File: | nnc/ccv_nnc_graph_run.c |
Warning: | line 879, column 4 Array access (from variable 'outputs') results in a null pointer dereference |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
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_nnc_graph.h" | ||||
6 | #include "_ccv_nnc_stream.h" | ||||
7 | #ifdef HAVE_CUDA1 | ||||
8 | #include "gpu/ccv_nnc_compat.h" | ||||
9 | #elif defined(HAVE_MPS) | ||||
10 | #include "mps/ccv_nnc_mps.h" | ||||
11 | #endif | ||||
12 | |||||
13 | // MARK - Level-2 API | ||||
14 | |||||
15 | static void _ccv_nnc_unwrap_tensor_wrap(const ccv_nnc_graph_t* const graph, const int64_t count, const int64_t reverse_count, ccv_nnc_graph_tensor_wrap_t* const tensor_wrap) | ||||
16 | { | ||||
17 | ccv_nnc_tensor_t* tensor = tensor_wrap->tensors[tensor_wrap->index]; | ||||
18 | while (CCV_IS_TENSOR_MULTIVIEW(tensor)((*(int*)(tensor)) & CCV_TENSOR_MULTIVIEW) && | ||||
19 | (((ccv_nnc_tensor_multiview_t*)tensor)->anchor == (intptr_t)graph || | ||||
20 | ((ccv_nnc_tensor_multiview_t*)tensor)->anchor == (intptr_t)graph->pair)) | ||||
21 | { | ||||
22 | // If the anchor is from the pair, we use the reverse_count instead (we are looking it up). | ||||
23 | const int i = (int)((((ccv_nnc_tensor_multiview_t*)tensor)->anchor == (intptr_t)graph) ? count : reverse_count); | ||||
24 | ccv_nnc_tensor_multiview_t* mv = (ccv_nnc_tensor_multiview_t*)tensor; | ||||
25 | const int off = mv->kind; | ||||
26 | const int mod = mv->repeat; | ||||
27 | tensor = CCV_NNC_MULTIVIEW_DATA(mv)((mv)->_heap_data ? (mv)->_heap_data : (mv)->_inline_data )[i >= off ? ((i - off) % mod) + off : i]; // Unwrap. | ||||
28 | // If reached the root. | ||||
29 | if (!CCV_IS_TENSOR_MULTIVIEW(tensor)((*(int*)(tensor)) & CCV_TENSOR_MULTIVIEW)) | ||||
30 | tensor_wrap->update_required = 1; // Need to update tensor updates. | ||||
31 | ++tensor_wrap->index; | ||||
32 | tensor_wrap->tensors[tensor_wrap->index] = tensor; | ||||
33 | assert(tensor_wrap->index < tensor_wrap->count)((void) sizeof ((tensor_wrap->index < tensor_wrap->count ) ? 1 : 0), __extension__ ({ if (tensor_wrap->index < tensor_wrap ->count) ; else __assert_fail ("tensor_wrap->index < tensor_wrap->count" , "ccv_nnc_graph_run.c", 33, __extension__ __PRETTY_FUNCTION__ ); })); | ||||
34 | } | ||||
35 | } | ||||
36 | |||||
37 | static void _ccv_nnc_graph_unwrap_sub_graph(const ccv_nnc_graph_t* const graph, const int64_t count, const int64_t reverse_count, const ccv_nnc_graph_t* const sub_graph) | ||||
38 | { | ||||
39 | int i; | ||||
40 | if (sub_graph->carry_overs) | ||||
41 | for (i = 0; i < sub_graph->carry_overs->rnum; i++) | ||||
42 | { | ||||
43 | ccv_nnc_graph_tensor_carry_over_t* const carry_over = (ccv_nnc_graph_tensor_carry_over_t*)ccv_array_get(sub_graph->carry_overs, i)((void*)(((char*)((sub_graph->carry_overs)->data)) + (size_t )(sub_graph->carry_overs)->rsize * (size_t)(i))); | ||||
44 | _ccv_nnc_unwrap_tensor_wrap(graph, count, reverse_count, carry_over->from); | ||||
45 | _ccv_nnc_unwrap_tensor_wrap(graph, count, reverse_count, carry_over->to); | ||||
46 | } | ||||
47 | if (sub_graph->sub_graphs) | ||||
48 | for (i = 0; i < sub_graph->sub_graphs->rnum; i++) | ||||
49 | _ccv_nnc_graph_unwrap_sub_graph(graph, count, reverse_count, *(ccv_nnc_graph_t**)ccv_array_get(sub_graph->sub_graphs, i)((void*)(((char*)((sub_graph->sub_graphs)->data)) + (size_t )(sub_graph->sub_graphs)->rsize * (size_t)(i)))); | ||||
50 | } | ||||
51 | |||||
52 | static void _ccv_nnc_graph_unwrap(const ccv_nnc_graph_t* const graph, const int64_t count, const int64_t reverse_count) | ||||
53 | { | ||||
54 | if (!graph->tensor_wraps_refs) | ||||
55 | return; | ||||
56 | int i, j; | ||||
57 | for (i = 0; i < graph->tensor_wraps_refs->rnum; i++) | ||||
58 | { | ||||
59 | const ccv_nnc_graph_tensor_wraps_ref_t* const tensor_wraps_ref = (const ccv_nnc_graph_tensor_wraps_ref_t*)ccv_array_get(graph->tensor_wraps_refs, i)((void*)(((char*)((graph->tensor_wraps_refs)->data)) + ( size_t)(graph->tensor_wraps_refs)->rsize * (size_t)(i)) ); | ||||
60 | const ccv_nnc_graph_t* const sub_graph = tensor_wraps_ref->graph; | ||||
61 | ccv_nnc_graph_tensor_wrap_array_t* const tensor_wrap_array = *(ccv_nnc_graph_tensor_wrap_array_t**)ccv_array_get(sub_graph->tensor_wraps, tensor_wraps_ref->d)((void*)(((char*)((sub_graph->tensor_wraps)->data)) + ( size_t)(sub_graph->tensor_wraps)->rsize * (size_t)(tensor_wraps_ref ->d))); | ||||
62 | if (tensor_wrap_array) | ||||
63 | for (j = 0; j < tensor_wrap_array->size; j++) | ||||
64 | { | ||||
65 | ccv_nnc_graph_tensor_wrap_t* const tensor_wrap = tensor_wrap_array->tensor_wraps[j]; | ||||
66 | if (!tensor_wrap) | ||||
67 | continue; | ||||
68 | _ccv_nnc_unwrap_tensor_wrap(graph, count, reverse_count, tensor_wrap); | ||||
69 | } | ||||
70 | } | ||||
71 | _ccv_nnc_graph_unwrap_sub_graph(graph, count, reverse_count, graph); | ||||
72 | } | ||||
73 | |||||
74 | static void _ccv_nnc_graph_transit_move_to(const ccv_nnc_graph_t* const graph) | ||||
75 | { | ||||
76 | int i; | ||||
77 | if (graph->carry_overs) | ||||
78 | for (i = 0; i < graph->carry_overs->rnum; i++) | ||||
79 | { | ||||
80 | ccv_nnc_graph_tensor_carry_over_t* const carry_over = (ccv_nnc_graph_tensor_carry_over_t*)ccv_array_get(graph->carry_overs, i)((void*)(((char*)((graph->carry_overs)->data)) + (size_t )(graph->carry_overs)->rsize * (size_t)(i))); | ||||
81 | ccv_nnc_tensor_t* it = (ccv_nnc_tensor_t*)(carry_over->to->tensors[carry_over->to->index]); | ||||
82 | assert(!CCV_IS_TENSOR_MULTIVIEW(it))((void) sizeof ((!((*(int*)(it)) & CCV_TENSOR_MULTIVIEW)) ? 1 : 0), __extension__ ({ if (!((*(int*)(it)) & CCV_TENSOR_MULTIVIEW )) ; else __assert_fail ("!CCV_IS_TENSOR_MULTIVIEW(it)", "ccv_nnc_graph_run.c" , 82, __extension__ __PRETTY_FUNCTION__); })); | ||||
83 | it->data = carry_over->transit; | ||||
84 | } | ||||
85 | } | ||||
86 | |||||
87 | static void _ccv_nnc_graph_from_move_transit(const ccv_nnc_graph_t* const graph) | ||||
88 | { | ||||
89 | int i; | ||||
90 | if (graph->carry_overs) | ||||
91 | for (i = 0; i < graph->carry_overs->rnum; i++) | ||||
92 | { | ||||
93 | ccv_nnc_graph_tensor_carry_over_t* const carry_over = (ccv_nnc_graph_tensor_carry_over_t*)ccv_array_get(graph->carry_overs, i)((void*)(((char*)((graph->carry_overs)->data)) + (size_t )(graph->carry_overs)->rsize * (size_t)(i))); | ||||
94 | ccv_nnc_tensor_t* it = (ccv_nnc_tensor_t*)(carry_over->from->tensors[carry_over->from->index]); | ||||
95 | assert(!CCV_IS_TENSOR_MULTIVIEW(it))((void) sizeof ((!((*(int*)(it)) & CCV_TENSOR_MULTIVIEW)) ? 1 : 0), __extension__ ({ if (!((*(int*)(it)) & CCV_TENSOR_MULTIVIEW )) ; else __assert_fail ("!CCV_IS_TENSOR_MULTIVIEW(it)", "ccv_nnc_graph_run.c" , 95, __extension__ __PRETTY_FUNCTION__); })); | ||||
96 | carry_over->transit = it->data; | ||||
97 | } | ||||
98 | } | ||||
99 | |||||
100 | static void _ccv_nnc_rewrap_tensor_wrap(const ccv_nnc_graph_t* const graph, ccv_nnc_graph_tensor_wrap_t* const tensor_wrap) | ||||
101 | { | ||||
102 | while (tensor_wrap->index > 0 && CCV_IS_TENSOR_MULTIVIEW(tensor_wrap->tensors[tensor_wrap->index - 1])((*(int*)(tensor_wrap->tensors[tensor_wrap->index - 1]) ) & CCV_TENSOR_MULTIVIEW) && | ||||
103 | (((ccv_nnc_tensor_multiview_t*)tensor_wrap->tensors[tensor_wrap->index - 1])->anchor == (intptr_t)graph || | ||||
104 | ((ccv_nnc_tensor_multiview_t*)tensor_wrap->tensors[tensor_wrap->index - 1])->anchor == (intptr_t)graph->pair)) | ||||
105 | --tensor_wrap->index; | ||||
106 | } | ||||
107 | |||||
108 | static void _ccv_nnc_graph_rewrap_sub_graph(const ccv_nnc_graph_t* const graph, const ccv_nnc_graph_t* const sub_graph) | ||||
109 | { | ||||
110 | int i; | ||||
111 | if (sub_graph->carry_overs) | ||||
112 | for (i = 0; i < sub_graph->carry_overs->rnum; i++) | ||||
113 | { | ||||
114 | ccv_nnc_graph_tensor_carry_over_t* const carry_over = (ccv_nnc_graph_tensor_carry_over_t*)ccv_array_get(sub_graph->carry_overs, i)((void*)(((char*)((sub_graph->carry_overs)->data)) + (size_t )(sub_graph->carry_overs)->rsize * (size_t)(i))); | ||||
115 | _ccv_nnc_rewrap_tensor_wrap(graph, carry_over->from); | ||||
116 | _ccv_nnc_rewrap_tensor_wrap(graph, carry_over->to); | ||||
117 | } | ||||
118 | if (sub_graph->sub_graphs) | ||||
119 | for (i = 0; i < sub_graph->sub_graphs->rnum; i++) | ||||
120 | _ccv_nnc_graph_rewrap_sub_graph(graph, *(ccv_nnc_graph_t**)ccv_array_get(sub_graph->sub_graphs, i)((void*)(((char*)((sub_graph->sub_graphs)->data)) + (size_t )(sub_graph->sub_graphs)->rsize * (size_t)(i)))); | ||||
121 | } | ||||
122 | |||||
123 | static void _ccv_nnc_graph_rewrap(const ccv_nnc_graph_t* const graph) // Call this method at the end to roll the wrap_ptr back | ||||
124 | { | ||||
125 | if (!graph->tensor_wraps_refs) | ||||
126 | return; | ||||
127 | int i, j; | ||||
128 | for (i = 0; i < graph->tensor_wraps_refs->rnum; i++) | ||||
129 | { | ||||
130 | const ccv_nnc_graph_tensor_wraps_ref_t* const tensor_wraps_ref = (const ccv_nnc_graph_tensor_wraps_ref_t*)ccv_array_get(graph->tensor_wraps_refs, i)((void*)(((char*)((graph->tensor_wraps_refs)->data)) + ( size_t)(graph->tensor_wraps_refs)->rsize * (size_t)(i)) ); | ||||
131 | const ccv_nnc_graph_t* const sub_graph = tensor_wraps_ref->graph; | ||||
132 | ccv_nnc_graph_tensor_wrap_array_t* const tensor_wrap_array = *(ccv_nnc_graph_tensor_wrap_array_t**)ccv_array_get(sub_graph->tensor_wraps, tensor_wraps_ref->d)((void*)(((char*)((sub_graph->tensor_wraps)->data)) + ( size_t)(sub_graph->tensor_wraps)->rsize * (size_t)(tensor_wraps_ref ->d))); | ||||
133 | if (tensor_wrap_array) | ||||
134 | for (j = 0; j < tensor_wrap_array->size; j++) | ||||
135 | { | ||||
136 | ccv_nnc_graph_tensor_wrap_t* const tensor_wrap = tensor_wrap_array->tensor_wraps[j]; | ||||
137 | if (!tensor_wrap) | ||||
138 | continue; | ||||
139 | _ccv_nnc_rewrap_tensor_wrap(graph, tensor_wrap); | ||||
140 | } | ||||
141 | } | ||||
142 | _ccv_nnc_graph_rewrap_sub_graph(graph, graph); | ||||
143 | } | ||||
144 | |||||
145 | static void _ccv_nnc_graph_exec_unwrap_io(const ccv_nnc_graph_t* const graph, ccv_nnc_graph_exec_info_t* const node) | ||||
146 | { | ||||
147 | if (!node->tensor_wraps_ref) | ||||
148 | return; | ||||
149 | int i; | ||||
150 | ccv_nnc_graph_tensor_wrap_array_t* const tensor_wrap_array = *(ccv_nnc_graph_tensor_wrap_array_t**)ccv_array_get(graph->tensor_wraps, node->tensor_wraps_ref - 1)((void*)(((char*)((graph->tensor_wraps)->data)) + (size_t )(graph->tensor_wraps)->rsize * (size_t)(node->tensor_wraps_ref - 1))); | ||||
151 | ccv_nnc_graph_tensor_wrap_t** const tensor_wraps = tensor_wrap_array->tensor_wraps; | ||||
152 | for (i = 0; i < tensor_wrap_array->size; i++) | ||||
153 | if (tensor_wraps[i]) | ||||
154 | { | ||||
155 | assert(tensor_wraps[i]->index > 0)((void) sizeof ((tensor_wraps[i]->index > 0) ? 1 : 0), __extension__ ({ if (tensor_wraps[i]->index > 0) ; else __assert_fail ("tensor_wraps[i]->index > 0", "ccv_nnc_graph_run.c", 155 , __extension__ __PRETTY_FUNCTION__); })); | ||||
156 | ccv_nnc_tensor_multiview_t* mv = (ccv_nnc_tensor_multiview_t*)(tensor_wraps[i]->tensors[tensor_wraps[i]->index - 1]); | ||||
157 | assert(CCV_IS_TENSOR_MULTIVIEW(mv))((void) sizeof ((((*(int*)(mv)) & CCV_TENSOR_MULTIVIEW)) ? 1 : 0), __extension__ ({ if (((*(int*)(mv)) & CCV_TENSOR_MULTIVIEW )) ; else __assert_fail ("CCV_IS_TENSOR_MULTIVIEW(mv)", "ccv_nnc_graph_run.c" , 157, __extension__ __PRETTY_FUNCTION__); })); | ||||
158 | // Only now set the mv->it, because now this node is about to get executed. | ||||
159 | mv->it = tensor_wraps[i]->tensors[tensor_wraps[i]->index]; | ||||
160 | assert(!CCV_IS_TENSOR_MULTIVIEW(mv->it))((void) sizeof ((!((*(int*)(mv->it)) & CCV_TENSOR_MULTIVIEW )) ? 1 : 0), __extension__ ({ if (!((*(int*)(mv->it)) & CCV_TENSOR_MULTIVIEW)) ; else __assert_fail ("!CCV_IS_TENSOR_MULTIVIEW(mv->it)" , "ccv_nnc_graph_run.c", 160, __extension__ __PRETTY_FUNCTION__ ); })); | ||||
161 | } | ||||
162 | for (i = 0; i < node->input_size; i++) | ||||
163 | if (tensor_wraps[i]) | ||||
164 | node->inputs[i] = tensor_wraps[i]->tensors[tensor_wraps[i]->index]; | ||||
165 | const int d = node->input_size; | ||||
166 | for (i = 0; i < node->output_size; i++) | ||||
167 | if (tensor_wraps[d + i]) | ||||
168 | node->outputs[i] = tensor_wraps[d + i]->tensors[tensor_wraps[d + i]->index]; | ||||
169 | } | ||||
170 | |||||
171 | static void _ccv_nnc_graph_exec_unwrap_while_expr(const ccv_nnc_graph_t* const graph, ccv_nnc_graph_exec_info_t* const node) | ||||
172 | { | ||||
173 | assert(node->flags & CCV_NNC_GRAPH_EXEC_P_WHILE)((void) sizeof ((node->flags & CCV_NNC_GRAPH_EXEC_P_WHILE ) ? 1 : 0), __extension__ ({ if (node->flags & CCV_NNC_GRAPH_EXEC_P_WHILE ) ; else __assert_fail ("node->flags & CCV_NNC_GRAPH_EXEC_P_WHILE" , "ccv_nnc_graph_run.c", 173, __extension__ __PRETTY_FUNCTION__ ); })); | ||||
174 | if (!node->p_while.tensor_wraps_ref) | ||||
175 | return; | ||||
176 | int i; | ||||
177 | ccv_nnc_graph_tensor_wrap_array_t* const tensor_wrap_array = *(ccv_nnc_graph_tensor_wrap_array_t**)ccv_array_get(graph->tensor_wraps, node->p_while.tensor_wraps_ref - 1)((void*)(((char*)((graph->tensor_wraps)->data)) + (size_t )(graph->tensor_wraps)->rsize * (size_t)(node->p_while .tensor_wraps_ref - 1))); | ||||
178 | ccv_nnc_graph_tensor_wrap_t** const tensor_wraps = tensor_wrap_array->tensor_wraps; | ||||
179 | for (i = 0; i < tensor_wrap_array->size; i++) | ||||
180 | if (tensor_wraps[i]) | ||||
181 | { | ||||
182 | assert(tensor_wraps[i]->index > 0)((void) sizeof ((tensor_wraps[i]->index > 0) ? 1 : 0), __extension__ ({ if (tensor_wraps[i]->index > 0) ; else __assert_fail ("tensor_wraps[i]->index > 0", "ccv_nnc_graph_run.c", 182 , __extension__ __PRETTY_FUNCTION__); })); | ||||
183 | ccv_nnc_tensor_multiview_t* mv = (ccv_nnc_tensor_multiview_t*)(tensor_wraps[i]->tensors[tensor_wraps[i]->index - 1]); | ||||
184 | assert(CCV_IS_TENSOR_MULTIVIEW(mv))((void) sizeof ((((*(int*)(mv)) & CCV_TENSOR_MULTIVIEW)) ? 1 : 0), __extension__ ({ if (((*(int*)(mv)) & CCV_TENSOR_MULTIVIEW )) ; else __assert_fail ("CCV_IS_TENSOR_MULTIVIEW(mv)", "ccv_nnc_graph_run.c" , 184, __extension__ __PRETTY_FUNCTION__); })); | ||||
185 | // Only now set the mv->it, because now this node is about to get executed. | ||||
186 | mv->it = tensor_wraps[i]->tensors[tensor_wraps[i]->index]; | ||||
187 | assert(!CCV_IS_TENSOR_MULTIVIEW(mv->it))((void) sizeof ((!((*(int*)(mv->it)) & CCV_TENSOR_MULTIVIEW )) ? 1 : 0), __extension__ ({ if (!((*(int*)(mv->it)) & CCV_TENSOR_MULTIVIEW)) ; else __assert_fail ("!CCV_IS_TENSOR_MULTIVIEW(mv->it)" , "ccv_nnc_graph_run.c", 187, __extension__ __PRETTY_FUNCTION__ ); })); | ||||
188 | } | ||||
189 | for (i = 0; i < node->p_while.input_size; i++) | ||||
190 | if (tensor_wraps[i]) | ||||
191 | node->p_while.inputs[i] = tensor_wraps[i]->tensors[tensor_wraps[i]->index]; | ||||
192 | } | ||||
193 | |||||
194 | static void _ccv_nnc_graph_exec_unwrap_phi(ccv_nnc_graph_t* const graph, const ccv_nnc_graph_exec_info_t* const node, const int ref) | ||||
195 | { | ||||
196 | int i; | ||||
197 | // If the output tensor is a phi multi-view tensor, we update our selection to all the subscribers. | ||||
198 | for (i = 0; i < node->output_size; i++) | ||||
199 | if (CCV_IS_TENSOR_MULTIVIEW(node->outputs[i])((*(int*)(node->outputs[i])) & CCV_TENSOR_MULTIVIEW) && | ||||
200 | ((ccv_nnc_tensor_multiview_t*)node->outputs[i])->anchor == CCV_NNC_MULTIVIEW_PHI(intptr_t)0x1) | ||||
201 | { | ||||
202 | ccv_nnc_tensor_multiview_t* const mv = (ccv_nnc_tensor_multiview_t*)node->outputs[i]; | ||||
203 | mv->it = CCV_NNC_MULTIVIEW_DATA(mv)((mv)->_heap_data ? (mv)->_heap_data : (mv)->_inline_data )[ref >= 0]; | ||||
204 | ccv_nnc_tensor_multiview_synchronize(mv); | ||||
205 | } | ||||
206 | } | ||||
207 | |||||
208 | static void _ccv_nnc_graph_exec_begin_synchronize_multiviews(ccv_nnc_graph_t* const graph, ccv_nnc_graph_exec_info_t* const node) | ||||
209 | { | ||||
210 | if (!node->tensor_wraps_ref) | ||||
211 | return; | ||||
212 | int i; | ||||
213 | ccv_nnc_graph_tensor_wrap_array_t* const tensor_wrap_array = *(ccv_nnc_graph_tensor_wrap_array_t**)ccv_array_get(graph->tensor_wraps, node->tensor_wraps_ref - 1)((void*)(((char*)((graph->tensor_wraps)->data)) + (size_t )(graph->tensor_wraps)->rsize * (size_t)(node->tensor_wraps_ref - 1))); | ||||
214 | ccv_nnc_graph_tensor_wrap_t** const tensor_wraps = tensor_wrap_array->tensor_wraps; | ||||
215 | for (i = 0; i < tensor_wrap_array->size; i++) | ||||
216 | if (tensor_wraps[i] && tensor_wraps[i]->update_required) | ||||
217 | { | ||||
218 | assert(tensor_wraps[i]->index > 0)((void) sizeof ((tensor_wraps[i]->index > 0) ? 1 : 0), __extension__ ({ if (tensor_wraps[i]->index > 0) ; else __assert_fail ("tensor_wraps[i]->index > 0", "ccv_nnc_graph_run.c", 218 , __extension__ __PRETTY_FUNCTION__); })); | ||||
219 | ccv_nnc_tensor_multiview_t* const mv = (ccv_nnc_tensor_multiview_t*)(tensor_wraps[i]->tensors[tensor_wraps[i]->index - 1]); | ||||
220 | // Now update the final pointer. | ||||
221 | ccv_nnc_tensor_multiview_synchronize(mv); | ||||
222 | tensor_wraps[i]->update_required = 0; // Reset, no need to update. | ||||
223 | } | ||||
224 | } | ||||
225 | |||||
226 | void ccv_nnc_print_tensor_shape(const ccv_nnc_tensor_t* const tensor) | ||||
227 | { | ||||
228 | int i; | ||||
229 | PRINT(CCV_CLI_INFO, " [%d", tensor->info.dim[0])do { if ((CCV_CLI_INFO & ccv_cli_get_output_levels())) { printf (" [%d", tensor->info.dim[0]); fflush(stdout); } } while ( 0); | ||||
230 | for (i = 1; i < CCV_NNC_MAX_DIM_ALLOC(12) && tensor->info.dim[i]; i++) | ||||
231 | PRINT(CCV_CLI_INFO, "x%d", tensor->info.dim[i])do { if ((CCV_CLI_INFO & ccv_cli_get_output_levels())) { printf ("x%d", tensor->info.dim[i]); fflush(stdout); } } while (0 ); | ||||
232 | PRINT(CCV_CLI_INFO, "]")do { if ((CCV_CLI_INFO & ccv_cli_get_output_levels())) { printf ("]"); fflush(stdout); } } while (0); | ||||
233 | } | ||||
234 | |||||
235 | void ccv_nnc_print_tensor_info(const ccv_nnc_tensor_t* const tensor) | ||||
236 | { | ||||
237 | int i; | ||||
238 | PRINT(CCV_CLI_INFO, " [%d", tensor->info.dim[0])do { if ((CCV_CLI_INFO & ccv_cli_get_output_levels())) { printf (" [%d", tensor->info.dim[0]); fflush(stdout); } } while ( 0); | ||||
239 | for (i = 1; i < CCV_NNC_MAX_DIM_ALLOC(12) && tensor->info.dim[i]; i++) | ||||
240 | PRINT(CCV_CLI_INFO, "x%d", tensor->info.dim[i])do { if ((CCV_CLI_INFO & ccv_cli_get_output_levels())) { printf ("x%d", tensor->info.dim[i]); fflush(stdout); } } while (0 ); | ||||
241 | PRINT(CCV_CLI_INFO, "]")do { if ((CCV_CLI_INFO & ccv_cli_get_output_levels())) { printf ("]"); fflush(stdout); } } while (0); | ||||
242 | if (!CCV_CLI_OUTPUT_LEVEL_IS(CCV_CLI_VERBOSE)(CCV_CLI_VERBOSE & ccv_cli_get_output_levels()) || tensor->info.dim[0] <= 0) | ||||
243 | return; | ||||
244 | const int nd = ccv_nnc_tensor_nd(tensor->info.dim); | ||||
245 | const int len = ccv_min(tensor->info.dim[nd - 1], 3)({ typeof (tensor->info.dim[nd - 1]) _a = (tensor->info .dim[nd - 1]); typeof (3) _b = (3); (_a < _b) ? _a : _b; } ); | ||||
246 | if (CCV_TENSOR_GET_MEMORY(tensor->info.type)((tensor->info.type) & 0x3) == CCV_TENSOR_GPU_MEMORY) | ||||
247 | { | ||||
248 | #ifdef HAVE_CUDA1 | ||||
249 | switch (tensor->info.datatype) | ||||
250 | { | ||||
251 | case CCV_16F: { | ||||
252 | uint16_t data[len]; | ||||
253 | cumemcpy(data, CCV_TENSOR_CPU_MEMORY, tensor->data.f16, tensor->info.type, len * sizeof(uint16_t)); | ||||
254 | float fp32[len]; | ||||
255 | ccv_half_precision_to_float(data, fp32, len); | ||||
256 | for (i = 0; i < len; i++) | ||||
257 | PRINT(CCV_CLI_VERBOSE, " %f", fp32[i])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf(" %f", fp32[i]); fflush(stdout); } } while (0); | ||||
258 | break; | ||||
259 | } | ||||
260 | case CCV_32F: { | ||||
261 | float data[len]; | ||||
262 | cumemcpy(data, CCV_TENSOR_CPU_MEMORY, tensor->data.f32, tensor->info.type, len * sizeof(float)); | ||||
263 | for (i = 0; i < len; i++) | ||||
264 | PRINT(CCV_CLI_VERBOSE, " %f", data[i])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf(" %f", data[i]); fflush(stdout); } } while (0); | ||||
265 | break; | ||||
266 | } | ||||
267 | case CCV_64F: { | ||||
268 | double data[len]; | ||||
269 | cumemcpy(data, CCV_TENSOR_CPU_MEMORY, tensor->data.f64, tensor->info.type, len * sizeof(double)); | ||||
270 | for (i = 0; i < len; i++) | ||||
271 | PRINT(CCV_CLI_VERBOSE, " %f", data[i])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf(" %f", data[i]); fflush(stdout); } } while (0); | ||||
272 | break; | ||||
273 | } | ||||
274 | case CCV_32S: { | ||||
275 | int data[len]; | ||||
276 | cumemcpy(data, CCV_TENSOR_CPU_MEMORY, tensor->data.i32, tensor->info.type, len * sizeof(int)); | ||||
277 | for (i = 0; i < len; i++) | ||||
278 | PRINT(CCV_CLI_VERBOSE, " %d", data[i])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf(" %d", data[i]); fflush(stdout); } } while (0); | ||||
279 | break; | ||||
280 | } | ||||
281 | case CCV_64S: { | ||||
282 | int64_t data[len]; | ||||
283 | cumemcpy(data, CCV_TENSOR_CPU_MEMORY, tensor->data.i64, tensor->info.type, len * sizeof(int64_t)); | ||||
284 | for (i = 0; i < len; i++) | ||||
285 | PRINT(CCV_CLI_VERBOSE, " %lld", (long long)data[i])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf(" %lld", (long long)data[i]); fflush(stdout); } } while (0); | ||||
286 | break; | ||||
287 | } | ||||
288 | case CCV_8U: { | ||||
289 | uint8_t data[len]; | ||||
290 | cumemcpy(data, CCV_TENSOR_CPU_MEMORY, tensor->data.u8, tensor->info.type, len * sizeof(uint8_t)); | ||||
291 | for (i = 0; i < len; i++) | ||||
292 | PRINT(CCV_CLI_VERBOSE, " %d", (int)data[i])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf(" %d", (int)data[i]); fflush(stdout); } } while (0); | ||||
293 | break; | ||||
294 | } | ||||
295 | } | ||||
296 | if (ccv_nnc_tensor_count(tensor->info) > 3) | ||||
297 | PRINT(CCV_CLI_VERBOSE, " ..")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf(" .."); fflush(stdout); } } while (0); | ||||
298 | #elif defined(HAVE_MPS) | ||||
299 | switch (tensor->info.datatype) | ||||
300 | { | ||||
301 | case CCV_16F: { | ||||
302 | uint16_t data[len]; | ||||
303 | mpmemcpy(data, 0, CCV_TENSOR_CPU_MEMORY, tensor->data.f16, tensor->dataof, tensor->info.type, len * sizeof(uint16_t)); | ||||
304 | float fp32[len]; | ||||
305 | ccv_half_precision_to_float(data, fp32, len); | ||||
306 | for (i = 0; i < len; i++) | ||||
307 | PRINT(CCV_CLI_VERBOSE, " %f", fp32[i])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf(" %f", fp32[i]); fflush(stdout); } } while (0); | ||||
308 | break; | ||||
309 | } | ||||
310 | case CCV_32F: { | ||||
311 | float data[len]; | ||||
312 | mpmemcpy(data, 0, CCV_TENSOR_CPU_MEMORY, tensor->data.f32, tensor->dataof, tensor->info.type, len * sizeof(float)); | ||||
313 | for (i = 0; i < len; i++) | ||||
314 | PRINT(CCV_CLI_VERBOSE, " %f", data[i])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf(" %f", data[i]); fflush(stdout); } } while (0); | ||||
315 | break; | ||||
316 | } | ||||
317 | case CCV_64F: { | ||||
318 | double data[len]; | ||||
319 | mpmemcpy(data, 0, CCV_TENSOR_CPU_MEMORY, tensor->data.f64, tensor->dataof, tensor->info.type, len * sizeof(double)); | ||||
320 | for (i = 0; i < len; i++) | ||||
321 | PRINT(CCV_CLI_VERBOSE, " %f", data[i])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf(" %f", data[i]); fflush(stdout); } } while (0); | ||||
322 | break; | ||||
323 | } | ||||
324 | case CCV_32S: { | ||||
325 | int data[len]; | ||||
326 | mpmemcpy(data, 0, CCV_TENSOR_CPU_MEMORY, tensor->data.i32, tensor->dataof, tensor->info.type, len * sizeof(int)); | ||||
327 | for (i = 0; i < len; i++) | ||||
328 | PRINT(CCV_CLI_VERBOSE, " %d", data[i])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf(" %d", data[i]); fflush(stdout); } } while (0); | ||||
329 | break; | ||||
330 | } | ||||
331 | case CCV_64S: { | ||||
332 | int64_t data[len]; | ||||
333 | mpmemcpy(data, 0, CCV_TENSOR_CPU_MEMORY, tensor->data.i64, tensor->dataof, tensor->info.type, len * sizeof(int64_t)); | ||||
334 | for (i = 0; i < len; i++) | ||||
335 | PRINT(CCV_CLI_VERBOSE, " %lld", (long long)data[i])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf(" %lld", (long long)data[i]); fflush(stdout); } } while (0); | ||||
336 | break; | ||||
337 | } | ||||
338 | case CCV_8U: { | ||||
339 | uint8_t data[len]; | ||||
340 | mpmemcpy(data, 0, CCV_TENSOR_CPU_MEMORY, tensor->data.u8, tensor->dataof, tensor->info.type, len * sizeof(uint8_t)); | ||||
341 | for (i = 0; i < len; i++) | ||||
342 | PRINT(CCV_CLI_VERBOSE, " %d", (int)data[i])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf(" %d", (int)data[i]); fflush(stdout); } } while (0); | ||||
343 | break; | ||||
344 | } | ||||
345 | } | ||||
346 | if (ccv_nnc_tensor_count(tensor->info) > 3) | ||||
347 | PRINT(CCV_CLI_VERBOSE, " ..")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf(" .."); fflush(stdout); } } while (0); | ||||
348 | #endif | ||||
349 | } else if (CCV_TENSOR_GET_MEMORY(tensor->info.type)((tensor->info.type) & 0x3) == CCV_TENSOR_CPU_MEMORY) { | ||||
350 | switch (tensor->info.datatype) | ||||
351 | { | ||||
352 | case CCV_16F: { | ||||
353 | float fp32[len]; | ||||
354 | ccv_half_precision_to_float((uint16_t*)tensor->data.f16, fp32, len); | ||||
355 | for (i = 0; i < len; i++) | ||||
356 | PRINT(CCV_CLI_VERBOSE, " %f", fp32[i])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf(" %f", fp32[i]); fflush(stdout); } } while (0); | ||||
357 | break; | ||||
358 | } | ||||
359 | case CCV_32F: | ||||
360 | for (i = 0; i < len; i++) | ||||
361 | PRINT(CCV_CLI_VERBOSE, " %f", tensor->data.f32[i])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf(" %f", tensor->data.f32[i]); fflush(stdout); } } while (0); | ||||
362 | break; | ||||
363 | case CCV_64F: | ||||
364 | for (i = 0; i < len; i++) | ||||
365 | PRINT(CCV_CLI_VERBOSE, " %f", tensor->data.f64[i])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf(" %f", tensor->data.f64[i]); fflush(stdout); } } while (0); | ||||
366 | break; | ||||
367 | case CCV_32S: | ||||
368 | for (i = 0; i < len; i++) | ||||
369 | PRINT(CCV_CLI_VERBOSE, " %d", tensor->data.i32[i])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf(" %d", tensor->data.i32[i]); fflush(stdout); } } while (0); | ||||
370 | break; | ||||
371 | case CCV_64S: | ||||
372 | for (i = 0; i < len; i++) | ||||
373 | PRINT(CCV_CLI_VERBOSE, " %lld", (long long)tensor->data.i64[i])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf(" %lld", (long long)tensor->data.i64[i]); fflush (stdout); } } while (0); | ||||
374 | break; | ||||
375 | case CCV_8U: | ||||
376 | for (i = 0; i < len; i++) | ||||
377 | PRINT(CCV_CLI_VERBOSE, " %d", (int)tensor->data.u8[i])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf(" %d", (int)tensor->data.u8[i]); fflush(stdout); } } while (0); | ||||
378 | break; | ||||
379 | } | ||||
380 | if (ccv_nnc_tensor_count(tensor->info) > 3) | ||||
381 | PRINT(CCV_CLI_VERBOSE, " ..")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf(" .."); fflush(stdout); } } while (0); | ||||
382 | } | ||||
383 | } | ||||
384 | |||||
385 | static co_decl(_ccv_nnc_graph_topsorted_run_coro, (ccv_nnc_graph_t* const graph, const int exec_idx, const ccv_nnc_graph_static_schedule_t* const schedule, ccv_nnc_graph_exec_info_t* const exec, ccv_nnc_tensor_tape_t* const tensor_tape, ccv_nnc_stream_context_t* const stream_context, const int flags))co_state_t _ccv_nnc_graph_topsorted_run_coro(co_routine_t* const _self, void* const _privates_); struct _ccv_nnc_graph_topsorted_run_coro_param_s { struct { ccv_nnc_graph_t* const graph;const int exec_idx;const ccv_nnc_graph_static_schedule_t* const schedule;ccv_nnc_graph_exec_info_t * const exec;ccv_nnc_tensor_tape_t* const tensor_tape;ccv_nnc_stream_context_t * const stream_context;const int flags;;;;;;;;;;;;;;;;;;;;;;; ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; } _co_params; }; size_t _ccv_nnc_graph_topsorted_run_coro_stack_size (void);; | ||||
386 | |||||
387 | static co_decl_task(_ccv_nnc_graph_exec_cases_of_coro, (ccv_nnc_graph_t* const graph, const int exec_idx, ccv_nnc_graph_exec_info_t* const exec, const ccv_nnc_graph_exec_schedule_t* const schd, ccv_nnc_tensor_t* const* const inputs, ccv_nnc_tensor_tape_t* const tensor_tape, ccv_nnc_stream_context_t* const stream_context, int flags), private(co_state_t _ccv_nnc_graph_exec_cases_of_coro(co_routine_t* const _self, void* const _privates_); struct _ccv_nnc_graph_exec_cases_of_coro_param_s { struct { ccv_nnc_graph_t* const graph;const int exec_idx;ccv_nnc_graph_exec_info_t * const exec;const ccv_nnc_graph_exec_schedule_t* const schd; ccv_nnc_tensor_t* const* const inputs;ccv_nnc_tensor_tape_t* const tensor_tape;ccv_nnc_stream_context_t* const stream_context;int flags;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; ;; } _co_params; }; size_t _ccv_nnc_graph_exec_cases_of_coro_stack_size (void); struct _ccv_nnc_graph_exec_cases_of_coro_private_s { struct _ccv_nnc_graph_exec_cases_of_coro_param_s _co_params; int ref ; ccv_nnc_graph_t* sub_graph; }; size_t _ccv_nnc_graph_exec_cases_of_coro_stack_size (void) { return sizeof(struct _ccv_nnc_graph_exec_cases_of_coro_private_s ); } co_state_t _ccv_nnc_graph_exec_cases_of_coro(co_routine_t * const _self_, void* const _privates_) { struct _private_s { struct _ccv_nnc_graph_exec_cases_of_coro_param_s _co_params; int ref; ccv_nnc_graph_t* sub_graph; }; switch (_self_->line ) { case 0: | ||||
388 | int ref;co_state_t _ccv_nnc_graph_exec_cases_of_coro(co_routine_t* const _self, void* const _privates_); struct _ccv_nnc_graph_exec_cases_of_coro_param_s { struct { ccv_nnc_graph_t* const graph;const int exec_idx;ccv_nnc_graph_exec_info_t * const exec;const ccv_nnc_graph_exec_schedule_t* const schd; ccv_nnc_tensor_t* const* const inputs;ccv_nnc_tensor_tape_t* const tensor_tape;ccv_nnc_stream_context_t* const stream_context;int flags;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; ;; } _co_params; }; size_t _ccv_nnc_graph_exec_cases_of_coro_stack_size (void); struct _ccv_nnc_graph_exec_cases_of_coro_private_s { struct _ccv_nnc_graph_exec_cases_of_coro_param_s _co_params; int ref ; ccv_nnc_graph_t* sub_graph; }; size_t _ccv_nnc_graph_exec_cases_of_coro_stack_size (void) { return sizeof(struct _ccv_nnc_graph_exec_cases_of_coro_private_s ); } co_state_t _ccv_nnc_graph_exec_cases_of_coro(co_routine_t * const _self_, void* const _privates_) { struct _private_s { struct _ccv_nnc_graph_exec_cases_of_coro_param_s _co_params; int ref; ccv_nnc_graph_t* sub_graph; }; switch (_self_->line ) { case 0: | ||||
389 | ccv_nnc_graph_t* sub_graph;co_state_t _ccv_nnc_graph_exec_cases_of_coro(co_routine_t* const _self, void* const _privates_); struct _ccv_nnc_graph_exec_cases_of_coro_param_s { struct { ccv_nnc_graph_t* const graph;const int exec_idx;ccv_nnc_graph_exec_info_t * const exec;const ccv_nnc_graph_exec_schedule_t* const schd; ccv_nnc_tensor_t* const* const inputs;ccv_nnc_tensor_tape_t* const tensor_tape;ccv_nnc_stream_context_t* const stream_context;int flags;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; ;; } _co_params; }; size_t _ccv_nnc_graph_exec_cases_of_coro_stack_size (void); struct _ccv_nnc_graph_exec_cases_of_coro_private_s { struct _ccv_nnc_graph_exec_cases_of_coro_param_s _co_params; int ref ; ccv_nnc_graph_t* sub_graph; }; size_t _ccv_nnc_graph_exec_cases_of_coro_stack_size (void) { return sizeof(struct _ccv_nnc_graph_exec_cases_of_coro_private_s ); } co_state_t _ccv_nnc_graph_exec_cases_of_coro(co_routine_t * const _self_, void* const _privates_) { struct _private_s { struct _ccv_nnc_graph_exec_cases_of_coro_param_s _co_params; int ref; ccv_nnc_graph_t* sub_graph; }; switch (_self_->line ) { case 0: | ||||
390 | ))co_state_t _ccv_nnc_graph_exec_cases_of_coro(co_routine_t* const _self, void* const _privates_); struct _ccv_nnc_graph_exec_cases_of_coro_param_s { struct { ccv_nnc_graph_t* const graph;const int exec_idx;ccv_nnc_graph_exec_info_t * const exec;const ccv_nnc_graph_exec_schedule_t* const schd; ccv_nnc_tensor_t* const* const inputs;ccv_nnc_tensor_tape_t* const tensor_tape;ccv_nnc_stream_context_t* const stream_context;int flags;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; ;; } _co_params; }; size_t _ccv_nnc_graph_exec_cases_of_coro_stack_size (void); struct _ccv_nnc_graph_exec_cases_of_coro_private_s { struct _ccv_nnc_graph_exec_cases_of_coro_param_s _co_params; int ref ; ccv_nnc_graph_t* sub_graph; }; size_t _ccv_nnc_graph_exec_cases_of_coro_stack_size (void) { return sizeof(struct _ccv_nnc_graph_exec_cases_of_coro_private_s ); } co_state_t _ccv_nnc_graph_exec_cases_of_coro(co_routine_t * const _self_, void* const _privates_) { struct _private_s { struct _ccv_nnc_graph_exec_cases_of_coro_param_s _co_params; int ref; ccv_nnc_graph_t* sub_graph; }; switch (_self_->line ) { case 0: { | ||||
391 | // Wait until this stream context is done. | ||||
392 | co_stream_await(CO_P(stream_context))do { if (!_co_stream_await(_self_, (((struct _private_s*)(_privates_ ))->_co_params._co_params.stream_context))) { return (co_state_t ){ 392, 0 }; } case 392: ; } while (0); | ||||
393 | if (CO_P(exec)(((struct _private_s*)(_privates_))->_co_params._co_params .exec)->cmd.cmd == CCV_NNC_GRAPH_FORWARD) | ||||
394 | { | ||||
395 | CO_V(ref)(((struct _private_s*)(_privates_))->ref) = CO_P(exec)(((struct _private_s*)(_privates_))->_co_params._co_params .exec)->case_of.offset + CO_P(exec)(((struct _private_s*)(_privates_))->_co_params._co_params .exec)->case_of.expr(CO_P(inputs)(((struct _private_s*)(_privates_))->_co_params._co_params .inputs), CO_P(exec)(((struct _private_s*)(_privates_))->_co_params._co_params .exec)->input_size, CO_P(exec)(((struct _private_s*)(_privates_))->_co_params._co_params .exec)->case_of.data); | ||||
396 | if (CO_P(tensor_tape)(((struct _private_s*)(_privates_))->_co_params._co_params .tensor_tape)) | ||||
397 | ccv_nnc_tensor_tape_set_numbering(CO_P(tensor_tape)(((struct _private_s*)(_privates_))->_co_params._co_params .tensor_tape), CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph), (ccv_nnc_graph_exec_t){ | ||||
398 | .d = CO_P(exec_idx)(((struct _private_s*)(_privates_))->_co_params._co_params .exec_idx), | ||||
399 | .graph = CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph), | ||||
400 | }, CO_V(ref)(((struct _private_s*)(_privates_))->ref)); | ||||
401 | } else { | ||||
402 | assert(CO_P(exec)->cmd.cmd == CCV_NNC_GRAPH_BACKWARD)((void) sizeof (((((struct _private_s*)(_privates_))->_co_params ._co_params.exec)->cmd.cmd == CCV_NNC_GRAPH_BACKWARD) ? 1 : 0), __extension__ ({ if ((((struct _private_s*)(_privates_)) ->_co_params._co_params.exec)->cmd.cmd == CCV_NNC_GRAPH_BACKWARD ) ; else __assert_fail ("CO_P(exec)->cmd.cmd == CCV_NNC_GRAPH_BACKWARD" , "ccv_nnc_graph_run.c", 402, __extension__ __PRETTY_FUNCTION__ ); })); | ||||
403 | assert(CO_P(tensor_tape))((void) sizeof (((((struct _private_s*)(_privates_))->_co_params ._co_params.tensor_tape)) ? 1 : 0), __extension__ ({ if ((((struct _private_s*)(_privates_))->_co_params._co_params.tensor_tape )) ; else __assert_fail ("CO_P(tensor_tape)", "ccv_nnc_graph_run.c" , 403, __extension__ __PRETTY_FUNCTION__); })); | ||||
404 | CO_V(ref)(((struct _private_s*)(_privates_))->ref) = ccv_nnc_tensor_tape_numbering(CO_P(tensor_tape)(((struct _private_s*)(_privates_))->_co_params._co_params .tensor_tape), CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph), (ccv_nnc_graph_exec_t){ | ||||
405 | .d = CO_P(exec_idx)(((struct _private_s*)(_privates_))->_co_params._co_params .exec_idx), | ||||
406 | .graph = CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph), | ||||
407 | }); | ||||
408 | } | ||||
409 | if (CO_V(ref)(((struct _private_s*)(_privates_))->ref) >= 0) | ||||
410 | { | ||||
411 | assert(CO_V(ref) < CO_P(exec)->graph_ref_size)((void) sizeof (((((struct _private_s*)(_privates_))->ref) < (((struct _private_s*)(_privates_))->_co_params._co_params .exec)->graph_ref_size) ? 1 : 0), __extension__ ({ if (((( struct _private_s*)(_privates_))->ref) < (((struct _private_s *)(_privates_))->_co_params._co_params.exec)->graph_ref_size ) ; else __assert_fail ("CO_V(ref) < CO_P(exec)->graph_ref_size" , "ccv_nnc_graph_run.c", 411, __extension__ __PRETTY_FUNCTION__ ); })); | ||||
412 | CO_V(sub_graph)(((struct _private_s*)(_privates_))->sub_graph) = *(ccv_nnc_graph_t**)ccv_array_get(CO_P(graph)->sub_graphs, CCV_NNC_GRAPH_REF(CO_P(exec))[CO_V(ref)] - 1)((void*)(((char*)(((((struct _private_s*)(_privates_))->_co_params ._co_params.graph)->sub_graphs)->data)) + (size_t)((((struct _private_s*)(_privates_))->_co_params._co_params.graph)-> sub_graphs)->rsize * (size_t)((((((struct _private_s*)(_privates_ ))->_co_params._co_params.exec))->_heap_graph_ref ? ((( (struct _private_s*)(_privates_))->_co_params._co_params.exec ))->_heap_graph_ref : ((((struct _private_s*)(_privates_)) ->_co_params._co_params.exec))->_inline_graph_ref)[(((struct _private_s*)(_privates_))->ref)] - 1))); | ||||
413 | assert(CO_P(schd)->stream_size == 1)((void) sizeof (((((struct _private_s*)(_privates_))->_co_params ._co_params.schd)->stream_size == 1) ? 1 : 0), __extension__ ({ if ((((struct _private_s*)(_privates_))->_co_params._co_params .schd)->stream_size == 1) ; else __assert_fail ("CO_P(schd)->stream_size == 1" , "ccv_nnc_graph_run.c", 413, __extension__ __PRETTY_FUNCTION__ ); })); | ||||
414 | assert(CO_P(graph)->streams[SCHEDULE_STREAMS(*CO_P(schd))[0]] == CO_V(sub_graph)->streams[0])((void) sizeof (((((struct _private_s*)(_privates_))->_co_params ._co_params.graph)->streams[((*(((struct _private_s*)(_privates_ ))->_co_params._co_params.schd)).stream_size <= 1 ? (*( ((struct _private_s*)(_privates_))->_co_params._co_params. schd))._inline_streams : (*(((struct _private_s*)(_privates_) )->_co_params._co_params.schd))._heap_streams)[0]] == (((struct _private_s*)(_privates_))->sub_graph)->streams[0]) ? 1 : 0), __extension__ ({ if ((((struct _private_s*)(_privates_ ))->_co_params._co_params.graph)->streams[((*(((struct _private_s *)(_privates_))->_co_params._co_params.schd)).stream_size <= 1 ? (*(((struct _private_s*)(_privates_))->_co_params._co_params .schd))._inline_streams : (*(((struct _private_s*)(_privates_ ))->_co_params._co_params.schd))._heap_streams)[0]] == ((( struct _private_s*)(_privates_))->sub_graph)->streams[0 ]) ; else __assert_fail ("CO_P(graph)->streams[SCHEDULE_STREAMS(*CO_P(schd))[0]] == CO_V(sub_graph)->streams[0]" , "ccv_nnc_graph_run.c", 414, __extension__ __PRETTY_FUNCTION__ ); })); | ||||
415 | co_apply(_ccv_nnc_graph_topsorted_run_coro, (CO_V(sub_graph), CO_P(exec_idx), CO_V(sub_graph)->default_schedule, CO_P(exec), CO_P(tensor_tape), CO_P(graph)->streams[SCHEDULE_STREAMS(*CO_P(schd))[0]], CO_P(flags)))do { _self_->callee = ({ co_routine_t* const task = malloc ((sizeof(co_routine_t) + _ccv_nnc_graph_topsorted_run_coro_stack_size ())); do { struct _ccv_nnc_graph_topsorted_run_coro_param_s params = { ._co_params = { (((struct _private_s*)(_privates_))-> sub_graph), (((struct _private_s*)(_privates_))->_co_params ._co_params.exec_idx), (((struct _private_s*)(_privates_))-> sub_graph)->default_schedule, (((struct _private_s*)(_privates_ ))->_co_params._co_params.exec), (((struct _private_s*)(_privates_ ))->_co_params._co_params.tensor_tape), (((struct _private_s *)(_privates_))->_co_params._co_params.graph)->streams[ ((*(((struct _private_s*)(_privates_))->_co_params._co_params .schd)).stream_size <= 1 ? (*(((struct _private_s*)(_privates_ ))->_co_params._co_params.schd))._inline_streams : (*(((struct _private_s*)(_privates_))->_co_params._co_params.schd))._heap_streams )[0]], (((struct _private_s*)(_privates_))->_co_params._co_params .flags) } }; task->fn = _ccv_nnc_graph_topsorted_run_coro; task->line = 0; task->done = 0; task->root = 0; task ->other_size = 0; task->notify_any = 0; task->others = 0; task->caller = 0; task->callee = 0; if (sizeof(params ) > 0) memcpy(task + 1, ¶ms, sizeof(params)); } while (0); task; }); _co_apply(_self_, _self_->callee); return ( co_state_t){ 415, 0 }; case 415: co_free(_self_->callee); _self_ ->callee = 0; } while (0); | ||||
416 | } | ||||
417 | _ccv_nnc_graph_exec_unwrap_phi(CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph), CO_P(exec)(((struct _private_s*)(_privates_))->_co_params._co_params .exec), CO_V(ref)(((struct _private_s*)(_privates_))->ref)); | ||||
418 | } co_end()default: return (co_state_t){ 418, 1 }; } } | ||||
419 | |||||
420 | typedef struct { | ||||
421 | ccv_nnc_graph_t* graph; | ||||
422 | const ccv_nnc_graph_exec_schedule_t* node; | ||||
423 | ccv_nnc_stream_context_t* stream; | ||||
424 | } ccv_nnc_graph_neighbor_context_discovery_t; | ||||
425 | |||||
426 | static ccv_nnc_stream_context_t* _ccv_nnc_graph_neighbor_context_discovery(const int device_id, void* const context) | ||||
427 | { | ||||
428 | const ccv_nnc_graph_neighbor_context_discovery_t* const discovery = (ccv_nnc_graph_neighbor_context_discovery_t*)context; | ||||
429 | if (CCV_STREAM_GET_DEVICE_ID(ccv_nnc_stream_context_type(discovery->stream))(((ccv_nnc_stream_context_type(discovery->stream)) & 0xfff00 ) >> 8) == device_id) | ||||
430 | return discovery->stream; | ||||
431 | ccv_nnc_graph_t* const graph = discovery->graph; | ||||
432 | const ccv_nnc_graph_exec_schedule_t* const node = discovery->node; | ||||
433 | int i; | ||||
434 | // First try to find in other streams of the same node. | ||||
435 | for (i = 0; i < node->stream_size; i++) | ||||
436 | { | ||||
437 | ccv_nnc_stream_context_t* const stream = graph->streams[SCHEDULE_STREAMS(*node)((*node).stream_size <= 1 ? (*node)._inline_streams : (*node )._heap_streams)[i]]; | ||||
438 | if (CCV_STREAM_GET_DEVICE_ID(ccv_nnc_stream_context_type(stream))(((ccv_nnc_stream_context_type(stream)) & 0xfff00) >> 8) == device_id) | ||||
439 | return stream; | ||||
440 | } | ||||
441 | // If cannot find, try to find in all the wait streams. | ||||
442 | for (i = 0; i < node->wait_size; i++) | ||||
443 | { | ||||
444 | ccv_nnc_stream_context_t* stream_context = ccv_nnc_stream_signal_get_emitter(graph->signals[node->waits[i]]); | ||||
445 | if (stream_context && CCV_STREAM_GET_DEVICE_ID(ccv_nnc_stream_context_type(stream_context))(((ccv_nnc_stream_context_type(stream_context)) & 0xfff00 ) >> 8) == device_id) | ||||
446 | return stream_context; | ||||
447 | } | ||||
448 | return 0; | ||||
449 | } | ||||
450 | |||||
451 | static co_routine_t* _ccv_nnc_graph_exec_run_task(ccv_nnc_graph_t* const graph, ccv_nnc_graph_exec_info_t* const node, const ccv_nnc_graph_exec_schedule_t* const schd, const int idx, ccv_nnc_tensor_tape_t* const tensor_tape, const int flags) | ||||
452 | { | ||||
453 | _ccv_nnc_graph_exec_unwrap_io(graph, node); | ||||
454 | ccv_nnc_tensor_t** inputs = node->inputs; | ||||
455 | ccv_nnc_tensor_t** outputs = inputs ? inputs + node->input_size : 0; | ||||
456 | if (tensor_tape) | ||||
457 | ccv_nnc_tensor_tape_io(tensor_tape, graph, node->input_flags, inputs, node->input_size, node->output_flags, outputs, node->output_size); | ||||
458 | /* Broadcast the updates to all subscribed references for input / output, even though at th | ||||
459 | * time output is not written yet, propagate pointer change is still valid. */ | ||||
460 | _ccv_nnc_graph_exec_begin_synchronize_multiviews(graph, node); | ||||
461 | if (node->cmd.cmd == CCV_NNC_GRAPH_FORWARD || node->cmd.cmd == CCV_NNC_GRAPH_BACKWARD) | ||||
462 | { | ||||
463 | if (node->flags & CCV_NNC_GRAPH_EXEC_CASE_OF) | ||||
464 | { | ||||
465 | ccv_nnc_stream_context_t* const node_stream = graph->streams[SCHEDULE_STREAMS(*schd)((*schd).stream_size <= 1 ? (*schd)._inline_streams : (*schd )._heap_streams)[0]]; | ||||
466 | return co_new(_ccv_nnc_graph_exec_cases_of_coro, (graph, idx, node, schd, inputs, tensor_tape, node_stream, flags))({ co_routine_t* const task = malloc((sizeof(co_routine_t) + _ccv_nnc_graph_exec_cases_of_coro_stack_size ())); do { struct _ccv_nnc_graph_exec_cases_of_coro_param_s params = { ._co_params = { graph, idx, node, schd, inputs, tensor_tape , node_stream, flags } }; task->fn = _ccv_nnc_graph_exec_cases_of_coro ; task->line = 0; task->done = 0; task->root = 0; task ->other_size = 0; task->notify_any = 0; task->others = 0; task->caller = 0; task->callee = 0; if (sizeof(params ) > 0) memcpy(task + 1, ¶ms, sizeof(params)); } while (0); task; }); | ||||
467 | } else if (node->flags & CCV_NNC_GRAPH_EXEC_P_WHILE) { | ||||
468 | ccv_nnc_graph_t* sub_graph = *(ccv_nnc_graph_t**)ccv_array_get(graph->sub_graphs, CCV_NNC_GRAPH_REF(node)[0] - 1)((void*)(((char*)((graph->sub_graphs)->data)) + (size_t )(graph->sub_graphs)->rsize * (size_t)(((node)->_heap_graph_ref ? (node)->_heap_graph_ref : (node)->_inline_graph_ref) [0] - 1))); | ||||
469 | assert(graph->streams[SCHEDULE_STREAMS(*schd)[0]] == sub_graph->streams[0])((void) sizeof ((graph->streams[((*schd).stream_size <= 1 ? (*schd)._inline_streams : (*schd)._heap_streams)[0]] == sub_graph ->streams[0]) ? 1 : 0), __extension__ ({ if (graph->streams [((*schd).stream_size <= 1 ? (*schd)._inline_streams : (*schd )._heap_streams)[0]] == sub_graph->streams[0]) ; else __assert_fail ("graph->streams[SCHEDULE_STREAMS(*schd)[0]] == sub_graph->streams[0]" , "ccv_nnc_graph_run.c", 469, __extension__ __PRETTY_FUNCTION__ ); })); | ||||
470 | return co_new(_ccv_nnc_graph_topsorted_run_coro, (sub_graph, idx, sub_graph->default_schedule, node, tensor_tape, graph->streams[SCHEDULE_STREAMS(*schd)[0]], flags))({ co_routine_t* const task = malloc((sizeof(co_routine_t) + _ccv_nnc_graph_topsorted_run_coro_stack_size ())); do { struct _ccv_nnc_graph_topsorted_run_coro_param_s params = { ._co_params = { sub_graph, idx, sub_graph->default_schedule , node, tensor_tape, graph->streams[((*schd).stream_size <= 1 ? (*schd)._inline_streams : (*schd)._heap_streams)[0]], flags } }; task->fn = _ccv_nnc_graph_topsorted_run_coro; task-> line = 0; task->done = 0; task->root = 0; task->other_size = 0; task->notify_any = 0; task->others = 0; task-> caller = 0; task->callee = 0; if (sizeof(params) > 0) memcpy (task + 1, ¶ms, sizeof(params)); } while (0); task; } ); | ||||
471 | } | ||||
472 | } else { | ||||
473 | PRINT(CCV_CLI_INFO, "%s [%d]: [%d] -> [%d] (%d)\n", ccv_nnc_cmd_name(node->cmd.cmd), idx, node->input_size, node->output_size, SCHEDULE_STREAMS(*schd)[0])do { if ((CCV_CLI_INFO & ccv_cli_get_output_levels())) { printf ("%s [%d]: [%d] -> [%d] (%d)\n", ccv_nnc_cmd_name(node-> cmd.cmd), idx, node->input_size, node->output_size, ((* schd).stream_size <= 1 ? (*schd)._inline_streams : (*schd) ._heap_streams)[0]); fflush(stdout); } } while (0); | ||||
474 | int i, j; | ||||
475 | int flag = 0; | ||||
476 | for (i = 0; i < schd->stream_size; i++) | ||||
477 | { | ||||
478 | ccv_nnc_stream_context_t* const stream = graph->streams[SCHEDULE_STREAMS(*schd)((*schd).stream_size <= 1 ? (*schd)._inline_streams : (*schd )._heap_streams)[i]]; | ||||
479 | for (j = 0; j < schd->wait_size; j++) | ||||
480 | { | ||||
481 | ccv_nnc_stream_context_wait_signal(stream, graph->signals[schd->waits[j]]); | ||||
482 | if (!flag) | ||||
483 | { | ||||
484 | PRINT(CCV_CLI_INFO, "Wait: (%d, %d)", SCHEDULE_STREAMS(*schd)[i], schd->waits[j])do { if ((CCV_CLI_INFO & ccv_cli_get_output_levels())) { printf ("Wait: (%d, %d)", ((*schd).stream_size <= 1 ? (*schd)._inline_streams : (*schd)._heap_streams)[i], schd->waits[j]); fflush(stdout ); } } while (0); | ||||
485 | flag = 1; | ||||
486 | } else | ||||
487 | PRINT(CCV_CLI_INFO, ", (%d, %d)", SCHEDULE_STREAMS(*schd)[i], schd->waits[j])do { if ((CCV_CLI_INFO & ccv_cli_get_output_levels())) { printf (", (%d, %d)", ((*schd).stream_size <= 1 ? (*schd)._inline_streams : (*schd)._heap_streams)[i], schd->waits[j]); fflush(stdout ); } } while (0); | ||||
488 | } | ||||
489 | } | ||||
490 | if (flag) | ||||
491 | PRINT(CCV_CLI_INFO, "\n")do { if ((CCV_CLI_INFO & ccv_cli_get_output_levels())) { printf ("\n"); fflush(stdout); } } while (0); | ||||
492 | for (i = 0; i < node->input_size; i++) | ||||
493 | { | ||||
494 | PRINT(CCV_CLI_INFO, "|-> %d. %p (%p:%d)", i + 1, inputs[i], (inputs[i] ? inputs[i]->data.u8 : 0), (inputs[i] ? CCV_TENSOR_GET_DEVICE_ID(inputs[i]->info.type) : -1))do { if ((CCV_CLI_INFO & ccv_cli_get_output_levels())) { printf ("|-> %d. %p (%p:%d)", i + 1, inputs[i], (inputs[i] ? inputs [i]->data.u8 : 0), (inputs[i] ? (((inputs[i]->info.type ) & 0xfff00) >> 8) : -1)); fflush(stdout); } } while (0); | ||||
495 | if (inputs[i] && CCV_CLI_OUTPUT_LEVEL_IS(CCV_CLI_INFO)(CCV_CLI_INFO & ccv_cli_get_output_levels())) | ||||
496 | ccv_nnc_print_tensor_info(inputs[i]); | ||||
497 | PRINT(CCV_CLI_INFO, "\n")do { if ((CCV_CLI_INFO & ccv_cli_get_output_levels())) { printf ("\n"); fflush(stdout); } } while (0); | ||||
498 | } | ||||
499 | for (i = 0; i < node->output_size; i++) | ||||
500 | { | ||||
501 | PRINT(CCV_CLI_INFO, "|<- %d. %p (%p:%d)", i + 1, outputs[i], (outputs[i] ? outputs[i]->data.u8 : 0), (outputs[i] ? CCV_TENSOR_GET_DEVICE_ID(outputs[i]->info.type) : -1))do { if ((CCV_CLI_INFO & ccv_cli_get_output_levels())) { printf ("|<- %d. %p (%p:%d)", i + 1, outputs[i], (outputs[i] ? outputs [i]->data.u8 : 0), (outputs[i] ? (((outputs[i]->info.type ) & 0xfff00) >> 8) : -1)); fflush(stdout); } } while (0); | ||||
502 | if (outputs[i] && CCV_CLI_OUTPUT_LEVEL_IS(CCV_CLI_INFO)(CCV_CLI_INFO & ccv_cli_get_output_levels())) | ||||
503 | ccv_nnc_print_tensor_shape(outputs[i]); | ||||
504 | PRINT(CCV_CLI_INFO, "\n")do { if ((CCV_CLI_INFO & ccv_cli_get_output_levels())) { printf ("\n"); fflush(stdout); } } while (0); | ||||
505 | } | ||||
506 | ccv_nnc_stream_context_t* const node_stream = graph->streams[SCHEDULE_STREAMS(*schd)((*schd).stream_size <= 1 ? (*schd)._inline_streams : (*schd )._heap_streams)[0]]; | ||||
507 | ccv_nnc_graph_neighbor_context_discovery_t discovery_context = { | ||||
508 | .graph = graph, | ||||
509 | .node = schd, | ||||
510 | .stream = node_stream | ||||
511 | }; | ||||
512 | ccv_nnc_stream_context_set_neighbor_discovery(node_stream, _ccv_nnc_graph_neighbor_context_discovery, &discovery_context); | ||||
513 | const int status = ccv_nnc_cmd_exec(node->cmd, node->hint, flags, inputs, node->input_size, outputs, node->output_size, node_stream); | ||||
514 | if (status != 0) | ||||
515 | PRINT(CCV_CLI_INFO, "Invalid Status: %d\n", status)do { if ((CCV_CLI_INFO & ccv_cli_get_output_levels())) { printf ("Invalid Status: %d\n", status); fflush(stdout); } } while ( 0); | ||||
516 | if (CCV_CLI_OUTPUT_LEVEL_IS(CCV_CLI_VERBOSE)(CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) | ||||
517 | { | ||||
518 | for (i = 0; i < node->output_size; i++) | ||||
519 | { | ||||
520 | PRINT(CCV_CLI_VERBOSE, "POST: |<- %d. %p (%p:%d)", i + 1, outputs[i], (outputs[i] ? outputs[i]->data.u8 : 0), (outputs[i] ? CCV_TENSOR_GET_DEVICE_ID(outputs[i]->info.type) : -1))do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("POST: |<- %d. %p (%p:%d)", i + 1, outputs[i], ( outputs[i] ? outputs[i]->data.u8 : 0), (outputs[i] ? (((outputs [i]->info.type) & 0xfff00) >> 8) : -1)); fflush( stdout); } } while (0); | ||||
521 | if (outputs[i]) | ||||
522 | ccv_nnc_print_tensor_info(outputs[i]); | ||||
523 | PRINT(CCV_CLI_VERBOSE, "\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels())) { printf("\n"); fflush(stdout); } } while (0); | ||||
524 | } | ||||
525 | } | ||||
526 | flag = 0; | ||||
527 | for (i = 0; i < schd->stream_size; i++) | ||||
528 | if (SCHEDULE_SIGNALS(*schd)((*schd).stream_size <= 1 ? (*schd)._inline_signals : (*schd )._heap_signals)[i] >= 0) | ||||
529 | { | ||||
530 | ccv_nnc_stream_context_t* const stream = graph->streams[SCHEDULE_STREAMS(*schd)((*schd).stream_size <= 1 ? (*schd)._inline_streams : (*schd )._heap_streams)[i]]; | ||||
531 | ccv_nnc_stream_context_emit_signal(stream, graph->signals[SCHEDULE_SIGNALS(*schd)((*schd).stream_size <= 1 ? (*schd)._inline_signals : (*schd )._heap_signals)[i]]); | ||||
532 | if (!flag) | ||||
533 | { | ||||
534 | PRINT(CCV_CLI_INFO, "Emit: (%d, %d)", SCHEDULE_STREAMS(*schd)[i], SCHEDULE_SIGNALS(*schd)[i])do { if ((CCV_CLI_INFO & ccv_cli_get_output_levels())) { printf ("Emit: (%d, %d)", ((*schd).stream_size <= 1 ? (*schd)._inline_streams : (*schd)._heap_streams)[i], ((*schd).stream_size <= 1 ? ( *schd)._inline_signals : (*schd)._heap_signals)[i]); fflush(stdout ); } } while (0); | ||||
535 | flag = 1; | ||||
536 | } else | ||||
537 | PRINT(CCV_CLI_INFO, ", (%d, %d)", SCHEDULE_STREAMS(*schd)[i], SCHEDULE_SIGNALS(*schd)[i])do { if ((CCV_CLI_INFO & ccv_cli_get_output_levels())) { printf (", (%d, %d)", ((*schd).stream_size <= 1 ? (*schd)._inline_streams : (*schd)._heap_streams)[i], ((*schd).stream_size <= 1 ? ( *schd)._inline_signals : (*schd)._heap_signals)[i]); fflush(stdout ); } } while (0); | ||||
538 | } | ||||
539 | if (flag) | ||||
540 | PRINT(CCV_CLI_INFO, "\n")do { if ((CCV_CLI_INFO & ccv_cli_get_output_levels())) { printf ("\n"); fflush(stdout); } } while (0); | ||||
541 | } | ||||
542 | return 0; | ||||
543 | } | ||||
544 | |||||
545 | static void _ccv_nnc_graph_mark_outgoing_streams_blocked_by_task(ccv_nnc_graph_t* const graph, const ccv_nnc_graph_exec_schedule_t* const schd_info, ccv_nnc_graph_exec_info_t* const node, co_routine_t* const task) | ||||
546 | { | ||||
547 | int i, j; | ||||
548 | if (node->outgoings) | ||||
549 | for (i = 0; i < node->outgoings->rnum; i++) | ||||
550 | { | ||||
551 | const int outgoing_idx = *(int*)ccv_array_get(node->outgoings, i)((void*)(((char*)((node->outgoings)->data)) + (size_t)( node->outgoings)->rsize * (size_t)(i))); | ||||
552 | const ccv_nnc_graph_exec_schedule_t* const outgoing_schd = schd_info + outgoing_idx; | ||||
553 | // An outgoing stream can be blocked by multiple other tasks from other streams. But it is OK, | ||||
554 | // because on next round of execution, that one will be marked as blocked again. | ||||
555 | for (j = 0; j < outgoing_schd->stream_size; j++) | ||||
556 | graph->block_stream_tasks[SCHEDULE_STREAMS(*outgoing_schd)((*outgoing_schd).stream_size <= 1 ? (*outgoing_schd)._inline_streams : (*outgoing_schd)._heap_streams)[j]] = task; | ||||
557 | } | ||||
558 | } | ||||
559 | |||||
560 | static co_decl_task(_ccv_nnc_graph_wait_any_sub_tasks, (ccv_nnc_graph_t* const graph, co_routine_t* const* const sub_tasks, const int sub_task_size, const ccv_nnc_graph_exec_schedule_t* const schd_info, const int* const pending_nodes, const int pending_node_size), private(co_state_t _ccv_nnc_graph_wait_any_sub_tasks(co_routine_t* const _self, void* const _privates_); struct _ccv_nnc_graph_wait_any_sub_tasks_param_s { struct { ccv_nnc_graph_t* const graph;co_routine_t* const* const sub_tasks;const int sub_task_size;const ccv_nnc_graph_exec_schedule_t * const schd_info;const int* const pending_nodes;const int pending_node_size ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; } _co_params; }; size_t _ccv_nnc_graph_wait_any_sub_tasks_stack_size (void); struct _ccv_nnc_graph_wait_any_sub_tasks_private_s { struct _ccv_nnc_graph_wait_any_sub_tasks_param_s _co_params; }; size_t _ccv_nnc_graph_wait_any_sub_tasks_stack_size(void) { return sizeof (struct _ccv_nnc_graph_wait_any_sub_tasks_private_s); } co_state_t _ccv_nnc_graph_wait_any_sub_tasks(co_routine_t* const _self_ , void* const _privates_) { struct _private_s { struct _ccv_nnc_graph_wait_any_sub_tasks_param_s _co_params; }; switch (_self_->line) { case 0: | ||||
561 | ))co_state_t _ccv_nnc_graph_wait_any_sub_tasks(co_routine_t* const _self, void* const _privates_); struct _ccv_nnc_graph_wait_any_sub_tasks_param_s { struct { ccv_nnc_graph_t* const graph;co_routine_t* const* const sub_tasks;const int sub_task_size;const ccv_nnc_graph_exec_schedule_t * const schd_info;const int* const pending_nodes;const int pending_node_size ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; } _co_params; }; size_t _ccv_nnc_graph_wait_any_sub_tasks_stack_size (void); struct _ccv_nnc_graph_wait_any_sub_tasks_private_s { struct _ccv_nnc_graph_wait_any_sub_tasks_param_s _co_params; }; size_t _ccv_nnc_graph_wait_any_sub_tasks_stack_size(void) { return sizeof (struct _ccv_nnc_graph_wait_any_sub_tasks_private_s); } co_state_t _ccv_nnc_graph_wait_any_sub_tasks(co_routine_t* const _self_ , void* const _privates_) { struct _private_s { struct _ccv_nnc_graph_wait_any_sub_tasks_param_s _co_params; }; switch (_self_->line) { case 0: { | ||||
562 | assert(CO_P(sub_task_size) > 0)((void) sizeof (((((struct _private_s*)(_privates_))->_co_params ._co_params.sub_task_size) > 0) ? 1 : 0), __extension__ ({ if ((((struct _private_s*)(_privates_))->_co_params._co_params .sub_task_size) > 0) ; else __assert_fail ("CO_P(sub_task_size) > 0" , "ccv_nnc_graph_run.c", 562, __extension__ __PRETTY_FUNCTION__ ); })); | ||||
563 | co_await_any(CO_P(sub_tasks), CO_P(sub_task_size))do { if (!_co_await_any(_self_, (((struct _private_s*)(_privates_ ))->_co_params._co_params.sub_tasks), (((struct _private_s *)(_privates_))->_co_params._co_params.sub_task_size))) { return (co_state_t){ 563, 0 }; } case 563: ; } while (0); | ||||
564 | // This is not good, these local variables need to be in the private section. | ||||
565 | // I got away with it because there is no yield or resume or apply or any after await above. | ||||
566 | int i, j, k; | ||||
567 | for (i = 0; i < CO_P(sub_task_size)(((struct _private_s*)(_privates_))->_co_params._co_params .sub_task_size); i++) | ||||
568 | if (co_is_done(CO_P(sub_tasks)(((struct _private_s*)(_privates_))->_co_params._co_params .sub_tasks)[i])) | ||||
569 | { | ||||
570 | for (j = 0; j < CO_P(pending_node_size)(((struct _private_s*)(_privates_))->_co_params._co_params .pending_node_size); j++) | ||||
571 | { | ||||
572 | const ccv_nnc_graph_exec_schedule_t* const node = CO_P(schd_info)(((struct _private_s*)(_privates_))->_co_params._co_params .schd_info) + CO_P(pending_nodes)(((struct _private_s*)(_privates_))->_co_params._co_params .pending_nodes)[j]; | ||||
573 | for (k = 0; k < node->stream_size; k++) | ||||
574 | if (CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph)->block_stream_tasks[SCHEDULE_STREAMS(*node)((*node).stream_size <= 1 ? (*node)._inline_streams : (*node )._heap_streams)[k]] == CO_P(sub_tasks)(((struct _private_s*)(_privates_))->_co_params._co_params .sub_tasks)[i]) | ||||
575 | CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph)->block_stream_tasks[SCHEDULE_STREAMS(*node)((*node).stream_size <= 1 ? (*node)._inline_streams : (*node )._heap_streams)[k]] = 0; | ||||
576 | } | ||||
577 | co_free(CO_P(sub_tasks)(((struct _private_s*)(_privates_))->_co_params._co_params .sub_tasks)[i]); | ||||
578 | } | ||||
579 | } co_end()default: return (co_state_t){ 579, 1 }; } } | ||||
580 | |||||
581 | static co_decl_task(_ccv_nnc_graph_exec_run_loop, (ccv_nnc_graph_t* const graph, ccv_nnc_graph_exec_info_t* const exec_info, const ccv_nnc_graph_exec_schedule_t* const schd_info, const int* const psort, const int start_index, const int exec_info_size, ccv_nnc_tensor_tape_t* const tensor_tape, const int flags), private(co_state_t _ccv_nnc_graph_exec_run_loop(co_routine_t* const _self , void* const _privates_); struct _ccv_nnc_graph_exec_run_loop_param_s { struct { ccv_nnc_graph_t* const graph;ccv_nnc_graph_exec_info_t * const exec_info;const ccv_nnc_graph_exec_schedule_t* const schd_info ;const int* const psort;const int start_index;const int exec_info_size ;ccv_nnc_tensor_tape_t* const tensor_tape;const int flags;;;; ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; } _co_params ; }; size_t _ccv_nnc_graph_exec_run_loop_stack_size(void); struct _ccv_nnc_graph_exec_run_loop_private_s { struct _ccv_nnc_graph_exec_run_loop_param_s _co_params; int i, p, q; int sub_task_size; co_routine_t** sub_tasks ; int* pending_nodes[2]; int pending_node_size[2]; int idx; ccv_nnc_graph_exec_info_t * node; const ccv_nnc_graph_exec_schedule_t* schd; co_routine_t * task; }; size_t _ccv_nnc_graph_exec_run_loop_stack_size(void ) { return sizeof(struct _ccv_nnc_graph_exec_run_loop_private_s ); } co_state_t _ccv_nnc_graph_exec_run_loop(co_routine_t* const _self_, void* const _privates_) { struct _private_s { struct _ccv_nnc_graph_exec_run_loop_param_s _co_params; int i, p, q ; int sub_task_size; co_routine_t** sub_tasks; int* pending_nodes [2]; int pending_node_size[2]; int idx; ccv_nnc_graph_exec_info_t * node; const ccv_nnc_graph_exec_schedule_t* schd; co_routine_t * task; }; switch (_self_->line) { case 0: | ||||
582 | int i, p, q;co_state_t _ccv_nnc_graph_exec_run_loop(co_routine_t* const _self , void* const _privates_); struct _ccv_nnc_graph_exec_run_loop_param_s { struct { ccv_nnc_graph_t* const graph;ccv_nnc_graph_exec_info_t * const exec_info;const ccv_nnc_graph_exec_schedule_t* const schd_info ;const int* const psort;const int start_index;const int exec_info_size ;ccv_nnc_tensor_tape_t* const tensor_tape;const int flags;;;; ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; } _co_params ; }; size_t _ccv_nnc_graph_exec_run_loop_stack_size(void); struct _ccv_nnc_graph_exec_run_loop_private_s { struct _ccv_nnc_graph_exec_run_loop_param_s _co_params; int i, p, q; int sub_task_size; co_routine_t** sub_tasks ; int* pending_nodes[2]; int pending_node_size[2]; int idx; ccv_nnc_graph_exec_info_t * node; const ccv_nnc_graph_exec_schedule_t* schd; co_routine_t * task; }; size_t _ccv_nnc_graph_exec_run_loop_stack_size(void ) { return sizeof(struct _ccv_nnc_graph_exec_run_loop_private_s ); } co_state_t _ccv_nnc_graph_exec_run_loop(co_routine_t* const _self_, void* const _privates_) { struct _private_s { struct _ccv_nnc_graph_exec_run_loop_param_s _co_params; int i, p, q ; int sub_task_size; co_routine_t** sub_tasks; int* pending_nodes [2]; int pending_node_size[2]; int idx; ccv_nnc_graph_exec_info_t * node; const ccv_nnc_graph_exec_schedule_t* schd; co_routine_t * task; }; switch (_self_->line) { case 0: | ||||
583 | int sub_task_size;co_state_t _ccv_nnc_graph_exec_run_loop(co_routine_t* const _self , void* const _privates_); struct _ccv_nnc_graph_exec_run_loop_param_s { struct { ccv_nnc_graph_t* const graph;ccv_nnc_graph_exec_info_t * const exec_info;const ccv_nnc_graph_exec_schedule_t* const schd_info ;const int* const psort;const int start_index;const int exec_info_size ;ccv_nnc_tensor_tape_t* const tensor_tape;const int flags;;;; ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; } _co_params ; }; size_t _ccv_nnc_graph_exec_run_loop_stack_size(void); struct _ccv_nnc_graph_exec_run_loop_private_s { struct _ccv_nnc_graph_exec_run_loop_param_s _co_params; int i, p, q; int sub_task_size; co_routine_t** sub_tasks ; int* pending_nodes[2]; int pending_node_size[2]; int idx; ccv_nnc_graph_exec_info_t * node; const ccv_nnc_graph_exec_schedule_t* schd; co_routine_t * task; }; size_t _ccv_nnc_graph_exec_run_loop_stack_size(void ) { return sizeof(struct _ccv_nnc_graph_exec_run_loop_private_s ); } co_state_t _ccv_nnc_graph_exec_run_loop(co_routine_t* const _self_, void* const _privates_) { struct _private_s { struct _ccv_nnc_graph_exec_run_loop_param_s _co_params; int i, p, q ; int sub_task_size; co_routine_t** sub_tasks; int* pending_nodes [2]; int pending_node_size[2]; int idx; ccv_nnc_graph_exec_info_t * node; const ccv_nnc_graph_exec_schedule_t* schd; co_routine_t * task; }; switch (_self_->line) { case 0: | ||||
584 | co_routine_t** sub_tasks;co_state_t _ccv_nnc_graph_exec_run_loop(co_routine_t* const _self , void* const _privates_); struct _ccv_nnc_graph_exec_run_loop_param_s { struct { ccv_nnc_graph_t* const graph;ccv_nnc_graph_exec_info_t * const exec_info;const ccv_nnc_graph_exec_schedule_t* const schd_info ;const int* const psort;const int start_index;const int exec_info_size ;ccv_nnc_tensor_tape_t* const tensor_tape;const int flags;;;; ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; } _co_params ; }; size_t _ccv_nnc_graph_exec_run_loop_stack_size(void); struct _ccv_nnc_graph_exec_run_loop_private_s { struct _ccv_nnc_graph_exec_run_loop_param_s _co_params; int i, p, q; int sub_task_size; co_routine_t** sub_tasks ; int* pending_nodes[2]; int pending_node_size[2]; int idx; ccv_nnc_graph_exec_info_t * node; const ccv_nnc_graph_exec_schedule_t* schd; co_routine_t * task; }; size_t _ccv_nnc_graph_exec_run_loop_stack_size(void ) { return sizeof(struct _ccv_nnc_graph_exec_run_loop_private_s ); } co_state_t _ccv_nnc_graph_exec_run_loop(co_routine_t* const _self_, void* const _privates_) { struct _private_s { struct _ccv_nnc_graph_exec_run_loop_param_s _co_params; int i, p, q ; int sub_task_size; co_routine_t** sub_tasks; int* pending_nodes [2]; int pending_node_size[2]; int idx; ccv_nnc_graph_exec_info_t * node; const ccv_nnc_graph_exec_schedule_t* schd; co_routine_t * task; }; switch (_self_->line) { case 0: | ||||
585 | int* pending_nodes[2];co_state_t _ccv_nnc_graph_exec_run_loop(co_routine_t* const _self , void* const _privates_); struct _ccv_nnc_graph_exec_run_loop_param_s { struct { ccv_nnc_graph_t* const graph;ccv_nnc_graph_exec_info_t * const exec_info;const ccv_nnc_graph_exec_schedule_t* const schd_info ;const int* const psort;const int start_index;const int exec_info_size ;ccv_nnc_tensor_tape_t* const tensor_tape;const int flags;;;; ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; } _co_params ; }; size_t _ccv_nnc_graph_exec_run_loop_stack_size(void); struct _ccv_nnc_graph_exec_run_loop_private_s { struct _ccv_nnc_graph_exec_run_loop_param_s _co_params; int i, p, q; int sub_task_size; co_routine_t** sub_tasks ; int* pending_nodes[2]; int pending_node_size[2]; int idx; ccv_nnc_graph_exec_info_t * node; const ccv_nnc_graph_exec_schedule_t* schd; co_routine_t * task; }; size_t _ccv_nnc_graph_exec_run_loop_stack_size(void ) { return sizeof(struct _ccv_nnc_graph_exec_run_loop_private_s ); } co_state_t _ccv_nnc_graph_exec_run_loop(co_routine_t* const _self_, void* const _privates_) { struct _private_s { struct _ccv_nnc_graph_exec_run_loop_param_s _co_params; int i, p, q ; int sub_task_size; co_routine_t** sub_tasks; int* pending_nodes [2]; int pending_node_size[2]; int idx; ccv_nnc_graph_exec_info_t * node; const ccv_nnc_graph_exec_schedule_t* schd; co_routine_t * task; }; switch (_self_->line) { case 0: | ||||
586 | int pending_node_size[2];co_state_t _ccv_nnc_graph_exec_run_loop(co_routine_t* const _self , void* const _privates_); struct _ccv_nnc_graph_exec_run_loop_param_s { struct { ccv_nnc_graph_t* const graph;ccv_nnc_graph_exec_info_t * const exec_info;const ccv_nnc_graph_exec_schedule_t* const schd_info ;const int* const psort;const int start_index;const int exec_info_size ;ccv_nnc_tensor_tape_t* const tensor_tape;const int flags;;;; ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; } _co_params ; }; size_t _ccv_nnc_graph_exec_run_loop_stack_size(void); struct _ccv_nnc_graph_exec_run_loop_private_s { struct _ccv_nnc_graph_exec_run_loop_param_s _co_params; int i, p, q; int sub_task_size; co_routine_t** sub_tasks ; int* pending_nodes[2]; int pending_node_size[2]; int idx; ccv_nnc_graph_exec_info_t * node; const ccv_nnc_graph_exec_schedule_t* schd; co_routine_t * task; }; size_t _ccv_nnc_graph_exec_run_loop_stack_size(void ) { return sizeof(struct _ccv_nnc_graph_exec_run_loop_private_s ); } co_state_t _ccv_nnc_graph_exec_run_loop(co_routine_t* const _self_, void* const _privates_) { struct _private_s { struct _ccv_nnc_graph_exec_run_loop_param_s _co_params; int i, p, q ; int sub_task_size; co_routine_t** sub_tasks; int* pending_nodes [2]; int pending_node_size[2]; int idx; ccv_nnc_graph_exec_info_t * node; const ccv_nnc_graph_exec_schedule_t* schd; co_routine_t * task; }; switch (_self_->line) { case 0: | ||||
587 | int idx;co_state_t _ccv_nnc_graph_exec_run_loop(co_routine_t* const _self , void* const _privates_); struct _ccv_nnc_graph_exec_run_loop_param_s { struct { ccv_nnc_graph_t* const graph;ccv_nnc_graph_exec_info_t * const exec_info;const ccv_nnc_graph_exec_schedule_t* const schd_info ;const int* const psort;const int start_index;const int exec_info_size ;ccv_nnc_tensor_tape_t* const tensor_tape;const int flags;;;; ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; } _co_params ; }; size_t _ccv_nnc_graph_exec_run_loop_stack_size(void); struct _ccv_nnc_graph_exec_run_loop_private_s { struct _ccv_nnc_graph_exec_run_loop_param_s _co_params; int i, p, q; int sub_task_size; co_routine_t** sub_tasks ; int* pending_nodes[2]; int pending_node_size[2]; int idx; ccv_nnc_graph_exec_info_t * node; const ccv_nnc_graph_exec_schedule_t* schd; co_routine_t * task; }; size_t _ccv_nnc_graph_exec_run_loop_stack_size(void ) { return sizeof(struct _ccv_nnc_graph_exec_run_loop_private_s ); } co_state_t _ccv_nnc_graph_exec_run_loop(co_routine_t* const _self_, void* const _privates_) { struct _private_s { struct _ccv_nnc_graph_exec_run_loop_param_s _co_params; int i, p, q ; int sub_task_size; co_routine_t** sub_tasks; int* pending_nodes [2]; int pending_node_size[2]; int idx; ccv_nnc_graph_exec_info_t * node; const ccv_nnc_graph_exec_schedule_t* schd; co_routine_t * task; }; switch (_self_->line) { case 0: | ||||
588 | ccv_nnc_graph_exec_info_t* node;co_state_t _ccv_nnc_graph_exec_run_loop(co_routine_t* const _self , void* const _privates_); struct _ccv_nnc_graph_exec_run_loop_param_s { struct { ccv_nnc_graph_t* const graph;ccv_nnc_graph_exec_info_t * const exec_info;const ccv_nnc_graph_exec_schedule_t* const schd_info ;const int* const psort;const int start_index;const int exec_info_size ;ccv_nnc_tensor_tape_t* const tensor_tape;const int flags;;;; ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; } _co_params ; }; size_t _ccv_nnc_graph_exec_run_loop_stack_size(void); struct _ccv_nnc_graph_exec_run_loop_private_s { struct _ccv_nnc_graph_exec_run_loop_param_s _co_params; int i, p, q; int sub_task_size; co_routine_t** sub_tasks ; int* pending_nodes[2]; int pending_node_size[2]; int idx; ccv_nnc_graph_exec_info_t * node; const ccv_nnc_graph_exec_schedule_t* schd; co_routine_t * task; }; size_t _ccv_nnc_graph_exec_run_loop_stack_size(void ) { return sizeof(struct _ccv_nnc_graph_exec_run_loop_private_s ); } co_state_t _ccv_nnc_graph_exec_run_loop(co_routine_t* const _self_, void* const _privates_) { struct _private_s { struct _ccv_nnc_graph_exec_run_loop_param_s _co_params; int i, p, q ; int sub_task_size; co_routine_t** sub_tasks; int* pending_nodes [2]; int pending_node_size[2]; int idx; ccv_nnc_graph_exec_info_t * node; const ccv_nnc_graph_exec_schedule_t* schd; co_routine_t * task; }; switch (_self_->line) { case 0: | ||||
589 | const ccv_nnc_graph_exec_schedule_t* schd;co_state_t _ccv_nnc_graph_exec_run_loop(co_routine_t* const _self , void* const _privates_); struct _ccv_nnc_graph_exec_run_loop_param_s { struct { ccv_nnc_graph_t* const graph;ccv_nnc_graph_exec_info_t * const exec_info;const ccv_nnc_graph_exec_schedule_t* const schd_info ;const int* const psort;const int start_index;const int exec_info_size ;ccv_nnc_tensor_tape_t* const tensor_tape;const int flags;;;; ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; } _co_params ; }; size_t _ccv_nnc_graph_exec_run_loop_stack_size(void); struct _ccv_nnc_graph_exec_run_loop_private_s { struct _ccv_nnc_graph_exec_run_loop_param_s _co_params; int i, p, q; int sub_task_size; co_routine_t** sub_tasks ; int* pending_nodes[2]; int pending_node_size[2]; int idx; ccv_nnc_graph_exec_info_t * node; const ccv_nnc_graph_exec_schedule_t* schd; co_routine_t * task; }; size_t _ccv_nnc_graph_exec_run_loop_stack_size(void ) { return sizeof(struct _ccv_nnc_graph_exec_run_loop_private_s ); } co_state_t _ccv_nnc_graph_exec_run_loop(co_routine_t* const _self_, void* const _privates_) { struct _private_s { struct _ccv_nnc_graph_exec_run_loop_param_s _co_params; int i, p, q ; int sub_task_size; co_routine_t** sub_tasks; int* pending_nodes [2]; int pending_node_size[2]; int idx; ccv_nnc_graph_exec_info_t * node; const ccv_nnc_graph_exec_schedule_t* schd; co_routine_t * task; }; switch (_self_->line) { case 0: | ||||
590 | co_routine_t* task;co_state_t _ccv_nnc_graph_exec_run_loop(co_routine_t* const _self , void* const _privates_); struct _ccv_nnc_graph_exec_run_loop_param_s { struct { ccv_nnc_graph_t* const graph;ccv_nnc_graph_exec_info_t * const exec_info;const ccv_nnc_graph_exec_schedule_t* const schd_info ;const int* const psort;const int start_index;const int exec_info_size ;ccv_nnc_tensor_tape_t* const tensor_tape;const int flags;;;; ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; } _co_params ; }; size_t _ccv_nnc_graph_exec_run_loop_stack_size(void); struct _ccv_nnc_graph_exec_run_loop_private_s { struct _ccv_nnc_graph_exec_run_loop_param_s _co_params; int i, p, q; int sub_task_size; co_routine_t** sub_tasks ; int* pending_nodes[2]; int pending_node_size[2]; int idx; ccv_nnc_graph_exec_info_t * node; const ccv_nnc_graph_exec_schedule_t* schd; co_routine_t * task; }; size_t _ccv_nnc_graph_exec_run_loop_stack_size(void ) { return sizeof(struct _ccv_nnc_graph_exec_run_loop_private_s ); } co_state_t _ccv_nnc_graph_exec_run_loop(co_routine_t* const _self_, void* const _privates_) { struct _private_s { struct _ccv_nnc_graph_exec_run_loop_param_s _co_params; int i, p, q ; int sub_task_size; co_routine_t** sub_tasks; int* pending_nodes [2]; int pending_node_size[2]; int idx; ccv_nnc_graph_exec_info_t * node; const ccv_nnc_graph_exec_schedule_t* schd; co_routine_t * task; }; switch (_self_->line) { case 0: | ||||
591 | ))co_state_t _ccv_nnc_graph_exec_run_loop(co_routine_t* const _self , void* const _privates_); struct _ccv_nnc_graph_exec_run_loop_param_s { struct { ccv_nnc_graph_t* const graph;ccv_nnc_graph_exec_info_t * const exec_info;const ccv_nnc_graph_exec_schedule_t* const schd_info ;const int* const psort;const int start_index;const int exec_info_size ;ccv_nnc_tensor_tape_t* const tensor_tape;const int flags;;;; ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; } _co_params ; }; size_t _ccv_nnc_graph_exec_run_loop_stack_size(void); struct _ccv_nnc_graph_exec_run_loop_private_s { struct _ccv_nnc_graph_exec_run_loop_param_s _co_params; int i, p, q; int sub_task_size; co_routine_t** sub_tasks ; int* pending_nodes[2]; int pending_node_size[2]; int idx; ccv_nnc_graph_exec_info_t * node; const ccv_nnc_graph_exec_schedule_t* schd; co_routine_t * task; }; size_t _ccv_nnc_graph_exec_run_loop_stack_size(void ) { return sizeof(struct _ccv_nnc_graph_exec_run_loop_private_s ); } co_state_t _ccv_nnc_graph_exec_run_loop(co_routine_t* const _self_, void* const _privates_) { struct _private_s { struct _ccv_nnc_graph_exec_run_loop_param_s _co_params; int i, p, q ; int sub_task_size; co_routine_t** sub_tasks; int* pending_nodes [2]; int pending_node_size[2]; int idx; ccv_nnc_graph_exec_info_t * node; const ccv_nnc_graph_exec_schedule_t* schd; co_routine_t * task; }; switch (_self_->line) { case 0: { | ||||
592 | CO_V(sub_task_size)(((struct _private_s*)(_privates_))->sub_task_size) = 0; | ||||
593 | CO_V(sub_tasks)(((struct _private_s*)(_privates_))->sub_tasks) = (co_routine_t**)ccv_nnc_graph_buffer(CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph), sizeof(co_routine_t*) * (CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph)->sub_graphs ? CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph)->sub_graphs->rnum : 0) + sizeof(int) * CO_P(exec_info_size)(((struct _private_s*)(_privates_))->_co_params._co_params .exec_info_size) * 2); | ||||
594 | CO_V(pending_nodes)(((struct _private_s*)(_privates_))->pending_nodes)[0] = (int*)(CO_V(sub_tasks)(((struct _private_s*)(_privates_))->sub_tasks) + (CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph)->sub_graphs ? CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph)->sub_graphs->rnum : 0)); | ||||
595 | CO_V(pending_nodes)(((struct _private_s*)(_privates_))->pending_nodes)[1] = CO_V(pending_nodes)(((struct _private_s*)(_privates_))->pending_nodes)[0] + CO_P(exec_info_size)(((struct _private_s*)(_privates_))->_co_params._co_params .exec_info_size); | ||||
596 | CO_V(pending_node_size)(((struct _private_s*)(_privates_))->pending_node_size)[0] = 0; | ||||
597 | CO_V(pending_node_size)(((struct _private_s*)(_privates_))->pending_node_size)[1] = 0; | ||||
598 | for (CO_V(i)(((struct _private_s*)(_privates_))->i) = CO_P(start_index)(((struct _private_s*)(_privates_))->_co_params._co_params .start_index); CO_V(i)(((struct _private_s*)(_privates_))->i) < CO_P(exec_info_size)(((struct _private_s*)(_privates_))->_co_params._co_params .exec_info_size); CO_V(i)(((struct _private_s*)(_privates_))->i)++) | ||||
599 | { | ||||
600 | CO_V(idx)(((struct _private_s*)(_privates_))->idx) = CO_P(psort)(((struct _private_s*)(_privates_))->_co_params._co_params .psort) ? CO_P(psort)(((struct _private_s*)(_privates_))->_co_params._co_params .psort)[CO_V(i)(((struct _private_s*)(_privates_))->i)] : CO_V(i)(((struct _private_s*)(_privates_))->i); | ||||
601 | CO_V(node)(((struct _private_s*)(_privates_))->node) = CO_P(exec_info)(((struct _private_s*)(_privates_))->_co_params._co_params .exec_info) + CO_V(idx)(((struct _private_s*)(_privates_))->idx); | ||||
602 | CO_V(schd)(((struct _private_s*)(_privates_))->schd) = CO_P(schd_info)(((struct _private_s*)(_privates_))->_co_params._co_params .schd_info) + CO_V(idx)(((struct _private_s*)(_privates_))->idx); | ||||
603 | // If stream is blocked by but not blocked by current executing task. | ||||
604 | int blocked = 0, j; | ||||
605 | for (j = 0; j < CO_V(schd)(((struct _private_s*)(_privates_))->schd)->stream_size; j++) | ||||
606 | if (CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph)->block_stream_tasks[SCHEDULE_STREAMS(*CO_V(schd))((*(((struct _private_s*)(_privates_))->schd)).stream_size <= 1 ? (*(((struct _private_s*)(_privates_))->schd))._inline_streams : (*(((struct _private_s*)(_privates_))->schd))._heap_streams )[j]]) | ||||
607 | { | ||||
608 | CO_V(pending_nodes)(((struct _private_s*)(_privates_))->pending_nodes)[0][CO_V(pending_node_size)(((struct _private_s*)(_privates_))->pending_node_size)[0]++] = CO_V(idx)(((struct _private_s*)(_privates_))->idx); | ||||
609 | _ccv_nnc_graph_mark_outgoing_streams_blocked_by_task(CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph), CO_P(schd_info)(((struct _private_s*)(_privates_))->_co_params._co_params .schd_info), CO_V(node)(((struct _private_s*)(_privates_))->node), CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph)->block_stream_tasks[SCHEDULE_STREAMS(*CO_V(schd))((*(((struct _private_s*)(_privates_))->schd)).stream_size <= 1 ? (*(((struct _private_s*)(_privates_))->schd))._inline_streams : (*(((struct _private_s*)(_privates_))->schd))._heap_streams )[j]]); | ||||
610 | blocked = 1; | ||||
611 | } | ||||
612 | if (blocked) | ||||
613 | continue; | ||||
614 | CO_V(task)(((struct _private_s*)(_privates_))->task) = _ccv_nnc_graph_exec_run_task(CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph), CO_V(node)(((struct _private_s*)(_privates_))->node), CO_V(schd)(((struct _private_s*)(_privates_))->schd), CO_V(idx)(((struct _private_s*)(_privates_))->idx), CO_P(tensor_tape)(((struct _private_s*)(_privates_))->_co_params._co_params .tensor_tape), CO_P(flags)(((struct _private_s*)(_privates_))->_co_params._co_params .flags)); | ||||
615 | if (CO_V(task)(((struct _private_s*)(_privates_))->task)) | ||||
616 | { | ||||
617 | co_resume(CO_V(task))do { _co_resume(_self_, (((struct _private_s*)(_privates_))-> task)); return (co_state_t){ 617, 0 }; case 617: _self_->callee = 0; } while (0); | ||||
618 | if (!co_is_done(CO_V(task)(((struct _private_s*)(_privates_))->task))) | ||||
619 | { | ||||
620 | CO_V(sub_tasks)(((struct _private_s*)(_privates_))->sub_tasks)[CO_V(sub_task_size)(((struct _private_s*)(_privates_))->sub_task_size)++] = CO_V(task)(((struct _private_s*)(_privates_))->task); | ||||
621 | int j; | ||||
622 | for (j = 0; j < CO_V(schd)(((struct _private_s*)(_privates_))->schd)->stream_size; j++) | ||||
623 | CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph)->block_stream_tasks[SCHEDULE_STREAMS(*CO_V(schd))((*(((struct _private_s*)(_privates_))->schd)).stream_size <= 1 ? (*(((struct _private_s*)(_privates_))->schd))._inline_streams : (*(((struct _private_s*)(_privates_))->schd))._heap_streams )[j]] = CO_V(task)(((struct _private_s*)(_privates_))->task); | ||||
624 | _ccv_nnc_graph_mark_outgoing_streams_blocked_by_task(CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph), CO_P(schd_info)(((struct _private_s*)(_privates_))->_co_params._co_params .schd_info), CO_V(node)(((struct _private_s*)(_privates_))->node), CO_V(task)(((struct _private_s*)(_privates_))->task)); | ||||
625 | } else | ||||
626 | co_free(CO_V(task)(((struct _private_s*)(_privates_))->task)); | ||||
627 | } | ||||
628 | } | ||||
629 | if (CO_V(sub_task_size)(((struct _private_s*)(_privates_))->sub_task_size)) | ||||
630 | co_apply(_ccv_nnc_graph_wait_any_sub_tasks, (CO_P(graph), CO_V(sub_tasks), CO_V(sub_task_size), CO_P(schd_info), CO_V(pending_nodes)[0], CO_V(pending_node_size)[0]))do { _self_->callee = ({ co_routine_t* const task = malloc ((sizeof(co_routine_t) + _ccv_nnc_graph_wait_any_sub_tasks_stack_size ())); do { struct _ccv_nnc_graph_wait_any_sub_tasks_param_s params = { ._co_params = { (((struct _private_s*)(_privates_))-> _co_params._co_params.graph), (((struct _private_s*)(_privates_ ))->sub_tasks), (((struct _private_s*)(_privates_))->sub_task_size ), (((struct _private_s*)(_privates_))->_co_params._co_params .schd_info), (((struct _private_s*)(_privates_))->pending_nodes )[0], (((struct _private_s*)(_privates_))->pending_node_size )[0] } }; task->fn = _ccv_nnc_graph_wait_any_sub_tasks; task ->line = 0; task->done = 0; task->root = 0; task-> other_size = 0; task->notify_any = 0; task->others = 0; task->caller = 0; task->callee = 0; if (sizeof(params) > 0) memcpy(task + 1, ¶ms, sizeof(params)); } while (0); task; }); _co_apply(_self_, _self_->callee); return ( co_state_t){ 630, 0 }; case 630: co_free(_self_->callee); _self_ ->callee = 0; } while (0); | ||||
631 | CO_V(p)(((struct _private_s*)(_privates_))->p) = 0; | ||||
632 | CO_V(q)(((struct _private_s*)(_privates_))->q) = 1; | ||||
633 | while (CO_V(pending_node_size)(((struct _private_s*)(_privates_))->pending_node_size)[CO_V(p)(((struct _private_s*)(_privates_))->p)] > 0) | ||||
634 | { | ||||
635 | CO_V(pending_node_size)(((struct _private_s*)(_privates_))->pending_node_size)[CO_V(q)(((struct _private_s*)(_privates_))->q)] = 0; | ||||
636 | CO_V(sub_task_size)(((struct _private_s*)(_privates_))->sub_task_size) = 0; | ||||
637 | for (CO_V(i)(((struct _private_s*)(_privates_))->i) = 0; CO_V(i)(((struct _private_s*)(_privates_))->i) < CO_V(pending_node_size)(((struct _private_s*)(_privates_))->pending_node_size)[CO_V(p)(((struct _private_s*)(_privates_))->p)]; CO_V(i)(((struct _private_s*)(_privates_))->i)++) | ||||
638 | { | ||||
639 | CO_V(idx)(((struct _private_s*)(_privates_))->idx) = CO_V(pending_nodes)(((struct _private_s*)(_privates_))->pending_nodes)[CO_V(p)(((struct _private_s*)(_privates_))->p)][CO_V(i)(((struct _private_s*)(_privates_))->i)]; | ||||
640 | CO_V(node)(((struct _private_s*)(_privates_))->node) = CO_P(exec_info)(((struct _private_s*)(_privates_))->_co_params._co_params .exec_info) + CO_V(idx)(((struct _private_s*)(_privates_))->idx); | ||||
641 | CO_V(schd)(((struct _private_s*)(_privates_))->schd) = CO_P(schd_info)(((struct _private_s*)(_privates_))->_co_params._co_params .schd_info) + CO_V(idx)(((struct _private_s*)(_privates_))->idx); | ||||
642 | int blocked = 0, j; | ||||
643 | for (j = 0; j < CO_V(schd)(((struct _private_s*)(_privates_))->schd)->stream_size; j++) | ||||
644 | if (CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph)->block_stream_tasks[SCHEDULE_STREAMS(*CO_V(schd))((*(((struct _private_s*)(_privates_))->schd)).stream_size <= 1 ? (*(((struct _private_s*)(_privates_))->schd))._inline_streams : (*(((struct _private_s*)(_privates_))->schd))._heap_streams )[j]]) | ||||
645 | { | ||||
646 | _ccv_nnc_graph_mark_outgoing_streams_blocked_by_task(CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph), CO_P(schd_info)(((struct _private_s*)(_privates_))->_co_params._co_params .schd_info), CO_V(node)(((struct _private_s*)(_privates_))->node), CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph)->block_stream_tasks[SCHEDULE_STREAMS(*CO_V(schd))((*(((struct _private_s*)(_privates_))->schd)).stream_size <= 1 ? (*(((struct _private_s*)(_privates_))->schd))._inline_streams : (*(((struct _private_s*)(_privates_))->schd))._heap_streams )[j]]); | ||||
647 | CO_V(pending_nodes)(((struct _private_s*)(_privates_))->pending_nodes)[CO_V(q)(((struct _private_s*)(_privates_))->q)][CO_V(pending_node_size)(((struct _private_s*)(_privates_))->pending_node_size)[CO_V(q)(((struct _private_s*)(_privates_))->q)]++] = CO_V(idx)(((struct _private_s*)(_privates_))->idx); | ||||
648 | blocked = 1; | ||||
649 | } | ||||
650 | if (blocked) | ||||
651 | continue; | ||||
652 | CO_V(task)(((struct _private_s*)(_privates_))->task) = _ccv_nnc_graph_exec_run_task(CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph), CO_V(node)(((struct _private_s*)(_privates_))->node), CO_V(schd)(((struct _private_s*)(_privates_))->schd), CO_V(idx)(((struct _private_s*)(_privates_))->idx), CO_P(tensor_tape)(((struct _private_s*)(_privates_))->_co_params._co_params .tensor_tape), CO_P(flags)(((struct _private_s*)(_privates_))->_co_params._co_params .flags)); | ||||
653 | if (CO_V(task)(((struct _private_s*)(_privates_))->task)) | ||||
654 | { | ||||
655 | co_resume(CO_V(task))do { _co_resume(_self_, (((struct _private_s*)(_privates_))-> task)); return (co_state_t){ 655, 0 }; case 655: _self_->callee = 0; } while (0); | ||||
656 | if (!co_is_done(CO_V(task)(((struct _private_s*)(_privates_))->task))) | ||||
657 | { | ||||
658 | CO_V(sub_tasks)(((struct _private_s*)(_privates_))->sub_tasks)[CO_V(sub_task_size)(((struct _private_s*)(_privates_))->sub_task_size)++] = CO_V(task)(((struct _private_s*)(_privates_))->task); | ||||
659 | for (j = 0; j < CO_V(schd)(((struct _private_s*)(_privates_))->schd)->stream_size; j++) | ||||
660 | CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph)->block_stream_tasks[SCHEDULE_STREAMS(*CO_V(schd))((*(((struct _private_s*)(_privates_))->schd)).stream_size <= 1 ? (*(((struct _private_s*)(_privates_))->schd))._inline_streams : (*(((struct _private_s*)(_privates_))->schd))._heap_streams )[j]] = CO_V(task)(((struct _private_s*)(_privates_))->task); | ||||
661 | _ccv_nnc_graph_mark_outgoing_streams_blocked_by_task(CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph), CO_P(schd_info)(((struct _private_s*)(_privates_))->_co_params._co_params .schd_info), CO_V(node)(((struct _private_s*)(_privates_))->node), CO_V(task)(((struct _private_s*)(_privates_))->task)); | ||||
662 | } else | ||||
663 | co_free(CO_V(task)(((struct _private_s*)(_privates_))->task)); | ||||
664 | } | ||||
665 | } | ||||
666 | int t; | ||||
667 | CCV_SWAP(CO_V(p), CO_V(q), t)((t) = ((((struct _private_s*)(_privates_))->p)), ((((struct _private_s*)(_privates_))->p)) = ((((struct _private_s*)( _privates_))->q)), ((((struct _private_s*)(_privates_))-> q)) = (t)); | ||||
668 | if (CO_V(sub_task_size)(((struct _private_s*)(_privates_))->sub_task_size)) | ||||
669 | co_apply(_ccv_nnc_graph_wait_any_sub_tasks, (CO_P(graph), CO_V(sub_tasks), CO_V(sub_task_size), CO_P(schd_info), CO_V(pending_nodes)[CO_V(p)], CO_V(pending_node_size)[CO_V(p)]))do { _self_->callee = ({ co_routine_t* const task = malloc ((sizeof(co_routine_t) + _ccv_nnc_graph_wait_any_sub_tasks_stack_size ())); do { struct _ccv_nnc_graph_wait_any_sub_tasks_param_s params = { ._co_params = { (((struct _private_s*)(_privates_))-> _co_params._co_params.graph), (((struct _private_s*)(_privates_ ))->sub_tasks), (((struct _private_s*)(_privates_))->sub_task_size ), (((struct _private_s*)(_privates_))->_co_params._co_params .schd_info), (((struct _private_s*)(_privates_))->pending_nodes )[(((struct _private_s*)(_privates_))->p)], (((struct _private_s *)(_privates_))->pending_node_size)[(((struct _private_s*) (_privates_))->p)] } }; task->fn = _ccv_nnc_graph_wait_any_sub_tasks ; task->line = 0; task->done = 0; task->root = 0; task ->other_size = 0; task->notify_any = 0; task->others = 0; task->caller = 0; task->callee = 0; if (sizeof(params ) > 0) memcpy(task + 1, ¶ms, sizeof(params)); } while (0); task; }); _co_apply(_self_, _self_->callee); return ( co_state_t){ 669, 0 }; case 669: co_free(_self_->callee); _self_ ->callee = 0; } while (0); | ||||
670 | } | ||||
671 | } co_end()default: return (co_state_t){ 671, 1 }; } } | ||||
672 | |||||
673 | co_task(_ccv_nnc_graph_topsorted_run_coro, (ccv_nnc_graph_t* const graph, const int exec_idx, const ccv_nnc_graph_static_schedule_t* const schedule, ccv_nnc_graph_exec_info_t* const exec, ccv_nnc_tensor_tape_t* const tensor_tape, ccv_nnc_stream_context_t* const stream_context, const int flags), private(struct _ccv_nnc_graph_topsorted_run_coro_private_s { struct _ccv_nnc_graph_topsorted_run_coro_param_s _co_params; ccv_nnc_graph_exec_info_t* exec_info; const ccv_nnc_graph_exec_schedule_t * schd_info; co_routine_t* previous_main; int stream_0; int64_t count, reverse_count; int graph_breakpoint_size; int i, j; } ; size_t _ccv_nnc_graph_topsorted_run_coro_stack_size(void) { return sizeof(struct _ccv_nnc_graph_topsorted_run_coro_private_s ); } co_state_t _ccv_nnc_graph_topsorted_run_coro(co_routine_t * const _self_, void* const _privates_) { struct _private_s { struct _ccv_nnc_graph_topsorted_run_coro_param_s _co_params; ccv_nnc_graph_exec_info_t* exec_info; const ccv_nnc_graph_exec_schedule_t * schd_info; co_routine_t* previous_main; int stream_0; int64_t count, reverse_count; int graph_breakpoint_size; int i, j; } ; switch (_self_->line) { case 0: | ||||
674 | ccv_nnc_graph_exec_info_t* exec_info;struct _ccv_nnc_graph_topsorted_run_coro_private_s { struct _ccv_nnc_graph_topsorted_run_coro_param_s _co_params; ccv_nnc_graph_exec_info_t* exec_info; const ccv_nnc_graph_exec_schedule_t * schd_info; co_routine_t* previous_main; int stream_0; int64_t count, reverse_count; int graph_breakpoint_size; int i, j; } ; size_t _ccv_nnc_graph_topsorted_run_coro_stack_size(void) { return sizeof(struct _ccv_nnc_graph_topsorted_run_coro_private_s ); } co_state_t _ccv_nnc_graph_topsorted_run_coro(co_routine_t * const _self_, void* const _privates_) { struct _private_s { struct _ccv_nnc_graph_topsorted_run_coro_param_s _co_params; ccv_nnc_graph_exec_info_t* exec_info; const ccv_nnc_graph_exec_schedule_t * schd_info; co_routine_t* previous_main; int stream_0; int64_t count, reverse_count; int graph_breakpoint_size; int i, j; } ; switch (_self_->line) { case 0: | ||||
675 | const ccv_nnc_graph_exec_schedule_t* schd_info;struct _ccv_nnc_graph_topsorted_run_coro_private_s { struct _ccv_nnc_graph_topsorted_run_coro_param_s _co_params; ccv_nnc_graph_exec_info_t* exec_info; const ccv_nnc_graph_exec_schedule_t * schd_info; co_routine_t* previous_main; int stream_0; int64_t count, reverse_count; int graph_breakpoint_size; int i, j; } ; size_t _ccv_nnc_graph_topsorted_run_coro_stack_size(void) { return sizeof(struct _ccv_nnc_graph_topsorted_run_coro_private_s ); } co_state_t _ccv_nnc_graph_topsorted_run_coro(co_routine_t * const _self_, void* const _privates_) { struct _private_s { struct _ccv_nnc_graph_topsorted_run_coro_param_s _co_params; ccv_nnc_graph_exec_info_t* exec_info; const ccv_nnc_graph_exec_schedule_t * schd_info; co_routine_t* previous_main; int stream_0; int64_t count, reverse_count; int graph_breakpoint_size; int i, j; } ; switch (_self_->line) { case 0: | ||||
676 | co_routine_t* previous_main;struct _ccv_nnc_graph_topsorted_run_coro_private_s { struct _ccv_nnc_graph_topsorted_run_coro_param_s _co_params; ccv_nnc_graph_exec_info_t* exec_info; const ccv_nnc_graph_exec_schedule_t * schd_info; co_routine_t* previous_main; int stream_0; int64_t count, reverse_count; int graph_breakpoint_size; int i, j; } ; size_t _ccv_nnc_graph_topsorted_run_coro_stack_size(void) { return sizeof(struct _ccv_nnc_graph_topsorted_run_coro_private_s ); } co_state_t _ccv_nnc_graph_topsorted_run_coro(co_routine_t * const _self_, void* const _privates_) { struct _private_s { struct _ccv_nnc_graph_topsorted_run_coro_param_s _co_params; ccv_nnc_graph_exec_info_t* exec_info; const ccv_nnc_graph_exec_schedule_t * schd_info; co_routine_t* previous_main; int stream_0; int64_t count, reverse_count; int graph_breakpoint_size; int i, j; } ; switch (_self_->line) { case 0: | ||||
677 | int stream_0;struct _ccv_nnc_graph_topsorted_run_coro_private_s { struct _ccv_nnc_graph_topsorted_run_coro_param_s _co_params; ccv_nnc_graph_exec_info_t* exec_info; const ccv_nnc_graph_exec_schedule_t * schd_info; co_routine_t* previous_main; int stream_0; int64_t count, reverse_count; int graph_breakpoint_size; int i, j; } ; size_t _ccv_nnc_graph_topsorted_run_coro_stack_size(void) { return sizeof(struct _ccv_nnc_graph_topsorted_run_coro_private_s ); } co_state_t _ccv_nnc_graph_topsorted_run_coro(co_routine_t * const _self_, void* const _privates_) { struct _private_s { struct _ccv_nnc_graph_topsorted_run_coro_param_s _co_params; ccv_nnc_graph_exec_info_t* exec_info; const ccv_nnc_graph_exec_schedule_t * schd_info; co_routine_t* previous_main; int stream_0; int64_t count, reverse_count; int graph_breakpoint_size; int i, j; } ; switch (_self_->line) { case 0: | ||||
678 | // while loopstruct _ccv_nnc_graph_topsorted_run_coro_private_s { struct _ccv_nnc_graph_topsorted_run_coro_param_s _co_params; ccv_nnc_graph_exec_info_t* exec_info; const ccv_nnc_graph_exec_schedule_t * schd_info; co_routine_t* previous_main; int stream_0; int64_t count, reverse_count; int graph_breakpoint_size; int i, j; } ; size_t _ccv_nnc_graph_topsorted_run_coro_stack_size(void) { return sizeof(struct _ccv_nnc_graph_topsorted_run_coro_private_s ); } co_state_t _ccv_nnc_graph_topsorted_run_coro(co_routine_t * const _self_, void* const _privates_) { struct _private_s { struct _ccv_nnc_graph_topsorted_run_coro_param_s _co_params; ccv_nnc_graph_exec_info_t* exec_info; const ccv_nnc_graph_exec_schedule_t * schd_info; co_routine_t* previous_main; int stream_0; int64_t count, reverse_count; int graph_breakpoint_size; int i, j; } ; switch (_self_->line) { case 0: | ||||
679 | int64_t count, reverse_count;struct _ccv_nnc_graph_topsorted_run_coro_private_s { struct _ccv_nnc_graph_topsorted_run_coro_param_s _co_params; ccv_nnc_graph_exec_info_t* exec_info; const ccv_nnc_graph_exec_schedule_t * schd_info; co_routine_t* previous_main; int stream_0; int64_t count, reverse_count; int graph_breakpoint_size; int i, j; } ; size_t _ccv_nnc_graph_topsorted_run_coro_stack_size(void) { return sizeof(struct _ccv_nnc_graph_topsorted_run_coro_private_s ); } co_state_t _ccv_nnc_graph_topsorted_run_coro(co_routine_t * const _self_, void* const _privates_) { struct _private_s { struct _ccv_nnc_graph_topsorted_run_coro_param_s _co_params; ccv_nnc_graph_exec_info_t* exec_info; const ccv_nnc_graph_exec_schedule_t * schd_info; co_routine_t* previous_main; int stream_0; int64_t count, reverse_count; int graph_breakpoint_size; int i, j; } ; switch (_self_->line) { case 0: | ||||
680 | int graph_breakpoint_size;struct _ccv_nnc_graph_topsorted_run_coro_private_s { struct _ccv_nnc_graph_topsorted_run_coro_param_s _co_params; ccv_nnc_graph_exec_info_t* exec_info; const ccv_nnc_graph_exec_schedule_t * schd_info; co_routine_t* previous_main; int stream_0; int64_t count, reverse_count; int graph_breakpoint_size; int i, j; } ; size_t _ccv_nnc_graph_topsorted_run_coro_stack_size(void) { return sizeof(struct _ccv_nnc_graph_topsorted_run_coro_private_s ); } co_state_t _ccv_nnc_graph_topsorted_run_coro(co_routine_t * const _self_, void* const _privates_) { struct _private_s { struct _ccv_nnc_graph_topsorted_run_coro_param_s _co_params; ccv_nnc_graph_exec_info_t* exec_info; const ccv_nnc_graph_exec_schedule_t * schd_info; co_routine_t* previous_main; int stream_0; int64_t count, reverse_count; int graph_breakpoint_size; int i, j; } ; switch (_self_->line) { case 0: | ||||
681 | int i, j;struct _ccv_nnc_graph_topsorted_run_coro_private_s { struct _ccv_nnc_graph_topsorted_run_coro_param_s _co_params; ccv_nnc_graph_exec_info_t* exec_info; const ccv_nnc_graph_exec_schedule_t * schd_info; co_routine_t* previous_main; int stream_0; int64_t count, reverse_count; int graph_breakpoint_size; int i, j; } ; size_t _ccv_nnc_graph_topsorted_run_coro_stack_size(void) { return sizeof(struct _ccv_nnc_graph_topsorted_run_coro_private_s ); } co_state_t _ccv_nnc_graph_topsorted_run_coro(co_routine_t * const _self_, void* const _privates_) { struct _private_s { struct _ccv_nnc_graph_topsorted_run_coro_param_s _co_params; ccv_nnc_graph_exec_info_t* exec_info; const ccv_nnc_graph_exec_schedule_t * schd_info; co_routine_t* previous_main; int stream_0; int64_t count, reverse_count; int graph_breakpoint_size; int i, j; } ; switch (_self_->line) { case 0: | ||||
682 | ))struct _ccv_nnc_graph_topsorted_run_coro_private_s { struct _ccv_nnc_graph_topsorted_run_coro_param_s _co_params; ccv_nnc_graph_exec_info_t* exec_info; const ccv_nnc_graph_exec_schedule_t * schd_info; co_routine_t* previous_main; int stream_0; int64_t count, reverse_count; int graph_breakpoint_size; int i, j; } ; size_t _ccv_nnc_graph_topsorted_run_coro_stack_size(void) { return sizeof(struct _ccv_nnc_graph_topsorted_run_coro_private_s ); } co_state_t _ccv_nnc_graph_topsorted_run_coro(co_routine_t * const _self_, void* const _privates_) { struct _private_s { struct _ccv_nnc_graph_topsorted_run_coro_param_s _co_params; ccv_nnc_graph_exec_info_t* exec_info; const ccv_nnc_graph_exec_schedule_t * schd_info; co_routine_t* previous_main; int stream_0; int64_t count, reverse_count; int graph_breakpoint_size; int i, j; } ; switch (_self_->line) { case 0: { | ||||
683 | assert(CO_P(graph)->stream_size > 0)((void) sizeof (((((struct _private_s*)(_privates_))->_co_params ._co_params.graph)->stream_size > 0) ? 1 : 0), __extension__ ({ if ((((struct _private_s*)(_privates_))->_co_params._co_params .graph)->stream_size > 0) ; else __assert_fail ("CO_P(graph)->stream_size > 0" , "ccv_nnc_graph_run.c", 683, __extension__ __PRETTY_FUNCTION__ ); })); | ||||
684 | int i; | ||||
685 | // Assign the resource container pointer. | ||||
686 | for (i = 0; i < CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph)->stream_size; i++) | ||||
687 | CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph)->streams[i]->resource_container = CO_P(stream_context)(((struct _private_s*)(_privates_))->_co_params._co_params .stream_context)->_inline_container; | ||||
688 | CO_V(exec_info)(((struct _private_s*)(_privates_))->exec_info) = (ccv_nnc_graph_exec_info_t*)ccv_array_get(CO_P(graph)->exec_info, 0)((void*)(((char*)(((((struct _private_s*)(_privates_))->_co_params ._co_params.graph)->exec_info)->data)) + (size_t)((((struct _private_s*)(_privates_))->_co_params._co_params.graph)-> exec_info)->rsize * (size_t)(0))); | ||||
689 | CO_V(schd_info)(((struct _private_s*)(_privates_))->schd_info) = CO_P(schedule)(((struct _private_s*)(_privates_))->_co_params._co_params .schedule)->exec_info; | ||||
690 | CO_V(stream_0)(((struct _private_s*)(_privates_))->stream_0) = CO_P(schedule)(((struct _private_s*)(_privates_))->_co_params._co_params .schedule)->stream_0; | ||||
691 | if (CO_P(exec_idx)(((struct _private_s*)(_privates_))->_co_params._co_params .exec_idx) == -1) | ||||
692 | { | ||||
693 | if (CO_P(stream_context)(((struct _private_s*)(_privates_))->_co_params._co_params .stream_context)->main) | ||||
694 | { | ||||
695 | CO_V(previous_main)(((struct _private_s*)(_privates_))->previous_main) = CO_P(stream_context)(((struct _private_s*)(_privates_))->_co_params._co_params .stream_context)->main; | ||||
696 | CO_P(stream_context)(((struct _private_s*)(_privates_))->_co_params._co_params .stream_context)->main = co_self()(_self_); | ||||
697 | // Wait the previous task to be done. This makes sure that our graph run is serial on the same stream. | ||||
698 | assert(!co_is_done(CO_V(previous_main)))((void) sizeof ((!co_is_done((((struct _private_s*)(_privates_ ))->previous_main))) ? 1 : 0), __extension__ ({ if (!co_is_done ((((struct _private_s*)(_privates_))->previous_main))) ; else __assert_fail ("!co_is_done(CO_V(previous_main))", "ccv_nnc_graph_run.c" , 698, __extension__ __PRETTY_FUNCTION__); })); | ||||
699 | co_await(CO_V(previous_main))do { if (!_co_await_any(_self_, &((((struct _private_s*)( _privates_))->previous_main)), 1)) { return (co_state_t){ 699 , 0 }; } case 699: ; } while (0); | ||||
700 | } else | ||||
701 | CO_P(stream_context)(((struct _private_s*)(_privates_))->_co_params._co_params .stream_context)->main = co_self()(_self_); | ||||
702 | PRINT(CCV_CLI_INFO, "Graph Stream %d Begin", CO_V(stream_0))do { if ((CCV_CLI_INFO & ccv_cli_get_output_levels())) { printf ("Graph Stream %d Begin", (((struct _private_s*)(_privates_)) ->stream_0)); fflush(stdout); } } while (0); | ||||
703 | ccv_nnc_stream_signal_t* stream_0_signal; | ||||
704 | if (CO_P(stream_context)(((struct _private_s*)(_privates_))->_co_params._co_params .stream_context) != CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph)->streams[CO_V(stream_0)(((struct _private_s*)(_privates_))->stream_0)]) | ||||
705 | { | ||||
706 | // Make sure when we start work on streams[0], the current stream context is done. | ||||
707 | stream_0_signal = ccv_nnc_stream_context_emit_signal_new(CO_P(stream_context)(((struct _private_s*)(_privates_))->_co_params._co_params .stream_context)); | ||||
708 | ccv_nnc_stream_context_wait_signal(CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph)->streams[CO_V(stream_0)(((struct _private_s*)(_privates_))->stream_0)], stream_0_signal); | ||||
709 | } else if (CO_P(schedule)(((struct _private_s*)(_privates_))->_co_params._co_params .schedule)->stream_1_size) { | ||||
710 | ccv_nnc_stream_context_emit_signal(CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph)->streams[CO_V(stream_0)(((struct _private_s*)(_privates_))->stream_0)], CO_P(schedule)(((struct _private_s*)(_privates_))->_co_params._co_params .schedule)->begin); | ||||
711 | stream_0_signal = CO_P(schedule)(((struct _private_s*)(_privates_))->_co_params._co_params .schedule)->begin; | ||||
712 | } | ||||
713 | int i, flag = 0; | ||||
714 | for (i = 0; i < CO_P(schedule)(((struct _private_s*)(_privates_))->_co_params._co_params .schedule)->stream_1_size; i++) | ||||
715 | { | ||||
716 | ccv_nnc_stream_context_wait_signal(CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph)->streams[CO_P(schedule)(((struct _private_s*)(_privates_))->_co_params._co_params .schedule)->stream_1s[i]], stream_0_signal); | ||||
717 | if (!flag) | ||||
718 | { | ||||
719 | PRINT(CCV_CLI_INFO, ", Wait: %d", CO_P(schedule)->stream_1s[i])do { if ((CCV_CLI_INFO & ccv_cli_get_output_levels())) { printf (", Wait: %d", (((struct _private_s*)(_privates_))->_co_params ._co_params.schedule)->stream_1s[i]); fflush(stdout); } } while (0); | ||||
720 | flag = 1; | ||||
721 | } else | ||||
722 | PRINT(CCV_CLI_INFO, ", %d", CO_P(schedule)->stream_1s[i])do { if ((CCV_CLI_INFO & ccv_cli_get_output_levels())) { printf (", %d", (((struct _private_s*)(_privates_))->_co_params._co_params .schedule)->stream_1s[i]); fflush(stdout); } } while (0); | ||||
723 | } | ||||
724 | PRINT(CCV_CLI_INFO, "\n")do { if ((CCV_CLI_INFO & ccv_cli_get_output_levels())) { printf ("\n"); fflush(stdout); } } while (0); | ||||
725 | } else { | ||||
726 | assert(CO_P(stream_context) == CO_P(graph)->streams[0])((void) sizeof (((((struct _private_s*)(_privates_))->_co_params ._co_params.stream_context) == (((struct _private_s*)(_privates_ ))->_co_params._co_params.graph)->streams[0]) ? 1 : 0), __extension__ ({ if ((((struct _private_s*)(_privates_))-> _co_params._co_params.stream_context) == (((struct _private_s *)(_privates_))->_co_params._co_params.graph)->streams[ 0]) ; else __assert_fail ("CO_P(stream_context) == CO_P(graph)->streams[0]" , "ccv_nnc_graph_run.c", 726, __extension__ __PRETTY_FUNCTION__ ); })); | ||||
727 | } | ||||
728 | if (CO_P(exec)(((struct _private_s*)(_privates_))->_co_params._co_params .exec) && (CO_P(exec)(((struct _private_s*)(_privates_))->_co_params._co_params .exec)->flags & CCV_NNC_GRAPH_EXEC_P_WHILE)) | ||||
729 | { | ||||
730 | assert(CO_P(schedule) == CO_P(graph)->default_schedule)((void) sizeof (((((struct _private_s*)(_privates_))->_co_params ._co_params.schedule) == (((struct _private_s*)(_privates_))-> _co_params._co_params.graph)->default_schedule) ? 1 : 0), __extension__ ({ if ((((struct _private_s*)(_privates_))->_co_params._co_params .schedule) == (((struct _private_s*)(_privates_))->_co_params ._co_params.graph)->default_schedule) ; else __assert_fail ("CO_P(schedule) == CO_P(graph)->default_schedule", "ccv_nnc_graph_run.c" , 730, __extension__ __PRETTY_FUNCTION__); })); | ||||
731 | assert(CO_P(exec)->p_while.expr)((void) sizeof (((((struct _private_s*)(_privates_))->_co_params ._co_params.exec)->p_while.expr) ? 1 : 0), __extension__ ( { if ((((struct _private_s*)(_privates_))->_co_params._co_params .exec)->p_while.expr) ; else __assert_fail ("CO_P(exec)->p_while.expr" , "ccv_nnc_graph_run.c", 731, __extension__ __PRETTY_FUNCTION__ ); })); | ||||
732 | CO_V(count)(((struct _private_s*)(_privates_))->count) = 0; | ||||
733 | // This is a forward while loop. Backward while loop will just consult its pairing part. | ||||
734 | if (CO_P(exec)(((struct _private_s*)(_privates_))->_co_params._co_params .exec)->cmd.cmd == CCV_NNC_GRAPH_FORWARD) | ||||
735 | { | ||||
736 | CO_V(graph_breakpoint_size)(((struct _private_s*)(_privates_))->graph_breakpoint_size ) = CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph)->breakpoint_offset + CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph)->breakpoint_size; | ||||
737 | for (;; ++CO_V(count)(((struct _private_s*)(_privates_))->count)) | ||||
738 | { | ||||
739 | CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph)->while_count = CO_V(count)(((struct _private_s*)(_privates_))->count); | ||||
740 | if (CO_P(tensor_tape)(((struct _private_s*)(_privates_))->_co_params._co_params .tensor_tape)) | ||||
741 | ccv_nnc_tensor_tape_set_numbering(CO_P(tensor_tape)(((struct _private_s*)(_privates_))->_co_params._co_params .tensor_tape), CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph)->p, (ccv_nnc_graph_exec_t){ | ||||
742 | .d = CO_P(exec_idx)(((struct _private_s*)(_privates_))->_co_params._co_params .exec_idx), | ||||
743 | .graph = CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph)->p, | ||||
744 | }, CO_V(count)(((struct _private_s*)(_privates_))->count)); | ||||
745 | _ccv_nnc_graph_unwrap(CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph), CO_V(count)(((struct _private_s*)(_privates_))->count), 0); | ||||
746 | if (CO_V(count)(((struct _private_s*)(_privates_))->count) > 0) | ||||
747 | _ccv_nnc_graph_transit_move_to(CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph)); | ||||
748 | co_apply(_ccv_nnc_graph_exec_run_loop, (CO_P(graph), CO_V(exec_info), CO_V(schd_info), 0, 0, CO_V(graph_breakpoint_size), CO_P(tensor_tape), CO_P(flags)))do { _self_->callee = ({ co_routine_t* const task = malloc ((sizeof(co_routine_t) + _ccv_nnc_graph_exec_run_loop_stack_size ())); do { struct _ccv_nnc_graph_exec_run_loop_param_s params = { ._co_params = { (((struct _private_s*)(_privates_))-> _co_params._co_params.graph), (((struct _private_s*)(_privates_ ))->exec_info), (((struct _private_s*)(_privates_))->schd_info ), 0, 0, (((struct _private_s*)(_privates_))->graph_breakpoint_size ), (((struct _private_s*)(_privates_))->_co_params._co_params .tensor_tape), (((struct _private_s*)(_privates_))->_co_params ._co_params.flags) } }; task->fn = _ccv_nnc_graph_exec_run_loop ; task->line = 0; task->done = 0; task->root = 0; task ->other_size = 0; task->notify_any = 0; task->others = 0; task->caller = 0; task->callee = 0; if (sizeof(params ) > 0) memcpy(task + 1, ¶ms, sizeof(params)); } while (0); task; }); _co_apply(_self_, _self_->callee); return ( co_state_t){ 748, 0 }; case 748: co_free(_self_->callee); _self_ ->callee = 0; } while (0); | ||||
749 | // Reached breakpoints, now check the breakpoint, if not met, break out. | ||||
750 | // Wait until everything on the stream is executed. | ||||
751 | for (CO_V(i)(((struct _private_s*)(_privates_))->i) = CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph)->breakpoint_offset; CO_V(i)(((struct _private_s*)(_privates_))->i) < CO_V(graph_breakpoint_size)(((struct _private_s*)(_privates_))->graph_breakpoint_size ); CO_V(i)(((struct _private_s*)(_privates_))->i)++) | ||||
752 | for (CO_V(j)(((struct _private_s*)(_privates_))->j) = 0; CO_V(j)(((struct _private_s*)(_privates_))->j) < CO_V(schd_info)(((struct _private_s*)(_privates_))->schd_info)[CO_V(i)(((struct _private_s*)(_privates_))->i)].stream_size; CO_V(j)(((struct _private_s*)(_privates_))->j)++) | ||||
753 | co_stream_await(CO_P(graph)->streams[SCHEDULE_STREAMS(CO_V(schd_info)[CO_V(i)])[CO_V(j)]])do { if (!_co_stream_await(_self_, (((struct _private_s*)(_privates_ ))->_co_params._co_params.graph)->streams[(((((struct _private_s *)(_privates_))->schd_info)[(((struct _private_s*)(_privates_ ))->i)]).stream_size <= 1 ? ((((struct _private_s*)(_privates_ ))->schd_info)[(((struct _private_s*)(_privates_))->i)] )._inline_streams : ((((struct _private_s*)(_privates_))-> schd_info)[(((struct _private_s*)(_privates_))->i)])._heap_streams )[(((struct _private_s*)(_privates_))->j)]])) { return (co_state_t ){ 753, 0 }; } case 753: ; } while (0); | ||||
754 | _ccv_nnc_graph_exec_unwrap_while_expr(CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph), CO_P(exec)(((struct _private_s*)(_privates_))->_co_params._co_params .exec)); | ||||
755 | if (!CO_P(exec)(((struct _private_s*)(_privates_))->_co_params._co_params .exec)->p_while.expr(CO_P(exec)(((struct _private_s*)(_privates_))->_co_params._co_params .exec)->p_while.inputs, CO_P(exec)(((struct _private_s*)(_privates_))->_co_params._co_params .exec)->p_while.input_size, CO_P(exec)(((struct _private_s*)(_privates_))->_co_params._co_params .exec)->p_while.data)) | ||||
756 | { | ||||
757 | _ccv_nnc_graph_rewrap(CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph)); | ||||
758 | // If we break from here, it is ok because all the streams are waited. | ||||
759 | break; | ||||
760 | } | ||||
761 | co_apply(_ccv_nnc_graph_exec_run_loop, (CO_P(graph), CO_V(exec_info), CO_V(schd_info), 0, CO_V(graph_breakpoint_size), CO_P(graph)->exec_info->rnum, CO_P(tensor_tape), CO_P(flags)))do { _self_->callee = ({ co_routine_t* const task = malloc ((sizeof(co_routine_t) + _ccv_nnc_graph_exec_run_loop_stack_size ())); do { struct _ccv_nnc_graph_exec_run_loop_param_s params = { ._co_params = { (((struct _private_s*)(_privates_))-> _co_params._co_params.graph), (((struct _private_s*)(_privates_ ))->exec_info), (((struct _private_s*)(_privates_))->schd_info ), 0, (((struct _private_s*)(_privates_))->graph_breakpoint_size ), (((struct _private_s*)(_privates_))->_co_params._co_params .graph)->exec_info->rnum, (((struct _private_s*)(_privates_ ))->_co_params._co_params.tensor_tape), (((struct _private_s *)(_privates_))->_co_params._co_params.flags) } }; task-> fn = _ccv_nnc_graph_exec_run_loop; task->line = 0; task-> done = 0; task->root = 0; task->other_size = 0; task-> notify_any = 0; task->others = 0; task->caller = 0; task ->callee = 0; if (sizeof(params) > 0) memcpy(task + 1, & params, sizeof(params)); } while (0); task; }); _co_apply(_self_ , _self_->callee); return (co_state_t){ 761, 0 }; case 761 : co_free(_self_->callee); _self_->callee = 0; } while ( 0); | ||||
762 | _ccv_nnc_graph_from_move_transit(CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph)); | ||||
763 | _ccv_nnc_graph_rewrap(CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph)); | ||||
764 | } | ||||
765 | } else { | ||||
766 | // For backward graph, no need to evaluate the while expr. | ||||
767 | assert(CO_P(exec)->cmd.cmd == CCV_NNC_GRAPH_BACKWARD)((void) sizeof (((((struct _private_s*)(_privates_))->_co_params ._co_params.exec)->cmd.cmd == CCV_NNC_GRAPH_BACKWARD) ? 1 : 0), __extension__ ({ if ((((struct _private_s*)(_privates_)) ->_co_params._co_params.exec)->cmd.cmd == CCV_NNC_GRAPH_BACKWARD ) ; else __assert_fail ("CO_P(exec)->cmd.cmd == CCV_NNC_GRAPH_BACKWARD" , "ccv_nnc_graph_run.c", 767, __extension__ __PRETTY_FUNCTION__ ); })); | ||||
768 | assert(CO_P(graph)->pair)((void) sizeof (((((struct _private_s*)(_privates_))->_co_params ._co_params.graph)->pair) ? 1 : 0), __extension__ ({ if (( ((struct _private_s*)(_privates_))->_co_params._co_params. graph)->pair) ; else __assert_fail ("CO_P(graph)->pair" , "ccv_nnc_graph_run.c", 768, __extension__ __PRETTY_FUNCTION__ ); })); | ||||
769 | assert(CO_P(tensor_tape))((void) sizeof (((((struct _private_s*)(_privates_))->_co_params ._co_params.tensor_tape)) ? 1 : 0), __extension__ ({ if ((((struct _private_s*)(_privates_))->_co_params._co_params.tensor_tape )) ; else __assert_fail ("CO_P(tensor_tape)", "ccv_nnc_graph_run.c" , 769, __extension__ __PRETTY_FUNCTION__); })); | ||||
770 | CO_V(count)(((struct _private_s*)(_privates_))->count) = 0; | ||||
771 | CO_V(reverse_count)(((struct _private_s*)(_privates_))->reverse_count) = CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph)->while_count = ccv_nnc_tensor_tape_numbering(CO_P(tensor_tape)(((struct _private_s*)(_privates_))->_co_params._co_params .tensor_tape), CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph)->p, (ccv_nnc_graph_exec_t){ | ||||
772 | .d = CO_P(exec_idx)(((struct _private_s*)(_privates_))->_co_params._co_params .exec_idx), | ||||
773 | .graph = CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph)->p, | ||||
774 | }); | ||||
775 | _ccv_nnc_graph_unwrap(CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph), CO_V(count)(((struct _private_s*)(_privates_))->count), CO_V(reverse_count)(((struct _private_s*)(_privates_))->reverse_count)); | ||||
776 | co_apply(_ccv_nnc_graph_exec_run_loop, (CO_P(graph), CO_V(exec_info), CO_V(schd_info), 0, CO_P(graph)->breakpoint_offset, CO_P(graph)->exec_info->rnum, CO_P(tensor_tape), CO_P(flags)))do { _self_->callee = ({ co_routine_t* const task = malloc ((sizeof(co_routine_t) + _ccv_nnc_graph_exec_run_loop_stack_size ())); do { struct _ccv_nnc_graph_exec_run_loop_param_s params = { ._co_params = { (((struct _private_s*)(_privates_))-> _co_params._co_params.graph), (((struct _private_s*)(_privates_ ))->exec_info), (((struct _private_s*)(_privates_))->schd_info ), 0, (((struct _private_s*)(_privates_))->_co_params._co_params .graph)->breakpoint_offset, (((struct _private_s*)(_privates_ ))->_co_params._co_params.graph)->exec_info->rnum, ( ((struct _private_s*)(_privates_))->_co_params._co_params. tensor_tape), (((struct _private_s*)(_privates_))->_co_params ._co_params.flags) } }; task->fn = _ccv_nnc_graph_exec_run_loop ; task->line = 0; task->done = 0; task->root = 0; task ->other_size = 0; task->notify_any = 0; task->others = 0; task->caller = 0; task->callee = 0; if (sizeof(params ) > 0) memcpy(task + 1, ¶ms, sizeof(params)); } while (0); task; }); _co_apply(_self_, _self_->callee); return ( co_state_t){ 776, 0 }; case 776: co_free(_self_->callee); _self_ ->callee = 0; } while (0); | ||||
777 | _ccv_nnc_graph_from_move_transit(CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph)); | ||||
778 | _ccv_nnc_graph_rewrap(CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph)); | ||||
779 | for (CO_V(count)(((struct _private_s*)(_privates_))->count) = 1; CO_V(reverse_count)(((struct _private_s*)(_privates_))->reverse_count) > 0; ++CO_V(count)(((struct _private_s*)(_privates_))->count)) | ||||
780 | { | ||||
781 | CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph)->while_count = --CO_V(reverse_count)(((struct _private_s*)(_privates_))->reverse_count); | ||||
782 | _ccv_nnc_graph_unwrap(CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph), CO_V(count)(((struct _private_s*)(_privates_))->count), CO_V(reverse_count)(((struct _private_s*)(_privates_))->reverse_count)); | ||||
783 | _ccv_nnc_graph_transit_move_to(CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph)); | ||||
784 | co_apply(_ccv_nnc_graph_exec_run_loop, (CO_P(graph), CO_V(exec_info), CO_V(schd_info), 0, 0, CO_P(graph)->exec_info->rnum, CO_P(tensor_tape), CO_P(flags)))do { _self_->callee = ({ co_routine_t* const task = malloc ((sizeof(co_routine_t) + _ccv_nnc_graph_exec_run_loop_stack_size ())); do { struct _ccv_nnc_graph_exec_run_loop_param_s params = { ._co_params = { (((struct _private_s*)(_privates_))-> _co_params._co_params.graph), (((struct _private_s*)(_privates_ ))->exec_info), (((struct _private_s*)(_privates_))->schd_info ), 0, 0, (((struct _private_s*)(_privates_))->_co_params._co_params .graph)->exec_info->rnum, (((struct _private_s*)(_privates_ ))->_co_params._co_params.tensor_tape), (((struct _private_s *)(_privates_))->_co_params._co_params.flags) } }; task-> fn = _ccv_nnc_graph_exec_run_loop; task->line = 0; task-> done = 0; task->root = 0; task->other_size = 0; task-> notify_any = 0; task->others = 0; task->caller = 0; task ->callee = 0; if (sizeof(params) > 0) memcpy(task + 1, & params, sizeof(params)); } while (0); task; }); _co_apply(_self_ , _self_->callee); return (co_state_t){ 784, 0 }; case 784 : co_free(_self_->callee); _self_->callee = 0; } while ( 0); | ||||
785 | _ccv_nnc_graph_from_move_transit(CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph)); | ||||
786 | _ccv_nnc_graph_rewrap(CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph)); | ||||
787 | } | ||||
788 | } | ||||
789 | assert(CO_V(stream_0) == 0)((void) sizeof (((((struct _private_s*)(_privates_))->stream_0 ) == 0) ? 1 : 0), __extension__ ({ if ((((struct _private_s*) (_privates_))->stream_0) == 0) ; else __assert_fail ("CO_V(stream_0) == 0" , "ccv_nnc_graph_run.c", 789, __extension__ __PRETTY_FUNCTION__ ); })); | ||||
790 | int i; | ||||
791 | for (i = 0; i < CO_P(schedule)(((struct _private_s*)(_privates_))->_co_params._co_params .schedule)->wait_size; i++) | ||||
792 | ccv_nnc_stream_context_wait_signal(CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph)->streams[0], CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph)->signals[CO_P(schedule)(((struct _private_s*)(_privates_))->_co_params._co_params .schedule)->waits[i]]); | ||||
793 | } else { | ||||
794 | CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph)->while_count = 0; | ||||
795 | co_apply(_ccv_nnc_graph_exec_run_loop, (CO_P(graph), CO_V(exec_info), CO_V(schd_info), CO_P(schedule)->psort, 0, CO_P(schedule)->psort ? CO_P(schedule)->psort_size : CO_P(schedule)->exec_info_size, CO_P(tensor_tape), CO_P(flags)))do { _self_->callee = ({ co_routine_t* const task = malloc ((sizeof(co_routine_t) + _ccv_nnc_graph_exec_run_loop_stack_size ())); do { struct _ccv_nnc_graph_exec_run_loop_param_s params = { ._co_params = { (((struct _private_s*)(_privates_))-> _co_params._co_params.graph), (((struct _private_s*)(_privates_ ))->exec_info), (((struct _private_s*)(_privates_))->schd_info ), (((struct _private_s*)(_privates_))->_co_params._co_params .schedule)->psort, 0, (((struct _private_s*)(_privates_))-> _co_params._co_params.schedule)->psort ? (((struct _private_s *)(_privates_))->_co_params._co_params.schedule)->psort_size : (((struct _private_s*)(_privates_))->_co_params._co_params .schedule)->exec_info_size, (((struct _private_s*)(_privates_ ))->_co_params._co_params.tensor_tape), (((struct _private_s *)(_privates_))->_co_params._co_params.flags) } }; task-> fn = _ccv_nnc_graph_exec_run_loop; task->line = 0; task-> done = 0; task->root = 0; task->other_size = 0; task-> notify_any = 0; task->others = 0; task->caller = 0; task ->callee = 0; if (sizeof(params) > 0) memcpy(task + 1, & params, sizeof(params)); } while (0); task; }); _co_apply(_self_ , _self_->callee); return (co_state_t){ 795, 0 }; case 795 : co_free(_self_->callee); _self_->callee = 0; } while ( 0); | ||||
796 | PRINT(CCV_CLI_INFO, "Graph Stream %d End", CO_V(stream_0))do { if ((CCV_CLI_INFO & ccv_cli_get_output_levels())) { printf ("Graph Stream %d End", (((struct _private_s*)(_privates_))-> stream_0)); fflush(stdout); } } while (0); | ||||
797 | int i, flag = 0; | ||||
798 | for (i = 0; i < CO_P(schedule)(((struct _private_s*)(_privates_))->_co_params._co_params .schedule)->wait_size; i++) | ||||
799 | { | ||||
800 | ccv_nnc_stream_context_wait_signal(CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph)->streams[CO_V(stream_0)(((struct _private_s*)(_privates_))->stream_0)], CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph)->signals[CO_P(schedule)(((struct _private_s*)(_privates_))->_co_params._co_params .schedule)->waits[i]]); | ||||
801 | if (!flag) | ||||
802 | { | ||||
803 | PRINT(CCV_CLI_INFO, ", Wait: %d", CO_P(schedule)->waits[i])do { if ((CCV_CLI_INFO & ccv_cli_get_output_levels())) { printf (", Wait: %d", (((struct _private_s*)(_privates_))->_co_params ._co_params.schedule)->waits[i]); fflush(stdout); } } while (0); | ||||
804 | flag = 1; | ||||
805 | } else | ||||
806 | PRINT(CCV_CLI_INFO, ", %d", CO_P(schedule)->waits[i])do { if ((CCV_CLI_INFO & ccv_cli_get_output_levels())) { printf (", %d", (((struct _private_s*)(_privates_))->_co_params._co_params .schedule)->waits[i]); fflush(stdout); } } while (0); | ||||
807 | } | ||||
808 | PRINT(CCV_CLI_INFO, "\n")do { if ((CCV_CLI_INFO & ccv_cli_get_output_levels())) { printf ("\n"); fflush(stdout); } } while (0); | ||||
809 | } | ||||
810 | if (CO_P(stream_context)(((struct _private_s*)(_privates_))->_co_params._co_params .stream_context) != CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph)->streams[CO_V(stream_0)(((struct _private_s*)(_privates_))->stream_0)]) | ||||
811 | { | ||||
812 | assert(CO_P(exec_idx) == -1)((void) sizeof (((((struct _private_s*)(_privates_))->_co_params ._co_params.exec_idx) == -1) ? 1 : 0), __extension__ ({ if (( ((struct _private_s*)(_privates_))->_co_params._co_params. exec_idx) == -1) ; else __assert_fail ("CO_P(exec_idx) == -1" , "ccv_nnc_graph_run.c", 812, __extension__ __PRETTY_FUNCTION__ ); })); | ||||
813 | ccv_nnc_stream_context_emit_signal(CO_P(graph)(((struct _private_s*)(_privates_))->_co_params._co_params .graph)->streams[CO_V(stream_0)(((struct _private_s*)(_privates_))->stream_0)], CO_P(schedule)(((struct _private_s*)(_privates_))->_co_params._co_params .schedule)->end); | ||||
814 | ccv_nnc_stream_context_wait_signal(CO_P(stream_context)(((struct _private_s*)(_privates_))->_co_params._co_params .stream_context), CO_P(schedule)(((struct _private_s*)(_privates_))->_co_params._co_params .schedule)->end); | ||||
815 | } | ||||
816 | // Reset main to 0 if it is current me. | ||||
817 | if (CO_P(exec_idx)(((struct _private_s*)(_privates_))->_co_params._co_params .exec_idx) == -1 && CO_P(stream_context)(((struct _private_s*)(_privates_))->_co_params._co_params .stream_context)->main == co_self()(_self_)) | ||||
818 | CO_P(stream_context)(((struct _private_s*)(_privates_))->_co_params._co_params .stream_context)->main = 0; | ||||
819 | } co_end()default: return (co_state_t){ 819, 1 }; } } | ||||
820 | |||||
821 | static int _ccv_nnc_graph_run(ccv_nnc_graph_t* const graph, const int exec_idx, ccv_nnc_graph_exec_info_t* const exec, ccv_nnc_tensor_t* const* const inputs, const int input_size, ccv_nnc_tensor_t* const* const outputs, const int output_size, const int flags, const ccv_nnc_graph_exec_t* const sources, const int source_size, const ccv_nnc_graph_exec_t* const destinations, const int destination_size, ccv_nnc_tensor_tape_t* const tensor_tape, ccv_nnc_stream_context_t* const stream_context); | ||||
822 | |||||
823 | static inline void _ccv_nnc_graph_exec_run(ccv_nnc_graph_t* const graph, ccv_nnc_graph_exec_info_t* const node, const int idx, ccv_nnc_tensor_tape_t* const tensor_tape, ccv_nnc_stream_context_t* const stream_context, const int flags) | ||||
824 | { | ||||
825 | int i; | ||||
826 | _ccv_nnc_graph_exec_unwrap_io(graph, node); | ||||
827 | ccv_nnc_tensor_t** inputs = node->inputs; | ||||
828 | ccv_nnc_tensor_t** outputs = inputs ? inputs + node->input_size : 0; | ||||
829 | if (tensor_tape) | ||||
830 | ccv_nnc_tensor_tape_io(tensor_tape, graph, node->input_flags, inputs, node->input_size, node->output_flags, outputs, node->output_size); | ||||
831 | /* Broadcast the updates to all subscribed references for input / output, even though at th | ||||
832 | * time output is not written yet, propagate pointer change is still valid. */ | ||||
833 | _ccv_nnc_graph_exec_begin_synchronize_multiviews(graph, node); | ||||
834 | if (node->cmd.cmd == CCV_NNC_GRAPH_FORWARD || node->cmd.cmd == CCV_NNC_GRAPH_BACKWARD) | ||||
835 | { | ||||
836 | assert(!stream_context)((void) sizeof ((!stream_context) ? 1 : 0), __extension__ ({ if (!stream_context) ; else __assert_fail ("!stream_context", "ccv_nnc_graph_run.c" , 836, __extension__ __PRETTY_FUNCTION__); })); // This doesn't work properly with stream context. | ||||
837 | if (node->flags & CCV_NNC_GRAPH_EXEC_CASE_OF) | ||||
838 | { | ||||
839 | int ref; | ||||
840 | if (node->cmd.cmd == CCV_NNC_GRAPH_FORWARD) | ||||
841 | { | ||||
842 | ref = node->case_of.offset + node->case_of.expr(inputs, node->input_size, node->case_of.data); | ||||
843 | if (tensor_tape) | ||||
844 | ccv_nnc_tensor_tape_set_numbering(tensor_tape, graph, (ccv_nnc_graph_exec_t){ | ||||
845 | .d = idx, | ||||
846 | .graph = graph, | ||||
847 | }, ref); | ||||
848 | } else { | ||||
849 | assert(node->cmd.cmd == CCV_NNC_GRAPH_BACKWARD)((void) sizeof ((node->cmd.cmd == CCV_NNC_GRAPH_BACKWARD) ? 1 : 0), __extension__ ({ if (node->cmd.cmd == CCV_NNC_GRAPH_BACKWARD ) ; else __assert_fail ("node->cmd.cmd == CCV_NNC_GRAPH_BACKWARD" , "ccv_nnc_graph_run.c", 849, __extension__ __PRETTY_FUNCTION__ ); })); | ||||
850 | assert(tensor_tape)((void) sizeof ((tensor_tape) ? 1 : 0), __extension__ ({ if ( tensor_tape) ; else __assert_fail ("tensor_tape", "ccv_nnc_graph_run.c" , 850, __extension__ __PRETTY_FUNCTION__); })); | ||||
851 | ref = ccv_nnc_tensor_tape_numbering(tensor_tape, graph, (ccv_nnc_graph_exec_t){ | ||||
852 | .d = idx, | ||||
853 | .graph = graph, | ||||
854 | }); | ||||
855 | } | ||||
856 | if (ref >= 0) | ||||
857 | { | ||||
858 | assert(ref < node->graph_ref_size)((void) sizeof ((ref < node->graph_ref_size) ? 1 : 0), __extension__ ({ if (ref < node->graph_ref_size) ; else __assert_fail ("ref < node->graph_ref_size", "ccv_nnc_graph_run.c", 858 , __extension__ __PRETTY_FUNCTION__); })); | ||||
859 | ccv_nnc_graph_t* sub_graph = *(ccv_nnc_graph_t**)ccv_array_get(graph->sub_graphs, CCV_NNC_GRAPH_REF(node)[ref] - 1)((void*)(((char*)((graph->sub_graphs)->data)) + (size_t )(graph->sub_graphs)->rsize * (size_t)(((node)->_heap_graph_ref ? (node)->_heap_graph_ref : (node)->_inline_graph_ref) [ref] - 1))); | ||||
860 | _ccv_nnc_graph_run(sub_graph, idx, node, inputs, node->input_size, outputs, node->output_size, flags, 0, 0, 0, 0, tensor_tape, stream_context); | ||||
861 | } | ||||
862 | _ccv_nnc_graph_exec_unwrap_phi(graph, node, ref); | ||||
863 | } else if (node->flags & CCV_NNC_GRAPH_EXEC_P_WHILE) { | ||||
864 | ccv_nnc_graph_t* sub_graph = *(ccv_nnc_graph_t**)ccv_array_get(graph->sub_graphs, CCV_NNC_GRAPH_REF(node)[0] - 1)((void*)(((char*)((graph->sub_graphs)->data)) + (size_t )(graph->sub_graphs)->rsize * (size_t)(((node)->_heap_graph_ref ? (node)->_heap_graph_ref : (node)->_inline_graph_ref) [0] - 1))); | ||||
865 | _ccv_nnc_graph_run(sub_graph, idx, node, inputs, node->input_size, outputs, node->output_size, flags, 0, 0, 0, 0, tensor_tape, stream_context); | ||||
866 | } | ||||
867 | } else { | ||||
868 | PRINT(CCV_CLI_INFO, "%s [%d]: [%d] -> [%d]\n", ccv_nnc_cmd_name(node->cmd.cmd), idx, node->input_size, node->output_size)do { if ((CCV_CLI_INFO & ccv_cli_get_output_levels())) { printf ("%s [%d]: [%d] -> [%d]\n", ccv_nnc_cmd_name(node->cmd. cmd), idx, node->input_size, node->output_size); fflush (stdout); } } while (0); | ||||
869 | for (i = 0; i < node->input_size; i++) | ||||
870 | { | ||||
871 | PRINT(CCV_CLI_INFO, "|-> %d. %p (%p:%d)", i + 1, inputs[i], (inputs[i] ? inputs[i]->data.u8 : 0), (inputs[i] ? CCV_TENSOR_GET_DEVICE_ID(inputs[i]->info.type) : -1))do { if ((CCV_CLI_INFO & ccv_cli_get_output_levels())) { printf ("|-> %d. %p (%p:%d)", i + 1, inputs[i], (inputs[i] ? inputs [i]->data.u8 : 0), (inputs[i] ? (((inputs[i]->info.type ) & 0xfff00) >> 8) : -1)); fflush(stdout); } } while (0); | ||||
872 | if (inputs[i] && CCV_CLI_OUTPUT_LEVEL_IS(CCV_CLI_INFO)(CCV_CLI_INFO & ccv_cli_get_output_levels())) | ||||
873 | ccv_nnc_print_tensor_info(inputs[i]); | ||||
874 | PRINT(CCV_CLI_INFO, "\n")do { if ((CCV_CLI_INFO & ccv_cli_get_output_levels())) { printf ("\n"); fflush(stdout); } } while (0); | ||||
875 | } | ||||
876 | ccv_nnc_cmd_exec(node->cmd, node->hint, flags, inputs, node->input_size, outputs, node->output_size, stream_context); | ||||
877 | for (i = 0; i < node->output_size; i++) | ||||
878 | { | ||||
879 | PRINT(CCV_CLI_INFO, "|<- %d. %p (%p:%d)", i + 1, outputs[i], (outputs[i] ? outputs[i]->data.u8 : 0), (outputs[i] ? CCV_TENSOR_GET_DEVICE_ID(outputs[i]->info.type) : -1))do { if ((CCV_CLI_INFO & ccv_cli_get_output_levels())) { printf ("|<- %d. %p (%p:%d)", i + 1, outputs[i], (outputs[i] ? outputs [i]->data.u8 : 0), (outputs[i] ? (((outputs[i]->info.type ) & 0xfff00) >> 8) : -1)); fflush(stdout); } } while (0); | ||||
| |||||
880 | if (outputs[i] && CCV_CLI_OUTPUT_LEVEL_IS(CCV_CLI_INFO)(CCV_CLI_INFO & ccv_cli_get_output_levels())) | ||||
881 | ccv_nnc_print_tensor_info(outputs[i]); | ||||
882 | PRINT(CCV_CLI_INFO, "\n")do { if ((CCV_CLI_INFO & ccv_cli_get_output_levels())) { printf ("\n"); fflush(stdout); } } while (0); | ||||
883 | } | ||||
884 | } | ||||
885 | } | ||||
886 | |||||
887 | static inline void _ccv_nnc_graph_topsorted_run(ccv_nnc_graph_t* const graph, const int exec_idx, ccv_nnc_graph_exec_info_t* const exec, const int flags, ccv_nnc_tensor_tape_t* const tensor_tape, ccv_nnc_stream_context_t* const stream_context) | ||||
888 | { | ||||
889 | int i; | ||||
890 | if (exec
| ||||
891 | { | ||||
892 | assert(!stream_context)((void) sizeof ((!stream_context) ? 1 : 0), __extension__ ({ if (!stream_context) ; else __assert_fail ("!stream_context", "ccv_nnc_graph_run.c" , 892, __extension__ __PRETTY_FUNCTION__); })); // This doesn't work properly with stream context. | ||||
893 | assert(exec->p_while.expr)((void) sizeof ((exec->p_while.expr) ? 1 : 0), __extension__ ({ if (exec->p_while.expr) ; else __assert_fail ("exec->p_while.expr" , "ccv_nnc_graph_run.c", 893, __extension__ __PRETTY_FUNCTION__ ); })); | ||||
894 | int64_t count = 0; | ||||
895 | // This is a forward while loop. Backward while loop will just consult its pairing part. | ||||
896 | if (exec->cmd.cmd == CCV_NNC_GRAPH_FORWARD) | ||||
897 | { | ||||
898 | const int graph_breakpoint_size = graph->breakpoint_offset + graph->breakpoint_size; | ||||
899 | for (;; ++count) | ||||
900 | { | ||||
901 | graph->while_count = count; | ||||
902 | if (tensor_tape) | ||||
903 | ccv_nnc_tensor_tape_set_numbering(tensor_tape, graph->p, (ccv_nnc_graph_exec_t){ | ||||
904 | .d = exec_idx, | ||||
905 | .graph = graph->p, | ||||
906 | }, count); | ||||
907 | _ccv_nnc_graph_unwrap(graph, count, 0); | ||||
908 | if (count > 0) | ||||
909 | _ccv_nnc_graph_transit_move_to(graph); | ||||
910 | for (i = 0; i < graph_breakpoint_size; i++) | ||||
911 | _ccv_nnc_graph_exec_run(graph, (ccv_nnc_graph_exec_info_t*)ccv_array_get(graph->exec_info, i)((void*)(((char*)((graph->exec_info)->data)) + (size_t) (graph->exec_info)->rsize * (size_t)(i))), i, tensor_tape, stream_context, flags); | ||||
912 | _ccv_nnc_graph_exec_unwrap_while_expr(graph, exec); | ||||
913 | // Reached breakpoints, now check the breakpoint, if not met, break out. | ||||
914 | if (!exec->p_while.expr(exec->p_while.inputs, exec->p_while.input_size, exec->p_while.data)) | ||||
915 | { | ||||
916 | _ccv_nnc_graph_rewrap(graph); | ||||
917 | break; | ||||
918 | } | ||||
919 | for (i = graph_breakpoint_size; i < graph->exec_info->rnum; i++) | ||||
920 | _ccv_nnc_graph_exec_run(graph, (ccv_nnc_graph_exec_info_t*)ccv_array_get(graph->exec_info, i)((void*)(((char*)((graph->exec_info)->data)) + (size_t) (graph->exec_info)->rsize * (size_t)(i))), i, tensor_tape, stream_context, flags); | ||||
921 | _ccv_nnc_graph_from_move_transit(graph); | ||||
922 | _ccv_nnc_graph_rewrap(graph); | ||||
923 | } | ||||
924 | } else { | ||||
925 | // For backward graph, no need to evaluate the while expr. | ||||
926 | assert(exec->cmd.cmd == CCV_NNC_GRAPH_BACKWARD)((void) sizeof ((exec->cmd.cmd == CCV_NNC_GRAPH_BACKWARD) ? 1 : 0), __extension__ ({ if (exec->cmd.cmd == CCV_NNC_GRAPH_BACKWARD ) ; else __assert_fail ("exec->cmd.cmd == CCV_NNC_GRAPH_BACKWARD" , "ccv_nnc_graph_run.c", 926, __extension__ __PRETTY_FUNCTION__ ); })); | ||||
927 | assert(graph->pair)((void) sizeof ((graph->pair) ? 1 : 0), __extension__ ({ if (graph->pair) ; else __assert_fail ("graph->pair", "ccv_nnc_graph_run.c" , 927, __extension__ __PRETTY_FUNCTION__); })); | ||||
928 | assert(tensor_tape)((void) sizeof ((tensor_tape) ? 1 : 0), __extension__ ({ if ( tensor_tape) ; else __assert_fail ("tensor_tape", "ccv_nnc_graph_run.c" , 928, __extension__ __PRETTY_FUNCTION__); })); | ||||
929 | count = 0; | ||||
930 | int64_t reverse_count = graph->while_count = ccv_nnc_tensor_tape_numbering(tensor_tape, graph->p, (ccv_nnc_graph_exec_t){ | ||||
931 | .d = exec_idx, | ||||
932 | .graph = graph->p, | ||||
933 | }); | ||||
934 | _ccv_nnc_graph_unwrap(graph, count, reverse_count); | ||||
935 | for (i = graph->breakpoint_offset; i < graph->exec_info->rnum; i++) | ||||
936 | _ccv_nnc_graph_exec_run(graph, (ccv_nnc_graph_exec_info_t*)ccv_array_get(graph->exec_info, i)((void*)(((char*)((graph->exec_info)->data)) + (size_t) (graph->exec_info)->rsize * (size_t)(i))), i, tensor_tape, stream_context, flags); | ||||
937 | _ccv_nnc_graph_from_move_transit(graph); | ||||
938 | _ccv_nnc_graph_rewrap(graph); | ||||
939 | for (count = 1; reverse_count > 0; ++count) | ||||
940 | { | ||||
941 | graph->while_count = --reverse_count; | ||||
942 | _ccv_nnc_graph_unwrap(graph, count, reverse_count); | ||||
943 | _ccv_nnc_graph_transit_move_to(graph); | ||||
944 | for (i = 0; i < graph->exec_info->rnum; i++) | ||||
945 | _ccv_nnc_graph_exec_run(graph, (ccv_nnc_graph_exec_info_t*)ccv_array_get(graph->exec_info, i)((void*)(((char*)((graph->exec_info)->data)) + (size_t) (graph->exec_info)->rsize * (size_t)(i))), i, tensor_tape, stream_context, flags); | ||||
946 | _ccv_nnc_graph_from_move_transit(graph); | ||||
947 | _ccv_nnc_graph_rewrap(graph); | ||||
948 | } | ||||
949 | } | ||||
950 | } else { | ||||
951 | graph->while_count = 0; | ||||
952 | for (i = 0; i < graph->exec_info->rnum; i++) | ||||
953 | _ccv_nnc_graph_exec_run(graph, (ccv_nnc_graph_exec_info_t*)ccv_array_get(graph->exec_info, i)((void*)(((char*)((graph->exec_info)->data)) + (size_t) (graph->exec_info)->rsize * (size_t)(i))), i, tensor_tape, stream_context, flags); | ||||
954 | } | ||||
955 | } | ||||
956 | |||||
957 | static inline void _ccv_nnc_graph_run_slow_path(ccv_nnc_graph_t* const graph, const int exec_idx, ccv_nnc_graph_exec_info_t* const exec, ccv_nnc_tensor_t* const* const inputs, const int input_size, ccv_nnc_tensor_t* const* const outputs, const int output_size, const int flags, const ccv_nnc_graph_exec_t* const sources, const int source_size, const ccv_nnc_graph_exec_t* const destinations, const int destination_size, ccv_nnc_tensor_tape_t* const tensor_tape, ccv_nnc_stream_context_t* const stream_context) | ||||
958 | { | ||||
959 | int i, j; | ||||
960 | const ccv_nnc_graph_exec_t* const graph_sources = sources ? sources : (ccv_nnc_graph_exec_t*)ccv_array_get(graph->sources, 0)((void*)(((char*)((graph->sources)->data)) + (size_t)(graph ->sources)->rsize * (size_t)(0))); | ||||
961 | const int graph_source_size = source_size ? source_size : graph->sources->rnum; | ||||
962 | const ccv_nnc_graph_exec_t* const graph_destinations = destinations ? destinations : (ccv_nnc_graph_exec_t*)ccv_array_get(graph->destinations, 0)((void*)(((char*)((graph->destinations)->data)) + (size_t )(graph->destinations)->rsize * (size_t)(0))); | ||||
963 | const int graph_destination_size = destination_size ? destination_size : graph->destinations->rnum; | ||||
964 | #define visitor(node, idx, ...) \ | ||||
965 | _ccv_nnc_graph_exec_run(graph, node, idx, tensor_tape, stream_context, flags) | ||||
966 | if (exec && (exec->flags & CCV_NNC_GRAPH_EXEC_P_WHILE)) | ||||
967 | { | ||||
968 | assert(!stream_context)((void) sizeof ((!stream_context) ? 1 : 0), __extension__ ({ if (!stream_context) ; else __assert_fail ("!stream_context", "ccv_nnc_graph_run.c" , 968, __extension__ __PRETTY_FUNCTION__); })); // This doesn't work properly with stream context. | ||||
969 | assert(exec->p_while.expr)((void) sizeof ((exec->p_while.expr) ? 1 : 0), __extension__ ({ if (exec->p_while.expr) ; else __assert_fail ("exec->p_while.expr" , "ccv_nnc_graph_run.c", 969, __extension__ __PRETTY_FUNCTION__ ); })); | ||||
970 | int64_t count = 0; | ||||
971 | // This is a forward while loop. Backward while loop will just consult its pairing part. | ||||
972 | if (exec->cmd.cmd == CCV_NNC_GRAPH_FORWARD) | ||||
973 | { | ||||
974 | ccv_array_t* follows = ccv_array_new(sizeof(ccv_nnc_graph_exec_t), graph->breakpoint_size, 0); | ||||
975 | for (i = 0; i < graph->breakpoint_size; i++) | ||||
976 | { | ||||
977 | const ccv_nnc_graph_exec_info_t* const exec_info = (const ccv_nnc_graph_exec_info_t*)ccv_array_get(graph->exec_info, graph->breakpoints->d)((void*)(((char*)((graph->exec_info)->data)) + (size_t) (graph->exec_info)->rsize * (size_t)(graph->breakpoints ->d))); | ||||
978 | if (exec_info->outgoings) | ||||
979 | for (j = 0; j < exec_info->outgoings->rnum; j++) | ||||
980 | { | ||||
981 | const ccv_nnc_graph_exec_t exec = { | ||||
982 | .d = *(int*)ccv_array_get(exec_info->outgoings, j)((void*)(((char*)((exec_info->outgoings)->data)) + (size_t )(exec_info->outgoings)->rsize * (size_t)(j))), | ||||
983 | .graph = graph, | ||||
984 | }; | ||||
985 | ccv_array_push(follows, &exec); | ||||
986 | } | ||||
987 | } | ||||
988 | for (;; ++count) | ||||
989 | { | ||||
990 | graph->while_count = count; | ||||
991 | if (tensor_tape) | ||||
992 | ccv_nnc_tensor_tape_set_numbering(tensor_tape, graph->p, (ccv_nnc_graph_exec_t){ | ||||
993 | .d = exec_idx, | ||||
994 | .graph = graph->p, | ||||
995 | }, count); | ||||
996 | _ccv_nnc_graph_unwrap(graph, count, 0); | ||||
997 | if (count > 0) | ||||
998 | _ccv_nnc_graph_transit_move_to(graph); | ||||
999 | CCV_NNC_GRAPH_VISIT(graph, (ccv_nnc_graph_exec_info_t*)ccv_array_get(graph->exec_info, 0), graph->exec_info->rnum, graph_sources, graph_source_size, graph->breakpoints, graph->breakpoint_size, 0, visitor)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_ < (graph->exec_info->rnum); _i_ ++) _incoming_edges_ += (((ccv_nnc_graph_exec_info_t*)((void* )(((char*)((graph->exec_info)->data)) + (size_t)(graph-> exec_info)->rsize * (size_t)(0))))[_i_].outgoings) ? ((ccv_nnc_graph_exec_info_t *)((void*)(((char*)((graph->exec_info)->data)) + (size_t )(graph->exec_info)->rsize * (size_t)(0))))[_i_].outgoings ->rnum : 0; const int _heap_mem_ = ((graph->exec_info-> rnum) + _incoming_edges_ > 1024); ccv_nnc_incoming_t* _incomings_ ; if (_heap_mem_) _incomings_ = (ccv_nnc_incoming_t*)malloc(sizeof (ccv_nnc_incoming_t) * (graph->exec_info->rnum) + sizeof (int32_t) * ((graph->exec_info->rnum) * 2 + _incoming_edges_ )); else _incomings_ = (ccv_nnc_incoming_t*)__builtin_alloca ( sizeof(ccv_nnc_incoming_t) * (graph->exec_info->rnum) + sizeof(int32_t) * ((graph->exec_info->rnum) * 2 + _incoming_edges_ )); memset(_incomings_, 0, sizeof(ccv_nnc_incoming_t) * (graph ->exec_info->rnum)); int32_t* _exists_[2] = { (int32_t* )(_incomings_ + (graph->exec_info->rnum)), (int32_t*)(_incomings_ + (graph->exec_info->rnum)) + (graph->exec_info-> rnum), }; int32_t* const _edges_ = _exists_[1] + (graph->exec_info ->rnum); for (_i_ = 0; _i_ < (graph_source_size); _i_++ ) { ((void) sizeof (((graph_sources)[_i_].graph == graph) ? 1 : 0), __extension__ ({ if ((graph_sources)[_i_].graph == graph ) ; else __assert_fail ("(graph_sources)[_i_].graph == graph" , "ccv_nnc_graph_run.c", 999, __extension__ __PRETTY_FUNCTION__ ); })); _incomings_[(graph_sources)[_i_].d].r = 1; _exists_[0 ][_i_] = (graph_sources)[_i_].d; } int _exist_size_[2] = { (graph_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 (((ccv_nnc_graph_exec_info_t*)((void*)(((char*)((graph-> exec_info)->data)) + (size_t)(graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings) for (_j_ = 0; _j_ < (( ccv_nnc_graph_exec_info_t*)((void*)(((char*)((graph->exec_info )->data)) + (size_t)(graph->exec_info)->rsize * (size_t )(0))))[_idx_].outgoings->rnum; _j_++) { const int d = *(int *)((void*)(((char*)((((ccv_nnc_graph_exec_info_t*)((void*)((( char*)((graph->exec_info)->data)) + (size_t)(graph-> exec_info)->rsize * (size_t)(0))))[_idx_].outgoings)->data )) + (size_t)(((ccv_nnc_graph_exec_info_t*)((void*)(((char*)( (graph->exec_info)->data)) + (size_t)(graph->exec_info )->rsize * (size_t)(0))))[_idx_].outgoings)->rsize * (size_t )(_j_))); ++_incomings_[d].c; if (_incomings_[d].r != 0) continue ; _incomings_[d].r = 1; ((void) sizeof ((_exist_size_[_q_] < (graph->exec_info->rnum)) ? 1 : 0), __extension__ ({ if (_exist_size_[_q_] < (graph->exec_info->rnum)) ; else __assert_fail ("_exist_size_[_q_] < (graph->exec_info->rnum)" , "ccv_nnc_graph_run.c", 999, __extension__ __PRETTY_FUNCTION__ ); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[ _q_]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for (_i_ = 0; _i_ < (graph_source_size); _i_++) { ((void) sizeof (((graph_sources)[_i_].graph == graph) ? 1 : 0), __extension__ ({ if ((graph_sources)[_i_].graph == graph) ; else __assert_fail ("(graph_sources)[_i_].graph == graph", "ccv_nnc_graph_run.c" , 999, __extension__ __PRETTY_FUNCTION__); })); _incomings_[( graph_sources)[_i_].d].r = 3; _exists_[0][_i_] = (graph_sources )[_i_].d; } _exist_size_[0] = (graph_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 (((ccv_nnc_graph_exec_info_t*)((void*)(((char *)((graph->exec_info)->data)) + (size_t)(graph->exec_info )->rsize * (size_t)(0))))[_idx_].outgoings) for (_j_ = 0; _j_ < ((ccv_nnc_graph_exec_info_t*)((void*)(((char*)((graph-> exec_info)->data)) + (size_t)(graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings->rnum; _j_++) { const int d = *(int*)((void*)(((char*)((((ccv_nnc_graph_exec_info_t*)( (void*)(((char*)((graph->exec_info)->data)) + (size_t)( graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings )->data)) + (size_t)(((ccv_nnc_graph_exec_info_t*)((void*) (((char*)((graph->exec_info)->data)) + (size_t)(graph-> exec_info)->rsize * (size_t)(0))))[_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_] < (graph->exec_info->rnum)) ? 1 : 0), __extension__ ({ if (_exist_size_[_q_] < (graph->exec_info->rnum)) ; else __assert_fail ("_exist_size_[_q_] < (graph->exec_info->rnum)" , "ccv_nnc_graph_run.c", 999, __extension__ __PRETTY_FUNCTION__ ); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[ _q_]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for (_i_ = 0; _i_ < (graph->breakpoint_size); _i_++) { ((void ) sizeof (((graph->breakpoints)[_i_].graph == graph) ? 1 : 0), __extension__ ({ if ((graph->breakpoints)[_i_].graph == graph) ; else __assert_fail ("(graph->breakpoints)[_i_].graph == graph" , "ccv_nnc_graph_run.c", 999, __extension__ __PRETTY_FUNCTION__ ); })); _incomings_[(graph->breakpoints)[_i_].d].r = 5; _exists_ [0][_i_] = (graph->breakpoints)[_i_].d; } _exist_size_[0] = (graph->breakpoint_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_] < (graph->exec_info->rnum)) ? 1 : 0 ), __extension__ ({ if (_exist_size_[_q_] < (graph->exec_info ->rnum)) ; else __assert_fail ("_exist_size_[_q_] < (graph->exec_info->rnum)" , "ccv_nnc_graph_run.c", 999, __extension__ __PRETTY_FUNCTION__ ); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[ _q_]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for (_i_ = 0; _i_ < (graph->breakpoint_size); _i_++) { ((void ) sizeof (((graph->breakpoints)[_i_].graph == graph) ? 1 : 0), __extension__ ({ if ((graph->breakpoints)[_i_].graph == graph) ; else __assert_fail ("(graph->breakpoints)[_i_].graph == graph" , "ccv_nnc_graph_run.c", 999, __extension__ __PRETTY_FUNCTION__ ); })); _incomings_[(graph->breakpoints)[_i_].d].d = 1; } for (_i_ = 0; _i_ < (graph_source_size); _i_++) { ((void) sizeof (((graph_sources)[_i_].graph == graph) ? 1 : 0), __extension__ ({ if ((graph_sources)[_i_].graph == graph) ; else __assert_fail ("(graph_sources)[_i_].graph == graph", "ccv_nnc_graph_run.c" , 999, __extension__ __PRETTY_FUNCTION__); })); _exists_[0][_i_ ] = (graph_sources)[_i_].d; } _p_ = 0; _q_ = 1; _exist_size_[ 0] = (graph_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_]; visitor((((ccv_nnc_graph_exec_info_t*)((void*)((( char*)((graph->exec_info)->data)) + (size_t)(graph-> exec_info)->rsize * (size_t)(0)))) + _idx_), (_idx_), (_incomings_ [_idx_].d)); if (_incomings_[_idx_].d) { ++_d_; _incomings_[_idx_ ].r = 7; } if (((ccv_nnc_graph_exec_info_t*)((void*)(((char*) ((graph->exec_info)->data)) + (size_t)(graph->exec_info )->rsize * (size_t)(0))))[_idx_].outgoings) { if (((ccv_nnc_graph_exec_info_t *)((void*)(((char*)((graph->exec_info)->data)) + (size_t )(graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings ->rnum == 1) { const int d = *(int*)((void*)(((char*)((((ccv_nnc_graph_exec_info_t *)((void*)(((char*)((graph->exec_info)->data)) + (size_t )(graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings )->data)) + (size_t)(((ccv_nnc_graph_exec_info_t*)((void*) (((char*)((graph->exec_info)->data)) + (size_t)(graph-> exec_info)->rsize * (size_t)(0))))[_idx_].outgoings)->rsize * (size_t)(0))); --_incomings_[d].c; if (_incomings_[d].c == 0 && _incomings_[d].r == 6 && _d_ < (graph ->breakpoint_size)) { _exists_[_p_][_i_] = d; continue; } } else for (_j_ = 0; _j_ < ((ccv_nnc_graph_exec_info_t*)((void *)(((char*)((graph->exec_info)->data)) + (size_t)(graph ->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings-> rnum; _j_++) { const int d = *(int*)((void*)(((char*)((((ccv_nnc_graph_exec_info_t *)((void*)(((char*)((graph->exec_info)->data)) + (size_t )(graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings )->data)) + (size_t)(((ccv_nnc_graph_exec_info_t*)((void*) (((char*)((graph->exec_info)->data)) + (size_t)(graph-> exec_info)->rsize * (size_t)(0))))[_idx_].outgoings)->rsize * (size_t)(_j_))); --_incomings_[d].c; if (_incomings_[d].c == 0 && _incomings_[d].r == 6 && _d_ < (graph ->breakpoint_size)) { ((void) sizeof ((_exist_size_[_q_] < (graph->exec_info->rnum)) ? 1 : 0), __extension__ ({ if (_exist_size_[_q_] < (graph->exec_info->rnum)) ; else __assert_fail ("_exist_size_[_q_] < (graph->exec_info->rnum)" , "ccv_nnc_graph_run.c", 999, __extension__ __PRETTY_FUNCTION__ ); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[ _q_]; } } } ++_i_; } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = ( _i_)); } for (_i_ = 0; _i_ < (graph->breakpoint_size); _i_ ++) { ((void) sizeof (((graph->breakpoints)[_i_].graph == graph ) ? 1 : 0), __extension__ ({ if ((graph->breakpoints)[_i_] .graph == graph) ; else __assert_fail ("(graph->breakpoints)[_i_].graph == graph" , "ccv_nnc_graph_run.c", 999, __extension__ __PRETTY_FUNCTION__ ); })); if (_incomings_[(graph->breakpoints)[_i_].d].r == 7 ) continue; if (!(0)) { ((void) sizeof ((_incomings_[(graph-> breakpoints)[_i_].d].c == 0) ? 1 : 0), __extension__ ({ if (_incomings_ [(graph->breakpoints)[_i_].d].c == 0) ; else __assert_fail ("_incomings_[(graph->breakpoints)[_i_].d].c == 0", "ccv_nnc_graph_run.c" , 999, __extension__ __PRETTY_FUNCTION__); })); } else if (_incomings_ [(graph->breakpoints)[_i_].d].c > 0) continue; visitor( (((ccv_nnc_graph_exec_info_t*)((void*)(((char*)((graph->exec_info )->data)) + (size_t)(graph->exec_info)->rsize * (size_t )(0)))) + (graph->breakpoints)[_i_].d), ((graph->breakpoints )[_i_].d), (_incomings_[(graph->breakpoints)[_i_].d].d)); } if (_heap_mem_) free(_incomings_); } while (0);; | ||||
1000 | _ccv_nnc_graph_exec_unwrap_while_expr(graph, exec); | ||||
1001 | // Reached breakpoints, now check the breakpoint, if not met, break out. | ||||
1002 | if (!exec->p_while.expr(exec->p_while.inputs, exec->p_while.input_size, exec->p_while.data)) | ||||
1003 | { | ||||
1004 | _ccv_nnc_graph_rewrap(graph); | ||||
1005 | break; | ||||
1006 | } | ||||
1007 | if (follows->rnum > 0) | ||||
1008 | CCV_NNC_GRAPH_VISIT(graph, (ccv_nnc_graph_exec_info_t*)ccv_array_get(graph->exec_info, 0), graph->exec_info->rnum, (ccv_nnc_graph_exec_t*)ccv_array_get(follows, 0), follows->rnum, graph_destinations, graph_destination_size, 0, visitor)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_ < (graph->exec_info->rnum); _i_ ++) _incoming_edges_ += (((ccv_nnc_graph_exec_info_t*)((void* )(((char*)((graph->exec_info)->data)) + (size_t)(graph-> exec_info)->rsize * (size_t)(0))))[_i_].outgoings) ? ((ccv_nnc_graph_exec_info_t *)((void*)(((char*)((graph->exec_info)->data)) + (size_t )(graph->exec_info)->rsize * (size_t)(0))))[_i_].outgoings ->rnum : 0; const int _heap_mem_ = ((graph->exec_info-> rnum) + _incoming_edges_ > 1024); ccv_nnc_incoming_t* _incomings_ ; if (_heap_mem_) _incomings_ = (ccv_nnc_incoming_t*)malloc(sizeof (ccv_nnc_incoming_t) * (graph->exec_info->rnum) + sizeof (int32_t) * ((graph->exec_info->rnum) * 2 + _incoming_edges_ )); else _incomings_ = (ccv_nnc_incoming_t*)__builtin_alloca ( sizeof(ccv_nnc_incoming_t) * (graph->exec_info->rnum) + sizeof(int32_t) * ((graph->exec_info->rnum) * 2 + _incoming_edges_ )); memset(_incomings_, 0, sizeof(ccv_nnc_incoming_t) * (graph ->exec_info->rnum)); int32_t* _exists_[2] = { (int32_t* )(_incomings_ + (graph->exec_info->rnum)), (int32_t*)(_incomings_ + (graph->exec_info->rnum)) + (graph->exec_info-> rnum), }; int32_t* const _edges_ = _exists_[1] + (graph->exec_info ->rnum); for (_i_ = 0; _i_ < (follows->rnum); _i_++) { ((void) sizeof ((((ccv_nnc_graph_exec_t*)((void*)(((char*) ((follows)->data)) + (size_t)(follows)->rsize * (size_t )(0))))[_i_].graph == graph) ? 1 : 0), __extension__ ({ if (( (ccv_nnc_graph_exec_t*)((void*)(((char*)((follows)->data)) + (size_t)(follows)->rsize * (size_t)(0))))[_i_].graph == graph) ; else __assert_fail ("((ccv_nnc_graph_exec_t*)((void*)(((char*)((follows)->data)) + (size_t)(follows)->rsize * (size_t)(0))))[_i_].graph == graph" , "ccv_nnc_graph_run.c", 1008, __extension__ __PRETTY_FUNCTION__ ); })); _incomings_[((ccv_nnc_graph_exec_t*)((void*)(((char*) ((follows)->data)) + (size_t)(follows)->rsize * (size_t )(0))))[_i_].d].r = 1; _exists_[0][_i_] = ((ccv_nnc_graph_exec_t *)((void*)(((char*)((follows)->data)) + (size_t)(follows)-> rsize * (size_t)(0))))[_i_].d; } int _exist_size_[2] = { (follows ->rnum), 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 (((ccv_nnc_graph_exec_info_t*)((void*)(((char*)((graph-> exec_info)->data)) + (size_t)(graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings) for (_j_ = 0; _j_ < (( ccv_nnc_graph_exec_info_t*)((void*)(((char*)((graph->exec_info )->data)) + (size_t)(graph->exec_info)->rsize * (size_t )(0))))[_idx_].outgoings->rnum; _j_++) { const int d = *(int *)((void*)(((char*)((((ccv_nnc_graph_exec_info_t*)((void*)((( char*)((graph->exec_info)->data)) + (size_t)(graph-> exec_info)->rsize * (size_t)(0))))[_idx_].outgoings)->data )) + (size_t)(((ccv_nnc_graph_exec_info_t*)((void*)(((char*)( (graph->exec_info)->data)) + (size_t)(graph->exec_info )->rsize * (size_t)(0))))[_idx_].outgoings)->rsize * (size_t )(_j_))); ++_incomings_[d].c; if (_incomings_[d].r != 0) continue ; _incomings_[d].r = 1; ((void) sizeof ((_exist_size_[_q_] < (graph->exec_info->rnum)) ? 1 : 0), __extension__ ({ if (_exist_size_[_q_] < (graph->exec_info->rnum)) ; else __assert_fail ("_exist_size_[_q_] < (graph->exec_info->rnum)" , "ccv_nnc_graph_run.c", 1008, __extension__ __PRETTY_FUNCTION__ ); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[ _q_]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for (_i_ = 0; _i_ < (follows->rnum); _i_++) { ((void) sizeof ((((ccv_nnc_graph_exec_t*)((void*)(((char*)((follows)->data )) + (size_t)(follows)->rsize * (size_t)(0))))[_i_].graph == graph) ? 1 : 0), __extension__ ({ if (((ccv_nnc_graph_exec_t *)((void*)(((char*)((follows)->data)) + (size_t)(follows)-> rsize * (size_t)(0))))[_i_].graph == graph) ; else __assert_fail ("((ccv_nnc_graph_exec_t*)((void*)(((char*)((follows)->data)) + (size_t)(follows)->rsize * (size_t)(0))))[_i_].graph == graph" , "ccv_nnc_graph_run.c", 1008, __extension__ __PRETTY_FUNCTION__ ); })); _incomings_[((ccv_nnc_graph_exec_t*)((void*)(((char*) ((follows)->data)) + (size_t)(follows)->rsize * (size_t )(0))))[_i_].d].r = 3; _exists_[0][_i_] = ((ccv_nnc_graph_exec_t *)((void*)(((char*)((follows)->data)) + (size_t)(follows)-> rsize * (size_t)(0))))[_i_].d; } _exist_size_[0] = (follows-> rnum); _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 (((ccv_nnc_graph_exec_info_t*) ((void*)(((char*)((graph->exec_info)->data)) + (size_t) (graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings ) for (_j_ = 0; _j_ < ((ccv_nnc_graph_exec_info_t*)((void* )(((char*)((graph->exec_info)->data)) + (size_t)(graph-> exec_info)->rsize * (size_t)(0))))[_idx_].outgoings->rnum ; _j_++) { const int d = *(int*)((void*)(((char*)((((ccv_nnc_graph_exec_info_t *)((void*)(((char*)((graph->exec_info)->data)) + (size_t )(graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings )->data)) + (size_t)(((ccv_nnc_graph_exec_info_t*)((void*) (((char*)((graph->exec_info)->data)) + (size_t)(graph-> exec_info)->rsize * (size_t)(0))))[_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_] < (graph->exec_info->rnum)) ? 1 : 0), __extension__ ({ if (_exist_size_[_q_] < (graph->exec_info->rnum)) ; else __assert_fail ("_exist_size_[_q_] < (graph->exec_info->rnum)" , "ccv_nnc_graph_run.c", 1008, __extension__ __PRETTY_FUNCTION__ ); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[ _q_]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for (_i_ = 0; _i_ < (graph_destination_size); _i_++) { ((void ) sizeof (((graph_destinations)[_i_].graph == graph) ? 1 : 0) , __extension__ ({ if ((graph_destinations)[_i_].graph == graph ) ; else __assert_fail ("(graph_destinations)[_i_].graph == graph" , "ccv_nnc_graph_run.c", 1008, __extension__ __PRETTY_FUNCTION__ ); })); _incomings_[(graph_destinations)[_i_].d].r = 5; _exists_ [0][_i_] = (graph_destinations)[_i_].d; } _exist_size_[0] = ( graph_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_] < (graph->exec_info->rnum)) ? 1 : 0 ), __extension__ ({ if (_exist_size_[_q_] < (graph->exec_info ->rnum)) ; else __assert_fail ("_exist_size_[_q_] < (graph->exec_info->rnum)" , "ccv_nnc_graph_run.c", 1008, __extension__ __PRETTY_FUNCTION__ ); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[ _q_]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for (_i_ = 0; _i_ < (graph_destination_size); _i_++) { ((void ) sizeof (((graph_destinations)[_i_].graph == graph) ? 1 : 0) , __extension__ ({ if ((graph_destinations)[_i_].graph == graph ) ; else __assert_fail ("(graph_destinations)[_i_].graph == graph" , "ccv_nnc_graph_run.c", 1008, __extension__ __PRETTY_FUNCTION__ ); })); _incomings_[(graph_destinations)[_i_].d].d = 1; } for (_i_ = 0; _i_ < (follows->rnum); _i_++) { ((void) sizeof ((((ccv_nnc_graph_exec_t*)((void*)(((char*)((follows)->data )) + (size_t)(follows)->rsize * (size_t)(0))))[_i_].graph == graph) ? 1 : 0), __extension__ ({ if (((ccv_nnc_graph_exec_t *)((void*)(((char*)((follows)->data)) + (size_t)(follows)-> rsize * (size_t)(0))))[_i_].graph == graph) ; else __assert_fail ("((ccv_nnc_graph_exec_t*)((void*)(((char*)((follows)->data)) + (size_t)(follows)->rsize * (size_t)(0))))[_i_].graph == graph" , "ccv_nnc_graph_run.c", 1008, __extension__ __PRETTY_FUNCTION__ ); })); _exists_[0][_i_] = ((ccv_nnc_graph_exec_t*)((void*)(( (char*)((follows)->data)) + (size_t)(follows)->rsize * ( size_t)(0))))[_i_].d; } _p_ = 0; _q_ = 1; _exist_size_[0] = ( follows->rnum); _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_ ]; visitor((((ccv_nnc_graph_exec_info_t*)((void*)(((char*)((graph ->exec_info)->data)) + (size_t)(graph->exec_info)-> rsize * (size_t)(0)))) + _idx_), (_idx_), (_incomings_[_idx_] .d)); if (_incomings_[_idx_].d) { ++_d_; _incomings_[_idx_].r = 7; } if (((ccv_nnc_graph_exec_info_t*)((void*)(((char*)((graph ->exec_info)->data)) + (size_t)(graph->exec_info)-> rsize * (size_t)(0))))[_idx_].outgoings) { if (((ccv_nnc_graph_exec_info_t *)((void*)(((char*)((graph->exec_info)->data)) + (size_t )(graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings ->rnum == 1) { const int d = *(int*)((void*)(((char*)((((ccv_nnc_graph_exec_info_t *)((void*)(((char*)((graph->exec_info)->data)) + (size_t )(graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings )->data)) + (size_t)(((ccv_nnc_graph_exec_info_t*)((void*) (((char*)((graph->exec_info)->data)) + (size_t)(graph-> exec_info)->rsize * (size_t)(0))))[_idx_].outgoings)->rsize * (size_t)(0))); --_incomings_[d].c; if (_incomings_[d].c == 0 && _incomings_[d].r == 6 && _d_ < (graph_destination_size )) { _exists_[_p_][_i_] = d; continue; } } else for (_j_ = 0; _j_ < ((ccv_nnc_graph_exec_info_t*)((void*)(((char*)((graph ->exec_info)->data)) + (size_t)(graph->exec_info)-> rsize * (size_t)(0))))[_idx_].outgoings->rnum; _j_++) { const int d = *(int*)((void*)(((char*)((((ccv_nnc_graph_exec_info_t *)((void*)(((char*)((graph->exec_info)->data)) + (size_t )(graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings )->data)) + (size_t)(((ccv_nnc_graph_exec_info_t*)((void*) (((char*)((graph->exec_info)->data)) + (size_t)(graph-> exec_info)->rsize * (size_t)(0))))[_idx_].outgoings)->rsize * (size_t)(_j_))); --_incomings_[d].c; if (_incomings_[d].c == 0 && _incomings_[d].r == 6 && _d_ < (graph_destination_size )) { ((void) sizeof ((_exist_size_[_q_] < (graph->exec_info ->rnum)) ? 1 : 0), __extension__ ({ if (_exist_size_[_q_] < (graph->exec_info->rnum)) ; else __assert_fail ("_exist_size_[_q_] < (graph->exec_info->rnum)" , "ccv_nnc_graph_run.c", 1008, __extension__ __PRETTY_FUNCTION__ ); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[ _q_]; } } } ++_i_; } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = ( _i_)); } for (_i_ = 0; _i_ < (graph_destination_size); _i_ ++) { ((void) sizeof (((graph_destinations)[_i_].graph == graph ) ? 1 : 0), __extension__ ({ if ((graph_destinations)[_i_].graph == graph) ; else __assert_fail ("(graph_destinations)[_i_].graph == graph" , "ccv_nnc_graph_run.c", 1008, __extension__ __PRETTY_FUNCTION__ ); })); if (_incomings_[(graph_destinations)[_i_].d].r == 7) continue ; if (!(0)) { ((void) sizeof ((_incomings_[(graph_destinations )[_i_].d].c == 0) ? 1 : 0), __extension__ ({ if (_incomings_[ (graph_destinations)[_i_].d].c == 0) ; else __assert_fail ("_incomings_[(graph_destinations)[_i_].d].c == 0" , "ccv_nnc_graph_run.c", 1008, __extension__ __PRETTY_FUNCTION__ ); })); } else if (_incomings_[(graph_destinations)[_i_].d].c > 0) continue; visitor((((ccv_nnc_graph_exec_info_t*)((void *)(((char*)((graph->exec_info)->data)) + (size_t)(graph ->exec_info)->rsize * (size_t)(0)))) + (graph_destinations )[_i_].d), ((graph_destinations)[_i_].d), (_incomings_[(graph_destinations )[_i_].d].d)); } if (_heap_mem_) free(_incomings_); } while ( 0);; | ||||
1009 | _ccv_nnc_graph_from_move_transit(graph); | ||||
1010 | _ccv_nnc_graph_rewrap(graph); | ||||
1011 | } | ||||
1012 | ccv_array_free(follows); | ||||
1013 | } else { | ||||
1014 | // For backward graph, no need to evaluate the while expr. | ||||
1015 | assert(exec->cmd.cmd == CCV_NNC_GRAPH_BACKWARD)((void) sizeof ((exec->cmd.cmd == CCV_NNC_GRAPH_BACKWARD) ? 1 : 0), __extension__ ({ if (exec->cmd.cmd == CCV_NNC_GRAPH_BACKWARD ) ; else __assert_fail ("exec->cmd.cmd == CCV_NNC_GRAPH_BACKWARD" , "ccv_nnc_graph_run.c", 1015, __extension__ __PRETTY_FUNCTION__ ); })); | ||||
1016 | assert(graph->pair)((void) sizeof ((graph->pair) ? 1 : 0), __extension__ ({ if (graph->pair) ; else __assert_fail ("graph->pair", "ccv_nnc_graph_run.c" , 1016, __extension__ __PRETTY_FUNCTION__); })); | ||||
1017 | assert(tensor_tape)((void) sizeof ((tensor_tape) ? 1 : 0), __extension__ ({ if ( tensor_tape) ; else __assert_fail ("tensor_tape", "ccv_nnc_graph_run.c" , 1017, __extension__ __PRETTY_FUNCTION__); })); | ||||
1018 | count = 0; | ||||
1019 | int64_t reverse_count = graph->while_count = ccv_nnc_tensor_tape_numbering(tensor_tape, graph->p, (ccv_nnc_graph_exec_t){ | ||||
1020 | .d = exec_idx, | ||||
1021 | .graph = graph->p, | ||||
1022 | }); | ||||
1023 | _ccv_nnc_graph_unwrap(graph, count, reverse_count); | ||||
1024 | CCV_NNC_GRAPH_VISIT(graph, (ccv_nnc_graph_exec_info_t*)ccv_array_get(graph->exec_info, 0), graph->exec_info->rnum, graph->breakpoints, graph->breakpoint_size, graph_destinations, graph_destination_size, 1, visitor)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_ < (graph->exec_info->rnum); _i_ ++) _incoming_edges_ += (((ccv_nnc_graph_exec_info_t*)((void* )(((char*)((graph->exec_info)->data)) + (size_t)(graph-> exec_info)->rsize * (size_t)(0))))[_i_].outgoings) ? ((ccv_nnc_graph_exec_info_t *)((void*)(((char*)((graph->exec_info)->data)) + (size_t )(graph->exec_info)->rsize * (size_t)(0))))[_i_].outgoings ->rnum : 0; const int _heap_mem_ = ((graph->exec_info-> rnum) + _incoming_edges_ > 1024); ccv_nnc_incoming_t* _incomings_ ; if (_heap_mem_) _incomings_ = (ccv_nnc_incoming_t*)malloc(sizeof (ccv_nnc_incoming_t) * (graph->exec_info->rnum) + sizeof (int32_t) * ((graph->exec_info->rnum) * 2 + _incoming_edges_ )); else _incomings_ = (ccv_nnc_incoming_t*)__builtin_alloca ( sizeof(ccv_nnc_incoming_t) * (graph->exec_info->rnum) + sizeof(int32_t) * ((graph->exec_info->rnum) * 2 + _incoming_edges_ )); memset(_incomings_, 0, sizeof(ccv_nnc_incoming_t) * (graph ->exec_info->rnum)); int32_t* _exists_[2] = { (int32_t* )(_incomings_ + (graph->exec_info->rnum)), (int32_t*)(_incomings_ + (graph->exec_info->rnum)) + (graph->exec_info-> rnum), }; int32_t* const _edges_ = _exists_[1] + (graph->exec_info ->rnum); for (_i_ = 0; _i_ < (graph->breakpoint_size ); _i_++) { ((void) sizeof (((graph->breakpoints)[_i_].graph == graph) ? 1 : 0), __extension__ ({ if ((graph->breakpoints )[_i_].graph == graph) ; else __assert_fail ("(graph->breakpoints)[_i_].graph == graph" , "ccv_nnc_graph_run.c", 1024, __extension__ __PRETTY_FUNCTION__ ); })); _incomings_[(graph->breakpoints)[_i_].d].r = 1; _exists_ [0][_i_] = (graph->breakpoints)[_i_].d; } int _exist_size_ [2] = { (graph->breakpoint_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 (((ccv_nnc_graph_exec_info_t*) ((void*)(((char*)((graph->exec_info)->data)) + (size_t) (graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings ) for (_j_ = 0; _j_ < ((ccv_nnc_graph_exec_info_t*)((void* )(((char*)((graph->exec_info)->data)) + (size_t)(graph-> exec_info)->rsize * (size_t)(0))))[_idx_].outgoings->rnum ; _j_++) { const int d = *(int*)((void*)(((char*)((((ccv_nnc_graph_exec_info_t *)((void*)(((char*)((graph->exec_info)->data)) + (size_t )(graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings )->data)) + (size_t)(((ccv_nnc_graph_exec_info_t*)((void*) (((char*)((graph->exec_info)->data)) + (size_t)(graph-> exec_info)->rsize * (size_t)(0))))[_idx_].outgoings)->rsize * (size_t)(_j_))); ++_incomings_[d].c; if (_incomings_[d].r != 0) continue; _incomings_[d].r = 1; ((void) sizeof ((_exist_size_ [_q_] < (graph->exec_info->rnum)) ? 1 : 0), __extension__ ({ if (_exist_size_[_q_] < (graph->exec_info->rnum) ) ; else __assert_fail ("_exist_size_[_q_] < (graph->exec_info->rnum)" , "ccv_nnc_graph_run.c", 1024, __extension__ __PRETTY_FUNCTION__ ); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[ _q_]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for (_i_ = 0; _i_ < (graph->breakpoint_size); _i_++) { ((void ) sizeof (((graph->breakpoints)[_i_].graph == graph) ? 1 : 0), __extension__ ({ if ((graph->breakpoints)[_i_].graph == graph) ; else __assert_fail ("(graph->breakpoints)[_i_].graph == graph" , "ccv_nnc_graph_run.c", 1024, __extension__ __PRETTY_FUNCTION__ ); })); _incomings_[(graph->breakpoints)[_i_].d].r = 3; _exists_ [0][_i_] = (graph->breakpoints)[_i_].d; } _exist_size_[0] = (graph->breakpoint_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 (((ccv_nnc_graph_exec_info_t *)((void*)(((char*)((graph->exec_info)->data)) + (size_t )(graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings ) for (_j_ = 0; _j_ < ((ccv_nnc_graph_exec_info_t*)((void* )(((char*)((graph->exec_info)->data)) + (size_t)(graph-> exec_info)->rsize * (size_t)(0))))[_idx_].outgoings->rnum ; _j_++) { const int d = *(int*)((void*)(((char*)((((ccv_nnc_graph_exec_info_t *)((void*)(((char*)((graph->exec_info)->data)) + (size_t )(graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings )->data)) + (size_t)(((ccv_nnc_graph_exec_info_t*)((void*) (((char*)((graph->exec_info)->data)) + (size_t)(graph-> exec_info)->rsize * (size_t)(0))))[_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_] < (graph->exec_info->rnum)) ? 1 : 0), __extension__ ({ if (_exist_size_[_q_] < (graph->exec_info->rnum)) ; else __assert_fail ("_exist_size_[_q_] < (graph->exec_info->rnum)" , "ccv_nnc_graph_run.c", 1024, __extension__ __PRETTY_FUNCTION__ ); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[ _q_]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for (_i_ = 0; _i_ < (graph_destination_size); _i_++) { ((void ) sizeof (((graph_destinations)[_i_].graph == graph) ? 1 : 0) , __extension__ ({ if ((graph_destinations)[_i_].graph == graph ) ; else __assert_fail ("(graph_destinations)[_i_].graph == graph" , "ccv_nnc_graph_run.c", 1024, __extension__ __PRETTY_FUNCTION__ ); })); _incomings_[(graph_destinations)[_i_].d].r = 5; _exists_ [0][_i_] = (graph_destinations)[_i_].d; } _exist_size_[0] = ( graph_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_] < (graph->exec_info->rnum)) ? 1 : 0 ), __extension__ ({ if (_exist_size_[_q_] < (graph->exec_info ->rnum)) ; else __assert_fail ("_exist_size_[_q_] < (graph->exec_info->rnum)" , "ccv_nnc_graph_run.c", 1024, __extension__ __PRETTY_FUNCTION__ ); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[ _q_]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for (_i_ = 0; _i_ < (graph_destination_size); _i_++) { ((void ) sizeof (((graph_destinations)[_i_].graph == graph) ? 1 : 0) , __extension__ ({ if ((graph_destinations)[_i_].graph == graph ) ; else __assert_fail ("(graph_destinations)[_i_].graph == graph" , "ccv_nnc_graph_run.c", 1024, __extension__ __PRETTY_FUNCTION__ ); })); _incomings_[(graph_destinations)[_i_].d].d = 1; } for (_i_ = 0; _i_ < (graph->breakpoint_size); _i_++) { ((void ) sizeof (((graph->breakpoints)[_i_].graph == graph) ? 1 : 0), __extension__ ({ if ((graph->breakpoints)[_i_].graph == graph) ; else __assert_fail ("(graph->breakpoints)[_i_].graph == graph" , "ccv_nnc_graph_run.c", 1024, __extension__ __PRETTY_FUNCTION__ ); })); _exists_[0][_i_] = (graph->breakpoints)[_i_].d; } _p_ = 0; _q_ = 1; _exist_size_[0] = (graph->breakpoint_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_]; visitor(( ((ccv_nnc_graph_exec_info_t*)((void*)(((char*)((graph->exec_info )->data)) + (size_t)(graph->exec_info)->rsize * (size_t )(0)))) + _idx_), (_idx_), (_incomings_[_idx_].d)); if (_incomings_ [_idx_].d) { ++_d_; _incomings_[_idx_].r = 7; } if (((ccv_nnc_graph_exec_info_t *)((void*)(((char*)((graph->exec_info)->data)) + (size_t )(graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings ) { if (((ccv_nnc_graph_exec_info_t*)((void*)(((char*)((graph ->exec_info)->data)) + (size_t)(graph->exec_info)-> rsize * (size_t)(0))))[_idx_].outgoings->rnum == 1) { const int d = *(int*)((void*)(((char*)((((ccv_nnc_graph_exec_info_t *)((void*)(((char*)((graph->exec_info)->data)) + (size_t )(graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings )->data)) + (size_t)(((ccv_nnc_graph_exec_info_t*)((void*) (((char*)((graph->exec_info)->data)) + (size_t)(graph-> exec_info)->rsize * (size_t)(0))))[_idx_].outgoings)->rsize * (size_t)(0))); --_incomings_[d].c; if (_incomings_[d].c == 0 && _incomings_[d].r == 6 && _d_ < (graph_destination_size )) { _exists_[_p_][_i_] = d; continue; } } else for (_j_ = 0; _j_ < ((ccv_nnc_graph_exec_info_t*)((void*)(((char*)((graph ->exec_info)->data)) + (size_t)(graph->exec_info)-> rsize * (size_t)(0))))[_idx_].outgoings->rnum; _j_++) { const int d = *(int*)((void*)(((char*)((((ccv_nnc_graph_exec_info_t *)((void*)(((char*)((graph->exec_info)->data)) + (size_t )(graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings )->data)) + (size_t)(((ccv_nnc_graph_exec_info_t*)((void*) (((char*)((graph->exec_info)->data)) + (size_t)(graph-> exec_info)->rsize * (size_t)(0))))[_idx_].outgoings)->rsize * (size_t)(_j_))); --_incomings_[d].c; if (_incomings_[d].c == 0 && _incomings_[d].r == 6 && _d_ < (graph_destination_size )) { ((void) sizeof ((_exist_size_[_q_] < (graph->exec_info ->rnum)) ? 1 : 0), __extension__ ({ if (_exist_size_[_q_] < (graph->exec_info->rnum)) ; else __assert_fail ("_exist_size_[_q_] < (graph->exec_info->rnum)" , "ccv_nnc_graph_run.c", 1024, __extension__ __PRETTY_FUNCTION__ ); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[ _q_]; } } } ++_i_; } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = ( _i_)); } for (_i_ = 0; _i_ < (graph_destination_size); _i_ ++) { ((void) sizeof (((graph_destinations)[_i_].graph == graph ) ? 1 : 0), __extension__ ({ if ((graph_destinations)[_i_].graph == graph) ; else __assert_fail ("(graph_destinations)[_i_].graph == graph" , "ccv_nnc_graph_run.c", 1024, __extension__ __PRETTY_FUNCTION__ ); })); if (_incomings_[(graph_destinations)[_i_].d].r == 7) continue ; if (!(1)) { ((void) sizeof ((_incomings_[(graph_destinations )[_i_].d].c == 0) ? 1 : 0), __extension__ ({ if (_incomings_[ (graph_destinations)[_i_].d].c == 0) ; else __assert_fail ("_incomings_[(graph_destinations)[_i_].d].c == 0" , "ccv_nnc_graph_run.c", 1024, __extension__ __PRETTY_FUNCTION__ ); })); } else if (_incomings_[(graph_destinations)[_i_].d].c > 0) continue; visitor((((ccv_nnc_graph_exec_info_t*)((void *)(((char*)((graph->exec_info)->data)) + (size_t)(graph ->exec_info)->rsize * (size_t)(0)))) + (graph_destinations )[_i_].d), ((graph_destinations)[_i_].d), (_incomings_[(graph_destinations )[_i_].d].d)); } if (_heap_mem_) free(_incomings_); } while ( 0);; | ||||
1025 | _ccv_nnc_graph_from_move_transit(graph); | ||||
1026 | _ccv_nnc_graph_rewrap(graph); | ||||
1027 | for (count = 1; reverse_count > 0; ++count) | ||||
1028 | { | ||||
1029 | graph->while_count = --reverse_count; | ||||
1030 | _ccv_nnc_graph_unwrap(graph, count, reverse_count); | ||||
1031 | _ccv_nnc_graph_transit_move_to(graph); | ||||
1032 | CCV_NNC_GRAPH_VISIT(graph, (ccv_nnc_graph_exec_info_t*)ccv_array_get(graph->exec_info, 0), graph->exec_info->rnum, graph_sources, graph_source_size, graph_destinations, graph_destination_size, 0, visitor)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_ < (graph->exec_info->rnum); _i_ ++) _incoming_edges_ += (((ccv_nnc_graph_exec_info_t*)((void* )(((char*)((graph->exec_info)->data)) + (size_t)(graph-> exec_info)->rsize * (size_t)(0))))[_i_].outgoings) ? ((ccv_nnc_graph_exec_info_t *)((void*)(((char*)((graph->exec_info)->data)) + (size_t )(graph->exec_info)->rsize * (size_t)(0))))[_i_].outgoings ->rnum : 0; const int _heap_mem_ = ((graph->exec_info-> rnum) + _incoming_edges_ > 1024); ccv_nnc_incoming_t* _incomings_ ; if (_heap_mem_) _incomings_ = (ccv_nnc_incoming_t*)malloc(sizeof (ccv_nnc_incoming_t) * (graph->exec_info->rnum) + sizeof (int32_t) * ((graph->exec_info->rnum) * 2 + _incoming_edges_ )); else _incomings_ = (ccv_nnc_incoming_t*)__builtin_alloca ( sizeof(ccv_nnc_incoming_t) * (graph->exec_info->rnum) + sizeof(int32_t) * ((graph->exec_info->rnum) * 2 + _incoming_edges_ )); memset(_incomings_, 0, sizeof(ccv_nnc_incoming_t) * (graph ->exec_info->rnum)); int32_t* _exists_[2] = { (int32_t* )(_incomings_ + (graph->exec_info->rnum)), (int32_t*)(_incomings_ + (graph->exec_info->rnum)) + (graph->exec_info-> rnum), }; int32_t* const _edges_ = _exists_[1] + (graph->exec_info ->rnum); for (_i_ = 0; _i_ < (graph_source_size); _i_++ ) { ((void) sizeof (((graph_sources)[_i_].graph == graph) ? 1 : 0), __extension__ ({ if ((graph_sources)[_i_].graph == graph ) ; else __assert_fail ("(graph_sources)[_i_].graph == graph" , "ccv_nnc_graph_run.c", 1032, __extension__ __PRETTY_FUNCTION__ ); })); _incomings_[(graph_sources)[_i_].d].r = 1; _exists_[0 ][_i_] = (graph_sources)[_i_].d; } int _exist_size_[2] = { (graph_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 (((ccv_nnc_graph_exec_info_t*)((void*)(((char*)((graph-> exec_info)->data)) + (size_t)(graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings) for (_j_ = 0; _j_ < (( ccv_nnc_graph_exec_info_t*)((void*)(((char*)((graph->exec_info )->data)) + (size_t)(graph->exec_info)->rsize * (size_t )(0))))[_idx_].outgoings->rnum; _j_++) { const int d = *(int *)((void*)(((char*)((((ccv_nnc_graph_exec_info_t*)((void*)((( char*)((graph->exec_info)->data)) + (size_t)(graph-> exec_info)->rsize * (size_t)(0))))[_idx_].outgoings)->data )) + (size_t)(((ccv_nnc_graph_exec_info_t*)((void*)(((char*)( (graph->exec_info)->data)) + (size_t)(graph->exec_info )->rsize * (size_t)(0))))[_idx_].outgoings)->rsize * (size_t )(_j_))); ++_incomings_[d].c; if (_incomings_[d].r != 0) continue ; _incomings_[d].r = 1; ((void) sizeof ((_exist_size_[_q_] < (graph->exec_info->rnum)) ? 1 : 0), __extension__ ({ if (_exist_size_[_q_] < (graph->exec_info->rnum)) ; else __assert_fail ("_exist_size_[_q_] < (graph->exec_info->rnum)" , "ccv_nnc_graph_run.c", 1032, __extension__ __PRETTY_FUNCTION__ ); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[ _q_]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for (_i_ = 0; _i_ < (graph_source_size); _i_++) { ((void) sizeof (((graph_sources)[_i_].graph == graph) ? 1 : 0), __extension__ ({ if ((graph_sources)[_i_].graph == graph) ; else __assert_fail ("(graph_sources)[_i_].graph == graph", "ccv_nnc_graph_run.c" , 1032, __extension__ __PRETTY_FUNCTION__); })); _incomings_[ (graph_sources)[_i_].d].r = 3; _exists_[0][_i_] = (graph_sources )[_i_].d; } _exist_size_[0] = (graph_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 (((ccv_nnc_graph_exec_info_t*)((void*)(((char *)((graph->exec_info)->data)) + (size_t)(graph->exec_info )->rsize * (size_t)(0))))[_idx_].outgoings) for (_j_ = 0; _j_ < ((ccv_nnc_graph_exec_info_t*)((void*)(((char*)((graph-> exec_info)->data)) + (size_t)(graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings->rnum; _j_++) { const int d = *(int*)((void*)(((char*)((((ccv_nnc_graph_exec_info_t*)( (void*)(((char*)((graph->exec_info)->data)) + (size_t)( graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings )->data)) + (size_t)(((ccv_nnc_graph_exec_info_t*)((void*) (((char*)((graph->exec_info)->data)) + (size_t)(graph-> exec_info)->rsize * (size_t)(0))))[_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_] < (graph->exec_info->rnum)) ? 1 : 0), __extension__ ({ if (_exist_size_[_q_] < (graph->exec_info->rnum)) ; else __assert_fail ("_exist_size_[_q_] < (graph->exec_info->rnum)" , "ccv_nnc_graph_run.c", 1032, __extension__ __PRETTY_FUNCTION__ ); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[ _q_]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for (_i_ = 0; _i_ < (graph_destination_size); _i_++) { ((void ) sizeof (((graph_destinations)[_i_].graph == graph) ? 1 : 0) , __extension__ ({ if ((graph_destinations)[_i_].graph == graph ) ; else __assert_fail ("(graph_destinations)[_i_].graph == graph" , "ccv_nnc_graph_run.c", 1032, __extension__ __PRETTY_FUNCTION__ ); })); _incomings_[(graph_destinations)[_i_].d].r = 5; _exists_ [0][_i_] = (graph_destinations)[_i_].d; } _exist_size_[0] = ( graph_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_] < (graph->exec_info->rnum)) ? 1 : 0 ), __extension__ ({ if (_exist_size_[_q_] < (graph->exec_info ->rnum)) ; else __assert_fail ("_exist_size_[_q_] < (graph->exec_info->rnum)" , "ccv_nnc_graph_run.c", 1032, __extension__ __PRETTY_FUNCTION__ ); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[ _q_]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for (_i_ = 0; _i_ < (graph_destination_size); _i_++) { ((void ) sizeof (((graph_destinations)[_i_].graph == graph) ? 1 : 0) , __extension__ ({ if ((graph_destinations)[_i_].graph == graph ) ; else __assert_fail ("(graph_destinations)[_i_].graph == graph" , "ccv_nnc_graph_run.c", 1032, __extension__ __PRETTY_FUNCTION__ ); })); _incomings_[(graph_destinations)[_i_].d].d = 1; } for (_i_ = 0; _i_ < (graph_source_size); _i_++) { ((void) sizeof (((graph_sources)[_i_].graph == graph) ? 1 : 0), __extension__ ({ if ((graph_sources)[_i_].graph == graph) ; else __assert_fail ("(graph_sources)[_i_].graph == graph", "ccv_nnc_graph_run.c" , 1032, __extension__ __PRETTY_FUNCTION__); })); _exists_[0][ _i_] = (graph_sources)[_i_].d; } _p_ = 0; _q_ = 1; _exist_size_ [0] = (graph_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_]; visitor((((ccv_nnc_graph_exec_info_t*)((void*)((( char*)((graph->exec_info)->data)) + (size_t)(graph-> exec_info)->rsize * (size_t)(0)))) + _idx_), (_idx_), (_incomings_ [_idx_].d)); if (_incomings_[_idx_].d) { ++_d_; _incomings_[_idx_ ].r = 7; } if (((ccv_nnc_graph_exec_info_t*)((void*)(((char*) ((graph->exec_info)->data)) + (size_t)(graph->exec_info )->rsize * (size_t)(0))))[_idx_].outgoings) { if (((ccv_nnc_graph_exec_info_t *)((void*)(((char*)((graph->exec_info)->data)) + (size_t )(graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings ->rnum == 1) { const int d = *(int*)((void*)(((char*)((((ccv_nnc_graph_exec_info_t *)((void*)(((char*)((graph->exec_info)->data)) + (size_t )(graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings )->data)) + (size_t)(((ccv_nnc_graph_exec_info_t*)((void*) (((char*)((graph->exec_info)->data)) + (size_t)(graph-> exec_info)->rsize * (size_t)(0))))[_idx_].outgoings)->rsize * (size_t)(0))); --_incomings_[d].c; if (_incomings_[d].c == 0 && _incomings_[d].r == 6 && _d_ < (graph_destination_size )) { _exists_[_p_][_i_] = d; continue; } } else for (_j_ = 0; _j_ < ((ccv_nnc_graph_exec_info_t*)((void*)(((char*)((graph ->exec_info)->data)) + (size_t)(graph->exec_info)-> rsize * (size_t)(0))))[_idx_].outgoings->rnum; _j_++) { const int d = *(int*)((void*)(((char*)((((ccv_nnc_graph_exec_info_t *)((void*)(((char*)((graph->exec_info)->data)) + (size_t )(graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings )->data)) + (size_t)(((ccv_nnc_graph_exec_info_t*)((void*) (((char*)((graph->exec_info)->data)) + (size_t)(graph-> exec_info)->rsize * (size_t)(0))))[_idx_].outgoings)->rsize * (size_t)(_j_))); --_incomings_[d].c; if (_incomings_[d].c == 0 && _incomings_[d].r == 6 && _d_ < (graph_destination_size )) { ((void) sizeof ((_exist_size_[_q_] < (graph->exec_info ->rnum)) ? 1 : 0), __extension__ ({ if (_exist_size_[_q_] < (graph->exec_info->rnum)) ; else __assert_fail ("_exist_size_[_q_] < (graph->exec_info->rnum)" , "ccv_nnc_graph_run.c", 1032, __extension__ __PRETTY_FUNCTION__ ); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[ _q_]; } } } ++_i_; } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = ( _i_)); } for (_i_ = 0; _i_ < (graph_destination_size); _i_ ++) { ((void) sizeof (((graph_destinations)[_i_].graph == graph ) ? 1 : 0), __extension__ ({ if ((graph_destinations)[_i_].graph == graph) ; else __assert_fail ("(graph_destinations)[_i_].graph == graph" , "ccv_nnc_graph_run.c", 1032, __extension__ __PRETTY_FUNCTION__ ); })); if (_incomings_[(graph_destinations)[_i_].d].r == 7) continue ; if (!(0)) { ((void) sizeof ((_incomings_[(graph_destinations )[_i_].d].c == 0) ? 1 : 0), __extension__ ({ if (_incomings_[ (graph_destinations)[_i_].d].c == 0) ; else __assert_fail ("_incomings_[(graph_destinations)[_i_].d].c == 0" , "ccv_nnc_graph_run.c", 1032, __extension__ __PRETTY_FUNCTION__ ); })); } else if (_incomings_[(graph_destinations)[_i_].d].c > 0) continue; visitor((((ccv_nnc_graph_exec_info_t*)((void *)(((char*)((graph->exec_info)->data)) + (size_t)(graph ->exec_info)->rsize * (size_t)(0)))) + (graph_destinations )[_i_].d), ((graph_destinations)[_i_].d), (_incomings_[(graph_destinations )[_i_].d].d)); } if (_heap_mem_) free(_incomings_); } while ( 0);; | ||||
1033 | _ccv_nnc_graph_from_move_transit(graph); | ||||
1034 | _ccv_nnc_graph_rewrap(graph); | ||||
1035 | } | ||||
1036 | } | ||||
1037 | } else { | ||||
1038 | graph->while_count = 0; | ||||
1039 | CCV_NNC_GRAPH_VISIT(graph, (ccv_nnc_graph_exec_info_t*)ccv_array_get(graph->exec_info, 0), graph->exec_info->rnum, graph_sources, graph_source_size, graph_destinations, graph_destination_size, 0, visitor)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_ < (graph->exec_info->rnum); _i_ ++) _incoming_edges_ += (((ccv_nnc_graph_exec_info_t*)((void* )(((char*)((graph->exec_info)->data)) + (size_t)(graph-> exec_info)->rsize * (size_t)(0))))[_i_].outgoings) ? ((ccv_nnc_graph_exec_info_t *)((void*)(((char*)((graph->exec_info)->data)) + (size_t )(graph->exec_info)->rsize * (size_t)(0))))[_i_].outgoings ->rnum : 0; const int _heap_mem_ = ((graph->exec_info-> rnum) + _incoming_edges_ > 1024); ccv_nnc_incoming_t* _incomings_ ; if (_heap_mem_) _incomings_ = (ccv_nnc_incoming_t*)malloc(sizeof (ccv_nnc_incoming_t) * (graph->exec_info->rnum) + sizeof (int32_t) * ((graph->exec_info->rnum) * 2 + _incoming_edges_ )); else _incomings_ = (ccv_nnc_incoming_t*)__builtin_alloca ( sizeof(ccv_nnc_incoming_t) * (graph->exec_info->rnum) + sizeof(int32_t) * ((graph->exec_info->rnum) * 2 + _incoming_edges_ )); memset(_incomings_, 0, sizeof(ccv_nnc_incoming_t) * (graph ->exec_info->rnum)); int32_t* _exists_[2] = { (int32_t* )(_incomings_ + (graph->exec_info->rnum)), (int32_t*)(_incomings_ + (graph->exec_info->rnum)) + (graph->exec_info-> rnum), }; int32_t* const _edges_ = _exists_[1] + (graph->exec_info ->rnum); for (_i_ = 0; _i_ < (graph_source_size); _i_++ ) { ((void) sizeof (((graph_sources)[_i_].graph == graph) ? 1 : 0), __extension__ ({ if ((graph_sources)[_i_].graph == graph ) ; else __assert_fail ("(graph_sources)[_i_].graph == graph" , "ccv_nnc_graph_run.c", 1039, __extension__ __PRETTY_FUNCTION__ ); })); _incomings_[(graph_sources)[_i_].d].r = 1; _exists_[0 ][_i_] = (graph_sources)[_i_].d; } int _exist_size_[2] = { (graph_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 (((ccv_nnc_graph_exec_info_t*)((void*)(((char*)((graph-> exec_info)->data)) + (size_t)(graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings) for (_j_ = 0; _j_ < (( ccv_nnc_graph_exec_info_t*)((void*)(((char*)((graph->exec_info )->data)) + (size_t)(graph->exec_info)->rsize * (size_t )(0))))[_idx_].outgoings->rnum; _j_++) { const int d = *(int *)((void*)(((char*)((((ccv_nnc_graph_exec_info_t*)((void*)((( char*)((graph->exec_info)->data)) + (size_t)(graph-> exec_info)->rsize * (size_t)(0))))[_idx_].outgoings)->data )) + (size_t)(((ccv_nnc_graph_exec_info_t*)((void*)(((char*)( (graph->exec_info)->data)) + (size_t)(graph->exec_info )->rsize * (size_t)(0))))[_idx_].outgoings)->rsize * (size_t )(_j_))); ++_incomings_[d].c; if (_incomings_[d].r != 0) continue ; _incomings_[d].r = 1; ((void) sizeof ((_exist_size_[_q_] < (graph->exec_info->rnum)) ? 1 : 0), __extension__ ({ if (_exist_size_[_q_] < (graph->exec_info->rnum)) ; else __assert_fail ("_exist_size_[_q_] < (graph->exec_info->rnum)" , "ccv_nnc_graph_run.c", 1039, __extension__ __PRETTY_FUNCTION__ ); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[ _q_]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for (_i_ = 0; _i_ < (graph_source_size); _i_++) { ((void) sizeof (((graph_sources)[_i_].graph == graph) ? 1 : 0), __extension__ ({ if ((graph_sources)[_i_].graph == graph) ; else __assert_fail ("(graph_sources)[_i_].graph == graph", "ccv_nnc_graph_run.c" , 1039, __extension__ __PRETTY_FUNCTION__); })); _incomings_[ (graph_sources)[_i_].d].r = 3; _exists_[0][_i_] = (graph_sources )[_i_].d; } _exist_size_[0] = (graph_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 (((ccv_nnc_graph_exec_info_t*)((void*)(((char *)((graph->exec_info)->data)) + (size_t)(graph->exec_info )->rsize * (size_t)(0))))[_idx_].outgoings) for (_j_ = 0; _j_ < ((ccv_nnc_graph_exec_info_t*)((void*)(((char*)((graph-> exec_info)->data)) + (size_t)(graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings->rnum; _j_++) { const int d = *(int*)((void*)(((char*)((((ccv_nnc_graph_exec_info_t*)( (void*)(((char*)((graph->exec_info)->data)) + (size_t)( graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings )->data)) + (size_t)(((ccv_nnc_graph_exec_info_t*)((void*) (((char*)((graph->exec_info)->data)) + (size_t)(graph-> exec_info)->rsize * (size_t)(0))))[_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_] < (graph->exec_info->rnum)) ? 1 : 0), __extension__ ({ if (_exist_size_[_q_] < (graph->exec_info->rnum)) ; else __assert_fail ("_exist_size_[_q_] < (graph->exec_info->rnum)" , "ccv_nnc_graph_run.c", 1039, __extension__ __PRETTY_FUNCTION__ ); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[ _q_]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for (_i_ = 0; _i_ < (graph_destination_size); _i_++) { ((void ) sizeof (((graph_destinations)[_i_].graph == graph) ? 1 : 0) , __extension__ ({ if ((graph_destinations)[_i_].graph == graph ) ; else __assert_fail ("(graph_destinations)[_i_].graph == graph" , "ccv_nnc_graph_run.c", 1039, __extension__ __PRETTY_FUNCTION__ ); })); _incomings_[(graph_destinations)[_i_].d].r = 5; _exists_ [0][_i_] = (graph_destinations)[_i_].d; } _exist_size_[0] = ( graph_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_] < (graph->exec_info->rnum)) ? 1 : 0 ), __extension__ ({ if (_exist_size_[_q_] < (graph->exec_info ->rnum)) ; else __assert_fail ("_exist_size_[_q_] < (graph->exec_info->rnum)" , "ccv_nnc_graph_run.c", 1039, __extension__ __PRETTY_FUNCTION__ ); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[ _q_]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for (_i_ = 0; _i_ < (graph_destination_size); _i_++) { ((void ) sizeof (((graph_destinations)[_i_].graph == graph) ? 1 : 0) , __extension__ ({ if ((graph_destinations)[_i_].graph == graph ) ; else __assert_fail ("(graph_destinations)[_i_].graph == graph" , "ccv_nnc_graph_run.c", 1039, __extension__ __PRETTY_FUNCTION__ ); })); _incomings_[(graph_destinations)[_i_].d].d = 1; } for (_i_ = 0; _i_ < (graph_source_size); _i_++) { ((void) sizeof (((graph_sources)[_i_].graph == graph) ? 1 : 0), __extension__ ({ if ((graph_sources)[_i_].graph == graph) ; else __assert_fail ("(graph_sources)[_i_].graph == graph", "ccv_nnc_graph_run.c" , 1039, __extension__ __PRETTY_FUNCTION__); })); _exists_[0][ _i_] = (graph_sources)[_i_].d; } _p_ = 0; _q_ = 1; _exist_size_ [0] = (graph_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_]; visitor((((ccv_nnc_graph_exec_info_t*)((void*)((( char*)((graph->exec_info)->data)) + (size_t)(graph-> exec_info)->rsize * (size_t)(0)))) + _idx_), (_idx_), (_incomings_ [_idx_].d)); if (_incomings_[_idx_].d) { ++_d_; _incomings_[_idx_ ].r = 7; } if (((ccv_nnc_graph_exec_info_t*)((void*)(((char*) ((graph->exec_info)->data)) + (size_t)(graph->exec_info )->rsize * (size_t)(0))))[_idx_].outgoings) { if (((ccv_nnc_graph_exec_info_t *)((void*)(((char*)((graph->exec_info)->data)) + (size_t )(graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings ->rnum == 1) { const int d = *(int*)((void*)(((char*)((((ccv_nnc_graph_exec_info_t *)((void*)(((char*)((graph->exec_info)->data)) + (size_t )(graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings )->data)) + (size_t)(((ccv_nnc_graph_exec_info_t*)((void*) (((char*)((graph->exec_info)->data)) + (size_t)(graph-> exec_info)->rsize * (size_t)(0))))[_idx_].outgoings)->rsize * (size_t)(0))); --_incomings_[d].c; if (_incomings_[d].c == 0 && _incomings_[d].r == 6 && _d_ < (graph_destination_size )) { _exists_[_p_][_i_] = d; continue; } } else for (_j_ = 0; _j_ < ((ccv_nnc_graph_exec_info_t*)((void*)(((char*)((graph ->exec_info)->data)) + (size_t)(graph->exec_info)-> rsize * (size_t)(0))))[_idx_].outgoings->rnum; _j_++) { const int d = *(int*)((void*)(((char*)((((ccv_nnc_graph_exec_info_t *)((void*)(((char*)((graph->exec_info)->data)) + (size_t )(graph->exec_info)->rsize * (size_t)(0))))[_idx_].outgoings )->data)) + (size_t)(((ccv_nnc_graph_exec_info_t*)((void*) (((char*)((graph->exec_info)->data)) + (size_t)(graph-> exec_info)->rsize * (size_t)(0))))[_idx_].outgoings)->rsize * (size_t)(_j_))); --_incomings_[d].c; if (_incomings_[d].c == 0 && _incomings_[d].r == 6 && _d_ < (graph_destination_size )) { ((void) sizeof ((_exist_size_[_q_] < (graph->exec_info ->rnum)) ? 1 : 0), __extension__ ({ if (_exist_size_[_q_] < (graph->exec_info->rnum)) ; else __assert_fail ("_exist_size_[_q_] < (graph->exec_info->rnum)" , "ccv_nnc_graph_run.c", 1039, __extension__ __PRETTY_FUNCTION__ ); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[ _q_]; } } } ++_i_; } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = ( _i_)); } for (_i_ = 0; _i_ < (graph_destination_size); _i_ ++) { ((void) sizeof (((graph_destinations)[_i_].graph == graph ) ? 1 : 0), __extension__ ({ if ((graph_destinations)[_i_].graph == graph) ; else __assert_fail ("(graph_destinations)[_i_].graph == graph" , "ccv_nnc_graph_run.c", 1039, __extension__ __PRETTY_FUNCTION__ ); })); if (_incomings_[(graph_destinations)[_i_].d].r == 7) continue ; if (!(0)) { ((void) sizeof ((_incomings_[(graph_destinations )[_i_].d].c == 0) ? 1 : 0), __extension__ ({ if (_incomings_[ (graph_destinations)[_i_].d].c == 0) ; else __assert_fail ("_incomings_[(graph_destinations)[_i_].d].c == 0" , "ccv_nnc_graph_run.c", 1039, __extension__ __PRETTY_FUNCTION__ ); })); } else if (_incomings_[(graph_destinations)[_i_].d].c > 0) continue; visitor((((ccv_nnc_graph_exec_info_t*)((void *)(((char*)((graph->exec_info)->data)) + (size_t)(graph ->exec_info)->rsize * (size_t)(0)))) + (graph_destinations )[_i_].d), ((graph_destinations)[_i_].d), (_incomings_[(graph_destinations )[_i_].d].d)); } if (_heap_mem_) free(_incomings_); } while ( 0);; | ||||
1040 | } | ||||
1041 | #undef visitor | ||||
1042 | } | ||||
1043 | |||||
1044 | static int _ccv_nnc_graph_run(ccv_nnc_graph_t* const graph, const int exec_idx, ccv_nnc_graph_exec_info_t* const exec, ccv_nnc_tensor_t* const* const inputs, const int input_size, ccv_nnc_tensor_t* const* const outputs, const int output_size, const int flags, const ccv_nnc_graph_exec_t* const sources, const int source_size, const ccv_nnc_graph_exec_t* const destinations, const int destination_size, ccv_nnc_tensor_tape_t* const tensor_tape, ccv_nnc_stream_context_t* const stream_context) | ||||
1045 | { | ||||
1046 | assert((sources == 0 && source_size == 0) || (sources && source_size))((void) sizeof (((sources == 0 && source_size == 0) || (sources && source_size)) ? 1 : 0), __extension__ ({ if ((sources == 0 && source_size == 0) || (sources && source_size)) ; else __assert_fail ("(sources == 0 && source_size == 0) || (sources && source_size)" , "ccv_nnc_graph_run.c", 1046, __extension__ __PRETTY_FUNCTION__ ); })); | ||||
1047 | assert((destinations == 0 && destination_size == 0) || (destinations && destination_size))((void) sizeof (((destinations == 0 && destination_size == 0) || (destinations && destination_size)) ? 1 : 0 ), __extension__ ({ if ((destinations == 0 && destination_size == 0) || (destinations && destination_size)) ; else __assert_fail ("(destinations == 0 && destination_size == 0) || (destinations && destination_size)" , "ccv_nnc_graph_run.c", 1047, __extension__ __PRETTY_FUNCTION__ ); })); | ||||
1048 | const ccv_nnc_graph_exec_t* const graph_sources = sources
->sources)->rsize * (size_t)(0))); | ||||
1049 | const int graph_source_size = source_size
| ||||
1050 | const ccv_nnc_graph_exec_t* const graph_destinations = destinations
)(graph->destinations)->rsize * (size_t)(0))); | ||||
1051 | const int graph_destination_size = destination_size
| ||||
1052 | int i; | ||||
1053 | for (i = 0; i < graph_source_size; i++) | ||||
1054 | if (graph_sources[i].graph != graph) | ||||
1055 | return CCV_NNC_EXEC_INVALID; | ||||
1056 | for (i = 0; i < graph_destination_size; i++) | ||||
1057 | if (graph_destinations[i].graph != graph) | ||||
1058 | return CCV_NNC_EXEC_INVALID; | ||||
1059 | // When topsorted is true, there is no memory allocation when run the graph. | ||||
1060 | const int topsorted = (!sources
| ||||
1061 | if (topsorted) | ||||
1062 | _ccv_nnc_graph_topsorted_run(graph, exec_idx, exec, flags, tensor_tape, stream_context); | ||||
1063 | else | ||||
1064 | _ccv_nnc_graph_run_slow_path(graph, exec_idx, exec, inputs, input_size, outputs, output_size, flags, sources, source_size, destinations, destination_size, tensor_tape, stream_context); | ||||
1065 | return CCV_NNC_EXEC_SUCCESS; | ||||
1066 | } | ||||
1067 | |||||
1068 | int ccv_nnc_graph_run(ccv_nnc_graph_t* const graph, const int flags, const ccv_nnc_graph_exec_t* const sources, const int source_size, const ccv_nnc_graph_exec_t* const destinations, const int destination_size, ccv_nnc_tensor_tape_t* const tensor_tape, ccv_nnc_stream_context_t* const stream_context) | ||||
1069 | { | ||||
1070 | if (stream_context && graph->topsorted && graph->stream_size > 0 && graph->default_schedule && source_size == 0 && destination_size == 0) | ||||
| |||||
1071 | { | ||||
1072 | co_scheduler_t* const scheduler = ccv_nnc_stream_context_get_scheduler(stream_context); | ||||
1073 | co_routine_t* const task = co_new(_ccv_nnc_graph_topsorted_run_coro, (graph, -1, graph->default_schedule, 0, tensor_tape, stream_context, flags))({ co_routine_t* const task = malloc((sizeof(co_routine_t) + _ccv_nnc_graph_topsorted_run_coro_stack_size ())); do { struct _ccv_nnc_graph_topsorted_run_coro_param_s params = { ._co_params = { graph, -1, graph->default_schedule, 0 , tensor_tape, stream_context, flags } }; task->fn = _ccv_nnc_graph_topsorted_run_coro ; task->line = 0; task->done = 0; task->root = 0; task ->other_size = 0; task->notify_any = 0; task->others = 0; task->caller = 0; task->callee = 0; if (sizeof(params ) > 0) memcpy(task + 1, ¶ms, sizeof(params)); } while (0); task; }); | ||||
1074 | co_schedule(scheduler, task); | ||||
1075 | // I don't need to worry about freeing this task, it will free itself at the end. | ||||
1076 | return CCV_NNC_EXEC_SUCCESS; | ||||
1077 | } else | ||||
1078 | return _ccv_nnc_graph_run(graph, -1, 0, 0, 0, 0, 0, flags, sources, source_size, destinations, destination_size, tensor_tape, 0 /* In this case, we don't support stream context yet. */); | ||||
1079 | } | ||||
1080 | |||||
1081 | int ccv_nnc_graph_run_with_schedule(ccv_nnc_graph_t* const graph, const int flags, const ccv_nnc_graph_static_schedule_t* const _schedule, ccv_nnc_tensor_tape_t* const tensor_tape, ccv_nnc_stream_context_t* const _stream_context) | ||||
1082 | { | ||||
1083 | assert(graph->topsorted)((void) sizeof ((graph->topsorted) ? 1 : 0), __extension__ ({ if (graph->topsorted) ; else __assert_fail ("graph->topsorted" , "ccv_nnc_graph_run.c", 1083, __extension__ __PRETTY_FUNCTION__ ); })); | ||||
1084 | if (graph->exec_info->rnum == 0) | ||||
1085 | return CCV_NNC_EXEC_SUCCESS; | ||||
1086 | assert(graph->stream_size > 0)((void) sizeof ((graph->stream_size > 0) ? 1 : 0), __extension__ ({ if (graph->stream_size > 0) ; else __assert_fail ("graph->stream_size > 0" , "ccv_nnc_graph_run.c", 1086, __extension__ __PRETTY_FUNCTION__ ); })); | ||||
1087 | const ccv_nnc_graph_static_schedule_t* const schedule = _schedule ? _schedule : graph->default_schedule; | ||||
1088 | assert(schedule)((void) sizeof ((schedule) ? 1 : 0), __extension__ ({ if (schedule ) ; else __assert_fail ("schedule", "ccv_nnc_graph_run.c", 1088 , __extension__ __PRETTY_FUNCTION__); })); | ||||
1089 | assert(schedule->stream_0 < graph->stream_size)((void) sizeof ((schedule->stream_0 < graph->stream_size ) ? 1 : 0), __extension__ ({ if (schedule->stream_0 < graph ->stream_size) ; else __assert_fail ("schedule->stream_0 < graph->stream_size" , "ccv_nnc_graph_run.c", 1089, __extension__ __PRETTY_FUNCTION__ ); })); | ||||
1090 | ccv_nnc_stream_context_t* const stream_context = _stream_context ? _stream_context : graph->streams[schedule->stream_0]; | ||||
1091 | co_scheduler_t* const scheduler = ccv_nnc_stream_context_get_scheduler(stream_context); | ||||
1092 | co_routine_t* const task = co_new(_ccv_nnc_graph_topsorted_run_coro, (graph, -1, schedule, 0, tensor_tape, stream_context, flags))({ co_routine_t* const task = malloc((sizeof(co_routine_t) + _ccv_nnc_graph_topsorted_run_coro_stack_size ())); do { struct _ccv_nnc_graph_topsorted_run_coro_param_s params = { ._co_params = { graph, -1, schedule, 0, tensor_tape, stream_context , flags } }; task->fn = _ccv_nnc_graph_topsorted_run_coro; task->line = 0; task->done = 0; task->root = 0; task ->other_size = 0; task->notify_any = 0; task->others = 0; task->caller = 0; task->callee = 0; if (sizeof(params ) > 0) memcpy(task + 1, ¶ms, sizeof(params)); } while (0); task; }); | ||||
1093 | co_schedule(scheduler, task); | ||||
1094 | // I don't need to worry about freeing this task, it will free itself at the end. | ||||
1095 | if (!_stream_context) // If no stream context provided, this is a sync operation. | ||||
1096 | ccv_nnc_stream_context_wait(stream_context); | ||||
1097 | return CCV_NNC_EXEC_SUCCESS; | ||||
1098 | } |