ccv_cnnp_model.c

Bug Summary

File:	nnc/ccv_cnnp_model.c
Warning:	line 2527, column 8 Array access (via field 'vals') results in a null pointer dereference
Annotated Source Code

Press '?' to see keyboard shortcuts
Show analyzer invocation
clang -cc1 -cc1 -triple x86_64-unknown-linux-gnu -analyze -disable-free -clear-ast-before-backend -disable-llvm-verifier -discard-value-names -main-file-name ccv_cnnp_model.c -analyzer-checker=core -analyzer-checker=apiModeling -analyzer-checker=unix -analyzer-checker=deadcode -analyzer-checker=security.insecureAPI.UncheckedReturn -analyzer-checker=security.insecureAPI.getpw -analyzer-checker=security.insecureAPI.gets -analyzer-checker=security.insecureAPI.mktemp -analyzer-checker=security.insecureAPI.mkstemp -analyzer-checker=security.insecureAPI.vfork -analyzer-checker=nullability.NullPassedToNonnull -analyzer-checker=nullability.NullReturnedFromNonnull -analyzer-output plist -w -setup-static-analyzer -mrelocation-model pic -pic-level 2 -pic-is-pie -mframe-pointer=none -fmath-errno -ffp-contract=on -fno-rounding-math -mconstructor-aliases -funwind-tables=2 -target-cpu x86-64 -target-feature +sse2 -tune-cpu generic -debugger-tuning=gdb -fdebug-compilation-dir=/home/liu/actions-runner/_work/ccv/ccv/lib/nnc -fcoverage-compilation-dir=/home/liu/actions-runner/_work/ccv/ccv/lib/nnc -resource-dir /usr/local/lib/clang/19 -I ../ -I /usr/local/cuda/include -D HAVE_CBLAS -D HAVE_LIBPNG -D HAVE_LIBJPEG -D HAVE_FFTW3 -D HAVE_PTHREAD -D HAVE_LIBLINEAR -D HAVE_TESSERACT -D HAVE_AVCODEC -D HAVE_AVFORMAT -D HAVE_AVUTIL -D HAVE_SWSCALE -D HAVE_SSE2 -D HAVE_GSL -D HAVE_CUDA -D HAVE_CUDNN -D HAVE_NCCL -D USE_SYSTEM_CUB -I /usr/local/include -internal-isystem /usr/local/lib/clang/19/include -internal-isystem /usr/local/include -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/12/../../../../x86_64-linux-gnu/include -internal-externc-isystem /usr/include/x86_64-linux-gnu -internal-externc-isystem /include -internal-externc-isystem /usr/include -O3 -ferror-limit 19 -fgnuc-version=4.2.1 -fskip-odr-check-in-gmf -vectorize-loops -vectorize-slp -analyzer-output=html -faddrsig -D__GCC_HAVE_DWARF2_CFI_ASM=1 -o /home/liu/actions-runner/_work/ccv/ccv/_analyze/2026-05-23-114332-2944924-1 -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	if (loss.cmd != CCV_NNC_NOOP)
407	{
408		ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
409		ccv_nnc_symbolic_graph_simplify(model->graph,
410			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.
411			0, 0, // No need to provide binds at this point.
412			compiled_data->f, model->output_size,
413			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));
414		ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
415	}
416	// If inputs are from GPU, stream type is GPU.
417	compiled_data->parameters = parameters;
418	compiled_data->parameter_flags = parameter_flags;
419	compiled_data->internals = internals;
420	compiled_data->ids.parameters = parameter_ids;
421	compiled_data->ids.internals = internal_ids;
422	ccv_cnnp_model_gradient_checkpoints_cleanup_after_build(compiled_data, model->graph);
423}
424 
425static 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)
426{
427	ccv_array_t* const stack = (ccv_array_t*)context;
428	ccv_array_push(stack, &symbol.d);
429}
430 
431static 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)
432{
433	const ccv_nnc_tensor_symbol_t src_symbol = {
434		.d = src_index,
435		.graph = src_graph
436	};
437	const ccv_nnc_tensor_symbol_t dest_symbol = {
438		.d = dest_index,
439		.graph = dest_graph
440	};
441	const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(src_graph, src_symbol);
442	ccv_nnc_tensor_symbol_set(dest_graph, dest_symbol, params);
443	int ofs[CCV_NNC_MAX_DIM_ALLOC(12)];
444	int inc[CCV_NNC_MAX_DIM_ALLOC(12)];
445	if (0 == ccv_nnc_tensor_symbol_alias_params(src_graph, src_symbol, ofs, inc))
446		ccv_nnc_tensor_symbol_alias_set(dest_graph, dest_symbol, ofs, inc);
447}
448 
449static 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)
450{
451	const ccv_nnc_tensor_symbol_t src_symbol = {
452		.d = src_index,
453		.graph = src_graph
454	};
455	const ccv_nnc_tensor_param_t src_params = ccv_nnc_tensor_symbol_params(src_graph, src_symbol);
456	const ccv_nnc_tensor_symbol_t dest_symbol = {
457		.d = dest_index,
458		.graph = dest_graph
459	};
460	const ccv_nnc_tensor_param_t dest_params = ccv_nnc_tensor_symbol_params(dest_graph, dest_symbol);
461	if (src_params.dim[0] == 0 || dest_params.dim[0] == 0)
462		return 1;
463	return memcmp(src_params.dim, dest_params.dim, sizeof(src_params.dim)) == 0;
464}
465 
466static 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);
467static void _ccv_cnnp_compiled_data_graph_free(ccv_cnnp_compiled_data_t* const compiled_data);
468 
469typedef struct {
470	int parallel_count;
471	ccv_nnc_symbolic_graph_t* graph;
472	ccv_nnc_graph_exec_arena_t* graph_exec_arena;
473} ccv_nnc_graph_exec_update_t;
474 
475static 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)
476{
477	ccv_nnc_graph_exec_update_t* const graph_exec_update = (ccv_nnc_graph_exec_update_t*)context;
478	ccv_nnc_graph_exec_arena_t* const graph_exec_arena = graph_exec_update->graph_exec_arena;
479	ccv_nnc_graph_exec_t graph_exec = ccv_nnc_graph_exec_from_symbol(graph_exec_arena, symbol);
480	ccv_nnc_graph_exec_set(graph_exec.graph, graph_exec, cmd);
481	ccv_nnc_graph_exec_set_hint(graph_exec.graph, graph_exec, hint);
482	const ccv_nnc_symbolic_graph_t* const graph = graph_exec_update->graph;
483	const int parallel_count = graph_exec_update->parallel_count;
484	int i;
485	for (i = 1; i < parallel_count; i++)
486	{
487		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));
488		if (!CCV_NO_GRAPH_EXEC(copy)((copy).graph == 0))
489		{
490			ccv_nnc_graph_exec_set(copy.graph, copy, cmd);
491			ccv_nnc_graph_exec_set_hint(copy.graph, copy, hint);
492		}
493	}
494}
495 
496void 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)
497{
498	assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if
 (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c"
, 498, __extension__ __PRETTY_FUNCTION__); }));
499	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", 499, __extension__ __PRETTY_FUNCTION__)
; }));
500	assert(!init->graph)((void) sizeof ((!init->graph) ? 1 : 0), __extension__ ({ if
 (!init->graph) ; else __assert_fail ("!init->graph", "ccv_cnnp_model.c"
, 500, __extension__ __PRETTY_FUNCTION__); }));
501	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
502	init->graph = ccv_nnc_symbolic_graph_new();
503	ccv_array_t* const stack = ccv_array_new(sizeof(int), 0, 0);
504	ccv_nnc_graph_exec_symbol_new_hook(init->graph, _ccv_cnnp_graph_push_graph_exec_symbol, stack, 0);
505	_ccv_cnnp_model_compile(init, inputs, input_size, compiled_data->loss);
506	init->parallel_count = model->parallel_count;
507	init->memory_compression = model->memory_compression;
508	init->memory_reduction = model->memory_reduction;
509	init->gradient_checkpointing = model->gradient_checkpointing;
510	init->compiled_data->stream_type = model->compiled_data->stream_type;
511	init->compiled_data->minimize.minimizer = model->compiled_data->minimize.minimizer;
512	init->compiled_data->minimize.max_saved_aux_size = model->compiled_data->minimize.max_saved_aux_size;
513	if (model->compiled_data->gradient_mode != CCV_CNNP_COMPILED_DATA_GRADIENT_NONE)
514		_ccv_cnnp_model_gradient_init(init, model->compiled_data->gradient_mode, model->compiled_data->disable_outgrad, 0, 0);
515	ccv_nnc_graph_exec_symbol_new_hook(init->graph, 0, 0, 0);
516	ccv_nnc_symbolic_graph_tensor_auto(init->graph, TRAVERSE_FULL0,0,0,0);
517	int i, j;
518	// Verify parameters, internals and saved_aux in both graph has the same dimensionality.
519	for (i = 0; i < compiled_data->parameters->rnum; i++)
520	{
521		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;
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	for (i = 0; i < compiled_data->internals->rnum; i++)
525	{
526		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;
527		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", 527, __extension__ __PRETTY_FUNCTION__)
; }));
528	}
529	// Update inputs.
530	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", 530, __extension__ __PRETTY_FUNCTION__)
; }));
531	for (i = 0; i < model->input_size; i++)
532		if (model->inputs[i].d >= 0)
533		{
534			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", 534, __extension__ __PRETTY_FUNCTION__)
; }));
535			_ccv_nnc_tensor_symbol_reinit(init->graph, model->graph, init->inputs[i].d, model->inputs[i].d);
536		}
537	// Update outputs.
538	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", 538, __extension__ __PRETTY_FUNCTION__)
; }));
539	for (i = 0; i < model->output_size; i++)
540	{
541		if (model->outputs[i].d >= 0)
542		{
543			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", 543, __extension__
 __PRETTY_FUNCTION__); }));
544			_ccv_nnc_tensor_symbol_reinit(init->graph, model->graph, init->outputs[i].d, model->outputs[i].d);
545		}
546		if (model->outputs[i].d != model->compiled_data->f[i].d)
547		{
548			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", 548, __extension__ __PRETTY_FUNCTION__)
; }));
549			if (model->compiled_data->f[i].d >= 0)
550			{
551				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", 551, __extension__ __PRETTY_FUNCTION__)
; }));
552				_ccv_nnc_tensor_symbol_reinit(init->graph, model->graph, init->compiled_data->f[i].d, model->compiled_data->f[i].d);
553			}
554		}
555	}
556	// Go through the graph to set tensor on matching symbols
557	for (i = 0; i < stack->rnum; i++)
558	{
559		const int d = *(int*)ccv_array_get(stack, i)((void*)(((char*)((stack)->data)) + (size_t)(stack)->rsize
 * (size_t)(i)));
560		// If exceed range, skip.
561		if (d >= ccv_nnc_graph_exec_symbol_count(init->graph) ||
562			d >= ccv_nnc_graph_exec_symbol_count(model->graph))
563			continue;
564		const ccv_nnc_graph_exec_symbol_t src_symbol = {
565			.d = d,
566			.graph = init->graph
567		};
568		const ccv_nnc_graph_exec_symbol_t dest_symbol = {
569			.d = d,
570			.graph = model->graph
571		};
572		const ccv_nnc_cmd_t src_cmd = ccv_nnc_graph_exec_symbol_cmd(init->graph, src_symbol);
573		const ccv_nnc_cmd_t dest_cmd = ccv_nnc_graph_exec_symbol_cmd(model->graph, dest_symbol);
574		// If the name doesn't match, skip.
575		if (dest_cmd.cmd != src_cmd.cmd && src_cmd.cmd != CCV_NNC_NOOP)
576			continue;
577		// Now get all the inputs and outputs, if matches, set them.
578		const int* src_inputs;
579		int src_input_size;
580		const int* src_outputs;
581		int src_output_size;
582		ccv_nnc_graph_exec_symbol_io(init->graph, src_symbol, &src_inputs, &src_input_size, &src_outputs, &src_output_size);
583		const int* dest_inputs;
584		int dest_input_size;
585		const int* dest_outputs;
586		int dest_output_size;
587		ccv_nnc_graph_exec_symbol_io(model->graph, dest_symbol, &dest_inputs, &dest_input_size, &dest_outputs, &dest_output_size);
588		// We may have unmatched input / output size because this is the minimizer and it has
589		// different saved_aux (for example, when we shrunk with CMD_NOOP).
590		if (src_input_size != dest_input_size)
591			continue;
592		if (src_output_size != dest_output_size)
593			continue;
594		ccv_nnc_graph_exec_symbol_set(model->graph, dest_symbol, src_cmd);
595		// There may be mismatches of the source tensor symbols and destination tensor symbols. The reason is because
596		// we may later passed-in the minimizer, therefore, we may allocate tensors for minimizer later in the original
597		// graph whereas in the newly created graph, it is streamlined (the minimizer exists from the beginning). That
598		// will make the order of tensor symbols creation different, therefore, exact which tensor is which wrong as
599		// well. However, set a new minimizer won't change the exec symbol ordering, because we never create new exec
600		// symbols after gradient init step. Changing a new minimizer just updated that exec symbols setting, it is not
601		// a new exec symbol.
602		for (j = 0; j < src_input_size; j++)
603			if (src_inputs[j] >= 0)
604				_ccv_nnc_tensor_symbol_reinit(init->graph, model->graph, src_inputs[j], dest_inputs[j]);
605		for (j = 0; j < src_output_size; j++)
606			if (src_outputs[j] >= 0)
607				_ccv_nnc_tensor_symbol_reinit(init->graph, model->graph, src_outputs[j], dest_outputs[j]);
608	}
609	ccv_array_free(stack);
610	// After this, we get all tensors in the model graph resolved through tensor_auto.
611	ccv_nnc_symbolic_graph_tensor_auto(model->graph, TRAVERSE_FULL0,0,0,0);
612	// Verify symbols we get matches.
613	const int parameter_size = compiled_data->parameters->rnum;
614	for (i = 0; i < parameter_size; i++)
615		{ 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", 615, __extension__ __PRETTY_FUNCTION__)
; })); }
616	const int internal_size = compiled_data->internals->rnum;
617	for (i = 0; i < internal_size; i++)
618		{ 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", 618, __extension__ __PRETTY_FUNCTION__)
; })); }
619	// Go through compiled data.
620	if (compiled_data->tensor_arena)
621	{
622		const int flag = ccv_nnc_tensor_arena_reinit(compiled_data->tensor_arena, model->graph);
623		if (flag == 0 && compiled_data->graph_exec_arena)
624		{
625			ccv_nnc_graph_exec_reinit(compiled_data->graph_exec_arena, compiled_data->graph, model->graph);
626			// Since we will reinit, if we previously set is_test, we need to set it again.
627			if (compiled_data->is_test)
628			{
629				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; });
630				ccv_nnc_graph_exec_update_t update = {
631					.parallel_count = parallel_count,
632					.graph = model->graph,
633					.graph_exec_arena = compiled_data->graph_exec_arena,
634				};
635				ccv_cnnp_model_set_is_test(model, 1, _ccv_cnnp_cmd_update_for_execs, &update);
636			}
637		} else
638			// Free-up tensor arena & graph exec arena.
639			_ccv_cnnp_compiled_data_graph_free(compiled_data);
640	}
641	// There are other compiled graphs, for accum and apply gradients.
642	// However, the main conclusion is, these absorb operations shouldn't impact parameters.
643	// Thus, it won't impact the shape of gradients (only outgrad). Since for outgrad, we
644	// don't allocate ourselves, it is not a concern. For normal gradients, the shape cannot
645	// be changed otherwise parameters' shape will be meaningless. The same goes to internals.
646	// That is why we don't update these compiled graphs at all this point.
647	// Free the model, we've already "absorbed" it.
648	ccv_cnnp_model_free(init);
649}
650 
651void 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)
652{
653	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", 653, __extension__ __PRETTY_FUNCTION__)
; }));
654	if (model->input_size == 0)
655		model->input_size = input_size;
656	if (!model->graph) // The graph is not compiled yet.
657	{
658		model->graph = ccv_nnc_symbolic_graph_new();
659		_ccv_cnnp_model_compile(model, inputs, input_size, loss);
660		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", 660, __extension__ __PRETTY_FUNCTION__)
; }));
661		int i, flag = 0;
662		for (i = 0; !flag && i < input_size; i++)
663			flag = (CCV_TENSOR_GET_MEMORY(inputs[i].type)((inputs[i].type) & 0x3) == CCV_TENSOR_GPU_MEMORY);
664		// If inputs are from GPU, stream type is GPU.
665		model->compiled_data->stream_type = flag ? CCV_STREAM_CONTEXT_GPU : CCV_STREAM_CONTEXT_CPU;
666		model->compiled_data->minimize.minimizer = minimizer;
667		model->compiled_data->minimize.max_saved_aux_size = ccv_nnc_minimizer_saved_aux_size(minimizer);
668	} else {
669		// Now, finally fill in this part. If the graph is already compiled, we make a copy of the model.
670		// And then absorb the "new model" to the old one.
671		ccv_cnnp_model_t* const init = ccv_cnnp_model_copy(model, model->is_trainable);
672		ccv_cnnp_model_absorb(model, init, inputs, input_size);
673		// Reset minimizer.
674		ccv_cnnp_model_set_minimizer(model, minimizer, 1, 0, 0);
675	}
676}
677 
678ccv_cnnp_model_t* ccv_cnnp_model_copy(const ccv_cnnp_model_t* const model, const int is_trainable)
679{
680	ccv_cnnp_model_t* const new_model = _ccv_cnnp_model_copy(model, 0);
681	new_model->is_trainable = is_trainable;
682	return new_model;
683}
684 
685void ccv_cnnp_model_tensor_auto(ccv_cnnp_model_t* const model, ccv_nnc_tensor_param_t* const outputs, const int output_size)
686{
687	assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if
 (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c"
, 687, __extension__ __PRETTY_FUNCTION__); }));
688	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", 688, __extension__ __PRETTY_FUNCTION__)
; }));
689	ccv_nnc_symbolic_graph_t* const graph = model->graph;
690	ccv_nnc_symbolic_graph_tensor_auto(graph, TRAVERSE_FULL0,0,0,0);
691	int i;
692	for (i = 0; i < output_size; i++)
693	{
694		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", 694, __extension__ __PRETTY_FUNCTION__)
; }));
695		outputs[i] = ccv_nnc_tensor_symbol_params(graph, model->outputs[i]);
696	}
697}
698 
699void ccv_cnnp_model_set_workspace_size(ccv_cnnp_model_t* const model, size_t workspace_size)
700{
701	if (workspace_size == model->workspace_size)
702		return;
703	model->workspace_size = workspace_size;
704	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
705	if (compiled_data && compiled_data->graph)
706		ccv_nnc_graph_autotune(compiled_data->graph, workspace_size, 0, TRAVERSE_FULL0,0,0,0);
707}
708 
709size_t ccv_cnnp_model_workspace_size(ccv_cnnp_model_t* const model)
710{
711	return model->workspace_size;
712}
713 
714void ccv_cnnp_model_set_data_parallel(ccv_cnnp_model_t* const model, const int parallel)
715{
716	if (parallel == 0)
717		model->parallel_count = ccv_nnc_device_count(CCV_STREAM_CONTEXT_GPU);
718	else
719		model->parallel_count = parallel;
720	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
721	if (compiled_data)
722		{ 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", 722, __extension__ __PRETTY_FUNCTION__)
; })); }
723}
724 
725void ccv_cnnp_model_set_max_concurrency(ccv_cnnp_model_t* const model, const int max_stream_count)
726{
727	model->max_stream_count = max_stream_count;
728	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
729	if (compiled_data)
730		{ 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", 730, __extension__ __PRETTY_FUNCTION__)
; })); }
731}
732 
733void ccv_cnnp_model_set_memory_compression(ccv_cnnp_model_t* const model, const int memory_compression)
734{
735	model->memory_compression = memory_compression;
736	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
737	if (compiled_data)
738		{ 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", 738, __extension__ __PRETTY_FUNCTION__)
; })); }
739}
740 
741void ccv_cnnp_model_set_memory_reduction(ccv_cnnp_model_t* const model, const int memory_reduction)
742{
743	model->memory_reduction = memory_reduction;
744	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
745	if (compiled_data)
746		{ 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", 746, __extension__ __PRETTY_FUNCTION__)
; })); }
747}
748 
749void ccv_cnnp_model_set_gradient_checkpointing(ccv_cnnp_model_t* const model, const int gradient_checkpointing)
750{
751	model->gradient_checkpointing = gradient_checkpointing;
752}
753 
754int ccv_cnnp_model_gradient_checkpointing(ccv_cnnp_model_t* const model)
755{
756	return model->gradient_checkpointing;
757}
758 
759typedef struct {
760	int parallel_count;
761	ccv_nnc_symbolic_graph_t* graph;
762	ccv_cnnp_compiled_data_t* compiled_data;
763	ccv_nnc_tensor_arena_t* tensor_arena;
764} ccv_nnc_tensor_init_states_t;
765 
766static int _ccv_cnnp_any_to_init(const ccv_cnnp_compiled_data_t* const compiled_data)
767{
768	int i;
769	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));
770	for (i = 0; i < compiled_data->parameters->rnum; i++)
771	{
772		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;
773		if (!(init_v[d >> 5] & (1u << (d & 0x1f))))
774			return 1;
775	}
776	for (i = 0; i < compiled_data->internals->rnum; i++)
777	{
778		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;
779		if (!(init_v[d >> 5] & (1u << (d & 0x1f))))
780			return 1;
781	}
782	return 0;
783}
784 
785static 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)
786{
787	ccv_nnc_tensor_init_states_t* const tensor_init_states = (ccv_nnc_tensor_init_states_t*)context;
788	ccv_nnc_tensor_arena_t* const tensor_arena = tensor_init_states->tensor_arena;
789	ccv_nnc_tensor_t* const output_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, output_symbol);
790	if (!output_tensor)
791		return;
792	const int d = output_symbol.d;
793	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", 793, __extension__ __PRETTY_FUNCTION__)
; }));
794  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));
795	if (init_v[d >> 5] & (1u << (d & 0x1f)))
796		return;
797	init_v[d >> 5] |= (1u << (d & 0x1f));
798	ccv_nnc_cmd_exec(cmd, hint, flags, &input, input ? 1 : 0, &output_tensor, 1, 0);
799	const ccv_nnc_symbolic_graph_t* const graph = tensor_init_states->graph;
800	const int parallel_count = tensor_init_states->parallel_count;
801	int i;
802	for (i = 1; i < parallel_count; i++)
803	{
804		ccv_nnc_tensor_t* const copy = ccv_nnc_tensor_from_symbol(tensor_arena, ccv_nnc_tensor_symbol_copy(graph, output_symbol, i));
805		if (copy)
806			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);
807	}
808}
809 
810// This method can only handle cases we added new tensors and exec, never delete. This invariant is true because
811// we setup everything (including calling simplify method) in ccv_cnnp_model_compile method, before this rewind setup.
812static void _ccv_cnnp_model_rewind_graph(ccv_cnnp_model_t* const model)
813{
814	assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if
 (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c"
, 814, __extension__ __PRETTY_FUNCTION__); }));
815	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", 815, __extension__ __PRETTY_FUNCTION__)
; }));
816	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
817	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", 817, __extension__
 __PRETTY_FUNCTION__); }));
818	int i;
819	for (i = 0; i < compiled_data->rewindables->rnum; i++)
820	{
821		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)));
822		if (rewind_symbol->type == CCV_CNNP_REWIND_GRAPH_EXEC)
823			ccv_nnc_graph_exec_symbol_free(model->graph, rewind_symbol->graph_exec);
824		else if (rewind_symbol->type == CCV_CNNP_REWIND_TENSOR)
825			ccv_nnc_tensor_symbol_free(model->graph, rewind_symbol->tensor);
826	}
827	ccv_array_clear(compiled_data->rewindables);
828	ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
829}
830 
831static 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)
832{
833	const ccv_cnnp_rewind_symbol_t rewind_symbol = {
834		.type = CCV_CNNP_REWIND_TENSOR,
835		.tensor = symbol
836	};
837	ccv_array_t* const rewind_symbols = (ccv_array_t*)context;
838	ccv_array_push(rewind_symbols, &rewind_symbol);
839}
840 
841static 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)
842{
843	const ccv_cnnp_rewind_symbol_t rewind_symbol = {
844		.type = CCV_CNNP_REWIND_TENSOR,
845		.tensor = symbol
846	};
847	ccv_array_t* const rewind_symbols = (ccv_array_t*)context;
848	ccv_array_push(rewind_symbols, &rewind_symbol);
849}
850 
851static 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)
852{
853	const ccv_cnnp_rewind_symbol_t rewind_symbol = {
854		.type = CCV_CNNP_REWIND_GRAPH_EXEC,
855		.graph_exec = symbol
856	};
857	ccv_array_t* const rewind_symbols = (ccv_array_t*)context;
858	ccv_array_push(rewind_symbols, &rewind_symbol);
859}
860 
861static 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)
862{
863	ccv_nnc_graph_exec_t const update_exec = ccv_nnc_graph_exec_from_symbol(graph_exec_arena, exec_symbol);
864	if (!CCV_NO_GRAPH_EXEC(update_exec)((update_exec).graph == 0))
865		ccv_nnc_graph_exec_set(update_exec.graph, update_exec, cmd);
866	int i;
867	for (i = 1; i < parallel_count; i++)
868	{
869		ccv_nnc_graph_exec_symbol_t copy_symbol = ccv_nnc_graph_exec_symbol_copy(symbolic_graph, exec_symbol, i);
870		const ccv_nnc_graph_exec_t copy = ccv_nnc_graph_exec_from_symbol(graph_exec_arena, copy_symbol);
871		if (!CCV_NO_GRAPH_EXEC(copy)((copy).graph == 0))
872			ccv_nnc_graph_exec_set(copy.graph, copy, cmd);
873	}
874}
875 
876static 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)
877{
878	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 878, __extension__ __PRETTY_FUNCTION__); }));
879	assert(symbolic_graph)((void) sizeof ((symbolic_graph) ? 1 : 0), __extension__ ({ if
 (symbolic_graph) ; else __assert_fail ("symbolic_graph", "ccv_cnnp_model.c"
, 879, __extension__ __PRETTY_FUNCTION__); }));
880	ccv_nnc_graph_exec_symbol_set(symbolic_graph, exec_symbol, cmd);
881	int i;
882	for (i = 1; i < parallel_count; i++)
883	{
884		ccv_nnc_graph_exec_symbol_t copy_symbol = ccv_nnc_graph_exec_symbol_copy(symbolic_graph, exec_symbol, i);
885		if (copy_symbol.graph)
886			ccv_nnc_graph_exec_symbol_set(symbolic_graph, copy_symbol, cmd);
887	}
888	ccv_nnc_graph_exec_arena_t* const graph_exec_arena = compiled_data->graph_exec_arena;
889	if (graph_exec_arena)
890		_ccv_cnnp_model_graph_symbol_exec_set_for_graph_exec_arena(graph_exec_arena, parallel_count, exec_symbol, cmd, symbolic_graph);
891	// Skip backward graph exec arena because it is for a specific accum symbolic graph, not the main graph (model->graph)
892	ccv_nnc_graph_exec_arena_t* const gradient_graph_exec_arena = compiled_data->apply_gradients.graph_exec_arena;
893	if (gradient_graph_exec_arena)
894		_ccv_cnnp_model_graph_symbol_exec_set_for_graph_exec_arena(gradient_graph_exec_arena, parallel_count, exec_symbol, cmd, symbolic_graph);
895}
896 
897static 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)
898{
899	int this_parameter_flag = 0;
900	if (update_nodes[parameter_indice].d == CCV_NNC_NO_TENSOR_SYMBOL)
901		return this_parameter_flag;
902	const ccv_nnc_cmd_t old_minimizer = ccv_nnc_graph_exec_symbol_cmd(graph, update_nodes[parameter_indice]);
903	int j, k;
904	// For no-op, we can preserve previous saved_aux_size.
905	if (old_minimizer.cmd != minimizer.cmd && minimizer.cmd != CCV_NNC_NOOP)
906	{
907		// If the old minimizer is a noop, then the old_saved_aux_size should be whatever its previous
908		// saved_aux_size is, otherwise we will reinit the saved_aux repeatedly if you switch between
909		// noop and a minimizer. We don't want that because we do that in high-level frameworks to
910		// make sure some model parameters don't update if we don't want them to.
911		int old_saved_aux_size;
912		if (old_minimizer.cmd == CCV_NNC_NOOP)
913		{
914			int input_size;
915			ccv_nnc_graph_exec_symbol_io(graph, update_nodes[parameter_indice], 0, &input_size, 0, 0);
916			if (input_size < 2) // This is not legit.
917				old_saved_aux_size = ccv_nnc_minimizer_saved_aux_size(old_minimizer);
918			else // See ccv_nnc_minimizer_saved_aux_size, the saved_aux is inputs excluding gradients and parameters.
919				old_saved_aux_size = input_size - 2;
920		} else
921			old_saved_aux_size = ccv_nnc_minimizer_saved_aux_size(old_minimizer);
922		if (old_saved_aux_size != saved_aux_size)
923		{
924			this_parameter_flag = 1;
925			if (saved_aux_size > old_saved_aux_size)
926			{
927				// Allocate new tensor symbols.
928				const ccv_nnc_tensor_param_t info = ccv_nnc_tensor_symbol_params(graph, updated_parameters[parameter_indice]);
929				for (j = old_saved_aux_size; j < saved_aux_size; j++)
930				{
931					saved_aux[parameter_indice * max_saved_aux_size + j].source = ccv_nnc_tensor_symbol_new(graph, info, 0);
932					saved_aux[parameter_indice * max_saved_aux_size + j].destination = ccv_nnc_tensor_symbol_new(graph, info, 0);
933					const int device_id = CCV_TENSOR_GET_DEVICE_ID(info.type)(((info.type) & 0xfff00) >> 8);
934					for (k = 1; k < parallel_count; k++)
935					{
936						ccv_nnc_tensor_param_t dev_info = info;
937						if (k != device_id)
938							CCV_TENSOR_SET_DEVICE_ID(dev_info.type, k)(dev_info.type) = (((dev_info.type) & ~0xfff00) | (((k) &
 0xfff) << 8));
939						else
940							CCV_TENSOR_SET_DEVICE_ID(dev_info.type, 0)(dev_info.type) = (((dev_info.type) & ~0xfff00) | (((0) &
 0xfff) << 8));
941						const ccv_nnc_tensor_symbol_t src_copy = ccv_nnc_tensor_symbol_new(graph, dev_info, 0);
942						const ccv_nnc_tensor_symbol_t dest_copy = ccv_nnc_tensor_symbol_new(graph, dev_info, 0);
943						ccv_nnc_tensor_symbol_set_copy(graph, saved_aux[parameter_indice * max_saved_aux_size + j].source, k, src_copy);
944						ccv_nnc_tensor_symbol_set_copy(graph, saved_aux[parameter_indice * max_saved_aux_size + j].destination, k, dest_copy);
945					}
946				}
947			} else {
948				for (j = saved_aux_size; j < old_saved_aux_size; j++)
949				{
950					for (k = 1; k < parallel_count; k++)
951					{
952						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);
953						if (src_copy.d >= 0)
954						{
955							ccv_nnc_tensor_symbol_free(graph, src_copy);
956							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
});
957						}
958						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);
959						if (dest_copy.d >= 0)
960						{
961							ccv_nnc_tensor_symbol_free(graph, dest_copy);
962							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
});
963						}
964					}
965					ccv_nnc_tensor_symbol_free(graph, saved_aux[parameter_indice * max_saved_aux_size + j].source);
966					ccv_nnc_tensor_symbol_free(graph, saved_aux[parameter_indice * max_saved_aux_size + j].destination);
967					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
};
968				}
969			}
970		}
971	}
972	_ccv_cnnp_model_graph_exec_symbol_set(graph, compiled_data, parallel_count, update_nodes[parameter_indice], minimizer);
973	if (this_parameter_flag)
974	{
975		ccv_nnc_tensor_symbol_t update_inputs[saved_aux_size + 2];
976		ccv_nnc_tensor_symbol_t update_outputs[saved_aux_size + 1];
977		const int* inputs = 0;
978		int input_size = 0;
979		ccv_nnc_graph_exec_symbol_io(graph, update_nodes[parameter_indice], &inputs, &input_size, 0, 0);
980		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", 980, __extension__ __PRETTY_FUNCTION__)
; }));
981		update_inputs[0].d = inputs[0];
982		update_inputs[0].graph = graph;
983		update_inputs[1].d = inputs[1];
984		update_inputs[1].graph = graph;
985		update_outputs[0] = updated_parameters[parameter_indice];
986		for (j = 0; j < saved_aux_size; j++)
987		{
988			update_inputs[j + 2] = saved_aux[parameter_indice * max_saved_aux_size + j].source;
989			update_outputs[j + 1] = saved_aux[parameter_indice * max_saved_aux_size + j].destination;
990		}
991		ccv_nnc_graph_exec_symbol_set_io(graph, update_nodes[parameter_indice], update_inputs, saved_aux_size + 2, update_outputs, saved_aux_size + 1);
992		for (k = 1; k < parallel_count; k++)
993		{
994			const ccv_nnc_graph_exec_symbol_t copy = ccv_nnc_graph_exec_symbol_copy(graph, update_nodes[parameter_indice], k);
995			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"
, 995, __extension__ __PRETTY_FUNCTION__); }));
996			ccv_nnc_graph_exec_symbol_io(graph, copy, &inputs, &input_size, 0, 0);
997			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", 997, __extension__ __PRETTY_FUNCTION__)
; }));
998			update_inputs[0].d = inputs[0];
999			update_inputs[0].graph = graph;
1000			update_inputs[1].d = inputs[1];
1001			update_inputs[1].graph = graph;
1002			update_outputs[0] = ccv_nnc_tensor_symbol_copy(graph, updated_parameters[parameter_indice], k);
1003			for (j = 0; j < saved_aux_size; j++)
1004			{
1005				update_inputs[j + 2] = ccv_nnc_tensor_symbol_copy(graph, saved_aux[parameter_indice * max_saved_aux_size + j].source, k);
1006				update_outputs[j + 1] = ccv_nnc_tensor_symbol_copy(graph, saved_aux[parameter_indice * max_saved_aux_size + j].destination, k);
1007			}
1008			ccv_nnc_graph_exec_symbol_set_io(graph, copy, update_inputs, saved_aux_size + 2, update_outputs, saved_aux_size + 1);
1009		}
1010	}
1011	return this_parameter_flag;
1012}
1013 
1014typedef struct {
1015	int parameter_size;
1016	ccv_nnc_cmd_t minimizer;
1017	ccv_cnnp_model_io_t parameters[1];
1018} ccv_cnnp_set_minimizer_for_parameter_t;
1019 
1020static int _ccv_cnnp_apply_parameters_with_minimizer(ccv_cnnp_model_t* const model)
1021{
1022	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1023	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 1023, __extension__ __PRETTY_FUNCTION__); }));
1024	const int max_saved_aux_size = compiled_data->minimize.max_saved_aux_size;
1025	// We update all parameters, at this point, we have one minimizer.
1026	const int parameter_size = compiled_data->parameters->rnum;
1027	ccv_nnc_graph_exec_symbol_t* const update_nodes = compiled_data->update_nodes;
1028	ccv_nnc_symbolic_graph_t* const symbolic_graph = model->graph;
1029	assert(symbolic_graph)((void) sizeof ((symbolic_graph) ? 1 : 0), __extension__ ({ if
 (symbolic_graph) ; else __assert_fail ("symbolic_graph", "ccv_cnnp_model.c"
, 1029, __extension__ __PRETTY_FUNCTION__); }));
1030	const int parallel_count = _ccv_cnnp_model_root_parallel_count(model);
1031	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", 1031, __extension__ __PRETTY_FUNCTION__
); }));
1032	ccv_array_t* const parameters = compiled_data->minimize.parameters;
1033	ccv_array_t* const parameter_indices = ccv_array_new(sizeof(int), 0, 0);
1034	int i, j, flag = 0;
1035	for (i = 0; i < parameters->rnum; i++)
1036	{
1037		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)));
1038		for (j = 0; j < set_minimizer_for_parameter->parameter_size; j++)
1039		{
1040			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;
1041			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", 1041, __extension__ __PRETTY_FUNCTION__
); }));
1042			const int old_rnum = parameter_indices->rnum;
1043			ccv_cnnp_model_add_to_parameter_indices(set_minimizer_for_parameter->parameters[j]->model, param_sel, parameter_indices);
1044			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;
1045			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", 1045, __extension__ __PRETTY_FUNCTION__
); }));
1046			if (param_ref >= 0)
1047			{
1048				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", 1048, __extension__ __PRETTY_FUNCTION__
); }));
1049				*(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)));
1050				parameter_indices->rnum = old_rnum + 1;
1051			}
1052		}
1053		const int saved_aux_size = ccv_nnc_minimizer_saved_aux_size(set_minimizer_for_parameter->minimizer);
1054		// We may have duplicated indices, but that is OK, we will set it twice.
1055		for (j = 0; j < parameter_indices->rnum; j++)
1056		{
1057			const int d = *(int*)ccv_array_get(parameter_indices, j)((void*)(((char*)((parameter_indices)->data)) + (size_t)(parameter_indices
)->rsize * (size_t)(j)));
1058			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", 1058, __extension__ __PRETTY_FUNCTION__
); }));
1059			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))
1060				flag = 1;
1061		}
1062		ccv_array_clear(parameter_indices);
1063	}
1064	ccv_array_free(parameter_indices);
1065	return flag;
1066}
1067 
1068static 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)
1069{
1070	if (new_saved_aux_size == old_saved_aux_size)
1071		return;
1072	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", 1072, __extension__ __PRETTY_FUNCTION__
); }));
1073	int i, j;
1074	for (i = parameter_size - 1; i >= 0; i--)
1075	{
1076		for (j = new_saved_aux_size - 1; j >= old_saved_aux_size; j--)
1077			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
};
1078		for (j = old_saved_aux_size - 1; j >= 0; j--)
1079			saved_aux[i * new_saved_aux_size + j] = saved_aux[i * old_saved_aux_size + j];
1080	}
1081}
1082 
1083static void _ccv_cnnp_model_set_rewindables(ccv_cnnp_model_t* const model)
1084{
1085	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1086	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 1086, __extension__ __PRETTY_FUNCTION__); }));
1087	if (!compiled_data->rewindables)
1088		compiled_data->rewindables = ccv_array_new(sizeof(ccv_cnnp_rewind_symbol_t), 0, 0);
1089	ccv_nnc_tensor_symbol_new_hook(model->graph, _ccv_cnnp_model_tensor_symbol_new_hook, compiled_data->rewindables, 0);
1090	ccv_nnc_tensor_symbol_alias_new_hook(model->graph, _ccv_cnnp_model_tensor_symbol_alias_new_hook, compiled_data->rewindables, 0);
1091	ccv_nnc_graph_exec_symbol_new_hook(model->graph, _ccv_cnnp_model_graph_exec_symbol_new_hook, compiled_data->rewindables, 0);
1092}
1093 
1094static 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)
1095{
1096	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1097	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", 1097, __extension__ __PRETTY_FUNCTION__
); }));
1098	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", 1098, __extension__ __PRETTY_FUNCTION__
); }));
1099	const int evaluate_to_size = compiled_data->evaluate.to_size;
1100	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", 1100, __extension__ __PRETTY_FUNCTION__
); }));
1101	const int parallel_count = _ccv_cnnp_model_root_parallel_count(model);
1102	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", 1102, __extension__ __PRETTY_FUNCTION__
); }));
1103	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);
1104	compiled_data->evaluate.to_ops = (ccv_nnc_graph_exec_t*)(compiled_data->evaluate.tos + evaluate_to_size * parallel_count);
1105	int i, j;
1106	const int output_size = model->output_size;
1107	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", 1107, __extension__ __PRETTY_FUNCTION__
); }));
1108	if (fits)
1109		for (i = 0; i < output_size; i++)
1110			ccv_nnc_tensor_symbol_set(model->graph, compiled_data->fits[i], fits[i]->info);
1111	const int max_saved_aux_size = compiled_data->minimize.max_saved_aux_size;
1112	const int parameter_size = compiled_data->parameters->rnum;
1113	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);
1114	compiled_data->update_nodes = (ccv_nnc_graph_exec_symbol_t*)(compiled_data->updated_parameters + parameter_size);
1115	compiled_data->saved_aux = (ccv_nnc_tensor_symbol_map_t*)(compiled_data->update_nodes + parameter_size);
1116	int parameter_size_maybe_more = parameter_size;
1117	compiled_data->disable_outgrad = disable_outgrad;
1118	int outgrad_size;
1119	if (gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES || model->input_size == 0)
1120		outgrad_size = 0;
1121	else if (disable_outgrad == CCV_CNNP_DISABLE_OUTGRAD_NONE) // Compute minimize with gradients including inputs.
1122		outgrad_size = model->input_size;
1123	else {
1124		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", 1124, __extension__ __PRETTY_FUNCTION__
); })); // If it is disable all, gradient mode won't be this.
1125		outgrad_size = 0;
1126		for (i = 0; i < model->input_size; i++)
1127			if (!(disable_outgrad & ((uint64_t)1 << i)))
1128				++outgrad_size;
1129	}
1130	compiled_data->outgrad_size = outgrad_size;
1131	parameter_size_maybe_more += outgrad_size;
1132	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);
1133	compiled_data->outgrads = parameter_size_maybe_more > parameter_size ? compiled_data->gradients + parameter_size : 0;
1134	compiled_data->backward.tos = (ccv_nnc_graph_exec_symbol_t*)(compiled_data->gradients + parameter_size_maybe_more);
1135	compiled_data->backward.to_size = parameter_size_maybe_more;
1136	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)));
1137	if (compiled_data->parameter_flags)
1138	{
1139		parameters = (ccv_nnc_tensor_symbol_t*)ccmallocmalloc(sizeof(ccv_nnc_tensor_symbol_t) * parameter_size);
1140		for (i = 0; i < parameter_size; i++)
1141			if (compiled_data->parameter_flags[i >> 6] & ((uint64_t)1 << (i & 63)))
1142				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)));
1143			else
1144				parameters[i] = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
1145	}
1146	if (gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES || model->input_size == 0)
1147		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);
1148	else if (disable_outgrad == CCV_CNNP_DISABLE_OUTGRAD_NONE) // Compute minimize with gradients including inputs.
1149		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);
1150	else { // Compute minimize with gradients including selected inputs.
1151		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", 1151, __extension__ __PRETTY_FUNCTION__
); }));
1152		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", 1152, __extension__ __PRETTY_FUNCTION__
); })); // If it is disable all, gradient mode won't be this.
1153		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", 1153, __extension__ __PRETTY_FUNCTION__
); }));
1154		ccv_nnc_tensor_symbol_t outgrads[outgrad_size];
1155		j = 0;
1156		for (i = 0; i < model->input_size; i++)
1157			if (!(disable_outgrad & ((uint64_t)1 << i)))
1158				outgrads[j++] = model->inputs[i];
1159		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);
1160	}
1161	if (compiled_data->parameter_flags)
1162		ccfreefree(parameters);
1163	_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);
1164	if (compiled_data->minimize.parameters)
1165		_ccv_cnnp_apply_parameters_with_minimizer(model);
1166	// Go through gradient checkpoints to generate tensor inputs for backward pass just before executing the backward pass.
1167	ccv_cnnp_model_apply_gradient_checkpoints(compiled_data, model->graph);
1168	for (i = 0; i < output_size; i++)
1169	{
1170		const ccv_nnc_tensor_symbol_t df = ccv_nnc_tensor_symbol_for_backward(model->graph, compiled_data->f[i]);
1171		// Init this to 1 so we can backprop.
1172		ccv_nnc_tensor_symbol_set_flags(model->graph, df, CCV_NNC_TENSOR_SYMBOL_INIT_ONES);
1173	}
1174	compiled_data->backward.to_size = 0;
1175	for (i = 0; i < parameter_size_maybe_more; i++)
1176		if (compiled_data->gradients[i].d != CCV_NNC_NO_TENSOR_SYMBOL)
1177			compiled_data->backward.tos[compiled_data->backward.to_size++] = ccv_nnc_graph_exec_symbol_for_backward(model->graph, compiled_data->gradients[i]);
1178	ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS);
1179	ccv_nnc_symbolic_graph_set_destinations(model->graph, compiled_data->update_nodes, parameter_size);
1180	for (i = 0; i < parameter_size_maybe_more - parameter_size; i++)
1181	{
1182		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.
1183			continue;
1184		const ccv_nnc_graph_exec_symbol_t outgrad = ccv_nnc_graph_exec_symbol_for_backward(model->graph, compiled_data->outgrads[i]);
1185		const int* tos;
1186		int to_size;
1187		ccv_nnc_graph_exec_symbol_to(model->graph, outgrad, &tos, &to_size);
1188		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.
1189		{
1190			const ccv_nnc_graph_exec_symbol_t* destinations = ccv_nnc_symbolic_graph_destinations(model->graph);
1191			const int destination_count = ccv_nnc_symbolic_graph_destination_size(model->graph);
1192			int flag = 0;
1193			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; });
1194			for (j = i - 1; !flag && j >= 0; j--)
1195				if (j + outgrad_destination_start < destination_count)
1196					flag = (destinations[j + outgrad_destination_start].d == outgrad.d);
1197			if (!flag) // Only if we cannot find it, we add it.
1198				ccv_nnc_symbolic_graph_add_destination(model->graph, outgrad);
1199		}
1200	}
1201	if (parallel_count > 1)
1202	{
1203		ccv_nnc_symbolic_graph_data_parallel(model->graph, parallel_count,
1204			0, 0,
1205			compiled_data->gradients, parameter_size /* No need to deal with outgrads, we don't allreduce outgrads */,
1206			compiled_data->gradients /* We only care about gradients before allreduce, thus, update our current pointers */,
1207			0, 0, 0,
1208			CCV_NNC_PARALLEL_REDUCE_OP_SUM,
1209			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));
1210		ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
1211		for (i = 0; i < evaluate_to_size; i++)
1212			for (j = 1; j < parallel_count; j++)
1213			{
1214				const ccv_nnc_graph_exec_symbol_t copy = ccv_nnc_graph_exec_symbol_copy(model->graph, compiled_data->evaluate.tos[i], j);
1215				if (copy.d != CCV_NNC_NO_GRAPH_EXEC_SYMBOL)
1216					compiled_data->evaluate.tos[compiled_data->evaluate.to_size++] = copy;
1217			}
1218		const int backward_to_size = compiled_data->backward.to_size;
1219		for (i = 0; i < backward_to_size; i++)
1220			for (j = 1; j < parallel_count; j++)
1221			{
1222				const ccv_nnc_graph_exec_symbol_t copy = ccv_nnc_graph_exec_symbol_copy(model->graph, compiled_data->backward.tos[i], j);
1223				if (copy.d != CCV_NNC_NO_GRAPH_EXEC_SYMBOL)
1224					compiled_data->backward.tos[compiled_data->backward.to_size++] = copy;
1225			}
1226	}
1227	// Only use memory compression if we are in gradient parameter mode.
1228	if (gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES || gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES_AND_INPUTS)
1229	{
1230		if (model->memory_compression)
1231			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));
1232		if (model->memory_reduction)
1233			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));
1234	}
1235	compiled_data->backward.to_size = _ccv_nnc_array_dedup_graph_exec_symbols(compiled_data->backward.tos, compiled_data->backward.to_size);
1236	compiled_data->gradient_mode = gradient_mode;
1237}
1238 
1239void ccv_cnnp_model_tensors_init_0(const ccv_cnnp_model_t* const model, ccv_cnnp_compiled_data_t* const compiled_data)
1240{
1241	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", 1241, __extension__ __PRETTY_FUNCTION__
); }));
1242	const int parameter_size = compiled_data->parameters->rnum;
1243	const int parallel_count = _ccv_cnnp_model_effective_parallel_count(model);
1244	compiled_data->parallel_count = parallel_count;
1245	const int internal_size = compiled_data->internals->rnum;
1246	compiled_data->tensors_init.size = ccv_nnc_tensor_symbol_count(model->graph);
1247	compiled_data->tensors_init.v = cccalloccalloc(((compiled_data->tensors_init.size + 31) >> 5), sizeof(uint32_t));
1248	compiled_data->tensors.parameters = (ccv_nnc_tensor_t**)cccalloccalloc((parameter_size + internal_size) * parallel_count, sizeof(ccv_nnc_tensor_t*));
1249	compiled_data->tensors.internals = compiled_data->tensors.parameters + parameter_size * parallel_count;
1250}
1251 
1252int ccv_cnnp_model_tensors_any_to_alloc(const ccv_cnnp_model_t* const model, ccv_cnnp_compiled_data_t* const compiled_data)
1253{
1254	int i, j;
1255	const int parameter_size = compiled_data->parameters->rnum;
1256	const int parallel_count = _ccv_cnnp_compiled_data_parallel_count(model, compiled_data);
1257	const int internal_size = compiled_data->internals->rnum;
1258	for (i = 0; i < parameter_size; i++)
1259	{
1260		// parameters has to be allocated all together.
1261		if (compiled_data->tensors.parameters[i])
1262		{
1263			for (j = 1; j < parallel_count; j++)
1264				{ 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", 1264, __extension__ __PRETTY_FUNCTION__
); })); }
1265			continue;
1266		}
1267		return 1;
1268	}
1269	for (i = 0; i < internal_size; i++)
1270	{
1271		if (!compiled_data->tensors.internals[i])
1272			return 1;
1273		for (j = 1; j < parallel_count; j++)
1274			if (!compiled_data->tensors.internals[i + j * internal_size])
1275				return 1;
1276	}
1277	return 0;
1278}
1279 
1280void ccv_cnnp_model_tensors_init_1(const ccv_cnnp_model_t* const model, ccv_cnnp_compiled_data_t* const compiled_data)
1281{
1282	int i, j;
1283	const int parameter_size = compiled_data->parameters->rnum;
1284	const int parallel_count = _ccv_cnnp_compiled_data_parallel_count(model, compiled_data);
1285	compiled_data->parallel_count = parallel_count;
1286	const int internal_size = compiled_data->internals->rnum;
1287	for (i = 0; i < parameter_size; i++)
1288	{
1289		// parameters has to be allocated all together.
1290		if (compiled_data->tensors.parameters[i])
1291		{
1292			for (j = 1; j < parallel_count; j++)
1293				{ 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", 1293, __extension__ __PRETTY_FUNCTION__
); })); }
1294			continue;
1295		}
1296		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)));
1297		ccv_nnc_tensor_param_t info = ccv_nnc_tensor_symbol_params(parameter.graph, parameter);
1298		if (CCV_TENSOR_GET_DEVICE(info.type)((info.type) & 0xfff00) == CCV_COMPUTE_DEVICE_ANY)
1299			CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff
) << 8));
1300		const int device_id = CCV_TENSOR_GET_DEVICE_ID(info.type)(((info.type) & 0xfff00) >> 8);
1301		compiled_data->tensors.parameters[i] = ccv_nnc_tensor_new(0, info, 0);
1302		for (j = 1; j < parallel_count; j++)
1303		{
1304			if (j != device_id)
1305				CCV_TENSOR_SET_DEVICE_ID(info.type, j)(info.type) = (((info.type) & ~0xfff00) | (((j) & 0xfff
) << 8));
1306			else
1307				CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff
) << 8));
1308			compiled_data->tensors.parameters[i + j * parameter_size] = ccv_nnc_tensor_new(0, info, 0);
1309		}
1310	}
1311	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));
1312	for (i = 0; i < internal_size; i++)
1313	{
1314		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))
);
1315		const int d = retained.d;
1316		if (init_v[d >> 5] & (1u << (d & 0x1f)))
1317			continue;
1318		ccv_nnc_tensor_param_t info = ccv_nnc_tensor_symbol_params(retained.graph, retained);
1319		if (CCV_TENSOR_GET_DEVICE(info.type)((info.type) & 0xfff00) == CCV_COMPUTE_DEVICE_ANY)
1320			CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff
) << 8));
1321		const int device_id = CCV_TENSOR_GET_DEVICE_ID(info.type)(((info.type) & 0xfff00) >> 8);
1322		if (!compiled_data->tensors.internals[i])
1323			compiled_data->tensors.internals[i] = ccv_nnc_tensor_new(0, info, 0);
1324		for (j = 1; j < parallel_count; j++)
1325		{
1326			if (j != device_id)
1327				CCV_TENSOR_SET_DEVICE_ID(info.type, j)(info.type) = (((info.type) & ~0xfff00) | (((j) & 0xfff
) << 8));
1328			else
1329				CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff
) << 8));
1330			if (!compiled_data->tensors.internals[i + j * internal_size])
1331				compiled_data->tensors.internals[i + j * internal_size] = ccv_nnc_tensor_new(0, info, 0);
1332		}
1333	}
1334	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.
1335}
1336 
1337static void _ccv_cnnp_model_tensors_init(const ccv_cnnp_model_t* const model, ccv_cnnp_compiled_data_t* const compiled_data)
1338{
1339	ccv_cnnp_model_tensors_init_0(model, compiled_data);
1340	ccv_cnnp_model_tensors_init_1(model, compiled_data);
1341}
1342 
1343static 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)
1344{
1345	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", 1345, __extension__ __PRETTY_FUNCTION__
); }));
1346	int i, j;
1347	for (i = 0; i < tensor_size; i++)
1348	{
1349		if (!tensors[i])
1350			continue;
1351		const int d = tensor_symbols[i].d;
1352		if (!(tensors_init[d >> 5] & (1u << (d & 0x1f))))
1353			continue;
1354		for (j = 1; j < parallel_count; j++)
1355			if (tensors[i + j * tensor_size])
1356			{
1357				ccv_nnc_tensor_t* const input = CCV_NNC_TENSOR(tensors[i])((ccv_nnc_tensor_t*)((uintptr_t)(tensors[i]) & ~(uintptr_t
)1));
1358				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));
1359				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);
1360			}
1361	}
1362}
1363 
1364static 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)
1365{
1366	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", 1366, __extension__ __PRETTY_FUNCTION__
); }));
1367	int i, j;
1368	for (i = 0; i < tensor_size; i++)
1369	{
1370		const ccv_nnc_tensor_symbol_t tensor_symbol = tensor_symbols[i];
1371		for (j = 1; j < parallel_count; j++)
1372		{
1373			const ccv_nnc_tensor_symbol_t copy = ccv_nnc_tensor_symbol_copy(graph, tensor_symbol, j);
1374			ccv_nnc_tensor_t* copy_tensor = tensors[i + j * tensor_size];
1375			if (copy_tensor && copy.d == CCV_NNC_NO_TENSOR_SYMBOL)
1376			{ // We shouldn't allocate this, free it up.
1377				ccv_nnc_tensor_free(tensors[i + j * tensor_size]);
1378				tensors[i + j * tensor_size] = 0;
1379			}
1380		}
1381	}
1382}
1383 
1384static 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)
1385{
1386	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", 1386, __extension__ __PRETTY_FUNCTION__
); }));
1387	int i, j;
1388	for (i = 0; i < tensor_size; i++)
1389	{
1390		ccv_nnc_tensor_symbol_t tensor_symbol = tensor_symbols[i];
1391		if (tensor_symbol.d == CCV_NNC_NO_TENSOR_SYMBOL)
1392			continue;
1393		if (graph)
1394		{
1395			const ccv_nnc_tensor_symbol_t alias_to = ccv_nnc_tensor_symbol_alias_to(graph, tensor_symbol);
1396			if (alias_to.d != CCV_NNC_NO_TENSOR_SYMBOL)
1397				tensor_symbol = alias_to;
1398		}
1399		ccv_nnc_tensor_t* const tensor = CCV_NNC_TENSOR(tensors[i])((ccv_nnc_tensor_t*)((uintptr_t)(tensors[i]) & ~(uintptr_t
)1));
1400		if (tensor && tensor_symbol.d != CCV_NNC_NO_TENSOR_SYMBOL)
1401		{
1402			const ccv_nnc_tensor_bind_t retained_bind = {
1403				.symbol = tensor_symbol,
1404				.tensor = tensor
1405			};
1406			ccv_array_push(tensor_binds, &retained_bind);
1407		}
1408		for (j = 1; j < parallel_count; j++)
1409		{
1410			const ccv_nnc_tensor_symbol_t copy = ccv_nnc_tensor_symbol_copy(graph, tensor_symbol, j);
1411			ccv_nnc_tensor_t* copy_tensor = tensors[i + j * tensor_size];
1412			if (copy_tensor && copy.d != CCV_NNC_NO_TENSOR_SYMBOL)
1413			{
1414				const ccv_nnc_tensor_bind_t bind = {
1415					.symbol = copy,
1416					.tensor = tensors[i + j * tensor_size]
1417				};
1418				ccv_array_push(tensor_binds, &bind);
1419			}
1420		}
1421	}
1422}
1423 
1424static void _ccv_cnnp_compiled_data_graph_free(ccv_cnnp_compiled_data_t* const compiled_data)
1425{
1426	if (compiled_data->graph)
1427		ccv_nnc_graph_free(compiled_data->graph);
1428	compiled_data->graph = 0;
1429	compiled_data->is_test = 0;
1430	if (compiled_data->tensor_arena)
1431		ccv_nnc_tensor_arena_free(compiled_data->tensor_arena);
1432	compiled_data->tensor_arena = 0;
1433	if (compiled_data->graph_exec_arena)
1434		ccv_nnc_graph_exec_arena_free(compiled_data->graph_exec_arena);
1435	compiled_data->graph_exec_arena = 0;
1436	if (compiled_data->backward.from_ops)
1437		ccfreefree(compiled_data->backward.from_ops);
1438	compiled_data->backward.from_ops = 0;
1439	if (compiled_data->evaluate.schedule)
1440		ccv_nnc_graph_static_schedule_free(compiled_data->evaluate.schedule);
1441	compiled_data->evaluate.schedule = 0;
1442	if (compiled_data->backward.schedule)
1443		ccv_nnc_graph_static_schedule_free(compiled_data->backward.schedule);
1444	compiled_data->backward.schedule = 0;
1445}
1446 
1447static void _ccv_cnnp_compiled_data_gradient_free(ccv_cnnp_compiled_data_t* const compiled_data)
1448{
1449	if (compiled_data->gradients)
1450		ccfreefree(compiled_data->gradients);
1451	compiled_data->gradients = 0;
1452	if (compiled_data->updated_parameters)
1453		ccfreefree(compiled_data->updated_parameters);
1454	compiled_data->updated_parameters = 0;
1455	compiled_data->update_nodes = 0;
1456	compiled_data->saved_aux = 0;
1457}
1458 
1459static void _ccv_cnnp_compiled_data_backward_free(ccv_cnnp_compiled_data_t* const compiled_data)
1460{
1461	if (compiled_data->backward.gradients)
1462		ccfreefree(compiled_data->backward.gradients);
1463	compiled_data->backward.gradients = 0;
1464	if (compiled_data->backward.accum)
1465		ccv_nnc_graph_free(compiled_data->backward.accum);
1466	compiled_data->backward.accum = 0;
1467	if (compiled_data->backward.tensor_arena)
1468		ccv_nnc_tensor_arena_free(compiled_data->backward.tensor_arena);
1469	compiled_data->backward.tensor_arena = 0;
1470	if (compiled_data->backward.graph_exec_arena)
1471		ccv_nnc_graph_exec_arena_free(compiled_data->backward.graph_exec_arena);
1472	compiled_data->backward.graph_exec_arena = 0;
1473}
1474 
1475static void _ccv_cnnp_compiled_data_apply_gradients_free(ccv_cnnp_compiled_data_t* const compiled_data)
1476{
1477	if (compiled_data->apply_gradients.graph)
1478		ccv_nnc_graph_free(compiled_data->apply_gradients.graph);
1479	compiled_data->apply_gradients.graph = 0;
1480	if (compiled_data->apply_gradients.tensor_arena)
1481		ccv_nnc_tensor_arena_free(compiled_data->apply_gradients.tensor_arena);
1482	compiled_data->apply_gradients.tensor_arena = 0;
1483	if (compiled_data->apply_gradients.graph_exec_arena)
1484		ccv_nnc_graph_exec_arena_free(compiled_data->apply_gradients.graph_exec_arena);
1485	compiled_data->apply_gradients.graph_exec_arena = 0;
1486}
1487 
1488// Compile the graph to run ccv_cnnp_model_fit
1489static 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)
1490{
1491	int i, j;
1492	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1493	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", 1493, __extension__ __PRETTY_FUNCTION__
); }));
1494	compiled_data->graph_mode = CCV_CNNP_MODEL_GRAPH_FIT_MODE;
1495	const int parallel_count = _ccv_cnnp_model_root_parallel_count(model);
1496	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", 1496, __extension__ __PRETTY_FUNCTION__
); }));
1497	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", 1497
, __extension__ __PRETTY_FUNCTION__); }));
1498	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", 1498, __extension__ __PRETTY_FUNCTION__
); }));
1499	if (compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_NONE)
1500	{
1501		_ccv_cnnp_model_set_rewindables(model);
1502		_ccv_cnnp_model_gradient_init(model, CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES, CCV_CNNP_DISABLE_OUTGRAD_ALL, fits, fit_size);
1503	} else if (compiled_data->gradient_mode != CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES) {
1504		_ccv_cnnp_model_rewind_graph(model);
1505		_ccv_cnnp_compiled_data_gradient_free(compiled_data);
1506		compiled_data->gradient_mode = CCV_CNNP_COMPILED_DATA_GRADIENT_NONE;
1507		_ccv_cnnp_model_gradient_init(model, CCV_CNNP_COMPILED_DATA_GRADIENT_TRAINABLES, CCV_CNNP_DISABLE_OUTGRAD_ALL, fits, fit_size);
1508	}
1509	const int tensors_init = !!compiled_data->tensors_init.v;
1510	if (!tensors_init)
1511		_ccv_cnnp_model_tensors_init(model, compiled_data);
1512	else if ((uintptr_t)compiled_data->tensors_init.v & (uintptr_t)1)
1513	// Check if it is not fully allocated, if it is not, init_1.
1514		ccv_cnnp_model_tensors_init_1(model, compiled_data);
1515	ccv_array_t* const tensor_binds = ccv_array_new(sizeof(ccv_nnc_tensor_bind_t), 0, 0);
1516	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"
, 1516, __extension__ __PRETTY_FUNCTION__); }));
1517	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"
, 1517, __extension__ __PRETTY_FUNCTION__); }));
1518	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", 1518
, __extension__ __PRETTY_FUNCTION__); }));
1519	const int input_size_per_p = input_size / parallel_count;
1520	_ccv_cnnp_model_bind_tensors(model->graph, model->inputs, inputs, input_size_per_p, parallel_count, tensor_binds);
1521	const int output_size_per_p = output_size / parallel_count;
1522	_ccv_cnnp_model_bind_tensors(model->graph, model->outputs, outputs, output_size_per_p, parallel_count, tensor_binds);
1523	const int fit_size_per_p = fit_size / parallel_count;
1524	_ccv_cnnp_model_bind_tensors(model->graph, compiled_data->fits, fits, fit_size_per_p, parallel_count, tensor_binds);
1525	const int parameter_size = compiled_data->parameters->rnum;
1526	_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);
1527	_ccv_cnnp_model_bind_tensors(model->graph, compiled_data->updated_parameters, compiled_data->tensors.parameters, parameter_size, parallel_count, tensor_binds);
1528	const int internal_size = compiled_data->internals->rnum;
1529	_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);
1530	_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);
1531	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);
1532	ccv_array_free(tensor_binds);
1533	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));
1534	if (tensors_init && parallel_count > 1)
1535		_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);
1536	// If tensor is not init'ed, we need to init states first.
1537	if (_ccv_cnnp_any_to_init(compiled_data))
1538	{
1539		ccv_nnc_tensor_init_states_t tensor_init_states = {
1540			.parallel_count = parallel_count,
1541			.graph = model->graph,
1542			.compiled_data = compiled_data,
1543			.tensor_arena = compiled_data->tensor_arena
1544		};
1545		ccv_cnnp_model_init_states(model, model->graph, _ccv_cnnp_init_states_for_tensors, &tensor_init_states);
1546	}
1547	compiled_data->is_test = 0;
1548	const int saved_aux_size = ccv_nnc_minimizer_saved_aux_size(compiled_data->minimize.minimizer);
1549	// No need to set because it is default to training mode.
1550	// ccv_cnnp_model_set_is_test(model, 0, _ccv_cnnp_cmd_update_for_execs, &update);
1551	for (i = 0; i < saved_aux_size * parameter_size; i++)
1552	{
1553		if (compiled_data->saved_aux[i].source.d == CCV_NNC_NO_TENSOR_SYMBOL)
1554			continue;
1555		ccv_nnc_tensor_t* const tensor = ccv_nnc_tensor_from_symbol(compiled_data->tensor_arena, compiled_data->saved_aux[i].source);
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, &tensor, 1, 0);
1557		for (j = 1; j < parallel_count; j++)
1558		{
1559			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));
1560			if (copy)
1561				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);
1562		}
1563	}
1564	const int evaluate_to_size = compiled_data->evaluate.to_size;
1565	compiled_data->evaluate.to_op_size = 0;
1566	for (i = 0; i < evaluate_to_size; i++)
1567	{
1568		ccv_nnc_graph_exec_t const to = ccv_nnc_graph_exec_from_symbol(compiled_data->graph_exec_arena, compiled_data->evaluate.tos[i]);
1569		if (to.graph)
1570			compiled_data->evaluate.to_ops[compiled_data->evaluate.to_op_size++] = to;
1571	}
1572	ccv_nnc_graph_set_default_static_schedule(compiled_data->graph, compiled_data->stream_type, model->max_stream_count);
1573	ccv_nnc_graph_autotune(compiled_data->graph, model->workspace_size, 0, TRAVERSE_FULL0,0,0,0);
1574}
1575 
1576ccv_nnc_stream_context_t* ccv_cnnp_model_default_stream(const ccv_cnnp_model_t* const model)
1577{
1578	const ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1579	if (!compiled_data || !compiled_data->graph)
1580		return 0;
1581	return ccv_nnc_graph_default_stream(compiled_data->graph);
1582}
1583 
1584uint64_t ccv_cnnp_model_memory_size(const ccv_cnnp_model_t* const model)
1585{
1586	const ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1587	if (!compiled_data || !compiled_data->tensor_arena)
1588		return 0;
1589	return ccv_nnc_tensor_arena_size(compiled_data->tensor_arena);
1590}
1591 
1592static 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)
1593{
1594	int i, j;
1595	for (i = 0; i < tensor_size; i++)
1596	{
1597		ccv_nnc_tensor_symbol_t tensor_symbol = tensor_symbols[i];
1598		if (tensor_symbol.d == CCV_NNC_NO_TENSOR_SYMBOL)
1599			continue;
1600		if (graph)
1601		{
1602			const ccv_nnc_tensor_symbol_t alias_to = ccv_nnc_tensor_symbol_alias_to(graph, tensor_symbol);
1603			if (alias_to.d != CCV_NNC_NO_TENSOR_SYMBOL)
1604				tensor_symbol = alias_to;
1605		}
1606		ccv_nnc_tensor_bind_symbol(tensor_arena, tensor_symbol, tensors[i]);
1607		for (j = 1; j < parallel_count; j++)
1608		{
1609			const ccv_nnc_tensor_symbol_t copy = ccv_nnc_tensor_symbol_copy(graph, tensor_symbol, j);
1610			if (copy.d != CCV_NNC_NO_TENSOR_SYMBOL)
1611				ccv_nnc_tensor_bind_symbol(tensor_arena, copy, tensors[i + tensor_size * j]);
1612		}
1613	}
1614}
1615 
1616void 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)
1617{
1618	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1619	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 1619, __extension__ __PRETTY_FUNCTION__); }));
1620	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; });
1621	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", 1621, __extension__ __PRETTY_FUNCTION__
); }));
1622	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", 1622, __extension__ __PRETTY_FUNCTION__
); }));
1623	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", 1623
, __extension__ __PRETTY_FUNCTION__); }));
1624	assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if
 (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c"
, 1624, __extension__ __PRETTY_FUNCTION__); }));
1625	if (!compiled_data->graph || compiled_data->graph_mode != CCV_CNNP_MODEL_GRAPH_FIT_MODE)
1626	{
1627		_ccv_cnnp_compiled_data_graph_free(compiled_data);
1628		_ccv_cnnp_compiled_data_backward_free(compiled_data);
1629		_ccv_cnnp_compiled_data_apply_gradients_free(compiled_data);
1630		// Compile the symbolic graph down only when needed.
1631		_ccv_cnnp_model_fit_jit(model, inputs, input_size, fits, fit_size, outputs, output_size);
1632	} else {
1633		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"
, 1633, __extension__ __PRETTY_FUNCTION__); }));
1634		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"
, 1634, __extension__ __PRETTY_FUNCTION__); }));
1635		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", 1635
, __extension__ __PRETTY_FUNCTION__); }));
1636		const int input_size_per_p = input_size / parallel_count;
1637		_ccv_cnnp_bind_tensors_to_arena(compiled_data->tensor_arena, model->graph, model->inputs, inputs, input_size_per_p, parallel_count);
1638		const int output_size_per_p = output_size / parallel_count;
1639		_ccv_cnnp_bind_tensors_to_arena(compiled_data->tensor_arena, model->graph, model->outputs, outputs, output_size_per_p, parallel_count);
1640		const int fit_size_per_p = fit_size / parallel_count;
1641		_ccv_cnnp_bind_tensors_to_arena(compiled_data->tensor_arena, model->graph, compiled_data->fits, fits, fit_size_per_p, parallel_count);
1642	}
1643	if (compiled_data->is_test)
1644	{
1645		compiled_data->is_test = 0;
1646		ccv_nnc_graph_exec_update_t update = {
1647			.parallel_count = parallel_count,
1648			.graph = model->graph,
1649			.graph_exec_arena = compiled_data->graph_exec_arena,
1650		};
1651		ccv_cnnp_model_set_is_test(model, 0, _ccv_cnnp_cmd_update_for_execs, &update);
1652	}
1653	ccv_nnc_graph_run_with_schedule(compiled_data->graph, 0, 0, tensor_tape, stream_context);
1654}
1655 
1656// Compile the graph to run ccv_cnnp_model_evaluate with require_grad = false (MULTISTAGE_MODE_NO_GRAD).
1657static 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)
1658{
1659	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1660	compiled_data->graph_mode = CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE_NO_GRAD;
1661	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; });
1662	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", 1662, __extension__ __PRETTY_FUNCTION__
); }));
1663	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", 1663, __extension__ __PRETTY_FUNCTION__
); }));
1664	// If the gradient is not initialized, continue to setup parallel process. We don't init gradient here, but rather,
1665	// we setup proper rewindables so the graph can be rewinded to previous state before we run data parallel.
1666	if (parallel_count > 1 && compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_NONE)
1667	{
1668		const int evaluate_to_size = compiled_data->evaluate.to_size;
1669		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);
1670		_ccv_cnnp_model_set_rewindables(model);
1671		ccv_nnc_symbolic_graph_data_parallel(model->graph, parallel_count,
1672			0, 0,
1673			0, 0, 0,
1674			0, 0, 0,
1675			CCV_NNC_PARALLEL_REDUCE_OP_SUM,
1676			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));
1677		ccv_nnc_graph_exec_symbol_autogen(model->graph, 0, 0, CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
1678		int i, j;
1679		for (i = 0; i < evaluate_to_size; i++)
1680			for (j = 1; j < parallel_count; j++)
1681			{
1682				const ccv_nnc_graph_exec_symbol_t copy = ccv_nnc_graph_exec_symbol_copy(model->graph, compiled_data->evaluate.tos[i], j);
1683				if (copy.d != CCV_NNC_NO_GRAPH_EXEC_SYMBOL)
1684					compiled_data->evaluate.tos[compiled_data->evaluate.to_size++] = copy;
1685			}
1686	}
1687	const int tensors_init = !!compiled_data->tensors_init.v;
1688	if (!tensors_init)
1689		_ccv_cnnp_model_tensors_init(model, compiled_data);
1690	else if ((uintptr_t)compiled_data->tensors_init.v & (uintptr_t)1)
1691	// Check if it is not fully allocated, if it is not, init_1.
1692		ccv_cnnp_model_tensors_init_1(model, compiled_data);
1693	const int tensor_parallel_count = _ccv_cnnp_compiled_data_parallel_count(model, compiled_data);
1694	ccv_array_t* const tensor_binds = ccv_array_new(sizeof(ccv_nnc_tensor_bind_t), 0, 0);
1695	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"
, 1695, __extension__ __PRETTY_FUNCTION__); }));
1696	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"
, 1696, __extension__ __PRETTY_FUNCTION__); }));
1697	const int input_size_per_p = input_size / parallel_count;
1698	_ccv_cnnp_model_bind_tensors(model->graph, model->inputs, inputs, input_size_per_p, parallel_count, tensor_binds);
1699	const int output_size_per_p = output_size / parallel_count;
1700	_ccv_cnnp_model_bind_tensors(model->graph, model->outputs, outputs, output_size_per_p, parallel_count, tensor_binds);
1701	const int parameter_size = compiled_data->parameters->rnum;
1702	_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);
1703	const int internal_size = compiled_data->internals->rnum;
1704	_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);
1705	_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);
1706	// If we generated gradient for the graph, only compile part of the graph because the rest is irrelevant for evaluation.
1707	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);
1708	ccv_array_free(tensor_binds);
1709	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));
1710	// If tensor is not init'ed, we need to init states first.
1711	if (tensors_init && tensor_parallel_count > 1)
1712		_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);
1713	if (_ccv_cnnp_any_to_init(compiled_data))
1714	{
1715		ccv_nnc_tensor_init_states_t tensor_init_states = {
1716			.parallel_count = tensor_parallel_count,
1717			.graph = model->graph,
1718			.compiled_data = compiled_data,
1719			.tensor_arena = compiled_data->tensor_arena
1720		};
1721		ccv_cnnp_model_init_states(model, model->graph, _ccv_cnnp_init_states_for_tensors, &tensor_init_states);
1722	}
1723	compiled_data->is_test = 1;
1724	ccv_nnc_graph_exec_update_t update = {
1725		.parallel_count = parallel_count,
1726		.graph = model->graph,
1727		.graph_exec_arena = compiled_data->graph_exec_arena,
1728	};
1729	ccv_cnnp_model_set_is_test(model, 1, _ccv_cnnp_cmd_update_for_execs, &update);
1730	ccv_nnc_graph_set_default_static_schedule(compiled_data->graph, compiled_data->stream_type, model->max_stream_count);
1731	ccv_nnc_graph_autotune(compiled_data->graph, model->workspace_size, 0, TRAVERSE_FULL0,0,0,0);
1732}
1733 
1734static void _ccv_cnnp_model_gradient_tensors_init(const ccv_cnnp_model_t* const model, ccv_cnnp_compiled_data_t* const compiled_data)
1735{
1736	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", 1736, __extension__ __PRETTY_FUNCTION__
); }));
1737	const int parameter_size = compiled_data->parameters->rnum;
1738	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; });
1739	compiled_data->tensors.gradients = (ccv_nnc_tensor_t**)ccmallocmalloc(sizeof(ccv_nnc_tensor_t*) * parameter_size * 2 * parallel_count);
1740	compiled_data->tensors.accum_gradients = compiled_data->tensors.gradients + parameter_size * parallel_count;
1741	int i, j;
1742	for (i = 0; i < parameter_size; i++)
1743	{
1744		if (compiled_data->parameter_flags && !(compiled_data->parameter_flags[i >> 6] & ((uint64_t)1 << (i & 63))))
1745		{
1746			compiled_data->tensors.gradients[i] = 0;
1747			compiled_data->tensors.accum_gradients[i] = 0;
1748			for (j = 1; j < parallel_count; j++)
1749			{
1750				compiled_data->tensors.gradients[i + j * parameter_size] = 0;
1751				compiled_data->tensors.accum_gradients[i + j * parameter_size] = 0;
1752			}
1753			continue;
1754		}
1755		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)));
1756		ccv_nnc_tensor_param_t info = ccv_nnc_tensor_symbol_params(parameter.graph, parameter);
1757		if (CCV_TENSOR_GET_DEVICE(info.type)((info.type) & 0xfff00) == CCV_COMPUTE_DEVICE_ANY)
1758			CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff
) << 8));
1759		const int device_id = CCV_TENSOR_GET_DEVICE_ID(info.type)(((info.type) & 0xfff00) >> 8);
1760		compiled_data->tensors.gradients[i] = ccv_nnc_tensor_new(0, info, 0);
1761		compiled_data->tensors.accum_gradients[i] = 0; // delay the accumulated gradient allocation until when we need it.
1762		for (j = 1; j < parallel_count; j++)
1763		{
1764			if (j != device_id)
1765				CCV_TENSOR_SET_DEVICE_ID(info.type, j)(info.type) = (((info.type) & ~0xfff00) | (((j) & 0xfff
) << 8));
1766			else
1767				CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff
) << 8));
1768			compiled_data->tensors.gradients[i + j * parameter_size] = ccv_nnc_tensor_new(0, info, 0);
1769			compiled_data->tensors.accum_gradients[i + j * parameter_size] = 0;
1770		}
1771	}
1772}
1773 
1774static int _ccv_cnnp_is_disable_outgrad_all(const uint64_t disable_outgrad, const int input_size)
1775{
1776	if (disable_outgrad == CCV_CNNP_DISABLE_OUTGRAD_ALL)
1777		return 1;
1778	if (disable_outgrad == CCV_CNNP_DISABLE_OUTGRAD_NONE)
1779		return 0;
1780	int i;
1781	for (i = 0; i < input_size; i++)
1782		if (!(disable_outgrad & ((uint64_t)1 << i)))
1783			return 0;
1784	return 1;
1785}
1786 
1787// Compile the graph to run ccv_cnnp_model_evaluate with requires_grad = true (MULTISTAGE_MODE).
1788// Particularly, this method compiles the evaluation and backprop graph (the main graph).
1789static 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)
1790{
1791	int i, j;
1792	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1793	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;
1794	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", 1794, __extension__ __PRETTY_FUNCTION__
); }));
1795	compiled_data->graph_mode = CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE;
1796	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; });
1797	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", 1797, __extension__ __PRETTY_FUNCTION__
); }));
1798	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", 1798, __extension__ __PRETTY_FUNCTION__
); }));
1799	// There shouldn't be a loss function if we evaluate with multistage jit.
1800	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", 1800, __extension__ __PRETTY_FUNCTION__
); }));
1801	if (compiled_data->gradient_mode == CCV_CNNP_COMPILED_DATA_GRADIENT_NONE)
1802	{
1803		_ccv_cnnp_model_set_rewindables(model);
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	} else if (compiled_data->gradient_mode != target_gradient_mode) {
1806		_ccv_cnnp_model_rewind_graph(model);
1807		_ccv_cnnp_compiled_data_gradient_free(compiled_data);
1808		compiled_data->gradient_mode = CCV_CNNP_COMPILED_DATA_GRADIENT_NONE;
1809		_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.
1810	}
1811	const int tensors_init = !!compiled_data->tensors_init.v;
1812	if (!tensors_init)
1813		_ccv_cnnp_model_tensors_init(model, compiled_data);
1814	else if ((uintptr_t)compiled_data->tensors_init.v & (uintptr_t)1)
1815	// Check if it is not fully allocated, if it is not, init_1.
1816		ccv_cnnp_model_tensors_init_1(model, compiled_data);
1817	ccv_array_t* const tensor_binds = ccv_array_new(sizeof(ccv_nnc_tensor_bind_t), 0, 0);
1818	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"
, 1818, __extension__ __PRETTY_FUNCTION__); }));
1819	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"
, 1819, __extension__ __PRETTY_FUNCTION__); }));
1820	const int input_size_per_p = input_size / parallel_count;
1821	_ccv_cnnp_model_bind_tensors(model->graph, model->inputs, inputs, input_size_per_p, parallel_count, tensor_binds);
1822	const int output_size_per_p = output_size / parallel_count;
1823	_ccv_cnnp_model_bind_tensors(model->graph, model->outputs, outputs, output_size_per_p, parallel_count, tensor_binds);
1824	const int parameter_size = compiled_data->parameters->rnum;
1825	_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);
1826	const int internal_size = compiled_data->internals->rnum;
1827	_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);
1828	_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);
1829	if (!compiled_data->tensors.gradients)
1830		_ccv_cnnp_model_gradient_tensors_init(model, compiled_data);
1831	_ccv_cnnp_model_bind_tensors(model->graph, compiled_data->gradients, compiled_data->tensors.gradients, parameter_size, parallel_count, tensor_binds);
1832	if (compiled_data->backward.to_size > 0)
1833		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);
1834	else
1835		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);
1836	ccv_array_free(tensor_binds);
1837	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));
1838	if (tensors_init && parallel_count > 1)
1839		_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);
1840	// If tensor is not init'ed, we need to init states first.
1841	if (_ccv_cnnp_any_to_init(compiled_data))
1842	{
1843		ccv_nnc_tensor_init_states_t tensor_init_states = {
1844			.parallel_count = parallel_count,
1845			.graph = model->graph,
1846			.compiled_data = compiled_data,
1847			.tensor_arena = compiled_data->tensor_arena
1848		};
1849		ccv_cnnp_model_init_states(model, model->graph, _ccv_cnnp_init_states_for_tensors, &tensor_init_states);
1850	}
1851	compiled_data->is_test = is_test;
1852	ccv_nnc_graph_exec_update_t update = {
1853		.parallel_count = parallel_count,
1854		.graph = model->graph,
1855		.graph_exec_arena = compiled_data->graph_exec_arena,
1856	};
1857	ccv_cnnp_model_set_is_test(model, is_test, _ccv_cnnp_cmd_update_for_execs, &update);
1858	const int evaluate_to_size = compiled_data->evaluate.to_size;
1859	compiled_data->evaluate.to_op_size = 0;
1860	ccv_array_t* const backward_from = ccv_array_new(sizeof(int), 0, 0);
1861	for (i = 0; i < evaluate_to_size; i++)
1862	{
1863		ccv_nnc_graph_exec_t const to_op = ccv_nnc_graph_exec_from_symbol(compiled_data->graph_exec_arena, compiled_data->evaluate.tos[i]);
1864		if (to_op.graph)
1865			compiled_data->evaluate.to_ops[compiled_data->evaluate.to_op_size++] = to_op;
1866		const int* tos;
1867		int to_size;
1868		ccv_nnc_graph_exec_symbol_to(model->graph, compiled_data->evaluate.tos[i], &tos, &to_size);
1869		for (j = 0; j < to_size; j++)
1870		{
1871			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){
1872				.d = tos[j],
1873				.graph = model->graph
1874			});
1875			if (to_op.graph)
1876				ccv_array_add_unique_int(backward_from, to_op.d);
1877		}
1878	}
1879	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", 1879, __extension__
 __PRETTY_FUNCTION__); }));
1880	compiled_data->backward.from_op_size = backward_from->rnum;
1881	compiled_data->backward.from_ops = (ccv_nnc_graph_exec_t*)ccmallocmalloc(sizeof(ccv_nnc_graph_exec_t) * backward_from->rnum);
1882	for (i = 0; i < backward_from->rnum; i++)
1883		compiled_data->backward.from_ops[i] = (ccv_nnc_graph_exec_t){
1884			.d = *(int*)ccv_array_get(backward_from, i)((void*)(((char*)((backward_from)->data)) + (size_t)(backward_from
)->rsize * (size_t)(i))),
1885			.graph = compiled_data->graph,
1886		};
1887	// 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.
1888	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)));
1889	const int exec_info_size = compiled_data->graph->exec_info->rnum;
1890	uint32_t* const visited = cccalloccalloc((exec_info_size + 31) >> 5, sizeof(uint32_t));
1891	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)));
1892	const int source_size = compiled_data->graph->sources->rnum;
1893	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", 1893, __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", 1893, __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", 1893, __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", 1893, __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", 1893, __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", 1893, __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", 1893, __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", 1893, __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", 1893, __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", 1893, __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", 1893, __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"
, 1893, __extension__ __PRETTY_FUNCTION__); })); _visit_; });
1894	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; {
1895		visited[(idx >> 5)] |= (1u << (idx & 31));
1896	} ccv_nnc_graph_visit_endfor} }
1897	ccv_nnc_graph_visit_free(visit);
1898	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)));
1899	const int destination_size = compiled_data->graph->destinations->rnum;
1900	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", 1900, __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", 1900, __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", 1900, __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"
, 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_ < (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", 1900, __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", 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	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; {
1902		visited[(idx >> 5)] |= (1u << (idx & 31));
1903	} ccv_nnc_graph_visit_endfor} }
1904	ccv_nnc_graph_visit_free(visit);
1905	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", 1905, __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", 1905, __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", 1905, __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", 1905, __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", 1905, __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", 1905, __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", 1905, __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", 1905, __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", 1905, __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", 1905, __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", 1905, __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", 1905, __extension__ __PRETTY_FUNCTION__
); })); _visit_; });
1906	// Find any missing nodes to be added as source. Right now, these are only set nodes.
1907	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; {
1908		if (!(visited[(idx >> 5)] & (1u << (idx & 31))))
1909		{
1910			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", 1910, __extension__ __PRETTY_FUNCTION__
); }));
1911			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.
1912				ccv_array_add_unique_int(backward_from, idx);
1913		}
1914	} ccv_nnc_graph_visit_endfor} }
1915	ccv_nnc_graph_visit_free(visit);
1916	ccfreefree(visited);
1917	if (backward_from->rnum != compiled_data->backward.from_op_size) // If it doesn't match, need to redo this.
1918	{
1919		compiled_data->backward.from_op_size = backward_from->rnum;
1920		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);
1921		for (i = 0; i < backward_from->rnum; i++)
1922			compiled_data->backward.from_ops[i] = (ccv_nnc_graph_exec_t){
1923				.d = *(int*)ccv_array_get(backward_from, i)((void*)(((char*)((backward_from)->data)) + (size_t)(backward_from
)->rsize * (size_t)(i))),
1924				.graph = compiled_data->graph,
1925			};
1926	}
1927	ccv_array_free(backward_from);
1928	ccv_nnc_graph_set_default_static_schedule(compiled_data->graph, compiled_data->stream_type, model->max_stream_count);
1929	ccv_nnc_graph_autotune(compiled_data->graph, model->workspace_size, 0, TRAVERSE_FULL0,0,0,0);
1930}
1931 
1932void 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)
1933{
1934	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1935	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 1935, __extension__ __PRETTY_FUNCTION__); }));
1936	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; });
1937	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", 1937, __extension__ __PRETTY_FUNCTION__
); }));
1938	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", 1938, __extension__ __PRETTY_FUNCTION__
); }));
1939	assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if
 (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c"
, 1939, __extension__ __PRETTY_FUNCTION__); }));
1940	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;
1941	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));
1942	if (!compiled_data->graph || mode_mismatch)
1943	{
1944		_ccv_cnnp_compiled_data_graph_free(compiled_data);
1945		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.
1946			_ccv_cnnp_compiled_data_backward_free(compiled_data);
1947		if (params.requires_grad)
1948			_ccv_cnnp_model_multistage_jit_0(model, params.disable_outgrad, params.is_test, inputs, input_size, outputs, output_size);
1949		else
1950			_ccv_cnnp_model_multistage_no_grad_jit(model, inputs, input_size, outputs, output_size);
1951	} else {
1952		ccv_nnc_tensor_arena_clear_bindings(compiled_data->tensor_arena);
1953		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"
, 1953, __extension__ __PRETTY_FUNCTION__); }));
1954		const int input_size_per_p = input_size / parallel_count;
1955		_ccv_cnnp_bind_tensors_to_arena(compiled_data->tensor_arena, model->graph, model->inputs, inputs, input_size_per_p, parallel_count);
1956		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"
, 1956, __extension__ __PRETTY_FUNCTION__); }));
1957		const int output_size_per_p = output_size / parallel_count;
1958		_ccv_cnnp_bind_tensors_to_arena(compiled_data->tensor_arena, model->graph, model->outputs, outputs, output_size_per_p, parallel_count);
1959	}
1960	if (compiled_data->is_test != params.is_test)
1961	{
1962		compiled_data->is_test = params.is_test;
1963		ccv_nnc_graph_exec_update_t update = {
1964			.parallel_count = parallel_count,
1965			.graph = model->graph,
1966			.graph_exec_arena = compiled_data->graph_exec_arena,
1967		};
1968		ccv_cnnp_model_set_is_test(model, params.is_test, _ccv_cnnp_cmd_update_for_execs, &update);
1969	}
1970}
1971 
1972void 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)
1973{
1974	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1975	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 1975, __extension__ __PRETTY_FUNCTION__); }));
1976	ccv_cnnp_model_dry_run(model, params, inputs, input_size, outputs, output_size);
1977	if (compiled_data->graph_mode == CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE_NO_GRAD)
1978		ccv_nnc_graph_run_with_schedule(compiled_data->graph, 0, 0, tensor_tape, stream_context);
1979	else {
1980		if (!compiled_data->evaluate.schedule)
1981			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);
1982		ccv_nnc_graph_run_with_schedule(compiled_data->graph, 0, compiled_data->evaluate.schedule, tensor_tape, stream_context);
1983	}
1984}
1985 
1986// Compile the graph to run ccv_cnnp_model_backward after ccv_cnnp_model_evaluate with requires_grad = true (MULTISTAGE_MODE).
1987// Particularly, this method compiles the accumulator graph.
1988static void _ccv_cnnp_model_multistage_jit_1(ccv_cnnp_model_t* const model)
1989{
1990	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
1991	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 1991, __extension__ __PRETTY_FUNCTION__); }));
1992	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", 1992, __extension__ __PRETTY_FUNCTION__
); }));
1993	ccv_nnc_symbolic_graph_t* accum = ccv_nnc_symbolic_graph_new();
1994	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; });
1995	const int parameter_size = compiled_data->parameters->rnum;
1996	int i, j;
1997	compiled_data->backward.gradients = (ccv_nnc_tensor_symbol_t*)ccmallocmalloc(sizeof(ccv_nnc_tensor_symbol_t) * parameter_size * parallel_count * 3);
1998	compiled_data->backward.accum_gradients = compiled_data->backward.gradients + parameter_size * parallel_count;
1999	compiled_data->backward.updated_accum_gradients = compiled_data->backward.accum_gradients + parameter_size * parallel_count;
2000	for (i = 0; i < parameter_size; i++)
2001		for (j = 0; j < parallel_count; j++)
2002			if (compiled_data->tensors.gradients[i + j * parameter_size])
2003			{
2004				const ccv_nnc_tensor_param_t info = compiled_data->tensors.gradients[i + j * parameter_size]->info;
2005				// Now, the old gradient is the accumulated gradient, getting new gradient tensor setup so we can collect them.
2006				compiled_data->tensors.accum_gradients[i + j * parameter_size] = compiled_data->tensors.gradients[i + j * parameter_size];
2007				compiled_data->tensors.gradients[i + j * parameter_size] = ccv_nnc_tensor_new(0, info, 0);
2008				ccv_nnc_tensor_symbol_t inputs[2];
2009				inputs[0] = compiled_data->backward.accum_gradients[i + j * parameter_size] = ccv_nnc_tensor_symbol_new(accum, info, 0);
2010				inputs[1] = compiled_data->backward.gradients[i + j * parameter_size] = ccv_nnc_tensor_symbol_new(accum, info, 0);
2011				ccv_nnc_tensor_symbol_t output = compiled_data->backward.updated_accum_gradients[i + j * parameter_size] = ccv_nnc_tensor_symbol_new(accum, info, 0);
2012				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);
2013			} else {
2014				compiled_data->backward.accum_gradients[i + j * parameter_size] = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
2015				compiled_data->backward.gradients[i + j * parameter_size] = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
2016				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
};
2017			}
2018	ccv_nnc_graph_exec_symbol_autogen(accum, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
2019	if (ccv_nnc_symbolic_graph_source_size(accum) == 0)
2020	{
2021		ccv_nnc_symbolic_graph_free(accum);
2022		// Create empty graph.
2023		compiled_data->backward.accum = ccv_nnc_graph_new();
2024		ccv_nnc_graph_topsort(compiled_data->backward.accum, 0, 0);
2025		return;
2026	}
2027	ccv_array_t* const tensor_binds = ccv_array_new(sizeof(ccv_nnc_tensor_bind_t), 0, 0);
2028	_ccv_cnnp_model_bind_tensors(accum, compiled_data->backward.accum_gradients, compiled_data->tensors.accum_gradients, parameter_size * parallel_count, 1, tensor_binds);
2029	_ccv_cnnp_model_bind_tensors(accum, compiled_data->backward.gradients, compiled_data->tensors.gradients, parameter_size * parallel_count, 1, tensor_binds);
2030	_ccv_cnnp_model_bind_tensors(accum, compiled_data->backward.updated_accum_gradients, compiled_data->tensors.accum_gradients, parameter_size * parallel_count, 1, tensor_binds);
2031	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);
2032	ccv_nnc_symbolic_graph_free(accum);
2033	ccv_array_free(tensor_binds);
2034	ccv_nnc_graph_set_default_static_schedule(compiled_data->backward.accum, compiled_data->stream_type, model->max_stream_count);
2035}
2036 
2037void 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)
2038{
2039	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2040	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 2040, __extension__ __PRETTY_FUNCTION__); }));
2041	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", 2041, __extension__ __PRETTY_FUNCTION__
); }));
2042	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; });
2043	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", 2043, __extension__ __PRETTY_FUNCTION__
); }));
2044	if (outgrad_size > 0)
2045		{ 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", 2045, __extension__ __PRETTY_FUNCTION__
); })); }
2046	assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if
 (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c"
, 2046, __extension__ __PRETTY_FUNCTION__); }));
2047	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", 2047, __extension__ __PRETTY_FUNCTION__
); }));
2048	const int parameter_size = compiled_data->parameters->rnum;
2049	// If we need to accumulate the gradients now, do jit on accumulator.
2050	if (compiled_data->backward.count > 0)
2051	{
2052		if (!compiled_data->backward.accum)
2053			_ccv_cnnp_model_multistage_jit_1(model);
2054		else if (compiled_data->backward.count == 1) {
2055			//  On this round, we need to switch accumulated gradients with gradients (so we can do accumulation properly).
2056			int i;
2057			for (i = 0; i < parameter_size * parallel_count; i++)
2058			{
2059				ccv_nnc_tensor_t* tensor;
2060				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));
2061			}
2062			if (compiled_data->backward.tensor_arena)
2063			{
2064				ccv_nnc_tensor_arena_clear_bindings(compiled_data->backward.tensor_arena);
2065				// Do rebind in case we messed up the binding (we switch accum_gradients and gradients).
2066				_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);
2067				_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);
2068				_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);
2069			}
2070		}
2071	}
2072	const int ingrad_size_per_p = model->output_size;
2073	const int outgrad_size_per_p = compiled_data->outgrad_size;
2074	int i, j;
2075	for (i = 0; i < ingrad_size_per_p; i++)
2076	{
2077		const ccv_nnc_tensor_symbol_t ingrad = ccv_nnc_tensor_symbol_for_backward(model->graph, compiled_data->f[i]);
2078		if (!ingrad_size || !ingrads || ingrads[i] == 0)
2079		{
2080			// Set it to 1 if it is not specified.
2081			ccv_nnc_tensor_t* const ingrad_tensor = ccv_nnc_tensor_from_symbol(compiled_data->tensor_arena, ingrad);
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			for (j = 1; j < parallel_count; j++)
2085			{
2086				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));
2087				if (ingrad_tensor)
2088					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);
2089			}
2090		} else {
2091			// Make sure the length matches, in case it is an alias.
2092			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", 2092, __extension__ __PRETTY_FUNCTION__
); }));
2093			ccv_nnc_tensor_bind_symbol(compiled_data->tensor_arena, ingrad, ingrads[i]);
2094			for (j = 1; j < parallel_count; j++)
2095				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]);
2096		}
2097	}
2098	if (outgrad_size > 0)
2099	{
2100		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", 2100, __extension__ __PRETTY_FUNCTION__
); }));
2101		for (i = 0; i < outgrad_size_per_p; i++)
2102			if (outgrads[i])
2103			{
2104				const ccv_nnc_tensor_symbol_t outgrad = compiled_data->outgrads[i];
2105				ccv_nnc_tensor_bind_symbol(compiled_data->tensor_arena, outgrad, outgrads[i]);
2106				for (j = 1; j < parallel_count; j++)
2107					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]);
2108			}
2109	} else {
2110		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", 2111, __extension__ __PRETTY_FUNCTION__
); }))
2111			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", 2111, __extension__ __PRETTY_FUNCTION__
); }));
2112	}
2113	// We need to rebind here because in ccv_cnnp_evaluate, we clear bindings, that will reset all bindings for the gradients.
2114	// For parameters and internals these are fine because when we clear bindings, it restores to original bindings, which are these
2115	// parameters and internals. The same cannot be said for gradients due to the accum_gradients switching.
2116	_ccv_cnnp_bind_tensors_to_arena(compiled_data->tensor_arena, model->graph, compiled_data->gradients, compiled_data->tensors.gradients, parameter_size, parallel_count);
2117	if (!compiled_data->backward.schedule)
2118		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);
2119	// Run the backward pass.
2120	ccv_nnc_graph_run_with_schedule(compiled_data->graph, 0, compiled_data->backward.schedule, tensor_tape, stream_context);
2121	// If we need to run accumulation round, do that now.
2122	if (compiled_data->backward.count > 0)
2123		ccv_nnc_graph_run_with_schedule(compiled_data->backward.accum, 0, 0, 0, stream_context);
2124	// Update the count, this determines whether we need to accumulate or not.
2125	++compiled_data->backward.count;
2126}
2127 
2128// Compile the graph to run ccv_cnnp_model_apply_gradients after ccv_cnnp_model_backward (MULTISTAGE_MODE).
2129// Particularly, this method compiles the parameter update graph.
2130static void _ccv_cnnp_model_multistage_jit_2(ccv_cnnp_model_t* const model)
2131{
2132	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2133	assert(compiled_data->graph_mode == CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE)((void) sizeof ((compiled_data->graph_mode == CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE
) ? 1 : 0), __extension__ ({ if (compiled_data->graph_mode
 == CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE) ; else __assert_fail
 ("compiled_data->graph_mode == CCV_CNNP_MODEL_GRAPH_MULTISTAGE_MODE"
, "ccv_cnnp_model.c", 2133, __extension__ __PRETTY_FUNCTION__
); }));
2134	const int parallel_count = ccv_max(model->parallel_count, 1)({ typeof (model->parallel_count) _a = (model->parallel_count
); typeof (1) _b = (1); (_a > _b) ? _a : _b; });
2135	const int parameter_size = compiled_data->parameters->rnum;
2136	ccv_array_t* const tensor_binds = ccv_array_new(sizeof(ccv_nnc_tensor_bind_t), 0, 0);
2137	_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);
2138	_ccv_cnnp_model_bind_tensors(model->graph, compiled_data->updated_parameters, compiled_data->tensors.parameters, parameter_size, parallel_count, tensor_binds);
2139	// Bind accumulated gradients.
2140	if (compiled_data->backward.count > 1)
2141		_ccv_cnnp_model_bind_tensors(model->graph, compiled_data->gradients, compiled_data->tensors.accum_gradients, parameter_size, parallel_count, tensor_binds);
2142	else
2143		_ccv_cnnp_model_bind_tensors(model->graph, compiled_data->gradients, compiled_data->tensors.gradients, parameter_size, parallel_count, tensor_binds);
2144	ccv_array_t* const apply_gradients_from = ccv_array_new(sizeof(int), 0, 0);
2145	int i, j;
2146	for (i = 0; i < compiled_data->backward.to_size; i++)
2147	{
2148		const int* tos;
2149		int to_size;
2150		ccv_nnc_graph_exec_symbol_to(model->graph, compiled_data->backward.tos[i], &tos, &to_size);
2151		for (j = 0; j < to_size; j++)
2152		{
2153			// Check if this is already show up in the backward graph, if that is the case, it won't be in the apply
2154			// gradients graph.
2155			const ccv_nnc_graph_exec_t exec = ccv_nnc_graph_exec_from_symbol(compiled_data->graph_exec_arena, (ccv_nnc_graph_exec_symbol_t){
2156				.d = tos[j],
2157				.graph = model->graph,
2158			});
2159			if (!exec.graph)
2160				ccv_array_add_unique_int(apply_gradients_from, tos[j]);
2161		}
2162	}
2163	const int from_size = apply_gradients_from->rnum;
2164	if (from_size == 0)
2165	{
2166		ccv_array_free(apply_gradients_from);
2167		ccv_array_free(tensor_binds);
2168		return;
2169	}
2170	ccv_nnc_graph_exec_symbol_t* const froms = (ccv_nnc_graph_exec_symbol_t*)ccmallocmalloc(sizeof(ccv_nnc_graph_exec_symbol_t) * from_size);
2171	for (i = 0; i < from_size; i++)
2172		froms[i] = (ccv_nnc_graph_exec_symbol_t){
2173			.d = *(int*)ccv_array_get(apply_gradients_from, i)((void*)(((char*)((apply_gradients_from)->data)) + (size_t
)(apply_gradients_from)->rsize * (size_t)(i))),
2174			.graph = model->graph
2175		};
2176	ccv_array_free(apply_gradients_from);
2177	// It can only ends with updates on the parameters.
2178	ccv_array_t* const tos = ccv_array_new(sizeof(ccv_nnc_graph_exec_symbol_t), parameter_size * parallel_count, 0);
2179	for (i = 0;  i < parameter_size; i++)
2180	{
2181		if (compiled_data->update_nodes[i].d == CCV_NNC_NO_TENSOR_SYMBOL)
2182			continue;
2183		ccv_array_push(tos, &compiled_data->update_nodes[i]);
2184		for (j = 1; j < parallel_count; j++)
2185		{
2186			const ccv_nnc_graph_exec_symbol_t copy = ccv_nnc_graph_exec_symbol_copy(model->graph, compiled_data->update_nodes[i], j);
2187			ccv_array_push(tos, &copy);
2188		}
2189	}
2190	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);
2191	ccv_array_free(tos);
2192	ccv_array_free(tensor_binds);
2193	ccfreefree(froms);
2194	const int max_saved_aux_size = compiled_data->minimize.max_saved_aux_size;
2195	for (i = 0; i < max_saved_aux_size * parameter_size; i++)
2196	{
2197		// Skip on no tensor.
2198		if (compiled_data->saved_aux[i].source.d == CCV_NNC_NO_TENSOR_SYMBOL)
2199			continue;
2200		ccv_nnc_tensor_t* const tensor = ccv_nnc_tensor_from_symbol(compiled_data->apply_gradients.tensor_arena, compiled_data->saved_aux[i].source);
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, &tensor, 1, 0);
2202		for (j = 1; j < parallel_count; j++)
2203		{
2204			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));
2205			if (copy)
2206				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);
2207		}
2208	}
2209	ccv_nnc_graph_set_default_static_schedule(compiled_data->apply_gradients.graph, compiled_data->stream_type, model->max_stream_count);
2210}
2211 
2212void ccv_cnnp_model_apply_gradients(ccv_cnnp_model_t* const model, ccv_nnc_stream_context_t* const stream_context)
2213{
2214	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2215	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 2215, __extension__ __PRETTY_FUNCTION__); }));
2216	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", 2216, __extension__ __PRETTY_FUNCTION__
); }));
2217	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; });
2218	assert(model->graph)((void) sizeof ((model->graph) ? 1 : 0), __extension__ ({ if
 (model->graph) ; else __assert_fail ("model->graph", "ccv_cnnp_model.c"
, 2218, __extension__ __PRETTY_FUNCTION__); }));
2219	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", 2219, __extension__ __PRETTY_FUNCTION__
); }));
2220	// Skip if there is no backward pass.
2221	if (compiled_data->backward.count <= 0)
2222		return;
2223	// Skip if there is no parameters.
2224	if (compiled_data->parameters->rnum == 0)
2225	{
2226		compiled_data->backward.count = 0;
2227		return;
2228	}
2229	if (!compiled_data->apply_gradients.graph)
2230		_ccv_cnnp_model_multistage_jit_2(model);
2231	else {
2232		const int parameter_size = compiled_data->parameters->rnum;
2233		ccv_nnc_tensor_arena_clear_bindings(compiled_data->apply_gradients.tensor_arena);
2234		// Change to bind accum_gradients if we do gradient accumulation (run backward more than once).
2235		if (compiled_data->backward.count > 1)
2236			_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);
2237		else
2238			_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);
2239	}
2240	if (compiled_data->apply_gradients.graph)
2241		ccv_nnc_graph_run_with_schedule(compiled_data->apply_gradients.graph, 0, 0, 0, stream_context);
2242	// Reset backward count to 0.
2243	compiled_data->backward.count = 0;
2244}
2245 
2246void 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)
2247{
2248	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2249	const int param_sel = parameter->param_sel > 0 ? parameter->param_sel - 1 : parameter->param_sel;
2250	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", 2250, __extension__ __PRETTY_FUNCTION__
); }));
2251	const int tensors_init = !!compiled_data->tensors_init.v;
2252	int this_tensor_init = tensors_init;
2253	if (!tensors_init)
2254		ccv_cnnp_model_tensors_init_0(model, compiled_data);
2255	else if ((uintptr_t)compiled_data->tensors_init.v & (uintptr_t)1)
2256	// Check if it is not fully allocated, if it is not, init_1.
2257		this_tensor_init = 0;
2258	ccv_array_t* const parameter_indices = ccv_array_new(sizeof(int), 0, 0);
2259	ccv_cnnp_model_add_to_parameter_indices(parameter->model, param_sel, parameter_indices);
2260	const int param_ref = parameter->param_ref > 0 ? parameter->param_ref - 1 : parameter->param_ref;
2261	if (param_ref < 0)
2262		{ assert(parameter_indices->rnum == 1)((void) sizeof ((parameter_indices->rnum == 1) ? 1 : 0), __extension__
 ({ if (parameter_indices->rnum == 1) ; else __assert_fail
 ("parameter_indices->rnum == 1", "ccv_cnnp_model.c", 2262
, __extension__ __PRETTY_FUNCTION__); })); }
2263	else
2264		{ assert(param_ref < parameter_indices->rnum)((void) sizeof ((param_ref < parameter_indices->rnum) ?
 1 : 0), __extension__ ({ if (param_ref < parameter_indices
->rnum) ; else __assert_fail ("param_ref < parameter_indices->rnum"
, "ccv_cnnp_model.c", 2264, __extension__ __PRETTY_FUNCTION__
); })); }
2265	const int d = *(int*)ccv_array_get(parameter_indices, param_ref >= 0 ? param_ref : 0)((void*)(((char*)((parameter_indices)->data)) + (size_t)(parameter_indices
)->rsize * (size_t)(param_ref >= 0 ? param_ref : 0)));
2266	ccv_array_free(parameter_indices);
2267	const int parameter_size = compiled_data->parameters->rnum;
2268	assert(d >= 0)((void) sizeof ((d >= 0) ? 1 : 0), __extension__ ({ if (d >=
 0) ; else __assert_fail ("d >= 0", "ccv_cnnp_model.c", 2268
, __extension__ __PRETTY_FUNCTION__); }));
2269	assert(d < parameter_size)((void) sizeof ((d < parameter_size) ? 1 : 0), __extension__
 ({ if (d < parameter_size) ; else __assert_fail ("d < parameter_size"
, "ccv_cnnp_model.c", 2269, __extension__ __PRETTY_FUNCTION__
); }));
2270	const int parallel_count = _ccv_cnnp_compiled_data_parallel_count(model, compiled_data);
2271	int i;
2272	if (!this_tensor_init)
2273	{
2274		if (compiled_data->tensors.parameters[d])
2275		{
2276			for (i = 1; i < parallel_count; i++)
2277				{ 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", 2277, __extension__ __PRETTY_FUNCTION__
); })); }
2278			this_tensor_init = 1;
2279		} else {
2280			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)));
2281			ccv_nnc_tensor_param_t info = ccv_nnc_tensor_symbol_params(parameter.graph, parameter);
2282			if (CCV_TENSOR_GET_DEVICE(info.type)((info.type) & 0xfff00) == CCV_COMPUTE_DEVICE_ANY)
2283				CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff
) << 8));
2284			const int device_id = CCV_TENSOR_GET_DEVICE_ID(info.type)(((info.type) & 0xfff00) >> 8);
2285			compiled_data->tensors.parameters[d] = ccv_nnc_tensor_new(0, info, 0);
2286			for (i = 1; i < parallel_count; i++)
2287			{
2288				if (i != device_id)
2289					CCV_TENSOR_SET_DEVICE_ID(info.type, i)(info.type) = (((info.type) & ~0xfff00) | (((i) & 0xfff
) << 8));
2290				else
2291					CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff
) << 8));
2292				compiled_data->tensors.parameters[d + i * parameter_size] = ccv_nnc_tensor_new(0, info, 0);
2293			}
2294		}
2295	}
2296	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));
2297	assert(dest)((void) sizeof ((dest) ? 1 : 0), __extension__ ({ if (dest) ;
 else __assert_fail ("dest", "ccv_cnnp_model.c", 2297, __extension__
 __PRETTY_FUNCTION__); }));
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((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);
2299	for (i = 1; i < parallel_count; i++)
2300	{
2301		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));
2302		if (copy_tensor)
2303			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);
2304	}
2305	// Mark this symbol as init'ed.
2306	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;
2307	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));
2308	init_v[s >> 5] |= (1u << (s & 0x1f));
2309	// If we just allocated this tensor, now it is time to check if we need to mark it as fully allocated.
2310	if (!this_tensor_init)
2311	{
2312		if (ccv_cnnp_model_tensors_any_to_alloc(model, compiled_data))
2313			compiled_data->tensors_init.v = (uint32_t*)((uintptr_t)compiled_data->tensors_init.v | (uintptr_t)1);
2314		else // Remove the flag.
2315			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));
2316	}
2317}
2318 
2319void ccv_cnnp_model_parameter_copy(ccv_cnnp_model_t* const model, const ccv_cnnp_model_io_t parameter, ccv_nnc_tensor_t* const tensor)
2320{
2321	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2322	const int param_sel = parameter->param_sel > 0 ? parameter->param_sel - 1 : parameter->param_sel;
2323	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", 2323, __extension__ __PRETTY_FUNCTION__
); }));
2324	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", 2324, __extension__ __PRETTY_FUNCTION__
); }));
2325	ccv_array_t* const parameter_indices = ccv_array_new(sizeof(int), 0, 0);
2326	ccv_cnnp_model_add_to_parameter_indices(parameter->model, param_sel, parameter_indices);
2327	const int param_ref = parameter->param_ref > 0 ? parameter->param_ref - 1 : parameter->param_ref;
2328	if (param_ref < 0)
2329		{ 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", 2329
, __extension__ __PRETTY_FUNCTION__); })); }
2330	else
2331		{ 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", 2331, __extension__ __PRETTY_FUNCTION__
); })); }
2332	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)));
2333	ccv_array_free(parameter_indices);
2334	const int parameter_size = compiled_data->parameters->rnum;
2335	assert(d >= 0)((void) sizeof ((d >= 0) ? 1 : 0), __extension__ ({ if (d >=
 0) ; else __assert_fail ("d >= 0", "ccv_cnnp_model.c", 2335
, __extension__ __PRETTY_FUNCTION__); }));
2336	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", 2336, __extension__ __PRETTY_FUNCTION__
); }));
2337	// We don't need to consider parallel_count, every parameter on each device is identical.
2338	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));
2339	assert(src)((void) sizeof ((src) ? 1 : 0), __extension__ ({ if (src) ; else
 __assert_fail ("src", "ccv_cnnp_model.c", 2339, __extension__
 __PRETTY_FUNCTION__); }));
2340	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);
2341}
2342 
2343ccv_nnc_tensor_param_t ccv_cnnp_model_parameter_tensor_params(ccv_cnnp_model_t* const model, const ccv_cnnp_model_io_t parameter)
2344{
2345	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2346	const int param_sel = parameter->param_sel > 0 ? parameter->param_sel - 1 : parameter->param_sel;
2347	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", 2347, __extension__ __PRETTY_FUNCTION__
); }));
2348	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", 2348, __extension__ __PRETTY_FUNCTION__
); }));
2349	ccv_array_t* const parameter_indices = ccv_array_new(sizeof(int), 0, 0);
2350	ccv_cnnp_model_add_to_parameter_indices(parameter->model, param_sel, parameter_indices);
2351	const int param_ref = parameter->param_ref > 0 ? parameter->param_ref - 1 : parameter->param_ref;
2352	if (param_ref < 0)
2353		{ 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", 2353
, __extension__ __PRETTY_FUNCTION__); })); }
2354	else
2355		{ 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", 2355, __extension__ __PRETTY_FUNCTION__
); })); }
2356	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)));
2357	ccv_array_free(parameter_indices);
2358	const int parameter_size = compiled_data->parameters->rnum;
2359	assert(d >= 0)((void) sizeof ((d >= 0) ? 1 : 0), __extension__ ({ if (d >=
 0) ; else __assert_fail ("d >= 0", "ccv_cnnp_model.c", 2359
, __extension__ __PRETTY_FUNCTION__); }));
2360	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", 2360, __extension__ __PRETTY_FUNCTION__
); }));
2361	// We don't need to consider parallel_count, every parameter on each device is identical.
2362	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));
2363	assert(tensor)((void) sizeof ((tensor) ? 1 : 0), __extension__ ({ if (tensor
) ; else __assert_fail ("tensor", "ccv_cnnp_model.c", 2363, __extension__
 __PRETTY_FUNCTION__); }));
2364	return tensor->info;
2365}
2366 
2367const char* ccv_cnnp_model_parameter_name(ccv_cnnp_model_t* const model, const ccv_cnnp_model_io_t parameter)
2368{
2369	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2370	const int param_sel = parameter->param_sel > 0 ? parameter->param_sel - 1 : parameter->param_sel;
2371	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", 2371, __extension__ __PRETTY_FUNCTION__
); }));
2372	ccv_array_t* const parameter_indices = ccv_array_new(sizeof(int), 0, 0);
2373	ccv_cnnp_model_add_to_parameter_indices(parameter->model, param_sel, parameter_indices);
2374	const int param_ref = parameter->param_ref > 0 ? parameter->param_ref - 1 : parameter->param_ref;
2375	if (param_ref < 0)
2376		{ 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", 2376
, __extension__ __PRETTY_FUNCTION__); })); }
2377	else
2378		{ 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", 2378, __extension__ __PRETTY_FUNCTION__
); })); }
2379	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)));
2380	ccv_array_free(parameter_indices);
2381	const int parameter_size = compiled_data->parameters->rnum;
2382	assert(d >= 0)((void) sizeof ((d >= 0) ? 1 : 0), __extension__ ({ if (d >=
 0) ; else __assert_fail ("d >= 0", "ccv_cnnp_model.c", 2382
, __extension__ __PRETTY_FUNCTION__); }));
2383	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", 2383, __extension__ __PRETTY_FUNCTION__
); }));
2384	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)));
2385}
2386 
2387int ccv_cnnp_model_parameter_count(ccv_cnnp_model_t* const model)
2388{
2389	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", 2389, __extension__ __PRETTY_FUNCTION__
); }));
2390	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2391	return compiled_data->parameters->rnum;
2392}
2393 
2394uint64_t ccv_cnnp_model_parameters_size(ccv_cnnp_model_t* const model)
2395{
2396	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", 2396, __extension__ __PRETTY_FUNCTION__
); }));
2397	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2398	const int parameter_size = compiled_data->parameters->rnum;
2399	int i;
2400	const ccv_nnc_symbolic_graph_t* const graph = model->graph;
2401	uint64_t size = 0;
2402	const int tensors_init = !!compiled_data->tensors_init.v;
2403	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;
2404	for (i = 0; i < parameter_size; i++)
2405	{
2406		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;
2407		if (tensors_init && compiled_data->tensors.parameters && (init_v[d >> 5] | (1u << (d & 0x1f))) && compiled_data->tensors.parameters[i])
2408		{
2409			ccv_nnc_tensor_param_t params = compiled_data->tensors.parameters[i]->info;
2410			size += ccv_nnc_tensor_data_size(params);
2411			continue;
2412		}
2413		ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, (ccv_nnc_tensor_symbol_t){
2414			.graph = graph,
2415			.d = d
2416		});
2417		size += ccv_nnc_tensor_data_size(params);
2418	}
2419	return size;
2420}
2421 
2422int 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)
2423{
2424	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", 2424, __extension__ __PRETTY_FUNCTION__
); }));
2425	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2426	if (count != compiled_data->parameters->rnum)
2427		return 0;
2428	if (CCV_TENSOR_GET_DEVICE(type)((type) & 0xfff00) == CCV_COMPUTE_DEVICE_ANY)
2429		CCV_TENSOR_SET_DEVICE_ID(type, 0)(type) = (((type) & ~0xfff00) | (((0) & 0xfff) <<
 8));
2430	int i;
2431	// We don't need to consider parallel_count, every parameter on each device is identical.
2432	for (i = 0; i < count; i++)
2433	{
2434		ccv_nnc_tensor_t* tensor = compiled_data->tensors.parameters[i];
2435		if ((uintptr_t)tensor & (uintptr_t)1) // If it is not owned. We don't do anything.
2436		{
2437			tensors[i] = 0;
2438			continue;
2439		}
2440		tensor = CCV_NNC_TENSOR(tensor)((ccv_nnc_tensor_t*)((uintptr_t)(tensor) & ~(uintptr_t)1)
);
2441		if (tensor->info.type == type)
2442			tensors[i] = tensor;
2443		else {
2444			ccv_nnc_tensor_param_t info = tensor->info;
2445			info.type = type;
2446			tensors[i] = ccv_nnc_tensor_new(0, info, 0); // Create this tensor, don't initiate copy yet.
2447		}
2448	}
2449	for (i = 0; i < count; i++)
2450	{
2451		ccv_nnc_tensor_t* tensor = compiled_data->tensors.parameters[i];
2452		if ((uintptr_t)tensor & (uintptr_t)1) // If it is not owned. We don't do anything.
2453			continue;
2454		tensor = CCV_NNC_TENSOR(tensor)((ccv_nnc_tensor_t*)((uintptr_t)(tensor) & ~(uintptr_t)1)
);
2455		// Now initiate transfer. We should do this one on a stream.
2456		if (tensor->info.type != type)
2457			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);
2458	}
2459	// Copy names and remove parameters.
2460	for (i = 0; i < count; i++)
2461	{
2462		ccv_nnc_tensor_t* const tensor = compiled_data->tensors.parameters[i];
2463		if ((uintptr_t)tensor & (uintptr_t)1) // If it is not owned. We don't do anything.
2464		{
2465			names[i] = 0;
2466			continue;
2467		}
2468		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)));
2469		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; });
2470		names[i] = ccmallocmalloc(name_len + 1);
2471		names[i][name_len] = 0;
2472		memcpy(names[i], name, name_len);
2473		if (tensor->info.type == type)
2474			compiled_data->tensors.parameters[i] = 0; // Only move when it is moved.
2475	}
2476	return 1;
2477}
2478 
2479KHASH_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; } }
6
←
Null pointer value stored to field 'vals'→
2480 
2481void 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)
2482{
2483	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", 2483, __extension__ __PRETTY_FUNCTION__
); }));
1
Assuming field 'compiled_data' is non-null→
2
←
Taking true branch→
2484	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2485	int i;
2486	khash_t(ccv_cnnp_parameter_id)kh_ccv_cnnp_parameter_id_t* id_map = 0;
2487	if (count != compiled_data->parameters->rnum)
3
←
Assuming 'count' is not equal to field 'rnum'→
4
←
Taking true branch→
2488	{
2489		id_map = kh_init(ccv_cnnp_parameter_id)kh_init_ccv_cnnp_parameter_id();
5
←
Calling 'kh_init_ccv_cnnp_parameter_id'→
7
←
Returning from 'kh_init_ccv_cnnp_parameter_id'→
2490		// Build the map between name and the index.
2491		for (i = 0; i < count; i++)
8
←
Assuming 'i' is >= 'count'→
9
←
Loop condition is false. Execution continues on line 2499→
2492		{
2493			int ret;
2494			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);
2495			assert(ret != 0)((void) sizeof ((ret != 0) ? 1 : 0), __extension__ ({ if (ret
 != 0) ; else __assert_fail ("ret != 0", "ccv_cnnp_model.c", 2495
, __extension__ __PRETTY_FUNCTION__); }));
2496			kh_val(id_map, k)((id_map)->vals[k]) = i;
2497		}
2498	}
2499	const int parameter_size = compiled_data->parameters->rnum;
2500	int* copy_back = 0;
2501	const int tensors_init = !!compiled_data->tensors_init.v;
10
←
Assuming field 'v' is non-null→
2502	if (!tensors_init10.1
'tensors_init' is 1
)
11
←
Taking false branch→
2503		ccv_cnnp_model_tensors_init_0(model, compiled_data);
2504	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; });
12
←
Assuming '_a' is <= '_b'→
13
←
'?' condition is false→
2505	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));
2506	for (i = 0; i < parameter_size; i++)
14
←
Assuming 'i' is < 'parameter_size'→
15
←
Loop condition is true.  Entering loop body→
2507	{
2508		int j = i;
2509		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)));
2510		if (i15.1
'i' is >= 0
 >= 0 || strncmp(name, names[i], 1023) != 0)
2511		{
2512			// Build the map.
2513			if (id_map == 0)
16
←
Assuming 'id_map' is not equal to null→
17
←
Taking false branch→
2514			{
2515				id_map = kh_init(ccv_cnnp_parameter_id)kh_init_ccv_cnnp_parameter_id();
2516				for (j = 0; j < count; j++)
2517				{
2518					int ret;
2519					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);
2520					assert(ret != 0)((void) sizeof ((ret != 0) ? 1 : 0), __extension__ ({ if (ret
 != 0) ; else __assert_fail ("ret != 0", "ccv_cnnp_model.c", 2520
, __extension__ __PRETTY_FUNCTION__); }));
2521					kh_val(id_map, k)((id_map)->vals[k]) = j;
2522				}
2523			}
2524			const khiter_t k = kh_get(ccv_cnnp_parameter_id, id_map, name)kh_get_ccv_cnnp_parameter_id(id_map, name);
2525			if (k == kh_end(id_map)((id_map)->n_buckets)) // Cannot find the name, skip.
18
←
Assuming 'k' is not equal to field 'n_buckets'→
19
←
Taking false branch→
2526				continue;
2527			j = kh_val(id_map, k)((id_map)->vals[k]);
20
←
Array access (via field 'vals') results in a null pointer dereference
2528		}
2529		if (compiled_data->tensors.parameters[i]) // Cannot be a shared parameter to read.
2530			{ 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", 2530, __extension__ __PRETTY_FUNCTION__
); })); }
2531		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)));
2532		ccv_nnc_tensor_param_t info = ccv_nnc_tensor_symbol_params(parameter.graph, parameter);
2533		if (CCV_TENSOR_GET_DEVICE(info.type)((info.type) & 0xfff00) == CCV_COMPUTE_DEVICE_ANY)
2534			CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff
) << 8));
2535		const int d = parameter.d;
2536		if (info.type == tensors[j]->info.type && invalidates) // Can move.
2537		{
2538			// Deallocate it if needed.
2539			if (!((uintptr_t)compiled_data->tensors.parameters[i] & (uintptr_t)1))
2540				if (compiled_data->tensors.parameters[i])
2541					ccv_nnc_tensor_free(compiled_data->tensors.parameters[i]);
2542			compiled_data->tensors.parameters[i] = tensors[j];
2543			tensors[j] = 0;
2544		} else {
2545			if (!compiled_data->tensors.parameters[i])
2546			{ // Not allocated, to allocate first.
2547				// Create new one, make sure we create this by having the right parameters.
2548				const int type = info.type;
2549				info = tensors[j]->info;
2550				info.type = type; // Revert back the type.
2551				compiled_data->tensors.parameters[i] = ccv_nnc_tensor_new(0, info, 0);
2552			}
2553			if (!copy_back)
2554				copy_back = (int*)cccalloccalloc(parameter_size, sizeof(int));
2555			copy_back[i] = j + 1;
2556		}
2557		init_v[d >> 5] |= (1u << (d & 0x1f));
2558		// Create this tensor for other data parallel allocations.
2559		info = compiled_data->tensors.parameters[i]->info; // In case we loaded a different info.
2560		const int device_id = CCV_TENSOR_GET_DEVICE_ID(info.type)(((info.type) & 0xfff00) >> 8);
2561		for (j = 1; j < parallel_count; j++)
2562			if (!compiled_data->tensors.parameters[i + j * parameter_size])
2563			{
2564				if (j != device_id)
2565					CCV_TENSOR_SET_DEVICE_ID(info.type, j)(info.type) = (((info.type) & ~0xfff00) | (((j) & 0xfff
) << 8));
2566				else
2567					CCV_TENSOR_SET_DEVICE_ID(info.type, 0)(info.type) = (((info.type) & ~0xfff00) | (((0) & 0xfff
) << 8));
2568				compiled_data->tensors.parameters[i + j * parameter_size] = ccv_nnc_tensor_new(0, info, 0);
2569			}
2570			// No need to copy over, this is done in ccv_cnnp_model.c's copy_tensors method.
2571	}
2572	if (id_map)
2573		kh_destroy(ccv_cnnp_parameter_id, id_map)kh_destroy_ccv_cnnp_parameter_id(id_map);
2574	// Now do the transfer.
2575	if (copy_back)
2576	{
2577		for (i = 0; i < parameter_size; i++)
2578		{
2579			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));
2580			if (copy_back[i] == 0)
2581				continue;
2582			const int j = copy_back[i] - 1;
2583			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);
2584		}
2585		ccfreefree(copy_back);
2586	}
2587}
2588 
2589ccv_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)
2590{
2591	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2592	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 2592, __extension__ __PRETTY_FUNCTION__); }));
2593	const int parameter_size = compiled_data->parameters->rnum;
2594	int i;
2595	for (i = 0; i < parameter_size; i++)
2596	{
2597		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)));
2598		if (first(model, name, context))
2599			return ccv_cnnp_model_parameters(model, -1, i);
2600	}
2601	return 0;
2602}
2603 
2604ccv_array_t* ccv_cnnp_model_parameters_filter(ccv_cnnp_model_t* const model, ccv_cnnp_model_parameters_filter_f filter, void* const context)
2605{
2606	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2607	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 2607, __extension__ __PRETTY_FUNCTION__); }));
2608	ccv_array_t* const parameters = ccv_array_new(sizeof(ccv_cnnp_model_io_t), 0, 0);
2609	const int parameter_size = compiled_data->parameters->rnum;
2610	int i;
2611	for (i = 0; i < parameter_size; i++)
2612	{
2613		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)));
2614		if (filter(model, name, context))
2615		{
2616			ccv_cnnp_model_io_t parameter = ccv_cnnp_model_parameters(model, -1, i);
2617			ccv_array_push(parameters, &parameter);
2618		}
2619	}
2620	return parameters;
2621 
2622}
2623 
2624CCV_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)
2625{
2626	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
2627	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 2627, __extension__ __PRETTY_FUNCTION__); }));
2628	const int tensors_init = !!compiled_data->tensors_init.v;
2629	if (!tensors_init) // If nothing initialized, we return parameter 0.
2630		return ccv_cnnp_model_parameters(model, -1, 0);
2631	const int parameter_size = compiled_data->parameters->rnum;
2632	int i;
2633	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));
2634	for (i = 0; i < parameter_size; i++)
2635	{
2636		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;
2637		if (!(init_v[d >> 5] & (1u << (d & 0x1f))))
2638			return ccv_cnnp_model_parameters(model, -1, i);
2639	}
2640	return 0;
2641}
2642 
2643static 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)
2644{
2645	const int to_param_sel = parameters->param_sel > 0 ? parameters->param_sel - 1 : parameters->param_sel;
2646	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", 2646, __extension__
 __PRETTY_FUNCTION__); }));
2647	ccv_array_t* const to_parameter_indices = ccv_array_new(sizeof(int), 0, 0);
2648	ccv_cnnp_model_add_to_parameter_indices(parameters->model, to_param_sel, to_parameter_indices);
2649	*param_ref = parameters->param_ref > 0 ? parameters->param_ref - 1 : parameters->param_ref;
2650	return to_parameter_indices;
2651}
2652 
2653static 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)
2654{
2655	// If the model is not compiled yet. Compile them now.
2656	if (!model->graph)
2657	{
2658		model->graph = ccv_nnc_symbolic_graph_new();
2659		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", 2659, __extension__ __PRETTY_FUNCTION__
); }));
2660		const int input_size = from_model->input_size;
2661		ccv_nnc_tensor_param_t input_params[input_size];
2662		int i;
2663		for (i = 0; i < input_size; i++)
2664			input_params[i] = ccv_nnc_tensor_symbol_params(from_model->graph, from_model->inputs[i]);
2665		_ccv_cnnp_model_compile(model, input_params, input_size, from_model->compiled_data->loss);
2666		model->parallel_count = from_model->parallel_count;
2667		model->memory_compression = from_model->memory_compression;
2668		model->memory_reduction = from_model->memory_reduction;
2669		model->gradient_checkpointing = from_model->gradient_checkpointing;
2670		model->compiled_data->stream_type = from_model->compiled_data->stream_type;
2671		model->compiled_data->minimize.minimizer = from_model->compiled_data->minimize.minimizer;
2672		model->compiled_data->minimize.max_saved_aux_size = from_model->compiled_data->minimize.max_saved_aux_size;
2673	}
2674	ccv_cnnp_compiled_data_t* const to_compiled_data = model->compiled_data;
2675	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", 2675, __extension__ __PRETTY_FUNCTION__
); }));
2676	const int to_tensors_init = !!to_compiled_data->tensors_init.v;
2677	if (!to_tensors_init)
2678	{
2679		if (only_init_0)
2680			ccv_cnnp_model_tensors_init_0(model, to_compiled_data);
2681		else
2682			_ccv_cnnp_model_tensors_init(model, to_compiled_data);
2683	} else if (!only_init_0 && (uintptr_t)to_compiled_data->tensors_init.v & (uintptr_t)1)
2684		// Check if it is not fully allocated, if it is not, init_1.
2685			ccv_cnnp_model_tensors_init_1(model, to_compiled_data);
2686	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", 2686, __extension__ __PRETTY_FUNCTION__
); }));
2687	*parameter_indices = _ccv_cnnp_model_parameter_indices(model, parameters, param_ref);
2688	*from_parameter_indices = _ccv_cnnp_model_parameter_indices(from_model, from_parameters, from_param_ref);
2689	if (*from_param_ref < 0 && *param_ref >= 0)
2690		{ 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", 2690, __extension__ __PRETTY_FUNCTION__
); })); }
2691	else if (*from_param_ref >= 0)
2692		{ 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", 2692, __extension__ __PRETTY_FUNCTION__
); })); }
2693	if (*param_ref < 0 && *from_param_ref >= 0)
2694		{ 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"
, 2694, __extension__ __PRETTY_FUNCTION__); })); }
2695	else if (*param_ref >= 0)
2696		{ 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", 2696, __extension__ __PRETTY_FUNCTION__
); })); }
2697}
2698 
2699void 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)
2700{
2701	ccv_array_t* to_parameter_indices;
2702	int to_param_ref;
2703	ccv_array_t* from_parameter_indices;
2704	int from_param_ref;
2705	_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);
2706	// Should be exactly the same tensor.
2707	if (to_param_ref < 0 && from_param_ref < 0)
2708		{ 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", 2708, __extension__ __PRETTY_FUNCTION__
); })); }
2709	// To models.
2710	ccv_cnnp_compiled_data_t* const to_compiled_data = model->compiled_data;
2711	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", 2711, __extension__ __PRETTY_FUNCTION__
); }));
2712	// From models.
2713	const ccv_cnnp_compiled_data_t* const from_compiled_data = from_model->compiled_data;
2714	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; });
2715	const int to_parameter_size = to_compiled_data->parameters->rnum;
2716	const int rnum = (to_param_ref < 0 && from_param_ref < 0) ? from_parameter_indices->rnum : 1;
2717	int i, j;
2718	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));
2719  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));
2720	for (i = 0; i < rnum; i++)
2721	{
2722		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)));
2723		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"
, 2723, __extension__ __PRETTY_FUNCTION__); }));
2724		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", 2724, __extension__ __PRETTY_FUNCTION__
); }));
2725		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;
2726		// If the original is not init'ed. We cannot copy from.
2727		if (!(from_init_v[s >> 5] & (1u << (s & 0x1f))))
2728			continue;
2729		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)));
2730		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"
, 2730, __extension__ __PRETTY_FUNCTION__); }));
2731		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", 2731, __extension__ __PRETTY_FUNCTION__
); }));
2732		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));
2733		assert(src)((void) sizeof ((src) ? 1 : 0), __extension__ ({ if (src) ; else
 __assert_fail ("src", "ccv_cnnp_model.c", 2733, __extension__
 __PRETTY_FUNCTION__); }));
2734		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));
2735		assert(dest)((void) sizeof ((dest) ? 1 : 0), __extension__ ({ if (dest) ;
 else __assert_fail ("dest", "ccv_cnnp_model.c", 2735, __extension__
 __PRETTY_FUNCTION__); }));
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(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);
2737		for (j = 1; j < parallel_count; j++)
2738		{
2739			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));
2740			if (copy_tensor)
2741				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);
2742		}
2743		// Mark this symbol as init'ed.
2744		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;
2745		to_init_v[d >> 5] |= (1u << (d & 0x1f));
2746	}
2747	ccv_array_free(to_parameter_indices);
2748	ccv_array_free(from_parameter_indices);
2749}
2750 
2751void 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)
2752{
2753	ccv_array_t* to_parameter_indices;
2754	int to_param_ref;
2755	ccv_array_t* from_parameter_indices;
2756	int from_param_ref;
2757	_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);
2758	// Should be exactly the same tensor.
2759	if (renamer == 0 && to_param_ref < 0 && from_param_ref < 0)
2760		{ 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", 2760, __extension__ __PRETTY_FUNCTION__
); })); }
2761	// To models.
2762	ccv_cnnp_compiled_data_t* const to_compiled_data = model->compiled_data;
2763	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", 2763, __extension__ __PRETTY_FUNCTION__
); }));
2764	// From models.
2765	const ccv_cnnp_compiled_data_t* const from_compiled_data = from_model->compiled_data;
2766	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; });
2767	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", 2767, __extension__ __PRETTY_FUNCTION__
); })); // Should have the same parallel count can share parameters.
2768	const int from_parameter_size = from_compiled_data->parameters->rnum;
2769	const int to_parameter_size = to_compiled_data->parameters->rnum;
2770	const int rnum = (to_param_ref < 0 && from_param_ref < 0) ? to_parameter_indices->rnum : 1;
2771	int i, j;
2772	khash_t(ccv_cnnp_parameter_id)kh_ccv_cnnp_parameter_id_t* id_map = 0;
2773	char* updated_name = 0;
2774	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));
2775  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));
2776	for (i = 0; i < rnum; i++)
2777	{
2778		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;
2779		// Need to figure out how to use the renamer here.
2780		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)));
2781		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"
, 2781, __extension__ __PRETTY_FUNCTION__); }));
2782		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", 2782, __extension__
 __PRETTY_FUNCTION__); }));
2783		if (renamer)
2784		{
2785			const char* const src_name = (src_d < 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;
2786			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)));
2787			if (!updated_name)
2788				updated_name = (char*)ccmallocmalloc(1024);
2789			const size_t src_name_len = src_name == 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; });
2790			if (src_name_len > 0)
2791				memcpy(updated_name, src_name, src_name_len);
2792			updated_name[src_name_len] = 0;
2793			if (renamer(context, dest_name, updated_name, 1024) != 0)
2794				continue; // Skip this.
2795			if (src_name != 0 && memcmp(updated_name, src_name, src_name_len) == 0 && strnlen(updated_name, 1023) == src_name_len)
2796			{
2797				// Nothing changed.
2798			} else {
2799				if (!id_map)
2800				{
2801					id_map = kh_init(ccv_cnnp_parameter_id)kh_init_ccv_cnnp_parameter_id();
2802					for (j = 0; j < from_parameter_size; j++)
2803					{
2804						int ret;
2805						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);
2806						assert(ret != 0)((void) sizeof ((ret != 0) ? 1 : 0), __extension__ ({ if (ret
 != 0) ; else __assert_fail ("ret != 0", "ccv_cnnp_model.c", 2806
, __extension__ __PRETTY_FUNCTION__); }));
2807						kh_val(id_map, k)((id_map)->vals[k]) = j;
2808					}
2809				}
2810				const khiter_t k = kh_get(ccv_cnnp_parameter_id, id_map, updated_name)kh_get_ccv_cnnp_parameter_id(id_map, updated_name);
2811				if (k == kh_end(id_map)((id_map)->n_buckets)) // Cannot find the name, skip.
2812					continue;
2813				src_d = kh_val(id_map, k)((id_map)->vals[k]);
2814				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"
, 2814, __extension__ __PRETTY_FUNCTION__); }));
2815				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", 2815, __extension__
 __PRETTY_FUNCTION__); }));
2816			}
2817		}
2818		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"
, 2818, __extension__ __PRETTY_FUNCTION__); }));
2819		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", 2819, __extension__
 __PRETTY_FUNCTION__); }));
2820		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;
2821		// If the original is not init'ed. We cannot share from.
2822		if (!(from_init_v[s >> 5] & (1u << (s & 0x1f))))
2823			continue;
2824		for (j = 0; j < parallel_count; j++)
2825		{
2826			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));
2827			assert(src)((void) sizeof ((src) ? 1 : 0), __extension__ ({ if (src) ; else
 __assert_fail ("src", "ccv_cnnp_model.c", 2827, __extension__
 __PRETTY_FUNCTION__); }));
2828			ccv_nnc_tensor_t* const dest = to_compiled_data->tensors.parameters[dest_d + j * to_parameter_size];
2829			if (dest && !((uintptr_t)dest & (uintptr_t)1))
2830				ccv_nnc_tensor_free(dest);
2831			to_compiled_data->tensors.parameters[dest_d + j * to_parameter_size] = (ccv_nnc_tensor_t*)((uintptr_t)src | (uintptr_t)1);
2832		}
2833		// Mark this symbol as init'ed.
2834		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;
2835		to_init_v[d >> 5] |= (1u << (d & 0x1f));
2836	}
2837	ccv_array_free(to_parameter_indices);
2838	ccv_array_free(from_parameter_indices);
2839	if (id_map)
2840		kh_destroy(ccv_cnnp_parameter_id, id_map)kh_destroy_ccv_cnnp_parameter_id(id_map);
2841	if (updated_name)
2842		ccfreefree(updated_name);
2843	// Mark it as incomplete so we will call init_1.
2844	if (ccv_cnnp_model_tensors_any_to_alloc(model, to_compiled_data))
2845		to_compiled_data->tensors_init.v = (uint32_t*)((uintptr_t)to_compiled_data->tensors_init.v | (uintptr_t)1);
2846	else // Remove the flag.
2847		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));
2848}
2849 
2850ccv_nnc_stream_context_t* ccv_cnnp_compiled_data_get_stream(ccv_cnnp_compiled_data_t* const compiled_data, const int type)
2851{
2852	if (!compiled_data->stream_map)
2853		compiled_data->stream_map = kh_init(stream_map)kh_init_stream_map();
2854	int ret = 0;
2855	khiter_t k = kh_put(stream_map, compiled_data->stream_map, type, &ret)kh_put_stream_map(compiled_data->stream_map, type, &ret
);
2856	assert(ret >= 0)((void) sizeof ((ret >= 0) ? 1 : 0), __extension__ ({ if (
ret >= 0) ; else __assert_fail ("ret >= 0", "ccv_cnnp_model.c"
, 2856, __extension__ __PRETTY_FUNCTION__); }));
2857	ccv_nnc_stream_context_t* stream = kh_val(compiled_data->stream_map, k)((compiled_data->stream_map)->vals[k]);
2858	// If ret == 0, the key already exist, we can return directly, otherwise, create and return.
2859	if (ret != 0)
2860	{
2861		stream = ccv_nnc_stream_context_new(type);
2862		kh_val(compiled_data->stream_map, k)((compiled_data->stream_map)->vals[k]) = stream;
2863	}
2864	return stream;
2865}
2866 
2867void 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)
2868{
2869	ccv_array_t* to_parameter_indices;
2870	int to_param_ref;
2871	ccv_array_t* from_parameter_indices;
2872	int from_param_ref;
2873	_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);
2874	// Should be exactly the same tensor.
2875	if (to_param_ref < 0 && from_param_ref < 0)
2876		{ 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", 2876, __extension__ __PRETTY_FUNCTION__
); })); }
2877	// To models.
2878	ccv_cnnp_compiled_data_t* const to_compiled_data = model->compiled_data;
2879	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", 2879, __extension__ __PRETTY_FUNCTION__
); }));
2880	// From models.
2881	const ccv_cnnp_compiled_data_t* const from_compiled_data = from_model->compiled_data;
2882	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; });
2883	const int to_parameter_size = to_compiled_data->parameters->rnum;
2884	const int rnum = (to_param_ref < 0 && from_param_ref < 0) ? from_parameter_indices->rnum : 1;
2885	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", 2885, __extension__ __PRETTY_FUNCTION__
); }));
2886	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", 2886, __extension__ __PRETTY_FUNCTION__
); }));
2887	int i, j;
2888	ccv_nnc_tensor_t* inputs[aux_in_size + 2];
2889	ccv_nnc_tensor_t* outputs[aux_out_size + 1];
2890	for (i = 0; i < aux_in_size; i++)
2891		inputs[i + 2] = aux_ins[i];
2892	for (i = 0; i < aux_out_size; i++)
2893		outputs[i + 1] = aux_outs[i];
2894	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));
2895	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));
2896	for (i = 0; i < rnum; i++)
2897	{
2898		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)));
2899		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"
, 2899, __extension__ __PRETTY_FUNCTION__); }));
2900		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", 2900, __extension__ __PRETTY_FUNCTION__
); }));
2901		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;
2902		// If the original is not init'ed. We cannot copy from.
2903		if (!(from_init_v[s >> 5] & (1u << (s & 0x1f))))
2904			continue;
2905		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)));
2906		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"
, 2906, __extension__ __PRETTY_FUNCTION__); }));
2907		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", 2907, __extension__ __PRETTY_FUNCTION__
); }));
2908		if (parallel_count > 1)
2909		{
2910			ccv_nnc_stream_context_t* streams[parallel_count];
2911			ccv_nnc_stream_signal_t* signal;
2912			if (stream_context)
2913				signal = ccv_nnc_stream_context_emit_signal_new(stream_context);
2914			for (j = 0; j < parallel_count; j++)
2915			{
2916				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));
2917				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));
2918				if (!dest || !src)
2919				{
2920					streams[j] = 0;
2921					continue;
2922				}
2923				// At the moment, can only handle them on the same device.
2924				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", 2924, __extension__ __PRETTY_FUNCTION__
); }));
2925				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", 2925, __extension__ __PRETTY_FUNCTION__
); }));
2926				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;
2927				const int device_id = CCV_TENSOR_GET_DEVICE_ID(src->info.type)(((src->info.type) & 0xfff00) >> 8);
2928				int type = stream_type;
2929				CCV_STREAM_SET_DEVICE_ID(type, device_id)(type) = (((type) & ~0xfff00) | (((device_id) & 0xfff
) << 8));
2930				ccv_nnc_stream_context_t* const stream_0 = ccv_cnnp_compiled_data_get_stream(to_compiled_data, type);
2931				// Wait signal to finish.
2932				if (stream_context)
2933					ccv_nnc_stream_context_wait_signal(stream_0, signal);
2934				inputs[0] = outputs[0] = dest;
2935				inputs[1] = src;
2936				ccv_nnc_cmd_exec(cmd, hint, flags, inputs, aux_in_size + 2, outputs, aux_out_size + 1, stream_0);
2937				if (stream_context)
2938				{
2939					ccv_nnc_stream_signal_t* const signal = ccv_nnc_stream_context_emit_signal_new(stream_0);
2940					ccv_nnc_stream_context_wait_signal(stream_context, signal);
2941				}
2942				streams[j] = stream_0;
2943			}
2944			// If this should be blocking, blocking it.
2945			if (!stream_context)
2946				for (j = 0; j < parallel_count; j++)
2947					if (streams[j])
2948						ccv_nnc_stream_context_wait(streams[j]);
2949		} else {
2950			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));
2951			assert(src)((void) sizeof ((src) ? 1 : 0), __extension__ ({ if (src) ; else
 __assert_fail ("src", "ccv_cnnp_model.c", 2951, __extension__
 __PRETTY_FUNCTION__); }));
2952			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));
2953			assert(dest)((void) sizeof ((dest) ? 1 : 0), __extension__ ({ if (dest) ;
 else __assert_fail ("dest", "ccv_cnnp_model.c", 2953, __extension__
 __PRETTY_FUNCTION__); }));
2954			inputs[0] = outputs[0] = dest;
2955			inputs[1] = src;
2956			ccv_nnc_cmd_exec(cmd, hint, flags, inputs, aux_in_size + 2, outputs, aux_out_size + 1, stream_context);
2957		}
2958		// Mark this symbol as init'ed.
2959		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;
2960		to_init_v[d >> 5] |= (1u << (d & 0x1f));
2961	}
2962	ccv_array_free(to_parameter_indices);
2963	ccv_array_free(from_parameter_indices);
2964}
2965 
2966void 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)
2967{
2968	int to_param_ref;
2969	ccv_array_t* const to_parameter_indices = _ccv_cnnp_model_parameter_indices(model, parameters, &to_param_ref);
2970	// To models.
2971	ccv_cnnp_compiled_data_t* const to_compiled_data = model->compiled_data;
2972	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", 2972, __extension__ __PRETTY_FUNCTION__
); }));
2973	// Tensor has to be inited already.
2974	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", 2974, __extension__ __PRETTY_FUNCTION__
); }));
2975	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", 2975, __extension__ __PRETTY_FUNCTION__
); }));
2976	// From models.
2977	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; });
2978	const int to_parameter_size = to_compiled_data->parameters->rnum;
2979	const int rnum = (to_param_ref < 0) ? to_parameter_indices->rnum : 1;
2980	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", 2980, __extension__ __PRETTY_FUNCTION__
); }));
2981	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", 2981, __extension__ __PRETTY_FUNCTION__
); }));
2982	int i, j;
2983	ccv_nnc_tensor_t* inputs[aux_in_size + 1];
2984	ccv_nnc_tensor_t* outputs[aux_out_size + 1];
2985	for (i = 0; i < aux_in_size; i++)
2986		inputs[i + 1] = aux_ins[i];
2987	for (i = 0; i < aux_out_size; i++)
2988		outputs[i + 1] = aux_outs[i];
2989	for (i = 0; i < rnum; i++)
2990	{
2991		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)));
2992		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"
, 2992, __extension__ __PRETTY_FUNCTION__); }));
2993		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", 2993, __extension__ __PRETTY_FUNCTION__
); }));
2994		if (parallel_count > 1)
2995		{
2996			ccv_nnc_stream_context_t* streams[parallel_count];
2997			ccv_nnc_stream_signal_t* signal;
2998			if (stream_context)
2999				signal = ccv_nnc_stream_context_emit_signal_new(stream_context);
3000			for (j = 0; j < parallel_count; j++)
3001			{
3002				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));
3003				if (!dest)
3004				{
3005					streams[j] = 0;
3006					continue;
3007				}
3008				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;
3009				const int device_id = CCV_TENSOR_GET_DEVICE_ID(dest->info.type)(((dest->info.type) & 0xfff00) >> 8);
3010				int type = stream_type;
3011				CCV_STREAM_SET_DEVICE_ID(type, device_id)(type) = (((type) & ~0xfff00) | (((device_id) & 0xfff
) << 8));
3012				ccv_nnc_stream_context_t* const stream_0 = ccv_cnnp_compiled_data_get_stream(to_compiled_data, type);
3013				// Wait signal to finish.
3014				if (stream_context)
3015					ccv_nnc_stream_context_wait_signal(stream_0, signal);
3016				inputs[0] = outputs[0] = dest;
3017				ccv_nnc_cmd_exec(cmd, hint, flags, inputs, aux_in_size + 1, outputs, aux_out_size + 1, stream_0);
3018				if (stream_context)
3019				{
3020					ccv_nnc_stream_signal_t* const signal = ccv_nnc_stream_context_emit_signal_new(stream_0);
3021					ccv_nnc_stream_context_wait_signal(stream_context, signal);
3022				}
3023				streams[j] = stream_0;
3024			}
3025			// If this should be blocking, blocking it.
3026			if (!stream_context)
3027				for (j = 0; j < parallel_count; j++)
3028					if (streams[j])
3029						ccv_nnc_stream_context_wait(streams[j]);
3030		} else {
3031			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));
3032			assert(dest)((void) sizeof ((dest) ? 1 : 0), __extension__ ({ if (dest) ;
 else __assert_fail ("dest", "ccv_cnnp_model.c", 3032, __extension__
 __PRETTY_FUNCTION__); }));
3033			inputs[0] = outputs[0] = dest;
3034			ccv_nnc_cmd_exec(cmd, hint, flags, inputs, aux_in_size + 1, outputs, aux_out_size + 1, stream_context);
3035		}
3036		// No need to mark this symbol as init'ed, it is already.
3037	}
3038	ccv_array_free(to_parameter_indices);
3039}
3040 
3041void 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)
3042{
3043	int to_param_ref;
3044	ccv_array_t* const to_parameter_indices = _ccv_cnnp_model_parameter_indices(model, parameters, &to_param_ref);
3045	// To models.
3046	ccv_cnnp_compiled_data_t* const to_compiled_data = model->compiled_data;
3047	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", 3047, __extension__ __PRETTY_FUNCTION__
); }));
3048	// Tensor has to be inited already.
3049	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", 3049, __extension__ __PRETTY_FUNCTION__
); }));
3050	ccv_nnc_tensor_t** tensor_gradients;
3051	if (to_compiled_data->backward.count > 1)
3052		tensor_gradients = to_compiled_data->tensors.accum_gradients;
3053	else
3054		tensor_gradients = to_compiled_data->tensors.gradients;
3055	assert(tensor_gradients)((void) sizeof ((tensor_gradients) ? 1 : 0), __extension__ ({
 if (tensor_gradients) ; else __assert_fail ("tensor_gradients"
, "ccv_cnnp_model.c", 3055, __extension__ __PRETTY_FUNCTION__
); }));
3056	// From models.
3057	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; });
3058	const int to_parameter_size = to_compiled_data->parameters->rnum;
3059	const int rnum = (to_param_ref < 0) ? to_parameter_indices->rnum : 1;
3060	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", 3060, __extension__ __PRETTY_FUNCTION__
); }));
3061	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", 3061, __extension__ __PRETTY_FUNCTION__
); }));
3062	int i, j;
3063	ccv_nnc_tensor_t* inputs[aux_in_size + 1];
3064	ccv_nnc_tensor_t* outputs[aux_out_size + 1];
3065	for (i = 0; i < aux_in_size; i++)
3066		inputs[i + 1] = aux_ins[i];
3067	for (i = 0; i < aux_out_size; i++)
3068		outputs[i + 1] = aux_outs[i];
3069	for (i = 0; i < rnum; i++)
3070	{
3071		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)));
3072		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"
, 3072, __extension__ __PRETTY_FUNCTION__); }));
3073		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", 3073, __extension__ __PRETTY_FUNCTION__
); }));
3074		if (parallel_count > 1)
3075		{
3076			ccv_nnc_stream_context_t* streams[parallel_count];
3077			ccv_nnc_stream_signal_t* signal;
3078			if (stream_context)
3079				signal = ccv_nnc_stream_context_emit_signal_new(stream_context);
3080			for (j = 0; j < parallel_count; j++)
3081			{
3082				ccv_nnc_tensor_t* const dest = tensor_gradients[dest_d + j * to_parameter_size];
3083				if (!dest)
3084				{
3085					streams[j] = 0;
3086					continue;
3087				}
3088				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;
3089				const int device_id = CCV_TENSOR_GET_DEVICE_ID(dest->info.type)(((dest->info.type) & 0xfff00) >> 8);
3090				int type = stream_type;
3091				CCV_STREAM_SET_DEVICE_ID(type, device_id)(type) = (((type) & ~0xfff00) | (((device_id) & 0xfff
) << 8));
3092				ccv_nnc_stream_context_t* const stream_0 = ccv_cnnp_compiled_data_get_stream(to_compiled_data, type);
3093				// Wait signal to finish.
3094				if (stream_context)
3095					ccv_nnc_stream_context_wait_signal(stream_0, signal);
3096				inputs[0] = outputs[0] = dest;
3097				ccv_nnc_cmd_exec(cmd, hint, flags, inputs, aux_in_size + 1, outputs, aux_out_size + 1, stream_0);
3098				if (stream_context)
3099				{
3100					ccv_nnc_stream_signal_t* const signal = ccv_nnc_stream_context_emit_signal_new(stream_0);
3101					ccv_nnc_stream_context_wait_signal(stream_context, signal);
3102				}
3103				streams[j] = stream_0;
3104			}
3105			// If this should be blocking, blocking it.
3106			if (!stream_context)
3107				for (j = 0; j < parallel_count; j++)
3108					if (streams[j])
3109						ccv_nnc_stream_context_wait(streams[j]);
3110		} else {
3111			ccv_nnc_tensor_t* const dest = tensor_gradients[dest_d];
3112			if (!dest)
3113				continue;
3114			assert(dest)((void) sizeof ((dest) ? 1 : 0), __extension__ ({ if (dest) ;
 else __assert_fail ("dest", "ccv_cnnp_model.c", 3114, __extension__
 __PRETTY_FUNCTION__); }));
3115			inputs[0] = outputs[0] = dest;
3116			ccv_nnc_cmd_exec(cmd, hint, flags, inputs, aux_in_size + 1, outputs, aux_out_size + 1, stream_context);
3117		}
3118		// No need to mark this symbol as init'ed, it is already.
3119	}
3120	ccv_array_free(to_parameter_indices);
3121}
3122 
3123void 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)
3124{
3125	// Only CUDA backend has this feature.
3126#ifdef HAVE_CUDA1
3127	int to_param_ref;
3128	ccv_array_t* const to_parameter_indices = _ccv_cnnp_model_parameter_indices(model, parameters, &to_param_ref);
3129	// To models.
3130	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
3131	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 3131, __extension__ __PRETTY_FUNCTION__); }));
3132	// Tensor has to be inited already.
3133	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"
, 3133, __extension__ __PRETTY_FUNCTION__); }));
3134	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", 3134, __extension__ __PRETTY_FUNCTION__
); }));
3135	// From models.
3136	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; });
3137	const int rnum = (to_param_ref < 0) ? to_parameter_indices->rnum : 1;
3138	int i;
3139	for (i = 0; i < rnum; i++)
3140	{
3141		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)));
3142		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"
, 3142, __extension__ __PRETTY_FUNCTION__); }));
3143		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", 3143, __extension__ __PRETTY_FUNCTION__
); }));
3144		if (parallel_count > 1)
3145		{
3146			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", 3146, __extension__ __PRETTY_FUNCTION__
); }));
3147		} else {
3148			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));
3149			assert(src)((void) sizeof ((src) ? 1 : 0), __extension__ ({ if (src) ; else
 __assert_fail ("src", "ccv_cnnp_model.c", 3149, __extension__
 __PRETTY_FUNCTION__); }));
3150			ccv_nnc_tensor_param_t params = src->info;
3151			if (CCV_TENSOR_GET_MEMORY(params.type)((params.type) & 0x3) != CCV_TENSOR_GPU_MEMORY)
3152				continue;
3153			const size_t size = ccv_nnc_tensor_data_size(params);
3154			if (size <= 0)
3155				continue;
3156			const int should_free = !((uintptr_t)compiled_data->tensors.parameters[dest_d] & (uintptr_t)1);
3157			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);
3158			ccv_nnc_tensor_t* const tensor = (ccv_nnc_tensor_t*)ccmallocmalloc(sizeof(ccv_nnc_tensor_t));
3159			tensor->dataof = 0;
3160			tensor->alias_ref = 0;
3161			tensor->sig = 0;
3162			tensor->refcount = 1;
3163			tensor->info = params;
3164			if (tfb)
3165			{
3166				tensor->type = CCV_NO_DATA_ALLOC | CCV_MATRIX_DENSE | CCV_GET_DATA_TYPE(params.datatype)((params.datatype) & 0xFF000) | params.dim[2];
3167				// This corresponding to mat->step
3168				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);
3169			} else // This won't be recognized by ccv_dense_matrix_t
3170				tensor->type = CCV_NO_DATA_ALLOC | CCV_MATRIX_DENSE | CCV_GET_DATA_TYPE(params.datatype)((params.datatype) & 0xFF000);
3171			// Remove this flag so it can be deallocated as usual.
3172			tensor->type &= ~CCV_NO_DATA_ALLOC;
3173			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", 3173, __extension__ __PRETTY_FUNCTION__
); }));
3174			void* ptr = cumallocmanaged(CCV_TENSOR_GET_DEVICE_ID(params.type)(((params.type) & 0xfff00) >> 8), size);
3175			if (ptr) // If allocated successfully. Otherwise we go through the fallback path.
3176			{
3177				tensor->data.u8 = (uint8_t*)ptr;
3178				tensor->type |= CCV_MAPPED_MEM; // This denotes the tensor is mapped to CPU, and would prefer a explicit prefetch call.
3179			} else {
3180				// Allocation failed.
3181				ccfreefree(tensor);
3182				continue;
3183			}
3184			// TODO: Cannot run this on the stream context yet, due to allocation and deallocations.
3185			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);
3186			cumemadvisereadmostly(CCV_TENSOR_GET_DEVICE_ID(params.type)(((params.type) & 0xfff00) >> 8), tensor->data.u8, size);
3187			compiled_data->tensors.parameters[dest_d] = tensor;
3188			// Can free out the old one.
3189			if (should_free)
3190				ccv_nnc_tensor_free(src);
3191		}
3192		// No need to mark this symbol as init'ed, it is already.
3193	}
3194	ccv_array_free(to_parameter_indices);
3195#endif
3196}
3197 
3198ccv_nnc_cmd_t ccv_cnnp_model_minimizer(ccv_cnnp_model_t* const model)
3199{
3200	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
3201	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 3201, __extension__ __PRETTY_FUNCTION__); }));
3202	return compiled_data->minimize.minimizer;
3203}
3204 
3205void 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)
3206{
3207	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
3208	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 3208, __extension__ __PRETTY_FUNCTION__); }));
3209	const int parameter_size = compiled_data->parameters->rnum;
3210	if (parameter_size == 0)
3211		return;
3212	if (reset)
3213		{ 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", 3213, __extension__ __PRETTY_FUNCTION__
); })); }
3214	const int old_max_saved_aux_size = compiled_data->minimize.max_saved_aux_size;
3215	const int saved_aux_size = ccv_nnc_minimizer_saved_aux_size(minimizer);
3216	if (saved_aux_size > compiled_data->minimize.max_saved_aux_size)
3217		compiled_data->minimize.max_saved_aux_size = saved_aux_size;
3218	const int max_saved_aux_size = compiled_data->minimize.max_saved_aux_size;
3219	// We update all parameters, at this point, we have one minimizer.
3220	if (set_parameters == 0 || set_parameter_size == 0)
3221		compiled_data->minimize.minimizer = minimizer;
3222	int i;
3223	if (set_parameters && set_parameter_size)
3224	{
3225		// I need to save what's the minimizer along with this.
3226		if (!compiled_data->minimize.parameters)
3227			compiled_data->minimize.parameters = ccv_array_new(sizeof(ccv_cnnp_set_minimizer_for_parameter_t*), 1, 0);
3228		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));
3229		set_minimizer_for_parameter->minimizer = minimizer;
3230		set_minimizer_for_parameter->parameter_size = set_parameter_size;
3231		memcpy(set_minimizer_for_parameter->parameters, set_parameters, sizeof(ccv_cnnp_model_io_t) * set_parameter_size);
3232		ccv_array_push(compiled_data->minimize.parameters, &set_minimizer_for_parameter);
3233	}
3234	// If reset is true, clear the parameters array.
3235	if (reset && compiled_data->minimize.parameters)
3236	{
3237		for (i = 0; i < compiled_data->minimize.parameters->rnum; i++)
3238			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))));
3239		ccv_array_clear(compiled_data->minimize.parameters);
3240	}
3241	if (!compiled_data->update_nodes)
3242		return;
3243	ccv_nnc_symbolic_graph_t* const symbolic_graph = model->graph;
3244	assert(symbolic_graph)((void) sizeof ((symbolic_graph) ? 1 : 0), __extension__ ({ if
 (symbolic_graph) ; else __assert_fail ("symbolic_graph", "ccv_cnnp_model.c"
, 3244, __extension__ __PRETTY_FUNCTION__); }));
3245	if (saved_aux_size > old_max_saved_aux_size)
3246	{
3247		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", 3247, __extension__ __PRETTY_FUNCTION__
); }));
3248		// Reallocate first, move them around later.
3249		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);
3250		compiled_data->update_nodes = (ccv_nnc_graph_exec_symbol_t*)(compiled_data->updated_parameters + parameter_size);
3251		compiled_data->saved_aux = (ccv_nnc_tensor_symbol_map_t*)(compiled_data->update_nodes + parameter_size);
3252		// We need to do this from back to front because saved_aux_size > old_saved_aux_size, it could overlap.
3253		_ccv_cnnp_scatter_saved_aux(compiled_data->saved_aux, parameter_size, old_max_saved_aux_size, saved_aux_size);
3254	}
3255	int flag = 0;
3256	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; });
3257	if (set_parameters && set_parameter_size)
3258	{
3259		ccv_array_t* const parameter_indices = ccv_array_new(sizeof(int), 0, 0);
3260		for (i = 0; i < set_parameter_size; i++)
3261		{
3262			const int param_sel = set_parameters[i]->param_sel > 0 ? set_parameters[i]->param_sel - 1 : set_parameters[i]->param_sel;
3263			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", 3263, __extension__ __PRETTY_FUNCTION__
); }));
3264			const int old_rnum = parameter_indices->rnum;
3265			ccv_cnnp_model_add_to_parameter_indices(set_parameters[i]->model, param_sel, parameter_indices);
3266			const int param_ref = set_parameters[i]->param_ref > 0 ? set_parameters[i]->param_ref - 1 : set_parameters[i]->param_ref;
3267			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", 3267, __extension__ __PRETTY_FUNCTION__
); }));
3268			if (param_ref >= 0)
3269			{
3270				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", 3270, __extension__ __PRETTY_FUNCTION__
); }));
3271				*(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)));
3272				parameter_indices->rnum = old_rnum + 1;
3273			}
3274		}
3275		// We may have duplicated indices, but that is OK, we will set it twice.
3276		for (i = 0; i < parameter_indices->rnum; i++)
3277		{
3278			const int d = *(int*)ccv_array_get(parameter_indices, i)((void*)(((char*)((parameter_indices)->data)) + (size_t)(parameter_indices
)->rsize * (size_t)(i)));
3279			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))
3280				flag = 1;
3281		}
3282		ccv_array_free(parameter_indices);
3283	} else {
3284		for (i = 0; i < parameter_size; i++)
3285			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))
3286				flag = 1;
3287		if (compiled_data->minimize.parameters)
3288			if (_ccv_cnnp_apply_parameters_with_minimizer(model))
3289				flag = 1;
3290	}
3291	if (flag)
3292	{
3293		// 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.
3294		if (compiled_data->graph_mode == CCV_CNNP_MODEL_GRAPH_FIT_MODE)
3295			_ccv_cnnp_compiled_data_graph_free(compiled_data);
3296		_ccv_cnnp_compiled_data_apply_gradients_free(compiled_data);
3297	}
3298}
3299 
3300void ccv_cnnp_model_set_compile_params(ccv_cnnp_model_t* const model, const ccv_nnc_symbolic_graph_compile_param_t compile_params)
3301{
3302	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
3303	assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
 (compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model.c"
, 3303, __extension__ __PRETTY_FUNCTION__); }));
3304	compiled_data->compile_params = compile_params;
3305}
3306 
3307void ccv_cnnp_model_dot(const ccv_cnnp_model_t* const model, const int flags, FILE** const outs, const int out_size)
3308{
3309	if (model->graph && out_size > 0)
3310		ccv_nnc_symbolic_graph_dot(model->graph, flags, outs[0]);
3311	if (model->compiled_data && model->compiled_data->graph && out_size > 1)
3312		ccv_nnc_graph_dot(model->compiled_data->graph, flags, outs[1]);
3313	if (model->compiled_data && model->compiled_data->backward.accum && out_size > 2)
3314		ccv_nnc_graph_dot(model->compiled_data->backward.accum, flags, outs[2]);
3315	if (model->compiled_data && model->compiled_data->apply_gradients.graph && out_size > 3)
3316		ccv_nnc_graph_dot(model->compiled_data->apply_gradients.graph, flags, outs[3]);
3317}
3318 
3319void ccv_cnnp_model_format(const ccv_cnnp_model_t* const model, const ccv_nnc_symbolic_graph_format_f format_fn, void* const context)
3320{
3321	if (model->graph)
3322		ccv_nnc_symbolic_graph_format(model->graph, 0, 0, 0, 0, format_fn, context);
3323}
3324 
3325static void _ccv_cnnp_compiled_data_free(const ccv_cnnp_model_t* const model, ccv_cnnp_compiled_data_t* const compiled_data)
3326{
3327	int i;
3328	const int parameter_size = compiled_data->parameters->rnum;
3329	ccv_array_free(compiled_data->parameters);
3330	if (compiled_data->parameter_flags)
3331		ccfreefree(compiled_data->parameter_flags);
3332	const int internal_size = compiled_data->internals->rnum;
3333	ccv_array_free(compiled_data->internals);
3334	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", 3334, __extension__ __PRETTY_FUNCTION__
); }));
3335	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", 3335, __extension__ __PRETTY_FUNCTION__
); }));
3336	for (i = 0; i < parameter_size; i++)
3337		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))));
3338	ccv_array_free(compiled_data->ids.parameters);
3339	for (i = 0; i < internal_size; i++)
3340		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))));
3341	ccv_array_free(compiled_data->ids.internals);
3342	const int parallel_count = compiled_data->parallel_count > 0 ? compiled_data->parallel_count : _ccv_cnnp_model_root_parallel_count(model);
3343	if (compiled_data->tensors.parameters)
3344	{
3345		for (i = 0; i < parameter_size * parallel_count; i++)
3346			// If it is not marked as not belonging, we can free it.
3347			if (!((uintptr_t)compiled_data->tensors.parameters[i] & (uintptr_t)1))
3348				if (compiled_data->tensors.parameters[i])
3349					ccv_nnc_tensor_free(compiled_data->tensors.parameters[i]);
3350		for (i = 0; i < internal_size * parallel_count; i++)
3351			if (compiled_data->tensors.internals[i])
3352				ccv_nnc_tensor_free(compiled_data->tensors.internals[i]);
3353		ccfreefree(compiled_data->tensors.parameters);
3354	}
3355	if (compiled_data->tensors.gradients)
3356	{
3357		for (i = 0; i < parameter_size * parallel_count; i++)
3358		{
3359			if (compiled_data->tensors.gradients[i])
3360				ccv_nnc_tensor_free(compiled_data->tensors.gradients[i]);
3361			if (compiled_data->tensors.accum_gradients[i])
3362				ccv_nnc_tensor_free(compiled_data->tensors.accum_gradients[i]);
3363		}
3364		ccfreefree(compiled_data->tensors.gradients);
3365	}
3366	if (compiled_data->minimize.parameters)
3367	{
3368		for (i = 0; i < compiled_data->minimize.parameters->rnum; i++)
3369			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))));
3370		ccv_array_free(compiled_data->minimize.parameters);
3371	}
3372	if (compiled_data->rewindables)
3373		ccv_array_free(compiled_data->rewindables);
3374	if (compiled_data->tensors_init.v)
3375		ccfreefree(CCV_NNC_INIT_V(compiled_data->tensors_init.v)((uint32_t*)((uintptr_t)(compiled_data->tensors_init.v) &
 ~(uintptr_t)1)));
3376	if (compiled_data->evaluate.tos)
3377		ccfreefree(compiled_data->evaluate.tos);
3378	compiled_data->evaluate.tos = 0;
3379	if (compiled_data->stream_map)
3380	{
3381		khiter_t k;
3382		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)
3383		{
3384			if (!kh_exist(compiled_data->stream_map, k)(!(((compiled_data->stream_map)->flags[(k)>>4]>>
(((k)&0xfU)<<1))&3)))
3385				continue;
3386			ccv_nnc_stream_context_t* const stream = kh_val(compiled_data->stream_map, k)((compiled_data->stream_map)->vals[k]);
3387			ccv_nnc_stream_context_free(stream);
3388		}
3389		kh_destroy(stream_map, compiled_data->stream_map)kh_destroy_stream_map(compiled_data->stream_map);
3390	}
3391	_ccv_cnnp_compiled_data_graph_free(compiled_data);
3392	_ccv_cnnp_compiled_data_gradient_free(compiled_data);
3393	_ccv_cnnp_compiled_data_backward_free(compiled_data);
3394	_ccv_cnnp_compiled_data_apply_gradients_free(compiled_data);
3395	if (compiled_data->gradient_checkpoints)
3396	{
3397		for (i = 0; i < compiled_data->gradient_checkpoints->rnum; i++)
3398		{
3399			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)));
3400			assert(checkpoint->inputs)((void) sizeof ((checkpoint->inputs) ? 1 : 0), __extension__
 ({ if (checkpoint->inputs) ; else __assert_fail ("checkpoint->inputs"
, "ccv_cnnp_model.c", 3400, __extension__ __PRETTY_FUNCTION__
); }));
3401			ccfreefree(checkpoint->inputs);
3402			ccv_array_free(checkpoint->tensor_symbols);
3403		}
3404		ccv_array_free(compiled_data->gradient_checkpoints);
3405	}
3406	ccv_nnc_xpu_alloc_destroy(&compiled_data->xpu_alloc);
3407	ccfreefree(compiled_data);
3408}
3409 
3410void ccv_cnnp_model_free(ccv_cnnp_model_t* const model)
3411{
3412	ccv_cnnp_model_deinit(model);
3413	if (model->isa->dealloc)
3414		model->isa->dealloc(model);
3415	if (model->io)
3416	{
3417		int i;
3418		for (i = 0; i < model->io->rnum; i++)
3419		{
3420			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)));
3421			if (model_io->outgoings)
3422				ccv_array_free(model_io->outgoings);
3423			if (model_io->incomings)
3424				ccv_array_free(model_io->incomings);
3425			if (model_io->dependencies)
3426				ccv_array_free(model_io->dependencies);
3427			ccfreefree(model_io);
3428		}
3429		ccv_array_free(model->io);
3430	}
3431	if (model->parameter_indices)
3432		ccv_array_free(model->parameter_indices);
3433	if (model->inputs)
3434		ccfreefree(model->inputs);
3435	if (model->graph)
3436		ccv_nnc_symbolic_graph_free(model->graph);
3437	if (model->compiled_data)
3438		_ccv_cnnp_compiled_data_free(model, model->compiled_data);
3439	if (model->name)
3440		ccfreefree(model->name);
3441	ccfreefree(model);
3442}
3443 
3444void ccv_cnnp_model_cancel(ccv_cnnp_model_t* const model)
3445{
3446	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
3447	if (!compiled_data)
3448		return;
3449	if (compiled_data->graph)
3450		ccv_nnc_graph_cancel(compiled_data->graph);
3451	if (compiled_data->apply_gradients.graph)
3452		ccv_nnc_graph_cancel(compiled_data->apply_gradients.graph);
3453}
3454 
3455void ccv_cnnp_model_async_enter(ccv_cnnp_model_t* const model)
3456{
3457	ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
3458	if (!compiled_data)
3459		return;
3460	if (compiled_data->graph)
3461		ccv_nnc_graph_async_enter(compiled_data->graph);
3462	if (compiled_data->apply_gradients.graph)
3463		ccv_nnc_graph_async_enter(compiled_data->apply_gradients.graph);
3464}
3465 
3466void ccv_cnnp_model_set_flags(ccv_cnnp_model_t* const model, const int flags)
3467{
3468	model->exec_flags = flags;
3469}
3470 
3471int ccv_cnnp_model_flags(ccv_cnnp_model_t* const model)
3472{
3473	return model->exec_flags;
3474}