File: | nnc/ccv_nnc_symbolic_graph_backward.c |
Warning: | line 731, column 4 Assigned value is garbage or undefined |
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_symbolic_graph.h" | |||
6 | ||||
7 | // MARK - Level-3.5 API | |||
8 | ||||
9 | typedef struct { | |||
10 | int f_wrt; // Check if both f_symbols and wrt_symbols flow through this node. | |||
11 | ccv_array_t* outgoings; // backward traverse nodes. | |||
12 | uint64_t* input_bitmasks; | |||
13 | int input_bitmask_size; | |||
14 | uint64_t* output_bitmasks; | |||
15 | int output_bitmask_size; | |||
16 | } ccv_nnc_graph_backward_info_t; | |||
17 | ||||
18 | typedef struct { | |||
19 | int input_size; | |||
20 | int* inputs; | |||
21 | int output; | |||
22 | ccv_array_t* outgoings; | |||
23 | float value; | |||
24 | ccv_nnc_graph_exec_symbol_t symbol; | |||
25 | } ccv_nnc_sum_or_set_graph_exec_symbol_t; | |||
26 | ||||
27 | typedef struct { | |||
28 | int input_size; | |||
29 | int output_size; | |||
30 | int* inputs; | |||
31 | int* outputs; | |||
32 | ccv_array_t* outgoings; | |||
33 | ccv_nnc_cmd_t cmd; | |||
34 | ccv_nnc_graph_exec_symbol_t symbol; | |||
35 | } ccv_nnc_autograd_graph_exec_symbol_t; | |||
36 | ||||
37 | typedef struct { | |||
38 | int d; // The pointer to the forward level object. | |||
39 | int alias_ref; // The alias ref to itself (autograd_tensor_symbols array). | |||
40 | int flags; // Flags for this symbol. | |||
41 | ccv_nnc_tensor_symbol_t symbol; | |||
42 | } ccv_nnc_autograd_tensor_symbol_t; | |||
43 | ||||
44 | typedef struct { | |||
45 | int d; // The tensor symbol ref. | |||
46 | int x; // The exec symbol ref. | |||
47 | ccv_array_t* exec_registry; // Additional exec symbol refs, similar to x, only useful for aliasing. | |||
48 | ccv_array_t* alias_registry; // int point to all the alias (if this is not an alias). The alias is the object in autograd_tensor_symbols, you need another level of indirection to get the actual forward level alias. | |||
49 | } ccv_nnc_tensor_ref_t; | |||
50 | ||||
51 | typedef struct { | |||
52 | int c; // The start non-accumulated version. | |||
53 | ccv_array_t* ref_version; // tensor ref point to the reverse tensor symbol. | |||
54 | } ccv_nnc_autograd_tensor_version_t; | |||
55 | ||||
56 | typedef struct { | |||
57 | int d; | |||
58 | int alias_ref; | |||
59 | } ccv_nnc_sum_variable_t; | |||
60 | ||||
61 | // This method tries to figure out if a set of aliases can cover the whole tensor dim. | |||
62 | // This is not a precise implementation though. The requirement is to answer this question | |||
63 | // with a given memory constraint, therefore, only allow up to 65536 different tensor locations. | |||
64 | // If you have more than that, it will assume that it doesn't have fully assigned aliases, | |||
65 | // and will return 0. | |||
66 | ||||
67 | // Return 1 if inserted successfully. | |||
68 | static inline int _ccv_nnc_try_mix(int* const md, const int ins, const int c) | |||
69 | { | |||
70 | if (!c) | |||
71 | { | |||
72 | md[0] = ins; | |||
73 | return 1; | |||
74 | } | |||
75 | int ll = 0, uu = c - 1; | |||
76 | int mm; | |||
77 | do { | |||
78 | mm = ll + ((uu - ll) >> 1); | |||
79 | if (ins == md[mm]) | |||
80 | return 0; | |||
81 | else if (ins < md[mm]) | |||
82 | uu = mm - 1; | |||
83 | else if (ins > md[mm]) | |||
84 | ll = mm + 1; | |||
85 | } while (ll <= uu); | |||
86 | if (ll < c) | |||
87 | memmove(md + ll + 1, md + ll, sizeof(int) * (c - ll)); | |||
88 | md[ll] = ins; | |||
89 | return 1; | |||
90 | } | |||
91 | ||||
92 | static inline int _ccv_nnc_mix_idx(const int* const md, const int ins, const int c) | |||
93 | { | |||
94 | if (c <= 1) | |||
95 | return 0; | |||
96 | int ll = 0, uu = c - 1; | |||
97 | int mm; | |||
98 | do { | |||
99 | mm = ll + ((uu - ll) >> 1); | |||
100 | if (ins == md[mm]) | |||
101 | return mm; | |||
102 | else if (ins < md[mm]) | |||
103 | uu = mm - 1; | |||
104 | else if (ins > md[mm]) | |||
105 | ll = mm + 1; | |||
106 | } while (ll <= uu); | |||
107 | assert(0 && "Shouldn't reach here")((void) sizeof ((0 && "Shouldn't reach here") ? 1 : 0 ), __extension__ ({ if (0 && "Shouldn't reach here") ; else __assert_fail ("0 && \"Shouldn't reach here\"", "ccv_nnc_symbolic_graph_backward.c", 107, __extension__ __PRETTY_FUNCTION__ ); })); | |||
108 | return -1; | |||
109 | } | |||
110 | ||||
111 | static inline void _ccv_nnc_try_set_pix_0(const int* const ofs, const int* const dim, const int* const tensor_dim, int* const* const scmd, const int* const cube_dim, const int* const cube_step, uint32_t* const cube, int offset) | |||
112 | { | |||
113 | const int s = (ofs[0] == 0) ? 0 : _ccv_nnc_mix_idx(scmd[0], ofs[0], cube_dim[0]) + 1; | |||
114 | const int d = ((ofs[0] + dim[0] == tensor_dim[0]) ? cube_dim[0] : _ccv_nnc_mix_idx(scmd[0], ofs[0] + ccv_max(1, dim[0])({ typeof (1) _a = (1); typeof (dim[0]) _b = (dim[0]); (_a > _b) ? _a : _b; }), cube_dim[0])) + 1; | |||
115 | assert(s >= 0 && d > s)((void) sizeof ((s >= 0 && d > s) ? 1 : 0), __extension__ ({ if (s >= 0 && d > s) ; else __assert_fail ( "s >= 0 && d > s", "ccv_nnc_symbolic_graph_backward.c" , 115, __extension__ __PRETTY_FUNCTION__); })); | |||
116 | int i; | |||
117 | for (i = s; i < d; i++) | |||
118 | // Fill this pix. I can make this faster by loop through full ones (divided by 8), but too lazy. | |||
119 | cube[(offset + i) >> 5] |= (1u << ((offset + i) & 0x1f)); | |||
120 | } | |||
121 | ||||
122 | static inline void _ccv_nnc_try_set_pix_1(const int* const ofs, const int* const dim, const int* const tensor_dim, int* const* const scmd, const int* const cube_dim, const int* const cube_step, uint32_t* const cube, int offset) | |||
123 | { | |||
124 | const int s0 = (ofs[0] == 0) ? 0 : _ccv_nnc_mix_idx(scmd[0], ofs[0], cube_dim[0]) + 1; | |||
125 | const int d0 = ((ofs[0] + dim[0] == tensor_dim[0]) ? cube_dim[0] : _ccv_nnc_mix_idx(scmd[0], ofs[0] + ccv_max(1, dim[0])({ typeof (1) _a = (1); typeof (dim[0]) _b = (dim[0]); (_a > _b) ? _a : _b; }), cube_dim[0])) + 1; | |||
126 | assert(s0 >= 0 && d0 > s0)((void) sizeof ((s0 >= 0 && d0 > s0) ? 1 : 0), __extension__ ({ if (s0 >= 0 && d0 > s0) ; else __assert_fail ("s0 >= 0 && d0 > s0", "ccv_nnc_symbolic_graph_backward.c" , 126, __extension__ __PRETTY_FUNCTION__); })); | |||
127 | const int s1 = (ofs[1] == 0) ? 0 : _ccv_nnc_mix_idx(scmd[1], ofs[1], cube_dim[1]) + 1; | |||
128 | const int d1 = ((ofs[1] + dim[1] == tensor_dim[1]) ? cube_dim[1] : _ccv_nnc_mix_idx(scmd[1], ofs[1] + ccv_max(1, dim[1])({ typeof (1) _a = (1); typeof (dim[1]) _b = (dim[1]); (_a > _b) ? _a : _b; }), cube_dim[1])) + 1; | |||
129 | assert(s1 >= 0 && d1 > s1)((void) sizeof ((s1 >= 0 && d1 > s1) ? 1 : 0), __extension__ ({ if (s1 >= 0 && d1 > s1) ; else __assert_fail ("s1 >= 0 && d1 > s1", "ccv_nnc_symbolic_graph_backward.c" , 129, __extension__ __PRETTY_FUNCTION__); })); | |||
130 | int i, j; | |||
131 | const int step1 = cube_step[1]; | |||
132 | if (step1 == d0 - s0) | |||
133 | { | |||
134 | // Faster one, we can simply loop through. | |||
135 | for (i = s1 * step1; i < d1 * step1; i++) | |||
136 | cube[(offset + i) >> 5] |= (1u << ((offset + i) & 0x1f)); | |||
137 | } else { | |||
138 | offset += s1 * step1; | |||
139 | // There are gaps, slow one. | |||
140 | for (i = s1; i < d1; i++, offset += step1) | |||
141 | for (j = s0; j < d0; j++) | |||
142 | cube[(offset + j) >> 5] |= (1u << ((offset + j) & 0x1f)); | |||
143 | } | |||
144 | } | |||
145 | ||||
146 | static inline void _ccv_nnc_try_set_pix(const int* const ofs, const int* const dim, const int* const tensor_dim, int* const* const scmd, const int* const cube_dim, const int* const cube_step, uint32_t* const cube, int offset, const int dim_idx) | |||
147 | { | |||
148 | switch (dim_idx) | |||
149 | { | |||
150 | case 1: | |||
151 | _ccv_nnc_try_set_pix_1(ofs, dim, tensor_dim, scmd, cube_dim, cube_step, cube, offset); | |||
152 | return; | |||
153 | case 0: | |||
154 | _ccv_nnc_try_set_pix_0(ofs, dim, tensor_dim, scmd, cube_dim, cube_step, cube, offset); | |||
155 | return; | |||
156 | } | |||
157 | int i; | |||
158 | const int s = (ofs[dim_idx] == 0) ? 0 : _ccv_nnc_mix_idx(scmd[dim_idx], ofs[dim_idx], cube_dim[dim_idx]) + 1; | |||
159 | const int d = ((ofs[dim_idx] + dim[dim_idx] == tensor_dim[dim_idx]) ? cube_dim[dim_idx] : _ccv_nnc_mix_idx(scmd[dim_idx], ofs[dim_idx] + ccv_max(1, dim[dim_idx])({ typeof (1) _a = (1); typeof (dim[dim_idx]) _b = (dim[dim_idx ]); (_a > _b) ? _a : _b; }), cube_dim[dim_idx])) + 1; | |||
160 | assert(s >= 0 && d > s)((void) sizeof ((s >= 0 && d > s) ? 1 : 0), __extension__ ({ if (s >= 0 && d > s) ; else __assert_fail ( "s >= 0 && d > s", "ccv_nnc_symbolic_graph_backward.c" , 160, __extension__ __PRETTY_FUNCTION__); })); | |||
161 | for (i = s; i < d; i++) | |||
162 | _ccv_nnc_try_set_pix(ofs, dim, tensor_dim, scmd, cube_dim, cube_step, cube, offset + i * cube_step[dim_idx], dim_idx - 1); | |||
163 | } | |||
164 | ||||
165 | static int _ccv_nnc_tensor_ref_fully_assigned_with_aliases(const ccv_nnc_tensor_ref_t* const tensor_ref, const ccv_array_t* const autograd_tensor_symbols, const ccv_nnc_tensor_symbol_info_t* const tensor_symbol_info) | |||
166 | { | |||
167 | // Only work with tensor_ref of aliases. | |||
168 | assert(tensor_ref->alias_registry)((void) sizeof ((tensor_ref->alias_registry) ? 1 : 0), __extension__ ({ if (tensor_ref->alias_registry) ; else __assert_fail ( "tensor_ref->alias_registry", "ccv_nnc_symbolic_graph_backward.c" , 168, __extension__ __PRETTY_FUNCTION__); })); | |||
169 | const ccv_nnc_autograd_tensor_symbol_t* autograd = (ccv_nnc_autograd_tensor_symbol_t*)ccv_array_get(autograd_tensor_symbols, tensor_ref->d)((void*)(((char*)((autograd_tensor_symbols)->data)) + (size_t )(autograd_tensor_symbols)->rsize * (size_t)(tensor_ref-> d))); | |||
170 | assert(tensor_symbol_info[autograd->d].alias_ref == 0)((void) sizeof ((tensor_symbol_info[autograd->d].alias_ref == 0) ? 1 : 0), __extension__ ({ if (tensor_symbol_info[autograd ->d].alias_ref == 0) ; else __assert_fail ("tensor_symbol_info[autograd->d].alias_ref == 0" , "ccv_nnc_symbolic_graph_backward.c", 170, __extension__ __PRETTY_FUNCTION__ ); })); | |||
171 | const int* tensor_dim = tensor_symbol_info[autograd->d].info.dim; | |||
172 | const int tensor_count = ccv_nnc_dimension_count(tensor_dim); | |||
173 | int i, j; | |||
174 | for (i = 0; i < tensor_ref->alias_registry->rnum; i++) | |||
175 | { | |||
176 | const int d = *(int*)ccv_array_get(tensor_ref->alias_registry, i)((void*)(((char*)((tensor_ref->alias_registry)->data)) + (size_t)(tensor_ref->alias_registry)->rsize * (size_t) (i))); | |||
177 | assert(d < autograd_tensor_symbols->rnum)((void) sizeof ((d < autograd_tensor_symbols->rnum) ? 1 : 0), __extension__ ({ if (d < autograd_tensor_symbols-> rnum) ; else __assert_fail ("d < autograd_tensor_symbols->rnum" , "ccv_nnc_symbolic_graph_backward.c", 177, __extension__ __PRETTY_FUNCTION__ ); })); | |||
178 | const ccv_nnc_autograd_tensor_symbol_t* autograd = (ccv_nnc_autograd_tensor_symbol_t*)ccv_array_get(autograd_tensor_symbols, d)((void*)(((char*)((autograd_tensor_symbols)->data)) + (size_t )(autograd_tensor_symbols)->rsize * (size_t)(d))); | |||
179 | assert(tensor_symbol_info[autograd->d].alias_ref)((void) sizeof ((tensor_symbol_info[autograd->d].alias_ref ) ? 1 : 0), __extension__ ({ if (tensor_symbol_info[autograd-> d].alias_ref) ; else __assert_fail ("tensor_symbol_info[autograd->d].alias_ref" , "ccv_nnc_symbolic_graph_backward.c", 179, __extension__ __PRETTY_FUNCTION__ ); })); | |||
180 | const int* stride = tensor_symbol_info[autograd->d].stride; | |||
181 | // If this is just reshaped (i.e., dimension is the same, and inc covers the whole). We have fully assigned. | |||
182 | if (ccv_nnc_is_tensor_stride_packed(stride, tensor_symbol_info[autograd->d].info.dim) && ccv_nnc_dimension_count(tensor_symbol_info[autograd->d].info.dim) == tensor_count) | |||
183 | return 1; | |||
184 | // Otherwise if inc doesn't match original dim, it is not covered. | |||
185 | if (!ccv_nnc_is_tensor_stride_packed(stride, tensor_dim)) | |||
186 | return 0; | |||
187 | } | |||
188 | /* We need a solid cube (potentially hyper dimensional) to compute if there are overlaps. | |||
189 | * To make this cube as small as possible, we need to map the actual tensor dimension | |||
190 | * (therefore, we don't actually allocate the whole tensor to compute overlaps) to a smaller | |||
191 | * cube given the ofs and dim size of its aliases. | |||
192 | * | |||
193 | * The following code generated the dimension mapping (using scratch space) with binary search + insertion | |||
194 | * and then we fill the cube with a given tensor alias's dimensional information (ofs, dim). | |||
195 | * Afterwards, we simply need to check if the cube is totally filled up to know if this tensor | |||
196 | * is fully assigned with its aliases (if that is the case, we can skip zeroing for this tensor). | |||
197 | * | |||
198 | * There are several restrictions though to make this faster: 1). I cannot handle any cube that all side | |||
199 | * lengths combined larger than 1023 (scm only have 1024 scratch space). 2). I cannot handle any cube | |||
200 | * that the total volume is larger than 2048 * 8 (I only allocate 2K on stack for this). | |||
201 | * */ | |||
202 | int scm[1024]; // Having 1024 int scratch space for mapping dimensions. (Or sparse coordinate mapping). | |||
203 | int cube_dim[CCV_NNC_MAX_DIM_ALLOC(12)] = {}; // Mapping dimension size. | |||
204 | int cube_size = 1; | |||
205 | int* scmptr = scm; | |||
206 | for (i = 0; i < CCV_NNC_MAX_DIM_ALLOC(12) && tensor_dim[i]; i++) | |||
207 | { | |||
208 | int head = 0, tail = 0; // Note that we touched both the head and tail (otherwise this dimension is not fully covered). | |||
209 | int len = 0; | |||
210 | for (j = 0; j < tensor_ref->alias_registry->rnum; j++) | |||
211 | { | |||
212 | const int d = *(int*)ccv_array_get(tensor_ref->alias_registry, j)((void*)(((char*)((tensor_ref->alias_registry)->data)) + (size_t)(tensor_ref->alias_registry)->rsize * (size_t) (j))); | |||
213 | assert(d < autograd_tensor_symbols->rnum)((void) sizeof ((d < autograd_tensor_symbols->rnum) ? 1 : 0), __extension__ ({ if (d < autograd_tensor_symbols-> rnum) ; else __assert_fail ("d < autograd_tensor_symbols->rnum" , "ccv_nnc_symbolic_graph_backward.c", 213, __extension__ __PRETTY_FUNCTION__ ); })); | |||
214 | const ccv_nnc_autograd_tensor_symbol_t* autograd = (ccv_nnc_autograd_tensor_symbol_t*)ccv_array_get(autograd_tensor_symbols, d)((void*)(((char*)((autograd_tensor_symbols)->data)) + (size_t )(autograd_tensor_symbols)->rsize * (size_t)(d))); | |||
215 | assert(tensor_symbol_info[autograd->d].alias_ref)((void) sizeof ((tensor_symbol_info[autograd->d].alias_ref ) ? 1 : 0), __extension__ ({ if (tensor_symbol_info[autograd-> d].alias_ref) ; else __assert_fail ("tensor_symbol_info[autograd->d].alias_ref" , "ccv_nnc_symbolic_graph_backward.c", 215, __extension__ __PRETTY_FUNCTION__ ); })); | |||
216 | const int* ofs = tensor_symbol_info[autograd->d].ofs; | |||
217 | const int* dim = tensor_symbol_info[autograd->d].info.dim; | |||
218 | head = head || (ofs[i] == 0); | |||
219 | tail = tail || (ofs[i] + ccv_max(1, dim[i])({ typeof (1) _a = (1); typeof (dim[i]) _b = (dim[i]); (_a > _b) ? _a : _b; }) == tensor_dim[i]); | |||
220 | if (ofs[i] != 0) | |||
221 | len += _ccv_nnc_try_mix(scmptr, ofs[i], len); | |||
222 | if (scmptr - scm + len >= 1024) // Cannot handle that much, abort. | |||
223 | return 0; | |||
224 | if (ofs[i] + ccv_max(1, dim[i])({ typeof (1) _a = (1); typeof (dim[i]) _b = (dim[i]); (_a > _b) ? _a : _b; }) < tensor_dim[i]) | |||
225 | len += _ccv_nnc_try_mix(scmptr, ofs[i] + ccv_max(1, dim[i])({ typeof (1) _a = (1); typeof (dim[i]) _b = (dim[i]); (_a > _b) ? _a : _b; }), len); | |||
226 | if (scmptr - scm + len >= 1024) // Cannot handle that much, abort. | |||
227 | return 0; | |||
228 | } | |||
229 | if (!head || !tail) | |||
230 | return 0; | |||
231 | cube_size *= (len + 1); | |||
232 | cube_dim[i] = len; | |||
233 | scmptr += len; // Moving to next level. | |||
234 | } | |||
235 | // The cube map is too large, cannot do the computation, assume it is not fully assigned. | |||
236 | if (cube_size > 2048 * 8) | |||
237 | return 0; | |||
238 | // binary map to see if it fills up. | |||
239 | uint32_t cube[(cube_size + 31) >> 5]; | |||
240 | memset(cube, 0, sizeof(uint32_t) * ((cube_size + 31) >> 5)); | |||
241 | int* scmd[CCV_NNC_MAX_DIM_ALLOC(12)] = {}; // Sparse coordinate map at dimension x. | |||
242 | int cube_step[CCV_NNC_MAX_DIM_ALLOC(12)] = {}; | |||
243 | for (i = 0; i < CCV_NNC_MAX_DIM_ALLOC(12) && tensor_dim[i]; i++) | |||
244 | { | |||
245 | cube_step[i] = (i > 0) ? cube_step[i - 1] * (cube_dim[i - 1] + 1) : 1; | |||
246 | scmd[i] = (i > 0) ? scmd[i - 1] + cube_dim[i - 1] : scm; | |||
247 | } | |||
248 | const int max_dim = i; | |||
249 | for (i = 0; i < tensor_ref->alias_registry->rnum; i++) | |||
250 | { | |||
251 | const int d = *(int*)ccv_array_get(tensor_ref->alias_registry, i)((void*)(((char*)((tensor_ref->alias_registry)->data)) + (size_t)(tensor_ref->alias_registry)->rsize * (size_t) (i))); | |||
252 | assert(d < autograd_tensor_symbols->rnum)((void) sizeof ((d < autograd_tensor_symbols->rnum) ? 1 : 0), __extension__ ({ if (d < autograd_tensor_symbols-> rnum) ; else __assert_fail ("d < autograd_tensor_symbols->rnum" , "ccv_nnc_symbolic_graph_backward.c", 252, __extension__ __PRETTY_FUNCTION__ ); })); | |||
253 | const ccv_nnc_autograd_tensor_symbol_t* autograd = (ccv_nnc_autograd_tensor_symbol_t*)ccv_array_get(autograd_tensor_symbols, d)((void*)(((char*)((autograd_tensor_symbols)->data)) + (size_t )(autograd_tensor_symbols)->rsize * (size_t)(d))); | |||
254 | assert(tensor_symbol_info[autograd->d].alias_ref)((void) sizeof ((tensor_symbol_info[autograd->d].alias_ref ) ? 1 : 0), __extension__ ({ if (tensor_symbol_info[autograd-> d].alias_ref) ; else __assert_fail ("tensor_symbol_info[autograd->d].alias_ref" , "ccv_nnc_symbolic_graph_backward.c", 254, __extension__ __PRETTY_FUNCTION__ ); })); | |||
255 | const int* ofs = tensor_symbol_info[autograd->d].ofs; | |||
256 | const int* dim = tensor_symbol_info[autograd->d].info.dim; | |||
257 | _ccv_nnc_try_set_pix(ofs, dim, tensor_dim, scmd, cube_dim, cube_step, cube, 0, max_dim - 1); | |||
258 | } | |||
259 | // Compare to see now if the binary map filled up. If it filled up, we know it is fully assigned. | |||
260 | for (i = 0; i < (cube_size >> 5); i++) | |||
261 | if (cube[i] < 0xffffffff) | |||
262 | return 0; | |||
263 | if ((cube_size & 0x1f) > 0) | |||
264 | { | |||
265 | // Fetch the rest. | |||
266 | uint32_t r = 0; | |||
267 | for (i = 0; i < (cube_size & 0x1f); i++) | |||
268 | r |= (1u << i); | |||
269 | assert(cube[((cube_size + 31) >> 5) - 1] <= r)((void) sizeof ((cube[((cube_size + 31) >> 5) - 1] <= r) ? 1 : 0), __extension__ ({ if (cube[((cube_size + 31) >> 5) - 1] <= r) ; else __assert_fail ("cube[((cube_size + 31) >> 5) - 1] <= r" , "ccv_nnc_symbolic_graph_backward.c", 269, __extension__ __PRETTY_FUNCTION__ ); })); | |||
270 | if (cube[((cube_size + 31) >> 5) - 1] < r) | |||
271 | return 0; | |||
272 | } | |||
273 | return 1; | |||
274 | } | |||
275 | ||||
276 | static int _ccv_nnc_tensor_ref_version_find_init(const ccv_nnc_autograd_tensor_version_t* const tensor_ver) | |||
277 | { | |||
278 | int i; | |||
279 | for (i = 0; i < tensor_ver->ref_version->rnum; i++) | |||
280 | if (((ccv_nnc_tensor_ref_t*)ccv_array_get(tensor_ver->ref_version, i)((void*)(((char*)((tensor_ver->ref_version)->data)) + ( size_t)(tensor_ver->ref_version)->rsize * (size_t)(i))))->x < 0) | |||
281 | return i; | |||
282 | return -1; | |||
283 | } | |||
284 | ||||
285 | static void _ccv_nnc_graph_sum_autograd_tensor_versions(const int idx, const int d, const int exec_symbol_info_size, const ccv_nnc_tensor_symbol_info_t* const tensor_symbol_info, ccv_nnc_autograd_tensor_version_t* const tensor_ver, ccv_nnc_autograd_graph_exec_symbol_t* const autograd_execs, ccv_array_t* const autograd_tensor_symbols, ccv_array_t* const sum_or_set_execs) | |||
286 | { | |||
287 | int i, j; | |||
288 | assert(tensor_ver->c < tensor_ver->ref_version->rnum)((void) sizeof ((tensor_ver->c < tensor_ver->ref_version ->rnum) ? 1 : 0), __extension__ ({ if (tensor_ver->c < tensor_ver->ref_version->rnum) ; else __assert_fail ("tensor_ver->c < tensor_ver->ref_version->rnum" , "ccv_nnc_symbolic_graph_backward.c", 288, __extension__ __PRETTY_FUNCTION__ ); })); | |||
289 | const int input_size = tensor_ver->ref_version->rnum - tensor_ver->c; | |||
290 | int* inputs = (int*)ccmallocmalloc(sizeof(int) * input_size); | |||
291 | for (i = tensor_ver->c; i < tensor_ver->ref_version->rnum; i++) | |||
292 | inputs[i] = ((ccv_nnc_tensor_ref_t*)ccv_array_get(tensor_ver->ref_version, i)((void*)(((char*)((tensor_ver->ref_version)->data)) + ( size_t)(tensor_ver->ref_version)->rsize * (size_t)(i))))->d; | |||
293 | const ccv_nnc_autograd_tensor_symbol_t tensor_sym = { | |||
294 | .d = d | |||
295 | }; | |||
296 | ccv_array_push(autograd_tensor_symbols, &tensor_sym); | |||
297 | ccv_nnc_sum_or_set_graph_exec_symbol_t sum_exec = { | |||
298 | .input_size = input_size, | |||
299 | .inputs = inputs, | |||
300 | .output = autograd_tensor_symbols->rnum - 1 | |||
301 | }; | |||
302 | if (idx >= 0) | |||
303 | { | |||
304 | sum_exec.outgoings = ccv_array_new(sizeof(int), 1, 0); | |||
305 | ccv_array_push(sum_exec.outgoings, &idx); | |||
306 | } | |||
307 | ccv_array_push(sum_or_set_execs, &sum_exec); | |||
308 | const int outgoing = exec_symbol_info_size + sum_or_set_execs->rnum - 1; | |||
309 | for (i = tensor_ver->c; i < tensor_ver->ref_version->rnum; i++) | |||
310 | { | |||
311 | const ccv_nnc_tensor_ref_t* tensor_ref = (ccv_nnc_tensor_ref_t*)ccv_array_get(tensor_ver->ref_version, i)((void*)(((char*)((tensor_ver->ref_version)->data)) + ( size_t)(tensor_ver->ref_version)->rsize * (size_t)(i))); | |||
312 | const int x = tensor_ref->x; | |||
313 | if (x < 0) /* This is initialization tensor, it has to be occurred before the execution anyway. */ | |||
314 | { | |||
315 | // No alias. | |||
316 | assert(!tensor_ref->alias_registry)((void) sizeof ((!tensor_ref->alias_registry) ? 1 : 0), __extension__ ({ if (!tensor_ref->alias_registry) ; else __assert_fail ( "!tensor_ref->alias_registry", "ccv_nnc_symbolic_graph_backward.c" , 316, __extension__ __PRETTY_FUNCTION__); })); | |||
317 | // No associated additional execs. | |||
318 | assert(!tensor_ref->exec_registry)((void) sizeof ((!tensor_ref->exec_registry) ? 1 : 0), __extension__ ({ if (!tensor_ref->exec_registry) ; else __assert_fail ( "!tensor_ref->exec_registry", "ccv_nnc_symbolic_graph_backward.c" , 318, __extension__ __PRETTY_FUNCTION__); })); | |||
319 | continue; | |||
320 | } | |||
321 | if (x < exec_symbol_info_size) | |||
322 | { | |||
323 | ccv_nnc_autograd_graph_exec_symbol_t* back_exec = autograd_execs + x; | |||
324 | if (!back_exec->outgoings) | |||
325 | back_exec->outgoings = ccv_array_new(sizeof(int), 1, 0); | |||
326 | ccv_array_replace_unique_int(back_exec->outgoings, idx, outgoing); | |||
327 | } else { | |||
328 | // This tensor_ref is generated by the sum operation. | |||
329 | ccv_nnc_sum_or_set_graph_exec_symbol_t* sum_or_set = (ccv_nnc_sum_or_set_graph_exec_symbol_t*)ccv_array_get(sum_or_set_execs, x - exec_symbol_info_size)((void*)(((char*)((sum_or_set_execs)->data)) + (size_t)(sum_or_set_execs )->rsize * (size_t)(x - exec_symbol_info_size))); | |||
330 | ccv_array_replace_unique_int(sum_or_set->outgoings, idx, outgoing); | |||
331 | } | |||
332 | // If this tensor have associated alias, we need to init it to zeros when it is allocated (we only need to set a flag here) | |||
333 | // it is handled at compilation phase. | |||
334 | if (tensor_ref->alias_registry && | |||
335 | // Loop over to see if this tensor is fully occupied to avoid extra zero step. | |||
336 | !_ccv_nnc_tensor_ref_fully_assigned_with_aliases(tensor_ref, autograd_tensor_symbols, tensor_symbol_info)) | |||
337 | { | |||
338 | ccv_nnc_autograd_tensor_symbol_t* tensor_sym = (ccv_nnc_autograd_tensor_symbol_t*)ccv_array_get(autograd_tensor_symbols, tensor_ref->d)((void*)(((char*)((autograd_tensor_symbols)->data)) + (size_t )(autograd_tensor_symbols)->rsize * (size_t)(tensor_ref-> d))); | |||
339 | // By having alias_registry, what this symbol represents must not by an alias. | |||
340 | assert(tensor_sym->alias_ref == 0)((void) sizeof ((tensor_sym->alias_ref == 0) ? 1 : 0), __extension__ ({ if (tensor_sym->alias_ref == 0) ; else __assert_fail ( "tensor_sym->alias_ref == 0", "ccv_nnc_symbolic_graph_backward.c" , 340, __extension__ __PRETTY_FUNCTION__); })); | |||
341 | tensor_sym->flags = CCV_NNC_TENSOR_SYMBOL_INIT_ZEROS; | |||
342 | } | |||
343 | if (tensor_ref->exec_registry) | |||
344 | for (j = 0; j < tensor_ref->exec_registry->rnum; j++) | |||
345 | { | |||
346 | const int x = *(int*)ccv_array_get(tensor_ref->exec_registry, j)((void*)(((char*)((tensor_ref->exec_registry)->data)) + (size_t)(tensor_ref->exec_registry)->rsize * (size_t)( j))); | |||
347 | assert(x >= 0)((void) sizeof ((x >= 0) ? 1 : 0), __extension__ ({ if (x >= 0) ; else __assert_fail ("x >= 0", "ccv_nnc_symbolic_graph_backward.c" , 347, __extension__ __PRETTY_FUNCTION__); })); | |||
348 | // The exec_registry can only be generated by alias registry, therefore, it cannot reference to a sum operation. | |||
349 | assert(x < exec_symbol_info_size)((void) sizeof ((x < exec_symbol_info_size) ? 1 : 0), __extension__ ({ if (x < exec_symbol_info_size) ; else __assert_fail ("x < exec_symbol_info_size" , "ccv_nnc_symbolic_graph_backward.c", 349, __extension__ __PRETTY_FUNCTION__ ); })); | |||
350 | ccv_nnc_autograd_graph_exec_symbol_t* back_exec = autograd_execs + x; | |||
351 | if (!back_exec->outgoings) | |||
352 | back_exec->outgoings = ccv_array_new(sizeof(int), 1, 0); | |||
353 | ccv_array_replace_unique_int(back_exec->outgoings, idx, outgoing); | |||
354 | } | |||
355 | } | |||
356 | const ccv_nnc_tensor_ref_t tensor_ref = { | |||
357 | .d = autograd_tensor_symbols->rnum - 1, | |||
358 | .x = outgoing | |||
359 | }; | |||
360 | ccv_array_push(tensor_ver->ref_version, &tensor_ref); | |||
361 | /* Move the c pointer up to the latest summed result. */ | |||
362 | tensor_ver->c = tensor_ver->ref_version->rnum - 1; | |||
363 | } | |||
364 | ||||
365 | static int _ccv_nnc_tensor_ref_version_involve_alias(const ccv_nnc_tensor_ref_t* const tensor_ref, const ccv_array_t* const autograd_tensor_symbols, const ccv_nnc_tensor_symbol_info_t* const tensor_symbol_info, const ccv_nnc_tensor_symbol_info_t* const alias) | |||
366 | { | |||
367 | assert(alias->alias_ref > 0)((void) sizeof ((alias->alias_ref > 0) ? 1 : 0), __extension__ ({ if (alias->alias_ref > 0) ; else __assert_fail ("alias->alias_ref > 0" , "ccv_nnc_symbolic_graph_backward.c", 367, __extension__ __PRETTY_FUNCTION__ ); })); | |||
368 | // No alias_registry, must conflict (owns the whole band). | |||
369 | if (!tensor_ref->alias_registry) | |||
370 | return 1; | |||
371 | int i; | |||
372 | for (i = 0; i < tensor_ref->alias_registry->rnum; i++) | |||
373 | { | |||
374 | const int d = *(int*)ccv_array_get(tensor_ref->alias_registry, i)((void*)(((char*)((tensor_ref->alias_registry)->data)) + (size_t)(tensor_ref->alias_registry)->rsize * (size_t) (i))); | |||
375 | assert(d < autograd_tensor_symbols->rnum)((void) sizeof ((d < autograd_tensor_symbols->rnum) ? 1 : 0), __extension__ ({ if (d < autograd_tensor_symbols-> rnum) ; else __assert_fail ("d < autograd_tensor_symbols->rnum" , "ccv_nnc_symbolic_graph_backward.c", 375, __extension__ __PRETTY_FUNCTION__ ); })); | |||
376 | ccv_nnc_autograd_tensor_symbol_t* autograd = (ccv_nnc_autograd_tensor_symbol_t*)ccv_array_get(autograd_tensor_symbols, d)((void*)(((char*)((autograd_tensor_symbols)->data)) + (size_t )(autograd_tensor_symbols)->rsize * (size_t)(d))); | |||
377 | if (ccv_nnc_over_tensor_symbol_aliases(tensor_symbol_info + autograd->d, alias)) | |||
378 | return 1; | |||
379 | } | |||
380 | // All aliases referenced by this ref_version doesn't overlap with the provided one, thus, there is no conflict at all. | |||
381 | return 0; | |||
382 | } | |||
383 | ||||
384 | static int _ccv_nnc_tensor_ref_version_find_alias(const ccv_nnc_tensor_ref_t* const tensor_ref, const ccv_array_t* const autograd_tensor_symbols, const ccv_nnc_tensor_symbol_info_t* const tensor_symbol_info, const ccv_nnc_tensor_symbol_info_t* const alias) | |||
385 | { | |||
386 | assert(alias->alias_ref > 0)((void) sizeof ((alias->alias_ref > 0) ? 1 : 0), __extension__ ({ if (alias->alias_ref > 0) ; else __assert_fail ("alias->alias_ref > 0" , "ccv_nnc_symbolic_graph_backward.c", 386, __extension__ __PRETTY_FUNCTION__ ); })); | |||
387 | // No alias_registry, thus, cannot find the exact matched alias. | |||
388 | if (!tensor_ref->alias_registry) | |||
389 | return -1; | |||
390 | int i; | |||
391 | for (i = 0; i < tensor_ref->alias_registry->rnum; i++) | |||
392 | { | |||
393 | const int d = *(int*)ccv_array_get(tensor_ref->alias_registry, i)((void*)(((char*)((tensor_ref->alias_registry)->data)) + (size_t)(tensor_ref->alias_registry)->rsize * (size_t) (i))); | |||
394 | assert(d < autograd_tensor_symbols->rnum)((void) sizeof ((d < autograd_tensor_symbols->rnum) ? 1 : 0), __extension__ ({ if (d < autograd_tensor_symbols-> rnum) ; else __assert_fail ("d < autograd_tensor_symbols->rnum" , "ccv_nnc_symbolic_graph_backward.c", 394, __extension__ __PRETTY_FUNCTION__ ); })); | |||
395 | ccv_nnc_autograd_tensor_symbol_t* autograd = (ccv_nnc_autograd_tensor_symbol_t*)ccv_array_get(autograd_tensor_symbols, d)((void*)(((char*)((autograd_tensor_symbols)->data)) + (size_t )(autograd_tensor_symbols)->rsize * (size_t)(d))); | |||
396 | // This must reference to an alias. | |||
397 | assert(tensor_symbol_info[autograd->d].alias_ref)((void) sizeof ((tensor_symbol_info[autograd->d].alias_ref ) ? 1 : 0), __extension__ ({ if (tensor_symbol_info[autograd-> d].alias_ref) ; else __assert_fail ("tensor_symbol_info[autograd->d].alias_ref" , "ccv_nnc_symbolic_graph_backward.c", 397, __extension__ __PRETTY_FUNCTION__ ); })); | |||
398 | const int* stride = tensor_symbol_info[autograd->d].stride; | |||
399 | const int* ofs = tensor_symbol_info[autograd->d].ofs; | |||
400 | const int* dim = tensor_symbol_info[autograd->d].info.dim; | |||
401 | // If everything matches, this is the required alias. | |||
402 | if (memcmp(stride, alias->stride, sizeof(int) * CCV_NNC_MAX_DIM_ALLOC(12)) == 0 && | |||
403 | memcmp(ofs, alias->ofs, sizeof(int) * CCV_NNC_MAX_DIM_ALLOC(12)) == 0 && | |||
404 | memcmp(dim, alias->info.dim, sizeof(int) * CCV_NNC_MAX_DIM_ALLOC(12)) == 0) | |||
405 | return d; | |||
406 | } | |||
407 | return -1; | |||
408 | } | |||
409 | ||||
410 | static int _ccv_nnc_tensor_ref_version_has_this_alias_exclusively(const ccv_nnc_tensor_ref_t* const tensor_ref, const ccv_array_t* const autograd_tensor_symbols, const ccv_nnc_tensor_symbol_info_t* const tensor_symbol_info, const ccv_nnc_tensor_symbol_info_t* const alias) | |||
411 | { | |||
412 | assert(alias->alias_ref > 0)((void) sizeof ((alias->alias_ref > 0) ? 1 : 0), __extension__ ({ if (alias->alias_ref > 0) ; else __assert_fail ("alias->alias_ref > 0" , "ccv_nnc_symbolic_graph_backward.c", 412, __extension__ __PRETTY_FUNCTION__ ); })); | |||
413 | // No alias_registry, thus, cannot find the exact matched alias. | |||
414 | if (!tensor_ref->alias_registry) | |||
415 | return 0; | |||
416 | int i; | |||
417 | for (i = 0; i < tensor_ref->alias_registry->rnum; i++) | |||
418 | { | |||
419 | const int d = *(int*)ccv_array_get(tensor_ref->alias_registry, i)((void*)(((char*)((tensor_ref->alias_registry)->data)) + (size_t)(tensor_ref->alias_registry)->rsize * (size_t) (i))); | |||
420 | assert(d < autograd_tensor_symbols->rnum)((void) sizeof ((d < autograd_tensor_symbols->rnum) ? 1 : 0), __extension__ ({ if (d < autograd_tensor_symbols-> rnum) ; else __assert_fail ("d < autograd_tensor_symbols->rnum" , "ccv_nnc_symbolic_graph_backward.c", 420, __extension__ __PRETTY_FUNCTION__ ); })); | |||
421 | ccv_nnc_autograd_tensor_symbol_t* autograd = (ccv_nnc_autograd_tensor_symbol_t*)ccv_array_get(autograd_tensor_symbols, d)((void*)(((char*)((autograd_tensor_symbols)->data)) + (size_t )(autograd_tensor_symbols)->rsize * (size_t)(d))); | |||
422 | // This must reference to an alias. | |||
423 | assert(tensor_symbol_info[autograd->d].alias_ref)((void) sizeof ((tensor_symbol_info[autograd->d].alias_ref ) ? 1 : 0), __extension__ ({ if (tensor_symbol_info[autograd-> d].alias_ref) ; else __assert_fail ("tensor_symbol_info[autograd->d].alias_ref" , "ccv_nnc_symbolic_graph_backward.c", 423, __extension__ __PRETTY_FUNCTION__ ); })); | |||
424 | const int* stride = tensor_symbol_info[autograd->d].stride; | |||
425 | const int* ofs = tensor_symbol_info[autograd->d].ofs; | |||
426 | const int* dim = tensor_symbol_info[autograd->d].info.dim; | |||
427 | if (memcmp(stride, alias->stride, sizeof(int) * CCV_NNC_MAX_DIM_ALLOC(12)) != 0 || | |||
428 | memcmp(ofs, alias->ofs, sizeof(int) * CCV_NNC_MAX_DIM_ALLOC(12)) != 0 || | |||
429 | memcmp(dim, alias->info.dim, sizeof(int) * CCV_NNC_MAX_DIM_ALLOC(12)) != 0) | |||
430 | return 0; | |||
431 | } | |||
432 | // If everything matches for every alias in registry, we can use any of the alias directly. | |||
433 | return 1; | |||
434 | } | |||
435 | ||||
436 | static int _ccv_nnc_graph_sum_autograd_tensor_versions_alias(const int idx, const int d, const ccv_nnc_tensor_symbol_info_t* const tensor_symbol_info, const int exec_symbol_info_size, const ccv_nnc_tensor_symbol_info_t* const alias, ccv_nnc_autograd_tensor_version_t* const tensor_ver, ccv_nnc_autograd_graph_exec_symbol_t* const autograd_execs, ccv_array_t* const autograd_tensor_symbols, ccv_array_t* const sum_or_set_execs) | |||
437 | { | |||
438 | assert(tensor_ver->c < tensor_ver->ref_version->rnum)((void) sizeof ((tensor_ver->c < tensor_ver->ref_version ->rnum) ? 1 : 0), __extension__ ({ if (tensor_ver->c < tensor_ver->ref_version->rnum) ; else __assert_fail ("tensor_ver->c < tensor_ver->ref_version->rnum" , "ccv_nnc_symbolic_graph_backward.c", 438, __extension__ __PRETTY_FUNCTION__ ); })); | |||
439 | int i, j = 0; | |||
440 | struct { | |||
441 | int k; | |||
442 | int i; | |||
443 | } kd[tensor_ver->ref_version->rnum - tensor_ver->c]; | |||
444 | for (i = tensor_ver->c; i < tensor_ver->ref_version->rnum; i++) | |||
445 | { | |||
446 | ccv_nnc_tensor_ref_t* tensor_ref = (ccv_nnc_tensor_ref_t*)ccv_array_get(tensor_ver->ref_version, i)((void*)(((char*)((tensor_ver->ref_version)->data)) + ( size_t)(tensor_ver->ref_version)->rsize * (size_t)(i))); | |||
447 | const int k = _ccv_nnc_tensor_ref_version_find_alias(tensor_ref, autograd_tensor_symbols, tensor_symbol_info, alias); | |||
448 | if (k >= 0) | |||
449 | kd[j++] = (typeof(kd[0])){ | |||
450 | .k = k, .i = i | |||
451 | }; | |||
452 | else if (_ccv_nnc_tensor_ref_version_involve_alias(tensor_ref, autograd_tensor_symbols, tensor_symbol_info, alias)) | |||
453 | kd[j++] = (typeof(kd[0])) { | |||
454 | .k = -1, .i = i // It has dependency to the original tensor (non-alias) now, label this with highest bit. | |||
455 | }; | |||
456 | } | |||
457 | // Can only find one. This is the easy case, we can simply return that symbol (or its alias). | |||
458 | if (j == 1) | |||
459 | { | |||
460 | if (kd[0].k >= 0) | |||
461 | return kd[0].k; // Only can find one alias, that is the one. | |||
462 | // Otherwise, need to create a new alias. | |||
463 | ccv_nnc_tensor_ref_t* tensor_ref = (ccv_nnc_tensor_ref_t*)ccv_array_get(tensor_ver->ref_version, kd[0].i)((void*)(((char*)((tensor_ver->ref_version)->data)) + ( size_t)(tensor_ver->ref_version)->rsize * (size_t)(kd[0 ].i))); | |||
464 | ccv_nnc_autograd_tensor_symbol_t* ref = (ccv_nnc_autograd_tensor_symbol_t*)ccv_array_get(autograd_tensor_symbols, tensor_ref->d)((void*)(((char*)((autograd_tensor_symbols)->data)) + (size_t )(autograd_tensor_symbols)->rsize * (size_t)(tensor_ref-> d))); | |||
465 | // Since we create new alias, we need to set the referenced one to be allocated with 0s. | |||
466 | if (ref->alias_ref) // If this is an alias, it has to be zero initialized. | |||
467 | { | |||
468 | ref = (ccv_nnc_autograd_tensor_symbol_t*)ccv_array_get(autograd_tensor_symbols, ref->alias_ref - 1)((void*)(((char*)((autograd_tensor_symbols)->data)) + (size_t )(autograd_tensor_symbols)->rsize * (size_t)(ref->alias_ref - 1))); | |||
469 | assert(ref->alias_ref == 0)((void) sizeof ((ref->alias_ref == 0) ? 1 : 0), __extension__ ({ if (ref->alias_ref == 0) ; else __assert_fail ("ref->alias_ref == 0" , "ccv_nnc_symbolic_graph_backward.c", 469, __extension__ __PRETTY_FUNCTION__ ); })); // This is original. | |||
470 | ref->flags = CCV_NNC_TENSOR_SYMBOL_INIT_ZEROS; | |||
471 | } else if (tensor_ref->alias_registry && // Otherwise, to see if this symbol is fully occupied. | |||
472 | // Loop over to see if this tensor is fully occupied to avoid extra zero step. | |||
473 | !_ccv_nnc_tensor_ref_fully_assigned_with_aliases(tensor_ref, autograd_tensor_symbols, tensor_symbol_info)) { | |||
474 | ref->flags = CCV_NNC_TENSOR_SYMBOL_INIT_ZEROS; | |||
475 | } | |||
476 | ccv_nnc_autograd_tensor_symbol_t tensor_sym = { | |||
477 | .d = d, | |||
478 | .alias_ref = tensor_ref->d + 1 | |||
479 | }; | |||
480 | ccv_array_push(autograd_tensor_symbols, &tensor_sym); | |||
481 | const int ad = autograd_tensor_symbols->rnum - 1; | |||
482 | if (tensor_ref->alias_registry) // Only push this when it has an alias registry (otherwise it already conflict with everyone). | |||
483 | ccv_array_push(tensor_ref->alias_registry, &ad); | |||
484 | if (tensor_ref->x >= exec_symbol_info_size && idx >= 0) | |||
485 | { | |||
486 | ccv_nnc_sum_or_set_graph_exec_symbol_t* const sum_or_set_exec = (ccv_nnc_sum_or_set_graph_exec_symbol_t*)ccv_array_get(sum_or_set_execs, tensor_ref->x - exec_symbol_info_size)((void*)(((char*)((sum_or_set_execs)->data)) + (size_t)(sum_or_set_execs )->rsize * (size_t)(tensor_ref->x - exec_symbol_info_size ))); | |||
487 | // This may be summed, thus, we need to create a connection between this and the sum. | |||
488 | if (!sum_or_set_exec->outgoings) | |||
489 | sum_or_set_exec->outgoings = ccv_array_new(sizeof(int), 1, 0); | |||
490 | ccv_array_push(sum_or_set_exec->outgoings, &idx); | |||
491 | } | |||
492 | // The newly inserted tensor symbol. | |||
493 | return ad; | |||
494 | } | |||
495 | // Otherwise, we need to create the sum operation out of these. | |||
496 | const int input_size = j; | |||
497 | int has_this_alias_exclusively = 1; | |||
498 | int* inputs = input_size > 0 ? (int*)ccmallocmalloc(sizeof(int) * input_size) : 0; | |||
499 | for (i = 0; i < input_size; i++) | |||
500 | { | |||
501 | ccv_nnc_tensor_ref_t* tensor_ref = (ccv_nnc_tensor_ref_t*)ccv_array_get(tensor_ver->ref_version, kd[i].i)((void*)(((char*)((tensor_ver->ref_version)->data)) + ( size_t)(tensor_ver->ref_version)->rsize * (size_t)(kd[i ].i))); | |||
502 | // Can take a fast path if every ref involved has the same alias, our sum operation can be faster (using alias directly). | |||
503 | if (has_this_alias_exclusively && kd[i].k >= 0 && _ccv_nnc_tensor_ref_version_has_this_alias_exclusively(tensor_ref, autograd_tensor_symbols, tensor_symbol_info, alias)) | |||
504 | inputs[i] = *(int*)ccv_array_get(tensor_ref->alias_registry, 0)((void*)(((char*)((tensor_ref->alias_registry)->data)) + (size_t)(tensor_ref->alias_registry)->rsize * (size_t) (0))); // Assigning the alias. | |||
505 | else { | |||
506 | if (has_this_alias_exclusively) | |||
507 | { | |||
508 | has_this_alias_exclusively = 0; | |||
509 | for (j = 0; j < i; j++) | |||
510 | inputs[j] = ((ccv_nnc_tensor_ref_t*)ccv_array_get(tensor_ver->ref_version, kd[j].i)((void*)(((char*)((tensor_ver->ref_version)->data)) + ( size_t)(tensor_ver->ref_version)->rsize * (size_t)(kd[j ].i))))->d; | |||
511 | } | |||
512 | inputs[i] = tensor_ref->d; | |||
513 | } | |||
514 | } | |||
515 | ccv_nnc_autograd_tensor_symbol_t tensor_sym = { | |||
516 | .d = alias->alias_ref - 1 | |||
517 | }; | |||
518 | ccv_array_push(autograd_tensor_symbols, &tensor_sym); | |||
519 | const int tensor_ref_d = autograd_tensor_symbols->rnum - 1; | |||
520 | tensor_sym.d = d; | |||
521 | tensor_sym.alias_ref = tensor_ref_d + 1; | |||
522 | ccv_array_push(autograd_tensor_symbols, &tensor_sym); | |||
523 | const int ad = autograd_tensor_symbols->rnum - 1; | |||
524 | ccv_nnc_sum_or_set_graph_exec_symbol_t sum_exec = { | |||
525 | .input_size = input_size, | |||
526 | .inputs = inputs, | |||
527 | .output = has_this_alias_exclusively ? ad : tensor_ref_d /* If has this alias exclusively, the output should be alias as well. Otherwise the output is the real tensor. */ | |||
528 | }; | |||
529 | if (idx >= 0) | |||
530 | { | |||
531 | sum_exec.outgoings = ccv_array_new(sizeof(int), 1, 0); | |||
532 | ccv_array_push(sum_exec.outgoings, &idx); | |||
533 | } | |||
534 | ccv_array_push(sum_or_set_execs, &sum_exec); | |||
535 | const int outgoing = exec_symbol_info_size + sum_or_set_execs->rnum - 1; | |||
536 | int no_alias_registry = 0; | |||
537 | for (i = 0; i < input_size; i++) | |||
538 | { | |||
539 | ccv_nnc_tensor_ref_t* tensor_ref = (ccv_nnc_tensor_ref_t*)ccv_array_get(tensor_ver->ref_version, kd[i].i)((void*)(((char*)((tensor_ver->ref_version)->data)) + ( size_t)(tensor_ver->ref_version)->rsize * (size_t)(kd[i ].i))); | |||
540 | if (!has_this_alias_exclusively) | |||
541 | { | |||
542 | // If the sum operation is not operating on one alias. I need to zero this tensor out when it is first | |||
543 | // allocated (see discussions around the flags I use). | |||
544 | ccv_nnc_autograd_tensor_symbol_t* tensor_sym = (ccv_nnc_autograd_tensor_symbol_t*)ccv_array_get(autograd_tensor_symbols, tensor_ref->d)((void*)(((char*)((autograd_tensor_symbols)->data)) + (size_t )(autograd_tensor_symbols)->rsize * (size_t)(tensor_ref-> d))); | |||
545 | if (tensor_sym->alias_ref) | |||
546 | { | |||
547 | // Find the original tensor_sym and set its flags (I prefer to set flags on its original). | |||
548 | ccv_nnc_autograd_tensor_symbol_t* ref = (ccv_nnc_autograd_tensor_symbol_t*)ccv_array_get(autograd_tensor_symbols, tensor_sym->alias_ref - 1)((void*)(((char*)((autograd_tensor_symbols)->data)) + (size_t )(autograd_tensor_symbols)->rsize * (size_t)(tensor_sym-> alias_ref - 1))); | |||
549 | assert(ref->alias_ref == 0)((void) sizeof ((ref->alias_ref == 0) ? 1 : 0), __extension__ ({ if (ref->alias_ref == 0) ; else __assert_fail ("ref->alias_ref == 0" , "ccv_nnc_symbolic_graph_backward.c", 549, __extension__ __PRETTY_FUNCTION__ ); })); // This is original. | |||
550 | ref->flags = CCV_NNC_TENSOR_SYMBOL_INIT_ZEROS; | |||
551 | } else if (tensor_ref->alias_registry && // Otherwise, to see if this symbol is fully occupied. | |||
552 | // Loop over to see if this tensor is fully occupied to avoid extra zero step. | |||
553 | !_ccv_nnc_tensor_ref_fully_assigned_with_aliases(tensor_ref, autograd_tensor_symbols, tensor_symbol_info)) { | |||
554 | tensor_sym->flags = CCV_NNC_TENSOR_SYMBOL_INIT_ZEROS; | |||
555 | } | |||
556 | } | |||
557 | // Check to see if any of these tensors doesn't have alias. | |||
558 | no_alias_registry |= (!tensor_ref->alias_registry); | |||
559 | const int x = tensor_ref->x; | |||
560 | assert(x >= 0)((void) sizeof ((x >= 0) ? 1 : 0), __extension__ ({ if (x >= 0) ; else __assert_fail ("x >= 0", "ccv_nnc_symbolic_graph_backward.c" , 560, __extension__ __PRETTY_FUNCTION__); })); /* Otherwise, this is initialization tensor, which is impossible to be summed up by. */ | |||
561 | if (x < exec_symbol_info_size) | |||
562 | { | |||
563 | ccv_nnc_autograd_graph_exec_symbol_t* back_exec = autograd_execs + x; | |||
564 | if (!back_exec->outgoings) | |||
565 | back_exec->outgoings = ccv_array_new(sizeof(int), 1, 0); | |||
566 | ccv_array_push(back_exec->outgoings, &outgoing); | |||
567 | } else { | |||
568 | ccv_nnc_sum_or_set_graph_exec_symbol_t* sum_or_set = (ccv_nnc_sum_or_set_graph_exec_symbol_t*)ccv_array_get(sum_or_set_execs, x - exec_symbol_info_size)((void*)(((char*)((sum_or_set_execs)->data)) + (size_t)(sum_or_set_execs )->rsize * (size_t)(x - exec_symbol_info_size))); | |||
569 | ccv_array_push(sum_or_set->outgoings, &outgoing); | |||
570 | } | |||
571 | if (tensor_ref->exec_registry) | |||
572 | for (j = 0; j < tensor_ref->exec_registry->rnum; j++) | |||
573 | { | |||
574 | const int x = *(int*)ccv_array_get(tensor_ref->exec_registry, j)((void*)(((char*)((tensor_ref->exec_registry)->data)) + (size_t)(tensor_ref->exec_registry)->rsize * (size_t)( j))); | |||
575 | assert(x >= 0)((void) sizeof ((x >= 0) ? 1 : 0), __extension__ ({ if (x >= 0) ; else __assert_fail ("x >= 0", "ccv_nnc_symbolic_graph_backward.c" , 575, __extension__ __PRETTY_FUNCTION__); })); /* Otherwise, this is initialization tensor, which is impossible to be summed up by. */ | |||
576 | assert(x < exec_symbol_info_size)((void) sizeof ((x < exec_symbol_info_size) ? 1 : 0), __extension__ ({ if (x < exec_symbol_info_size) ; else __assert_fail ("x < exec_symbol_info_size" , "ccv_nnc_symbolic_graph_backward.c", 576, __extension__ __PRETTY_FUNCTION__ ); })); // exec_registry is only used by alias_registry, it simply cannot reference to a sum operation. | |||
577 | ccv_nnc_autograd_graph_exec_symbol_t* back_exec = autograd_execs + x; | |||
578 | if (!back_exec->outgoings) | |||
579 | back_exec->outgoings = ccv_array_new(sizeof(int), 1, 0); | |||
580 | ccv_array_push(back_exec->outgoings, &outgoing); | |||
581 | } | |||
582 | } | |||
583 | const ccv_nnc_tensor_ref_t tensor_ref = { | |||
584 | .d = tensor_ref_d, | |||
585 | .x = outgoing, | |||
586 | .exec_registry = 0, // I don't need to take execution dependencies because this tensor is generated by sum, therefore, we already take that dependency. | |||
587 | .alias_registry = !no_alias_registry || has_this_alias_exclusively ? ccv_array_new(sizeof(int), 1, 0) : 0 | |||
588 | }; | |||
589 | // If there is no alias registry, then we take the whole tensor ref as one. | |||
590 | if (!no_alias_registry || has_this_alias_exclusively) | |||
591 | { | |||
592 | // If this tensor ref contains multiple different types of alias, have to add them together (otherwise | |||
593 | // the computation for if there is an empty slot in this tensor ref is not correct without all the | |||
594 | // occupancy availability information). | |||
595 | if (!has_this_alias_exclusively) | |||
596 | for (i = 0; i < input_size; i++) | |||
597 | { | |||
598 | ccv_nnc_tensor_ref_t* ref = (ccv_nnc_tensor_ref_t*)ccv_array_get(tensor_ver->ref_version, kd[i].i)((void*)(((char*)((tensor_ver->ref_version)->data)) + ( size_t)(tensor_ver->ref_version)->rsize * (size_t)(kd[i ].i))); | |||
599 | assert(ref->alias_registry)((void) sizeof ((ref->alias_registry) ? 1 : 0), __extension__ ({ if (ref->alias_registry) ; else __assert_fail ("ref->alias_registry" , "ccv_nnc_symbolic_graph_backward.c", 599, __extension__ __PRETTY_FUNCTION__ ); })); | |||
600 | // It may get duplicates. But whatever, won't matter the computation. | |||
601 | for (j = 0; j < ref->alias_registry->rnum; j++) | |||
602 | ccv_array_push(tensor_ref.alias_registry, ccv_array_get(ref->alias_registry, j)((void*)(((char*)((ref->alias_registry)->data)) + (size_t )(ref->alias_registry)->rsize * (size_t)(j)))); | |||
603 | } | |||
604 | ccv_array_push(tensor_ref.alias_registry, &ad); | |||
605 | } | |||
606 | assert(input_size <= tensor_ver->ref_version->rnum - tensor_ver->c)((void) sizeof ((input_size <= tensor_ver->ref_version-> rnum - tensor_ver->c) ? 1 : 0), __extension__ ({ if (input_size <= tensor_ver->ref_version->rnum - tensor_ver->c ) ; else __assert_fail ("input_size <= tensor_ver->ref_version->rnum - tensor_ver->c" , "ccv_nnc_symbolic_graph_backward.c", 606, __extension__ __PRETTY_FUNCTION__ ); })); | |||
607 | ccv_nnc_tensor_ref_t x; | |||
608 | for (i = 0; i < input_size; i++) | |||
609 | // If the current one (i + tensor_ver->c) is smaller than the one referenced to, exchange. | |||
610 | if (kd[i].i > i + tensor_ver->c) | |||
611 | CCV_SWAP(*(ccv_nnc_tensor_ref_t*)ccv_array_get(tensor_ver->ref_version, i + tensor_ver->c), *(ccv_nnc_tensor_ref_t*)ccv_array_get(tensor_ver->ref_version, kd[i].i), x)((x) = (*(ccv_nnc_tensor_ref_t*)((void*)(((char*)((tensor_ver ->ref_version)->data)) + (size_t)(tensor_ver->ref_version )->rsize * (size_t)(i + tensor_ver->c)))), (*(ccv_nnc_tensor_ref_t *)((void*)(((char*)((tensor_ver->ref_version)->data)) + (size_t)(tensor_ver->ref_version)->rsize * (size_t)(i + tensor_ver->c)))) = (*(ccv_nnc_tensor_ref_t*)((void*)(((char *)((tensor_ver->ref_version)->data)) + (size_t)(tensor_ver ->ref_version)->rsize * (size_t)(kd[i].i)))), (*(ccv_nnc_tensor_ref_t *)((void*)(((char*)((tensor_ver->ref_version)->data)) + (size_t)(tensor_ver->ref_version)->rsize * (size_t)(kd [i].i)))) = (x)); | |||
612 | ccv_array_push(tensor_ver->ref_version, &tensor_ref); | |||
613 | // We've consumed input_size tensor refs, now move c up to the pointer of non-consumed tensors. | |||
614 | tensor_ver->c += input_size; | |||
615 | return ad; | |||
616 | } | |||
617 | ||||
618 | typedef struct ccv_nnc_symbolic_graph_backward_prep_s { | |||
619 | int exec_symbol_info_size; // Number of graph exec symbols before adding any new symbols related to automatic differentiation. | |||
620 | int tensor_symbol_info_size; // Number of tensor symbols before adding anything new. | |||
621 | int sub_prep_size; | |||
622 | ccv_nnc_graph_exec_symbol_info_t* exec_symbol_info; | |||
623 | ccv_nnc_tensor_symbol_info_t* tensor_symbol_info; | |||
624 | ccv_nnc_graph_backward_info_t* backward_info; // Corresponding to forward graph exec symbol info, it is exactly in reverse. | |||
625 | ccv_nnc_graph_visit_t* forward_visit; // The visitor structure (top sorted index) when doing traversal. | |||
626 | ccv_nnc_graph_visit_t* backward_visit; // The visitor structure (top sorted index) when doing reverse traversal. | |||
627 | ccv_nnc_autograd_graph_exec_symbol_t* autograd_execs; // The graph exec symbols we need for automatic differentiation. This is a 1:1 mapping for forward graph exec symbols, however, unlike backward_info, its outgoings may be more complex (may contain outgoing flows to sum nodes). | |||
628 | ccv_nnc_autograd_tensor_version_t* autograd_tensor_versions; // Corresponding to forward tensor symbols, each may contain multiple versions (due to multi-write). | |||
629 | ccv_array_t* autograd_tensor_symbols; // The tensor symbols we need for automatic differentiation (it may not be 1:1 mapping). | |||
630 | ccv_array_t* sum_or_set_execs; // The sum nodes, because in reverse mode, a tensor could have multiple versions, we need to sum them up before use. | |||
631 | struct ccv_nnc_symbolic_graph_backward_prep_s* sub_preps; // The preps of its sub-graphs. | |||
632 | // Pointers not managed by this struct | |||
633 | ccv_nnc_symbolic_graph_t* graph; | |||
634 | } ccv_nnc_symbolic_graph_backward_prep_t; | |||
635 | ||||
636 | static ccv_nnc_symbolic_graph_backward_prep_t _ccv_nnc_symbolic_graph_backward_prep(const ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_graph_exec_symbol_t* const sources, const int source_size, const ccv_nnc_graph_exec_symbol_t* const destinations, const int destination_size) | |||
637 | { | |||
638 | const int exec_symbol_info_size = graph->exec_symbol_info->rnum; | |||
639 | assert(exec_symbol_info_size > 0)((void) sizeof ((exec_symbol_info_size > 0) ? 1 : 0), __extension__ ({ if (exec_symbol_info_size > 0) ; else __assert_fail ("exec_symbol_info_size > 0" , "ccv_nnc_symbolic_graph_backward.c", 639, __extension__ __PRETTY_FUNCTION__ ); })); | |||
640 | const int tensor_symbol_info_size = graph->tensor_symbol_info->rnum; | |||
641 | assert(tensor_symbol_info_size > 0)((void) sizeof ((tensor_symbol_info_size > 0) ? 1 : 0), __extension__ ({ if (tensor_symbol_info_size > 0) ; else __assert_fail ( "tensor_symbol_info_size > 0", "ccv_nnc_symbolic_graph_backward.c" , 641, __extension__ __PRETTY_FUNCTION__); })); | |||
642 | ccv_nnc_graph_exec_symbol_info_t* exec_symbol_info = (ccv_nnc_graph_exec_symbol_info_t*)ccmallocmalloc(sizeof(ccv_nnc_graph_exec_symbol_info_t) * exec_symbol_info_size); | |||
643 | ccv_nnc_tensor_symbol_info_t* tensor_symbol_info = (ccv_nnc_tensor_symbol_info_t*)ccmallocmalloc(sizeof(ccv_nnc_tensor_symbol_info_t) * tensor_symbol_info_size); | |||
644 | ccv_nnc_graph_visit_t* forward_visit = ccv_nnc_graph_visit_new(graph, (ccv_nnc_graph_exec_symbol_info_t*)ccv_array_get(graph->exec_symbol_info, 0), exec_symbol_info_size, sources, source_size, destinations, destination_size, 0)({ ccv_nnc_graph_visit_t* _visit_ = (ccv_nnc_graph_visit_t*)malloc (sizeof(ccv_nnc_graph_visit_t) + sizeof(_visit_->node[0]) * ((exec_symbol_info_size) - 1)); _visit_->size = 0; do { typedef struct { int8_t d; int8_t r; uint16_t c; int32_t edges; } ccv_nnc_incoming_t ; int _i_, _j_; int _incoming_edges_ = 0; for (_i_ = 0; _i_ < (exec_symbol_info_size); _i_++) _incoming_edges_ += (((ccv_nnc_graph_exec_symbol_info_t *)((void*)(((char*)((graph->exec_symbol_info)->data)) + (size_t)(graph->exec_symbol_info)->rsize * (size_t)(0) )))[_i_].outgoings) ? ((ccv_nnc_graph_exec_symbol_info_t*)((void *)(((char*)((graph->exec_symbol_info)->data)) + (size_t )(graph->exec_symbol_info)->rsize * (size_t)(0))))[_i_] .outgoings->rnum : 0; const int _heap_mem_ = ((exec_symbol_info_size ) + _incoming_edges_ > 1024); ccv_nnc_incoming_t* _incomings_ ; if (_heap_mem_) _incomings_ = (ccv_nnc_incoming_t*)malloc(sizeof (ccv_nnc_incoming_t) * (exec_symbol_info_size) + sizeof(int32_t ) * ((exec_symbol_info_size) * 2 + _incoming_edges_)); else _incomings_ = (ccv_nnc_incoming_t*)__builtin_alloca (sizeof(ccv_nnc_incoming_t ) * (exec_symbol_info_size) + sizeof(int32_t) * ((exec_symbol_info_size ) * 2 + _incoming_edges_)); memset(_incomings_, 0, sizeof(ccv_nnc_incoming_t ) * (exec_symbol_info_size)); int32_t* _exists_[2] = { (int32_t *)(_incomings_ + (exec_symbol_info_size)), (int32_t*)(_incomings_ + (exec_symbol_info_size)) + (exec_symbol_info_size), }; int32_t * const _edges_ = _exists_[1] + (exec_symbol_info_size); for ( _i_ = 0; _i_ < (source_size); _i_++) { ((void) sizeof (((sources )[_i_].graph == graph) ? 1 : 0), __extension__ ({ if ((sources )[_i_].graph == graph) ; else __assert_fail ("(sources)[_i_].graph == graph" , "ccv_nnc_symbolic_graph_backward.c", 644, __extension__ __PRETTY_FUNCTION__ ); })); _incomings_[(sources)[_i_].d].r = 1; _exists_[0][_i_] = (sources)[_i_].d; } int _exist_size_[2] = { (source_size), 0, }; int _p_ = 0, _q_ = 1; while (_exist_size_[_p_] > 0) { _exist_size_[_q_] = 0; for (_i_ = 0; _i_ < _exist_size_ [_p_]; _i_++) { const int32_t _idx_ = _exists_[_p_][_i_]; if ( _incomings_[_idx_].r != 1) continue; _incomings_[_idx_].r = 2 ; if (((ccv_nnc_graph_exec_symbol_info_t*)((void*)(((char*)(( graph->exec_symbol_info)->data)) + (size_t)(graph->exec_symbol_info )->rsize * (size_t)(0))))[_idx_].outgoings) for (_j_ = 0; _j_ < ((ccv_nnc_graph_exec_symbol_info_t*)((void*)(((char*)(( graph->exec_symbol_info)->data)) + (size_t)(graph->exec_symbol_info )->rsize * (size_t)(0))))[_idx_].outgoings->rnum; _j_++ ) { const int d = *(int*)((void*)(((char*)((((ccv_nnc_graph_exec_symbol_info_t *)((void*)(((char*)((graph->exec_symbol_info)->data)) + (size_t)(graph->exec_symbol_info)->rsize * (size_t)(0) )))[_idx_].outgoings)->data)) + (size_t)(((ccv_nnc_graph_exec_symbol_info_t *)((void*)(((char*)((graph->exec_symbol_info)->data)) + (size_t)(graph->exec_symbol_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_] < (exec_symbol_info_size )) ? 1 : 0), __extension__ ({ if (_exist_size_[_q_] < (exec_symbol_info_size )) ; else __assert_fail ("_exist_size_[_q_] < (exec_symbol_info_size)" , "ccv_nnc_symbolic_graph_backward.c", 644, __extension__ __PRETTY_FUNCTION__ ); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[ _q_]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for (_i_ = 0; _i_ < (source_size); _i_++) { ((void) sizeof (( (sources)[_i_].graph == graph) ? 1 : 0), __extension__ ({ if ( (sources)[_i_].graph == graph) ; else __assert_fail ("(sources)[_i_].graph == graph" , "ccv_nnc_symbolic_graph_backward.c", 644, __extension__ __PRETTY_FUNCTION__ ); })); _incomings_[(sources)[_i_].d].r = 3; _exists_[0][_i_] = (sources)[_i_].d; } _exist_size_[0] = (source_size); _exist_size_ [1] = 0; _p_ = 0, _q_ = 1; int _bump_ = 1; while (_exist_size_ [_p_] > 0) { _exist_size_[_q_] = 0; for (_i_ = 0; _i_ < _exist_size_[_p_]; _i_++) { const int32_t _idx_ = _exists_[_p_ ][_i_]; if (_incomings_[_idx_].r != 3) continue; _incomings_[ _idx_].r = 4; if (((ccv_nnc_graph_exec_symbol_info_t*)((void* )(((char*)((graph->exec_symbol_info)->data)) + (size_t) (graph->exec_symbol_info)->rsize * (size_t)(0))))[_idx_ ].outgoings) for (_j_ = 0; _j_ < ((ccv_nnc_graph_exec_symbol_info_t *)((void*)(((char*)((graph->exec_symbol_info)->data)) + (size_t)(graph->exec_symbol_info)->rsize * (size_t)(0) )))[_idx_].outgoings->rnum; _j_++) { const int d = *(int*) ((void*)(((char*)((((ccv_nnc_graph_exec_symbol_info_t*)((void *)(((char*)((graph->exec_symbol_info)->data)) + (size_t )(graph->exec_symbol_info)->rsize * (size_t)(0))))[_idx_ ].outgoings)->data)) + (size_t)(((ccv_nnc_graph_exec_symbol_info_t *)((void*)(((char*)((graph->exec_symbol_info)->data)) + (size_t)(graph->exec_symbol_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_] < (exec_symbol_info_size)) ? 1 : 0), __extension__ ({ if (_exist_size_[_q_] < (exec_symbol_info_size)) ; else __assert_fail ("_exist_size_[_q_] < (exec_symbol_info_size)" , "ccv_nnc_symbolic_graph_backward.c", 644, __extension__ __PRETTY_FUNCTION__ ); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[ _q_]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for (_i_ = 0; _i_ < (destination_size); _i_++) { ((void) sizeof (((destinations)[_i_].graph == graph) ? 1 : 0), __extension__ ({ if ((destinations)[_i_].graph == graph) ; else __assert_fail ("(destinations)[_i_].graph == graph", "ccv_nnc_symbolic_graph_backward.c" , 644, __extension__ __PRETTY_FUNCTION__); })); _incomings_[( destinations)[_i_].d].r = 5; _exists_[0][_i_] = (destinations )[_i_].d; } _exist_size_[0] = (destination_size); _exist_size_ [1] = 0; _p_ = 0, _q_ = 1; while (_exist_size_[_p_] > 0) { _exist_size_[_q_] = 0; for (_i_ = 0; _i_ < _exist_size_[_p_ ]; _i_++) { const int32_t _idx_ = _exists_[_p_][_i_]; if (_incomings_ [_idx_].r != 5) continue; _incomings_[_idx_].r = 6; if (_incomings_ [_idx_].edges > 0) for (_j_ = 0; _j_ < _incomings_[_idx_ ].c; _j_++) { const int d = _edges_[_incomings_[_idx_].edges - 1 + _j_]; if (_incomings_[d].r != 4) continue; _incomings_[d ].r = 5; ((void) sizeof ((_exist_size_[_q_] < (exec_symbol_info_size )) ? 1 : 0), __extension__ ({ if (_exist_size_[_q_] < (exec_symbol_info_size )) ; else __assert_fail ("_exist_size_[_q_] < (exec_symbol_info_size)" , "ccv_nnc_symbolic_graph_backward.c", 644, __extension__ __PRETTY_FUNCTION__ ); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[ _q_]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for (_i_ = 0; _i_ < (destination_size); _i_++) { ((void) sizeof (((destinations)[_i_].graph == graph) ? 1 : 0), __extension__ ({ if ((destinations)[_i_].graph == graph) ; else __assert_fail ("(destinations)[_i_].graph == graph", "ccv_nnc_symbolic_graph_backward.c" , 644, __extension__ __PRETTY_FUNCTION__); })); _incomings_[( destinations)[_i_].d].d = 1; } for (_i_ = 0; _i_ < (source_size ); _i_++) { ((void) sizeof (((sources)[_i_].graph == graph) ? 1 : 0), __extension__ ({ if ((sources)[_i_].graph == graph) ; else __assert_fail ("(sources)[_i_].graph == graph", "ccv_nnc_symbolic_graph_backward.c" , 644, __extension__ __PRETTY_FUNCTION__); })); _exists_[0][_i_ ] = (sources)[_i_].d; } _p_ = 0; _q_ = 1; _exist_size_[0] = ( source_size); _exist_size_[1] = 0; int _d_ = 0; while (_exist_size_ [_p_] > 0) { _exist_size_[_q_] = 0; for (_i_ = 0; _i_ < _exist_size_[_p_];) { const int32_t _idx_ = _exists_[_p_][_i_ ]; _visit_->node[_visit_->size].index = ((_idx_)); _visit_ ->node[_visit_->size].term = ((_incomings_[_idx_].d)); ++ _visit_->size;; if (_incomings_[_idx_].d) { ++_d_; _incomings_ [_idx_].r = 7; } if (((ccv_nnc_graph_exec_symbol_info_t*)((void *)(((char*)((graph->exec_symbol_info)->data)) + (size_t )(graph->exec_symbol_info)->rsize * (size_t)(0))))[_idx_ ].outgoings) { if (((ccv_nnc_graph_exec_symbol_info_t*)((void *)(((char*)((graph->exec_symbol_info)->data)) + (size_t )(graph->exec_symbol_info)->rsize * (size_t)(0))))[_idx_ ].outgoings->rnum == 1) { const int d = *(int*)((void*)((( char*)((((ccv_nnc_graph_exec_symbol_info_t*)((void*)(((char*) ((graph->exec_symbol_info)->data)) + (size_t)(graph-> exec_symbol_info)->rsize * (size_t)(0))))[_idx_].outgoings )->data)) + (size_t)(((ccv_nnc_graph_exec_symbol_info_t*)( (void*)(((char*)((graph->exec_symbol_info)->data)) + (size_t )(graph->exec_symbol_info)->rsize * (size_t)(0))))[_idx_ ].outgoings)->rsize * (size_t)(0))); --_incomings_[d].c; if (_incomings_[d].c == 0 && _incomings_[d].r == 6 && _d_ < (destination_size)) { _exists_[_p_][_i_] = d; continue ; } } else for (_j_ = 0; _j_ < ((ccv_nnc_graph_exec_symbol_info_t *)((void*)(((char*)((graph->exec_symbol_info)->data)) + (size_t)(graph->exec_symbol_info)->rsize * (size_t)(0) )))[_idx_].outgoings->rnum; _j_++) { const int d = *(int*) ((void*)(((char*)((((ccv_nnc_graph_exec_symbol_info_t*)((void *)(((char*)((graph->exec_symbol_info)->data)) + (size_t )(graph->exec_symbol_info)->rsize * (size_t)(0))))[_idx_ ].outgoings)->data)) + (size_t)(((ccv_nnc_graph_exec_symbol_info_t *)((void*)(((char*)((graph->exec_symbol_info)->data)) + (size_t)(graph->exec_symbol_info)->rsize * (size_t)(0) )))[_idx_].outgoings)->rsize * (size_t)(_j_))); --_incomings_ [d].c; if (_incomings_[d].c == 0 && _incomings_[d].r == 6 && _d_ < (destination_size)) { ((void) sizeof ( (_exist_size_[_q_] < (exec_symbol_info_size)) ? 1 : 0), __extension__ ({ if (_exist_size_[_q_] < (exec_symbol_info_size)) ; else __assert_fail ("_exist_size_[_q_] < (exec_symbol_info_size)" , "ccv_nnc_symbolic_graph_backward.c", 644, __extension__ __PRETTY_FUNCTION__ ); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[ _q_]; } } } ++_i_; } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = ( _i_)); } for (_i_ = 0; _i_ < (destination_size); _i_++) { ( (void) sizeof (((destinations)[_i_].graph == graph) ? 1 : 0), __extension__ ({ if ((destinations)[_i_].graph == graph) ; else __assert_fail ("(destinations)[_i_].graph == graph", "ccv_nnc_symbolic_graph_backward.c" , 644, __extension__ __PRETTY_FUNCTION__); })); if (_incomings_ [(destinations)[_i_].d].r == 7) continue; if (!(0)) { ((void) sizeof ((_incomings_[(destinations)[_i_].d].c == 0) ? 1 : 0) , __extension__ ({ if (_incomings_[(destinations)[_i_].d].c == 0) ; else __assert_fail ("_incomings_[(destinations)[_i_].d].c == 0" , "ccv_nnc_symbolic_graph_backward.c", 644, __extension__ __PRETTY_FUNCTION__ ); })); } else if (_incomings_[(destinations)[_i_].d].c > 0 ) continue; _visit_->node[_visit_->size].index = (((destinations )[_i_].d)); _visit_->node[_visit_->size].term = ((_incomings_ [(destinations)[_i_].d].d)); ++_visit_->size;; } if (_heap_mem_ ) free(_incomings_); } while (0);; ((void) sizeof ((_visit_-> size <= (exec_symbol_info_size)) ? 1 : 0), __extension__ ( { if (_visit_->size <= (exec_symbol_info_size)) ; else __assert_fail ("_visit_->size <= (exec_symbol_info_size)", "ccv_nnc_symbolic_graph_backward.c" , 644, __extension__ __PRETTY_FUNCTION__); })); _visit_; }); | |||
645 | ccv_nnc_symbolic_graph_symbol_infer(graph, forward_visit, sources, source_size, destinations, destination_size, 0, 0, tensor_symbol_info, exec_symbol_info); | |||
646 | int i; | |||
647 | // Now, for each one of these, find a reverse graph. | |||
648 | ccv_nnc_graph_backward_info_t* backward_info = (ccv_nnc_graph_backward_info_t*)cccalloccalloc(exec_symbol_info_size, sizeof(ccv_nnc_graph_backward_info_t)); | |||
649 | ccv_nnc_graph_visit_for(forward_visit, exec_symbol_info, node, idx){ int _i_; for (_i_ = 0; _i_ < (forward_visit)->size; _i_ ++) { const int idx __attribute__((unused)) = (forward_visit) ->node[_i_].index; const int _node_unused_ __attribute__(( unused)) = (forward_visit)->node[_i_].term; typeof ((exec_symbol_info )) const node __attribute__((unused)) = (exec_symbol_info) + idx ; { | |||
650 | assert(ccv_nnc_cmd_is_forward(node->cmd) || node->cmd.cmd == CCV_NNC_NOOP)((void) sizeof ((ccv_nnc_cmd_is_forward(node->cmd) || node ->cmd.cmd == CCV_NNC_NOOP) ? 1 : 0), __extension__ ({ if ( ccv_nnc_cmd_is_forward(node->cmd) || node->cmd.cmd == CCV_NNC_NOOP ) ; else __assert_fail ("ccv_nnc_cmd_is_forward(node->cmd) || node->cmd.cmd == CCV_NNC_NOOP" , "ccv_nnc_symbolic_graph_backward.c", 650, __extension__ __PRETTY_FUNCTION__ ); })); | |||
651 | if (node->outgoings) | |||
652 | for (i = 0; i < node->outgoings->rnum; i++) | |||
653 | { | |||
654 | int d = *(int*)ccv_array_get(node->outgoings, i)((void*)(((char*)((node->outgoings)->data)) + (size_t)( node->outgoings)->rsize * (size_t)(i))); | |||
655 | if (!backward_info[d].outgoings) | |||
656 | backward_info[d].outgoings = ccv_array_new(sizeof(int32_t), 1, 0); | |||
657 | ccv_array_push(backward_info[d].outgoings, &idx); | |||
658 | } | |||
659 | } ccv_nnc_graph_visit_endfor} } | |||
660 | // Also mark only the output bits that we use. | |||
661 | for (i = 0; i < exec_symbol_info_size; i++) | |||
662 | { | |||
663 | backward_info[i].input_bitmask_size = ((exec_symbol_info[i].output_size * 2 + exec_symbol_info[i].input_size + 63) >> 6); | |||
664 | backward_info[i].output_bitmask_size = ((exec_symbol_info[i].input_size + 63) >> 6); | |||
665 | // Allocate input / output bitmasks | |||
666 | if (backward_info[i].input_bitmask_size + backward_info[i].output_bitmask_size > 0) | |||
667 | { | |||
668 | backward_info[i].input_bitmasks = (uint64_t*)cccalloccalloc(backward_info[i].input_bitmask_size + backward_info[i].output_bitmask_size, sizeof(uint64_t)); | |||
669 | if (backward_info[i].output_bitmask_size) | |||
670 | backward_info[i].output_bitmasks = backward_info[i].input_bitmasks + backward_info[i].input_bitmask_size; | |||
671 | } | |||
672 | } | |||
673 | ccv_nnc_graph_visit_t* backward_visit = ccv_nnc_graph_visit_new(graph, backward_info, exec_symbol_info_size, destinations, destination_size, sources, source_size, 0)({ ccv_nnc_graph_visit_t* _visit_ = (ccv_nnc_graph_visit_t*)malloc (sizeof(ccv_nnc_graph_visit_t) + sizeof(_visit_->node[0]) * ((exec_symbol_info_size) - 1)); _visit_->size = 0; do { typedef struct { int8_t d; int8_t r; uint16_t c; int32_t edges; } ccv_nnc_incoming_t ; int _i_, _j_; int _incoming_edges_ = 0; for (_i_ = 0; _i_ < (exec_symbol_info_size); _i_++) _incoming_edges_ += ((backward_info )[_i_].outgoings) ? (backward_info)[_i_].outgoings->rnum : 0; const int _heap_mem_ = ((exec_symbol_info_size) + _incoming_edges_ > 1024); ccv_nnc_incoming_t* _incomings_; if (_heap_mem_) _incomings_ = (ccv_nnc_incoming_t*)malloc(sizeof(ccv_nnc_incoming_t ) * (exec_symbol_info_size) + sizeof(int32_t) * ((exec_symbol_info_size ) * 2 + _incoming_edges_)); else _incomings_ = (ccv_nnc_incoming_t *)__builtin_alloca (sizeof(ccv_nnc_incoming_t) * (exec_symbol_info_size ) + sizeof(int32_t) * ((exec_symbol_info_size) * 2 + _incoming_edges_ )); memset(_incomings_, 0, sizeof(ccv_nnc_incoming_t) * (exec_symbol_info_size )); int32_t* _exists_[2] = { (int32_t*)(_incomings_ + (exec_symbol_info_size )), (int32_t*)(_incomings_ + (exec_symbol_info_size)) + (exec_symbol_info_size ), }; int32_t* const _edges_ = _exists_[1] + (exec_symbol_info_size ); for (_i_ = 0; _i_ < (destination_size); _i_++) { ((void ) sizeof (((destinations)[_i_].graph == graph) ? 1 : 0), __extension__ ({ if ((destinations)[_i_].graph == graph) ; else __assert_fail ("(destinations)[_i_].graph == graph", "ccv_nnc_symbolic_graph_backward.c" , 673, __extension__ __PRETTY_FUNCTION__); })); _incomings_[( destinations)[_i_].d].r = 1; _exists_[0][_i_] = (destinations )[_i_].d; } int _exist_size_[2] = { (destination_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 ((backward_info )[_idx_].outgoings) for (_j_ = 0; _j_ < (backward_info)[_idx_ ].outgoings->rnum; _j_++) { const int d = *(int*)((void*)( ((char*)(((backward_info)[_idx_].outgoings)->data)) + (size_t )((backward_info)[_idx_].outgoings)->rsize * (size_t)(_j_) )); ++_incomings_[d].c; if (_incomings_[d].r != 0) continue; _incomings_ [d].r = 1; ((void) sizeof ((_exist_size_[_q_] < (exec_symbol_info_size )) ? 1 : 0), __extension__ ({ if (_exist_size_[_q_] < (exec_symbol_info_size )) ; else __assert_fail ("_exist_size_[_q_] < (exec_symbol_info_size)" , "ccv_nnc_symbolic_graph_backward.c", 673, __extension__ __PRETTY_FUNCTION__ ); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[ _q_]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for (_i_ = 0; _i_ < (destination_size); _i_++) { ((void) sizeof (((destinations)[_i_].graph == graph) ? 1 : 0), __extension__ ({ if ((destinations)[_i_].graph == graph) ; else __assert_fail ("(destinations)[_i_].graph == graph", "ccv_nnc_symbolic_graph_backward.c" , 673, __extension__ __PRETTY_FUNCTION__); })); _incomings_[( destinations)[_i_].d].r = 3; _exists_[0][_i_] = (destinations )[_i_].d; } _exist_size_[0] = (destination_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 ((backward_info)[_idx_].outgoings) for (_j_ = 0; _j_ < (backward_info)[_idx_].outgoings->rnum; _j_++ ) { const int d = *(int*)((void*)(((char*)(((backward_info)[_idx_ ].outgoings)->data)) + (size_t)((backward_info)[_idx_].outgoings )->rsize * (size_t)(_j_))); if (_incomings_[d].edges == 0) { _incomings_[d].edges = _bump_; _bump_ += _incomings_[d].c; _incomings_[d].c = 0; } _edges_[_incomings_[d].edges - 1 + _incomings_ [d].c] = _idx_; ++_incomings_[d].c; if (_incomings_[d].r != 2 ) continue; _incomings_[d].r = 3; ((void) sizeof ((_exist_size_ [_q_] < (exec_symbol_info_size)) ? 1 : 0), __extension__ ( { if (_exist_size_[_q_] < (exec_symbol_info_size)) ; else __assert_fail ("_exist_size_[_q_] < (exec_symbol_info_size)", "ccv_nnc_symbolic_graph_backward.c" , 673, __extension__ __PRETTY_FUNCTION__); })); _exists_[_q_] [_exist_size_[_q_]] = d; ++_exist_size_[_q_]; } } ((_i_) = (_p_ ), (_p_) = (_q_), (_q_) = (_i_)); } for (_i_ = 0; _i_ < (source_size ); _i_++) { ((void) sizeof (((sources)[_i_].graph == graph) ? 1 : 0), __extension__ ({ if ((sources)[_i_].graph == graph) ; else __assert_fail ("(sources)[_i_].graph == graph", "ccv_nnc_symbolic_graph_backward.c" , 673, __extension__ __PRETTY_FUNCTION__); })); _incomings_[( sources)[_i_].d].r = 5; _exists_[0][_i_] = (sources)[_i_].d; } _exist_size_[0] = (source_size); _exist_size_[1] = 0; _p_ = 0 , _q_ = 1; while (_exist_size_[_p_] > 0) { _exist_size_[_q_ ] = 0; for (_i_ = 0; _i_ < _exist_size_[_p_]; _i_++) { const int32_t _idx_ = _exists_[_p_][_i_]; if (_incomings_[_idx_].r != 5) continue; _incomings_[_idx_].r = 6; if (_incomings_[_idx_ ].edges > 0) for (_j_ = 0; _j_ < _incomings_[_idx_].c; _j_ ++) { const int d = _edges_[_incomings_[_idx_].edges - 1 + _j_ ]; if (_incomings_[d].r != 4) continue; _incomings_[d].r = 5; ((void) sizeof ((_exist_size_[_q_] < (exec_symbol_info_size )) ? 1 : 0), __extension__ ({ if (_exist_size_[_q_] < (exec_symbol_info_size )) ; else __assert_fail ("_exist_size_[_q_] < (exec_symbol_info_size)" , "ccv_nnc_symbolic_graph_backward.c", 673, __extension__ __PRETTY_FUNCTION__ ); })); _exists_[_q_][_exist_size_[_q_]] = d; ++_exist_size_[ _q_]; } } ((_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for (_i_ = 0; _i_ < (source_size); _i_++) { ((void) sizeof (( (sources)[_i_].graph == graph) ? 1 : 0), __extension__ ({ if ( (sources)[_i_].graph == graph) ; else __assert_fail ("(sources)[_i_].graph == graph" , "ccv_nnc_symbolic_graph_backward.c", 673, __extension__ __PRETTY_FUNCTION__ ); })); _incomings_[(sources)[_i_].d].d = 1; } for (_i_ = 0; _i_ < (destination_size); _i_++) { ((void) sizeof (((destinations )[_i_].graph == graph) ? 1 : 0), __extension__ ({ if ((destinations )[_i_].graph == graph) ; else __assert_fail ("(destinations)[_i_].graph == graph" , "ccv_nnc_symbolic_graph_backward.c", 673, __extension__ __PRETTY_FUNCTION__ ); })); _exists_[0][_i_] = (destinations)[_i_].d; } _p_ = 0; _q_ = 1; _exist_size_[0] = (destination_size); _exist_size_[1] = 0; int _d_ = 0; while (_exist_size_[_p_] > 0) { _exist_size_ [_q_] = 0; for (_i_ = 0; _i_ < _exist_size_[_p_];) { const int32_t _idx_ = _exists_[_p_][_i_]; _visit_->node[_visit_ ->size].index = ((_idx_)); _visit_->node[_visit_->size ].term = ((_incomings_[_idx_].d)); ++_visit_->size;; if (_incomings_ [_idx_].d) { ++_d_; _incomings_[_idx_].r = 7; } if ((backward_info )[_idx_].outgoings) { if ((backward_info)[_idx_].outgoings-> rnum == 1) { const int d = *(int*)((void*)(((char*)(((backward_info )[_idx_].outgoings)->data)) + (size_t)((backward_info)[_idx_ ].outgoings)->rsize * (size_t)(0))); --_incomings_[d].c; if (_incomings_[d].c == 0 && _incomings_[d].r == 6 && _d_ < (source_size)) { _exists_[_p_][_i_] = d; continue; } } else for (_j_ = 0; _j_ < (backward_info)[_idx_].outgoings ->rnum; _j_++) { const int d = *(int*)((void*)(((char*)((( backward_info)[_idx_].outgoings)->data)) + (size_t)((backward_info )[_idx_].outgoings)->rsize * (size_t)(_j_))); --_incomings_ [d].c; if (_incomings_[d].c == 0 && _incomings_[d].r == 6 && _d_ < (source_size)) { ((void) sizeof ((_exist_size_ [_q_] < (exec_symbol_info_size)) ? 1 : 0), __extension__ ( { if (_exist_size_[_q_] < (exec_symbol_info_size)) ; else __assert_fail ("_exist_size_[_q_] < (exec_symbol_info_size)", "ccv_nnc_symbolic_graph_backward.c" , 673, __extension__ __PRETTY_FUNCTION__); })); _exists_[_q_] [_exist_size_[_q_]] = d; ++_exist_size_[_q_]; } } } ++_i_; } ( (_i_) = (_p_), (_p_) = (_q_), (_q_) = (_i_)); } for (_i_ = 0; _i_ < (source_size); _i_++) { ((void) sizeof (((sources)[ _i_].graph == graph) ? 1 : 0), __extension__ ({ if ((sources) [_i_].graph == graph) ; else __assert_fail ("(sources)[_i_].graph == graph" , "ccv_nnc_symbolic_graph_backward.c", 673, __extension__ __PRETTY_FUNCTION__ ); })); if (_incomings_[(sources)[_i_].d].r == 7) continue; if (!(0)) { ((void) sizeof ((_incomings_[(sources)[_i_].d].c == 0) ? 1 : 0), __extension__ ({ if (_incomings_[(sources)[_i_] .d].c == 0) ; else __assert_fail ("_incomings_[(sources)[_i_].d].c == 0" , "ccv_nnc_symbolic_graph_backward.c", 673, __extension__ __PRETTY_FUNCTION__ ); })); } else if (_incomings_[(sources)[_i_].d].c > 0) continue ; _visit_->node[_visit_->size].index = (((sources)[_i_] .d)); _visit_->node[_visit_->size].term = ((_incomings_ [(sources)[_i_].d].d)); ++_visit_->size;; } if (_heap_mem_ ) free(_incomings_); } while (0);; ((void) sizeof ((_visit_-> size <= (exec_symbol_info_size)) ? 1 : 0), __extension__ ( { if (_visit_->size <= (exec_symbol_info_size)) ; else __assert_fail ("_visit_->size <= (exec_symbol_info_size)", "ccv_nnc_symbolic_graph_backward.c" , 673, __extension__ __PRETTY_FUNCTION__); })); _visit_; }); | |||
674 | const int sub_prep_size = graph->sub_graphs ? graph->sub_graphs->rnum : 0; | |||
675 | ccv_nnc_symbolic_graph_backward_prep_t* sub_preps = sub_prep_size > 0 ? (ccv_nnc_symbolic_graph_backward_prep_t*)cccalloccalloc(sub_prep_size, sizeof(ccv_nnc_symbolic_graph_backward_prep_t)) : 0; | |||
676 | for (i = 0; i < sub_prep_size; i++) | |||
677 | { | |||
678 | const ccv_nnc_symbolic_graph_t* const sub_graph = *(ccv_nnc_symbolic_graph_t**)ccv_array_get(graph->sub_graphs, i)((void*)(((char*)((graph->sub_graphs)->data)) + (size_t )(graph->sub_graphs)->rsize * (size_t)(i))); | |||
679 | sub_preps[i] = _ccv_nnc_symbolic_graph_backward_prep(sub_graph, ccv_nnc_symbolic_graph_sources(sub_graph), ccv_nnc_symbolic_graph_source_size(sub_graph), ccv_nnc_symbolic_graph_destinations(sub_graph), ccv_nnc_symbolic_graph_destination_size(sub_graph)); | |||
680 | } | |||
681 | return (ccv_nnc_symbolic_graph_backward_prep_t){ | |||
682 | .exec_symbol_info_size = exec_symbol_info_size, | |||
683 | .tensor_symbol_info_size = tensor_symbol_info_size, | |||
684 | .sub_prep_size = sub_prep_size, | |||
685 | .exec_symbol_info = exec_symbol_info, | |||
686 | .tensor_symbol_info = tensor_symbol_info, | |||
687 | .backward_info = backward_info, | |||
688 | .forward_visit = forward_visit, | |||
689 | .backward_visit = backward_visit, | |||
690 | .sub_preps = sub_preps, | |||
691 | .graph = (ccv_nnc_symbolic_graph_t*)graph, | |||
692 | }; | |||
693 | } | |||
694 | ||||
695 | static void _ccv_nnc_symbolic_graph_backward_exec_io(const ccv_nnc_graph_exec_symbol_info_t* const node, int** const back_input_map, int** const back_output_map, int* const back_input_size, int* const back_output_size) | |||
696 | { | |||
697 | int i; | |||
698 | if (node->flags & CCV_NNC_GRAPH_EXEC_CASE_OF) | |||
699 | { | |||
700 | *back_input_map = node->outputs; | |||
701 | *back_input_size = node->output_size; | |||
702 | for (i = 0; i < node->case_of.argument.offset; i++) | |||
703 | (*back_output_map)[i] = node->inputs[i]; | |||
704 | const int argument_offset = node->case_of.argument.offset; | |||
705 | const int argument_size = node->case_of.argument.size; | |||
706 | // Skip the argument range. | |||
707 | for (i = argument_offset + argument_size; i < node->input_size; i++) | |||
708 | (*back_output_map)[i - argument_size] = node->inputs[i]; | |||
709 | *back_output_size = node->input_size - node->case_of.argument.size; | |||
710 | } else { // if (node->flags & CCV_NNC_GRAPH_EXEC_P_WHILE) { | |||
711 | *back_input_map = node->outputs; | |||
712 | *back_input_size = node->output_size; | |||
713 | *back_output_map = node->inputs; | |||
714 | *back_output_size = node->input_size; | |||
715 | } | |||
716 | } | |||
717 | ||||
718 | static void _ccv_nnc_symbolic_graph_backward_prep_sub_f_wrt_symbols(const ccv_nnc_graph_exec_symbol_info_t* const forw_exec, const ccv_nnc_symbolic_graph_t* const sub_graph, const int graph_ref, const ccv_nnc_tensor_symbol_info_t* const tensor_symbol_info, const uint64_t* const input_bitmasks, const uint64_t* const output_bitmasks, ccv_array_t* const sub_f_symbols, ccv_array_t* const sub_wrt_symbols) | |||
719 | { | |||
720 | int i, j; | |||
721 | ccv_array_clear(sub_wrt_symbols); | |||
722 | int forw_outputs[ccv_max(1, forw_exec->output_size)({ typeof (1) _a = (1); typeof (forw_exec->output_size) _b = (forw_exec->output_size); (_a > _b) ? _a : _b; })]; | |||
723 | int forw_inputs[ccv_max(1, forw_exec->input_size)({ typeof (1) _a = (1); typeof (forw_exec->input_size) _b = (forw_exec->input_size); (_a > _b) ? _a : _b; })]; | |||
724 | int* back_input_map = forw_outputs; | |||
725 | int* back_output_map = forw_inputs; | |||
726 | int back_input_size, back_output_size; | |||
727 | _ccv_nnc_symbolic_graph_backward_exec_io(forw_exec, &back_input_map, &back_output_map, &back_input_size, &back_output_size); | |||
728 | for (i = 0; i < back_output_size; i++) | |||
729 | if (output_bitmasks[i >> 6] & ((uint64_t)1 << (i & 63))) | |||
730 | { | |||
731 | const int d = back_output_map[i]; | |||
| ||||
732 | const ccv_array_t* const s_refs = tensor_symbol_info[d].s_ref; | |||
733 | const int s_ref = s_refs && s_refs->rnum > graph_ref ? *(int*)ccv_array_get(s_refs, graph_ref)((void*)(((char*)((s_refs)->data)) + (size_t)(s_refs)-> rsize * (size_t)(graph_ref))) - 1 : -1; | |||
734 | if (s_ref >= 0) | |||
735 | { | |||
736 | ccv_nnc_tensor_symbol_t sub_wrt_symbol = { | |||
737 | .d = s_ref, | |||
738 | .graph = sub_graph, | |||
739 | }; | |||
740 | ccv_array_push(sub_wrt_symbols, &sub_wrt_symbol); | |||
741 | } else | |||
742 | ccv_array_push(sub_wrt_symbols, &NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL }); | |||
743 | } | |||
744 | ccv_array_clear(sub_f_symbols); | |||
745 | for (i = 0; i < back_input_size; i++) | |||
746 | if (input_bitmasks[i >> 6] & ((uint64_t)1 << (i & 63))) | |||
747 | { | |||
748 | const int d = back_input_map[i]; | |||
749 | ccv_nnc_tensor_symbol_t sub_f_symbol = { | |||
750 | .d = *(int*)ccv_array_get(tensor_symbol_info[d].s_ref, graph_ref)((void*)(((char*)((tensor_symbol_info[d].s_ref)->data)) + ( size_t)(tensor_symbol_info[d].s_ref)->rsize * (size_t)(graph_ref ))) - 1, | |||
751 | .graph = sub_graph, | |||
752 | }; | |||
753 | ccv_array_push(sub_f_symbols, &sub_f_symbol); | |||
754 | } | |||
755 | // Go through all its assignments (parameterized loop), making them either wrt or f. | |||
756 | // The reason is these must flow through the graph, otherwise we cannot form a full | |||
757 | // enclosed loop. Also because they are the additional f / wrt symbols, there is | |||
758 | // no case that we cannot find their corresponding gradients in the backward sub graphs | |||
759 | // (these gradients have to be parameterized to form an enclosed loop as well). | |||
760 | for (i = 0; i < sub_graph->tensor_symbol_info->rnum; i++) | |||
761 | { | |||
762 | const ccv_nnc_tensor_symbol_info_t* const tensor_symbol_info = (ccv_nnc_tensor_symbol_info_t*)ccv_array_get(sub_graph->tensor_symbol_info, i)((void*)(((char*)((sub_graph->tensor_symbol_info)->data )) + (size_t)(sub_graph->tensor_symbol_info)->rsize * ( size_t)(i))); | |||
763 | if (tensor_symbol_info->assign_ref) | |||
764 | { | |||
765 | const int assign_ref = tensor_symbol_info->assign_ref - 1; | |||
766 | // i is the wrt, assign_ref is the f. | |||
767 | int flag = 0; | |||
768 | for (j = 0; !flag && j < sub_wrt_symbols->rnum; j++) | |||
769 | flag = (((ccv_nnc_tensor_symbol_t*)ccv_array_get(sub_wrt_symbols, j)((void*)(((char*)((sub_wrt_symbols)->data)) + (size_t)(sub_wrt_symbols )->rsize * (size_t)(j))))->d == i); | |||
770 | if (!flag) | |||
771 | { | |||
772 | ccv_nnc_tensor_symbol_t sub_wrt_symbol = { | |||
773 | .d = i, | |||
774 | .graph = sub_graph, | |||
775 | }; | |||
776 | ccv_array_push(sub_wrt_symbols, &sub_wrt_symbol); | |||
777 | } | |||
778 | flag = 0; | |||
779 | for (j = 0; !flag && j < sub_f_symbols->rnum; j++) | |||
780 | flag = (((ccv_nnc_tensor_symbol_t*)ccv_array_get(sub_f_symbols, j)((void*)(((char*)((sub_f_symbols)->data)) + (size_t)(sub_f_symbols )->rsize * (size_t)(j))))->d == assign_ref); | |||
781 | if (!flag) | |||
782 | { | |||
783 | ccv_nnc_tensor_symbol_t sub_f_symbol = { | |||
784 | .d = assign_ref, | |||
785 | .graph = sub_graph, | |||
786 | }; | |||
787 | ccv_array_push(sub_f_symbols, &sub_f_symbol); | |||
788 | } | |||
789 | } | |||
790 | } | |||
791 | } | |||
792 | ||||
793 | // Check whether for a given f_symbol, we can compute wrt_symbols at all, if we can, tag the minimal io and ops (some ops can be replaced with noop) required to do so. | |||
794 | static int _ccv_nnc_symbolic_graph_backward_prep_prune_ops(const ccv_nnc_symbolic_graph_backward_prep_t* const backward_prep, const ccv_nnc_tensor_symbol_t* const f_symbols, const int f_symbol_size, const ccv_nnc_tensor_symbol_t* const wrt_symbols, const int wrt_symbol_size, const ccv_nnc_graph_exec_symbol_t* const sources, const int source_size, const ccv_nnc_graph_exec_symbol_t* const destinations, const int destination_size) | |||
795 | { | |||
796 | int i, j, p; | |||
797 | const int tensor_symbol_info_size = backward_prep->tensor_symbol_info_size; | |||
798 | const ccv_nnc_graph_exec_symbol_info_t* const exec_symbol_info = backward_prep->exec_symbol_info; | |||
799 | const ccv_nnc_tensor_symbol_info_t* const tensor_symbol_info =backward_prep->tensor_symbol_info; | |||
800 | const ccv_nnc_graph_visit_t* const forward_visit = backward_prep->forward_visit; | |||
801 | // Now, for each one of these, find a reverse graph. | |||
802 | ccv_nnc_graph_backward_info_t* const backward_info = backward_prep->backward_info; | |||
803 | const ccv_nnc_graph_visit_t* const backward_visit = backward_prep->backward_visit; | |||
804 | // Find the f_symbols, and tag its flows. | |||
805 | ccv_nnc_graph_visit_for(backward_visit, backward_info, node, idx){ int _i_; for (_i_ = 0; _i_ < (backward_visit)->size; _i_ ++) { const int idx __attribute__((unused)) = (backward_visit )->node[_i_].index; const int _node_unused_ __attribute__( (unused)) = (backward_visit)->node[_i_].term; typeof ((backward_info )) const node __attribute__((unused)) = (backward_info) + idx ; { | |||
806 | int f = node->f_wrt & 0x1; | |||
807 | for (i = 0; i < exec_symbol_info[idx].output_size && !f; i++) | |||
808 | { | |||
809 | int d = exec_symbol_info[idx].outputs[i]; | |||
810 | if (d < 0) | |||
811 | continue; | |||
812 | while (tensor_symbol_info[d].alias_ref) | |||
813 | d = tensor_symbol_info[d].alias_ref - 1; | |||
814 | for (j = 0; j < f_symbol_size && !f; j++) | |||
815 | if (d == f_symbols[j].d) | |||
816 | f = 1; | |||
817 | } | |||
818 | if (f) | |||
819 | { | |||
820 | node->f_wrt |= f; | |||
821 | if (node->outgoings) | |||
822 | for (i = 0; i < node->outgoings->rnum; i++) | |||
823 | { | |||
824 | int d = *(int*)ccv_array_get(node->outgoings, i)((void*)(((char*)((node->outgoings)->data)) + (size_t)( node->outgoings)->rsize * (size_t)(i))); | |||
825 | backward_info[d].f_wrt |= f; | |||
826 | } | |||
827 | } | |||
828 | } ccv_nnc_graph_visit_endfor} } | |||
829 | // Find the wrt_symbols, and tag its flows. | |||
830 | ccv_nnc_graph_visit_for(forward_visit, exec_symbol_info, node, idx){ int _i_; for (_i_ = 0; _i_ < (forward_visit)->size; _i_ ++) { const int idx __attribute__((unused)) = (forward_visit) ->node[_i_].index; const int _node_unused_ __attribute__(( unused)) = (forward_visit)->node[_i_].term; typeof ((exec_symbol_info )) const node __attribute__((unused)) = (exec_symbol_info) + idx ; { | |||
831 | int wrt = backward_info[idx].f_wrt & 0x2; | |||
832 | for (i = 0; i < node->input_size && !wrt; i++) | |||
833 | { | |||
834 | int d = node->inputs[i]; | |||
835 | if (d < 0) | |||
836 | continue; | |||
837 | while (tensor_symbol_info[d].alias_ref) | |||
838 | d = tensor_symbol_info[d].alias_ref - 1; | |||
839 | for (j = 0; j < wrt_symbol_size && !wrt; j++) | |||
840 | { | |||
841 | int wrt_d = wrt_symbols[j].d; | |||
842 | if (wrt_d < 0) | |||
843 | continue; | |||
844 | // Find the root of this tensor alias. | |||
845 | if (tensor_symbol_info[wrt_d].alias_ref) | |||
846 | wrt_d = tensor_symbol_info[wrt_d].alias_ref - 1; | |||
847 | if (d == wrt_d) | |||
848 | wrt = 0x2; | |||
849 | } | |||
850 | } | |||
851 | if (wrt) | |||
852 | { | |||
853 | backward_info[idx].f_wrt |= wrt; | |||
854 | if (node->outgoings) | |||
855 | for (i = 0; i < node->outgoings->rnum; i++) | |||
856 | { | |||
857 | int d = *(int*)ccv_array_get(node->outgoings, i)((void*)(((char*)((node->outgoings)->data)) + (size_t)( node->outgoings)->rsize * (size_t)(i))); | |||
858 | backward_info[d].f_wrt |= wrt; | |||
859 | } | |||
860 | } | |||
861 | } ccv_nnc_graph_visit_endfor} } | |||
862 | enum { | |||
863 | WRT_SYMBOL_USE = 1, | |||
864 | F_SYMBOL_USE = 2 | |||
865 | }; | |||
866 | uint8_t* used_grad = (uint8_t*)cccalloccalloc(tensor_symbol_info_size, sizeof(uint8_t)); | |||
867 | // First, all f_symbols and wrt_symbols are used. | |||
868 | for (i = 0; i < f_symbol_size; i++) | |||
869 | if (f_symbols[i].d >= 0) | |||
870 | used_grad[tensor_symbol_info[f_symbols[i].d].alias_ref ? tensor_symbol_info[f_symbols[i].d].alias_ref - 1 : f_symbols[i].d] |= F_SYMBOL_USE; | |||
871 | for (i = 0; i < wrt_symbol_size; i++) | |||
872 | if (wrt_symbols[i].d >= 0) | |||
873 | used_grad[tensor_symbol_info[wrt_symbols[i].d].alias_ref ? tensor_symbol_info[wrt_symbols[i].d].alias_ref - 1 : wrt_symbols[i].d] |= WRT_SYMBOL_USE; | |||
874 | // Do optimistic assumption, and then compute used_grad | |||
875 | ccv_nnc_graph_visit_for(forward_visit, exec_symbol_info, _, idx){ int _i_; for (_i_ = 0; _i_ < (forward_visit)->size; _i_ ++) { const int idx __attribute__((unused)) = (forward_visit) ->node[_i_].index; const int _node_unused_ __attribute__(( unused)) = (forward_visit)->node[_i_].term; typeof ((exec_symbol_info )) const _ __attribute__((unused)) = (exec_symbol_info) + idx ; { | |||
876 | ccv_nnc_graph_backward_info_t* node = backward_info + idx; | |||
877 | /* Only interested in the ones on the f / wrt flow */ | |||
878 | if ((node->f_wrt & 0x3) == 0x3) | |||
879 | { | |||
880 | const ccv_nnc_graph_exec_symbol_info_t* forw_exec = exec_symbol_info + idx; | |||
881 | ccv_nnc_cmd_t cmd = forw_exec->cmd; | |||
882 | if (cmd.cmd != CCV_NNC_NOOP) | |||
883 | cmd.cmd += 1; /* Backward command is the one after forward command. */ | |||
884 | assert(ccv_nnc_cmd_is_backward(cmd) || cmd.cmd == CCV_NNC_NOOP)((void) sizeof ((ccv_nnc_cmd_is_backward(cmd) || cmd.cmd == CCV_NNC_NOOP ) ? 1 : 0), __extension__ ({ if (ccv_nnc_cmd_is_backward(cmd) || cmd.cmd == CCV_NNC_NOOP) ; else __assert_fail ("ccv_nnc_cmd_is_backward(cmd) || cmd.cmd == CCV_NNC_NOOP" , "ccv_nnc_symbolic_graph_backward.c", 884, __extension__ __PRETTY_FUNCTION__ ); })); | |||
885 | for (i = 0; i < forw_exec->output_size * 2 + forw_exec->input_size; i++) | |||
886 | if (!(i >= forw_exec->output_size && i < forw_exec->output_size + forw_exec->input_size && | |||
887 | forw_exec->inputs[i - forw_exec->output_size] < 0) && // If the input is empty, no need. | |||
888 | !(i >= forw_exec->output_size + forw_exec->input_size && i < forw_exec->output_size * 2 + forw_exec->input_size && | |||
889 | forw_exec->outputs[i - forw_exec->output_size - forw_exec->input_size] < 0) && // If the output is empty, no need. | |||
890 | !(i < forw_exec->output_size && forw_exec->outputs[i] < 0)) // If the output is empty for gradient, no need. | |||
891 | node->input_bitmasks[i >> 6] |= ((uint64_t)1 << (i & 63)); | |||
892 | for (i = 0; i < forw_exec->input_size; i++) | |||
893 | if (!(forw_exec->inputs[i] < 0)) // If the inputs is empty, no need. | |||
894 | node->output_bitmasks[i >> 6] |= ((uint64_t)1 << (i & 63)); | |||
895 | int maybe_noop = 1; | |||
896 | for (i = 0; i < forw_exec->input_size; i++) | |||
897 | /* See if it is used as wrt, if not, no need to run this node at all. */ | |||
898 | if (forw_exec->inputs[i] >= 0 && used_grad[tensor_symbol_info[forw_exec->inputs[i]].alias_ref ? tensor_symbol_info[forw_exec->inputs[i]].alias_ref - 1 : forw_exec->inputs[i]] & WRT_SYMBOL_USE) | |||
899 | { | |||
900 | maybe_noop = 0; | |||
901 | break; | |||
902 | } | |||
903 | if (maybe_noop) | |||
904 | { | |||
905 | for (i = 0; i < node->input_bitmask_size; i++) | |||
906 | node->input_bitmasks[i] = 0; | |||
907 | for (i = 0; i < node->output_bitmask_size; i++) | |||
908 | node->output_bitmasks[i] = 0; | |||
909 | node->output_bitmask_size = 0; | |||
910 | } else if (cmd.cmd == CCV_NNC_GRAPH_FORWARD || cmd.cmd == CCV_NNC_GRAPH_BACKWARD) { | |||
911 | // Clear out all potential outputs if we think it is not a wrt symbols. | |||
912 | for (i = 0; i < forw_exec->input_size; i++) | |||
913 | if ((node->output_bitmasks[i >> 6] & ((uint64_t)1 << (i & 63))) && | |||
914 | !(used_grad[tensor_symbol_info[forw_exec->inputs[i]].alias_ref ? tensor_symbol_info[forw_exec->inputs[i]].alias_ref - 1 : forw_exec->inputs[i]] & WRT_SYMBOL_USE)) | |||
915 | node->output_bitmasks[i >> 6] &= ~((uint64_t)1 << (i & 63)); | |||
916 | // But for now, assuming we need all input gradients. | |||
917 | // Clear out all inputs / outputs from forward op. | |||
918 | for (i = forw_exec->output_size; i < forw_exec->output_size * 2 + forw_exec->input_size; i++) | |||
919 | node->input_bitmasks[i >> 6] &= ~((uint64_t)1 << (i & 63)); | |||
920 | } else if (ccv_nnc_cmd_bitmask(cmd, forw_exec->output_size * 2 + forw_exec->input_size, forw_exec->input_size, node->input_bitmasks, node->input_bitmask_size, node->output_bitmasks, node->output_bitmask_size)) { | |||
921 | int flag; /* Only continue if it changed */ | |||
922 | do { | |||
923 | flag = 0; | |||
924 | /* Check if the output first */ | |||
925 | for (i = 0; i < forw_exec->input_size; i++) | |||
926 | /* Only try to eliminate the one that is not used. */ | |||
927 | if ((node->output_bitmasks[i >> 6] & ((uint64_t)1 << (i & 63))) && | |||
928 | !(used_grad[tensor_symbol_info[forw_exec->inputs[i]].alias_ref ? tensor_symbol_info[forw_exec->inputs[i]].alias_ref - 1 : forw_exec->inputs[i]] & WRT_SYMBOL_USE)) | |||
929 | { | |||
930 | node->output_bitmasks[i >> 6] &= ~((uint64_t)1 << (i & 63)); | |||
931 | /* If it worked, mark it as flagged. */ | |||
932 | if (ccv_nnc_cmd_bitmask(cmd, forw_exec->output_size * 2 + forw_exec->input_size, forw_exec->input_size, node->input_bitmasks, node->input_bitmask_size, node->output_bitmasks, node->output_bitmask_size)) | |||
933 | flag = 1; | |||
934 | else /* Refit this with the bit back again. */ | |||
935 | node->output_bitmasks[i >> 6] |= ((uint64_t)1 << (i & 63)); | |||
936 | } | |||
937 | for (i = 0; i < forw_exec->output_size * 2 + forw_exec->input_size; i++) | |||
938 | if ((node->input_bitmasks[i >> 6] & ((uint64_t)1 << (i & 63))) && | |||
939 | (i >= forw_exec->output_size || | |||
940 | !(used_grad[tensor_symbol_info[forw_exec->outputs[i]].alias_ref ? tensor_symbol_info[forw_exec->outputs[i]].alias_ref - 1 : forw_exec->outputs[i]] & F_SYMBOL_USE))) | |||
941 | { /* Try to eliminate one of the input. */ | |||
942 | node->input_bitmasks[i >> 6] &= ~((uint64_t)1 << (i & 63)); | |||
943 | /* If it worked, mark it as flagged. */ | |||
944 | if (ccv_nnc_cmd_bitmask(cmd, forw_exec->output_size * 2 + forw_exec->input_size, forw_exec->input_size, node->input_bitmasks, node->input_bitmask_size, node->output_bitmasks, node->output_bitmask_size)) | |||
945 | flag = 1; | |||
946 | else /* Refit this with the bit back again. */ | |||
947 | node->input_bitmasks[i >> 6] |= ((uint64_t)1 << (i & 63)); | |||
948 | } | |||
949 | } while (flag); | |||
950 | } | |||
951 | for (i = 0; i < forw_exec->output_size; i++) | |||
952 | if (node->input_bitmasks[i >> 6] & ((uint64_t)1 << (i & 63))) | |||
953 | /* Mark it is used as wrt. */ | |||
954 | used_grad[tensor_symbol_info[forw_exec->outputs[i]].alias_ref ? tensor_symbol_info[forw_exec->outputs[i]].alias_ref - 1 : forw_exec->outputs[i]] |= WRT_SYMBOL_USE; | |||
955 | for (i = 0; i < forw_exec->input_size; i++) | |||
956 | /* Mark it is used as f. */ | |||
957 | if (node->output_bitmasks[i >> 6] & ((uint64_t)1 << (i & 63))) | |||
958 | used_grad[tensor_symbol_info[forw_exec->inputs[i]].alias_ref ? tensor_symbol_info[forw_exec->inputs[i]].alias_ref - 1 : forw_exec->inputs[i]] |= F_SYMBOL_USE; | |||
959 | } | |||
960 | } ccv_nnc_graph_visit_endfor} } | |||
961 | ccv_array_t* sub_f_symbols = 0; | |||
962 | ccv_array_t* sub_wrt_symbols = 0; | |||
963 | ccv_nnc_graph_visit_for(forward_visit, exec_symbol_info, _, idx){ int _i_; for (_i_ = 0; _i_ < (forward_visit)->size; _i_ ++) { const int idx __attribute__((unused)) = (forward_visit) ->node[_i_].index; const int _node_unused_ __attribute__(( unused)) = (forward_visit)->node[_i_].term; typeof ((exec_symbol_info )) const _ __attribute__((unused)) = (exec_symbol_info) + idx ; { | |||
964 | ccv_nnc_graph_backward_info_t* node = backward_info + idx; | |||
965 | const ccv_nnc_graph_exec_symbol_info_t* forw_exec = exec_symbol_info + idx; | |||
966 | /* Only interested in the ones on the f / wrt flow */ | |||
967 | if ((node->f_wrt & 0x3) == 0x3 && forw_exec->graph_ref_size > 0) | |||
968 | { | |||
969 | uint64_t stack_input_bitmasks1[node->input_bitmask_size]; | |||
970 | uint64_t stack_input_bitmasks2[node->input_bitmask_size]; | |||
971 | uint64_t* const input_bitmasks = forw_exec->graph_ref_size > 1 ? stack_input_bitmasks1 : node->input_bitmasks; | |||
972 | // We collect input masks into this location. | |||
973 | if (forw_exec->graph_ref_size > 1) | |||
974 | memset(stack_input_bitmasks2, 0, sizeof(uint64_t) * node->input_bitmask_size); | |||
975 | for (p = 0; p < forw_exec->graph_ref_size; p++) | |||
976 | { | |||
977 | // Reset the stack input bitmasks. | |||
978 | if (forw_exec->graph_ref_size > 1) | |||
979 | memcpy(stack_input_bitmasks1, node->input_bitmasks, sizeof(uint64_t) * node->input_bitmask_size); | |||
980 | // Now calling it recursively until we are sure no f_symbols can be removed. | |||
981 | const int graph_ref = CCV_NNC_GRAPH_REF(forw_exec)((forw_exec)->_heap_graph_ref ? (forw_exec)->_heap_graph_ref : (forw_exec)->_inline_graph_ref)[p] - 1; | |||
982 | ccv_nnc_symbolic_graph_backward_prep_t* const sub_prep = backward_prep->sub_preps + graph_ref; | |||
983 | if (!sub_wrt_symbols) | |||
984 | sub_wrt_symbols = ccv_array_new(sizeof(ccv_nnc_tensor_symbol_t), 0, 0); | |||
985 | else | |||
986 | ccv_array_clear(sub_wrt_symbols); | |||
987 | for (i = 0; i < forw_exec->input_size; i++) | |||
988 | if (node->output_bitmasks[i >> 6] & ((uint64_t)1 << (i & 63))) | |||
989 | { | |||
990 | const ccv_array_t* const s_refs = tensor_symbol_info[forw_exec->inputs[i]].s_ref; | |||
991 | const int s_ref = s_refs && s_refs->rnum > graph_ref ? *(int*)ccv_array_get(s_refs, graph_ref)((void*)(((char*)((s_refs)->data)) + (size_t)(s_refs)-> rsize * (size_t)(graph_ref))) - 1 : -1; | |||
992 | if (s_ref >= 0) | |||
993 | { | |||
994 | ccv_nnc_tensor_symbol_t sub_wrt_symbol = { | |||
995 | .d = s_ref, | |||
996 | .graph = sub_prep->graph, | |||
997 | }; | |||
998 | ccv_array_push(sub_wrt_symbols, &sub_wrt_symbol); | |||
999 | } | |||
1000 | } | |||
1001 | int flag; // Only continue if it changed */ | |||
1002 | do { | |||
1003 | flag = 0; | |||
1004 | for (i = 0; i < forw_exec->output_size; i++) | |||
1005 | // Try to reduce number of inputs for the backward graph. If it is not tagged as F_SYMBOL_USE, we can reduce it. | |||
1006 | // It is reducible because this sub graph may have multiple computation paths, therefore, some of these may not | |||
1007 | // involve our wrt symbols at all. | |||
1008 | if (!(used_grad[tensor_symbol_info[forw_exec->outputs[i]].alias_ref ? tensor_symbol_info[forw_exec->outputs[i]].alias_ref - 1 : forw_exec->outputs[i]] & F_SYMBOL_USE) && | |||
1009 | input_bitmasks[i >> 6] & ((uint64_t)1 << (i & 63))) | |||
1010 | { /* Try to eliminate one of the input. */ | |||
1011 | input_bitmasks[i >> 6] &= ~((uint64_t)1 << (i & 63)); | |||
1012 | if (!sub_f_symbols) | |||
1013 | sub_f_symbols = ccv_array_new(sizeof(ccv_nnc_tensor_symbol_t), 0, 0); | |||
1014 | else | |||
1015 | ccv_array_clear(sub_f_symbols); | |||
1016 | for (j = 0; j < forw_exec->output_size; j++) | |||
1017 | if (node->input_bitmasks[j >> 6] & ((uint64_t)1 << (j & 63))) | |||
1018 | { | |||
1019 | const int s_ref = *(int*)ccv_array_get(tensor_symbol_info[forw_exec->outputs[j]].s_ref, graph_ref)((void*)(((char*)((tensor_symbol_info[forw_exec->outputs[j ]].s_ref)->data)) + (size_t)(tensor_symbol_info[forw_exec-> outputs[j]].s_ref)->rsize * (size_t)(graph_ref))) - 1; | |||
1020 | assert(s_ref >= 0)((void) sizeof ((s_ref >= 0) ? 1 : 0), __extension__ ({ if (s_ref >= 0) ; else __assert_fail ("s_ref >= 0", "ccv_nnc_symbolic_graph_backward.c" , 1020, __extension__ __PRETTY_FUNCTION__); })); | |||
1021 | ccv_nnc_tensor_symbol_t sub_f_symbol = { | |||
1022 | .d = s_ref, | |||
1023 | .graph = sub_prep->graph, | |||
1024 | }; | |||
1025 | ccv_array_push(sub_f_symbols, &sub_f_symbol); | |||
1026 | } | |||
1027 | if (_ccv_nnc_symbolic_graph_backward_prep_prune_ops(sub_prep, (ccv_nnc_tensor_symbol_t*)ccv_array_get(sub_f_symbols, 0)((void*)(((char*)((sub_f_symbols)->data)) + (size_t)(sub_f_symbols )->rsize * (size_t)(0))), sub_f_symbols->rnum, (ccv_nnc_tensor_symbol_t*)ccv_array_get(sub_wrt_symbols, 0)((void*)(((char*)((sub_wrt_symbols)->data)) + (size_t)(sub_wrt_symbols )->rsize * (size_t)(0))), sub_wrt_symbols->rnum, ccv_nnc_symbolic_graph_sources(sub_prep->graph), ccv_nnc_symbolic_graph_source_size(sub_prep->graph), ccv_nnc_symbolic_graph_destinations(sub_prep->graph), ccv_nnc_symbolic_graph_destination_size(sub_prep->graph))) | |||
1028 | flag = 1; | |||
1029 | else /* Refit this with the bit back again. */ | |||
1030 | input_bitmasks[i >> 6] |= ((uint64_t)1 << (i & 63)); | |||
1031 | } | |||
1032 | } while (flag); | |||
1033 | // I am done, need to redo above for sub_prep, and it has to be successful now. | |||
1034 | if (!sub_f_symbols) | |||
1035 | sub_f_symbols = ccv_array_new(sizeof(ccv_nnc_tensor_symbol_t), 0, 0); | |||
1036 | else | |||
1037 | ccv_array_clear(sub_f_symbols); | |||
1038 | for (i = 0; i < forw_exec->output_size; i++) | |||
1039 | if (input_bitmasks[i >> 6] & ((uint64_t)1 << (i & 63))) | |||
1040 | { | |||
1041 | const int s_ref = *(int*)ccv_array_get(tensor_symbol_info[forw_exec->outputs[i]].s_ref, graph_ref)((void*)(((char*)((tensor_symbol_info[forw_exec->outputs[i ]].s_ref)->data)) + (size_t)(tensor_symbol_info[forw_exec-> outputs[i]].s_ref)->rsize * (size_t)(graph_ref))) - 1; | |||
1042 | assert(s_ref >= 0)((void) sizeof ((s_ref >= 0) ? 1 : 0), __extension__ ({ if (s_ref >= 0) ; else __assert_fail ("s_ref >= 0", "ccv_nnc_symbolic_graph_backward.c" , 1042, __extension__ __PRETTY_FUNCTION__); })); | |||
1043 | ccv_nnc_tensor_symbol_t sub_f_symbol = { | |||
1044 | .d = s_ref, | |||
1045 | .graph = sub_prep->graph, | |||
1046 | }; | |||
1047 | ccv_array_push(sub_f_symbols, &sub_f_symbol); | |||
1048 | } | |||
1049 | _ccv_nnc_symbolic_graph_backward_prep_prune_ops(sub_prep, (ccv_nnc_tensor_symbol_t*)ccv_array_get(sub_f_symbols, 0)((void*)(((char*)((sub_f_symbols)->data)) + (size_t)(sub_f_symbols )->rsize * (size_t)(0))), sub_f_symbols->rnum, (ccv_nnc_tensor_symbol_t*)ccv_array_get(sub_wrt_symbols, 0)((void*)(((char*)((sub_wrt_symbols)->data)) + (size_t)(sub_wrt_symbols )->rsize * (size_t)(0))), sub_wrt_symbols->rnum, ccv_nnc_symbolic_graph_sources(sub_prep->graph), ccv_nnc_symbolic_graph_source_size(sub_prep->graph), ccv_nnc_symbolic_graph_destinations(sub_prep->graph), ccv_nnc_symbolic_graph_destination_size(sub_prep->graph)); | |||
1050 | if (forw_exec->graph_ref_size > 1) | |||
1051 | for (i = 0; i < node->input_bitmask_size; i++) | |||
1052 | stack_input_bitmasks2[i] |= input_bitmasks[i]; | |||
1053 | } | |||
1054 | if (forw_exec->graph_ref_size > 1) | |||
1055 | memcpy(node->input_bitmasks, stack_input_bitmasks2, sizeof(uint64_t) * node->input_bitmask_size); | |||
1056 | } | |||
1057 | } ccv_nnc_graph_visit_endfor} } | |||
1058 | if (sub_f_symbols) | |||
1059 | ccv_array_free(sub_f_symbols); | |||
1060 | if (sub_wrt_symbols) | |||
1061 | ccv_array_free(sub_wrt_symbols); | |||
1062 | int flag = 1; | |||
1063 | for (i = 0; i < f_symbol_size && flag; i++) | |||
1064 | flag = (used_grad[tensor_symbol_info[f_symbols[i].d].alias_ref ? tensor_symbol_info[f_symbols[i].d].alias_ref - 1 : f_symbols[i].d] & WRT_SYMBOL_USE); | |||
1065 | ccfreefree(used_grad); | |||
1066 | return flag; | |||
1067 | } | |||
1068 | ||||
1069 | static void _ccv_nnc_symbolic_graph_backward_prep_gen(ccv_nnc_symbolic_graph_backward_prep_t* const backward_prep, const ccv_nnc_tensor_symbol_t* const f_symbols, const int f_symbol_size, const ccv_nnc_tensor_symbol_t* const wrt_symbols, const int wrt_symbol_size, const int is_while, const ccv_nnc_graph_exec_symbol_t* const sources, const int source_size, const ccv_nnc_graph_exec_symbol_t* const destinations, const int destination_size) | |||
1070 | { | |||
1071 | const int exec_symbol_info_size = backward_prep->exec_symbol_info_size; | |||
1072 | const int tensor_symbol_info_size = backward_prep->tensor_symbol_info_size; | |||
1073 | const ccv_nnc_graph_exec_symbol_info_t* const exec_symbol_info = backward_prep->exec_symbol_info; | |||
1074 | const ccv_nnc_tensor_symbol_info_t* const tensor_symbol_info =backward_prep->tensor_symbol_info; | |||
1075 | const ccv_nnc_graph_visit_t* const forward_visit = backward_prep->forward_visit; | |||
1076 | // Now, for each one of these, find a reverse graph. | |||
1077 | ccv_nnc_graph_backward_info_t* const backward_info = backward_prep->backward_info; | |||
1078 | const ccv_nnc_graph_visit_t* const backward_visit = backward_prep->backward_visit; | |||
1079 | int i, j; | |||
1080 | // Now, only the flow from f_symbols back to wrt_symbols are interested to us. | |||
1081 | // Visit the graph in reverse order, build the AD nodes. | |||
1082 | ccv_nnc_autograd_graph_exec_symbol_t* const autograd_execs = (ccv_nnc_autograd_graph_exec_symbol_t*)cccalloccalloc(exec_symbol_info_size, sizeof(ccv_nnc_autograd_graph_exec_symbol_t)); | |||
1083 | int max_forw_input_size = 0, max_forw_output_size = 0; | |||
1084 | for (i = 0; i < exec_symbol_info_size; i++) | |||
1085 | if ((backward_info[i].f_wrt & 0x3) == 0x3) | |||
1086 | { | |||
1087 | max_forw_input_size = ccv_max(max_forw_input_size, exec_symbol_info[i].input_size)({ typeof (max_forw_input_size) _a = (max_forw_input_size); typeof (exec_symbol_info[i].input_size) _b = (exec_symbol_info[i].input_size ); (_a > _b) ? _a : _b; }); | |||
1088 | max_forw_output_size = ccv_max(max_forw_output_size, exec_symbol_info[i].output_size)({ typeof (max_forw_output_size) _a = (max_forw_output_size); typeof (exec_symbol_info[i].output_size) _b = (exec_symbol_info [i].output_size); (_a > _b) ? _a : _b; }); | |||
1089 | if (backward_info[i].outgoings) | |||
1090 | { | |||
1091 | // Copy over the outgoing bits. | |||
1092 | autograd_execs[i].outgoings = ccv_array_new(sizeof(int), backward_info[i].outgoings->rnum, 0); | |||
1093 | for (j = 0; j < backward_info[i].outgoings->rnum; j++) | |||
1094 | { | |||
1095 | const int d = *(int*)ccv_array_get(backward_info[i].outgoings, j)((void*)(((char*)((backward_info[i].outgoings)->data)) + ( size_t)(backward_info[i].outgoings)->rsize * (size_t)(j))); | |||
1096 | // Only push the outgoing node if it is in the f_wrt path. | |||
1097 | if ((backward_info[d].f_wrt & 0x3) == 0x3) | |||
1098 | ccv_array_push(autograd_execs[i].outgoings, &d); | |||
1099 | } | |||
1100 | } | |||
1101 | } | |||
1102 | int max_forw_inputs[ccv_max(1, max_forw_input_size)({ typeof (1) _a = (1); typeof (max_forw_input_size) _b = (max_forw_input_size ); (_a > _b) ? _a : _b; })]; | |||
1103 | int max_forw_outputs[ccv_max(1, max_forw_output_size)({ typeof (1) _a = (1); typeof (max_forw_output_size) _b = (max_forw_output_size ); (_a > _b) ? _a : _b; })]; | |||
1104 | ccv_nnc_autograd_tensor_version_t* const autograd_tensor_versions = (ccv_nnc_autograd_tensor_version_t*)cccalloccalloc(tensor_symbol_info_size, sizeof(ccv_nnc_autograd_tensor_version_t)); | |||
1105 | ccv_array_t* autograd_tensor_symbols = ccv_array_new(sizeof(ccv_nnc_autograd_tensor_symbol_t), tensor_symbol_info_size, 0); | |||
1106 | ccv_array_t* sum_or_set_execs = ccv_array_new(sizeof(ccv_nnc_sum_or_set_graph_exec_symbol_t), 0, 0); | |||
1107 | ccv_nnc_graph_visit_for(backward_visit, backward_info, back_info_node, idx){ int _i_; for (_i_ = 0; _i_ < (backward_visit)->size; _i_ ++) { const int idx __attribute__((unused)) = (backward_visit )->node[_i_].index; const int _node_unused_ __attribute__( (unused)) = (backward_visit)->node[_i_].term; typeof ((backward_info )) const back_info_node __attribute__((unused)) = (backward_info ) + idx; { | |||
1108 | /* This is required by both f flow and wrt flow, therefore, an interest to us */ | |||
1109 | if ((back_info_node->f_wrt & 0x3) == 0x3) | |||
1110 | { | |||
1111 | const ccv_nnc_graph_exec_symbol_info_t* forw_exec = exec_symbol_info + idx; | |||
1112 | ccv_nnc_autograd_graph_exec_symbol_t* back_exec = autograd_execs + idx; | |||
1113 | back_exec->cmd = forw_exec->cmd; | |||
1114 | if (back_exec->cmd.cmd != CCV_NNC_NOOP) | |||
1115 | back_exec->cmd.cmd += 1; /* Backward command is the one after forward command. */ | |||
1116 | assert(ccv_nnc_cmd_is_backward(back_exec->cmd) || back_exec->cmd.cmd == CCV_NNC_NOOP)((void) sizeof ((ccv_nnc_cmd_is_backward(back_exec->cmd) || back_exec->cmd.cmd == CCV_NNC_NOOP) ? 1 : 0), __extension__ ({ if (ccv_nnc_cmd_is_backward(back_exec->cmd) || back_exec ->cmd.cmd == CCV_NNC_NOOP) ; else __assert_fail ("ccv_nnc_cmd_is_backward(back_exec->cmd) || back_exec->cmd.cmd == CCV_NNC_NOOP" , "ccv_nnc_symbolic_graph_backward.c", 1116, __extension__ __PRETTY_FUNCTION__ ); })); | |||
1117 | if (!back_info_node->output_bitmask_size) /* This has no output, can be a noop. */ | |||
1118 | back_exec->cmd.cmd = CCV_NNC_NOOP; | |||
1119 | else { | |||
1120 | int* back_input_map = max_forw_outputs; | |||
1121 | int* back_output_map = max_forw_inputs; | |||
1122 | _ccv_nnc_symbolic_graph_backward_exec_io(forw_exec, &back_input_map, &back_output_map, &back_exec->input_size, &back_exec->output_size); | |||
1123 | back_exec->inputs = ccmallocmalloc(sizeof(int) * (back_exec->input_size + back_exec->output_size)); | |||
1124 | back_exec->outputs = back_exec->inputs + back_exec->input_size; | |||
1125 | /* Need to compute input before we compute output */ | |||
1126 | for (i = 0; i < back_exec->input_size; i++) | |||
1127 | { | |||
1128 | /* If we can skip this input, do that. */ | |||
1129 | if (!(back_info_node->input_bitmasks[i >> 6] & ((uint64_t)1 << i))) | |||
1130 | continue; | |||
1131 | const int d = back_input_map[i]; | |||
1132 | const int alias_ref = tensor_symbol_info[d].alias_ref; | |||
1133 | ccv_nnc_autograd_tensor_version_t* tensor_ver = alias_ref ? autograd_tensor_versions + (alias_ref - 1) : autograd_tensor_versions + d; | |||
1134 | /* Initialization tensor, should corresponding to f symbols */ | |||
1135 | if (!tensor_ver->ref_version) | |||
1136 | { | |||
1137 | ccv_nnc_autograd_tensor_symbol_t tensor_sym = {}; | |||
1138 | if (!alias_ref) | |||
1139 | { | |||
1140 | tensor_sym.d = d; | |||
1141 | ccv_array_push(autograd_tensor_symbols, &tensor_sym); | |||
1142 | const ccv_nnc_tensor_ref_t tensor_ref = { | |||
1143 | .d = autograd_tensor_symbols->rnum - 1, | |||
1144 | .x = idx, | |||
1145 | .alias_registry = 0 | |||
1146 | }; | |||
1147 | tensor_ver->ref_version = ccv_array_new(sizeof(ccv_nnc_tensor_ref_t), 1, 0); | |||
1148 | ccv_array_push(tensor_ver->ref_version, &tensor_ref); | |||
1149 | } else { | |||
1150 | tensor_sym.d = alias_ref - 1; | |||
1151 | ccv_array_push(autograd_tensor_symbols, &tensor_sym); | |||
1152 | const ccv_nnc_tensor_ref_t tensor_ref = { | |||
1153 | .d = autograd_tensor_symbols->rnum - 1, | |||
1154 | .x = idx, | |||
1155 | .alias_registry = ccv_array_new(sizeof(int), 1, 0) | |||
1156 | }; | |||
1157 | tensor_ver->ref_version = ccv_array_new(sizeof(ccv_nnc_tensor_ref_t), 1, 0); | |||
1158 | ccv_array_push(tensor_ver->ref_version, &tensor_ref); | |||
1159 | tensor_sym.d = d; /* set back */ | |||
1160 | tensor_sym.alias_ref = tensor_ref.d + 1; | |||
1161 | ccv_array_push(autograd_tensor_symbols, &tensor_sym); | |||
1162 | const int ad = autograd_tensor_symbols->rnum - 1; | |||
1163 | ccv_array_push(tensor_ref.alias_registry, &ad); | |||
1164 | } | |||
1165 | } | |||
1166 | /* The simplest case (most common), it is not an alias. */ | |||
1167 | if (!alias_ref) | |||
1168 | { | |||
1169 | /* Even simpler, this only have one reference tensor, thus, pass this as input. */ | |||
1170 | if (tensor_ver->c == tensor_ver->ref_version->rnum - 1) | |||
1171 | { | |||
1172 | ccv_nnc_tensor_ref_t* tensor_ref = (ccv_nnc_tensor_ref_t*)ccv_array_get(tensor_ver->ref_version, tensor_ver->c)((void*)(((char*)((tensor_ver->ref_version)->data)) + ( size_t)(tensor_ver->ref_version)->rsize * (size_t)(tensor_ver ->c))); | |||
1173 | /* There are alias associated with this tensor ref, zero it out when this tensor is allocated. */ | |||
1174 | /* This is is required. Consider the case that we have an alias of this tensor used somehwere */ | |||
1175 | /* on forward pass, when we compute backward, we have that alias computed first, however, its */ | |||
1176 | /* underlying tensor is not zero initialized, and we will end up with garbage values here. */ | |||
1177 | if (tensor_ref->alias_registry && | |||
1178 | /* Loop over to see if this tensor is fully occupied to avoid extra zero step. */ | |||
1179 | !_ccv_nnc_tensor_ref_fully_assigned_with_aliases(tensor_ref, autograd_tensor_symbols, tensor_symbol_info)) | |||
1180 | { | |||
1181 | ccv_nnc_autograd_tensor_symbol_t* tensor_sym = (ccv_nnc_autograd_tensor_symbol_t*)ccv_array_get(autograd_tensor_symbols, tensor_ref->d)((void*)(((char*)((autograd_tensor_symbols)->data)) + (size_t )(autograd_tensor_symbols)->rsize * (size_t)(tensor_ref-> d))); | |||
1182 | assert(tensor_sym->alias_ref == 0)((void) sizeof ((tensor_sym->alias_ref == 0) ? 1 : 0), __extension__ ({ if (tensor_sym->alias_ref == 0) ; else __assert_fail ( "tensor_sym->alias_ref == 0", "ccv_nnc_symbolic_graph_backward.c" , 1182, __extension__ __PRETTY_FUNCTION__); })); | |||
1183 | tensor_sym->flags = CCV_NNC_TENSOR_SYMBOL_INIT_ZEROS; | |||
1184 | } | |||
1185 | back_exec->inputs[i] = tensor_ref->d; | |||
1186 | } else { | |||
1187 | /* Otherwise, we need to sum them up, and then pass the summed result to the computation. */ | |||
1188 | _ccv_nnc_graph_sum_autograd_tensor_versions(idx, d, exec_symbol_info_size, tensor_symbol_info, tensor_ver, autograd_execs, autograd_tensor_symbols, sum_or_set_execs); | |||
1189 | ccv_nnc_tensor_ref_t* tensor_ref = (ccv_nnc_tensor_ref_t*)ccv_array_get(tensor_ver->ref_version, tensor_ver->c)((void*)(((char*)((tensor_ver->ref_version)->data)) + ( size_t)(tensor_ver->ref_version)->rsize * (size_t)(tensor_ver ->c))); | |||
1190 | back_exec->inputs[i] = tensor_ref->d; | |||
1191 | } | |||
1192 | } else | |||
1193 | /* If this is an alias, go through all available tensor ref versions */ | |||
1194 | back_exec->inputs[i] = _ccv_nnc_graph_sum_autograd_tensor_versions_alias(idx, d, tensor_symbol_info, exec_symbol_info_size, tensor_symbol_info + d, tensor_ver, autograd_execs, autograd_tensor_symbols, sum_or_set_execs); | |||
1195 | } | |||
1196 | for (i = 0; i < back_exec->output_size; i++) | |||
1197 | { | |||
1198 | /* If we can skip this output, do that. */ | |||
1199 | if (!(back_info_node->output_bitmasks[i >> 6] & ((uint64_t)1 << i))) | |||
1200 | continue; | |||
1201 | const int d = back_output_map[i]; | |||
1202 | const int alias_ref = tensor_symbol_info[d].alias_ref; | |||
1203 | ccv_nnc_autograd_tensor_symbol_t tensor_sym = { | |||
1204 | .d = d | |||
1205 | }; | |||
1206 | /* The simplest case (most common), it is not an alias. */ | |||
1207 | if (!alias_ref) | |||
1208 | { | |||
1209 | ccv_array_push(autograd_tensor_symbols, &tensor_sym); | |||
1210 | const ccv_nnc_tensor_ref_t tensor_ref = { | |||
1211 | .d = autograd_tensor_symbols->rnum - 1, | |||
1212 | .x = idx, | |||
1213 | .exec_registry = 0, | |||
1214 | .alias_registry = 0 | |||
1215 | }; | |||
1216 | ccv_nnc_autograd_tensor_version_t* tensor_ver = autograd_tensor_versions + d; | |||
1217 | if (!tensor_ver->ref_version) | |||
1218 | tensor_ver->ref_version = ccv_array_new(sizeof(ccv_nnc_tensor_ref_t), 1, 0); | |||
1219 | ccv_array_push(tensor_ver->ref_version, &tensor_ref); | |||
1220 | back_exec->outputs[i] = tensor_ref.d; | |||
1221 | } else { | |||
1222 | /* Otherwise, in case that this is an alias, we try to find the existing one (in tensor_ver | |||
1223 | * see if can meet the need (thus, for the tensor info / ofs, it fits). */ | |||
1224 | ccv_nnc_autograd_tensor_version_t* tensor_ver = autograd_tensor_versions + (alias_ref - 1); | |||
1225 | if (!tensor_ver->ref_version) | |||
1226 | tensor_ver->ref_version = ccv_array_new(sizeof(ccv_nnc_tensor_ref_t), 1, 0); | |||
1227 | /* If already exists a ref version, check if any of these not-sealed tensors have free space. */ | |||
1228 | int found = 0; | |||
1229 | for (j = tensor_ver->c; !found && j < tensor_ver->ref_version->rnum; j++) | |||
1230 | { | |||
1231 | ccv_nnc_tensor_ref_t* tensor_ref = (ccv_nnc_tensor_ref_t*)ccv_array_get(tensor_ver->ref_version, j)((void*)(((char*)((tensor_ver->ref_version)->data)) + ( size_t)(tensor_ver->ref_version)->rsize * (size_t)(j))); | |||
1232 | if (!_ccv_nnc_tensor_ref_version_involve_alias(tensor_ref, autograd_tensor_symbols, tensor_symbol_info, tensor_symbol_info + d)) | |||
1233 | { | |||
1234 | tensor_sym.alias_ref = tensor_ref->d + 1; | |||
1235 | ccv_array_push(autograd_tensor_symbols, &tensor_sym); | |||
1236 | const int ad = autograd_tensor_symbols->rnum - 1; | |||
1237 | ccv_array_push(tensor_ref->alias_registry, &ad); | |||
1238 | if (!tensor_ref->exec_registry) | |||
1239 | tensor_ref->exec_registry = ccv_array_new(sizeof(int), 1, 0); | |||
1240 | ccv_array_push(tensor_ref->exec_registry, &idx); | |||
1241 | back_exec->outputs[i] = ad; | |||
1242 | found = 1; | |||
1243 | } | |||
1244 | } | |||
1245 | if (!found) /* Cannot find an tensor ref to insert, create one first */ | |||
1246 | { | |||
1247 | tensor_sym.d = alias_ref - 1; /* Reference back to the non-alias. */ | |||
1248 | ccv_array_push(autograd_tensor_symbols, &tensor_sym); | |||
1249 | const ccv_nnc_tensor_ref_t tensor_ref = { | |||
1250 | .d = autograd_tensor_symbols->rnum - 1, | |||
1251 | .x = idx, | |||
1252 | .exec_registry = 0, | |||
1253 | .alias_registry = ccv_array_new(sizeof(int), 1, 0) | |||
1254 | }; | |||
1255 | ccv_array_push(tensor_ver->ref_version, &tensor_ref); | |||
1256 | tensor_sym.d = d; /* set back */ | |||
1257 | tensor_sym.alias_ref = tensor_ref.d + 1; | |||
1258 | ccv_array_push(autograd_tensor_symbols, &tensor_sym); | |||
1259 | const int ad = autograd_tensor_symbols->rnum - 1; | |||
1260 | ccv_array_push(tensor_ref.alias_registry, &ad); | |||
1261 | back_exec->outputs[i] = ad; | |||
1262 | } | |||
1263 | } | |||
1264 | } | |||
1265 | } | |||
1266 | } | |||
1267 | } ccv_nnc_graph_visit_endfor} } | |||
1268 | // Find all relevant wrt symbols, generate sum for them if needed. | |||
1269 | for (i = 0; i < wrt_symbol_size; i++) | |||
1270 | { | |||
1271 | const int d = wrt_symbols[i].d; | |||
1272 | if (d < 0) | |||
1273 | continue; | |||
1274 | const int ref_d = (!tensor_symbol_info[d].alias_ref) ? d : tensor_symbol_info[d].alias_ref - 1; | |||
1275 | ccv_nnc_autograd_tensor_version_t* tensor_ver = autograd_tensor_versions + ref_d; | |||
1276 | if (!tensor_ver->ref_version) | |||
1277 | { | |||
1278 | // This wrt symbol is not available at all, for this case, we set its flag to init zero. | |||
1279 | const ccv_nnc_autograd_tensor_symbol_t tensor_sym = { | |||
1280 | .d = ref_d | |||
1281 | }; | |||
1282 | ccv_array_push(autograd_tensor_symbols, &tensor_sym); | |||
1283 | ccv_nnc_sum_or_set_graph_exec_symbol_t set_exec = { | |||
1284 | .value = 0, | |||
1285 | .output = autograd_tensor_symbols->rnum - 1, | |||
1286 | }; | |||
1287 | ccv_array_push(sum_or_set_execs, &set_exec); | |||
1288 | // Insert the one to be set to zero. | |||
1289 | const ccv_nnc_tensor_ref_t tensor_ref = { | |||
1290 | .d = autograd_tensor_symbols->rnum - 1, | |||
1291 | .x = exec_symbol_info_size + sum_or_set_execs->rnum - 1, | |||
1292 | }; | |||
1293 | tensor_ver->ref_version = ccv_array_new(sizeof(ccv_nnc_tensor_ref_t), 1, 0); | |||
1294 | ccv_array_push(tensor_ver->ref_version, &tensor_ref); | |||
1295 | continue; | |||
1296 | } | |||
1297 | // If it is a while loop, we need to insert an accumulator to the graph (this is expressed as a initialization tensor summed with existing results). | |||
1298 | // First, insert the initialization tensor if this wrt results is not used directly in next while loop (thus, it participates the computation, therefore, no need to accumulate). | |||
1299 | if (is_while && !tensor_symbol_info[ref_d].assign_ref && | |||
1300 | _ccv_nnc_tensor_ref_version_find_init(tensor_ver) < 0) // If the initialization tensor is not inserted yet. | |||
1301 | { | |||
1302 | const ccv_nnc_autograd_tensor_symbol_t tensor_sym = { | |||
1303 | .d = ref_d | |||
1304 | }; | |||
1305 | ccv_array_push(autograd_tensor_symbols, &tensor_sym); | |||
1306 | // Insert the one to be summed. | |||
1307 | const ccv_nnc_tensor_ref_t tensor_ref = { | |||
1308 | .d = autograd_tensor_symbols->rnum - 1, | |||
1309 | .x = -1, // This denotes it is an initialization vector. | |||
1310 | }; | |||
1311 | ccv_array_push(tensor_ver->ref_version, &tensor_ref); | |||
1312 | } | |||
1313 | // If there are more than one tensor in the list, it is possible to sum them up. | |||
1314 | if (tensor_ver->c < tensor_ver->ref_version->rnum - 1) | |||
1315 | _ccv_nnc_graph_sum_autograd_tensor_versions(-1, ref_d, exec_symbol_info_size, tensor_symbol_info, tensor_ver, autograd_execs, autograd_tensor_symbols, sum_or_set_execs); | |||
1316 | // The tensor version should have ref_version, and only one now (after sum up). | |||
1317 | assert(tensor_ver->c == tensor_ver->ref_version->rnum - 1)((void) sizeof ((tensor_ver->c == tensor_ver->ref_version ->rnum - 1) ? 1 : 0), __extension__ ({ if (tensor_ver-> c == tensor_ver->ref_version->rnum - 1) ; else __assert_fail ("tensor_ver->c == tensor_ver->ref_version->rnum - 1" , "ccv_nnc_symbolic_graph_backward.c", 1317, __extension__ __PRETTY_FUNCTION__ ); })); | |||
1318 | } | |||
1319 | // Adding additional fields to backward_prep now. | |||
1320 | backward_prep->autograd_execs = autograd_execs; | |||
1321 | backward_prep->autograd_tensor_versions = autograd_tensor_versions; | |||
1322 | backward_prep->autograd_tensor_symbols = autograd_tensor_symbols; | |||
1323 | backward_prep->sum_or_set_execs = sum_or_set_execs; | |||
1324 | ccv_array_t* sub_f_symbols = 0; | |||
1325 | ccv_array_t* sub_wrt_symbols = 0; | |||
1326 | ccv_nnc_graph_visit_for(forward_visit, exec_symbol_info, _, idx){ int _i_; for (_i_ = 0; _i_ < (forward_visit)->size; _i_ ++) { const int idx __attribute__((unused)) = (forward_visit) ->node[_i_].index; const int _node_unused_ __attribute__(( unused)) = (forward_visit)->node[_i_].term; typeof ((exec_symbol_info )) const _ __attribute__((unused)) = (exec_symbol_info) + idx ; { | |||
1327 | ccv_nnc_graph_backward_info_t* node = backward_info + idx; | |||
1328 | const ccv_nnc_graph_exec_symbol_info_t* forw_exec = exec_symbol_info + idx; | |||
1329 | /* Only interested in the ones on the f / wrt flow */ | |||
1330 | if ((node->f_wrt & 0x3) == 0x3) | |||
1331 | { | |||
1332 | const int is_while = (forw_exec->flags & CCV_NNC_GRAPH_EXEC_P_WHILE); | |||
1333 | for (i = 0; i < forw_exec->graph_ref_size; i++) | |||
1334 | { | |||
1335 | // Now calling it recursively until we are sure no f_symbols can be removed. | |||
1336 | const int graph_ref = CCV_NNC_GRAPH_REF(forw_exec)((forw_exec)->_heap_graph_ref ? (forw_exec)->_heap_graph_ref : (forw_exec)->_inline_graph_ref)[i] - 1; | |||
1337 | ccv_nnc_symbolic_graph_backward_prep_t* const sub_prep = backward_prep->sub_preps + graph_ref; | |||
1338 | if (!sub_wrt_symbols) | |||
1339 | sub_wrt_symbols = ccv_array_new(sizeof(ccv_nnc_tensor_symbol_t), 0, 0); | |||
1340 | if (!sub_f_symbols) | |||
1341 | sub_f_symbols = ccv_array_new(sizeof(ccv_nnc_tensor_symbol_t), 0, 0); | |||
1342 | _ccv_nnc_symbolic_graph_backward_prep_sub_f_wrt_symbols(forw_exec, sub_prep->graph, graph_ref, tensor_symbol_info, node->input_bitmasks, node->output_bitmasks, sub_f_symbols, sub_wrt_symbols); | |||
1343 | _ccv_nnc_symbolic_graph_backward_prep_gen(sub_prep, (ccv_nnc_tensor_symbol_t*)ccv_array_get(sub_f_symbols, 0)((void*)(((char*)((sub_f_symbols)->data)) + (size_t)(sub_f_symbols )->rsize * (size_t)(0))), sub_f_symbols->rnum, (ccv_nnc_tensor_symbol_t*)ccv_array_get(sub_wrt_symbols, 0)((void*)(((char*)((sub_wrt_symbols)->data)) + (size_t)(sub_wrt_symbols )->rsize * (size_t)(0))), sub_wrt_symbols->rnum, is_while, ccv_nnc_symbolic_graph_sources(sub_prep->graph), ccv_nnc_symbolic_graph_source_size(sub_prep->graph), ccv_nnc_symbolic_graph_destinations(sub_prep->graph), ccv_nnc_symbolic_graph_destination_size(sub_prep->graph)); | |||
1344 | } | |||
1345 | } | |||
1346 | } ccv_nnc_graph_visit_endfor} } | |||
1347 | if (sub_f_symbols) | |||
1348 | ccv_array_free(sub_f_symbols); | |||
1349 | if (sub_wrt_symbols) | |||
1350 | ccv_array_free(sub_wrt_symbols); | |||
1351 | } | |||
1352 | ||||
1353 | static void _ccv_nnc_symbolic_graph_backward_prep_free(const ccv_nnc_symbolic_graph_backward_prep_t backward_prep) | |||
1354 | { | |||
1355 | int i, j; | |||
1356 | const int exec_symbol_info_size = backward_prep.exec_symbol_info_size; | |||
1357 | const int tensor_symbol_info_size = backward_prep.tensor_symbol_info_size; | |||
1358 | ccv_nnc_autograd_graph_exec_symbol_t* const autograd_execs = backward_prep.autograd_execs; | |||
1359 | if (autograd_execs) | |||
1360 | { | |||
1361 | for (i = 0; i < exec_symbol_info_size; i++) | |||
1362 | { | |||
1363 | if (autograd_execs[i].inputs) | |||
1364 | ccfreefree(autograd_execs[i].inputs); | |||
1365 | if (autograd_execs[i].outgoings) | |||
1366 | ccv_array_free(autograd_execs[i].outgoings); | |||
1367 | } | |||
1368 | ccfreefree(autograd_execs); | |||
1369 | } | |||
1370 | ccv_nnc_autograd_tensor_version_t* const autograd_tensor_versions = backward_prep.autograd_tensor_versions; | |||
1371 | if (autograd_tensor_versions) | |||
1372 | { | |||
1373 | for (i = 0; i < tensor_symbol_info_size; i++) | |||
1374 | { | |||
1375 | if (autograd_tensor_versions[i].ref_version) | |||
1376 | { | |||
1377 | for (j = 0; j < autograd_tensor_versions[i].ref_version->rnum; j++) | |||
1378 | { | |||
1379 | ccv_nnc_tensor_ref_t* ref_version = (ccv_nnc_tensor_ref_t*)ccv_array_get(autograd_tensor_versions[i].ref_version, j)((void*)(((char*)((autograd_tensor_versions[i].ref_version)-> data)) + (size_t)(autograd_tensor_versions[i].ref_version)-> rsize * (size_t)(j))); | |||
1380 | if (ref_version->exec_registry) | |||
1381 | ccv_array_free(ref_version->exec_registry); | |||
1382 | if (ref_version->alias_registry) | |||
1383 | ccv_array_free(ref_version->alias_registry); | |||
1384 | } | |||
1385 | ccv_array_free(autograd_tensor_versions[i].ref_version); | |||
1386 | } | |||
1387 | } | |||
1388 | ccfreefree(autograd_tensor_versions); | |||
1389 | } | |||
1390 | if (backward_prep.autograd_tensor_symbols) | |||
1391 | ccv_array_free(backward_prep.autograd_tensor_symbols); | |||
1392 | ccv_array_t* const sum_or_set_execs = backward_prep.sum_or_set_execs; | |||
1393 | if (sum_or_set_execs) | |||
1394 | { | |||
1395 | for (i = 0; i < sum_or_set_execs->rnum; i++) | |||
1396 | { | |||
1397 | ccv_nnc_sum_or_set_graph_exec_symbol_t* sum_or_set = (ccv_nnc_sum_or_set_graph_exec_symbol_t*)ccv_array_get(sum_or_set_execs, i)((void*)(((char*)((sum_or_set_execs)->data)) + (size_t)(sum_or_set_execs )->rsize * (size_t)(i))); | |||
1398 | if (sum_or_set->inputs) | |||
1399 | ccfreefree(sum_or_set->inputs); | |||
1400 | if (sum_or_set->outgoings) | |||
1401 | ccv_array_free(sum_or_set->outgoings); | |||
1402 | } | |||
1403 | ccv_array_free(sum_or_set_execs); | |||
1404 | } | |||
1405 | // Now afterwards, these are mandatory. | |||
1406 | ccv_nnc_graph_backward_info_t* const backward_info = backward_prep.backward_info; | |||
1407 | for (i = 0; i < exec_symbol_info_size; i++) | |||
1408 | { | |||
1409 | if (backward_info[i].outgoings) | |||
1410 | ccv_array_free(backward_info[i].outgoings); | |||
1411 | if (backward_info[i].input_bitmasks) | |||
1412 | ccfreefree(backward_info[i].input_bitmasks); | |||
1413 | } | |||
1414 | ccfreefree(backward_info); | |||
1415 | ccv_nnc_graph_visit_free(backward_prep.backward_visit); | |||
1416 | ccv_nnc_graph_visit_free(backward_prep.forward_visit); | |||
1417 | ccfreefree(backward_prep.exec_symbol_info); | |||
1418 | ccfreefree(backward_prep.tensor_symbol_info); | |||
1419 | for (i = 0; i < backward_prep.sub_prep_size; i++) | |||
1420 | _ccv_nnc_symbolic_graph_backward_prep_free(backward_prep.sub_preps[i]); | |||
1421 | if (backward_prep.sub_preps) | |||
1422 | ccfreefree(backward_prep.sub_preps); | |||
1423 | } | |||
1424 | ||||
1425 | static void _ccv_nnc_add_backward_breakpoint_for_symbol(const ccv_nnc_symbolic_graph_backward_prep_t* const backward_prep, const ccv_nnc_graph_exec_symbol_t breakpoint, ccv_nnc_symbolic_graph_t* const graph, ccv_array_t* const sub_breakpoints) | |||
1426 | { | |||
1427 | const ccv_nnc_graph_exec_symbol_t noop = ccv_nnc_graph_exec_symbol_new(graph, ccv_nnc_cmd(CCV_NNC_NOOP, 0, CMD_GENERIC()((ccv_nnc_cmd_param_t){.size={.dim={1,1,1}}}), 0), 0, 0, 0, 0, 0); | |||
1428 | ccv_array_push(sub_breakpoints, &noop); | |||
1429 | // Now need to hook this up to the graph. | |||
1430 | const ccv_nnc_graph_exec_symbol_info_t* const exec_symbol_info = backward_prep->exec_symbol_info; | |||
1431 | const ccv_nnc_graph_visit_t* const forward_visit = backward_prep->forward_visit; | |||
1432 | // Now, for each one of these, find a reverse graph. | |||
1433 | ccv_nnc_graph_backward_info_t* const backward_info = backward_prep->backward_info; | |||
1434 | int i; | |||
1435 | // Clean up the high bit. | |||
1436 | for (i = 0; i < backward_prep->exec_symbol_info_size; i++) | |||
1437 | backward_info[i].f_wrt &= ~0x4; | |||
1438 | assert((backward_info[breakpoint.d].f_wrt & 0x3) != 0x3)((void) sizeof (((backward_info[breakpoint.d].f_wrt & 0x3 ) != 0x3) ? 1 : 0), __extension__ ({ if ((backward_info[breakpoint .d].f_wrt & 0x3) != 0x3) ; else __assert_fail ("(backward_info[breakpoint.d].f_wrt & 0x3) != 0x3" , "ccv_nnc_symbolic_graph_backward.c", 1438, __extension__ __PRETTY_FUNCTION__ ); })); | |||
1439 | backward_info[breakpoint.d].f_wrt |= 0x4; | |||
1440 | const ccv_nnc_graph_visit_t* const backward_visit = backward_prep->backward_visit; | |||
1441 | const ccv_nnc_autograd_graph_exec_symbol_t* const autograd_execs = backward_prep->autograd_execs; | |||
1442 | // Going forward to find whether this breakpoint is a source node to some f_wrt nodes. | |||
1443 | ccv_nnc_graph_visit_for(forward_visit, exec_symbol_info, forw_exec, idx){ int _i_; for (_i_ = 0; _i_ < (forward_visit)->size; _i_ ++) { const int idx __attribute__((unused)) = (forward_visit) ->node[_i_].index; const int _node_unused_ __attribute__(( unused)) = (forward_visit)->node[_i_].term; typeof ((exec_symbol_info )) const forw_exec __attribute__((unused)) = (exec_symbol_info ) + idx; { | |||
1444 | ccv_nnc_graph_backward_info_t* const node = backward_info + idx; | |||
1445 | // If it is tagged on breakpoint flow, but not as both f or wrt, flow through it. | |||
1446 | if ((node->f_wrt & 0x4) && (node->f_wrt & 0x3) != 0x3) | |||
1447 | for (i = 0; forw_exec->outgoings && i < forw_exec->outgoings->rnum; i++) | |||
1448 | { | |||
1449 | const int outgoing_idx = *(int*)ccv_array_get(forw_exec->outgoings, i)((void*)(((char*)((forw_exec->outgoings)->data)) + (size_t )(forw_exec->outgoings)->rsize * (size_t)(i))); | |||
1450 | ccv_nnc_graph_backward_info_t* const outgoing_node = backward_info + outgoing_idx; | |||
1451 | // If this is a f_wrt node. Concatenate. | |||
1452 | if (!(outgoing_node->f_wrt & 0x4) && (outgoing_node->f_wrt & 0x3) == 0x3) | |||
1453 | ccv_nnc_graph_exec_symbol_concat(graph, autograd_execs[outgoing_idx].symbol, noop); | |||
1454 | outgoing_node->f_wrt |= 0x4; | |||
1455 | } | |||
1456 | } ccv_nnc_graph_visit_endfor} } | |||
1457 | // Going backward to find whether this breakpoint is a destination node for some f_wrt_nodes. | |||
1458 | ccv_nnc_graph_visit_for(backward_visit, backward_info, node, idx){ int _i_; for (_i_ = 0; _i_ < (backward_visit)->size; _i_ ++) { const int idx __attribute__((unused)) = (backward_visit )->node[_i_].index; const int _node_unused_ __attribute__( (unused)) = (backward_visit)->node[_i_].term; typeof ((backward_info )) const node __attribute__((unused)) = (backward_info) + idx ; { | |||
1459 | if ((node->f_wrt & 0x4) && (node->f_wrt & 0x3) != 0x3) | |||
1460 | for (i = 0; node->outgoings && i < node->outgoings->rnum; i++) | |||
1461 | { | |||
1462 | const int outgoing_idx = *(int*)ccv_array_get(node->outgoings, i)((void*)(((char*)((node->outgoings)->data)) + (size_t)( node->outgoings)->rsize * (size_t)(i))); | |||
1463 | ccv_nnc_graph_backward_info_t* const outgoing_node = backward_info + outgoing_idx; | |||
1464 | // If this is a f_wrt node. Concatenate. | |||
1465 | if (!(outgoing_node->f_wrt & 0x4) && (outgoing_node->f_wrt & 0x3) == 0x3) | |||
1466 | ccv_nnc_graph_exec_symbol_concat(graph, noop, autograd_execs[outgoing_idx].symbol); | |||
1467 | outgoing_node->f_wrt |= 0x4; | |||
1468 | } | |||
1469 | } ccv_nnc_graph_visit_endfor} } | |||
1470 | } | |||
1471 | ||||
1472 | static ccv_nnc_autograd_tensor_symbol_t* _ccv_nnc_autograd_tensor_symbol_from_tensor_version(ccv_array_t* const autograd_tensor_symbols, const ccv_nnc_autograd_tensor_version_t* const tensor_ver) | |||
1473 | { | |||
1474 | assert(tensor_ver->ref_version)((void) sizeof ((tensor_ver->ref_version) ? 1 : 0), __extension__ ({ if (tensor_ver->ref_version) ; else __assert_fail ("tensor_ver->ref_version" , "ccv_nnc_symbolic_graph_backward.c", 1474, __extension__ __PRETTY_FUNCTION__ ); })); | |||
1475 | const ccv_nnc_tensor_ref_t* const tensor_ref = (ccv_nnc_tensor_ref_t*)ccv_array_get(tensor_ver->ref_version, tensor_ver->c)((void*)(((char*)((tensor_ver->ref_version)->data)) + ( size_t)(tensor_ver->ref_version)->rsize * (size_t)(tensor_ver ->c))); | |||
1476 | return (ccv_nnc_autograd_tensor_symbol_t*)ccv_array_get(autograd_tensor_symbols, tensor_ref->d)((void*)(((char*)((autograd_tensor_symbols)->data)) + (size_t )(autograd_tensor_symbols)->rsize * (size_t)(tensor_ref-> d))); | |||
1477 | } | |||
1478 | ||||
1479 | static void _ccv_nnc_symbolic_graph_set_backward_carry_overs(const ccv_nnc_symbolic_graph_backward_prep_t* const backward_prep, const ccv_nnc_tensor_symbol_t* const wrt_symbols, const int wrt_symbol_size, ccv_nnc_symbolic_graph_t* const graph) | |||
1480 | { | |||
1481 | int i; | |||
1482 | for (i = 0; i < backward_prep->graph->tensor_symbol_info->rnum; i++) | |||
1483 | { | |||
1484 | const ccv_nnc_tensor_symbol_info_t* const tensor_symbol_info = backward_prep->tensor_symbol_info + i; | |||
1485 | if (tensor_symbol_info->assign_ref) | |||
1486 | { | |||
1487 | const int assign_ref = tensor_symbol_info->assign_ref - 1; | |||
1488 | ccv_nnc_autograd_tensor_symbol_t* const destination_autograd_symbol = _ccv_nnc_autograd_tensor_symbol_from_tensor_version(backward_prep->autograd_tensor_symbols, backward_prep->autograd_tensor_versions + assign_ref); | |||
1489 | ccv_nnc_autograd_tensor_symbol_t* const source_autograd_symbol = _ccv_nnc_autograd_tensor_symbol_from_tensor_version(backward_prep->autograd_tensor_symbols, backward_prep->autograd_tensor_versions + i); | |||
1490 | ccv_nnc_symbolic_graph_set_carry_overs(graph, (ccv_nnc_tensor_symbol_map_t []){ | |||
1491 | { .source = source_autograd_symbol->symbol, .destination = destination_autograd_symbol->symbol } | |||
1492 | }, 1); | |||
1493 | } | |||
1494 | } | |||
1495 | for (i = 0; i < wrt_symbol_size; i++) | |||
1496 | { | |||
1497 | const int d = wrt_symbols[i].d; | |||
1498 | if (d < 0) | |||
1499 | continue; | |||
1500 | const int ref_d = (!backward_prep->tensor_symbol_info[d].alias_ref) ? d : backward_prep->tensor_symbol_info[d].alias_ref - 1; | |||
1501 | const ccv_nnc_autograd_tensor_version_t* const tensor_ver = backward_prep->autograd_tensor_versions + ref_d; | |||
1502 | const int init_ref_ver = _ccv_nnc_tensor_ref_version_find_init(tensor_ver); | |||
1503 | if (init_ref_ver >= 0) | |||
1504 | { | |||
1505 | const int init_d = ((ccv_nnc_tensor_ref_t*)ccv_array_get(tensor_ver->ref_version, init_ref_ver)((void*)(((char*)((tensor_ver->ref_version)->data)) + ( size_t)(tensor_ver->ref_version)->rsize * (size_t)(init_ref_ver ))))->d; | |||
1506 | ccv_nnc_autograd_tensor_symbol_t* const destination_autograd_symbol = (ccv_nnc_autograd_tensor_symbol_t*)ccv_array_get(backward_prep->autograd_tensor_symbols, init_d)((void*)(((char*)((backward_prep->autograd_tensor_symbols) ->data)) + (size_t)(backward_prep->autograd_tensor_symbols )->rsize * (size_t)(init_d))); | |||
1507 | ccv_nnc_autograd_tensor_symbol_t* const source_autograd_symbol = _ccv_nnc_autograd_tensor_symbol_from_tensor_version(backward_prep->autograd_tensor_symbols, backward_prep->autograd_tensor_versions + ref_d); | |||
1508 | ccv_nnc_symbolic_graph_set_carry_overs(graph, (ccv_nnc_tensor_symbol_map_t []){ | |||
1509 | { .source = source_autograd_symbol->symbol, .destination = destination_autograd_symbol->symbol } | |||
1510 | }, 1); | |||
1511 | } | |||
1512 | } | |||
1513 | } | |||
1514 | ||||
1515 | static void _ccv_nnc_symbolic_graph_add_init_zeros(const ccv_nnc_symbolic_graph_backward_prep_t* const sub_prep, const ccv_nnc_tensor_symbol_t* const wrt_symbols, const int wrt_symbol_size, ccv_nnc_symbolic_graph_t* const graph, ccv_nnc_symbolic_graph_t* const sub_graph, ccv_array_t* const symbols) | |||
1516 | { | |||
1517 | int i; | |||
1518 | for (i = 0; i < wrt_symbol_size; i++) | |||
1519 | { | |||
1520 | const int d = wrt_symbols[i].d; | |||
1521 | if (d < 0) | |||
1522 | continue; | |||
1523 | const int ref_d = (!sub_prep->tensor_symbol_info[d].alias_ref) ? d : sub_prep->tensor_symbol_info[d].alias_ref - 1; | |||
1524 | const ccv_nnc_autograd_tensor_version_t* const tensor_ver = sub_prep->autograd_tensor_versions + ref_d; | |||
1525 | const int init_ref_ver = _ccv_nnc_tensor_ref_version_find_init(tensor_ver); | |||
1526 | if (init_ref_ver >= 0) | |||
1527 | { | |||
1528 | // Need de-dup logic. | |||
1529 | const int init_d = ((ccv_nnc_tensor_ref_t*)ccv_array_get(tensor_ver->ref_version, init_ref_ver)((void*)(((char*)((tensor_ver->ref_version)->data)) + ( size_t)(tensor_ver->ref_version)->rsize * (size_t)(init_ref_ver ))))->d; | |||
1530 | ccv_nnc_autograd_tensor_symbol_t* const init_autograd_symbol = (ccv_nnc_autograd_tensor_symbol_t*)ccv_array_get(sub_prep->autograd_tensor_symbols, init_d)((void*)(((char*)((sub_prep->autograd_tensor_symbols)-> data)) + (size_t)(sub_prep->autograd_tensor_symbols)->rsize * (size_t)(init_d))); | |||
1531 | const ccv_nnc_tensor_symbol_info_t* const sub_init_symbol_info = (ccv_nnc_tensor_symbol_info_t*)ccv_array_get(sub_graph->tensor_symbol_info, init_autograd_symbol->symbol.d)((void*)(((char*)((sub_graph->tensor_symbol_info)->data )) + (size_t)(sub_graph->tensor_symbol_info)->rsize * ( size_t)(init_autograd_symbol->symbol.d))); | |||
1532 | // If it doesn't have a parent ref yet, create one. | |||
1533 | if (!sub_init_symbol_info->p_ref) | |||
1534 | { | |||
1535 | ccv_nnc_tensor_symbol_t new_symbol = ccv_nnc_tensor_symbol_new(graph, sub_prep->tensor_symbol_info[ref_d].info, 0); | |||
1536 | ccv_nnc_tensor_symbol_set_flags(graph, new_symbol, CCV_NNC_TENSOR_SYMBOL_INIT_ZEROS); | |||
1537 | ccv_array_push(symbols, &new_symbol); | |||
1538 | ccv_nnc_tensor_symbol_hookup(graph, sub_graph, new_symbol, init_autograd_symbol->symbol); | |||
1539 | } | |||
1540 | } | |||
1541 | } | |||
1542 | } | |||
1543 | ||||
1544 | static void _ccv_nnc_symbolic_graph_add_tape_vars(const ccv_nnc_symbolic_graph_backward_prep_t* const sub_prep, ccv_nnc_symbolic_graph_t* const root, ccv_nnc_symbolic_graph_t* const graph, ccv_nnc_symbolic_graph_t* const sub_graph, ccv_array_t* const symbols) | |||
1545 | { | |||
1546 | int i; | |||
1547 | for (i = 0; i < sub_graph->tensor_symbol_info->rnum; i++) | |||
1548 | { | |||
1549 | const ccv_nnc_tensor_symbol_info_t* const symbol_info = (ccv_nnc_tensor_symbol_info_t*)ccv_array_get(sub_graph->tensor_symbol_info, i)((void*)(((char*)((sub_graph->tensor_symbol_info)->data )) + (size_t)(sub_graph->tensor_symbol_info)->rsize * ( size_t)(i))); | |||
1550 | if ((symbol_info->flags & CCV_NNC_TENSOR_SYMBOL_TAPE_VAR) && symbol_info->pair_ref) | |||
1551 | { | |||
1552 | const int pair_ref = symbol_info->pair_ref - 1; | |||
1553 | const ccv_nnc_tensor_symbol_t root_symbol = ccv_nnc_tensor_symbol_resolve(root, (ccv_nnc_tensor_symbol_t){ | |||
1554 | .d = pair_ref, | |||
1555 | .graph = sub_prep->graph, | |||
1556 | }); | |||
1557 | if (root_symbol.d >= 0) | |||
1558 | { | |||
1559 | ccv_nnc_tensor_symbol_hookup(root, sub_graph, root_symbol, (ccv_nnc_tensor_symbol_t){ | |||
1560 | .d = i, | |||
1561 | .graph = sub_graph, | |||
1562 | }); | |||
1563 | if (symbols) | |||
1564 | { | |||
1565 | const ccv_nnc_tensor_symbol_t p_symbol = ccv_nnc_tensor_symbol_resolve(graph, (ccv_nnc_tensor_symbol_t){ | |||
1566 | .d = i, | |||
1567 | .graph = sub_graph, | |||
1568 | }); | |||
1569 | ccv_array_push(symbols, &p_symbol); | |||
1570 | } | |||
1571 | } | |||
1572 | } | |||
1573 | } | |||
1574 | } | |||
1575 | ||||
1576 | static void _ccv_nnc_symbolic_graph_backward_gen(const ccv_nnc_symbolic_graph_backward_prep_t* const backward_prep, const ccv_nnc_tensor_symbol_t* const f_symbols, const int f_symbol_size, const ccv_nnc_tensor_symbol_t* const wrt_symbols, const int wrt_symbol_size, ccv_nnc_symbolic_graph_t* const graph, ccv_nnc_symbolic_graph_t* const root) | |||
1577 | { | |||
1578 | assert(graph == backward_prep->graph || graph->pair == backward_prep->graph)((void) sizeof ((graph == backward_prep->graph || graph-> pair == backward_prep->graph) ? 1 : 0), __extension__ ({ if (graph == backward_prep->graph || graph->pair == backward_prep ->graph) ; else __assert_fail ("graph == backward_prep->graph || graph->pair == backward_prep->graph" , "ccv_nnc_symbolic_graph_backward.c", 1578, __extension__ __PRETTY_FUNCTION__ ); })); | |||
| ||||
1579 | const int exec_symbol_info_size = backward_prep->exec_symbol_info_size; | |||
1580 | const int tensor_symbol_info_size = backward_prep->tensor_symbol_info_size; | |||
1581 | const ccv_nnc_graph_exec_symbol_info_t* const exec_symbol_info = backward_prep->exec_symbol_info; | |||
1582 | const ccv_nnc_tensor_symbol_info_t* const tensor_symbol_info = backward_prep->tensor_symbol_info; | |||
1583 | int i, j, k, p; | |||
1584 | ccv_array_t* const autograd_tensor_symbols = backward_prep->autograd_tensor_symbols; | |||
1585 | // Generate required symbols based on the information gathered above. | |||
1586 | for (i = 0; i < autograd_tensor_symbols->rnum; i++) | |||
1587 | { | |||
1588 | ccv_nnc_autograd_tensor_symbol_t* symbol = (ccv_nnc_autograd_tensor_symbol_t*)ccv_array_get(autograd_tensor_symbols, i)((void*)(((char*)((autograd_tensor_symbols)->data)) + (size_t )(autograd_tensor_symbols)->rsize * (size_t)(i))); | |||
1589 | assert(symbol->d >= 0)((void) sizeof ((symbol->d >= 0) ? 1 : 0), __extension__ ({ if (symbol->d >= 0) ; else __assert_fail ("symbol->d >= 0" , "ccv_nnc_symbolic_graph_backward.c", 1589, __extension__ __PRETTY_FUNCTION__ ); })); | |||
1590 | assert(symbol->d < tensor_symbol_info_size)((void) sizeof ((symbol->d < tensor_symbol_info_size) ? 1 : 0), __extension__ ({ if (symbol->d < tensor_symbol_info_size ) ; else __assert_fail ("symbol->d < tensor_symbol_info_size" , "ccv_nnc_symbolic_graph_backward.c", 1590, __extension__ __PRETTY_FUNCTION__ ); })); | |||
1591 | const ccv_nnc_tensor_symbol_info_t* const forw_symbol = tensor_symbol_info + symbol->d; | |||
1592 | if (!symbol->alias_ref) | |||
1593 | { | |||
1594 | assert(!forw_symbol->alias_ref)((void) sizeof ((!forw_symbol->alias_ref) ? 1 : 0), __extension__ ({ if (!forw_symbol->alias_ref) ; else __assert_fail ("!forw_symbol->alias_ref" , "ccv_nnc_symbolic_graph_backward.c", 1594, __extension__ __PRETTY_FUNCTION__ ); })); | |||
1595 | symbol->symbol = ccv_nnc_tensor_symbol_new(graph, forw_symbol->info, 0); | |||
1596 | ccv_nnc_tensor_symbol_set_flags(graph, symbol->symbol, symbol->flags); | |||
1597 | } else { | |||
1598 | assert(forw_symbol->alias_ref)((void) sizeof ((forw_symbol->alias_ref) ? 1 : 0), __extension__ ({ if (forw_symbol->alias_ref) ; else __assert_fail ("forw_symbol->alias_ref" , "ccv_nnc_symbolic_graph_backward.c", 1598, __extension__ __PRETTY_FUNCTION__ ); })); | |||
1599 | assert(symbol->flags == 0)((void) sizeof ((symbol->flags == 0) ? 1 : 0), __extension__ ({ if (symbol->flags == 0) ; else __assert_fail ("symbol->flags == 0" , "ccv_nnc_symbolic_graph_backward.c", 1599, __extension__ __PRETTY_FUNCTION__ ); })); // We don't set flags on alias. | |||
1600 | // Due to our generation order, this must be after the original symbol is created. | |||
1601 | ccv_nnc_autograd_tensor_symbol_t* ref = (ccv_nnc_autograd_tensor_symbol_t*)ccv_array_get(autograd_tensor_symbols, symbol->alias_ref - 1)((void*)(((char*)((autograd_tensor_symbols)->data)) + (size_t )(autograd_tensor_symbols)->rsize * (size_t)(symbol->alias_ref - 1))); | |||
1602 | symbol->symbol = ccv_nnc_tensor_symbol_alias_new(graph, ref->symbol, forw_symbol->ofs, forw_symbol->stride, forw_symbol->info, 0); | |||
1603 | } | |||
1604 | } | |||
1605 | ccv_nnc_graph_backward_info_t* const backward_info = backward_prep->backward_info; | |||
1606 | ccv_nnc_autograd_graph_exec_symbol_t* const autograd_execs = backward_prep->autograd_execs; | |||
1607 | ccv_array_t* symbols = ccv_array_new(sizeof(ccv_nnc_tensor_symbol_t), 0, 0); | |||
1608 | ccv_array_t* symbol_map = ccv_array_new(sizeof(ccv_nnc_tensor_symbol_map_t), 0, 0); | |||
1609 | ccv_array_t* sub_f_symbols = 0; | |||
1610 | ccv_array_t* sub_wrt_symbols = 0; | |||
1611 | ccv_array_t* sub_execs = 0; | |||
1612 | for (i = 0; i < exec_symbol_info_size; i++) | |||
1613 | { | |||
1614 | // This is not going to be an interesting node. Skip. | |||
1615 | if ((backward_info[i].f_wrt & 0x3) != 0x3) | |||
1616 | continue; | |||
1617 | ccv_nnc_graph_backward_info_t* const back_info = backward_info + i; | |||
1618 | ccv_nnc_autograd_graph_exec_symbol_t* const back_exec = autograd_execs + i; | |||
1619 | if (back_exec->cmd.cmd == CCV_NNC_NOOP) | |||
1620 | { | |||
1621 | back_exec->symbol = ccv_nnc_graph_exec_symbol_new(graph, back_exec->cmd, 0, 0, 0, 0, 0); | |||
1622 | continue; | |||
1623 | } | |||
1624 | const ccv_nnc_graph_exec_symbol_info_t* const forw_exec = exec_symbol_info + i; | |||
1625 | if (forw_exec->flags & CCV_NNC_GRAPH_EXEC_P_WHILE) | |||
1626 | { | |||
1627 | ccv_array_clear(symbols); | |||
1628 | const int graph_ref = CCV_NNC_GRAPH_REF(forw_exec)((forw_exec)->_heap_graph_ref ? (forw_exec)->_heap_graph_ref : (forw_exec)->_inline_graph_ref)[0] - 1; | |||
1629 | ccv_nnc_symbolic_graph_backward_prep_t* sub_prep = backward_prep->sub_preps + graph_ref; | |||
1630 | ccv_nnc_symbolic_graph_t* sub_graph = ccv_nnc_symbolic_graph_new(); | |||
1631 | sub_graph->pair = sub_prep->graph; | |||
1632 | if (!sub_wrt_symbols
| |||
1633 | sub_wrt_symbols = ccv_array_new(sizeof(ccv_nnc_tensor_symbol_t), 0, 0); | |||
1634 | // I am done, need to redo above for sub_prep, and it has to be successful now. | |||
1635 | if (!sub_f_symbols
| |||
1636 | sub_f_symbols = ccv_array_new(sizeof(ccv_nnc_tensor_symbol_t), 0, 0); | |||
1637 | _ccv_nnc_symbolic_graph_backward_prep_sub_f_wrt_symbols(forw_exec, sub_prep->graph, graph_ref, tensor_symbol_info, back_info->input_bitmasks, back_info->output_bitmasks, sub_f_symbols, sub_wrt_symbols); | |||
1638 | _ccv_nnc_symbolic_graph_backward_gen(sub_prep, (ccv_nnc_tensor_symbol_t*)ccv_array_get(sub_f_symbols, 0)((void*)(((char*)((sub_f_symbols)->data)) + (size_t)(sub_f_symbols )->rsize * (size_t)(0))), sub_f_symbols->rnum, (ccv_nnc_tensor_symbol_t*)ccv_array_get(sub_wrt_symbols, 0)((void*)(((char*)((sub_wrt_symbols)->data)) + (size_t)(sub_wrt_symbols )->rsize * (size_t)(0))), sub_wrt_symbols->rnum, sub_graph, root); | |||
1639 | back_exec->symbol = ccv_nnc_symbolic_graph_while(graph, back_exec->cmd.cmd, sub_graph, forw_exec->name); | |||
1640 | if (!sub_execs) | |||
1641 | sub_execs = ccv_array_new(sizeof(ccv_nnc_graph_exec_symbol_t), 0, 0); | |||
1642 | ccv_array_clear(sub_execs); | |||
1643 | // Find the breakpoints in forward graph, creating the reverse one. | |||
1644 | for (j = 0; j < sub_prep->graph->breakpoint_size; j++) | |||
1645 | { | |||
1646 | const int d = sub_prep->graph->breakpoints[j].d; | |||
1647 | if (sub_prep->autograd_execs[d].symbol.graph) | |||
1648 | ccv_array_push(sub_execs, &sub_prep->autograd_execs[d].symbol); | |||
1649 | else | |||
1650 | _ccv_nnc_add_backward_breakpoint_for_symbol(sub_prep, sub_prep->graph->breakpoints[j], sub_graph, sub_execs); | |||
1651 | } | |||
1652 | ccv_nnc_symbolic_graph_set_while_expr(sub_graph, NOOP_GRAPH_WHILE_EXPR(ccv_nnc_graph_while_f)(1), 0, 0, 0, (ccv_nnc_graph_exec_symbol_t*)ccv_array_get(sub_execs, 0)((void*)(((char*)((sub_execs)->data)) + (size_t)(sub_execs )->rsize * (size_t)(0))), sub_execs->rnum); | |||
1653 | ccv_nnc_graph_exec_symbol_autogen(sub_graph, 0, 0, CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); | |||
1654 | _ccv_nnc_symbolic_graph_set_backward_carry_overs(sub_prep, (ccv_nnc_tensor_symbol_t*)ccv_array_get(sub_wrt_symbols, 0)((void*)(((char*)((sub_wrt_symbols)->data)) + (size_t)(sub_wrt_symbols )->rsize * (size_t)(0))), sub_wrt_symbols->rnum, sub_graph); | |||
1655 | for (j = 0; j < back_exec->input_size; j++) | |||
1656 | if (back_info->input_bitmasks[j >> 6] & ((uint64_t)1 << j)) | |||
1657 | ccv_array_push(symbols, &(((ccv_nnc_autograd_tensor_symbol_t*)ccv_array_get(autograd_tensor_symbols, back_exec->inputs[j])((void*)(((char*)((autograd_tensor_symbols)->data)) + (size_t )(autograd_tensor_symbols)->rsize * (size_t)(back_exec-> inputs[j]))))->symbol)); | |||
1658 | // Find whether in the wrt symbols, anything we need to init to zero, if there are, these need to be inputs here too. | |||
1659 | _ccv_nnc_symbolic_graph_add_init_zeros(sub_prep, (ccv_nnc_tensor_symbol_t*)ccv_array_get(sub_wrt_symbols, 0)((void*)(((char*)((sub_wrt_symbols)->data)) + (size_t)(sub_wrt_symbols )->rsize * (size_t)(0))), sub_wrt_symbols->rnum, graph, sub_graph, symbols); | |||
1660 | _ccv_nnc_symbolic_graph_add_tape_vars(sub_prep, root, graph, sub_graph, symbols); | |||
1661 | // input_size at this point, may be different from the back_exec->input_size, the reason is because we may added zeroing tensors as input tensors. | |||
1662 | const int input_size = symbols->rnum; | |||
1663 | for (j = 0; j < back_exec->output_size; j++) | |||
1664 | if (back_info->output_bitmasks[j >> 6] & ((uint64_t)1 << j)) | |||
1665 | ccv_array_push(symbols, &(((ccv_nnc_autograd_tensor_symbol_t*)ccv_array_get(autograd_tensor_symbols, back_exec->outputs[j])((void*)(((char*)((autograd_tensor_symbols)->data)) + (size_t )(autograd_tensor_symbols)->rsize * (size_t)(back_exec-> outputs[j]))))->symbol)); | |||
1666 | const int output_size = symbols->rnum - input_size; | |||
1667 | const int p_idx = sub_prep->graph->p_idx - 1; | |||
1668 | assert(back_exec->input_size == forw_exec->output_size)((void) sizeof ((back_exec->input_size == forw_exec->output_size ) ? 1 : 0), __extension__ ({ if (back_exec->input_size == forw_exec ->output_size) ; else __assert_fail ("back_exec->input_size == forw_exec->output_size" , "ccv_nnc_symbolic_graph_backward.c", 1668, __extension__ __PRETTY_FUNCTION__ ); })); | |||
1669 | k = 0; | |||
1670 | for (j = 0; j < back_exec->input_size; j++) | |||
1671 | if (back_info->input_bitmasks[j >> 6] & ((uint64_t)1 << j)) | |||
1672 | { | |||
1673 | const ccv_nnc_tensor_symbol_info_t* const info = tensor_symbol_info + forw_exec->outputs[j]; | |||
1674 | const int s_idx = *(int*)ccv_array_get(info->s_ref, p_idx)((void*)(((char*)((info->s_ref)->data)) + (size_t)(info ->s_ref)->rsize * (size_t)(p_idx))) - 1; | |||
1675 | assert(s_idx >= 0)((void) sizeof ((s_idx >= 0) ? 1 : 0), __extension__ ({ if (s_idx >= 0) ; else __assert_fail ("s_idx >= 0", "ccv_nnc_symbolic_graph_backward.c" , 1675, __extension__ __PRETTY_FUNCTION__); })); | |||
1676 | const ccv_nnc_autograd_tensor_symbol_t* const autograd_symbol = _ccv_nnc_autograd_tensor_symbol_from_tensor_version(sub_prep->autograd_tensor_symbols, sub_prep->autograd_tensor_versions + s_idx); | |||
1677 | ccv_nnc_tensor_symbol_hookup(graph, sub_graph, *(ccv_nnc_tensor_symbol_t*)ccv_array_get(symbols, k)((void*)(((char*)((symbols)->data)) + (size_t)(symbols)-> rsize * (size_t)(k))), autograd_symbol->symbol); | |||
1678 | ++k; | |||
1679 | } | |||
1680 | k = input_size; // Reset k, the symbol pass already set up by add_init_zeros. | |||
1681 | assert(back_exec->output_size == forw_exec->input_size)((void) sizeof ((back_exec->output_size == forw_exec->input_size ) ? 1 : 0), __extension__ ({ if (back_exec->output_size == forw_exec->input_size) ; else __assert_fail ("back_exec->output_size == forw_exec->input_size" , "ccv_nnc_symbolic_graph_backward.c", 1681, __extension__ __PRETTY_FUNCTION__ ); })); | |||
1682 | for (j = 0; j < back_exec->output_size; j++) | |||
1683 | if (back_info->output_bitmasks[j >> 6] & ((uint64_t)1 << j)) | |||
1684 | { | |||
1685 | const ccv_nnc_tensor_symbol_info_t* const info = tensor_symbol_info + forw_exec->inputs[j]; | |||
1686 | const int s_idx = *(int*)ccv_array_get(info->s_ref, p_idx)((void*)(((char*)((info->s_ref)->data)) + (size_t)(info ->s_ref)->rsize * (size_t)(p_idx))) - 1; | |||
1687 | assert(s_idx >= 0)((void) sizeof ((s_idx >= 0) ? 1 : 0), __extension__ ({ if (s_idx >= 0) ; else __assert_fail ("s_idx >= 0", "ccv_nnc_symbolic_graph_backward.c" , 1687, __extension__ __PRETTY_FUNCTION__); })); | |||
1688 | const ccv_nnc_autograd_tensor_symbol_t* const autograd_symbol = _ccv_nnc_autograd_tensor_symbol_from_tensor_version(sub_prep->autograd_tensor_symbols, sub_prep->autograd_tensor_versions + s_idx); | |||
1689 | ccv_nnc_tensor_symbol_hookup(graph, sub_graph, *(ccv_nnc_tensor_symbol_t*)ccv_array_get(symbols, k)((void*)(((char*)((symbols)->data)) + (size_t)(symbols)-> rsize * (size_t)(k))), autograd_symbol->symbol); | |||
1690 | ++k; | |||
1691 | } | |||
1692 | ccv_nnc_graph_exec_symbol_set_io(graph, back_exec->symbol, ccv_array_get(symbols, 0)((void*)(((char*)((symbols)->data)) + (size_t)(symbols)-> rsize * (size_t)(0))), input_size, ccv_array_get(symbols, input_size)((void*)(((char*)((symbols)->data)) + (size_t)(symbols)-> rsize * (size_t)(input_size))), output_size); | |||
1693 | } else if (forw_exec->flags & CCV_NNC_GRAPH_EXEC_CASE_OF) { | |||
1694 | ccv_array_clear(symbol_map); | |||
1695 | for (j = 0; j < back_exec->output_size; j++) | |||
1696 | if (back_info->output_bitmasks[j >> 6] & ((uint64_t)1 << j)) | |||
1697 | { | |||
1698 | ccv_nnc_tensor_symbol_map_t symbol = { | |||
1699 | .source = ((ccv_nnc_autograd_tensor_symbol_t*)ccv_array_get(autograd_tensor_symbols, back_exec->inputs[j])((void*)(((char*)((autograd_tensor_symbols)->data)) + (size_t )(autograd_tensor_symbols)->rsize * (size_t)(back_exec-> inputs[j]))))->symbol, | |||
1700 | .destination = ((ccv_nnc_autograd_tensor_symbol_t*)ccv_array_get(autograd_tensor_symbols, back_exec->outputs[j])((void*)(((char*)((autograd_tensor_symbols)->data)) + (size_t )(autograd_tensor_symbols)->rsize * (size_t)(back_exec-> outputs[j]))))->symbol, | |||
1701 | }; | |||
1702 | ccv_array_push(symbol_map, &symbol); | |||
1703 | } | |||
1704 | const int symbol_map_size = symbol_map->rnum; | |||
1705 | back_exec->symbol = ccv_nnc_symbolic_graph_case_of_new(graph, back_exec->cmd.cmd, 0, 0, ccv_array_get(symbol_map, 0)((void*)(((char*)((symbol_map)->data)) + (size_t)(symbol_map )->rsize * (size_t)(0))), symbol_map_size, forw_exec->name); | |||
1706 | ccv_nnc_symbolic_graph_set_case_of_expr(graph, back_exec->symbol, NOOP_GRAPH_CASE_OF_EXPR(ccv_nnc_graph_case_of_f)(1), 0); | |||
1707 | for (p = 0; p < forw_exec->graph_ref_size; p++) | |||
1708 | { | |||
1709 | const int graph_ref = CCV_NNC_GRAPH_REF(forw_exec)((forw_exec)->_heap_graph_ref ? (forw_exec)->_heap_graph_ref : (forw_exec)->_inline_graph_ref)[p] - 1; | |||
1710 | ccv_nnc_symbolic_graph_backward_prep_t* sub_prep = backward_prep->sub_preps + graph_ref; | |||
1711 | ccv_nnc_symbolic_graph_t* sub_graph = ccv_nnc_symbolic_graph_new(); | |||
1712 | sub_graph->pair = sub_prep->graph; | |||
1713 | if (!sub_wrt_symbols) | |||
1714 | sub_wrt_symbols = ccv_array_new(sizeof(ccv_nnc_tensor_symbol_t), 0, 0); | |||
1715 | // I am done, need to redo above for sub_prep, and it has to be successful now. | |||
1716 | if (!sub_f_symbols) | |||
1717 | sub_f_symbols = ccv_array_new(sizeof(ccv_nnc_tensor_symbol_t), 0, 0); | |||
1718 | _ccv_nnc_symbolic_graph_backward_prep_sub_f_wrt_symbols(forw_exec, sub_prep->graph, graph_ref, tensor_symbol_info, back_info->input_bitmasks, back_info->output_bitmasks, sub_f_symbols, sub_wrt_symbols); | |||
1719 | _ccv_nnc_symbolic_graph_backward_gen(sub_prep, (ccv_nnc_tensor_symbol_t*)ccv_array_get(sub_f_symbols, 0)((void*)(((char*)((sub_f_symbols)->data)) + (size_t)(sub_f_symbols )->rsize * (size_t)(0))), sub_f_symbols->rnum, (ccv_nnc_tensor_symbol_t*)ccv_array_get(sub_wrt_symbols, 0)((void*)(((char*)((sub_wrt_symbols)->data)) + (size_t)(sub_wrt_symbols )->rsize * (size_t)(0))), sub_wrt_symbols->rnum, sub_graph, root); | |||
1720 | ccv_array_clear(symbol_map); | |||
1721 | k = 0; | |||
1722 | for (j = 0; j < back_exec->output_size; j++) | |||
1723 | if (back_info->output_bitmasks[j >> 6] & ((uint64_t)1 << j)) | |||
1724 | { | |||
1725 | const int d = ((ccv_nnc_tensor_symbol_t*)ccv_array_get(sub_wrt_symbols, k)((void*)(((char*)((sub_wrt_symbols)->data)) + (size_t)(sub_wrt_symbols )->rsize * (size_t)(k))))->d; | |||
1726 | if (d >= 0) | |||
1727 | { | |||
1728 | const ccv_nnc_autograd_tensor_symbol_t* const autograd_symbol = _ccv_nnc_autograd_tensor_symbol_from_tensor_version(sub_prep->autograd_tensor_symbols, sub_prep->autograd_tensor_versions + d); | |||
1729 | ccv_nnc_tensor_symbol_map_t symbol = { | |||
1730 | .source = autograd_symbol->symbol, | |||
1731 | .destination = ((ccv_nnc_autograd_tensor_symbol_t*)ccv_array_get(autograd_tensor_symbols, back_exec->outputs[j])((void*)(((char*)((autograd_tensor_symbols)->data)) + (size_t )(autograd_tensor_symbols)->rsize * (size_t)(back_exec-> outputs[j]))))->symbol, | |||
1732 | }; | |||
1733 | ccv_array_push(symbol_map, &symbol); | |||
1734 | } else { | |||
1735 | // Create a new tensor in sub-graph and set it to be 0. | |||
1736 | const ccv_nnc_autograd_tensor_symbol_t* const autograd_symbol = (ccv_nnc_autograd_tensor_symbol_t*)ccv_array_get(autograd_tensor_symbols, back_exec->outputs[j])((void*)(((char*)((autograd_tensor_symbols)->data)) + (size_t )(autograd_tensor_symbols)->rsize * (size_t)(back_exec-> outputs[j]))); | |||
1737 | // autograd_symbol->d points to the corresponding forward tensor. | |||
1738 | ccv_nnc_tensor_symbol_t zero_symbol = ccv_nnc_tensor_symbol_new(sub_graph, tensor_symbol_info[autograd_symbol->d].info, 0); | |||
1739 | ccv_nnc_graph_exec_symbol_new(sub_graph, CMD_SET_FORWARD(0)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size ={.dim={1,1,1}},.blas={.a={0,}}}, 0), 0, 0, &zero_symbol, 1, 0); | |||
1740 | ccv_nnc_tensor_symbol_map_t symbol = { | |||
1741 | .source = zero_symbol, | |||
1742 | .destination = autograd_symbol->symbol, | |||
1743 | }; | |||
1744 | ccv_array_push(symbol_map, &symbol); | |||
1745 | } | |||
1746 | ++k; | |||
1747 | } | |||
1748 | ccv_nnc_graph_exec_symbol_autogen(sub_graph, 0, 0, CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS); | |||
1749 | const int symbol_map_size = symbol_map->rnum; | |||
1750 | ccv_nnc_symbolic_graph_set_case_of(graph, back_exec->symbol, sub_graph, p, ccv_array_get(symbol_map, 0)((void*)(((char*)((symbol_map)->data)) + (size_t)(symbol_map )->rsize * (size_t)(0))), symbol_map_size); | |||
1751 | // Hookup input only after this becomes a sub graph of the graph. | |||
1752 | k = 0; | |||
1753 | for (j = 0; j < back_exec->input_size; j++) | |||
1754 | if (back_info->input_bitmasks[j >> 6] & ((uint64_t)1 << j)) | |||
1755 | { | |||
1756 | const int d = ((ccv_nnc_tensor_symbol_t*)ccv_array_get(sub_f_symbols, k)((void*)(((char*)((sub_f_symbols)->data)) + (size_t)(sub_f_symbols )->rsize * (size_t)(k))))->d; | |||
1757 | assert(d >= 0)((void) sizeof ((d >= 0) ? 1 : 0), __extension__ ({ if (d >= 0) ; else __assert_fail ("d >= 0", "ccv_nnc_symbolic_graph_backward.c" , 1757, __extension__ __PRETTY_FUNCTION__); })); | |||
1758 | // No corresponding sub tensors allocated. Skip. | |||
1759 | if (!sub_prep->autograd_tensor_versions[d].ref_version || | |||
1760 | !sub_prep->autograd_tensor_versions[d].ref_version->rnum) | |||
1761 | continue; | |||
1762 | const ccv_nnc_autograd_tensor_symbol_t* const autograd_symbol = _ccv_nnc_autograd_tensor_symbol_from_tensor_version(sub_prep->autograd_tensor_symbols, sub_prep->autograd_tensor_versions + d); | |||
1763 | ccv_nnc_tensor_symbol_hookup(graph, sub_graph, ((ccv_nnc_autograd_tensor_symbol_t*)ccv_array_get(autograd_tensor_symbols, back_exec->inputs[j])((void*)(((char*)((autograd_tensor_symbols)->data)) + (size_t )(autograd_tensor_symbols)->rsize * (size_t)(back_exec-> inputs[j]))))->symbol, autograd_symbol->symbol); | |||
1764 | ++k; | |||
1765 | } | |||
1766 | // Need to make sure tape vars are hooked up. | |||
1767 | _ccv_nnc_symbolic_graph_add_tape_vars(sub_prep, root, graph, sub_graph, 0); | |||
1768 | } | |||
1769 | } else { | |||
1770 | ccv_array_clear(symbols); | |||
1771 | // Gradient inputs. | |||
1772 | for (j = 0; j < back_exec->input_size; j++) | |||
1773 | if (back_info->input_bitmasks[j >> 6] & ((uint64_t)1 << j)) | |||
1774 | ccv_array_push(symbols, &(((ccv_nnc_autograd_tensor_symbol_t*)ccv_array_get(autograd_tensor_symbols, back_exec->inputs[j])((void*)(((char*)((autograd_tensor_symbols)->data)) + (size_t )(autograd_tensor_symbols)->rsize * (size_t)(back_exec-> inputs[j]))))->symbol)); | |||
1775 | else | |||
1776 | ccv_array_push(symbols, &NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL }); | |||
1777 | // Inputs from forward function. | |||
1778 | for (j = 0; j < forw_exec->input_size; j++) | |||
1779 | if (!(back_info->input_bitmasks[(j + back_exec->input_size) >> 6] & ((uint64_t)1 << (j + back_exec->input_size)))) | |||
1780 | ccv_array_push(symbols, &NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL }); | |||
1781 | else { | |||
1782 | const ccv_nnc_tensor_symbol_t symbol = { | |||
1783 | .d = forw_exec->inputs[j], | |||
1784 | .graph = backward_prep->graph | |||
1785 | }; | |||
1786 | if (graph == backward_prep->graph) | |||
1787 | ccv_array_push(symbols, &symbol); | |||
1788 | else { // Otherwise, create a new symbol, and set its pair to the old symbol. | |||
1789 | const ccv_nnc_tensor_symbol_t new_symbol = ccv_nnc_tensor_symbol_new(graph, tensor_symbol_info[forw_exec->inputs[j]].info, tensor_symbol_info[forw_exec->inputs[j]].name); | |||
1790 | ccv_nnc_tensor_symbol_pair_with(graph, new_symbol, symbol); | |||
1791 | const int flags = ccv_nnc_tensor_symbol_flags(backward_prep->graph, symbol) | CCV_NNC_TENSOR_SYMBOL_TAPE_VAR; | |||
1792 | ccv_nnc_tensor_symbol_set_flags(graph, new_symbol, flags); | |||
1793 | ccv_nnc_tensor_symbol_set_flags(backward_prep->graph, symbol, flags); | |||
1794 | ccv_array_push(symbols, &new_symbol); | |||
1795 | } | |||
1796 | } | |||
1797 | // Outputs from forward function. | |||
1798 | for (j = 0; j < forw_exec->output_size; j++) | |||
1799 | if (!(back_info->input_bitmasks[(j + back_exec->input_size + forw_exec->input_size) >> 6] & ((uint64_t)1 << (j + back_exec->input_size + forw_exec->input_size)))) | |||
1800 | ccv_array_push(symbols, &NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL }); | |||
1801 | else { | |||
1802 | const ccv_nnc_tensor_symbol_t symbol = { | |||
1803 | .d = forw_exec->outputs[j], | |||
1804 | .graph = backward_prep->graph | |||
1805 | }; | |||
1806 | if (graph == backward_prep->graph) | |||
1807 | ccv_array_push(symbols, &symbol); | |||
1808 | else { // Otherwise, create a new symbol, and set its pair to the old symbol. | |||
1809 | const ccv_nnc_tensor_symbol_t new_symbol = ccv_nnc_tensor_symbol_new(graph, tensor_symbol_info[forw_exec->outputs[j]].info, tensor_symbol_info[forw_exec->outputs[j]].name); | |||
1810 | ccv_nnc_tensor_symbol_pair_with(graph, new_symbol, symbol); | |||
1811 | const int flags = ccv_nnc_tensor_symbol_flags(backward_prep->graph, symbol) | CCV_NNC_TENSOR_SYMBOL_TAPE_VAR; | |||
1812 | ccv_nnc_tensor_symbol_set_flags(graph, new_symbol, flags); | |||
1813 | ccv_nnc_tensor_symbol_set_flags(backward_prep->graph, symbol, flags); | |||
1814 | ccv_array_push(symbols, &new_symbol); | |||
1815 | } | |||
1816 | } | |||
1817 | for (j = 0; j < back_exec->output_size; j++) | |||
1818 | if (back_info->output_bitmasks[j >> 6] & ((uint64_t)1 << j)) | |||
1819 | ccv_array_push(symbols, &(((ccv_nnc_autograd_tensor_symbol_t*)ccv_array_get(autograd_tensor_symbols, back_exec->outputs[j])((void*)(((char*)((autograd_tensor_symbols)->data)) + (size_t )(autograd_tensor_symbols)->rsize * (size_t)(back_exec-> outputs[j]))))->symbol)); | |||
1820 | else | |||
1821 | ccv_array_push(symbols, &NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL }); | |||
1822 | back_exec->symbol = ccv_nnc_graph_exec_symbol_new(graph, back_exec->cmd, ccv_array_get(symbols, 0)((void*)(((char*)((symbols)->data)) + (size_t)(symbols)-> rsize * (size_t)(0))), back_exec->input_size + forw_exec->input_size + forw_exec->output_size, ccv_array_get(symbols, back_exec->input_size + forw_exec->input_size + forw_exec->output_size)((void*)(((char*)((symbols)->data)) + (size_t)(symbols)-> rsize * (size_t)(back_exec->input_size + forw_exec->input_size + forw_exec->output_size))), back_exec->output_size, 0); | |||
1823 | ccv_nnc_graph_exec_symbol_set_hint(graph, back_exec->symbol, exec_symbol_info[i].hint); | |||
1824 | ccv_nnc_graph_exec_symbol_pair_with(graph, back_exec->symbol, (ccv_nnc_graph_exec_symbol_t){ | |||
1825 | .d = i, | |||
1826 | .graph = backward_prep->graph, | |||
1827 | }); | |||
1828 | } | |||
1829 | } | |||
1830 | if (sub_f_symbols) | |||
1831 | ccv_array_free(sub_f_symbols); | |||
1832 | if (sub_wrt_symbols) | |||
1833 | ccv_array_free(sub_wrt_symbols); | |||
1834 | if (sub_execs) | |||
1835 | ccv_array_free(sub_execs); | |||
1836 | ccv_array_t* const sum_or_set_execs = backward_prep->sum_or_set_execs; | |||
1837 | for (i = 0; i < sum_or_set_execs->rnum; i++) | |||
1838 | { | |||
1839 | ccv_nnc_sum_or_set_graph_exec_symbol_t* sum_or_set_exec = (ccv_nnc_sum_or_set_graph_exec_symbol_t*)ccv_array_get(sum_or_set_execs, i)((void*)(((char*)((sum_or_set_execs)->data)) + (size_t)(sum_or_set_execs )->rsize * (size_t)(i))); | |||
1840 | // It is sum, set don't have inputs. | |||
1841 | if (sum_or_set_exec->input_size) | |||
1842 | { | |||
1843 | ccv_array_clear(symbols); | |||
1844 | // This is to sum. | |||
1845 | for (j = 0; j < sum_or_set_exec->input_size; j++) | |||
1846 | ccv_array_push(symbols, &(((ccv_nnc_autograd_tensor_symbol_t*)ccv_array_get(autograd_tensor_symbols, sum_or_set_exec->inputs[j])((void*)(((char*)((autograd_tensor_symbols)->data)) + (size_t )(autograd_tensor_symbols)->rsize * (size_t)(sum_or_set_exec ->inputs[j]))))->symbol)); | |||
1847 | ccv_nnc_cmd_t cmd = ccv_nnc_cmd(CCV_NNC_EWSUM_FORWARD, 0, CMD_GENERIC()((ccv_nnc_cmd_param_t){.size={.dim={1,1,1}}}), 0); | |||
1848 | sum_or_set_exec->symbol = ccv_nnc_graph_exec_symbol_new(graph, cmd, ccv_array_get(symbols, 0)((void*)(((char*)((symbols)->data)) + (size_t)(symbols)-> rsize * (size_t)(0))), sum_or_set_exec->input_size, &(((ccv_nnc_autograd_tensor_symbol_t*)ccv_array_get(autograd_tensor_symbols, sum_or_set_exec->output)((void*)(((char*)((autograd_tensor_symbols)->data)) + (size_t )(autograd_tensor_symbols)->rsize * (size_t)(sum_or_set_exec ->output))))->symbol), 1, 0); | |||
1849 | } else | |||
1850 | sum_or_set_exec->symbol = ccv_nnc_graph_exec_symbol_new(graph, CMD_SET_FORWARD(sum_or_set_exec->value)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size ={.dim={1,1,1}},.blas={.a={sum_or_set_exec->value,}}}, 0), 0, 0, &(((ccv_nnc_autograd_tensor_symbol_t*)ccv_array_get(autograd_tensor_symbols, sum_or_set_exec->output)((void*)(((char*)((autograd_tensor_symbols)->data)) + (size_t )(autograd_tensor_symbols)->rsize * (size_t)(sum_or_set_exec ->output))))->symbol), 1, 0); | |||
1851 | } | |||
1852 | ccv_array_free(symbol_map); | |||
1853 | ccv_array_free(symbols); | |||
1854 | for (i = 0; i < exec_symbol_info_size; i++) | |||
1855 | { | |||
1856 | // This is not going to be an interesting node. Skip. | |||
1857 | if ((backward_info[i].f_wrt & 0x3) != 0x3) | |||
1858 | continue; | |||
1859 | ccv_nnc_autograd_graph_exec_symbol_t* const back_exec = autograd_execs + i; | |||
1860 | // If on the same graph, we cannot decide whether it is before or after the forw_exec, enforcing it is after forw_exec. | |||
1861 | if (graph == backward_prep->graph) | |||
1862 | ccv_nnc_graph_exec_symbol_concat(graph, (ccv_nnc_graph_exec_symbol_t){ | |||
1863 | .d = i, | |||
1864 | .graph = graph | |||
1865 | }, back_exec->symbol); | |||
1866 | if (back_exec->outgoings) | |||
1867 | for (j = 0; j < back_exec->outgoings->rnum; j++) | |||
1868 | { | |||
1869 | int d = *(int*)ccv_array_get(back_exec->outgoings, j)((void*)(((char*)((back_exec->outgoings)->data)) + (size_t )(back_exec->outgoings)->rsize * (size_t)(j))); | |||
1870 | if (d < exec_symbol_info_size) | |||
1871 | ccv_nnc_graph_exec_symbol_concat(graph, back_exec->symbol, autograd_execs[d].symbol); | |||
1872 | else | |||
1873 | ccv_nnc_graph_exec_symbol_concat(graph, back_exec->symbol, ((ccv_nnc_sum_or_set_graph_exec_symbol_t*)ccv_array_get(sum_or_set_execs, d - exec_symbol_info_size)((void*)(((char*)((sum_or_set_execs)->data)) + (size_t)(sum_or_set_execs )->rsize * (size_t)(d - exec_symbol_info_size))))->symbol); | |||
1874 | } | |||
1875 | } | |||
1876 | for (i = 0; i < sum_or_set_execs->rnum; i++) | |||
1877 | { | |||
1878 | ccv_nnc_sum_or_set_graph_exec_symbol_t* exec = (ccv_nnc_sum_or_set_graph_exec_symbol_t*)ccv_array_get(sum_or_set_execs, i)((void*)(((char*)((sum_or_set_execs)->data)) + (size_t)(sum_or_set_execs )->rsize * (size_t)(i))); | |||
1879 | if (exec->outgoings) | |||
1880 | for (j = 0; j < exec->outgoings->rnum; j++) | |||
1881 | { | |||
1882 | int d = *(int*)ccv_array_get(exec->outgoings, j)((void*)(((char*)((exec->outgoings)->data)) + (size_t)( exec->outgoings)->rsize * (size_t)(j))); | |||
1883 | if (d < exec_symbol_info_size) | |||
1884 | ccv_nnc_graph_exec_symbol_concat(graph, exec->symbol, autograd_execs[d].symbol); | |||
1885 | else | |||
1886 | ccv_nnc_graph_exec_symbol_concat(graph, exec->symbol, ((ccv_nnc_sum_or_set_graph_exec_symbol_t*)ccv_array_get(sum_or_set_execs, d - exec_symbol_info_size)((void*)(((char*)((sum_or_set_execs)->data)) + (size_t)(sum_or_set_execs )->rsize * (size_t)(d - exec_symbol_info_size))))->symbol); | |||
1887 | } | |||
1888 | } | |||
1889 | // Now, everything is done, set the metadata on graph so that we can lookup later for backward symbols | |||
1890 | if (graph->backward.tensor_symbol_idx) | |||
1891 | graph->backward.tensor_symbol_idx = (int*)ccreallocrealloc(graph->backward.tensor_symbol_idx, sizeof(int) * (graph->tensor_symbol_info->rnum + tensor_symbol_info_size)); | |||
1892 | else | |||
1893 | graph->backward.tensor_symbol_idx = (int*)ccmallocmalloc(sizeof(int) * (graph->tensor_symbol_info->rnum + tensor_symbol_info_size)); | |||
1894 | graph->backward.tensor_symbol_size = tensor_symbol_info_size; | |||
1895 | graph->backward.exec_symbol_idx = graph->backward.tensor_symbol_idx + tensor_symbol_info_size; | |||
1896 | graph->backward.exec_symbol_size = graph->tensor_symbol_info->rnum; | |||
1897 | for (i = 0; i < tensor_symbol_info_size; i++) | |||
1898 | graph->backward.tensor_symbol_idx[i] = -1; | |||
1899 | for (i = 0; i < graph->backward.exec_symbol_size; i++) | |||
1900 | graph->backward.exec_symbol_idx[i] = -1; | |||
1901 | ccv_nnc_autograd_tensor_version_t* const autograd_tensor_versions = backward_prep->autograd_tensor_versions; | |||
1902 | // Assigning for wrt symbols. | |||
1903 | for (i = 0; i < wrt_symbol_size; i++) | |||
1904 | { | |||
1905 | const int d = wrt_symbols[i].d; | |||
1906 | if (d < 0) | |||
1907 | continue; | |||
1908 | assert(d < tensor_symbol_info_size)((void) sizeof ((d < tensor_symbol_info_size) ? 1 : 0), __extension__ ({ if (d < tensor_symbol_info_size) ; else __assert_fail ( "d < tensor_symbol_info_size", "ccv_nnc_symbolic_graph_backward.c" , 1908, __extension__ __PRETTY_FUNCTION__); })); | |||
1909 | const ccv_nnc_tensor_symbol_info_t* const forw_symbol = tensor_symbol_info + d; | |||
1910 | ccv_nnc_autograd_tensor_version_t* const tensor_ver = autograd_tensor_versions + ((!forw_symbol->alias_ref) ? d : forw_symbol->alias_ref - 1); | |||
1911 | assert(tensor_ver->ref_version)((void) sizeof ((tensor_ver->ref_version) ? 1 : 0), __extension__ ({ if (tensor_ver->ref_version) ; else __assert_fail ("tensor_ver->ref_version" , "ccv_nnc_symbolic_graph_backward.c", 1911, __extension__ __PRETTY_FUNCTION__ ); })); | |||
1912 | ccv_nnc_tensor_ref_t* const tensor_ref = (ccv_nnc_tensor_ref_t*)ccv_array_get(tensor_ver->ref_version, tensor_ver->c)((void*)(((char*)((tensor_ver->ref_version)->data)) + ( size_t)(tensor_ver->ref_version)->rsize * (size_t)(tensor_ver ->c))); | |||
1913 | ccv_nnc_autograd_tensor_symbol_t* const autograd_symbol = (ccv_nnc_autograd_tensor_symbol_t*)ccv_array_get(autograd_tensor_symbols, tensor_ref->d)((void*)(((char*)((autograd_tensor_symbols)->data)) + (size_t )(autograd_tensor_symbols)->rsize * (size_t)(tensor_ref-> d))); | |||
1914 | // If this wrt symbol is an alias, create extra alias for this. | |||
1915 | if (!forw_symbol->alias_ref) | |||
1916 | graph->backward.tensor_symbol_idx[d] = autograd_symbol->symbol.d; | |||
1917 | else // We create new alias, and this cannot be referenced from exec_symbol_idx because its size limited to previous tensor symbol size. | |||
1918 | graph->backward.tensor_symbol_idx[d] = ccv_nnc_tensor_symbol_alias_new(graph, autograd_symbol->symbol, forw_symbol->ofs, forw_symbol->stride, forw_symbol->info, 0).d; | |||
1919 | const int dd = autograd_symbol->symbol.d; | |||
1920 | const int x = tensor_ref->x; | |||
1921 | if (tensor_ref->exec_registry && tensor_ref->exec_registry->rnum) // Create no-op node. | |||
1922 | { | |||
1923 | ccv_nnc_graph_exec_symbol_t noop = ccv_nnc_graph_exec_symbol_new(graph, ccv_nnc_cmd(CCV_NNC_NOOP, 0, CMD_GENERIC()((ccv_nnc_cmd_param_t){.size={.dim={1,1,1}}}), 0), 0, 0, 0, 0, 0); | |||
1924 | if (x < exec_symbol_info_size) | |||
1925 | ccv_nnc_graph_exec_symbol_concat(graph, autograd_execs[x].symbol, noop); | |||
1926 | else | |||
1927 | ccv_nnc_graph_exec_symbol_concat(graph, ((ccv_nnc_sum_or_set_graph_exec_symbol_t*)ccv_array_get(sum_or_set_execs, x - exec_symbol_info_size)((void*)(((char*)((sum_or_set_execs)->data)) + (size_t)(sum_or_set_execs )->rsize * (size_t)(x - exec_symbol_info_size))))->symbol, noop); | |||
1928 | for (j = 0; j < tensor_ref->exec_registry->rnum; j++) | |||
1929 | { | |||
1930 | const int x = *(int*)ccv_array_get(tensor_ref->exec_registry, j)((void*)(((char*)((tensor_ref->exec_registry)->data)) + (size_t)(tensor_ref->exec_registry)->rsize * (size_t)( j))); | |||
1931 | assert(x >= 0)((void) sizeof ((x >= 0) ? 1 : 0), __extension__ ({ if (x >= 0) ; else __assert_fail ("x >= 0", "ccv_nnc_symbolic_graph_backward.c" , 1931, __extension__ __PRETTY_FUNCTION__); })); /* Otherwise, this is initialization tensor, which is impossible to be summed up by. */ | |||
1932 | assert(x < exec_symbol_info_size)((void) sizeof ((x < exec_symbol_info_size) ? 1 : 0), __extension__ ({ if (x < exec_symbol_info_size) ; else __assert_fail ("x < exec_symbol_info_size" , "ccv_nnc_symbolic_graph_backward.c", 1932, __extension__ __PRETTY_FUNCTION__ ); })); // exec_registry is only used by alias_registry, it simply cannot reference to a sum operation. | |||
1933 | ccv_nnc_graph_exec_symbol_concat(graph, autograd_execs[x].symbol, noop); | |||
1934 | } | |||
1935 | graph->backward.exec_symbol_idx[dd] = noop.d; | |||
1936 | } else { | |||
1937 | if (x < exec_symbol_info_size) | |||
1938 | graph->backward.exec_symbol_idx[dd] = autograd_execs[x].symbol.d; | |||
1939 | else | |||
1940 | graph->backward.exec_symbol_idx[dd] = ((ccv_nnc_sum_or_set_graph_exec_symbol_t*)ccv_array_get(sum_or_set_execs, x - exec_symbol_info_size)((void*)(((char*)((sum_or_set_execs)->data)) + (size_t)(sum_or_set_execs )->rsize * (size_t)(x - exec_symbol_info_size))))->symbol.d; | |||
1941 | } | |||
1942 | } | |||
1943 | // Assigning for f symbols. | |||
1944 | for (i = 0; i < f_symbol_size; i++) | |||
1945 | { | |||
1946 | const int d = f_symbols[i].d; | |||
1947 | assert(d >= 0)((void) sizeof ((d >= 0) ? 1 : 0), __extension__ ({ if (d >= 0) ; else __assert_fail ("d >= 0", "ccv_nnc_symbolic_graph_backward.c" , 1947, __extension__ __PRETTY_FUNCTION__); })); | |||
1948 | assert(d < tensor_symbol_info_size)((void) sizeof ((d < tensor_symbol_info_size) ? 1 : 0), __extension__ ({ if (d < tensor_symbol_info_size) ; else __assert_fail ( "d < tensor_symbol_info_size", "ccv_nnc_symbolic_graph_backward.c" , 1948, __extension__ __PRETTY_FUNCTION__); })); | |||
1949 | const ccv_nnc_autograd_tensor_version_t* const tensor_ver = autograd_tensor_versions + d; | |||
1950 | if (tensor_ver->ref_version) | |||
1951 | { | |||
1952 | // We don't use _ccv_nnc_autograd_tensor_symbol_from_tensor_version because that select the last version, but for us, we need the first version. | |||
1953 | const ccv_nnc_tensor_ref_t* const tensor_ref = (ccv_nnc_tensor_ref_t*)ccv_array_get(tensor_ver->ref_version, 0)((void*)(((char*)((tensor_ver->ref_version)->data)) + ( size_t)(tensor_ver->ref_version)->rsize * (size_t)(0))); | |||
1954 | const ccv_nnc_autograd_tensor_symbol_t* const autograd_symbol = (ccv_nnc_autograd_tensor_symbol_t*)ccv_array_get(autograd_tensor_symbols, tensor_ref->d)((void*)(((char*)((autograd_tensor_symbols)->data)) + (size_t )(autograd_tensor_symbols)->rsize * (size_t)(tensor_ref-> d))); | |||
1955 | graph->backward.tensor_symbol_idx[d] = autograd_symbol->symbol.d; | |||
1956 | // Cannot find relevant backward exec symbols for f, it could be many. | |||
1957 | } | |||
1958 | } | |||
1959 | } | |||
1960 | ||||
1961 | void ccv_nnc_symbolic_graph_backward(ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const f_symbols, const int f_symbol_size, const ccv_nnc_tensor_symbol_t* const wrt_symbols, const int wrt_symbol_size, const ccv_nnc_graph_exec_symbol_t* const sources, const int source_size, const ccv_nnc_graph_exec_symbol_t* const destinations, const int destination_size) | |||
1962 | { | |||
1963 | int i; | |||
1964 | // f symbols cannot be alias. | |||
1965 | for (i = 0; i < f_symbol_size; i++) | |||
1966 | if (f_symbols[i].d >= 0) | |||
1967 | { | |||
1968 | assert(f_symbols[i].graph == graph)((void) sizeof ((f_symbols[i].graph == graph) ? 1 : 0), __extension__ ({ if (f_symbols[i].graph == graph) ; else __assert_fail ("f_symbols[i].graph == graph" , "ccv_nnc_symbolic_graph_backward.c", 1968, __extension__ __PRETTY_FUNCTION__ ); })); // f symbol has to be in the current graph. | |||
1969 | assert(!((ccv_nnc_tensor_symbol_info_t*)ccv_array_get(graph->tensor_symbol_info, f_symbols[i].d))->alias_ref)((void) sizeof ((!((ccv_nnc_tensor_symbol_info_t*)((void*)((( char*)((graph->tensor_symbol_info)->data)) + (size_t)(graph ->tensor_symbol_info)->rsize * (size_t)(f_symbols[i].d) )))->alias_ref) ? 1 : 0), __extension__ ({ if (!((ccv_nnc_tensor_symbol_info_t *)((void*)(((char*)((graph->tensor_symbol_info)->data)) + (size_t)(graph->tensor_symbol_info)->rsize * (size_t )(f_symbols[i].d))))->alias_ref) ; else __assert_fail ("!((ccv_nnc_tensor_symbol_info_t*)ccv_array_get(graph->tensor_symbol_info, f_symbols[i].d))->alias_ref" , "ccv_nnc_symbolic_graph_backward.c", 1969, __extension__ __PRETTY_FUNCTION__ ); })); | |||
1970 | } | |||
1971 | for (i = 0; i < wrt_symbol_size; i++) | |||
1972 | if (wrt_symbols[i].d >= 0) | |||
1973 | { | |||
1974 | assert(wrt_symbols[i].graph == graph)((void) sizeof ((wrt_symbols[i].graph == graph) ? 1 : 0), __extension__ ({ if (wrt_symbols[i].graph == graph) ; else __assert_fail ( "wrt_symbols[i].graph == graph", "ccv_nnc_symbolic_graph_backward.c" , 1974, __extension__ __PRETTY_FUNCTION__); })); | |||
1975 | // This is not an alias, or what it refers to is not an alias. | |||
1976 | assert(!((ccv_nnc_tensor_symbol_info_t*)ccv_array_get(graph->tensor_symbol_info, wrt_symbols[i].d))->alias_ref || !((ccv_nnc_tensor_symbol_info_t*)ccv_array_get(graph->tensor_symbol_info, ((ccv_nnc_tensor_symbol_info_t*)ccv_array_get(graph->tensor_symbol_info, wrt_symbols[i].d))->alias_ref - 1))->alias_ref)((void) sizeof ((!((ccv_nnc_tensor_symbol_info_t*)((void*)((( char*)((graph->tensor_symbol_info)->data)) + (size_t)(graph ->tensor_symbol_info)->rsize * (size_t)(wrt_symbols[i]. d))))->alias_ref || !((ccv_nnc_tensor_symbol_info_t*)((void *)(((char*)((graph->tensor_symbol_info)->data)) + (size_t )(graph->tensor_symbol_info)->rsize * (size_t)(((ccv_nnc_tensor_symbol_info_t *)((void*)(((char*)((graph->tensor_symbol_info)->data)) + (size_t)(graph->tensor_symbol_info)->rsize * (size_t )(wrt_symbols[i].d))))->alias_ref - 1))))->alias_ref) ? 1 : 0), __extension__ ({ if (!((ccv_nnc_tensor_symbol_info_t *)((void*)(((char*)((graph->tensor_symbol_info)->data)) + (size_t)(graph->tensor_symbol_info)->rsize * (size_t )(wrt_symbols[i].d))))->alias_ref || !((ccv_nnc_tensor_symbol_info_t *)((void*)(((char*)((graph->tensor_symbol_info)->data)) + (size_t)(graph->tensor_symbol_info)->rsize * (size_t )(((ccv_nnc_tensor_symbol_info_t*)((void*)(((char*)((graph-> tensor_symbol_info)->data)) + (size_t)(graph->tensor_symbol_info )->rsize * (size_t)(wrt_symbols[i].d))))->alias_ref - 1 ))))->alias_ref) ; else __assert_fail ("!((ccv_nnc_tensor_symbol_info_t*)ccv_array_get(graph->tensor_symbol_info, wrt_symbols[i].d))->alias_ref || !((ccv_nnc_tensor_symbol_info_t*)ccv_array_get(graph->tensor_symbol_info, ((ccv_nnc_tensor_symbol_info_t*)ccv_array_get(graph->tensor_symbol_info, wrt_symbols[i].d))->alias_ref - 1))->alias_ref" , "ccv_nnc_symbolic_graph_backward.c", 1976, __extension__ __PRETTY_FUNCTION__ ); })); | |||
1977 | } | |||
1978 | const int exec_symbol_info_size = graph->exec_symbol_info->rnum; | |||
1979 | const int tensor_symbol_info_size = graph->tensor_symbol_info->rnum; | |||
1980 | assert(exec_symbol_info_size > 0)((void) sizeof ((exec_symbol_info_size > 0) ? 1 : 0), __extension__ ({ if (exec_symbol_info_size > 0) ; else __assert_fail ("exec_symbol_info_size > 0" , "ccv_nnc_symbolic_graph_backward.c", 1980, __extension__ __PRETTY_FUNCTION__ ); })); | |||
1981 | assert(tensor_symbol_info_size > 0)((void) sizeof ((tensor_symbol_info_size > 0) ? 1 : 0), __extension__ ({ if (tensor_symbol_info_size > 0) ; else __assert_fail ( "tensor_symbol_info_size > 0", "ccv_nnc_symbolic_graph_backward.c" , 1981, __extension__ __PRETTY_FUNCTION__); })); | |||
1982 | ccv_nnc_symbolic_graph_backward_prep_t backward_prep = _ccv_nnc_symbolic_graph_backward_prep(graph, sources, source_size, destinations, destination_size); | |||
1983 | _ccv_nnc_symbolic_graph_backward_prep_prune_ops(&backward_prep, f_symbols, f_symbol_size, wrt_symbols, wrt_symbol_size, sources, source_size, destinations, destination_size); | |||
1984 | _ccv_nnc_symbolic_graph_backward_prep_gen(&backward_prep, f_symbols, f_symbol_size, wrt_symbols, wrt_symbol_size, 0, sources, source_size, destinations, destination_size); | |||
1985 | _ccv_nnc_symbolic_graph_backward_gen(&backward_prep, f_symbols, f_symbol_size, wrt_symbols, wrt_symbol_size, graph, graph); | |||
1986 | _ccv_nnc_symbolic_graph_backward_prep_free(backward_prep); | |||
1987 | } | |||
1988 | ||||
1989 | ccv_nnc_tensor_symbol_t ccv_nnc_tensor_symbol_for_backward(const ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t symbol) | |||
1990 | { | |||
1991 | assert(symbol.d >= 0)((void) sizeof ((symbol.d >= 0) ? 1 : 0), __extension__ ({ if (symbol.d >= 0) ; else __assert_fail ("symbol.d >= 0" , "ccv_nnc_symbolic_graph_backward.c", 1991, __extension__ __PRETTY_FUNCTION__ ); })); | |||
1992 | assert(symbol.d < graph->backward.tensor_symbol_size)((void) sizeof ((symbol.d < graph->backward.tensor_symbol_size ) ? 1 : 0), __extension__ ({ if (symbol.d < graph->backward .tensor_symbol_size) ; else __assert_fail ("symbol.d < graph->backward.tensor_symbol_size" , "ccv_nnc_symbolic_graph_backward.c", 1992, __extension__ __PRETTY_FUNCTION__ ); })); | |||
1993 | if (graph->backward.tensor_symbol_idx[symbol.d] < 0) | |||
1994 | return NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL }; | |||
1995 | ccv_nnc_tensor_symbol_t tensor = { | |||
1996 | .d = graph->backward.tensor_symbol_idx[symbol.d], | |||
1997 | .graph = graph, | |||
1998 | }; | |||
1999 | return tensor; | |||
2000 | } | |||
2001 | ||||
2002 | ccv_nnc_graph_exec_symbol_t ccv_nnc_graph_exec_symbol_for_backward(const ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t symbol) | |||
2003 | { | |||
2004 | assert(symbol.d >= 0)((void) sizeof ((symbol.d >= 0) ? 1 : 0), __extension__ ({ if (symbol.d >= 0) ; else __assert_fail ("symbol.d >= 0" , "ccv_nnc_symbolic_graph_backward.c", 2004, __extension__ __PRETTY_FUNCTION__ ); })); | |||
2005 | assert(symbol.d < graph->tensor_symbol_info->rnum)((void) sizeof ((symbol.d < graph->tensor_symbol_info-> rnum) ? 1 : 0), __extension__ ({ if (symbol.d < graph-> tensor_symbol_info->rnum) ; else __assert_fail ("symbol.d < graph->tensor_symbol_info->rnum" , "ccv_nnc_symbolic_graph_backward.c", 2005, __extension__ __PRETTY_FUNCTION__ ); })); | |||
2006 | int dd = symbol.d; | |||
2007 | // Check if this is an alias. Use the original if it is. | |||
2008 | ccv_nnc_tensor_symbol_info_t* const symbol_info = (ccv_nnc_tensor_symbol_info_t*)ccv_array_get(graph->tensor_symbol_info, dd)((void*)(((char*)((graph->tensor_symbol_info)->data)) + (size_t)(graph->tensor_symbol_info)->rsize * (size_t)( dd))); | |||
2009 | if (symbol_info->alias_ref) | |||
2010 | dd = symbol_info->alias_ref - 1; | |||
2011 | assert(dd >= 0)((void) sizeof ((dd >= 0) ? 1 : 0), __extension__ ({ if (dd >= 0) ; else __assert_fail ("dd >= 0", "ccv_nnc_symbolic_graph_backward.c" , 2011, __extension__ __PRETTY_FUNCTION__); })); | |||
2012 | assert(dd < graph->backward.exec_symbol_size)((void) sizeof ((dd < graph->backward.exec_symbol_size) ? 1 : 0), __extension__ ({ if (dd < graph->backward.exec_symbol_size ) ; else __assert_fail ("dd < graph->backward.exec_symbol_size" , "ccv_nnc_symbolic_graph_backward.c", 2012, __extension__ __PRETTY_FUNCTION__ ); })); | |||
2013 | if (graph->backward.exec_symbol_idx[dd] < 0) | |||
2014 | return (ccv_nnc_graph_exec_symbol_t){ | |||
2015 | .graph = 0, | |||
2016 | .d = CCV_NNC_NO_GRAPH_EXEC_SYMBOL | |||
2017 | }; | |||
2018 | ccv_nnc_graph_exec_symbol_t exec = { | |||
2019 | .d = graph->backward.exec_symbol_idx[dd], | |||
2020 | .graph = graph | |||
2021 | }; | |||
2022 | return exec; | |||
2023 | } |