ccv_cnnp_model.c

Bug Summary

File:	nnc/ccv_cnnp_model.c
Warning:	line 2504, 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/2024-11-08-093000-160011-1 -x c ccv_cnnp_model.c
1#include "ccv_nnc.h"
2#include "ccv_nnc_easy.h"
3#include "ccv_nnc_internal.h"
4#include "ccv_internal.h"
5#include "_ccv_cnnp_model.h"
6#include "_ccv_nnc_graph.h"
7 
8// MARK - Level-5 API
9 
10ccv_cnnp_model_io_t ccv_cnnp_model_apply(ccv_cnnp_model_t* const model, const ccv_cnnp_model_io_t* const inputs, const int input_size)
11{
12	if (!model->io)
13		model->io = ccv_array_new(sizeof(ccv_cnnp_model_io_t), 1, 0);
14	ccv_cnnp_model_io_t model_io = ccmallocmalloc(sizeof(struct ccv_cnnp_model_io_s) + sizeof(ccv_nnc_tensor_symbol_t) * model->output_size);
15	model_io->param_ref = 0;
16	model_io->param_sel = 0;
17	model_io->visit = 0;
18	model_io->model = model;
19	model_io->dependencies = 0;
20	model_io->dependents = 0;
21	model_io->outgoings = 0;
22	model_io->outputs = (ccv_nnc_tensor_symbol_t*)(model_io + 1);
23	ccv_array_push(model->io, &model_io);
24	if (input_size > 0)
25	{
26		model_io->incomings = ccv_array_new(sizeof(ccv_cnnp_model_io_t), input_size, 0);
27		ccv_array_resize(model_io->incomings, input_size);
28		int i;
29		memcpy(ccv_array_get(model_io->incomings, 0)((void*)(((char*)((model_io->incomings)->data)) + (size_t
)(model_io->incomings)->rsize * (size_t)(0))), inputs, sizeof(ccv_cnnp_model_io_t) * input_size);
30		for (i = 0; i < input_size; i++)
31		{
32			if (!inputs[i]->outgoings)
33				inputs[i]->outgoings = ccv_array_new(sizeof(ccv_cnnp_model_io_t), 1, 0);
34			ccv_array_push(inputs[i]->outgoings, &model_io);
35		}
36	} else {
37		model_io->incomings = 0;
38	}
39	return model_io;
40}
41 
42void ccv_cnnp_model_add_dependencies(ccv_cnnp_model_io_t model_io, const ccv_cnnp_model_io_t* const dependencies, const int dependency_size)
43{
44	assert(dependency_size > 0)((void) sizeof ((dependency_size > 0) ? 1 : 0), __extension__
 ({ if (dependency_size > 0) ; else __assert_fail ("dependency_size > 0"
, "ccv_cnnp_model.c", 44, __extension__ __PRETTY_FUNCTION__);
 }));
45	if (!model_io->dependencies)
46		model_io->dependencies = ccv_array_new(sizeof(ccv_cnnp_model_io_t), dependency_size, 0);
47	int i, j;
48	for (i = 0; i < dependency_size; i++)
49	{
50		int flag = 0;
51		// Check if it is already exist or not.
52		for (j = 0; !flag && j < model_io->dependencies->rnum; j++)
53			if (*(ccv_cnnp_model_io_t*)ccv_array_get(model_io->dependencies, j)((void*)(((char*)((model_io->dependencies)->data)) + (size_t
)(model_io->dependencies)->rsize * (size_t)(j))) == dependencies[i])
54				flag = 1;
55		if (flag)
56			continue;
57		ccv_array_push(model_io->dependencies, dependencies + i);
58		++dependencies[i]->dependents;
59	}
60}
61 
62int ccv_cnnp_model_output_size(const ccv_cnnp_model_t* const model)
63{
64	return model->output_size;
65}
66 
67int ccv_cnnp_model_is_trainable(const ccv_cnnp_model_t* const model)
68{
69	// If the model is compiled, it is default to 1 unless it is not.
70	if (model->compiled_data)
71		return model->is_trainable >= 0 ? model->is_trainable : 1;
72	return model->is_trainable;
73}
74 
75ccv_cnnp_model_io_t ccv_cnnp_model_parameters(ccv_cnnp_model_t* const model, const int selector, const int index)
76{
77	if (!model->io)
78		model->io = ccv_array_new(sizeof(ccv_cnnp_model_io_t), 1, 0);
79	ccv_cnnp_model_io_t model_io = ccmallocmalloc(sizeof(struct ccv_cnnp_model_io_s));
80	model_io->param_ref = index >= 0 ? index + 1 : ALL_PARAMETERS-1;
81	model_io->param_sel = selector >= 0 ? selector + 1 : ALL_PARAMETERS-1;
82	model_io->visit = 0;
83	model_io->model = model;
84	model_io->outputs = 0;
85	model_io->dependencies = 0;
86	model_io->dependents = 0;
87	model_io->incomings = 0;
88	model_io->outgoings = 0;
89	ccv_array_push(model->io, &model_io);
90	return model_io;
91}
92 
93void ccv_cnnp_model_notify_hook(ccv_cnnp_model_t* const model, ccv_cnnp_model_notify_f func, void* const context)
94{
95	model->notify_hook.func = func;
96	model->notify_hook.context = context;
97}
98 
99void ccv_cnnp_model_notify(const ccv_cnnp_model_t* const model, const int tag, void* const payload)
100{
101	if (model->notify_hook.func)
102		model->notify_hook.func(model, tag, payload, model->notify_hook.context);
103	if (model->isa->notify)
104		model->isa->notify(model, tag, payload);
105}
106 
107static int _ccv_nnc_array_dedup_graph_exec_symbols(ccv_nnc_graph_exec_symbol_t* const graph_exec_symbols, int graph_exec_symbol_size)
108{
109	int i, j;
110	for (i = 0; i < graph_exec_symbol_size; i++)
111	{
112		ccv_nnc_graph_exec_symbol_t* const graph_exec_symbol = graph_exec_symbols + i;
113		// Check whether this tensor symbol has any duplicate.
114		for (j = i + 1; j < graph_exec_symbol_size;)
115		{
116			ccv_nnc_graph_exec_symbol_t* const other_symbol = graph_exec_symbols + j;
117			// If there is a same tensor symbol, remove it.
118			if (other_symbol->d == graph_exec_symbol->d && other_symbol->graph == graph_exec_symbol->graph)
119			{
120				if (j + 1 < graph_exec_symbol_size)
121					*other_symbol = graph_exec_symbols[graph_exec_symbol_size - 1];
122				--graph_exec_symbol_size;
123				continue;
124			}
125			++j;
126		}
127	}
128	return graph_exec_symbol_size;
129}
130 
131void ccv_cnnp_model_add_to_array(void* const context, const ccv_nnc_tensor_symbol_t symbol, const int is_trainable)
132{
133	ccv_cnnp_model_add_to_array_context_t* const add_to_array_context = (ccv_cnnp_model_add_to_array_context_t*)context;
134	ccv_cnnp_model_t* const model = add_to_array_context->sequence->model;
135	int i;
136	if (!model->parameter_indices)
137		model->parameter_indices = ccv_array_new(sizeof(int), 0, 0);
138	for (i = 0; i < add_to_array_context->symbols->rnum; i++)
139	{
140		const ccv_nnc_tensor_symbol_t other_symbol = *(ccv_nnc_tensor_symbol_t*)ccv_array_get(add_to_array_context->symbols, i)((void*)(((char*)((add_to_array_context->symbols)->data
)) + (size_t)(add_to_array_context->symbols)->rsize * (
size_t)(i)));
141		if (other_symbol.d == symbol.d && other_symbol.graph == symbol.graph)
142		{
143			// Only add to parameter_indices if it is trainable.
144			if (add_to_array_context->prefix == 't')
145				ccv_array_add_unique_int(model->parameter_indices, i);
146			// Found it, return, don't add it.
147			return;
148		}
149	}
150	// Only add to parameter_indices if it is trainable.
151	if (add_to_array_context->prefix == 't')
152		ccv_array_push(model->parameter_indices, &add_to_array_context->symbols->rnum);
153	// This is a new one, no need to add_unique_int, it is unique.
154	ccv_array_push(add_to_array_context->symbols, &symbol);
155	if (add_to_array_context->trainables)
156		ccv_array_push(add_to_array_context->trainables, &is_trainable);
157	char id[2048];
158	id[0] = add_to_array_context->prefix;
159	id[1] = '-';
160	int total_len = 2;
161	for (i = 0; i < add_to_array_context->sequence->sequences->rnum; i++)
162	{
163		const ccv_cnnp_model_name_t* const name = (ccv_cnnp_model_name_t*)ccv_array_get(add_to_array_context->sequence->sequences, i)((void*)(((char*)((add_to_array_context->sequence->sequences
)->data)) + (size_t)(add_to_array_context->sequence->
sequences)->rsize * (size_t)(i)));
164		int len;
165		if (name->name && name->name[0] != '\0')
166			len = snprintf(id + total_len, 2048 - total_len, "%s-%d-", name->name, name->sequence);
167		else
168			len = snprintf(id + total_len, 2048 - total_len, "%d-", name->sequence);
169		total_len += len;
170		if (total_len >= 2047)
171			break;
172	}
173	if (total_len < 2047)
174		total_len += snprintf(id + total_len, 2048 - total_len, "%d", add_to_array_context->sequence->it);
175	assert(total_len < 2048)((void) sizeof ((total_len < 2048) ? 1 : 0), __extension__
 ({ if (total_len < 2048) ; else __assert_fail ("total_len < 2048"
, "ccv_cnnp_model.c", 175, __extension__ __PRETTY_FUNCTION__)
; }));
176	char *heap_id = (char*)ccmallocmalloc(total_len + 1);
177	memcpy(heap_id, id, total_len + 1);
178	ccv_array_push(add_to_array_context->ids, &heap_id);
179	++add_to_array_context->sequence->it;
180}
181 
182static void _ccv_cnnp_compiled_data_init(ccv_cnnp_compiled_data_t* const compiled_data, const int output_size, ccv_array_t* const gradient_checkpoints)
183{
184	compiled_data->f = compiled_data->fits + output_size;
185	compiled_data->xpu_alloc.mp_hdr = -1;
186	compiled_data->xpu_alloc.freed = kh_init(dy_str)kh_init_dy_str();
187	compiled_data->xpu_alloc.allocd = kh_init(dy_alloc)kh_init_dy_alloc();
188	compiled_data->gradient_checkpoints = gradient_checkpoints;
189}
190 
191static void _ccv_cnnp_model_compile(ccv_cnnp_model_t* const model, const ccv_nnc_tensor_param_t* const inputs, const int input_size, const ccv_nnc_cmd_t loss)
192{
193	assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if
 (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c"
, 193, __extension__ __PRETTY_FUNCTION__); }));
194	model->inputs = ccmallocmalloc(sizeof(ccv_nnc_tensor_symbol_t) * input_size);
195	int i;
196	for (i = 0; i < input_size; i++)
197		model->inputs[i] = ccv_nnc_tensor_symbol_new(model->graph, inputs[i], 0);
198	ccv_array_t* const parameters = ccv_array_new(sizeof(ccv_nnc_tensor_symbol_t), 0, 0);
199	ccv_array_t* const parameter_ids = ccv_array_new(sizeof(char*), 0, 0);
200	ccv_array_t* const parameter_trainables = ccv_array_new(sizeof(int), 0, 0);
201	ccv_cnnp_model_sequence_t model_sequence = {
202		.bank = kh_init(ccv_cnnp_model_name_bank)kh_init_ccv_cnnp_model_name_bank()
203	};
204	ccv_cnnp_model_add_to_array_context_t add_to_parameter_context = {
205		.sequence = &model_sequence,
206		.prefix = 't',
207		.symbols = parameters,
208		.ids = parameter_ids,
209		.trainables = parameter_trainables,
210	};
211	ccv_array_t* const internals = ccv_array_new(sizeof(ccv_nnc_tensor_symbol_t), 0, 0);
212	ccv_array_t* const internal_ids = ccv_array_new(sizeof(char*), 0, 0);
213	ccv_cnnp_model_add_to_array_context_t add_to_output_context = {
214		.sequence = &model_sequence,
215		.prefix = 'r',
216		.symbols = internals,
217		.ids = internal_ids,
218		.trainables = 0,
219	};
220	ccv_cnnp_model_build_data_t build_data = {
221		.is_trainable = model->is_trainable >= 0 ? model->is_trainable : 1,
222		.model_sequence = &model_sequence,
223		.add_to_array = ccv_cnnp_model_add_to_array,
224		.parameters = parameters,
225		.context = {
226			.add_to_parameter = &add_to_parameter_context,
227			.add_to_output = &add_to_output_context,
228		},
229		.gradient_checkpoints = 0,
230	};
231	model->data = &build_data;
232	ccv_cnnp_model_build(model, model->graph, model->inputs, input_size, 0, 0);
233	for (i = 0; i < model->output_size; i++)
234	{
235		const ccv_nnc_tensor_symbol_t output = model->outputs[i];
236		const ccv_nnc_tensor_symbol_t alias_to = ccv_nnc_tensor_symbol_alias_to(model->graph, output);
237		if (alias_to.d == CCV_NNC_NO_TENSOR_SYMBOL)
238			continue;
239		// 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
240		// 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
241		// honest, because we cannot handle cases of alias is part of the original tensor but bind differently).
242		const ccv_nnc_tensor_param_t output_params = ccv_nnc_tensor_symbol_params(model->graph, output);
243		model->outputs[i] = ccv_nnc_tensor_symbol_new(model->graph, output_params, 0);
244		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");
245		ccv_nnc_graph_exec_symbol_set_flags(model->graph, make_contiguous, CCV_NNC_GRAPH_EXEC_DISABLE_OPT);
246	}
247	model->data = 0;
248	kh_destroy(ccv_cnnp_model_name_bank, model_sequence.bank)kh_destroy_ccv_cnnp_model_name_bank(model_sequence.bank);
249	if (model_sequence.sequences)
250		ccv_array_free(model_sequence.sequences);
251	// Check if there are parameters that are not trainables. If there are, we will allocate uint64 bitmap to record that.
252	int not_trainables = 0;
253	// Assert no parameter is alias.
254	for (i = 0; i < parameters->rnum; i++)
255	{
256		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)));
257		const ccv_nnc_tensor_symbol_t alias_to = ccv_nnc_tensor_symbol_alias_to(parameter.graph, parameter);
258		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", 258, __extension__ __PRETTY_FUNCTION__)
; })); // Cannot find the one alias to.
259		if (*(int*)ccv_array_get(parameter_trainables, i)((void*)(((char*)((parameter_trainables)->data)) + (size_t
)(parameter_trainables)->rsize * (size_t)(i))) == 0)
260			not_trainables = 1;
261	}
262	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", 262, __extension__ __PRETTY_FUNCTION__)
; }));
263	uint64_t* parameter_flags = 0;
264	if (not_trainables)
265	{
266		parameter_flags = (uint64_t*)cccalloccalloc(((parameters->rnum + 63) >> 6), sizeof(uint64_t));
267		for (i = 0; i < parameter_trainables->rnum; i++)
268			if (*(int*)ccv_array_get(parameter_trainables, i)((void*)(((char*)((parameter_trainables)->data)) + (size_t
)(parameter_trainables)->rsize * (size_t)(i))))
269				parameter_flags[i >> 6] |= ((uint64_t)1 << (i & 63));
270	}
271	ccv_array_free(parameter_trainables);
272	// Assert no internal is alias.
273	for (i = 0; i < internals->rnum; i++)
274	{
275		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)));
276		const ccv_nnc_tensor_symbol_t alias_to = ccv_nnc_tensor_symbol_alias_to(internal.graph, internal);
277		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", 277, __extension__ __PRETTY_FUNCTION__)
; })); // Cannot find the one alias to.
278	}
279	const int output_size = model->output_size;
280	ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
281	const int parameters_rnum = parameters->rnum;
282	if (input_size > 0)
283	{
284		ccv_array_resize(parameters, parameters_rnum + input_size);
285		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));
286	}
287	ccv_nnc_symbolic_graph_simplify(model->graph,
288		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)
289			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)
290			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)
291			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),
292		ccv_array_get(parameters, 0)((void*)(((char*)((parameters)->data)) + (size_t)(parameters
)->rsize * (size_t)(0))), parameters_rnum + input_size,
293		model->outputs, output_size,
294		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));
295	ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
296	// Size it down.
297	parameters->rnum = parameters_rnum;
298	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));
299	_ccv_cnnp_compiled_data_init(compiled_data, output_size, build_data.gradient_checkpoints);
300	const int evaluate_to_size = compiled_data->evaluate.to_size = ccv_nnc_symbolic_graph_destination_size(model->graph);
301	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", 301, __extension__ __PRETTY_FUNCTION__)
; }));
302	compiled_data->evaluate.tos = ccmallocmalloc(sizeof(ccv_nnc_graph_exec_symbol_t) * evaluate_to_size);
303	memcpy(compiled_data->evaluate.tos, ccv_nnc_symbolic_graph_destinations(model->graph), sizeof(ccv_nnc_graph_exec_symbol_t) * evaluate_to_size);
304	compiled_data->loss = loss;
305	if (loss.cmd == CCV_NNC_NOOP)
306	{
307		// If no loss function provided, there is no fits.
308		for (i = 0; i < output_size; i++)
309		{
310			compiled_data->fits[i] = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
311			const ccv_nnc_tensor_symbol_t alias_to = ccv_nnc_tensor_symbol_alias_to(model->graph, model->outputs[i]);
312			if (alias_to.d < 0)
313				compiled_data->f[i] = model->outputs[i];
314			else { // We cannot differentiate against an alias, therefore, we have to verify this output is full, and we can diff against the original.
315				int ofs[CCV_NNC_MAX_DIM_ALLOC(12)];
316				int inc[CCV_NNC_MAX_DIM_ALLOC(12)];
317				ccv_nnc_tensor_symbol_alias_params(model->graph, model->outputs[i], ofs, inc);
318				int j;
319				for (j = 0; j < CCV_NNC_MAX_DIM_ALLOC(12); j++)
320					{ 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"
, 320, __extension__ __PRETTY_FUNCTION__); })); } // There is no ofs.
321				compiled_data->f[i] = alias_to; // Unfortunately, I cannot assert the size yet.
322			}
323		}
324	} else {
325		for (i = 0; i < output_size; i++)
326		{
327			const ccv_nnc_tensor_param_t info = ccv_nnc_tensor_symbol_params(model->graph, model->outputs[i]);
328			const ccv_nnc_tensor_symbol_t fit = compiled_data->fits[i] = ccv_nnc_tensor_symbol_new(model->graph, info, 0);
329			compiled_data->f[i] = ccv_nnc_tensor_symbol_new(model->graph, ccv_nnc_tensor_auto, 0);
330			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);
331		}
332	}
333	ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
334	ccv_nnc_symbolic_graph_simplify(model->graph,
335		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.
336		0, 0, // No need to provide binds at this point.
337		compiled_data->f, model->output_size,
338		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));
339	ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
340	// If inputs are from GPU, stream type is GPU.
341	compiled_data->parameters = parameters;
342	compiled_data->parameter_flags = parameter_flags;
343	compiled_data->internals = internals;
344	compiled_data->ids.parameters = parameter_ids;
345	compiled_data->ids.internals = internal_ids;
346	ccv_cnnp_model_gradient_checkpoints_cleanup_after_build(compiled_data, model->graph);
347}
348 
349static 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)
350{
351	ccv_array_t* const stack = (ccv_array_t*)context;
352	ccv_array_push(stack, &symbol.d);
353}
354 
355static 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)
356{
357	const ccv_nnc_tensor_symbol_t src_symbol = {
358		.d = src_index,
359		.graph = src_graph
360	};
361	const ccv_nnc_tensor_symbol_t dest_symbol = {
362		.d = dest_index,
363		.graph = dest_graph
364	};
365	const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(src_graph, src_symbol);
366	ccv_nnc_tensor_symbol_set(dest_graph, dest_symbol, params);
367	int ofs[CCV_NNC_MAX_DIM_ALLOC(12)];
368	int inc[CCV_NNC_MAX_DIM_ALLOC(12)];
369	if (0 == ccv_nnc_tensor_symbol_alias_params(src_graph, src_symbol, ofs, inc))
370		ccv_nnc_tensor_symbol_alias_set(dest_graph, dest_symbol, ofs, inc);
371}
372 
373static 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)
374{
375	const ccv_nnc_tensor_symbol_t src_symbol = {
376		.d = src_index,
377		.graph = src_graph
378	};
379	const ccv_nnc_tensor_param_t src_params = ccv_nnc_tensor_symbol_params(src_graph, src_symbol);
380	const ccv_nnc_tensor_symbol_t dest_symbol = {
381		.d = dest_index,
382		.graph = dest_graph
383	};
384	const ccv_nnc_tensor_param_t dest_params = ccv_nnc_tensor_symbol_params(dest_graph, dest_symbol);
385	return memcmp(src_params.dim, dest_params.dim, sizeof(src_params.dim)) == 0;
386}
387 
388static 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);
389static void _ccv_cnnp_compiled_data_graph_free(ccv_cnnp_compiled_data_t* const compiled_data);
390 
391typedef struct {
392	int parallel_count;
393	ccv_nnc_symbolic_graph_t* graph;
394	ccv_nnc_graph_exec_arena_t* graph_exec_arena;
395} ccv_nnc_graph_exec_update_t;
396 
397static 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)
398{
399	ccv_nnc_graph_exec_update_t* const graph_exec_update = (ccv_nnc_graph_exec_update_t*)context;
400	ccv_nnc_graph_exec_arena_t* const graph_exec_arena = graph_exec_update->graph_exec_arena;
401	ccv_nnc_graph_exec_t graph_exec = ccv_nnc_graph_exec_from_symbol(graph_exec_arena, symbol);
402	ccv_nnc_graph_exec_set(graph_exec.graph, graph_exec, cmd);
403	ccv_nnc_graph_exec_set_hint(graph_exec.graph, graph_exec, hint);
404	const ccv_nnc_symbolic_graph_t* const graph = graph_exec_update->graph;
405	const int parallel_count = graph_exec_update->parallel_count;
406	int i;
407	for (i = 1; i < parallel_count; i++)
408	{
409		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));
410		if (!CCV_NO_GRAPH_EXEC(copy)((copy).graph == 0))
411		{
412			ccv_nnc_graph_exec_set(copy.graph, copy, cmd);
413			ccv_nnc_graph_exec_set_hint(copy.graph, copy, hint);
414		}
415	}
416}
417 
418void 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)
419{
420	assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if
 (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c"
, 420, __extension__ __PRETTY_FUNCTION__); }));
421	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", 421, __extension__ __PRETTY_FUNCTION__)
; }));
422	assert(!init->graph)((void) sizeof ((!init->graph) ? 1 : 0), __extension__ ({ if
 (!init->graph) ; else __assert_fail ("!init->graph", "ccv_cnnp_model.c"
, 422, __extension__ __PRETTY_FUNCTION__); }));
423	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
424	init->graph = ccv_nnc_symbolic_graph_new();
425	ccv_array_t* const stack = ccv_array_new(sizeof(int), 0, 0);
426	ccv_nnc_graph_exec_symbol_new_hook(init->graph, _ccv_cnnp_graph_push_graph_exec_symbol, stack, 0);
427	_ccv_cnnp_model_compile(init, inputs, input_size, compiled_data->loss);
428	init->parallel_count = model->parallel_count;
429	init->memory_compression = model->memory_compression;
430	init->memory_reduction = model->memory_reduction;
431	init->gradient_checkpointing = model->gradient_checkpointing;
432	init->compiled_data->stream_type = model->compiled_data->stream_type;
433	init->compiled_data->minimize.minimizer = model->compiled_data->minimize.minimizer;
434	init->compiled_data->minimize.max_saved_aux_size = model->compiled_data->minimize.max_saved_aux_size;
435	if (model->compiled_data->gradient_mode != CCV_CNNP_COMPILED_DATA_GRADIENT_NONE)
436		_ccv_cnnp_model_gradient_init(init, model->compiled_data->gradient_mode, model->compiled_data->disable_outgrad, 0, 0);
437	ccv_nnc_graph_exec_symbol_new_hook(init->graph, 0, 0, 0);
438	ccv_nnc_symbolic_graph_tensor_auto(init->graph, TRAVERSE_FULL0,0,0,0);
439	int i, j;
440	// Verify parameters, internals and saved_aux in both graph has the same dimensionality.
441	for (i = 0; i < compiled_data->parameters->rnum; i++)
442	{
443		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;
444		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", 444, __extension__ __PRETTY_FUNCTION__)
; }));
445	}
446	for (i = 0; i < compiled_data->internals->rnum; i++)
447	{
448		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;
449		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", 449, __extension__ __PRETTY_FUNCTION__)
; }));
450	}
451	// Update inputs.
452	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", 452, __extension__ __PRETTY_FUNCTION__)
; }));
453	for (i = 0; i < model->input_size; i++)
454		if (model->inputs[i].d >= 0)
455		{
456			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", 456, __extension__ __PRETTY_FUNCTION__)
; }));
457			_ccv_nnc_tensor_symbol_reinit(init->graph, model->graph, init->inputs[i].d, model->inputs[i].d);
458		}
459	// Update outputs.
460	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", 460, __extension__ __PRETTY_FUNCTION__)
; }));
461	for (i = 0; i < model->output_size; i++)
462	{
463		if (model->outputs[i].d >= 0)
464		{
465			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", 465, __extension__
 __PRETTY_FUNCTION__); }));
466			_ccv_nnc_tensor_symbol_reinit(init->graph, model->graph, init->outputs[i].d, model->outputs[i].d);
467		}
468		if (model->outputs[i].d != model->compiled_data->f[i].d)
469		{
470			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", 470, __extension__ __PRETTY_FUNCTION__)
; }));
471			if (model->compiled_data->f[i].d >= 0)
472			{
473				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", 473, __extension__ __PRETTY_FUNCTION__)
; }));
474				_ccv_nnc_tensor_symbol_reinit(init->graph, model->graph, init->compiled_data->f[i].d, model->compiled_data->f[i].d);
475			}
476		}
477	}
478	// Go through the graph to set tensor on matching symbols
479	for (i = 0; i < stack->rnum; i++)
480	{
481		const int d = *(int*)ccv_array_get(stack, i)((void*)(((char*)((stack)->data)) + (size_t)(stack)->rsize
 * (size_t)(i)));
482		// If exceed range, skip.
483		if (d >= ccv_nnc_graph_exec_symbol_count(init->graph) ||
484			d >= ccv_nnc_graph_exec_symbol_count(model->graph))
485			continue;
486		const ccv_nnc_graph_exec_symbol_t src_symbol = {
487			.d = d,
488			.graph = init->graph
489		};
490		const ccv_nnc_graph_exec_symbol_t dest_symbol = {
491			.d = d,
492			.graph = model->graph
493		};
494		const ccv_nnc_cmd_t src_cmd = ccv_nnc_graph_exec_symbol_cmd(init->graph, src_symbol);
495		const ccv_nnc_cmd_t dest_cmd = ccv_nnc_graph_exec_symbol_cmd(model->graph, dest_symbol);
496		// If the name doesn't match, skip.
497		if (dest_cmd.cmd != src_cmd.cmd && src_cmd.cmd != CCV_NNC_NOOP)
498			continue;
499		// Now get all the inputs and outputs, if matches, set them.
500		const int* src_inputs;
501		int src_input_size;
502		const int* src_outputs;
503		int src_output_size;
504		ccv_nnc_graph_exec_symbol_io(init->graph, src_symbol, &src_inputs, &src_input_size, &src_outputs, &src_output_size);
505		const int* dest_inputs;
506		int dest_input_size;
507		const int* dest_outputs;
508		int dest_output_size;
509		ccv_nnc_graph_exec_symbol_io(model->graph, dest_symbol, &dest_inputs, &dest_input_size, &dest_outputs, &dest_output_size);
510		// We may have unmatched input / output size because this is the minimizer and it has
511		// different saved_aux (for example, when we shrunk with CMD_NOOP).
512		if (src_input_size != dest_input_size)
513			continue;
514		if (src_output_size != dest_output_size)
515			continue;
516		ccv_nnc_graph_exec_symbol_set(model->graph, dest_symbol, src_cmd);
517		// There may be mismatches of the source tensor symbols and destination tensor symbols. The reason is because
518		// we may later passed-in the minimizer, therefore, we may allocate tensors for minimizer later in the original
519		// graph whereas in the newly created graph, it is streamlined (the minimizer exists from the beginning). That
520		// will make the order of tensor symbols creation different, therefore, exact which tensor is which wrong as
521		// well. However, set a new minimizer won't change the exec symbol ordering, because we never create new exec
522		// symbols after gradient init step. Changing a new minimizer just updated that exec symbols setting, it is not
523		// a new exec symbol.
524		for (j = 0; j < src_input_size; j++)
525			if (src_inputs[j] >= 0)
526				_ccv_nnc_tensor_symbol_reinit(init->graph, model->graph, src_inputs[j], dest_inputs[j]);
527		for (j = 0; j < src_output_size; j++)
528			if (src_outputs[j] >= 0)
529				_ccv_nnc_tensor_symbol_reinit(init->graph, model->graph, src_outputs[j], dest_outputs[j]);
530	}
531	ccv_array_free(stack);
532	// After this, we get all tensors in the model graph resolved through tensor_auto.
533	ccv_nnc_symbolic_graph_tensor_auto(model->graph, TRAVERSE_FULL0,0,0,0);
534	// Verify symbols we get matches.
535	const int parameter_size = compiled_data->parameters->rnum;
536	for (i = 0; i < parameter_size; i++)
537		{ 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", 537, __extension__ __PRETTY_FUNCTION__)
; })); }
538	const int internal_size = compiled_data->internals->rnum;
539	for (i = 0; i < internal_size; i++)
540		{ 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", 540, __extension__ __PRETTY_FUNCTION__)
; })); }
541	// Go through compiled data.
542	if (compiled_data->tensor_arena)
543	{
544		const int flag = ccv_nnc_tensor_arena_reinit(compiled_data->tensor_arena, model->graph);
545		if (flag == 0 && compiled_data->graph_exec_arena)
546		{
547			ccv_nnc_graph_exec_reinit(compiled_data->graph_exec_arena, compiled_data->graph, model->graph);
548			// Since we will reinit, if we previously set is_test, we need to set it again.
549			if (compiled_data->is_test)
550			{
551				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; });
552				ccv_nnc_graph_exec_update_t update = {
553					.parallel_count = parallel_count,
554					.graph = model->graph,
555					.graph_exec_arena = compiled_data->graph_exec_arena,
556				};
557				ccv_cnnp_model_set_is_test(model, 1, _ccv_cnnp_cmd_update_for_execs, &update);
558			}
559		} else
560			// Free-up tensor arena & graph exec arena.
561			_ccv_cnnp_compiled_data_graph_free(compiled_data);
562	}
563	// There are other compiled graphs, for accum and apply gradients.
564	// However, the main conclusion is, these absorb operations shouldn't impact parameters.
565	// Thus, it won't impact the shape of gradients (only outgrad). Since for outgrad, we
566	// don't allocate ourselves, it is not a concern. For normal gradients, the shape cannot
567	// be changed otherwise parameters' shape will be meaningless. The same goes to internals.
568	// That is why we don't update these compiled graphs at all this point.
569	// Free the model, we've already "absorbed" it.
570	ccv_cnnp_model_free(init);
571}
572 
573void 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)
574{
575	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", 575, __extension__ __PRETTY_FUNCTION__)
; }));
576	if (model->input_size == 0)
577		model->input_size = input_size;
578	if (!model->graph) // The graph is not compiled yet.
579	{
580		model->graph = ccv_nnc_symbolic_graph_new();
581		_ccv_cnnp_model_compile(model, inputs, input_size, loss);
582		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", 582, __extension__ __PRETTY_FUNCTION__)
; }));
583		int i, flag = 0;
584		for (i = 0; !flag && i < input_size; i++)
585			flag = (CCV_TENSOR_GET_MEMORY(inputs[i].type)((inputs[i].type) & 0x3) == CCV_TENSOR_GPU_MEMORY);
586		// If inputs are from GPU, stream type is GPU.
587		model->compiled_data->stream_type = flag ? CCV_STREAM_CONTEXT_GPU : CCV_STREAM_CONTEXT_CPU;
588		model->compiled_data->minimize.minimizer = minimizer;
589		model->compiled_data->minimize.max_saved_aux_size = ccv_nnc_minimizer_saved_aux_size(minimizer);
590	} else {
591		// Now, finally fill in this part. If the graph is already compiled, we make a copy of the model.
592		// And then absorb the "new model" to the old one.
593		ccv_cnnp_model_t* const init = ccv_cnnp_model_copy(model, model->is_trainable);
594		ccv_cnnp_model_absorb(model, init, inputs, input_size);
595		// Reset minimizer.
596		ccv_cnnp_model_set_minimizer(model, minimizer, 1, 0, 0);
597	}
598}
599 
600ccv_cnnp_model_t* ccv_cnnp_model_copy(const ccv_cnnp_model_t* const model, const int is_trainable)
601{
602	ccv_cnnp_model_t* const new_model = _ccv_cnnp_model_copy(model, 0);
603	new_model->is_trainable = is_trainable;
604	return new_model;
605}
606 
607void ccv_cnnp_model_tensor_auto(ccv_cnnp_model_t* const model, ccv_nnc_tensor_param_t* const outputs, const int output_size)
608{
609	assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if
 (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c"
, 609, __extension__ __PRETTY_FUNCTION__); }));
610	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", 610, __extension__ __PRETTY_FUNCTION__)
; }));
611	ccv_nnc_symbolic_graph_t* const graph = model->graph;
612	ccv_nnc_symbolic_graph_tensor_auto(graph, TRAVERSE_FULL0,0,0,0);
613	int i;
614	for (i = 0; i < output_size; i++)
615	{
616		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", 616, __extension__ __PRETTY_FUNCTION__)
; }));
617		outputs[i] = ccv_nnc_tensor_symbol_params(graph, model->outputs[i]);
618	}
619}
620 
621void ccv_cnnp_model_set_workspace_size(ccv_cnnp_model_t* const model, size_t workspace_size)
622{
623	if (workspace_size == model->workspace_size)
624		return;
625	model->workspace_size = workspace_size;
626	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
627	if (compiled_data && compiled_data->graph)
628		ccv_nnc_graph_autotune(compiled_data->graph, workspace_size, 0, TRAVERSE_FULL0,0,0,0);
629}
630 
631size_t ccv_cnnp_model_workspace_size(ccv_cnnp_model_t* const model)
632{
633	return model->workspace_size;
634}
635 
636void ccv_cnnp_model_set_data_parallel(ccv_cnnp_model_t* const model, const int parallel)
637{
638	if (parallel == 0)
639		model->parallel_count = ccv_nnc_device_count(CCV_STREAM_CONTEXT_GPU);
640	else
641		model->parallel_count = parallel;
642	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
643	if (compiled_data)
644		{ 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", 644, __extension__ __PRETTY_FUNCTION__)
; })); }
645}
646 
647void ccv_cnnp_model_set_max_concurrency(ccv_cnnp_model_t* const model, const int max_stream_count)
648{
649	model->max_stream_count = max_stream_count;
650	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
651	if (compiled_data)
652		{ 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", 652, __extension__ __PRETTY_FUNCTION__)
; })); }
653}
654 
655void ccv_cnnp_model_set_memory_compression(ccv_cnnp_model_t* const model, const int memory_compression)
656{
657	model->memory_compression = memory_compression;
658	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
659	if (compiled_data)
660		{ 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", 660, __extension__ __PRETTY_FUNCTION__)
; })); }
661}
662 
663void ccv_cnnp_model_set_memory_reduction(ccv_cnnp_model_t* const model, const int memory_reduction)
664{
665	model->memory_reduction = memory_reduction;
666	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
667	if (compiled_data)
668		{ 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", 668, __extension__ __PRETTY_FUNCTION__)
; })); }
669}
670 
671void ccv_cnnp_model_set_gradient_checkpointing(ccv_cnnp_model_t* const model, const int gradient_checkpointing)
672{
673	model->gradient_checkpointing = gradient_checkpointing;
674}
675 
676int ccv_cnnp_model_gradient_checkpointing(ccv_cnnp_model_t* const model)
677{
678	return model->gradient_checkpointing;
679}
680 
681typedef struct {
682	int parallel_count;
683	ccv_nnc_symbolic_graph_t* graph;
684	ccv_cnnp_compiled_data_t* compiled_data;
685	ccv_nnc_tensor_arena_t* tensor_arena;
686} ccv_nnc_tensor_init_states_t;
687 
688static int _ccv_cnnp_any_to_init(const ccv_cnnp_compiled_data_t* const compiled_data)
689{
690	int i;
691	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));
692	for (i = 0; i < compiled_data->parameters->rnum; i++)
693	{
694		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;
695		if (!(init_v[d >> 5] & (1u << (d & 0x1f))))
696			return 1;
697	}
698	for (i = 0; i < compiled_data->internals->rnum; i++)
699	{
700		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;
701		if (!(init_v[d >> 5] & (1u << (d & 0x1f))))
702			return 1;
703	}
704	return 0;
705}
706 
707static 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)
708{
709	ccv_nnc_tensor_init_states_t* const tensor_init_states = (ccv_nnc_tensor_init_states_t*)context;
710	ccv_nnc_tensor_arena_t* const tensor_arena = tensor_init_states->tensor_arena;
711	ccv_nnc_tensor_t* const output_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, output_symbol);
712	if (!output_tensor)
713		return;
714	const int d = output_symbol.d;
715	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", 715, __extension__ __PRETTY_FUNCTION__)
; }));
716  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));
717	if (init_v[d >> 5] & (1u << (d & 0x1f)))
718		return;
719	init_v[d >> 5] |= (1u << (d & 0x1f));
720	ccv_nnc_cmd_exec(cmd, hint, flags, &input, input ? 1 : 0, &output_tensor, 1, 0);
721	const ccv_nnc_symbolic_graph_t* const graph = tensor_init_states->graph;
722	const int parallel_count = tensor_init_states->parallel_count;
723	int i;
724	for (i = 1; i < parallel_count; i++)
725	{
726		ccv_nnc_tensor_t* const copy = ccv_nnc_tensor_from_symbol(tensor_arena, ccv_nnc_tensor_symbol_copy(graph, output_symbol, i));
727		if (copy)
728			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);
729	}
730}
731 
732// This method can only handle cases we added new tensors and exec, never delete. This invariant is true because
733// we setup everything (including calling simplify method) in ccv_cnnp_model_compile method, before this rewind setup.
734static void _ccv_cnnp_model_rewind_graph(ccv_cnnp_model_t* const model)
735{
736	assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if
 (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c"
, 736, __extension__ __PRETTY_FUNCTION__); }));
737	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", 737, __extension__ __PRETTY_FUNCTION__)
; }));
738	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
739	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", 739, __extension__
 __PRETTY_FUNCTION__); }));
740	int i;
741	for (i = 0; i < compiled_data->rewindables->rnum; i++)
742	{
743		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)));
744		if (rewind_symbol->type == CCV_CNNP_REWIND_GRAPH_EXEC)
745			ccv_nnc_graph_exec_symbol_free(model->graph, rewind_symbol->graph_exec);
746		else if (rewind_symbol->type == CCV_CNNP_REWIND_TENSOR)
747			ccv_nnc_tensor_symbol_free(model->graph, rewind_symbol->tensor);
748	}
749	ccv_array_clear(compiled_data->rewindables);
750	ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
751}
752 
753static 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)
754{
755	const ccv_cnnp_rewind_symbol_t rewind_symbol = {
756		.type = CCV_CNNP_REWIND_TENSOR,
757		.tensor = symbol
758	};
759	ccv_array_t* const rewind_symbols = (ccv_array_t*)context;
760	ccv_array_push(rewind_symbols, &rewind_symbol);
761}
762 
763static 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)
764{
765	const ccv_cnnp_rewind_symbol_t rewind_symbol = {
766		.type = CCV_CNNP_REWIND_TENSOR,
767		.tensor = symbol
768	};
769	ccv_array_t* const rewind_symbols = (ccv_array_t*)context;
770	ccv_array_push(rewind_symbols, &rewind_symbol);
771}
772 
773static 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)
774{
775	const ccv_cnnp_rewind_symbol_t rewind_symbol = {
776		.type = CCV_CNNP_REWIND_GRAPH_EXEC,
777		.graph_exec = symbol
778	};
779	ccv_array_t* const rewind_symbols = (ccv_array_t*)context;
780	ccv_array_push(rewind_symbols, &rewind_symbol);
781}
782 
783static 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)
784{
785	ccv_nnc_graph_exec_t const update_exec = ccv_nnc_graph_exec_from_symbol(graph_exec_arena, exec_symbol);
786	if (!CCV_NO_GRAPH_EXEC(update_exec)((update_exec).graph == 0))
787		ccv_nnc_graph_exec_set(update_exec.graph, update_exec, cmd);
788	int i;
789	for (i = 1; i < parallel_count; i++)
790	{
791		ccv_nnc_graph_exec_symbol_t copy_symbol = ccv_nnc_graph_exec_symbol_copy(symbolic_graph, exec_symbol, i);
792		const ccv_nnc_graph_exec_t copy = ccv_nnc_graph_exec_from_symbol(graph_exec_arena, copy_symbol);
793		if (!CCV_NO_GRAPH_EXEC(copy)((copy).graph == 0))
794			ccv_nnc_graph_exec_set(copy.graph, copy, cmd);
795	}
796}
797 
798static 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)
799{
800	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 800, __extension__ __PRETTY_FUNCTION__); }));
801	assert(symbolic_graph)((void) sizeof ((symbolic_graph) ? 1 : 0), __extension__ ({ if
 (symbolic_graph) ; else __assert_fail ("symbolic_graph", "ccv_cnnp_model.c"
, 801, __extension__ __PRETTY_FUNCTION__); }));
802	ccv_nnc_graph_exec_symbol_set(symbolic_graph, exec_symbol, cmd);
803	int i;
804	for (i = 1; i < parallel_count; i++)
805	{
806		ccv_nnc_graph_exec_symbol_t copy_symbol = ccv_nnc_graph_exec_symbol_copy(symbolic_graph, exec_symbol, i);
807		if (copy_symbol.graph)
808			ccv_nnc_graph_exec_symbol_set(symbolic_graph, copy_symbol, cmd);
809	}
810	ccv_nnc_graph_exec_arena_t* const graph_exec_arena = compiled_data->graph_exec_arena;
811	if (graph_exec_arena)
812		_ccv_cnnp_model_graph_symbol_exec_set_for_graph_exec_arena(graph_exec_arena, parallel_count, exec_symbol, cmd, symbolic_graph);
813	// Skip backward graph exec arena because it is for a specific accum symbolic graph, not the main graph (model->graph)
814	ccv_nnc_graph_exec_arena_t* const gradient_graph_exec_arena = compiled_data->apply_gradients.graph_exec_arena;
815	if (gradient_graph_exec_arena)
816		_ccv_cnnp_model_graph_symbol_exec_set_for_graph_exec_arena(gradient_graph_exec_arena, parallel_count, exec_symbol, cmd, symbolic_graph);
817}
818 
819static 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)
820{
821	int this_parameter_flag = 0;
822	if (update_nodes[parameter_indice].d == CCV_NNC_NO_TENSOR_SYMBOL)
823		return this_parameter_flag;
824	const ccv_nnc_cmd_t old_minimizer = ccv_nnc_graph_exec_symbol_cmd(graph, update_nodes[parameter_indice]);
825	int j, k;
826	// For no-op, we can preserve previous saved_aux_size.
827	if (old_minimizer.cmd != minimizer.cmd && minimizer.cmd != CCV_NNC_NOOP)
828	{
829		// If the old minimizer is a noop, then the old_saved_aux_size should be whatever its previous
830		// saved_aux_size is, otherwise we will reinit the saved_aux repeatedly if you switch between
831		// noop and a minimizer. We don't want that because we do that in high-level frameworks to
832		// make sure some model parameters don't update if we don't want them to.
833		int old_saved_aux_size;
834		if (old_minimizer.cmd == CCV_NNC_NOOP)
835		{
836			int input_size;
837			ccv_nnc_graph_exec_symbol_io(graph, update_nodes[parameter_indice], 0, &input_size, 0, 0);
838			if (input_size < 2) // This is not legit.
839				old_saved_aux_size = ccv_nnc_minimizer_saved_aux_size(old_minimizer);
840			else // See ccv_nnc_minimizer_saved_aux_size, the saved_aux is inputs excluding gradients and parameters.
841				old_saved_aux_size = input_size - 2;
842		} else
843			old_saved_aux_size = ccv_nnc_minimizer_saved_aux_size(old_minimizer);
844		if (old_saved_aux_size != saved_aux_size)
845		{
846			this_parameter_flag = 1;
847			if (saved_aux_size > old_saved_aux_size)
848			{
849				// Allocate new tensor symbols.
850				const ccv_nnc_tensor_param_t info = ccv_nnc_tensor_symbol_params(graph, updated_parameters[parameter_indice]);
851				for (j = old_saved_aux_size; j < saved_aux_size; j++)
852				{
853					saved_aux[parameter_indice * max_saved_aux_size + j].source = ccv_nnc_tensor_symbol_new(graph, info, 0);
854					saved_aux[parameter_indice * max_saved_aux_size + j].destination = ccv_nnc_tensor_symbol_new(graph, info, 0);
855					const int device_id = CCV_TENSOR_GET_DEVICE_ID(info.type)(((info.type) & 0xfff00) >> 8);
856					for (k = 1; k < parallel_count; k++)
857					{
858						ccv_nnc_tensor_param_t dev_info = info;
859						if (k != device_id)
860							CCV_TENSOR_SET_DEVICE_ID(dev_info.type, k)(dev_info.type) = (((dev_info.type) & ~0xfff00) | (((k) &
 0xfff) << 8));
861						else
862							CCV_TENSOR_SET_DEVICE_ID(dev_info.type, 0)(dev_info.type) = (((dev_info.type) & ~0xfff00) | (((0) &
 0xfff) << 8));
863						const ccv_nnc_tensor_symbol_t src_copy = ccv_nnc_tensor_symbol_new(graph, dev_info, 0);
864						const ccv_nnc_tensor_symbol_t dest_copy = ccv_nnc_tensor_symbol_new(graph, dev_info, 0);
865						ccv_nnc_tensor_symbol_set_copy(graph, saved_aux[parameter_indice * max_saved_aux_size + j].source, k, src_copy);
866						ccv_nnc_tensor_symbol_set_copy(graph, saved_aux[parameter_indice * max_saved_aux_size + j].destination, k, dest_copy);
867					}
868				}
869			} else {
870				for (j = saved_aux_size; j < old_saved_aux_size; j++)
871				{
872					for (k = 1; k < parallel_count; k++)
873					{
874						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);
875						if (src_copy.d >= 0)
876						{
877							ccv_nnc_tensor_symbol_free(graph, src_copy);
878							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
});
879						}
880						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);
881						if (dest_copy.d >= 0)
882						{
883							ccv_nnc_tensor_symbol_free(graph, dest_copy);
884							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
});
885						}
886					}
887					ccv_nnc_tensor_symbol_free(graph, saved_aux[parameter_indice * max_saved_aux_size + j].source);
888					ccv_nnc_tensor_symbol_free(graph, saved_aux[parameter_indice * max_saved_aux_size + j].destination);
889					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
};
890				}
891			}
892		}
893	}
894	_ccv_cnnp_model_graph_exec_symbol_set(graph, compiled_data, parallel_count, update_nodes[parameter_indice], minimizer);
895	if (this_parameter_flag)
896	{
897		ccv_nnc_tensor_symbol_t update_inputs[saved_aux_size + 2];
898		ccv_nnc_tensor_symbol_t update_outputs[saved_aux_size + 1];
899		const int* inputs = 0;
900		int input_size = 0;
901		ccv_nnc_graph_exec_symbol_io(graph, update_nodes[parameter_indice], &inputs, &input_size, 0, 0);
902		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", 902, __extension__ __PRETTY_FUNCTION__)
; }));
903		update_inputs[0].d = inputs[0];
904		update_inputs[0].graph = graph;
905		update_inputs[1].d = inputs[1];
906		update_inputs[1].graph = graph;
907		update_outputs[0] = updated_parameters[parameter_indice];
908		for (j = 0; j < saved_aux_size; j++)
909		{
910			update_inputs[j + 2] = saved_aux[parameter_indice * max_saved_aux_size + j].source;
911			update_outputs[j + 1] = saved_aux[parameter_indice * max_saved_aux_size + j].destination;
912		}
913		ccv_nnc_graph_exec_symbol_set_io(graph, update_nodes[parameter_indice], update_inputs, saved_aux_size + 2, update_outputs, saved_aux_size + 1);
914		for (k = 1; k < parallel_count; k++)
915		{
916			const ccv_nnc_graph_exec_symbol_t copy = ccv_nnc_graph_exec_symbol_copy(graph, update_nodes[parameter_indice], k);
917			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"
, 917, __extension__ __PRETTY_FUNCTION__); }));
918			ccv_nnc_graph_exec_symbol_io(graph, copy, &inputs, &input_size, 0, 0);
919			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", 919, __extension__ __PRETTY_FUNCTION__)
; }));
920			update_inputs[0].d = inputs[0];
921			update_inputs[0].graph = graph;
922			update_inputs[1].d = inputs[1];
923			update_inputs[1].graph = graph;
924			update_outputs[0] = ccv_nnc_tensor_symbol_copy(graph, updated_parameters[parameter_indice], k);
925			for (j = 0; j < saved_aux_size; j++)
926			{
927				update_inputs[j + 2] = ccv_nnc_tensor_symbol_copy(graph, saved_aux[parameter_indice * max_saved_aux_size + j].source, k);
928				update_outputs[j + 1] = ccv_nnc_tensor_symbol_copy(graph, saved_aux[parameter_indice * max_saved_aux_size + j].destination, k);
929			}
930			ccv_nnc_graph_exec_symbol_set_io(graph, copy, update_inputs, saved_aux_size + 2, update_outputs, saved_aux_size + 1);
931		}
932	}
933	return this_parameter_flag;
934}
935 
936typedef struct {
937	int parameter_size;
938	ccv_nnc_cmd_t minimizer;
939	ccv_cnnp_model_io_t parameters[1];
940} ccv_cnnp_set_minimizer_for_parameter_t;
941 
942static int _ccv_cnnp_apply_parameters_with_minimizer(ccv_cnnp_model_t* const model)
943{
944	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
945	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 945, __extension__ __PRETTY_FUNCTION__); }));
946	const int max_saved_aux_size = compiled_data->minimize.max_saved_aux_size;
947	// We update all parameters, at this point, we have one minimizer.
948	const int parameter_size = compiled_data->parameters->rnum;
949	ccv_nnc_graph_exec_symbol_t* const update_nodes = compiled_data->update_nodes;
950	ccv_nnc_symbolic_graph_t* const symbolic_graph = model->graph;
951	assert(symbolic_graph)((void) sizeof ((symbolic_graph) ? 1 : 0), __extension__ ({ if
 (symbolic_graph) ; else __assert_fail ("symbolic_graph", "ccv_cnnp_model.c"
, 951, __extension__ __PRETTY_FUNCTION__); }));
952	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; });
953	ccv_array_t* const parameters = compiled_data->minimize.parameters;
954	ccv_array_t* const parameter_indices = ccv_array_new(sizeof(int), 0, 0);
955	int i, j, flag = 0;
956	for (i = 0; i < parameters->rnum; i++)
957	{
958		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)));
959		for (j = 0; j < set_minimizer_for_parameter->parameter_size; j++)
960		{
961			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;
962			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", 962, __extension__ __PRETTY_FUNCTION__)
; }));
963			const int old_rnum = parameter_indices->rnum;
964			ccv_cnnp_model_add_to_parameter_indices(set_minimizer_for_parameter->parameters[j]->model, param_sel, parameter_indices);
965			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;
966			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", 966, __extension__ __PRETTY_FUNCTION__)
; }));
967			if (param_ref >= 0)
968			{
969				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", 969, __extension__ __PRETTY_FUNCTION__)
; }));
970				*(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)));
971				parameter_indices->rnum = old_rnum + 1;
972			}
973		}
974		const int saved_aux_size = ccv_nnc_minimizer_saved_aux_size(set_minimizer_for_parameter->minimizer);
975		// We may have duplicated indices, but that is OK, we will set it twice.
976		for (j = 0; j < parameter_indices->rnum; j++)
977		{
978			const int d = *(int*)ccv_array_get(parameter_indices, j)((void*)(((char*)((parameter_indices)->data)) + (size_t)(parameter_indices
)->rsize * (size_t)(j)));
979			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", 979, __extension__ __PRETTY_FUNCTION__)
; }));
980			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))
981				flag = 1;
982		}
983		ccv_array_clear(parameter_indices);
984	}
985	ccv_array_free(parameter_indices);
986	return flag;
987}
988 
989static 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)
990{
991	if (new_saved_aux_size == old_saved_aux_size)
992		return;
993	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", 993, __extension__ __PRETTY_FUNCTION__)
; }));
994	int i, j;
995	for (i = parameter_size - 1; i >= 0; i--)
996	{
997		for (j = new_saved_aux_size - 1; j >= old_saved_aux_size; j--)
998			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
};
999		for (j = old_saved_aux_size - 1; j >= 0; j--)
1000			saved_aux[i * new_saved_aux_size + j] = saved_aux[i * old_saved_aux_size + j];
1001	}
1002}
1003 
1004static void _ccv_cnnp_model_set_rewindables(ccv_cnnp_model_t* const model)
1005{
1006	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1007	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 1007, __extension__ __PRETTY_FUNCTION__); }));
1008	if (!compiled_data->rewindables)
1009		compiled_data->rewindables = ccv_array_new(sizeof(ccv_cnnp_rewind_symbol_t), 0, 0);
1010	ccv_nnc_tensor_symbol_new_hook(model->graph, _ccv_cnnp_model_tensor_symbol_new_hook, compiled_data->rewindables, 0);
1011	ccv_nnc_tensor_symbol_alias_new_hook(model->graph, _ccv_cnnp_model_tensor_symbol_alias_new_hook, compiled_data->rewindables, 0);
1012	ccv_nnc_graph_exec_symbol_new_hook(model->graph, _ccv_cnnp_model_graph_exec_symbol_new_hook, compiled_data->rewindables, 0);
1013}
1014 
1015static 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)
1016{
1017	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1018	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", 1018, __extension__ __PRETTY_FUNCTION__
); }));
1019	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", 1019, __extension__ __PRETTY_FUNCTION__
); }));
1020	const int evaluate_to_size = compiled_data->evaluate.to_size;
1021	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", 1021, __extension__ __PRETTY_FUNCTION__
); }));
1022	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; });
1023	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);
1024	compiled_data->evaluate.to_ops = (ccv_nnc_graph_exec_t*)(compiled_data->evaluate.tos + evaluate_to_size * parallel_count);
1025	int i, j;
1026	const int output_size = model->output_size;
1027	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", 1027, __extension__ __PRETTY_FUNCTION__
); }));
1028	if (fits)
1029		for (i = 0; i < output_size; i++)
1030			ccv_nnc_tensor_symbol_set(model->graph, compiled_data->fits[i], fits[i]->info);
1031	const int max_saved_aux_size = compiled_data->minimize.max_saved_aux_size;
1032	const int parameter_size = compiled_data->parameters->rnum;
1033	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);
1034	compiled_data->update_nodes = (ccv_nnc_graph_exec_symbol_t*)(compiled_data->updated_parameters + parameter_size);
1035	compiled_data->saved_aux = (ccv_nnc_tensor_symbol_map_t*)(compiled_data->update_nodes + parameter_size);
1036	int parameter_size_maybe_more = parameter_size;
1037	compiled_data->disable_outgrad = disable_outgrad;
1038	int outgrad_size;
1039	if (gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES || model->input_size == 0)
1040		outgrad_size = 0;
1041	else if (disable_outgrad == CCV_CNNP_DISABLE_OUTGRAD_NONE) // Compute minimize with gradients including inputs.
1042		outgrad_size = model->input_size;
1043	else {
1044		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", 1044, __extension__ __PRETTY_FUNCTION__
); })); // If it is disable all, gradient mode won't be this.
1045		outgrad_size = 0;
1046		for (i = 0; i < model->input_size; i++)
1047			if (!(disable_outgrad & ((uint64_t)1 << i)))
1048				++outgrad_size;
1049	}
1050	compiled_data->outgrad_size = outgrad_size;
1051	parameter_size_maybe_more += outgrad_size;
1052	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);
1053	compiled_data->outgrads = parameter_size_maybe_more > parameter_size ? compiled_data->gradients + parameter_size : 0;
1054	compiled_data->backward.tos = (ccv_nnc_graph_exec_symbol_t*)(compiled_data->gradients + parameter_size_maybe_more);
1055	compiled_data->backward.to_size = parameter_size_maybe_more;
1056	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)));
1057	if (compiled_data->parameter_flags)
1058	{
1059		parameters = (ccv_nnc_tensor_symbol_t*)ccmallocmalloc(sizeof(ccv_nnc_tensor_symbol_t) * parameter_size);
1060		for (i = 0; i < parameter_size; i++)
1061			if (compiled_data->parameter_flags[i >> 6] & ((uint64_t)1 << (i & 63)))
1062				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)));
1063			else
1064				parameters[i] = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
1065	}
1066	if (gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES || model->input_size == 0)
1067		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);
1068	else if (disable_outgrad == CCV_CNNP_DISABLE_OUTGRAD_NONE) // Compute minimize with gradients including inputs.
1069		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);
1070	else { // Compute minimize with gradients including selected inputs.
1071		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", 1071, __extension__ __PRETTY_FUNCTION__
); }));
1072		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", 1072, __extension__ __PRETTY_FUNCTION__
); })); // If it is disable all, gradient mode won't be this.
1073		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", 1073, __extension__ __PRETTY_FUNCTION__
); }));
1074		ccv_nnc_tensor_symbol_t outgrads[outgrad_size];
1075		j = 0;
1076		for (i = 0; i < model->input_size; i++)
1077			if (!(disable_outgrad & ((uint64_t)1 << i)))
1078				outgrads[j++] = model->inputs[i];
1079		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);
1080	}
1081	if (compiled_data->parameter_flags)
1082		ccfreefree(parameters);
1083	_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);
1084	if (compiled_data->minimize.parameters)
1085		_ccv_cnnp_apply_parameters_with_minimizer(model);
1086	// Go through gradient checkpoints to generate tensor inputs for backward pass just before executing the backward pass.
1087	ccv_cnnp_model_apply_gradient_checkpoints(compiled_data, model->graph);
1088	for (i = 0; i < output_size; i++)
1089	{
1090		const ccv_nnc_tensor_symbol_t df = ccv_nnc_tensor_symbol_for_backward(model->graph, compiled_data->f[i]);
1091		// Init this to 1 so we can backprop.
1092		ccv_nnc_tensor_symbol_set_flags(model->graph, df, CCV_NNC_TENSOR_SYMBOL_INIT_ONES);
1093	}
1094	compiled_data->backward.to_size = 0;
1095	for (i = 0; i < parameter_size_maybe_more; i++)
1096		if (compiled_data->gradients[i].d != CCV_NNC_NO_TENSOR_SYMBOL)
1097			compiled_data->backward.tos[compiled_data->backward.to_size++] = ccv_nnc_graph_exec_symbol_for_backward(model->graph, compiled_data->gradients[i]);
1098	ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS);
1099	ccv_nnc_symbolic_graph_set_destinations(model->graph, compiled_data->update_nodes, parameter_size);
1100	for (i = 0; i < parameter_size_maybe_more - parameter_size; i++)
1101	{
1102		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.
1103			continue;
1104		const ccv_nnc_graph_exec_symbol_t outgrad = ccv_nnc_graph_exec_symbol_for_backward(model->graph, compiled_data->outgrads[i]);
1105		const int* tos;
1106		int to_size;
1107		ccv_nnc_graph_exec_symbol_to(model->graph, outgrad, &tos, &to_size);
1108		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.
1109		{
1110			const ccv_nnc_graph_exec_symbol_t* destinations = ccv_nnc_symbolic_graph_destinations(model->graph);
1111			const int destination_count = ccv_nnc_symbolic_graph_destination_size(model->graph);
1112			int flag = 0;
1113			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; });
1114			for (j = i - 1; !flag && j >= 0; j--)
1115				if (j + outgrad_destination_start < destination_count)
1116					flag = (destinations[j + outgrad_destination_start].d == outgrad.d);
1117			if (!flag) // Only if we cannot find it, we add it.
1118				ccv_nnc_symbolic_graph_add_destination(model->graph, outgrad);
1119		}
1120	}
1121	if (parallel_count > 1)
1122	{
1123		ccv_nnc_symbolic_graph_data_parallel(model->graph, parallel_count,
1124			0, 0,
1125			compiled_data->gradients, parameter_size /* No need to deal with outgrads, we don't allreduce outgrads */,
1126			compiled_data->gradients /* We only care about gradients before allreduce, thus, update our current pointers */,
1127			0, 0, 0,
1128			CCV_NNC_PARALLEL_REDUCE_OP_SUM,
1129			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));
1130		ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
1131		for (i = 0; i < evaluate_to_size; i++)
1132			for (j = 1; j < parallel_count; j++)
1133			{
1134				const ccv_nnc_graph_exec_symbol_t copy = ccv_nnc_graph_exec_symbol_copy(model->graph, compiled_data->evaluate.tos[i], j);
1135				if (copy.d != CCV_NNC_NO_GRAPH_EXEC_SYMBOL)
1136					compiled_data->evaluate.tos[compiled_data->evaluate.to_size++] = copy;
1137			}
1138		const int backward_to_size = compiled_data->backward.to_size;
1139		for (i = 0; i < backward_to_size; i++)
1140			for (j = 1; j < parallel_count; j++)
1141			{
1142				const ccv_nnc_graph_exec_symbol_t copy = ccv_nnc_graph_exec_symbol_copy(model->graph, compiled_data->backward.tos[i], j);
1143				if (copy.d != CCV_NNC_NO_GRAPH_EXEC_SYMBOL)
1144					compiled_data->backward.tos[compiled_data->backward.to_size++] = copy;
1145			}
1146	}
1147	// Only use memory compression if we are in gradient parameter mode.
1148	if (gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES || gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES_AND_INPUTS)
1149	{
1150		if (model->memory_compression)
1151			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));
1152		if (model->memory_reduction)
1153			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));
1154	}
1155	compiled_data->backward.to_size = _ccv_nnc_array_dedup_graph_exec_symbols(compiled_data->backward.tos, compiled_data->backward.to_size);
1156	compiled_data->gradient_mode = gradient_mode;
1157}
1158 
1159void ccv_cnnp_model_tensors_init_0(const ccv_cnnp_model_t* const model, ccv_cnnp_compiled_data_t* const compiled_data)
1160{
1161	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", 1161, __extension__ __PRETTY_FUNCTION__
); }));
1162	const int parameter_size = compiled_data->parameters->rnum;
1163	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; });
1164	const int internal_size = compiled_data->internals->rnum;
1165	compiled_data->tensors_init.size = ccv_nnc_tensor_symbol_count(model->graph);
1166	compiled_data->tensors_init.v = cccalloccalloc(((compiled_data->tensors_init.size + 31) >> 5), sizeof(uint32_t));
1167	compiled_data->tensors.parameters = (ccv_nnc_tensor_t**)cccalloccalloc((parameter_size + internal_size) * parallel_count, sizeof(ccv_nnc_tensor_t*));
1168	compiled_data->tensors.internals = compiled_data->tensors.parameters + parameter_size * parallel_count;
1169}
1170 
1171int ccv_cnnp_model_tensors_any_to_alloc(const ccv_cnnp_model_t* const model, ccv_cnnp_compiled_data_t* const compiled_data)
1172{
1173	int i, j;
1174	const int parameter_size = compiled_data->parameters->rnum;
1175	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; });
1176	const int internal_size = compiled_data->internals->rnum;
1177	for (i = 0; i < parameter_size; i++)
1178	{
1179		// parameters has to be allocated all together.
1180		if (compiled_data->tensors.parameters[i])
1181		{
1182			for (j = 1; j < parallel_count; j++)
1183				{ 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", 1183, __extension__ __PRETTY_FUNCTION__
); })); }
1184			continue;
1185		}
1186		return 1;
1187	}
1188	for (i = 0; i < internal_size; i++)
1189	{
1190		if (!compiled_data->tensors.internals[i])
1191			return 1;
1192		for (j = 1; j < parallel_count; j++)
1193			if (!compiled_data->tensors.internals[i + j * internal_size])
1194				return 1;
1195	}
1196	return 0;
1197}
1198 
1199void ccv_cnnp_model_tensors_init_1(const ccv_cnnp_model_t* const model, ccv_cnnp_compiled_data_t* const compiled_data)
1200{
1201	int i, j;
1202	const int parameter_size = compiled_data->parameters->rnum;
1203	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; });
1204	const int internal_size = compiled_data->internals->rnum;
1205	for (i = 0; i < parameter_size; i++)
1206	{
1207		// parameters has to be allocated all together.
1208		if (compiled_data->tensors.parameters[i])
1209		{
1210			for (j = 1; j < parallel_count; j++)
1211				{ 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", 1211, __extension__ __PRETTY_FUNCTION__
); })); }
1212			continue;
1213		}
1214		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)));
1215		ccv_nnc_tensor_param_t info = ccv_nnc_tensor_symbol_params(parameter.graph, parameter);
1216		if (CCV_TENSOR_GET_DEVICE(info.type)((info.type) & 0xfff00) == CCV_COMPUTE_DEVICE_ANY)
1217			CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff
) << 8));
1218		const int device_id = CCV_TENSOR_GET_DEVICE_ID(info.type)(((info.type) & 0xfff00) >> 8);
1219		compiled_data->tensors.parameters[i] = ccv_nnc_tensor_new(0, info, 0);
1220		for (j = 1; j < parallel_count; j++)
1221		{
1222			if (j != device_id)
1223				CCV_TENSOR_SET_DEVICE_ID(info.type, j)(info.type) = (((info.type) & ~0xfff00) | (((j) & 0xfff
) << 8));
1224			else
1225				CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff
) << 8));
1226			compiled_data->tensors.parameters[i + j * parameter_size] = ccv_nnc_tensor_new(0, info, 0);
1227		}
1228	}
1229	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));
1230	for (i = 0; i < internal_size; i++)
1231	{
1232		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))
);
1233		const int d = retained.d;
1234		if (init_v[d >> 5] & (1u << (d & 0x1f)))
1235			continue;
1236		ccv_nnc_tensor_param_t info = ccv_nnc_tensor_symbol_params(retained.graph, retained);
1237		if (CCV_TENSOR_GET_DEVICE(info.type)((info.type) & 0xfff00) == CCV_COMPUTE_DEVICE_ANY)
1238			CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff
) << 8));
1239		const int device_id = CCV_TENSOR_GET_DEVICE_ID(info.type)(((info.type) & 0xfff00) >> 8);
1240		if (!compiled_data->tensors.internals[i])
1241			compiled_data->tensors.internals[i] = ccv_nnc_tensor_new(0, info, 0);
1242		for (j = 1; j < parallel_count; j++)
1243		{
1244			if (j != device_id)
1245				CCV_TENSOR_SET_DEVICE_ID(info.type, j)(info.type) = (((info.type) & ~0xfff00) | (((j) & 0xfff
) << 8));
1246			else
1247				CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff
) << 8));
1248			if (!compiled_data->tensors.internals[i + j * internal_size])
1249				compiled_data->tensors.internals[i + j * internal_size] = ccv_nnc_tensor_new(0, info, 0);
1250		}
1251	}
1252	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.
1253}
1254 
1255static void _ccv_cnnp_model_tensors_init(const ccv_cnnp_model_t* const model, ccv_cnnp_compiled_data_t* const compiled_data)
1256{
1257	ccv_cnnp_model_tensors_init_0(model, compiled_data);
1258	ccv_cnnp_model_tensors_init_1(model, compiled_data);
1259}
1260 
1261static 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)
1262{
1263	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", 1263, __extension__ __PRETTY_FUNCTION__
); }));
1264	int i, j;
1265	for (i = 0; i < tensor_size; i++)
1266	{
1267		if (!tensors[i])
1268			continue;
1269		const int d = tensor_symbols[i].d;
1270		if (!(tensors_init[d >> 5] & (1u << (d & 0x1f))))
1271			continue;
1272		for (j = 1; j < parallel_count; j++)
1273			if (tensors[i + j * tensor_size])
1274			{
1275				ccv_nnc_tensor_t* const input = CCV_NNC_TENSOR(tensors[i])((ccv_nnc_tensor_t*)((uintptr_t)(tensors[i]) & ~(uintptr_t
)1));
1276				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));
1277				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);
1278			}
1279	}
1280}
1281 
1282static 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)
1283{
1284	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", 1284, __extension__ __PRETTY_FUNCTION__
); }));
1285	int i, j;
1286	for (i = 0; i < tensor_size; i++)
1287	{
1288		const ccv_nnc_tensor_symbol_t tensor_symbol = tensor_symbols[i];
1289		for (j = 1; j < parallel_count; j++)
1290		{
1291			const ccv_nnc_tensor_symbol_t copy = ccv_nnc_tensor_symbol_copy(graph, tensor_symbol, j);
1292			ccv_nnc_tensor_t* copy_tensor = tensors[i + j * tensor_size];
1293			if (copy_tensor && copy.d == CCV_NNC_NO_TENSOR_SYMBOL)
1294			{ // We shouldn't allocate this, free it up.
1295				ccv_nnc_tensor_free(tensors[i + j * tensor_size]);
1296				tensors[i + j * tensor_size] = 0;
1297			}
1298		}
1299	}
1300}
1301 
1302static 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)
1303{
1304	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", 1304, __extension__ __PRETTY_FUNCTION__
); }));
1305	int i, j;
1306	for (i = 0; i < tensor_size; i++)
1307	{
1308		ccv_nnc_tensor_symbol_t tensor_symbol = tensor_symbols[i];
1309		if (tensor_symbol.d == CCV_NNC_NO_TENSOR_SYMBOL)
1310			continue;
1311		if (graph)
1312		{
1313			const ccv_nnc_tensor_symbol_t alias_to = ccv_nnc_tensor_symbol_alias_to(graph, tensor_symbol);
1314			if (alias_to.d != CCV_NNC_NO_TENSOR_SYMBOL)
1315				tensor_symbol = alias_to;
1316		}
1317		ccv_nnc_tensor_t* const tensor = CCV_NNC_TENSOR(tensors[i])((ccv_nnc_tensor_t*)((uintptr_t)(tensors[i]) & ~(uintptr_t
)1));
1318		if (tensor && tensor_symbol.d != CCV_NNC_NO_TENSOR_SYMBOL)
1319		{
1320			const ccv_nnc_tensor_bind_t retained_bind = {
1321				.symbol = tensor_symbol,
1322				.tensor = tensor
1323			};
1324			ccv_array_push(tensor_binds, &retained_bind);
1325		}
1326		for (j = 1; j < parallel_count; j++)
1327		{
1328			const ccv_nnc_tensor_symbol_t copy = ccv_nnc_tensor_symbol_copy(graph, tensor_symbol, j);
1329			ccv_nnc_tensor_t* copy_tensor = tensors[i + j * tensor_size];
1330			if (copy_tensor && copy.d != CCV_NNC_NO_TENSOR_SYMBOL)
1331			{
1332				const ccv_nnc_tensor_bind_t bind = {
1333					.symbol = copy,
1334					.tensor = tensors[i + j * tensor_size]
1335				};
1336				ccv_array_push(tensor_binds, &bind);
1337			}
1338		}
1339	}
1340}
1341 
1342static void _ccv_cnnp_compiled_data_graph_free(ccv_cnnp_compiled_data_t* const compiled_data)
1343{
1344	if (compiled_data->graph)
1345		ccv_nnc_graph_free(compiled_data->graph);
1346	compiled_data->graph = 0;
1347	compiled_data->is_test = 0;
1348	if (compiled_data->tensor_arena)
1349		ccv_nnc_tensor_arena_free(compiled_data->tensor_arena);
1350	compiled_data->tensor_arena = 0;
1351	if (compiled_data->graph_exec_arena)
1352		ccv_nnc_graph_exec_arena_free(compiled_data->graph_exec_arena);
1353	compiled_data->graph_exec_arena = 0;
1354	if (compiled_data->backward.from_ops)
1355		ccfreefree(compiled_data->backward.from_ops);
1356	compiled_data->backward.from_ops = 0;
1357	if (compiled_data->evaluate.schedule)
1358		ccv_nnc_graph_static_schedule_free(compiled_data->evaluate.schedule);
1359	compiled_data->evaluate.schedule = 0;
1360	if (compiled_data->backward.schedule)
1361		ccv_nnc_graph_static_schedule_free(compiled_data->backward.schedule);
1362	compiled_data->backward.schedule = 0;
1363}
1364 
1365static void _ccv_cnnp_compiled_data_gradient_free(ccv_cnnp_compiled_data_t* const compiled_data)
1366{
1367	if (compiled_data->gradients)
1368		ccfreefree(compiled_data->gradients);
1369	compiled_data->gradients = 0;
1370	if (compiled_data->updated_parameters)
1371		ccfreefree(compiled_data->updated_parameters);
1372	compiled_data->updated_parameters = 0;
1373	compiled_data->update_nodes = 0;
1374	compiled_data->saved_aux = 0;
1375}
1376 
1377static void _ccv_cnnp_compiled_data_backward_free(ccv_cnnp_compiled_data_t* const compiled_data)
1378{
1379	if (compiled_data->backward.gradients)
1380		ccfreefree(compiled_data->backward.gradients);
1381	compiled_data->backward.gradients = 0;
1382	if (compiled_data->backward.accum)
1383		ccv_nnc_graph_free(compiled_data->backward.accum);
1384	compiled_data->backward.accum = 0;
1385	if (compiled_data->backward.tensor_arena)
1386		ccv_nnc_tensor_arena_free(compiled_data->backward.tensor_arena);
1387	compiled_data->backward.tensor_arena = 0;
1388	if (compiled_data->backward.graph_exec_arena)
1389		ccv_nnc_graph_exec_arena_free(compiled_data->backward.graph_exec_arena);
1390	compiled_data->backward.graph_exec_arena = 0;
1391}
1392 
1393static void _ccv_cnnp_compiled_data_apply_gradients_free(ccv_cnnp_compiled_data_t* const compiled_data)
1394{
1395	if (compiled_data->apply_gradients.graph)
1396		ccv_nnc_graph_free(compiled_data->apply_gradients.graph);
1397	compiled_data->apply_gradients.graph = 0;
1398	if (compiled_data->apply_gradients.tensor_arena)
1399		ccv_nnc_tensor_arena_free(compiled_data->apply_gradients.tensor_arena);
1400	compiled_data->apply_gradients.tensor_arena = 0;
1401	if (compiled_data->apply_gradients.graph_exec_arena)
1402		ccv_nnc_graph_exec_arena_free(compiled_data->apply_gradients.graph_exec_arena);
1403	compiled_data->apply_gradients.graph_exec_arena = 0;
1404}
1405 
1406// Compile the graph to run ccv_cnnp_model_fit
1407static 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)
1408{
1409	int i, j;
1410	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1411	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", 1411, __extension__ __PRETTY_FUNCTION__
); }));
1412	compiled_data->graph_mode = CCV_CNNP_MODEL_GRAPH_FIT_MODE;
1413	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; });
1414	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", 1414, __extension__ __PRETTY_FUNCTION__
); }));
1415	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", 1415
, __extension__ __PRETTY_FUNCTION__); }));
1416	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", 1416, __extension__ __PRETTY_FUNCTION__
); }));
1417	if (compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_NONE)
1418	{
1419		_ccv_cnnp_model_set_rewindables(model);
1420		_ccv_cnnp_model_gradient_init(model, CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES, CCV_CNNP_DISABLE_OUTGRAD_ALL, fits, fit_size);
1421	} else if (compiled_data->gradient_mode != CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES) {
1422		_ccv_cnnp_model_rewind_graph(model);
1423		_ccv_cnnp_compiled_data_gradient_free(compiled_data);
1424		compiled_data->gradient_mode = CCV_CNNP_COMPILED_DATA_GRADIENT_NONE;
1425		_ccv_cnnp_model_gradient_init(model, CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES, CCV_CNNP_DISABLE_OUTGRAD_ALL, fits, fit_size);
1426	}
1427	const int tensors_init = !!compiled_data->tensors_init.v;
1428	if (!tensors_init)
1429		_ccv_cnnp_model_tensors_init(model, compiled_data);
1430	else if ((uintptr_t)compiled_data->tensors_init.v & (uintptr_t)1)
1431	// Check if it is not fully allocated, if it is not, init_1.
1432		ccv_cnnp_model_tensors_init_1(model, compiled_data);
1433	ccv_array_t* const tensor_binds = ccv_array_new(sizeof(ccv_nnc_tensor_bind_t), 0, 0);
1434	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"
, 1434, __extension__ __PRETTY_FUNCTION__); }));
1435	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"
, 1435, __extension__ __PRETTY_FUNCTION__); }));
1436	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", 1436
, __extension__ __PRETTY_FUNCTION__); }));
1437	const int input_size_per_p = input_size / parallel_count;
1438	_ccv_cnnp_model_bind_tensors(model->graph, model->inputs, inputs, input_size_per_p, parallel_count, tensor_binds);
1439	const int output_size_per_p = output_size / parallel_count;
1440	_ccv_cnnp_model_bind_tensors(model->graph, model->outputs, outputs, output_size_per_p, parallel_count, tensor_binds);
1441	const int fit_size_per_p = fit_size / parallel_count;
1442	_ccv_cnnp_model_bind_tensors(model->graph, compiled_data->fits, fits, fit_size_per_p, parallel_count, tensor_binds);
1443	const int parameter_size = compiled_data->parameters->rnum;
1444	_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);
1445	_ccv_cnnp_model_bind_tensors(model->graph, compiled_data->updated_parameters, compiled_data->tensors.parameters, parameter_size, parallel_count, tensor_binds);
1446	const int internal_size = compiled_data->internals->rnum;
1447	_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);
1448	_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);
1449	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);
1450	ccv_array_free(tensor_binds);
1451	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));
1452	if (tensors_init && parallel_count > 1)
1453		_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);
1454	// If tensor is not init'ed, we need to init states first.
1455	if (_ccv_cnnp_any_to_init(compiled_data))
1456	{
1457		ccv_nnc_tensor_init_states_t tensor_init_states = {
1458			.parallel_count = parallel_count,
1459			.graph = model->graph,
1460			.compiled_data = compiled_data,
1461			.tensor_arena = compiled_data->tensor_arena
1462		};
1463		ccv_cnnp_model_init_states(model, model->graph, _ccv_cnnp_init_states_for_tensors, &tensor_init_states);
1464	}
1465	compiled_data->is_test = 0;
1466	const int saved_aux_size = ccv_nnc_minimizer_saved_aux_size(compiled_data->minimize.minimizer);
1467	// No need to set because it is default to training mode.
1468	// ccv_cnnp_model_set_is_test(model, 0, _ccv_cnnp_cmd_update_for_execs, &update);
1469	for (i = 0; i < saved_aux_size * parameter_size; i++)
1470	{
1471		if (compiled_data->saved_aux[i].source.d == CCV_NNC_NO_TENSOR_SYMBOL)
1472			continue;
1473		ccv_nnc_tensor_t* const tensor = ccv_nnc_tensor_from_symbol(compiled_data->tensor_arena, compiled_data->saved_aux[i].source);
1474		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);
1475		for (j = 1; j < parallel_count; j++)
1476		{
1477			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));
1478			if (copy)
1479				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);
1480		}
1481	}
1482	const int evaluate_to_size = compiled_data->evaluate.to_size;
1483	compiled_data->evaluate.to_op_size = 0;
1484	for (i = 0; i < evaluate_to_size; i++)
1485	{
1486		ccv_nnc_graph_exec_t const to = ccv_nnc_graph_exec_from_symbol(compiled_data->graph_exec_arena, compiled_data->evaluate.tos[i]);
1487		if (to.graph)
1488			compiled_data->evaluate.to_ops[compiled_data->evaluate.to_op_size++] = to;
1489	}
1490	ccv_nnc_graph_set_default_static_schedule(compiled_data->graph, compiled_data->stream_type, model->max_stream_count);
1491	ccv_nnc_graph_autotune(compiled_data->graph, model->workspace_size, 0, TRAVERSE_FULL0,0,0,0);
1492}
1493 
1494ccv_nnc_stream_context_t* ccv_cnnp_model_default_stream(const ccv_cnnp_model_t* const model)
1495{
1496	const ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1497	if (!compiled_data || !compiled_data->graph)
1498		return 0;
1499	return ccv_nnc_graph_default_stream(compiled_data->graph);
1500}
1501 
1502uint64_t ccv_cnnp_model_memory_size(const ccv_cnnp_model_t* const model)
1503{
1504	const ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1505	if (!compiled_data || !compiled_data->tensor_arena)
1506		return 0;
1507	return ccv_nnc_tensor_arena_size(compiled_data->tensor_arena);
1508}
1509 
1510static 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)
1511{
1512	int i, j;
1513	for (i = 0; i < tensor_size; i++)
1514	{
1515		ccv_nnc_tensor_symbol_t tensor_symbol = tensor_symbols[i];
1516		if (tensor_symbol.d == CCV_NNC_NO_TENSOR_SYMBOL)
1517			continue;
1518		if (graph)
1519		{
1520			const ccv_nnc_tensor_symbol_t alias_to = ccv_nnc_tensor_symbol_alias_to(graph, tensor_symbol);
1521			if (alias_to.d != CCV_NNC_NO_TENSOR_SYMBOL)
1522				tensor_symbol = alias_to;
1523		}
1524		ccv_nnc_tensor_bind_symbol(tensor_arena, tensor_symbol, tensors[i]);
1525		for (j = 1; j < parallel_count; j++)
1526		{
1527			const ccv_nnc_tensor_symbol_t copy = ccv_nnc_tensor_symbol_copy(graph, tensor_symbol, j);
1528			if (copy.d != CCV_NNC_NO_TENSOR_SYMBOL)
1529				ccv_nnc_tensor_bind_symbol(tensor_arena, copy, tensors[i + tensor_size * j]);
1530		}
1531	}
1532}
1533 
1534void 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)
1535{
1536	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1537	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 1537, __extension__ __PRETTY_FUNCTION__); }));
1538	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; });
1539	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", 1539, __extension__ __PRETTY_FUNCTION__
); }));
1540	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", 1540, __extension__ __PRETTY_FUNCTION__
); }));
1541	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", 1541
, __extension__ __PRETTY_FUNCTION__); }));
1542	assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if
 (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c"
, 1542, __extension__ __PRETTY_FUNCTION__); }));
1543	if (!compiled_data->graph || compiled_data->graph_mode != CCV_CNNP_MODEL_GRAPH_FIT_MODE)
1544	{
1545		_ccv_cnnp_compiled_data_graph_free(compiled_data);
1546		_ccv_cnnp_compiled_data_backward_free(compiled_data);
1547		_ccv_cnnp_compiled_data_apply_gradients_free(compiled_data);
1548		// Compile the symbolic graph down only when needed.
1549		_ccv_cnnp_model_fit_jit(model, inputs, input_size, fits, fit_size, outputs, output_size);
1550	} else {
1551		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"
, 1551, __extension__ __PRETTY_FUNCTION__); }));
1552		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"
, 1552, __extension__ __PRETTY_FUNCTION__); }));
1553		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", 1553
, __extension__ __PRETTY_FUNCTION__); }));
1554		const int input_size_per_p = input_size / parallel_count;
1555		_ccv_cnnp_bind_tensors_to_arena(compiled_data->tensor_arena, model->graph, model->inputs, inputs, input_size_per_p, parallel_count);
1556		const int output_size_per_p = output_size / parallel_count;
1557		_ccv_cnnp_bind_tensors_to_arena(compiled_data->tensor_arena, model->graph, model->outputs, outputs, output_size_per_p, parallel_count);
1558		const int fit_size_per_p = fit_size / parallel_count;
1559		_ccv_cnnp_bind_tensors_to_arena(compiled_data->tensor_arena, model->graph, compiled_data->fits, fits, fit_size_per_p, parallel_count);
1560	}
1561	if (compiled_data->is_test)
1562	{
1563		compiled_data->is_test = 0;
1564		ccv_nnc_graph_exec_update_t update = {
1565			.parallel_count = parallel_count,
1566			.graph = model->graph,
1567			.graph_exec_arena = compiled_data->graph_exec_arena,
1568		};
1569		ccv_cnnp_model_set_is_test(model, 0, _ccv_cnnp_cmd_update_for_execs, &update);
1570	}
1571	ccv_nnc_graph_run_with_schedule(compiled_data->graph, 0, 0, tensor_tape, stream_context);
1572}
1573 
1574// Compile the graph to run ccv_cnnp_model_evaluate with require_grad = false (MULTISTAGE_MODE_NO_GRAD).
1575static 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)
1576{
1577	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1578	compiled_data->graph_mode = CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE_NO_GRAD;
1579	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; });
1580	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", 1580, __extension__ __PRETTY_FUNCTION__
); }));
1581	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", 1581, __extension__ __PRETTY_FUNCTION__
); }));
1582	// If the gradient is not initialized, continue to setup parallel process. We don't init gradient here, but rather,
1583	// we setup proper rewindables so the graph can be rewinded to previous state before we run data parallel.
1584	if (parallel_count > 1 && compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_NONE)
1585	{
1586		const int evaluate_to_size = compiled_data->evaluate.to_size;
1587		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);
1588		_ccv_cnnp_model_set_rewindables(model);
1589		ccv_nnc_symbolic_graph_data_parallel(model->graph, parallel_count,
1590			0, 0,
1591			0, 0, 0,
1592			0, 0, 0,
1593			CCV_NNC_PARALLEL_REDUCE_OP_SUM,
1594			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));
1595		ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
1596		int i, j;
1597		for (i = 0; i < evaluate_to_size; i++)
1598			for (j = 1; j < parallel_count; j++)
1599			{
1600				const ccv_nnc_graph_exec_symbol_t copy = ccv_nnc_graph_exec_symbol_copy(model->graph, compiled_data->evaluate.tos[i], j);
1601				if (copy.d != CCV_NNC_NO_GRAPH_EXEC_SYMBOL)
1602					compiled_data->evaluate.tos[compiled_data->evaluate.to_size++] = copy;
1603			}
1604	}
1605	const int tensors_init = !!compiled_data->tensors_init.v;
1606	if (!tensors_init)
1607		_ccv_cnnp_model_tensors_init(model, compiled_data);
1608	else if ((uintptr_t)compiled_data->tensors_init.v & (uintptr_t)1)
1609	// Check if it is not fully allocated, if it is not, init_1.
1610		ccv_cnnp_model_tensors_init_1(model, compiled_data);
1611	ccv_array_t* const tensor_binds = ccv_array_new(sizeof(ccv_nnc_tensor_bind_t), 0, 0);
1612	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"
, 1612, __extension__ __PRETTY_FUNCTION__); }));
1613	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"
, 1613, __extension__ __PRETTY_FUNCTION__); }));
1614	const int input_size_per_p = input_size / parallel_count;
1615	_ccv_cnnp_model_bind_tensors(model->graph, model->inputs, inputs, input_size_per_p, parallel_count, tensor_binds);
1616	const int output_size_per_p = output_size / parallel_count;
1617	_ccv_cnnp_model_bind_tensors(model->graph, model->outputs, outputs, output_size_per_p, parallel_count, tensor_binds);
1618	const int parameter_size = compiled_data->parameters->rnum;
1619	_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);
1620	const int internal_size = compiled_data->internals->rnum;
1621	_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);
1622	_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);
1623	// If we generated gradient for the graph, only compile part of the graph because the rest is irrelevant for evaluation.
1624	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);
1625	ccv_array_free(tensor_binds);
1626	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));
1627	// If tensor is not init'ed, we need to init states first.
1628	if (tensors_init && parallel_count > 1)
1629		_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);
1630	if (_ccv_cnnp_any_to_init(compiled_data))
1631	{
1632		ccv_nnc_tensor_init_states_t tensor_init_states = {
1633			.parallel_count = parallel_count,
1634			.graph = model->graph,
1635			.compiled_data = compiled_data,
1636			.tensor_arena = compiled_data->tensor_arena
1637		};
1638		ccv_cnnp_model_init_states(model, model->graph, _ccv_cnnp_init_states_for_tensors, &tensor_init_states);
1639	}
1640	compiled_data->is_test = 1;
1641	ccv_nnc_graph_exec_update_t update = {
1642		.parallel_count = parallel_count,
1643		.graph = model->graph,
1644		.graph_exec_arena = compiled_data->graph_exec_arena,
1645	};
1646	ccv_cnnp_model_set_is_test(model, 1, _ccv_cnnp_cmd_update_for_execs, &update);
1647	ccv_nnc_graph_set_default_static_schedule(compiled_data->graph, compiled_data->stream_type, model->max_stream_count);
1648	ccv_nnc_graph_autotune(compiled_data->graph, model->workspace_size, 0, TRAVERSE_FULL0,0,0,0);
1649}
1650 
1651static void _ccv_cnnp_model_gradient_tensors_init(const ccv_cnnp_model_t* const model, ccv_cnnp_compiled_data_t* const compiled_data)
1652{
1653	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", 1653, __extension__ __PRETTY_FUNCTION__
); }));
1654	const int parameter_size = compiled_data->parameters->rnum;
1655	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; });
1656	compiled_data->tensors.gradients = (ccv_nnc_tensor_t**)ccmallocmalloc(sizeof(ccv_nnc_tensor_t*) * parameter_size * 2 * parallel_count);
1657	compiled_data->tensors.accum_gradients = compiled_data->tensors.gradients + parameter_size * parallel_count;
1658	int i, j;
1659	for (i = 0; i < parameter_size; i++)
1660	{
1661		if (compiled_data->parameter_flags && !(compiled_data->parameter_flags[i >> 6] & ((uint64_t)1 << (i & 63))))
1662		{
1663			compiled_data->tensors.gradients[i] = 0;
1664			compiled_data->tensors.accum_gradients[i] = 0;
1665			for (j = 1; j < parallel_count; j++)
1666			{
1667				compiled_data->tensors.gradients[i + j * parameter_size] = 0;
1668				compiled_data->tensors.accum_gradients[i + j * parameter_size] = 0;
1669			}
1670			continue;
1671		}
1672		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)));
1673		ccv_nnc_tensor_param_t info = ccv_nnc_tensor_symbol_params(parameter.graph, parameter);
1674		if (CCV_TENSOR_GET_DEVICE(info.type)((info.type) & 0xfff00) == CCV_COMPUTE_DEVICE_ANY)
1675			CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff
) << 8));
1676		const int device_id = CCV_TENSOR_GET_DEVICE_ID(info.type)(((info.type) & 0xfff00) >> 8);
1677		compiled_data->tensors.gradients[i] = ccv_nnc_tensor_new(0, info, 0);
1678		compiled_data->tensors.accum_gradients[i] = 0; // delay the accumulated gradient allocation until when we need it.
1679		for (j = 1; j < parallel_count; j++)
1680		{
1681			if (j != device_id)
1682				CCV_TENSOR_SET_DEVICE_ID(info.type, j)(info.type) = (((info.type) & ~0xfff00) | (((j) & 0xfff
) << 8));
1683			else
1684				CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff
) << 8));
1685			compiled_data->tensors.gradients[i + j * parameter_size] = ccv_nnc_tensor_new(0, info, 0);
1686			compiled_data->tensors.accum_gradients[i + j * parameter_size] = 0;
1687		}
1688	}
1689}
1690 
1691static int _ccv_cnnp_is_disable_outgrad_all(const uint64_t disable_outgrad, const int input_size)
1692{
1693	if (disable_outgrad == CCV_CNNP_DISABLE_OUTGRAD_ALL)
1694		return 1;
1695	if (disable_outgrad == CCV_CNNP_DISABLE_OUTGRAD_NONE)
1696		return 0;
1697	int i;
1698	for (i = 0; i < input_size; i++)
1699		if (!(disable_outgrad & ((uint64_t)1 << i)))
1700			return 0;
1701	return 1;
1702}
1703 
1704// Compile the graph to run ccv_cnnp_model_evaluate with requires_grad = true (MULTISTAGE_MODE).
1705// Particularly, this method compiles the evaluation and backprop graph (the main graph).
1706static 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)
1707{
1708	int i, j;
1709	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1710	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;
1711	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", 1711, __extension__ __PRETTY_FUNCTION__
); }));
1712	compiled_data->graph_mode = CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE;
1713	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; });
1714	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", 1714, __extension__ __PRETTY_FUNCTION__
); }));
1715	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", 1715, __extension__ __PRETTY_FUNCTION__
); }));
1716	// There shouldn't be a loss function if we evaluate with multistage jit.
1717	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", 1717, __extension__ __PRETTY_FUNCTION__
); }));
1718	if (compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_NONE)
1719	{
1720		_ccv_cnnp_model_set_rewindables(model);
1721		_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.
1722	} else if (compiled_data->gradient_mode != target_gradient_mode) {
1723		_ccv_cnnp_model_rewind_graph(model);
1724		_ccv_cnnp_compiled_data_gradient_free(compiled_data);
1725		compiled_data->gradient_mode = CCV_CNNP_COMPILED_DATA_GRADIENT_NONE;
1726		_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.
1727	}
1728	const int tensors_init = !!compiled_data->tensors_init.v;
1729	if (!tensors_init)
1730		_ccv_cnnp_model_tensors_init(model, compiled_data);
1731	else if ((uintptr_t)compiled_data->tensors_init.v & (uintptr_t)1)
1732	// Check if it is not fully allocated, if it is not, init_1.
1733		ccv_cnnp_model_tensors_init_1(model, compiled_data);
1734	ccv_array_t* const tensor_binds = ccv_array_new(sizeof(ccv_nnc_tensor_bind_t), 0, 0);
1735	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"
, 1735, __extension__ __PRETTY_FUNCTION__); }));
1736	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"
, 1736, __extension__ __PRETTY_FUNCTION__); }));
1737	const int input_size_per_p = input_size / parallel_count;
1738	_ccv_cnnp_model_bind_tensors(model->graph, model->inputs, inputs, input_size_per_p, parallel_count, tensor_binds);
1739	const int output_size_per_p = output_size / parallel_count;
1740	_ccv_cnnp_model_bind_tensors(model->graph, model->outputs, outputs, output_size_per_p, parallel_count, tensor_binds);
1741	const int parameter_size = compiled_data->parameters->rnum;
1742	_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);
1743	const int internal_size = compiled_data->internals->rnum;
1744	_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);
1745	_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);
1746	if (!compiled_data->tensors.gradients)
1747		_ccv_cnnp_model_gradient_tensors_init(model, compiled_data);
1748	_ccv_cnnp_model_bind_tensors(model->graph, compiled_data->gradients, compiled_data->tensors.gradients, parameter_size, parallel_count, tensor_binds);
1749	if (compiled_data->backward.to_size > 0)
1750		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);
1751	else
1752		ccv_nnc_symbolic_graph_compile(model->graph, compiled_data->compile_params, (ccv_nnc_tensor_bind_t*)ccv_array_get(tensor_binds, 0)((void*)(((char*)((tensor_binds)->data)) + (size_t)(tensor_binds
)->rsize * (size_t)(0))), tensor_binds->rnum, 0, 0, SYMBOLIC_GRAPH_SOURCES(model->graph)ccv_nnc_symbolic_graph_sources(model->graph), ccv_nnc_symbolic_graph_source_size
(model->graph), compiled_data->evaluate.tos, compiled_data->evaluate.to_size, &compiled_data->graph, &compiled_data->tensor_arena, &compiled_data->graph_exec_arena);
1753	ccv_array_free(tensor_binds);
1754	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));
1755	if (tensors_init && parallel_count > 1)
1756		_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);
1757	// If tensor is not init'ed, we need to init states first.
1758	if (_ccv_cnnp_any_to_init(compiled_data))
1759	{
1760		ccv_nnc_tensor_init_states_t tensor_init_states = {
1761			.parallel_count = parallel_count,
1762			.graph = model->graph,
1763			.compiled_data = compiled_data,
1764			.tensor_arena = compiled_data->tensor_arena
1765		};
1766		ccv_cnnp_model_init_states(model, model->graph, _ccv_cnnp_init_states_for_tensors, &tensor_init_states);
1767	}
1768	compiled_data->is_test = is_test;
1769	ccv_nnc_graph_exec_update_t update = {
1770		.parallel_count = parallel_count,
1771		.graph = model->graph,
1772		.graph_exec_arena = compiled_data->graph_exec_arena,
1773	};
1774	ccv_cnnp_model_set_is_test(model, is_test, _ccv_cnnp_cmd_update_for_execs, &update);
1775	const int evaluate_to_size = compiled_data->evaluate.to_size;
1776	compiled_data->evaluate.to_op_size = 0;
1777	ccv_array_t* const backward_from = ccv_array_new(sizeof(int), 0, 0);
1778	for (i = 0; i < evaluate_to_size; i++)
1779	{
1780		ccv_nnc_graph_exec_t const to_op = ccv_nnc_graph_exec_from_symbol(compiled_data->graph_exec_arena, compiled_data->evaluate.tos[i]);
1781		if (to_op.graph)
1782			compiled_data->evaluate.to_ops[compiled_data->evaluate.to_op_size++] = to_op;
1783		const int* tos;
1784		int to_size;
1785		ccv_nnc_graph_exec_symbol_to(model->graph, compiled_data->evaluate.tos[i], &tos, &to_size);
1786		for (j = 0; j < to_size; j++)
1787		{
1788			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){
1789				.d = tos[j],
1790				.graph = model->graph
1791			});
1792			if (to_op.graph)
1793				ccv_array_add_unique_int(backward_from, to_op.d);
1794		}
1795	}
1796	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", 1796, __extension__
 __PRETTY_FUNCTION__); }));
1797	compiled_data->backward.from_op_size = backward_from->rnum;
1798	compiled_data->backward.from_ops = (ccv_nnc_graph_exec_t*)ccmallocmalloc(sizeof(ccv_nnc_graph_exec_t) * backward_from->rnum);
1799	for (i = 0; i < backward_from->rnum; i++)
1800		compiled_data->backward.from_ops[i] = (ccv_nnc_graph_exec_t){
1801			.d = *(int*)ccv_array_get(backward_from, i)((void*)(((char*)((backward_from)->data)) + (size_t)(backward_from
)->rsize * (size_t)(i))),
1802			.graph = compiled_data->graph,
1803		};
1804	// If there are any set node (to set some tensors to 0) inserted through backward pass, these won't be executed if we just do sources -> evaluate.to_ops, backward.from_ops -> destinations. We need this logic to find out these nodes and explicitly adding them to backward.from_ops.
1805	ccv_nnc_graph_exec_info_t* const exec_info = (ccv_nnc_graph_exec_info_t*)ccv_array_get(compiled_data->graph->exec_info, 0)((void*)(((char*)((compiled_data->graph->exec_info)->
data)) + (size_t)(compiled_data->graph->exec_info)->
rsize * (size_t)(0)));
1806	const int exec_info_size = compiled_data->graph->exec_info->rnum;
1807	uint32_t* const visited = cccalloccalloc((exec_info_size + 31) >> 5, sizeof(uint32_t));
1808	const ccv_nnc_graph_exec_t* const sources = (ccv_nnc_graph_exec_t*)ccv_array_get(compiled_data->graph->sources, 0)((void*)(((char*)((compiled_data->graph->sources)->data
)) + (size_t)(compiled_data->graph->sources)->rsize *
 (size_t)(0)));
1809	const int source_size = compiled_data->graph->sources->rnum;
1810	ccv_nnc_graph_visit_t* visit = ccv_nnc_graph_visit_new(compiled_data->graph, exec_info, exec_info_size, sources, source_size, compiled_data->evaluate.to_ops, compiled_data->evaluate.to_op_size, 0)({ ccv_nnc_graph_visit_t* _visit_ = (ccv_nnc_graph_visit_t*)malloc
(sizeof(ccv_nnc_graph_visit_t) + sizeof(_visit_->node[0]) *
 ((exec_info_size) - 1)); _visit_->size = 0; do { typedef struct
 { int8_t d; int8_t r; uint16_t c; int32_t edges; } ccv_nnc_incoming_t
; int _i_, _j_; int _incoming_edges_ = 0; for (_i_ = 0; _i_ <
 (exec_info_size); _i_++) _incoming_edges_ += ((exec_info)[_i_
].outgoings) ? (exec_info)[_i_].outgoings->rnum : 0; const
 int _heap_mem_ = ((exec_info_size) + _incoming_edges_ > 1024
); ccv_nnc_incoming_t* _incomings_; if (_heap_mem_) _incomings_
 = (ccv_nnc_incoming_t*)malloc(sizeof(ccv_nnc_incoming_t) * (
exec_info_size) + sizeof(int32_t) * ((exec_info_size) * 2 + _incoming_edges_
)); else _incomings_ = (ccv_nnc_incoming_t*)__builtin_alloca (
sizeof(ccv_nnc_incoming_t) * (exec_info_size) + sizeof(int32_t
) * ((exec_info_size) * 2 + _incoming_edges_)); memset(_incomings_
, 0, sizeof(ccv_nnc_incoming_t) * (exec_info_size)); int32_t*
 _exists_[2] = { (int32_t*)(_incomings_ + (exec_info_size)), (
int32_t*)(_incomings_ + (exec_info_size)) + (exec_info_size),
 }; int32_t* const _edges_ = _exists_[1] + (exec_info_size); for
 (_i_ = 0; _i_ < (source_size); _i_++) { ((void) sizeof ((
(sources)[_i_].graph == compiled_data->graph) ? 1 : 0), __extension__
 ({ if ((sources)[_i_].graph == compiled_data->graph) ; else
 __assert_fail ("(sources)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1810, __extension__ __PRETTY_FUNCTION__
); })); _incomings_[(sources)[_i_].d].r = 1; _exists_[0][_i_]
 = (sources)[_i_].d; } int _exist_size_[2] = { (source_size),
 0, }; int _p_ = 0, _q_ = 1; while (_exist_size_[_p_] > 0)
 { _exist_size_[_q_] = 0; for (_i_ = 0; _i_ < _exist_size_
[_p_]; _i_++) { const int32_t _idx_ = _exists_[_p_][_i_]; if (
_incomings_[_idx_].r != 1) continue; _incomings_[_idx_].r = 2
; if ((exec_info)[_idx_].outgoings) for (_j_ = 0; _j_ < (exec_info
)[_idx_].outgoings->rnum; _j_++) { const int d = *(int*)((
void*)(((char*)(((exec_info)[_idx_].outgoings)->data)) + (
size_t)((exec_info)[_idx_].outgoings)->rsize * (size_t)(_j_
))); ++_incomings_[d].c; if (_incomings_[d].r != 0) continue;
 _incomings_[d].r = 1; ((void) sizeof ((_exist_size_[_q_] <
 (exec_info_size)) ? 1 : 0), __extension__ ({ if (_exist_size_
[_q_] < (exec_info_size)) ; else __assert_fail ("_exist_size_[_q_] < (exec_info_size)"
, "ccv_cnnp_model.c", 1810, __extension__ __PRETTY_FUNCTION__
); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[
_q_]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for
 (_i_ = 0; _i_ < (source_size); _i_++) { ((void) sizeof ((
(sources)[_i_].graph == compiled_data->graph) ? 1 : 0), __extension__
 ({ if ((sources)[_i_].graph == compiled_data->graph) ; else
 __assert_fail ("(sources)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1810, __extension__ __PRETTY_FUNCTION__
); })); _incomings_[(sources)[_i_].d].r = 3; _exists_[0][_i_]
 = (sources)[_i_].d; } _exist_size_[0] = (source_size); _exist_size_
[1] = 0; _p_ = 0, _q_ = 1; int _bump_ = 1; while (_exist_size_
[_p_] > 0) { _exist_size_[_q_] = 0; for (_i_ = 0; _i_ <
 _exist_size_[_p_]; _i_++) { const int32_t _idx_ = _exists_[_p_
][_i_]; if (_incomings_[_idx_].r != 3) continue; _incomings_[
_idx_].r = 4; if ((exec_info)[_idx_].outgoings) for (_j_ = 0;
 _j_ < (exec_info)[_idx_].outgoings->rnum; _j_++) { const
 int d = *(int*)((void*)(((char*)(((exec_info)[_idx_].outgoings
)->data)) + (size_t)((exec_info)[_idx_].outgoings)->rsize
 * (size_t)(_j_))); if (_incomings_[d].edges == 0) { _incomings_
[d].edges = _bump_; _bump_ += _incomings_[d].c; _incomings_[d
].c = 0; } _edges_[_incomings_[d].edges - 1 + _incomings_[d].
c] = _idx_; ++_incomings_[d].c; if (_incomings_[d].r != 2) continue
; _incomings_[d].r = 3; ((void) sizeof ((_exist_size_[_q_] <
 (exec_info_size)) ? 1 : 0), __extension__ ({ if (_exist_size_
[_q_] < (exec_info_size)) ; else __assert_fail ("_exist_size_[_q_] < (exec_info_size)"
, "ccv_cnnp_model.c", 1810, __extension__ __PRETTY_FUNCTION__
); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[
_q_]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for
 (_i_ = 0; _i_ < (compiled_data->evaluate.to_op_size); _i_
++) { ((void) sizeof (((compiled_data->evaluate.to_ops)[_i_
].graph == compiled_data->graph) ? 1 : 0), __extension__ (
{ if ((compiled_data->evaluate.to_ops)[_i_].graph == compiled_data
->graph) ; else __assert_fail ("(compiled_data->evaluate.to_ops)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1810, __extension__ __PRETTY_FUNCTION__
); })); _incomings_[(compiled_data->evaluate.to_ops)[_i_].
d].r = 5; _exists_[0][_i_] = (compiled_data->evaluate.to_ops
)[_i_].d; } _exist_size_[0] = (compiled_data->evaluate.to_op_size
); _exist_size_[1] = 0; _p_ = 0, _q_ = 1; while (_exist_size_
[_p_] > 0) { _exist_size_[_q_] = 0; for (_i_ = 0; _i_ <
 _exist_size_[_p_]; _i_++) { const int32_t _idx_ = _exists_[_p_
][_i_]; if (_incomings_[_idx_].r != 5) continue; _incomings_[
_idx_].r = 6; if (_incomings_[_idx_].edges > 0) for (_j_ =
 0; _j_ < _incomings_[_idx_].c; _j_++) { const int d = _edges_
[_incomings_[_idx_].edges - 1 + _j_]; if (_incomings_[d].r !=
 4) continue; _incomings_[d].r = 5; ((void) sizeof ((_exist_size_
[_q_] < (exec_info_size)) ? 1 : 0), __extension__ ({ if (_exist_size_
[_q_] < (exec_info_size)) ; else __assert_fail ("_exist_size_[_q_] < (exec_info_size)"
, "ccv_cnnp_model.c", 1810, __extension__ __PRETTY_FUNCTION__
); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[
_q_]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for
 (_i_ = 0; _i_ < (compiled_data->evaluate.to_op_size); _i_
++) { ((void) sizeof (((compiled_data->evaluate.to_ops)[_i_
].graph == compiled_data->graph) ? 1 : 0), __extension__ (
{ if ((compiled_data->evaluate.to_ops)[_i_].graph == compiled_data
->graph) ; else __assert_fail ("(compiled_data->evaluate.to_ops)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1810, __extension__ __PRETTY_FUNCTION__
); })); _incomings_[(compiled_data->evaluate.to_ops)[_i_].
d].d = 1; } for (_i_ = 0; _i_ < (source_size); _i_++) { ((
void) sizeof (((sources)[_i_].graph == compiled_data->graph
) ? 1 : 0), __extension__ ({ if ((sources)[_i_].graph == compiled_data
->graph) ; else __assert_fail ("(sources)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1810, __extension__ __PRETTY_FUNCTION__
); })); _exists_[0][_i_] = (sources)[_i_].d; } _p_ = 0; _q_ =
 1; _exist_size_[0] = (source_size); _exist_size_[1] = 0; int
 _d_ = 0; while (_exist_size_[_p_] > 0) { _exist_size_[_q_
] = 0; for (_i_ = 0; _i_ < _exist_size_[_p_];) { const int32_t
 _idx_ = _exists_[_p_][_i_]; _visit_->node[_visit_->size
].index = ((_idx_)); _visit_->node[_visit_->size].term =
 ((_incomings_[_idx_].d)); ++_visit_->size;; if (_incomings_
[_idx_].d) { ++_d_; _incomings_[_idx_].r = 7; } if ((exec_info
)[_idx_].outgoings) { if ((exec_info)[_idx_].outgoings->rnum
 == 1) { const int d = *(int*)((void*)(((char*)(((exec_info)[
_idx_].outgoings)->data)) + (size_t)((exec_info)[_idx_].outgoings
)->rsize * (size_t)(0))); --_incomings_[d].c; if (_incomings_
[d].c == 0 && _incomings_[d].r == 6 && _d_ <
 (compiled_data->evaluate.to_op_size)) { _exists_[_p_][_i_
] = d; continue; } } else for (_j_ = 0; _j_ < (exec_info)[
_idx_].outgoings->rnum; _j_++) { const int d = *(int*)((void
*)(((char*)(((exec_info)[_idx_].outgoings)->data)) + (size_t
)((exec_info)[_idx_].outgoings)->rsize * (size_t)(_j_))); --
_incomings_[d].c; if (_incomings_[d].c == 0 && _incomings_
[d].r == 6 && _d_ < (compiled_data->evaluate.to_op_size
)) { ((void) sizeof ((_exist_size_[_q_] < (exec_info_size)
) ? 1 : 0), __extension__ ({ if (_exist_size_[_q_] < (exec_info_size
)) ; else __assert_fail ("_exist_size_[_q_] < (exec_info_size)"
, "ccv_cnnp_model.c", 1810, __extension__ __PRETTY_FUNCTION__
); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[
_q_]; } } } ++_i_; } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (
_i_)); } for (_i_ = 0; _i_ < (compiled_data->evaluate.to_op_size
); _i_++) { ((void) sizeof (((compiled_data->evaluate.to_ops
)[_i_].graph == compiled_data->graph) ? 1 : 0), __extension__
 ({ if ((compiled_data->evaluate.to_ops)[_i_].graph == compiled_data
->graph) ; else __assert_fail ("(compiled_data->evaluate.to_ops)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1810, __extension__ __PRETTY_FUNCTION__
); })); if (_incomings_[(compiled_data->evaluate.to_ops)[_i_
].d].r == 7) continue; if (!(0)) { ((void) sizeof ((_incomings_
[(compiled_data->evaluate.to_ops)[_i_].d].c == 0) ? 1 : 0)
, __extension__ ({ if (_incomings_[(compiled_data->evaluate
.to_ops)[_i_].d].c == 0) ; else __assert_fail ("_incomings_[(compiled_data->evaluate.to_ops)[_i_].d].c == 0"
, "ccv_cnnp_model.c", 1810, __extension__ __PRETTY_FUNCTION__
); })); } else if (_incomings_[(compiled_data->evaluate.to_ops
)[_i_].d].c > 0) continue; _visit_->node[_visit_->size
].index = (((compiled_data->evaluate.to_ops)[_i_].d)); _visit_
->node[_visit_->size].term = ((_incomings_[(compiled_data
->evaluate.to_ops)[_i_].d].d)); ++_visit_->size;; } if (
_heap_mem_) free(_incomings_); } while (0);; ((void) sizeof (
(_visit_->size <= (exec_info_size)) ? 1 : 0), __extension__
 ({ if (_visit_->size <= (exec_info_size)) ; else __assert_fail
 ("_visit_->size <= (exec_info_size)", "ccv_cnnp_model.c"
, 1810, __extension__ __PRETTY_FUNCTION__); })); _visit_; });
1811	ccv_nnc_graph_visit_for(visit, exec_info, node, idx){ int _i_; for (_i_ = 0; _i_ < (visit)->size; _i_++) { const
 int idx __attribute__((unused)) = (visit)->node[_i_].index
; const int _node_unused_ __attribute__((unused)) = (visit)->
node[_i_].term; typeof ((exec_info)) const node __attribute__
((unused)) = (exec_info) + idx; {
1812		visited[(idx >> 5)] |= (1u << (idx & 31));
1813	} ccv_nnc_graph_visit_endfor} }
1814	ccv_nnc_graph_visit_free(visit);
1815	const ccv_nnc_graph_exec_t* const destinations = (ccv_nnc_graph_exec_t*)ccv_array_get(compiled_data->graph->destinations, 0)((void*)(((char*)((compiled_data->graph->destinations)->
data)) + (size_t)(compiled_data->graph->destinations)->
rsize * (size_t)(0)));
1816	const int destination_size = compiled_data->graph->destinations->rnum;
1817	visit = ccv_nnc_graph_visit_new(compiled_data->graph, exec_info, exec_info_size, compiled_data->backward.from_ops, compiled_data->backward.from_op_size, destinations, destination_size, 0)({ ccv_nnc_graph_visit_t* _visit_ = (ccv_nnc_graph_visit_t*)malloc
(sizeof(ccv_nnc_graph_visit_t) + sizeof(_visit_->node[0]) *
 ((exec_info_size) - 1)); _visit_->size = 0; do { typedef struct
 { int8_t d; int8_t r; uint16_t c; int32_t edges; } ccv_nnc_incoming_t
; int _i_, _j_; int _incoming_edges_ = 0; for (_i_ = 0; _i_ <
 (exec_info_size); _i_++) _incoming_edges_ += ((exec_info)[_i_
].outgoings) ? (exec_info)[_i_].outgoings->rnum : 0; const
 int _heap_mem_ = ((exec_info_size) + _incoming_edges_ > 1024
); ccv_nnc_incoming_t* _incomings_; if (_heap_mem_) _incomings_
 = (ccv_nnc_incoming_t*)malloc(sizeof(ccv_nnc_incoming_t) * (
exec_info_size) + sizeof(int32_t) * ((exec_info_size) * 2 + _incoming_edges_
)); else _incomings_ = (ccv_nnc_incoming_t*)__builtin_alloca (
sizeof(ccv_nnc_incoming_t) * (exec_info_size) + sizeof(int32_t
) * ((exec_info_size) * 2 + _incoming_edges_)); memset(_incomings_
, 0, sizeof(ccv_nnc_incoming_t) * (exec_info_size)); int32_t*
 _exists_[2] = { (int32_t*)(_incomings_ + (exec_info_size)), (
int32_t*)(_incomings_ + (exec_info_size)) + (exec_info_size),
 }; int32_t* const _edges_ = _exists_[1] + (exec_info_size); for
 (_i_ = 0; _i_ < (compiled_data->backward.from_op_size)
; _i_++) { ((void) sizeof (((compiled_data->backward.from_ops
)[_i_].graph == compiled_data->graph) ? 1 : 0), __extension__
 ({ if ((compiled_data->backward.from_ops)[_i_].graph == compiled_data
->graph) ; else __assert_fail ("(compiled_data->backward.from_ops)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1817, __extension__ __PRETTY_FUNCTION__
); })); _incomings_[(compiled_data->backward.from_ops)[_i_
].d].r = 1; _exists_[0][_i_] = (compiled_data->backward.from_ops
)[_i_].d; } int _exist_size_[2] = { (compiled_data->backward
.from_op_size), 0, }; int _p_ = 0, _q_ = 1; while (_exist_size_
[_p_] > 0) { _exist_size_[_q_] = 0; for (_i_ = 0; _i_ <
 _exist_size_[_p_]; _i_++) { const int32_t _idx_ = _exists_[_p_
][_i_]; if (_incomings_[_idx_].r != 1) continue; _incomings_[
_idx_].r = 2; if ((exec_info)[_idx_].outgoings) for (_j_ = 0;
 _j_ < (exec_info)[_idx_].outgoings->rnum; _j_++) { const
 int d = *(int*)((void*)(((char*)(((exec_info)[_idx_].outgoings
)->data)) + (size_t)((exec_info)[_idx_].outgoings)->rsize
 * (size_t)(_j_))); ++_incomings_[d].c; if (_incomings_[d].r !=
 0) continue; _incomings_[d].r = 1; ((void) sizeof ((_exist_size_
[_q_] < (exec_info_size)) ? 1 : 0), __extension__ ({ if (_exist_size_
[_q_] < (exec_info_size)) ; else __assert_fail ("_exist_size_[_q_] < (exec_info_size)"
, "ccv_cnnp_model.c", 1817, __extension__ __PRETTY_FUNCTION__
); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[
_q_]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for
 (_i_ = 0; _i_ < (compiled_data->backward.from_op_size)
; _i_++) { ((void) sizeof (((compiled_data->backward.from_ops
)[_i_].graph == compiled_data->graph) ? 1 : 0), __extension__
 ({ if ((compiled_data->backward.from_ops)[_i_].graph == compiled_data
->graph) ; else __assert_fail ("(compiled_data->backward.from_ops)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1817, __extension__ __PRETTY_FUNCTION__
); })); _incomings_[(compiled_data->backward.from_ops)[_i_
].d].r = 3; _exists_[0][_i_] = (compiled_data->backward.from_ops
)[_i_].d; } _exist_size_[0] = (compiled_data->backward.from_op_size
); _exist_size_[1] = 0; _p_ = 0, _q_ = 1; int _bump_ = 1; while
 (_exist_size_[_p_] > 0) { _exist_size_[_q_] = 0; for (_i_
 = 0; _i_ < _exist_size_[_p_]; _i_++) { const int32_t _idx_
 = _exists_[_p_][_i_]; if (_incomings_[_idx_].r != 3) continue
; _incomings_[_idx_].r = 4; if ((exec_info)[_idx_].outgoings)
 for (_j_ = 0; _j_ < (exec_info)[_idx_].outgoings->rnum
; _j_++) { const int d = *(int*)((void*)(((char*)(((exec_info
)[_idx_].outgoings)->data)) + (size_t)((exec_info)[_idx_].
outgoings)->rsize * (size_t)(_j_))); if (_incomings_[d].edges
 == 0) { _incomings_[d].edges = _bump_; _bump_ += _incomings_
[d].c; _incomings_[d].c = 0; } _edges_[_incomings_[d].edges -
 1 + _incomings_[d].c] = _idx_; ++_incomings_[d].c; if (_incomings_
[d].r != 2) continue; _incomings_[d].r = 3; ((void) sizeof ((
_exist_size_[_q_] < (exec_info_size)) ? 1 : 0), __extension__
 ({ if (_exist_size_[_q_] < (exec_info_size)) ; else __assert_fail
 ("_exist_size_[_q_] < (exec_info_size)", "ccv_cnnp_model.c"
, 1817, __extension__ __PRETTY_FUNCTION__); })); _exists_[_q_
][_exist_size_[_q_]] = d; ++_exist_size_[_q_]; } } ((_i_) = (
_p_), (_p_) = (_q_), (_q_) = (_i_)); } for (_i_ = 0; _i_ <
 (destination_size); _i_++) { ((void) sizeof (((destinations)
[_i_].graph == compiled_data->graph) ? 1 : 0), __extension__
 ({ if ((destinations)[_i_].graph == compiled_data->graph)
 ; else __assert_fail ("(destinations)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1817, __extension__ __PRETTY_FUNCTION__
); })); _incomings_[(destinations)[_i_].d].r = 5; _exists_[0]
[_i_] = (destinations)[_i_].d; } _exist_size_[0] = (destination_size
); _exist_size_[1] = 0; _p_ = 0, _q_ = 1; while (_exist_size_
[_p_] > 0) { _exist_size_[_q_] = 0; for (_i_ = 0; _i_ <
 _exist_size_[_p_]; _i_++) { const int32_t _idx_ = _exists_[_p_
][_i_]; if (_incomings_[_idx_].r != 5) continue; _incomings_[
_idx_].r = 6; if (_incomings_[_idx_].edges > 0) for (_j_ =
 0; _j_ < _incomings_[_idx_].c; _j_++) { const int d = _edges_
[_incomings_[_idx_].edges - 1 + _j_]; if (_incomings_[d].r !=
 4) continue; _incomings_[d].r = 5; ((void) sizeof ((_exist_size_
[_q_] < (exec_info_size)) ? 1 : 0), __extension__ ({ if (_exist_size_
[_q_] < (exec_info_size)) ; else __assert_fail ("_exist_size_[_q_] < (exec_info_size)"
, "ccv_cnnp_model.c", 1817, __extension__ __PRETTY_FUNCTION__
); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[
_q_]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for
 (_i_ = 0; _i_ < (destination_size); _i_++) { ((void) sizeof
 (((destinations)[_i_].graph == compiled_data->graph) ? 1 :
 0), __extension__ ({ if ((destinations)[_i_].graph == compiled_data
->graph) ; else __assert_fail ("(destinations)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1817, __extension__ __PRETTY_FUNCTION__
); })); _incomings_[(destinations)[_i_].d].d = 1; } for (_i_ =
 0; _i_ < (compiled_data->backward.from_op_size); _i_++
) { ((void) sizeof (((compiled_data->backward.from_ops)[_i_
].graph == compiled_data->graph) ? 1 : 0), __extension__ (
{ if ((compiled_data->backward.from_ops)[_i_].graph == compiled_data
->graph) ; else __assert_fail ("(compiled_data->backward.from_ops)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1817, __extension__ __PRETTY_FUNCTION__
); })); _exists_[0][_i_] = (compiled_data->backward.from_ops
)[_i_].d; } _p_ = 0; _q_ = 1; _exist_size_[0] = (compiled_data
->backward.from_op_size); _exist_size_[1] = 0; int _d_ = 0
; while (_exist_size_[_p_] > 0) { _exist_size_[_q_] = 0; for
 (_i_ = 0; _i_ < _exist_size_[_p_];) { const int32_t _idx_
 = _exists_[_p_][_i_]; _visit_->node[_visit_->size].index
 = ((_idx_)); _visit_->node[_visit_->size].term = ((_incomings_
[_idx_].d)); ++_visit_->size;; if (_incomings_[_idx_].d) {
 ++_d_; _incomings_[_idx_].r = 7; } if ((exec_info)[_idx_].outgoings
) { if ((exec_info)[_idx_].outgoings->rnum == 1) { const int
 d = *(int*)((void*)(((char*)(((exec_info)[_idx_].outgoings)->
data)) + (size_t)((exec_info)[_idx_].outgoings)->rsize * (
size_t)(0))); --_incomings_[d].c; if (_incomings_[d].c == 0 &&
 _incomings_[d].r == 6 && _d_ < (destination_size)
) { _exists_[_p_][_i_] = d; continue; } } else for (_j_ = 0; _j_
 < (exec_info)[_idx_].outgoings->rnum; _j_++) { const int
 d = *(int*)((void*)(((char*)(((exec_info)[_idx_].outgoings)->
data)) + (size_t)((exec_info)[_idx_].outgoings)->rsize * (
size_t)(_j_))); --_incomings_[d].c; if (_incomings_[d].c == 0
 && _incomings_[d].r == 6 && _d_ < (destination_size
)) { ((void) sizeof ((_exist_size_[_q_] < (exec_info_size)
) ? 1 : 0), __extension__ ({ if (_exist_size_[_q_] < (exec_info_size
)) ; else __assert_fail ("_exist_size_[_q_] < (exec_info_size)"
, "ccv_cnnp_model.c", 1817, __extension__ __PRETTY_FUNCTION__
); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[
_q_]; } } } ++_i_; } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (
_i_)); } for (_i_ = 0; _i_ < (destination_size); _i_++) { (
(void) sizeof (((destinations)[_i_].graph == compiled_data->
graph) ? 1 : 0), __extension__ ({ if ((destinations)[_i_].graph
 == compiled_data->graph) ; else __assert_fail ("(destinations)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1817, __extension__ __PRETTY_FUNCTION__
); })); if (_incomings_[(destinations)[_i_].d].r == 7) continue
; if (!(0)) { ((void) sizeof ((_incomings_[(destinations)[_i_
].d].c == 0) ? 1 : 0), __extension__ ({ if (_incomings_[(destinations
)[_i_].d].c == 0) ; else __assert_fail ("_incomings_[(destinations)[_i_].d].c == 0"
, "ccv_cnnp_model.c", 1817, __extension__ __PRETTY_FUNCTION__
); })); } else if (_incomings_[(destinations)[_i_].d].c > 0
) continue; _visit_->node[_visit_->size].index = (((destinations
)[_i_].d)); _visit_->node[_visit_->size].term = ((_incomings_
[(destinations)[_i_].d].d)); ++_visit_->size;; } if (_heap_mem_
) free(_incomings_); } while (0);; ((void) sizeof ((_visit_->
size <= (exec_info_size)) ? 1 : 0), __extension__ ({ if (_visit_
->size <= (exec_info_size)) ; else __assert_fail ("_visit_->size <= (exec_info_size)"
, "ccv_cnnp_model.c", 1817, __extension__ __PRETTY_FUNCTION__
); })); _visit_; });
1818	ccv_nnc_graph_visit_for(visit, exec_info, node, idx){ int _i_; for (_i_ = 0; _i_ < (visit)->size; _i_++) { const
 int idx __attribute__((unused)) = (visit)->node[_i_].index
; const int _node_unused_ __attribute__((unused)) = (visit)->
node[_i_].term; typeof ((exec_info)) const node __attribute__
((unused)) = (exec_info) + idx; {
1819		visited[(idx >> 5)] |= (1u << (idx & 31));
1820	} ccv_nnc_graph_visit_endfor} }
1821	ccv_nnc_graph_visit_free(visit);
1822	visit = ccv_nnc_graph_visit_new(compiled_data->graph, exec_info, exec_info_size, sources, source_size, destinations, destination_size, 0)({ ccv_nnc_graph_visit_t* _visit_ = (ccv_nnc_graph_visit_t*)malloc
(sizeof(ccv_nnc_graph_visit_t) + sizeof(_visit_->node[0]) *
 ((exec_info_size) - 1)); _visit_->size = 0; do { typedef struct
 { int8_t d; int8_t r; uint16_t c; int32_t edges; } ccv_nnc_incoming_t
; int _i_, _j_; int _incoming_edges_ = 0; for (_i_ = 0; _i_ <
 (exec_info_size); _i_++) _incoming_edges_ += ((exec_info)[_i_
].outgoings) ? (exec_info)[_i_].outgoings->rnum : 0; const
 int _heap_mem_ = ((exec_info_size) + _incoming_edges_ > 1024
); ccv_nnc_incoming_t* _incomings_; if (_heap_mem_) _incomings_
 = (ccv_nnc_incoming_t*)malloc(sizeof(ccv_nnc_incoming_t) * (
exec_info_size) + sizeof(int32_t) * ((exec_info_size) * 2 + _incoming_edges_
)); else _incomings_ = (ccv_nnc_incoming_t*)__builtin_alloca (
sizeof(ccv_nnc_incoming_t) * (exec_info_size) + sizeof(int32_t
) * ((exec_info_size) * 2 + _incoming_edges_)); memset(_incomings_
, 0, sizeof(ccv_nnc_incoming_t) * (exec_info_size)); int32_t*
 _exists_[2] = { (int32_t*)(_incomings_ + (exec_info_size)), (
int32_t*)(_incomings_ + (exec_info_size)) + (exec_info_size),
 }; int32_t* const _edges_ = _exists_[1] + (exec_info_size); for
 (_i_ = 0; _i_ < (source_size); _i_++) { ((void) sizeof ((
(sources)[_i_].graph == compiled_data->graph) ? 1 : 0), __extension__
 ({ if ((sources)[_i_].graph == compiled_data->graph) ; else
 __assert_fail ("(sources)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1822, __extension__ __PRETTY_FUNCTION__
); })); _incomings_[(sources)[_i_].d].r = 1; _exists_[0][_i_]
 = (sources)[_i_].d; } int _exist_size_[2] = { (source_size),
 0, }; int _p_ = 0, _q_ = 1; while (_exist_size_[_p_] > 0)
 { _exist_size_[_q_] = 0; for (_i_ = 0; _i_ < _exist_size_
[_p_]; _i_++) { const int32_t _idx_ = _exists_[_p_][_i_]; if (
_incomings_[_idx_].r != 1) continue; _incomings_[_idx_].r = 2
; if ((exec_info)[_idx_].outgoings) for (_j_ = 0; _j_ < (exec_info
)[_idx_].outgoings->rnum; _j_++) { const int d = *(int*)((
void*)(((char*)(((exec_info)[_idx_].outgoings)->data)) + (
size_t)((exec_info)[_idx_].outgoings)->rsize * (size_t)(_j_
))); ++_incomings_[d].c; if (_incomings_[d].r != 0) continue;
 _incomings_[d].r = 1; ((void) sizeof ((_exist_size_[_q_] <
 (exec_info_size)) ? 1 : 0), __extension__ ({ if (_exist_size_
[_q_] < (exec_info_size)) ; else __assert_fail ("_exist_size_[_q_] < (exec_info_size)"
, "ccv_cnnp_model.c", 1822, __extension__ __PRETTY_FUNCTION__
); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[
_q_]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for
 (_i_ = 0; _i_ < (source_size); _i_++) { ((void) sizeof ((
(sources)[_i_].graph == compiled_data->graph) ? 1 : 0), __extension__
 ({ if ((sources)[_i_].graph == compiled_data->graph) ; else
 __assert_fail ("(sources)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1822, __extension__ __PRETTY_FUNCTION__
); })); _incomings_[(sources)[_i_].d].r = 3; _exists_[0][_i_]
 = (sources)[_i_].d; } _exist_size_[0] = (source_size); _exist_size_
[1] = 0; _p_ = 0, _q_ = 1; int _bump_ = 1; while (_exist_size_
[_p_] > 0) { _exist_size_[_q_] = 0; for (_i_ = 0; _i_ <
 _exist_size_[_p_]; _i_++) { const int32_t _idx_ = _exists_[_p_
][_i_]; if (_incomings_[_idx_].r != 3) continue; _incomings_[
_idx_].r = 4; if ((exec_info)[_idx_].outgoings) for (_j_ = 0;
 _j_ < (exec_info)[_idx_].outgoings->rnum; _j_++) { const
 int d = *(int*)((void*)(((char*)(((exec_info)[_idx_].outgoings
)->data)) + (size_t)((exec_info)[_idx_].outgoings)->rsize
 * (size_t)(_j_))); if (_incomings_[d].edges == 0) { _incomings_
[d].edges = _bump_; _bump_ += _incomings_[d].c; _incomings_[d
].c = 0; } _edges_[_incomings_[d].edges - 1 + _incomings_[d].
c] = _idx_; ++_incomings_[d].c; if (_incomings_[d].r != 2) continue
; _incomings_[d].r = 3; ((void) sizeof ((_exist_size_[_q_] <
 (exec_info_size)) ? 1 : 0), __extension__ ({ if (_exist_size_
[_q_] < (exec_info_size)) ; else __assert_fail ("_exist_size_[_q_] < (exec_info_size)"
, "ccv_cnnp_model.c", 1822, __extension__ __PRETTY_FUNCTION__
); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[
_q_]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for
 (_i_ = 0; _i_ < (destination_size); _i_++) { ((void) sizeof
 (((destinations)[_i_].graph == compiled_data->graph) ? 1 :
 0), __extension__ ({ if ((destinations)[_i_].graph == compiled_data
->graph) ; else __assert_fail ("(destinations)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1822, __extension__ __PRETTY_FUNCTION__
); })); _incomings_[(destinations)[_i_].d].r = 5; _exists_[0]
[_i_] = (destinations)[_i_].d; } _exist_size_[0] = (destination_size
); _exist_size_[1] = 0; _p_ = 0, _q_ = 1; while (_exist_size_
[_p_] > 0) { _exist_size_[_q_] = 0; for (_i_ = 0; _i_ <
 _exist_size_[_p_]; _i_++) { const int32_t _idx_ = _exists_[_p_
][_i_]; if (_incomings_[_idx_].r != 5) continue; _incomings_[
_idx_].r = 6; if (_incomings_[_idx_].edges > 0) for (_j_ =
 0; _j_ < _incomings_[_idx_].c; _j_++) { const int d = _edges_
[_incomings_[_idx_].edges - 1 + _j_]; if (_incomings_[d].r !=
 4) continue; _incomings_[d].r = 5; ((void) sizeof ((_exist_size_
[_q_] < (exec_info_size)) ? 1 : 0), __extension__ ({ if (_exist_size_
[_q_] < (exec_info_size)) ; else __assert_fail ("_exist_size_[_q_] < (exec_info_size)"
, "ccv_cnnp_model.c", 1822, __extension__ __PRETTY_FUNCTION__
); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[
_q_]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for
 (_i_ = 0; _i_ < (destination_size); _i_++) { ((void) sizeof
 (((destinations)[_i_].graph == compiled_data->graph) ? 1 :
 0), __extension__ ({ if ((destinations)[_i_].graph == compiled_data
->graph) ; else __assert_fail ("(destinations)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1822, __extension__ __PRETTY_FUNCTION__
); })); _incomings_[(destinations)[_i_].d].d = 1; } for (_i_ =
 0; _i_ < (source_size); _i_++) { ((void) sizeof (((sources
)[_i_].graph == compiled_data->graph) ? 1 : 0), __extension__
 ({ if ((sources)[_i_].graph == compiled_data->graph) ; else
 __assert_fail ("(sources)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1822, __extension__ __PRETTY_FUNCTION__
); })); _exists_[0][_i_] = (sources)[_i_].d; } _p_ = 0; _q_ =
 1; _exist_size_[0] = (source_size); _exist_size_[1] = 0; int
 _d_ = 0; while (_exist_size_[_p_] > 0) { _exist_size_[_q_
] = 0; for (_i_ = 0; _i_ < _exist_size_[_p_];) { const int32_t
 _idx_ = _exists_[_p_][_i_]; _visit_->node[_visit_->size
].index = ((_idx_)); _visit_->node[_visit_->size].term =
 ((_incomings_[_idx_].d)); ++_visit_->size;; if (_incomings_
[_idx_].d) { ++_d_; _incomings_[_idx_].r = 7; } if ((exec_info
)[_idx_].outgoings) { if ((exec_info)[_idx_].outgoings->rnum
 == 1) { const int d = *(int*)((void*)(((char*)(((exec_info)[
_idx_].outgoings)->data)) + (size_t)((exec_info)[_idx_].outgoings
)->rsize * (size_t)(0))); --_incomings_[d].c; if (_incomings_
[d].c == 0 && _incomings_[d].r == 6 && _d_ <
 (destination_size)) { _exists_[_p_][_i_] = d; continue; } } else
 for (_j_ = 0; _j_ < (exec_info)[_idx_].outgoings->rnum
; _j_++) { const int d = *(int*)((void*)(((char*)(((exec_info
)[_idx_].outgoings)->data)) + (size_t)((exec_info)[_idx_].
outgoings)->rsize * (size_t)(_j_))); --_incomings_[d].c; if
 (_incomings_[d].c == 0 && _incomings_[d].r == 6 &&
 _d_ < (destination_size)) { ((void) sizeof ((_exist_size_
[_q_] < (exec_info_size)) ? 1 : 0), __extension__ ({ if (_exist_size_
[_q_] < (exec_info_size)) ; else __assert_fail ("_exist_size_[_q_] < (exec_info_size)"
, "ccv_cnnp_model.c", 1822, __extension__ __PRETTY_FUNCTION__
); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[
_q_]; } } } ++_i_; } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (
_i_)); } for (_i_ = 0; _i_ < (destination_size); _i_++) { (
(void) sizeof (((destinations)[_i_].graph == compiled_data->
graph) ? 1 : 0), __extension__ ({ if ((destinations)[_i_].graph
 == compiled_data->graph) ; else __assert_fail ("(destinations)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1822, __extension__ __PRETTY_FUNCTION__
); })); if (_incomings_[(destinations)[_i_].d].r == 7) continue
; if (!(0)) { ((void) sizeof ((_incomings_[(destinations)[_i_
].d].c == 0) ? 1 : 0), __extension__ ({ if (_incomings_[(destinations
)[_i_].d].c == 0) ; else __assert_fail ("_incomings_[(destinations)[_i_].d].c == 0"
, "ccv_cnnp_model.c", 1822, __extension__ __PRETTY_FUNCTION__
); })); } else if (_incomings_[(destinations)[_i_].d].c > 0
) continue; _visit_->node[_visit_->size].index = (((destinations
)[_i_].d)); _visit_->node[_visit_->size].term = ((_incomings_
[(destinations)[_i_].d].d)); ++_visit_->size;; } if (_heap_mem_
) free(_incomings_); } while (0);; ((void) sizeof ((_visit_->
size <= (exec_info_size)) ? 1 : 0), __extension__ ({ if (_visit_
->size <= (exec_info_size)) ; else __assert_fail ("_visit_->size <= (exec_info_size)"
, "ccv_cnnp_model.c", 1822, __extension__ __PRETTY_FUNCTION__
); })); _visit_; });
1823	// Find any missing nodes to be added as source. Right now, these are only set nodes.
1824	ccv_nnc_graph_visit_for(visit, exec_info, node, idx){ int _i_; for (_i_ = 0; _i_ < (visit)->size; _i_++) { const
 int idx __attribute__((unused)) = (visit)->node[_i_].index
; const int _node_unused_ __attribute__((unused)) = (visit)->
node[_i_].term; typeof ((exec_info)) const node __attribute__
((unused)) = (exec_info) + idx; {
1825		if (!(visited[(idx >> 5)] & (1u << (idx & 31))))
1826		{
1827			assert(exec_info[idx].cmd.cmd == CCV_NNC_SET_FORWARD)((void) sizeof ((exec_info[idx].cmd.cmd == CCV_NNC_SET_FORWARD
) ? 1 : 0), __extension__ ({ if (exec_info[idx].cmd.cmd == CCV_NNC_SET_FORWARD
) ; else __assert_fail ("exec_info[idx].cmd.cmd == CCV_NNC_SET_FORWARD"
, "ccv_cnnp_model.c", 1827, __extension__ __PRETTY_FUNCTION__
); }));
1828			if (exec_info[idx].cmd.info.blas.a[0] == 0) // Special-casing for empty out the tensor set function, not for the set grad to 1 one.
1829				ccv_array_add_unique_int(backward_from, idx);
1830		}
1831	} ccv_nnc_graph_visit_endfor} }
1832	ccv_nnc_graph_visit_free(visit);
1833	ccfreefree(visited);
1834	if (backward_from->rnum != compiled_data->backward.from_op_size) // If it doesn't match, need to redo this.
1835	{
1836		compiled_data->backward.from_op_size = backward_from->rnum;
1837		compiled_data->backward.from_ops = (ccv_nnc_graph_exec_t*)ccreallocrealloc(compiled_data->backward.from_ops, sizeof(ccv_nnc_graph_exec_t) * backward_from->rnum);
1838		for (i = 0; i < backward_from->rnum; i++)
1839			compiled_data->backward.from_ops[i] = (ccv_nnc_graph_exec_t){
1840				.d = *(int*)ccv_array_get(backward_from, i)((void*)(((char*)((backward_from)->data)) + (size_t)(backward_from
)->rsize * (size_t)(i))),
1841				.graph = compiled_data->graph,
1842			};
1843	}
1844	ccv_array_free(backward_from);
1845	ccv_nnc_graph_set_default_static_schedule(compiled_data->graph, compiled_data->stream_type, model->max_stream_count);
1846	ccv_nnc_graph_autotune(compiled_data->graph, model->workspace_size, 0, TRAVERSE_FULL0,0,0,0);
1847}
1848 
1849void 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)
1850{
1851	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1852	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 1852, __extension__ __PRETTY_FUNCTION__); }));
1853	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; });
1854	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", 1854, __extension__ __PRETTY_FUNCTION__
); }));
1855	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", 1855, __extension__ __PRETTY_FUNCTION__
); }));
1856	assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if
 (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c"
, 1856, __extension__ __PRETTY_FUNCTION__); }));
1857	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;
1858	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));
1859	if (!compiled_data->graph || mode_mismatch)
1860	{
1861		_ccv_cnnp_compiled_data_graph_free(compiled_data);
1862		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.
1863			_ccv_cnnp_compiled_data_backward_free(compiled_data);
1864		if (params.requires_grad)
1865			_ccv_cnnp_model_multistage_jit_0(model, params.disable_outgrad, params.is_test, inputs, input_size, outputs, output_size);
1866		else
1867			_ccv_cnnp_model_multistage_no_grad_jit(model, inputs, input_size, outputs, output_size);
1868	} else {
1869		ccv_nnc_tensor_arena_clear_bindings(compiled_data->tensor_arena);
1870		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"
, 1870, __extension__ __PRETTY_FUNCTION__); }));
1871		const int input_size_per_p = input_size / parallel_count;
1872		_ccv_cnnp_bind_tensors_to_arena(compiled_data->tensor_arena, model->graph, model->inputs, inputs, input_size_per_p, parallel_count);
1873		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"
, 1873, __extension__ __PRETTY_FUNCTION__); }));
1874		const int output_size_per_p = output_size / parallel_count;
1875		_ccv_cnnp_bind_tensors_to_arena(compiled_data->tensor_arena, model->graph, model->outputs, outputs, output_size_per_p, parallel_count);
1876	}
1877	if (compiled_data->is_test != params.is_test)
1878	{
1879		compiled_data->is_test = params.is_test;
1880		ccv_nnc_graph_exec_update_t update = {
1881			.parallel_count = parallel_count,
1882			.graph = model->graph,
1883			.graph_exec_arena = compiled_data->graph_exec_arena,
1884		};
1885		ccv_cnnp_model_set_is_test(model, params.is_test, _ccv_cnnp_cmd_update_for_execs, &update);
1886	}
1887}
1888 
1889void 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)
1890{
1891	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1892	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 1892, __extension__ __PRETTY_FUNCTION__); }));
1893	ccv_cnnp_model_dry_run(model, params, inputs, input_size, outputs, output_size);
1894	if (compiled_data->graph_mode == CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE_NO_GRAD)
1895		ccv_nnc_graph_run_with_schedule(compiled_data->graph, 0, 0, tensor_tape, stream_context);
1896	else {
1897		if (!compiled_data->evaluate.schedule)
1898			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);
1899		ccv_nnc_graph_run_with_schedule(compiled_data->graph, 0, compiled_data->evaluate.schedule, tensor_tape, stream_context);
1900	}
1901}
1902 
1903// Compile the graph to run ccv_cnnp_model_backward after ccv_cnnp_model_evaluate with requires_grad = true (MULTISTAGE_MODE).
1904// Particularly, this method compiles the accumulator graph.
1905static void _ccv_cnnp_model_multistage_jit_1(ccv_cnnp_model_t* const model)
1906{
1907	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1908	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 1908, __extension__ __PRETTY_FUNCTION__); }));
1909	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", 1909, __extension__ __PRETTY_FUNCTION__
); }));
1910	ccv_nnc_symbolic_graph_t* accum = ccv_nnc_symbolic_graph_new();
1911	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; });
1912	const int parameter_size = compiled_data->parameters->rnum;
1913	int i, j;
1914	compiled_data->backward.gradients = (ccv_nnc_tensor_symbol_t*)ccmallocmalloc(sizeof(ccv_nnc_tensor_symbol_t) * parameter_size * parallel_count * 3);
1915	compiled_data->backward.accum_gradients = compiled_data->backward.gradients + parameter_size * parallel_count;
1916	compiled_data->backward.updated_accum_gradients = compiled_data->backward.accum_gradients + parameter_size * parallel_count;
1917	for (i = 0; i < parameter_size; i++)
1918		for (j = 0; j < parallel_count; j++)
1919			if (compiled_data->tensors.gradients[i + j * parameter_size])
1920			{
1921				const ccv_nnc_tensor_param_t info = compiled_data->tensors.gradients[i + j * parameter_size]->info;
1922				// Now, the old gradient is the accumulated gradient, getting new gradient tensor setup so we can collect them.
1923				compiled_data->tensors.accum_gradients[i + j * parameter_size] = compiled_data->tensors.gradients[i + j * parameter_size];
1924				compiled_data->tensors.gradients[i + j * parameter_size] = ccv_nnc_tensor_new(0, info, 0);
1925				ccv_nnc_tensor_symbol_t inputs[2];
1926				inputs[0] = compiled_data->backward.accum_gradients[i + j * parameter_size] = ccv_nnc_tensor_symbol_new(accum, info, 0);
1927				inputs[1] = compiled_data->backward.gradients[i + j * parameter_size] = ccv_nnc_tensor_symbol_new(accum, info, 0);
1928				ccv_nnc_tensor_symbol_t output = compiled_data->backward.updated_accum_gradients[i + j * parameter_size] = ccv_nnc_tensor_symbol_new(accum, info, 0);
1929				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);
1930			} else {
1931				compiled_data->backward.accum_gradients[i + j * parameter_size] = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
1932				compiled_data->backward.gradients[i + j * parameter_size] = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
1933				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
};
1934			}
1935	ccv_nnc_graph_exec_symbol_autogen(accum, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
1936	if (ccv_nnc_symbolic_graph_source_size(accum) == 0)
1937	{
1938		ccv_nnc_symbolic_graph_free(accum);
1939		// Create empty graph.
1940		compiled_data->backward.accum = ccv_nnc_graph_new();
1941		ccv_nnc_graph_topsort(compiled_data->backward.accum, 0, 0);
1942		return;
1943	}
1944	ccv_array_t* const tensor_binds = ccv_array_new(sizeof(ccv_nnc_tensor_bind_t), 0, 0);
1945	_ccv_cnnp_model_bind_tensors(accum, compiled_data->backward.accum_gradients, compiled_data->tensors.accum_gradients, parameter_size * parallel_count, 1, tensor_binds);
1946	_ccv_cnnp_model_bind_tensors(accum, compiled_data->backward.gradients, compiled_data->tensors.gradients, parameter_size * parallel_count, 1, tensor_binds);
1947	_ccv_cnnp_model_bind_tensors(accum, compiled_data->backward.updated_accum_gradients, compiled_data->tensors.accum_gradients, parameter_size * parallel_count, 1, tensor_binds);
1948	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);
1949	ccv_nnc_symbolic_graph_free(accum);
1950	ccv_array_free(tensor_binds);
1951	ccv_nnc_graph_set_default_static_schedule(compiled_data->backward.accum, compiled_data->stream_type, model->max_stream_count);
1952}
1953 
1954void 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)
1955{
1956	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1957	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 1957, __extension__ __PRETTY_FUNCTION__); }));
1958	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", 1958, __extension__ __PRETTY_FUNCTION__
); }));
1959	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; });
1960	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", 1960, __extension__ __PRETTY_FUNCTION__
); }));
1961	if (outgrad_size > 0)
1962		{ 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", 1962, __extension__ __PRETTY_FUNCTION__
); })); }
1963	assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if
 (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c"
, 1963, __extension__ __PRETTY_FUNCTION__); }));
1964	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", 1964, __extension__ __PRETTY_FUNCTION__
); }));
1965	const int parameter_size = compiled_data->parameters->rnum;
1966	// If we need to accumulate the gradients now, do jit on accumulator.
1967	if (compiled_data->backward.count > 0)
1968	{
1969		if (!compiled_data->backward.accum)
1970			_ccv_cnnp_model_multistage_jit_1(model);
1971		else if (compiled_data->backward.count == 1) {
1972			//  On this round, we need to switch accumulated gradients with gradients (so we can do accumulation properly).
1973			int i;
1974			for (i = 0; i < parameter_size * parallel_count; i++)
1975			{
1976				ccv_nnc_tensor_t* tensor;
1977				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));
1978			}
1979			if (compiled_data->backward.tensor_arena)
1980			{
1981				ccv_nnc_tensor_arena_clear_bindings(compiled_data->backward.tensor_arena);
1982				// Do rebind in case we messed up the binding (we switch accum_gradients and gradients).
1983				_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);
1984				_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);
1985				_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);
1986			}
1987		}
1988	}
1989	const int ingrad_size_per_p = model->output_size;
1990	const int outgrad_size_per_p = compiled_data->outgrad_size;
1991	int i, j;
1992	for (i = 0; i < ingrad_size_per_p; i++)
1993	{
1994		const ccv_nnc_tensor_symbol_t ingrad = ccv_nnc_tensor_symbol_for_backward(model->graph, compiled_data->f[i]);
1995		if (!ingrad_size || !ingrads || ingrads[i] == 0)
1996		{
1997			// Set it to 1 if it is not specified.
1998			ccv_nnc_tensor_t* const ingrad_tensor = ccv_nnc_tensor_from_symbol(compiled_data->tensor_arena, ingrad);
1999			if (ingrad_tensor)
2000				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);
2001			for (j = 1; j < parallel_count; j++)
2002			{
2003				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));
2004				if (ingrad_tensor)
2005					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);
2006			}
2007		} else {
2008			// Make sure the length matches, in case it is an alias.
2009			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", 2009, __extension__ __PRETTY_FUNCTION__
); }));
2010			ccv_nnc_tensor_bind_symbol(compiled_data->tensor_arena, ingrad, ingrads[i]);
2011			for (j = 1; j < parallel_count; j++)
2012				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]);
2013		}
2014	}
2015	if (outgrad_size > 0)
2016	{
2017		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", 2017, __extension__ __PRETTY_FUNCTION__
); }));
2018		for (i = 0; i < outgrad_size_per_p; i++)
2019			if (outgrads[i])
2020			{
2021				const ccv_nnc_tensor_symbol_t outgrad = compiled_data->outgrads[i];
2022				ccv_nnc_tensor_bind_symbol(compiled_data->tensor_arena, outgrad, outgrads[i]);
2023				for (j = 1; j < parallel_count; j++)
2024					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]);
2025			}
2026	} else {
2027		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", 2028, __extension__ __PRETTY_FUNCTION__
); }))
2028			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", 2028, __extension__ __PRETTY_FUNCTION__
); }));
2029	}
2030	// We need to rebind here because in ccv_cnnp_evaluate, we clear bindings, that will reset all bindings for the gradients.
2031	// For parameters and internals these are fine because when we clear bindings, it restores to original bindings, which are these
2032	// parameters and internals. The same cannot be said for gradients due to the accum_gradients switching.
2033	_ccv_cnnp_bind_tensors_to_arena(compiled_data->tensor_arena, model->graph, compiled_data->gradients, compiled_data->tensors.gradients, parameter_size, parallel_count);
2034	if (!compiled_data->backward.schedule)
2035		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);
2036	// Run the backward pass.
2037	ccv_nnc_graph_run_with_schedule(compiled_data->graph, 0, compiled_data->backward.schedule, tensor_tape, stream_context);
2038	// If we need to run accumulation round, do that now.
2039	if (compiled_data->backward.count > 0)
2040		ccv_nnc_graph_run_with_schedule(compiled_data->backward.accum, 0, 0, 0, stream_context);
2041	// Update the count, this determines whether we need to accumulate or not.
2042	++compiled_data->backward.count;
2043}
2044 
2045// Compile the graph to run ccv_cnnp_model_apply_gradients after ccv_cnnp_model_backward (MULTISTAGE_MODE).
2046// Particularly, this method compiles the parameter update graph.
2047static void _ccv_cnnp_model_multistage_jit_2(ccv_cnnp_model_t* const model)
2048{
2049	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2050	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", 2050, __extension__ __PRETTY_FUNCTION__
); }));
2051	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; });
2052	const int parameter_size = compiled_data->parameters->rnum;
2053	ccv_array_t* const tensor_binds = ccv_array_new(sizeof(ccv_nnc_tensor_bind_t), 0, 0);
2054	_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);
2055	_ccv_cnnp_model_bind_tensors(model->graph, compiled_data->updated_parameters, compiled_data->tensors.parameters, parameter_size, parallel_count, tensor_binds);
2056	// Bind accumulated gradients.
2057	if (compiled_data->backward.count > 1)
2058		_ccv_cnnp_model_bind_tensors(model->graph, compiled_data->gradients, compiled_data->tensors.accum_gradients, parameter_size, parallel_count, tensor_binds);
2059	else
2060		_ccv_cnnp_model_bind_tensors(model->graph, compiled_data->gradients, compiled_data->tensors.gradients, parameter_size, parallel_count, tensor_binds);
2061	ccv_array_t* const apply_gradients_from = ccv_array_new(sizeof(int), 0, 0);
2062	int i, j;
2063	for (i = 0; i < compiled_data->backward.to_size; i++)
2064	{
2065		const int* tos;
2066		int to_size;
2067		ccv_nnc_graph_exec_symbol_to(model->graph, compiled_data->backward.tos[i], &tos, &to_size);
2068		for (j = 0; j < to_size; j++)
2069		{
2070			// Check if this is already show up in the backward graph, if that is the case, it won't be in the apply
2071			// gradients graph.
2072			const ccv_nnc_graph_exec_t exec = ccv_nnc_graph_exec_from_symbol(compiled_data->graph_exec_arena, (ccv_nnc_graph_exec_symbol_t){
2073				.d = tos[j],
2074				.graph = model->graph,
2075			});
2076			if (!exec.graph)
2077				ccv_array_add_unique_int(apply_gradients_from, tos[j]);
2078		}
2079	}
2080	const int from_size = apply_gradients_from->rnum;
2081	if (from_size == 0)
2082	{
2083		ccv_array_free(apply_gradients_from);
2084		ccv_array_free(tensor_binds);
2085		return;
2086	}
2087	ccv_nnc_graph_exec_symbol_t* const froms = (ccv_nnc_graph_exec_symbol_t*)ccmallocmalloc(sizeof(ccv_nnc_graph_exec_symbol_t) * from_size);
2088	for (i = 0; i < from_size; i++)
2089		froms[i] = (ccv_nnc_graph_exec_symbol_t){
2090			.d = *(int*)ccv_array_get(apply_gradients_from, i)((void*)(((char*)((apply_gradients_from)->data)) + (size_t
)(apply_gradients_from)->rsize * (size_t)(i))),
2091			.graph = model->graph
2092		};
2093	ccv_array_free(apply_gradients_from);
2094	// It can only ends with updates on the parameters.
2095	ccv_array_t* const tos = ccv_array_new(sizeof(ccv_nnc_graph_exec_symbol_t), parameter_size * parallel_count, 0);
2096	for (i = 0;  i < parameter_size; i++)
2097	{
2098		if (compiled_data->update_nodes[i].d == CCV_NNC_NO_TENSOR_SYMBOL)
2099			continue;
2100		ccv_array_push(tos, &compiled_data->update_nodes[i]);
2101		for (j = 1; j < parallel_count; j++)
2102		{
2103			const ccv_nnc_graph_exec_symbol_t copy = ccv_nnc_graph_exec_symbol_copy(model->graph, compiled_data->update_nodes[i], j);
2104			ccv_array_push(tos, &copy);
2105		}
2106	}
2107	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);
2108	ccv_array_free(tos);
2109	ccv_array_free(tensor_binds);
2110	ccfreefree(froms);
2111	const int max_saved_aux_size = compiled_data->minimize.max_saved_aux_size;
2112	for (i = 0; i < max_saved_aux_size * parameter_size; i++)
2113	{
2114		// Skip on no tensor.
2115		if (compiled_data->saved_aux[i].source.d == CCV_NNC_NO_TENSOR_SYMBOL)
2116			continue;
2117		ccv_nnc_tensor_t* const tensor = ccv_nnc_tensor_from_symbol(compiled_data->apply_gradients.tensor_arena, compiled_data->saved_aux[i].source);
2118		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);
2119		for (j = 1; j < parallel_count; j++)
2120		{
2121			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));
2122			if (copy)
2123				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);
2124		}
2125	}
2126	ccv_nnc_graph_set_default_static_schedule(compiled_data->apply_gradients.graph, compiled_data->stream_type, model->max_stream_count);
2127}
2128 
2129void ccv_cnnp_model_apply_gradients(ccv_cnnp_model_t* const model, ccv_nnc_stream_context_t* const stream_context)
2130{
2131	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2132	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 2132, __extension__ __PRETTY_FUNCTION__); }));
2133	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", 2133, __extension__ __PRETTY_FUNCTION__
); }));
2134	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; });
2135	assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if
 (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c"
, 2135, __extension__ __PRETTY_FUNCTION__); }));
2136	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", 2136, __extension__ __PRETTY_FUNCTION__
); }));
2137	// Skip if there is no backward pass.
2138	if (compiled_data->backward.count <= 0)
2139		return;
2140	// Skip if there is no parameters.
2141	if (compiled_data->parameters->rnum == 0)
2142	{
2143		compiled_data->backward.count = 0;
2144		return;
2145	}
2146	if (!compiled_data->apply_gradients.graph)
2147		_ccv_cnnp_model_multistage_jit_2(model);
2148	else {
2149		const int parameter_size = compiled_data->parameters->rnum;
2150		ccv_nnc_tensor_arena_clear_bindings(compiled_data->apply_gradients.tensor_arena);
2151		// Change to bind accum_gradients if we do gradient accumulation (run backward more than once).
2152		if (compiled_data->backward.count > 1)
2153			_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);
2154		else
2155			_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);
2156	}
2157	if (compiled_data->apply_gradients.graph)
2158		ccv_nnc_graph_run_with_schedule(compiled_data->apply_gradients.graph, 0, 0, 0, stream_context);
2159	// Reset backward count to 0.
2160	compiled_data->backward.count = 0;
2161}
2162 
2163void 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)
2164{
2165	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2166	const int param_sel = parameter->param_sel > 0 ? parameter->param_sel - 1 : parameter->param_sel;
2167	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", 2167, __extension__ __PRETTY_FUNCTION__
); }));
2168	const int tensors_init = !!compiled_data->tensors_init.v;
2169	if (!tensors_init)
2170		_ccv_cnnp_model_tensors_init(model, compiled_data);
2171	else if ((uintptr_t)compiled_data->tensors_init.v & (uintptr_t)1)
2172	// Check if it is not fully allocated, if it is not, init_1.
2173		ccv_cnnp_model_tensors_init_1(model, compiled_data);
2174	ccv_array_t* const parameter_indices = ccv_array_new(sizeof(int), 0, 0);
2175	ccv_cnnp_model_add_to_parameter_indices(parameter->model, param_sel, parameter_indices);
2176	const int param_ref = parameter->param_ref > 0 ? parameter->param_ref - 1 : parameter->param_ref;
2177	if (param_ref < 0)
2178		{ 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", 2178
, __extension__ __PRETTY_FUNCTION__); })); }
2179	else
2180		{ 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", 2180, __extension__ __PRETTY_FUNCTION__
); })); }
2181	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)));
2182	ccv_array_free(parameter_indices);
2183	const int parameter_size = compiled_data->parameters->rnum;
2184	assert(d >= 0)((void) sizeof ((d >= 0) ? 1 : 0), __extension__ ({ if (d >=
 0) ; else __assert_fail ("d >= 0", "ccv_cnnp_model.c", 2184
, __extension__ __PRETTY_FUNCTION__); }));
2185	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", 2185, __extension__ __PRETTY_FUNCTION__
); }));
2186	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; });
2187	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));
2188	assert(dest)((void) sizeof ((dest) ? 1 : 0), __extension__ ({ if (dest) ;
 else __assert_fail ("dest", "ccv_cnnp_model.c", 2188, __extension__
 __PRETTY_FUNCTION__); }));
2189	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);
2190	int i;
2191	for (i = 1; i < parallel_count; i++)
2192	{
2193		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));
2194		if (copy_tensor)
2195			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);
2196	}
2197	// Mark this symbol as init'ed.
2198	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;
2199  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));
2200	init_v[s >> 5] |= (1u << (s & 0x1f));
2201}
2202 
2203void ccv_cnnp_model_parameter_copy(ccv_cnnp_model_t* const model, const ccv_cnnp_model_io_t parameter, ccv_nnc_tensor_t* const tensor)
2204{
2205	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2206	const int param_sel = parameter->param_sel > 0 ? parameter->param_sel - 1 : parameter->param_sel;
2207	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", 2207, __extension__ __PRETTY_FUNCTION__
); }));
2208	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", 2208, __extension__ __PRETTY_FUNCTION__
); }));
2209	ccv_array_t* const parameter_indices = ccv_array_new(sizeof(int), 0, 0);
2210	ccv_cnnp_model_add_to_parameter_indices(parameter->model, param_sel, parameter_indices);
2211	const int param_ref = parameter->param_ref > 0 ? parameter->param_ref - 1 : parameter->param_ref;
2212	if (param_ref < 0)
2213		{ 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", 2213
, __extension__ __PRETTY_FUNCTION__); })); }
2214	else
2215		{ 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", 2215, __extension__ __PRETTY_FUNCTION__
); })); }
2216	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)));
2217	ccv_array_free(parameter_indices);
2218	const int parameter_size = compiled_data->parameters->rnum;
2219	assert(d >= 0)((void) sizeof ((d >= 0) ? 1 : 0), __extension__ ({ if (d >=
 0) ; else __assert_fail ("d >= 0", "ccv_cnnp_model.c", 2219
, __extension__ __PRETTY_FUNCTION__); }));
2220	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", 2220, __extension__ __PRETTY_FUNCTION__
); }));
2221	// We don't need to consider parallel_count, every parameter on each device is identical.
2222	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));
2223	assert(src)((void) sizeof ((src) ? 1 : 0), __extension__ ({ if (src) ; else
 __assert_fail ("src", "ccv_cnnp_model.c", 2223, __extension__
 __PRETTY_FUNCTION__); }));
2224	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);
2225}
2226 
2227ccv_nnc_tensor_param_t ccv_cnnp_model_parameter_tensor_params(ccv_cnnp_model_t* const model, const ccv_cnnp_model_io_t parameter)
2228{
2229	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2230	const int param_sel = parameter->param_sel > 0 ? parameter->param_sel - 1 : parameter->param_sel;
2231	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", 2231, __extension__ __PRETTY_FUNCTION__
); }));
2232	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", 2232, __extension__ __PRETTY_FUNCTION__
); }));
2233	ccv_array_t* const parameter_indices = ccv_array_new(sizeof(int), 0, 0);
2234	ccv_cnnp_model_add_to_parameter_indices(parameter->model, param_sel, parameter_indices);
2235	const int param_ref = parameter->param_ref > 0 ? parameter->param_ref - 1 : parameter->param_ref;
2236	if (param_ref < 0)
2237		{ 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", 2237
, __extension__ __PRETTY_FUNCTION__); })); }
2238	else
2239		{ 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", 2239, __extension__ __PRETTY_FUNCTION__
); })); }
2240	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)));
2241	ccv_array_free(parameter_indices);
2242	const int parameter_size = compiled_data->parameters->rnum;
2243	assert(d >= 0)((void) sizeof ((d >= 0) ? 1 : 0), __extension__ ({ if (d >=
 0) ; else __assert_fail ("d >= 0", "ccv_cnnp_model.c", 2243
, __extension__ __PRETTY_FUNCTION__); }));
2244	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", 2244, __extension__ __PRETTY_FUNCTION__
); }));
2245	// We don't need to consider parallel_count, every parameter on each device is identical.
2246	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));
2247	assert(tensor)((void) sizeof ((tensor) ? 1 : 0), __extension__ ({ if (tensor
) ; else __assert_fail ("tensor", "ccv_cnnp_model.c", 2247, __extension__
 __PRETTY_FUNCTION__); }));
2248	return tensor->info;
2249}
2250 
2251const char* ccv_cnnp_model_parameter_name(ccv_cnnp_model_t* const model, const ccv_cnnp_model_io_t parameter)
2252{
2253	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2254	const int param_sel = parameter->param_sel > 0 ? parameter->param_sel - 1 : parameter->param_sel;
2255	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", 2255, __extension__ __PRETTY_FUNCTION__
); }));
2256	ccv_array_t* const parameter_indices = ccv_array_new(sizeof(int), 0, 0);
2257	ccv_cnnp_model_add_to_parameter_indices(parameter->model, param_sel, parameter_indices);
2258	const int param_ref = parameter->param_ref > 0 ? parameter->param_ref - 1 : parameter->param_ref;
2259	if (param_ref < 0)
2260		{ 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", 2260
, __extension__ __PRETTY_FUNCTION__); })); }
2261	else
2262		{ 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", 2262, __extension__ __PRETTY_FUNCTION__
); })); }
2263	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)));
2264	ccv_array_free(parameter_indices);
2265	const int parameter_size = compiled_data->parameters->rnum;
2266	assert(d >= 0)((void) sizeof ((d >= 0) ? 1 : 0), __extension__ ({ if (d >=
 0) ; else __assert_fail ("d >= 0", "ccv_cnnp_model.c", 2266
, __extension__ __PRETTY_FUNCTION__); }));
2267	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", 2267, __extension__ __PRETTY_FUNCTION__
); }));
2268	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)));
2269}
2270 
2271int ccv_cnnp_model_parameter_count(ccv_cnnp_model_t* const model)
2272{
2273	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", 2273, __extension__ __PRETTY_FUNCTION__
); }));
2274	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2275	return compiled_data->parameters->rnum;
2276}
2277 
2278ccv_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)
2279{
2280	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2281	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 2281, __extension__ __PRETTY_FUNCTION__); }));
2282	const int parameter_size = compiled_data->parameters->rnum;
2283	int i;
2284	for (i = 0; i < parameter_size; i++)
2285	{
2286		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)));
2287		if (first(model, name, context))
2288			return ccv_cnnp_model_parameters(model, -1, i);
2289	}
2290	return 0;
2291}
2292 
2293ccv_array_t* ccv_cnnp_model_parameters_filter(ccv_cnnp_model_t* const model, ccv_cnnp_model_parameters_filter_f filter, void* const context)
2294{
2295	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2296	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 2296, __extension__ __PRETTY_FUNCTION__); }));
2297	ccv_array_t* const parameters = ccv_array_new(sizeof(ccv_cnnp_model_io_t), 0, 0);
2298	const int parameter_size = compiled_data->parameters->rnum;
2299	int i;
2300	for (i = 0; i < parameter_size; i++)
2301	{
2302		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)));
2303		if (filter(model, name, context))
2304		{
2305			ccv_cnnp_model_io_t parameter = ccv_cnnp_model_parameters(model, -1, i);
2306			ccv_array_push(parameters, &parameter);
2307		}
2308	}
2309	return parameters;
2310 
2311}
2312 
2313CCV_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)
2314{
2315	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2316	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 2316, __extension__ __PRETTY_FUNCTION__); }));
2317	const int tensors_init = !!compiled_data->tensors_init.v;
2318	if (!tensors_init) // If nothing initialized, we return parameter 0.
2319		return ccv_cnnp_model_parameters(model, -1, 0);
2320	const int parameter_size = compiled_data->parameters->rnum;
2321	int i;
2322	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));
2323	for (i = 0; i < parameter_size; i++)
2324	{
2325		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;
2326		if (!(init_v[d >> 5] & (1u << (d & 0x1f))))
2327			return ccv_cnnp_model_parameters(model, -1, i);
2328	}
2329	return 0;
2330}
2331 
2332static 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)
2333{
2334	const int to_param_sel = parameters->param_sel > 0 ? parameters->param_sel - 1 : parameters->param_sel;
2335	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", 2335, __extension__
 __PRETTY_FUNCTION__); }));
2336	ccv_array_t* const to_parameter_indices = ccv_array_new(sizeof(int), 0, 0);
2337	ccv_cnnp_model_add_to_parameter_indices(parameters->model, to_param_sel, to_parameter_indices);
2338	*param_ref = parameters->param_ref > 0 ? parameters->param_ref - 1 : parameters->param_ref;
2339	return to_parameter_indices;
2340}
2341 
2342static 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)
2343{
2344	// If the model is not compiled yet. Compile them now.
2345	if (!model->graph)
2346	{
2347		model->graph = ccv_nnc_symbolic_graph_new();
2348		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", 2348, __extension__ __PRETTY_FUNCTION__
); }));
2349		const int input_size = from_model->input_size;
2350		ccv_nnc_tensor_param_t input_params[input_size];
2351		int i;
2352		for (i = 0; i < input_size; i++)
2353			input_params[i] = ccv_nnc_tensor_symbol_params(from_model->graph, from_model->inputs[i]);
2354		_ccv_cnnp_model_compile(model, input_params, input_size, from_model->compiled_data->loss);
2355		model->parallel_count = from_model->parallel_count;
2356		model->memory_compression = from_model->memory_compression;
2357		model->memory_reduction = from_model->memory_reduction;
2358		model->gradient_checkpointing = from_model->gradient_checkpointing;
2359		model->compiled_data->stream_type = from_model->compiled_data->stream_type;
2360		model->compiled_data->minimize.minimizer = from_model->compiled_data->minimize.minimizer;
2361		model->compiled_data->minimize.max_saved_aux_size = from_model->compiled_data->minimize.max_saved_aux_size;
2362	}
2363	ccv_cnnp_compiled_data_t* const to_compiled_data = model->compiled_data;
2364	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", 2364, __extension__ __PRETTY_FUNCTION__
); }));
2365	const int to_tensors_init = !!to_compiled_data->tensors_init.v;
2366	if (!to_tensors_init)
2367	{
2368		if (only_init_0)
2369			ccv_cnnp_model_tensors_init_0(model, to_compiled_data);
2370		else
2371			_ccv_cnnp_model_tensors_init(model, to_compiled_data);
2372	} else if (!only_init_0 && (uintptr_t)to_compiled_data->tensors_init.v & (uintptr_t)1)
2373		// Check if it is not fully allocated, if it is not, init_1.
2374			ccv_cnnp_model_tensors_init_1(model, to_compiled_data);
2375	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", 2375, __extension__ __PRETTY_FUNCTION__
); }));
2376	*parameter_indices = _ccv_cnnp_model_parameter_indices(model, parameters, param_ref);
2377	*from_parameter_indices = _ccv_cnnp_model_parameter_indices(from_model, from_parameters, from_param_ref);
2378	if (*from_param_ref < 0 && *param_ref >= 0)
2379		{ 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", 2379, __extension__ __PRETTY_FUNCTION__
); })); }
2380	else if (*from_param_ref >= 0)
2381		{ 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", 2381, __extension__ __PRETTY_FUNCTION__
); })); }
2382	if (*param_ref < 0 && *from_param_ref >= 0)
2383		{ 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"
, 2383, __extension__ __PRETTY_FUNCTION__); })); }
2384	else if (*param_ref >= 0)
2385		{ 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", 2385, __extension__ __PRETTY_FUNCTION__
); })); }
2386}
2387 
2388void 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)
2389{
2390	ccv_array_t* to_parameter_indices;
2391	int to_param_ref;
2392	ccv_array_t* from_parameter_indices;
2393	int from_param_ref;
2394	_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);
2395	// Should be exactly the same tensor.
2396	if (to_param_ref < 0 && from_param_ref < 0)
2397		{ 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", 2397, __extension__ __PRETTY_FUNCTION__
); })); }
2398	// To models.
2399	ccv_cnnp_compiled_data_t* const to_compiled_data = model->compiled_data;
2400	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", 2400, __extension__ __PRETTY_FUNCTION__
); }));
2401	// From models.
2402	const ccv_cnnp_compiled_data_t* const from_compiled_data = from_model->compiled_data;
2403	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; });
2404	const int to_parameter_size = to_compiled_data->parameters->rnum;
2405	const int rnum = (to_param_ref < 0 && from_param_ref < 0) ? from_parameter_indices->rnum : 1;
2406	int i, j;
2407	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));
2408  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));
2409	for (i = 0; i < rnum; i++)
2410	{
2411		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)));
2412		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"
, 2412, __extension__ __PRETTY_FUNCTION__); }));
2413		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", 2413, __extension__ __PRETTY_FUNCTION__
); }));
2414		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;
2415		// If the original is not init'ed. We cannot copy from.
2416		if (!(from_init_v[s >> 5] & (1u << (s & 0x1f))))
2417			continue;
2418		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)));
2419		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"
, 2419, __extension__ __PRETTY_FUNCTION__); }));
2420		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", 2420, __extension__ __PRETTY_FUNCTION__
); }));
2421		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));
2422		assert(src)((void) sizeof ((src) ? 1 : 0), __extension__ ({ if (src) ; else
 __assert_fail ("src", "ccv_cnnp_model.c", 2422, __extension__
 __PRETTY_FUNCTION__); }));
2423		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));
2424		assert(dest)((void) sizeof ((dest) ? 1 : 0), __extension__ ({ if (dest) ;
 else __assert_fail ("dest", "ccv_cnnp_model.c", 2424, __extension__
 __PRETTY_FUNCTION__); }));
2425		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);
2426		for (j = 1; j < parallel_count; j++)
2427		{
2428			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));
2429			if (copy_tensor)
2430				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);
2431		}
2432		// Mark this symbol as init'ed.
2433		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;
2434		to_init_v[d >> 5] |= (1u << (d & 0x1f));
2435	}
2436	ccv_array_free(to_parameter_indices);
2437	ccv_array_free(from_parameter_indices);
2438}
2439 
2440KHASH_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'→
2441 
2442void 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)
2443{
2444	ccv_array_t* to_parameter_indices;
2445	int to_param_ref;
2446	ccv_array_t* from_parameter_indices;
2447	int from_param_ref;
2448	_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);
2449	// Should be exactly the same tensor.
2450	if (renamer == 0 && to_param_ref < 0 && from_param_ref < 0)
1
Assuming 'renamer' is not equal to null→
2451		{ 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", 2451, __extension__ __PRETTY_FUNCTION__
); })); }
2452	// To models.
2453	ccv_cnnp_compiled_data_t* const to_compiled_data = model->compiled_data;
2454	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", 2454, __extension__ __PRETTY_FUNCTION__
); }));
2
←
Assuming 'to_compiled_data' is non-null→
3
←
Taking true branch→
2455	// From models.
2456	const ccv_cnnp_compiled_data_t* const from_compiled_data = from_model->compiled_data;
2457	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→
2458	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", 2458, __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→
2459	const int from_parameter_size = from_compiled_data->parameters->rnum;
2460	const int to_parameter_size = to_compiled_data->parameters->rnum;
2461	const int rnum = (to_param_ref < 0 && from_param_ref < 0) ? to_parameter_indices->rnum : 1;
9
←
Assuming 'to_param_ref' is >= 0→
2462	int i, j;
2463	khash_t(ccv_cnnp_parameter_id)kh_ccv_cnnp_parameter_id_t* id_map = 0;
2464	char* updated_name = 0;
2465	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));
2466  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));
2467	for (i = 0; i < rnum; i++)
2468	{
2469		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→
2470		// Need to figure out how to use the renamer here.
2471		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→
2472		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"
, 2472, __extension__ __PRETTY_FUNCTION__); }));
15
←
Assuming 'dest_d' is >= 0→
16
←
Taking true branch→
2473		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", 2473, __extension__
 __PRETTY_FUNCTION__); }));
17
←
Assuming 'dest_d' is < 'to_parameter_size'→
18
←
Taking true branch→
2474		if (renamer18.1
'renamer' is non-null
)
2475		{
2476			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;
2477			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)));
2478			if (!updated_name18.3
'updated_name' is null
)
19
←
Taking true branch→
2479				updated_name = (char*)ccmallocmalloc(1024);
2480			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→
2481			if (src_name_len20.1
'src_name_len' is <= 0
 > 0)
21
←
Taking false branch→
2482				memcpy(updated_name, src_name, src_name_len);
2483			updated_name[src_name_len] = 0;
2484			if (renamer(context, dest_name, updated_name, 1024) != 0)
22
←
Assuming the condition is false→
2485				continue; // Skip this.
2486			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)
2487			{
2488				// Nothing changed.
2489			} else {
2490				if (!id_map22.2
'id_map' is null
)
23
←
Taking true branch→
2491				{
2492					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'→
2493					for (j = 0; j < from_parameter_size; j++)
27
←
Assuming 'j' is >= 'from_parameter_size'→
28
←
Loop condition is false. Execution continues on line 2501→
2494					{
2495						int ret;
2496						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);
2497						assert(ret != 0)((void) sizeof ((ret != 0) ? 1 : 0), __extension__ ({ if (ret
 != 0) ; else __assert_fail ("ret != 0", "ccv_cnnp_model.c", 2497
, __extension__ __PRETTY_FUNCTION__); }));
2498						kh_val(id_map, k)((id_map)->vals[k]) = j;
2499					}
2500				}
2501				const khiter_t k = kh_get(ccv_cnnp_parameter_id, id_map, updated_name)kh_get_ccv_cnnp_parameter_id(id_map, updated_name);
2502				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→
2503					continue;
2504				src_d = kh_val(id_map, k)((id_map)->vals[k]);
31
←
Array access (via field 'vals') results in a null pointer dereference
2505				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"
, 2505, __extension__ __PRETTY_FUNCTION__); }));
2506				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", 2506, __extension__
 __PRETTY_FUNCTION__); }));
2507			}
2508		}
2509		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"
, 2509, __extension__ __PRETTY_FUNCTION__); }));
2510		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", 2510, __extension__
 __PRETTY_FUNCTION__); }));
2511		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;
2512		// If the original is not init'ed. We cannot share from.
2513		if (!(from_init_v[s >> 5] & (1u << (s & 0x1f))))
2514			continue;
2515		for (j = 0; j < parallel_count; j++)
2516		{
2517			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));
2518			assert(src)((void) sizeof ((src) ? 1 : 0), __extension__ ({ if (src) ; else
 __assert_fail ("src", "ccv_cnnp_model.c", 2518, __extension__
 __PRETTY_FUNCTION__); }));
2519			ccv_nnc_tensor_t* const dest = to_compiled_data->tensors.parameters[dest_d + j * to_parameter_size];
2520			if (dest && !((uintptr_t)dest & (uintptr_t)1))
2521				ccv_nnc_tensor_free(dest);
2522			to_compiled_data->tensors.parameters[dest_d + j * to_parameter_size] = (ccv_nnc_tensor_t*)((uintptr_t)src | (uintptr_t)1);
2523		}
2524		// Mark this symbol as init'ed.
2525		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;
2526		to_init_v[d >> 5] |= (1u << (d & 0x1f));
2527	}
2528	ccv_array_free(to_parameter_indices);
2529	ccv_array_free(from_parameter_indices);
2530	if (id_map)
2531		kh_destroy(ccv_cnnp_parameter_id, id_map)kh_destroy_ccv_cnnp_parameter_id(id_map);
2532	if (updated_name)
2533		ccfreefree(updated_name);
2534	// Mark it as incomplete so we will call init_1.
2535	if (ccv_cnnp_model_tensors_any_to_alloc(model, to_compiled_data))
2536		to_compiled_data->tensors_init.v = (uint32_t*)((uintptr_t)to_compiled_data->tensors_init.v | (uintptr_t)1);
2537	else // Remove the flag.
2538		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));
2539}
2540 
2541ccv_nnc_stream_context_t* ccv_cnnp_compiled_data_get_stream(ccv_cnnp_compiled_data_t* const compiled_data, const int type)
2542{
2543	if (!compiled_data->stream_map)
2544		compiled_data->stream_map = kh_init(stream_map)kh_init_stream_map();
2545	int ret = 0;
2546	khiter_t k = kh_put(stream_map, compiled_data->stream_map, type, &ret)kh_put_stream_map(compiled_data->stream_map, type, &ret
);
2547	assert(ret >= 0)((void) sizeof ((ret >= 0) ? 1 : 0), __extension__ ({ if (
ret >= 0) ; else __assert_fail ("ret >= 0", "ccv_cnnp_model.c"
, 2547, __extension__ __PRETTY_FUNCTION__); }));
2548	ccv_nnc_stream_context_t* stream = kh_val(compiled_data->stream_map, k)((compiled_data->stream_map)->vals[k]);
2549	// If ret == 0, the key already exist, we can return directly, otherwise, create and return.
2550	if (ret != 0)
2551	{
2552		stream = ccv_nnc_stream_context_new(type);
2553		kh_val(compiled_data->stream_map, k)((compiled_data->stream_map)->vals[k]) = stream;
2554	}
2555	return stream;
2556}
2557 
2558void 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)
2559{
2560	ccv_array_t* to_parameter_indices;
2561	int to_param_ref;
2562	ccv_array_t* from_parameter_indices;
2563	int from_param_ref;
2564	_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);
2565	// Should be exactly the same tensor.
2566	if (to_param_ref < 0 && from_param_ref < 0)
2567		{ 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", 2567, __extension__ __PRETTY_FUNCTION__
); })); }
2568	// To models.
2569	ccv_cnnp_compiled_data_t* const to_compiled_data = model->compiled_data;
2570	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", 2570, __extension__ __PRETTY_FUNCTION__
); }));
2571	// From models.
2572	const ccv_cnnp_compiled_data_t* const from_compiled_data = from_model->compiled_data;
2573	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; });
2574	const int to_parameter_size = to_compiled_data->parameters->rnum;
2575	const int rnum = (to_param_ref < 0 && from_param_ref < 0) ? from_parameter_indices->rnum : 1;
2576	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", 2576, __extension__ __PRETTY_FUNCTION__
); }));
2577	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", 2577, __extension__ __PRETTY_FUNCTION__
); }));
2578	int i, j;
2579	ccv_nnc_tensor_t* inputs[aux_in_size + 2];
2580	ccv_nnc_tensor_t* outputs[aux_out_size + 1];
2581	for (i = 0; i < aux_in_size; i++)
2582		inputs[i + 2] = aux_ins[i];
2583	for (i = 0; i < aux_out_size; i++)
2584		outputs[i + 1] = aux_outs[i];
2585	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));
2586  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));
2587	for (i = 0; i < rnum; i++)
2588	{
2589		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)));
2590		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"
, 2590, __extension__ __PRETTY_FUNCTION__); }));
2591		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", 2591, __extension__ __PRETTY_FUNCTION__
); }));
2592		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;
2593		// If the original is not init'ed. We cannot copy from.
2594		if (!(from_init_v[s >> 5] & (1u << (s & 0x1f))))
2595			continue;
2596		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)));
2597		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"
, 2597, __extension__ __PRETTY_FUNCTION__); }));
2598		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", 2598, __extension__ __PRETTY_FUNCTION__
); }));
2599		if (parallel_count > 1)
2600		{
2601			ccv_nnc_stream_context_t* streams[parallel_count];
2602			ccv_nnc_stream_signal_t* signal;
2603			if (stream_context)
2604				signal = ccv_nnc_stream_context_emit_signal_new(stream_context);
2605			for (j = 0; j < parallel_count; j++)
2606			{
2607				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));
2608				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));
2609				if (!dest || !src)
2610				{
2611					streams[j] = 0;
2612					continue;
2613				}
2614				// At the moment, can only handle them on the same device.
2615				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", 2615, __extension__ __PRETTY_FUNCTION__
); }));
2616				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", 2616, __extension__ __PRETTY_FUNCTION__
); }));
2617				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;
2618				const int device_id = CCV_TENSOR_GET_DEVICE_ID(src->info.type)(((src->info.type) & 0xfff00) >> 8);
2619				int type = stream_type;
2620				CCV_STREAM_SET_DEVICE_ID(type, device_id)(type) = (((type) & ~0xfff00) | (((device_id) & 0xfff
) << 8));
2621				ccv_nnc_stream_context_t* const stream_0 = ccv_cnnp_compiled_data_get_stream(to_compiled_data, type);
2622				// Wait signal to finish.
2623				if (stream_context)
2624					ccv_nnc_stream_context_wait_signal(stream_0, signal);
2625				inputs[0] = outputs[0] = dest;
2626				inputs[1] = src;
2627				ccv_nnc_cmd_exec(cmd, hint, flags, inputs, aux_in_size + 2, outputs, aux_out_size + 1, stream_0);
2628				if (stream_context)
2629				{
2630					ccv_nnc_stream_signal_t* const signal = ccv_nnc_stream_context_emit_signal_new(stream_0);
2631					ccv_nnc_stream_context_wait_signal(stream_context, signal);
2632				}
2633				streams[j] = stream_0;
2634			}
2635			// If this should be blocking, blocking it.
2636			if (!stream_context)
2637				for (j = 0; j < parallel_count; j++)
2638					if (streams[j])
2639						ccv_nnc_stream_context_wait(streams[j]);
2640		} else {
2641			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));
2642			assert(src)((void) sizeof ((src) ? 1 : 0), __extension__ ({ if (src) ; else
 __assert_fail ("src", "ccv_cnnp_model.c", 2642, __extension__
 __PRETTY_FUNCTION__); }));
2643			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));
2644			assert(dest)((void) sizeof ((dest) ? 1 : 0), __extension__ ({ if (dest) ;
 else __assert_fail ("dest", "ccv_cnnp_model.c", 2644, __extension__
 __PRETTY_FUNCTION__); }));
2645			inputs[0] = outputs[0] = dest;
2646			inputs[1] = src;
2647			ccv_nnc_cmd_exec(cmd, hint, flags, inputs, aux_in_size + 2, outputs, aux_out_size + 1, stream_context);
2648		}
2649		// Mark this symbol as init'ed.
2650		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;
2651		to_init_v[d >> 5] |= (1u << (d & 0x1f));
2652	}
2653	ccv_array_free(to_parameter_indices);
2654	ccv_array_free(from_parameter_indices);
2655}
2656 
2657void 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)
2658{
2659	int to_param_ref;
2660	ccv_array_t* const to_parameter_indices = _ccv_cnnp_model_parameter_indices(model, parameters, &to_param_ref);
2661	// To models.
2662	ccv_cnnp_compiled_data_t* const to_compiled_data = model->compiled_data;
2663	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", 2663, __extension__ __PRETTY_FUNCTION__
); }));
2664	// Tensor has to be inited already.
2665	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", 2665, __extension__ __PRETTY_FUNCTION__
); }));
2666	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", 2666, __extension__ __PRETTY_FUNCTION__
); }));
2667	// From models.
2668	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; });
2669	const int to_parameter_size = to_compiled_data->parameters->rnum;
2670	const int rnum = (to_param_ref < 0) ? to_parameter_indices->rnum : 1;
2671	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", 2671, __extension__ __PRETTY_FUNCTION__
); }));
2672	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", 2672, __extension__ __PRETTY_FUNCTION__
); }));
2673	int i, j;
2674	ccv_nnc_tensor_t* inputs[aux_in_size + 1];
2675	ccv_nnc_tensor_t* outputs[aux_out_size + 1];
2676	for (i = 0; i < aux_in_size; i++)
2677		inputs[i + 1] = aux_ins[i];
2678	for (i = 0; i < aux_out_size; i++)
2679		outputs[i + 1] = aux_outs[i];
2680	for (i = 0; i < rnum; i++)
2681	{
2682		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)));
2683		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"
, 2683, __extension__ __PRETTY_FUNCTION__); }));
2684		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", 2684, __extension__ __PRETTY_FUNCTION__
); }));
2685		if (parallel_count > 1)
2686		{
2687			ccv_nnc_stream_context_t* streams[parallel_count];
2688			ccv_nnc_stream_signal_t* signal;
2689			if (stream_context)
2690				signal = ccv_nnc_stream_context_emit_signal_new(stream_context);
2691			for (j = 0; j < parallel_count; j++)
2692			{
2693				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));
2694				if (!dest)
2695				{
2696					streams[j] = 0;
2697					continue;
2698				}
2699				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;
2700				const int device_id = CCV_TENSOR_GET_DEVICE_ID(dest->info.type)(((dest->info.type) & 0xfff00) >> 8);
2701				int type = stream_type;
2702				CCV_STREAM_SET_DEVICE_ID(type, device_id)(type) = (((type) & ~0xfff00) | (((device_id) & 0xfff
) << 8));
2703				ccv_nnc_stream_context_t* const stream_0 = ccv_cnnp_compiled_data_get_stream(to_compiled_data, type);
2704				// Wait signal to finish.
2705				if (stream_context)
2706					ccv_nnc_stream_context_wait_signal(stream_0, signal);
2707				inputs[0] = outputs[0] = dest;
2708				ccv_nnc_cmd_exec(cmd, hint, flags, inputs, aux_in_size + 1, outputs, aux_out_size + 1, stream_0);
2709				if (stream_context)
2710				{
2711					ccv_nnc_stream_signal_t* const signal = ccv_nnc_stream_context_emit_signal_new(stream_0);
2712					ccv_nnc_stream_context_wait_signal(stream_context, signal);
2713				}
2714				streams[j] = stream_0;
2715			}
2716			// If this should be blocking, blocking it.
2717			if (!stream_context)
2718				for (j = 0; j < parallel_count; j++)
2719					if (streams[j])
2720						ccv_nnc_stream_context_wait(streams[j]);
2721		} else {
2722			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));
2723			assert(dest)((void) sizeof ((dest) ? 1 : 0), __extension__ ({ if (dest) ;
 else __assert_fail ("dest", "ccv_cnnp_model.c", 2723, __extension__
 __PRETTY_FUNCTION__); }));
2724			inputs[0] = outputs[0] = dest;
2725			ccv_nnc_cmd_exec(cmd, hint, flags, inputs, aux_in_size + 1, outputs, aux_out_size + 1, stream_context);
2726		}
2727		// No need to mark this symbol as init'ed, it is already.
2728	}
2729	ccv_array_free(to_parameter_indices);
2730}
2731 
2732void 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)
2733{
2734	int to_param_ref;
2735	ccv_array_t* const to_parameter_indices = _ccv_cnnp_model_parameter_indices(model, parameters, &to_param_ref);
2736	// To models.
2737	ccv_cnnp_compiled_data_t* const to_compiled_data = model->compiled_data;
2738	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", 2738, __extension__ __PRETTY_FUNCTION__
); }));
2739	// Tensor has to be inited already.
2740	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", 2740, __extension__ __PRETTY_FUNCTION__
); }));
2741	ccv_nnc_tensor_t** tensor_gradients;
2742	if (to_compiled_data->backward.count > 1)
2743		tensor_gradients = to_compiled_data->tensors.accum_gradients;
2744	else
2745		tensor_gradients = to_compiled_data->tensors.gradients;
2746	assert(tensor_gradients)((void) sizeof ((tensor_gradients) ? 1 : 0), __extension__ ({
 if (tensor_gradients) ; else __assert_fail ("tensor_gradients"
, "ccv_cnnp_model.c", 2746, __extension__ __PRETTY_FUNCTION__
); }));
2747	// From models.
2748	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; });
2749	const int to_parameter_size = to_compiled_data->parameters->rnum;
2750	const int rnum = (to_param_ref < 0) ? to_parameter_indices->rnum : 1;
2751	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", 2751, __extension__ __PRETTY_FUNCTION__
); }));
2752	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", 2752, __extension__ __PRETTY_FUNCTION__
); }));
2753	int i, j;
2754	ccv_nnc_tensor_t* inputs[aux_in_size + 1];
2755	ccv_nnc_tensor_t* outputs[aux_out_size + 1];
2756	for (i = 0; i < aux_in_size; i++)
2757		inputs[i + 1] = aux_ins[i];
2758	for (i = 0; i < aux_out_size; i++)
2759		outputs[i + 1] = aux_outs[i];
2760	for (i = 0; i < rnum; i++)
2761	{
2762		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)));
2763		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"
, 2763, __extension__ __PRETTY_FUNCTION__); }));
2764		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", 2764, __extension__ __PRETTY_FUNCTION__
); }));
2765		if (parallel_count > 1)
2766		{
2767			ccv_nnc_stream_context_t* streams[parallel_count];
2768			ccv_nnc_stream_signal_t* signal;
2769			if (stream_context)
2770				signal = ccv_nnc_stream_context_emit_signal_new(stream_context);
2771			for (j = 0; j < parallel_count; j++)
2772			{
2773				ccv_nnc_tensor_t* const dest = tensor_gradients[dest_d + j * to_parameter_size];
2774				if (!dest)
2775				{
2776					streams[j] = 0;
2777					continue;
2778				}
2779				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;
2780				const int device_id = CCV_TENSOR_GET_DEVICE_ID(dest->info.type)(((dest->info.type) & 0xfff00) >> 8);
2781				int type = stream_type;
2782				CCV_STREAM_SET_DEVICE_ID(type, device_id)(type) = (((type) & ~0xfff00) | (((device_id) & 0xfff
) << 8));
2783				ccv_nnc_stream_context_t* const stream_0 = ccv_cnnp_compiled_data_get_stream(to_compiled_data, type);
2784				// Wait signal to finish.
2785				if (stream_context)
2786					ccv_nnc_stream_context_wait_signal(stream_0, signal);
2787				inputs[0] = outputs[0] = dest;
2788				ccv_nnc_cmd_exec(cmd, hint, flags, inputs, aux_in_size + 1, outputs, aux_out_size + 1, stream_0);
2789				if (stream_context)
2790				{
2791					ccv_nnc_stream_signal_t* const signal = ccv_nnc_stream_context_emit_signal_new(stream_0);
2792					ccv_nnc_stream_context_wait_signal(stream_context, signal);
2793				}
2794				streams[j] = stream_0;
2795			}
2796			// If this should be blocking, blocking it.
2797			if (!stream_context)
2798				for (j = 0; j < parallel_count; j++)
2799					if (streams[j])
2800						ccv_nnc_stream_context_wait(streams[j]);
2801		} else {
2802			ccv_nnc_tensor_t* const dest = tensor_gradients[dest_d];
2803			if (!dest)
2804				continue;
2805			assert(dest)((void) sizeof ((dest) ? 1 : 0), __extension__ ({ if (dest) ;
 else __assert_fail ("dest", "ccv_cnnp_model.c", 2805, __extension__
 __PRETTY_FUNCTION__); }));
2806			inputs[0] = outputs[0] = dest;
2807			ccv_nnc_cmd_exec(cmd, hint, flags, inputs, aux_in_size + 1, outputs, aux_out_size + 1, stream_context);
2808		}
2809		// No need to mark this symbol as init'ed, it is already.
2810	}
2811	ccv_array_free(to_parameter_indices);
2812}
2813 
2814ccv_nnc_cmd_t ccv_cnnp_model_minimizer(ccv_cnnp_model_t* const model)
2815{
2816	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2817	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 2817, __extension__ __PRETTY_FUNCTION__); }));
2818	return compiled_data->minimize.minimizer;
2819}
2820 
2821void 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)
2822{
2823	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2824	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 2824, __extension__ __PRETTY_FUNCTION__); }));
2825	const int parameter_size = compiled_data->parameters->rnum;
2826	if (parameter_size == 0)
2827		return;
2828	if (reset)
2829		{ 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", 2829, __extension__ __PRETTY_FUNCTION__
); })); }
2830	const int old_max_saved_aux_size = compiled_data->minimize.max_saved_aux_size;
2831	const int saved_aux_size = ccv_nnc_minimizer_saved_aux_size(minimizer);
2832	if (saved_aux_size > compiled_data->minimize.max_saved_aux_size)
2833		compiled_data->minimize.max_saved_aux_size = saved_aux_size;
2834	const int max_saved_aux_size = compiled_data->minimize.max_saved_aux_size;
2835	// We update all parameters, at this point, we have one minimizer.
2836	if (set_parameters == 0 || set_parameter_size == 0)
2837		compiled_data->minimize.minimizer = minimizer;
2838	int i;
2839	if (set_parameters && set_parameter_size)
2840	{
2841		// I need to save what's the minimizer along with this.
2842		if (!compiled_data->minimize.parameters)
2843			compiled_data->minimize.parameters = ccv_array_new(sizeof(ccv_cnnp_set_minimizer_for_parameter_t*), 1, 0);
2844		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));
2845		set_minimizer_for_parameter->minimizer = minimizer;
2846		set_minimizer_for_parameter->parameter_size = set_parameter_size;
2847		memcpy(set_minimizer_for_parameter->parameters, set_parameters, sizeof(ccv_cnnp_model_io_t) * set_parameter_size);
2848		ccv_array_push(compiled_data->minimize.parameters, &set_minimizer_for_parameter);
2849	}
2850	// If reset is true, clear the parameters array.
2851	if (reset && compiled_data->minimize.parameters)
2852	{
2853		for (i = 0; i < compiled_data->minimize.parameters->rnum; i++)
2854			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))));
2855		ccv_array_clear(compiled_data->minimize.parameters);
2856	}
2857	if (!compiled_data->update_nodes)
2858		return;
2859	ccv_nnc_symbolic_graph_t* const symbolic_graph = model->graph;
2860	assert(symbolic_graph)((void) sizeof ((symbolic_graph) ? 1 : 0), __extension__ ({ if
 (symbolic_graph) ; else __assert_fail ("symbolic_graph", "ccv_cnnp_model.c"
, 2860, __extension__ __PRETTY_FUNCTION__); }));
2861	if (saved_aux_size > old_max_saved_aux_size)
2862	{
2863		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", 2863, __extension__ __PRETTY_FUNCTION__
); }));
2864		// Reallocate first, move them around later.
2865		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);
2866		compiled_data->update_nodes = (ccv_nnc_graph_exec_symbol_t*)(compiled_data->updated_parameters + parameter_size);
2867		compiled_data->saved_aux = (ccv_nnc_tensor_symbol_map_t*)(compiled_data->update_nodes + parameter_size);
2868		// We need to do this from back to front because saved_aux_size > old_saved_aux_size, it could overlap.
2869		_ccv_cnnp_scatter_saved_aux(compiled_data->saved_aux, parameter_size, old_max_saved_aux_size, saved_aux_size);
2870	}
2871	int flag = 0;
2872	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; });
2873	if (set_parameters && set_parameter_size)
2874	{
2875		ccv_array_t* const parameter_indices = ccv_array_new(sizeof(int), 0, 0);
2876		for (i = 0; i < set_parameter_size; i++)
2877		{
2878			const int param_sel = set_parameters[i]->param_sel > 0 ? set_parameters[i]->param_sel - 1 : set_parameters[i]->param_sel;
2879			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", 2879, __extension__ __PRETTY_FUNCTION__
); }));
2880			const int old_rnum = parameter_indices->rnum;
2881			ccv_cnnp_model_add_to_parameter_indices(set_parameters[i]->model, param_sel, parameter_indices);
2882			const int param_ref = set_parameters[i]->param_ref > 0 ? set_parameters[i]->param_ref - 1 : set_parameters[i]->param_ref;
2883			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", 2883, __extension__ __PRETTY_FUNCTION__
); }));
2884			if (param_ref >= 0)
2885			{
2886				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", 2886, __extension__ __PRETTY_FUNCTION__
); }));
2887				*(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)));
2888				parameter_indices->rnum = old_rnum + 1;
2889			}
2890		}
2891		// We may have duplicated indices, but that is OK, we will set it twice.
2892		for (i = 0; i < parameter_indices->rnum; i++)
2893		{
2894			const int d = *(int*)ccv_array_get(parameter_indices, i)((void*)(((char*)((parameter_indices)->data)) + (size_t)(parameter_indices
)->rsize * (size_t)(i)));
2895			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))
2896				flag = 1;
2897		}
2898		ccv_array_free(parameter_indices);
2899	} else {
2900		for (i = 0; i < parameter_size; i++)
2901			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))
2902				flag = 1;
2903		if (compiled_data->minimize.parameters)
2904			if (_ccv_cnnp_apply_parameters_with_minimizer(model))
2905				flag = 1;
2906	}
2907	if (flag)
2908	{
2909		// 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.
2910		if (compiled_data->graph_mode == CCV_CNNP_MODEL_GRAPH_FIT_MODE)
2911			_ccv_cnnp_compiled_data_graph_free(compiled_data);
2912		_ccv_cnnp_compiled_data_apply_gradients_free(compiled_data);
2913	}
2914}
2915 
2916void ccv_cnnp_model_set_compile_params(ccv_cnnp_model_t* const model, const ccv_nnc_symbolic_graph_compile_param_t compile_params)
2917{
2918	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2919	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 2919, __extension__ __PRETTY_FUNCTION__); }));
2920	compiled_data->compile_params = compile_params;
2921}
2922 
2923void ccv_cnnp_model_dot(const ccv_cnnp_model_t* const model, const int flags, FILE** const outs, const int out_size)
2924{
2925	if (model->graph && out_size > 0)
2926		ccv_nnc_symbolic_graph_dot(model->graph, flags, outs[0]);
2927	if (model->compiled_data && model->compiled_data->graph && out_size > 1)
2928		ccv_nnc_graph_dot(model->compiled_data->graph, flags, outs[1]);
2929	if (model->compiled_data && model->compiled_data->backward.accum && out_size > 2)
2930		ccv_nnc_graph_dot(model->compiled_data->backward.accum, flags, outs[2]);
2931	if (model->compiled_data && model->compiled_data->apply_gradients.graph && out_size > 3)
2932		ccv_nnc_graph_dot(model->compiled_data->apply_gradients.graph, flags, outs[3]);
2933}
2934 
2935void ccv_cnnp_model_format(const ccv_cnnp_model_t* const model, const ccv_nnc_symbolic_graph_format_f format_fn, void* const context)
2936{
2937	if (model->graph)
2938		ccv_nnc_symbolic_graph_format(model->graph, 0, 0, 0, 0, format_fn, context);
2939}
2940 
2941static void _ccv_cnnp_compiled_data_free(const ccv_cnnp_model_t* const model, ccv_cnnp_compiled_data_t* const compiled_data)
2942{
2943	int i;
2944	const int parameter_size = compiled_data->parameters->rnum;
2945	ccv_array_free(compiled_data->parameters);
2946	if (compiled_data->parameter_flags)
2947		ccfreefree(compiled_data->parameter_flags);
2948	const int internal_size = compiled_data->internals->rnum;
2949	ccv_array_free(compiled_data->internals);
2950	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", 2950, __extension__ __PRETTY_FUNCTION__
); }));
2951	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", 2951, __extension__ __PRETTY_FUNCTION__
); }));
2952	for (i = 0; i < parameter_size; i++)
2953		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))));
2954	ccv_array_free(compiled_data->ids.parameters);
2955	for (i = 0; i < internal_size; i++)
2956		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))));
2957	ccv_array_free(compiled_data->ids.internals);
2958	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; });
2959	if (compiled_data->tensors.parameters)
2960	{
2961		for (i = 0; i < parameter_size * parallel_count; i++)
2962			// If it is not marked as not belonging, we can free it.
2963			if (!((uintptr_t)compiled_data->tensors.parameters[i] & (uintptr_t)1))
2964				if (compiled_data->tensors.parameters[i])
2965					ccv_nnc_tensor_free(compiled_data->tensors.parameters[i]);
2966		for (i = 0; i < internal_size * parallel_count; i++)
2967			if (compiled_data->tensors.internals[i])
2968				ccv_nnc_tensor_free(compiled_data->tensors.internals[i]);
2969		ccfreefree(compiled_data->tensors.parameters);
2970	}
2971	if (compiled_data->tensors.gradients)
2972	{
2973		for (i = 0; i < parameter_size * parallel_count; i++)
2974		{
2975			if (compiled_data->tensors.gradients[i])
2976				ccv_nnc_tensor_free(compiled_data->tensors.gradients[i]);
2977			if (compiled_data->tensors.accum_gradients[i])
2978				ccv_nnc_tensor_free(compiled_data->tensors.accum_gradients[i]);
2979		}
2980		ccfreefree(compiled_data->tensors.gradients);
2981	}
2982	if (compiled_data->minimize.parameters)
2983	{
2984		for (i = 0; i < compiled_data->minimize.parameters->rnum; i++)
2985			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))));
2986		ccv_array_free(compiled_data->minimize.parameters);
2987	}
2988	if (compiled_data->rewindables)
2989		ccv_array_free(compiled_data->rewindables);
2990	if (compiled_data->tensors_init.v)
2991		ccfreefree(CCV_NNC_INIT_V(compiled_data->tensors_init.v)((uint32_t*)((uintptr_t)(compiled_data->tensors_init.v) &
 ~(uintptr_t)1)));
2992	if (compiled_data->evaluate.tos)
2993		ccfreefree(compiled_data->evaluate.tos);
2994	compiled_data->evaluate.tos = 0;
2995	if (compiled_data->stream_map)
2996	{
2997		khiter_t k;
2998		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)
2999		{
3000			if (!kh_exist(compiled_data->stream_map, k)(!(((compiled_data->stream_map)->flags[(k)>>4]>>
(((k)&0xfU)<<1))&3)))
3001				continue;
3002			ccv_nnc_stream_context_t* const stream = kh_val(compiled_data->stream_map, k)((compiled_data->stream_map)->vals[k]);
3003			ccv_nnc_stream_context_free(stream);
3004		}
3005		kh_destroy(stream_map, compiled_data->stream_map)kh_destroy_stream_map(compiled_data->stream_map);
3006	}
3007	_ccv_cnnp_compiled_data_graph_free(compiled_data);
3008	_ccv_cnnp_compiled_data_gradient_free(compiled_data);
3009	_ccv_cnnp_compiled_data_backward_free(compiled_data);
3010	_ccv_cnnp_compiled_data_apply_gradients_free(compiled_data);
3011	if (compiled_data->gradient_checkpoints)
3012	{
3013		for (i = 0; i < compiled_data->gradient_checkpoints->rnum; i++)
3014		{
3015			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)));
3016			assert(checkpoint->inputs)((void) sizeof ((checkpoint->inputs) ? 1 : 0), __extension__
 ({ if (checkpoint->inputs) ; else __assert_fail ("checkpoint->inputs"
, "ccv_cnnp_model.c", 3016, __extension__ __PRETTY_FUNCTION__
); }));
3017			ccfreefree(checkpoint->inputs);
3018			ccv_array_free(checkpoint->tensor_symbols);
3019		}
3020		ccv_array_free(compiled_data->gradient_checkpoints);
3021	}
3022	ccv_nnc_xpu_alloc_destroy(&compiled_data->xpu_alloc);
3023	ccfreefree(compiled_data);
3024}
3025 
3026void ccv_cnnp_model_free(ccv_cnnp_model_t* const model)
3027{
3028	if (model->isa->deinit)
3029		model->isa->deinit(model);
3030	if (model->io)
3031	{
3032		int i;
3033		for (i = 0; i < model->io->rnum; i++)
3034		{
3035			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)));
3036			if (model_io->outgoings)
3037				ccv_array_free(model_io->outgoings);
3038			if (model_io->incomings)
3039				ccv_array_free(model_io->incomings);
3040			if (model_io->dependencies)
3041				ccv_array_free(model_io->dependencies);
3042			ccfreefree(model_io);
3043		}
3044		ccv_array_free(model->io);
3045	}
3046	if (model->parameter_indices)
3047		ccv_array_free(model->parameter_indices);
3048	if (model->inputs)
3049		ccfreefree(model->inputs);
3050	if (model->graph)
3051		ccv_nnc_symbolic_graph_free(model->graph);
3052	if (model->compiled_data)
3053		_ccv_cnnp_compiled_data_free(model, model->compiled_data);
3054	if (model->name)
3055		ccfreefree(model->name);
3056	ccfreefree(model);
3057}
3058 
3059void ccv_cnnp_model_cancel(ccv_cnnp_model_t* const model)
3060{
3061	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
3062	if (!compiled_data)
3063		return;
3064	if (compiled_data->graph)
3065		ccv_nnc_graph_cancel(compiled_data->graph);
3066	if (compiled_data->apply_gradients.graph)
3067		ccv_nnc_graph_cancel(compiled_data->apply_gradients.graph);
3068}