ccv_array_push(inputs[i]->outgoings, &model_io);
34		}
35	} else {
36		model_io->incomings = 0;
37	}
38	return model_io;
39}
40 
41void ccv_cnnp_model_add_dependencies(ccv_cnnp_model_io_t model_io, const ccv_cnnp_model_io_t* const dependencies, const int dependency_size)
42{
43	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", 43, __extension__ __PRETTY_FUNCTION__);
 }));
44	if (!model_io->dependencies)
45		model_io->dependencies = ccv_array_new(sizeof(ccv_cnnp_model_io_t), dependency_size, 0);
46	int i, j;
47	for (i = 0; i < dependency_size; i++)
48	{
49		int flag = 0;
50		// Check if it is already exist or not.
51		for (j = 0; !flag && j < model_io->dependencies->rnum; j++)
52			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])
53				flag = 1;
54		if (flag)
55			continue;
56		ccv_array_push(model_io->dependencies, dependencies + i);
57		++dependencies[i]->dependents;
58	}
59}
60 
61int ccv_cnnp_model_output_size(const ccv_cnnp_model_t* const model)
62{
63	return model->output_size;
64}
65 
66int ccv_cnnp_model_is_trainable(const ccv_cnnp_model_t* const model)
67{
68	// If the model is compiled, it is default to 1 unless it is not.
69	if (model->compiled_data)
70		return model->is_trainable >= 0 ? model->is_trainable : 1;
71	return model->is_trainable;
72}
73 
74ccv_cnnp_model_io_t ccv_cnnp_model_parameters(ccv_cnnp_model_t* const model, const int selector, const int index)
75{
76	if (!model->io)
77		model->io = ccv_array_new(sizeof(ccv_cnnp_model_io_t), 1, 0);
78	ccv_cnnp_model_io_t model_io = ccmallocmalloc(sizeof(struct ccv_cnnp_model_io_s));
79	model_io->param_ref = index >= 0 ? index + 1 : ALL_PARAMETERS-1;
80	model_io->param_sel = selector >= 0 ? selector + 1 : ALL_PARAMETERS-1;
81	model_io->visit = 0;
82	model_io->model = model;
83	model_io->outputs = 0;
84	model_io->dependencies = 0;
85	model_io->dependents = 0;
86	model_io->incomings = 0;
87	model_io->outgoings = 0;
88	ccv_array_push(model->io, &model_io);
89	return model_io;
90}
91 
92void ccv_cnnp_model_notify_hook(ccv_cnnp_model_t* const model, ccv_cnnp_model_notify_f func, void* const context)
93{
94	model->notify_hook.func = func;
95	model->notify_hook.context = context;
96}
97 
98void ccv_cnnp_model_notify(const ccv_cnnp_model_t* const model, const int tag, void* const payload)
99{
100	if (model->notify_hook.func)
101		model->notify_hook.func(model, tag, payload, model->notify_hook.context);
102	if (model->isa->notify)
103		model->isa->notify(model, tag, payload);
104}
105 
106static int _ccv_nnc_array_dedup_graph_exec_symbols(ccv_nnc_graph_exec_symbol_t* const graph_exec_symbols, int graph_exec_symbol_size)
107{
108	int i, j;
109	for (i = 0; i < graph_exec_symbol_size; i++)
110	{
111		ccv_nnc_graph_exec_symbol_t* const graph_exec_symbol = graph_exec_symbols + i;
112		// Check whether this tensor symbol has any duplicate.
113		for (j = i + 1; j < graph_exec_symbol_size;)
114		{
115			ccv_nnc_graph_exec_symbol_t* const other_symbol = graph_exec_symbols + j;
116			// If there is a same tensor symbol, remove it.
117			if (other_symbol->d == graph_exec_symbol->d && other_symbol->graph == graph_exec_symbol->graph)
118			{
119				if (j + 1 < graph_exec_symbol_size)
120					*other_symbol = graph_exec_symbols[graph_exec_symbol_size - 1];
121				--graph_exec_symbol_size;
122				continue;
123			}
124			++j;
125		}
126	}
127	return graph_exec_symbol_size;
128}
129 
130void ccv_cnnp_model_add_to_array(void* const context, const ccv_nnc_tensor_symbol_t symbol, const int is_trainable)
131{
132	ccv_cnnp_model_add_to_array_context_t* const add_to_array_context = (ccv_cnnp_model_add_to_array_context_t*)context;
133	ccv_cnnp_model_t* const model = add_to_array_context->sequence->model;
134	int i;
135	if (!model->parameter_indices)
136		model->parameter_indices = ccv_array_new(sizeof(int), 0, 0);
137	for (i = 0; i < add_to_array_context->symbols->rnum; i++)
138	{
139		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)));
140		if (other_symbol.d == symbol.d && other_symbol.graph == symbol.graph)
141		{
142			// Only add to parameter_indices if it is trainable.
143			if (add_to_array_context->prefix == 't')
144				ccv_array_add_unique_int(model->parameter_indices, i);
145			// Found it, return, don't add it.
146			return;
147		}
148	}
149	// Only add to parameter_indices if it is trainable.
150	if (add_to_array_context->prefix == 't')
151		ccv_array_push(model->parameter_indices, &add_to_array_context->symbols->rnum);
152	// This is a new one, no need to add_unique_int, it is unique.
153	ccv_array_push(add_to_array_context->symbols, &symbol);
154	if (add_to_array_context->trainables)
155		ccv_array_push(add_to_array_context->trainables, &is_trainable);
156	char id[2048];
157	id[0] = add_to_array_context->prefix;
158	id[1] = '-';
159	int total_len = 2;
160	for (i = 0; i < add_to_array_context->sequence->sequences->rnum; i++)
161	{
162		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)));
163		int len;
164		if (name->name && name->name[0] != '\0')
165			len = snprintf(id + total_len, 2048 - total_len, "%s-%d-", name->name, name->sequence);
166		else
167			len = snprintf(id + total_len, 2048 - total_len, "%d-", name->sequence);
168		total_len += len;
169		if (total_len >= 2047)
170			break;
171	}
172	if (total_len < 2047)
173		total_len += snprintf(id + total_len, 2048 - total_len, "%d", add_to_array_context->sequence->it);
174	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", 174, __extension__ __PRETTY_FUNCTION__)
; }));
175	char *heap_id = (char*)ccmallocmalloc(total_len + 1);
176	memcpy(heap_id, id, total_len + 1);
177	ccv_array_push(add_to_array_context->ids, &heap_id);
178	++add_to_array_context->sequence->it;
179}
180 
181static void _ccv_cnnp_compiled_data_init(ccv_cnnp_compiled_data_t* const compiled_data, const int output_size, ccv_array_t* const gradient_checkpoints)
182{
183	compiled_data->f = compiled_data->fits + output_size;
184	compiled_data->xpu_alloc.mp_hdr = -1;
185	compiled_data->xpu_alloc.freed = kh_init(dy_str)kh_init_dy_str();
186	compiled_data->xpu_alloc.allocd = kh_init(dy_alloc)kh_init_dy_alloc();
187	compiled_data->gradient_checkpoints = gradient_checkpoints;
188}
189 
190static 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)
191{
192	assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if
 (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c"
, 192, __extension__ __PRETTY_FUNCTION__); }));
193	model->inputs = ccmallocmalloc(sizeof(ccv_nnc_tensor_symbol_t) * input_size);
194	int i;
195	for (i = 0; i < input_size; i++)
196		model->inputs[i] = ccv_nnc_tensor_symbol_new(model->graph, inputs[i], 0);
197	ccv_array_t* const parameters = ccv_array_new(sizeof(ccv_nnc_tensor_symbol_t), 0, 0);
198	ccv_array_t* const parameter_ids = ccv_array_new(sizeof(char*), 0, 0);
199	ccv_array_t* const parameter_trainables = ccv_array_new(sizeof(int), 0, 0);
200	ccv_cnnp_model_sequence_t model_sequence = {
201		.bank = kh_init(ccv_cnnp_model_name_bank)kh_init_ccv_cnnp_model_name_bank()
202	};
203	ccv_cnnp_model_add_to_array_context_t add_to_parameter_context = {
204		.sequence = &model_sequence,
205		.prefix = 't',
206		.symbols = parameters,
207		.ids = parameter_ids,
208		.trainables = parameter_trainables,
209	};
210	ccv_array_t* const internals = ccv_array_new(sizeof(ccv_nnc_tensor_symbol_t), 0, 0);
211	ccv_array_t* const internal_ids = ccv_array_new(sizeof(char*), 0, 0);
212	ccv_cnnp_model_add_to_array_context_t add_to_output_context = {
213		.sequence = &model_sequence,
214		.prefix = 'r',
215		.symbols = internals,
216		.ids = internal_ids,
217		.trainables = 0,
218	};
219	ccv_cnnp_model_build_data_t build_data = {
220		.is_trainable = model->is_trainable >= 0 ? model->is_trainable : 1,
221		.model_sequence = &model_sequence,
222		.add_to_array = ccv_cnnp_model_add_to_array,
223		.parameters = parameters,
224		.context = {
225			.add_to_parameter = &add_to_parameter_context,
226			.add_to_output = &add_to_output_context,
227		},
228		.gradient_checkpoints = 0,
229	};
230	model->data = &build_data;
231	ccv_cnnp_model_build(model, model->graph, model->inputs, input_size, 0, 0);
232	for (i = 0; i < model->output_size; i++)
233	{
234		const ccv_nnc_tensor_symbol_t output = model->outputs[i];
235		const ccv_nnc_tensor_symbol_t alias_to = ccv_nnc_tensor_symbol_alias_to(model->graph, output);
236		if (alias_to.d == CCV_NNC_NO_TENSOR_SYMBOL)
237			continue;
238		// 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
239		// 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
240		// honest, because we cannot handle cases of alias is part of the original tensor but bind differently).
241		const ccv_nnc_tensor_param_t output_params = ccv_nnc_tensor_symbol_params(model->graph, output);
242		model->outputs[i] = ccv_nnc_tensor_symbol_new(model->graph, output_params, 0);
243		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");
244		ccv_nnc_graph_exec_symbol_set_flags(model->graph, make_contiguous, CCV_NNC_GRAPH_EXEC_DISABLE_OPT);
245	}
246	model->data = 0;
247	kh_destroy(ccv_cnnp_model_name_bank, model_sequence.bank)kh_destroy_ccv_cnnp_model_name_bank(model_sequence.bank);
248	if (model_sequence.sequences)
249		ccv_array_free(model_sequence.sequences);
250	// Check if there are parameters that are not trainables. If there are, we will allocate uint64 bitmap to record that.
251	int not_trainables = 0;
252	// Assert no parameter is alias.
253	for (i = 0; i < parameters->rnum; i++)
254	{
255		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)));
256		const ccv_nnc_tensor_symbol_t alias_to = ccv_nnc_tensor_symbol_alias_to(parameter.graph, parameter);
257		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", 257, __extension__ __PRETTY_FUNCTION__)
; })); // Cannot find the one alias to.
258		if (*(int*)ccv_array_get(parameter_trainables, i)((void*)(((char*)((parameter_trainables)->data)) + (size_t
)(parameter_trainables)->rsize * (size_t)(i))) == 0)
259			not_trainables = 1;
260	}
261	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", 261, __extension__ __PRETTY_FUNCTION__)
; }));
262	uint64_t* parameter_flags = 0;
263	if (not_trainables)
264	{
265		parameter_flags = (uint64_t*)cccalloccalloc(((parameters->rnum + 63) >> 6), sizeof(uint64_t));
266		for (i = 0; i < parameter_trainables->rnum; i++)
267			if (*(int*)ccv_array_get(parameter_trainables, i)((void*)(((char*)((parameter_trainables)->data)) + (size_t
)(parameter_trainables)->rsize * (size_t)(i))))
268				parameter_flags[i >> 6] |= ((uint64_t)1 << (i & 63));
269	}
270	ccv_array_free(parameter_trainables);
271	// Assert no internal is alias.
272	for (i = 0; i < internals->rnum; i++)
273	{
274		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)));
275		const ccv_nnc_tensor_symbol_t alias_to = ccv_nnc_tensor_symbol_alias_to(internal.graph, internal);
276		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", 276, __extension__ __PRETTY_FUNCTION__)
; })); // Cannot find the one alias to.
277	}
278	const int output_size = model->output_size;
279	ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
280	const int parameters_rnum = parameters->rnum;
281	if (input_size > 0)
282	{
283		ccv_array_resize(parameters, parameters_rnum + input_size);
284		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));
285	}
286	ccv_nnc_symbolic_graph_simplify(model->graph,
287		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)
288			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)
289			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)
290			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),
291		ccv_array_get(parameters, 0)((void*)(((char*)((parameters)->data)) + (size_t)(parameters
)->rsize * (size_t)(0))), parameters_rnum + input_size,
292		model->outputs, output_size,
293		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));
294	ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
295	// Size it down.
296	parameters->rnum = parameters_rnum;
297	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));
298	_ccv_cnnp_compiled_data_init(compiled_data, output_size, build_data.gradient_checkpoints);
299	const int evaluate_to_size = compiled_data->evaluate.to_size = ccv_nnc_symbolic_graph_destination_size(model->graph);
300	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", 300, __extension__ __PRETTY_FUNCTION__)
; }));
301	compiled_data->evaluate.tos = ccmallocmalloc(sizeof(ccv_nnc_graph_exec_symbol_t) * evaluate_to_size);
302	memcpy(compiled_data->evaluate.tos, ccv_nnc_symbolic_graph_destinations(model->graph), sizeof(ccv_nnc_graph_exec_symbol_t) * evaluate_to_size);
303	compiled_data->loss = loss;
304	if (loss.cmd == CCV_NNC_NOOP)
305	{
306		// If no loss function provided, there is no fits.
307		for (i = 0; i < output_size; i++)
308		{
309			compiled_data->fits[i] = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
310			const ccv_nnc_tensor_symbol_t alias_to = ccv_nnc_tensor_symbol_alias_to(model->graph, model->outputs[i]);
311			if (alias_to.d < 0)
312				compiled_data->f[i] = model->outputs[i];
313			else { // We cannot differentiate against an alias, therefore, we have to verify this output is full, and we can diff against the original.
314				int ofs[CCV_NNC_MAX_DIM_ALLOC(12)];
315				int inc[CCV_NNC_MAX_DIM_ALLOC(12)];
316				ccv_nnc_tensor_symbol_alias_params(model->graph, model->outputs[i], ofs, inc);
317				int j;
318				for (j = 0; j < CCV_NNC_MAX_DIM_ALLOC(12); j++)
319					{ 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"
, 319, __extension__ __PRETTY_FUNCTION__); })); } // There is no ofs.
320				compiled_data->f[i] = alias_to; // Unfortunately, I cannot assert the size yet.
321			}
322		}
323	} else {
324		for (i = 0; i < output_size; i++)
325		{
326			const ccv_nnc_tensor_param_t info = ccv_nnc_tensor_symbol_params(model->graph, model->outputs[i]);
327			const ccv_nnc_tensor_symbol_t fit = compiled_data->fits[i] = ccv_nnc_tensor_symbol_new(model->graph, info, 0);
328			compiled_data->f[i] = ccv_nnc_tensor_symbol_new(model->graph, ccv_nnc_tensor_auto, 0);
329			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);
330		}
331	}
332	ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
333	ccv_nnc_symbolic_graph_simplify(model->graph,
334		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.
335		0, 0, // No need to provide binds at this point.
336		compiled_data->f, model->output_size,
337		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));
338	ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
339	// If inputs are from GPU, stream type is GPU.
340	compiled_data->parameters = parameters;
341	compiled_data->parameter_flags = parameter_flags;
342	compiled_data->internals = internals;
343	compiled_data->ids.parameters = parameter_ids;
344	compiled_data->ids.internals = internal_ids;
345	ccv_cnnp_model_gradient_checkpoints_cleanup_after_build(compiled_data, model->graph);
346}
347 
348static 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)
349{
350	ccv_array_t* const stack = (ccv_array_t*)context;
351	ccv_array_push(stack, &symbol.d);
352}
353 
354static 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)
355{
356	const ccv_nnc_tensor_symbol_t src_symbol = {
357		.d = src_index,
358		.graph = src_graph
359	};
360	const ccv_nnc_tensor_symbol_t dest_symbol = {
361		.d = dest_index,
362		.graph = dest_graph
363	};
364	const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(src_graph, src_symbol);
365	ccv_nnc_tensor_symbol_set(dest_graph, dest_symbol, params);
366	int ofs[CCV_NNC_MAX_DIM_ALLOC(12)];
367	int inc[CCV_NNC_MAX_DIM_ALLOC(12)];
368	if (0 == ccv_nnc_tensor_symbol_alias_params(src_graph, src_symbol, ofs, inc))
369		ccv_nnc_tensor_symbol_alias_set(dest_graph, dest_symbol, ofs, inc);
370}
371 
372static 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)
373{
374	const ccv_nnc_tensor_symbol_t src_symbol = {
375		.d = src_index,
376		.graph = src_graph
377	};
378	const ccv_nnc_tensor_param_t src_params = ccv_nnc_tensor_symbol_params(src_graph, src_symbol);
379	const ccv_nnc_tensor_symbol_t dest_symbol = {
380		.d = dest_index,
381		.graph = dest_graph
382	};
383	const ccv_nnc_tensor_param_t dest_params = ccv_nnc_tensor_symbol_params(dest_graph, dest_symbol);
384	return memcmp(src_params.dim, dest_params.dim, sizeof(src_params.dim)) == 0;
385}
386 
387static 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);
388static void _ccv_cnnp_compiled_data_graph_free(ccv_cnnp_compiled_data_t* const compiled_data);
389 
390typedef struct {
391	int parallel_count;
392	ccv_nnc_symbolic_graph_t* graph;
393	ccv_nnc_graph_exec_arena_t* graph_exec_arena;
394} ccv_nnc_graph_exec_update_t;
395 
396static 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)
397{
398	ccv_nnc_graph_exec_update_t* const graph_exec_update = (ccv_nnc_graph_exec_update_t*)context;
399	ccv_nnc_graph_exec_arena_t* const graph_exec_arena = graph_exec_update->graph_exec_arena;
400	ccv_nnc_graph_exec_t graph_exec = ccv_nnc_graph_exec_from_symbol(graph_exec_arena, symbol);
401	ccv_nnc_graph_exec_set(graph_exec.graph, graph_exec, cmd);
402	ccv_nnc_graph_exec_set_hint(graph_exec.graph, graph_exec, hint);
403	const ccv_nnc_symbolic_graph_t* const graph = graph_exec_update->graph;
404	const int parallel_count = graph_exec_update->parallel_count;
405	int i;
406	for (i = 1; i < parallel_count; i++)
407	{
408		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));
409		if (!CCV_NO_GRAPH_EXEC(copy)((copy).graph == 0))
410		{
411			ccv_nnc_graph_exec_set(copy.graph, copy, cmd);
412			ccv_nnc_graph_exec_set_hint(copy.graph, copy, hint);
413		}
414	}
415}
416 
417void 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)
418{
419	assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if
 (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c"
, 419, __extension__ __PRETTY_FUNCTION__); }));
420	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", 420, __extension__ __PRETTY_FUNCTION__)
; }));
421	assert(!init->graph)((void) sizeof ((!init->graph) ? 1 : 0), __extension__ ({ if
 (!init->graph) ; else __assert_fail ("!init->graph", "ccv_cnnp_model.c"
, 421, __extension__ __PRETTY_FUNCTION__); }));
422	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
423	init->graph = ccv_nnc_symbolic_graph_new();
424	ccv_array_t* const stack = ccv_array_new(sizeof(int), 0, 0);
425	ccv_nnc_graph_exec_symbol_new_hook(init->graph, _ccv_cnnp_graph_push_graph_exec_symbol, stack, 0);
426	_ccv_cnnp_model_compile(init, inputs, input_size, compiled_data->loss);
427	init->parallel_count = model->parallel_count;
428	init->memory_compression = model->memory_compression;
429	init->memory_reduction = model->memory_reduction;
430	init->gradient_checkpointing = model->gradient_checkpointing;
431	init->compiled_data->stream_type = model->compiled_data->stream_type;
432	init->compiled_data->minimize.minimizer = model->compiled_data->minimize.minimizer;
433	init->compiled_data->minimize.max_saved_aux_size = model->compiled_data->minimize.max_saved_aux_size;
434	if (model->compiled_data->gradient_mode != CCV_CNNP_COMPILED_DATA_GRADIENT_NONE)
435		_ccv_cnnp_model_gradient_init(init, model->compiled_data->gradient_mode, model->compiled_data->disable_outgrad, 0, 0);
436	ccv_nnc_graph_exec_symbol_new_hook(init->graph, 0, 0, 0);
437	ccv_nnc_symbolic_graph_tensor_auto(init->graph, TRAVERSE_FULL0,0,0,0);
438	int i, j;
439	// Verify parameters, internals and saved_aux in both graph has the same dimensionality.
440	for (i = 0; i < compiled_data->parameters->rnum; i++)
441	{
442		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;
443		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", 443, __extension__ __PRETTY_FUNCTION__)
; }));
444	}
445	for (i = 0; i < compiled_data->internals->rnum; i++)
446	{
447		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;
448		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", 448, __extension__ __PRETTY_FUNCTION__)
; }));
449	}
450	// Update inputs.
451	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", 451, __extension__ __PRETTY_FUNCTION__)
; }));
452	for (i = 0; i < model->input_size; i++)
453		if (model->inputs[i].d >= 0)
454		{
455			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", 455, __extension__ __PRETTY_FUNCTION__)
; }));
456			_ccv_nnc_tensor_symbol_reinit(init->graph, model->graph, init->inputs[i].d, model->inputs[i].d);
457		}
458	// Update outputs.
459	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", 459, __extension__ __PRETTY_FUNCTION__)
; }));
460	for (i = 0; i < model->output_size; i++)
461	{
462		if (model->outputs[i].d >= 0)
463		{
464			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", 464, __extension__
 __PRETTY_FUNCTION__); }));
465			_ccv_nnc_tensor_symbol_reinit(init->graph, model->graph, init->outputs[i].d, model->outputs[i].d);
466		}
467		if (model->outputs[i].d != model->compiled_data->f[i].d)
468		{
469			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", 469, __extension__ __PRETTY_FUNCTION__)
; }));
470			if (model->compiled_data->f[i].d >= 0)
471			{
472				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", 472, __extension__ __PRETTY_FUNCTION__)
; }));
473				_ccv_nnc_tensor_symbol_reinit(init->graph, model->graph, init->compiled_data->f[i].d, model->compiled_data->f[i].d);
474			}
475		}
476	}
477	// Go through the graph to set tensor on matching symbols
478	for (i = 0; i < stack->rnum; i++)
479	{
480		const int d = *(int*)ccv_array_get(stack, i)((void*)(((char*)((stack)->data)) + (size_t)(stack)->rsize
 * (size_t)(i)));
481		// If exceed range, skip.
482		if (d >= ccv_nnc_graph_exec_symbol_count(init->graph) ||
483			d >= ccv_nnc_graph_exec_symbol_count(model->graph))
484			continue;
485		const ccv_nnc_graph_exec_symbol_t src_symbol = {
486			.d = d,
487			.graph = init->graph
488		};
489		const ccv_nnc_graph_exec_symbol_t dest_symbol = {
490			.d = d,
491			.graph = model->graph
492		};
493		const ccv_nnc_cmd_t src_cmd = ccv_nnc_graph_exec_symbol_cmd(init->graph, src_symbol);
494		const ccv_nnc_cmd_t dest_cmd = ccv_nnc_graph_exec_symbol_cmd(model->graph, dest_symbol);
495		// If the name doesn't match, skip.
496		if (dest_cmd.cmd != src_cmd.cmd && src_cmd.cmd != CCV_NNC_NOOP)
497			continue;
498		// Now get all the inputs and outputs, if matches, set them.
499		const int* src_inputs;
500		int src_input_size;
501		const int* src_outputs;
502		int src_output_size;
503		ccv_nnc_graph_exec_symbol_io(init->graph, src_symbol, &src_inputs, &src_input_size, &src_outputs, &src_output_size);
504		const int* dest_inputs;
505		int dest_input_size;
506		const int* dest_outputs;
507		int dest_output_size;
508		ccv_nnc_graph_exec_symbol_io(model->graph, dest_symbol, &dest_inputs, &dest_input_size, &dest_outputs, &dest_output_size);
509		// We may have unmatched input / output size because this is the minimizer and it has
510		// different saved_aux (for example, when we shrunk with CMD_NOOP).
511		if (src_input_size != dest_input_size)
512			continue;
513		if (src_output_size != dest_output_size)
514			continue;
515		ccv_nnc_graph_exec_symbol_set(model->graph, dest_symbol, src_cmd);
516		// There may be mismatches of the source tensor symbols and destination tensor symbols. The reason is because
517		// we may later passed-in the minimizer, therefore, we may allocate tensors for minimizer later in the original
518		// graph whereas in the newly created graph, it is streamlined (the minimizer exists from the beginning). That
519		// will make the order of tensor symbols creation different, therefore, exact which tensor is which wrong as
520		// well. However, set a new minimizer won't change the exec symbol ordering, because we never create new exec
521		// symbols after gradient init step. Changing a new minimizer just updated that exec symbols setting, it is not
522		// a new exec symbol.
523		for (j = 0; j < src_input_size; j++)
524			if (src_inputs[j] >= 0)
525				_ccv_nnc_tensor_symbol_reinit(init->graph, model->graph, src_inputs[j], dest_inputs[j]);
526		for (j = 0; j < src_output_size; j++)
527			if (src_outputs[j] >= 0)
528				_ccv_nnc_tensor_symbol_reinit(init->graph, model->graph, src_outputs[j], dest_outputs[j]);
529	}
530	ccv_array_free(stack);
531	// After this, we get all tensors in the model graph resolved through tensor_auto.
532	ccv_nnc_symbolic_graph_tensor_auto(model->graph, TRAVERSE_FULL0,0,0,0);
533	// Verify symbols we get matches.
534	const int parameter_size = compiled_data->parameters->rnum;
535	for (i = 0; i < parameter_size; i++)
536		{ 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", 536, __extension__ __PRETTY_FUNCTION__)
; })); }
537	const int internal_size = compiled_data->internals->rnum;
538	for (i = 0; i < internal_size; i++)
539		{ 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", 539, __extension__ __PRETTY_FUNCTION__)
; })); }
540	// Go through compiled data.
541	if (compiled_data->tensor_arena)
542	{
543		const int flag = ccv_nnc_tensor_arena_reinit(compiled_data->tensor_arena, model->graph);
544		if (flag == 0 && compiled_data->graph_exec_arena)
545		{
546			ccv_nnc_graph_exec_reinit(compiled_data->graph_exec_arena, compiled_data->graph, model->graph);
547			// Since we will reinit, if we previously set is_test, we need to set it again.
548			if (compiled_data->is_test)
549			{
550				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; });
551				ccv_nnc_graph_exec_update_t update = {
552					.parallel_count = parallel_count,
553					.graph = model->graph,
554					.graph_exec_arena = compiled_data->graph_exec_arena,
555				};
556				ccv_cnnp_model_set_is_test(model, 1, _ccv_cnnp_cmd_update_for_execs, &update);
557			}
558		} else
559			// Free-up tensor arena & graph exec arena.
560			_ccv_cnnp_compiled_data_graph_free(compiled_data);
561	}
562	// There are other compiled graphs, for accum and apply gradients.
563	// However, the main conclusion is, these absorb operations shouldn't impact parameters.
564	// Thus, it won't impact the shape of gradients (only outgrad). Since for outgrad, we
565	// don't allocate ourselves, it is not a concern. For normal gradients, the shape cannot
566	// be changed otherwise parameters' shape will be meaningless. The same goes to internals.
567	// That is why we don't update these compiled graphs at all this point.
568	// Free the model, we've already "absorbed" it.
569	ccv_cnnp_model_free(init);
570}
571 
572void 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)
573{
574	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", 574, __extension__ __PRETTY_FUNCTION__)
; }));
575	if (model->input_size == 0)
576		model->input_size = input_size;
577	if (!model->graph) // The graph is not compiled yet.
578	{
579		model->graph = ccv_nnc_symbolic_graph_new();
580		_ccv_cnnp_model_compile(model, inputs, input_size, loss);
581		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", 581, __extension__ __PRETTY_FUNCTION__)
; }));
582		int i, flag = 0;
583		for (i = 0; !flag && i < input_size; i++)
584			flag = (CCV_TENSOR_GET_MEMORY(inputs[i].type)((inputs[i].type) & 0x3) == CCV_TENSOR_GPU_MEMORY);
585		// If inputs are from GPU, stream type is GPU.
586		model->compiled_data->stream_type = flag ? CCV_STREAM_CONTEXT_GPU : CCV_STREAM_CONTEXT_CPU;
587		model->compiled_data->minimize.minimizer = minimizer;
588		model->compiled_data->minimize.max_saved_aux_size = ccv_nnc_minimizer_saved_aux_size(minimizer);
589	} else {
590		// Now, finally fill in this part. If the graph is already compiled, we make a copy of the model.
591		// And then absorb the "new model" to the old one.
592		ccv_cnnp_model_t* const init = ccv_cnnp_model_copy(model, model->is_trainable);
593		ccv_cnnp_model_absorb(model, init, inputs, input_size);
594		// Reset minimizer.
595		ccv_cnnp_model_set_minimizer(model, minimizer, 1, 0, 0);
596	}
597}
598 
599ccv_cnnp_model_t* ccv_cnnp_model_copy(const ccv_cnnp_model_t* const model, const int is_trainable)
600{
601	ccv_cnnp_model_t* const new_model = _ccv_cnnp_model_copy(model, 0);
602	new_model->is_trainable = is_trainable;
603	return new_model;
604}
605 
606void ccv_cnnp_model_tensor_auto(ccv_cnnp_model_t* const model, ccv_nnc_tensor_param_t* const outputs, const int output_size)
607{
608	assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if
 (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c"
, 608, __extension__ __PRETTY_FUNCTION__); }));
609	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", 609, __extension__ __PRETTY_FUNCTION__)
; }));
610	ccv_nnc_symbolic_graph_t* const graph = model->graph;
611	ccv_nnc_symbolic_graph_tensor_auto(graph, TRAVERSE_FULL0,0,0,0);
612	int i;
613	for (i = 0; i < output_size; i++)
614	{
615		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", 615, __extension__ __PRETTY_FUNCTION__)
; }));
616		outputs[i] = ccv_nnc_tensor_symbol_params(graph, model->outputs[i]);
617	}
618}
619 
620void ccv_cnnp_model_set_workspace_size(ccv_cnnp_model_t* const model, size_t workspace_size)
621{
622	if (workspace_size == model->workspace_size)
623		return;
624	model->workspace_size = workspace_size;
625	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
626	if (compiled_data && compiled_data->graph)
627		ccv_nnc_graph_autotune(compiled_data->graph, workspace_size, 0, TRAVERSE_FULL0,0,0,0);
628}
629 
630size_t ccv_cnnp_model_workspace_size(ccv_cnnp_model_t* const model)
631{
632	return model->workspace_size;
633}
634 
635void ccv_cnnp_model_set_data_parallel(ccv_cnnp_model_t* const model, const int parallel)
636{
637	if (parallel == 0)
638		model->parallel_count = ccv_nnc_device_count(CCV_STREAM_CONTEXT_GPU);
639	else
640		model->parallel_count = parallel;
641	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
642	if (compiled_data)
643		{ 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", 643, __extension__ __PRETTY_FUNCTION__)
; })); }
644}
645 
646void ccv_cnnp_model_set_max_concurrency(ccv_cnnp_model_t* const model, const int max_stream_count)
647{
648	model->max_stream_count = max_stream_count;
649	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
650	if (compiled_data)
651		{ 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", 651, __extension__ __PRETTY_FUNCTION__)
; })); }
652}
653 
654void ccv_cnnp_model_set_memory_compression(ccv_cnnp_model_t* const model, const int memory_compression)
655{
656	model->memory_compression = memory_compression;
657	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
658	if (compiled_data)
659		{ 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", 659, __extension__ __PRETTY_FUNCTION__)
; })); }
660}
661 
662void ccv_cnnp_model_set_memory_reduction(ccv_cnnp_model_t* const model, const int memory_reduction)
663{
664	model->memory_reduction = memory_reduction;
665	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
666	if (compiled_data)
667		{ 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", 667, __extension__ __PRETTY_FUNCTION__)
; })); }
668}
669 
670void ccv_cnnp_model_set_gradient_checkpointing(ccv_cnnp_model_t* const model, const int gradient_checkpointing)
671{
672	model->gradient_checkpointing = gradient_checkpointing;
673}
674 
675int ccv_cnnp_model_gradient_checkpointing(ccv_cnnp_model_t* const model)
676{
677	return model->gradient_checkpointing;
678}
679 
680typedef struct {
681	int parallel_count;
682	ccv_nnc_symbolic_graph_t* graph;
683	ccv_cnnp_compiled_data_t* compiled_data;
684	ccv_nnc_tensor_arena_t* tensor_arena;
685} ccv_nnc_tensor_init_states_t;
686 
687static int _ccv_cnnp_any_to_init(const ccv_cnnp_compiled_data_t* const compiled_data)
688{
689	int i;
690	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));
691	for (i = 0; i < compiled_data->parameters->rnum; i++)
692	{
693		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;
694		if (!(init_v[d >> 5] & (1u << (d & 0x1f))))
695			return 1;
696	}
697	for (i = 0; i < compiled_data->internals->rnum; i++)
698	{
699		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;
700		if (!(init_v[d >> 5] & (1u << (d & 0x1f))))
701			return 1;
702	}
703	return 0;
704}
705 
706static 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)
707{
708	ccv_nnc_tensor_init_states_t* const tensor_init_states = (ccv_nnc_tensor_init_states_t*)context;
709	ccv_nnc_tensor_arena_t* const tensor_arena = tensor_init_states->tensor_arena;
710	ccv_nnc_tensor_t* const output_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, output_symbol);
711	if (!output_tensor)
712		return;
713	const int d = output_symbol.d;
714	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", 714, __extension__ __PRETTY_FUNCTION__)
; }));
715  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));
716	if (init_v[d >> 5] & (1u << (d & 0x1f)))
717		return;
718	init_v[d >> 5] |= (1u << (d & 0x1f));
719	ccv_nnc_cmd_exec(cmd, hint, flags, &input, input ? 1 : 0, &output_tensor, 1, 0);
720	const ccv_nnc_symbolic_graph_t* const graph = tensor_init_states->graph;
721	const int parallel_count = tensor_init_states->parallel_count;
722	int i;
723	for (i = 1; i < parallel_count; i++)
724	{
725		ccv_nnc_tensor_t* const copy = ccv_nnc_tensor_from_symbol(tensor_arena, ccv_nnc_tensor_symbol_copy(graph, output_symbol, i));
726		if (copy)
727			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);
728	}
729}
730 
731// This method can only handle cases we added new tensors and exec, never delete. This invariant is true because
732// we setup everything (including calling simplify method) in ccv_cnnp_model_compile method, before this rewind setup.
733static void _ccv_cnnp_model_rewind_graph(ccv_cnnp_model_t* const model)
734{
735	assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if
 (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c"
, 735, __extension__ __PRETTY_FUNCTION__); }));
736	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", 736, __extension__ __PRETTY_FUNCTION__)
; }));
737	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
738	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", 738, __extension__
 __PRETTY_FUNCTION__); }));
739	int i;
740	for (i = 0; i < compiled_data->rewindables->rnum; i++)
741	{
742		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)));
743		if (rewind_symbol->type == CCV_CNNP_REWIND_GRAPH_EXEC)
744			ccv_nnc_graph_exec_symbol_free(model->graph, rewind_symbol->graph_exec);
745		else if (rewind_symbol->type == CCV_CNNP_REWIND_TENSOR)
746			ccv_nnc_tensor_symbol_free(model->graph, rewind_symbol->tensor);
747	}
748	ccv_array_clear(compiled_data->rewindables);
749	ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
750}
751 
752static 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)
753{
754	const ccv_cnnp_rewind_symbol_t rewind_symbol = {
755		.type = CCV_CNNP_REWIND_TENSOR,
756		.tensor = symbol
757	};
758	ccv_array_t* const rewind_symbols = (ccv_array_t*)context;
759	ccv_array_push(rewind_symbols, &rewind_symbol);
760}
761 
762static 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)
763{
764	const ccv_cnnp_rewind_symbol_t rewind_symbol = {
765		.type = CCV_CNNP_REWIND_TENSOR,
766		.tensor = symbol
767	};
768	ccv_array_t* const rewind_symbols = (ccv_array_t*)context;
769	ccv_array_push(rewind_symbols, &rewind_symbol);
770}
771 
772static 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)
773{
774	const ccv_cnnp_rewind_symbol_t rewind_symbol = {
775		.type = CCV_CNNP_REWIND_GRAPH_EXEC,
776		.graph_exec = symbol
777	};
778	ccv_array_t* const rewind_symbols = (ccv_array_t*)context;
779	ccv_array_push(rewind_symbols, &rewind_symbol);
780}
781 
782static 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)
783{
784	ccv_nnc_graph_exec_t const update_exec = ccv_nnc_graph_exec_from_symbol(graph_exec_arena, exec_symbol);
785	if (!CCV_NO_GRAPH_EXEC(update_exec)((update_exec).graph == 0))
786		ccv_nnc_graph_exec_set(update_exec.graph, update_exec, cmd);
787	int i;
788	for (i = 1; i < parallel_count; i++)
789	{
790		ccv_nnc_graph_exec_symbol_t copy_symbol = ccv_nnc_graph_exec_symbol_copy(symbolic_graph, exec_symbol, i);
791		const ccv_nnc_graph_exec_t copy = ccv_nnc_graph_exec_from_symbol(graph_exec_arena, copy_symbol);
792		if (!CCV_NO_GRAPH_EXEC(copy)((copy).graph == 0))
793			ccv_nnc_graph_exec_set(copy.graph, copy, cmd);
794	}
795}
796 
797static 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)
798{
799	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 799, __extension__ __PRETTY_FUNCTION__); }));
800	assert(symbolic_graph)((void) sizeof ((symbolic_graph) ? 1 : 0), __extension__ ({ if
 (symbolic_graph) ; else __assert_fail ("symbolic_graph", "ccv_cnnp_model.c"
, 800, __extension__ __PRETTY_FUNCTION__); }));
801	ccv_nnc_graph_exec_symbol_set(symbolic_graph, exec_symbol, cmd);
802	int i;
803	for (i = 1; i < parallel_count; i++)
804	{
805		ccv_nnc_graph_exec_symbol_t copy_symbol = ccv_nnc_graph_exec_symbol_copy(symbolic_graph, exec_symbol, i);
806		if (copy_symbol.graph)
807			ccv_nnc_graph_exec_symbol_set(symbolic_graph, copy_symbol, cmd);
808	}
809	ccv_nnc_graph_exec_arena_t* const graph_exec_arena = compiled_data->graph_exec_arena;
810	if (graph_exec_arena)
811		_ccv_cnnp_model_graph_symbol_exec_set_for_graph_exec_arena(graph_exec_arena, parallel_count, exec_symbol, cmd, symbolic_graph);
812	// Skip backward graph exec arena because it is for a specific accum symbolic graph, not the main graph (model->graph)
813	ccv_nnc_graph_exec_arena_t* const gradient_graph_exec_arena = compiled_data->apply_gradients.graph_exec_arena;
814	if (gradient_graph_exec_arena)
815		_ccv_cnnp_model_graph_symbol_exec_set_for_graph_exec_arena(gradient_graph_exec_arena, parallel_count, exec_symbol, cmd, symbolic_graph);
816}
817 
818static 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)
819{
820	int this_parameter_flag = 0;
821	if (update_nodes[parameter_indice].d == CCV_NNC_NO_TENSOR_SYMBOL)
822		return this_parameter_flag;
823	const ccv_nnc_cmd_t old_minimizer = ccv_nnc_graph_exec_symbol_cmd(graph, update_nodes[parameter_indice]);
824	int j, k;
825	// For no-op, we can preserve previous saved_aux_size.
826	if (old_minimizer.cmd != minimizer.cmd && minimizer.cmd != CCV_NNC_NOOP)
827	{
828		// If the old minimizer is a noop, then the old_saved_aux_size should be whatever its previous
829		// saved_aux_size is, otherwise we will reinit the saved_aux repeatedly if you switch between
830		// noop and a minimizer. We don't want that because we do that in high-level frameworks to
831		// make sure some model parameters don't update if we don't want them to.
832		int old_saved_aux_size;
833		if (old_minimizer.cmd == CCV_NNC_NOOP)
834		{
835			int input_size;
836			ccv_nnc_graph_exec_symbol_io(graph, update_nodes[parameter_indice], 0, &input_size, 0, 0);
837			if (input_size < 2) // This is not legit.
838				old_saved_aux_size = ccv_nnc_minimizer_saved_aux_size(old_minimizer);
839			else // See ccv_nnc_minimizer_saved_aux_size, the saved_aux is inputs excluding gradients and parameters.
840				old_saved_aux_size = input_size - 2;
841		} else
842			old_saved_aux_size = ccv_nnc_minimizer_saved_aux_size(old_minimizer);
843		if (old_saved_aux_size != saved_aux_size)
844		{
845			this_parameter_flag = 1;
846			if (saved_aux_size > old_saved_aux_size)
847			{
848				// Allocate new tensor symbols.
849				const ccv_nnc_tensor_param_t info = ccv_nnc_tensor_symbol_params(graph, updated_parameters[parameter_indice]);
850				for (j = old_saved_aux_size; j < saved_aux_size; j++)
851				{
852					saved_aux[parameter_indice * max_saved_aux_size + j].source = ccv_nnc_tensor_symbol_new(graph, info, 0);
853					saved_aux[parameter_indice * max_saved_aux_size + j].destination = ccv_nnc_tensor_symbol_new(graph, info, 0);
854					const int device_id = CCV_TENSOR_GET_DEVICE_ID(info.type)(((info.type) & 0xfff00) >> 8);
855					for (k = 1; k < parallel_count; k++)
856					{
857						ccv_nnc_tensor_param_t dev_info = info;
858						if (k != device_id)
859							CCV_TENSOR_SET_DEVICE_ID(dev_info.type, k)(dev_info.type) = (((dev_info.type) & ~0xfff00) | (((k) &
 0xfff) << 8));
860						else
861							CCV_TENSOR_SET_DEVICE_ID(dev_info.type, 0)(dev_info.type) = (((dev_info.type) & ~0xfff00) | (((0) &
 0xfff) << 8));
862						const ccv_nnc_tensor_symbol_t src_copy = ccv_nnc_tensor_symbol_new(graph, dev_info, 0);
863						const ccv_nnc_tensor_symbol_t dest_copy = ccv_nnc_tensor_symbol_new(graph, dev_info, 0);
864						ccv_nnc_tensor_symbol_set_copy(graph, saved_aux[parameter_indice * max_saved_aux_size + j].source, k, src_copy);
865						ccv_nnc_tensor_symbol_set_copy(graph, saved_aux[parameter_indice * max_saved_aux_size + j].destination, k, dest_copy);
866					}
867				}
868			} else {
869				for (j = saved_aux_size; j < old_saved_aux_size; j++)
870				{
871					for (k = 1; k < parallel_count; k++)
872					{
873						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);
874						if (src_copy.d >= 0)
875						{
876							ccv_nnc_tensor_symbol_free(graph, src_copy);
877							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
});
878						}
879						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);
880						if (dest_copy.d >= 0)
881						{
882							ccv_nnc_tensor_symbol_free(graph, dest_copy);
883							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
});
884						}
885					}
886					ccv_nnc_tensor_symbol_free(graph, saved_aux[parameter_indice * max_saved_aux_size + j].source);
887					ccv_nnc_tensor_symbol_free(graph, saved_aux[parameter_indice * max_saved_aux_size + j].destination);
888					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
};
889				}
890			}
891		}
892	}
893	_ccv_cnnp_model_graph_exec_symbol_set(graph, compiled_data, parallel_count, update_nodes[parameter_indice], minimizer);
894	if (this_parameter_flag)
895	{
896		ccv_nnc_tensor_symbol_t update_inputs[saved_aux_size + 2];
897		ccv_nnc_tensor_symbol_t update_outputs[saved_aux_size + 1];
898		const int* inputs = 0;
899		int input_size = 0;
900		ccv_nnc_graph_exec_symbol_io(graph, update_nodes[parameter_indice], &inputs, &input_size, 0, 0);
901		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", 901, __extension__ __PRETTY_FUNCTION__)
; }));
902		update_inputs[0].d = inputs[0];
903		update_inputs[0].graph = graph;
904		update_inputs[1].d = inputs[1];
905		update_inputs[1].graph = graph;
906		update_outputs[0] = updated_parameters[parameter_indice];
907		for (j = 0; j < saved_aux_size; j++)
908		{
909			update_inputs[j + 2] = saved_aux[parameter_indice * max_saved_aux_size + j].source;
910			update_outputs[j + 1] = saved_aux[parameter_indice * max_saved_aux_size + j].destination;
911		}
912		ccv_nnc_graph_exec_symbol_set_io(graph, update_nodes[parameter_indice], update_inputs, saved_aux_size + 2, update_outputs, saved_aux_size + 1);
913		for (k = 1; k < parallel_count; k++)
914		{
915			const ccv_nnc_graph_exec_symbol_t copy = ccv_nnc_graph_exec_symbol_copy(graph, update_nodes[parameter_indice], k);
916			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"
, 916, __extension__ __PRETTY_FUNCTION__); }));
917			ccv_nnc_graph_exec_symbol_io(graph, copy, &inputs, &input_size, 0, 0);
918			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", 918, __extension__ __PRETTY_FUNCTION__)
; }));
919			update_inputs[0].d = inputs[0];
920			update_inputs[0].graph = graph;
921			update_inputs[1].d = inputs[1];
922			update_inputs[1].graph = graph;
923			update_outputs[0] = ccv_nnc_tensor_symbol_copy(graph, updated_parameters[parameter_indice], k);
924			for (j = 0; j < saved_aux_size; j++)
925			{
926				update_inputs[j + 2] = ccv_nnc_tensor_symbol_copy(graph, saved_aux[parameter_indice * max_saved_aux_size + j].source, k);
927				update_outputs[j + 1] = ccv_nnc_tensor_symbol_copy(graph, saved_aux[parameter_indice * max_saved_aux_size + j].destination, k);
928			}
929			ccv_nnc_graph_exec_symbol_set_io(graph, copy, update_inputs, saved_aux_size + 2, update_outputs, saved_aux_size + 1);
930		}
931	}
932	return this_parameter_flag;
933}
934 
935typedef struct {
936	int parameter_size;
937	ccv_nnc_cmd_t minimizer;
938	ccv_cnnp_model_io_t parameters[1];
939} ccv_cnnp_set_minimizer_for_parameter_t;
940 
941static int _ccv_cnnp_apply_parameters_with_minimizer(ccv_cnnp_model_t* const model)
942{
943	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
944	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 944, __extension__ __PRETTY_FUNCTION__); }));
945	const int max_saved_aux_size = compiled_data->minimize.max_saved_aux_size;
946	// We update all parameters, at this point, we have one minimizer.
947	const int parameter_size = compiled_data->parameters->rnum;
948	ccv_nnc_graph_exec_symbol_t* const update_nodes = compiled_data->update_nodes;
949	ccv_nnc_symbolic_graph_t* const symbolic_graph = model->graph;
950	assert(symbolic_graph)((void) sizeof ((symbolic_graph) ? 1 : 0), __extension__ ({ if
 (symbolic_graph) ; else __assert_fail ("symbolic_graph", "ccv_cnnp_model.c"
, 950, __extension__ __PRETTY_FUNCTION__); }));
951	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; });
952	ccv_array_t* const parameters = compiled_data->minimize.parameters;
953	ccv_array_t* const parameter_indices = ccv_array_new(sizeof(int), 0, 0);
954	int i, j, flag = 0;
955	for (i = 0; i < parameters->rnum; i++)
956	{
957		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)));
958		for (j = 0; j < set_minimizer_for_parameter->parameter_size; j++)
959		{
960			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;
961			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", 961, __extension__ __PRETTY_FUNCTION__)
; }));
962			const int old_rnum = parameter_indices->rnum;
963			ccv_cnnp_model_add_to_parameter_indices(set_minimizer_for_parameter->parameters[j]->model, param_sel, parameter_indices);
964			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;
965			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", 965, __extension__ __PRETTY_FUNCTION__)
; }));
966			if (param_ref >= 0)
967			{
968				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", 968, __extension__ __PRETTY_FUNCTION__)
; }));
969				*(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)));
970				parameter_indices->rnum = old_rnum + 1;
971			}
972		}
973		const int saved_aux_size = ccv_nnc_minimizer_saved_aux_size(set_minimizer_for_parameter->minimizer);
974		// We may have duplicated indices, but that is OK, we will set it twice.
975		for (j = 0; j < parameter_indices->rnum; j++)
976		{
977			const int d = *(int*)ccv_array_get(parameter_indices, j)((void*)(((char*)((parameter_indices)->data)) + (size_t)(parameter_indices
)->rsize * (size_t)(j)));
978			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", 978, __extension__ __PRETTY_FUNCTION__)
; }));
979			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))
980				flag = 1;
981		}
982		ccv_array_clear(parameter_indices);
983	}
984	ccv_array_free(parameter_indices);
985	return flag;
986}
987 
988static 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)
989{
990	if (new_saved_aux_size == old_saved_aux_size)
991		return;
992	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", 992, __extension__ __PRETTY_FUNCTION__)
; }));
993	int i, j;
994	for (i = parameter_size - 1; i >= 0; i--)
995	{
996		for (j = new_saved_aux_size - 1; j >= old_saved_aux_size; j--)
997			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
};
998		for (j = old_saved_aux_size - 1; j >= 0; j--)
999			saved_aux[i * new_saved_aux_size + j] = saved_aux[i * old_saved_aux_size + j];
1000	}
1001}
1002 
1003static void _ccv_cnnp_model_set_rewindables(ccv_cnnp_model_t* const model)
1004{
1005	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1006	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 1006, __extension__ __PRETTY_FUNCTION__); }));
1007	if (!compiled_data->rewindables)
1008		compiled_data->rewindables = ccv_array_new(sizeof(ccv_cnnp_rewind_symbol_t), 0, 0);
1009	ccv_nnc_tensor_symbol_new_hook(model->graph, _ccv_cnnp_model_tensor_symbol_new_hook, compiled_data->rewindables, 0);
1010	ccv_nnc_tensor_symbol_alias_new_hook(model->graph, _ccv_cnnp_model_tensor_symbol_alias_new_hook, compiled_data->rewindables, 0);
1011	ccv_nnc_graph_exec_symbol_new_hook(model->graph, _ccv_cnnp_model_graph_exec_symbol_new_hook, compiled_data->rewindables, 0);
1012}
1013 
1014static 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)
1015{
1016	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1017	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", 1017, __extension__ __PRETTY_FUNCTION__
); }));
1018	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", 1018, __extension__ __PRETTY_FUNCTION__
); }));
1019	const int evaluate_to_size = compiled_data->evaluate.to_size;
1020	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", 1020, __extension__ __PRETTY_FUNCTION__
); }));
1021	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; });
1022	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);
1023	compiled_data->evaluate.to_ops = (ccv_nnc_graph_exec_t*)(compiled_data->evaluate.tos + evaluate_to_size * parallel_count);
1024	int i, j;
1025	const int output_size = model->output_size;
1026	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", 1026, __extension__ __PRETTY_FUNCTION__
); }));
1027	if (fits)
1028		for (i = 0; i < output_size; i++)
1029			ccv_nnc_tensor_symbol_set(model->graph, compiled_data->fits[i], fits[i]->info);
1030	const int max_saved_aux_size = compiled_data->minimize.max_saved_aux_size;
1031	const int parameter_size = compiled_data->parameters->rnum;
1032	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);
1033	compiled_data->update_nodes = (ccv_nnc_graph_exec_symbol_t*)(compiled_data->updated_parameters + parameter_size);
1034	compiled_data->saved_aux = (ccv_nnc_tensor_symbol_map_t*)(compiled_data->update_nodes + parameter_size);
1035	int parameter_size_maybe_more = parameter_size;
1036	compiled_data->disable_outgrad = disable_outgrad;
1037	int outgrad_size;
1038	if (gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES || model->input_size == 0)
1039		outgrad_size = 0;
1040	else if (disable_outgrad == CCV_CNNP_DISABLE_OUTGRAD_NONE) // Compute minimize with gradients including inputs.
1041		outgrad_size = model->input_size;
1042	else {
1043		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", 1043, __extension__ __PRETTY_FUNCTION__
); })); // If it is disable all, gradient mode won't be this.
1044		outgrad_size = 0;
1045		for (i = 0; i < model->input_size; i++)
1046			if (!(disable_outgrad & ((uint64_t)1 << i)))
1047				++outgrad_size;
1048	}
1049	compiled_data->outgrad_size = outgrad_size;
1050	parameter_size_maybe_more += outgrad_size;
1051	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);
1052	compiled_data->outgrads = parameter_size_maybe_more > parameter_size ? compiled_data->gradients + parameter_size : 0;
1053	compiled_data->backward.tos = (ccv_nnc_graph_exec_symbol_t*)(compiled_data->gradients + parameter_size_maybe_more);
1054	compiled_data->backward.to_size = parameter_size_maybe_more;
1055	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)));
1056	if (compiled_data->parameter_flags)
1057	{
1058		parameters = (ccv_nnc_tensor_symbol_t*)ccmallocmalloc(sizeof(ccv_nnc_tensor_symbol_t) * parameter_size);
1059		for (i = 0; i < parameter_size; i++)
1060			if (compiled_data->parameter_flags[i >> 6] & ((uint64_t)1 << (i & 63)))
1061				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)));
1062			else
1063				parameters[i] = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
1064	}
1065	if (gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES || model->input_size == 0)
1066		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);
1067	else if (disable_outgrad == CCV_CNNP_DISABLE_OUTGRAD_NONE) // Compute minimize with gradients including inputs.
1068		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);
1069	else { // Compute minimize with gradients including selected inputs.
1070		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", 1070, __extension__ __PRETTY_FUNCTION__
); }));
1071		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", 1071, __extension__ __PRETTY_FUNCTION__
); })); // If it is disable all, gradient mode won't be this.
1072		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", 1072, __extension__ __PRETTY_FUNCTION__
); }));
1073		ccv_nnc_tensor_symbol_t outgrads[outgrad_size];
1074		j = 0;
1075		for (i = 0; i < model->input_size; i++)
1076			if (!(disable_outgrad & ((uint64_t)1 << i)))
1077				outgrads[j++] = model->inputs[i];
1078		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);
1079	}
1080	if (compiled_data->parameter_flags)
1081		ccfreefree(parameters);
1082	_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);
1083	if (compiled_data->minimize.parameters)
1084		_ccv_cnnp_apply_parameters_with_minimizer(model);
1085	// Go through gradient checkpoints to generate tensor inputs for backward pass just before executing the backward pass.
1086	ccv_cnnp_model_apply_gradient_checkpoints(compiled_data, model->graph);
1087	for (i = 0; i < output_size; i++)
1088	{
1089		const ccv_nnc_tensor_symbol_t df = ccv_nnc_tensor_symbol_for_backward(model->graph, compiled_data->f[i]);
1090		// Init this to 1 so we can backprop.
1091		ccv_nnc_tensor_symbol_set_flags(model->graph, df, CCV_NNC_TENSOR_SYMBOL_INIT_ONES);
1092	}
1093	compiled_data->backward.to_size = 0;
1094	for (i = 0; i < parameter_size_maybe_more; i++)
1095		if (compiled_data->gradients[i].d != CCV_NNC_NO_TENSOR_SYMBOL)
1096			compiled_data->backward.tos[compiled_data->backward.to_size++] = ccv_nnc_graph_exec_symbol_for_backward(model->graph, compiled_data->gradients[i]);
1097	ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS);
1098	ccv_nnc_symbolic_graph_set_destinations(model->graph, compiled_data->update_nodes, parameter_size);
1099	for (i = 0; i < parameter_size_maybe_more - parameter_size; i++)
1100	{
1101		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.
1102			continue;
1103		const ccv_nnc_graph_exec_symbol_t outgrad = ccv_nnc_graph_exec_symbol_for_backward(model->graph, compiled_data->outgrads[i]);
1104		const int* tos;
1105		int to_size;
1106		ccv_nnc_graph_exec_symbol_to(model->graph, outgrad, &tos, &to_size);
1107		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.
1108		{
1109			const ccv_nnc_graph_exec_symbol_t* destinations = ccv_nnc_symbolic_graph_destinations(model->graph);
1110			const int destination_count = ccv_nnc_symbolic_graph_destination_size(model->graph);
1111			int flag = 0;
1112			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; });
1113			for (j = i - 1; !flag && j >= 0; j--)
1114				if (j + outgrad_destination_start < destination_count)
1115					flag = (destinations[j + outgrad_destination_start].d == outgrad.d);
1116			if (!flag) // Only if we cannot find it, we add it.
1117				ccv_nnc_symbolic_graph_add_destination(model->graph, outgrad);
1118		}
1119	}
1120	if (parallel_count > 1)
1121	{
1122		ccv_nnc_symbolic_graph_data_parallel(model->graph, parallel_count,
1123			0, 0,
1124			compiled_data->gradients, parameter_size /* No need to deal with outgrads, we don't allreduce outgrads */,
1125			compiled_data->gradients /* We only care about gradients before allreduce, thus, update our current pointers */,
1126			0, 0, 0,
1127			CCV_NNC_PARALLEL_REDUCE_OP_SUM,
1128			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));
1129		ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
1130		for (i = 0; i < evaluate_to_size; i++)
1131			for (j = 1; j < parallel_count; j++)
1132			{
1133				const ccv_nnc_graph_exec_symbol_t copy = ccv_nnc_graph_exec_symbol_copy(model->graph, compiled_data->evaluate.tos[i], j);
1134				if (copy.d != CCV_NNC_NO_GRAPH_EXEC_SYMBOL)
1135					compiled_data->evaluate.tos[compiled_data->evaluate.to_size++] = copy;
1136			}
1137		const int backward_to_size = compiled_data->backward.to_size;
1138		for (i = 0; i < backward_to_size; i++)
1139			for (j = 1; j < parallel_count; j++)
1140			{
1141				const ccv_nnc_graph_exec_symbol_t copy = ccv_nnc_graph_exec_symbol_copy(model->graph, compiled_data->backward.tos[i], j);
1142				if (copy.d != CCV_NNC_NO_GRAPH_EXEC_SYMBOL)
1143					compiled_data->backward.tos[compiled_data->backward.to_size++] = copy;
1144			}
1145	}
1146	// Only use memory compression if we are in gradient parameter mode.
1147	if (gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES || gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES_AND_INPUTS)
1148	{
1149		if (model->memory_compression)
1150			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));
1151		if (model->memory_reduction)
1152			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));
1153	}
1154	compiled_data->backward.to_size = _ccv_nnc_array_dedup_graph_exec_symbols(compiled_data->backward.tos, compiled_data->backward.to_size);
1155	compiled_data->gradient_mode = gradient_mode;
1156}
1157 
1158void ccv_cnnp_model_tensors_init_0(const ccv_cnnp_model_t* const model, ccv_cnnp_compiled_data_t* const compiled_data)
1159{
1160	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", 1160, __extension__ __PRETTY_FUNCTION__
); }));
1161	const int parameter_size = compiled_data->parameters->rnum;
1162	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; });
1163	const int internal_size = compiled_data->internals->rnum;
1164	compiled_data->tensors_init.size = ccv_nnc_tensor_symbol_count(model->graph);
1165	compiled_data->tensors_init.v = cccalloccalloc(((compiled_data->tensors_init.size + 31) >> 5), sizeof(uint32_t));
1166	compiled_data->tensors.parameters = (ccv_nnc_tensor_t**)cccalloccalloc((parameter_size + internal_size) * parallel_count, sizeof(ccv_nnc_tensor_t*));
1167	compiled_data->tensors.internals = compiled_data->tensors.parameters + parameter_size * parallel_count;
1168}
1169 
1170int ccv_cnnp_model_tensors_any_to_alloc(const ccv_cnnp_model_t* const model, ccv_cnnp_compiled_data_t* const compiled_data)
1171{
1172	int i, j;
1173	const int parameter_size = compiled_data->parameters->rnum;
1174	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; });
1175	const int internal_size = compiled_data->internals->rnum;
1176	for (i = 0; i < parameter_size; i++)
1177	{
1178		// parameters has to be allocated all together.
1179		if (compiled_data->tensors.parameters[i])
1180		{
1181			for (j = 1; j < parallel_count; j++)
1182				{ 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", 1182, __extension__ __PRETTY_FUNCTION__
); })); }
1183			continue;
1184		}
1185		return 1;
1186	}
1187	for (i = 0; i < internal_size; i++)
1188	{
1189		if (!compiled_data->tensors.internals[i])
1190			return 1;
1191		for (j = 1; j < parallel_count; j++)
1192			if (!compiled_data->tensors.internals[i + j * internal_size])
1193				return 1;
1194	}
1195	return 0;
1196}
1197 
1198void ccv_cnnp_model_tensors_init_1(const ccv_cnnp_model_t* const model, ccv_cnnp_compiled_data_t* const compiled_data)
1199{
1200	int i, j;
1201	const int parameter_size = compiled_data->parameters->rnum;
1202	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; });
1203	const int internal_size = compiled_data->internals->rnum;
1204	for (i = 0; i < parameter_size; i++)
1205	{
1206		// parameters has to be allocated all together.
1207		if (compiled_data->tensors.parameters[i])
1208		{
1209			for (j = 1; j < parallel_count; j++)
1210				{ 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", 1210, __extension__ __PRETTY_FUNCTION__
); })); }
1211			continue;
1212		}
1213		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)));
1214		ccv_nnc_tensor_param_t info = ccv_nnc_tensor_symbol_params(parameter.graph, parameter);
1215		if (CCV_TENSOR_GET_DEVICE(info.type)((info.type) & 0xfff00) == CCV_COMPUTE_DEVICE_ANY)
1216			CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff
) << 8));
1217		const int device_id = CCV_TENSOR_GET_DEVICE_ID(info.type)(((info.type) & 0xfff00) >> 8);
1218		compiled_data->tensors.parameters[i] = ccv_nnc_tensor_new(0, info, 0);
1219		for (j = 1; j < parallel_count; j++)
1220		{
1221			if (j != device_id)
1222				CCV_TENSOR_SET_DEVICE_ID(info.type, j)(info.type) = (((info.type) & ~0xfff00) | (((j) & 0xfff
) << 8));
1223			else
1224				CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff
) << 8));
1225			compiled_data->tensors.parameters[i + j * parameter_size] = ccv_nnc_tensor_new(0, info, 0);
1226		}
1227	}
1228	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));
1229	for (i = 0; i < internal_size; i++)
1230	{
1231		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))
);
1232		const int d = retained.d;
1233		if (init_v[d >> 5] & (1u << (d & 0x1f)))
1234			continue;
1235		ccv_nnc_tensor_param_t info = ccv_nnc_tensor_symbol_params(retained.graph, retained);
1236		if (CCV_TENSOR_GET_DEVICE(info.type)((info.type) & 0xfff00) == CCV_COMPUTE_DEVICE_ANY)
1237			CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff
) << 8));
1238		const int device_id = CCV_TENSOR_GET_DEVICE_ID(info.type)(((info.type) & 0xfff00) >> 8);
1239		if (!compiled_data->tensors.internals[i])
1240			compiled_data->tensors.internals[i] = ccv_nnc_tensor_new(0, info, 0);
1241		for (j = 1; j < parallel_count; j++)
1242		{
1243			if (j != device_id)
1244				CCV_TENSOR_SET_DEVICE_ID(info.type, j)(info.type) = (((info.type) & ~0xfff00) | (((j) & 0xfff
) << 8));
1245			else
1246				CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff
) << 8));
1247			if (!compiled_data->tensors.internals[i + j * internal_size])
1248				compiled_data->tensors.internals[i + j * internal_size] = ccv_nnc_tensor_new(0, info, 0);
1249		}
1250	}
1251	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.
1252}
1253 
1254static void _ccv_cnnp_model_tensors_init(const ccv_cnnp_model_t* const model, ccv_cnnp_compiled_data_t* const compiled_data)
1255{
1256	ccv_cnnp_model_tensors_init_0(model, compiled_data);
1257	ccv_cnnp_model_tensors_init_1(model, compiled_data);
1258}
1259 
1260static 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)
1261{
1262	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", 1262, __extension__ __PRETTY_FUNCTION__
); }));
1263	int i, j;
1264	for (i = 0; i < tensor_size; i++)
1265	{
1266		if (!tensors[i])
1267			continue;
1268		const int d = tensor_symbols[i].d;
1269		if (!(tensors_init[d >> 5] & (1u << (d & 0x1f))))
1270			continue;
1271		for (j = 1; j < parallel_count; j++)
1272			if (tensors[i + j * tensor_size])
1273			{
1274				ccv_nnc_tensor_t* const input = CCV_NNC_TENSOR(tensors[i])((ccv_nnc_tensor_t*)((uintptr_t)(tensors[i]) & ~(uintptr_t
)1));
1275				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));
1276				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);
1277			}
1278	}
1279}
1280 
1281static 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)
1282{
1283	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", 1283, __extension__ __PRETTY_FUNCTION__
); }));
1284	int i, j;
1285	for (i = 0; i < tensor_size; i++)
1286	{
1287		const ccv_nnc_tensor_symbol_t tensor_symbol = tensor_symbols[i];
1288		for (j = 1; j < parallel_count; j++)
1289		{
1290			const ccv_nnc_tensor_symbol_t copy = ccv_nnc_tensor_symbol_copy(graph, tensor_symbol, j);
1291			ccv_nnc_tensor_t* copy_tensor = tensors[i + j * tensor_size];
1292			if (copy_tensor && copy.d == CCV_NNC_NO_TENSOR_SYMBOL)
1293			{ // We shouldn't allocate this, free it up.
1294				ccv_nnc_tensor_free(tensors[i + j * tensor_size]);
1295				tensors[i + j * tensor_size] = 0;
1296			}
1297		}
1298	}
1299}
1300 
1301static 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)
1302{
1303	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", 1303, __extension__ __PRETTY_FUNCTION__
); }));
1304	int i, j;
1305	for (i = 0; i < tensor_size; i++)
1306	{
1307		ccv_nnc_tensor_symbol_t tensor_symbol = tensor_symbols[i];
1308		if (tensor_symbol.d == CCV_NNC_NO_TENSOR_SYMBOL)
1309			continue;
1310		if (graph)
1311		{
1312			const ccv_nnc_tensor_symbol_t alias_to = ccv_nnc_tensor_symbol_alias_to(graph, tensor_symbol);
1313			if (alias_to.d != CCV_NNC_NO_TENSOR_SYMBOL)
1314				tensor_symbol = alias_to;
1315		}
1316		ccv_nnc_tensor_t* const tensor = CCV_NNC_TENSOR(tensors[i])((ccv_nnc_tensor_t*)((uintptr_t)(tensors[i]) & ~(uintptr_t
)1));
1317		if (tensor && tensor_symbol.d != CCV_NNC_NO_TENSOR_SYMBOL)
1318		{
1319			const ccv_nnc_tensor_bind_t retained_bind = {
1320				.symbol = tensor_symbol,
1321				.tensor = tensor
1322			};
1323			ccv_array_push(tensor_binds, &retained_bind);
1324		}
1325		for (j = 1; j < parallel_count; j++)
1326		{
1327			const ccv_nnc_tensor_symbol_t copy = ccv_nnc_tensor_symbol_copy(graph, tensor_symbol, j);
1328			ccv_nnc_tensor_t* copy_tensor = tensors[i + j * tensor_size];
1329			if (copy_tensor && copy.d != CCV_NNC_NO_TENSOR_SYMBOL)
1330			{
1331				const ccv_nnc_tensor_bind_t bind = {
1332					.symbol = copy,
1333					.tensor = tensors[i + j * tensor_size]
1334				};
1335				ccv_array_push(tensor_binds, &bind);
1336			}
1337		}
1338	}
1339}
1340 
1341static void _ccv_cnnp_compiled_data_graph_free(ccv_cnnp_compiled_data_t* const compiled_data)
1342{
1343	if (compiled_data->graph)
1344		ccv_nnc_graph_free(compiled_data->graph);
1345	compiled_data->graph = 0;
1346	compiled_data->is_test = 0;
1347	if (compiled_data->tensor_arena)
1348		ccv_nnc_tensor_arena_free(compiled_data->tensor_arena);
1349	compiled_data->tensor_arena = 0;
1350	if (compiled_data->graph_exec_arena)
1351		ccv_nnc_graph_exec_arena_free(compiled_data->graph_exec_arena);
1352	compiled_data->graph_exec_arena = 0;
1353	if (compiled_data->backward.from_ops)
1354		ccfreefree(compiled_data->backward.from_ops);
1355	compiled_data->backward.from_ops = 0;
1356	if (compiled_data->evaluate.schedule)
1357		ccv_nnc_graph_static_schedule_free(compiled_data->evaluate.schedule);
1358	compiled_data->evaluate.schedule = 0;
1359	if (compiled_data->backward.schedule)
1360		ccv_nnc_graph_static_schedule_free(compiled_data->backward.schedule);
1361	compiled_data->backward.schedule = 0;
1362}
1363 
1364static void _ccv_cnnp_compiled_data_gradient_free(ccv_cnnp_compiled_data_t* const compiled_data)
1365{
1366	if (compiled_data->gradients)
1367		ccfreefree(compiled_data->gradients);
1368	compiled_data->gradients = 0;
1369	if (compiled_data->updated_parameters)
1370		ccfreefree(compiled_data->updated_parameters);
1371	compiled_data->updated_parameters = 0;
1372	compiled_data->update_nodes = 0;
1373	compiled_data->saved_aux = 0;
1374}
1375 
1376static void _ccv_cnnp_compiled_data_backward_free(ccv_cnnp_compiled_data_t* const compiled_data)
1377{
1378	if (compiled_data->backward.gradients)
1379		ccfreefree(compiled_data->backward.gradients);
1380	compiled_data->backward.gradients = 0;
1381	if (compiled_data->backward.accum)
1382		ccv_nnc_graph_free(compiled_data->backward.accum);
1383	compiled_data->backward.accum = 0;
1384	if (compiled_data->backward.tensor_arena)
1385		ccv_nnc_tensor_arena_free(compiled_data->backward.tensor_arena);
1386	compiled_data->backward.tensor_arena = 0;
1387	if (compiled_data->backward.graph_exec_arena)
1388		ccv_nnc_graph_exec_arena_free(compiled_data->backward.graph_exec_arena);
1389	compiled_data->backward.graph_exec_arena = 0;
1390}
1391 
1392static void _ccv_cnnp_compiled_data_apply_gradients_free(ccv_cnnp_compiled_data_t* const compiled_data)
1393{
1394	if (compiled_data->apply_gradients.graph)
1395		ccv_nnc_graph_free(compiled_data->apply_gradients.graph);
1396	compiled_data->apply_gradients.graph = 0;
1397	if (compiled_data->apply_gradients.tensor_arena)
1398		ccv_nnc_tensor_arena_free(compiled_data->apply_gradients.tensor_arena);
1399	compiled_data->apply_gradients.tensor_arena = 0;
1400	if (compiled_data->apply_gradients.graph_exec_arena)
1401		ccv_nnc_graph_exec_arena_free(compiled_data->apply_gradients.graph_exec_arena);
1402	compiled_data->apply_gradients.graph_exec_arena = 0;
1403}
1404 
1405// Compile the graph to run ccv_cnnp_model_fit
1406static 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)
1407{
1408	int i, j;
1409	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1410	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", 1410, __extension__ __PRETTY_FUNCTION__
); }));
1411	compiled_data->graph_mode = CCV_CNNP_MODEL_GRAPH_FIT_MODE;
1412	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; });
1413	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", 1413, __extension__ __PRETTY_FUNCTION__
); }));
1414	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", 1414
, __extension__ __PRETTY_FUNCTION__); }));
1415	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", 1415, __extension__ __PRETTY_FUNCTION__
); }));
1416	if (compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_NONE)
1417	{
1418		_ccv_cnnp_model_set_rewindables(model);
1419		_ccv_cnnp_model_gradient_init(model, CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES, CCV_CNNP_DISABLE_OUTGRAD_ALL, fits, fit_size);
1420	} else if (compiled_data->gradient_mode != CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES) {
1421		_ccv_cnnp_model_rewind_graph(model);
1422		_ccv_cnnp_compiled_data_gradient_free(compiled_data);
1423		compiled_data->gradient_mode = CCV_CNNP_COMPILED_DATA_GRADIENT_NONE;
1424		_ccv_cnnp_model_gradient_init(model, CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES, CCV_CNNP_DISABLE_OUTGRAD_ALL, fits, fit_size);
1425	}
1426	const int tensors_init = !!compiled_data->tensors_init.v;
1427	if (!tensors_init)
1428		_ccv_cnnp_model_tensors_init(model, compiled_data);
1429	else if ((uintptr_t)compiled_data->tensors_init.v & (uintptr_t)1)
1430	// Check if it is not fully allocated, if it is not, init_1.
1431		ccv_cnnp_model_tensors_init_1(model, compiled_data);
1432	ccv_array_t* const tensor_binds = ccv_array_new(sizeof(ccv_nnc_tensor_bind_t), 0, 0);
1433	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"
, 1433, __extension__ __PRETTY_FUNCTION__); }));
1434	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"
, 1434, __extension__ __PRETTY_FUNCTION__); }));
1435	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", 1435
, __extension__ __PRETTY_FUNCTION__); }));
1436	const int input_size_per_p = input_size / parallel_count;
1437	_ccv_cnnp_model_bind_tensors(model->graph, model->inputs, inputs, input_size_per_p, parallel_count, tensor_binds);
1438	const int output_size_per_p = output_size / parallel_count;
1439	_ccv_cnnp_model_bind_tensors(model->graph, model->outputs, outputs, output_size_per_p, parallel_count, tensor_binds);
1440	const int fit_size_per_p = fit_size / parallel_count;
1441	_ccv_cnnp_model_bind_tensors(model->graph, compiled_data->fits, fits, fit_size_per_p, parallel_count, tensor_binds);
1442	const int parameter_size = compiled_data->parameters->rnum;
1443	_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);
1444	_ccv_cnnp_model_bind_tensors(model->graph, compiled_data->updated_parameters, compiled_data->tensors.parameters, parameter_size, parallel_count, tensor_binds);
1445	const int internal_size = compiled_data->internals->rnum;
1446	_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);
1447	_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);
1448	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);
1449	ccv_array_free(tensor_binds);
1450	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));
1451	if (tensors_init && parallel_count > 1)
1452		_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);
1453	// If tensor is not init'ed, we need to init states first.
1454	if (_ccv_cnnp_any_to_init(compiled_data))
1455	{
1456		ccv_nnc_tensor_init_states_t tensor_init_states = {
1457			.parallel_count = parallel_count,
1458			.graph = model->graph,
1459			.compiled_data = compiled_data,
1460			.tensor_arena = compiled_data->tensor_arena
1461		};
1462		ccv_cnnp_model_init_states(model, model->graph, _ccv_cnnp_init_states_for_tensors, &tensor_init_states);
1463	}
1464	compiled_data->is_test = 0;
1465	const int saved_aux_size = ccv_nnc_minimizer_saved_aux_size(compiled_data->minimize.minimizer);
1466	// No need to set because it is default to training mode.
1467	// ccv_cnnp_model_set_is_test(model, 0, _ccv_cnnp_cmd_update_for_execs, &update);
1468	for (i = 0; i < saved_aux_size * parameter_size; i++)
1469	{
1470		if (compiled_data->saved_aux[i].source.d == CCV_NNC_NO_TENSOR_SYMBOL)
1471			continue;
1472		ccv_nnc_tensor_t* const tensor = ccv_nnc_tensor_from_symbol(compiled_data->tensor_arena, compiled_data->saved_aux[i].source);
1473		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);
1474		for (j = 1; j < parallel_count; j++)
1475		{
1476			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));
1477			if (copy)
1478				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);
1479		}
1480	}
1481	const int evaluate_to_size = compiled_data->evaluate.to_size;
1482	compiled_data->evaluate.to_op_size = 0;
1483	for (i = 0; i < evaluate_to_size; i++)
1484	{
1485		ccv_nnc_graph_exec_t const to = ccv_nnc_graph_exec_from_symbol(compiled_data->graph_exec_arena, compiled_data->evaluate.tos[i]);
1486		if (to.graph)
1487			compiled_data->evaluate.to_ops[compiled_data->evaluate.to_op_size++] = to;
1488	}
1489	ccv_nnc_graph_set_default_static_schedule(compiled_data->graph, compiled_data->stream_type, model->max_stream_count);
1490	ccv_nnc_graph_autotune(compiled_data->graph, model->workspace_size, 0, TRAVERSE_FULL0,0,0,0);
1491}
1492 
1493ccv_nnc_stream_context_t* ccv_cnnp_model_default_stream(const ccv_cnnp_model_t* const model)
1494{
1495	const ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1496	if (!compiled_data || !compiled_data->graph)
1497		return 0;
1498	return ccv_nnc_graph_default_stream(compiled_data->graph);
1499}
1500 
1501uint64_t ccv_cnnp_model_memory_size(const ccv_cnnp_model_t* const model)
1502{
1503	const ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1504	if (!compiled_data || !compiled_data->tensor_arena)
1505		return 0;
1506	return ccv_nnc_tensor_arena_size(compiled_data->tensor_arena);
1507}
1508 
1509static 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)
1510{
1511	int i, j;
1512	for (i = 0; i < tensor_size; i++)
1513	{
1514		ccv_nnc_tensor_symbol_t tensor_symbol = tensor_symbols[i];
1515		if (tensor_symbol.d == CCV_NNC_NO_TENSOR_SYMBOL)
1516			continue;
1517		if (graph)
1518		{
1519			const ccv_nnc_tensor_symbol_t alias_to = ccv_nnc_tensor_symbol_alias_to(graph, tensor_symbol);
1520			if (alias_to.d != CCV_NNC_NO_TENSOR_SYMBOL)
1521				tensor_symbol = alias_to;
1522		}
1523		ccv_nnc_tensor_bind_symbol(tensor_arena, tensor_symbol, tensors[i]);
1524		for (j = 1; j < parallel_count; j++)
1525		{
1526			const ccv_nnc_tensor_symbol_t copy = ccv_nnc_tensor_symbol_copy(graph, tensor_symbol, j);
1527			if (copy.d != CCV_NNC_NO_TENSOR_SYMBOL)
1528				ccv_nnc_tensor_bind_symbol(tensor_arena, copy, tensors[i + tensor_size * j]);
1529		}
1530	}
1531}
1532 
1533void 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)
1534{
1535	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1536	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 1536, __extension__ __PRETTY_FUNCTION__); }));
1537	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; });
1538	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", 1538, __extension__ __PRETTY_FUNCTION__
); }));
1539	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", 1539, __extension__ __PRETTY_FUNCTION__
); }));
1540	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", 1540
, __extension__ __PRETTY_FUNCTION__); }));
1541	assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if
 (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c"
, 1541, __extension__ __PRETTY_FUNCTION__); }));
1542	if (!compiled_data->graph || compiled_data->graph_mode != CCV_CNNP_MODEL_GRAPH_FIT_MODE)
1543	{
1544		_ccv_cnnp_compiled_data_graph_free(compiled_data);
1545		_ccv_cnnp_compiled_data_backward_free(compiled_data);
1546		_ccv_cnnp_compiled_data_apply_gradients_free(compiled_data);
1547		// Compile the symbolic graph down only when needed.
1548		_ccv_cnnp_model_fit_jit(model, inputs, input_size, fits, fit_size, outputs, output_size);
1549	} else {
1550		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"
, 1550, __extension__ __PRETTY_FUNCTION__); }));
1551		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"
, 1551, __extension__ __PRETTY_FUNCTION__); }));
1552		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", 1552
, __extension__ __PRETTY_FUNCTION__); }));
1553		const int input_size_per_p = input_size / parallel_count;
1554		_ccv_cnnp_bind_tensors_to_arena(compiled_data->tensor_arena, model->graph, model->inputs, inputs, input_size_per_p, parallel_count);
1555		const int output_size_per_p = output_size / parallel_count;
1556		_ccv_cnnp_bind_tensors_to_arena(compiled_data->tensor_arena, model->graph, model->outputs, outputs, output_size_per_p, parallel_count);
1557		const int fit_size_per_p = fit_size / parallel_count;
1558		_ccv_cnnp_bind_tensors_to_arena(compiled_data->tensor_arena, model->graph, compiled_data->fits, fits, fit_size_per_p, parallel_count);
1559	}
1560	if (compiled_data->is_test)
1561	{
1562		compiled_data->is_test = 0;
1563		ccv_nnc_graph_exec_update_t update = {
1564			.parallel_count = parallel_count,
1565			.graph = model->graph,
1566			.graph_exec_arena = compiled_data->graph_exec_arena,
1567		};
1568		ccv_cnnp_model_set_is_test(model, 0, _ccv_cnnp_cmd_update_for_execs, &update);
1569	}
1570	ccv_nnc_graph_run_with_schedule(compiled_data->graph, 0, 0, tensor_tape, stream_context);
1571}
1572 
1573// Compile the graph to run ccv_cnnp_model_evaluate with require_grad = false (MULTISTAGE_MODE_NO_GRAD).
1574static 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)
1575{
1576	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1577	compiled_data->graph_mode = CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE_NO_GRAD;
1578	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; });
1579	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", 1579, __extension__ __PRETTY_FUNCTION__
); }));
1580	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", 1580, __extension__ __PRETTY_FUNCTION__
); }));
1581	// If the gradient is not initialized, continue to setup parallel process. We don't init gradient here, but rather,
1582	// we setup proper rewindables so the graph can be rewinded to previous state before we run data parallel.
1583	if (parallel_count > 1 && compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_NONE)
1584	{
1585		const int evaluate_to_size = compiled_data->evaluate.to_size;
1586		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);
1587		_ccv_cnnp_model_set_rewindables(model);
1588		ccv_nnc_symbolic_graph_data_parallel(model->graph, parallel_count,
1589			0, 0,
1590			0, 0, 0,
1591			0, 0, 0,
1592			CCV_NNC_PARALLEL_REDUCE_OP_SUM,
1593			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));
1594		ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
1595		int i, j;
1596		for (i = 0; i < evaluate_to_size; i++)
1597			for (j = 1; j < parallel_count; j++)
1598			{
1599				const ccv_nnc_graph_exec_symbol_t copy = ccv_nnc_graph_exec_symbol_copy(model->graph, compiled_data->evaluate.tos[i], j);
1600				if (copy.d != CCV_NNC_NO_GRAPH_EXEC_SYMBOL)
1601					compiled_data->evaluate.tos[compiled_data->evaluate.to_size++] = copy;
1602			}
1603	}
1604	const int tensors_init = !!compiled_data->tensors_init.v;
1605	if (!tensors_init)
1606		_ccv_cnnp_model_tensors_init(model, compiled_data);
1607	else if ((uintptr_t)compiled_data->tensors_init.v & (uintptr_t)1)
1608	// Check if it is not fully allocated, if it is not, init_1.
1609		ccv_cnnp_model_tensors_init_1(model, compiled_data);
1610	ccv_array_t* const tensor_binds = ccv_array_new(sizeof(ccv_nnc_tensor_bind_t), 0, 0);
1611	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"
, 1611, __extension__ __PRETTY_FUNCTION__); }));
1612	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"
, 1612, __extension__ __PRETTY_FUNCTION__); }));
1613	const int input_size_per_p = input_size / parallel_count;
1614	_ccv_cnnp_model_bind_tensors(model->graph, model->inputs, inputs, input_size_per_p, parallel_count, tensor_binds);
1615	const int output_size_per_p = output_size / parallel_count;
1616	_ccv_cnnp_model_bind_tensors(model->graph, model->outputs, outputs, output_size_per_p, parallel_count, tensor_binds);
1617	const int parameter_size = compiled_data->parameters->rnum;
1618	_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);
1619	const int internal_size = compiled_data->internals->rnum;
1620	_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);
1621	_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);
1622	// If we generated gradient for the graph, only compile part of the graph because the rest is irrelevant for evaluation.
1623	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);
1624	ccv_array_free(tensor_binds);
1625	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));
1626	// If tensor is not init'ed, we need to init states first.
1627	if (tensors_init && parallel_count > 1)
1628		_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);
1629	if (_ccv_cnnp_any_to_init(compiled_data))
1630	{
1631		ccv_nnc_tensor_init_states_t tensor_init_states = {
1632			.parallel_count = parallel_count,
1633			.graph = model->graph,
1634			.compiled_data = compiled_data,
1635			.tensor_arena = compiled_data->tensor_arena
1636		};
1637		ccv_cnnp_model_init_states(model, model->graph, _ccv_cnnp_init_states_for_tensors, &tensor_init_states);
1638	}
1639	compiled_data->is_test = 1;
1640	ccv_nnc_graph_exec_update_t update = {
1641		.parallel_count = parallel_count,
1642		.graph = model->graph,
1643		.graph_exec_arena = compiled_data->graph_exec_arena,
1644	};
1645	ccv_cnnp_model_set_is_test(model, 1, _ccv_cnnp_cmd_update_for_execs, &update);
1646	ccv_nnc_graph_set_default_static_schedule(compiled_data->graph, compiled_data->stream_type, model->max_stream_count);
1647	ccv_nnc_graph_autotune(compiled_data->graph, model->workspace_size, 0, TRAVERSE_FULL0,0,0,0);
1648}
1649 
1650static void _ccv_cnnp_model_gradient_tensors_init(const ccv_cnnp_model_t* const model, ccv_cnnp_compiled_data_t* const compiled_data)
1651{
1652	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", 1652, __extension__ __PRETTY_FUNCTION__
); }));
1653	const int parameter_size = compiled_data->parameters->rnum;
1654	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; });
1655	compiled_data->tensors.gradients = (ccv_nnc_tensor_t**)ccmallocmalloc(sizeof(ccv_nnc_tensor_t*) * parameter_size * 2 * parallel_count);
1656	compiled_data->tensors.accum_gradients = compiled_data->tensors.gradients + parameter_size * parallel_count;
1657	int i, j;
1658	for (i = 0; i < parameter_size; i++)
1659	{
1660		if (compiled_data->parameter_flags && !(compiled_data->parameter_flags[i >> 6] & ((uint64_t)1 << (i & 63))))
1661		{
1662			compiled_data->tensors.gradients[i] = 0;
1663			compiled_data->tensors.accum_gradients[i] = 0;
1664			for (j = 1; j < parallel_count; j++)
1665			{
1666				compiled_data->tensors.gradients[i + j * parameter_size] = 0;
1667				compiled_data->tensors.accum_gradients[i + j * parameter_size] = 0;
1668			}
1669			continue;
1670		}
1671		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)));
1672		ccv_nnc_tensor_param_t info = ccv_nnc_tensor_symbol_params(parameter.graph, parameter);
1673		if (CCV_TENSOR_GET_DEVICE(info.type)((info.type) & 0xfff00) == CCV_COMPUTE_DEVICE_ANY)
1674			CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff
) << 8));
1675		const int device_id = CCV_TENSOR_GET_DEVICE_ID(info.type)(((info.type) & 0xfff00) >> 8);
1676		compiled_data->tensors.gradients[i] = ccv_nnc_tensor_new(0, info, 0);
1677		compiled_data->tensors.accum_gradients[i] = 0; // delay the accumulated gradient allocation until when we need it.
1678		for (j = 1; j < parallel_count; j++)
1679		{
1680			if (j != device_id)
1681				CCV_TENSOR_SET_DEVICE_ID(info.type, j)(info.type) = (((info.type) & ~0xfff00) | (((j) & 0xfff
) << 8));
1682			else
1683				CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff
) << 8));
1684			compiled_data->tensors.gradients[i + j * parameter_size] = ccv_nnc_tensor_new(0, info, 0);
1685			compiled_data->tensors.accum_gradients[i + j * parameter_size] = 0;
1686		}
1687	}
1688}
1689 
1690static int _ccv_cnnp_is_disable_outgrad_all(const uint64_t disable_outgrad, const int input_size)
1691{
1692	if (disable_outgrad == CCV_CNNP_DISABLE_OUTGRAD_ALL)
1693		return 1;
1694	if (disable_outgrad == CCV_CNNP_DISABLE_OUTGRAD_NONE)
1695		return 0;
1696	int i;
1697	for (i = 0; i < input_size; i++)
1698		if (!(disable_outgrad & ((uint64_t)1 << i)))
1699			return 0;
1700	return 1;
1701}
1702 
1703// Compile the graph to run ccv_cnnp_model_evaluate with requires_grad = true (MULTISTAGE_MODE).
1704// Particularly, this method compiles the evaluation and backprop graph (the main graph).
1705static 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)
1706{
1707	int i, j;
1708	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1709	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;
1710	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", 1710, __extension__ __PRETTY_FUNCTION__
); }));
1711	compiled_data->graph_mode = CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE;
1712	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; });
1713	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", 1713, __extension__ __PRETTY_FUNCTION__
); }));
1714	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", 1714, __extension__ __PRETTY_FUNCTION__
); }));
1715	// There shouldn't be a loss function if we evaluate with multistage jit.
1716	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", 1716, __extension__ __PRETTY_FUNCTION__
); }));
1717	if (compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_NONE)
1718	{
1719		_ccv_cnnp_model_set_rewindables(model);
1720		_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.
1721	} else if (compiled_data->gradient_mode != target_gradient_mode) {
1722		_ccv_cnnp_model_rewind_graph(model);
1723		_ccv_cnnp_compiled_data_gradient_free(compiled_data);
1724		compiled_data->gradient_mode = CCV_CNNP_COMPILED_DATA_GRADIENT_NONE;
1725		_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.
1726	}
1727	const int tensors_init = !!compiled_data->tensors_init.v;
1728	if (!tensors_init)
1729		_ccv_cnnp_model_tensors_init(model, compiled_data);
1730	else if ((uintptr_t)compiled_data->tensors_init.v & (uintptr_t)1)
1731	// Check if it is not fully allocated, if it is not, init_1.
1732		ccv_cnnp_model_tensors_init_1(model, compiled_data);
1733	ccv_array_t* const tensor_binds = ccv_array_new(sizeof(ccv_nnc_tensor_bind_t), 0, 0);
1734	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"
, 1734, __extension__ __PRETTY_FUNCTION__); }));
1735	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"
, 1735, __extension__ __PRETTY_FUNCTION__); }));
1736	const int input_size_per_p = input_size / parallel_count;
1737	_ccv_cnnp_model_bind_tensors(model->graph, model->inputs, inputs, input_size_per_p, parallel_count, tensor_binds);
1738	const int output_size_per_p = output_size / parallel_count;
1739	_ccv_cnnp_model_bind_tensors(model->graph, model->outputs, outputs, output_size_per_p, parallel_count, tensor_binds);
1740	const int parameter_size = compiled_data->parameters->rnum;
1741	_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);
1742	const int internal_size = compiled_data->internals->rnum;
1743	_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);
1744	_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);
1745	if (!compiled_data->tensors.gradients)
1746		_ccv_cnnp_model_gradient_tensors_init(model, compiled_data);
1747	_ccv_cnnp_model_bind_tensors(model->graph, compiled_data->gradients, compiled_data->tensors.gradients, parameter_size, parallel_count, tensor_binds);
1748	if (compiled_data->backward.to_size > 0)
1749		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);
1750	else
1751		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);
1752	ccv_array_free(tensor_binds);
1753	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));
1754	if (tensors_init && parallel_count > 1)
1755		_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);
1756	// If tensor is not init'ed, we need to init states first.
1757	if (_ccv_cnnp_any_to_init(compiled_data))
1758	{
1759		ccv_nnc_tensor_init_states_t tensor_init_states = {
1760			.parallel_count = parallel_count,
1761			.graph = model->graph,
1762			.compiled_data = compiled_data,
1763			.tensor_arena = compiled_data->tensor_arena
1764		};
1765		ccv_cnnp_model_init_states(model, model->graph, _ccv_cnnp_init_states_for_tensors, &tensor_init_states);
1766	}
1767	compiled_data->is_test = is_test;
1768	ccv_nnc_graph_exec_update_t update = {
1769		.parallel_count = parallel_count,
1770		.graph = model->graph,
1771		.graph_exec_arena = compiled_data->graph_exec_arena,
1772	};
1773	ccv_cnnp_model_set_is_test(model, is_test, _ccv_cnnp_cmd_update_for_execs, &update);
1774	const int evaluate_to_size = compiled_data->evaluate.to_size;
1775	compiled_data->evaluate.to_op_size = 0;
1776	ccv_array_t* const backward_from = ccv_array_new(sizeof(int), 0, 0);
1777	for (i = 0; i < evaluate_to_size; i++)
1778	{
1779		ccv_nnc_graph_exec_t const to_op = ccv_nnc_graph_exec_from_symbol(compiled_data->graph_exec_arena, compiled_data->evaluate.tos[i]);
1780		if (to_op.graph)
1781			compiled_data->evaluate.to_ops[compiled_data->evaluate.to_op_size++] = to_op;
1782		const int* tos;
1783		int to_size;
1784		ccv_nnc_graph_exec_symbol_to(model->graph, compiled_data->evaluate.tos[i], &tos, &to_size);
1785		for (j = 0; j < to_size; j++)
1786		{
1787			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){
1788				.d = tos[j],
1789				.graph = model->graph
1790			});
1791			if (to_op.graph)
1792				ccv_array_add_unique_int(backward_from, to_op.d);
1793		}
1794	}
1795	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", 1795, __extension__
 __PRETTY_FUNCTION__); }));
1796	compiled_data->backward.from_op_size = backward_from->rnum;
1797	compiled_data->backward.from_ops = (ccv_nnc_graph_exec_t*)ccmallocmalloc(sizeof(ccv_nnc_graph_exec_t) * backward_from->rnum);
1798	for (i = 0; i < backward_from->rnum; i++)
1799		compiled_data->backward.from_ops[i] = (ccv_nnc_graph_exec_t){
1800			.d = *(int*)ccv_array_get(backward_from, i)((void*)(((char*)((backward_from)->data)) + (size_t)(backward_from
)->rsize * (size_t)(i))),
1801			.graph = compiled_data->graph,
1802		};
1803	ccv_array_free(backward_from);
1804	ccv_nnc_graph_set_default_static_schedule(compiled_data->graph, compiled_data->stream_type, model->max_stream_count);
1805	ccv_nnc_graph_autotune(compiled_data->graph, model->workspace_size, 0, TRAVERSE_FULL0,0,0,0);
1806}
1807 
1808void 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)
1809{
1810	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1811	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 1811, __extension__ __PRETTY_FUNCTION__); }));
1812	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; });
1813	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", 1813, __extension__ __PRETTY_FUNCTION__
); }));
1814	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", 1814, __extension__ __PRETTY_FUNCTION__
); }));
1815	assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if
 (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c"
, 1815, __extension__ __PRETTY_FUNCTION__); }));
1816	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;
1817	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));
1818	if (!compiled_data->graph || mode_mismatch)
1819	{
1820		_ccv_cnnp_compiled_data_graph_free(compiled_data);
1821		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.
1822			_ccv_cnnp_compiled_data_backward_free(compiled_data);
1823		if (params.requires_grad)
1824			_ccv_cnnp_model_multistage_jit_0(model, params.disable_outgrad, params.is_test, inputs, input_size, outputs, output_size);
1825		else
1826			_ccv_cnnp_model_multistage_no_grad_jit(model, inputs, input_size, outputs, output_size);
1827	} else {
1828		ccv_nnc_tensor_arena_clear_bindings(compiled_data->tensor_arena);
1829		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"
, 1829, __extension__ __PRETTY_FUNCTION__); }));
1830		const int input_size_per_p = input_size / parallel_count;
1831		_ccv_cnnp_bind_tensors_to_arena(compiled_data->tensor_arena, model->graph, model->inputs, inputs, input_size_per_p, parallel_count);
1832		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"
, 1832, __extension__ __PRETTY_FUNCTION__); }));
1833		const int output_size_per_p = output_size / parallel_count;
1834		_ccv_cnnp_bind_tensors_to_arena(compiled_data->tensor_arena, model->graph, model->outputs, outputs, output_size_per_p, parallel_count);
1835	}
1836	if (compiled_data->is_test != params.is_test)
1837	{
1838		compiled_data->is_test = params.is_test;
1839		ccv_nnc_graph_exec_update_t update = {
1840			.parallel_count = parallel_count,
1841			.graph = model->graph,
1842			.graph_exec_arena = compiled_data->graph_exec_arena,
1843		};
1844		ccv_cnnp_model_set_is_test(model, params.is_test, _ccv_cnnp_cmd_update_for_execs, &update);
1845	}
1846}
1847 
1848void 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)
1849{
1850	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1851	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 1851, __extension__ __PRETTY_FUNCTION__); }));
1852	ccv_cnnp_model_dry_run(model, params, inputs, input_size, outputs, output_size);
1853	if (compiled_data->graph_mode == CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE_NO_GRAD)
1854		ccv_nnc_graph_run_with_schedule(compiled_data->graph, 0, 0, tensor_tape, stream_context);
1855	else {
1856		if (!compiled_data->evaluate.schedule)
1857			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);
1858		ccv_nnc_graph_run_with_schedule(compiled_data->graph, 0, compiled_data->evaluate.schedule, tensor_tape, stream_context);
1859	}
1860}
1861 
1862// Compile the graph to run ccv_cnnp_model_backward after ccv_cnnp_model_evaluate with requires_grad = true (MULTISTAGE_MODE).
1863// Particularly, this method compiles the accumulator graph.
1864static void _ccv_cnnp_model_multistage_jit_1(ccv_cnnp_model_t* const model)
1865{
1866	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1867	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 1867, __extension__ __PRETTY_FUNCTION__); }));
1868	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", 1868, __extension__ __PRETTY_FUNCTION__
); }));
1869	ccv_nnc_symbolic_graph_t* accum = ccv_nnc_symbolic_graph_new();
1870	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; });
1871	const int parameter_size = compiled_data->parameters->rnum;
1872	int i, j;
1873	compiled_data->backward.gradients = (ccv_nnc_tensor_symbol_t*)ccmallocmalloc(sizeof(ccv_nnc_tensor_symbol_t) * parameter_size * parallel_count * 3);
1874	compiled_data->backward.accum_gradients = compiled_data->backward.gradients + parameter_size * parallel_count;
1875	compiled_data->backward.updated_accum_gradients = compiled_data->backward.accum_gradients + parameter_size * parallel_count;
1876	for (i = 0; i < parameter_size; i++)
1877		for (j = 0; j < parallel_count; j++)
1878			if (compiled_data->tensors.gradients[i + j * parameter_size])
1879			{
1880				const ccv_nnc_tensor_param_t info = compiled_data->tensors.gradients[i + j * parameter_size]->info;
1881				// Now, the old gradient is the accumulated gradient, getting new gradient tensor setup so we can collect them.
1882				compiled_data->tensors.accum_gradients[i + j * parameter_size] = compiled_data->tensors.gradients[i + j * parameter_size];
1883				compiled_data->tensors.gradients[i + j * parameter_size] = ccv_nnc_tensor_new(0, info, 0);
1884				ccv_nnc_tensor_symbol_t inputs[2];
1885				inputs[0] = compiled_data->backward.accum_gradients[i + j * parameter_size] = ccv_nnc_tensor_symbol_new(accum, info, 0);
1886				inputs[1] = compiled_data->backward.gradients[i + j * parameter_size] = ccv_nnc_tensor_symbol_new(accum, info, 0);
1887				ccv_nnc_tensor_symbol_t output = compiled_data->backward.updated_accum_gradients[i + j * parameter_size] = ccv_nnc_tensor_symbol_new(accum, info, 0);
1888				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);
1889			} else {
1890				compiled_data->backward.accum_gradients[i + j * parameter_size] = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
1891				compiled_data->backward.gradients[i + j * parameter_size] = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
1892				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
};
1893			}
1894	ccv_nnc_graph_exec_symbol_autogen(accum, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
1895	if (ccv_nnc_symbolic_graph_source_size(accum) == 0)
1896	{
1897		ccv_nnc_symbolic_graph_free(accum);
1898		// Create empty graph.
1899		compiled_data->backward.accum = ccv_nnc_graph_new();
1900		ccv_nnc_graph_topsort(compiled_data->backward.accum, 0, 0);
1901		return;
1902	}
1903	ccv_array_t* const tensor_binds = ccv_array_new(sizeof(ccv_nnc_tensor_bind_t), 0, 0);
1904	_ccv_cnnp_model_bind_tensors(accum, compiled_data->backward.accum_gradients, compiled_data->tensors.accum_gradients, parameter_size * parallel_count, 1, tensor_binds);
1905	_ccv_cnnp_model_bind_tensors(accum, compiled_data->backward.gradients, compiled_data->tensors.gradients, parameter_size * parallel_count, 1, tensor_binds);
1906	_ccv_cnnp_model_bind_tensors(accum, compiled_data->backward.updated_accum_gradients, compiled_data->tensors.accum_gradients, parameter_size * parallel_count, 1, tensor_binds);
1907	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);
1908	ccv_nnc_symbolic_graph_free(accum);
1909	ccv_array_free(tensor_binds);
1910	ccv_nnc_graph_set_default_static_schedule(compiled_data->backward.accum, compiled_data->stream_type, model->max_stream_count);
1911}
1912 
1913void 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)
1914{
1915	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1916	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 1916, __extension__ __PRETTY_FUNCTION__); }));
1917	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", 1917, __extension__ __PRETTY_FUNCTION__
); }));
1918	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; });
1919	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", 1919, __extension__ __PRETTY_FUNCTION__
); }));
1920	if (outgrad_size > 0)
1921		{ 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", 1921, __extension__ __PRETTY_FUNCTION__
); })); }
1922	assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if
 (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c"
, 1922, __extension__ __PRETTY_FUNCTION__); }));
1923	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", 1923, __extension__ __PRETTY_FUNCTION__
); }));
1924	const int parameter_size = compiled_data->parameters->rnum;
1925	// If we need to accumulate the gradients now, do jit on accumulator.
1926	if (compiled_data->backward.count > 0)
1927	{
1928		if (!compiled_data->backward.accum)
1929			_ccv_cnnp_model_multistage_jit_1(model);
1930		else if (compiled_data->backward.count == 1) {
1931			//  On this round, we need to switch accumulated gradients with gradients (so we can do accumulation properly).
1932			int i;
1933			for (i = 0; i < parameter_size * parallel_count; i++)
1934			{
1935				ccv_nnc_tensor_t* tensor;
1936				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));
1937			}
1938			if (compiled_data->backward.tensor_arena)
1939			{
1940				ccv_nnc_tensor_arena_clear_bindings(compiled_data->backward.tensor_arena);
1941				// Do rebind in case we messed up the binding (we switch accum_gradients and gradients).
1942				_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);
1943				_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);
1944				_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);
1945			}
1946		}
1947	}
1948	const int ingrad_size_per_p = model->output_size;
1949	const int outgrad_size_per_p = compiled_data->outgrad_size;
1950	int i, j;
1951	for (i = 0; i < ingrad_size_per_p; i++)
1952	{
1953		const ccv_nnc_tensor_symbol_t ingrad = ccv_nnc_tensor_symbol_for_backward(model->graph, compiled_data->f[i]);
1954		if (!ingrad_size || !ingrads || ingrads[i] == 0)
1955		{
1956			// Set it to 1 if it is not specified.
1957			ccv_nnc_tensor_t* const ingrad_tensor = ccv_nnc_tensor_from_symbol(compiled_data->tensor_arena, ingrad);
1958			if (ingrad_tensor)
1959				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);
1960			for (j = 1; j < parallel_count; j++)
1961			{
1962				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));
1963				if (ingrad_tensor)
1964					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);
1965			}
1966		} else {
1967			// Make sure the length matches, in case it is an alias.
1968			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", 1968, __extension__ __PRETTY_FUNCTION__
); }));
1969			ccv_nnc_tensor_bind_symbol(compiled_data->tensor_arena, ingrad, ingrads[i]);
1970			for (j = 1; j < parallel_count; j++)
1971				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]);
1972		}
1973	}
1974	if (outgrad_size > 0)
1975	{
1976		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", 1976, __extension__ __PRETTY_FUNCTION__
); }));
1977		for (i = 0; i < outgrad_size_per_p; i++)
1978			if (outgrads[i])
1979			{
1980				const ccv_nnc_tensor_symbol_t outgrad = compiled_data->outgrads[i];
1981				ccv_nnc_tensor_bind_symbol(compiled_data->tensor_arena, outgrad, outgrads[i]);
1982				for (j = 1; j < parallel_count; j++)
1983					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]);
1984			}
1985	} else {
1986		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", 1987, __extension__ __PRETTY_FUNCTION__
); }))
1987			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", 1987, __extension__ __PRETTY_FUNCTION__
); }));
1988	}
1989	// We need to rebind here because in ccv_cnnp_evaluate, we clear bindings, that will reset all bindings for the gradients.
1990	// For parameters and internals these are fine because when we clear bindings, it restores to original bindings, which are these
1991	// parameters and internals. The same cannot be said for gradients due to the accum_gradients switching.
1992	_ccv_cnnp_bind_tensors_to_arena(compiled_data->tensor_arena, model->graph, compiled_data->gradients, compiled_data->tensors.gradients, parameter_size, parallel_count);
1993	if (!compiled_data->backward.schedule)
1994		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);
1995	// Run the backward pass.
1996	ccv_nnc_graph_run_with_schedule(compiled_data->graph, 0, compiled_data->backward.schedule, tensor_tape, stream_context);
1997	// If we need to run accumulation round, do that now.
1998	if (compiled_data->backward.count > 0)
1999		ccv_nnc_graph_run_with_schedule(compiled_data->backward.accum, 0, 0, 0, stream_context);
2000	// Update the count, this determines whether we need to accumulate or not.
2001	++compiled_data->backward.count;
2002}
2003 
2004// Compile the graph to run ccv_cnnp_model_apply_gradients after ccv_cnnp_model_backward (MULTISTAGE_MODE).
2005// Particularly, this method compiles the parameter update graph.
2006static void _ccv_cnnp_model_multistage_jit_2(ccv_cnnp_model_t* const model)
2007{
2008	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2009	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", 2009, __extension__ __PRETTY_FUNCTION__
); }));
2010	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; });
2011	const int parameter_size = compiled_data->parameters->rnum;
2012	ccv_array_t* const tensor_binds = ccv_array_new(sizeof(ccv_nnc_tensor_bind_t), 0, 0);
2013	_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);
2014	_ccv_cnnp_model_bind_tensors(model->graph, compiled_data->updated_parameters, compiled_data->tensors.parameters, parameter_size, parallel_count, tensor_binds);
2015	// Bind accumulated gradients.
2016	if (compiled_data->backward.count > 1)
2017		_ccv_cnnp_model_bind_tensors(model->graph, compiled_data->gradients, compiled_data->tensors.accum_gradients, parameter_size, parallel_count, tensor_binds);
2018	else
2019		_ccv_cnnp_model_bind_tensors(model->graph, compiled_data->gradients, compiled_data->tensors.gradients, parameter_size, parallel_count, tensor_binds);
2020	ccv_array_t* const apply_gradients_from = ccv_array_new(sizeof(int), 0, 0);
2021	int i, j;
2022	for (i = 0; i < compiled_data->backward.to_size; i++)
2023	{
2024		const int* tos;
2025		int to_size;
2026		ccv_nnc_graph_exec_symbol_to(model->graph, compiled_data->backward.tos[i], &tos, &to_size);
2027		for (j = 0; j < to_size; j++)
2028		{
2029			// Check if this is already show up in the backward graph, if that is the case, it won't be in the apply
2030			// gradients graph.
2031			const ccv_nnc_graph_exec_t exec = ccv_nnc_graph_exec_from_symbol(compiled_data->graph_exec_arena, (ccv_nnc_graph_exec_symbol_t){
2032				.d = tos[j],
2033				.graph = model->graph,
2034			});
2035			if (!exec.graph)
2036				ccv_array_add_unique_int(apply_gradients_from, tos[j]);
2037		}
2038	}
2039	const int from_size = apply_gradients_from->rnum;
2040	if (from_size == 0)
2041	{
2042		ccv_array_free(apply_gradients_from);
2043		ccv_array_free(tensor_binds);
2044		return;
2045	}
2046	ccv_nnc_graph_exec_symbol_t* const froms = (ccv_nnc_graph_exec_symbol_t*)ccmallocmalloc(sizeof(ccv_nnc_graph_exec_symbol_t) * from_size);
2047	for (i = 0; i < from_size; i++)
2048		froms[i] = (ccv_nnc_graph_exec_symbol_t){
2049			.d = *(int*)ccv_array_get(apply_gradients_from, i)((void*)(((char*)((apply_gradients_from)->data)) + (size_t
)(apply_gradients_from)->rsize * (size_t)(i))),
2050			.graph = model->graph
2051		};
2052	ccv_array_free(apply_gradients_from);
2053	// It can only ends with updates on the parameters.
2054	ccv_array_t* const tos = ccv_array_new(sizeof(ccv_nnc_graph_exec_symbol_t), parameter_size * parallel_count, 0);
2055	for (i = 0;  i < parameter_size; i++)
2056	{
2057		if (compiled_data->update_nodes[i].d == CCV_NNC_NO_TENSOR_SYMBOL)
2058			continue;
2059		ccv_array_push(tos, &compiled_data->update_nodes[i]);
2060		for (j = 1; j < parallel_count; j++)
2061		{
2062			const ccv_nnc_graph_exec_symbol_t copy = ccv_nnc_graph_exec_symbol_copy(model->graph, compiled_data->update_nodes[i], j);
2063			ccv_array_push(tos, &copy);
2064		}
2065	}
2066	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);
2067	ccv_array_free(tos);
2068	ccv_array_free(tensor_binds);
2069	ccfreefree(froms);
2070	const int max_saved_aux_size = compiled_data->minimize.max_saved_aux_size;
2071	for (i = 0; i < max_saved_aux_size * parameter_size; i++)
2072	{
2073		// Skip on no tensor.
2074		if (compiled_data->saved_aux[i].source.d == CCV_NNC_NO_TENSOR_SYMBOL)
2075			continue;
2076		ccv_nnc_tensor_t* const tensor = ccv_nnc_tensor_from_symbol(compiled_data->apply_gradients.tensor_arena, compiled_data->saved_aux[i].source);
2077		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);
2078		for (j = 1; j < parallel_count; j++)
2079		{
2080			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));
2081			if (copy)
2082				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);
2083		}
2084	}
2085	ccv_nnc_graph_set_default_static_schedule(compiled_data->apply_gradients.graph, compiled_data->stream_type, model->max_stream_count);
2086}
2087 
2088void ccv_cnnp_model_apply_gradients(ccv_cnnp_model_t* const model, ccv_nnc_stream_context_t* const stream_context)
2089{
2090	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2091	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 2091, __extension__ __PRETTY_FUNCTION__); }));
2092	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", 2092, __extension__ __PRETTY_FUNCTION__
); }));
2093	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; });
2094	assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if
 (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c"
, 2094, __extension__ __PRETTY_FUNCTION__); }));
2095	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", 2095, __extension__ __PRETTY_FUNCTION__
); }));
2096	// Skip if there is no backward pass.
2097	if (compiled_data->backward.count <= 0)
2098		return;
2099	// Skip if there is no parameters.
2100	if (compiled_data->parameters->rnum == 0)
2101	{
2102		compiled_data->backward.count = 0;
2103		return;
2104	}
2105	if (!compiled_data->apply_gradients.graph)
2106		_ccv_cnnp_model_multistage_jit_2(model);
2107	else {
2108		const int parameter_size = compiled_data->parameters->rnum;
2109		ccv_nnc_tensor_arena_clear_bindings(compiled_data->apply_gradients.tensor_arena);
2110		// Change to bind accum_gradients if we do gradient accumulation (run backward more than once).
2111		if (compiled_data->backward.count > 1)
2112			_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);
2113		else
2114			_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);
2115	}
2116	if (compiled_data->apply_gradients.graph)
2117		ccv_nnc_graph_run_with_schedule(compiled_data->apply_gradients.graph, 0, 0, 0, stream_context);
2118	// Reset backward count to 0.
2119	compiled_data->backward.count = 0;
2120}
2121 
2122void 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)
2123{
2124	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2125	const int param_sel = parameter->param_sel > 0 ? parameter->param_sel - 1 : parameter->param_sel;
2126	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", 2126, __extension__ __PRETTY_FUNCTION__
); }));
2127	const int tensors_init = !!compiled_data->tensors_init.v;
2128	if (!tensors_init)
2129		_ccv_cnnp_model_tensors_init(model, compiled_data);
2130	else if ((uintptr_t)compiled_data->tensors_init.v & (uintptr_t)1)
2131	// Check if it is not fully allocated, if it is not, init_1.
2132		ccv_cnnp_model_tensors_init_1(model, compiled_data);
2133	ccv_array_t* const parameter_indices = ccv_array_new(sizeof(int), 0, 0);
2134	ccv_cnnp_model_add_to_parameter_indices(parameter->model, param_sel, parameter_indices);
2135	const int param_ref = parameter->param_ref > 0 ? parameter->param_ref - 1 : parameter->param_ref;
2136	if (param_ref < 0)
2137		{ 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", 2137
, __extension__ __PRETTY_FUNCTION__); })); }
2138	else
2139		{ 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", 2139, __extension__ __PRETTY_FUNCTION__
); })); }
2140	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)));
2141	ccv_array_free(parameter_indices);
2142	const int parameter_size = compiled_data->parameters->rnum;
2143	assert(d >= 0)((void) sizeof ((d >= 0) ? 1 : 0), __extension__ ({ if (d >=
 0) ; else __assert_fail ("d >= 0", "ccv_cnnp_model.c", 2143
, __extension__ __PRETTY_FUNCTION__); }));
2144	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", 2144, __extension__ __PRETTY_FUNCTION__
); }));
2145	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; });
2146	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));
2147	assert(dest)((void) sizeof ((dest) ? 1 : 0), __extension__ ({ if (dest) ;
 else __assert_fail ("dest", "ccv_cnnp_model.c", 2147, __extension__
 __PRETTY_FUNCTION__); }));
2148	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);
2149	int i;
2150	for (i = 1; i < parallel_count; i++)
2151	{
2152		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));
2153		if (copy_tensor)
2154			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);
2155	}
2156	// Mark this symbol as init'ed.
2157	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;
2158  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));
2159	init_v[s >> 5] |= (1u << (s & 0x1f));
2160}
2161 
2162void ccv_cnnp_model_parameter_copy(ccv_cnnp_model_t* const model, const ccv_cnnp_model_io_t parameter, ccv_nnc_tensor_t* const tensor)
2163{
2164	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2165	const int param_sel = parameter->param_sel > 0 ? parameter->param_sel - 1 : parameter->param_sel;
2166	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", 2166, __extension__ __PRETTY_FUNCTION__
); }));
2167	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", 2167, __extension__ __PRETTY_FUNCTION__
); }));
2168	ccv_array_t* const parameter_indices = ccv_array_new(sizeof(int), 0, 0);
2169	ccv_cnnp_model_add_to_parameter_indices(parameter->model, param_sel, parameter_indices);
2170	const int param_ref = parameter->param_ref > 0 ? parameter->param_ref - 1 : parameter->param_ref;
2171	if (param_ref < 0)
2172		{ 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", 2172
, __extension__ __PRETTY_FUNCTION__); })); }
2173	else
2174		{ 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", 2174, __extension__ __PRETTY_FUNCTION__
); })); }
2175	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)));
2176	ccv_array_free(parameter_indices);
2177	const int parameter_size = compiled_data->parameters->rnum;
2178	assert(d >= 0)((void) sizeof ((d >= 0) ? 1 : 0), __extension__ ({ if (d >=
 0) ; else __assert_fail ("d >= 0", "ccv_cnnp_model.c", 2178
, __extension__ __PRETTY_FUNCTION__); }));
2179	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", 2179, __extension__ __PRETTY_FUNCTION__
); }));
2180	// We don't need to consider parallel_count, every parameter on each device is identical.
2181	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));
2182	assert(src)((void) sizeof ((src) ? 1 : 0), __extension__ ({ if (src) ; else
 __assert_fail ("src", "ccv_cnnp_model.c", 2182, __extension__
 __PRETTY_FUNCTION__); }));
2183	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);
2184}
2185 
2186ccv_nnc_tensor_param_t ccv_cnnp_model_parameter_tensor_params(ccv_cnnp_model_t* const model, const ccv_cnnp_model_io_t parameter)
2187{
2188	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2189	const int param_sel = parameter->param_sel > 0 ? parameter->param_sel - 1 : parameter->param_sel;
2190	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", 2190, __extension__ __PRETTY_FUNCTION__
); }));
2191	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", 2191, __extension__ __PRETTY_FUNCTION__
); }));
2192	ccv_array_t* const parameter_indices = ccv_array_new(sizeof(int), 0, 0);
2193	ccv_cnnp_model_add_to_parameter_indices(parameter->model, param_sel, parameter_indices);
2194	const int param_ref = parameter->param_ref > 0 ? parameter->param_ref - 1 : parameter->param_ref;
2195	if (param_ref < 0)
2196		{ 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", 2196
, __extension__ __PRETTY_FUNCTION__); })); }
2197	else
2198		{ 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", 2198, __extension__ __PRETTY_FUNCTION__
); })); }
2199	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)));
2200	ccv_array_free(parameter_indices);
2201	const int parameter_size = compiled_data->parameters->rnum;
2202	assert(d >= 0)((void) sizeof ((d >= 0) ? 1 : 0), __extension__ ({ if (d >=
 0) ; else __assert_fail ("d >= 0", "ccv_cnnp_model.c", 2202
, __extension__ __PRETTY_FUNCTION__); }));
2203	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", 2203, __extension__ __PRETTY_FUNCTION__
); }));
2204	// We don't need to consider parallel_count, every parameter on each device is identical.
2205	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));
2206	assert(tensor)((void) sizeof ((tensor) ? 1 : 0), __extension__ ({ if (tensor
) ; else __assert_fail ("tensor", "ccv_cnnp_model.c", 2206, __extension__
 __PRETTY_FUNCTION__); }));
2207	return tensor->info;
2208}
2209 
2210const char* ccv_cnnp_model_parameter_name(ccv_cnnp_model_t* const model, const ccv_cnnp_model_io_t parameter)
2211{
2212	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2213	const int param_sel = parameter->param_sel > 0 ? parameter->param_sel - 1 : parameter->param_sel;
2214	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", 2214, __extension__ __PRETTY_FUNCTION__
); }));
2215	ccv_array_t* const parameter_indices = ccv_array_new(sizeof(int), 0, 0);
2216	ccv_cnnp_model_add_to_parameter_indices(parameter->model, param_sel, parameter_indices);
2217	const int param_ref = parameter->param_ref > 0 ? parameter->param_ref - 1 : parameter->param_ref;
2218	if (param_ref < 0)
2219		{ 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", 2219
, __extension__ __PRETTY_FUNCTION__); })); }
2220	else
2221		{ 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", 2221, __extension__ __PRETTY_FUNCTION__
); })); }
2222	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)));
2223	ccv_array_free(parameter_indices);
2224	const int parameter_size = compiled_data->parameters->rnum;
2225	assert(d >= 0)((void) sizeof ((d >= 0) ? 1 : 0), __extension__ ({ if (d >=
 0) ; else __assert_fail ("d >= 0", "ccv_cnnp_model.c", 2225
, __extension__ __PRETTY_FUNCTION__); }));
2226	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", 2226, __extension__ __PRETTY_FUNCTION__
); }));
2227	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)));
2228}
2229 
2230int ccv_cnnp_model_parameter_count(ccv_cnnp_model_t* const model)
2231{
2232	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", 2232, __extension__ __PRETTY_FUNCTION__
); }));
2233	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2234	return compiled_data->parameters->rnum;
2235}
2236 
2237ccv_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)
2238{
2239	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2240	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 2240, __extension__ __PRETTY_FUNCTION__); }));
2241	const int parameter_size = compiled_data->parameters->rnum;
2242	int i;
2243	for (i = 0; i < parameter_size; i++)
2244	{
2245		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)));
2246		if (first(model, name, context))
2247			return ccv_cnnp_model_parameters(model, -1, i);
2248	}
2249	return 0;
2250}
2251 
2252ccv_array_t* ccv_cnnp_model_parameters_filter(ccv_cnnp_model_t* const model, ccv_cnnp_model_parameters_filter_f filter, void* const context)
2253{
2254	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2255	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 2255, __extension__ __PRETTY_FUNCTION__); }));
2256	ccv_array_t* const parameters = ccv_array_new(sizeof(ccv_cnnp_model_io_t), 0, 0);
2257	const int parameter_size = compiled_data->parameters->rnum;
2258	int i;
2259	for (i = 0; i < parameter_size; i++)
2260	{
2261		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)));
2262		if (filter(model, name, context))
2263		{
2264			ccv_cnnp_model_io_t parameter = ccv_cnnp_model_parameters(model, -1, i);
2265			ccv_array_push(parameters, &parameter);
2266		}
2267	}
2268	return parameters;
2269 
2270}
2271 
2272CCV_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)
2273{
2274	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2275	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 2275, __extension__ __PRETTY_FUNCTION__); }));
2276	const int tensors_init = !!compiled_data->tensors_init.v;
2277	if (!tensors_init) // If nothing initialized, we return parameter 0.
2278		return ccv_cnnp_model_parameters(model, -1, 0);
2279	const int parameter_size = compiled_data->parameters->rnum;
2280	int i;
2281	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));
2282	for (i = 0; i < parameter_size; i++)
2283	{
2284		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;
2285		if (!(init_v[d >> 5] & (1u << (d & 0x1f))))
2286			return ccv_cnnp_model_parameters(model, -1, i);
2287	}
2288	return 0;
2289}
2290 
2291static 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)
2292{
2293	const int to_param_sel = parameters->param_sel > 0 ? parameters->param_sel - 1 : parameters->param_sel;
2294	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", 2294, __extension__
 __PRETTY_FUNCTION__); }));
2295	ccv_array_t* const to_parameter_indices = ccv_array_new(sizeof(int), 0, 0);
2296	ccv_cnnp_model_add_to_parameter_indices(parameters->model, to_param_sel, to_parameter_indices);
2297	*param_ref = parameters->param_ref > 0 ? parameters->param_ref - 1 : parameters->param_ref;
2298	return to_parameter_indices;
2299}
2300 
2301static 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)
2302{
2303	// If the model is not compiled yet. Compile them now.
2304	if (!model->graph)
2305	{
2306		model->graph = ccv_nnc_symbolic_graph_new();
2307		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", 2307, __extension__ __PRETTY_FUNCTION__
); }));
2308		const int input_size = from_model->input_size;
2309		ccv_nnc_tensor_param_t input_params[input_size];
2310		int i;
2311		for (i = 0; i < input_size; i++)
2312			input_params[i] = ccv_nnc_tensor_symbol_params(from_model->graph, from_model->inputs[i]);
2313		_ccv_cnnp_model_compile(model, input_params, input_size, from_model->compiled_data->loss);
2314		model->parallel_count = from_model->parallel_count;
2315		model->memory_compression = from_model->memory_compression;
2316		model->memory_reduction = from_model->memory_reduction;
2317		model->gradient_checkpointing = from_model->gradient_checkpointing;
2318		model->compiled_data->stream_type = from_model->compiled_data->stream_type;
2319		model->compiled_data->minimize.minimizer = from_model->compiled_data->minimize.minimizer;
2320		model->compiled_data->minimize.max_saved_aux_size = from_model->compiled_data->minimize.max_saved_aux_size;
2321	}
2322	ccv_cnnp_compiled_data_t* const to_compiled_data = model->compiled_data;
2323	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", 2323, __extension__ __PRETTY_FUNCTION__
); }));
2324	const int to_tensors_init = !!to_compiled_data->tensors_init.v;
2325	if (!to_tensors_init)
2326	{
2327		if (only_init_0)
2328			ccv_cnnp_model_tensors_init_0(model, to_compiled_data);
2329		else
2330			_ccv_cnnp_model_tensors_init(model, to_compiled_data);
2331	} else if (!only_init_0 && (uintptr_t)to_compiled_data->tensors_init.v & (uintptr_t)1)
2332		// Check if it is not fully allocated, if it is not, init_1.
2333			ccv_cnnp_model_tensors_init_1(model, to_compiled_data);
2334	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", 2334, __extension__ __PRETTY_FUNCTION__
); }));
2335	*parameter_indices = _ccv_cnnp_model_parameter_indices(model, parameters, param_ref);
2336	*from_parameter_indices = _ccv_cnnp_model_parameter_indices(from_model, from_parameters, from_param_ref);
2337	if (*from_param_ref < 0 && *param_ref >= 0)
2338		{ 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", 2338, __extension__ __PRETTY_FUNCTION__
); })); }
2339	else if (*from_param_ref >= 0)
2340		{ 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", 2340, __extension__ __PRETTY_FUNCTION__
); })); }
2341	if (*param_ref < 0 && *from_param_ref >= 0)
2342		{ 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"
, 2342, __extension__ __PRETTY_FUNCTION__); })); }
2343	else if (*param_ref >= 0)
2344		{ 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", 2344, __extension__ __PRETTY_FUNCTION__
); })); }
2345}
2346 
2347void 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)
2348{
2349	ccv_array_t* to_parameter_indices;
2350	int to_param_ref;
2351	ccv_array_t* from_parameter_indices;
2352	int from_param_ref;
2353	_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);
2354	// Should be exactly the same tensor.
2355	if (to_param_ref < 0 && from_param_ref < 0)
2356		{ 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", 2356, __extension__ __PRETTY_FUNCTION__
); })); }
2357	// To models.
2358	ccv_cnnp_compiled_data_t* const to_compiled_data = model->compiled_data;
2359	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", 2359, __extension__ __PRETTY_FUNCTION__
); }));
2360	// From models.
2361	const ccv_cnnp_compiled_data_t* const from_compiled_data = from_model->compiled_data;
2362	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; });
2363	const int to_parameter_size = to_compiled_data->parameters->rnum;
2364	const int rnum = (to_param_ref < 0 && from_param_ref < 0) ? from_parameter_indices->rnum : 1;
2365	int i, j;
2366	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));
2367  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));
2368	for (i = 0; i < rnum; i++)
2369	{
2370		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)));
2371		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"
, 2371, __extension__ __PRETTY_FUNCTION__); }));
2372		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", 2372, __extension__ __PRETTY_FUNCTION__
); }));
2373		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;
2374		// If the original is not init'ed. We cannot copy from.
2375		if (!(from_init_v[s >> 5] & (1u << (s & 0x1f))))
2376			continue;
2377		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)));
2378		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"
, 2378, __extension__ __PRETTY_FUNCTION__); }));
2379		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", 2379, __extension__ __PRETTY_FUNCTION__
); }));
2380		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));
2381		assert(src)((void) sizeof ((src) ? 1 : 0), __extension__ ({ if (src) ; else
 __assert_fail ("src", "ccv_cnnp_model.c", 2381, __extension__
 __PRETTY_FUNCTION__); }));
2382		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));
2383		assert(dest)((void) sizeof ((dest) ? 1 : 0), __extension__ ({ if (dest) ;
 else __assert_fail ("dest", "ccv_cnnp_model.c", 2383, __extension__
 __PRETTY_FUNCTION__); }));
2384		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);
2385		for (j = 1; j < parallel_count; j++)
2386		{
2387			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));
2388			if (copy_tensor)
2389				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);
2390		}
2391		// Mark this symbol as init'ed.
2392		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;
2393		to_init_v[d >> 5] |= (1u << (d & 0x1f));
2394	}
2395	ccv_array_free(to_parameter_indices);
2396	ccv_array_free(from_parameter_indices);
2397}
2398 
2399KHASH_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; } }
25
←
Null pointer value stored to field 'vals'→
2400 
2401void 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)
2402{
2403	ccv_array_t* to_parameter_indices;
2404	int to_param_ref;
2405	ccv_array_t* from_parameter_indices;
2406	int from_param_ref;
2407	_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);
2408	// Should be exactly the same tensor.
2409	if (renamer == 0 && to_param_ref < 0 && from_param_ref < 0)
1
Assuming 'renamer' is not equal to null→
2410		{ 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", 2410, __extension__ __PRETTY_FUNCTION__
); })); }
2411	// To models.
2412	ccv_cnnp_compiled_data_t* const to_compiled_data = model->compiled_data;
2413	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", 2413, __extension__ __PRETTY_FUNCTION__
); }));
2
←
Assuming 'to_compiled_data' is non-null→
3
←
Taking true branch→
2414	// From models.
2415	const ccv_cnnp_compiled_data_t* const from_compiled_data = from_model->compiled_data;
2416	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→
2417	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", 2417, __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→
2418	const int from_parameter_size = from_compiled_data->parameters->rnum;
2419	const int to_parameter_size = to_compiled_data->parameters->rnum;
2420	const int rnum = (to_param_ref < 0 && from_param_ref < 0) ? to_parameter_indices->rnum : 1;
9
←
Assuming 'to_param_ref' is >= 0→
2421	int i, j;
2422	khash_t(ccv_cnnp_parameter_id)kh_ccv_cnnp_parameter_id_t* id_map = 0;
2423	char* updated_name = 0;
2424	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));
2425  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));
2426	for (i = 0; i < rnum; i++)
2427	{
2428		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 true→
12
←
Assuming the condition is false→
13
←
'?' condition is false→
2429		// Need to figure out how to use the renamer here.
2430		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→
2431		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"
, 2431, __extension__ __PRETTY_FUNCTION__); }));
15
←
Assuming 'dest_d' is >= 0→
16
←
Taking true branch→
2432		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", 2432, __extension__
 __PRETTY_FUNCTION__); }));
17
←
Assuming 'dest_d' is < 'to_parameter_size'→
18
←
Taking true branch→
2433		if (renamer18.1
'renamer' is non-null
)
2434		{
2435			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;
2436			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)));
2437			if (!updated_name18.3
'updated_name' is null
)
19
←
Taking true branch→
2438				updated_name = (char*)ccmallocmalloc(1024);
2439			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→
2440			if (src_name_len20.1
'src_name_len' is <= 0
 > 0)
21
←
Taking false branch→
2441				memcpy(updated_name, src_name, src_name_len);
2442			updated_name[src_name_len] = 0;
2443			if (renamer(context, dest_name, updated_name, 1024) != 0)
22
←
Assuming the condition is false→
2444				continue; // Skip this.
2445			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)
2446			{
2447				// Nothing changed.
2448			} else {
2449				if (!id_map22.2
'id_map' is null
)
23
←
Taking true branch→
2450				{
2451					id_map = kh_init(ccv_cnnp_parameter_id)kh_init_ccv_cnnp_parameter_id();
24
←
Calling 'kh_init_ccv_cnnp_parameter_id'→
26
←
Returning from 'kh_init_ccv_cnnp_parameter_id'→
2452					for (j = 0; j < from_parameter_size; j++)
27
←
Assuming 'j' is >= 'from_parameter_size'→
28
←
Loop condition is false. Execution continues on line 2460→
2453					{
2454						int ret;
2455						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);
2456						assert(ret != 0)((void) sizeof ((ret != 0) ? 1 : 0), __extension__ ({ if (ret
 != 0) ; else __assert_fail ("ret != 0", "ccv_cnnp_model.c", 2456
, __extension__ __PRETTY_FUNCTION__); }));
2457						kh_val(id_map, k)((id_map)->vals[k]) = j;
2458					}
2459				}
2460				const khiter_t k = kh_get(ccv_cnnp_parameter_id, id_map, updated_name)kh_get_ccv_cnnp_parameter_id(id_map, updated_name);
2461				if (k == kh_end(id_map)((id_map)->n_buckets)) // Cannot find the name, skip.
29
←
Assuming 'k' is not equal to field 'n_buckets'→
30
←
Taking false branch→
2462					continue;
2463				src_d = kh_val(id_map, k)((id_map)->vals[k]);
31
←
Array access (via field 'vals') results in a null pointer dereference
2464				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"
, 2464, __extension__ __PRETTY_FUNCTION__); }));
2465				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", 2465, __extension__
 __PRETTY_FUNCTION__); }));
2466			}
2467		}
2468		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"
, 2468, __extension__ __PRETTY_FUNCTION__); }));
2469		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", 2469, __extension__
 __PRETTY_FUNCTION__); }));
2470		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;
2471		// If the original is not init'ed. We cannot share from.
2472		if (!(from_init_v[s >> 5] & (1u << (s & 0x1f))))
2473			continue;
2474		for (j = 0; j < parallel_count; j++)
2475		{
2476			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));
2477			assert(src)((void) sizeof ((src) ? 1 : 0), __extension__ ({ if (src) ; else
 __assert_fail ("src", "ccv_cnnp_model.c", 2477, __extension__
 __PRETTY_FUNCTION__); }));
2478			ccv_nnc_tensor_t* const dest = to_compiled_data->tensors.parameters[dest_d + j * to_parameter_size];
2479			if (dest && !((uintptr_t)dest & (uintptr_t)1))
2480				ccv_nnc_tensor_free(dest);
2481			to_compiled_data->tensors.parameters[dest_d + j * to_parameter_size] = (ccv_nnc_tensor_t*)((uintptr_t)src | (uintptr_t)1);
2482		}
2483		// Mark this symbol as init'ed.
2484		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;
2485		to_init_v[d >> 5] |= (1u << (d & 0x1f));
2486	}
2487	ccv_array_free(to_parameter_indices);
2488	ccv_array_free(from_parameter_indices);
2489	if (id_map)
2490		kh_destroy(ccv_cnnp_parameter_id, id_map)kh_destroy_ccv_cnnp_parameter_id(id_map);
2491	if (updated_name)
2492		ccfreefree(updated_name);
2493	// Mark it as incomplete so we will call init_1.
2494	if (ccv_cnnp_model_tensors_any_to_alloc(model, to_compiled_data))
2495		to_compiled_data->tensors_init.v = (uint32_t*)((uintptr_t)to_compiled_data->tensors_init.v | (uintptr_t)1);
2496	else // Remove the flag.
2497		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));
2498}
2499 
2500ccv_nnc_stream_context_t* ccv_cnnp_compiled_data_get_stream(ccv_cnnp_compiled_data_t* const compiled_data, const int type)
2501{
2502	if (!compiled_data->stream_map)
2503		compiled_data->stream_map = kh_init(stream_map)kh_init_stream_map();
2504	int ret = 0;
2505	khiter_t k = kh_put(stream_map, compiled_data->stream_map, type, &ret)kh_put_stream_map(compiled_data->stream_map, type, &ret
);
2506	assert(ret >= 0)((void) sizeof ((ret >= 0) ? 1 : 0), __extension__ ({ if (
ret >= 0) ; else __assert_fail ("ret >= 0", "ccv_cnnp_model.c"
, 2506, __extension__ __PRETTY_FUNCTION__); }));
2507	ccv_nnc_stream_context_t* stream = kh_val(compiled_data->stream_map, k)((compiled_data->stream_map)->vals[k]);
2508	// If ret == 0, the key already exist, we can return directly, otherwise, create and return.
2509	if (ret != 0)
2510	{
2511		stream = ccv_nnc_stream_context_new(type);
2512		kh_val(compiled_data->stream_map, k)((compiled_data->stream_map)->vals[k]) = stream;
2513	}
2514	return stream;
2515}
2516 
2517void 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)
2518{
2519	ccv_array_t* to_parameter_indices;
2520	int to_param_ref;
2521	ccv_array_t* from_parameter_indices;
2522	int from_param_ref;
2523	_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);
2524	// Should be exactly the same tensor.
2525	if (to_param_ref < 0 && from_param_ref < 0)
2526		{ 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", 2526, __extension__ __PRETTY_FUNCTION__
); })); }
2527	// To models.
2528	ccv_cnnp_compiled_data_t* const to_compiled_data = model->compiled_data;
2529	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", 2529, __extension__ __PRETTY_FUNCTION__
); }));
2530	// From models.
2531	const ccv_cnnp_compiled_data_t* const from_compiled_data = from_model->compiled_data;
2532	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; });
2533	const int to_parameter_size = to_compiled_data->parameters->rnum;
2534	const int rnum = (to_param_ref < 0 && from_param_ref < 0) ? from_parameter_indices->rnum : 1;
2535	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", 2535, __extension__ __PRETTY_FUNCTION__
); }));
2536	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", 2536, __extension__ __PRETTY_FUNCTION__
); }));
2537	int i, j;
2538	ccv_nnc_tensor_t* inputs[aux_in_size + 2];
2539	ccv_nnc_tensor_t* outputs[aux_out_size + 1];
2540	for (i = 0; i < aux_in_size; i++)
2541		inputs[i + 2] = aux_ins[i];
2542	for (i = 0; i < aux_out_size; i++)
2543		outputs[i + 1] = aux_outs[i];
2544	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));
2545  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));
2546	for (i = 0; i < rnum; i++)
2547	{
2548		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)));
2549		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"
, 2549, __extension__ __PRETTY_FUNCTION__); }));
2550		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", 2550, __extension__ __PRETTY_FUNCTION__
); }));
2551		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;
2552		// If the original is not init'ed. We cannot copy from.
2553		if (!(from_init_v[s >> 5] & (1u << (s & 0x1f))))
2554			continue;
2555		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)));
2556		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"
, 2556, __extension__ __PRETTY_FUNCTION__); }));
2557		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", 2557, __extension__ __PRETTY_FUNCTION__
); }));
2558		if (parallel_count > 1)
2559		{
2560			ccv_nnc_stream_context_t* streams[parallel_count];
2561			ccv_nnc_stream_signal_t* signal;
2562			if (stream_context)
2563				signal = ccv_nnc_stream_context_emit_signal_new(stream_context);
2564			for (j = 0; j < parallel_count; j++)
2565			{
2566				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));
2567				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));
2568				if (!dest || !src)
2569				{
2570					streams[j] = 0;
2571					continue;
2572				}
2573				// At the moment, can only handle them on the same device.
2574				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", 2574, __extension__ __PRETTY_FUNCTION__
); }));
2575				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", 2575, __extension__ __PRETTY_FUNCTION__
); }));
2576				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;
2577				const int device_id = CCV_TENSOR_GET_DEVICE_ID(src->info.type)(((src->info.type) & 0xfff00) >> 8);
2578				int type = stream_type;
2579				CCV_STREAM_SET_DEVICE_ID(type, device_id)(type) = (((type) & ~0xfff00) | (((device_id) & 0xfff
) << 8));
2580				ccv_nnc_stream_context_t* const stream_0 = ccv_cnnp_compiled_data_get_stream(to_compiled_data, type);
2581				// Wait signal to finish.
2582				if (stream_context)
2583					ccv_nnc_stream_context_wait_signal(stream_0, signal);
2584				inputs[0] = outputs[0] = dest;
2585				inputs[1] = src;
2586				ccv_nnc_cmd_exec(cmd, hint, flags, inputs, aux_in_size + 2, outputs, aux_out_size + 1, stream_0);
2587				if (stream_context)
2588				{
2589					ccv_nnc_stream_signal_t* const signal = ccv_nnc_stream_context_emit_signal_new(stream_0);
2590					ccv_nnc_stream_context_wait_signal(stream_context, signal);
2591				}
2592				streams[j] = stream_0;
2593			}
2594			// If this should be blocking, blocking it.
2595			if (!stream_context)
2596				for (j = 0; j < parallel_count; j++)
2597					if (streams[j])
2598						ccv_nnc_stream_context_wait(streams[j]);
2599		} else {
2600			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));
2601			assert(src)((void) sizeof ((src) ? 1 : 0), __extension__ ({ if (src) ; else
 __assert_fail ("src", "ccv_cnnp_model.c", 2601, __extension__
 __PRETTY_FUNCTION__); }));
2602			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));
2603			assert(dest)((void) sizeof ((dest) ? 1 : 0), __extension__ ({ if (dest) ;
 else __assert_fail ("dest", "ccv_cnnp_model.c", 2603, __extension__
 __PRETTY_FUNCTION__); }));
2604			inputs[0] = outputs[0] = dest;
2605			inputs[1] = src;
2606			ccv_nnc_cmd_exec(cmd, hint, flags, inputs, aux_in_size + 2, outputs, aux_out_size + 1, stream_context);
2607		}
2608		// Mark this symbol as init'ed.
2609		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;
2610		to_init_v[d >> 5] |= (1u << (d & 0x1f));
2611	}
2612	ccv_array_free(to_parameter_indices);
2613	ccv_array_free(from_parameter_indices);
2614}
2615 
2616void 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)
2617{
2618	int to_param_ref;
2619	ccv_array_t* const to_parameter_indices = _ccv_cnnp_model_parameter_indices(model, parameters, &to_param_ref);
2620	// To models.
2621	ccv_cnnp_compiled_data_t* const to_compiled_data = model->compiled_data;
2622	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", 2622, __extension__ __PRETTY_FUNCTION__
); }));
2623	// Tensor has to be inited already.
2624	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", 2624, __extension__ __PRETTY_FUNCTION__
); }));
2625	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", 2625, __extension__ __PRETTY_FUNCTION__
); }));
2626	// From models.
2627	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; });
2628	const int to_parameter_size = to_compiled_data->parameters->rnum;
2629	const int rnum = (to_param_ref < 0) ? to_parameter_indices->rnum : 1;
2630	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", 2630, __extension__ __PRETTY_FUNCTION__
); }));
2631	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", 2631, __extension__ __PRETTY_FUNCTION__
); }));
2632	int i, j;
2633	ccv_nnc_tensor_t* inputs[aux_in_size + 1];
2634	ccv_nnc_tensor_t* outputs[aux_out_size + 1];
2635	for (i = 0; i < aux_in_size; i++)
2636		inputs[i + 1] = aux_ins[i];
2637	for (i = 0; i < aux_out_size; i++)
2638		outputs[i + 1] = aux_outs[i];
2639	for (i = 0; i < rnum; i++)
2640	{
2641		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)));
2642		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"
, 2642, __extension__ __PRETTY_FUNCTION__); }));
2643		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", 2643, __extension__ __PRETTY_FUNCTION__
); }));
2644		if (parallel_count > 1)
2645		{
2646			ccv_nnc_stream_context_t* streams[parallel_count];
2647			ccv_nnc_stream_signal_t* signal;
2648			if (stream_context)
2649				signal = ccv_nnc_stream_context_emit_signal_new(stream_context);
2650			for (j = 0; j < parallel_count; j++)
2651			{
2652				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));
2653				if (!dest)
2654				{
2655					streams[j] = 0;
2656					continue;
2657				}
2658				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;
2659				const int device_id = CCV_TENSOR_GET_DEVICE_ID(dest->info.type)(((dest->info.type) & 0xfff00) >> 8);
2660				int type = stream_type;
2661				CCV_STREAM_SET_DEVICE_ID(type, device_id)(type) = (((type) & ~0xfff00) | (((device_id) & 0xfff
) << 8));
2662				ccv_nnc_stream_context_t* const stream_0 = ccv_cnnp_compiled_data_get_stream(to_compiled_data, type);
2663				// Wait signal to finish.
2664				if (stream_context)
2665					ccv_nnc_stream_context_wait_signal(stream_0, signal);
2666				inputs[0] = outputs[0] = dest;
2667				ccv_nnc_cmd_exec(cmd, hint, flags, inputs, aux_in_size + 1, outputs, aux_out_size + 1, stream_0);
2668				if (stream_context)
2669				{
2670					ccv_nnc_stream_signal_t* const signal = ccv_nnc_stream_context_emit_signal_new(stream_0);
2671					ccv_nnc_stream_context_wait_signal(stream_context, signal);
2672				}
2673				streams[j] = stream_0;
2674			}
2675			// If this should be blocking, blocking it.
2676			if (!stream_context)
2677				for (j = 0; j < parallel_count; j++)
2678					if (streams[j])
2679						ccv_nnc_stream_context_wait(streams[j]);
2680		} else {
2681			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));
2682			assert(dest)((void) sizeof ((dest) ? 1 : 0), __extension__ ({ if (dest) ;
 else __assert_fail ("dest", "ccv_cnnp_model.c", 2682, __extension__
 __PRETTY_FUNCTION__); }));
2683			inputs[0] = outputs[0] = dest;
2684			ccv_nnc_cmd_exec(cmd, hint, flags, inputs, aux_in_size + 1, outputs, aux_out_size + 1, stream_context);
2685		}
2686		// No need to mark this symbol as init'ed, it is already.
2687	}
2688	ccv_array_free(to_parameter_indices);
2689}
2690 
2691void 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)
2692{
2693	int to_param_ref;
2694	ccv_array_t* const to_parameter_indices = _ccv_cnnp_model_parameter_indices(model, parameters, &to_param_ref);
2695	// To models.
2696	ccv_cnnp_compiled_data_t* const to_compiled_data = model->compiled_data;
2697	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", 2697, __extension__ __PRETTY_FUNCTION__
); }));
2698	// Tensor has to be inited already.
2699	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", 2699, __extension__ __PRETTY_FUNCTION__
); }));
2700	ccv_nnc_tensor_t** tensor_gradients;
2701	if (to_compiled_data->backward.count > 1)
2702		tensor_gradients = to_compiled_data->tensors.accum_gradients;
2703	else
2704		tensor_gradients = to_compiled_data->tensors.gradients;
2705	assert(tensor_gradients)((void) sizeof ((tensor_gradients) ? 1 : 0), __extension__ ({
 if (tensor_gradients) ; else __assert_fail ("tensor_gradients"
, "ccv_cnnp_model.c", 2705, __extension__ __PRETTY_FUNCTION__
); }));
2706	// From models.
2707	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; });
2708	const int to_parameter_size = to_compiled_data->parameters->rnum;
2709	const int rnum = (to_param_ref < 0) ? to_parameter_indices->rnum : 1;
2710	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", 2710, __extension__ __PRETTY_FUNCTION__
); }));
2711	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", 2711, __extension__ __PRETTY_FUNCTION__
); }));
2712	int i, j;
2713	ccv_nnc_tensor_t* inputs[aux_in_size + 1];
2714	ccv_nnc_tensor_t* outputs[aux_out_size + 1];
2715	for (i = 0; i < aux_in_size; i++)
2716		inputs[i + 1] = aux_ins[i];
2717	for (i = 0; i < aux_out_size; i++)
2718		outputs[i + 1] = aux_outs[i];
2719	for (i = 0; i < rnum; i++)
2720	{
2721		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)));
2722		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"
, 2722, __extension__ __PRETTY_FUNCTION__); }));
2723		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", 2723, __extension__ __PRETTY_FUNCTION__
); }));
2724		if (parallel_count > 1)
2725		{
2726			ccv_nnc_stream_context_t* streams[parallel_count];
2727			ccv_nnc_stream_signal_t* signal;
2728			if (stream_context)
2729				signal = ccv_nnc_stream_context_emit_signal_new(stream_context);
2730			for (j = 0; j < parallel_count; j++)
2731			{
2732				ccv_nnc_tensor_t* const dest = tensor_gradients[dest_d + j * to_parameter_size];
2733				if (!dest)
2734				{
2735					streams[j] = 0;
2736					continue;
2737				}
2738				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;
2739				const int device_id = CCV_TENSOR_GET_DEVICE_ID(dest->info.type)(((dest->info.type) & 0xfff00) >> 8);
2740				int type = stream_type;
2741				CCV_STREAM_SET_DEVICE_ID(type, device_id)(type) = (((type) & ~0xfff00) | (((device_id) & 0xfff
) << 8));
2742				ccv_nnc_stream_context_t* const stream_0 = ccv_cnnp_compiled_data_get_stream(to_compiled_data, type);
2743				// Wait signal to finish.
2744				if (stream_context)
2745					ccv_nnc_stream_context_wait_signal(stream_0, signal);
2746				inputs[0] = outputs[0] = dest;
2747				ccv_nnc_cmd_exec(cmd, hint, flags, inputs, aux_in_size + 1, outputs, aux_out_size + 1, stream_0);
2748				if (stream_context)
2749				{
2750					ccv_nnc_stream_signal_t* const signal = ccv_nnc_stream_context_emit_signal_new(stream_0);
2751					ccv_nnc_stream_context_wait_signal(stream_context, signal);
2752				}
2753				streams[j] = stream_0;
2754			}
2755			// If this should be blocking, blocking it.
2756			if (!stream_context)
2757				for (j = 0; j < parallel_count; j++)
2758					if (streams[j])
2759						ccv_nnc_stream_context_wait(streams[j]);
2760		} else {
2761			ccv_nnc_tensor_t* const dest = tensor_gradients[dest_d];
2762			if (!dest)
2763				continue;
2764			assert(dest)((void) sizeof ((dest) ? 1 : 0), __extension__ ({ if (dest) ;
 else __assert_fail ("dest", "ccv_cnnp_model.c", 2764, __extension__
 __PRETTY_FUNCTION__); }));
2765			inputs[0] = outputs[0] = dest;
2766			ccv_nnc_cmd_exec(cmd, hint, flags, inputs, aux_in_size + 1, outputs, aux_out_size + 1, stream_context);
2767		}
2768		// No need to mark this symbol as init'ed, it is already.
2769	}
2770	ccv_array_free(to_parameter_indices);
2771}
2772 
2773ccv_nnc_cmd_t ccv_cnnp_model_minimizer(ccv_cnnp_model_t* const model)
2774{
2775	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2776	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 2776, __extension__ __PRETTY_FUNCTION__); }));
2777	return compiled_data->minimize.minimizer;
2778}
2779 
2780void 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)
2781{
2782	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2783	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 2783, __extension__ __PRETTY_FUNCTION__); }));
2784	const int parameter_size = compiled_data->parameters->rnum;
2785	if (parameter_size == 0)
2786		return;
2787	if (reset)
2788		{ 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", 2788, __extension__ __PRETTY_FUNCTION__
); })); }
2789	const int old_max_saved_aux_size = compiled_data->minimize.max_saved_aux_size;
2790	const int saved_aux_size = ccv_nnc_minimizer_saved_aux_size(minimizer);
2791	if (saved_aux_size > compiled_data->minimize.max_saved_aux_size)
2792		compiled_data->minimize.max_saved_aux_size = saved_aux_size;
2793	const int max_saved_aux_size = compiled_data->minimize.max_saved_aux_size;
2794	// We update all parameters, at this point, we have one minimizer.
2795	if (set_parameters == 0 || set_parameter_size == 0)
2796		compiled_data->minimize.minimizer = minimizer;
2797	int i;
2798	if (set_parameters && set_parameter_size)
2799	{
2800		// I need to save what's the minimizer along with this.
2801		if (!compiled_data->minimize.parameters)
2802			compiled_data->minimize.parameters = ccv_array_new(sizeof(ccv_cnnp_set_minimizer_for_parameter_t*), 1, 0);
2803		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));
2804		set_minimizer_for_parameter->minimizer = minimizer;
2805		set_minimizer_for_parameter->parameter_size = set_parameter_size;
2806		memcpy(set_minimizer_for_parameter->parameters, set_parameters, sizeof(ccv_cnnp_model_io_t) * set_parameter_size);
2807		ccv_array_push(compiled_data->minimize.parameters, &set_minimizer_for_parameter);
2808	}
2809	// If reset is true, clear the parameters array.
2810	if (reset && compiled_data->minimize.parameters)
2811	{
2812		for (i = 0; i < compiled_data->minimize.parameters->rnum; i++)
2813			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))));
2814		ccv_array_clear(compiled_data->minimize.parameters);
2815	}
2816	if (!compiled_data->update_nodes)
2817		return;
2818	ccv_nnc_symbolic_graph_t* const symbolic_graph = model->graph;
2819	assert(symbolic_graph)((void) sizeof ((symbolic_graph) ? 1 : 0), __extension__ ({ if
 (symbolic_graph) ; else __assert_fail ("symbolic_graph", "ccv_cnnp_model.c"
, 2819, __extension__ __PRETTY_FUNCTION__); }));
2820	if (saved_aux_size > old_max_saved_aux_size)
2821	{
2822		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", 2822, __extension__ __PRETTY_FUNCTION__
); }));
2823		// Reallocate first, move them around later.
2824		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);
2825		compiled_data->update_nodes = (ccv_nnc_graph_exec_symbol_t*)(compiled_data->updated_parameters + parameter_size);
2826		compiled_data->saved_aux = (ccv_nnc_tensor_symbol_map_t*)(compiled_data->update_nodes + parameter_size);
2827		// We need to do this from back to front because saved_aux_size > old_saved_aux_size, it could overlap.
2828		_ccv_cnnp_scatter_saved_aux(compiled_data->saved_aux, parameter_size, old_max_saved_aux_size, saved_aux_size);
2829	}
2830	int flag = 0;
2831	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; });
2832	if (set_parameters && set_parameter_size)
2833	{
2834		ccv_array_t* const parameter_indices = ccv_array_new(sizeof(int), 0, 0);
2835		for (i = 0; i < set_parameter_size; i++)
2836		{
2837			const int param_sel = set_parameters[i]->param_sel > 0 ? set_parameters[i]->param_sel - 1 : set_parameters[i]->param_sel;
2838			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", 2838, __extension__ __PRETTY_FUNCTION__
); }));
2839			const int old_rnum = parameter_indices->rnum;
2840			ccv_cnnp_model_add_to_parameter_indices(set_parameters[i]->model, param_sel, parameter_indices);
2841			const int param_ref = set_parameters[i]->param_ref > 0 ? set_parameters[i]->param_ref - 1 : set_parameters[i]->param_ref;
2842			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", 2842, __extension__ __PRETTY_FUNCTION__
); }));
2843			if (param_ref >= 0)
2844			{
2845				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", 2845, __extension__ __PRETTY_FUNCTION__
); }));
2846				*(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)));
2847				parameter_indices->rnum = old_rnum + 1;
2848			}
2849		}
2850		// We may have duplicated indices, but that is OK, we will set it twice.
2851		for (i = 0; i < parameter_indices->rnum; i++)
2852		{
2853			const int d = *(int*)ccv_array_get(parameter_indices, i)((void*)(((char*)((parameter_indices)->data)) + (size_t)(parameter_indices
)->rsize * (size_t)(i)));
2854			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))
2855				flag = 1;
2856		}
2857		ccv_array_free(parameter_indices);
2858	} else {
2859		for (i = 0; i < parameter_size; i++)
2860			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))
2861				flag = 1;
2862		if (compiled_data->minimize.parameters)
2863			if (_ccv_cnnp_apply_parameters_with_minimizer(model))
2864				flag = 1;
2865	}
2866	if (flag)
2867	{
2868		// 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.
2869		if (compiled_data->graph_mode == CCV_CNNP_MODEL_GRAPH_FIT_MODE)
2870			_ccv_cnnp_compiled_data_graph_free(compiled_data);
2871		_ccv_cnnp_compiled_data_apply_gradients_free(compiled_data);
2872	}
2873}
2874 
2875void ccv_cnnp_model_set_compile_params(ccv_cnnp_model_t* const model, const ccv_nnc_symbolic_graph_compile_param_t compile_params)
2876{
2877	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2878	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 2878, __extension__ __PRETTY_FUNCTION__); }));
2879	compiled_data->compile_params = compile_params;
2880}
2881 
2882void ccv_cnnp_model_dot(const ccv_cnnp_model_t* const model, const int flags, FILE** const outs, const int out_size)
2883{
2884	if (model->graph && out_size > 0)
2885		ccv_nnc_symbolic_graph_dot(model->graph, flags, outs[0]);
2886	if (model->compiled_data && model->compiled_data->graph && out_size > 1)
2887		ccv_nnc_graph_dot(model->compiled_data->graph, flags, outs[1]);
2888	if (model->compiled_data && model->compiled_data->backward.accum && out_size > 2)
2889		ccv_nnc_graph_dot(model->compiled_data->backward.accum, flags, outs[2]);
2890	if (model->compiled_data && model->compiled_data->apply_gradients.graph && out_size > 3)
2891		ccv_nnc_graph_dot(model->compiled_data->apply_gradients.graph, flags, outs[3]);
2892}
2893 
2894void ccv_cnnp_model_format(const ccv_cnnp_model_t* const model, const ccv_nnc_symbolic_graph_format_f format_fn, void* const context)
2895{
2896	if (model->graph)
2897		ccv_nnc_symbolic_graph_format(model->graph, 0, 0, 0, 0, format_fn, context);
2898}
2899 
2900static void _ccv_cnnp_compiled_data_free(const ccv_cnnp_model_t* const model, ccv_cnnp_compiled_data_t* const compiled_data)
2901{
2902	int i;
2903	const int parameter_size = compiled_data->parameters->rnum;
2904	ccv_array_free(compiled_data->parameters);
2905	if (compiled_data->parameter_flags)
2906		ccfreefree(compiled_data->parameter_flags);
2907	const int internal_size = compiled_data->internals->rnum;
2908	ccv_array_free(compiled_data->internals);
2909	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", 2909, __extension__ __PRETTY_FUNCTION__
); }));
2910	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", 2910, __extension__ __PRETTY_FUNCTION__
); }));
2911	for (i = 0; i < parameter_size; i++)
2912		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))));
2913	ccv_array_free(compiled_data->ids.parameters);
2914	for (i = 0; i < internal_size; i++)
2915		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))));
2916	ccv_array_free(compiled_data->ids.internals);
2917	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; });
2918	if (compiled_data->tensors.parameters)
2919	{
2920		for (i = 0; i < parameter_size * parallel_count; i++)
2921			// If it is not marked as not belonging, we can free it.
2922			if (!((uintptr_t)compiled_data->tensors.parameters[i] & (uintptr_t)1))
2923				if (compiled_data->tensors.parameters[i])
2924					ccv_nnc_tensor_free(compiled_data->tensors.parameters[i]);
2925		for (i = 0; i < internal_size * parallel_count; i++)
2926			if (compiled_data->tensors.internals[i])
2927				ccv_nnc_tensor_free(compiled_data->tensors.internals[i]);
2928		ccfreefree(compiled_data->tensors.parameters);
2929	}
2930	if (compiled_data->tensors.gradients)
2931	{
2932		for (i = 0; i < parameter_size * parallel_count; i++)
2933		{
2934			if (compiled_data->tensors.gradients[i])
2935				ccv_nnc_tensor_free(compiled_data->tensors.gradients[i]);
2936			if (compiled_data->tensors.accum_gradients[i])
2937				ccv_nnc_tensor_free(compiled_data->tensors.accum_gradients[i]);
2938		}
2939		ccfreefree(compiled_data->tensors.gradients);
2940	}
2941	if (compiled_data->minimize.parameters)
2942	{
2943		for (i = 0; i < compiled_data->minimize.parameters->rnum; i++)
2944			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))));
2945		ccv_array_free(compiled_data->minimize.parameters);
2946	}
2947	if (compiled_data->rewindables)
2948		ccv_array_free(compiled_data->rewindables);
2949	if (compiled_data->tensors_init.v)
2950		ccfreefree(CCV_NNC_INIT_V(compiled_data->tensors_init.v)((uint32_t*)((uintptr_t)(compiled_data->tensors_init.v) &
 ~(uintptr_t)1)));
2951	if (compiled_data->evaluate.tos)
2952		ccfreefree(compiled_data->evaluate.tos);
2953	compiled_data->evaluate.tos = 0;
2954	if (compiled_data->stream_map)
2955	{
2956		khiter_t k;
2957		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)
2958		{
2959			if (!kh_exist(compiled_data->stream_map, k)(!(((compiled_data->stream_map)->flags[(k)>>4]>>
(((k)&0xfU)<<1))&3)))
2960				continue;
2961			ccv_nnc_stream_context_t* const stream = kh_val(compiled_data->stream_map, k)((compiled_data->stream_map)->vals[k]);
2962			ccv_nnc_stream_context_free(stream);
2963		}
2964		kh_destroy(stream_map, compiled_data->stream_map)kh_destroy_stream_map(compiled_data->stream_map);
2965	}
2966	_ccv_cnnp_compiled_data_graph_free(compiled_data);
2967	_ccv_cnnp_compiled_data_gradient_free(compiled_data);
2968	_ccv_cnnp_compiled_data_backward_free(compiled_data);
2969	_ccv_cnnp_compiled_data_apply_gradients_free(compiled_data);
2970	if (compiled_data->gradient_checkpoints)
2971	{
2972		for (i = 0; i < compiled_data->gradient_checkpoints->rnum; i++)
2973		{
2974			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)));
2975			assert(checkpoint->inputs)((void) sizeof ((checkpoint->inputs) ? 1 : 0), __extension__
 ({ if (checkpoint->inputs) ; else __assert_fail ("checkpoint->inputs"
, "ccv_cnnp_model.c", 2975, __extension__ __PRETTY_FUNCTION__
); }));
2976			ccfreefree(checkpoint->inputs);
2977			ccv_array_free(checkpoint->tensor_symbols);
2978		}
2979		ccv_array_free(compiled_data->gradient_checkpoints);
2980	}
2981	ccv_nnc_xpu_alloc_destroy(&compiled_data->xpu_alloc);
2982	ccfreefree(compiled_data);
2983}
2984 
2985void ccv_cnnp_model_free(ccv_cnnp_model_t* const model)
2986{
2987	if (model->isa->deinit)
2988		model->isa->deinit(model);
2989	if (model->io)
2990	{
2991		int i;
2992		for (i = 0; i < model->io->rnum; i++)
2993		{
2994			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)));
2995			if (model_io->outgoings)
2996				ccv_array_free(model_io->outgoings);
2997			if (model_io->incomings)
2998				ccv_array_free(model_io->incomings);
2999			if (model_io->dependencies)
3000				ccv_array_free(model_io->dependencies);
3001			ccfreefree(model_io);
3002		}
3003		ccv_array_free(model->io);
3004	}
3005	if (model->parameter_indices)
3006		ccv_array_free(model->parameter_indices);
3007	if (model->inputs)
3008		ccfreefree(model->inputs);
3009	if (model->graph)
3010		ccv_nnc_symbolic_graph_free(model->graph);
3011	if (model->compiled_data)
3012		_ccv_cnnp_compiled_data_free(model, model->compiled_data);
3013	if (model->name)
3014		ccfreefree(model->name);
3015	ccfreefree(model);
3016}
3017 
3018void ccv_cnnp_model_cancel(ccv_cnnp_model_t* const model)
3019{
3020	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
3021	if (!compiled_data)
3022		return;
3023	if (compiled_data->graph)
3024		ccv_nnc_graph_cancel(compiled_data->graph);
3025	if (compiled_data->apply_gradients.graph)
3026		ccv_nnc_graph_cancel(compiled_data->apply_gradients.graph);
3027}