ccv_cnnp_model.c

Bug Summary

File:	nnc/ccv_cnnp_model.c
Warning:	line 2379, column 1 1st function call argument is an uninitialized value
Annotated Source Code

Press '?' to see keyboard shortcuts
Show analyzer invocation
clang -cc1 -cc1 -triple x86_64-unknown-linux-gnu -analyze -disable-free -clear-ast-before-backend -disable-llvm-verifier -discard-value-names -main-file-name ccv_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/18 -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 -D HAVE_CUDA_SM80 -I /usr/local/include -internal-isystem /usr/local/lib/clang/18/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/2024-07-02-165253-122966-1 -x c ccv_cnnp_model.c
1#include "ccv_nnc.h"
2#include "ccv_nnc_easy.h"
3#include "ccv_nnc_internal.h"
4#include "ccv_internal.h"
5#include "_ccv_cnnp_model.h"
6 
7// MARK - Level-5 API
8 
9ccv_cnnp_model_io_t ccv_cnnp_model_apply(ccv_cnnp_model_t* const model, const ccv_cnnp_model_io_t* const inputs, const int input_size)
10{
11	if (!model->io)
12		model->io = ccv_array_new(sizeof(ccv_cnnp_model_io_t), 1, 0);
13	ccv_cnnp_model_io_t model_io = ccmallocmalloc(sizeof(struct ccv_cnnp_model_io_s) + sizeof(ccv_nnc_tensor_symbol_t) * model->output_size);
14	model_io->param_ref = 0;
15	model_io->param_sel = 0;
16	model_io->visit = 0;
17	model_io->model = model;
18	model_io->dependencies = 0;
19	model_io->dependents = 0;
20	model_io->outgoings = 0;
21	model_io->outputs = (ccv_nnc_tensor_symbol_t*)(model_io + 1);
22	ccv_array_push(model->io, &model_io);
23	if (input_size > 0)
24	{
25		model_io->incomings = ccv_array_new(sizeof(ccv_cnnp_model_io_t), input_size, 0);
26		ccv_array_resize(model_io->incomings, input_size);
27		int i;
28		memcpy(ccv_array_get(model_io->incomings, 0)((void*)(((char*)((model_io->incomings)->data)) + (size_t
)(model_io->incomings)->rsize * (size_t)(0))), inputs, sizeof(ccv_cnnp_model_io_t) * input_size);
29		for (i = 0; i < input_size; i++)
30		{
31			if (!inputs[i]->outgoings)
32				inputs[i]->outgoings = ccv_array_new(sizeof(ccv_cnnp_model_io_t), 1, 0);
33			ccv_array_push(inputs[i]->outgoings, &model_io);
34		}
35	} else {
36		model_io->incomings = 0;
37	}
38	return model_io;
39}
40 
41void ccv_cnnp_model_add_dependencies(ccv_cnnp_model_io_t model_io, const ccv_cnnp_model_io_t* const dependencies, const int dependency_size)
42{
43	assert(dependency_size > 0)((void) sizeof ((dependency_size > 0) ? 1 : 0), __extension__
 ({ if (dependency_size > 0) ; else __assert_fail ("dependency_size > 0"
, "ccv_cnnp_model.c", 43, __extension__ __PRETTY_FUNCTION__);
 }));
44	if (!model_io->dependencies)
45		model_io->dependencies = ccv_array_new(sizeof(ccv_cnnp_model_io_t), dependency_size, 0);
46	int i, j;
47	for (i = 0; i < dependency_size; i++)
48	{
49		int flag = 0;
50		// Check if it is already exist or not.
51		for (j = 0; !flag && j < model_io->dependencies->rnum; j++)
52			if (*(ccv_cnnp_model_io_t*)ccv_array_get(model_io->dependencies, j)((void*)(((char*)((model_io->dependencies)->data)) + (size_t
)(model_io->dependencies)->rsize * (size_t)(j))) == dependencies[i])
53				flag = 1;
54		if (flag)
55			continue;
56		ccv_array_push(model_io->dependencies, dependencies + i);
57		++dependencies[i]->dependents;
58	}
59}
60 
61int ccv_cnnp_model_output_size(const ccv_cnnp_model_t* const model)
62{
63	return model->output_size;
64}
65 
66int ccv_cnnp_model_is_trainable(const ccv_cnnp_model_t* const model)
67{
68	// If the model is compiled, it is default to 1 unless it is not.
69	if (model->compiled_data)
70		return model->is_trainable >= 0 ? model->is_trainable : 1;
71	return model->is_trainable;
72}
73 
74ccv_cnnp_model_io_t ccv_cnnp_model_parameters(ccv_cnnp_model_t* const model, const int selector, const int index)
75{
76	if (!model->io)
77		model->io = ccv_array_new(sizeof(ccv_cnnp_model_io_t), 1, 0);
78	ccv_cnnp_model_io_t model_io = ccmallocmalloc(sizeof(struct ccv_cnnp_model_io_s));
79	model_io->param_ref = index >= 0 ? index + 1 : ALL_PARAMETERS-1;
80	model_io->param_sel = selector >= 0 ? selector + 1 : ALL_PARAMETERS-1;
81	model_io->visit = 0;
82	model_io->model = model;
83	model_io->outputs = 0;
84	model_io->dependencies = 0;
85	model_io->dependents = 0;
86	model_io->incomings = 0;
87	model_io->outgoings = 0;
88	ccv_array_push(model->io, &model_io);
89	return model_io;
90}
91 
92void ccv_cnnp_model_notify_hook(ccv_cnnp_model_t* const model, ccv_cnnp_model_notify_f func, void* const context)
93{
94	model->notify_hook.func = func;
95	model->notify_hook.context = context;
96}
97 
98void ccv_cnnp_model_notify(const ccv_cnnp_model_t* const model, const int tag, void* const payload)
99{
100	if (model->notify_hook.func)
101		model->notify_hook.func(model, tag, payload, model->notify_hook.context);
102	if (model->isa->notify)
103		model->isa->notify(model, tag, payload);
104}
105 
106static int _ccv_nnc_array_dedup_graph_exec_symbols(ccv_nnc_graph_exec_symbol_t* const graph_exec_symbols, int graph_exec_symbol_size)
107{
108	int i, j;
109	for (i = 0; i < graph_exec_symbol_size; i++)
110	{
111		ccv_nnc_graph_exec_symbol_t* const graph_exec_symbol = graph_exec_symbols + i;
112		// Check whether this tensor symbol has any duplicate.
113		for (j = i + 1; j < graph_exec_symbol_size;)
114		{
115			ccv_nnc_graph_exec_symbol_t* const other_symbol = graph_exec_symbols + j;
116			// If there is a same tensor symbol, remove it.
117			if (other_symbol->d == graph_exec_symbol->d && other_symbol->graph == graph_exec_symbol->graph)
118			{
119				if (j + 1 < graph_exec_symbol_size)
120					*other_symbol = graph_exec_symbols[graph_exec_symbol_size - 1];
121				--graph_exec_symbol_size;
122				continue;
123			}
124			++j;
125		}
126	}
127	return graph_exec_symbol_size;
128}
129 
130void ccv_cnnp_model_add_to_array(void* const context, const ccv_nnc_tensor_symbol_t symbol, const int is_trainable)
131{
132	ccv_cnnp_model_add_to_array_context_t* const add_to_array_context = (ccv_cnnp_model_add_to_array_context_t*)context;
133	ccv_cnnp_model_t* const model = add_to_array_context->sequence->model;
134	int i;
135	if (!model->parameter_indices)
136		model->parameter_indices = ccv_array_new(sizeof(int), 0, 0);
137	for (i = 0; i < add_to_array_context->symbols->rnum; i++)
138	{
139		const ccv_nnc_tensor_symbol_t other_symbol = *(ccv_nnc_tensor_symbol_t*)ccv_array_get(add_to_array_context->symbols, i)((void*)(((char*)((add_to_array_context->symbols)->data
)) + (size_t)(add_to_array_context->symbols)->rsize * (
size_t)(i)));
140		if (other_symbol.d == symbol.d && other_symbol.graph == symbol.graph)
141		{
142			// Only add to parameter_indices if it is trainable.
143			if (add_to_array_context->prefix == 't')
144				ccv_array_add_unique_int(model->parameter_indices, i);
145			// Found it, return, don't add it.
146			return;
147		}
148	}
149	// Only add to parameter_indices if it is trainable.
150	if (add_to_array_context->prefix == 't')
151		ccv_array_push(model->parameter_indices, &add_to_array_context->symbols->rnum);
152	// This is a new one, no need to add_unique_int, it is unique.
153	ccv_array_push(add_to_array_context->symbols, &symbol);
154	if (add_to_array_context->trainables)
155		ccv_array_push(add_to_array_context->trainables, &is_trainable);
156	char id[2048];
157	id[0] = add_to_array_context->prefix;
158	id[1] = '-';
159	int total_len = 2;
160	for (i = 0; i < add_to_array_context->sequence->sequences->rnum; i++)
161	{
162		const ccv_cnnp_model_name_t* const name = (ccv_cnnp_model_name_t*)ccv_array_get(add_to_array_context->sequence->sequences, i)((void*)(((char*)((add_to_array_context->sequence->sequences
)->data)) + (size_t)(add_to_array_context->sequence->
sequences)->rsize * (size_t)(i)));
163		int len;
164		if (name->name && name->name[0] != '\0')
165			len = snprintf(id + total_len, 2048 - total_len, "%s-%d-", name->name, name->sequence);
166		else
167			len = snprintf(id + total_len, 2048 - total_len, "%d-", name->sequence);
168		total_len += len;
169		if (total_len >= 2047)
170			break;
171	}
172	if (total_len < 2047)
173		total_len += snprintf(id + total_len, 2048 - total_len, "%d", add_to_array_context->sequence->it);
174	assert(total_len < 2048)((void) sizeof ((total_len < 2048) ? 1 : 0), __extension__
 ({ if (total_len < 2048) ; else __assert_fail ("total_len < 2048"
, "ccv_cnnp_model.c", 174, __extension__ __PRETTY_FUNCTION__)
; }));
175	char *heap_id = (char*)ccmallocmalloc(total_len + 1);
176	memcpy(heap_id, id, total_len + 1);
177	ccv_array_push(add_to_array_context->ids, &heap_id);
178	++add_to_array_context->sequence->it;
179}
180 
181static void _ccv_cnnp_compiled_data_init(ccv_cnnp_compiled_data_t* const compiled_data, const int output_size, ccv_array_t* const gradient_checkpoints)
182{
183	compiled_data->f = compiled_data->fits + output_size;
184	compiled_data->xpu_alloc.mp_hdr = -1;
185	compiled_data->xpu_alloc.freed = kh_init(dy_str)kh_init_dy_str();
186	compiled_data->xpu_alloc.allocd = kh_init(dy_alloc)kh_init_dy_alloc();
187	compiled_data->gradient_checkpoints = gradient_checkpoints;
188}
189 
190static void _ccv_cnnp_model_compile(ccv_cnnp_model_t* const model, const ccv_nnc_tensor_param_t* const inputs, const int input_size, const ccv_nnc_cmd_t loss)
191{
192	assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if
 (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c"
, 192, __extension__ __PRETTY_FUNCTION__); }));
193	model->inputs = ccmallocmalloc(sizeof(ccv_nnc_tensor_symbol_t) * input_size);
194	int i;
195	for (i = 0; i < input_size; i++)
196		model->inputs[i] = ccv_nnc_tensor_symbol_new(model->graph, inputs[i], 0);
197	ccv_array_t* const parameters = ccv_array_new(sizeof(ccv_nnc_tensor_symbol_t), 0, 0);
198	ccv_array_t* const parameter_ids = ccv_array_new(sizeof(char*), 0, 0);
199	ccv_array_t* const parameter_trainables = ccv_array_new(sizeof(int), 0, 0);
200	ccv_cnnp_model_sequence_t model_sequence = {
201		.bank = kh_init(ccv_cnnp_model_name_bank)kh_init_ccv_cnnp_model_name_bank()
202	};
203	ccv_cnnp_model_add_to_array_context_t add_to_parameter_context = {
204		.sequence = &model_sequence,
205		.prefix = 't',
206		.symbols = parameters,
207		.ids = parameter_ids,
208		.trainables = parameter_trainables,
209	};
210	ccv_array_t* const internals = ccv_array_new(sizeof(ccv_nnc_tensor_symbol_t), 0, 0);
211	ccv_array_t* const internal_ids = ccv_array_new(sizeof(char*), 0, 0);
212	ccv_cnnp_model_add_to_array_context_t add_to_output_context = {
213		.sequence = &model_sequence,
214		.prefix = 'r',
215		.symbols = internals,
216		.ids = internal_ids,
217		.trainables = 0,
218	};
219	ccv_cnnp_model_build_data_t build_data = {
220		.is_trainable = model->is_trainable >= 0 ? model->is_trainable : 1,
221		.model_sequence = &model_sequence,
222		.add_to_array = ccv_cnnp_model_add_to_array,
223		.parameters = parameters,
224		.context = {
225			.add_to_parameter = &add_to_parameter_context,
226			.add_to_output = &add_to_output_context,
227		},
228		.gradient_checkpoints = 0,
229	};
230	model->data = &build_data;
231	ccv_cnnp_model_build(model, model->graph, model->inputs, input_size, 0, 0);
232	for (i = 0; i < model->output_size; i++)
233	{
234		const ccv_nnc_tensor_symbol_t output = model->outputs[i];
235		const ccv_nnc_tensor_symbol_t alias_to = ccv_nnc_tensor_symbol_alias_to(model->graph, output);
236		if (alias_to.d == CCV_NNC_NO_TENSOR_SYMBOL)
237			continue;
238		// If output is an alias, insert data transform regardless for result correctness (we cannot bind an alias). You can check ccv_nnc_tensor_bind_symbol method
239		// to see that we can correctly bind a tensor which from it, has aliases, but we cannot bind an alias tensor correctly (this is expected, sort of, to be
240		// honest, because we cannot handle cases of alias is part of the original tensor but bind differently).
241		const ccv_nnc_tensor_param_t output_params = ccv_nnc_tensor_symbol_params(model->graph, output);
242		model->outputs[i] = ccv_nnc_tensor_symbol_new(model->graph, output_params, 0);
243		ccv_nnc_graph_exec_symbol_t make_contiguous = ccv_nnc_graph_exec_symbol_new(model->graph, CMD_FORMAT_TRANSFORM_FORWARD()ccv_nnc_cmd(CCV_NNC_FORMAT_TRANSFORM_FORWARD, 0, ccv_nnc_cmd_auto
, 0), &output, 1, model->outputs + i, 1, "contiguous");
244		ccv_nnc_graph_exec_symbol_set_flags(model->graph, make_contiguous, CCV_NNC_GRAPH_EXEC_DISABLE_OPT);
245	}
246	model->data = 0;
247	kh_destroy(ccv_cnnp_model_name_bank, model_sequence.bank)kh_destroy_ccv_cnnp_model_name_bank(model_sequence.bank);
248	if (model_sequence.sequences)
249		ccv_array_free(model_sequence.sequences);
250	// Check if there are parameters that are not trainables. If there are, we will allocate uint64 bitmap to record that.
251	int not_trainables = 0;
252	// Assert no parameter is alias.
253	for (i = 0; i < parameters->rnum; i++)
254	{
255		const ccv_nnc_tensor_symbol_t parameter = *(ccv_nnc_tensor_symbol_t*)ccv_array_get(parameters, i)((void*)(((char*)((parameters)->data)) + (size_t)(parameters
)->rsize * (size_t)(i)));
256		const ccv_nnc_tensor_symbol_t alias_to = ccv_nnc_tensor_symbol_alias_to(parameter.graph, parameter);
257		assert(alias_to.graph == 0)((void) sizeof ((alias_to.graph == 0) ? 1 : 0), __extension__
 ({ if (alias_to.graph == 0) ; else __assert_fail ("alias_to.graph == 0"
, "ccv_cnnp_model.c", 257, __extension__ __PRETTY_FUNCTION__)
; })); // Cannot find the one alias to.
258		if (*(int*)ccv_array_get(parameter_trainables, i)((void*)(((char*)((parameter_trainables)->data)) + (size_t
)(parameter_trainables)->rsize * (size_t)(i))) == 0)
259			not_trainables = 1;
260	}
261	assert(parameters->rnum == parameter_trainables->rnum)((void) sizeof ((parameters->rnum == parameter_trainables->
rnum) ? 1 : 0), __extension__ ({ if (parameters->rnum == parameter_trainables
->rnum) ; else __assert_fail ("parameters->rnum == parameter_trainables->rnum"
, "ccv_cnnp_model.c", 261, __extension__ __PRETTY_FUNCTION__)
; }));
262	uint64_t* parameter_flags = 0;
263	if (not_trainables)
264	{
265		parameter_flags = (uint64_t*)cccalloccalloc(((parameters->rnum + 63) >> 6), sizeof(uint64_t));
266		for (i = 0; i < parameter_trainables->rnum; i++)
267			if (*(int*)ccv_array_get(parameter_trainables, i)((void*)(((char*)((parameter_trainables)->data)) + (size_t
)(parameter_trainables)->rsize * (size_t)(i))))
268				parameter_flags[i >> 6] |= ((uint64_t)1 << (i & 63));
269	}
270	ccv_array_free(parameter_trainables);
271	// Assert no internal is alias.
272	for (i = 0; i < internals->rnum; i++)
273	{
274		const ccv_nnc_tensor_symbol_t internal = *(ccv_nnc_tensor_symbol_t*)ccv_array_get(internals, i)((void*)(((char*)((internals)->data)) + (size_t)(internals
)->rsize * (size_t)(i)));
275		const ccv_nnc_tensor_symbol_t alias_to = ccv_nnc_tensor_symbol_alias_to(internal.graph, internal);
276		assert(alias_to.graph == 0)((void) sizeof ((alias_to.graph == 0) ? 1 : 0), __extension__
 ({ if (alias_to.graph == 0) ; else __assert_fail ("alias_to.graph == 0"
, "ccv_cnnp_model.c", 276, __extension__ __PRETTY_FUNCTION__)
; })); // Cannot find the one alias to.
277	}
278	const int output_size = model->output_size;
279	ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
280	const int parameters_rnum = parameters->rnum;
281	if (input_size > 0)
282	{
283		ccv_array_resize(parameters, parameters_rnum + input_size);
284		memcpy(ccv_array_get(parameters, parameters_rnum)((void*)(((char*)((parameters)->data)) + (size_t)(parameters
)->rsize * (size_t)(parameters_rnum))), model->inputs, input_size * sizeof(ccv_nnc_tensor_symbol_t));
285	}
286	ccv_nnc_symbolic_graph_simplify(model->graph,
287		SYMBOLIC_GRAPH_PASSES(CCV_NNC_SIMPLIFY_COMMON_SUBEXPRESSION_ELIMINATION,(const int []){CCV_NNC_SIMPLIFY_COMMON_SUBEXPRESSION_ELIMINATION
, CCV_NNC_SIMPLIFY_DATA_TRANSFER_OPT, CCV_NNC_SIMPLIFY_OPS_FUSION
, CCV_NNC_SIMPLIFY_GRAPH_PRUNING}, (1 +1 +1 +1 +1 +0 +0 +0 +0
 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1)
288			CCV_NNC_SIMPLIFY_DATA_TRANSFER_OPT,(const int []){CCV_NNC_SIMPLIFY_COMMON_SUBEXPRESSION_ELIMINATION
, CCV_NNC_SIMPLIFY_DATA_TRANSFER_OPT, CCV_NNC_SIMPLIFY_OPS_FUSION
, CCV_NNC_SIMPLIFY_GRAPH_PRUNING}, (1 +1 +1 +1 +1 +0 +0 +0 +0
 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1)
289			CCV_NNC_SIMPLIFY_OPS_FUSION,(const int []){CCV_NNC_SIMPLIFY_COMMON_SUBEXPRESSION_ELIMINATION
, CCV_NNC_SIMPLIFY_DATA_TRANSFER_OPT, CCV_NNC_SIMPLIFY_OPS_FUSION
, CCV_NNC_SIMPLIFY_GRAPH_PRUNING}, (1 +1 +1 +1 +1 +0 +0 +0 +0
 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1)
290			CCV_NNC_SIMPLIFY_GRAPH_PRUNING)(const int []){CCV_NNC_SIMPLIFY_COMMON_SUBEXPRESSION_ELIMINATION
, CCV_NNC_SIMPLIFY_DATA_TRANSFER_OPT, CCV_NNC_SIMPLIFY_OPS_FUSION
, CCV_NNC_SIMPLIFY_GRAPH_PRUNING}, (1 +1 +1 +1 +1 +0 +0 +0 +0
 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1),
291		ccv_array_get(parameters, 0)((void*)(((char*)((parameters)->data)) + (size_t)(parameters
)->rsize * (size_t)(0))), parameters_rnum + input_size,
292		model->outputs, output_size,
293		SYMBOLIC_GRAPH_SOURCES(model->graph)ccv_nnc_symbolic_graph_sources(model->graph), ccv_nnc_symbolic_graph_source_size
(model->graph), SYMBOLIC_GRAPH_DESTINATIONS(model->graph)ccv_nnc_symbolic_graph_destinations(model->graph), ccv_nnc_symbolic_graph_destination_size
(model->graph));
294	ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
295	// Size it down.
296	parameters->rnum = parameters_rnum;
297	ccv_cnnp_compiled_data_t* compiled_data = model->compiled_data = cccalloccalloc(1, sizeof(ccv_cnnp_compiled_data_t) + sizeof(ccv_nnc_tensor_symbol_t) * (output_size * 2 - 1));
298	_ccv_cnnp_compiled_data_init(compiled_data, output_size, build_data.gradient_checkpoints);
299	const int evaluate_to_size = compiled_data->evaluate.to_size = ccv_nnc_symbolic_graph_destination_size(model->graph);
300	assert(evaluate_to_size > 0)((void) sizeof ((evaluate_to_size > 0) ? 1 : 0), __extension__
 ({ if (evaluate_to_size > 0) ; else __assert_fail ("evaluate_to_size > 0"
, "ccv_cnnp_model.c", 300, __extension__ __PRETTY_FUNCTION__)
; }));
301	compiled_data->evaluate.tos = ccmallocmalloc(sizeof(ccv_nnc_graph_exec_symbol_t) * evaluate_to_size);
302	memcpy(compiled_data->evaluate.tos, ccv_nnc_symbolic_graph_destinations(model->graph), sizeof(ccv_nnc_graph_exec_symbol_t) * evaluate_to_size);
303	compiled_data->loss = loss;
304	if (loss.cmd == CCV_NNC_NOOP)
305	{
306		// If no loss function provided, there is no fits.
307		for (i = 0; i < output_size; i++)
308		{
309			compiled_data->fits[i] = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
310			const ccv_nnc_tensor_symbol_t alias_to = ccv_nnc_tensor_symbol_alias_to(model->graph, model->outputs[i]);
311			if (alias_to.d < 0)
312				compiled_data->f[i] = model->outputs[i];
313			else { // We cannot differentiate against an alias, therefore, we have to verify this output is full, and we can diff against the original.
314				int ofs[CCV_NNC_MAX_DIM_ALLOC(12)];
315				int inc[CCV_NNC_MAX_DIM_ALLOC(12)];
316				ccv_nnc_tensor_symbol_alias_params(model->graph, model->outputs[i], ofs, inc);
317				int j;
318				for (j = 0; j < CCV_NNC_MAX_DIM_ALLOC(12); j++)
319					{ assert(ofs[j] == 0)((void) sizeof ((ofs[j] == 0) ? 1 : 0), __extension__ ({ if (
ofs[j] == 0) ; else __assert_fail ("ofs[j] == 0", "ccv_cnnp_model.c"
, 319, __extension__ __PRETTY_FUNCTION__); })); } // There is no ofs.
320				compiled_data->f[i] = alias_to; // Unfortunately, I cannot assert the size yet.
321			}
322		}
323	} else {
324		for (i = 0; i < output_size; i++)
325		{
326			const ccv_nnc_tensor_param_t info = ccv_nnc_tensor_symbol_params(model->graph, model->outputs[i]);
327			const ccv_nnc_tensor_symbol_t fit = compiled_data->fits[i] = ccv_nnc_tensor_symbol_new(model->graph, info, 0);
328			compiled_data->f[i] = ccv_nnc_tensor_symbol_new(model->graph, ccv_nnc_tensor_auto, 0);
329			ccv_nnc_graph_exec_symbol_new(model->graph, loss, TENSOR_SYMBOL_LIST(model->outputs[i], fit)(const ccv_nnc_tensor_symbol_t []){model->outputs[i], fit}
, (1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(compiled_data->f[i])(const ccv_nnc_tensor_symbol_t []){compiled_data->f[i]}, (
1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
 +0 +0 -1), 0);
330		}
331	}
332	ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
333	ccv_nnc_symbolic_graph_simplify(model->graph,
334		SYMBOLIC_GRAPH_PASSES(CCV_NNC_SIMPLIFY_OPS_FUSION)(const int []){CCV_NNC_SIMPLIFY_OPS_FUSION}, (1 +1 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), // Only do Ops fusion, in this way, we can fuse the loss function.
335		0, 0, // No need to provide binds at this point.
336		compiled_data->f, model->output_size,
337		SYMBOLIC_GRAPH_SOURCES(model->graph)ccv_nnc_symbolic_graph_sources(model->graph), ccv_nnc_symbolic_graph_source_size
(model->graph), SYMBOLIC_GRAPH_DESTINATIONS(model->graph)ccv_nnc_symbolic_graph_destinations(model->graph), ccv_nnc_symbolic_graph_destination_size
(model->graph));
338	ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
339	// If inputs are from GPU, stream type is GPU.
340	compiled_data->parameters = parameters;
341	compiled_data->parameter_flags = parameter_flags;
342	compiled_data->internals = internals;
343	compiled_data->ids.parameters = parameter_ids;
344	compiled_data->ids.internals = internal_ids;
345}
346 
347static 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)
348{
349	ccv_array_t* const stack = (ccv_array_t*)context;
350	ccv_array_push(stack, &symbol.d);
351}
352 
353static 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)
354{
355	const ccv_nnc_tensor_symbol_t src_symbol = {
356		.d = src_index,
357		.graph = src_graph
358	};
359	const ccv_nnc_tensor_symbol_t dest_symbol = {
360		.d = dest_index,
361		.graph = dest_graph
362	};
363	const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(src_graph, src_symbol);
364	ccv_nnc_tensor_symbol_set(dest_graph, dest_symbol, params);
365	int ofs[CCV_NNC_MAX_DIM_ALLOC(12)];
366	int inc[CCV_NNC_MAX_DIM_ALLOC(12)];
367	if (0 == ccv_nnc_tensor_symbol_alias_params(src_graph, src_symbol, ofs, inc))
368		ccv_nnc_tensor_symbol_alias_set(dest_graph, dest_symbol, ofs, inc);
369}
370 
371static 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)
372{
373	const ccv_nnc_tensor_symbol_t src_symbol = {
374		.d = src_index,
375		.graph = src_graph
376	};
377	const ccv_nnc_tensor_param_t src_params = ccv_nnc_tensor_symbol_params(src_graph, src_symbol);
378	const ccv_nnc_tensor_symbol_t dest_symbol = {
379		.d = dest_index,
380		.graph = dest_graph
381	};
382	const ccv_nnc_tensor_param_t dest_params = ccv_nnc_tensor_symbol_params(dest_graph, dest_symbol);
383	return memcmp(src_params.dim, dest_params.dim, sizeof(src_params.dim)) == 0;
384}
385 
386static 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);
387static void _ccv_cnnp_compiled_data_graph_free(ccv_cnnp_compiled_data_t* const compiled_data);
388 
389typedef struct {
390	int parallel_count;
391	ccv_nnc_symbolic_graph_t* graph;
392	ccv_nnc_graph_exec_arena_t* graph_exec_arena;
393} ccv_nnc_graph_exec_update_t;
394 
395static 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)
396{
397	ccv_nnc_graph_exec_update_t* const graph_exec_update = (ccv_nnc_graph_exec_update_t*)context;
398	ccv_nnc_graph_exec_arena_t* const graph_exec_arena = graph_exec_update->graph_exec_arena;
399	ccv_nnc_graph_exec_t graph_exec = ccv_nnc_graph_exec_from_symbol(graph_exec_arena, symbol);
400	ccv_nnc_graph_exec_set(graph_exec.graph, graph_exec, cmd);
401	ccv_nnc_graph_exec_set_hint(graph_exec.graph, graph_exec, hint);
402	const ccv_nnc_symbolic_graph_t* const graph = graph_exec_update->graph;
403	const int parallel_count = graph_exec_update->parallel_count;
404	int i;
405	for (i = 1; i < parallel_count; i++)
406	{
407		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));
408		if (!CCV_NO_GRAPH_EXEC(copy)((copy).graph == 0))
409		{
410			ccv_nnc_graph_exec_set(copy.graph, copy, cmd);
411			ccv_nnc_graph_exec_set_hint(copy.graph, copy, hint);
412		}
413	}
414}
415 
416void 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)
417{
418	assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if
 (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c"
, 418, __extension__ __PRETTY_FUNCTION__); }));
419	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", 419, __extension__ __PRETTY_FUNCTION__)
; }));
420	assert(!init->graph)((void) sizeof ((!init->graph) ? 1 : 0), __extension__ ({ if
 (!init->graph) ; else __assert_fail ("!init->graph", "ccv_cnnp_model.c"
, 420, __extension__ __PRETTY_FUNCTION__); }));
421	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
422	init->graph = ccv_nnc_symbolic_graph_new();
423	ccv_array_t* const stack = ccv_array_new(sizeof(int), 0, 0);
424	ccv_nnc_graph_exec_symbol_new_hook(init->graph, _ccv_cnnp_graph_push_graph_exec_symbol, stack, 0);
425	_ccv_cnnp_model_compile(init, inputs, input_size, compiled_data->loss);
426	init->parallel_count = model->parallel_count;
427	init->memory_compression = model->memory_compression;
428	init->memory_reduction = model->memory_reduction;
429	init->gradient_checkpointing = model->gradient_checkpointing;
430	init->compiled_data->stream_type = model->compiled_data->stream_type;
431	init->compiled_data->minimize.minimizer = model->compiled_data->minimize.minimizer;
432	init->compiled_data->minimize.max_saved_aux_size = model->compiled_data->minimize.max_saved_aux_size;
433	if (model->compiled_data->gradient_mode != CCV_CNNP_COMPILED_DATA_GRADIENT_NONE)
434		_ccv_cnnp_model_gradient_init(init, model->compiled_data->gradient_mode, model->compiled_data->disable_outgrad, 0, 0);
435	ccv_nnc_graph_exec_symbol_new_hook(init->graph, 0, 0, 0);
436	ccv_nnc_symbolic_graph_tensor_auto(init->graph, TRAVERSE_FULL0,0,0,0);
437	int i, j;
438	// Verify parameters, internals and saved_aux in both graph has the same dimensionality.
439	for (i = 0; i < compiled_data->parameters->rnum; i++)
440	{
441		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;
442		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", 442, __extension__ __PRETTY_FUNCTION__)
; }));
443	}
444	for (i = 0; i < compiled_data->internals->rnum; i++)
445	{
446		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;
447		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", 447, __extension__ __PRETTY_FUNCTION__)
; }));
448	}
449	// Update inputs.
450	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", 450, __extension__ __PRETTY_FUNCTION__)
; }));
451	for (i = 0; i < model->input_size; i++)
452		if (model->inputs[i].d >= 0)
453		{
454			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", 454, __extension__ __PRETTY_FUNCTION__)
; }));
455			_ccv_nnc_tensor_symbol_reinit(init->graph, model->graph, init->inputs[i].d, model->inputs[i].d);
456		}
457	// Update outputs.
458	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", 458, __extension__ __PRETTY_FUNCTION__)
; }));
459	for (i = 0; i < model->output_size; i++)
460	{
461		if (model->outputs[i].d >= 0)
462		{
463			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", 463, __extension__
 __PRETTY_FUNCTION__); }));
464			_ccv_nnc_tensor_symbol_reinit(init->graph, model->graph, init->outputs[i].d, model->outputs[i].d);
465		}
466		if (model->outputs[i].d != model->compiled_data->f[i].d)
467		{
468			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", 468, __extension__ __PRETTY_FUNCTION__)
; }));
469			if (model->compiled_data->f[i].d >= 0)
470			{
471				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", 471, __extension__ __PRETTY_FUNCTION__)
; }));
472				_ccv_nnc_tensor_symbol_reinit(init->graph, model->graph, init->compiled_data->f[i].d, model->compiled_data->f[i].d);
473			}
474		}
475	}
476	// Go through the graph to set tensor on matching symbols
477	for (i = 0; i < stack->rnum; i++)
478	{
479		const int d = *(int*)ccv_array_get(stack, i)((void*)(((char*)((stack)->data)) + (size_t)(stack)->rsize
 * (size_t)(i)));
480		// If exceed range, skip.
481		if (d >= ccv_nnc_graph_exec_symbol_count(init->graph) ||
482			d >= ccv_nnc_graph_exec_symbol_count(model->graph))
483			continue;
484		const ccv_nnc_graph_exec_symbol_t src_symbol = {
485			.d = d,
486			.graph = init->graph
487		};
488		const ccv_nnc_graph_exec_symbol_t dest_symbol = {
489			.d = d,
490			.graph = model->graph
491		};
492		const ccv_nnc_cmd_t src_cmd = ccv_nnc_graph_exec_symbol_cmd(init->graph, src_symbol);
493		const ccv_nnc_cmd_t dest_cmd = ccv_nnc_graph_exec_symbol_cmd(model->graph, dest_symbol);
494		// If the name doesn't match, skip.
495		if (dest_cmd.cmd != src_cmd.cmd && src_cmd.cmd != CCV_NNC_NOOP)
496			continue;
497		// Now get all the inputs and outputs, if matches, set them.
498		const int* src_inputs;
499		int src_input_size;
500		const int* src_outputs;
501		int src_output_size;
502		ccv_nnc_graph_exec_symbol_io(init->graph, src_symbol, &src_inputs, &src_input_size, &src_outputs, &src_output_size);
503		const int* dest_inputs;
504		int dest_input_size;
505		const int* dest_outputs;
506		int dest_output_size;
507		ccv_nnc_graph_exec_symbol_io(model->graph, dest_symbol, &dest_inputs, &dest_input_size, &dest_outputs, &dest_output_size);
508		// We may have unmatched input / output size because this is the minimizer and it has
509		// different saved_aux (for example, when we shrunk with CMD_NOOP).
510		if (src_input_size != dest_input_size)
511			continue;
512		if (src_output_size != dest_output_size)
513			continue;
514		ccv_nnc_graph_exec_symbol_set(model->graph, dest_symbol, src_cmd);
515		// There may be mismatches of the source tensor symbols and destination tensor symbols. The reason is because
516		// we may later passed-in the minimizer, therefore, we may allocate tensors for minimizer later in the original
517		// graph whereas in the newly created graph, it is streamlined (the minimizer exists from the beginning). That
518		// will make the order of tensor symbols creation different, therefore, exact which tensor is which wrong as
519		// well. However, set a new minimizer won't change the exec symbol ordering, because we never create new exec
520		// symbols after gradient init step. Changing a new minimizer just updated that exec symbols setting, it is not
521		// a new exec symbol.
522		for (j = 0; j < src_input_size; j++)
523			if (src_inputs[j] >= 0)
524				_ccv_nnc_tensor_symbol_reinit(init->graph, model->graph, src_inputs[j], dest_inputs[j]);
525		for (j = 0; j < src_output_size; j++)
526			if (src_outputs[j] >= 0)
527				_ccv_nnc_tensor_symbol_reinit(init->graph, model->graph, src_outputs[j], dest_outputs[j]);
528	}
529	ccv_array_free(stack);
530	// After this, we get all tensors in the model graph resolved through tensor_auto.
531	ccv_nnc_symbolic_graph_tensor_auto(model->graph, TRAVERSE_FULL0,0,0,0);
532	// Verify symbols we get matches.
533	const int parameter_size = compiled_data->parameters->rnum;
534	for (i = 0; i < parameter_size; i++)
535		{ 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", 535, __extension__ __PRETTY_FUNCTION__)
; })); }
536	const int internal_size = compiled_data->internals->rnum;
537	for (i = 0; i < internal_size; i++)
538		{ 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", 538, __extension__ __PRETTY_FUNCTION__)
; })); }
539	// Go through compiled data.
540	if (compiled_data->tensor_arena)
541	{
542		const int flag = ccv_nnc_tensor_arena_reinit(compiled_data->tensor_arena, model->graph);
543		if (flag == 0 && compiled_data->graph_exec_arena)
544		{
545			ccv_nnc_graph_exec_reinit(compiled_data->graph_exec_arena, compiled_data->graph, model->graph);
546			// Since we will reinit, if we previously set is_test, we need to set it again.
547			if (compiled_data->is_test)
548			{
549				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; });
550				ccv_nnc_graph_exec_update_t update = {
551					.parallel_count = parallel_count,
552					.graph = model->graph,
553					.graph_exec_arena = compiled_data->graph_exec_arena,
554				};
555				ccv_cnnp_model_set_is_test(model, 1, _ccv_cnnp_cmd_update_for_execs, &update);
556			}
557		} else
558			// Free-up tensor arena & graph exec arena.
559			_ccv_cnnp_compiled_data_graph_free(compiled_data);
560	}
561	// There are other compiled graphs, for accum and apply gradients.
562	// However, the main conclusion is, these absorb operations shouldn't impact parameters.
563	// Thus, it won't impact the shape of gradients (only outgrad). Since for outgrad, we
564	// don't allocate ourselves, it is not a concern. For normal gradients, the shape cannot
565	// be changed otherwise parameters' shape will be meaningless. The same goes to internals.
566	// That is why we don't update these compiled graphs at all this point.
567	// Free the model, we've already "absorbed" it.
568	ccv_cnnp_model_free(init);
569}
570 
571void 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)
572{
573	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", 573, __extension__ __PRETTY_FUNCTION__)
; }));
574	if (model->input_size == 0)
575		model->input_size = input_size;
576	if (!model->graph) // The graph is not compiled yet.
577	{
578		model->graph = ccv_nnc_symbolic_graph_new();
579		_ccv_cnnp_model_compile(model, inputs, input_size, loss);
580		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", 580, __extension__ __PRETTY_FUNCTION__)
; }));
581		int i, flag = 0;
582		for (i = 0; !flag && i < input_size; i++)
583			flag = (CCV_TENSOR_GET_MEMORY(inputs[i].type)((inputs[i].type) & 0x3) == CCV_TENSOR_GPU_MEMORY);
584		// If inputs are from GPU, stream type is GPU.
585		model->compiled_data->stream_type = flag ? CCV_STREAM_CONTEXT_GPU : CCV_STREAM_CONTEXT_CPU;
586		model->compiled_data->minimize.minimizer = minimizer;
587		model->compiled_data->minimize.max_saved_aux_size = ccv_nnc_minimizer_saved_aux_size(minimizer);
588	} else {
589		// Now, finally fill in this part. If the graph is already compiled, we make a copy of the model.
590		// And then absorb the "new model" to the old one.
591		ccv_cnnp_model_t* const init = ccv_cnnp_model_copy(model, model->is_trainable);
592		ccv_cnnp_model_absorb(model, init, inputs, input_size);
593		// Reset minimizer.
594		ccv_cnnp_model_set_minimizer(model, minimizer, 1, 0, 0);
595	}
596}
597 
598ccv_cnnp_model_t* ccv_cnnp_model_copy(const ccv_cnnp_model_t* const model, const int is_trainable)
599{
600	ccv_cnnp_model_t* const new_model = _ccv_cnnp_model_copy(model, 0);
601	new_model->is_trainable = is_trainable;
602	return new_model;
603}
604 
605void ccv_cnnp_model_tensor_auto(ccv_cnnp_model_t* const model, ccv_nnc_tensor_param_t* const outputs, const int output_size)
606{
607	assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if
 (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c"
, 607, __extension__ __PRETTY_FUNCTION__); }));
608	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", 608, __extension__ __PRETTY_FUNCTION__)
; }));
609	ccv_nnc_symbolic_graph_t* const graph = model->graph;
610	ccv_nnc_symbolic_graph_tensor_auto(graph, TRAVERSE_FULL0,0,0,0);
611	int i;
612	for (i = 0; i < output_size; i++)
613	{
614		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", 614, __extension__ __PRETTY_FUNCTION__)
; }));
615		outputs[i] = ccv_nnc_tensor_symbol_params(graph, model->outputs[i]);
616	}
617}
618 
619void ccv_cnnp_model_set_workspace_size(ccv_cnnp_model_t* const model, size_t workspace_size)
620{
621	if (workspace_size == model->workspace_size)
622		return;
623	model->workspace_size = workspace_size;
624	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
625	if (compiled_data && compiled_data->graph)
626		ccv_nnc_graph_autotune(compiled_data->graph, workspace_size, 0, TRAVERSE_FULL0,0,0,0);
627}
628 
629size_t ccv_cnnp_model_workspace_size(ccv_cnnp_model_t* const model)
630{
631	return model->workspace_size;
632}
633 
634void ccv_cnnp_model_set_data_parallel(ccv_cnnp_model_t* const model, const int parallel)
635{
636	if (parallel == 0)
637		model->parallel_count = ccv_nnc_device_count(CCV_STREAM_CONTEXT_GPU);
638	else
639		model->parallel_count = parallel;
640	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
641	if (compiled_data)
642		{ 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", 642, __extension__ __PRETTY_FUNCTION__)
; })); }
643}
644 
645void ccv_cnnp_model_set_max_concurrency(ccv_cnnp_model_t* const model, const int max_stream_count)
646{
647	model->max_stream_count = max_stream_count;
648	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
649	if (compiled_data)
650		{ 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", 650, __extension__ __PRETTY_FUNCTION__)
; })); }
651}
652 
653void ccv_cnnp_model_set_memory_compression(ccv_cnnp_model_t* const model, const int memory_compression)
654{
655	model->memory_compression = memory_compression;
656	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
657	if (compiled_data)
658		{ 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", 658, __extension__ __PRETTY_FUNCTION__)
; })); }
659}
660 
661void ccv_cnnp_model_set_memory_reduction(ccv_cnnp_model_t* const model, const int memory_reduction)
662{
663	model->memory_reduction = memory_reduction;
664	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
665	if (compiled_data)
666		{ 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", 666, __extension__ __PRETTY_FUNCTION__)
; })); }
667}
668 
669void ccv_cnnp_model_set_gradient_checkpointing(ccv_cnnp_model_t* const model, const int gradient_checkpointing)
670{
671	model->gradient_checkpointing = gradient_checkpointing;
672}
673 
674int ccv_cnnp_model_gradient_checkpointing(ccv_cnnp_model_t* const model)
675{
676	return model->gradient_checkpointing;
677}
678 
679typedef struct {
680	int parallel_count;
681	ccv_nnc_symbolic_graph_t* graph;
682	ccv_cnnp_compiled_data_t* compiled_data;
683	ccv_nnc_tensor_arena_t* tensor_arena;
684} ccv_nnc_tensor_init_states_t;
685 
686static int _ccv_cnnp_any_to_init(const ccv_cnnp_compiled_data_t* const compiled_data)
687{
688	int i;
689	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));
690	for (i = 0; i < compiled_data->parameters->rnum; i++)
691	{
692		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;
693		if (!(init_v[d >> 5] & (1u << (d & 0x1f))))
694			return 1;
695	}
696	for (i = 0; i < compiled_data->internals->rnum; i++)
697	{
698		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;
699		if (!(init_v[d >> 5] & (1u << (d & 0x1f))))
700			return 1;
701	}
702	return 0;
703}
704 
705static 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)
706{
707	ccv_nnc_tensor_init_states_t* const tensor_init_states = (ccv_nnc_tensor_init_states_t*)context;
708	ccv_nnc_tensor_arena_t* const tensor_arena = tensor_init_states->tensor_arena;
709	ccv_nnc_tensor_t* const output_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, output_symbol);
710	if (!output_tensor)
711		return;
712	const int d = output_symbol.d;
713	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", 713, __extension__ __PRETTY_FUNCTION__)
; }));
714  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));
715	if (init_v[d >> 5] & (1u << (d & 0x1f)))
716		return;
717	init_v[d >> 5] |= (1u << (d & 0x1f));
718	ccv_nnc_cmd_exec(cmd, hint, flags, &input, input ? 1 : 0, &output_tensor, 1, 0);
719	const ccv_nnc_symbolic_graph_t* const graph = tensor_init_states->graph;
720	const int parallel_count = tensor_init_states->parallel_count;
721	int i;
722	for (i = 1; i < parallel_count; i++)
723	{
724		ccv_nnc_tensor_t* const copy = ccv_nnc_tensor_from_symbol(tensor_arena, ccv_nnc_tensor_symbol_copy(graph, output_symbol, i));
725		if (copy)
726			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);
727	}
728}
729 
730// This method can only handle cases we added new tensors and exec, never delete. This invariant is true because
731// we setup everything (including calling simplify method) in ccv_cnnp_model_compile method, before this rewind setup.
732static void _ccv_cnnp_model_rewind_graph(ccv_cnnp_model_t* const model)
733{
734	assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if
 (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c"
, 734, __extension__ __PRETTY_FUNCTION__); }));
735	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", 735, __extension__ __PRETTY_FUNCTION__)
; }));
736	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
737	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", 737, __extension__
 __PRETTY_FUNCTION__); }));
738	int i;
739	for (i = 0; i < compiled_data->rewindables->rnum; i++)
740	{
741		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)));
742		if (rewind_symbol->type == CCV_CNNP_REWIND_GRAPH_EXEC)
743			ccv_nnc_graph_exec_symbol_free(model->graph, rewind_symbol->graph_exec);
744		else if (rewind_symbol->type == CCV_CNNP_REWIND_TENSOR)
745			ccv_nnc_tensor_symbol_free(model->graph, rewind_symbol->tensor);
746	}
747	ccv_array_clear(compiled_data->rewindables);
748	ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
749}
750 
751static 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)
752{
753	const ccv_cnnp_rewind_symbol_t rewind_symbol = {
754		.type = CCV_CNNP_REWIND_TENSOR,
755		.tensor = symbol
756	};
757	ccv_array_t* const rewind_symbols = (ccv_array_t*)context;
758	ccv_array_push(rewind_symbols, &rewind_symbol);
759}
760 
761static 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)
762{
763	const ccv_cnnp_rewind_symbol_t rewind_symbol = {
764		.type = CCV_CNNP_REWIND_TENSOR,
765		.tensor = symbol
766	};
767	ccv_array_t* const rewind_symbols = (ccv_array_t*)context;
768	ccv_array_push(rewind_symbols, &rewind_symbol);
769}
770 
771static 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)
772{
773	const ccv_cnnp_rewind_symbol_t rewind_symbol = {
774		.type = CCV_CNNP_REWIND_GRAPH_EXEC,
775		.graph_exec = symbol
776	};
777	ccv_array_t* const rewind_symbols = (ccv_array_t*)context;
778	ccv_array_push(rewind_symbols, &rewind_symbol);
779}
780 
781static 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)
782{
783	ccv_nnc_graph_exec_t const update_exec = ccv_nnc_graph_exec_from_symbol(graph_exec_arena, exec_symbol);
784	if (!CCV_NO_GRAPH_EXEC(update_exec)((update_exec).graph == 0))
785		ccv_nnc_graph_exec_set(update_exec.graph, update_exec, cmd);
786	int i;
787	for (i = 1; i < parallel_count; i++)
788	{
789		ccv_nnc_graph_exec_symbol_t copy_symbol = ccv_nnc_graph_exec_symbol_copy(symbolic_graph, exec_symbol, i);
790		const ccv_nnc_graph_exec_t copy = ccv_nnc_graph_exec_from_symbol(graph_exec_arena, copy_symbol);
791		if (!CCV_NO_GRAPH_EXEC(copy)((copy).graph == 0))
792			ccv_nnc_graph_exec_set(copy.graph, copy, cmd);
793	}
794}
795 
796static 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)
797{
798	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 798, __extension__ __PRETTY_FUNCTION__); }));
799	assert(symbolic_graph)((void) sizeof ((symbolic_graph) ? 1 : 0), __extension__ ({ if
 (symbolic_graph) ; else __assert_fail ("symbolic_graph", "ccv_cnnp_model.c"
, 799, __extension__ __PRETTY_FUNCTION__); }));
800	ccv_nnc_graph_exec_symbol_set(symbolic_graph, exec_symbol, cmd);
801	int i;
802	for (i = 1; i < parallel_count; i++)
803	{
804		ccv_nnc_graph_exec_symbol_t copy_symbol = ccv_nnc_graph_exec_symbol_copy(symbolic_graph, exec_symbol, i);
805		if (copy_symbol.graph)
806			ccv_nnc_graph_exec_symbol_set(symbolic_graph, copy_symbol, cmd);
807	}
808	ccv_nnc_graph_exec_arena_t* const graph_exec_arena = compiled_data->graph_exec_arena;
809	if (graph_exec_arena)
810		_ccv_cnnp_model_graph_symbol_exec_set_for_graph_exec_arena(graph_exec_arena, parallel_count, exec_symbol, cmd, symbolic_graph);
811	// Skip backward graph exec arena because it is for a specific accum symbolic graph, not the main graph (model->graph)
812	ccv_nnc_graph_exec_arena_t* const gradient_graph_exec_arena = compiled_data->apply_gradients.graph_exec_arena;
813	if (gradient_graph_exec_arena)
814		_ccv_cnnp_model_graph_symbol_exec_set_for_graph_exec_arena(gradient_graph_exec_arena, parallel_count, exec_symbol, cmd, symbolic_graph);
815}
816 
817static 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)
818{
819	int this_parameter_flag = 0;
820	if (update_nodes[parameter_indice].d == CCV_NNC_NO_TENSOR_SYMBOL)
821		return this_parameter_flag;
822	const ccv_nnc_cmd_t old_minimizer = ccv_nnc_graph_exec_symbol_cmd(graph, update_nodes[parameter_indice]);
823	int j, k;
824	// For no-op, we can preserve previous saved_aux_size.
825	if (old_minimizer.cmd != minimizer.cmd && minimizer.cmd != CCV_NNC_NOOP)
826	{
827		// If the old minimizer is a noop, then the old_saved_aux_size should be whatever its previous
828		// saved_aux_size is, otherwise we will reinit the saved_aux repeatedly if you switch between
829		// noop and a minimizer. We don't want that because we do that in high-level frameworks to
830		// make sure some model parameters don't update if we don't want them to.
831		int old_saved_aux_size;
832		if (old_minimizer.cmd == CCV_NNC_NOOP)
833		{
834			int input_size;
835			ccv_nnc_graph_exec_symbol_io(graph, update_nodes[parameter_indice], 0, &input_size, 0, 0);
836			if (input_size < 2) // This is not legit.
837				old_saved_aux_size = ccv_nnc_minimizer_saved_aux_size(old_minimizer);
838			else // See ccv_nnc_minimizer_saved_aux_size, the saved_aux is inputs excluding gradients and parameters.
839				old_saved_aux_size = input_size - 2;
840		} else
841			old_saved_aux_size = ccv_nnc_minimizer_saved_aux_size(old_minimizer);
842		if (old_saved_aux_size != saved_aux_size)
843		{
844			this_parameter_flag = 1;
845			if (saved_aux_size > old_saved_aux_size)
846			{
847				// Allocate new tensor symbols.
848				const ccv_nnc_tensor_param_t info = ccv_nnc_tensor_symbol_params(graph, updated_parameters[parameter_indice]);
849				for (j = old_saved_aux_size; j < saved_aux_size; j++)
850				{
851					saved_aux[parameter_indice * max_saved_aux_size + j].source = ccv_nnc_tensor_symbol_new(graph, info, 0);
852					saved_aux[parameter_indice * max_saved_aux_size + j].destination = ccv_nnc_tensor_symbol_new(graph, info, 0);
853					const int device_id = CCV_TENSOR_GET_DEVICE_ID(info.type)(((info.type) & 0xfff00) >> 8);
854					for (k = 1; k < parallel_count; k++)
855					{
856						ccv_nnc_tensor_param_t dev_info = info;
857						if (k != device_id)
858							CCV_TENSOR_SET_DEVICE_ID(dev_info.type, k)(dev_info.type) = (((dev_info.type) & ~0xfff00) | (((k) &
 0xfff) << 8));
859						else
860							CCV_TENSOR_SET_DEVICE_ID(dev_info.type, 0)(dev_info.type) = (((dev_info.type) & ~0xfff00) | (((0) &
 0xfff) << 8));
861						const ccv_nnc_tensor_symbol_t src_copy = ccv_nnc_tensor_symbol_new(graph, dev_info, 0);
862						const ccv_nnc_tensor_symbol_t dest_copy = ccv_nnc_tensor_symbol_new(graph, dev_info, 0);
863						ccv_nnc_tensor_symbol_set_copy(graph, saved_aux[parameter_indice * max_saved_aux_size + j].source, k, src_copy);
864						ccv_nnc_tensor_symbol_set_copy(graph, saved_aux[parameter_indice * max_saved_aux_size + j].destination, k, dest_copy);
865					}
866				}
867			} else {
868				for (j = saved_aux_size; j < old_saved_aux_size; j++)
869				{
870					for (k = 1; k < parallel_count; k++)
871					{
872						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);
873						if (src_copy.d >= 0)
874						{
875							ccv_nnc_tensor_symbol_free(graph, src_copy);
876							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
});
877						}
878						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);
879						if (dest_copy.d >= 0)
880						{
881							ccv_nnc_tensor_symbol_free(graph, dest_copy);
882							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
});
883						}
884					}
885					ccv_nnc_tensor_symbol_free(graph, saved_aux[parameter_indice * max_saved_aux_size + j].source);
886					ccv_nnc_tensor_symbol_free(graph, saved_aux[parameter_indice * max_saved_aux_size + j].destination);
887					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
};
888				}
889			}
890		}
891	}
892	_ccv_cnnp_model_graph_exec_symbol_set(graph, compiled_data, parallel_count, update_nodes[parameter_indice], minimizer);
893	if (this_parameter_flag)
894	{
895		ccv_nnc_tensor_symbol_t update_inputs[saved_aux_size + 2];
896		ccv_nnc_tensor_symbol_t update_outputs[saved_aux_size + 1];
897		const int* inputs = 0;
898		int input_size = 0;
899		ccv_nnc_graph_exec_symbol_io(graph, update_nodes[parameter_indice], &inputs, &input_size, 0, 0);
900		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", 900, __extension__ __PRETTY_FUNCTION__)
; }));
901		update_inputs[0].d = inputs[0];
902		update_inputs[0].graph = graph;
903		update_inputs[1].d = inputs[1];
904		update_inputs[1].graph = graph;
905		update_outputs[0] = updated_parameters[parameter_indice];
906		for (j = 0; j < saved_aux_size; j++)
907		{
908			update_inputs[j + 2] = saved_aux[parameter_indice * max_saved_aux_size + j].source;
909			update_outputs[j + 1] = saved_aux[parameter_indice * max_saved_aux_size + j].destination;
910		}
911		ccv_nnc_graph_exec_symbol_set_io(graph, update_nodes[parameter_indice], update_inputs, saved_aux_size + 2, update_outputs, saved_aux_size + 1);
912		for (k = 1; k < parallel_count; k++)
913		{
914			const ccv_nnc_graph_exec_symbol_t copy = ccv_nnc_graph_exec_symbol_copy(graph, update_nodes[parameter_indice], k);
915			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"
, 915, __extension__ __PRETTY_FUNCTION__); }));
916			ccv_nnc_graph_exec_symbol_io(graph, copy, &inputs, &input_size, 0, 0);
917			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", 917, __extension__ __PRETTY_FUNCTION__)
; }));
918			update_inputs[0].d = inputs[0];
919			update_inputs[0].graph = graph;
920			update_inputs[1].d = inputs[1];
921			update_inputs[1].graph = graph;
922			update_outputs[0] = ccv_nnc_tensor_symbol_copy(graph, updated_parameters[parameter_indice], k);
923			for (j = 0; j < saved_aux_size; j++)
924			{
925				update_inputs[j + 2] = ccv_nnc_tensor_symbol_copy(graph, saved_aux[parameter_indice * max_saved_aux_size + j].source, k);
926				update_outputs[j + 1] = ccv_nnc_tensor_symbol_copy(graph, saved_aux[parameter_indice * max_saved_aux_size + j].destination, k);
927			}
928			ccv_nnc_graph_exec_symbol_set_io(graph, copy, update_inputs, saved_aux_size + 2, update_outputs, saved_aux_size + 1);
929		}
930	}
931	return this_parameter_flag;
932}
933 
934typedef struct {
935	int parameter_size;
936	ccv_nnc_cmd_t minimizer;
937	ccv_cnnp_model_io_t parameters[1];
938} ccv_cnnp_set_minimizer_for_parameter_t;
939 
940static int _ccv_cnnp_apply_parameters_with_minimizer(ccv_cnnp_model_t* const model)
941{
942	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
943	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 943, __extension__ __PRETTY_FUNCTION__); }));
944	const int max_saved_aux_size = compiled_data->minimize.max_saved_aux_size;
945	// We update all parameters, at this point, we have one minimizer.
946	const int parameter_size = compiled_data->parameters->rnum;
947	ccv_nnc_graph_exec_symbol_t* const update_nodes = compiled_data->update_nodes;
948	ccv_nnc_symbolic_graph_t* const symbolic_graph = model->graph;
949	assert(symbolic_graph)((void) sizeof ((symbolic_graph) ? 1 : 0), __extension__ ({ if
 (symbolic_graph) ; else __assert_fail ("symbolic_graph", "ccv_cnnp_model.c"
, 949, __extension__ __PRETTY_FUNCTION__); }));
950	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; });
951	ccv_array_t* const parameters = compiled_data->minimize.parameters;
952	ccv_array_t* const parameter_indices = ccv_array_new(sizeof(int), 0, 0);
953	int i, j, flag = 0;
954	for (i = 0; i < parameters->rnum; i++)
955	{
956		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)));
957		for (j = 0; j < set_minimizer_for_parameter->parameter_size; j++)
958		{
959			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;
960			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", 960, __extension__ __PRETTY_FUNCTION__)
; }));
961			const int old_rnum = parameter_indices->rnum;
962			ccv_cnnp_model_add_to_parameter_indices(set_minimizer_for_parameter->parameters[j]->model, param_sel, parameter_indices);
963			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;
964			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", 964, __extension__ __PRETTY_FUNCTION__)
; }));
965			if (param_ref >= 0)
966			{
967				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", 967, __extension__ __PRETTY_FUNCTION__)
; }));
968				*(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)));
969				parameter_indices->rnum = old_rnum + 1;
970			}
971		}
972		const int saved_aux_size = ccv_nnc_minimizer_saved_aux_size(set_minimizer_for_parameter->minimizer);
973		// We may have duplicated indices, but that is OK, we will set it twice.
974		for (j = 0; j < parameter_indices->rnum; j++)
975		{
976			const int d = *(int*)ccv_array_get(parameter_indices, j)((void*)(((char*)((parameter_indices)->data)) + (size_t)(parameter_indices
)->rsize * (size_t)(j)));
977			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", 977, __extension__ __PRETTY_FUNCTION__)
; }));
978			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))
979				flag = 1;
980		}
981		ccv_array_clear(parameter_indices);
982	}
983	ccv_array_free(parameter_indices);
984	return flag;
985}
986 
987static 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)
988{
989	if (new_saved_aux_size == old_saved_aux_size)
990		return;
991	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", 991, __extension__ __PRETTY_FUNCTION__)
; }));
992	int i, j;
993	for (i = parameter_size - 1; i >= 0; i--)
994	{
995		for (j = new_saved_aux_size - 1; j >= old_saved_aux_size; j--)
996			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
};
997		for (j = old_saved_aux_size - 1; j >= 0; j--)
998			saved_aux[i * new_saved_aux_size + j] = saved_aux[i * old_saved_aux_size + j];
999	}
1000}
1001 
1002static void _ccv_cnnp_model_set_rewindables(ccv_cnnp_model_t* const model)
1003{
1004	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1005	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 1005, __extension__ __PRETTY_FUNCTION__); }));
1006	if (!compiled_data->rewindables)
1007		compiled_data->rewindables = ccv_array_new(sizeof(ccv_cnnp_rewind_symbol_t), 0, 0);
1008	ccv_nnc_tensor_symbol_new_hook(model->graph, _ccv_cnnp_model_tensor_symbol_new_hook, compiled_data->rewindables, 0);
1009	ccv_nnc_tensor_symbol_alias_new_hook(model->graph, _ccv_cnnp_model_tensor_symbol_alias_new_hook, compiled_data->rewindables, 0);
1010	ccv_nnc_graph_exec_symbol_new_hook(model->graph, _ccv_cnnp_model_graph_exec_symbol_new_hook, compiled_data->rewindables, 0);
1011}
1012 
1013static 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)
1014{
1015	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1016	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", 1016, __extension__ __PRETTY_FUNCTION__
); }));
1017	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", 1017, __extension__ __PRETTY_FUNCTION__
); }));
1018	const int evaluate_to_size = compiled_data->evaluate.to_size;
1019	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", 1019, __extension__ __PRETTY_FUNCTION__
); }));
1020	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; });
1021	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);
1022	compiled_data->evaluate.to_ops = (ccv_nnc_graph_exec_t*)(compiled_data->evaluate.tos + evaluate_to_size * parallel_count);
1023	int i, j;
1024	const int output_size = model->output_size;
1025	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", 1025, __extension__ __PRETTY_FUNCTION__
); }));
1026	if (fits)
1027		for (i = 0; i < output_size; i++)
1028			ccv_nnc_tensor_symbol_set(model->graph, compiled_data->fits[i], fits[i]->info);
1029	const int max_saved_aux_size = compiled_data->minimize.max_saved_aux_size;
1030	const int parameter_size = compiled_data->parameters->rnum;
1031	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);
1032	compiled_data->update_nodes = (ccv_nnc_graph_exec_symbol_t*)(compiled_data->updated_parameters + parameter_size);
1033	compiled_data->saved_aux = (ccv_nnc_tensor_symbol_map_t*)(compiled_data->update_nodes + parameter_size);
1034	int parameter_size_maybe_more = parameter_size;
1035	compiled_data->disable_outgrad = disable_outgrad;
1036	int outgrad_size;
1037	if (gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES || model->input_size == 0)
1038		outgrad_size = 0;
1039	else if (disable_outgrad == CCV_CNNP_DISABLE_OUTGRAD_NONE) // Compute minimize with gradients including inputs.
1040		outgrad_size = model->input_size;
1041	else {
1042		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", 1042, __extension__ __PRETTY_FUNCTION__
); })); // If it is disable all, gradient mode won't be this.
1043		outgrad_size = 0;
1044		for (i = 0; i < model->input_size; i++)
1045			if (!(disable_outgrad & ((uint64_t)1 << i)))
1046				++outgrad_size;
1047	}
1048	compiled_data->outgrad_size = outgrad_size;
1049	parameter_size_maybe_more += outgrad_size;
1050	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);
1051	compiled_data->outgrads = parameter_size_maybe_more > parameter_size ? compiled_data->gradients + parameter_size : 0;
1052	compiled_data->backward.tos = (ccv_nnc_graph_exec_symbol_t*)(compiled_data->gradients + parameter_size_maybe_more);
1053	compiled_data->backward.to_size = parameter_size_maybe_more;
1054	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)));
1055	if (compiled_data->parameter_flags)
1056	{
1057		parameters = (ccv_nnc_tensor_symbol_t*)ccmallocmalloc(sizeof(ccv_nnc_tensor_symbol_t) * parameter_size);
1058		for (i = 0; i < parameter_size; i++)
1059			if (compiled_data->parameter_flags[i >> 6] & ((uint64_t)1 << (i & 63)))
1060				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)));
1061			else
1062				parameters[i] = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
1063	}
1064	if (gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES || model->input_size == 0)
1065		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);
1066	else if (disable_outgrad == CCV_CNNP_DISABLE_OUTGRAD_NONE) // Compute minimize with gradients including inputs.
1067		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);
1068	else { // Compute minimize with gradients including selected inputs.
1069		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", 1069, __extension__ __PRETTY_FUNCTION__
); }));
1070		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", 1070, __extension__ __PRETTY_FUNCTION__
); })); // If it is disable all, gradient mode won't be this.
1071		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", 1071, __extension__ __PRETTY_FUNCTION__
); }));
1072		ccv_nnc_tensor_symbol_t outgrads[outgrad_size];
1073		j = 0;
1074		for (i = 0; i < model->input_size; i++)
1075			if (!(disable_outgrad & ((uint64_t)1 << i)))
1076				outgrads[j++] = model->inputs[i];
1077		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);
1078	}
1079	if (compiled_data->parameter_flags)
1080		ccfreefree(parameters);
1081	_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);
1082	if (compiled_data->minimize.parameters)
1083		_ccv_cnnp_apply_parameters_with_minimizer(model);
1084	// Go through gradient checkpoints to generate tensor inputs for backward pass just before executing the backward pass.
1085	ccv_cnnp_model_apply_gradient_checkpoints(compiled_data, model->graph);
1086	for (i = 0; i < output_size; i++)
1087	{
1088		const ccv_nnc_tensor_symbol_t df = ccv_nnc_tensor_symbol_for_backward(model->graph, compiled_data->f[i]);
1089		// Init this to 1 so we can backprop.
1090		ccv_nnc_tensor_symbol_set_flags(model->graph, df, CCV_NNC_TENSOR_SYMBOL_INIT_ONES);
1091	}
1092	compiled_data->backward.to_size = 0;
1093	for (i = 0; i < parameter_size_maybe_more; i++)
1094		if (compiled_data->gradients[i].d != CCV_NNC_NO_TENSOR_SYMBOL)
1095			compiled_data->backward.tos[compiled_data->backward.to_size++] = ccv_nnc_graph_exec_symbol_for_backward(model->graph, compiled_data->gradients[i]);
1096	ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS);
1097	ccv_nnc_symbolic_graph_set_destinations(model->graph, compiled_data->update_nodes, parameter_size);
1098	for (i = 0; i < parameter_size_maybe_more - parameter_size; i++)
1099	{
1100		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.
1101			continue;
1102		const ccv_nnc_graph_exec_symbol_t outgrad = ccv_nnc_graph_exec_symbol_for_backward(model->graph, compiled_data->outgrads[i]);
1103		const int* tos;
1104		int to_size;
1105		ccv_nnc_graph_exec_symbol_to(model->graph, outgrad, &tos, &to_size);
1106		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.
1107		{
1108			const ccv_nnc_graph_exec_symbol_t* destinations = ccv_nnc_symbolic_graph_destinations(model->graph);
1109			const int destination_count = ccv_nnc_symbolic_graph_destination_size(model->graph);
1110			int flag = 0;
1111			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; });
1112			for (j = i - 1; !flag && j >= 0; j--)
1113				if (j + outgrad_destination_start < destination_count)
1114					flag = (destinations[j + outgrad_destination_start].d == outgrad.d);
1115			if (!flag) // Only if we cannot find it, we add it.
1116				ccv_nnc_symbolic_graph_add_destination(model->graph, outgrad);
1117		}
1118	}
1119	if (parallel_count > 1)
1120	{
1121		ccv_nnc_symbolic_graph_data_parallel(model->graph, parallel_count,
1122			0, 0,
1123			compiled_data->gradients, parameter_size /* No need to deal with outgrads, we don't allreduce outgrads */,
1124			compiled_data->gradients /* We only care about gradients before allreduce, thus, update our current pointers */,
1125			0, 0, 0,
1126			CCV_NNC_PARALLEL_REDUCE_OP_SUM,
1127			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));
1128		ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
1129		for (i = 0; i < evaluate_to_size; i++)
1130			for (j = 1; j < parallel_count; j++)
1131			{
1132				const ccv_nnc_graph_exec_symbol_t copy = ccv_nnc_graph_exec_symbol_copy(model->graph, compiled_data->evaluate.tos[i], j);
1133				if (copy.d != CCV_NNC_NO_GRAPH_EXEC_SYMBOL)
1134					compiled_data->evaluate.tos[compiled_data->evaluate.to_size++] = copy;
1135			}
1136		const int backward_to_size = compiled_data->backward.to_size;
1137		for (i = 0; i < backward_to_size; i++)
1138			for (j = 1; j < parallel_count; j++)
1139			{
1140				const ccv_nnc_graph_exec_symbol_t copy = ccv_nnc_graph_exec_symbol_copy(model->graph, compiled_data->backward.tos[i], j);
1141				if (copy.d != CCV_NNC_NO_GRAPH_EXEC_SYMBOL)
1142					compiled_data->backward.tos[compiled_data->backward.to_size++] = copy;
1143			}
1144	}
1145	// Only use memory compression if we are in gradient parameter mode.
1146	if (gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES || gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES_AND_INPUTS)
1147	{
1148		if (model->memory_compression)
1149			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));
1150		if (model->memory_reduction)
1151			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));
1152	}
1153	compiled_data->backward.to_size = _ccv_nnc_array_dedup_graph_exec_symbols(compiled_data->backward.tos, compiled_data->backward.to_size);
1154	compiled_data->gradient_mode = gradient_mode;
1155}
1156 
1157void ccv_cnnp_model_tensors_init_0(const ccv_cnnp_model_t* const model, ccv_cnnp_compiled_data_t* const compiled_data)
1158{
1159	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", 1159, __extension__ __PRETTY_FUNCTION__
); }));
1160	const int parameter_size = compiled_data->parameters->rnum;
1161	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; });
1162	const int internal_size = compiled_data->internals->rnum;
1163	compiled_data->tensors_init.size = ccv_nnc_tensor_symbol_count(model->graph);
1164	compiled_data->tensors_init.v = cccalloccalloc(((compiled_data->tensors_init.size + 31) >> 5), sizeof(uint32_t));
1165	compiled_data->tensors.parameters = (ccv_nnc_tensor_t**)cccalloccalloc((parameter_size + internal_size) * parallel_count, sizeof(ccv_nnc_tensor_t*));
1166	compiled_data->tensors.internals = compiled_data->tensors.parameters + parameter_size * parallel_count;
1167}
1168 
1169int ccv_cnnp_model_tensors_any_to_alloc(const ccv_cnnp_model_t* const model, ccv_cnnp_compiled_data_t* const compiled_data)
1170{
1171	int i, j;
1172	const int parameter_size = compiled_data->parameters->rnum;
1173	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; });
1174	const int internal_size = compiled_data->internals->rnum;
1175	for (i = 0; i < parameter_size; i++)
1176	{
1177		// parameters has to be allocated all together.
1178		if (compiled_data->tensors.parameters[i])
1179		{
1180			for (j = 1; j < parallel_count; j++)
1181				{ 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", 1181, __extension__ __PRETTY_FUNCTION__
); })); }
1182			continue;
1183		}
1184		return 1;
1185	}
1186	for (i = 0; i < internal_size; i++)
1187	{
1188		if (!compiled_data->tensors.internals[i])
1189			return 1;
1190		for (j = 1; j < parallel_count; j++)
1191			if (!compiled_data->tensors.internals[i + j * internal_size])
1192				return 1;
1193	}
1194	return 0;
1195}
1196 
1197void ccv_cnnp_model_tensors_init_1(const ccv_cnnp_model_t* const model, ccv_cnnp_compiled_data_t* const compiled_data)
1198{
1199	int i, j;
1200	const int parameter_size = compiled_data->parameters->rnum;
1201	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; });
1202	const int internal_size = compiled_data->internals->rnum;
1203	for (i = 0; i < parameter_size; i++)
1204	{
1205		// parameters has to be allocated all together.
1206		if (compiled_data->tensors.parameters[i])
1207		{
1208			for (j = 1; j < parallel_count; j++)
1209				{ 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", 1209, __extension__ __PRETTY_FUNCTION__
); })); }
1210			continue;
1211		}
1212		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)));
1213		ccv_nnc_tensor_param_t info = ccv_nnc_tensor_symbol_params(parameter.graph, parameter);
1214		if (CCV_TENSOR_GET_DEVICE(info.type)((info.type) & 0xfff00) == CCV_COMPUTE_DEVICE_ANY)
1215			CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff
) << 8));
1216		const int device_id = CCV_TENSOR_GET_DEVICE_ID(info.type)(((info.type) & 0xfff00) >> 8);
1217		compiled_data->tensors.parameters[i] = ccv_nnc_tensor_new(0, info, 0);
1218		for (j = 1; j < parallel_count; j++)
1219		{
1220			if (j != device_id)
1221				CCV_TENSOR_SET_DEVICE_ID(info.type, j)(info.type) = (((info.type) & ~0xfff00) | (((j) & 0xfff
) << 8));
1222			else
1223				CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff
) << 8));
1224			compiled_data->tensors.parameters[i + j * parameter_size] = ccv_nnc_tensor_new(0, info, 0);
1225		}
1226	}
1227	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));
1228	for (i = 0; i < internal_size; i++)
1229	{
1230		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))
);
1231		const int d = retained.d;
1232		if (init_v[d >> 5] & (1u << (d & 0x1f)))
1233			continue;
1234		ccv_nnc_tensor_param_t info = ccv_nnc_tensor_symbol_params(retained.graph, retained);
1235		if (CCV_TENSOR_GET_DEVICE(info.type)((info.type) & 0xfff00) == CCV_COMPUTE_DEVICE_ANY)
1236			CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff
) << 8));
1237		const int device_id = CCV_TENSOR_GET_DEVICE_ID(info.type)(((info.type) & 0xfff00) >> 8);
1238		if (!compiled_data->tensors.internals[i])
1239			compiled_data->tensors.internals[i] = ccv_nnc_tensor_new(0, info, 0);
1240		for (j = 1; j < parallel_count; j++)
1241		{
1242			if (j != device_id)
1243				CCV_TENSOR_SET_DEVICE_ID(info.type, j)(info.type) = (((info.type) & ~0xfff00) | (((j) & 0xfff
) << 8));
1244			else
1245				CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff
) << 8));
1246			if (!compiled_data->tensors.internals[i + j * internal_size])
1247				compiled_data->tensors.internals[i + j * internal_size] = ccv_nnc_tensor_new(0, info, 0);
1248		}
1249	}
1250	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.
1251}
1252 
1253static void _ccv_cnnp_model_tensors_init(const ccv_cnnp_model_t* const model, ccv_cnnp_compiled_data_t* const compiled_data)
1254{
1255	ccv_cnnp_model_tensors_init_0(model, compiled_data);
1256	ccv_cnnp_model_tensors_init_1(model, compiled_data);
1257}
1258 
1259static 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)
1260{
1261	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", 1261, __extension__ __PRETTY_FUNCTION__
); }));
1262	int i, j;
1263	for (i = 0; i < tensor_size; i++)
1264	{
1265		if (!tensors[i])
1266			continue;
1267		const int d = tensor_symbols[i].d;
1268		if (!(tensors_init[d >> 5] & (1u << (d & 0x1f))))
1269			continue;
1270		for (j = 1; j < parallel_count; j++)
1271			if (tensors[i + j * tensor_size])
1272			{
1273				ccv_nnc_tensor_t* const input = CCV_NNC_TENSOR(tensors[i])((ccv_nnc_tensor_t*)((uintptr_t)(tensors[i]) & ~(uintptr_t
)1));
1274				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));
1275				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);
1276			}
1277	}
1278}
1279 
1280static 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)
1281{
1282	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", 1282, __extension__ __PRETTY_FUNCTION__
); }));
1283	int i, j;
1284	for (i = 0; i < tensor_size; i++)
1285	{
1286		const ccv_nnc_tensor_symbol_t tensor_symbol = tensor_symbols[i];
1287		for (j = 1; j < parallel_count; j++)
1288		{
1289			const ccv_nnc_tensor_symbol_t copy = ccv_nnc_tensor_symbol_copy(graph, tensor_symbol, j);
1290			ccv_nnc_tensor_t* copy_tensor = tensors[i + j * tensor_size];
1291			if (copy_tensor && copy.d == CCV_NNC_NO_TENSOR_SYMBOL)
1292			{ // We shouldn't allocate this, free it up.
1293				ccv_nnc_tensor_free(tensors[i + j * tensor_size]);
1294				tensors[i + j * tensor_size] = 0;
1295			}
1296		}
1297	}
1298}
1299 
1300static 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)
1301{
1302	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", 1302, __extension__ __PRETTY_FUNCTION__
); }));
1303	int i, j;
1304	for (i = 0; i < tensor_size; i++)
1305	{
1306		ccv_nnc_tensor_symbol_t tensor_symbol = tensor_symbols[i];
1307		if (tensor_symbol.d == CCV_NNC_NO_TENSOR_SYMBOL)
1308			continue;
1309		if (graph)
1310		{
1311			const ccv_nnc_tensor_symbol_t alias_to = ccv_nnc_tensor_symbol_alias_to(graph, tensor_symbol);
1312			if (alias_to.d != CCV_NNC_NO_TENSOR_SYMBOL)
1313				tensor_symbol = alias_to;
1314		}
1315		ccv_nnc_tensor_t* const tensor = CCV_NNC_TENSOR(tensors[i])((ccv_nnc_tensor_t*)((uintptr_t)(tensors[i]) & ~(uintptr_t
)1));
1316		if (tensor && tensor_symbol.d != CCV_NNC_NO_TENSOR_SYMBOL)
1317		{
1318			const ccv_nnc_tensor_bind_t retained_bind = {
1319				.symbol = tensor_symbol,
1320				.tensor = tensor
1321			};
1322			ccv_array_push(tensor_binds, &retained_bind);
1323		}
1324		for (j = 1; j < parallel_count; j++)
1325		{
1326			const ccv_nnc_tensor_symbol_t copy = ccv_nnc_tensor_symbol_copy(graph, tensor_symbol, j);
1327			ccv_nnc_tensor_t* copy_tensor = tensors[i + j * tensor_size];
1328			if (copy_tensor && copy.d != CCV_NNC_NO_TENSOR_SYMBOL)
1329			{
1330				const ccv_nnc_tensor_bind_t bind = {
1331					.symbol = copy,
1332					.tensor = tensors[i + j * tensor_size]
1333				};
1334				ccv_array_push(tensor_binds, &bind);
1335			}
1336		}
1337	}
1338}
1339 
1340static void _ccv_cnnp_compiled_data_graph_free(ccv_cnnp_compiled_data_t* const compiled_data)
1341{
1342	if (compiled_data->graph)
1343		ccv_nnc_graph_free(compiled_data->graph);
1344	compiled_data->graph = 0;
1345	compiled_data->is_test = 0;
1346	if (compiled_data->tensor_arena)
1347		ccv_nnc_tensor_arena_free(compiled_data->tensor_arena);
1348	compiled_data->tensor_arena = 0;
1349	if (compiled_data->graph_exec_arena)
1350		ccv_nnc_graph_exec_arena_free(compiled_data->graph_exec_arena);
1351	compiled_data->graph_exec_arena = 0;
1352	if (compiled_data->backward.from_ops)
1353		ccfreefree(compiled_data->backward.from_ops);
1354	compiled_data->backward.from_ops = 0;
1355	if (compiled_data->evaluate.schedule)
1356		ccv_nnc_graph_static_schedule_free(compiled_data->evaluate.schedule);
1357	compiled_data->evaluate.schedule = 0;
1358	if (compiled_data->backward.schedule)
1359		ccv_nnc_graph_static_schedule_free(compiled_data->backward.schedule);
1360	compiled_data->backward.schedule = 0;
1361}
1362 
1363static void _ccv_cnnp_compiled_data_gradient_free(ccv_cnnp_compiled_data_t* const compiled_data)
1364{
1365	if (compiled_data->gradients)
1366		ccfreefree(compiled_data->gradients);
1367	compiled_data->gradients = 0;
1368	if (compiled_data->updated_parameters)
1369		ccfreefree(compiled_data->updated_parameters);
1370	compiled_data->updated_parameters = 0;
1371	compiled_data->update_nodes = 0;
1372	compiled_data->saved_aux = 0;
1373}
1374 
1375static void _ccv_cnnp_compiled_data_backward_free(ccv_cnnp_compiled_data_t* const compiled_data)
1376{
1377	if (compiled_data->backward.gradients)
1378		ccfreefree(compiled_data->backward.gradients);
1379	compiled_data->backward.gradients = 0;
1380	if (compiled_data->backward.accum)
1381		ccv_nnc_graph_free(compiled_data->backward.accum);
1382	compiled_data->backward.accum = 0;
1383	if (compiled_data->backward.tensor_arena)
1384		ccv_nnc_tensor_arena_free(compiled_data->backward.tensor_arena);
1385	compiled_data->backward.tensor_arena = 0;
1386	if (compiled_data->backward.graph_exec_arena)
1387		ccv_nnc_graph_exec_arena_free(compiled_data->backward.graph_exec_arena);
1388	compiled_data->backward.graph_exec_arena = 0;
1389}
1390 
1391static void _ccv_cnnp_compiled_data_apply_gradients_free(ccv_cnnp_compiled_data_t* const compiled_data)
1392{
1393	if (compiled_data->apply_gradients.graph)
1394		ccv_nnc_graph_free(compiled_data->apply_gradients.graph);
1395	compiled_data->apply_gradients.graph = 0;
1396	if (compiled_data->apply_gradients.tensor_arena)
1397		ccv_nnc_tensor_arena_free(compiled_data->apply_gradients.tensor_arena);
1398	compiled_data->apply_gradients.tensor_arena = 0;
1399	if (compiled_data->apply_gradients.graph_exec_arena)
1400		ccv_nnc_graph_exec_arena_free(compiled_data->apply_gradients.graph_exec_arena);
1401	compiled_data->apply_gradients.graph_exec_arena = 0;
1402}
1403 
1404// Compile the graph to run ccv_cnnp_model_fit
1405static 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)
1406{
1407	int i, j;
1408	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1409	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", 1409, __extension__ __PRETTY_FUNCTION__
); }));
1410	compiled_data->graph_mode = CCV_CNNP_MODEL_GRAPH_FIT_MODE;
1411	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; });
1412	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", 1412, __extension__ __PRETTY_FUNCTION__
); }));
1413	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", 1413
, __extension__ __PRETTY_FUNCTION__); }));
1414	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", 1414, __extension__ __PRETTY_FUNCTION__
); }));
1415	if (compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_NONE)
1416	{
1417		_ccv_cnnp_model_set_rewindables(model);
1418		_ccv_cnnp_model_gradient_init(model, CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES, CCV_CNNP_DISABLE_OUTGRAD_ALL, fits, fit_size);
1419	} else if (compiled_data->gradient_mode != CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES) {
1420		_ccv_cnnp_model_rewind_graph(model);
1421		_ccv_cnnp_compiled_data_gradient_free(compiled_data);
1422		compiled_data->gradient_mode = CCV_CNNP_COMPILED_DATA_GRADIENT_NONE;
1423		_ccv_cnnp_model_gradient_init(model, CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES, CCV_CNNP_DISABLE_OUTGRAD_ALL, fits, fit_size);
1424	}
1425	const int tensors_init = !!compiled_data->tensors_init.v;
1426	if (!tensors_init)
1427		_ccv_cnnp_model_tensors_init(model, compiled_data);
1428	else if ((uintptr_t)compiled_data->tensors_init.v & (uintptr_t)1)
1429	// Check if it is not fully allocated, if it is not, init_1.
1430		ccv_cnnp_model_tensors_init_1(model, compiled_data);
1431	ccv_array_t* const tensor_binds = ccv_array_new(sizeof(ccv_nnc_tensor_bind_t), 0, 0);
1432	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"
, 1432, __extension__ __PRETTY_FUNCTION__); }));
1433	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"
, 1433, __extension__ __PRETTY_FUNCTION__); }));
1434	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", 1434
, __extension__ __PRETTY_FUNCTION__); }));
1435	const int input_size_per_p = input_size / parallel_count;
1436	_ccv_cnnp_model_bind_tensors(model->graph, model->inputs, inputs, input_size_per_p, parallel_count, tensor_binds);
1437	const int output_size_per_p = output_size / parallel_count;
1438	_ccv_cnnp_model_bind_tensors(model->graph, model->outputs, outputs, output_size_per_p, parallel_count, tensor_binds);
1439	const int fit_size_per_p = fit_size / parallel_count;
1440	_ccv_cnnp_model_bind_tensors(model->graph, compiled_data->fits, fits, fit_size_per_p, parallel_count, tensor_binds);
1441	const int parameter_size = compiled_data->parameters->rnum;
1442	_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);
1443	_ccv_cnnp_model_bind_tensors(model->graph, compiled_data->updated_parameters, compiled_data->tensors.parameters, parameter_size, parallel_count, tensor_binds);
1444	const int internal_size = compiled_data->internals->rnum;
1445	_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);
1446	_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);
1447	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);
1448	ccv_array_free(tensor_binds);
1449	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));
1450	if (tensors_init && parallel_count > 1)
1451		_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);
1452	// If tensor is not init'ed, we need to init states first.
1453	if (_ccv_cnnp_any_to_init(compiled_data))
1454	{
1455		ccv_nnc_tensor_init_states_t tensor_init_states = {
1456			.parallel_count = parallel_count,
1457			.graph = model->graph,
1458			.compiled_data = compiled_data,
1459			.tensor_arena = compiled_data->tensor_arena
1460		};
1461		ccv_cnnp_model_init_states(model, model->graph, _ccv_cnnp_init_states_for_tensors, &tensor_init_states);
1462	}
1463	compiled_data->is_test = 0;
1464	const int saved_aux_size = ccv_nnc_minimizer_saved_aux_size(compiled_data->minimize.minimizer);
1465	// No need to set because it is default to training mode.
1466	// ccv_cnnp_model_set_is_test(model, 0, _ccv_cnnp_cmd_update_for_execs, &update);
1467	for (i = 0; i < saved_aux_size * parameter_size; i++)
1468	{
1469		if (compiled_data->saved_aux[i].source.d == CCV_NNC_NO_TENSOR_SYMBOL)
1470			continue;
1471		ccv_nnc_tensor_t* const tensor = ccv_nnc_tensor_from_symbol(compiled_data->tensor_arena, compiled_data->saved_aux[i].source);
1472		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);
1473		for (j = 1; j < parallel_count; j++)
1474		{
1475			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));
1476			if (copy)
1477				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);
1478		}
1479	}
1480	const int evaluate_to_size = compiled_data->evaluate.to_size;
1481	compiled_data->evaluate.to_op_size = 0;
1482	for (i = 0; i < evaluate_to_size; i++)
1483	{
1484		ccv_nnc_graph_exec_t const to = ccv_nnc_graph_exec_from_symbol(compiled_data->graph_exec_arena, compiled_data->evaluate.tos[i]);
1485		if (to.graph)
1486			compiled_data->evaluate.to_ops[compiled_data->evaluate.to_op_size++] = to;
1487	}
1488	ccv_nnc_graph_set_default_static_schedule(compiled_data->graph, compiled_data->stream_type, model->max_stream_count);
1489	ccv_nnc_graph_autotune(compiled_data->graph, model->workspace_size, 0, TRAVERSE_FULL0,0,0,0);
1490}
1491 
1492ccv_nnc_stream_context_t* ccv_cnnp_model_default_stream(const ccv_cnnp_model_t* const model)
1493{
1494	const ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1495	if (!compiled_data || !compiled_data->graph)
1496		return 0;
1497	return ccv_nnc_graph_default_stream(compiled_data->graph);
1498}
1499 
1500uint64_t ccv_cnnp_model_memory_size(const ccv_cnnp_model_t* const model)
1501{
1502	const ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1503	if (!compiled_data || !compiled_data->tensor_arena)
1504		return 0;
1505	return ccv_nnc_tensor_arena_size(compiled_data->tensor_arena);
1506}
1507 
1508static 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)
1509{
1510	int i, j;
1511	for (i = 0; i < tensor_size; i++)
1512	{
1513		ccv_nnc_tensor_symbol_t tensor_symbol = tensor_symbols[i];
1514		if (tensor_symbol.d == CCV_NNC_NO_TENSOR_SYMBOL)
1515			continue;
1516		if (graph)
1517		{
1518			const ccv_nnc_tensor_symbol_t alias_to = ccv_nnc_tensor_symbol_alias_to(graph, tensor_symbol);
1519			if (alias_to.d != CCV_NNC_NO_TENSOR_SYMBOL)
1520				tensor_symbol = alias_to;
1521		}
1522		ccv_nnc_tensor_bind_symbol(tensor_arena, tensor_symbol, tensors[i]);
1523		for (j = 1; j < parallel_count; j++)
1524		{
1525			const ccv_nnc_tensor_symbol_t copy = ccv_nnc_tensor_symbol_copy(graph, tensor_symbol, j);
1526			if (copy.d != CCV_NNC_NO_TENSOR_SYMBOL)
1527				ccv_nnc_tensor_bind_symbol(tensor_arena, copy, tensors[i + tensor_size * j]);
1528		}
1529	}
1530}
1531 
1532void 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)
1533{
1534	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1535	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 1535, __extension__ __PRETTY_FUNCTION__); }));
1536	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; });
1537	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", 1537, __extension__ __PRETTY_FUNCTION__
); }));
1538	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", 1538, __extension__ __PRETTY_FUNCTION__
); }));
1539	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", 1539
, __extension__ __PRETTY_FUNCTION__); }));
1540	assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if
 (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c"
, 1540, __extension__ __PRETTY_FUNCTION__); }));
1541	if (!compiled_data->graph || compiled_data->graph_mode != CCV_CNNP_MODEL_GRAPH_FIT_MODE)
1542	{
1543		_ccv_cnnp_compiled_data_graph_free(compiled_data);
1544		_ccv_cnnp_compiled_data_backward_free(compiled_data);
1545		_ccv_cnnp_compiled_data_apply_gradients_free(compiled_data);
1546		// Compile the symbolic graph down only when needed.
1547		_ccv_cnnp_model_fit_jit(model, inputs, input_size, fits, fit_size, outputs, output_size);
1548	} else {
1549		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"
, 1549, __extension__ __PRETTY_FUNCTION__); }));
1550		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"
, 1550, __extension__ __PRETTY_FUNCTION__); }));
1551		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", 1551
, __extension__ __PRETTY_FUNCTION__); }));
1552		const int input_size_per_p = input_size / parallel_count;
1553		_ccv_cnnp_bind_tensors_to_arena(compiled_data->tensor_arena, model->graph, model->inputs, inputs, input_size_per_p, parallel_count);
1554		const int output_size_per_p = output_size / parallel_count;
1555		_ccv_cnnp_bind_tensors_to_arena(compiled_data->tensor_arena, model->graph, model->outputs, outputs, output_size_per_p, parallel_count);
1556		const int fit_size_per_p = fit_size / parallel_count;
1557		_ccv_cnnp_bind_tensors_to_arena(compiled_data->tensor_arena, model->graph, compiled_data->fits, fits, fit_size_per_p, parallel_count);
1558	}
1559	if (compiled_data->is_test)
1560	{
1561		compiled_data->is_test = 0;
1562		ccv_nnc_graph_exec_update_t update = {
1563			.parallel_count = parallel_count,
1564			.graph = model->graph,
1565			.graph_exec_arena = compiled_data->graph_exec_arena,
1566		};
1567		ccv_cnnp_model_set_is_test(model, 0, _ccv_cnnp_cmd_update_for_execs, &update);
1568	}
1569	ccv_nnc_graph_run_with_schedule(compiled_data->graph, 0, 0, tensor_tape, stream_context);
1570}
1571 
1572// Compile the graph to run ccv_cnnp_model_evaluate with require_grad = false (MULTISTAGE_MODE_NO_GRAD).
1573static 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)
1574{
1575	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1576	compiled_data->graph_mode = CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE_NO_GRAD;
1577	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; });
1578	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", 1578, __extension__ __PRETTY_FUNCTION__
); }));
1579	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", 1579, __extension__ __PRETTY_FUNCTION__
); }));
1580	// If the gradient is not initialized, continue to setup parallel process. We don't init gradient here, but rather,
1581	// we setup proper rewindables so the graph can be rewinded to previous state before we run data parallel.
1582	if (parallel_count > 1 && compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_NONE)
1583	{
1584		const int evaluate_to_size = compiled_data->evaluate.to_size;
1585		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);
1586		_ccv_cnnp_model_set_rewindables(model);
1587		ccv_nnc_symbolic_graph_data_parallel(model->graph, parallel_count,
1588			0, 0,
1589			0, 0, 0,
1590			0, 0, 0,
1591			CCV_NNC_PARALLEL_REDUCE_OP_SUM,
1592			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));
1593		ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
1594		int i, j;
1595		for (i = 0; i < evaluate_to_size; i++)
1596			for (j = 1; j < parallel_count; j++)
1597			{
1598				const ccv_nnc_graph_exec_symbol_t copy = ccv_nnc_graph_exec_symbol_copy(model->graph, compiled_data->evaluate.tos[i], j);
1599				if (copy.d != CCV_NNC_NO_GRAPH_EXEC_SYMBOL)
1600					compiled_data->evaluate.tos[compiled_data->evaluate.to_size++] = copy;
1601			}
1602	}
1603	const int tensors_init = !!compiled_data->tensors_init.v;
1604	if (!tensors_init)
1605		_ccv_cnnp_model_tensors_init(model, compiled_data);
1606	else if ((uintptr_t)compiled_data->tensors_init.v & (uintptr_t)1)
1607	// Check if it is not fully allocated, if it is not, init_1.
1608		ccv_cnnp_model_tensors_init_1(model, compiled_data);
1609	ccv_array_t* const tensor_binds = ccv_array_new(sizeof(ccv_nnc_tensor_bind_t), 0, 0);
1610	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"
, 1610, __extension__ __PRETTY_FUNCTION__); }));
1611	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"
, 1611, __extension__ __PRETTY_FUNCTION__); }));
1612	const int input_size_per_p = input_size / parallel_count;
1613	_ccv_cnnp_model_bind_tensors(model->graph, model->inputs, inputs, input_size_per_p, parallel_count, tensor_binds);
1614	const int output_size_per_p = output_size / parallel_count;
1615	_ccv_cnnp_model_bind_tensors(model->graph, model->outputs, outputs, output_size_per_p, parallel_count, tensor_binds);
1616	const int parameter_size = compiled_data->parameters->rnum;
1617	_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);
1618	const int internal_size = compiled_data->internals->rnum;
1619	_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);
1620	_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);
1621	// If we generated gradient for the graph, only compile part of the graph because the rest is irrelevant for evaluation.
1622	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);
1623	ccv_array_free(tensor_binds);
1624	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));
1625	// If tensor is not init'ed, we need to init states first.
1626	if (tensors_init && parallel_count > 1)
1627		_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);
1628	if (_ccv_cnnp_any_to_init(compiled_data))
1629	{
1630		ccv_nnc_tensor_init_states_t tensor_init_states = {
1631			.parallel_count = parallel_count,
1632			.graph = model->graph,
1633			.compiled_data = compiled_data,
1634			.tensor_arena = compiled_data->tensor_arena
1635		};
1636		ccv_cnnp_model_init_states(model, model->graph, _ccv_cnnp_init_states_for_tensors, &tensor_init_states);
1637	}
1638	compiled_data->is_test = 1;
1639	ccv_nnc_graph_exec_update_t update = {
1640		.parallel_count = parallel_count,
1641		.graph = model->graph,
1642		.graph_exec_arena = compiled_data->graph_exec_arena,
1643	};
1644	ccv_cnnp_model_set_is_test(model, 1, _ccv_cnnp_cmd_update_for_execs, &update);
1645	ccv_nnc_graph_set_default_static_schedule(compiled_data->graph, compiled_data->stream_type, model->max_stream_count);
1646	ccv_nnc_graph_autotune(compiled_data->graph, model->workspace_size, 0, TRAVERSE_FULL0,0,0,0);
1647}
1648 
1649static void _ccv_cnnp_model_gradient_tensors_init(const ccv_cnnp_model_t* const model, ccv_cnnp_compiled_data_t* const compiled_data)
1650{
1651	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", 1651, __extension__ __PRETTY_FUNCTION__
); }));
1652	const int parameter_size = compiled_data->parameters->rnum;
1653	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; });
1654	compiled_data->tensors.gradients = (ccv_nnc_tensor_t**)ccmallocmalloc(sizeof(ccv_nnc_tensor_t*) * parameter_size * 2 * parallel_count);
1655	compiled_data->tensors.accum_gradients = compiled_data->tensors.gradients + parameter_size * parallel_count;
1656	int i, j;
1657	for (i = 0; i < parameter_size; i++)
1658	{
1659		if (compiled_data->parameter_flags && !(compiled_data->parameter_flags[i >> 6] & ((uint64_t)1 << (i & 63))))
1660		{
1661			compiled_data->tensors.gradients[i] = 0;
1662			compiled_data->tensors.accum_gradients[i] = 0;
1663			for (j = 1; j < parallel_count; j++)
1664			{
1665				compiled_data->tensors.gradients[i + j * parameter_size] = 0;
1666				compiled_data->tensors.accum_gradients[i + j * parameter_size] = 0;
1667			}
1668			continue;
1669		}
1670		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)));
1671		ccv_nnc_tensor_param_t info = ccv_nnc_tensor_symbol_params(parameter.graph, parameter);
1672		if (CCV_TENSOR_GET_DEVICE(info.type)((info.type) & 0xfff00) == CCV_COMPUTE_DEVICE_ANY)
1673			CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff
) << 8));
1674		const int device_id = CCV_TENSOR_GET_DEVICE_ID(info.type)(((info.type) & 0xfff00) >> 8);
1675		compiled_data->tensors.gradients[i] = ccv_nnc_tensor_new(0, info, 0);
1676		compiled_data->tensors.accum_gradients[i] = 0; // delay the accumulated gradient allocation until when we need it.
1677		for (j = 1; j < parallel_count; j++)
1678		{
1679			if (j != device_id)
1680				CCV_TENSOR_SET_DEVICE_ID(info.type, j)(info.type) = (((info.type) & ~0xfff00) | (((j) & 0xfff
) << 8));
1681			else
1682				CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff
) << 8));
1683			compiled_data->tensors.gradients[i + j * parameter_size] = ccv_nnc_tensor_new(0, info, 0);
1684			compiled_data->tensors.accum_gradients[i + j * parameter_size] = 0;
1685		}
1686	}
1687}
1688 
1689static int _ccv_cnnp_is_disable_outgrad_all(const uint64_t disable_outgrad, const int input_size)
1690{
1691	if (disable_outgrad == CCV_CNNP_DISABLE_OUTGRAD_ALL)
1692		return 1;
1693	if (disable_outgrad == CCV_CNNP_DISABLE_OUTGRAD_NONE)
1694		return 0;
1695	int i;
1696	for (i = 0; i < input_size; i++)
1697		if (!(disable_outgrad & ((uint64_t)1 << i)))
1698			return 0;
1699	return 1;
1700}
1701 
1702// Compile the graph to run ccv_cnnp_model_evaluate with requires_grad = true (MULTISTAGE_MODE).
1703// Particularly, this method compiles the evaluation and backprop graph (the main graph).
1704static 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)
1705{
1706	int i, j;
1707	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1708	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;
1709	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", 1709, __extension__ __PRETTY_FUNCTION__
); }));
1710	compiled_data->graph_mode = CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE;
1711	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; });
1712	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", 1712, __extension__ __PRETTY_FUNCTION__
); }));
1713	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", 1713, __extension__ __PRETTY_FUNCTION__
); }));
1714	// There shouldn't be a loss function if we evaluate with multistage jit.
1715	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", 1715, __extension__ __PRETTY_FUNCTION__
); }));
1716	if (compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_NONE)
1717	{
1718		_ccv_cnnp_model_set_rewindables(model);
1719		_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.
1720	} else if (compiled_data->gradient_mode != target_gradient_mode) {
1721		_ccv_cnnp_model_rewind_graph(model);
1722		_ccv_cnnp_compiled_data_gradient_free(compiled_data);
1723		compiled_data->gradient_mode = CCV_CNNP_COMPILED_DATA_GRADIENT_NONE;
1724		_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.
1725	}
1726	const int tensors_init = !!compiled_data->tensors_init.v;
1727	if (!tensors_init)
1728		_ccv_cnnp_model_tensors_init(model, compiled_data);
1729	else if ((uintptr_t)compiled_data->tensors_init.v & (uintptr_t)1)
1730	// Check if it is not fully allocated, if it is not, init_1.
1731		ccv_cnnp_model_tensors_init_1(model, compiled_data);
1732	ccv_array_t* const tensor_binds = ccv_array_new(sizeof(ccv_nnc_tensor_bind_t), 0, 0);
1733	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"
, 1733, __extension__ __PRETTY_FUNCTION__); }));
1734	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"
, 1734, __extension__ __PRETTY_FUNCTION__); }));
1735	const int input_size_per_p = input_size / parallel_count;
1736	_ccv_cnnp_model_bind_tensors(model->graph, model->inputs, inputs, input_size_per_p, parallel_count, tensor_binds);
1737	const int output_size_per_p = output_size / parallel_count;
1738	_ccv_cnnp_model_bind_tensors(model->graph, model->outputs, outputs, output_size_per_p, parallel_count, tensor_binds);
1739	const int parameter_size = compiled_data->parameters->rnum;
1740	_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);
1741	const int internal_size = compiled_data->internals->rnum;
1742	_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);
1743	_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);
1744	if (!compiled_data->tensors.gradients)
1745		_ccv_cnnp_model_gradient_tensors_init(model, compiled_data);
1746	_ccv_cnnp_model_bind_tensors(model->graph, compiled_data->gradients, compiled_data->tensors.gradients, parameter_size, parallel_count, tensor_binds);
1747	if (compiled_data->backward.to_size > 0)
1748		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);
1749	else
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->evaluate.tos, compiled_data->evaluate.to_size, &compiled_data->graph, &compiled_data->tensor_arena, &compiled_data->graph_exec_arena);
1751	ccv_array_free(tensor_binds);
1752	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));
1753	if (tensors_init && parallel_count > 1)
1754		_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);
1755	// If tensor is not init'ed, we need to init states first.
1756	if (_ccv_cnnp_any_to_init(compiled_data))
1757	{
1758		ccv_nnc_tensor_init_states_t tensor_init_states = {
1759			.parallel_count = parallel_count,
1760			.graph = model->graph,
1761			.compiled_data = compiled_data,
1762			.tensor_arena = compiled_data->tensor_arena
1763		};
1764		ccv_cnnp_model_init_states(model, model->graph, _ccv_cnnp_init_states_for_tensors, &tensor_init_states);
1765	}
1766	compiled_data->is_test = is_test;
1767	ccv_nnc_graph_exec_update_t update = {
1768		.parallel_count = parallel_count,
1769		.graph = model->graph,
1770		.graph_exec_arena = compiled_data->graph_exec_arena,
1771	};
1772	ccv_cnnp_model_set_is_test(model, is_test, _ccv_cnnp_cmd_update_for_execs, &update);
1773	const int evaluate_to_size = compiled_data->evaluate.to_size;
1774	compiled_data->evaluate.to_op_size = 0;
1775	ccv_array_t* const backward_from = ccv_array_new(sizeof(int), 0, 0);
1776	for (i = 0; i < evaluate_to_size; i++)
1777	{
1778		ccv_nnc_graph_exec_t const to_op = ccv_nnc_graph_exec_from_symbol(compiled_data->graph_exec_arena, compiled_data->evaluate.tos[i]);
1779		if (to_op.graph)
1780			compiled_data->evaluate.to_ops[compiled_data->evaluate.to_op_size++] = to_op;
1781		const int* tos;
1782		int to_size;
1783		ccv_nnc_graph_exec_symbol_to(model->graph, compiled_data->evaluate.tos[i], &tos, &to_size);
1784		for (j = 0; j < to_size; j++)
1785		{
1786			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){
1787				.d = tos[j],
1788				.graph = model->graph
1789			});
1790			if (to_op.graph)
1791				ccv_array_add_unique_int(backward_from, to_op.d);
1792		}
1793	}
1794	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", 1794, __extension__
 __PRETTY_FUNCTION__); }));
1795	compiled_data->backward.from_op_size = backward_from->rnum;
1796	compiled_data->backward.from_ops = (ccv_nnc_graph_exec_t*)ccmallocmalloc(sizeof(ccv_nnc_graph_exec_t) * backward_from->rnum);
1797	for (i = 0; i < backward_from->rnum; i++)
1798		compiled_data->backward.from_ops[i] = (ccv_nnc_graph_exec_t){
1799			.d = *(int*)ccv_array_get(backward_from, i)((void*)(((char*)((backward_from)->data)) + (size_t)(backward_from
)->rsize * (size_t)(i))),
1800			.graph = compiled_data->graph,
1801		};
1802	ccv_array_free(backward_from);
1803	ccv_nnc_graph_set_default_static_schedule(compiled_data->graph, compiled_data->stream_type, model->max_stream_count);
1804	ccv_nnc_graph_autotune(compiled_data->graph, model->workspace_size, 0, TRAVERSE_FULL0,0,0,0);
1805}
1806 
1807void 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)
1808{
1809	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1810	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 1810, __extension__ __PRETTY_FUNCTION__); }));
1811	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; });
1812	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", 1812, __extension__ __PRETTY_FUNCTION__
); }));
1813	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", 1813, __extension__ __PRETTY_FUNCTION__
); }));
1814	assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if
 (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c"
, 1814, __extension__ __PRETTY_FUNCTION__); }));
1815	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;
1816	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));
1817	if (!compiled_data->graph || mode_mismatch)
1818	{
1819		_ccv_cnnp_compiled_data_graph_free(compiled_data);
1820		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.
1821			_ccv_cnnp_compiled_data_backward_free(compiled_data);
1822		if (params.requires_grad)
1823			_ccv_cnnp_model_multistage_jit_0(model, params.disable_outgrad, params.is_test, inputs, input_size, outputs, output_size);
1824		else
1825			_ccv_cnnp_model_multistage_no_grad_jit(model, inputs, input_size, outputs, output_size);
1826	} else {
1827		ccv_nnc_tensor_arena_clear_bindings(compiled_data->tensor_arena);
1828		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"
, 1828, __extension__ __PRETTY_FUNCTION__); }));
1829		const int input_size_per_p = input_size / parallel_count;
1830		_ccv_cnnp_bind_tensors_to_arena(compiled_data->tensor_arena, model->graph, model->inputs, inputs, input_size_per_p, parallel_count);
1831		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"
, 1831, __extension__ __PRETTY_FUNCTION__); }));
1832		const int output_size_per_p = output_size / parallel_count;
1833		_ccv_cnnp_bind_tensors_to_arena(compiled_data->tensor_arena, model->graph, model->outputs, outputs, output_size_per_p, parallel_count);
1834	}
1835	if (compiled_data->is_test != params.is_test)
1836	{
1837		compiled_data->is_test = params.is_test;
1838		ccv_nnc_graph_exec_update_t update = {
1839			.parallel_count = parallel_count,
1840			.graph = model->graph,
1841			.graph_exec_arena = compiled_data->graph_exec_arena,
1842		};
1843		ccv_cnnp_model_set_is_test(model, params.is_test, _ccv_cnnp_cmd_update_for_execs, &update);
1844	}
1845}
1846 
1847void 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)
1848{
1849	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1850	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 1850, __extension__ __PRETTY_FUNCTION__); }));
1851	ccv_cnnp_model_dry_run(model, params, inputs, input_size, outputs, output_size);
1852	if (compiled_data->graph_mode == CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE_NO_GRAD)
1853		ccv_nnc_graph_run_with_schedule(compiled_data->graph, 0, 0, tensor_tape, stream_context);
1854	else {
1855		if (!compiled_data->evaluate.schedule)
1856			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);
1857		ccv_nnc_graph_run_with_schedule(compiled_data->graph, 0, compiled_data->evaluate.schedule, tensor_tape, stream_context);
1858	}
1859}
1860 
1861// Compile the graph to run ccv_cnnp_model_backward after ccv_cnnp_model_evaluate with requires_grad = true (MULTISTAGE_MODE).
1862// Particularly, this method compiles the accumulator graph.
1863static void _ccv_cnnp_model_multistage_jit_1(ccv_cnnp_model_t* const model)
1864{
1865	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1866	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 1866, __extension__ __PRETTY_FUNCTION__); }));
1867	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", 1867, __extension__ __PRETTY_FUNCTION__
); }));
1868	ccv_nnc_symbolic_graph_t* accum = ccv_nnc_symbolic_graph_new();
1869	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; });
1870	const int parameter_size = compiled_data->parameters->rnum;
1871	int i, j;
1872	compiled_data->backward.gradients = (ccv_nnc_tensor_symbol_t*)ccmallocmalloc(sizeof(ccv_nnc_tensor_symbol_t) * parameter_size * parallel_count * 3);
1873	compiled_data->backward.accum_gradients = compiled_data->backward.gradients + parameter_size * parallel_count;
1874	compiled_data->backward.updated_accum_gradients = compiled_data->backward.accum_gradients + parameter_size * parallel_count;
1875	for (i = 0; i < parameter_size; i++)
1876		for (j = 0; j < parallel_count; j++)
1877			if (compiled_data->tensors.gradients[i + j * parameter_size])
1878			{
1879				const ccv_nnc_tensor_param_t info = compiled_data->tensors.gradients[i + j * parameter_size]->info;
1880				// Now, the old gradient is the accumulated gradient, getting new gradient tensor setup so we can collect them.
1881				compiled_data->tensors.accum_gradients[i + j * parameter_size] = compiled_data->tensors.gradients[i + j * parameter_size];
1882				compiled_data->tensors.gradients[i + j * parameter_size] = ccv_nnc_tensor_new(0, info, 0);
1883				ccv_nnc_tensor_symbol_t inputs[2];
1884				inputs[0] = compiled_data->backward.accum_gradients[i + j * parameter_size] = ccv_nnc_tensor_symbol_new(accum, info, 0);
1885				inputs[1] = compiled_data->backward.gradients[i + j * parameter_size] = ccv_nnc_tensor_symbol_new(accum, info, 0);
1886				ccv_nnc_tensor_symbol_t output = compiled_data->backward.updated_accum_gradients[i + j * parameter_size] = ccv_nnc_tensor_symbol_new(accum, info, 0);
1887				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);
1888			} else {
1889				compiled_data->backward.accum_gradients[i + j * parameter_size] = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
1890				compiled_data->backward.gradients[i + j * parameter_size] = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
1891				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
};
1892			}
1893	ccv_nnc_graph_exec_symbol_autogen(accum, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
1894	if (ccv_nnc_symbolic_graph_source_size(accum) == 0)
1895	{
1896		ccv_nnc_symbolic_graph_free(accum);
1897		// Create empty graph.
1898		compiled_data->backward.accum = ccv_nnc_graph_new();
1899		ccv_nnc_graph_topsort(compiled_data->backward.accum, 0, 0);
1900		return;
1901	}
1902	ccv_array_t* const tensor_binds = ccv_array_new(sizeof(ccv_nnc_tensor_bind_t), 0, 0);
1903	_ccv_cnnp_model_bind_tensors(accum, compiled_data->backward.accum_gradients, compiled_data->tensors.accum_gradients, parameter_size * parallel_count, 1, tensor_binds);
1904	_ccv_cnnp_model_bind_tensors(accum, compiled_data->backward.gradients, compiled_data->tensors.gradients, parameter_size * parallel_count, 1, tensor_binds);
1905	_ccv_cnnp_model_bind_tensors(accum, compiled_data->backward.updated_accum_gradients, compiled_data->tensors.accum_gradients, parameter_size * parallel_count, 1, tensor_binds);
1906	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);
1907	ccv_nnc_symbolic_graph_free(accum);
1908	ccv_array_free(tensor_binds);
1909	ccv_nnc_graph_set_default_static_schedule(compiled_data->backward.accum, compiled_data->stream_type, model->max_stream_count);
1910}
1911 
1912void 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)
1913{
1914	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1915	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 1915, __extension__ __PRETTY_FUNCTION__); }));
1916	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", 1916, __extension__ __PRETTY_FUNCTION__
); }));
1917	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; });
1918	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", 1918, __extension__ __PRETTY_FUNCTION__
); }));
1919	if (outgrad_size > 0)
1920		{ 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", 1920, __extension__ __PRETTY_FUNCTION__
); })); }
1921	assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if
 (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c"
, 1921, __extension__ __PRETTY_FUNCTION__); }));
1922	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", 1922, __extension__ __PRETTY_FUNCTION__
); }));
1923	const int parameter_size = compiled_data->parameters->rnum;
1924	// If we need to accumulate the gradients now, do jit on accumulator.
1925	if (compiled_data->backward.count > 0)
1926	{
1927		if (!compiled_data->backward.accum)
1928			_ccv_cnnp_model_multistage_jit_1(model);
1929		else if (compiled_data->backward.count == 1) {
1930			//  On this round, we need to switch accumulated gradients with gradients (so we can do accumulation properly).
1931			int i;
1932			for (i = 0; i < parameter_size * parallel_count; i++)
1933			{
1934				ccv_nnc_tensor_t* tensor;
1935				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));
1936			}
1937			if (compiled_data->backward.tensor_arena)
1938			{
1939				ccv_nnc_tensor_arena_clear_bindings(compiled_data->backward.tensor_arena);
1940				// Do rebind in case we messed up the binding (we switch accum_gradients and gradients).
1941				_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);
1942				_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);
1943				_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);
1944			}
1945		}
1946	}
1947	const int ingrad_size_per_p = model->output_size;
1948	const int outgrad_size_per_p = compiled_data->outgrad_size;
1949	int i, j;
1950	for (i = 0; i < ingrad_size_per_p; i++)
1951	{
1952		const ccv_nnc_tensor_symbol_t ingrad = ccv_nnc_tensor_symbol_for_backward(model->graph, compiled_data->f[i]);
1953		if (!ingrad_size || !ingrads || ingrads[i] == 0)
1954		{
1955			// Set it to 1 if it is not specified.
1956			ccv_nnc_tensor_t* const ingrad_tensor = ccv_nnc_tensor_from_symbol(compiled_data->tensor_arena, ingrad);
1957			if (ingrad_tensor)
1958				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);
1959			for (j = 1; j < parallel_count; j++)
1960			{
1961				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));
1962				if (ingrad_tensor)
1963					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);
1964			}
1965		} else {
1966			// Make sure the length matches, in case it is an alias.
1967			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", 1967, __extension__ __PRETTY_FUNCTION__
); }));
1968			ccv_nnc_tensor_bind_symbol(compiled_data->tensor_arena, ingrad, ingrads[i]);
1969			for (j = 1; j < parallel_count; j++)
1970				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]);
1971		}
1972	}
1973	if (outgrad_size > 0)
1974	{
1975		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", 1975, __extension__ __PRETTY_FUNCTION__
); }));
1976		for (i = 0; i < outgrad_size_per_p; i++)
1977			if (outgrads[i])
1978			{
1979				const ccv_nnc_tensor_symbol_t outgrad = compiled_data->outgrads[i];
1980				ccv_nnc_tensor_bind_symbol(compiled_data->tensor_arena, outgrad, outgrads[i]);
1981				for (j = 1; j < parallel_count; j++)
1982					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]);
1983			}
1984	} else {
1985		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", 1986, __extension__ __PRETTY_FUNCTION__
); }))
1986			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", 1986, __extension__ __PRETTY_FUNCTION__
); }));
1987	}
1988	// We need to rebind here because in ccv_cnnp_evaluate, we clear bindings, that will reset all bindings for the gradients.
1989	// For parameters and internals these are fine because when we clear bindings, it restores to original bindings, which are these
1990	// parameters and internals. The same cannot be said for gradients due to the accum_gradients switching.
1991	_ccv_cnnp_bind_tensors_to_arena(compiled_data->tensor_arena, model->graph, compiled_data->gradients, compiled_data->tensors.gradients, parameter_size, parallel_count);
1992	if (!compiled_data->backward.schedule)
1993		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);
1994	// Run the backward pass.
1995	ccv_nnc_graph_run_with_schedule(compiled_data->graph, 0, compiled_data->backward.schedule, tensor_tape, stream_context);
1996	// If we need to run accumulation round, do that now.
1997	if (compiled_data->backward.count > 0)
1998		ccv_nnc_graph_run_with_schedule(compiled_data->backward.accum, 0, 0, 0, stream_context);
1999	// Update the count, this determines whether we need to accumulate or not.
2000	++compiled_data->backward.count;
2001}
2002 
2003// Compile the graph to run ccv_cnnp_model_apply_gradients after ccv_cnnp_model_backward (MULTISTAGE_MODE).
2004// Particularly, this method compiles the parameter update graph.
2005static void _ccv_cnnp_model_multistage_jit_2(ccv_cnnp_model_t* const model)
2006{
2007	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2008	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", 2008, __extension__ __PRETTY_FUNCTION__
); }));
2009	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; });
2010	const int parameter_size = compiled_data->parameters->rnum;
2011	ccv_array_t* const tensor_binds = ccv_array_new(sizeof(ccv_nnc_tensor_bind_t), 0, 0);
2012	_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);
2013	_ccv_cnnp_model_bind_tensors(model->graph, compiled_data->updated_parameters, compiled_data->tensors.parameters, parameter_size, parallel_count, tensor_binds);
2014	// Bind accumulated gradients.
2015	if (compiled_data->backward.count > 1)
2016		_ccv_cnnp_model_bind_tensors(model->graph, compiled_data->gradients, compiled_data->tensors.accum_gradients, parameter_size, parallel_count, tensor_binds);
2017	else
2018		_ccv_cnnp_model_bind_tensors(model->graph, compiled_data->gradients, compiled_data->tensors.gradients, parameter_size, parallel_count, tensor_binds);
2019	ccv_array_t* const apply_gradients_from = ccv_array_new(sizeof(int), 0, 0);
2020	int i, j;
2021	for (i = 0; i < compiled_data->backward.to_size; i++)
2022	{
2023		const int* tos;
2024		int to_size;
2025		ccv_nnc_graph_exec_symbol_to(model->graph, compiled_data->backward.tos[i], &tos, &to_size);
2026		for (j = 0; j < to_size; j++)
2027		{
2028			// Check if this is already show up in the backward graph, if that is the case, it won't be in the apply
2029			// gradients graph.
2030			const ccv_nnc_graph_exec_t exec = ccv_nnc_graph_exec_from_symbol(compiled_data->graph_exec_arena, (ccv_nnc_graph_exec_symbol_t){
2031				.d = tos[j],
2032				.graph = model->graph,
2033			});
2034			if (!exec.graph)
2035				ccv_array_add_unique_int(apply_gradients_from, tos[j]);
2036		}
2037	}
2038	const int from_size = apply_gradients_from->rnum;
2039	if (from_size == 0)
2040	{
2041		ccv_array_free(apply_gradients_from);
2042		ccv_array_free(tensor_binds);
2043		return;
2044	}
2045	ccv_nnc_graph_exec_symbol_t* const froms = (ccv_nnc_graph_exec_symbol_t*)ccmallocmalloc(sizeof(ccv_nnc_graph_exec_symbol_t) * from_size);
2046	for (i = 0; i < from_size; i++)
2047		froms[i] = (ccv_nnc_graph_exec_symbol_t){
2048			.d = *(int*)ccv_array_get(apply_gradients_from, i)((void*)(((char*)((apply_gradients_from)->data)) + (size_t
)(apply_gradients_from)->rsize * (size_t)(i))),
2049			.graph = model->graph
2050		};
2051	ccv_array_free(apply_gradients_from);
2052	// It can only ends with updates on the parameters.
2053	ccv_array_t* const tos = ccv_array_new(sizeof(ccv_nnc_graph_exec_symbol_t), parameter_size * parallel_count, 0);
2054	for (i = 0;  i < parameter_size; i++)
2055	{
2056		if (compiled_data->update_nodes[i].d == CCV_NNC_NO_TENSOR_SYMBOL)
2057			continue;
2058		ccv_array_push(tos, &compiled_data->update_nodes[i]);
2059		for (j = 1; j < parallel_count; j++)
2060		{
2061			const ccv_nnc_graph_exec_symbol_t copy = ccv_nnc_graph_exec_symbol_copy(model->graph, compiled_data->update_nodes[i], j);
2062			ccv_array_push(tos, &copy);
2063		}
2064	}
2065	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);
2066	ccv_array_free(tos);
2067	ccv_array_free(tensor_binds);
2068	ccfreefree(froms);
2069	const int max_saved_aux_size = compiled_data->minimize.max_saved_aux_size;
2070	for (i = 0; i < max_saved_aux_size * parameter_size; i++)
2071	{
2072		// Skip on no tensor.
2073		if (compiled_data->saved_aux[i].source.d == CCV_NNC_NO_TENSOR_SYMBOL)
2074			continue;
2075		ccv_nnc_tensor_t* const tensor = ccv_nnc_tensor_from_symbol(compiled_data->apply_gradients.tensor_arena, compiled_data->saved_aux[i].source);
2076		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);
2077		for (j = 1; j < parallel_count; j++)
2078		{
2079			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));
2080			if (copy)
2081				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);
2082		}
2083	}
2084	ccv_nnc_graph_set_default_static_schedule(compiled_data->apply_gradients.graph, compiled_data->stream_type, model->max_stream_count);
2085}
2086 
2087void ccv_cnnp_model_apply_gradients(ccv_cnnp_model_t* const model, ccv_nnc_stream_context_t* const stream_context)
2088{
2089	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2090	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 2090, __extension__ __PRETTY_FUNCTION__); }));
2091	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", 2091, __extension__ __PRETTY_FUNCTION__
); }));
2092	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; });
2093	assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if
 (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c"
, 2093, __extension__ __PRETTY_FUNCTION__); }));
2094	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", 2094, __extension__ __PRETTY_FUNCTION__
); }));
2095	// Skip if there is no backward pass.
2096	if (compiled_data->backward.count <= 0)
2097		return;
2098	// Skip if there is no parameters.
2099	if (compiled_data->parameters->rnum == 0)
2100	{
2101		compiled_data->backward.count = 0;
2102		return;
2103	}
2104	if (!compiled_data->apply_gradients.graph)
2105		_ccv_cnnp_model_multistage_jit_2(model);
2106	else {
2107		const int parameter_size = compiled_data->parameters->rnum;
2108		ccv_nnc_tensor_arena_clear_bindings(compiled_data->apply_gradients.tensor_arena);
2109		// Change to bind accum_gradients if we do gradient accumulation (run backward more than once).
2110		if (compiled_data->backward.count > 1)
2111			_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);
2112		else
2113			_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);
2114	}
2115	if (compiled_data->apply_gradients.graph)
2116		ccv_nnc_graph_run_with_schedule(compiled_data->apply_gradients.graph, 0, 0, 0, stream_context);
2117	// Reset backward count to 0.
2118	compiled_data->backward.count = 0;
2119}
2120 
2121void 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)
2122{
2123	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2124	const int param_sel = parameter->param_sel > 0 ? parameter->param_sel - 1 : parameter->param_sel;
2125	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", 2125, __extension__ __PRETTY_FUNCTION__
); }));
2126	const int tensors_init = !!compiled_data->tensors_init.v;
2127	if (!tensors_init)
2128		_ccv_cnnp_model_tensors_init(model, compiled_data);
2129	else if ((uintptr_t)compiled_data->tensors_init.v & (uintptr_t)1)
2130	// Check if it is not fully allocated, if it is not, init_1.
2131		ccv_cnnp_model_tensors_init_1(model, compiled_data);
2132	ccv_array_t* const parameter_indices = ccv_array_new(sizeof(int), 0, 0);
2133	ccv_cnnp_model_add_to_parameter_indices(parameter->model, param_sel, parameter_indices);
2134	const int param_ref = parameter->param_ref > 0 ? parameter->param_ref - 1 : parameter->param_ref;
2135	if (param_ref < 0)
2136		{ 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", 2136
, __extension__ __PRETTY_FUNCTION__); })); }
2137	else
2138		{ 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", 2138, __extension__ __PRETTY_FUNCTION__
); })); }
2139	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)));
2140	ccv_array_free(parameter_indices);
2141	const int parameter_size = compiled_data->parameters->rnum;
2142	assert(d >= 0)((void) sizeof ((d >= 0) ? 1 : 0), __extension__ ({ if (d >=
 0) ; else __assert_fail ("d >= 0", "ccv_cnnp_model.c", 2142
, __extension__ __PRETTY_FUNCTION__); }));
2143	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", 2143, __extension__ __PRETTY_FUNCTION__
); }));
2144	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; });
2145	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));
2146	assert(dest)((void) sizeof ((dest) ? 1 : 0), __extension__ ({ if (dest) ;
 else __assert_fail ("dest", "ccv_cnnp_model.c", 2146, __extension__
 __PRETTY_FUNCTION__); }));
2147	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);
2148	int i;
2149	for (i = 1; i < parallel_count; i++)
2150	{
2151		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));
2152		if (copy_tensor)
2153			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);
2154	}
2155	// Mark this symbol as init'ed.
2156	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;
2157  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));
2158	init_v[s >> 5] |= (1u << (s & 0x1f));
2159}
2160 
2161void ccv_cnnp_model_parameter_copy(ccv_cnnp_model_t* const model, const ccv_cnnp_model_io_t parameter, ccv_nnc_tensor_t* const tensor)
2162{
2163	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2164	const int param_sel = parameter->param_sel > 0 ? parameter->param_sel - 1 : parameter->param_sel;
2165	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", 2165, __extension__ __PRETTY_FUNCTION__
); }));
2166	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", 2166, __extension__ __PRETTY_FUNCTION__
); }));
2167	ccv_array_t* const parameter_indices = ccv_array_new(sizeof(int), 0, 0);
2168	ccv_cnnp_model_add_to_parameter_indices(parameter->model, param_sel, parameter_indices);
2169	const int param_ref = parameter->param_ref > 0 ? parameter->param_ref - 1 : parameter->param_ref;
2170	if (param_ref < 0)
2171		{ 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", 2171
, __extension__ __PRETTY_FUNCTION__); })); }
2172	else
2173		{ 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", 2173, __extension__ __PRETTY_FUNCTION__
); })); }
2174	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)));
2175	ccv_array_free(parameter_indices);
2176	const int parameter_size = compiled_data->parameters->rnum;
2177	assert(d >= 0)((void) sizeof ((d >= 0) ? 1 : 0), __extension__ ({ if (d >=
 0) ; else __assert_fail ("d >= 0", "ccv_cnnp_model.c", 2177
, __extension__ __PRETTY_FUNCTION__); }));
2178	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", 2178, __extension__ __PRETTY_FUNCTION__
); }));
2179	// We don't need to consider parallel_count, every parameter on each device is identical.
2180	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));
2181	assert(src)((void) sizeof ((src) ? 1 : 0), __extension__ ({ if (src) ; else
 __assert_fail ("src", "ccv_cnnp_model.c", 2181, __extension__
 __PRETTY_FUNCTION__); }));
2182	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);
2183}
2184 
2185ccv_nnc_tensor_param_t ccv_cnnp_model_parameter_tensor_params(ccv_cnnp_model_t* const model, const ccv_cnnp_model_io_t parameter)
2186{
2187	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2188	const int param_sel = parameter->param_sel > 0 ? parameter->param_sel - 1 : parameter->param_sel;
2189	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", 2189, __extension__ __PRETTY_FUNCTION__
); }));
2190	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", 2190, __extension__ __PRETTY_FUNCTION__
); }));
2191	ccv_array_t* const parameter_indices = ccv_array_new(sizeof(int), 0, 0);
2192	ccv_cnnp_model_add_to_parameter_indices(parameter->model, param_sel, parameter_indices);
2193	const int param_ref = parameter->param_ref > 0 ? parameter->param_ref - 1 : parameter->param_ref;
2194	if (param_ref < 0)
2195		{ 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", 2195
, __extension__ __PRETTY_FUNCTION__); })); }
2196	else
2197		{ 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", 2197, __extension__ __PRETTY_FUNCTION__
); })); }
2198	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)));
2199	ccv_array_free(parameter_indices);
2200	const int parameter_size = compiled_data->parameters->rnum;
2201	assert(d >= 0)((void) sizeof ((d >= 0) ? 1 : 0), __extension__ ({ if (d >=
 0) ; else __assert_fail ("d >= 0", "ccv_cnnp_model.c", 2201
, __extension__ __PRETTY_FUNCTION__); }));
2202	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", 2202, __extension__ __PRETTY_FUNCTION__
); }));
2203	// We don't need to consider parallel_count, every parameter on each device is identical.
2204	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));
2205	assert(tensor)((void) sizeof ((tensor) ? 1 : 0), __extension__ ({ if (tensor
) ; else __assert_fail ("tensor", "ccv_cnnp_model.c", 2205, __extension__
 __PRETTY_FUNCTION__); }));
2206	return tensor->info;
2207}
2208 
2209const char* ccv_cnnp_model_parameter_name(ccv_cnnp_model_t* const model, const ccv_cnnp_model_io_t parameter)
2210{
2211	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2212	const int param_sel = parameter->param_sel > 0 ? parameter->param_sel - 1 : parameter->param_sel;
2213	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", 2213, __extension__ __PRETTY_FUNCTION__
); }));
2214	ccv_array_t* const parameter_indices = ccv_array_new(sizeof(int), 0, 0);
2215	ccv_cnnp_model_add_to_parameter_indices(parameter->model, param_sel, parameter_indices);
2216	const int param_ref = parameter->param_ref > 0 ? parameter->param_ref - 1 : parameter->param_ref;
2217	if (param_ref < 0)
2218		{ 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", 2218
, __extension__ __PRETTY_FUNCTION__); })); }
2219	else
2220		{ 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", 2220, __extension__ __PRETTY_FUNCTION__
); })); }
2221	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)));
2222	ccv_array_free(parameter_indices);
2223	const int parameter_size = compiled_data->parameters->rnum;
2224	assert(d >= 0)((void) sizeof ((d >= 0) ? 1 : 0), __extension__ ({ if (d >=
 0) ; else __assert_fail ("d >= 0", "ccv_cnnp_model.c", 2224
, __extension__ __PRETTY_FUNCTION__); }));
2225	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", 2225, __extension__ __PRETTY_FUNCTION__
); }));
2226	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)));
2227}
2228 
2229int ccv_cnnp_model_parameter_count(ccv_cnnp_model_t* const model)
2230{
2231	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", 2231, __extension__ __PRETTY_FUNCTION__
); }));
2232	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2233	return compiled_data->parameters->rnum;
2234}
2235 
2236ccv_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)
2237{
2238	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2239	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 2239, __extension__ __PRETTY_FUNCTION__); }));
2240	const int parameter_size = compiled_data->parameters->rnum;
2241	int i;
2242	for (i = 0; i < parameter_size; i++)
2243	{
2244		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)));
2245		if (first(model, name, context))
2246			return ccv_cnnp_model_parameters(model, -1, i);
2247	}
2248	return 0;
2249}
2250 
2251ccv_array_t* ccv_cnnp_model_parameters_filter(ccv_cnnp_model_t* const model, ccv_cnnp_model_parameters_filter_f filter, void* const context)
2252{
2253	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2254	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 2254, __extension__ __PRETTY_FUNCTION__); }));
2255	ccv_array_t* const parameters = ccv_array_new(sizeof(ccv_cnnp_model_io_t), 0, 0);
2256	const int parameter_size = compiled_data->parameters->rnum;
2257	int i;
2258	for (i = 0; i < parameter_size; i++)
2259	{
2260		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)));
2261		if (filter(model, name, context))
2262		{
2263			ccv_cnnp_model_io_t parameter = ccv_cnnp_model_parameters(model, -1, i);
2264			ccv_array_push(parameters, &parameter);
2265		}
2266	}
2267	return parameters;
2268 
2269}
2270 
2271static 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)
2272{
2273	const int to_param_sel = parameters->param_sel > 0 ? parameters->param_sel - 1 : parameters->param_sel;
2274	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", 2274, __extension__
 __PRETTY_FUNCTION__); }));
2275	ccv_array_t* const to_parameter_indices = ccv_array_new(sizeof(int), 0, 0);
2276	ccv_cnnp_model_add_to_parameter_indices(parameters->model, to_param_sel, to_parameter_indices);
2277	*param_ref = parameters->param_ref > 0 ? parameters->param_ref - 1 : parameters->param_ref;
2278	return to_parameter_indices;
2279}
2280 
2281static 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)
2282{
2283	// If the model is not compiled yet. Compile them now.
2284	if (!model->graph)
2285	{
2286		model->graph = ccv_nnc_symbolic_graph_new();
2287		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", 2287, __extension__ __PRETTY_FUNCTION__
); }));
2288		const int input_size = from_model->input_size;
2289		ccv_nnc_tensor_param_t input_params[input_size];
2290		int i;
2291		for (i = 0; i < input_size; i++)
2292			input_params[i] = ccv_nnc_tensor_symbol_params(from_model->graph, from_model->inputs[i]);
2293		_ccv_cnnp_model_compile(model, input_params, input_size, from_model->compiled_data->loss);
2294		model->parallel_count = from_model->parallel_count;
2295		model->memory_compression = from_model->memory_compression;
2296		model->memory_reduction = from_model->memory_reduction;
2297		model->gradient_checkpointing = from_model->gradient_checkpointing;
2298		model->compiled_data->stream_type = from_model->compiled_data->stream_type;
2299		model->compiled_data->minimize.minimizer = from_model->compiled_data->minimize.minimizer;
2300		model->compiled_data->minimize.max_saved_aux_size = from_model->compiled_data->minimize.max_saved_aux_size;
2301	}
2302	ccv_cnnp_compiled_data_t* const to_compiled_data = model->compiled_data;
2303	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", 2303, __extension__ __PRETTY_FUNCTION__
); }));
2304	const int to_tensors_init = !!to_compiled_data->tensors_init.v;
2305	if (!to_tensors_init)
2306	{
2307		if (only_init_0)
2308			ccv_cnnp_model_tensors_init_0(model, to_compiled_data);
2309		else
2310			_ccv_cnnp_model_tensors_init(model, to_compiled_data);
2311	} else if (!only_init_0 && (uintptr_t)to_compiled_data->tensors_init.v & (uintptr_t)1)
2312		// Check if it is not fully allocated, if it is not, init_1.
2313			ccv_cnnp_model_tensors_init_1(model, to_compiled_data);
2314	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", 2314, __extension__ __PRETTY_FUNCTION__
); }));
2315	*parameter_indices = _ccv_cnnp_model_parameter_indices(model, parameters, param_ref);
2316	*from_parameter_indices = _ccv_cnnp_model_parameter_indices(from_model, from_parameters, from_param_ref);
2317	if (*from_param_ref < 0 && *param_ref >= 0)
2318		{ 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", 2318, __extension__ __PRETTY_FUNCTION__
); })); }
2319	else if (*from_param_ref >= 0)
2320		{ 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", 2320, __extension__ __PRETTY_FUNCTION__
); })); }
2321	if (*param_ref < 0 && *from_param_ref >= 0)
2322		{ 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"
, 2322, __extension__ __PRETTY_FUNCTION__); })); }
2323	else if (*param_ref >= 0)
2324		{ 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", 2324, __extension__ __PRETTY_FUNCTION__
); })); }
2325}
2326 
2327void 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)
2328{
2329	ccv_array_t* to_parameter_indices;
2330	int to_param_ref;
2331	ccv_array_t* from_parameter_indices;
2332	int from_param_ref;
2333	_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);
2334	// Should be exactly the same tensor.
2335	if (to_param_ref < 0 && from_param_ref < 0)
2336		{ 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", 2336, __extension__ __PRETTY_FUNCTION__
); })); }
2337	// To models.
2338	ccv_cnnp_compiled_data_t* const to_compiled_data = model->compiled_data;
2339	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", 2339, __extension__ __PRETTY_FUNCTION__
); }));
2340	// From models.
2341	const ccv_cnnp_compiled_data_t* const from_compiled_data = from_model->compiled_data;
2342	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; });
2343	const int to_parameter_size = to_compiled_data->parameters->rnum;
2344	const int rnum = (to_param_ref < 0 && from_param_ref < 0) ? from_parameter_indices->rnum : 1;
2345	int i, j;
2346	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));
2347  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));
2348	for (i = 0; i < rnum; i++)
2349	{
2350		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)));
2351		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"
, 2351, __extension__ __PRETTY_FUNCTION__); }));
2352		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", 2352, __extension__ __PRETTY_FUNCTION__
); }));
2353		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;
2354		// If the original is not init'ed. We cannot copy from.
2355		if (!(from_init_v[s >> 5] & (1u << (s & 0x1f))))
2356			continue;
2357		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)));
2358		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"
, 2358, __extension__ __PRETTY_FUNCTION__); }));
2359		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", 2359, __extension__ __PRETTY_FUNCTION__
); }));
2360		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));
2361		assert(src)((void) sizeof ((src) ? 1 : 0), __extension__ ({ if (src) ; else
 __assert_fail ("src", "ccv_cnnp_model.c", 2361, __extension__
 __PRETTY_FUNCTION__); }));
2362		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));
2363		assert(dest)((void) sizeof ((dest) ? 1 : 0), __extension__ ({ if (dest) ;
 else __assert_fail ("dest", "ccv_cnnp_model.c", 2363, __extension__
 __PRETTY_FUNCTION__); }));
2364		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);
2365		for (j = 1; j < parallel_count; j++)
2366		{
2367			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));
2368			if (copy_tensor)
2369				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);
2370		}
2371		// Mark this symbol as init'ed.
2372		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;
2373		to_init_v[d >> 5] |= (1u << (d & 0x1f));
2374	}
2375	ccv_array_free(to_parameter_indices);
2376	ccv_array_free(from_parameter_indices);
2377}
2378 
2379KHASH_MAP_INIT_STR(ccv_cnnp_parameter_id, int)typedef struct kh_ccv_cnnp_parameter_id_s { khint_t n_buckets
, size, n_occupied, upper_bound; khint32_t *flags; kh_cstr_t *
keys; int *vals; } kh_ccv_cnnp_parameter_id_t; static inline __attribute__
 ((__unused__)) kh_ccv_cnnp_parameter_id_t *kh_init_ccv_cnnp_parameter_id
(void) { return (kh_ccv_cnnp_parameter_id_t*)calloc(1,sizeof(
kh_ccv_cnnp_parameter_id_t)); } static inline __attribute__ (
(__unused__)) void kh_destroy_ccv_cnnp_parameter_id(kh_ccv_cnnp_parameter_id_t
 *h) { if (h) { free((void *)h->keys); free(h->flags); free
((void *)h->vals); free(h); } } static inline __attribute__
 ((__unused__)) void kh_clear_ccv_cnnp_parameter_id(kh_ccv_cnnp_parameter_id_t
 *h) { if (h && h->flags) { memset(h->flags, 0xaa
, ((h->n_buckets) < 16? 1 : (h->n_buckets)>>4)
 * sizeof(khint32_t)); h->size = h->n_occupied = 0; } }
 static inline __attribute__ ((__unused__)) khint_t kh_get_ccv_cnnp_parameter_id
(const kh_ccv_cnnp_parameter_id_t *h, kh_cstr_t key) { if (h->
n_buckets) { khint_t k, i, last, mask, step = 0; mask = h->
n_buckets - 1; k = __ac_X31_hash_string(key); i = k & mask
; last = i; while (!((h->flags[i>>4]>>((i&
0xfU)<<1))&2) && (((h->flags[i>>4]
>>((i&0xfU)<<1))&1) || !(strcmp(h->keys
[i], key) == 0))) { i = (i + (++step)) & mask; if (i == last
) return h->n_buckets; } return ((h->flags[i>>4]>>
((i&0xfU)<<1))&3)? h->n_buckets : i; } else return
 0; } static inline __attribute__ ((__unused__)) int kh_resize_ccv_cnnp_parameter_id
(kh_ccv_cnnp_parameter_id_t *h, khint_t new_n_buckets) { khint32_t
 *new_flags = 0; khint_t j = 1; { (--(new_n_buckets), (new_n_buckets
)|=(new_n_buckets)>>1, (new_n_buckets)|=(new_n_buckets)
>>2, (new_n_buckets)|=(new_n_buckets)>>4, (new_n_buckets
)|=(new_n_buckets)>>8, (new_n_buckets)|=(new_n_buckets)
>>16, ++(new_n_buckets)); if (new_n_buckets < 4) new_n_buckets
 = 4; if (h->size >= (khint_t)(new_n_buckets * __ac_HASH_UPPER
 + 0.5)) j = 0; else { new_flags = (khint32_t*)malloc(((new_n_buckets
) < 16? 1 : (new_n_buckets)>>4) * sizeof(khint32_t))
; if (!new_flags) return -1; memset(new_flags, 0xaa, ((new_n_buckets
) < 16? 1 : (new_n_buckets)>>4) * sizeof(khint32_t))
; if (h->n_buckets < new_n_buckets) { kh_cstr_t *new_keys
 = (kh_cstr_t*)realloc((void *)h->keys,new_n_buckets * sizeof
(kh_cstr_t)); if (!new_keys) { free(new_flags); return -1; } h
->keys = new_keys; if (1) { int *new_vals = (int*)realloc(
(void *)h->vals,new_n_buckets * sizeof(int)); if (!new_vals
) { free(new_flags); return -1; } h->vals = new_vals; } } }
 } if (j) { for (j = 0; j != h->n_buckets; ++j) { if (((h->
flags[j>>4]>>((j&0xfU)<<1))&3) == 0
) { kh_cstr_t key = h->keys[j]; int val; khint_t new_mask;
 new_mask = new_n_buckets - 1; if (1) val = h->vals[j]; (h
->flags[j>>4]|=1ul<<((j&0xfU)<<1)); while
 (1) { khint_t k, i, step = 0; k = __ac_X31_hash_string(key);
 i = k & new_mask; while (!((new_flags[i>>4]>>
((i&0xfU)<<1))&2)) i = (i + (++step)) & new_mask
; (new_flags[i>>4]&=~(2ul<<((i&0xfU)<<
1))); if (i < h->n_buckets && ((h->flags[i>>
4]>>((i&0xfU)<<1))&3) == 0) { { kh_cstr_t
 tmp = h->keys[i]; h->keys[i] = key; key = tmp; } if (1
) { int tmp = h->vals[i]; h->vals[i] = val; val = tmp; }
 (h->flags[i>>4]|=1ul<<((i&0xfU)<<1)
); } else { h->keys[i] = key; if (1) h->vals[i] = val; break
; } } } } if (h->n_buckets > new_n_buckets) { h->keys
 = (kh_cstr_t*)realloc((void *)h->keys,new_n_buckets * sizeof
(kh_cstr_t)); if (1) h->vals = (int*)realloc((void *)h->
vals,new_n_buckets * sizeof(int)); } free(h->flags); h->
flags = new_flags; h->n_buckets = new_n_buckets; h->n_occupied
 = h->size; h->upper_bound = (khint_t)(h->n_buckets *
 __ac_HASH_UPPER + 0.5); } return 0; } static inline __attribute__
 ((__unused__)) khint_t kh_put_ccv_cnnp_parameter_id(kh_ccv_cnnp_parameter_id_t
 *h, kh_cstr_t key, int *ret) { khint_t x; if (h->n_occupied
 >= h->upper_bound) { if (h->n_buckets > (h->size
<<1)) { if (kh_resize_ccv_cnnp_parameter_id(h, h->n_buckets
 - 1) < 0) { *ret = -1; return h->n_buckets; } } else if
 (kh_resize_ccv_cnnp_parameter_id(h, h->n_buckets + 1) <
 0) { *ret = -1; return h->n_buckets; } } { khint_t k, i, site
, last, mask = h->n_buckets - 1, step = 0; x = site = h->
n_buckets; k = __ac_X31_hash_string(key); i = k & mask; if
 (((h->flags[i>>4]>>((i&0xfU)<<1))&
2)) x = i; else { last = i; while (!((h->flags[i>>4]
>>((i&0xfU)<<1))&2) && (((h->flags
[i>>4]>>((i&0xfU)<<1))&1) || !(strcmp
(h->keys[i], key) == 0))) { if (((h->flags[i>>4]>>
((i&0xfU)<<1))&1)) site = i; i = (i + (++step))
 & mask; if (i == last) { x = site; break; } } if (x == h
->n_buckets) { if (((h->flags[i>>4]>>((i&
0xfU)<<1))&2) && site != h->n_buckets) x
 = site; else x = i; } } } if (((h->flags[x>>4]>>
((x&0xfU)<<1))&2)) { h->keys[x] = key; (h->
flags[x>>4]&=~(3ul<<((x&0xfU)<<1)))
; ++h->size; ++h->n_occupied; *ret = 1; } else if (((h->
flags[x>>4]>>((x&0xfU)<<1))&1)) { h
->keys[x] = key; (h->flags[x>>4]&=~(3ul<<
((x&0xfU)<<1))); ++h->size; *ret = 2; } else *ret
 = 0; return x; } static inline __attribute__ ((__unused__)) void
 kh_del_ccv_cnnp_parameter_id(kh_ccv_cnnp_parameter_id_t *h, khint_t
 x) { if (x != h->n_buckets && !((h->flags[x>>
4]>>((x&0xfU)<<1))&3)) { (h->flags[x>>
4]|=1ul<<((x&0xfU)<<1)); --h->size; } }
27
←
Taking true branch→
28
←
Taking false branch→
29
←
Calling 'kh_resize_ccv_cnnp_parameter_id'→
30
←
Taking true branch→
31
←
Assuming the condition is false→
32
←
Taking false branch→
33
←
'?' condition is true→
34
←
Assuming 'new_flags' is non-null→
35
←
Taking false branch→
36
←
'?' condition is true→
37
←
Taking true branch→
38
←
Storing uninitialized value→
39
←
Assuming 'new_keys' is non-null→
40
←
Taking false branch→
41
←
Taking true branch→
42
←
Assuming 'new_vals' is non-null→
43
←
Taking false branch→
44
←
Taking true branch→
45
←
Loop condition is false. Execution continues on line 2379→
46
←
Taking false branch→
47
←
Returning from 'kh_resize_ccv_cnnp_parameter_id'→
48
←
Taking false branch→
49
←
Assuming the condition is true→
50
←
Taking true branch→
51
←
Taking true branch→
57
←
Taking true branch→
58
←
Assuming the condition is true→
59
←
Assuming the condition is true→
60
←
The value 1 is assigned to 'i'→
61
←
Taking false branch→
62
←
Assuming the condition is true→
63
←
Assuming the condition is false→
64
←
1st function call argument is an uninitialized value
2380 
2381void 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)
2382{
2383	ccv_array_t* to_parameter_indices;
2384	int to_param_ref;
2385	ccv_array_t* from_parameter_indices;
2386	int from_param_ref;
2387	_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);
2388	// Should be exactly the same tensor.
2389	if (renamer == 0 && to_param_ref < 0 && from_param_ref < 0)
1
Assuming 'renamer' is not equal to null→
2390		{ 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", 2390, __extension__ __PRETTY_FUNCTION__
); })); }
2391	// To models.
2392	ccv_cnnp_compiled_data_t* const to_compiled_data = model->compiled_data;
2393	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", 2393, __extension__ __PRETTY_FUNCTION__
); }));
2
←
Assuming 'to_compiled_data' is non-null→
3
←
Taking true branch→
2394	// From models.
2395	const ccv_cnnp_compiled_data_t* const from_compiled_data = from_model->compiled_data;
2396	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→
2397	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", 2397, __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→
2398	const int from_parameter_size = from_compiled_data->parameters->rnum;
2399	const int to_parameter_size = to_compiled_data->parameters->rnum;
2400	const int rnum = (to_param_ref < 0 && from_param_ref < 0) ? to_parameter_indices->rnum : 1;
9
←
Assuming 'to_param_ref' is >= 0→
2401	int i, j;
2402	khash_t(ccv_cnnp_parameter_id)kh_ccv_cnnp_parameter_id_t* id_map = 0;
2403	char* updated_name = 0;
2404	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));
2405  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));
2406	for (i = 0; i < rnum; i++)
2407	{
2408		int src_d = (from_param_ref >= 0 ? from_param_ref : i) < from_parameter_indices->rnum ? *(int*)ccv_array_get(from_parameter_indices,from_param_ref >= 0 ? from_param_ref : i)((void*)(((char*)((from_parameter_indices)->data)) + (size_t
)(from_parameter_indices)->rsize * (size_t)(from_param_ref
 >= 0 ? from_param_ref : i))) : from_parameter_size;
10
←
Assuming 'from_param_ref' is < 0→
11
←
'?' condition is false→
12
←
Assuming the condition is false→
13
←
'?' condition is false→
2409		// Need to figure out how to use the renamer here.
2410		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→
2411		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"
, 2411, __extension__ __PRETTY_FUNCTION__); }));
15
←
Assuming 'dest_d' is >= 0→
16
←
Taking true branch→
2412		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", 2412, __extension__
 __PRETTY_FUNCTION__); }));
17
←
Assuming 'dest_d' is < 'to_parameter_size'→
18
←
Taking true branch→
2413		if (renamer18.1
'renamer' is non-null
)
2414		{
2415			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;
2416			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)));
2417			if (!updated_name18.3
'updated_name' is null
)
19
←
Taking true branch→
2418				updated_name = (char*)ccmallocmalloc(1024);
2419			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→
2420			if (src_name_len20.1
'src_name_len' is <= 0
 > 0)
21
←
Taking false branch→
2421				memcpy(updated_name, src_name, src_name_len);
2422			updated_name[src_name_len] = 0;
2423			if (renamer(context, dest_name, updated_name, 1024) != 0)
22
←
Assuming the condition is false→
2424				continue; // Skip this.
2425			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)
2426			{
2427				// Nothing changed.
2428			} else {
2429				if (!id_map22.2
'id_map' is null
)
23
←
Taking true branch→
2430				{
2431					id_map = kh_init(ccv_cnnp_parameter_id)kh_init_ccv_cnnp_parameter_id();
2432					for (j = 0; j < from_parameter_size; j++)
24
←
Assuming 'j' is < 'from_parameter_size'→
25
←
Loop condition is true.  Entering loop body→
54
←
Assuming 'j' is >= 'from_parameter_size'→
55
←
Loop condition is false. Execution continues on line 2440→
2433					{
2434						int ret;
2435						const khiter_t k = kh_put(ccv_cnnp_parameter_id, id_map, *(char**)ccv_array_get(from_compiled_data->ids.parameters, j), &ret)kh_put_ccv_cnnp_parameter_id(id_map, *(char**)((void*)(((char
*)((from_compiled_data->ids.parameters)->data)) + (size_t
)(from_compiled_data->ids.parameters)->rsize * (size_t)
(j))), &ret);
26
←
Calling 'kh_put_ccv_cnnp_parameter_id'→
52
←
Returning from 'kh_put_ccv_cnnp_parameter_id'→
2436						assert(ret != 0)((void) sizeof ((ret != 0) ? 1 : 0), __extension__ ({ if (ret
 != 0) ; else __assert_fail ("ret != 0", "ccv_cnnp_model.c", 2436
, __extension__ __PRETTY_FUNCTION__); }));
53
←
Taking true branch→
2437						kh_val(id_map, k)((id_map)->vals[k]) = j;
2438					}
2439				}
2440				const khiter_t k = kh_get(ccv_cnnp_parameter_id, id_map, updated_name)kh_get_ccv_cnnp_parameter_id(id_map, updated_name);
56
←
Calling 'kh_get_ccv_cnnp_parameter_id'→
2441				if (k == kh_end(id_map)((id_map)->n_buckets)) // Cannot find the name, skip.
2442					continue;
2443				src_d = kh_val(id_map, k)((id_map)->vals[k]);
2444				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"
, 2444, __extension__ __PRETTY_FUNCTION__); }));
2445				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", 2445, __extension__
 __PRETTY_FUNCTION__); }));
2446			}
2447		}
2448		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"
, 2448, __extension__ __PRETTY_FUNCTION__); }));
2449		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", 2449, __extension__
 __PRETTY_FUNCTION__); }));
2450		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;
2451		// If the original is not init'ed. We cannot share from.
2452		if (!(from_init_v[s >> 5] & (1u << (s & 0x1f))))
2453			continue;
2454		for (j = 0; j < parallel_count; j++)
2455		{
2456			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));
2457			assert(src)((void) sizeof ((src) ? 1 : 0), __extension__ ({ if (src) ; else
 __assert_fail ("src", "ccv_cnnp_model.c", 2457, __extension__
 __PRETTY_FUNCTION__); }));
2458			ccv_nnc_tensor_t* const dest = to_compiled_data->tensors.parameters[dest_d + j * to_parameter_size];
2459			if (dest && !((uintptr_t)dest & (uintptr_t)1))
2460				ccv_nnc_tensor_free(dest);
2461			to_compiled_data->tensors.parameters[dest_d + j * to_parameter_size] = (ccv_nnc_tensor_t*)((uintptr_t)src | (uintptr_t)1);
2462		}
2463		// Mark this symbol as init'ed.
2464		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;
2465		to_init_v[d >> 5] |= (1u << (d & 0x1f));
2466	}
2467	ccv_array_free(to_parameter_indices);
2468	ccv_array_free(from_parameter_indices);
2469	if (id_map)
2470		kh_destroy(ccv_cnnp_parameter_id, id_map)kh_destroy_ccv_cnnp_parameter_id(id_map);
2471	if (updated_name)
2472		ccfreefree(updated_name);
2473	// Mark it as incomplete so we will call init_1.
2474	if (ccv_cnnp_model_tensors_any_to_alloc(model, to_compiled_data))
2475		to_compiled_data->tensors_init.v = (uint32_t*)((uintptr_t)to_compiled_data->tensors_init.v | (uintptr_t)1);
2476	else // Remove the flag.
2477		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));
2478}
2479 
2480ccv_nnc_stream_context_t* ccv_cnnp_compiled_data_get_stream(ccv_cnnp_compiled_data_t* const compiled_data, const int type)
2481{
2482	if (!compiled_data->stream_map)
2483		compiled_data->stream_map = kh_init(stream_map)kh_init_stream_map();
2484	int ret = 0;
2485	khiter_t k = kh_put(stream_map, compiled_data->stream_map, type, &ret)kh_put_stream_map(compiled_data->stream_map, type, &ret
);
2486	assert(ret >= 0)((void) sizeof ((ret >= 0) ? 1 : 0), __extension__ ({ if (
ret >= 0) ; else __assert_fail ("ret >= 0", "ccv_cnnp_model.c"
, 2486, __extension__ __PRETTY_FUNCTION__); }));
2487	ccv_nnc_stream_context_t* stream = kh_val(compiled_data->stream_map, k)((compiled_data->stream_map)->vals[k]);
2488	// If ret == 0, the key already exist, we can return directly, otherwise, create and return.
2489	if (ret != 0)
2490	{
2491		stream = ccv_nnc_stream_context_new(type);
2492		kh_val(compiled_data->stream_map, k)((compiled_data->stream_map)->vals[k]) = stream;
2493	}
2494	return stream;
2495}
2496 
2497void 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)
2498{
2499	ccv_array_t* to_parameter_indices;
2500	int to_param_ref;
2501	ccv_array_t* from_parameter_indices;
2502	int from_param_ref;
2503	_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);
2504	// Should be exactly the same tensor.
2505	if (to_param_ref < 0 && from_param_ref < 0)
2506		{ 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", 2506, __extension__ __PRETTY_FUNCTION__
); })); }
2507	// To models.
2508	ccv_cnnp_compiled_data_t* const to_compiled_data = model->compiled_data;
2509	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", 2509, __extension__ __PRETTY_FUNCTION__
); }));
2510	// From models.
2511	const ccv_cnnp_compiled_data_t* const from_compiled_data = from_model->compiled_data;
2512	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; });
2513	const int to_parameter_size = to_compiled_data->parameters->rnum;
2514	const int rnum = (to_param_ref < 0 && from_param_ref < 0) ? from_parameter_indices->rnum : 1;
2515	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", 2515, __extension__ __PRETTY_FUNCTION__
); }));
2516	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", 2516, __extension__ __PRETTY_FUNCTION__
); }));
2517	int i, j;
2518	ccv_nnc_tensor_t* inputs[aux_in_size + 2];
2519	ccv_nnc_tensor_t* outputs[aux_out_size + 1];
2520	for (i = 0; i < aux_in_size; i++)
2521		inputs[i + 2] = aux_ins[i];
2522	for (i = 0; i < aux_out_size; i++)
2523		outputs[i + 1] = aux_outs[i];
2524	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));
2525  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));
2526	for (i = 0; i < rnum; i++)
2527	{
2528		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)));
2529		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"
, 2529, __extension__ __PRETTY_FUNCTION__); }));
2530		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", 2530, __extension__ __PRETTY_FUNCTION__
); }));
2531		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;
2532		// If the original is not init'ed. We cannot copy from.
2533		if (!(from_init_v[s >> 5] & (1u << (s & 0x1f))))
2534			continue;
2535		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)));
2536		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"
, 2536, __extension__ __PRETTY_FUNCTION__); }));
2537		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", 2537, __extension__ __PRETTY_FUNCTION__
); }));
2538		if (parallel_count > 1)
2539		{
2540			ccv_nnc_stream_context_t* streams[parallel_count];
2541			ccv_nnc_stream_signal_t* signal;
2542			if (stream_context)
2543				signal = ccv_nnc_stream_context_emit_signal_new(stream_context);
2544			for (j = 0; j < parallel_count; j++)
2545			{
2546				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));
2547				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));
2548				if (!dest || !src)
2549				{
2550					streams[j] = 0;
2551					continue;
2552				}
2553				// At the moment, can only handle them on the same device.
2554				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", 2554, __extension__ __PRETTY_FUNCTION__
); }));
2555				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", 2555, __extension__ __PRETTY_FUNCTION__
); }));
2556				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;
2557				const int device_id = CCV_TENSOR_GET_DEVICE_ID(src->info.type)(((src->info.type) & 0xfff00) >> 8);
2558				int type = stream_type;
2559				CCV_STREAM_SET_DEVICE_ID(type, device_id)(type) = (((type) & ~0xfff00) | (((device_id) & 0xfff
) << 8));
2560				ccv_nnc_stream_context_t* const stream_0 = ccv_cnnp_compiled_data_get_stream(to_compiled_data, type);
2561				// Wait signal to finish.
2562				if (stream_context)
2563					ccv_nnc_stream_context_wait_signal(stream_0, signal);
2564				inputs[0] = outputs[0] = dest;
2565				inputs[1] = src;
2566				ccv_nnc_cmd_exec(cmd, hint, flags, inputs, aux_in_size + 2, outputs, aux_out_size + 1, stream_0);
2567				if (stream_context)
2568				{
2569					ccv_nnc_stream_signal_t* const signal = ccv_nnc_stream_context_emit_signal_new(stream_0);
2570					ccv_nnc_stream_context_wait_signal(stream_context, signal);
2571				}
2572				streams[j] = stream_0;
2573			}
2574			// If this should be blocking, blocking it.
2575			if (!stream_context)
2576				for (j = 0; j < parallel_count; j++)
2577					if (streams[j])
2578						ccv_nnc_stream_context_wait(streams[j]);
2579		} else {
2580			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));
2581			assert(src)((void) sizeof ((src) ? 1 : 0), __extension__ ({ if (src) ; else
 __assert_fail ("src", "ccv_cnnp_model.c", 2581, __extension__
 __PRETTY_FUNCTION__); }));
2582			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));
2583			assert(dest)((void) sizeof ((dest) ? 1 : 0), __extension__ ({ if (dest) ;
 else __assert_fail ("dest", "ccv_cnnp_model.c", 2583, __extension__
 __PRETTY_FUNCTION__); }));
2584			inputs[0] = outputs[0] = dest;
2585			inputs[1] = src;
2586			ccv_nnc_cmd_exec(cmd, hint, flags, inputs, aux_in_size + 2, outputs, aux_out_size + 1, stream_context);
2587		}
2588		// Mark this symbol as init'ed.
2589		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;
2590		to_init_v[d >> 5] |= (1u << (d & 0x1f));
2591	}
2592	ccv_array_free(to_parameter_indices);
2593	ccv_array_free(from_parameter_indices);
2594}
2595 
2596void 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)
2597{
2598	int to_param_ref;
2599	ccv_array_t* const to_parameter_indices = _ccv_cnnp_model_parameter_indices(model, parameters, &to_param_ref);
2600	// To models.
2601	ccv_cnnp_compiled_data_t* const to_compiled_data = model->compiled_data;
2602	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", 2602, __extension__ __PRETTY_FUNCTION__
); }));
2603	// Tensor has to be inited already.
2604	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", 2604, __extension__ __PRETTY_FUNCTION__
); }));
2605	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", 2605, __extension__ __PRETTY_FUNCTION__
); }));
2606	// From models.
2607	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; });
2608	const int to_parameter_size = to_compiled_data->parameters->rnum;
2609	const int rnum = (to_param_ref < 0) ? to_parameter_indices->rnum : 1;
2610	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", 2610, __extension__ __PRETTY_FUNCTION__
); }));
2611	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", 2611, __extension__ __PRETTY_FUNCTION__
); }));
2612	int i, j;
2613	ccv_nnc_tensor_t* inputs[aux_in_size + 1];
2614	ccv_nnc_tensor_t* outputs[aux_out_size + 1];
2615	for (i = 0; i < aux_in_size; i++)
2616		inputs[i + 1] = aux_ins[i];
2617	for (i = 0; i < aux_out_size; i++)
2618		outputs[i + 1] = aux_outs[i];
2619	for (i = 0; i < rnum; i++)
2620	{
2621		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)));
2622		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"
, 2622, __extension__ __PRETTY_FUNCTION__); }));
2623		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", 2623, __extension__ __PRETTY_FUNCTION__
); }));
2624		if (parallel_count > 1)
2625		{
2626			ccv_nnc_stream_context_t* streams[parallel_count];
2627			ccv_nnc_stream_signal_t* signal;
2628			if (stream_context)
2629				signal = ccv_nnc_stream_context_emit_signal_new(stream_context);
2630			for (j = 0; j < parallel_count; j++)
2631			{
2632				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));
2633				if (!dest)
2634				{
2635					streams[j] = 0;
2636					continue;
2637				}
2638				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;
2639				const int device_id = CCV_TENSOR_GET_DEVICE_ID(dest->info.type)(((dest->info.type) & 0xfff00) >> 8);
2640				int type = stream_type;
2641				CCV_STREAM_SET_DEVICE_ID(type, device_id)(type) = (((type) & ~0xfff00) | (((device_id) & 0xfff
) << 8));
2642				ccv_nnc_stream_context_t* const stream_0 = ccv_cnnp_compiled_data_get_stream(to_compiled_data, type);
2643				// Wait signal to finish.
2644				if (stream_context)
2645					ccv_nnc_stream_context_wait_signal(stream_0, signal);
2646				inputs[0] = outputs[0] = dest;
2647				ccv_nnc_cmd_exec(cmd, hint, flags, inputs, aux_in_size + 1, outputs, aux_out_size + 1, stream_0);
2648				if (stream_context)
2649				{
2650					ccv_nnc_stream_signal_t* const signal = ccv_nnc_stream_context_emit_signal_new(stream_0);
2651					ccv_nnc_stream_context_wait_signal(stream_context, signal);
2652				}
2653				streams[j] = stream_0;
2654			}
2655			// If this should be blocking, blocking it.
2656			if (!stream_context)
2657				for (j = 0; j < parallel_count; j++)
2658					if (streams[j])
2659						ccv_nnc_stream_context_wait(streams[j]);
2660		} else {
2661			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));
2662			assert(dest)((void) sizeof ((dest) ? 1 : 0), __extension__ ({ if (dest) ;
 else __assert_fail ("dest", "ccv_cnnp_model.c", 2662, __extension__
 __PRETTY_FUNCTION__); }));
2663			inputs[0] = outputs[0] = dest;
2664			ccv_nnc_cmd_exec(cmd, hint, flags, inputs, aux_in_size + 1, outputs, aux_out_size + 1, stream_context);
2665		}
2666		// No need to mark this symbol as init'ed, it is already.
2667	}
2668	ccv_array_free(to_parameter_indices);
2669}
2670 
2671void 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)
2672{
2673	int to_param_ref;
2674	ccv_array_t* const to_parameter_indices = _ccv_cnnp_model_parameter_indices(model, parameters, &to_param_ref);
2675	// To models.
2676	ccv_cnnp_compiled_data_t* const to_compiled_data = model->compiled_data;
2677	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", 2677, __extension__ __PRETTY_FUNCTION__
); }));
2678	// Tensor has to be inited already.
2679	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", 2679, __extension__ __PRETTY_FUNCTION__
); }));
2680	ccv_nnc_tensor_t** tensor_gradients;
2681	if (to_compiled_data->backward.count > 1)
2682		tensor_gradients = to_compiled_data->tensors.accum_gradients;
2683	else
2684		tensor_gradients = to_compiled_data->tensors.gradients;
2685	assert(tensor_gradients)((void) sizeof ((tensor_gradients) ? 1 : 0), __extension__ ({
 if (tensor_gradients) ; else __assert_fail ("tensor_gradients"
, "ccv_cnnp_model.c", 2685, __extension__ __PRETTY_FUNCTION__
); }));
2686	// From models.
2687	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; });
2688	const int to_parameter_size = to_compiled_data->parameters->rnum;
2689	const int rnum = (to_param_ref < 0) ? to_parameter_indices->rnum : 1;
2690	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", 2690, __extension__ __PRETTY_FUNCTION__
); }));
2691	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", 2691, __extension__ __PRETTY_FUNCTION__
); }));
2692	int i, j;
2693	ccv_nnc_tensor_t* inputs[aux_in_size + 1];
2694	ccv_nnc_tensor_t* outputs[aux_out_size + 1];
2695	for (i = 0; i < aux_in_size; i++)
2696		inputs[i + 1] = aux_ins[i];
2697	for (i = 0; i < aux_out_size; i++)
2698		outputs[i + 1] = aux_outs[i];
2699	for (i = 0; i < rnum; i++)
2700	{
2701		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)));
2702		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"
, 2702, __extension__ __PRETTY_FUNCTION__); }));
2703		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", 2703, __extension__ __PRETTY_FUNCTION__
); }));
2704		if (parallel_count > 1)
2705		{
2706			ccv_nnc_stream_context_t* streams[parallel_count];
2707			ccv_nnc_stream_signal_t* signal;
2708			if (stream_context)
2709				signal = ccv_nnc_stream_context_emit_signal_new(stream_context);
2710			for (j = 0; j < parallel_count; j++)
2711			{
2712				ccv_nnc_tensor_t* const dest = tensor_gradients[dest_d + j * to_parameter_size];
2713				if (!dest)
2714				{
2715					streams[j] = 0;
2716					continue;
2717				}
2718				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;
2719				const int device_id = CCV_TENSOR_GET_DEVICE_ID(dest->info.type)(((dest->info.type) & 0xfff00) >> 8);
2720				int type = stream_type;
2721				CCV_STREAM_SET_DEVICE_ID(type, device_id)(type) = (((type) & ~0xfff00) | (((device_id) & 0xfff
) << 8));
2722				ccv_nnc_stream_context_t* const stream_0 = ccv_cnnp_compiled_data_get_stream(to_compiled_data, type);
2723				// Wait signal to finish.
2724				if (stream_context)
2725					ccv_nnc_stream_context_wait_signal(stream_0, signal);
2726				inputs[0] = outputs[0] = dest;
2727				ccv_nnc_cmd_exec(cmd, hint, flags, inputs, aux_in_size + 1, outputs, aux_out_size + 1, stream_0);
2728				if (stream_context)
2729				{
2730					ccv_nnc_stream_signal_t* const signal = ccv_nnc_stream_context_emit_signal_new(stream_0);
2731					ccv_nnc_stream_context_wait_signal(stream_context, signal);
2732				}
2733				streams[j] = stream_0;
2734			}
2735			// If this should be blocking, blocking it.
2736			if (!stream_context)
2737				for (j = 0; j < parallel_count; j++)
2738					if (streams[j])
2739						ccv_nnc_stream_context_wait(streams[j]);
2740		} else {
2741			ccv_nnc_tensor_t* const dest = tensor_gradients[dest_d];
2742			if (!dest)
2743				continue;
2744			assert(dest)((void) sizeof ((dest) ? 1 : 0), __extension__ ({ if (dest) ;
 else __assert_fail ("dest", "ccv_cnnp_model.c", 2744, __extension__
 __PRETTY_FUNCTION__); }));
2745			inputs[0] = outputs[0] = dest;
2746			ccv_nnc_cmd_exec(cmd, hint, flags, inputs, aux_in_size + 1, outputs, aux_out_size + 1, stream_context);
2747		}
2748		// No need to mark this symbol as init'ed, it is already.
2749	}
2750	ccv_array_free(to_parameter_indices);
2751}
2752 
2753ccv_nnc_cmd_t ccv_cnnp_model_minimizer(ccv_cnnp_model_t* const model)
2754{
2755	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2756	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 2756, __extension__ __PRETTY_FUNCTION__); }));
2757	return compiled_data->minimize.minimizer;
2758}
2759 
2760void 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)
2761{
2762	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2763	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 2763, __extension__ __PRETTY_FUNCTION__); }));
2764	const int parameter_size = compiled_data->parameters->rnum;
2765	if (parameter_size == 0)
2766		return;
2767	if (reset)
2768		{ 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", 2768, __extension__ __PRETTY_FUNCTION__
); })); }
2769	const int old_max_saved_aux_size = compiled_data->minimize.max_saved_aux_size;
2770	const int saved_aux_size = ccv_nnc_minimizer_saved_aux_size(minimizer);
2771	if (saved_aux_size > compiled_data->minimize.max_saved_aux_size)
2772		compiled_data->minimize.max_saved_aux_size = saved_aux_size;
2773	const int max_saved_aux_size = compiled_data->minimize.max_saved_aux_size;
2774	// We update all parameters, at this point, we have one minimizer.
2775	if (set_parameters == 0 || set_parameter_size == 0)
2776		compiled_data->minimize.minimizer = minimizer;
2777	int i;
2778	if (set_parameters && set_parameter_size)
2779	{
2780		// I need to save what's the minimizer along with this.
2781		if (!compiled_data->minimize.parameters)
2782			compiled_data->minimize.parameters = ccv_array_new(sizeof(ccv_cnnp_set_minimizer_for_parameter_t*), 1, 0);
2783		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));
2784		set_minimizer_for_parameter->minimizer = minimizer;
2785		set_minimizer_for_parameter->parameter_size = set_parameter_size;
2786		memcpy(set_minimizer_for_parameter->parameters, set_parameters, sizeof(ccv_cnnp_model_io_t) * set_parameter_size);
2787		ccv_array_push(compiled_data->minimize.parameters, &set_minimizer_for_parameter);
2788	}
2789	// If reset is true, clear the parameters array.
2790	if (reset && compiled_data->minimize.parameters)
2791	{
2792		for (i = 0; i < compiled_data->minimize.parameters->rnum; i++)
2793			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))));
2794		ccv_array_clear(compiled_data->minimize.parameters);
2795	}
2796	if (!compiled_data->update_nodes)
2797		return;
2798	ccv_nnc_symbolic_graph_t* const symbolic_graph = model->graph;
2799	assert(symbolic_graph)((void) sizeof ((symbolic_graph) ? 1 : 0), __extension__ ({ if
 (symbolic_graph) ; else __assert_fail ("symbolic_graph", "ccv_cnnp_model.c"
, 2799, __extension__ __PRETTY_FUNCTION__); }));
2800	if (saved_aux_size > old_max_saved_aux_size)
2801	{
2802		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", 2802, __extension__ __PRETTY_FUNCTION__
); }));
2803		// Reallocate first, move them around later.
2804		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);
2805		compiled_data->update_nodes = (ccv_nnc_graph_exec_symbol_t*)(compiled_data->updated_parameters + parameter_size);
2806		compiled_data->saved_aux = (ccv_nnc_tensor_symbol_map_t*)(compiled_data->update_nodes + parameter_size);
2807		// We need to do this from back to front because saved_aux_size > old_saved_aux_size, it could overlap.
2808		_ccv_cnnp_scatter_saved_aux(compiled_data->saved_aux, parameter_size, old_max_saved_aux_size, saved_aux_size);
2809	}
2810	int flag = 0;
2811	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; });
2812	if (set_parameters && set_parameter_size)
2813	{
2814		ccv_array_t* const parameter_indices = ccv_array_new(sizeof(int), 0, 0);
2815		for (i = 0; i < set_parameter_size; i++)
2816		{
2817			const int param_sel = set_parameters[i]->param_sel > 0 ? set_parameters[i]->param_sel - 1 : set_parameters[i]->param_sel;
2818			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", 2818, __extension__ __PRETTY_FUNCTION__
); }));
2819			const int old_rnum = parameter_indices->rnum;
2820			ccv_cnnp_model_add_to_parameter_indices(set_parameters[i]->model, param_sel, parameter_indices);
2821			const int param_ref = set_parameters[i]->param_ref > 0 ? set_parameters[i]->param_ref - 1 : set_parameters[i]->param_ref;
2822			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", 2822, __extension__ __PRETTY_FUNCTION__
); }));
2823			if (param_ref >= 0)
2824			{
2825				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", 2825, __extension__ __PRETTY_FUNCTION__
); }));
2826				*(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)));
2827				parameter_indices->rnum = old_rnum + 1;
2828			}
2829		}
2830		// We may have duplicated indices, but that is OK, we will set it twice.
2831		for (i = 0; i < parameter_indices->rnum; i++)
2832		{
2833			const int d = *(int*)ccv_array_get(parameter_indices, i)((void*)(((char*)((parameter_indices)->data)) + (size_t)(parameter_indices
)->rsize * (size_t)(i)));
2834			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))
2835				flag = 1;
2836		}
2837		ccv_array_free(parameter_indices);
2838	} else {
2839		for (i = 0; i < parameter_size; i++)
2840			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))
2841				flag = 1;
2842		if (compiled_data->minimize.parameters)
2843			if (_ccv_cnnp_apply_parameters_with_minimizer(model))
2844				flag = 1;
2845	}
2846	if (flag)
2847	{
2848		// 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.
2849		if (compiled_data->graph_mode == CCV_CNNP_MODEL_GRAPH_FIT_MODE)
2850			_ccv_cnnp_compiled_data_graph_free(compiled_data);
2851		_ccv_cnnp_compiled_data_apply_gradients_free(compiled_data);
2852	}
2853}
2854 
2855void ccv_cnnp_model_set_compile_params(ccv_cnnp_model_t* const model, const ccv_nnc_symbolic_graph_compile_param_t compile_params)
2856{
2857	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2858	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 2858, __extension__ __PRETTY_FUNCTION__); }));
2859	compiled_data->compile_params = compile_params;
2860}
2861 
2862void ccv_cnnp_model_dot(const ccv_cnnp_model_t* const model, const int flags, FILE** const outs, const int out_size)
2863{
2864	if (model->graph && out_size > 0)
2865		ccv_nnc_symbolic_graph_dot(model->graph, flags, outs[0]);
2866	if (model->compiled_data && model->compiled_data->graph && out_size > 1)
2867		ccv_nnc_graph_dot(model->compiled_data->graph, flags, outs[1]);
2868	if (model->compiled_data && model->compiled_data->backward.accum && out_size > 2)
2869		ccv_nnc_graph_dot(model->compiled_data->backward.accum, flags, outs[2]);
2870	if (model->compiled_data && model->compiled_data->apply_gradients.graph && out_size > 3)
2871		ccv_nnc_graph_dot(model->compiled_data->apply_gradients.graph, flags, outs[3]);
2872}
2873 
2874void ccv_cnnp_model_format(const ccv_cnnp_model_t* const model, const ccv_nnc_symbolic_graph_format_f format_fn, void* const context)
2875{
2876	if (model->graph)
2877		ccv_nnc_symbolic_graph_format(model->graph, 0, 0, 0, 0, format_fn, context);
2878}
2879 
2880static void _ccv_cnnp_compiled_data_free(const ccv_cnnp_model_t* const model, ccv_cnnp_compiled_data_t* const compiled_data)
2881{
2882	int i;
2883	const int parameter_size = compiled_data->parameters->rnum;
2884	ccv_array_free(compiled_data->parameters);
2885	if (compiled_data->parameter_flags)
2886		ccfreefree(compiled_data->parameter_flags);
2887	const int internal_size = compiled_data->internals->rnum;
2888	ccv_array_free(compiled_data->internals);
2889	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", 2889, __extension__ __PRETTY_FUNCTION__
); }));
2890	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", 2890, __extension__ __PRETTY_FUNCTION__
); }));
2891	for (i = 0; i < parameter_size; i++)
2892		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))));
2893	ccv_array_free(compiled_data->ids.parameters);
2894	for (i = 0; i < internal_size; i++)
2895		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))));
2896	ccv_array_free(compiled_data->ids.internals);
2897	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; });
2898	if (compiled_data->tensors.parameters)
2899	{
2900		for (i = 0; i < parameter_size * parallel_count; i++)
2901			// If it is not marked as not belonging, we can free it.
2902			if (!((uintptr_t)compiled_data->tensors.parameters[i] & (uintptr_t)1))
2903				if (compiled_data->tensors.parameters[i])
2904					ccv_nnc_tensor_free(compiled_data->tensors.parameters[i]);
2905		for (i = 0; i < internal_size * parallel_count; i++)
2906			if (compiled_data->tensors.internals[i])
2907				ccv_nnc_tensor_free(compiled_data->tensors.internals[i]);
2908		ccfreefree(compiled_data->tensors.parameters);
2909	}
2910	if (compiled_data->tensors.gradients)
2911	{
2912		for (i = 0; i < parameter_size * parallel_count; i++)
2913		{
2914			if (compiled_data->tensors.gradients[i])
2915				ccv_nnc_tensor_free(compiled_data->tensors.gradients[i]);
2916			if (compiled_data->tensors.accum_gradients[i])
2917				ccv_nnc_tensor_free(compiled_data->tensors.accum_gradients[i]);
2918		}
2919		ccfreefree(compiled_data->tensors.gradients);
2920	}
2921	if (compiled_data->minimize.parameters)
2922	{
2923		for (i = 0; i < compiled_data->minimize.parameters->rnum; i++)
2924			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))));
2925		ccv_array_free(compiled_data->minimize.parameters);
2926	}
2927	if (compiled_data->rewindables)
2928		ccv_array_free(compiled_data->rewindables);
2929	if (compiled_data->tensors_init.v)
2930		ccfreefree(CCV_NNC_INIT_V(compiled_data->tensors_init.v)((uint32_t*)((uintptr_t)(compiled_data->tensors_init.v) &
 ~(uintptr_t)1)));
2931	if (compiled_data->evaluate.tos)
2932		ccfreefree(compiled_data->evaluate.tos);
2933	compiled_data->evaluate.tos = 0;
2934	if (compiled_data->stream_map)
2935	{
2936		khiter_t k;
2937		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)
2938		{
2939			if (!kh_exist(compiled_data->stream_map, k)(!(((compiled_data->stream_map)->flags[(k)>>4]>>
(((k)&0xfU)<<1))&3)))
2940				continue;
2941			ccv_nnc_stream_context_t* const stream = kh_val(compiled_data->stream_map, k)((compiled_data->stream_map)->vals[k]);
2942			ccv_nnc_stream_context_free(stream);
2943		}
2944		kh_destroy(stream_map, compiled_data->stream_map)kh_destroy_stream_map(compiled_data->stream_map);
2945	}
2946	_ccv_cnnp_compiled_data_graph_free(compiled_data);
2947	_ccv_cnnp_compiled_data_gradient_free(compiled_data);
2948	_ccv_cnnp_compiled_data_backward_free(compiled_data);
2949	_ccv_cnnp_compiled_data_apply_gradients_free(compiled_data);
2950	if (compiled_data->gradient_checkpoints)
2951	{
2952		for (i = 0; i < compiled_data->gradient_checkpoints->rnum; i++)
2953		{
2954			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)));
2955			assert(checkpoint->inputs)((void) sizeof ((checkpoint->inputs) ? 1 : 0), __extension__
 ({ if (checkpoint->inputs) ; else __assert_fail ("checkpoint->inputs"
, "ccv_cnnp_model.c", 2955, __extension__ __PRETTY_FUNCTION__
); }));
2956			ccfreefree(checkpoint->inputs);
2957			ccv_array_free(checkpoint->tensor_symbols);
2958		}
2959		ccv_array_free(compiled_data->gradient_checkpoints);
2960	}
2961	ccv_nnc_xpu_alloc_destroy(&compiled_data->xpu_alloc);
2962	ccfreefree(compiled_data);
2963}
2964 
2965void ccv_cnnp_model_free(ccv_cnnp_model_t* const model)
2966{
2967	if (model->isa->deinit)
2968		model->isa->deinit(model);
2969	if (model->io)
2970	{
2971		int i;
2972		for (i = 0; i < model->io->rnum; i++)
2973		{
2974			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)));
2975			if (model_io->outgoings)
2976				ccv_array_free(model_io->outgoings);
2977			if (model_io->incomings)
2978				ccv_array_free(model_io->incomings);
2979			if (model_io->dependencies)
2980				ccv_array_free(model_io->dependencies);
2981			ccfreefree(model_io);
2982		}
2983		ccv_array_free(model->io);
2984	}
2985	if (model->parameter_indices)
2986		ccv_array_free(model->parameter_indices);
2987	if (model->inputs)
2988		ccfreefree(model->inputs);
2989	if (model->graph)
2990		ccv_nnc_symbolic_graph_free(model->graph);
2991	if (model->compiled_data)
2992		_ccv_cnnp_compiled_data_free(model, model->compiled_data);
2993	if (model->name)
2994		ccfreefree(model->name);
2995	ccfreefree(model);
2996}