ccv_cnnp_model.c

Bug Summary

File:	nnc/ccv_cnnp_model.c
Warning:	line 2474, 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/19 -I ../ -I /usr/local/cuda/include -D HAVE_CBLAS -D HAVE_LIBPNG -D HAVE_LIBJPEG -D HAVE_FFTW3 -D HAVE_PTHREAD -D HAVE_LIBLINEAR -D HAVE_TESSERACT -D HAVE_AVCODEC -D HAVE_AVFORMAT -D HAVE_AVUTIL -D HAVE_SWSCALE -D HAVE_SSE2 -D HAVE_GSL -D HAVE_CUDA -D HAVE_CUDNN -D HAVE_NCCL -D USE_SYSTEM_CUB -I /usr/local/include -internal-isystem /usr/local/lib/clang/19/include -internal-isystem /usr/local/include -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/12/../../../../x86_64-linux-gnu/include -internal-externc-isystem /usr/include/x86_64-linux-gnu -internal-externc-isystem /include -internal-externc-isystem /usr/include -O3 -ferror-limit 19 -fgnuc-version=4.2.1 -fskip-odr-check-in-gmf -vectorize-loops -vectorize-slp -analyzer-output=html -faddrsig -D__GCC_HAVE_DWARF2_CFI_ASM=1 -o /home/liu/actions-runner/_work/ccv/ccv/_analyze/2026-05-01-233500-1493103-1 -x c ccv_cnnp_model.c
1#include "ccv_nnc.h"
2#include "ccv_nnc_easy.h"
3#include "ccv_nnc_internal.h"
4#include "ccv_internal.h"
5#include "_ccv_cnnp_model.h"
6#include "_ccv_nnc_graph.h"
7#include "_ccv_nnc_symbolic_graph.h"
8#ifdef HAVE_CUDA1
9#include "gpu/ccv_nnc_compat.h"
10#endif
11 
12// MARK - Level-5 API
13 
14ccv_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)
15{
16	if (!model->io)
17		model->io = ccv_array_new(sizeof(ccv_cnnp_model_io_t), 1, 0);
18	ccv_cnnp_model_io_t model_io = ccmallocmalloc(sizeof(struct ccv_cnnp_model_io_s) + sizeof(ccv_nnc_tensor_symbol_t) * model->output_size);
19	model_io->param_ref = 0;
20	model_io->param_sel = 0;
21	model_io->visit = 0;
22	model_io->model = model;
23	model_io->dependencies = 0;
24	model_io->dependents = 0;
25	model_io->outgoings = 0;
26	model_io->outputs = (ccv_nnc_tensor_symbol_t*)(model_io + 1);
27	ccv_array_push(model->io, &model_io);
28	if (input_size > 0)
29	{
30		model_io->incomings = ccv_array_new(sizeof(ccv_cnnp_model_io_t), input_size, 0);
31		ccv_array_resize(model_io->incomings, input_size);
32		int i;
33		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);
34		for (i = 0; i < input_size; i++)
35		{
36			if (!inputs[i]->outgoings)
37				inputs[i]->outgoings = ccv_array_new(sizeof(ccv_cnnp_model_io_t), 1, 0);
38			ccv_array_push(inputs[i]->outgoings, &model_io);
39		}
40	} else {
41		model_io->incomings = 0;
42	}
43	return model_io;
44}
45 
46void ccv_cnnp_model_add_dependencies(ccv_cnnp_model_io_t model_io, const ccv_cnnp_model_io_t* const dependencies, const int dependency_size)
47{
48	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", 48, __extension__ __PRETTY_FUNCTION__);
 }));
49	if (!model_io->dependencies)
50		model_io->dependencies = ccv_array_new(sizeof(ccv_cnnp_model_io_t), dependency_size, 0);
51	int i, j;
52	for (i = 0; i < dependency_size; i++)
53	{
54		int flag = 0;
55		// Check if it is already exist or not.
56		for (j = 0; !flag && j < model_io->dependencies->rnum; j++)
57			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])
58				flag = 1;
59		if (flag)
60			continue;
61		ccv_array_push(model_io->dependencies, dependencies + i);
62		++dependencies[i]->dependents;
63	}
64}
65 
66int ccv_cnnp_model_output_size(const ccv_cnnp_model_t* const model)
67{
68	return model->output_size;
69}
70 
71int ccv_cnnp_model_is_trainable(const ccv_cnnp_model_t* const model)
72{
73	// If the model is compiled, it is default to 1 unless it is not.
74	if (model->compiled_data)
75		return model->is_trainable >= 0 ? model->is_trainable : 1;
76	return model->is_trainable;
77}
78 
79ccv_cnnp_model_io_t ccv_cnnp_model_parameters(ccv_cnnp_model_t* const model, const int selector, const int index)
80{
81	if (!model->io)
82		model->io = ccv_array_new(sizeof(ccv_cnnp_model_io_t), 1, 0);
83	ccv_cnnp_model_io_t model_io = ccmallocmalloc(sizeof(struct ccv_cnnp_model_io_s));
84	model_io->param_ref = index >= 0 ? index + 1 : ALL_PARAMETERS-1;
85	model_io->param_sel = selector >= 0 ? selector + 1 : ALL_PARAMETERS-1;
86	model_io->visit = 0;
87	model_io->model = model;
88	model_io->outputs = 0;
89	model_io->dependencies = 0;
90	model_io->dependents = 0;
91	model_io->incomings = 0;
92	model_io->outgoings = 0;
93	ccv_array_push(model->io, &model_io);
94	return model_io;
95}
96 
97void ccv_cnnp_model_notify_hook(ccv_cnnp_model_t* const model, ccv_cnnp_model_notify_f func, void* const context)
98{
99	model->notify_hook.func = func;
100	model->notify_hook.context = context;
101}
102 
103void ccv_cnnp_model_notify(const ccv_cnnp_model_t* const model, const int tag, void* const payload)
104{
105	if (model->notify_hook.func)
106		model->notify_hook.func(model, tag, payload, model->notify_hook.context);
107	if (model->isa->notify)
108		model->isa->notify(model, tag, payload);
109}
110 
111static int _ccv_nnc_array_dedup_graph_exec_symbols(ccv_nnc_graph_exec_symbol_t* const graph_exec_symbols, int graph_exec_symbol_size)
112{
113	int i, j;
114	for (i = 0; i < graph_exec_symbol_size; i++)
115	{
116		ccv_nnc_graph_exec_symbol_t* const graph_exec_symbol = graph_exec_symbols + i;
117		// Check whether this tensor symbol has any duplicate.
118		for (j = i + 1; j < graph_exec_symbol_size;)
119		{
120			ccv_nnc_graph_exec_symbol_t* const other_symbol = graph_exec_symbols + j;
121			// If there is a same tensor symbol, remove it.
122			if (other_symbol->d == graph_exec_symbol->d && other_symbol->graph == graph_exec_symbol->graph)
123			{
124				if (j + 1 < graph_exec_symbol_size)
125					*other_symbol = graph_exec_symbols[graph_exec_symbol_size - 1];
126				--graph_exec_symbol_size;
127				continue;
128			}
129			++j;
130		}
131	}
132	return graph_exec_symbol_size;
133}
134 
135void ccv_cnnp_model_add_to_array(void* const context, const ccv_nnc_tensor_symbol_t symbol, const int is_trainable)
136{
137	ccv_cnnp_model_add_to_array_context_t* const add_to_array_context = (ccv_cnnp_model_add_to_array_context_t*)context;
138	ccv_cnnp_model_t* const model = add_to_array_context->sequence->model;
139	int i;
140	if (add_to_array_context->add_parameter_indices && !model->parameter_indices)
141		model->parameter_indices = ccv_array_new(sizeof(int), 0, 0);
142	for (i = 0; i < add_to_array_context->symbols->rnum; i++)
143	{
144		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)));
145		if (other_symbol.d == symbol.d && other_symbol.graph == symbol.graph)
146		{
147			// Only add to parameter_indices if it is trainable.
148			if (add_to_array_context->add_parameter_indices)
149				ccv_array_add_unique_int(model->parameter_indices, i);
150			// Found it, return, don't add it.
151			return;
152		}
153	}
154	// Only add to parameter_indices if it is trainable.
155	if (add_to_array_context->add_parameter_indices)
156		ccv_array_push(model->parameter_indices, &add_to_array_context->symbols->rnum);
157	// This is a new one, no need to add_unique_int, it is unique.
158	ccv_array_push(add_to_array_context->symbols, &symbol);
159	if (add_to_array_context->trainables)
160		ccv_array_push(add_to_array_context->trainables, &is_trainable);
161	char id[2048];
162	id[0] = add_to_array_context->prefix;
163	id[1] = '-';
164	int total_len = 2;
165	for (i = 0; i < add_to_array_context->sequence->sequences->rnum; i++)
166	{
167		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)));
168		int len;
169		if (name->name && name->name[0] != '\0')
170			len = snprintf(id + total_len, 2048 - total_len, "%s-%d-", name->name, name->sequence);
171		else
172			len = snprintf(id + total_len, 2048 - total_len, "%d-", name->sequence);
173		total_len += len;
174		if (total_len >= 2047)
175			break;
176	}
177	if (total_len < 2047)
178		total_len += snprintf(id + total_len, 2048 - total_len, "%d", add_to_array_context->sequence->it);
179	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", 179, __extension__ __PRETTY_FUNCTION__)
; }));
180	char *heap_id = (char*)ccmallocmalloc(total_len + 1);
181	memcpy(heap_id, id, total_len + 1);
182	ccv_array_push(add_to_array_context->ids, &heap_id);
183	++add_to_array_context->sequence->it;
184}
185 
186static void _ccv_cnnp_compiled_data_init(ccv_cnnp_compiled_data_t* const compiled_data, const int output_size, ccv_array_t* const gradient_checkpoints)
187{
188	compiled_data->f = compiled_data->fits + output_size;
189	compiled_data->xpu_alloc.mp_hdr = -1;
190	compiled_data->xpu_alloc.freed = kh_init(dy_str)kh_init_dy_str();
191	compiled_data->xpu_alloc.allocd = kh_init(dy_alloc)kh_init_dy_alloc();
192	compiled_data->gradient_checkpoints = gradient_checkpoints;
193}
194 
195static int _ccv_cnnp_model_root_parallel_count(const ccv_cnnp_model_t* const model)
196{
197	return ccv_max(model->parallel_count, 1)({ typeof (model->parallel_count) _a = (model->parallel_count
); typeof (1) _b = (1); (_a > _b) ? _a : _b; });
198}
199 
200static int _ccv_cnnp_model_effective_parallel_count(const ccv_cnnp_model_t* const model)
201{
202	int parallel_count = _ccv_cnnp_model_root_parallel_count(model);
203	if (model->graph && model->graph->data_parallel.count > parallel_count)
204		parallel_count = model->graph->data_parallel.count;
205	return parallel_count;
206}
207 
208static int _ccv_cnnp_compiled_data_parallel_count(const ccv_cnnp_model_t* const model, const ccv_cnnp_compiled_data_t* const compiled_data)
209{
210	return compiled_data->parallel_count > 0 ? compiled_data->parallel_count : _ccv_cnnp_model_effective_parallel_count(model);
211}
212 
213ccv_nnc_tensor_symbol_t ccv_cnnp_model_get_symbol(ccv_cnnp_model_t* const self, const ccv_nnc_tensor_symbol_t symbol)
214{
215	assert(self->data)((void) sizeof ((self->data) ? 1 : 0), __extension__ ({ if
 (self->data) ; else __assert_fail ("self->data", "ccv_cnnp_model.c"
, 215, __extension__ __PRETTY_FUNCTION__); }));
216	ccv_cnnp_model_build_data_t* const build_data = (ccv_cnnp_model_build_data_t*)self->data;
217	if (build_data->parallel_count <= 1 || build_data->parallel_rank == 0)
218		return symbol;
219	const int rank = build_data->parallel_rank;
220	assert(rank > 0)((void) sizeof ((rank > 0) ? 1 : 0), __extension__ ({ if (
rank > 0) ; else __assert_fail ("rank > 0", "ccv_cnnp_model.c"
, 220, __extension__ __PRETTY_FUNCTION__); }));
221	assert(rank < build_data->parallel_count)((void) sizeof ((rank < build_data->parallel_count) ? 1
 : 0), __extension__ ({ if (rank < build_data->parallel_count
) ; else __assert_fail ("rank < build_data->parallel_count"
, "ccv_cnnp_model.c", 221, __extension__ __PRETTY_FUNCTION__)
; }));
222	ccv_nnc_symbolic_graph_t* const graph = (ccv_nnc_symbolic_graph_t*)symbol.graph;
223	ccv_nnc_tensor_symbol_t copy = ccv_nnc_tensor_symbol_copy(graph, symbol, rank);
224	if (copy.d != CCV_NNC_NO_TENSOR_SYMBOL)
225		return copy;
226	ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, symbol);
227	if (CCV_TENSOR_GET_MEMORY(params.type)((params.type) & 0x3) == CCV_TENSOR_GPU_MEMORY)
228		CCV_TENSOR_SET_DEVICE_ID(params.type, rank)(params.type) = (((params.type) & ~0xfff00) | (((rank) &
 0xfff) << 8));
229	copy = ccv_nnc_tensor_symbol_new(graph, params, 0);
230	ccv_nnc_tensor_symbol_set_copy(graph, symbol, rank, copy);
231	return copy;
232}
233 
234typedef struct {
235	void* old_graph_exec_symbol_new_hook_context;
236	ccv_nnc_graph_exec_symbol_new_hook_f old_graph_exec_symbol_new_hook;
237	ccv_nnc_symbolic_graph_t* graph;
238	ccv_cnnp_model_build_data_t* build_data;
239} ccv_cnnp_model_set_exec_flags_context_t;
240 
241static void _ccv_cnnp_model_set_exec_flags(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)
242{
243	ccv_cnnp_model_set_exec_flags_context_t* flags_context = (ccv_cnnp_model_set_exec_flags_context_t*)context;
244	if (flags_context->build_data->exec_flags)
245		ccv_nnc_graph_exec_symbol_set_flags(flags_context->graph, symbol, flags_context->build_data->exec_flags);
246	if (flags_context->old_graph_exec_symbol_new_hook)
247		flags_context->old_graph_exec_symbol_new_hook(flags_context->old_graph_exec_symbol_new_hook_context, symbol, cmd, inputs, input_size, outputs, output_size, name);
248}
249 
250static 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)
251{
252	assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if
 (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c"
, 252, __extension__ __PRETTY_FUNCTION__); }));
253	model->inputs = ccmallocmalloc(sizeof(ccv_nnc_tensor_symbol_t) * input_size);
254	int i;
255	for (i = 0; i < input_size; i++)
256		model->inputs[i] = ccv_nnc_tensor_symbol_new(model->graph, inputs[i], 0);
257	ccv_array_t* const parameters = ccv_array_new(sizeof(ccv_nnc_tensor_symbol_t), 0, 0);
258	ccv_array_t* const parameter_ids = ccv_array_new(sizeof(char*), 0, 0);
259	ccv_array_t* const parameter_trainables = ccv_array_new(sizeof(int), 0, 0);
260	ccv_cnnp_model_sequence_t model_sequence = {
261		.bank = kh_init(ccv_cnnp_model_name_bank)kh_init_ccv_cnnp_model_name_bank()
262	};
263	ccv_cnnp_model_add_to_array_context_t add_to_parameter_context = {
264		.add_parameter_indices = 1,
265		.prefix = 't',
266		.sequence = &model_sequence,
267		.symbols = parameters,
268		.ids = parameter_ids,
269		.trainables = parameter_trainables,
270	};
271	ccv_array_t* const internals = ccv_array_new(sizeof(ccv_nnc_tensor_symbol_t), 0, 0);
272	ccv_array_t* const internal_ids = ccv_array_new(sizeof(char*), 0, 0);
273	ccv_cnnp_model_add_to_array_context_t add_to_output_context = {
274		.add_parameter_indices = 0,
275		.prefix = 'r',
276		.sequence = &model_sequence,
277		.symbols = internals,
278		.ids = internal_ids,
279		.trainables = 0,
280	};
281	ccv_cnnp_model_build_data_t build_data = {
282		.exec_flags = 0,
283		.is_trainable = model->is_trainable >= 0 ? model->is_trainable : 1,
284		.parallel_count = 1,
285		.parallel_rank = 0,
286		.model_sequence = &model_sequence,
287		.add_to_array = ccv_cnnp_model_add_to_array,
288		.parameters = parameters,
289		.context = {
290			.add_to_parameter = &add_to_parameter_context,
291			.add_to_output = &add_to_output_context,
292		},
293		.gradient_checkpoints = 0,
294	};
295	model->data = &build_data;
296	ccv_cnnp_model_set_exec_flags_context_t flags_context = {
297		.graph = model->graph,
298		.build_data = &build_data,
299		.old_graph_exec_symbol_new_hook = 0,
300		.old_graph_exec_symbol_new_hook_context = 0
301	};
302	flags_context.old_graph_exec_symbol_new_hook_context = ccv_nnc_graph_exec_symbol_new_hook(model->graph, _ccv_cnnp_model_set_exec_flags, &flags_context, &flags_context.old_graph_exec_symbol_new_hook);
303	ccv_cnnp_model_build(model, model->graph, model->inputs, input_size, 0, 0);
304	// Reset back to previous hook.
305	ccv_nnc_graph_exec_symbol_new_hook(model->graph, flags_context.old_graph_exec_symbol_new_hook, flags_context.old_graph_exec_symbol_new_hook_context, 0);
306	for (i = 0; i < model->output_size; i++)
307	{
308		const ccv_nnc_tensor_symbol_t output = model->outputs[i];
309		const ccv_nnc_tensor_symbol_t alias_to = ccv_nnc_tensor_symbol_alias_to(model->graph, output);
310		if (alias_to.d == CCV_NNC_NO_TENSOR_SYMBOL)
311			continue;
312		// 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
313		// 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
314		// honest, because we cannot handle cases of alias is part of the original tensor but bind differently).
315		const ccv_nnc_tensor_param_t output_params = ccv_nnc_tensor_symbol_params(model->graph, output);
316		model->outputs[i] = ccv_nnc_tensor_symbol_new(model->graph, output_params, 0);
317		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");
318		ccv_nnc_graph_exec_symbol_set_flags(model->graph, make_contiguous, CCV_NNC_GRAPH_EXEC_DISABLE_OPT);
319	}
320	model->data = 0;
321	kh_destroy(ccv_cnnp_model_name_bank, model_sequence.bank)kh_destroy_ccv_cnnp_model_name_bank(model_sequence.bank);
322	if (model_sequence.sequences)
323		ccv_array_free(model_sequence.sequences);
324	// Check if there are parameters that are not trainables. If there are, we will allocate uint64 bitmap to record that.
325	int not_trainables = 0;
326	// Assert no parameter is alias.
327	for (i = 0; i < parameters->rnum; i++)
328	{
329		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)));
330		const ccv_nnc_tensor_symbol_t alias_to = ccv_nnc_tensor_symbol_alias_to(parameter.graph, parameter);
331		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", 331, __extension__ __PRETTY_FUNCTION__)
; })); // Cannot find the one alias to.
332		if (*(int*)ccv_array_get(parameter_trainables, i)((void*)(((char*)((parameter_trainables)->data)) + (size_t
)(parameter_trainables)->rsize * (size_t)(i))) == 0)
333			not_trainables = 1;
334	}
335	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", 335, __extension__ __PRETTY_FUNCTION__)
; }));
336	uint64_t* parameter_flags = 0;
337	if (not_trainables)
338	{
339		parameter_flags = (uint64_t*)cccalloccalloc(((parameters->rnum + 63) >> 6), sizeof(uint64_t));
340		for (i = 0; i < parameter_trainables->rnum; i++)
341			if (*(int*)ccv_array_get(parameter_trainables, i)((void*)(((char*)((parameter_trainables)->data)) + (size_t
)(parameter_trainables)->rsize * (size_t)(i))))
342				parameter_flags[i >> 6] |= ((uint64_t)1 << (i & 63));
343	}
344	ccv_array_free(parameter_trainables);
345	// Assert no internal is alias.
346	for (i = 0; i < internals->rnum; i++)
347	{
348		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)));
349		const ccv_nnc_tensor_symbol_t alias_to = ccv_nnc_tensor_symbol_alias_to(internal.graph, internal);
350		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", 350, __extension__ __PRETTY_FUNCTION__)
; })); // Cannot find the one alias to.
351	}
352	const int output_size = model->output_size;
353	ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
354	const int parameters_rnum = parameters->rnum;
355	if (input_size > 0)
356	{
357		ccv_array_resize(parameters, parameters_rnum + input_size);
358		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));
359	}
360	ccv_nnc_symbolic_graph_simplify(model->graph,
361		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)
362			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)
363			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)
364			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),
365		ccv_array_get(parameters, 0)((void*)(((char*)((parameters)->data)) + (size_t)(parameters
)->rsize * (size_t)(0))), parameters_rnum + input_size,
366		model->outputs, output_size,
367		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));
368	ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
369	// Size it down.
370	parameters->rnum = parameters_rnum;
371	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));
372	_ccv_cnnp_compiled_data_init(compiled_data, output_size, build_data.gradient_checkpoints);
373	const int evaluate_to_size = compiled_data->evaluate.to_size = ccv_nnc_symbolic_graph_destination_size(model->graph);
374	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", 374, __extension__ __PRETTY_FUNCTION__)
; }));
375	compiled_data->evaluate.tos = ccmallocmalloc(sizeof(ccv_nnc_graph_exec_symbol_t) * evaluate_to_size);
376	memcpy(compiled_data->evaluate.tos, ccv_nnc_symbolic_graph_destinations(model->graph), sizeof(ccv_nnc_graph_exec_symbol_t) * evaluate_to_size);
377	compiled_data->loss = loss;
378	if (loss.cmd == CCV_NNC_NOOP)
379	{
380		// If no loss function provided, there is no fits.
381		for (i = 0; i < output_size; i++)
382		{
383			compiled_data->fits[i] = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
384			const ccv_nnc_tensor_symbol_t alias_to = ccv_nnc_tensor_symbol_alias_to(model->graph, model->outputs[i]);
385			if (alias_to.d < 0)
386				compiled_data->f[i] = model->outputs[i];
387			else { // We cannot differentiate against an alias, therefore, we have to verify this output is full, and we can diff against the original.
388				int ofs[CCV_NNC_MAX_DIM_ALLOC(12)];
389				int inc[CCV_NNC_MAX_DIM_ALLOC(12)];
390				ccv_nnc_tensor_symbol_alias_params(model->graph, model->outputs[i], ofs, inc);
391				int j;
392				for (j = 0; j < CCV_NNC_MAX_DIM_ALLOC(12); j++)
393					{ 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"
, 393, __extension__ __PRETTY_FUNCTION__); })); } // There is no ofs.
394				compiled_data->f[i] = alias_to; // Unfortunately, I cannot assert the size yet.
395			}
396		}
397	} else {
398		for (i = 0; i < output_size; i++)
399		{
400			const ccv_nnc_tensor_param_t info = ccv_nnc_tensor_symbol_params(model->graph, model->outputs[i]);
401			const ccv_nnc_tensor_symbol_t fit = compiled_data->fits[i] = ccv_nnc_tensor_symbol_new(model->graph, info, 0);
402			compiled_data->f[i] = ccv_nnc_tensor_symbol_new(model->graph, ccv_nnc_tensor_auto, 0);
403			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);
404		}
405	}
406	ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
407	ccv_nnc_symbolic_graph_simplify(model->graph,
408		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.
409		0, 0, // No need to provide binds at this point.
410		compiled_data->f, model->output_size,
411		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));
412	ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
413	// If inputs are from GPU, stream type is GPU.
414	compiled_data->parameters = parameters;
415	compiled_data->parameter_flags = parameter_flags;
416	compiled_data->internals = internals;
417	compiled_data->ids.parameters = parameter_ids;
418	compiled_data->ids.internals = internal_ids;
419	ccv_cnnp_model_gradient_checkpoints_cleanup_after_build(compiled_data, model->graph);
420}
421 
422static 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)
423{
424	ccv_array_t* const stack = (ccv_array_t*)context;
425	ccv_array_push(stack, &symbol.d);
426}
427 
428static 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)
429{
430	const ccv_nnc_tensor_symbol_t src_symbol = {
431		.d = src_index,
432		.graph = src_graph
433	};
434	const ccv_nnc_tensor_symbol_t dest_symbol = {
435		.d = dest_index,
436		.graph = dest_graph
437	};
438	const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(src_graph, src_symbol);
439	ccv_nnc_tensor_symbol_set(dest_graph, dest_symbol, params);
440	int ofs[CCV_NNC_MAX_DIM_ALLOC(12)];
441	int inc[CCV_NNC_MAX_DIM_ALLOC(12)];
442	if (0 == ccv_nnc_tensor_symbol_alias_params(src_graph, src_symbol, ofs, inc))
443		ccv_nnc_tensor_symbol_alias_set(dest_graph, dest_symbol, ofs, inc);
444}
445 
446static 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)
447{
448	const ccv_nnc_tensor_symbol_t src_symbol = {
449		.d = src_index,
450		.graph = src_graph
451	};
452	const ccv_nnc_tensor_param_t src_params = ccv_nnc_tensor_symbol_params(src_graph, src_symbol);
453	const ccv_nnc_tensor_symbol_t dest_symbol = {
454		.d = dest_index,
455		.graph = dest_graph
456	};
457	const ccv_nnc_tensor_param_t dest_params = ccv_nnc_tensor_symbol_params(dest_graph, dest_symbol);
458	return memcmp(src_params.dim, dest_params.dim, sizeof(src_params.dim)) == 0;
459}
460 
461static 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);
462static void _ccv_cnnp_compiled_data_graph_free(ccv_cnnp_compiled_data_t* const compiled_data);
463 
464typedef struct {
465	int parallel_count;
466	ccv_nnc_symbolic_graph_t* graph;
467	ccv_nnc_graph_exec_arena_t* graph_exec_arena;
468} ccv_nnc_graph_exec_update_t;
469 
470static 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)
471{
472	ccv_nnc_graph_exec_update_t* const graph_exec_update = (ccv_nnc_graph_exec_update_t*)context;
473	ccv_nnc_graph_exec_arena_t* const graph_exec_arena = graph_exec_update->graph_exec_arena;
474	ccv_nnc_graph_exec_t graph_exec = ccv_nnc_graph_exec_from_symbol(graph_exec_arena, symbol);
475	ccv_nnc_graph_exec_set(graph_exec.graph, graph_exec, cmd);
476	ccv_nnc_graph_exec_set_hint(graph_exec.graph, graph_exec, hint);
477	const ccv_nnc_symbolic_graph_t* const graph = graph_exec_update->graph;
478	const int parallel_count = graph_exec_update->parallel_count;
479	int i;
480	for (i = 1; i < parallel_count; i++)
481	{
482		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));
483		if (!CCV_NO_GRAPH_EXEC(copy)((copy).graph == 0))
484		{
485			ccv_nnc_graph_exec_set(copy.graph, copy, cmd);
486			ccv_nnc_graph_exec_set_hint(copy.graph, copy, hint);
487		}
488	}
489}
490 
491void 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)
492{
493	assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if
 (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c"
, 493, __extension__ __PRETTY_FUNCTION__); }));
494	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", 494, __extension__ __PRETTY_FUNCTION__)
; }));
495	assert(!init->graph)((void) sizeof ((!init->graph) ? 1 : 0), __extension__ ({ if
 (!init->graph) ; else __assert_fail ("!init->graph", "ccv_cnnp_model.c"
, 495, __extension__ __PRETTY_FUNCTION__); }));
496	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
497	init->graph = ccv_nnc_symbolic_graph_new();
498	ccv_array_t* const stack = ccv_array_new(sizeof(int), 0, 0);
499	ccv_nnc_graph_exec_symbol_new_hook(init->graph, _ccv_cnnp_graph_push_graph_exec_symbol, stack, 0);
500	_ccv_cnnp_model_compile(init, inputs, input_size, compiled_data->loss);
501	init->parallel_count = model->parallel_count;
502	init->memory_compression = model->memory_compression;
503	init->memory_reduction = model->memory_reduction;
504	init->gradient_checkpointing = model->gradient_checkpointing;
505	init->compiled_data->stream_type = model->compiled_data->stream_type;
506	init->compiled_data->minimize.minimizer = model->compiled_data->minimize.minimizer;
507	init->compiled_data->minimize.max_saved_aux_size = model->compiled_data->minimize.max_saved_aux_size;
508	if (model->compiled_data->gradient_mode != CCV_CNNP_COMPILED_DATA_GRADIENT_NONE)
509		_ccv_cnnp_model_gradient_init(init, model->compiled_data->gradient_mode, model->compiled_data->disable_outgrad, 0, 0);
510	ccv_nnc_graph_exec_symbol_new_hook(init->graph, 0, 0, 0);
511	ccv_nnc_symbolic_graph_tensor_auto(init->graph, TRAVERSE_FULL0,0,0,0);
512	int i, j;
513	// Verify parameters, internals and saved_aux in both graph has the same dimensionality.
514	for (i = 0; i < compiled_data->parameters->rnum; i++)
515	{
516		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;
517		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", 517, __extension__ __PRETTY_FUNCTION__)
; }));
518	}
519	for (i = 0; i < compiled_data->internals->rnum; i++)
520	{
521		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;
522		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", 522, __extension__ __PRETTY_FUNCTION__)
; }));
523	}
524	// Update inputs.
525	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", 525, __extension__ __PRETTY_FUNCTION__)
; }));
526	for (i = 0; i < model->input_size; i++)
527		if (model->inputs[i].d >= 0)
528		{
529			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", 529, __extension__ __PRETTY_FUNCTION__)
; }));
530			_ccv_nnc_tensor_symbol_reinit(init->graph, model->graph, init->inputs[i].d, model->inputs[i].d);
531		}
532	// Update outputs.
533	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", 533, __extension__ __PRETTY_FUNCTION__)
; }));
534	for (i = 0; i < model->output_size; i++)
535	{
536		if (model->outputs[i].d >= 0)
537		{
538			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", 538, __extension__
 __PRETTY_FUNCTION__); }));
539			_ccv_nnc_tensor_symbol_reinit(init->graph, model->graph, init->outputs[i].d, model->outputs[i].d);
540		}
541		if (model->outputs[i].d != model->compiled_data->f[i].d)
542		{
543			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", 543, __extension__ __PRETTY_FUNCTION__)
; }));
544			if (model->compiled_data->f[i].d >= 0)
545			{
546				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", 546, __extension__ __PRETTY_FUNCTION__)
; }));
547				_ccv_nnc_tensor_symbol_reinit(init->graph, model->graph, init->compiled_data->f[i].d, model->compiled_data->f[i].d);
548			}
549		}
550	}
551	// Go through the graph to set tensor on matching symbols
552	for (i = 0; i < stack->rnum; i++)
553	{
554		const int d = *(int*)ccv_array_get(stack, i)((void*)(((char*)((stack)->data)) + (size_t)(stack)->rsize
 * (size_t)(i)));
555		// If exceed range, skip.
556		if (d >= ccv_nnc_graph_exec_symbol_count(init->graph) ||
557			d >= ccv_nnc_graph_exec_symbol_count(model->graph))
558			continue;
559		const ccv_nnc_graph_exec_symbol_t src_symbol = {
560			.d = d,
561			.graph = init->graph
562		};
563		const ccv_nnc_graph_exec_symbol_t dest_symbol = {
564			.d = d,
565			.graph = model->graph
566		};
567		const ccv_nnc_cmd_t src_cmd = ccv_nnc_graph_exec_symbol_cmd(init->graph, src_symbol);
568		const ccv_nnc_cmd_t dest_cmd = ccv_nnc_graph_exec_symbol_cmd(model->graph, dest_symbol);
569		// If the name doesn't match, skip.
570		if (dest_cmd.cmd != src_cmd.cmd && src_cmd.cmd != CCV_NNC_NOOP)
571			continue;
572		// Now get all the inputs and outputs, if matches, set them.
573		const int* src_inputs;
574		int src_input_size;
575		const int* src_outputs;
576		int src_output_size;
577		ccv_nnc_graph_exec_symbol_io(init->graph, src_symbol, &src_inputs, &src_input_size, &src_outputs, &src_output_size);
578		const int* dest_inputs;
579		int dest_input_size;
580		const int* dest_outputs;
581		int dest_output_size;
582		ccv_nnc_graph_exec_symbol_io(model->graph, dest_symbol, &dest_inputs, &dest_input_size, &dest_outputs, &dest_output_size);
583		// We may have unmatched input / output size because this is the minimizer and it has
584		// different saved_aux (for example, when we shrunk with CMD_NOOP).
585		if (src_input_size != dest_input_size)
586			continue;
587		if (src_output_size != dest_output_size)
588			continue;
589		ccv_nnc_graph_exec_symbol_set(model->graph, dest_symbol, src_cmd);
590		// There may be mismatches of the source tensor symbols and destination tensor symbols. The reason is because
591		// we may later passed-in the minimizer, therefore, we may allocate tensors for minimizer later in the original
592		// graph whereas in the newly created graph, it is streamlined (the minimizer exists from the beginning). That
593		// will make the order of tensor symbols creation different, therefore, exact which tensor is which wrong as
594		// well. However, set a new minimizer won't change the exec symbol ordering, because we never create new exec
595		// symbols after gradient init step. Changing a new minimizer just updated that exec symbols setting, it is not
596		// a new exec symbol.
597		for (j = 0; j < src_input_size; j++)
598			if (src_inputs[j] >= 0)
599				_ccv_nnc_tensor_symbol_reinit(init->graph, model->graph, src_inputs[j], dest_inputs[j]);
600		for (j = 0; j < src_output_size; j++)
601			if (src_outputs[j] >= 0)
602				_ccv_nnc_tensor_symbol_reinit(init->graph, model->graph, src_outputs[j], dest_outputs[j]);
603	}
604	ccv_array_free(stack);
605	// After this, we get all tensors in the model graph resolved through tensor_auto.
606	ccv_nnc_symbolic_graph_tensor_auto(model->graph, TRAVERSE_FULL0,0,0,0);
607	// Verify symbols we get matches.
608	const int parameter_size = compiled_data->parameters->rnum;
609	for (i = 0; i < parameter_size; i++)
610		{ 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", 610, __extension__ __PRETTY_FUNCTION__)
; })); }
611	const int internal_size = compiled_data->internals->rnum;
612	for (i = 0; i < internal_size; i++)
613		{ 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", 613, __extension__ __PRETTY_FUNCTION__)
; })); }
614	// Go through compiled data.
615	if (compiled_data->tensor_arena)
616	{
617		const int flag = ccv_nnc_tensor_arena_reinit(compiled_data->tensor_arena, model->graph);
618		if (flag == 0 && compiled_data->graph_exec_arena)
619		{
620			ccv_nnc_graph_exec_reinit(compiled_data->graph_exec_arena, compiled_data->graph, model->graph);
621			// Since we will reinit, if we previously set is_test, we need to set it again.
622			if (compiled_data->is_test)
623			{
624				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; });
625				ccv_nnc_graph_exec_update_t update = {
626					.parallel_count = parallel_count,
627					.graph = model->graph,
628					.graph_exec_arena = compiled_data->graph_exec_arena,
629				};
630				ccv_cnnp_model_set_is_test(model, 1, _ccv_cnnp_cmd_update_for_execs, &update);
631			}
632		} else
633			// Free-up tensor arena & graph exec arena.
634			_ccv_cnnp_compiled_data_graph_free(compiled_data);
635	}
636	// There are other compiled graphs, for accum and apply gradients.
637	// However, the main conclusion is, these absorb operations shouldn't impact parameters.
638	// Thus, it won't impact the shape of gradients (only outgrad). Since for outgrad, we
639	// don't allocate ourselves, it is not a concern. For normal gradients, the shape cannot
640	// be changed otherwise parameters' shape will be meaningless. The same goes to internals.
641	// That is why we don't update these compiled graphs at all this point.
642	// Free the model, we've already "absorbed" it.
643	ccv_cnnp_model_free(init);
644}
645 
646void 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)
647{
648	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", 648, __extension__ __PRETTY_FUNCTION__)
; }));
649	if (model->input_size == 0)
650		model->input_size = input_size;
651	if (!model->graph) // The graph is not compiled yet.
652	{
653		model->graph = ccv_nnc_symbolic_graph_new();
654		_ccv_cnnp_model_compile(model, inputs, input_size, loss);
655		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", 655, __extension__ __PRETTY_FUNCTION__)
; }));
656		int i, flag = 0;
657		for (i = 0; !flag && i < input_size; i++)
658			flag = (CCV_TENSOR_GET_MEMORY(inputs[i].type)((inputs[i].type) & 0x3) == CCV_TENSOR_GPU_MEMORY);
659		// If inputs are from GPU, stream type is GPU.
660		model->compiled_data->stream_type = flag ? CCV_STREAM_CONTEXT_GPU : CCV_STREAM_CONTEXT_CPU;
661		model->compiled_data->minimize.minimizer = minimizer;
662		model->compiled_data->minimize.max_saved_aux_size = ccv_nnc_minimizer_saved_aux_size(minimizer);
663	} else {
664		// Now, finally fill in this part. If the graph is already compiled, we make a copy of the model.
665		// And then absorb the "new model" to the old one.
666		ccv_cnnp_model_t* const init = ccv_cnnp_model_copy(model, model->is_trainable);
667		ccv_cnnp_model_absorb(model, init, inputs, input_size);
668		// Reset minimizer.
669		ccv_cnnp_model_set_minimizer(model, minimizer, 1, 0, 0);
670	}
671}
672 
673ccv_cnnp_model_t* ccv_cnnp_model_copy(const ccv_cnnp_model_t* const model, const int is_trainable)
674{
675	ccv_cnnp_model_t* const new_model = _ccv_cnnp_model_copy(model, 0);
676	new_model->is_trainable = is_trainable;
677	return new_model;
678}
679 
680void ccv_cnnp_model_tensor_auto(ccv_cnnp_model_t* const model, ccv_nnc_tensor_param_t* const outputs, const int output_size)
681{
682	assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if
 (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c"
, 682, __extension__ __PRETTY_FUNCTION__); }));
683	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", 683, __extension__ __PRETTY_FUNCTION__)
; }));
684	ccv_nnc_symbolic_graph_t* const graph = model->graph;
685	ccv_nnc_symbolic_graph_tensor_auto(graph, TRAVERSE_FULL0,0,0,0);
686	int i;
687	for (i = 0; i < output_size; i++)
688	{
689		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", 689, __extension__ __PRETTY_FUNCTION__)
; }));
690		outputs[i] = ccv_nnc_tensor_symbol_params(graph, model->outputs[i]);
691	}
692}
693 
694void ccv_cnnp_model_set_workspace_size(ccv_cnnp_model_t* const model, size_t workspace_size)
695{
696	if (workspace_size == model->workspace_size)
697		return;
698	model->workspace_size = workspace_size;
699	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
700	if (compiled_data && compiled_data->graph)
701		ccv_nnc_graph_autotune(compiled_data->graph, workspace_size, 0, TRAVERSE_FULL0,0,0,0);
702}
703 
704size_t ccv_cnnp_model_workspace_size(ccv_cnnp_model_t* const model)
705{
706	return model->workspace_size;
707}
708 
709void ccv_cnnp_model_set_data_parallel(ccv_cnnp_model_t* const model, const int parallel)
710{
711	if (parallel == 0)
712		model->parallel_count = ccv_nnc_device_count(CCV_STREAM_CONTEXT_GPU);
713	else
714		model->parallel_count = parallel;
715	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
716	if (compiled_data)
717		{ 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", 717, __extension__ __PRETTY_FUNCTION__)
; })); }
718}
719 
720void ccv_cnnp_model_set_max_concurrency(ccv_cnnp_model_t* const model, const int max_stream_count)
721{
722	model->max_stream_count = max_stream_count;
723	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
724	if (compiled_data)
725		{ 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", 725, __extension__ __PRETTY_FUNCTION__)
; })); }
726}
727 
728void ccv_cnnp_model_set_memory_compression(ccv_cnnp_model_t* const model, const int memory_compression)
729{
730	model->memory_compression = memory_compression;
731	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
732	if (compiled_data)
733		{ 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", 733, __extension__ __PRETTY_FUNCTION__)
; })); }
734}
735 
736void ccv_cnnp_model_set_memory_reduction(ccv_cnnp_model_t* const model, const int memory_reduction)
737{
738	model->memory_reduction = memory_reduction;
739	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
740	if (compiled_data)
741		{ 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", 741, __extension__ __PRETTY_FUNCTION__)
; })); }
742}
743 
744void ccv_cnnp_model_set_gradient_checkpointing(ccv_cnnp_model_t* const model, const int gradient_checkpointing)
745{
746	model->gradient_checkpointing = gradient_checkpointing;
747}
748 
749int ccv_cnnp_model_gradient_checkpointing(ccv_cnnp_model_t* const model)
750{
751	return model->gradient_checkpointing;
752}
753 
754typedef struct {
755	int parallel_count;
756	ccv_nnc_symbolic_graph_t* graph;
757	ccv_cnnp_compiled_data_t* compiled_data;
758	ccv_nnc_tensor_arena_t* tensor_arena;
759} ccv_nnc_tensor_init_states_t;
760 
761static int _ccv_cnnp_any_to_init(const ccv_cnnp_compiled_data_t* const compiled_data)
762{
763	int i;
764	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));
765	for (i = 0; i < compiled_data->parameters->rnum; i++)
766	{
767		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;
768		if (!(init_v[d >> 5] & (1u << (d & 0x1f))))
769			return 1;
770	}
771	for (i = 0; i < compiled_data->internals->rnum; i++)
772	{
773		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;
774		if (!(init_v[d >> 5] & (1u << (d & 0x1f))))
775			return 1;
776	}
777	return 0;
778}
779 
780static 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)
781{
782	ccv_nnc_tensor_init_states_t* const tensor_init_states = (ccv_nnc_tensor_init_states_t*)context;
783	ccv_nnc_tensor_arena_t* const tensor_arena = tensor_init_states->tensor_arena;
784	ccv_nnc_tensor_t* const output_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, output_symbol);
785	if (!output_tensor)
786		return;
787	const int d = output_symbol.d;
788	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", 788, __extension__ __PRETTY_FUNCTION__)
; }));
789  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));
790	if (init_v[d >> 5] & (1u << (d & 0x1f)))
791		return;
792	init_v[d >> 5] |= (1u << (d & 0x1f));
793	ccv_nnc_cmd_exec(cmd, hint, flags, &input, input ? 1 : 0, &output_tensor, 1, 0);
794	const ccv_nnc_symbolic_graph_t* const graph = tensor_init_states->graph;
795	const int parallel_count = tensor_init_states->parallel_count;
796	int i;
797	for (i = 1; i < parallel_count; i++)
798	{
799		ccv_nnc_tensor_t* const copy = ccv_nnc_tensor_from_symbol(tensor_arena, ccv_nnc_tensor_symbol_copy(graph, output_symbol, i));
800		if (copy)
801			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);
802	}
803}
804 
805// This method can only handle cases we added new tensors and exec, never delete. This invariant is true because
806// we setup everything (including calling simplify method) in ccv_cnnp_model_compile method, before this rewind setup.
807static void _ccv_cnnp_model_rewind_graph(ccv_cnnp_model_t* const model)
808{
809	assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if
 (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c"
, 809, __extension__ __PRETTY_FUNCTION__); }));
810	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", 810, __extension__ __PRETTY_FUNCTION__)
; }));
811	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
812	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", 812, __extension__
 __PRETTY_FUNCTION__); }));
813	int i;
814	for (i = 0; i < compiled_data->rewindables->rnum; i++)
815	{
816		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)));
817		if (rewind_symbol->type == CCV_CNNP_REWIND_GRAPH_EXEC)
818			ccv_nnc_graph_exec_symbol_free(model->graph, rewind_symbol->graph_exec);
819		else if (rewind_symbol->type == CCV_CNNP_REWIND_TENSOR)
820			ccv_nnc_tensor_symbol_free(model->graph, rewind_symbol->tensor);
821	}
822	ccv_array_clear(compiled_data->rewindables);
823	ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
824}
825 
826static 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)
827{
828	const ccv_cnnp_rewind_symbol_t rewind_symbol = {
829		.type = CCV_CNNP_REWIND_TENSOR,
830		.tensor = symbol
831	};
832	ccv_array_t* const rewind_symbols = (ccv_array_t*)context;
833	ccv_array_push(rewind_symbols, &rewind_symbol);
834}
835 
836static 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)
837{
838	const ccv_cnnp_rewind_symbol_t rewind_symbol = {
839		.type = CCV_CNNP_REWIND_TENSOR,
840		.tensor = symbol
841	};
842	ccv_array_t* const rewind_symbols = (ccv_array_t*)context;
843	ccv_array_push(rewind_symbols, &rewind_symbol);
844}
845 
846static 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)
847{
848	const ccv_cnnp_rewind_symbol_t rewind_symbol = {
849		.type = CCV_CNNP_REWIND_GRAPH_EXEC,
850		.graph_exec = symbol
851	};
852	ccv_array_t* const rewind_symbols = (ccv_array_t*)context;
853	ccv_array_push(rewind_symbols, &rewind_symbol);
854}
855 
856static 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)
857{
858	ccv_nnc_graph_exec_t const update_exec = ccv_nnc_graph_exec_from_symbol(graph_exec_arena, exec_symbol);
859	if (!CCV_NO_GRAPH_EXEC(update_exec)((update_exec).graph == 0))
860		ccv_nnc_graph_exec_set(update_exec.graph, update_exec, cmd);
861	int i;
862	for (i = 1; i < parallel_count; i++)
863	{
864		ccv_nnc_graph_exec_symbol_t copy_symbol = ccv_nnc_graph_exec_symbol_copy(symbolic_graph, exec_symbol, i);
865		const ccv_nnc_graph_exec_t copy = ccv_nnc_graph_exec_from_symbol(graph_exec_arena, copy_symbol);
866		if (!CCV_NO_GRAPH_EXEC(copy)((copy).graph == 0))
867			ccv_nnc_graph_exec_set(copy.graph, copy, cmd);
868	}
869}
870 
871static 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)
872{
873	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 873, __extension__ __PRETTY_FUNCTION__); }));
874	assert(symbolic_graph)((void) sizeof ((symbolic_graph) ? 1 : 0), __extension__ ({ if
 (symbolic_graph) ; else __assert_fail ("symbolic_graph", "ccv_cnnp_model.c"
, 874, __extension__ __PRETTY_FUNCTION__); }));
875	ccv_nnc_graph_exec_symbol_set(symbolic_graph, exec_symbol, cmd);
876	int i;
877	for (i = 1; i < parallel_count; i++)
878	{
879		ccv_nnc_graph_exec_symbol_t copy_symbol = ccv_nnc_graph_exec_symbol_copy(symbolic_graph, exec_symbol, i);
880		if (copy_symbol.graph)
881			ccv_nnc_graph_exec_symbol_set(symbolic_graph, copy_symbol, cmd);
882	}
883	ccv_nnc_graph_exec_arena_t* const graph_exec_arena = compiled_data->graph_exec_arena;
884	if (graph_exec_arena)
885		_ccv_cnnp_model_graph_symbol_exec_set_for_graph_exec_arena(graph_exec_arena, parallel_count, exec_symbol, cmd, symbolic_graph);
886	// Skip backward graph exec arena because it is for a specific accum symbolic graph, not the main graph (model->graph)
887	ccv_nnc_graph_exec_arena_t* const gradient_graph_exec_arena = compiled_data->apply_gradients.graph_exec_arena;
888	if (gradient_graph_exec_arena)
889		_ccv_cnnp_model_graph_symbol_exec_set_for_graph_exec_arena(gradient_graph_exec_arena, parallel_count, exec_symbol, cmd, symbolic_graph);
890}
891 
892static 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)
893{
894	int this_parameter_flag = 0;
895	if (update_nodes[parameter_indice].d == CCV_NNC_NO_TENSOR_SYMBOL)
896		return this_parameter_flag;
897	const ccv_nnc_cmd_t old_minimizer = ccv_nnc_graph_exec_symbol_cmd(graph, update_nodes[parameter_indice]);
898	int j, k;
899	// For no-op, we can preserve previous saved_aux_size.
900	if (old_minimizer.cmd != minimizer.cmd && minimizer.cmd != CCV_NNC_NOOP)
901	{
902		// If the old minimizer is a noop, then the old_saved_aux_size should be whatever its previous
903		// saved_aux_size is, otherwise we will reinit the saved_aux repeatedly if you switch between
904		// noop and a minimizer. We don't want that because we do that in high-level frameworks to
905		// make sure some model parameters don't update if we don't want them to.
906		int old_saved_aux_size;
907		if (old_minimizer.cmd == CCV_NNC_NOOP)
908		{
909			int input_size;
910			ccv_nnc_graph_exec_symbol_io(graph, update_nodes[parameter_indice], 0, &input_size, 0, 0);
911			if (input_size < 2) // This is not legit.
912				old_saved_aux_size = ccv_nnc_minimizer_saved_aux_size(old_minimizer);
913			else // See ccv_nnc_minimizer_saved_aux_size, the saved_aux is inputs excluding gradients and parameters.
914				old_saved_aux_size = input_size - 2;
915		} else
916			old_saved_aux_size = ccv_nnc_minimizer_saved_aux_size(old_minimizer);
917		if (old_saved_aux_size != saved_aux_size)
918		{
919			this_parameter_flag = 1;
920			if (saved_aux_size > old_saved_aux_size)
921			{
922				// Allocate new tensor symbols.
923				const ccv_nnc_tensor_param_t info = ccv_nnc_tensor_symbol_params(graph, updated_parameters[parameter_indice]);
924				for (j = old_saved_aux_size; j < saved_aux_size; j++)
925				{
926					saved_aux[parameter_indice * max_saved_aux_size + j].source = ccv_nnc_tensor_symbol_new(graph, info, 0);
927					saved_aux[parameter_indice * max_saved_aux_size + j].destination = ccv_nnc_tensor_symbol_new(graph, info, 0);
928					const int device_id = CCV_TENSOR_GET_DEVICE_ID(info.type)(((info.type) & 0xfff00) >> 8);
929					for (k = 1; k < parallel_count; k++)
930					{
931						ccv_nnc_tensor_param_t dev_info = info;
932						if (k != device_id)
933							CCV_TENSOR_SET_DEVICE_ID(dev_info.type, k)(dev_info.type) = (((dev_info.type) & ~0xfff00) | (((k) &
 0xfff) << 8));
934						else
935							CCV_TENSOR_SET_DEVICE_ID(dev_info.type, 0)(dev_info.type) = (((dev_info.type) & ~0xfff00) | (((0) &
 0xfff) << 8));
936						const ccv_nnc_tensor_symbol_t src_copy = ccv_nnc_tensor_symbol_new(graph, dev_info, 0);
937						const ccv_nnc_tensor_symbol_t dest_copy = ccv_nnc_tensor_symbol_new(graph, dev_info, 0);
938						ccv_nnc_tensor_symbol_set_copy(graph, saved_aux[parameter_indice * max_saved_aux_size + j].source, k, src_copy);
939						ccv_nnc_tensor_symbol_set_copy(graph, saved_aux[parameter_indice * max_saved_aux_size + j].destination, k, dest_copy);
940					}
941				}
942			} else {
943				for (j = saved_aux_size; j < old_saved_aux_size; j++)
944				{
945					for (k = 1; k < parallel_count; k++)
946					{
947						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);
948						if (src_copy.d >= 0)
949						{
950							ccv_nnc_tensor_symbol_free(graph, src_copy);
951							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
});
952						}
953						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);
954						if (dest_copy.d >= 0)
955						{
956							ccv_nnc_tensor_symbol_free(graph, dest_copy);
957							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
});
958						}
959					}
960					ccv_nnc_tensor_symbol_free(graph, saved_aux[parameter_indice * max_saved_aux_size + j].source);
961					ccv_nnc_tensor_symbol_free(graph, saved_aux[parameter_indice * max_saved_aux_size + j].destination);
962					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
};
963				}
964			}
965		}
966	}
967	_ccv_cnnp_model_graph_exec_symbol_set(graph, compiled_data, parallel_count, update_nodes[parameter_indice], minimizer);
968	if (this_parameter_flag)
969	{
970		ccv_nnc_tensor_symbol_t update_inputs[saved_aux_size + 2];
971		ccv_nnc_tensor_symbol_t update_outputs[saved_aux_size + 1];
972		const int* inputs = 0;
973		int input_size = 0;
974		ccv_nnc_graph_exec_symbol_io(graph, update_nodes[parameter_indice], &inputs, &input_size, 0, 0);
975		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", 975, __extension__ __PRETTY_FUNCTION__)
; }));
976		update_inputs[0].d = inputs[0];
977		update_inputs[0].graph = graph;
978		update_inputs[1].d = inputs[1];
979		update_inputs[1].graph = graph;
980		update_outputs[0] = updated_parameters[parameter_indice];
981		for (j = 0; j < saved_aux_size; j++)
982		{
983			update_inputs[j + 2] = saved_aux[parameter_indice * max_saved_aux_size + j].source;
984			update_outputs[j + 1] = saved_aux[parameter_indice * max_saved_aux_size + j].destination;
985		}
986		ccv_nnc_graph_exec_symbol_set_io(graph, update_nodes[parameter_indice], update_inputs, saved_aux_size + 2, update_outputs, saved_aux_size + 1);
987		for (k = 1; k < parallel_count; k++)
988		{
989			const ccv_nnc_graph_exec_symbol_t copy = ccv_nnc_graph_exec_symbol_copy(graph, update_nodes[parameter_indice], k);
990			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"
, 990, __extension__ __PRETTY_FUNCTION__); }));
991			ccv_nnc_graph_exec_symbol_io(graph, copy, &inputs, &input_size, 0, 0);
992			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", 992, __extension__ __PRETTY_FUNCTION__)
; }));
993			update_inputs[0].d = inputs[0];
994			update_inputs[0].graph = graph;
995			update_inputs[1].d = inputs[1];
996			update_inputs[1].graph = graph;
997			update_outputs[0] = ccv_nnc_tensor_symbol_copy(graph, updated_parameters[parameter_indice], k);
998			for (j = 0; j < saved_aux_size; j++)
999			{
1000				update_inputs[j + 2] = ccv_nnc_tensor_symbol_copy(graph, saved_aux[parameter_indice * max_saved_aux_size + j].source, k);
1001				update_outputs[j + 1] = ccv_nnc_tensor_symbol_copy(graph, saved_aux[parameter_indice * max_saved_aux_size + j].destination, k);
1002			}
1003			ccv_nnc_graph_exec_symbol_set_io(graph, copy, update_inputs, saved_aux_size + 2, update_outputs, saved_aux_size + 1);
1004		}
1005	}
1006	return this_parameter_flag;
1007}
1008 
1009typedef struct {
1010	int parameter_size;
1011	ccv_nnc_cmd_t minimizer;
1012	ccv_cnnp_model_io_t parameters[1];
1013} ccv_cnnp_set_minimizer_for_parameter_t;
1014 
1015static int _ccv_cnnp_apply_parameters_with_minimizer(ccv_cnnp_model_t* const model)
1016{
1017	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1018	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 1018, __extension__ __PRETTY_FUNCTION__); }));
1019	const int max_saved_aux_size = compiled_data->minimize.max_saved_aux_size;
1020	// We update all parameters, at this point, we have one minimizer.
1021	const int parameter_size = compiled_data->parameters->rnum;
1022	ccv_nnc_graph_exec_symbol_t* const update_nodes = compiled_data->update_nodes;
1023	ccv_nnc_symbolic_graph_t* const symbolic_graph = model->graph;
1024	assert(symbolic_graph)((void) sizeof ((symbolic_graph) ? 1 : 0), __extension__ ({ if
 (symbolic_graph) ; else __assert_fail ("symbolic_graph", "ccv_cnnp_model.c"
, 1024, __extension__ __PRETTY_FUNCTION__); }));
1025	const int parallel_count = _ccv_cnnp_model_root_parallel_count(model);
1026	assert(_ccv_cnnp_model_effective_parallel_count(model) == parallel_count && "local replicated stateful models only support forward / no-grad evaluation for now")((void) sizeof ((_ccv_cnnp_model_effective_parallel_count(model
) == parallel_count && "local replicated stateful models only support forward / no-grad evaluation for now"
) ? 1 : 0), __extension__ ({ if (_ccv_cnnp_model_effective_parallel_count
(model) == parallel_count && "local replicated stateful models only support forward / no-grad evaluation for now"
) ; else __assert_fail ("_ccv_cnnp_model_effective_parallel_count(model) == parallel_count && \"local replicated stateful models only support forward / no-grad evaluation for now\""
, "ccv_cnnp_model.c", 1026, __extension__ __PRETTY_FUNCTION__
); }));
1027	ccv_array_t* const parameters = compiled_data->minimize.parameters;
1028	ccv_array_t* const parameter_indices = ccv_array_new(sizeof(int), 0, 0);
1029	int i, j, flag = 0;
1030	for (i = 0; i < parameters->rnum; i++)
1031	{
1032		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)));
1033		for (j = 0; j < set_minimizer_for_parameter->parameter_size; j++)
1034		{
1035			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;
1036			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", 1036, __extension__ __PRETTY_FUNCTION__
); }));
1037			const int old_rnum = parameter_indices->rnum;
1038			ccv_cnnp_model_add_to_parameter_indices(set_minimizer_for_parameter->parameters[j]->model, param_sel, parameter_indices);
1039			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;
1040			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", 1040, __extension__ __PRETTY_FUNCTION__
); }));
1041			if (param_ref >= 0)
1042			{
1043				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", 1043, __extension__ __PRETTY_FUNCTION__
); }));
1044				*(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)));
1045				parameter_indices->rnum = old_rnum + 1;
1046			}
1047		}
1048		const int saved_aux_size = ccv_nnc_minimizer_saved_aux_size(set_minimizer_for_parameter->minimizer);
1049		// We may have duplicated indices, but that is OK, we will set it twice.
1050		for (j = 0; j < parameter_indices->rnum; j++)
1051		{
1052			const int d = *(int*)ccv_array_get(parameter_indices, j)((void*)(((char*)((parameter_indices)->data)) + (size_t)(parameter_indices
)->rsize * (size_t)(j)));
1053			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", 1053, __extension__ __PRETTY_FUNCTION__
); }));
1054			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))
1055				flag = 1;
1056		}
1057		ccv_array_clear(parameter_indices);
1058	}
1059	ccv_array_free(parameter_indices);
1060	return flag;
1061}
1062 
1063static 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)
1064{
1065	if (new_saved_aux_size == old_saved_aux_size)
1066		return;
1067	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", 1067, __extension__ __PRETTY_FUNCTION__
); }));
1068	int i, j;
1069	for (i = parameter_size - 1; i >= 0; i--)
1070	{
1071		for (j = new_saved_aux_size - 1; j >= old_saved_aux_size; j--)
1072			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
};
1073		for (j = old_saved_aux_size - 1; j >= 0; j--)
1074			saved_aux[i * new_saved_aux_size + j] = saved_aux[i * old_saved_aux_size + j];
1075	}
1076}
1077 
1078static void _ccv_cnnp_model_set_rewindables(ccv_cnnp_model_t* const model)
1079{
1080	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1081	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 1081, __extension__ __PRETTY_FUNCTION__); }));
1082	if (!compiled_data->rewindables)
1083		compiled_data->rewindables = ccv_array_new(sizeof(ccv_cnnp_rewind_symbol_t), 0, 0);
1084	ccv_nnc_tensor_symbol_new_hook(model->graph, _ccv_cnnp_model_tensor_symbol_new_hook, compiled_data->rewindables, 0);
1085	ccv_nnc_tensor_symbol_alias_new_hook(model->graph, _ccv_cnnp_model_tensor_symbol_alias_new_hook, compiled_data->rewindables, 0);
1086	ccv_nnc_graph_exec_symbol_new_hook(model->graph, _ccv_cnnp_model_graph_exec_symbol_new_hook, compiled_data->rewindables, 0);
1087}
1088 
1089static 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)
1090{
1091	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1092	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", 1092, __extension__ __PRETTY_FUNCTION__
); }));
1093	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", 1093, __extension__ __PRETTY_FUNCTION__
); }));
1094	const int evaluate_to_size = compiled_data->evaluate.to_size;
1095	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", 1095, __extension__ __PRETTY_FUNCTION__
); }));
1096	const int parallel_count = _ccv_cnnp_model_root_parallel_count(model);
1097	assert(_ccv_cnnp_model_effective_parallel_count(model) == parallel_count && "local replicated stateful models only support forward / no-grad evaluation for now")((void) sizeof ((_ccv_cnnp_model_effective_parallel_count(model
) == parallel_count && "local replicated stateful models only support forward / no-grad evaluation for now"
) ? 1 : 0), __extension__ ({ if (_ccv_cnnp_model_effective_parallel_count
(model) == parallel_count && "local replicated stateful models only support forward / no-grad evaluation for now"
) ; else __assert_fail ("_ccv_cnnp_model_effective_parallel_count(model) == parallel_count && \"local replicated stateful models only support forward / no-grad evaluation for now\""
, "ccv_cnnp_model.c", 1097, __extension__ __PRETTY_FUNCTION__
); }));
1098	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);
1099	compiled_data->evaluate.to_ops = (ccv_nnc_graph_exec_t*)(compiled_data->evaluate.tos + evaluate_to_size * parallel_count);
1100	int i, j;
1101	const int output_size = model->output_size;
1102	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", 1102, __extension__ __PRETTY_FUNCTION__
); }));
1103	if (fits)
1104		for (i = 0; i < output_size; i++)
1105			ccv_nnc_tensor_symbol_set(model->graph, compiled_data->fits[i], fits[i]->info);
1106	const int max_saved_aux_size = compiled_data->minimize.max_saved_aux_size;
1107	const int parameter_size = compiled_data->parameters->rnum;
1108	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);
1109	compiled_data->update_nodes = (ccv_nnc_graph_exec_symbol_t*)(compiled_data->updated_parameters + parameter_size);
1110	compiled_data->saved_aux = (ccv_nnc_tensor_symbol_map_t*)(compiled_data->update_nodes + parameter_size);
1111	int parameter_size_maybe_more = parameter_size;
1112	compiled_data->disable_outgrad = disable_outgrad;
1113	int outgrad_size;
1114	if (gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES || model->input_size == 0)
1115		outgrad_size = 0;
1116	else if (disable_outgrad == CCV_CNNP_DISABLE_OUTGRAD_NONE) // Compute minimize with gradients including inputs.
1117		outgrad_size = model->input_size;
1118	else {
1119		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", 1119, __extension__ __PRETTY_FUNCTION__
); })); // If it is disable all, gradient mode won't be this.
1120		outgrad_size = 0;
1121		for (i = 0; i < model->input_size; i++)
1122			if (!(disable_outgrad & ((uint64_t)1 << i)))
1123				++outgrad_size;
1124	}
1125	compiled_data->outgrad_size = outgrad_size;
1126	parameter_size_maybe_more += outgrad_size;
1127	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);
1128	compiled_data->outgrads = parameter_size_maybe_more > parameter_size ? compiled_data->gradients + parameter_size : 0;
1129	compiled_data->backward.tos = (ccv_nnc_graph_exec_symbol_t*)(compiled_data->gradients + parameter_size_maybe_more);
1130	compiled_data->backward.to_size = parameter_size_maybe_more;
1131	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)));
1132	if (compiled_data->parameter_flags)
1133	{
1134		parameters = (ccv_nnc_tensor_symbol_t*)ccmallocmalloc(sizeof(ccv_nnc_tensor_symbol_t) * parameter_size);
1135		for (i = 0; i < parameter_size; i++)
1136			if (compiled_data->parameter_flags[i >> 6] & ((uint64_t)1 << (i & 63)))
1137				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)));
1138			else
1139				parameters[i] = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
1140	}
1141	if (gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES || model->input_size == 0)
1142		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);
1143	else if (disable_outgrad == CCV_CNNP_DISABLE_OUTGRAD_NONE) // Compute minimize with gradients including inputs.
1144		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);
1145	else { // Compute minimize with gradients including selected inputs.
1146		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", 1146, __extension__ __PRETTY_FUNCTION__
); }));
1147		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", 1147, __extension__ __PRETTY_FUNCTION__
); })); // If it is disable all, gradient mode won't be this.
1148		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", 1148, __extension__ __PRETTY_FUNCTION__
); }));
1149		ccv_nnc_tensor_symbol_t outgrads[outgrad_size];
1150		j = 0;
1151		for (i = 0; i < model->input_size; i++)
1152			if (!(disable_outgrad & ((uint64_t)1 << i)))
1153				outgrads[j++] = model->inputs[i];
1154		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);
1155	}
1156	if (compiled_data->parameter_flags)
1157		ccfreefree(parameters);
1158	_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);
1159	if (compiled_data->minimize.parameters)
1160		_ccv_cnnp_apply_parameters_with_minimizer(model);
1161	// Go through gradient checkpoints to generate tensor inputs for backward pass just before executing the backward pass.
1162	ccv_cnnp_model_apply_gradient_checkpoints(compiled_data, model->graph);
1163	for (i = 0; i < output_size; i++)
1164	{
1165		const ccv_nnc_tensor_symbol_t df = ccv_nnc_tensor_symbol_for_backward(model->graph, compiled_data->f[i]);
1166		// Init this to 1 so we can backprop.
1167		ccv_nnc_tensor_symbol_set_flags(model->graph, df, CCV_NNC_TENSOR_SYMBOL_INIT_ONES);
1168	}
1169	compiled_data->backward.to_size = 0;
1170	for (i = 0; i < parameter_size_maybe_more; i++)
1171		if (compiled_data->gradients[i].d != CCV_NNC_NO_TENSOR_SYMBOL)
1172			compiled_data->backward.tos[compiled_data->backward.to_size++] = ccv_nnc_graph_exec_symbol_for_backward(model->graph, compiled_data->gradients[i]);
1173	ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS);
1174	ccv_nnc_symbolic_graph_set_destinations(model->graph, compiled_data->update_nodes, parameter_size);
1175	for (i = 0; i < parameter_size_maybe_more - parameter_size; i++)
1176	{
1177		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.
1178			continue;
1179		const ccv_nnc_graph_exec_symbol_t outgrad = ccv_nnc_graph_exec_symbol_for_backward(model->graph, compiled_data->outgrads[i]);
1180		const int* tos;
1181		int to_size;
1182		ccv_nnc_graph_exec_symbol_to(model->graph, outgrad, &tos, &to_size);
1183		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.
1184		{
1185			const ccv_nnc_graph_exec_symbol_t* destinations = ccv_nnc_symbolic_graph_destinations(model->graph);
1186			const int destination_count = ccv_nnc_symbolic_graph_destination_size(model->graph);
1187			int flag = 0;
1188			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; });
1189			for (j = i - 1; !flag && j >= 0; j--)
1190				if (j + outgrad_destination_start < destination_count)
1191					flag = (destinations[j + outgrad_destination_start].d == outgrad.d);
1192			if (!flag) // Only if we cannot find it, we add it.
1193				ccv_nnc_symbolic_graph_add_destination(model->graph, outgrad);
1194		}
1195	}
1196	if (parallel_count > 1)
1197	{
1198		ccv_nnc_symbolic_graph_data_parallel(model->graph, parallel_count,
1199			0, 0,
1200			compiled_data->gradients, parameter_size /* No need to deal with outgrads, we don't allreduce outgrads */,
1201			compiled_data->gradients /* We only care about gradients before allreduce, thus, update our current pointers */,
1202			0, 0, 0,
1203			CCV_NNC_PARALLEL_REDUCE_OP_SUM,
1204			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));
1205		ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
1206		for (i = 0; i < evaluate_to_size; i++)
1207			for (j = 1; j < parallel_count; j++)
1208			{
1209				const ccv_nnc_graph_exec_symbol_t copy = ccv_nnc_graph_exec_symbol_copy(model->graph, compiled_data->evaluate.tos[i], j);
1210				if (copy.d != CCV_NNC_NO_GRAPH_EXEC_SYMBOL)
1211					compiled_data->evaluate.tos[compiled_data->evaluate.to_size++] = copy;
1212			}
1213		const int backward_to_size = compiled_data->backward.to_size;
1214		for (i = 0; i < backward_to_size; i++)
1215			for (j = 1; j < parallel_count; j++)
1216			{
1217				const ccv_nnc_graph_exec_symbol_t copy = ccv_nnc_graph_exec_symbol_copy(model->graph, compiled_data->backward.tos[i], j);
1218				if (copy.d != CCV_NNC_NO_GRAPH_EXEC_SYMBOL)
1219					compiled_data->backward.tos[compiled_data->backward.to_size++] = copy;
1220			}
1221	}
1222	// Only use memory compression if we are in gradient parameter mode.
1223	if (gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES || gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES_AND_INPUTS)
1224	{
1225		if (model->memory_compression)
1226			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));
1227		if (model->memory_reduction)
1228			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));
1229	}
1230	compiled_data->backward.to_size = _ccv_nnc_array_dedup_graph_exec_symbols(compiled_data->backward.tos, compiled_data->backward.to_size);
1231	compiled_data->gradient_mode = gradient_mode;
1232}
1233 
1234void ccv_cnnp_model_tensors_init_0(const ccv_cnnp_model_t* const model, ccv_cnnp_compiled_data_t* const compiled_data)
1235{
1236	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", 1236, __extension__ __PRETTY_FUNCTION__
); }));
1237	const int parameter_size = compiled_data->parameters->rnum;
1238	const int parallel_count = _ccv_cnnp_model_effective_parallel_count(model);
1239	compiled_data->parallel_count = parallel_count;
1240	const int internal_size = compiled_data->internals->rnum;
1241	compiled_data->tensors_init.size = ccv_nnc_tensor_symbol_count(model->graph);
1242	compiled_data->tensors_init.v = cccalloccalloc(((compiled_data->tensors_init.size + 31) >> 5), sizeof(uint32_t));
1243	compiled_data->tensors.parameters = (ccv_nnc_tensor_t**)cccalloccalloc((parameter_size + internal_size) * parallel_count, sizeof(ccv_nnc_tensor_t*));
1244	compiled_data->tensors.internals = compiled_data->tensors.parameters + parameter_size * parallel_count;
1245}
1246 
1247int ccv_cnnp_model_tensors_any_to_alloc(const ccv_cnnp_model_t* const model, ccv_cnnp_compiled_data_t* const compiled_data)
1248{
1249	int i, j;
1250	const int parameter_size = compiled_data->parameters->rnum;
1251	const int parallel_count = _ccv_cnnp_compiled_data_parallel_count(model, compiled_data);
1252	const int internal_size = compiled_data->internals->rnum;
1253	for (i = 0; i < parameter_size; i++)
1254	{
1255		// parameters has to be allocated all together.
1256		if (compiled_data->tensors.parameters[i])
1257		{
1258			for (j = 1; j < parallel_count; j++)
1259				{ 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", 1259, __extension__ __PRETTY_FUNCTION__
); })); }
1260			continue;
1261		}
1262		return 1;
1263	}
1264	for (i = 0; i < internal_size; i++)
1265	{
1266		if (!compiled_data->tensors.internals[i])
1267			return 1;
1268		for (j = 1; j < parallel_count; j++)
1269			if (!compiled_data->tensors.internals[i + j * internal_size])
1270				return 1;
1271	}
1272	return 0;
1273}
1274 
1275void ccv_cnnp_model_tensors_init_1(const ccv_cnnp_model_t* const model, ccv_cnnp_compiled_data_t* const compiled_data)
1276{
1277	int i, j;
1278	const int parameter_size = compiled_data->parameters->rnum;
1279	const int parallel_count = _ccv_cnnp_compiled_data_parallel_count(model, compiled_data);
1280	compiled_data->parallel_count = parallel_count;
1281	const int internal_size = compiled_data->internals->rnum;
1282	for (i = 0; i < parameter_size; i++)
1283	{
1284		// parameters has to be allocated all together.
1285		if (compiled_data->tensors.parameters[i])
1286		{
1287			for (j = 1; j < parallel_count; j++)
1288				{ 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", 1288, __extension__ __PRETTY_FUNCTION__
); })); }
1289			continue;
1290		}
1291		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)));
1292		ccv_nnc_tensor_param_t info = ccv_nnc_tensor_symbol_params(parameter.graph, parameter);
1293		if (CCV_TENSOR_GET_DEVICE(info.type)((info.type) & 0xfff00) == CCV_COMPUTE_DEVICE_ANY)
1294			CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff
) << 8));
1295		const int device_id = CCV_TENSOR_GET_DEVICE_ID(info.type)(((info.type) & 0xfff00) >> 8);
1296		compiled_data->tensors.parameters[i] = ccv_nnc_tensor_new(0, info, 0);
1297		for (j = 1; j < parallel_count; j++)
1298		{
1299			if (j != device_id)
1300				CCV_TENSOR_SET_DEVICE_ID(info.type, j)(info.type) = (((info.type) & ~0xfff00) | (((j) & 0xfff
) << 8));
1301			else
1302				CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff
) << 8));
1303			compiled_data->tensors.parameters[i + j * parameter_size] = ccv_nnc_tensor_new(0, info, 0);
1304		}
1305	}
1306	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));
1307	for (i = 0; i < internal_size; i++)
1308	{
1309		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))
);
1310		const int d = retained.d;
1311		if (init_v[d >> 5] & (1u << (d & 0x1f)))
1312			continue;
1313		ccv_nnc_tensor_param_t info = ccv_nnc_tensor_symbol_params(retained.graph, retained);
1314		if (CCV_TENSOR_GET_DEVICE(info.type)((info.type) & 0xfff00) == CCV_COMPUTE_DEVICE_ANY)
1315			CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff
) << 8));
1316		const int device_id = CCV_TENSOR_GET_DEVICE_ID(info.type)(((info.type) & 0xfff00) >> 8);
1317		if (!compiled_data->tensors.internals[i])
1318			compiled_data->tensors.internals[i] = ccv_nnc_tensor_new(0, info, 0);
1319		for (j = 1; j < parallel_count; j++)
1320		{
1321			if (j != device_id)
1322				CCV_TENSOR_SET_DEVICE_ID(info.type, j)(info.type) = (((info.type) & ~0xfff00) | (((j) & 0xfff
) << 8));
1323			else
1324				CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff
) << 8));
1325			if (!compiled_data->tensors.internals[i + j * internal_size])
1326				compiled_data->tensors.internals[i + j * internal_size] = ccv_nnc_tensor_new(0, info, 0);
1327		}
1328	}
1329	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.
1330}
1331 
1332static void _ccv_cnnp_model_tensors_init(const ccv_cnnp_model_t* const model, ccv_cnnp_compiled_data_t* const compiled_data)
1333{
1334	ccv_cnnp_model_tensors_init_0(model, compiled_data);
1335	ccv_cnnp_model_tensors_init_1(model, compiled_data);
1336}
1337 
1338static 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)
1339{
1340	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", 1340, __extension__ __PRETTY_FUNCTION__
); }));
1341	int i, j;
1342	for (i = 0; i < tensor_size; i++)
1343	{
1344		if (!tensors[i])
1345			continue;
1346		const int d = tensor_symbols[i].d;
1347		if (!(tensors_init[d >> 5] & (1u << (d & 0x1f))))
1348			continue;
1349		for (j = 1; j < parallel_count; j++)
1350			if (tensors[i + j * tensor_size])
1351			{
1352				ccv_nnc_tensor_t* const input = CCV_NNC_TENSOR(tensors[i])((ccv_nnc_tensor_t*)((uintptr_t)(tensors[i]) & ~(uintptr_t
)1));
1353				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));
1354				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);
1355			}
1356	}
1357}
1358 
1359static 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)
1360{
1361	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", 1361, __extension__ __PRETTY_FUNCTION__
); }));
1362	int i, j;
1363	for (i = 0; i < tensor_size; i++)
1364	{
1365		const ccv_nnc_tensor_symbol_t tensor_symbol = tensor_symbols[i];
1366		for (j = 1; j < parallel_count; j++)
1367		{
1368			const ccv_nnc_tensor_symbol_t copy = ccv_nnc_tensor_symbol_copy(graph, tensor_symbol, j);
1369			ccv_nnc_tensor_t* copy_tensor = tensors[i + j * tensor_size];
1370			if (copy_tensor && copy.d == CCV_NNC_NO_TENSOR_SYMBOL)
1371			{ // We shouldn't allocate this, free it up.
1372				ccv_nnc_tensor_free(tensors[i + j * tensor_size]);
1373				tensors[i + j * tensor_size] = 0;
1374			}
1375		}
1376	}
1377}
1378 
1379static 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)
1380{
1381	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", 1381, __extension__ __PRETTY_FUNCTION__
); }));
1382	int i, j;
1383	for (i = 0; i < tensor_size; i++)
1384	{
1385		ccv_nnc_tensor_symbol_t tensor_symbol = tensor_symbols[i];
1386		if (tensor_symbol.d == CCV_NNC_NO_TENSOR_SYMBOL)
1387			continue;
1388		if (graph)
1389		{
1390			const ccv_nnc_tensor_symbol_t alias_to = ccv_nnc_tensor_symbol_alias_to(graph, tensor_symbol);
1391			if (alias_to.d != CCV_NNC_NO_TENSOR_SYMBOL)
1392				tensor_symbol = alias_to;
1393		}
1394		ccv_nnc_tensor_t* const tensor = CCV_NNC_TENSOR(tensors[i])((ccv_nnc_tensor_t*)((uintptr_t)(tensors[i]) & ~(uintptr_t
)1));
1395		if (tensor && tensor_symbol.d != CCV_NNC_NO_TENSOR_SYMBOL)
1396		{
1397			const ccv_nnc_tensor_bind_t retained_bind = {
1398				.symbol = tensor_symbol,
1399				.tensor = tensor
1400			};
1401			ccv_array_push(tensor_binds, &retained_bind);
1402		}
1403		for (j = 1; j < parallel_count; j++)
1404		{
1405			const ccv_nnc_tensor_symbol_t copy = ccv_nnc_tensor_symbol_copy(graph, tensor_symbol, j);
1406			ccv_nnc_tensor_t* copy_tensor = tensors[i + j * tensor_size];
1407			if (copy_tensor && copy.d != CCV_NNC_NO_TENSOR_SYMBOL)
1408			{
1409				const ccv_nnc_tensor_bind_t bind = {
1410					.symbol = copy,
1411					.tensor = tensors[i + j * tensor_size]
1412				};
1413				ccv_array_push(tensor_binds, &bind);
1414			}
1415		}
1416	}
1417}
1418 
1419static void _ccv_cnnp_compiled_data_graph_free(ccv_cnnp_compiled_data_t* const compiled_data)
1420{
1421	if (compiled_data->graph)
1422		ccv_nnc_graph_free(compiled_data->graph);
1423	compiled_data->graph = 0;
1424	compiled_data->is_test = 0;
1425	if (compiled_data->tensor_arena)
1426		ccv_nnc_tensor_arena_free(compiled_data->tensor_arena);
1427	compiled_data->tensor_arena = 0;
1428	if (compiled_data->graph_exec_arena)
1429		ccv_nnc_graph_exec_arena_free(compiled_data->graph_exec_arena);
1430	compiled_data->graph_exec_arena = 0;
1431	if (compiled_data->backward.from_ops)
1432		ccfreefree(compiled_data->backward.from_ops);
1433	compiled_data->backward.from_ops = 0;
1434	if (compiled_data->evaluate.schedule)
1435		ccv_nnc_graph_static_schedule_free(compiled_data->evaluate.schedule);
1436	compiled_data->evaluate.schedule = 0;
1437	if (compiled_data->backward.schedule)
1438		ccv_nnc_graph_static_schedule_free(compiled_data->backward.schedule);
1439	compiled_data->backward.schedule = 0;
1440}
1441 
1442static void _ccv_cnnp_compiled_data_gradient_free(ccv_cnnp_compiled_data_t* const compiled_data)
1443{
1444	if (compiled_data->gradients)
1445		ccfreefree(compiled_data->gradients);
1446	compiled_data->gradients = 0;
1447	if (compiled_data->updated_parameters)
1448		ccfreefree(compiled_data->updated_parameters);
1449	compiled_data->updated_parameters = 0;
1450	compiled_data->update_nodes = 0;
1451	compiled_data->saved_aux = 0;
1452}
1453 
1454static void _ccv_cnnp_compiled_data_backward_free(ccv_cnnp_compiled_data_t* const compiled_data)
1455{
1456	if (compiled_data->backward.gradients)
1457		ccfreefree(compiled_data->backward.gradients);
1458	compiled_data->backward.gradients = 0;
1459	if (compiled_data->backward.accum)
1460		ccv_nnc_graph_free(compiled_data->backward.accum);
1461	compiled_data->backward.accum = 0;
1462	if (compiled_data->backward.tensor_arena)
1463		ccv_nnc_tensor_arena_free(compiled_data->backward.tensor_arena);
1464	compiled_data->backward.tensor_arena = 0;
1465	if (compiled_data->backward.graph_exec_arena)
1466		ccv_nnc_graph_exec_arena_free(compiled_data->backward.graph_exec_arena);
1467	compiled_data->backward.graph_exec_arena = 0;
1468}
1469 
1470static void _ccv_cnnp_compiled_data_apply_gradients_free(ccv_cnnp_compiled_data_t* const compiled_data)
1471{
1472	if (compiled_data->apply_gradients.graph)
1473		ccv_nnc_graph_free(compiled_data->apply_gradients.graph);
1474	compiled_data->apply_gradients.graph = 0;
1475	if (compiled_data->apply_gradients.tensor_arena)
1476		ccv_nnc_tensor_arena_free(compiled_data->apply_gradients.tensor_arena);
1477	compiled_data->apply_gradients.tensor_arena = 0;
1478	if (compiled_data->apply_gradients.graph_exec_arena)
1479		ccv_nnc_graph_exec_arena_free(compiled_data->apply_gradients.graph_exec_arena);
1480	compiled_data->apply_gradients.graph_exec_arena = 0;
1481}
1482 
1483// Compile the graph to run ccv_cnnp_model_fit
1484static 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)
1485{
1486	int i, j;
1487	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1488	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", 1488, __extension__ __PRETTY_FUNCTION__
); }));
1489	compiled_data->graph_mode = CCV_CNNP_MODEL_GRAPH_FIT_MODE;
1490	const int parallel_count = _ccv_cnnp_model_root_parallel_count(model);
1491	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", 1491, __extension__ __PRETTY_FUNCTION__
); }));
1492	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", 1492
, __extension__ __PRETTY_FUNCTION__); }));
1493	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", 1493, __extension__ __PRETTY_FUNCTION__
); }));
1494	if (compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_NONE)
1495	{
1496		_ccv_cnnp_model_set_rewindables(model);
1497		_ccv_cnnp_model_gradient_init(model, CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES, CCV_CNNP_DISABLE_OUTGRAD_ALL, fits, fit_size);
1498	} else if (compiled_data->gradient_mode != CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES) {
1499		_ccv_cnnp_model_rewind_graph(model);
1500		_ccv_cnnp_compiled_data_gradient_free(compiled_data);
1501		compiled_data->gradient_mode = CCV_CNNP_COMPILED_DATA_GRADIENT_NONE;
1502		_ccv_cnnp_model_gradient_init(model, CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES, CCV_CNNP_DISABLE_OUTGRAD_ALL, fits, fit_size);
1503	}
1504	const int tensors_init = !!compiled_data->tensors_init.v;
1505	if (!tensors_init)
1506		_ccv_cnnp_model_tensors_init(model, compiled_data);
1507	else if ((uintptr_t)compiled_data->tensors_init.v & (uintptr_t)1)
1508	// Check if it is not fully allocated, if it is not, init_1.
1509		ccv_cnnp_model_tensors_init_1(model, compiled_data);
1510	ccv_array_t* const tensor_binds = ccv_array_new(sizeof(ccv_nnc_tensor_bind_t), 0, 0);
1511	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"
, 1511, __extension__ __PRETTY_FUNCTION__); }));
1512	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"
, 1512, __extension__ __PRETTY_FUNCTION__); }));
1513	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", 1513
, __extension__ __PRETTY_FUNCTION__); }));
1514	const int input_size_per_p = input_size / parallel_count;
1515	_ccv_cnnp_model_bind_tensors(model->graph, model->inputs, inputs, input_size_per_p, parallel_count, tensor_binds);
1516	const int output_size_per_p = output_size / parallel_count;
1517	_ccv_cnnp_model_bind_tensors(model->graph, model->outputs, outputs, output_size_per_p, parallel_count, tensor_binds);
1518	const int fit_size_per_p = fit_size / parallel_count;
1519	_ccv_cnnp_model_bind_tensors(model->graph, compiled_data->fits, fits, fit_size_per_p, parallel_count, tensor_binds);
1520	const int parameter_size = compiled_data->parameters->rnum;
1521	_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);
1522	_ccv_cnnp_model_bind_tensors(model->graph, compiled_data->updated_parameters, compiled_data->tensors.parameters, parameter_size, parallel_count, tensor_binds);
1523	const int internal_size = compiled_data->internals->rnum;
1524	_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);
1525	_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);
1526	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);
1527	ccv_array_free(tensor_binds);
1528	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));
1529	if (tensors_init && parallel_count > 1)
1530		_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);
1531	// If tensor is not init'ed, we need to init states first.
1532	if (_ccv_cnnp_any_to_init(compiled_data))
1533	{
1534		ccv_nnc_tensor_init_states_t tensor_init_states = {
1535			.parallel_count = parallel_count,
1536			.graph = model->graph,
1537			.compiled_data = compiled_data,
1538			.tensor_arena = compiled_data->tensor_arena
1539		};
1540		ccv_cnnp_model_init_states(model, model->graph, _ccv_cnnp_init_states_for_tensors, &tensor_init_states);
1541	}
1542	compiled_data->is_test = 0;
1543	const int saved_aux_size = ccv_nnc_minimizer_saved_aux_size(compiled_data->minimize.minimizer);
1544	// No need to set because it is default to training mode.
1545	// ccv_cnnp_model_set_is_test(model, 0, _ccv_cnnp_cmd_update_for_execs, &update);
1546	for (i = 0; i < saved_aux_size * parameter_size; i++)
1547	{
1548		if (compiled_data->saved_aux[i].source.d == CCV_NNC_NO_TENSOR_SYMBOL)
1549			continue;
1550		ccv_nnc_tensor_t* const tensor = ccv_nnc_tensor_from_symbol(compiled_data->tensor_arena, compiled_data->saved_aux[i].source);
1551		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);
1552		for (j = 1; j < parallel_count; j++)
1553		{
1554			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));
1555			if (copy)
1556				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);
1557		}
1558	}
1559	const int evaluate_to_size = compiled_data->evaluate.to_size;
1560	compiled_data->evaluate.to_op_size = 0;
1561	for (i = 0; i < evaluate_to_size; i++)
1562	{
1563		ccv_nnc_graph_exec_t const to = ccv_nnc_graph_exec_from_symbol(compiled_data->graph_exec_arena, compiled_data->evaluate.tos[i]);
1564		if (to.graph)
1565			compiled_data->evaluate.to_ops[compiled_data->evaluate.to_op_size++] = to;
1566	}
1567	ccv_nnc_graph_set_default_static_schedule(compiled_data->graph, compiled_data->stream_type, model->max_stream_count);
1568	ccv_nnc_graph_autotune(compiled_data->graph, model->workspace_size, 0, TRAVERSE_FULL0,0,0,0);
1569}
1570 
1571ccv_nnc_stream_context_t* ccv_cnnp_model_default_stream(const ccv_cnnp_model_t* const model)
1572{
1573	const ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1574	if (!compiled_data || !compiled_data->graph)
1575		return 0;
1576	return ccv_nnc_graph_default_stream(compiled_data->graph);
1577}
1578 
1579uint64_t ccv_cnnp_model_memory_size(const ccv_cnnp_model_t* const model)
1580{
1581	const ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1582	if (!compiled_data || !compiled_data->tensor_arena)
1583		return 0;
1584	return ccv_nnc_tensor_arena_size(compiled_data->tensor_arena);
1585}
1586 
1587static 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)
1588{
1589	int i, j;
1590	for (i = 0; i < tensor_size; i++)
1591	{
1592		ccv_nnc_tensor_symbol_t tensor_symbol = tensor_symbols[i];
1593		if (tensor_symbol.d == CCV_NNC_NO_TENSOR_SYMBOL)
1594			continue;
1595		if (graph)
1596		{
1597			const ccv_nnc_tensor_symbol_t alias_to = ccv_nnc_tensor_symbol_alias_to(graph, tensor_symbol);
1598			if (alias_to.d != CCV_NNC_NO_TENSOR_SYMBOL)
1599				tensor_symbol = alias_to;
1600		}
1601		ccv_nnc_tensor_bind_symbol(tensor_arena, tensor_symbol, tensors[i]);
1602		for (j = 1; j < parallel_count; j++)
1603		{
1604			const ccv_nnc_tensor_symbol_t copy = ccv_nnc_tensor_symbol_copy(graph, tensor_symbol, j);
1605			if (copy.d != CCV_NNC_NO_TENSOR_SYMBOL)
1606				ccv_nnc_tensor_bind_symbol(tensor_arena, copy, tensors[i + tensor_size * j]);
1607		}
1608	}
1609}
1610 
1611void 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)
1612{
1613	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1614	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 1614, __extension__ __PRETTY_FUNCTION__); }));
1615	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; });
1616	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", 1616, __extension__ __PRETTY_FUNCTION__
); }));
1617	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", 1617, __extension__ __PRETTY_FUNCTION__
); }));
1618	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", 1618
, __extension__ __PRETTY_FUNCTION__); }));
1619	assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if
 (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c"
, 1619, __extension__ __PRETTY_FUNCTION__); }));
1620	if (!compiled_data->graph || compiled_data->graph_mode != CCV_CNNP_MODEL_GRAPH_FIT_MODE)
1621	{
1622		_ccv_cnnp_compiled_data_graph_free(compiled_data);
1623		_ccv_cnnp_compiled_data_backward_free(compiled_data);
1624		_ccv_cnnp_compiled_data_apply_gradients_free(compiled_data);
1625		// Compile the symbolic graph down only when needed.
1626		_ccv_cnnp_model_fit_jit(model, inputs, input_size, fits, fit_size, outputs, output_size);
1627	} else {
1628		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"
, 1628, __extension__ __PRETTY_FUNCTION__); }));
1629		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"
, 1629, __extension__ __PRETTY_FUNCTION__); }));
1630		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", 1630
, __extension__ __PRETTY_FUNCTION__); }));
1631		const int input_size_per_p = input_size / parallel_count;
1632		_ccv_cnnp_bind_tensors_to_arena(compiled_data->tensor_arena, model->graph, model->inputs, inputs, input_size_per_p, parallel_count);
1633		const int output_size_per_p = output_size / parallel_count;
1634		_ccv_cnnp_bind_tensors_to_arena(compiled_data->tensor_arena, model->graph, model->outputs, outputs, output_size_per_p, parallel_count);
1635		const int fit_size_per_p = fit_size / parallel_count;
1636		_ccv_cnnp_bind_tensors_to_arena(compiled_data->tensor_arena, model->graph, compiled_data->fits, fits, fit_size_per_p, parallel_count);
1637	}
1638	if (compiled_data->is_test)
1639	{
1640		compiled_data->is_test = 0;
1641		ccv_nnc_graph_exec_update_t update = {
1642			.parallel_count = parallel_count,
1643			.graph = model->graph,
1644			.graph_exec_arena = compiled_data->graph_exec_arena,
1645		};
1646		ccv_cnnp_model_set_is_test(model, 0, _ccv_cnnp_cmd_update_for_execs, &update);
1647	}
1648	ccv_nnc_graph_run_with_schedule(compiled_data->graph, 0, 0, tensor_tape, stream_context);
1649}
1650 
1651// Compile the graph to run ccv_cnnp_model_evaluate with require_grad = false (MULTISTAGE_MODE_NO_GRAD).
1652static 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)
1653{
1654	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1655	compiled_data->graph_mode = CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE_NO_GRAD;
1656	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; });
1657	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", 1657, __extension__ __PRETTY_FUNCTION__
); }));
1658	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", 1658, __extension__ __PRETTY_FUNCTION__
); }));
1659	// If the gradient is not initialized, continue to setup parallel process. We don't init gradient here, but rather,
1660	// we setup proper rewindables so the graph can be rewinded to previous state before we run data parallel.
1661	if (parallel_count > 1 && compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_NONE)
1662	{
1663		const int evaluate_to_size = compiled_data->evaluate.to_size;
1664		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);
1665		_ccv_cnnp_model_set_rewindables(model);
1666		ccv_nnc_symbolic_graph_data_parallel(model->graph, parallel_count,
1667			0, 0,
1668			0, 0, 0,
1669			0, 0, 0,
1670			CCV_NNC_PARALLEL_REDUCE_OP_SUM,
1671			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));
1672		ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
1673		int i, j;
1674		for (i = 0; i < evaluate_to_size; i++)
1675			for (j = 1; j < parallel_count; j++)
1676			{
1677				const ccv_nnc_graph_exec_symbol_t copy = ccv_nnc_graph_exec_symbol_copy(model->graph, compiled_data->evaluate.tos[i], j);
1678				if (copy.d != CCV_NNC_NO_GRAPH_EXEC_SYMBOL)
1679					compiled_data->evaluate.tos[compiled_data->evaluate.to_size++] = copy;
1680			}
1681	}
1682	const int tensors_init = !!compiled_data->tensors_init.v;
1683	if (!tensors_init)
1684		_ccv_cnnp_model_tensors_init(model, compiled_data);
1685	else if ((uintptr_t)compiled_data->tensors_init.v & (uintptr_t)1)
1686	// Check if it is not fully allocated, if it is not, init_1.
1687		ccv_cnnp_model_tensors_init_1(model, compiled_data);
1688	const int tensor_parallel_count = _ccv_cnnp_compiled_data_parallel_count(model, compiled_data);
1689	ccv_array_t* const tensor_binds = ccv_array_new(sizeof(ccv_nnc_tensor_bind_t), 0, 0);
1690	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"
, 1690, __extension__ __PRETTY_FUNCTION__); }));
1691	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"
, 1691, __extension__ __PRETTY_FUNCTION__); }));
1692	const int input_size_per_p = input_size / parallel_count;
1693	_ccv_cnnp_model_bind_tensors(model->graph, model->inputs, inputs, input_size_per_p, parallel_count, tensor_binds);
1694	const int output_size_per_p = output_size / parallel_count;
1695	_ccv_cnnp_model_bind_tensors(model->graph, model->outputs, outputs, output_size_per_p, parallel_count, tensor_binds);
1696	const int parameter_size = compiled_data->parameters->rnum;
1697	_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, tensor_parallel_count, tensor_binds);
1698	const int internal_size = compiled_data->internals->rnum;
1699	_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, tensor_parallel_count);
1700	_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, tensor_parallel_count, tensor_binds);
1701	// If we generated gradient for the graph, only compile part of the graph because the rest is irrelevant for evaluation.
1702	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);
1703	ccv_array_free(tensor_binds);
1704	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));
1705	// If tensor is not init'ed, we need to init states first.
1706	if (tensors_init && tensor_parallel_count > 1)
1707		_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, tensor_parallel_count);
1708	if (_ccv_cnnp_any_to_init(compiled_data))
1709	{
1710		ccv_nnc_tensor_init_states_t tensor_init_states = {
1711			.parallel_count = tensor_parallel_count,
1712			.graph = model->graph,
1713			.compiled_data = compiled_data,
1714			.tensor_arena = compiled_data->tensor_arena
1715		};
1716		ccv_cnnp_model_init_states(model, model->graph, _ccv_cnnp_init_states_for_tensors, &tensor_init_states);
1717	}
1718	compiled_data->is_test = 1;
1719	ccv_nnc_graph_exec_update_t update = {
1720		.parallel_count = parallel_count,
1721		.graph = model->graph,
1722		.graph_exec_arena = compiled_data->graph_exec_arena,
1723	};
1724	ccv_cnnp_model_set_is_test(model, 1, _ccv_cnnp_cmd_update_for_execs, &update);
1725	ccv_nnc_graph_set_default_static_schedule(compiled_data->graph, compiled_data->stream_type, model->max_stream_count);
1726	ccv_nnc_graph_autotune(compiled_data->graph, model->workspace_size, 0, TRAVERSE_FULL0,0,0,0);
1727}
1728 
1729static void _ccv_cnnp_model_gradient_tensors_init(const ccv_cnnp_model_t* const model, ccv_cnnp_compiled_data_t* const compiled_data)
1730{
1731	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", 1731, __extension__ __PRETTY_FUNCTION__
); }));
1732	const int parameter_size = compiled_data->parameters->rnum;
1733	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; });
1734	compiled_data->tensors.gradients = (ccv_nnc_tensor_t**)ccmallocmalloc(sizeof(ccv_nnc_tensor_t*) * parameter_size * 2 * parallel_count);
1735	compiled_data->tensors.accum_gradients = compiled_data->tensors.gradients + parameter_size * parallel_count;
1736	int i, j;
1737	for (i = 0; i < parameter_size; i++)
1738	{
1739		if (compiled_data->parameter_flags && !(compiled_data->parameter_flags[i >> 6] & ((uint64_t)1 << (i & 63))))
1740		{
1741			compiled_data->tensors.gradients[i] = 0;
1742			compiled_data->tensors.accum_gradients[i] = 0;
1743			for (j = 1; j < parallel_count; j++)
1744			{
1745				compiled_data->tensors.gradients[i + j * parameter_size] = 0;
1746				compiled_data->tensors.accum_gradients[i + j * parameter_size] = 0;
1747			}
1748			continue;
1749		}
1750		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)));
1751		ccv_nnc_tensor_param_t info = ccv_nnc_tensor_symbol_params(parameter.graph, parameter);
1752		if (CCV_TENSOR_GET_DEVICE(info.type)((info.type) & 0xfff00) == CCV_COMPUTE_DEVICE_ANY)
1753			CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff
) << 8));
1754		const int device_id = CCV_TENSOR_GET_DEVICE_ID(info.type)(((info.type) & 0xfff00) >> 8);
1755		compiled_data->tensors.gradients[i] = ccv_nnc_tensor_new(0, info, 0);
1756		compiled_data->tensors.accum_gradients[i] = 0; // delay the accumulated gradient allocation until when we need it.
1757		for (j = 1; j < parallel_count; j++)
1758		{
1759			if (j != device_id)
1760				CCV_TENSOR_SET_DEVICE_ID(info.type, j)(info.type) = (((info.type) & ~0xfff00) | (((j) & 0xfff
) << 8));
1761			else
1762				CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff
) << 8));
1763			compiled_data->tensors.gradients[i + j * parameter_size] = ccv_nnc_tensor_new(0, info, 0);
1764			compiled_data->tensors.accum_gradients[i + j * parameter_size] = 0;
1765		}
1766	}
1767}
1768 
1769static int _ccv_cnnp_is_disable_outgrad_all(const uint64_t disable_outgrad, const int input_size)
1770{
1771	if (disable_outgrad == CCV_CNNP_DISABLE_OUTGRAD_ALL)
1772		return 1;
1773	if (disable_outgrad == CCV_CNNP_DISABLE_OUTGRAD_NONE)
1774		return 0;
1775	int i;
1776	for (i = 0; i < input_size; i++)
1777		if (!(disable_outgrad & ((uint64_t)1 << i)))
1778			return 0;
1779	return 1;
1780}
1781 
1782// Compile the graph to run ccv_cnnp_model_evaluate with requires_grad = true (MULTISTAGE_MODE).
1783// Particularly, this method compiles the evaluation and backprop graph (the main graph).
1784static 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)
1785{
1786	int i, j;
1787	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1788	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;
1789	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", 1789, __extension__ __PRETTY_FUNCTION__
); }));
1790	compiled_data->graph_mode = CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE;
1791	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; });
1792	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", 1792, __extension__ __PRETTY_FUNCTION__
); }));
1793	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", 1793, __extension__ __PRETTY_FUNCTION__
); }));
1794	// There shouldn't be a loss function if we evaluate with multistage jit.
1795	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", 1795, __extension__ __PRETTY_FUNCTION__
); }));
1796	if (compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_NONE)
1797	{
1798		_ccv_cnnp_model_set_rewindables(model);
1799		_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.
1800	} else if (compiled_data->gradient_mode != target_gradient_mode) {
1801		_ccv_cnnp_model_rewind_graph(model);
1802		_ccv_cnnp_compiled_data_gradient_free(compiled_data);
1803		compiled_data->gradient_mode = CCV_CNNP_COMPILED_DATA_GRADIENT_NONE;
1804		_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.
1805	}
1806	const int tensors_init = !!compiled_data->tensors_init.v;
1807	if (!tensors_init)
1808		_ccv_cnnp_model_tensors_init(model, compiled_data);
1809	else if ((uintptr_t)compiled_data->tensors_init.v & (uintptr_t)1)
1810	// Check if it is not fully allocated, if it is not, init_1.
1811		ccv_cnnp_model_tensors_init_1(model, compiled_data);
1812	ccv_array_t* const tensor_binds = ccv_array_new(sizeof(ccv_nnc_tensor_bind_t), 0, 0);
1813	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"
, 1813, __extension__ __PRETTY_FUNCTION__); }));
1814	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"
, 1814, __extension__ __PRETTY_FUNCTION__); }));
1815	const int input_size_per_p = input_size / parallel_count;
1816	_ccv_cnnp_model_bind_tensors(model->graph, model->inputs, inputs, input_size_per_p, parallel_count, tensor_binds);
1817	const int output_size_per_p = output_size / parallel_count;
1818	_ccv_cnnp_model_bind_tensors(model->graph, model->outputs, outputs, output_size_per_p, parallel_count, tensor_binds);
1819	const int parameter_size = compiled_data->parameters->rnum;
1820	_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);
1821	const int internal_size = compiled_data->internals->rnum;
1822	_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);
1823	_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);
1824	if (!compiled_data->tensors.gradients)
1825		_ccv_cnnp_model_gradient_tensors_init(model, compiled_data);
1826	_ccv_cnnp_model_bind_tensors(model->graph, compiled_data->gradients, compiled_data->tensors.gradients, parameter_size, parallel_count, tensor_binds);
1827	if (compiled_data->backward.to_size > 0)
1828		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);
1829	else
1830		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);
1831	ccv_array_free(tensor_binds);
1832	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));
1833	if (tensors_init && parallel_count > 1)
1834		_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);
1835	// If tensor is not init'ed, we need to init states first.
1836	if (_ccv_cnnp_any_to_init(compiled_data))
1837	{
1838		ccv_nnc_tensor_init_states_t tensor_init_states = {
1839			.parallel_count = parallel_count,
1840			.graph = model->graph,
1841			.compiled_data = compiled_data,
1842			.tensor_arena = compiled_data->tensor_arena
1843		};
1844		ccv_cnnp_model_init_states(model, model->graph, _ccv_cnnp_init_states_for_tensors, &tensor_init_states);
1845	}
1846	compiled_data->is_test = is_test;
1847	ccv_nnc_graph_exec_update_t update = {
1848		.parallel_count = parallel_count,
1849		.graph = model->graph,
1850		.graph_exec_arena = compiled_data->graph_exec_arena,
1851	};
1852	ccv_cnnp_model_set_is_test(model, is_test, _ccv_cnnp_cmd_update_for_execs, &update);
1853	const int evaluate_to_size = compiled_data->evaluate.to_size;
1854	compiled_data->evaluate.to_op_size = 0;
1855	ccv_array_t* const backward_from = ccv_array_new(sizeof(int), 0, 0);
1856	for (i = 0; i < evaluate_to_size; i++)
1857	{
1858		ccv_nnc_graph_exec_t const to_op = ccv_nnc_graph_exec_from_symbol(compiled_data->graph_exec_arena, compiled_data->evaluate.tos[i]);
1859		if (to_op.graph)
1860			compiled_data->evaluate.to_ops[compiled_data->evaluate.to_op_size++] = to_op;
1861		const int* tos;
1862		int to_size;
1863		ccv_nnc_graph_exec_symbol_to(model->graph, compiled_data->evaluate.tos[i], &tos, &to_size);
1864		for (j = 0; j < to_size; j++)
1865		{
1866			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){
1867				.d = tos[j],
1868				.graph = model->graph
1869			});
1870			if (to_op.graph)
1871				ccv_array_add_unique_int(backward_from, to_op.d);
1872		}
1873	}
1874	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", 1874, __extension__
 __PRETTY_FUNCTION__); }));
1875	compiled_data->backward.from_op_size = backward_from->rnum;
1876	compiled_data->backward.from_ops = (ccv_nnc_graph_exec_t*)ccmallocmalloc(sizeof(ccv_nnc_graph_exec_t) * backward_from->rnum);
1877	for (i = 0; i < backward_from->rnum; i++)
1878		compiled_data->backward.from_ops[i] = (ccv_nnc_graph_exec_t){
1879			.d = *(int*)ccv_array_get(backward_from, i)((void*)(((char*)((backward_from)->data)) + (size_t)(backward_from
)->rsize * (size_t)(i))),
1880			.graph = compiled_data->graph,
1881		};
1882	// If there are any set node (to set some tensors to 0) inserted through backward pass, these won't be executed if we just do sources -> evaluate.to_ops, backward.from_ops -> destinations. We need this logic to find out these nodes and explicitly adding them to backward.from_ops.
1883	ccv_nnc_graph_exec_info_t* const exec_info = (ccv_nnc_graph_exec_info_t*)ccv_array_get(compiled_data->graph->exec_info, 0)((void*)(((char*)((compiled_data->graph->exec_info)->
data)) + (size_t)(compiled_data->graph->exec_info)->
rsize * (size_t)(0)));
1884	const int exec_info_size = compiled_data->graph->exec_info->rnum;
1885	uint32_t* const visited = cccalloccalloc((exec_info_size + 31) >> 5, sizeof(uint32_t));
1886	const ccv_nnc_graph_exec_t* const sources = (ccv_nnc_graph_exec_t*)ccv_array_get(compiled_data->graph->sources, 0)((void*)(((char*)((compiled_data->graph->sources)->data
)) + (size_t)(compiled_data->graph->sources)->rsize *
 (size_t)(0)));
1887	const int source_size = compiled_data->graph->sources->rnum;
1888	ccv_nnc_graph_visit_t* visit = ccv_nnc_graph_visit_new(compiled_data->graph, exec_info, exec_info_size, sources, source_size, compiled_data->evaluate.to_ops, compiled_data->evaluate.to_op_size, 0)({ ccv_nnc_graph_visit_t* _visit_ = (ccv_nnc_graph_visit_t*)malloc
(sizeof(ccv_nnc_graph_visit_t) + sizeof(_visit_->node[0]) *
 ((exec_info_size) - 1)); _visit_->size = 0; do { typedef struct
 { int8_t d; int8_t r; uint16_t c; int32_t edges; } ccv_nnc_incoming_t
; int _i_, _j_; int _incoming_edges_ = 0; for (_i_ = 0; _i_ <
 (exec_info_size); _i_++) _incoming_edges_ += ((exec_info)[_i_
].outgoings) ? (exec_info)[_i_].outgoings->rnum : 0; const
 int _heap_mem_ = ((exec_info_size) + _incoming_edges_ > 1024
); ccv_nnc_incoming_t* _incomings_; if (_heap_mem_) _incomings_
 = (ccv_nnc_incoming_t*)malloc(sizeof(ccv_nnc_incoming_t) * (
exec_info_size) + sizeof(int32_t) * ((exec_info_size) * 2 + _incoming_edges_
)); else _incomings_ = (ccv_nnc_incoming_t*)__builtin_alloca (
sizeof(ccv_nnc_incoming_t) * (exec_info_size) + sizeof(int32_t
) * ((exec_info_size) * 2 + _incoming_edges_)); memset(_incomings_
, 0, sizeof(ccv_nnc_incoming_t) * (exec_info_size)); int32_t*
 _exists_[2] = { (int32_t*)(_incomings_ + (exec_info_size)), (
int32_t*)(_incomings_ + (exec_info_size)) + (exec_info_size),
 }; int32_t* const _edges_ = _exists_[1] + (exec_info_size); for
 (_i_ = 0; _i_ < (source_size); _i_++) { ((void) sizeof ((
(sources)[_i_].graph == compiled_data->graph) ? 1 : 0), __extension__
 ({ if ((sources)[_i_].graph == compiled_data->graph) ; else
 __assert_fail ("(sources)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1888, __extension__ __PRETTY_FUNCTION__
); })); _incomings_[(sources)[_i_].d].r = 1; _exists_[0][_i_]
 = (sources)[_i_].d; } int _exist_size_[2] = { (source_size),
 0, }; int _p_ = 0, _q_ = 1; while (_exist_size_[_p_] > 0)
 { _exist_size_[_q_] = 0; for (_i_ = 0; _i_ < _exist_size_
[_p_]; _i_++) { const int32_t _idx_ = _exists_[_p_][_i_]; if (
_incomings_[_idx_].r != 1) continue; _incomings_[_idx_].r = 2
; if ((exec_info)[_idx_].outgoings) for (_j_ = 0; _j_ < (exec_info
)[_idx_].outgoings->rnum; _j_++) { const int d = *(int*)((
void*)(((char*)(((exec_info)[_idx_].outgoings)->data)) + (
size_t)((exec_info)[_idx_].outgoings)->rsize * (size_t)(_j_
))); ++_incomings_[d].c; if (_incomings_[d].r != 0) continue;
 _incomings_[d].r = 1; ((void) sizeof ((_exist_size_[_q_] <
 (exec_info_size)) ? 1 : 0), __extension__ ({ if (_exist_size_
[_q_] < (exec_info_size)) ; else __assert_fail ("_exist_size_[_q_] < (exec_info_size)"
, "ccv_cnnp_model.c", 1888, __extension__ __PRETTY_FUNCTION__
); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[
_q_]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for
 (_i_ = 0; _i_ < (source_size); _i_++) { ((void) sizeof ((
(sources)[_i_].graph == compiled_data->graph) ? 1 : 0), __extension__
 ({ if ((sources)[_i_].graph == compiled_data->graph) ; else
 __assert_fail ("(sources)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1888, __extension__ __PRETTY_FUNCTION__
); })); _incomings_[(sources)[_i_].d].r = 3; _exists_[0][_i_]
 = (sources)[_i_].d; } _exist_size_[0] = (source_size); _exist_size_
[1] = 0; _p_ = 0, _q_ = 1; int _bump_ = 1; while (_exist_size_
[_p_] > 0) { _exist_size_[_q_] = 0; for (_i_ = 0; _i_ <
 _exist_size_[_p_]; _i_++) { const int32_t _idx_ = _exists_[_p_
][_i_]; if (_incomings_[_idx_].r != 3) continue; _incomings_[
_idx_].r = 4; if ((exec_info)[_idx_].outgoings) for (_j_ = 0;
 _j_ < (exec_info)[_idx_].outgoings->rnum; _j_++) { const
 int d = *(int*)((void*)(((char*)(((exec_info)[_idx_].outgoings
)->data)) + (size_t)((exec_info)[_idx_].outgoings)->rsize
 * (size_t)(_j_))); if (_incomings_[d].edges == 0) { _incomings_
[d].edges = _bump_; _bump_ += _incomings_[d].c; _incomings_[d
].c = 0; } _edges_[_incomings_[d].edges - 1 + _incomings_[d].
c] = _idx_; ++_incomings_[d].c; if (_incomings_[d].r != 2) continue
; _incomings_[d].r = 3; ((void) sizeof ((_exist_size_[_q_] <
 (exec_info_size)) ? 1 : 0), __extension__ ({ if (_exist_size_
[_q_] < (exec_info_size)) ; else __assert_fail ("_exist_size_[_q_] < (exec_info_size)"
, "ccv_cnnp_model.c", 1888, __extension__ __PRETTY_FUNCTION__
); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[
_q_]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for
 (_i_ = 0; _i_ < (compiled_data->evaluate.to_op_size); _i_
++) { ((void) sizeof (((compiled_data->evaluate.to_ops)[_i_
].graph == compiled_data->graph) ? 1 : 0), __extension__ (
{ if ((compiled_data->evaluate.to_ops)[_i_].graph == compiled_data
->graph) ; else __assert_fail ("(compiled_data->evaluate.to_ops)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1888, __extension__ __PRETTY_FUNCTION__
); })); _incomings_[(compiled_data->evaluate.to_ops)[_i_].
d].r = 5; _exists_[0][_i_] = (compiled_data->evaluate.to_ops
)[_i_].d; } _exist_size_[0] = (compiled_data->evaluate.to_op_size
); _exist_size_[1] = 0; _p_ = 0, _q_ = 1; while (_exist_size_
[_p_] > 0) { _exist_size_[_q_] = 0; for (_i_ = 0; _i_ <
 _exist_size_[_p_]; _i_++) { const int32_t _idx_ = _exists_[_p_
][_i_]; if (_incomings_[_idx_].r != 5) continue; _incomings_[
_idx_].r = 6; if (_incomings_[_idx_].edges > 0) for (_j_ =
 0; _j_ < _incomings_[_idx_].c; _j_++) { const int d = _edges_
[_incomings_[_idx_].edges - 1 + _j_]; if (_incomings_[d].r !=
 4) continue; _incomings_[d].r = 5; ((void) sizeof ((_exist_size_
[_q_] < (exec_info_size)) ? 1 : 0), __extension__ ({ if (_exist_size_
[_q_] < (exec_info_size)) ; else __assert_fail ("_exist_size_[_q_] < (exec_info_size)"
, "ccv_cnnp_model.c", 1888, __extension__ __PRETTY_FUNCTION__
); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[
_q_]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for
 (_i_ = 0; _i_ < (compiled_data->evaluate.to_op_size); _i_
++) { ((void) sizeof (((compiled_data->evaluate.to_ops)[_i_
].graph == compiled_data->graph) ? 1 : 0), __extension__ (
{ if ((compiled_data->evaluate.to_ops)[_i_].graph == compiled_data
->graph) ; else __assert_fail ("(compiled_data->evaluate.to_ops)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1888, __extension__ __PRETTY_FUNCTION__
); })); _incomings_[(compiled_data->evaluate.to_ops)[_i_].
d].d = 1; } for (_i_ = 0; _i_ < (source_size); _i_++) { ((
void) sizeof (((sources)[_i_].graph == compiled_data->graph
) ? 1 : 0), __extension__ ({ if ((sources)[_i_].graph == compiled_data
->graph) ; else __assert_fail ("(sources)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1888, __extension__ __PRETTY_FUNCTION__
); })); _exists_[0][_i_] = (sources)[_i_].d; } _p_ = 0; _q_ =
 1; _exist_size_[0] = (source_size); _exist_size_[1] = 0; int
 _d_ = 0; while (_exist_size_[_p_] > 0) { _exist_size_[_q_
] = 0; for (_i_ = 0; _i_ < _exist_size_[_p_];) { const int32_t
 _idx_ = _exists_[_p_][_i_]; _visit_->node[_visit_->size
].index = ((_idx_)); _visit_->node[_visit_->size].term =
 ((_incomings_[_idx_].d)); ++_visit_->size;; if (_incomings_
[_idx_].d) { ++_d_; _incomings_[_idx_].r = 7; } if ((exec_info
)[_idx_].outgoings) { if ((exec_info)[_idx_].outgoings->rnum
 == 1) { const int d = *(int*)((void*)(((char*)(((exec_info)[
_idx_].outgoings)->data)) + (size_t)((exec_info)[_idx_].outgoings
)->rsize * (size_t)(0))); --_incomings_[d].c; if (_incomings_
[d].c == 0 && _incomings_[d].r == 6 && _d_ <
 (compiled_data->evaluate.to_op_size)) { _exists_[_p_][_i_
] = d; continue; } } else for (_j_ = 0; _j_ < (exec_info)[
_idx_].outgoings->rnum; _j_++) { const int d = *(int*)((void
*)(((char*)(((exec_info)[_idx_].outgoings)->data)) + (size_t
)((exec_info)[_idx_].outgoings)->rsize * (size_t)(_j_))); --
_incomings_[d].c; if (_incomings_[d].c == 0 && _incomings_
[d].r == 6 && _d_ < (compiled_data->evaluate.to_op_size
)) { ((void) sizeof ((_exist_size_[_q_] < (exec_info_size)
) ? 1 : 0), __extension__ ({ if (_exist_size_[_q_] < (exec_info_size
)) ; else __assert_fail ("_exist_size_[_q_] < (exec_info_size)"
, "ccv_cnnp_model.c", 1888, __extension__ __PRETTY_FUNCTION__
); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[
_q_]; } } } ++_i_; } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (
_i_)); } for (_i_ = 0; _i_ < (compiled_data->evaluate.to_op_size
); _i_++) { ((void) sizeof (((compiled_data->evaluate.to_ops
)[_i_].graph == compiled_data->graph) ? 1 : 0), __extension__
 ({ if ((compiled_data->evaluate.to_ops)[_i_].graph == compiled_data
->graph) ; else __assert_fail ("(compiled_data->evaluate.to_ops)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1888, __extension__ __PRETTY_FUNCTION__
); })); if (_incomings_[(compiled_data->evaluate.to_ops)[_i_
].d].r == 7) continue; if (!(0)) { ((void) sizeof ((_incomings_
[(compiled_data->evaluate.to_ops)[_i_].d].c == 0) ? 1 : 0)
, __extension__ ({ if (_incomings_[(compiled_data->evaluate
.to_ops)[_i_].d].c == 0) ; else __assert_fail ("_incomings_[(compiled_data->evaluate.to_ops)[_i_].d].c == 0"
, "ccv_cnnp_model.c", 1888, __extension__ __PRETTY_FUNCTION__
); })); } else if (_incomings_[(compiled_data->evaluate.to_ops
)[_i_].d].c > 0) continue; _visit_->node[_visit_->size
].index = (((compiled_data->evaluate.to_ops)[_i_].d)); _visit_
->node[_visit_->size].term = ((_incomings_[(compiled_data
->evaluate.to_ops)[_i_].d].d)); ++_visit_->size;; } if (
_heap_mem_) free(_incomings_); } while (0);; ((void) sizeof (
(_visit_->size <= (exec_info_size)) ? 1 : 0), __extension__
 ({ if (_visit_->size <= (exec_info_size)) ; else __assert_fail
 ("_visit_->size <= (exec_info_size)", "ccv_cnnp_model.c"
, 1888, __extension__ __PRETTY_FUNCTION__); })); _visit_; });
1889	ccv_nnc_graph_visit_for(visit, exec_info, node, idx){ int _i_; for (_i_ = 0; _i_ < (visit)->size; _i_++) { const
 int idx __attribute__((unused)) = (visit)->node[_i_].index
; const int _node_unused_ __attribute__((unused)) = (visit)->
node[_i_].term; typeof ((exec_info)) const node __attribute__
((unused)) = (exec_info) + idx; {
1890		visited[(idx >> 5)] |= (1u << (idx & 31));
1891	} ccv_nnc_graph_visit_endfor} }
1892	ccv_nnc_graph_visit_free(visit);
1893	const ccv_nnc_graph_exec_t* const destinations = (ccv_nnc_graph_exec_t*)ccv_array_get(compiled_data->graph->destinations, 0)((void*)(((char*)((compiled_data->graph->destinations)->
data)) + (size_t)(compiled_data->graph->destinations)->
rsize * (size_t)(0)));
1894	const int destination_size = compiled_data->graph->destinations->rnum;
1895	visit = ccv_nnc_graph_visit_new(compiled_data->graph, exec_info, exec_info_size, compiled_data->backward.from_ops, compiled_data->backward.from_op_size, destinations, destination_size, 0)({ ccv_nnc_graph_visit_t* _visit_ = (ccv_nnc_graph_visit_t*)malloc
(sizeof(ccv_nnc_graph_visit_t) + sizeof(_visit_->node[0]) *
 ((exec_info_size) - 1)); _visit_->size = 0; do { typedef struct
 { int8_t d; int8_t r; uint16_t c; int32_t edges; } ccv_nnc_incoming_t
; int _i_, _j_; int _incoming_edges_ = 0; for (_i_ = 0; _i_ <
 (exec_info_size); _i_++) _incoming_edges_ += ((exec_info)[_i_
].outgoings) ? (exec_info)[_i_].outgoings->rnum : 0; const
 int _heap_mem_ = ((exec_info_size) + _incoming_edges_ > 1024
); ccv_nnc_incoming_t* _incomings_; if (_heap_mem_) _incomings_
 = (ccv_nnc_incoming_t*)malloc(sizeof(ccv_nnc_incoming_t) * (
exec_info_size) + sizeof(int32_t) * ((exec_info_size) * 2 + _incoming_edges_
)); else _incomings_ = (ccv_nnc_incoming_t*)__builtin_alloca (
sizeof(ccv_nnc_incoming_t) * (exec_info_size) + sizeof(int32_t
) * ((exec_info_size) * 2 + _incoming_edges_)); memset(_incomings_
, 0, sizeof(ccv_nnc_incoming_t) * (exec_info_size)); int32_t*
 _exists_[2] = { (int32_t*)(_incomings_ + (exec_info_size)), (
int32_t*)(_incomings_ + (exec_info_size)) + (exec_info_size),
 }; int32_t* const _edges_ = _exists_[1] + (exec_info_size); for
 (_i_ = 0; _i_ < (compiled_data->backward.from_op_size)
; _i_++) { ((void) sizeof (((compiled_data->backward.from_ops
)[_i_].graph == compiled_data->graph) ? 1 : 0), __extension__
 ({ if ((compiled_data->backward.from_ops)[_i_].graph == compiled_data
->graph) ; else __assert_fail ("(compiled_data->backward.from_ops)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1895, __extension__ __PRETTY_FUNCTION__
); })); _incomings_[(compiled_data->backward.from_ops)[_i_
].d].r = 1; _exists_[0][_i_] = (compiled_data->backward.from_ops
)[_i_].d; } int _exist_size_[2] = { (compiled_data->backward
.from_op_size), 0, }; int _p_ = 0, _q_ = 1; while (_exist_size_
[_p_] > 0) { _exist_size_[_q_] = 0; for (_i_ = 0; _i_ <
 _exist_size_[_p_]; _i_++) { const int32_t _idx_ = _exists_[_p_
][_i_]; if (_incomings_[_idx_].r != 1) continue; _incomings_[
_idx_].r = 2; if ((exec_info)[_idx_].outgoings) for (_j_ = 0;
 _j_ < (exec_info)[_idx_].outgoings->rnum; _j_++) { const
 int d = *(int*)((void*)(((char*)(((exec_info)[_idx_].outgoings
)->data)) + (size_t)((exec_info)[_idx_].outgoings)->rsize
 * (size_t)(_j_))); ++_incomings_[d].c; if (_incomings_[d].r !=
 0) continue; _incomings_[d].r = 1; ((void) sizeof ((_exist_size_
[_q_] < (exec_info_size)) ? 1 : 0), __extension__ ({ if (_exist_size_
[_q_] < (exec_info_size)) ; else __assert_fail ("_exist_size_[_q_] < (exec_info_size)"
, "ccv_cnnp_model.c", 1895, __extension__ __PRETTY_FUNCTION__
); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[
_q_]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for
 (_i_ = 0; _i_ < (compiled_data->backward.from_op_size)
; _i_++) { ((void) sizeof (((compiled_data->backward.from_ops
)[_i_].graph == compiled_data->graph) ? 1 : 0), __extension__
 ({ if ((compiled_data->backward.from_ops)[_i_].graph == compiled_data
->graph) ; else __assert_fail ("(compiled_data->backward.from_ops)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1895, __extension__ __PRETTY_FUNCTION__
); })); _incomings_[(compiled_data->backward.from_ops)[_i_
].d].r = 3; _exists_[0][_i_] = (compiled_data->backward.from_ops
)[_i_].d; } _exist_size_[0] = (compiled_data->backward.from_op_size
); _exist_size_[1] = 0; _p_ = 0, _q_ = 1; int _bump_ = 1; while
 (_exist_size_[_p_] > 0) { _exist_size_[_q_] = 0; for (_i_
 = 0; _i_ < _exist_size_[_p_]; _i_++) { const int32_t _idx_
 = _exists_[_p_][_i_]; if (_incomings_[_idx_].r != 3) continue
; _incomings_[_idx_].r = 4; if ((exec_info)[_idx_].outgoings)
 for (_j_ = 0; _j_ < (exec_info)[_idx_].outgoings->rnum
; _j_++) { const int d = *(int*)((void*)(((char*)(((exec_info
)[_idx_].outgoings)->data)) + (size_t)((exec_info)[_idx_].
outgoings)->rsize * (size_t)(_j_))); if (_incomings_[d].edges
 == 0) { _incomings_[d].edges = _bump_; _bump_ += _incomings_
[d].c; _incomings_[d].c = 0; } _edges_[_incomings_[d].edges -
 1 + _incomings_[d].c] = _idx_; ++_incomings_[d].c; if (_incomings_
[d].r != 2) continue; _incomings_[d].r = 3; ((void) sizeof ((
_exist_size_[_q_] < (exec_info_size)) ? 1 : 0), __extension__
 ({ if (_exist_size_[_q_] < (exec_info_size)) ; else __assert_fail
 ("_exist_size_[_q_] < (exec_info_size)", "ccv_cnnp_model.c"
, 1895, __extension__ __PRETTY_FUNCTION__); })); _exists_[_q_
][_exist_size_[_q_]] = d; ++_exist_size_[_q_]; } } ((_i_) = (
_p_), (_p_) = (_q_), (_q_) = (_i_)); } for (_i_ = 0; _i_ <
 (destination_size); _i_++) { ((void) sizeof (((destinations)
[_i_].graph == compiled_data->graph) ? 1 : 0), __extension__
 ({ if ((destinations)[_i_].graph == compiled_data->graph)
 ; else __assert_fail ("(destinations)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1895, __extension__ __PRETTY_FUNCTION__
); })); _incomings_[(destinations)[_i_].d].r = 5; _exists_[0]
[_i_] = (destinations)[_i_].d; } _exist_size_[0] = (destination_size
); _exist_size_[1] = 0; _p_ = 0, _q_ = 1; while (_exist_size_
[_p_] > 0) { _exist_size_[_q_] = 0; for (_i_ = 0; _i_ <
 _exist_size_[_p_]; _i_++) { const int32_t _idx_ = _exists_[_p_
][_i_]; if (_incomings_[_idx_].r != 5) continue; _incomings_[
_idx_].r = 6; if (_incomings_[_idx_].edges > 0) for (_j_ =
 0; _j_ < _incomings_[_idx_].c; _j_++) { const int d = _edges_
[_incomings_[_idx_].edges - 1 + _j_]; if (_incomings_[d].r !=
 4) continue; _incomings_[d].r = 5; ((void) sizeof ((_exist_size_
[_q_] < (exec_info_size)) ? 1 : 0), __extension__ ({ if (_exist_size_
[_q_] < (exec_info_size)) ; else __assert_fail ("_exist_size_[_q_] < (exec_info_size)"
, "ccv_cnnp_model.c", 1895, __extension__ __PRETTY_FUNCTION__
); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[
_q_]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for
 (_i_ = 0; _i_ < (destination_size); _i_++) { ((void) sizeof
 (((destinations)[_i_].graph == compiled_data->graph) ? 1 :
 0), __extension__ ({ if ((destinations)[_i_].graph == compiled_data
->graph) ; else __assert_fail ("(destinations)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1895, __extension__ __PRETTY_FUNCTION__
); })); _incomings_[(destinations)[_i_].d].d = 1; } for (_i_ =
 0; _i_ < (compiled_data->backward.from_op_size); _i_++
) { ((void) sizeof (((compiled_data->backward.from_ops)[_i_
].graph == compiled_data->graph) ? 1 : 0), __extension__ (
{ if ((compiled_data->backward.from_ops)[_i_].graph == compiled_data
->graph) ; else __assert_fail ("(compiled_data->backward.from_ops)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1895, __extension__ __PRETTY_FUNCTION__
); })); _exists_[0][_i_] = (compiled_data->backward.from_ops
)[_i_].d; } _p_ = 0; _q_ = 1; _exist_size_[0] = (compiled_data
->backward.from_op_size); _exist_size_[1] = 0; int _d_ = 0
; while (_exist_size_[_p_] > 0) { _exist_size_[_q_] = 0; for
 (_i_ = 0; _i_ < _exist_size_[_p_];) { const int32_t _idx_
 = _exists_[_p_][_i_]; _visit_->node[_visit_->size].index
 = ((_idx_)); _visit_->node[_visit_->size].term = ((_incomings_
[_idx_].d)); ++_visit_->size;; if (_incomings_[_idx_].d) {
 ++_d_; _incomings_[_idx_].r = 7; } if ((exec_info)[_idx_].outgoings
) { if ((exec_info)[_idx_].outgoings->rnum == 1) { const int
 d = *(int*)((void*)(((char*)(((exec_info)[_idx_].outgoings)->
data)) + (size_t)((exec_info)[_idx_].outgoings)->rsize * (
size_t)(0))); --_incomings_[d].c; if (_incomings_[d].c == 0 &&
 _incomings_[d].r == 6 && _d_ < (destination_size)
) { _exists_[_p_][_i_] = d; continue; } } else for (_j_ = 0; _j_
 < (exec_info)[_idx_].outgoings->rnum; _j_++) { const int
 d = *(int*)((void*)(((char*)(((exec_info)[_idx_].outgoings)->
data)) + (size_t)((exec_info)[_idx_].outgoings)->rsize * (
size_t)(_j_))); --_incomings_[d].c; if (_incomings_[d].c == 0
 && _incomings_[d].r == 6 && _d_ < (destination_size
)) { ((void) sizeof ((_exist_size_[_q_] < (exec_info_size)
) ? 1 : 0), __extension__ ({ if (_exist_size_[_q_] < (exec_info_size
)) ; else __assert_fail ("_exist_size_[_q_] < (exec_info_size)"
, "ccv_cnnp_model.c", 1895, __extension__ __PRETTY_FUNCTION__
); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[
_q_]; } } } ++_i_; } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (
_i_)); } for (_i_ = 0; _i_ < (destination_size); _i_++) { (
(void) sizeof (((destinations)[_i_].graph == compiled_data->
graph) ? 1 : 0), __extension__ ({ if ((destinations)[_i_].graph
 == compiled_data->graph) ; else __assert_fail ("(destinations)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1895, __extension__ __PRETTY_FUNCTION__
); })); if (_incomings_[(destinations)[_i_].d].r == 7) continue
; if (!(0)) { ((void) sizeof ((_incomings_[(destinations)[_i_
].d].c == 0) ? 1 : 0), __extension__ ({ if (_incomings_[(destinations
)[_i_].d].c == 0) ; else __assert_fail ("_incomings_[(destinations)[_i_].d].c == 0"
, "ccv_cnnp_model.c", 1895, __extension__ __PRETTY_FUNCTION__
); })); } else if (_incomings_[(destinations)[_i_].d].c > 0
) continue; _visit_->node[_visit_->size].index = (((destinations
)[_i_].d)); _visit_->node[_visit_->size].term = ((_incomings_
[(destinations)[_i_].d].d)); ++_visit_->size;; } if (_heap_mem_
) free(_incomings_); } while (0);; ((void) sizeof ((_visit_->
size <= (exec_info_size)) ? 1 : 0), __extension__ ({ if (_visit_
->size <= (exec_info_size)) ; else __assert_fail ("_visit_->size <= (exec_info_size)"
, "ccv_cnnp_model.c", 1895, __extension__ __PRETTY_FUNCTION__
); })); _visit_; });
1896	ccv_nnc_graph_visit_for(visit, exec_info, node, idx){ int _i_; for (_i_ = 0; _i_ < (visit)->size; _i_++) { const
 int idx __attribute__((unused)) = (visit)->node[_i_].index
; const int _node_unused_ __attribute__((unused)) = (visit)->
node[_i_].term; typeof ((exec_info)) const node __attribute__
((unused)) = (exec_info) + idx; {
1897		visited[(idx >> 5)] |= (1u << (idx & 31));
1898	} ccv_nnc_graph_visit_endfor} }
1899	ccv_nnc_graph_visit_free(visit);
1900	visit = ccv_nnc_graph_visit_new(compiled_data->graph, exec_info, exec_info_size, sources, source_size, destinations, destination_size, 0)({ ccv_nnc_graph_visit_t* _visit_ = (ccv_nnc_graph_visit_t*)malloc
(sizeof(ccv_nnc_graph_visit_t) + sizeof(_visit_->node[0]) *
 ((exec_info_size) - 1)); _visit_->size = 0; do { typedef struct
 { int8_t d; int8_t r; uint16_t c; int32_t edges; } ccv_nnc_incoming_t
; int _i_, _j_; int _incoming_edges_ = 0; for (_i_ = 0; _i_ <
 (exec_info_size); _i_++) _incoming_edges_ += ((exec_info)[_i_
].outgoings) ? (exec_info)[_i_].outgoings->rnum : 0; const
 int _heap_mem_ = ((exec_info_size) + _incoming_edges_ > 1024
); ccv_nnc_incoming_t* _incomings_; if (_heap_mem_) _incomings_
 = (ccv_nnc_incoming_t*)malloc(sizeof(ccv_nnc_incoming_t) * (
exec_info_size) + sizeof(int32_t) * ((exec_info_size) * 2 + _incoming_edges_
)); else _incomings_ = (ccv_nnc_incoming_t*)__builtin_alloca (
sizeof(ccv_nnc_incoming_t) * (exec_info_size) + sizeof(int32_t
) * ((exec_info_size) * 2 + _incoming_edges_)); memset(_incomings_
, 0, sizeof(ccv_nnc_incoming_t) * (exec_info_size)); int32_t*
 _exists_[2] = { (int32_t*)(_incomings_ + (exec_info_size)), (
int32_t*)(_incomings_ + (exec_info_size)) + (exec_info_size),
 }; int32_t* const _edges_ = _exists_[1] + (exec_info_size); for
 (_i_ = 0; _i_ < (source_size); _i_++) { ((void) sizeof ((
(sources)[_i_].graph == compiled_data->graph) ? 1 : 0), __extension__
 ({ if ((sources)[_i_].graph == compiled_data->graph) ; else
 __assert_fail ("(sources)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1900, __extension__ __PRETTY_FUNCTION__
); })); _incomings_[(sources)[_i_].d].r = 1; _exists_[0][_i_]
 = (sources)[_i_].d; } int _exist_size_[2] = { (source_size),
 0, }; int _p_ = 0, _q_ = 1; while (_exist_size_[_p_] > 0)
 { _exist_size_[_q_] = 0; for (_i_ = 0; _i_ < _exist_size_
[_p_]; _i_++) { const int32_t _idx_ = _exists_[_p_][_i_]; if (
_incomings_[_idx_].r != 1) continue; _incomings_[_idx_].r = 2
; if ((exec_info)[_idx_].outgoings) for (_j_ = 0; _j_ < (exec_info
)[_idx_].outgoings->rnum; _j_++) { const int d = *(int*)((
void*)(((char*)(((exec_info)[_idx_].outgoings)->data)) + (
size_t)((exec_info)[_idx_].outgoings)->rsize * (size_t)(_j_
))); ++_incomings_[d].c; if (_incomings_[d].r != 0) continue;
 _incomings_[d].r = 1; ((void) sizeof ((_exist_size_[_q_] <
 (exec_info_size)) ? 1 : 0), __extension__ ({ if (_exist_size_
[_q_] < (exec_info_size)) ; else __assert_fail ("_exist_size_[_q_] < (exec_info_size)"
, "ccv_cnnp_model.c", 1900, __extension__ __PRETTY_FUNCTION__
); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[
_q_]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for
 (_i_ = 0; _i_ < (source_size); _i_++) { ((void) sizeof ((
(sources)[_i_].graph == compiled_data->graph) ? 1 : 0), __extension__
 ({ if ((sources)[_i_].graph == compiled_data->graph) ; else
 __assert_fail ("(sources)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1900, __extension__ __PRETTY_FUNCTION__
); })); _incomings_[(sources)[_i_].d].r = 3; _exists_[0][_i_]
 = (sources)[_i_].d; } _exist_size_[0] = (source_size); _exist_size_
[1] = 0; _p_ = 0, _q_ = 1; int _bump_ = 1; while (_exist_size_
[_p_] > 0) { _exist_size_[_q_] = 0; for (_i_ = 0; _i_ <
 _exist_size_[_p_]; _i_++) { const int32_t _idx_ = _exists_[_p_
][_i_]; if (_incomings_[_idx_].r != 3) continue; _incomings_[
_idx_].r = 4; if ((exec_info)[_idx_].outgoings) for (_j_ = 0;
 _j_ < (exec_info)[_idx_].outgoings->rnum; _j_++) { const
 int d = *(int*)((void*)(((char*)(((exec_info)[_idx_].outgoings
)->data)) + (size_t)((exec_info)[_idx_].outgoings)->rsize
 * (size_t)(_j_))); if (_incomings_[d].edges == 0) { _incomings_
[d].edges = _bump_; _bump_ += _incomings_[d].c; _incomings_[d
].c = 0; } _edges_[_incomings_[d].edges - 1 + _incomings_[d].
c] = _idx_; ++_incomings_[d].c; if (_incomings_[d].r != 2) continue
; _incomings_[d].r = 3; ((void) sizeof ((_exist_size_[_q_] <
 (exec_info_size)) ? 1 : 0), __extension__ ({ if (_exist_size_
[_q_] < (exec_info_size)) ; else __assert_fail ("_exist_size_[_q_] < (exec_info_size)"
, "ccv_cnnp_model.c", 1900, __extension__ __PRETTY_FUNCTION__
); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[
_q_]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for
 (_i_ = 0; _i_ < (destination_size); _i_++) { ((void) sizeof
 (((destinations)[_i_].graph == compiled_data->graph) ? 1 :
 0), __extension__ ({ if ((destinations)[_i_].graph == compiled_data
->graph) ; else __assert_fail ("(destinations)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1900, __extension__ __PRETTY_FUNCTION__
); })); _incomings_[(destinations)[_i_].d].r = 5; _exists_[0]
[_i_] = (destinations)[_i_].d; } _exist_size_[0] = (destination_size
); _exist_size_[1] = 0; _p_ = 0, _q_ = 1; while (_exist_size_
[_p_] > 0) { _exist_size_[_q_] = 0; for (_i_ = 0; _i_ <
 _exist_size_[_p_]; _i_++) { const int32_t _idx_ = _exists_[_p_
][_i_]; if (_incomings_[_idx_].r != 5) continue; _incomings_[
_idx_].r = 6; if (_incomings_[_idx_].edges > 0) for (_j_ =
 0; _j_ < _incomings_[_idx_].c; _j_++) { const int d = _edges_
[_incomings_[_idx_].edges - 1 + _j_]; if (_incomings_[d].r !=
 4) continue; _incomings_[d].r = 5; ((void) sizeof ((_exist_size_
[_q_] < (exec_info_size)) ? 1 : 0), __extension__ ({ if (_exist_size_
[_q_] < (exec_info_size)) ; else __assert_fail ("_exist_size_[_q_] < (exec_info_size)"
, "ccv_cnnp_model.c", 1900, __extension__ __PRETTY_FUNCTION__
); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[
_q_]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for
 (_i_ = 0; _i_ < (destination_size); _i_++) { ((void) sizeof
 (((destinations)[_i_].graph == compiled_data->graph) ? 1 :
 0), __extension__ ({ if ((destinations)[_i_].graph == compiled_data
->graph) ; else __assert_fail ("(destinations)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1900, __extension__ __PRETTY_FUNCTION__
); })); _incomings_[(destinations)[_i_].d].d = 1; } for (_i_ =
 0; _i_ < (source_size); _i_++) { ((void) sizeof (((sources
)[_i_].graph == compiled_data->graph) ? 1 : 0), __extension__
 ({ if ((sources)[_i_].graph == compiled_data->graph) ; else
 __assert_fail ("(sources)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1900, __extension__ __PRETTY_FUNCTION__
); })); _exists_[0][_i_] = (sources)[_i_].d; } _p_ = 0; _q_ =
 1; _exist_size_[0] = (source_size); _exist_size_[1] = 0; int
 _d_ = 0; while (_exist_size_[_p_] > 0) { _exist_size_[_q_
] = 0; for (_i_ = 0; _i_ < _exist_size_[_p_];) { const int32_t
 _idx_ = _exists_[_p_][_i_]; _visit_->node[_visit_->size
].index = ((_idx_)); _visit_->node[_visit_->size].term =
 ((_incomings_[_idx_].d)); ++_visit_->size;; if (_incomings_
[_idx_].d) { ++_d_; _incomings_[_idx_].r = 7; } if ((exec_info
)[_idx_].outgoings) { if ((exec_info)[_idx_].outgoings->rnum
 == 1) { const int d = *(int*)((void*)(((char*)(((exec_info)[
_idx_].outgoings)->data)) + (size_t)((exec_info)[_idx_].outgoings
)->rsize * (size_t)(0))); --_incomings_[d].c; if (_incomings_
[d].c == 0 && _incomings_[d].r == 6 && _d_ <
 (destination_size)) { _exists_[_p_][_i_] = d; continue; } } else
 for (_j_ = 0; _j_ < (exec_info)[_idx_].outgoings->rnum
; _j_++) { const int d = *(int*)((void*)(((char*)(((exec_info
)[_idx_].outgoings)->data)) + (size_t)((exec_info)[_idx_].
outgoings)->rsize * (size_t)(_j_))); --_incomings_[d].c; if
 (_incomings_[d].c == 0 && _incomings_[d].r == 6 &&
 _d_ < (destination_size)) { ((void) sizeof ((_exist_size_
[_q_] < (exec_info_size)) ? 1 : 0), __extension__ ({ if (_exist_size_
[_q_] < (exec_info_size)) ; else __assert_fail ("_exist_size_[_q_] < (exec_info_size)"
, "ccv_cnnp_model.c", 1900, __extension__ __PRETTY_FUNCTION__
); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[
_q_]; } } } ++_i_; } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (
_i_)); } for (_i_ = 0; _i_ < (destination_size); _i_++) { (
(void) sizeof (((destinations)[_i_].graph == compiled_data->
graph) ? 1 : 0), __extension__ ({ if ((destinations)[_i_].graph
 == compiled_data->graph) ; else __assert_fail ("(destinations)[_i_].graph == compiled_data->graph"
, "ccv_cnnp_model.c", 1900, __extension__ __PRETTY_FUNCTION__
); })); if (_incomings_[(destinations)[_i_].d].r == 7) continue
; if (!(0)) { ((void) sizeof ((_incomings_[(destinations)[_i_
].d].c == 0) ? 1 : 0), __extension__ ({ if (_incomings_[(destinations
)[_i_].d].c == 0) ; else __assert_fail ("_incomings_[(destinations)[_i_].d].c == 0"
, "ccv_cnnp_model.c", 1900, __extension__ __PRETTY_FUNCTION__
); })); } else if (_incomings_[(destinations)[_i_].d].c > 0
) continue; _visit_->node[_visit_->size].index = (((destinations
)[_i_].d)); _visit_->node[_visit_->size].term = ((_incomings_
[(destinations)[_i_].d].d)); ++_visit_->size;; } if (_heap_mem_
) free(_incomings_); } while (0);; ((void) sizeof ((_visit_->
size <= (exec_info_size)) ? 1 : 0), __extension__ ({ if (_visit_
->size <= (exec_info_size)) ; else __assert_fail ("_visit_->size <= (exec_info_size)"
, "ccv_cnnp_model.c", 1900, __extension__ __PRETTY_FUNCTION__
); })); _visit_; });
1901	// Find any missing nodes to be added as source. Right now, these are only set nodes.
1902	ccv_nnc_graph_visit_for(visit, exec_info, node, idx){ int _i_; for (_i_ = 0; _i_ < (visit)->size; _i_++) { const
 int idx __attribute__((unused)) = (visit)->node[_i_].index
; const int _node_unused_ __attribute__((unused)) = (visit)->
node[_i_].term; typeof ((exec_info)) const node __attribute__
((unused)) = (exec_info) + idx; {
1903		if (!(visited[(idx >> 5)] & (1u << (idx & 31))))
1904		{
1905			assert(exec_info[idx].cmd.cmd == CCV_NNC_SET_FORWARD)((void) sizeof ((exec_info[idx].cmd.cmd == CCV_NNC_SET_FORWARD
) ? 1 : 0), __extension__ ({ if (exec_info[idx].cmd.cmd == CCV_NNC_SET_FORWARD
) ; else __assert_fail ("exec_info[idx].cmd.cmd == CCV_NNC_SET_FORWARD"
, "ccv_cnnp_model.c", 1905, __extension__ __PRETTY_FUNCTION__
); }));
1906			if (exec_info[idx].cmd.info.blas.a[0] == 0) // Special-casing for empty out the tensor set function, not for the set grad to 1 one.
1907				ccv_array_add_unique_int(backward_from, idx);
1908		}
1909	} ccv_nnc_graph_visit_endfor} }
1910	ccv_nnc_graph_visit_free(visit);
1911	ccfreefree(visited);
1912	if (backward_from->rnum != compiled_data->backward.from_op_size) // If it doesn't match, need to redo this.
1913	{
1914		compiled_data->backward.from_op_size = backward_from->rnum;
1915		compiled_data->backward.from_ops = (ccv_nnc_graph_exec_t*)ccreallocrealloc(compiled_data->backward.from_ops, sizeof(ccv_nnc_graph_exec_t) * backward_from->rnum);
1916		for (i = 0; i < backward_from->rnum; i++)
1917			compiled_data->backward.from_ops[i] = (ccv_nnc_graph_exec_t){
1918				.d = *(int*)ccv_array_get(backward_from, i)((void*)(((char*)((backward_from)->data)) + (size_t)(backward_from
)->rsize * (size_t)(i))),
1919				.graph = compiled_data->graph,
1920			};
1921	}
1922	ccv_array_free(backward_from);
1923	ccv_nnc_graph_set_default_static_schedule(compiled_data->graph, compiled_data->stream_type, model->max_stream_count);
1924	ccv_nnc_graph_autotune(compiled_data->graph, model->workspace_size, 0, TRAVERSE_FULL0,0,0,0);
1925}
1926 
1927void 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)
1928{
1929	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1930	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 1930, __extension__ __PRETTY_FUNCTION__); }));
1931	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; });
1932	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", 1932, __extension__ __PRETTY_FUNCTION__
); }));
1933	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", 1933, __extension__ __PRETTY_FUNCTION__
); }));
1934	assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if
 (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c"
, 1934, __extension__ __PRETTY_FUNCTION__); }));
1935	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;
1936	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));
1937	if (!compiled_data->graph || mode_mismatch)
1938	{
1939		_ccv_cnnp_compiled_data_graph_free(compiled_data);
1940		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.
1941			_ccv_cnnp_compiled_data_backward_free(compiled_data);
1942		if (params.requires_grad)
1943			_ccv_cnnp_model_multistage_jit_0(model, params.disable_outgrad, params.is_test, inputs, input_size, outputs, output_size);
1944		else
1945			_ccv_cnnp_model_multistage_no_grad_jit(model, inputs, input_size, outputs, output_size);
1946	} else {
1947		ccv_nnc_tensor_arena_clear_bindings(compiled_data->tensor_arena);
1948		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"
, 1948, __extension__ __PRETTY_FUNCTION__); }));
1949		const int input_size_per_p = input_size / parallel_count;
1950		_ccv_cnnp_bind_tensors_to_arena(compiled_data->tensor_arena, model->graph, model->inputs, inputs, input_size_per_p, parallel_count);
1951		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"
, 1951, __extension__ __PRETTY_FUNCTION__); }));
1952		const int output_size_per_p = output_size / parallel_count;
1953		_ccv_cnnp_bind_tensors_to_arena(compiled_data->tensor_arena, model->graph, model->outputs, outputs, output_size_per_p, parallel_count);
1954	}
1955	if (compiled_data->is_test != params.is_test)
1956	{
1957		compiled_data->is_test = params.is_test;
1958		ccv_nnc_graph_exec_update_t update = {
1959			.parallel_count = parallel_count,
1960			.graph = model->graph,
1961			.graph_exec_arena = compiled_data->graph_exec_arena,
1962		};
1963		ccv_cnnp_model_set_is_test(model, params.is_test, _ccv_cnnp_cmd_update_for_execs, &update);
1964	}
1965}
1966 
1967void 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)
1968{
1969	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1970	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 1970, __extension__ __PRETTY_FUNCTION__); }));
1971	ccv_cnnp_model_dry_run(model, params, inputs, input_size, outputs, output_size);
1972	if (compiled_data->graph_mode == CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE_NO_GRAD)
1973		ccv_nnc_graph_run_with_schedule(compiled_data->graph, 0, 0, tensor_tape, stream_context);
1974	else {
1975		if (!compiled_data->evaluate.schedule)
1976			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);
1977		ccv_nnc_graph_run_with_schedule(compiled_data->graph, 0, compiled_data->evaluate.schedule, tensor_tape, stream_context);
1978	}
1979}
1980 
1981// Compile the graph to run ccv_cnnp_model_backward after ccv_cnnp_model_evaluate with requires_grad = true (MULTISTAGE_MODE).
1982// Particularly, this method compiles the accumulator graph.
1983static void _ccv_cnnp_model_multistage_jit_1(ccv_cnnp_model_t* const model)
1984{
1985	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1986	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 1986, __extension__ __PRETTY_FUNCTION__); }));
1987	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", 1987, __extension__ __PRETTY_FUNCTION__
); }));
1988	ccv_nnc_symbolic_graph_t* accum = ccv_nnc_symbolic_graph_new();
1989	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; });
1990	const int parameter_size = compiled_data->parameters->rnum;
1991	int i, j;
1992	compiled_data->backward.gradients = (ccv_nnc_tensor_symbol_t*)ccmallocmalloc(sizeof(ccv_nnc_tensor_symbol_t) * parameter_size * parallel_count * 3);
1993	compiled_data->backward.accum_gradients = compiled_data->backward.gradients + parameter_size * parallel_count;
1994	compiled_data->backward.updated_accum_gradients = compiled_data->backward.accum_gradients + parameter_size * parallel_count;
1995	for (i = 0; i < parameter_size; i++)
1996		for (j = 0; j < parallel_count; j++)
1997			if (compiled_data->tensors.gradients[i + j * parameter_size])
1998			{
1999				const ccv_nnc_tensor_param_t info = compiled_data->tensors.gradients[i + j * parameter_size]->info;
2000				// Now, the old gradient is the accumulated gradient, getting new gradient tensor setup so we can collect them.
2001				compiled_data->tensors.accum_gradients[i + j * parameter_size] = compiled_data->tensors.gradients[i + j * parameter_size];
2002				compiled_data->tensors.gradients[i + j * parameter_size] = ccv_nnc_tensor_new(0, info, 0);
2003				ccv_nnc_tensor_symbol_t inputs[2];
2004				inputs[0] = compiled_data->backward.accum_gradients[i + j * parameter_size] = ccv_nnc_tensor_symbol_new(accum, info, 0);
2005				inputs[1] = compiled_data->backward.gradients[i + j * parameter_size] = ccv_nnc_tensor_symbol_new(accum, info, 0);
2006				ccv_nnc_tensor_symbol_t output = compiled_data->backward.updated_accum_gradients[i + j * parameter_size] = ccv_nnc_tensor_symbol_new(accum, info, 0);
2007				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);
2008			} else {
2009				compiled_data->backward.accum_gradients[i + j * parameter_size] = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
2010				compiled_data->backward.gradients[i + j * parameter_size] = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
2011				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
};
2012			}
2013	ccv_nnc_graph_exec_symbol_autogen(accum, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
2014	if (ccv_nnc_symbolic_graph_source_size(accum) == 0)
2015	{
2016		ccv_nnc_symbolic_graph_free(accum);
2017		// Create empty graph.
2018		compiled_data->backward.accum = ccv_nnc_graph_new();
2019		ccv_nnc_graph_topsort(compiled_data->backward.accum, 0, 0);
2020		return;
2021	}
2022	ccv_array_t* const tensor_binds = ccv_array_new(sizeof(ccv_nnc_tensor_bind_t), 0, 0);
2023	_ccv_cnnp_model_bind_tensors(accum, compiled_data->backward.accum_gradients, compiled_data->tensors.accum_gradients, parameter_size * parallel_count, 1, tensor_binds);
2024	_ccv_cnnp_model_bind_tensors(accum, compiled_data->backward.gradients, compiled_data->tensors.gradients, parameter_size * parallel_count, 1, tensor_binds);
2025	_ccv_cnnp_model_bind_tensors(accum, compiled_data->backward.updated_accum_gradients, compiled_data->tensors.accum_gradients, parameter_size * parallel_count, 1, tensor_binds);
2026	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);
2027	ccv_nnc_symbolic_graph_free(accum);
2028	ccv_array_free(tensor_binds);
2029	ccv_nnc_graph_set_default_static_schedule(compiled_data->backward.accum, compiled_data->stream_type, model->max_stream_count);
2030}
2031 
2032void 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)
2033{
2034	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2035	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 2035, __extension__ __PRETTY_FUNCTION__); }));
2036	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", 2036, __extension__ __PRETTY_FUNCTION__
); }));
2037	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; });
2038	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", 2038, __extension__ __PRETTY_FUNCTION__
); }));
2039	if (outgrad_size > 0)
2040		{ 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", 2040, __extension__ __PRETTY_FUNCTION__
); })); }
2041	assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if
 (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c"
, 2041, __extension__ __PRETTY_FUNCTION__); }));
2042	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", 2042, __extension__ __PRETTY_FUNCTION__
); }));
2043	const int parameter_size = compiled_data->parameters->rnum;
2044	// If we need to accumulate the gradients now, do jit on accumulator.
2045	if (compiled_data->backward.count > 0)
2046	{
2047		if (!compiled_data->backward.accum)
2048			_ccv_cnnp_model_multistage_jit_1(model);
2049		else if (compiled_data->backward.count == 1) {
2050			//  On this round, we need to switch accumulated gradients with gradients (so we can do accumulation properly).
2051			int i;
2052			for (i = 0; i < parameter_size * parallel_count; i++)
2053			{
2054				ccv_nnc_tensor_t* tensor;
2055				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));
2056			}
2057			if (compiled_data->backward.tensor_arena)
2058			{
2059				ccv_nnc_tensor_arena_clear_bindings(compiled_data->backward.tensor_arena);
2060				// Do rebind in case we messed up the binding (we switch accum_gradients and gradients).
2061				_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);
2062				_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);
2063				_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);
2064			}
2065		}
2066	}
2067	const int ingrad_size_per_p = model->output_size;
2068	const int outgrad_size_per_p = compiled_data->outgrad_size;
2069	int i, j;
2070	for (i = 0; i < ingrad_size_per_p; i++)
2071	{
2072		const ccv_nnc_tensor_symbol_t ingrad = ccv_nnc_tensor_symbol_for_backward(model->graph, compiled_data->f[i]);
2073		if (!ingrad_size || !ingrads || ingrads[i] == 0)
2074		{
2075			// Set it to 1 if it is not specified.
2076			ccv_nnc_tensor_t* const ingrad_tensor = ccv_nnc_tensor_from_symbol(compiled_data->tensor_arena, ingrad);
2077			if (ingrad_tensor)
2078				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);
2079			for (j = 1; j < parallel_count; j++)
2080			{
2081				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));
2082				if (ingrad_tensor)
2083					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);
2084			}
2085		} else {
2086			// Make sure the length matches, in case it is an alias.
2087			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", 2087, __extension__ __PRETTY_FUNCTION__
); }));
2088			ccv_nnc_tensor_bind_symbol(compiled_data->tensor_arena, ingrad, ingrads[i]);
2089			for (j = 1; j < parallel_count; j++)
2090				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]);
2091		}
2092	}
2093	if (outgrad_size > 0)
2094	{
2095		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", 2095, __extension__ __PRETTY_FUNCTION__
); }));
2096		for (i = 0; i < outgrad_size_per_p; i++)
2097			if (outgrads[i])
2098			{
2099				const ccv_nnc_tensor_symbol_t outgrad = compiled_data->outgrads[i];
2100				ccv_nnc_tensor_bind_symbol(compiled_data->tensor_arena, outgrad, outgrads[i]);
2101				for (j = 1; j < parallel_count; j++)
2102					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]);
2103			}
2104	} else {
2105		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", 2106, __extension__ __PRETTY_FUNCTION__
); }))
2106			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", 2106, __extension__ __PRETTY_FUNCTION__
); }));
2107	}
2108	// We need to rebind here because in ccv_cnnp_evaluate, we clear bindings, that will reset all bindings for the gradients.
2109	// For parameters and internals these are fine because when we clear bindings, it restores to original bindings, which are these
2110	// parameters and internals. The same cannot be said for gradients due to the accum_gradients switching.
2111	_ccv_cnnp_bind_tensors_to_arena(compiled_data->tensor_arena, model->graph, compiled_data->gradients, compiled_data->tensors.gradients, parameter_size, parallel_count);
2112	if (!compiled_data->backward.schedule)
2113		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);
2114	// Run the backward pass.
2115	ccv_nnc_graph_run_with_schedule(compiled_data->graph, 0, compiled_data->backward.schedule, tensor_tape, stream_context);
2116	// If we need to run accumulation round, do that now.
2117	if (compiled_data->backward.count > 0)
2118		ccv_nnc_graph_run_with_schedule(compiled_data->backward.accum, 0, 0, 0, stream_context);
2119	// Update the count, this determines whether we need to accumulate or not.
2120	++compiled_data->backward.count;
2121}
2122 
2123// Compile the graph to run ccv_cnnp_model_apply_gradients after ccv_cnnp_model_backward (MULTISTAGE_MODE).
2124// Particularly, this method compiles the parameter update graph.
2125static void _ccv_cnnp_model_multistage_jit_2(ccv_cnnp_model_t* const model)
2126{
2127	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2128	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", 2128, __extension__ __PRETTY_FUNCTION__
); }));
2129	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; });
2130	const int parameter_size = compiled_data->parameters->rnum;
2131	ccv_array_t* const tensor_binds = ccv_array_new(sizeof(ccv_nnc_tensor_bind_t), 0, 0);
2132	_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);
2133	_ccv_cnnp_model_bind_tensors(model->graph, compiled_data->updated_parameters, compiled_data->tensors.parameters, parameter_size, parallel_count, tensor_binds);
2134	// Bind accumulated gradients.
2135	if (compiled_data->backward.count > 1)
2136		_ccv_cnnp_model_bind_tensors(model->graph, compiled_data->gradients, compiled_data->tensors.accum_gradients, parameter_size, parallel_count, tensor_binds);
2137	else
2138		_ccv_cnnp_model_bind_tensors(model->graph, compiled_data->gradients, compiled_data->tensors.gradients, parameter_size, parallel_count, tensor_binds);
2139	ccv_array_t* const apply_gradients_from = ccv_array_new(sizeof(int), 0, 0);
2140	int i, j;
2141	for (i = 0; i < compiled_data->backward.to_size; i++)
2142	{
2143		const int* tos;
2144		int to_size;
2145		ccv_nnc_graph_exec_symbol_to(model->graph, compiled_data->backward.tos[i], &tos, &to_size);
2146		for (j = 0; j < to_size; j++)
2147		{
2148			// Check if this is already show up in the backward graph, if that is the case, it won't be in the apply
2149			// gradients graph.
2150			const ccv_nnc_graph_exec_t exec = ccv_nnc_graph_exec_from_symbol(compiled_data->graph_exec_arena, (ccv_nnc_graph_exec_symbol_t){
2151				.d = tos[j],
2152				.graph = model->graph,
2153			});
2154			if (!exec.graph)
2155				ccv_array_add_unique_int(apply_gradients_from, tos[j]);
2156		}
2157	}
2158	const int from_size = apply_gradients_from->rnum;
2159	if (from_size == 0)
2160	{
2161		ccv_array_free(apply_gradients_from);
2162		ccv_array_free(tensor_binds);
2163		return;
2164	}
2165	ccv_nnc_graph_exec_symbol_t* const froms = (ccv_nnc_graph_exec_symbol_t*)ccmallocmalloc(sizeof(ccv_nnc_graph_exec_symbol_t) * from_size);
2166	for (i = 0; i < from_size; i++)
2167		froms[i] = (ccv_nnc_graph_exec_symbol_t){
2168			.d = *(int*)ccv_array_get(apply_gradients_from, i)((void*)(((char*)((apply_gradients_from)->data)) + (size_t
)(apply_gradients_from)->rsize * (size_t)(i))),
2169			.graph = model->graph
2170		};
2171	ccv_array_free(apply_gradients_from);
2172	// It can only ends with updates on the parameters.
2173	ccv_array_t* const tos = ccv_array_new(sizeof(ccv_nnc_graph_exec_symbol_t), parameter_size * parallel_count, 0);
2174	for (i = 0;  i < parameter_size; i++)
2175	{
2176		if (compiled_data->update_nodes[i].d == CCV_NNC_NO_TENSOR_SYMBOL)
2177			continue;
2178		ccv_array_push(tos, &compiled_data->update_nodes[i]);
2179		for (j = 1; j < parallel_count; j++)
2180		{
2181			const ccv_nnc_graph_exec_symbol_t copy = ccv_nnc_graph_exec_symbol_copy(model->graph, compiled_data->update_nodes[i], j);
2182			ccv_array_push(tos, &copy);
2183		}
2184	}
2185	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);
2186	ccv_array_free(tos);
2187	ccv_array_free(tensor_binds);
2188	ccfreefree(froms);
2189	const int max_saved_aux_size = compiled_data->minimize.max_saved_aux_size;
2190	for (i = 0; i < max_saved_aux_size * parameter_size; i++)
2191	{
2192		// Skip on no tensor.
2193		if (compiled_data->saved_aux[i].source.d == CCV_NNC_NO_TENSOR_SYMBOL)
2194			continue;
2195		ccv_nnc_tensor_t* const tensor = ccv_nnc_tensor_from_symbol(compiled_data->apply_gradients.tensor_arena, compiled_data->saved_aux[i].source);
2196		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);
2197		for (j = 1; j < parallel_count; j++)
2198		{
2199			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));
2200			if (copy)
2201				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);
2202		}
2203	}
2204	ccv_nnc_graph_set_default_static_schedule(compiled_data->apply_gradients.graph, compiled_data->stream_type, model->max_stream_count);
2205}
2206 
2207void ccv_cnnp_model_apply_gradients(ccv_cnnp_model_t* const model, ccv_nnc_stream_context_t* const stream_context)
2208{
2209	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2210	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 2210, __extension__ __PRETTY_FUNCTION__); }));
2211	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", 2211, __extension__ __PRETTY_FUNCTION__
); }));
2212	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; });
2213	assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if
 (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c"
, 2213, __extension__ __PRETTY_FUNCTION__); }));
2214	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", 2214, __extension__ __PRETTY_FUNCTION__
); }));
2215	// Skip if there is no backward pass.
2216	if (compiled_data->backward.count <= 0)
2217		return;
2218	// Skip if there is no parameters.
2219	if (compiled_data->parameters->rnum == 0)
2220	{
2221		compiled_data->backward.count = 0;
2222		return;
2223	}
2224	if (!compiled_data->apply_gradients.graph)
2225		_ccv_cnnp_model_multistage_jit_2(model);
2226	else {
2227		const int parameter_size = compiled_data->parameters->rnum;
2228		ccv_nnc_tensor_arena_clear_bindings(compiled_data->apply_gradients.tensor_arena);
2229		// Change to bind accum_gradients if we do gradient accumulation (run backward more than once).
2230		if (compiled_data->backward.count > 1)
2231			_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);
2232		else
2233			_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);
2234	}
2235	if (compiled_data->apply_gradients.graph)
2236		ccv_nnc_graph_run_with_schedule(compiled_data->apply_gradients.graph, 0, 0, 0, stream_context);
2237	// Reset backward count to 0.
2238	compiled_data->backward.count = 0;
2239}
2240 
2241void 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)
2242{
2243	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2244	const int param_sel = parameter->param_sel > 0 ? parameter->param_sel - 1 : parameter->param_sel;
2245	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", 2245, __extension__ __PRETTY_FUNCTION__
); }));
2246	const int tensors_init = !!compiled_data->tensors_init.v;
2247	int this_tensor_init = tensors_init;
2248	if (!tensors_init)
2249		ccv_cnnp_model_tensors_init_0(model, compiled_data);
2250	else if ((uintptr_t)compiled_data->tensors_init.v & (uintptr_t)1)
2251	// Check if it is not fully allocated, if it is not, init_1.
2252		this_tensor_init = 0;
2253	ccv_array_t* const parameter_indices = ccv_array_new(sizeof(int), 0, 0);
2254	ccv_cnnp_model_add_to_parameter_indices(parameter->model, param_sel, parameter_indices);
2255	const int param_ref = parameter->param_ref > 0 ? parameter->param_ref - 1 : parameter->param_ref;
2256	if (param_ref < 0)
2257		{ 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", 2257
, __extension__ __PRETTY_FUNCTION__); })); }
2258	else
2259		{ 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", 2259, __extension__ __PRETTY_FUNCTION__
); })); }
2260	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)));
2261	ccv_array_free(parameter_indices);
2262	const int parameter_size = compiled_data->parameters->rnum;
2263	assert(d >= 0)((void) sizeof ((d >= 0) ? 1 : 0), __extension__ ({ if (d >=
 0) ; else __assert_fail ("d >= 0", "ccv_cnnp_model.c", 2263
, __extension__ __PRETTY_FUNCTION__); }));
2264	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", 2264, __extension__ __PRETTY_FUNCTION__
); }));
2265	const int parallel_count = _ccv_cnnp_compiled_data_parallel_count(model, compiled_data);
2266	int i;
2267	if (!this_tensor_init)
2268	{
2269		if (compiled_data->tensors.parameters[d])
2270		{
2271			for (i = 1; i < parallel_count; i++)
2272				{ assert(compiled_data->tensors.parameters[d + i * parameter_size])((void) sizeof ((compiled_data->tensors.parameters[d + i *
 parameter_size]) ? 1 : 0), __extension__ ({ if (compiled_data
->tensors.parameters[d + i * parameter_size]) ; else __assert_fail
 ("compiled_data->tensors.parameters[d + i * parameter_size]"
, "ccv_cnnp_model.c", 2272, __extension__ __PRETTY_FUNCTION__
); })); }
2273			this_tensor_init = 1;
2274		} else {
2275			const ccv_nnc_tensor_symbol_t parameter = *(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)));
2276			ccv_nnc_tensor_param_t info = ccv_nnc_tensor_symbol_params(parameter.graph, parameter);
2277			if (CCV_TENSOR_GET_DEVICE(info.type)((info.type) & 0xfff00) == CCV_COMPUTE_DEVICE_ANY)
2278				CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff
) << 8));
2279			const int device_id = CCV_TENSOR_GET_DEVICE_ID(info.type)(((info.type) & 0xfff00) >> 8);
2280			compiled_data->tensors.parameters[d] = ccv_nnc_tensor_new(0, info, 0);
2281			for (i = 1; i < parallel_count; i++)
2282			{
2283				if (i != device_id)
2284					CCV_TENSOR_SET_DEVICE_ID(info.type, i)(info.type) = (((info.type) & ~0xfff00) | (((i) & 0xfff
) << 8));
2285				else
2286					CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff
) << 8));
2287				compiled_data->tensors.parameters[d + i * parameter_size] = ccv_nnc_tensor_new(0, info, 0);
2288			}
2289		}
2290	}
2291	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));
2292	assert(dest)((void) sizeof ((dest) ? 1 : 0), __extension__ ({ if (dest) ;
 else __assert_fail ("dest", "ccv_cnnp_model.c", 2292, __extension__
 __PRETTY_FUNCTION__); }));
2293	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);
2294	for (i = 1; i < parallel_count; i++)
2295	{
2296		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));
2297		if (copy_tensor)
2298			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);
2299	}
2300	// Mark this symbol as init'ed.
2301	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;
2302	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));
2303	init_v[s >> 5] |= (1u << (s & 0x1f));
2304	// If we just allocated this tensor, now it is time to check if we need to mark it as fully allocated.
2305	if (!this_tensor_init)
2306	{
2307		if (ccv_cnnp_model_tensors_any_to_alloc(model, compiled_data))
2308			compiled_data->tensors_init.v = (uint32_t*)((uintptr_t)compiled_data->tensors_init.v | (uintptr_t)1);
2309		else // Remove the flag.
2310			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));
2311	}
2312}
2313 
2314void ccv_cnnp_model_parameter_copy(ccv_cnnp_model_t* const model, const ccv_cnnp_model_io_t parameter, ccv_nnc_tensor_t* const tensor)
2315{
2316	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2317	const int param_sel = parameter->param_sel > 0 ? parameter->param_sel - 1 : parameter->param_sel;
2318	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", 2318, __extension__ __PRETTY_FUNCTION__
); }));
2319	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", 2319, __extension__ __PRETTY_FUNCTION__
); }));
2320	ccv_array_t* const parameter_indices = ccv_array_new(sizeof(int), 0, 0);
2321	ccv_cnnp_model_add_to_parameter_indices(parameter->model, param_sel, parameter_indices);
2322	const int param_ref = parameter->param_ref > 0 ? parameter->param_ref - 1 : parameter->param_ref;
2323	if (param_ref < 0)
2324		{ 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", 2324
, __extension__ __PRETTY_FUNCTION__); })); }
2325	else
2326		{ 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", 2326, __extension__ __PRETTY_FUNCTION__
); })); }
2327	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)));
2328	ccv_array_free(parameter_indices);
2329	const int parameter_size = compiled_data->parameters->rnum;
2330	assert(d >= 0)((void) sizeof ((d >= 0) ? 1 : 0), __extension__ ({ if (d >=
 0) ; else __assert_fail ("d >= 0", "ccv_cnnp_model.c", 2330
, __extension__ __PRETTY_FUNCTION__); }));
2331	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", 2331, __extension__ __PRETTY_FUNCTION__
); }));
2332	// We don't need to consider parallel_count, every parameter on each device is identical.
2333	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));
2334	assert(src)((void) sizeof ((src) ? 1 : 0), __extension__ ({ if (src) ; else
 __assert_fail ("src", "ccv_cnnp_model.c", 2334, __extension__
 __PRETTY_FUNCTION__); }));
2335	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);
2336}
2337 
2338ccv_nnc_tensor_param_t ccv_cnnp_model_parameter_tensor_params(ccv_cnnp_model_t* const model, const ccv_cnnp_model_io_t parameter)
2339{
2340	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2341	const int param_sel = parameter->param_sel > 0 ? parameter->param_sel - 1 : parameter->param_sel;
2342	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", 2342, __extension__ __PRETTY_FUNCTION__
); }));
2343	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", 2343, __extension__ __PRETTY_FUNCTION__
); }));
2344	ccv_array_t* const parameter_indices = ccv_array_new(sizeof(int), 0, 0);
2345	ccv_cnnp_model_add_to_parameter_indices(parameter->model, param_sel, parameter_indices);
2346	const int param_ref = parameter->param_ref > 0 ? parameter->param_ref - 1 : parameter->param_ref;
2347	if (param_ref < 0)
2348		{ 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", 2348
, __extension__ __PRETTY_FUNCTION__); })); }
2349	else
2350		{ 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", 2350, __extension__ __PRETTY_FUNCTION__
); })); }
2351	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)));
2352	ccv_array_free(parameter_indices);
2353	const int parameter_size = compiled_data->parameters->rnum;
2354	assert(d >= 0)((void) sizeof ((d >= 0) ? 1 : 0), __extension__ ({ if (d >=
 0) ; else __assert_fail ("d >= 0", "ccv_cnnp_model.c", 2354
, __extension__ __PRETTY_FUNCTION__); }));
2355	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", 2355, __extension__ __PRETTY_FUNCTION__
); }));
2356	// We don't need to consider parallel_count, every parameter on each device is identical.
2357	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));
2358	assert(tensor)((void) sizeof ((tensor) ? 1 : 0), __extension__ ({ if (tensor
) ; else __assert_fail ("tensor", "ccv_cnnp_model.c", 2358, __extension__
 __PRETTY_FUNCTION__); }));
2359	return tensor->info;
2360}
2361 
2362const char* ccv_cnnp_model_parameter_name(ccv_cnnp_model_t* const model, const ccv_cnnp_model_io_t parameter)
2363{
2364	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2365	const int param_sel = parameter->param_sel > 0 ? parameter->param_sel - 1 : parameter->param_sel;
2366	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", 2366, __extension__ __PRETTY_FUNCTION__
); }));
2367	ccv_array_t* const parameter_indices = ccv_array_new(sizeof(int), 0, 0);
2368	ccv_cnnp_model_add_to_parameter_indices(parameter->model, param_sel, parameter_indices);
2369	const int param_ref = parameter->param_ref > 0 ? parameter->param_ref - 1 : parameter->param_ref;
2370	if (param_ref < 0)
2371		{ 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", 2371
, __extension__ __PRETTY_FUNCTION__); })); }
2372	else
2373		{ 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", 2373, __extension__ __PRETTY_FUNCTION__
); })); }
2374	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)));
2375	ccv_array_free(parameter_indices);
2376	const int parameter_size = compiled_data->parameters->rnum;
2377	assert(d >= 0)((void) sizeof ((d >= 0) ? 1 : 0), __extension__ ({ if (d >=
 0) ; else __assert_fail ("d >= 0", "ccv_cnnp_model.c", 2377
, __extension__ __PRETTY_FUNCTION__); }));
2378	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", 2378, __extension__ __PRETTY_FUNCTION__
); }));
2379	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)));
2380}
2381 
2382int ccv_cnnp_model_parameter_count(ccv_cnnp_model_t* const model)
2383{
2384	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", 2384, __extension__ __PRETTY_FUNCTION__
); }));
2385	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2386	return compiled_data->parameters->rnum;
2387}
2388 
2389uint64_t ccv_cnnp_model_parameters_size(ccv_cnnp_model_t* const model)
2390{
2391	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", 2391, __extension__ __PRETTY_FUNCTION__
); }));
2392	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2393	const int parameter_size = compiled_data->parameters->rnum;
2394	int i;
2395	const ccv_nnc_symbolic_graph_t* const graph = model->graph;
2396	uint64_t size = 0;
2397	const int tensors_init = !!compiled_data->tensors_init.v;
2398	uint32_t* const init_v = tensors_init ? CCV_NNC_INIT_V(compiled_data->tensors_init.v)((uint32_t*)((uintptr_t)(compiled_data->tensors_init.v) &
 ~(uintptr_t)1)) : 0;
2399	for (i = 0; i < parameter_size; i++)
2400	{
2401		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;
2402		if (tensors_init && compiled_data->tensors.parameters && (init_v[d >> 5] | (1u << (d & 0x1f))) && compiled_data->tensors.parameters[i])
2403		{
2404			ccv_nnc_tensor_param_t params = compiled_data->tensors.parameters[i]->info;
2405			size += ccv_nnc_tensor_data_size(params);
2406			continue;
2407		}
2408		ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, (ccv_nnc_tensor_symbol_t){
2409			.graph = graph,
2410			.d = d
2411		});
2412		size += ccv_nnc_tensor_data_size(params);
2413	}
2414	return size;
2415}
2416 
2417int ccv_cnnp_model_parameters_move(ccv_cnnp_model_t* const model, char** const names, ccv_nnc_tensor_t** const tensors, const int count, int type)
2418{
2419	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", 2419, __extension__ __PRETTY_FUNCTION__
); }));
2420	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2421	if (count != compiled_data->parameters->rnum)
2422		return 0;
2423	if (CCV_TENSOR_GET_DEVICE(type)((type) & 0xfff00) == CCV_COMPUTE_DEVICE_ANY)
2424		CCV_TENSOR_SET_DEVICE_ID(type, 0)(type) = (((type) & ~0xfff00) | (((0) & 0xfff) <<
 8));
2425	int i;
2426	// We don't need to consider parallel_count, every parameter on each device is identical.
2427	for (i = 0; i < count; i++)
2428	{
2429		ccv_nnc_tensor_t* tensor = compiled_data->tensors.parameters[i];
2430		if ((uintptr_t)tensor & (uintptr_t)1) // If it is not owned. We don't do anything.
2431		{
2432			tensors[i] = 0;
2433			continue;
2434		}
2435		tensor = CCV_NNC_TENSOR(tensor)((ccv_nnc_tensor_t*)((uintptr_t)(tensor) & ~(uintptr_t)1)
);
2436		if (tensor->info.type == type)
2437			tensors[i] = tensor;
2438		else {
2439			ccv_nnc_tensor_param_t info = tensor->info;
2440			info.type = type;
2441			tensors[i] = ccv_nnc_tensor_new(0, info, 0); // Create this tensor, don't initiate copy yet.
2442		}
2443	}
2444	for (i = 0; i < count; i++)
2445	{
2446		ccv_nnc_tensor_t* tensor = compiled_data->tensors.parameters[i];
2447		if ((uintptr_t)tensor & (uintptr_t)1) // If it is not owned. We don't do anything.
2448			continue;
2449		tensor = CCV_NNC_TENSOR(tensor)((ccv_nnc_tensor_t*)((uintptr_t)(tensor) & ~(uintptr_t)1)
);
2450		// Now initiate transfer. We should do this one on a stream.
2451		if (tensor->info.type != type)
2452			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(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), TENSOR_LIST(tensors[i])(ccv_nnc_tensor_t* []){tensors[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);
2453	}
2454	// Copy names and remove parameters.
2455	for (i = 0; i < count; i++)
2456	{
2457		ccv_nnc_tensor_t* const tensor = compiled_data->tensors.parameters[i];
2458		if ((uintptr_t)tensor & (uintptr_t)1) // If it is not owned. We don't do anything.
2459		{
2460			names[i] = 0;
2461			continue;
2462		}
2463		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)));
2464		const size_t name_len = ccv_min(strnlen(name, 1023), 1023)({ typeof (strnlen(name, 1023)) _a = (strnlen(name, 1023)); typeof
 (1023) _b = (1023); (_a < _b) ? _a : _b; });
2465		names[i] = ccmallocmalloc(name_len + 1);
2466		names[i][name_len] = 0;
2467		memcpy(names[i], name, name_len);
2468		if (tensor->info.type == type)
2469			compiled_data->tensors.parameters[i] = 0; // Only move when it is moved.
2470	}
2471	return 1;
2472}
2473 
2474KHASH_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 2474→
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
2475 
2476void ccv_cnnp_model_set_parameters_from_key_values(ccv_cnnp_model_t* const model, char* const* const names, ccv_nnc_tensor_t** const tensors, const int count, const int invalidates)
2477{
2478	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", 2478, __extension__ __PRETTY_FUNCTION__
); }));
2479	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2480	int i;
2481	khash_t(ccv_cnnp_parameter_id)kh_ccv_cnnp_parameter_id_t* id_map = 0;
2482	if (count != compiled_data->parameters->rnum)
2483	{
2484		id_map = kh_init(ccv_cnnp_parameter_id)kh_init_ccv_cnnp_parameter_id();
2485		// Build the map between name and the index.
2486		for (i = 0; i < count; i++)
2487		{
2488			int ret;
2489			const khiter_t k = kh_put(ccv_cnnp_parameter_id, id_map, names[i], &ret)kh_put_ccv_cnnp_parameter_id(id_map, names[i], &ret);
2490			assert(ret != 0)((void) sizeof ((ret != 0) ? 1 : 0), __extension__ ({ if (ret
 != 0) ; else __assert_fail ("ret != 0", "ccv_cnnp_model.c", 2490
, __extension__ __PRETTY_FUNCTION__); }));
2491			kh_val(id_map, k)((id_map)->vals[k]) = i;
2492		}
2493	}
2494	const int parameter_size = compiled_data->parameters->rnum;
2495	int* copy_back = 0;
2496	const int tensors_init = !!compiled_data->tensors_init.v;
2497	if (!tensors_init)
2498		ccv_cnnp_model_tensors_init_0(model, compiled_data);
2499	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; });
2500	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));
2501	for (i = 0; i < parameter_size; i++)
2502	{
2503		int j = i;
2504		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)));
2505		if (i >= 0 || strncmp(name, names[i], 1023) != 0)
2506		{
2507			// Build the map.
2508			if (id_map == 0)
2509			{
2510				id_map = kh_init(ccv_cnnp_parameter_id)kh_init_ccv_cnnp_parameter_id();
2511				for (j = 0; j < count; j++)
2512				{
2513					int ret;
2514					const khiter_t k = kh_put(ccv_cnnp_parameter_id, id_map, names[j], &ret)kh_put_ccv_cnnp_parameter_id(id_map, names[j], &ret);
2515					assert(ret != 0)((void) sizeof ((ret != 0) ? 1 : 0), __extension__ ({ if (ret
 != 0) ; else __assert_fail ("ret != 0", "ccv_cnnp_model.c", 2515
, __extension__ __PRETTY_FUNCTION__); }));
2516					kh_val(id_map, k)((id_map)->vals[k]) = j;
2517				}
2518			}
2519			const khiter_t k = kh_get(ccv_cnnp_parameter_id, id_map, name)kh_get_ccv_cnnp_parameter_id(id_map, name);
2520			if (k == kh_end(id_map)((id_map)->n_buckets)) // Cannot find the name, skip.
2521				continue;
2522			j = kh_val(id_map, k)((id_map)->vals[k]);
2523		}
2524		if (compiled_data->tensors.parameters[i]) // Cannot be a shared parameter to read.
2525			{ assert(!((uintptr_t)compiled_data->tensors.parameters[i] & (uintptr_t)1))((void) sizeof ((!((uintptr_t)compiled_data->tensors.parameters
[i] & (uintptr_t)1)) ? 1 : 0), __extension__ ({ if (!((uintptr_t
)compiled_data->tensors.parameters[i] & (uintptr_t)1))
 ; else __assert_fail ("!((uintptr_t)compiled_data->tensors.parameters[i] & (uintptr_t)1)"
, "ccv_cnnp_model.c", 2525, __extension__ __PRETTY_FUNCTION__
); })); }
2526		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)));
2527		ccv_nnc_tensor_param_t info = ccv_nnc_tensor_symbol_params(parameter.graph, parameter);
2528		if (CCV_TENSOR_GET_DEVICE(info.type)((info.type) & 0xfff00) == CCV_COMPUTE_DEVICE_ANY)
2529			CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff
) << 8));
2530		const int d = parameter.d;
2531		if (info.type == tensors[j]->info.type && invalidates) // Can move.
2532		{
2533			// Deallocate it if needed.
2534			if (!((uintptr_t)compiled_data->tensors.parameters[i] & (uintptr_t)1))
2535				if (compiled_data->tensors.parameters[i])
2536					ccv_nnc_tensor_free(compiled_data->tensors.parameters[i]);
2537			compiled_data->tensors.parameters[i] = tensors[j];
2538			tensors[j] = 0;
2539		} else {
2540			if (!compiled_data->tensors.parameters[i])
2541			{ // Not allocated, to allocate first.
2542				// Create new one, make sure we create this by having the right parameters.
2543				const int type = info.type;
2544				info = tensors[j]->info;
2545				info.type = type; // Revert back the type.
2546				compiled_data->tensors.parameters[i] = ccv_nnc_tensor_new(0, info, 0);
2547			}
2548			if (!copy_back)
2549				copy_back = (int*)cccalloccalloc(parameter_size, sizeof(int));
2550			copy_back[i] = j + 1;
2551		}
2552		init_v[d >> 5] |= (1u << (d & 0x1f));
2553		// Create this tensor for other data parallel allocations.
2554		info = compiled_data->tensors.parameters[i]->info; // In case we loaded a different info.
2555		const int device_id = CCV_TENSOR_GET_DEVICE_ID(info.type)(((info.type) & 0xfff00) >> 8);
2556		for (j = 1; j < parallel_count; j++)
2557			if (!compiled_data->tensors.parameters[i + j * parameter_size])
2558			{
2559				if (j != device_id)
2560					CCV_TENSOR_SET_DEVICE_ID(info.type, j)(info.type) = (((info.type) & ~0xfff00) | (((j) & 0xfff
) << 8));
2561				else
2562					CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff
) << 8));
2563				compiled_data->tensors.parameters[i + j * parameter_size] = ccv_nnc_tensor_new(0, info, 0);
2564			}
2565			// No need to copy over, this is done in ccv_cnnp_model.c's copy_tensors method.
2566	}
2567	if (id_map)
2568		kh_destroy(ccv_cnnp_parameter_id, id_map)kh_destroy_ccv_cnnp_parameter_id(id_map);
2569	// Now do the transfer.
2570	if (copy_back)
2571	{
2572		for (i = 0; i < parameter_size; i++)
2573		{
2574			ccv_nnc_tensor_t* const tensor = CCV_NNC_TENSOR(compiled_data->tensors.parameters[i])((ccv_nnc_tensor_t*)((uintptr_t)(compiled_data->tensors.parameters
[i]) & ~(uintptr_t)1));
2575			if (copy_back[i] == 0)
2576				continue;
2577			const int j = copy_back[i] - 1;
2578			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(tensors[j])(ccv_nnc_tensor_t* []){tensors[j]}, (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);
2579		}
2580		ccfreefree(copy_back);
2581	}
2582}
2583 
2584ccv_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)
2585{
2586	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2587	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 2587, __extension__ __PRETTY_FUNCTION__); }));
2588	const int parameter_size = compiled_data->parameters->rnum;
2589	int i;
2590	for (i = 0; i < parameter_size; i++)
2591	{
2592		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)));
2593		if (first(model, name, context))
2594			return ccv_cnnp_model_parameters(model, -1, i);
2595	}
2596	return 0;
2597}
2598 
2599ccv_array_t* ccv_cnnp_model_parameters_filter(ccv_cnnp_model_t* const model, ccv_cnnp_model_parameters_filter_f filter, void* const context)
2600{
2601	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2602	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 2602, __extension__ __PRETTY_FUNCTION__); }));
2603	ccv_array_t* const parameters = ccv_array_new(sizeof(ccv_cnnp_model_io_t), 0, 0);
2604	const int parameter_size = compiled_data->parameters->rnum;
2605	int i;
2606	for (i = 0; i < parameter_size; i++)
2607	{
2608		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)));
2609		if (filter(model, name, context))
2610		{
2611			ccv_cnnp_model_io_t parameter = ccv_cnnp_model_parameters(model, -1, i);
2612			ccv_array_push(parameters, &parameter);
2613		}
2614	}
2615	return parameters;
2616 
2617}
2618 
2619CCV_WARN_UNUSED(ccv_cnnp_model_io_t)ccv_cnnp_model_io_t __attribute__((warn_unused_result)) ccv_cnnp_model_parameter_first_uninit(ccv_cnnp_model_t* const model)
2620{
2621	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2622	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 2622, __extension__ __PRETTY_FUNCTION__); }));
2623	const int tensors_init = !!compiled_data->tensors_init.v;
2624	if (!tensors_init) // If nothing initialized, we return parameter 0.
2625		return ccv_cnnp_model_parameters(model, -1, 0);
2626	const int parameter_size = compiled_data->parameters->rnum;
2627	int i;
2628	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));
2629	for (i = 0; i < parameter_size; i++)
2630	{
2631		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;
2632		if (!(init_v[d >> 5] & (1u << (d & 0x1f))))
2633			return ccv_cnnp_model_parameters(model, -1, i);
2634	}
2635	return 0;
2636}
2637 
2638static 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)
2639{
2640	const int to_param_sel = parameters->param_sel > 0 ? parameters->param_sel - 1 : parameters->param_sel;
2641	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", 2641, __extension__
 __PRETTY_FUNCTION__); }));
2642	ccv_array_t* const to_parameter_indices = ccv_array_new(sizeof(int), 0, 0);
2643	ccv_cnnp_model_add_to_parameter_indices(parameters->model, to_param_sel, to_parameter_indices);
2644	*param_ref = parameters->param_ref > 0 ? parameters->param_ref - 1 : parameters->param_ref;
2645	return to_parameter_indices;
2646}
2647 
2648static 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)
2649{
2650	// If the model is not compiled yet. Compile them now.
2651	if (!model->graph)
2652	{
2653		model->graph = ccv_nnc_symbolic_graph_new();
2654		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", 2654, __extension__ __PRETTY_FUNCTION__
); }));
2655		const int input_size = from_model->input_size;
2656		ccv_nnc_tensor_param_t input_params[input_size];
2657		int i;
2658		for (i = 0; i < input_size; i++)
2659			input_params[i] = ccv_nnc_tensor_symbol_params(from_model->graph, from_model->inputs[i]);
2660		_ccv_cnnp_model_compile(model, input_params, input_size, from_model->compiled_data->loss);
2661		model->parallel_count = from_model->parallel_count;
2662		model->memory_compression = from_model->memory_compression;
2663		model->memory_reduction = from_model->memory_reduction;
2664		model->gradient_checkpointing = from_model->gradient_checkpointing;
2665		model->compiled_data->stream_type = from_model->compiled_data->stream_type;
2666		model->compiled_data->minimize.minimizer = from_model->compiled_data->minimize.minimizer;
2667		model->compiled_data->minimize.max_saved_aux_size = from_model->compiled_data->minimize.max_saved_aux_size;
2668	}
2669	ccv_cnnp_compiled_data_t* const to_compiled_data = model->compiled_data;
2670	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", 2670, __extension__ __PRETTY_FUNCTION__
); }));
2671	const int to_tensors_init = !!to_compiled_data->tensors_init.v;
2672	if (!to_tensors_init)
2673	{
2674		if (only_init_0)
2675			ccv_cnnp_model_tensors_init_0(model, to_compiled_data);
2676		else
2677			_ccv_cnnp_model_tensors_init(model, to_compiled_data);
2678	} else if (!only_init_0 && (uintptr_t)to_compiled_data->tensors_init.v & (uintptr_t)1)
2679		// Check if it is not fully allocated, if it is not, init_1.
2680			ccv_cnnp_model_tensors_init_1(model, to_compiled_data);
2681	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", 2681, __extension__ __PRETTY_FUNCTION__
); }));
2682	*parameter_indices = _ccv_cnnp_model_parameter_indices(model, parameters, param_ref);
2683	*from_parameter_indices = _ccv_cnnp_model_parameter_indices(from_model, from_parameters, from_param_ref);
2684	if (*from_param_ref < 0 && *param_ref >= 0)
2685		{ 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", 2685, __extension__ __PRETTY_FUNCTION__
); })); }
2686	else if (*from_param_ref >= 0)
2687		{ 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", 2687, __extension__ __PRETTY_FUNCTION__
); })); }
2688	if (*param_ref < 0 && *from_param_ref >= 0)
2689		{ 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"
, 2689, __extension__ __PRETTY_FUNCTION__); })); }
2690	else if (*param_ref >= 0)
2691		{ 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", 2691, __extension__ __PRETTY_FUNCTION__
); })); }
2692}
2693 
2694void 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)
2695{
2696	ccv_array_t* to_parameter_indices;
2697	int to_param_ref;
2698	ccv_array_t* from_parameter_indices;
2699	int from_param_ref;
2700	_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);
2701	// Should be exactly the same tensor.
2702	if (to_param_ref < 0 && from_param_ref < 0)
2703		{ 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", 2703, __extension__ __PRETTY_FUNCTION__
); })); }
2704	// To models.
2705	ccv_cnnp_compiled_data_t* const to_compiled_data = model->compiled_data;
2706	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", 2706, __extension__ __PRETTY_FUNCTION__
); }));
2707	// From models.
2708	const ccv_cnnp_compiled_data_t* const from_compiled_data = from_model->compiled_data;
2709	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; });
2710	const int to_parameter_size = to_compiled_data->parameters->rnum;
2711	const int rnum = (to_param_ref < 0 && from_param_ref < 0) ? from_parameter_indices->rnum : 1;
2712	int i, j;
2713	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));
2714  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));
2715	for (i = 0; i < rnum; i++)
2716	{
2717		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)));
2718		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"
, 2718, __extension__ __PRETTY_FUNCTION__); }));
2719		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", 2719, __extension__ __PRETTY_FUNCTION__
); }));
2720		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;
2721		// If the original is not init'ed. We cannot copy from.
2722		if (!(from_init_v[s >> 5] & (1u << (s & 0x1f))))
2723			continue;
2724		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)));
2725		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"
, 2725, __extension__ __PRETTY_FUNCTION__); }));
2726		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", 2726, __extension__ __PRETTY_FUNCTION__
); }));
2727		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));
2728		assert(src)((void) sizeof ((src) ? 1 : 0), __extension__ ({ if (src) ; else
 __assert_fail ("src", "ccv_cnnp_model.c", 2728, __extension__
 __PRETTY_FUNCTION__); }));
2729		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));
2730		assert(dest)((void) sizeof ((dest) ? 1 : 0), __extension__ ({ if (dest) ;
 else __assert_fail ("dest", "ccv_cnnp_model.c", 2730, __extension__
 __PRETTY_FUNCTION__); }));
2731		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);
2732		for (j = 1; j < parallel_count; j++)
2733		{
2734			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));
2735			if (copy_tensor)
2736				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);
2737		}
2738		// Mark this symbol as init'ed.
2739		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;
2740		to_init_v[d >> 5] |= (1u << (d & 0x1f));
2741	}
2742	ccv_array_free(to_parameter_indices);
2743	ccv_array_free(from_parameter_indices);
2744}
2745 
2746void 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)
2747{
2748	ccv_array_t* to_parameter_indices;
2749	int to_param_ref;
2750	ccv_array_t* from_parameter_indices;
2751	int from_param_ref;
2752	_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);
2753	// Should be exactly the same tensor.
2754	if (renamer == 0 && to_param_ref < 0 && from_param_ref < 0)
1
Assuming 'renamer' is not equal to null→
2755		{ 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", 2755, __extension__ __PRETTY_FUNCTION__
); })); }
2756	// To models.
2757	ccv_cnnp_compiled_data_t* const to_compiled_data = model->compiled_data;
2758	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", 2758, __extension__ __PRETTY_FUNCTION__
); }));
2
←
Assuming 'to_compiled_data' is non-null→
3
←
Taking true branch→
2759	// From models.
2760	const ccv_cnnp_compiled_data_t* const from_compiled_data = from_model->compiled_data;
2761	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→
2762	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", 2762, __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→
2763	const int from_parameter_size = from_compiled_data->parameters->rnum;
2764	const int to_parameter_size = to_compiled_data->parameters->rnum;
2765	const int rnum = (to_param_ref < 0 && from_param_ref < 0) ? to_parameter_indices->rnum : 1;
9
←
Assuming 'to_param_ref' is >= 0→
2766	int i, j;
2767	khash_t(ccv_cnnp_parameter_id)kh_ccv_cnnp_parameter_id_t* id_map = 0;
2768	char* updated_name = 0;
2769	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));
2770  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));
2771	for (i = 0; i < rnum; i++)
2772	{
2773		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→
2774		// Need to figure out how to use the renamer here.
2775		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→
2776		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"
, 2776, __extension__ __PRETTY_FUNCTION__); }));
15
←
Assuming 'dest_d' is >= 0→
16
←
Taking true branch→
2777		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", 2777, __extension__
 __PRETTY_FUNCTION__); }));
17
←
Assuming 'dest_d' is < 'to_parameter_size'→
18
←
Taking true branch→
2778		if (renamer18.1
'renamer' is non-null
)
2779		{
2780			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;
2781			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)));
2782			if (!updated_name18.3
'updated_name' is null
)
19
←
Taking true branch→
2783				updated_name = (char*)ccmallocmalloc(1024);
2784			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→
2785			if (src_name_len20.1
'src_name_len' is <= 0
 > 0)
21
←
Taking false branch→
2786				memcpy(updated_name, src_name, src_name_len);
2787			updated_name[src_name_len] = 0;
2788			if (renamer(context, dest_name, updated_name, 1024) != 0)
22
←
Assuming the condition is false→
2789				continue; // Skip this.
2790			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)
2791			{
2792				// Nothing changed.
2793			} else {
2794				if (!id_map22.2
'id_map' is null
)
23
←
Taking true branch→
2795				{
2796					id_map = kh_init(ccv_cnnp_parameter_id)kh_init_ccv_cnnp_parameter_id();
2797					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 2805→
2798					{
2799						int ret;
2800						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'→
2801						assert(ret != 0)((void) sizeof ((ret != 0) ? 1 : 0), __extension__ ({ if (ret
 != 0) ; else __assert_fail ("ret != 0", "ccv_cnnp_model.c", 2801
, __extension__ __PRETTY_FUNCTION__); }));
53
←
Taking true branch→
2802						kh_val(id_map, k)((id_map)->vals[k]) = j;
2803					}
2804				}
2805				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'→
2806				if (k == kh_end(id_map)((id_map)->n_buckets)) // Cannot find the name, skip.
2807					continue;
2808				src_d = kh_val(id_map, k)((id_map)->vals[k]);
2809				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"
, 2809, __extension__ __PRETTY_FUNCTION__); }));
2810				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", 2810, __extension__
 __PRETTY_FUNCTION__); }));
2811			}
2812		}
2813		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"
, 2813, __extension__ __PRETTY_FUNCTION__); }));
2814		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", 2814, __extension__
 __PRETTY_FUNCTION__); }));
2815		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;
2816		// If the original is not init'ed. We cannot share from.
2817		if (!(from_init_v[s >> 5] & (1u << (s & 0x1f))))
2818			continue;
2819		for (j = 0; j < parallel_count; j++)
2820		{
2821			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));
2822			assert(src)((void) sizeof ((src) ? 1 : 0), __extension__ ({ if (src) ; else
 __assert_fail ("src", "ccv_cnnp_model.c", 2822, __extension__
 __PRETTY_FUNCTION__); }));
2823			ccv_nnc_tensor_t* const dest = to_compiled_data->tensors.parameters[dest_d + j * to_parameter_size];
2824			if (dest && !((uintptr_t)dest & (uintptr_t)1))
2825				ccv_nnc_tensor_free(dest);
2826			to_compiled_data->tensors.parameters[dest_d + j * to_parameter_size] = (ccv_nnc_tensor_t*)((uintptr_t)src | (uintptr_t)1);
2827		}
2828		// Mark this symbol as init'ed.
2829		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;
2830		to_init_v[d >> 5] |= (1u << (d & 0x1f));
2831	}
2832	ccv_array_free(to_parameter_indices);
2833	ccv_array_free(from_parameter_indices);
2834	if (id_map)
2835		kh_destroy(ccv_cnnp_parameter_id, id_map)kh_destroy_ccv_cnnp_parameter_id(id_map);
2836	if (updated_name)
2837		ccfreefree(updated_name);
2838	// Mark it as incomplete so we will call init_1.
2839	if (ccv_cnnp_model_tensors_any_to_alloc(model, to_compiled_data))
2840		to_compiled_data->tensors_init.v = (uint32_t*)((uintptr_t)to_compiled_data->tensors_init.v | (uintptr_t)1);
2841	else // Remove the flag.
2842		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));
2843}
2844 
2845ccv_nnc_stream_context_t* ccv_cnnp_compiled_data_get_stream(ccv_cnnp_compiled_data_t* const compiled_data, const int type)
2846{
2847	if (!compiled_data->stream_map)
2848		compiled_data->stream_map = kh_init(stream_map)kh_init_stream_map();
2849	int ret = 0;
2850	khiter_t k = kh_put(stream_map, compiled_data->stream_map, type, &ret)kh_put_stream_map(compiled_data->stream_map, type, &ret
);
2851	assert(ret >= 0)((void) sizeof ((ret >= 0) ? 1 : 0), __extension__ ({ if (
ret >= 0) ; else __assert_fail ("ret >= 0", "ccv_cnnp_model.c"
, 2851, __extension__ __PRETTY_FUNCTION__); }));
2852	ccv_nnc_stream_context_t* stream = kh_val(compiled_data->stream_map, k)((compiled_data->stream_map)->vals[k]);
2853	// If ret == 0, the key already exist, we can return directly, otherwise, create and return.
2854	if (ret != 0)
2855	{
2856		stream = ccv_nnc_stream_context_new(type);
2857		kh_val(compiled_data->stream_map, k)((compiled_data->stream_map)->vals[k]) = stream;
2858	}
2859	return stream;
2860}
2861 
2862void 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)
2863{
2864	ccv_array_t* to_parameter_indices;
2865	int to_param_ref;
2866	ccv_array_t* from_parameter_indices;
2867	int from_param_ref;
2868	_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);
2869	// Should be exactly the same tensor.
2870	if (to_param_ref < 0 && from_param_ref < 0)
2871		{ 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", 2871, __extension__ __PRETTY_FUNCTION__
); })); }
2872	// To models.
2873	ccv_cnnp_compiled_data_t* const to_compiled_data = model->compiled_data;
2874	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", 2874, __extension__ __PRETTY_FUNCTION__
); }));
2875	// From models.
2876	const ccv_cnnp_compiled_data_t* const from_compiled_data = from_model->compiled_data;
2877	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; });
2878	const int to_parameter_size = to_compiled_data->parameters->rnum;
2879	const int rnum = (to_param_ref < 0 && from_param_ref < 0) ? from_parameter_indices->rnum : 1;
2880	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", 2880, __extension__ __PRETTY_FUNCTION__
); }));
2881	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", 2881, __extension__ __PRETTY_FUNCTION__
); }));
2882	int i, j;
2883	ccv_nnc_tensor_t* inputs[aux_in_size + 2];
2884	ccv_nnc_tensor_t* outputs[aux_out_size + 1];
2885	for (i = 0; i < aux_in_size; i++)
2886		inputs[i + 2] = aux_ins[i];
2887	for (i = 0; i < aux_out_size; i++)
2888		outputs[i + 1] = aux_outs[i];
2889	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));
2890	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));
2891	for (i = 0; i < rnum; i++)
2892	{
2893		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)));
2894		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"
, 2894, __extension__ __PRETTY_FUNCTION__); }));
2895		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", 2895, __extension__ __PRETTY_FUNCTION__
); }));
2896		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;
2897		// If the original is not init'ed. We cannot copy from.
2898		if (!(from_init_v[s >> 5] & (1u << (s & 0x1f))))
2899			continue;
2900		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)));
2901		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"
, 2901, __extension__ __PRETTY_FUNCTION__); }));
2902		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", 2902, __extension__ __PRETTY_FUNCTION__
); }));
2903		if (parallel_count > 1)
2904		{
2905			ccv_nnc_stream_context_t* streams[parallel_count];
2906			ccv_nnc_stream_signal_t* signal;
2907			if (stream_context)
2908				signal = ccv_nnc_stream_context_emit_signal_new(stream_context);
2909			for (j = 0; j < parallel_count; j++)
2910			{
2911				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));
2912				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));
2913				if (!dest || !src)
2914				{
2915					streams[j] = 0;
2916					continue;
2917				}
2918				// At the moment, can only handle them on the same device.
2919				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", 2919, __extension__ __PRETTY_FUNCTION__
); }));
2920				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", 2920, __extension__ __PRETTY_FUNCTION__
); }));
2921				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;
2922				const int device_id = CCV_TENSOR_GET_DEVICE_ID(src->info.type)(((src->info.type) & 0xfff00) >> 8);
2923				int type = stream_type;
2924				CCV_STREAM_SET_DEVICE_ID(type, device_id)(type) = (((type) & ~0xfff00) | (((device_id) & 0xfff
) << 8));
2925				ccv_nnc_stream_context_t* const stream_0 = ccv_cnnp_compiled_data_get_stream(to_compiled_data, type);
2926				// Wait signal to finish.
2927				if (stream_context)
2928					ccv_nnc_stream_context_wait_signal(stream_0, signal);
2929				inputs[0] = outputs[0] = dest;
2930				inputs[1] = src;
2931				ccv_nnc_cmd_exec(cmd, hint, flags, inputs, aux_in_size + 2, outputs, aux_out_size + 1, stream_0);
2932				if (stream_context)
2933				{
2934					ccv_nnc_stream_signal_t* const signal = ccv_nnc_stream_context_emit_signal_new(stream_0);
2935					ccv_nnc_stream_context_wait_signal(stream_context, signal);
2936				}
2937				streams[j] = stream_0;
2938			}
2939			// If this should be blocking, blocking it.
2940			if (!stream_context)
2941				for (j = 0; j < parallel_count; j++)
2942					if (streams[j])
2943						ccv_nnc_stream_context_wait(streams[j]);
2944		} else {
2945			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));
2946			assert(src)((void) sizeof ((src) ? 1 : 0), __extension__ ({ if (src) ; else
 __assert_fail ("src", "ccv_cnnp_model.c", 2946, __extension__
 __PRETTY_FUNCTION__); }));
2947			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));
2948			assert(dest)((void) sizeof ((dest) ? 1 : 0), __extension__ ({ if (dest) ;
 else __assert_fail ("dest", "ccv_cnnp_model.c", 2948, __extension__
 __PRETTY_FUNCTION__); }));
2949			inputs[0] = outputs[0] = dest;
2950			inputs[1] = src;
2951			ccv_nnc_cmd_exec(cmd, hint, flags, inputs, aux_in_size + 2, outputs, aux_out_size + 1, stream_context);
2952		}
2953		// Mark this symbol as init'ed.
2954		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;
2955		to_init_v[d >> 5] |= (1u << (d & 0x1f));
2956	}
2957	ccv_array_free(to_parameter_indices);
2958	ccv_array_free(from_parameter_indices);
2959}
2960 
2961void 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)
2962{
2963	int to_param_ref;
2964	ccv_array_t* const to_parameter_indices = _ccv_cnnp_model_parameter_indices(model, parameters, &to_param_ref);
2965	// To models.
2966	ccv_cnnp_compiled_data_t* const to_compiled_data = model->compiled_data;
2967	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", 2967, __extension__ __PRETTY_FUNCTION__
); }));
2968	// Tensor has to be inited already.
2969	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", 2969, __extension__ __PRETTY_FUNCTION__
); }));
2970	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", 2970, __extension__ __PRETTY_FUNCTION__
); }));
2971	// From models.
2972	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; });
2973	const int to_parameter_size = to_compiled_data->parameters->rnum;
2974	const int rnum = (to_param_ref < 0) ? to_parameter_indices->rnum : 1;
2975	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", 2975, __extension__ __PRETTY_FUNCTION__
); }));
2976	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", 2976, __extension__ __PRETTY_FUNCTION__
); }));
2977	int i, j;
2978	ccv_nnc_tensor_t* inputs[aux_in_size + 1];
2979	ccv_nnc_tensor_t* outputs[aux_out_size + 1];
2980	for (i = 0; i < aux_in_size; i++)
2981		inputs[i + 1] = aux_ins[i];
2982	for (i = 0; i < aux_out_size; i++)
2983		outputs[i + 1] = aux_outs[i];
2984	for (i = 0; i < rnum; i++)
2985	{
2986		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)));
2987		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"
, 2987, __extension__ __PRETTY_FUNCTION__); }));
2988		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", 2988, __extension__ __PRETTY_FUNCTION__
); }));
2989		if (parallel_count > 1)
2990		{
2991			ccv_nnc_stream_context_t* streams[parallel_count];
2992			ccv_nnc_stream_signal_t* signal;
2993			if (stream_context)
2994				signal = ccv_nnc_stream_context_emit_signal_new(stream_context);
2995			for (j = 0; j < parallel_count; j++)
2996			{
2997				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));
2998				if (!dest)
2999				{
3000					streams[j] = 0;
3001					continue;
3002				}
3003				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;
3004				const int device_id = CCV_TENSOR_GET_DEVICE_ID(dest->info.type)(((dest->info.type) & 0xfff00) >> 8);
3005				int type = stream_type;
3006				CCV_STREAM_SET_DEVICE_ID(type, device_id)(type) = (((type) & ~0xfff00) | (((device_id) & 0xfff
) << 8));
3007				ccv_nnc_stream_context_t* const stream_0 = ccv_cnnp_compiled_data_get_stream(to_compiled_data, type);
3008				// Wait signal to finish.
3009				if (stream_context)
3010					ccv_nnc_stream_context_wait_signal(stream_0, signal);
3011				inputs[0] = outputs[0] = dest;
3012				ccv_nnc_cmd_exec(cmd, hint, flags, inputs, aux_in_size + 1, outputs, aux_out_size + 1, stream_0);
3013				if (stream_context)
3014				{
3015					ccv_nnc_stream_signal_t* const signal = ccv_nnc_stream_context_emit_signal_new(stream_0);
3016					ccv_nnc_stream_context_wait_signal(stream_context, signal);
3017				}
3018				streams[j] = stream_0;
3019			}
3020			// If this should be blocking, blocking it.
3021			if (!stream_context)
3022				for (j = 0; j < parallel_count; j++)
3023					if (streams[j])
3024						ccv_nnc_stream_context_wait(streams[j]);
3025		} else {
3026			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));
3027			assert(dest)((void) sizeof ((dest) ? 1 : 0), __extension__ ({ if (dest) ;
 else __assert_fail ("dest", "ccv_cnnp_model.c", 3027, __extension__
 __PRETTY_FUNCTION__); }));
3028			inputs[0] = outputs[0] = dest;
3029			ccv_nnc_cmd_exec(cmd, hint, flags, inputs, aux_in_size + 1, outputs, aux_out_size + 1, stream_context);
3030		}
3031		// No need to mark this symbol as init'ed, it is already.
3032	}
3033	ccv_array_free(to_parameter_indices);
3034}
3035 
3036void 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)
3037{
3038	int to_param_ref;
3039	ccv_array_t* const to_parameter_indices = _ccv_cnnp_model_parameter_indices(model, parameters, &to_param_ref);
3040	// To models.
3041	ccv_cnnp_compiled_data_t* const to_compiled_data = model->compiled_data;
3042	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", 3042, __extension__ __PRETTY_FUNCTION__
); }));
3043	// Tensor has to be inited already.
3044	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", 3044, __extension__ __PRETTY_FUNCTION__
); }));
3045	ccv_nnc_tensor_t** tensor_gradients;
3046	if (to_compiled_data->backward.count > 1)
3047		tensor_gradients = to_compiled_data->tensors.accum_gradients;
3048	else
3049		tensor_gradients = to_compiled_data->tensors.gradients;
3050	assert(tensor_gradients)((void) sizeof ((tensor_gradients) ? 1 : 0), __extension__ ({
 if (tensor_gradients) ; else __assert_fail ("tensor_gradients"
, "ccv_cnnp_model.c", 3050, __extension__ __PRETTY_FUNCTION__
); }));
3051	// From models.
3052	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; });
3053	const int to_parameter_size = to_compiled_data->parameters->rnum;
3054	const int rnum = (to_param_ref < 0) ? to_parameter_indices->rnum : 1;
3055	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", 3055, __extension__ __PRETTY_FUNCTION__
); }));
3056	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", 3056, __extension__ __PRETTY_FUNCTION__
); }));
3057	int i, j;
3058	ccv_nnc_tensor_t* inputs[aux_in_size + 1];
3059	ccv_nnc_tensor_t* outputs[aux_out_size + 1];
3060	for (i = 0; i < aux_in_size; i++)
3061		inputs[i + 1] = aux_ins[i];
3062	for (i = 0; i < aux_out_size; i++)
3063		outputs[i + 1] = aux_outs[i];
3064	for (i = 0; i < rnum; i++)
3065	{
3066		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)));
3067		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"
, 3067, __extension__ __PRETTY_FUNCTION__); }));
3068		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", 3068, __extension__ __PRETTY_FUNCTION__
); }));
3069		if (parallel_count > 1)
3070		{
3071			ccv_nnc_stream_context_t* streams[parallel_count];
3072			ccv_nnc_stream_signal_t* signal;
3073			if (stream_context)
3074				signal = ccv_nnc_stream_context_emit_signal_new(stream_context);
3075			for (j = 0; j < parallel_count; j++)
3076			{
3077				ccv_nnc_tensor_t* const dest = tensor_gradients[dest_d + j * to_parameter_size];
3078				if (!dest)
3079				{
3080					streams[j] = 0;
3081					continue;
3082				}
3083				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;
3084				const int device_id = CCV_TENSOR_GET_DEVICE_ID(dest->info.type)(((dest->info.type) & 0xfff00) >> 8);
3085				int type = stream_type;
3086				CCV_STREAM_SET_DEVICE_ID(type, device_id)(type) = (((type) & ~0xfff00) | (((device_id) & 0xfff
) << 8));
3087				ccv_nnc_stream_context_t* const stream_0 = ccv_cnnp_compiled_data_get_stream(to_compiled_data, type);
3088				// Wait signal to finish.
3089				if (stream_context)
3090					ccv_nnc_stream_context_wait_signal(stream_0, signal);
3091				inputs[0] = outputs[0] = dest;
3092				ccv_nnc_cmd_exec(cmd, hint, flags, inputs, aux_in_size + 1, outputs, aux_out_size + 1, stream_0);
3093				if (stream_context)
3094				{
3095					ccv_nnc_stream_signal_t* const signal = ccv_nnc_stream_context_emit_signal_new(stream_0);
3096					ccv_nnc_stream_context_wait_signal(stream_context, signal);
3097				}
3098				streams[j] = stream_0;
3099			}
3100			// If this should be blocking, blocking it.
3101			if (!stream_context)
3102				for (j = 0; j < parallel_count; j++)
3103					if (streams[j])
3104						ccv_nnc_stream_context_wait(streams[j]);
3105		} else {
3106			ccv_nnc_tensor_t* const dest = tensor_gradients[dest_d];
3107			if (!dest)
3108				continue;
3109			assert(dest)((void) sizeof ((dest) ? 1 : 0), __extension__ ({ if (dest) ;
 else __assert_fail ("dest", "ccv_cnnp_model.c", 3109, __extension__
 __PRETTY_FUNCTION__); }));
3110			inputs[0] = outputs[0] = dest;
3111			ccv_nnc_cmd_exec(cmd, hint, flags, inputs, aux_in_size + 1, outputs, aux_out_size + 1, stream_context);
3112		}
3113		// No need to mark this symbol as init'ed, it is already.
3114	}
3115	ccv_array_free(to_parameter_indices);
3116}
3117 
3118void ccv_cnnp_model_parameters_to_unified_memory(ccv_cnnp_model_t* const model, const ccv_cnnp_model_io_t parameters, ccv_nnc_stream_context_t* const stream_context)
3119{
3120	// Only CUDA backend has this feature.
3121#ifdef HAVE_CUDA1
3122	int to_param_ref;
3123	ccv_array_t* const to_parameter_indices = _ccv_cnnp_model_parameter_indices(model, parameters, &to_param_ref);
3124	// To models.
3125	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
3126	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 3126, __extension__ __PRETTY_FUNCTION__); }));
3127	// Tensor has to be inited already.
3128	assert(!!compiled_data->tensors_init.v)((void) sizeof ((!!compiled_data->tensors_init.v) ? 1 : 0)
, __extension__ ({ if (!!compiled_data->tensors_init.v) ; else
 __assert_fail ("!!compiled_data->tensors_init.v", "ccv_cnnp_model.c"
, 3128, __extension__ __PRETTY_FUNCTION__); }));
3129	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", 3129, __extension__ __PRETTY_FUNCTION__
); }));
3130	// From models.
3131	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; });
3132	const int rnum = (to_param_ref < 0) ? to_parameter_indices->rnum : 1;
3133	int i;
3134	for (i = 0; i < rnum; i++)
3135	{
3136		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)));
3137		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"
, 3137, __extension__ __PRETTY_FUNCTION__); }));
3138		assert(dest_d < compiled_data->parameters->rnum)((void) sizeof ((dest_d < compiled_data->parameters->
rnum) ? 1 : 0), __extension__ ({ if (dest_d < compiled_data
->parameters->rnum) ; else __assert_fail ("dest_d < compiled_data->parameters->rnum"
, "ccv_cnnp_model.c", 3138, __extension__ __PRETTY_FUNCTION__
); }));
3139		if (parallel_count > 1)
3140		{
3141			assert(0 && "Cannot support this when data parallel is in effect.")((void) sizeof ((0 && "Cannot support this when data parallel is in effect."
) ? 1 : 0), __extension__ ({ if (0 && "Cannot support this when data parallel is in effect."
) ; else __assert_fail ("0 && \"Cannot support this when data parallel is in effect.\""
, "ccv_cnnp_model.c", 3141, __extension__ __PRETTY_FUNCTION__
); }));
3142		} else {
3143			ccv_nnc_tensor_t* const src = CCV_NNC_TENSOR(compiled_data->tensors.parameters[dest_d])((ccv_nnc_tensor_t*)((uintptr_t)(compiled_data->tensors.parameters
[dest_d]) & ~(uintptr_t)1));
3144			assert(src)((void) sizeof ((src) ? 1 : 0), __extension__ ({ if (src) ; else
 __assert_fail ("src", "ccv_cnnp_model.c", 3144, __extension__
 __PRETTY_FUNCTION__); }));
3145			ccv_nnc_tensor_param_t params = src->info;
3146			if (CCV_TENSOR_GET_MEMORY(params.type)((params.type) & 0x3) != CCV_TENSOR_GPU_MEMORY)
3147				continue;
3148			const size_t size = ccv_nnc_tensor_data_size(params);
3149			if (size <= 0)
3150				continue;
3151			const int should_free = !((uintptr_t)compiled_data->tensors.parameters[dest_d] & (uintptr_t)1);
3152			const int tfb = (CCV_TENSOR_GET_MEMORY(params.type)((params.type) & 0x3) == CCV_TENSOR_CPU_MEMORY && params.format == CCV_TENSOR_FORMAT_NHWC && params.dim[2] > 0 && params.dim[2] <= CCV_MAX_CHANNEL(0xFFF) && params.dim[0] > 0 && params.dim[1] > 0 && params.dim[3] == 0);
3153			ccv_nnc_tensor_t* const tensor = (ccv_nnc_tensor_t*)ccmallocmalloc(sizeof(ccv_nnc_tensor_t));
3154			tensor->dataof = 0;
3155			tensor->alias_ref = 0;
3156			tensor->sig = 0;
3157			tensor->refcount = 1;
3158			tensor->info = params;
3159			if (tfb)
3160			{
3161				tensor->type = CCV_NO_DATA_ALLOC | CCV_MATRIX_DENSE | CCV_GET_DATA_TYPE(params.datatype)((params.datatype) & 0xFF000) | params.dim[2];
3162				// This corresponding to mat->step
3163				tensor->info.dim[4] = CCV_GET_STEP(params.dim[1], (CCV_GET_DATA_TYPE(params.datatype) | params.dim[2]))(((params.dim[1]) * _ccv_get_data_type_size[(((((params.datatype
) & 0xFF000) | params.dim[2])) & 0xFF000) >> 12
] * (((((params.datatype) & 0xFF000) | params.dim[2])) &
 0xFFF) + 3) & -4);
3164			} else // This won't be recognized by ccv_dense_matrix_t
3165				tensor->type = CCV_NO_DATA_ALLOC | CCV_MATRIX_DENSE | CCV_GET_DATA_TYPE(params.datatype)((params.datatype) & 0xFF000);
3166			// Remove this flag so it can be deallocated as usual.
3167			tensor->type &= ~CCV_NO_DATA_ALLOC;
3168			assert(CCV_TENSOR_GET_DEVICE(params.type) != CCV_COMPUTE_DEVICE_ANY)((void) sizeof ((((params.type) & 0xfff00) != CCV_COMPUTE_DEVICE_ANY
) ? 1 : 0), __extension__ ({ if (((params.type) & 0xfff00
) != CCV_COMPUTE_DEVICE_ANY) ; else __assert_fail ("CCV_TENSOR_GET_DEVICE(params.type) != CCV_COMPUTE_DEVICE_ANY"
, "ccv_cnnp_model.c", 3168, __extension__ __PRETTY_FUNCTION__
); }));
3169			void* ptr = cumallocmanaged(CCV_TENSOR_GET_DEVICE_ID(params.type)(((params.type) & 0xfff00) >> 8), size);
3170			if (ptr) // If allocated successfully. Otherwise we go through the fallback path.
3171			{
3172				tensor->data.u8 = (uint8_t*)ptr;
3173				tensor->type |= CCV_MAPPED_MEM; // This denotes the tensor is mapped to CPU, and would prefer a explicit prefetch call.
3174			} else {
3175				// Allocation failed.
3176				ccfreefree(tensor);
3177				continue;
3178			}
3179			// TODO: Cannot run this on the stream context yet, due to allocation and deallocations.
3180			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, &src, 1, &tensor, 1, 0);
3181			cumemadvisereadmostly(CCV_TENSOR_GET_DEVICE_ID(params.type)(((params.type) & 0xfff00) >> 8), tensor->data.u8, size);
3182			compiled_data->tensors.parameters[dest_d] = tensor;
3183			// Can free out the old one.
3184			if (should_free)
3185				ccv_nnc_tensor_free(src);
3186		}
3187		// No need to mark this symbol as init'ed, it is already.
3188	}
3189	ccv_array_free(to_parameter_indices);
3190#endif
3191}
3192 
3193ccv_nnc_cmd_t ccv_cnnp_model_minimizer(ccv_cnnp_model_t* const model)
3194{
3195	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
3196	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 3196, __extension__ __PRETTY_FUNCTION__); }));
3197	return compiled_data->minimize.minimizer;
3198}
3199 
3200void 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)
3201{
3202	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
3203	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 3203, __extension__ __PRETTY_FUNCTION__); }));
3204	const int parameter_size = compiled_data->parameters->rnum;
3205	if (parameter_size == 0)
3206		return;
3207	if (reset)
3208		{ 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", 3208, __extension__ __PRETTY_FUNCTION__
); })); }
3209	const int old_max_saved_aux_size = compiled_data->minimize.max_saved_aux_size;
3210	const int saved_aux_size = ccv_nnc_minimizer_saved_aux_size(minimizer);
3211	if (saved_aux_size > compiled_data->minimize.max_saved_aux_size)
3212		compiled_data->minimize.max_saved_aux_size = saved_aux_size;
3213	const int max_saved_aux_size = compiled_data->minimize.max_saved_aux_size;
3214	// We update all parameters, at this point, we have one minimizer.
3215	if (set_parameters == 0 || set_parameter_size == 0)
3216		compiled_data->minimize.minimizer = minimizer;
3217	int i;
3218	if (set_parameters && set_parameter_size)
3219	{
3220		// I need to save what's the minimizer along with this.
3221		if (!compiled_data->minimize.parameters)
3222			compiled_data->minimize.parameters = ccv_array_new(sizeof(ccv_cnnp_set_minimizer_for_parameter_t*), 1, 0);
3223		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));
3224		set_minimizer_for_parameter->minimizer = minimizer;
3225		set_minimizer_for_parameter->parameter_size = set_parameter_size;
3226		memcpy(set_minimizer_for_parameter->parameters, set_parameters, sizeof(ccv_cnnp_model_io_t) * set_parameter_size);
3227		ccv_array_push(compiled_data->minimize.parameters, &set_minimizer_for_parameter);
3228	}
3229	// If reset is true, clear the parameters array.
3230	if (reset && compiled_data->minimize.parameters)
3231	{
3232		for (i = 0; i < compiled_data->minimize.parameters->rnum; i++)
3233			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))));
3234		ccv_array_clear(compiled_data->minimize.parameters);
3235	}
3236	if (!compiled_data->update_nodes)
3237		return;
3238	ccv_nnc_symbolic_graph_t* const symbolic_graph = model->graph;
3239	assert(symbolic_graph)((void) sizeof ((symbolic_graph) ? 1 : 0), __extension__ ({ if
 (symbolic_graph) ; else __assert_fail ("symbolic_graph", "ccv_cnnp_model.c"
, 3239, __extension__ __PRETTY_FUNCTION__); }));
3240	if (saved_aux_size > old_max_saved_aux_size)
3241	{
3242		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", 3242, __extension__ __PRETTY_FUNCTION__
); }));
3243		// Reallocate first, move them around later.
3244		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);
3245		compiled_data->update_nodes = (ccv_nnc_graph_exec_symbol_t*)(compiled_data->updated_parameters + parameter_size);
3246		compiled_data->saved_aux = (ccv_nnc_tensor_symbol_map_t*)(compiled_data->update_nodes + parameter_size);
3247		// We need to do this from back to front because saved_aux_size > old_saved_aux_size, it could overlap.
3248		_ccv_cnnp_scatter_saved_aux(compiled_data->saved_aux, parameter_size, old_max_saved_aux_size, saved_aux_size);
3249	}
3250	int flag = 0;
3251	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; });
3252	if (set_parameters && set_parameter_size)
3253	{
3254		ccv_array_t* const parameter_indices = ccv_array_new(sizeof(int), 0, 0);
3255		for (i = 0; i < set_parameter_size; i++)
3256		{
3257			const int param_sel = set_parameters[i]->param_sel > 0 ? set_parameters[i]->param_sel - 1 : set_parameters[i]->param_sel;
3258			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", 3258, __extension__ __PRETTY_FUNCTION__
); }));
3259			const int old_rnum = parameter_indices->rnum;
3260			ccv_cnnp_model_add_to_parameter_indices(set_parameters[i]->model, param_sel, parameter_indices);
3261			const int param_ref = set_parameters[i]->param_ref > 0 ? set_parameters[i]->param_ref - 1 : set_parameters[i]->param_ref;
3262			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", 3262, __extension__ __PRETTY_FUNCTION__
); }));
3263			if (param_ref >= 0)
3264			{
3265				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", 3265, __extension__ __PRETTY_FUNCTION__
); }));
3266				*(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)));
3267				parameter_indices->rnum = old_rnum + 1;
3268			}
3269		}
3270		// We may have duplicated indices, but that is OK, we will set it twice.
3271		for (i = 0; i < parameter_indices->rnum; i++)
3272		{
3273			const int d = *(int*)ccv_array_get(parameter_indices, i)((void*)(((char*)((parameter_indices)->data)) + (size_t)(parameter_indices
)->rsize * (size_t)(i)));
3274			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))
3275				flag = 1;
3276		}
3277		ccv_array_free(parameter_indices);
3278	} else {
3279		for (i = 0; i < parameter_size; i++)
3280			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))
3281				flag = 1;
3282		if (compiled_data->minimize.parameters)
3283			if (_ccv_cnnp_apply_parameters_with_minimizer(model))
3284				flag = 1;
3285	}
3286	if (flag)
3287	{
3288		// 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.
3289		if (compiled_data->graph_mode == CCV_CNNP_MODEL_GRAPH_FIT_MODE)
3290			_ccv_cnnp_compiled_data_graph_free(compiled_data);
3291		_ccv_cnnp_compiled_data_apply_gradients_free(compiled_data);
3292	}
3293}
3294 
3295void ccv_cnnp_model_set_compile_params(ccv_cnnp_model_t* const model, const ccv_nnc_symbolic_graph_compile_param_t compile_params)
3296{
3297	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
3298	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 3298, __extension__ __PRETTY_FUNCTION__); }));
3299	compiled_data->compile_params = compile_params;
3300}
3301 
3302void ccv_cnnp_model_dot(const ccv_cnnp_model_t* const model, const int flags, FILE** const outs, const int out_size)
3303{
3304	if (model->graph && out_size > 0)
3305		ccv_nnc_symbolic_graph_dot(model->graph, flags, outs[0]);
3306	if (model->compiled_data && model->compiled_data->graph && out_size > 1)
3307		ccv_nnc_graph_dot(model->compiled_data->graph, flags, outs[1]);
3308	if (model->compiled_data && model->compiled_data->backward.accum && out_size > 2)
3309		ccv_nnc_graph_dot(model->compiled_data->backward.accum, flags, outs[2]);
3310	if (model->compiled_data && model->compiled_data->apply_gradients.graph && out_size > 3)
3311		ccv_nnc_graph_dot(model->compiled_data->apply_gradients.graph, flags, outs[3]);
3312}
3313 
3314void ccv_cnnp_model_format(const ccv_cnnp_model_t* const model, const ccv_nnc_symbolic_graph_format_f format_fn, void* const context)
3315{
3316	if (model->graph)
3317		ccv_nnc_symbolic_graph_format(model->graph, 0, 0, 0, 0, format_fn, context);
3318}
3319 
3320static void _ccv_cnnp_compiled_data_free(const ccv_cnnp_model_t* const model, ccv_cnnp_compiled_data_t* const compiled_data)
3321{
3322	int i;
3323	const int parameter_size = compiled_data->parameters->rnum;
3324	ccv_array_free(compiled_data->parameters);
3325	if (compiled_data->parameter_flags)
3326		ccfreefree(compiled_data->parameter_flags);
3327	const int internal_size = compiled_data->internals->rnum;
3328	ccv_array_free(compiled_data->internals);
3329	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", 3329, __extension__ __PRETTY_FUNCTION__
); }));
3330	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", 3330, __extension__ __PRETTY_FUNCTION__
); }));
3331	for (i = 0; i < parameter_size; i++)
3332		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))));
3333	ccv_array_free(compiled_data->ids.parameters);
3334	for (i = 0; i < internal_size; i++)
3335		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))));
3336	ccv_array_free(compiled_data->ids.internals);
3337	const int parallel_count = compiled_data->parallel_count > 0 ? compiled_data->parallel_count : _ccv_cnnp_model_root_parallel_count(model);
3338	if (compiled_data->tensors.parameters)
3339	{
3340		for (i = 0; i < parameter_size * parallel_count; i++)
3341			// If it is not marked as not belonging, we can free it.
3342			if (!((uintptr_t)compiled_data->tensors.parameters[i] & (uintptr_t)1))
3343				if (compiled_data->tensors.parameters[i])
3344					ccv_nnc_tensor_free(compiled_data->tensors.parameters[i]);
3345		for (i = 0; i < internal_size * parallel_count; i++)
3346			if (compiled_data->tensors.internals[i])
3347				ccv_nnc_tensor_free(compiled_data->tensors.internals[i]);
3348		ccfreefree(compiled_data->tensors.parameters);
3349	}
3350	if (compiled_data->tensors.gradients)
3351	{
3352		for (i = 0; i < parameter_size * parallel_count; i++)
3353		{
3354			if (compiled_data->tensors.gradients[i])
3355				ccv_nnc_tensor_free(compiled_data->tensors.gradients[i]);
3356			if (compiled_data->tensors.accum_gradients[i])
3357				ccv_nnc_tensor_free(compiled_data->tensors.accum_gradients[i]);
3358		}
3359		ccfreefree(compiled_data->tensors.gradients);
3360	}
3361	if (compiled_data->minimize.parameters)
3362	{
3363		for (i = 0; i < compiled_data->minimize.parameters->rnum; i++)
3364			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))));
3365		ccv_array_free(compiled_data->minimize.parameters);
3366	}
3367	if (compiled_data->rewindables)
3368		ccv_array_free(compiled_data->rewindables);
3369	if (compiled_data->tensors_init.v)
3370		ccfreefree(CCV_NNC_INIT_V(compiled_data->tensors_init.v)((uint32_t*)((uintptr_t)(compiled_data->tensors_init.v) &
 ~(uintptr_t)1)));
3371	if (compiled_data->evaluate.tos)
3372		ccfreefree(compiled_data->evaluate.tos);
3373	compiled_data->evaluate.tos = 0;
3374	if (compiled_data->stream_map)
3375	{
3376		khiter_t k;
3377		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)
3378		{
3379			if (!kh_exist(compiled_data->stream_map, k)(!(((compiled_data->stream_map)->flags[(k)>>4]>>
(((k)&0xfU)<<1))&3)))
3380				continue;
3381			ccv_nnc_stream_context_t* const stream = kh_val(compiled_data->stream_map, k)((compiled_data->stream_map)->vals[k]);
3382			ccv_nnc_stream_context_free(stream);
3383		}
3384		kh_destroy(stream_map, compiled_data->stream_map)kh_destroy_stream_map(compiled_data->stream_map);
3385	}
3386	_ccv_cnnp_compiled_data_graph_free(compiled_data);
3387	_ccv_cnnp_compiled_data_gradient_free(compiled_data);
3388	_ccv_cnnp_compiled_data_backward_free(compiled_data);
3389	_ccv_cnnp_compiled_data_apply_gradients_free(compiled_data);
3390	if (compiled_data->gradient_checkpoints)
3391	{
3392		for (i = 0; i < compiled_data->gradient_checkpoints->rnum; i++)
3393		{
3394			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)));
3395			assert(checkpoint->inputs)((void) sizeof ((checkpoint->inputs) ? 1 : 0), __extension__
 ({ if (checkpoint->inputs) ; else __assert_fail ("checkpoint->inputs"
, "ccv_cnnp_model.c", 3395, __extension__ __PRETTY_FUNCTION__
); }));
3396			ccfreefree(checkpoint->inputs);
3397			ccv_array_free(checkpoint->tensor_symbols);
3398		}
3399		ccv_array_free(compiled_data->gradient_checkpoints);
3400	}
3401	ccv_nnc_xpu_alloc_destroy(&compiled_data->xpu_alloc);
3402	ccfreefree(compiled_data);
3403}
3404 
3405void ccv_cnnp_model_free(ccv_cnnp_model_t* const model)
3406{
3407	ccv_cnnp_model_deinit(model);
3408	if (model->isa->dealloc)
3409		model->isa->dealloc(model);
3410	if (model->io)
3411	{
3412		int i;
3413		for (i = 0; i < model->io->rnum; i++)
3414		{
3415			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)));
3416			if (model_io->outgoings)
3417				ccv_array_free(model_io->outgoings);
3418			if (model_io->incomings)
3419				ccv_array_free(model_io->incomings);
3420			if (model_io->dependencies)
3421				ccv_array_free(model_io->dependencies);
3422			ccfreefree(model_io);
3423		}
3424		ccv_array_free(model->io);
3425	}
3426	if (model->parameter_indices)
3427		ccv_array_free(model->parameter_indices);
3428	if (model->inputs)
3429		ccfreefree(model->inputs);
3430	if (model->graph)
3431		ccv_nnc_symbolic_graph_free(model->graph);
3432	if (model->compiled_data)
3433		_ccv_cnnp_compiled_data_free(model, model->compiled_data);
3434	if (model->name)
3435		ccfreefree(model->name);
3436	ccfreefree(model);
3437}
3438 
3439void ccv_cnnp_model_cancel(ccv_cnnp_model_t* const model)
3440{
3441	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
3442	if (!compiled_data)
3443		return;
3444	if (compiled_data->graph)
3445		ccv_nnc_graph_cancel(compiled_data->graph);
3446	if (compiled_data->apply_gradients.graph)
3447		ccv_nnc_graph_cancel(compiled_data->apply_gradients.graph);
3448}
3449 
3450void ccv_cnnp_model_set_flags(ccv_cnnp_model_t* const model, const int flags)
3451{
3452	model->exec_flags = flags;
3453}
3454 
3455int ccv_cnnp_model_flags(ccv_cnnp_model_t* const model)
3456{
3457	return model->exec_flags;
3458}